Add the rt linux 4.1.3-rt3 as base

Import the rt linux 4.1.3-rt3 as OPNFV kvm base.

It's from git://git.kernel.org/pub/scm/linux/kernel/git/rt/linux-rt-devel.git linux-4.1.y-rt and
the base is:

commit 0917f823c59692d751951bf5ea699a2d1e2f26a2
Author: Sebastian Andrzej Siewior <bigeasy@linutronix.de>
Date:   Sat Jul 25 12:13:34 2015 +0200

    Prepare v4.1.3-rt3

    Signed-off-by: Sebastian Andrzej Siewior <bigeasy@linutronix.de>

We lose all the git history this way and it's not good. We
should apply another opnfv project repo in future.

Change-Id: I87543d81c9df70d99c5001fbdf646b202c19f423
Signed-off-by: Yunhong Jiang <yunhong.jiang@intel.com>

287 files changed, 69563 insertions, 0 deletions

diff --git a/kernel/arch/tile/Kbuild b/kernel/arch/tile/Kbuild
new file mode 100644
index 000000000..a9b922716
--- /dev/null
+++ b/kernel/arch/tile/Kbuild
@@ -0,0 +1,3 @@
+
+obj-y += kernel/
+obj-y += mm/
diff --git a/kernel/arch/tile/Kconfig b/kernel/arch/tile/Kconfig
new file mode 100644
index 000000000..a07e31b50
--- /dev/null
+++ b/kernel/arch/tile/Kconfig
@@ -0,0 +1,473 @@
+# For a description of the syntax of this configuration file,
+# see Documentation/kbuild/kconfig-language.txt.
+
+config TILE
+	def_bool y
+	select HAVE_PERF_EVENTS
+	select USE_PMC if PERF_EVENTS
+	select HAVE_DMA_ATTRS
+	select HAVE_DMA_API_DEBUG
+	select HAVE_KVM if !TILEGX
+	select GENERIC_FIND_FIRST_BIT
+	select SYSCTL_EXCEPTION_TRACE
+	select CC_OPTIMIZE_FOR_SIZE
+	select HAVE_DEBUG_KMEMLEAK
+	select GENERIC_IRQ_PROBE
+	select GENERIC_PENDING_IRQ if SMP
+	select GENERIC_IRQ_SHOW
+	select HAVE_DEBUG_BUGVERBOSE
+	select VIRT_TO_BUS
+	select SYS_HYPERVISOR
+	select ARCH_HAS_DEBUG_STRICT_USER_COPY_CHECKS
+	select ARCH_HAVE_NMI_SAFE_CMPXCHG
+	select GENERIC_CLOCKEVENTS
+	select MODULES_USE_ELF_RELA
+	select HAVE_ARCH_TRACEHOOK
+	select HAVE_SYSCALL_TRACEPOINTS
+	select ARCH_HAS_ATOMIC64_DEC_IF_POSITIVE
+	select HAVE_DEBUG_STACKOVERFLOW
+	select ARCH_WANT_FRAME_POINTERS
+	select HAVE_CONTEXT_TRACKING
+
+# FIXME: investigate whether we need/want these options.
+#	select HAVE_IOREMAP_PROT
+#	select HAVE_OPTPROBES
+#	select HAVE_REGS_AND_STACK_ACCESS_API
+#	select HAVE_HW_BREAKPOINT
+#	select PERF_EVENTS
+#	select HAVE_USER_RETURN_NOTIFIER
+#	config NO_BOOTMEM
+#	config ARCH_SUPPORTS_DEBUG_PAGEALLOC
+#	config HUGETLB_PAGE_SIZE_VARIABLE
+
+config MMU
+	def_bool y
+
+config GENERIC_CSUM
+	def_bool y
+
+config HAVE_ARCH_ALLOC_REMAP
+	def_bool y
+
+config HAVE_SETUP_PER_CPU_AREA
+	def_bool y
+
+config NEED_PER_CPU_PAGE_FIRST_CHUNK
+	def_bool y
+
+config SYS_SUPPORTS_HUGETLBFS
+	def_bool y
+
+# Support for additional huge page sizes besides HPAGE_SIZE.
+# The software support is currently only present in the TILE-Gx
+# hypervisor. TILEPro in any case does not support page sizes
+# larger than the default HPAGE_SIZE.
+config HUGETLB_SUPER_PAGES
+	depends on HUGETLB_PAGE && TILEGX
+	def_bool y
+
+config GENERIC_TIME_VSYSCALL
+	def_bool y
+
+# Enable PMC if PERF_EVENTS, OPROFILE, or WATCHPOINTS are enabled.
+config USE_PMC
+	bool
+
+# FIXME: tilegx can implement a more efficient rwsem.
+config RWSEM_GENERIC_SPINLOCK
+	def_bool y
+
+# We only support gcc 4.4 and above, so this should work.
+config ARCH_SUPPORTS_OPTIMIZED_INLINING
+	def_bool y
+
+config ARCH_PHYS_ADDR_T_64BIT
+	def_bool y
+
+config ARCH_DMA_ADDR_T_64BIT
+	def_bool y
+
+config NEED_DMA_MAP_STATE
+	def_bool y
+
+config ARCH_HAS_DMA_SET_COHERENT_MASK
+	bool
+
+config LOCKDEP_SUPPORT
+	def_bool y
+
+config STACKTRACE_SUPPORT
+	def_bool y
+	select STACKTRACE
+
+# We use discontigmem for now; at some point we may want to switch
+# to sparsemem (Tilera bug 7996).
+config ARCH_DISCONTIGMEM_ENABLE
+	def_bool y
+
+config ARCH_DISCONTIGMEM_DEFAULT
+	def_bool y
+
+config TRACE_IRQFLAGS_SUPPORT
+	def_bool y
+
+config STRICT_DEVMEM
+	def_bool y
+
+# SMP is required for Tilera Linux.
+config SMP
+	def_bool y
+
+config HVC_TILE
+	depends on TTY
+	select HVC_DRIVER
+	select HVC_IRQ if TILEGX
+	def_bool y
+
+config TILEGX
+	bool "Building for TILE-Gx (64-bit) processor"
+	select SPARSE_IRQ
+	select GENERIC_IRQ_LEGACY_ALLOC_HWIRQ
+	select HAVE_FUNCTION_TRACER
+	select HAVE_FUNCTION_GRAPH_TRACER
+	select HAVE_DYNAMIC_FTRACE
+	select HAVE_FTRACE_MCOUNT_RECORD
+	select HAVE_KPROBES
+	select HAVE_KRETPROBES
+	select HAVE_ARCH_KGDB
+	select ARCH_SUPPORTS_ATOMIC_RMW
+
+config TILEPRO
+	def_bool !TILEGX
+
+config 64BIT
+	def_bool TILEGX
+
+config ARCH_DEFCONFIG
+	string
+	default "arch/tile/configs/tilepro_defconfig" if !TILEGX
+	default "arch/tile/configs/tilegx_defconfig" if TILEGX
+
+config PGTABLE_LEVELS
+	int
+	default 3 if 64BIT
+	default 2
+
+source "init/Kconfig"
+
+source "kernel/Kconfig.freezer"
+
+menu "Tilera-specific configuration"
+
+config NR_CPUS
+	int "Maximum number of tiles (2-255)"
+	range 2 255
+	depends on SMP
+	default "64"
+	---help---
+	  Building with 64 is the recommended value, but a slightly
+	  smaller kernel memory footprint results from using a smaller
+	  value on chips with fewer tiles.
+
+if TILEGX
+
+choice
+	prompt "Kernel page size"
+	default PAGE_SIZE_64KB
+	help
+	  This lets you select the page size of the kernel.  For best
+	  performance on memory-intensive applications, a page size of 64KB
+	  is recommended.  For workloads involving many small files, many
+	  connections, etc., it may be better to select 16KB, which uses
+	  memory more efficiently at some cost in TLB performance.
+
+	  Note that this option is TILE-Gx specific; currently
+	  TILEPro page size is set by rebuilding the hypervisor.
+
+config PAGE_SIZE_16KB
+	bool "16KB"
+
+config PAGE_SIZE_64KB
+	bool "64KB"
+
+endchoice
+
+endif
+
+source "kernel/Kconfig.hz"
+
+config KEXEC
+	bool "kexec system call"
+	---help---
+	  kexec is a system call that implements the ability to shutdown your
+	  current kernel, and to start another kernel.  It is like a reboot
+	  but it is independent of the system firmware.   It is used
+	  to implement the "mboot" Tilera booter.
+
+	  The name comes from the similarity to the exec system call.
+
+config COMPAT
+	bool "Support 32-bit TILE-Gx binaries in addition to 64-bit"
+	depends on TILEGX
+	select COMPAT_BINFMT_ELF
+	default y
+	---help---
+	  If enabled, the kernel will support running TILE-Gx binaries
+	  that were built with the -m32 option.
+
+config SYSVIPC_COMPAT
+	def_bool y
+	depends on COMPAT && SYSVIPC
+
+# We do not currently support disabling HIGHMEM on tilepro.
+config HIGHMEM
+	bool # "Support for more than 512 MB of RAM"
+	default !TILEGX
+	---help---
+	  Linux can use the full amount of RAM in the system by
+	  default.  However, the address space of TILE processors is
+	  only 4 Gigabytes large. That means that, if you have a large
+	  amount of physical memory, not all of it can be "permanently
+	  mapped" by the kernel. The physical memory that's not
+	  permanently mapped is called "high memory".
+
+	  If you are compiling a kernel which will never run on a
+	  machine with more than 512 MB total physical RAM, answer
+	  "false" here. This will result in the kernel mapping all of
+	  physical memory into the top 1 GB of virtual memory space.
+
+	  If unsure, say "true".
+
+config ZONE_DMA
+	def_bool y
+
+config IOMMU_HELPER
+	bool
+
+config NEED_SG_DMA_LENGTH
+	bool
+
+config SWIOTLB
+	bool
+	default TILEGX
+	select IOMMU_HELPER
+	select NEED_SG_DMA_LENGTH
+	select ARCH_HAS_DMA_SET_COHERENT_MASK
+
+# We do not currently support disabling NUMA.
+config NUMA
+	bool # "NUMA Memory Allocation and Scheduler Support"
+	depends on SMP && DISCONTIGMEM
+	default y
+	---help---
+	  NUMA memory allocation is required for TILE processors
+	  unless booting with memory striping enabled in the
+	  hypervisor, or with only a single memory controller.
+	  It is recommended that this option always be enabled.
+
+config NODES_SHIFT
+	int "Log base 2 of the max number of memory controllers"
+	default 2
+	depends on NEED_MULTIPLE_NODES
+	---help---
+	  By default, 2, i.e. 2^2 == 4 DDR2 controllers.
+	  In a system with more controllers, this value should be raised.
+
+choice
+	depends on !TILEGX
+	prompt "Memory split" if EXPERT
+	default VMSPLIT_3G
+	---help---
+	  Select the desired split between kernel and user memory.
+
+	  If the address range available to the kernel is less than the
+	  physical memory installed, the remaining memory will be available
+	  as "high memory". Accessing high memory is a little more costly
+	  than low memory, as it needs to be mapped into the kernel first.
+	  Note that increasing the kernel address space limits the range
+	  available to user programs, making the address space there
+	  tighter.  Selecting anything other than the default 3G/1G split
+	  will also likely make your kernel incompatible with binary-only
+	  kernel modules.
+
+	  If you are not absolutely sure what you are doing, leave this
+	  option alone!
+
+	config VMSPLIT_3_75G
+		bool "3.75G/0.25G user/kernel split (no kernel networking)"
+	config VMSPLIT_3_5G
+		bool "3.5G/0.5G user/kernel split"
+	config VMSPLIT_3G
+		bool "3G/1G user/kernel split"
+	config VMSPLIT_2_75G
+		bool "2.75G/1.25G user/kernel split (for full 1G low memory)"
+	config VMSPLIT_2_5G
+		bool "2.5G/1.5G user/kernel split"
+	config VMSPLIT_2_25G
+		bool "2.25G/1.75G user/kernel split"
+	config VMSPLIT_2G
+		bool "2G/2G user/kernel split"
+	config VMSPLIT_1G
+		bool "1G/3G user/kernel split"
+endchoice
+
+config PAGE_OFFSET
+	hex
+	depends on !64BIT
+	default 0xF0000000 if VMSPLIT_3_75G
+	default 0xE0000000 if VMSPLIT_3_5G
+	default 0xB0000000 if VMSPLIT_2_75G
+	default 0xA0000000 if VMSPLIT_2_5G
+	default 0x90000000 if VMSPLIT_2_25G
+	default 0x80000000 if VMSPLIT_2G
+	default 0x40000000 if VMSPLIT_1G
+	default 0xC0000000
+
+source "mm/Kconfig"
+
+source "kernel/Kconfig.preempt"
+
+config CMDLINE_BOOL
+	bool "Built-in kernel command line"
+	default n
+	---help---
+	  Allow for specifying boot arguments to the kernel at
+	  build time.  On some systems (e.g. embedded ones), it is
+	  necessary or convenient to provide some or all of the
+	  kernel boot arguments with the kernel itself (that is,
+	  to not rely on the boot loader to provide them.)
+
+	  To compile command line arguments into the kernel,
+	  set this option to 'Y', then fill in the
+	  the boot arguments in CONFIG_CMDLINE.
+
+	  Systems with fully functional boot loaders (e.g. mboot, or
+	  if booting over PCI) should leave this option set to 'N'.
+
+config CMDLINE
+	string "Built-in kernel command string"
+	depends on CMDLINE_BOOL
+	default ""
+	---help---
+	  Enter arguments here that should be compiled into the kernel
+	  image and used at boot time.  If the boot loader provides a
+	  command line at boot time, it is appended to this string to
+	  form the full kernel command line, when the system boots.
+
+	  However, you can use the CONFIG_CMDLINE_OVERRIDE option to
+	  change this behavior.
+
+	  In most cases, the command line (whether built-in or provided
+	  by the boot loader) should specify the device for the root
+	  file system.
+
+config CMDLINE_OVERRIDE
+	bool "Built-in command line overrides boot loader arguments"
+	default n
+	depends on CMDLINE_BOOL
+	---help---
+	  Set this option to 'Y' to have the kernel ignore the boot loader
+	  command line, and use ONLY the built-in command line.
+
+	  This is used to work around broken boot loaders.  This should
+	  be set to 'N' under normal conditions.
+
+config VMALLOC_RESERVE
+	hex
+	default 0x2000000
+
+config HARDWALL
+	bool "Hardwall support to allow access to user dynamic network"
+	default y
+
+config KERNEL_PL
+	int "Processor protection level for kernel"
+	range 1 2
+	default 2 if TILEGX
+	default 1 if !TILEGX
+	---help---
+	  Since MDE 4.2, the Tilera hypervisor runs the kernel
+	  at PL2 by default.  If running under an older hypervisor,
+	  or as a KVM guest, you must run at PL1.  (The current
+	  hypervisor may also be recompiled with "make HV_PL=2" to
+	  allow it to run a kernel at PL1, but clients running at PL1
+	  are not expected to be supported indefinitely.)
+
+	  If you're not sure, don't change the default.
+
+source "arch/tile/gxio/Kconfig"
+
+endmenu  # Tilera-specific configuration
+
+menu "Bus options"
+
+config PCI
+	bool "PCI support"
+	default y
+	select PCI_DOMAINS
+	select GENERIC_PCI_IOMAP
+	select TILE_GXIO_TRIO if TILEGX
+	select PCI_MSI if TILEGX
+	---help---
+	  Enable PCI root complex support, so PCIe endpoint devices can
+	  be attached to the Tile chip.  Many, but not all, PCI devices
+	  are supported under Tilera's root complex driver.
+
+config PCI_DOMAINS
+	bool
+
+config NO_IOMEM
+	def_bool !PCI
+
+config NO_IOPORT_MAP
+	def_bool !PCI
+
+config TILE_PCI_IO
+	bool "PCI I/O space support"
+	default n
+	depends on PCI
+	depends on TILEGX
+	---help---
+	  Enable PCI I/O space support on TILEGx. Since the PCI I/O space
+	  is used by few modern PCIe endpoint devices, its support is disabled
+	  by default to save the TRIO PIO Region resource for other purposes.
+
+source "drivers/pci/Kconfig"
+
+source "drivers/pci/pcie/Kconfig"
+
+config TILE_USB
+	tristate "Tilera USB host adapter support"
+	default y
+	depends on USB
+	depends on TILEGX
+	select TILE_GXIO_USB_HOST
+	---help---
+	  Provides USB host adapter support for the built-in EHCI and OHCI
+	  interfaces on TILE-Gx chips.
+
+source "drivers/pci/hotplug/Kconfig"
+
+endmenu
+
+menu "Executable file formats"
+
+source "fs/Kconfig.binfmt"
+
+endmenu
+
+source "net/Kconfig"
+
+source "drivers/Kconfig"
+
+source "fs/Kconfig"
+
+source "arch/tile/Kconfig.debug"
+
+source "security/Kconfig"
+
+source "crypto/Kconfig"
+
+source "lib/Kconfig"
+
+source "arch/tile/kvm/Kconfig"
diff --git a/kernel/arch/tile/Kconfig.debug b/kernel/arch/tile/Kconfig.debug
new file mode 100644
index 000000000..19734d3ab
--- /dev/null
+++ b/kernel/arch/tile/Kconfig.debug
@@ -0,0 +1,25 @@
+menu "Kernel hacking"
+
+source "lib/Kconfig.debug"
+
+config EARLY_PRINTK
+	bool "Early printk" if EXPERT && DEBUG_KERNEL
+	default y
+	help
+	  Write kernel log output directly via the hypervisor console.
+
+	  This is useful for kernel debugging when your machine crashes very
+	  early before the console code is initialized. For normal operation
+	  it is not recommended because it looks ugly and doesn't cooperate
+	  with klogd/syslogd. You should normally N here,
+	  unless you want to debug such a crash.
+
+config TILE_HVGLUE_TRACE
+	bool "Provide wrapper functions for hypervisor ABI calls"
+	default n
+	help
+	  Provide wrapper functions for the hypervisor ABI calls
+	  defined in arch/tile/kernel/hvglue.S.  This allows tracing
+	  mechanisms, etc., to have visibility into those calls.
+
+endmenu
diff --git a/kernel/arch/tile/Makefile b/kernel/arch/tile/Makefile
new file mode 100644
index 000000000..4dc380a51
--- /dev/null
+++ b/kernel/arch/tile/Makefile
@@ -0,0 +1,75 @@
+#
+# This file is subject to the terms and conditions of the GNU General Public
+# License.  See the file "COPYING" in the main directory of this archive
+# for more details.
+#
+# This file is included by the global makefile so that you can add your own
+# architecture-specific flags and dependencies. Remember to do have actions
+# for "archclean" and "archdep" for cleaning up and making dependencies for
+# this architecture
+
+# If building with TILERA_ROOT set (i.e. using the Tilera Multicore
+# Development Environment) we can set CROSS_COMPILE based on that.
+# If we're not cross-compiling, make sure we're on the right architecture.
+# Only bother to test for a few common targets, to avoid useless errors.
+ifeq ($(CROSS_COMPILE),)
+  ifdef TILERA_ROOT
+    CROSS_COMPILE := $(TILERA_ROOT)/bin/tile-
+  else
+    goals := $(if $(MAKECMDGOALS), $(MAKECMDGOALS), all)
+    ifneq ($(strip $(filter vmlinux modules all,$(goals))),)
+      HOST_ARCH := $(shell uname -m)
+      ifneq ($(HOST_ARCH),$(ARCH))
+$(error Set TILERA_ROOT or CROSS_COMPILE when building $(ARCH) on $(HOST_ARCH))
+      endif
+    endif
+  endif
+endif
+
+# The tile compiler may emit .eh_frame information for backtracing.
+# In kernel modules, this causes load failures due to unsupported relocations.
+KBUILD_CFLAGS   += -fno-asynchronous-unwind-tables
+
+LIBGCC_PATH     := \
+  $(shell $(CC) $(KBUILD_CFLAGS) $(KCFLAGS) -print-libgcc-file-name)
+
+# Provide the path to use for "make defconfig".
+# We default to the newer TILE-Gx architecture if only "tile" is given.
+ifeq ($(ARCH),tile)
+        KBUILD_DEFCONFIG := tilegx_defconfig
+else
+        KBUILD_DEFCONFIG := $(ARCH)_defconfig
+endif
+
+# Used as a file extension when useful, e.g. head_$(BITS).o
+# Not needed for (e.g.) "$(CC) -m32" since the compiler automatically
+# uses the right default anyway.
+export BITS
+ifeq ($(CONFIG_TILEGX),y)
+BITS := 64
+else
+BITS := 32
+endif
+
+head-y		:= arch/tile/kernel/head_$(BITS).o
+
+libs-y		+= arch/tile/lib/
+libs-y		+= $(LIBGCC_PATH)
+
+# See arch/tile/Kbuild for content of core part of the kernel
+core-y		+= arch/tile/
+
+core-$(CONFIG_TILE_GXIO) += arch/tile/gxio/
+
+ifdef TILERA_ROOT
+INSTALL_PATH ?= $(TILERA_ROOT)/tile/boot
+endif
+
+install:
+	install -D -m 755 vmlinux $(INSTALL_PATH)/vmlinux-$(KERNELRELEASE)
+	install -D -m 644 .config $(INSTALL_PATH)/config-$(KERNELRELEASE)
+	install -D -m 644 System.map $(INSTALL_PATH)/System.map-$(KERNELRELEASE)
+
+define archhelp
+	echo '  install         - install kernel into $(INSTALL_PATH)'
+endef
diff --git a/kernel/arch/tile/configs/tilegx_defconfig b/kernel/arch/tile/configs/tilegx_defconfig
new file mode 100644
index 000000000..37dc9364c
--- /dev/null
+++ b/kernel/arch/tile/configs/tilegx_defconfig
@@ -0,0 +1,418 @@
+CONFIG_TILEGX=y
+CONFIG_SYSVIPC=y
+CONFIG_POSIX_MQUEUE=y
+CONFIG_FHANDLE=y
+CONFIG_AUDIT=y
+CONFIG_NO_HZ=y
+CONFIG_BSD_PROCESS_ACCT=y
+CONFIG_BSD_PROCESS_ACCT_V3=y
+CONFIG_TASKSTATS=y
+CONFIG_TASK_DELAY_ACCT=y
+CONFIG_TASK_XACCT=y
+CONFIG_TASK_IO_ACCOUNTING=y
+CONFIG_LOG_BUF_SHIFT=19
+CONFIG_CGROUPS=y
+CONFIG_CGROUP_DEBUG=y
+CONFIG_CGROUP_DEVICE=y
+CONFIG_CPUSETS=y
+CONFIG_CGROUP_CPUACCT=y
+CONFIG_RESOURCE_COUNTERS=y
+CONFIG_CGROUP_SCHED=y
+CONFIG_RT_GROUP_SCHED=y
+CONFIG_BLK_CGROUP=y
+CONFIG_NAMESPACES=y
+CONFIG_RELAY=y
+CONFIG_BLK_DEV_INITRD=y
+CONFIG_RD_XZ=y
+CONFIG_SYSCTL_SYSCALL=y
+CONFIG_EMBEDDED=y
+# CONFIG_COMPAT_BRK is not set
+CONFIG_PROFILING=y
+CONFIG_KPROBES=y
+CONFIG_MODULES=y
+CONFIG_MODULE_FORCE_LOAD=y
+CONFIG_MODULE_UNLOAD=y
+CONFIG_BLK_DEV_INTEGRITY=y
+CONFIG_PARTITION_ADVANCED=y
+CONFIG_OSF_PARTITION=y
+CONFIG_AMIGA_PARTITION=y
+CONFIG_MAC_PARTITION=y
+CONFIG_BSD_DISKLABEL=y
+CONFIG_MINIX_SUBPARTITION=y
+CONFIG_SOLARIS_X86_PARTITION=y
+CONFIG_UNIXWARE_DISKLABEL=y
+CONFIG_SGI_PARTITION=y
+CONFIG_SUN_PARTITION=y
+CONFIG_KARMA_PARTITION=y
+CONFIG_CFQ_GROUP_IOSCHED=y
+CONFIG_NR_CPUS=100
+CONFIG_HZ_100=y
+# CONFIG_COMPACTION is not set
+CONFIG_PREEMPT_VOLUNTARY=y
+CONFIG_TILE_PCI_IO=y
+CONFIG_PCI_DEBUG=y
+# CONFIG_CORE_DUMP_DEFAULT_ELF_HEADERS is not set
+CONFIG_BINFMT_MISC=y
+CONFIG_NET=y
+CONFIG_PACKET=y
+CONFIG_UNIX=y
+CONFIG_XFRM_USER=y
+CONFIG_XFRM_SUB_POLICY=y
+CONFIG_XFRM_STATISTICS=y
+CONFIG_NET_KEY=m
+CONFIG_NET_KEY_MIGRATE=y
+CONFIG_INET=y
+CONFIG_IP_MULTICAST=y
+CONFIG_IP_ADVANCED_ROUTER=y
+CONFIG_IP_MULTIPLE_TABLES=y
+CONFIG_IP_ROUTE_MULTIPATH=y
+CONFIG_IP_ROUTE_VERBOSE=y
+CONFIG_NET_IPIP=m
+CONFIG_IP_MROUTE=y
+CONFIG_IP_PIMSM_V1=y
+CONFIG_IP_PIMSM_V2=y
+CONFIG_SYN_COOKIES=y
+CONFIG_INET_AH=m
+CONFIG_INET_ESP=m
+CONFIG_INET_IPCOMP=m
+CONFIG_INET_XFRM_MODE_TRANSPORT=m
+CONFIG_INET_XFRM_MODE_TUNNEL=m
+CONFIG_INET_XFRM_MODE_BEET=m
+CONFIG_INET_DIAG=m
+CONFIG_TCP_CONG_ADVANCED=y
+CONFIG_TCP_CONG_HSTCP=m
+CONFIG_TCP_CONG_HYBLA=m
+CONFIG_TCP_CONG_SCALABLE=m
+CONFIG_TCP_CONG_LP=m
+CONFIG_TCP_CONG_VENO=m
+CONFIG_TCP_CONG_YEAH=m
+CONFIG_TCP_CONG_ILLINOIS=m
+CONFIG_TCP_MD5SIG=y
+CONFIG_IPV6=y
+CONFIG_IPV6_PRIVACY=y
+CONFIG_IPV6_ROUTER_PREF=y
+CONFIG_IPV6_ROUTE_INFO=y
+CONFIG_IPV6_OPTIMISTIC_DAD=y
+CONFIG_INET6_AH=m
+CONFIG_INET6_ESP=m
+CONFIG_INET6_IPCOMP=m
+CONFIG_IPV6_MIP6=m
+CONFIG_INET6_XFRM_MODE_TRANSPORT=m
+CONFIG_INET6_XFRM_MODE_TUNNEL=m
+CONFIG_INET6_XFRM_MODE_BEET=m
+CONFIG_INET6_XFRM_MODE_ROUTEOPTIMIZATION=m
+CONFIG_IPV6_SIT=m
+CONFIG_IPV6_TUNNEL=m
+CONFIG_IPV6_MULTIPLE_TABLES=y
+CONFIG_IPV6_MROUTE=y
+CONFIG_IPV6_PIMSM_V2=y
+CONFIG_NETLABEL=y
+CONFIG_RDS=m
+CONFIG_RDS_TCP=m
+CONFIG_BRIDGE=m
+CONFIG_VLAN_8021Q=m
+CONFIG_VLAN_8021Q_GVRP=y
+CONFIG_PHONET=m
+CONFIG_NET_SCHED=y
+CONFIG_NET_SCH_CBQ=m
+CONFIG_NET_SCH_HTB=m
+CONFIG_NET_SCH_HFSC=m
+CONFIG_NET_SCH_PRIO=m
+CONFIG_NET_SCH_MULTIQ=m
+CONFIG_NET_SCH_RED=m
+CONFIG_NET_SCH_SFQ=m
+CONFIG_NET_SCH_TEQL=m
+CONFIG_NET_SCH_TBF=m
+CONFIG_NET_SCH_GRED=m
+CONFIG_NET_SCH_DSMARK=m
+CONFIG_NET_SCH_NETEM=m
+CONFIG_NET_SCH_DRR=m
+CONFIG_NET_SCH_INGRESS=m
+CONFIG_NET_CLS_BASIC=m
+CONFIG_NET_CLS_TCINDEX=m
+CONFIG_NET_CLS_ROUTE4=m
+CONFIG_NET_CLS_FW=m
+CONFIG_NET_CLS_U32=m
+CONFIG_CLS_U32_PERF=y
+CONFIG_CLS_U32_MARK=y
+CONFIG_NET_CLS_RSVP=m
+CONFIG_NET_CLS_RSVP6=m
+CONFIG_NET_CLS_FLOW=m
+CONFIG_NET_CLS_CGROUP=y
+CONFIG_NET_EMATCH=y
+CONFIG_NET_EMATCH_CMP=m
+CONFIG_NET_EMATCH_NBYTE=m
+CONFIG_NET_EMATCH_U32=m
+CONFIG_NET_EMATCH_META=m
+CONFIG_NET_EMATCH_TEXT=m
+CONFIG_NET_CLS_ACT=y
+CONFIG_NET_ACT_POLICE=m
+CONFIG_NET_ACT_GACT=m
+CONFIG_GACT_PROB=y
+CONFIG_NET_ACT_MIRRED=m
+CONFIG_NET_ACT_NAT=m
+CONFIG_NET_ACT_PEDIT=m
+CONFIG_NET_ACT_SIMP=m
+CONFIG_NET_ACT_SKBEDIT=m
+CONFIG_NET_CLS_IND=y
+CONFIG_DCB=y
+CONFIG_DNS_RESOLVER=y
+# CONFIG_WIRELESS is not set
+CONFIG_UEVENT_HELPER_PATH="/sbin/hotplug"
+CONFIG_DEVTMPFS=y
+CONFIG_DEVTMPFS_MOUNT=y
+# CONFIG_FIRMWARE_IN_KERNEL is not set
+CONFIG_CONNECTOR=y
+CONFIG_BLK_DEV_LOOP=y
+CONFIG_BLK_DEV_CRYPTOLOOP=m
+CONFIG_BLK_DEV_SX8=m
+CONFIG_BLK_DEV_RAM=y
+CONFIG_BLK_DEV_RAM_SIZE=16384
+CONFIG_ATA_OVER_ETH=m
+CONFIG_RAID_ATTRS=m
+CONFIG_BLK_DEV_SD=y
+CONFIG_SCSI_CONSTANTS=y
+CONFIG_SCSI_LOGGING=y
+CONFIG_SCSI_SAS_ATA=y
+CONFIG_ISCSI_TCP=m
+CONFIG_SCSI_MVSAS=y
+# CONFIG_SCSI_MVSAS_DEBUG is not set
+CONFIG_SCSI_MVSAS_TASKLET=y
+CONFIG_ATA=y
+CONFIG_SATA_AHCI=y
+CONFIG_SATA_SIL24=y
+# CONFIG_ATA_SFF is not set
+CONFIG_MD=y
+CONFIG_BLK_DEV_MD=y
+CONFIG_MD_LINEAR=m
+CONFIG_MD_RAID0=m
+CONFIG_MD_RAID1=m
+CONFIG_MD_RAID10=m
+CONFIG_MD_RAID456=m
+CONFIG_MD_FAULTY=m
+CONFIG_BLK_DEV_DM=m
+CONFIG_DM_DEBUG=y
+CONFIG_DM_CRYPT=m
+CONFIG_DM_SNAPSHOT=m
+CONFIG_DM_MIRROR=m
+CONFIG_DM_LOG_USERSPACE=m
+CONFIG_DM_ZERO=m
+CONFIG_DM_MULTIPATH=m
+CONFIG_DM_MULTIPATH_QL=m
+CONFIG_DM_MULTIPATH_ST=m
+CONFIG_DM_DELAY=m
+CONFIG_DM_UEVENT=y
+CONFIG_TARGET_CORE=m
+CONFIG_TCM_IBLOCK=m
+CONFIG_TCM_FILEIO=m
+CONFIG_TCM_PSCSI=m
+CONFIG_LOOPBACK_TARGET=m
+CONFIG_ISCSI_TARGET=m
+CONFIG_FUSION=y
+CONFIG_FUSION_SAS=y
+CONFIG_NETDEVICES=y
+CONFIG_BONDING=m
+CONFIG_DUMMY=m
+CONFIG_IFB=m
+CONFIG_MACVLAN=m
+CONFIG_MACVTAP=m
+CONFIG_NETCONSOLE=m
+CONFIG_NETCONSOLE_DYNAMIC=y
+CONFIG_TUN=y
+CONFIG_VETH=m
+CONFIG_NET_DSA_MV88E6060=y
+CONFIG_NET_DSA_MV88E6131=y
+CONFIG_NET_DSA_MV88E6123_61_65=y
+CONFIG_SKY2=y
+CONFIG_PTP_1588_CLOCK_TILEGX=y
+# CONFIG_WLAN is not set
+# CONFIG_INPUT_MOUSEDEV is not set
+# CONFIG_INPUT_KEYBOARD is not set
+# CONFIG_INPUT_MOUSE is not set
+# CONFIG_SERIO is not set
+# CONFIG_VT is not set
+# CONFIG_LEGACY_PTYS is not set
+CONFIG_SERIAL_TILEGX=y
+CONFIG_HW_RANDOM=y
+CONFIG_HW_RANDOM_TIMERIOMEM=m
+CONFIG_I2C=y
+CONFIG_I2C_CHARDEV=y
+# CONFIG_HWMON is not set
+CONFIG_WATCHDOG=y
+CONFIG_WATCHDOG_NOWAYOUT=y
+# CONFIG_VGA_ARB is not set
+CONFIG_DRM=m
+CONFIG_DRM_TDFX=m
+CONFIG_DRM_R128=m
+CONFIG_DRM_MGA=m
+CONFIG_DRM_VIA=m
+CONFIG_DRM_SAVAGE=m
+CONFIG_USB=y
+CONFIG_USB_EHCI_HCD=y
+CONFIG_USB_OHCI_HCD=y
+CONFIG_USB_STORAGE=y
+CONFIG_EDAC=y
+CONFIG_EDAC_MM_EDAC=y
+CONFIG_RTC_CLASS=y
+CONFIG_RTC_DRV_TILE=y
+CONFIG_EXT2_FS=y
+CONFIG_EXT2_FS_XATTR=y
+CONFIG_EXT2_FS_POSIX_ACL=y
+CONFIG_EXT2_FS_SECURITY=y
+CONFIG_EXT2_FS_XIP=y
+CONFIG_EXT3_FS=y
+CONFIG_EXT3_FS_POSIX_ACL=y
+CONFIG_EXT3_FS_SECURITY=y
+CONFIG_EXT4_FS=y
+CONFIG_EXT4_FS_POSIX_ACL=y
+CONFIG_EXT4_FS_SECURITY=y
+CONFIG_XFS_FS=y
+CONFIG_XFS_QUOTA=y
+CONFIG_XFS_POSIX_ACL=y
+CONFIG_GFS2_FS=m
+CONFIG_GFS2_FS_LOCKING_DLM=y
+CONFIG_BTRFS_FS=m
+CONFIG_BTRFS_FS_POSIX_ACL=y
+CONFIG_QUOTA=y
+CONFIG_QUOTA_NETLINK_INTERFACE=y
+# CONFIG_PRINT_QUOTA_WARNING is not set
+CONFIG_QFMT_V2=y
+CONFIG_AUTOFS4_FS=m
+CONFIG_FUSE_FS=y
+CONFIG_CUSE=m
+CONFIG_FSCACHE=m
+CONFIG_FSCACHE_STATS=y
+CONFIG_CACHEFILES=m
+CONFIG_ISO9660_FS=m
+CONFIG_JOLIET=y
+CONFIG_ZISOFS=y
+CONFIG_UDF_FS=m
+CONFIG_MSDOS_FS=m
+CONFIG_VFAT_FS=m
+CONFIG_FAT_DEFAULT_IOCHARSET="ascii"
+CONFIG_PROC_KCORE=y
+CONFIG_TMPFS=y
+CONFIG_TMPFS_POSIX_ACL=y
+CONFIG_HUGETLBFS=y
+CONFIG_ECRYPT_FS=m
+CONFIG_CRAMFS=m
+CONFIG_SQUASHFS=m
+CONFIG_NFS_FS=m
+CONFIG_NFS_V3_ACL=y
+CONFIG_NFS_V4=m
+CONFIG_NFS_V4_1=y
+CONFIG_NFS_FSCACHE=y
+CONFIG_NFSD=m
+CONFIG_NFSD_V3_ACL=y
+CONFIG_NFSD_V4=y
+CONFIG_CIFS=m
+CONFIG_CIFS_STATS=y
+CONFIG_CIFS_WEAK_PW_HASH=y
+CONFIG_CIFS_UPCALL=y
+CONFIG_CIFS_XATTR=y
+CONFIG_CIFS_POSIX=y
+CONFIG_CIFS_DFS_UPCALL=y
+CONFIG_CIFS_FSCACHE=y
+CONFIG_NLS_DEFAULT="utf8"
+CONFIG_NLS_CODEPAGE_437=y
+CONFIG_NLS_CODEPAGE_737=m
+CONFIG_NLS_CODEPAGE_775=m
+CONFIG_NLS_CODEPAGE_850=m
+CONFIG_NLS_CODEPAGE_852=m
+CONFIG_NLS_CODEPAGE_855=m
+CONFIG_NLS_CODEPAGE_857=m
+CONFIG_NLS_CODEPAGE_860=m
+CONFIG_NLS_CODEPAGE_861=m
+CONFIG_NLS_CODEPAGE_862=m
+CONFIG_NLS_CODEPAGE_863=m
+CONFIG_NLS_CODEPAGE_864=m
+CONFIG_NLS_CODEPAGE_865=m
+CONFIG_NLS_CODEPAGE_866=m
+CONFIG_NLS_CODEPAGE_869=m
+CONFIG_NLS_CODEPAGE_936=m
+CONFIG_NLS_CODEPAGE_950=m
+CONFIG_NLS_CODEPAGE_932=m
+CONFIG_NLS_CODEPAGE_949=m
+CONFIG_NLS_CODEPAGE_874=m
+CONFIG_NLS_ISO8859_8=m
+CONFIG_NLS_CODEPAGE_1250=m
+CONFIG_NLS_CODEPAGE_1251=m
+CONFIG_NLS_ASCII=y
+CONFIG_NLS_ISO8859_1=m
+CONFIG_NLS_ISO8859_2=m
+CONFIG_NLS_ISO8859_3=m
+CONFIG_NLS_ISO8859_4=m
+CONFIG_NLS_ISO8859_5=m
+CONFIG_NLS_ISO8859_6=m
+CONFIG_NLS_ISO8859_7=m
+CONFIG_NLS_ISO8859_9=m
+CONFIG_NLS_ISO8859_13=m
+CONFIG_NLS_ISO8859_14=m
+CONFIG_NLS_ISO8859_15=m
+CONFIG_NLS_KOI8_R=m
+CONFIG_NLS_KOI8_U=m
+CONFIG_NLS_UTF8=m
+CONFIG_DLM=m
+CONFIG_DLM_DEBUG=y
+CONFIG_DYNAMIC_DEBUG=y
+CONFIG_DEBUG_INFO=y
+CONFIG_DEBUG_INFO_REDUCED=y
+# CONFIG_ENABLE_WARN_DEPRECATED is not set
+CONFIG_STRIP_ASM_SYMS=y
+CONFIG_DEBUG_FS=y
+CONFIG_HEADERS_CHECK=y
+# CONFIG_FRAME_POINTER is not set
+CONFIG_DEBUG_FORCE_WEAK_PER_CPU=y
+CONFIG_DEBUG_VM=y
+CONFIG_DEBUG_MEMORY_INIT=y
+CONFIG_DEBUG_STACKOVERFLOW=y
+CONFIG_LOCKUP_DETECTOR=y
+CONFIG_SCHEDSTATS=y
+CONFIG_TIMER_STATS=y
+CONFIG_DEBUG_LIST=y
+CONFIG_DEBUG_CREDENTIALS=y
+CONFIG_RCU_CPU_STALL_TIMEOUT=60
+CONFIG_ASYNC_RAID6_TEST=m
+CONFIG_KGDB=y
+CONFIG_KEYS_DEBUG_PROC_KEYS=y
+CONFIG_SECURITY=y
+CONFIG_SECURITYFS=y
+CONFIG_SECURITY_NETWORK=y
+CONFIG_SECURITY_NETWORK_XFRM=y
+CONFIG_SECURITY_SELINUX=y
+CONFIG_SECURITY_SELINUX_BOOTPARAM=y
+CONFIG_SECURITY_SELINUX_DISABLE=y
+CONFIG_CRYPTO_PCRYPT=m
+CONFIG_CRYPTO_CRYPTD=m
+CONFIG_CRYPTO_TEST=m
+CONFIG_CRYPTO_CCM=m
+CONFIG_CRYPTO_GCM=m
+CONFIG_CRYPTO_CTS=m
+CONFIG_CRYPTO_LRW=m
+CONFIG_CRYPTO_PCBC=m
+CONFIG_CRYPTO_XTS=m
+CONFIG_CRYPTO_HMAC=y
+CONFIG_CRYPTO_XCBC=m
+CONFIG_CRYPTO_VMAC=m
+CONFIG_CRYPTO_MICHAEL_MIC=m
+CONFIG_CRYPTO_RMD128=m
+CONFIG_CRYPTO_RMD160=m
+CONFIG_CRYPTO_RMD256=m
+CONFIG_CRYPTO_RMD320=m
+CONFIG_CRYPTO_SHA1=y
+CONFIG_CRYPTO_SHA512=m
+CONFIG_CRYPTO_TGR192=m
+CONFIG_CRYPTO_WP512=m
+CONFIG_CRYPTO_ANUBIS=m
+CONFIG_CRYPTO_BLOWFISH=m
+CONFIG_CRYPTO_CAMELLIA=m
+CONFIG_CRYPTO_CAST5=m
+CONFIG_CRYPTO_CAST6=m
+CONFIG_CRYPTO_FCRYPT=m
+CONFIG_CRYPTO_KHAZAD=m
+CONFIG_CRYPTO_SEED=m
+CONFIG_CRYPTO_SERPENT=m
+CONFIG_CRYPTO_TEA=m
+CONFIG_CRYPTO_TWOFISH=m
+CONFIG_CRYPTO_ZLIB=m
+CONFIG_CRYPTO_LZO=m
diff --git a/kernel/arch/tile/configs/tilepro_defconfig b/kernel/arch/tile/configs/tilepro_defconfig
new file mode 100644
index 000000000..76a2781de
--- /dev/null
+++ b/kernel/arch/tile/configs/tilepro_defconfig
@@ -0,0 +1,532 @@
+CONFIG_SYSVIPC=y
+CONFIG_POSIX_MQUEUE=y
+CONFIG_AUDIT=y
+CONFIG_NO_HZ=y
+CONFIG_HIGH_RES_TIMERS=y
+CONFIG_BSD_PROCESS_ACCT=y
+CONFIG_BSD_PROCESS_ACCT_V3=y
+CONFIG_TASKSTATS=y
+CONFIG_TASK_DELAY_ACCT=y
+CONFIG_TASK_XACCT=y
+CONFIG_TASK_IO_ACCOUNTING=y
+CONFIG_LOG_BUF_SHIFT=19
+CONFIG_CGROUPS=y
+CONFIG_CGROUP_DEBUG=y
+CONFIG_CGROUP_DEVICE=y
+CONFIG_CPUSETS=y
+CONFIG_CGROUP_CPUACCT=y
+CONFIG_RESOURCE_COUNTERS=y
+CONFIG_CGROUP_SCHED=y
+CONFIG_RT_GROUP_SCHED=y
+CONFIG_BLK_CGROUP=y
+CONFIG_NAMESPACES=y
+CONFIG_RELAY=y
+CONFIG_BLK_DEV_INITRD=y
+CONFIG_RD_XZ=y
+CONFIG_SYSCTL_SYSCALL=y
+CONFIG_EMBEDDED=y
+# CONFIG_COMPAT_BRK is not set
+CONFIG_PROFILING=y
+CONFIG_MODULES=y
+CONFIG_MODULE_FORCE_LOAD=y
+CONFIG_MODULE_UNLOAD=y
+CONFIG_BLK_DEV_INTEGRITY=y
+CONFIG_PARTITION_ADVANCED=y
+CONFIG_OSF_PARTITION=y
+CONFIG_AMIGA_PARTITION=y
+CONFIG_MAC_PARTITION=y
+CONFIG_BSD_DISKLABEL=y
+CONFIG_MINIX_SUBPARTITION=y
+CONFIG_SOLARIS_X86_PARTITION=y
+CONFIG_UNIXWARE_DISKLABEL=y
+CONFIG_SGI_PARTITION=y
+CONFIG_SUN_PARTITION=y
+CONFIG_KARMA_PARTITION=y
+CONFIG_CFQ_GROUP_IOSCHED=y
+CONFIG_HZ_100=y
+# CONFIG_COMPACTION is not set
+CONFIG_PREEMPT_VOLUNTARY=y
+CONFIG_PCI_DEBUG=y
+# CONFIG_CORE_DUMP_DEFAULT_ELF_HEADERS is not set
+CONFIG_BINFMT_MISC=y
+CONFIG_NET=y
+CONFIG_PACKET=y
+CONFIG_UNIX=y
+CONFIG_XFRM_USER=y
+CONFIG_XFRM_SUB_POLICY=y
+CONFIG_XFRM_STATISTICS=y
+CONFIG_NET_KEY=m
+CONFIG_NET_KEY_MIGRATE=y
+CONFIG_INET=y
+CONFIG_IP_MULTICAST=y
+CONFIG_IP_ADVANCED_ROUTER=y
+CONFIG_IP_MULTIPLE_TABLES=y
+CONFIG_IP_ROUTE_MULTIPATH=y
+CONFIG_IP_ROUTE_VERBOSE=y
+CONFIG_NET_IPIP=m
+CONFIG_IP_MROUTE=y
+CONFIG_IP_PIMSM_V1=y
+CONFIG_IP_PIMSM_V2=y
+CONFIG_SYN_COOKIES=y
+CONFIG_INET_AH=m
+CONFIG_INET_ESP=m
+CONFIG_INET_IPCOMP=m
+CONFIG_INET_XFRM_MODE_TRANSPORT=m
+CONFIG_INET_XFRM_MODE_TUNNEL=m
+CONFIG_INET_XFRM_MODE_BEET=m
+CONFIG_INET_DIAG=m
+CONFIG_TCP_CONG_ADVANCED=y
+CONFIG_TCP_CONG_HSTCP=m
+CONFIG_TCP_CONG_HYBLA=m
+CONFIG_TCP_CONG_SCALABLE=m
+CONFIG_TCP_CONG_LP=m
+CONFIG_TCP_CONG_VENO=m
+CONFIG_TCP_CONG_YEAH=m
+CONFIG_TCP_CONG_ILLINOIS=m
+CONFIG_TCP_MD5SIG=y
+CONFIG_IPV6=y
+CONFIG_IPV6_PRIVACY=y
+CONFIG_IPV6_ROUTER_PREF=y
+CONFIG_IPV6_ROUTE_INFO=y
+CONFIG_IPV6_OPTIMISTIC_DAD=y
+CONFIG_INET6_AH=m
+CONFIG_INET6_ESP=m
+CONFIG_INET6_IPCOMP=m
+CONFIG_IPV6_MIP6=m
+CONFIG_INET6_XFRM_MODE_TRANSPORT=m
+CONFIG_INET6_XFRM_MODE_TUNNEL=m
+CONFIG_INET6_XFRM_MODE_BEET=m
+CONFIG_INET6_XFRM_MODE_ROUTEOPTIMIZATION=m
+CONFIG_IPV6_SIT=m
+CONFIG_IPV6_TUNNEL=m
+CONFIG_IPV6_MULTIPLE_TABLES=y
+CONFIG_IPV6_MROUTE=y
+CONFIG_IPV6_PIMSM_V2=y
+CONFIG_NETLABEL=y
+CONFIG_NETFILTER=y
+CONFIG_NF_CONNTRACK=m
+CONFIG_NF_CONNTRACK_SECMARK=y
+CONFIG_NF_CONNTRACK_ZONES=y
+CONFIG_NF_CONNTRACK_EVENTS=y
+CONFIG_NF_CT_PROTO_DCCP=m
+CONFIG_NF_CT_PROTO_UDPLITE=m
+CONFIG_NF_CONNTRACK_AMANDA=m
+CONFIG_NF_CONNTRACK_FTP=m
+CONFIG_NF_CONNTRACK_H323=m
+CONFIG_NF_CONNTRACK_IRC=m
+CONFIG_NF_CONNTRACK_NETBIOS_NS=m
+CONFIG_NF_CONNTRACK_PPTP=m
+CONFIG_NF_CONNTRACK_SANE=m
+CONFIG_NF_CONNTRACK_SIP=m
+CONFIG_NF_CONNTRACK_TFTP=m
+CONFIG_NETFILTER_XT_TARGET_CLASSIFY=m
+CONFIG_NETFILTER_XT_TARGET_CONNMARK=m
+CONFIG_NETFILTER_XT_TARGET_CONNSECMARK=m
+CONFIG_NETFILTER_XT_TARGET_DSCP=m
+CONFIG_NETFILTER_XT_TARGET_IDLETIMER=m
+CONFIG_NETFILTER_XT_TARGET_MARK=m
+CONFIG_NETFILTER_XT_TARGET_NFLOG=m
+CONFIG_NETFILTER_XT_TARGET_NFQUEUE=m
+CONFIG_NETFILTER_XT_TARGET_NOTRACK=m
+CONFIG_NETFILTER_XT_TARGET_TEE=m
+CONFIG_NETFILTER_XT_TARGET_TPROXY=m
+CONFIG_NETFILTER_XT_TARGET_TRACE=m
+CONFIG_NETFILTER_XT_TARGET_SECMARK=m
+CONFIG_NETFILTER_XT_TARGET_TCPMSS=m
+CONFIG_NETFILTER_XT_TARGET_TCPOPTSTRIP=m
+CONFIG_NETFILTER_XT_MATCH_CLUSTER=m
+CONFIG_NETFILTER_XT_MATCH_COMMENT=m
+CONFIG_NETFILTER_XT_MATCH_CONNBYTES=m
+CONFIG_NETFILTER_XT_MATCH_CONNLIMIT=m
+CONFIG_NETFILTER_XT_MATCH_CONNMARK=m
+CONFIG_NETFILTER_XT_MATCH_CONNTRACK=m
+CONFIG_NETFILTER_XT_MATCH_DCCP=m
+CONFIG_NETFILTER_XT_MATCH_DSCP=m
+CONFIG_NETFILTER_XT_MATCH_ESP=m
+CONFIG_NETFILTER_XT_MATCH_HASHLIMIT=m
+CONFIG_NETFILTER_XT_MATCH_HELPER=m
+CONFIG_NETFILTER_XT_MATCH_IPRANGE=m
+CONFIG_NETFILTER_XT_MATCH_IPVS=m
+CONFIG_NETFILTER_XT_MATCH_LENGTH=m
+CONFIG_NETFILTER_XT_MATCH_LIMIT=m
+CONFIG_NETFILTER_XT_MATCH_MAC=m
+CONFIG_NETFILTER_XT_MATCH_MARK=m
+CONFIG_NETFILTER_XT_MATCH_MULTIPORT=m
+CONFIG_NETFILTER_XT_MATCH_OSF=m
+CONFIG_NETFILTER_XT_MATCH_OWNER=m
+CONFIG_NETFILTER_XT_MATCH_POLICY=m
+CONFIG_NETFILTER_XT_MATCH_PHYSDEV=m
+CONFIG_NETFILTER_XT_MATCH_PKTTYPE=m
+CONFIG_NETFILTER_XT_MATCH_QUOTA=m
+CONFIG_NETFILTER_XT_MATCH_RATEEST=m
+CONFIG_NETFILTER_XT_MATCH_REALM=m
+CONFIG_NETFILTER_XT_MATCH_RECENT=m
+CONFIG_NETFILTER_XT_MATCH_SOCKET=m
+CONFIG_NETFILTER_XT_MATCH_STATE=m
+CONFIG_NETFILTER_XT_MATCH_STATISTIC=m
+CONFIG_NETFILTER_XT_MATCH_STRING=m
+CONFIG_NETFILTER_XT_MATCH_TCPMSS=m
+CONFIG_NETFILTER_XT_MATCH_TIME=m
+CONFIG_NETFILTER_XT_MATCH_U32=m
+CONFIG_IP_VS=m
+CONFIG_IP_VS_IPV6=y
+CONFIG_IP_VS_PROTO_TCP=y
+CONFIG_IP_VS_PROTO_UDP=y
+CONFIG_IP_VS_PROTO_ESP=y
+CONFIG_IP_VS_PROTO_AH=y
+CONFIG_IP_VS_PROTO_SCTP=y
+CONFIG_IP_VS_RR=m
+CONFIG_IP_VS_WRR=m
+CONFIG_IP_VS_LC=m
+CONFIG_IP_VS_WLC=m
+CONFIG_IP_VS_LBLC=m
+CONFIG_IP_VS_LBLCR=m
+CONFIG_IP_VS_SED=m
+CONFIG_IP_VS_NQ=m
+CONFIG_NF_CONNTRACK_IPV4=m
+# CONFIG_NF_CONNTRACK_PROC_COMPAT is not set
+CONFIG_IP_NF_IPTABLES=y
+CONFIG_IP_NF_MATCH_AH=m
+CONFIG_IP_NF_MATCH_ECN=m
+CONFIG_IP_NF_MATCH_TTL=m
+CONFIG_IP_NF_FILTER=y
+CONFIG_IP_NF_TARGET_REJECT=y
+CONFIG_IP_NF_TARGET_ULOG=m
+CONFIG_IP_NF_MANGLE=m
+CONFIG_IP_NF_TARGET_ECN=m
+CONFIG_IP_NF_TARGET_TTL=m
+CONFIG_IP_NF_RAW=m
+CONFIG_IP_NF_SECURITY=m
+CONFIG_IP_NF_ARPTABLES=m
+CONFIG_IP_NF_ARPFILTER=m
+CONFIG_IP_NF_ARP_MANGLE=m
+CONFIG_NF_CONNTRACK_IPV6=m
+CONFIG_IP6_NF_MATCH_AH=m
+CONFIG_IP6_NF_MATCH_EUI64=m
+CONFIG_IP6_NF_MATCH_FRAG=m
+CONFIG_IP6_NF_MATCH_OPTS=m
+CONFIG_IP6_NF_MATCH_HL=m
+CONFIG_IP6_NF_MATCH_IPV6HEADER=m
+CONFIG_IP6_NF_MATCH_MH=m
+CONFIG_IP6_NF_MATCH_RT=m
+CONFIG_IP6_NF_TARGET_HL=m
+CONFIG_IP6_NF_FILTER=m
+CONFIG_IP6_NF_TARGET_REJECT=m
+CONFIG_IP6_NF_MANGLE=m
+CONFIG_IP6_NF_RAW=m
+CONFIG_IP6_NF_SECURITY=m
+CONFIG_BRIDGE_NF_EBTABLES=m
+CONFIG_BRIDGE_EBT_BROUTE=m
+CONFIG_BRIDGE_EBT_T_FILTER=m
+CONFIG_BRIDGE_EBT_T_NAT=m
+CONFIG_BRIDGE_EBT_802_3=m
+CONFIG_BRIDGE_EBT_AMONG=m
+CONFIG_BRIDGE_EBT_ARP=m
+CONFIG_BRIDGE_EBT_IP=m
+CONFIG_BRIDGE_EBT_IP6=m
+CONFIG_BRIDGE_EBT_LIMIT=m
+CONFIG_BRIDGE_EBT_MARK=m
+CONFIG_BRIDGE_EBT_PKTTYPE=m
+CONFIG_BRIDGE_EBT_STP=m
+CONFIG_BRIDGE_EBT_VLAN=m
+CONFIG_BRIDGE_EBT_ARPREPLY=m
+CONFIG_BRIDGE_EBT_DNAT=m
+CONFIG_BRIDGE_EBT_MARK_T=m
+CONFIG_BRIDGE_EBT_REDIRECT=m
+CONFIG_BRIDGE_EBT_SNAT=m
+CONFIG_BRIDGE_EBT_LOG=m
+CONFIG_BRIDGE_EBT_ULOG=m
+CONFIG_BRIDGE_EBT_NFLOG=m
+CONFIG_RDS=m
+CONFIG_RDS_TCP=m
+CONFIG_BRIDGE=m
+CONFIG_VLAN_8021Q=m
+CONFIG_VLAN_8021Q_GVRP=y
+CONFIG_PHONET=m
+CONFIG_NET_SCHED=y
+CONFIG_NET_SCH_CBQ=m
+CONFIG_NET_SCH_HTB=m
+CONFIG_NET_SCH_HFSC=m
+CONFIG_NET_SCH_PRIO=m
+CONFIG_NET_SCH_MULTIQ=m
+CONFIG_NET_SCH_RED=m
+CONFIG_NET_SCH_SFQ=m
+CONFIG_NET_SCH_TEQL=m
+CONFIG_NET_SCH_TBF=m
+CONFIG_NET_SCH_GRED=m
+CONFIG_NET_SCH_DSMARK=m
+CONFIG_NET_SCH_NETEM=m
+CONFIG_NET_SCH_DRR=m
+CONFIG_NET_SCH_INGRESS=m
+CONFIG_NET_CLS_BASIC=m
+CONFIG_NET_CLS_TCINDEX=m
+CONFIG_NET_CLS_ROUTE4=m
+CONFIG_NET_CLS_FW=m
+CONFIG_NET_CLS_U32=m
+CONFIG_CLS_U32_PERF=y
+CONFIG_CLS_U32_MARK=y
+CONFIG_NET_CLS_RSVP=m
+CONFIG_NET_CLS_RSVP6=m
+CONFIG_NET_CLS_FLOW=m
+CONFIG_NET_CLS_CGROUP=y
+CONFIG_NET_EMATCH=y
+CONFIG_NET_EMATCH_CMP=m
+CONFIG_NET_EMATCH_NBYTE=m
+CONFIG_NET_EMATCH_U32=m
+CONFIG_NET_EMATCH_META=m
+CONFIG_NET_EMATCH_TEXT=m
+CONFIG_NET_CLS_ACT=y
+CONFIG_NET_ACT_POLICE=m
+CONFIG_NET_ACT_GACT=m
+CONFIG_GACT_PROB=y
+CONFIG_NET_ACT_MIRRED=m
+CONFIG_NET_ACT_IPT=m
+CONFIG_NET_ACT_NAT=m
+CONFIG_NET_ACT_PEDIT=m
+CONFIG_NET_ACT_SIMP=m
+CONFIG_NET_ACT_SKBEDIT=m
+CONFIG_NET_CLS_IND=y
+CONFIG_DCB=y
+CONFIG_DNS_RESOLVER=y
+# CONFIG_WIRELESS is not set
+CONFIG_UEVENT_HELPER_PATH="/sbin/hotplug"
+CONFIG_DEVTMPFS=y
+CONFIG_DEVTMPFS_MOUNT=y
+# CONFIG_FIRMWARE_IN_KERNEL is not set
+CONFIG_CONNECTOR=y
+CONFIG_BLK_DEV_LOOP=y
+CONFIG_BLK_DEV_CRYPTOLOOP=m
+CONFIG_BLK_DEV_SX8=m
+CONFIG_BLK_DEV_RAM=y
+CONFIG_BLK_DEV_RAM_SIZE=16384
+CONFIG_ATA_OVER_ETH=m
+CONFIG_RAID_ATTRS=m
+CONFIG_BLK_DEV_SD=y
+CONFIG_SCSI_CONSTANTS=y
+CONFIG_SCSI_LOGGING=y
+CONFIG_ATA=y
+CONFIG_SATA_SIL24=y
+# CONFIG_ATA_SFF is not set
+CONFIG_MD=y
+CONFIG_BLK_DEV_MD=y
+CONFIG_MD_LINEAR=m
+CONFIG_MD_RAID0=m
+CONFIG_MD_RAID1=m
+CONFIG_MD_RAID10=m
+CONFIG_MD_RAID456=m
+CONFIG_MD_FAULTY=m
+CONFIG_BLK_DEV_DM=m
+CONFIG_DM_DEBUG=y
+CONFIG_DM_CRYPT=m
+CONFIG_DM_SNAPSHOT=m
+CONFIG_DM_MIRROR=m
+CONFIG_DM_LOG_USERSPACE=m
+CONFIG_DM_ZERO=m
+CONFIG_DM_MULTIPATH=m
+CONFIG_DM_MULTIPATH_QL=m
+CONFIG_DM_MULTIPATH_ST=m
+CONFIG_DM_DELAY=m
+CONFIG_DM_UEVENT=y
+CONFIG_FUSION=y
+CONFIG_FUSION_SAS=y
+CONFIG_NETDEVICES=y
+CONFIG_BONDING=m
+CONFIG_DUMMY=m
+CONFIG_IFB=m
+CONFIG_MACVLAN=m
+CONFIG_MACVTAP=m
+CONFIG_NETCONSOLE=m
+CONFIG_NETCONSOLE_DYNAMIC=y
+CONFIG_TUN=y
+CONFIG_VETH=m
+CONFIG_NET_DSA_MV88E6060=y
+CONFIG_NET_DSA_MV88E6131=y
+CONFIG_NET_DSA_MV88E6123_61_65=y
+# CONFIG_NET_VENDOR_3COM is not set
+CONFIG_E1000E=y
+# CONFIG_WLAN is not set
+# CONFIG_INPUT_MOUSEDEV is not set
+# CONFIG_INPUT_KEYBOARD is not set
+# CONFIG_INPUT_MOUSE is not set
+# CONFIG_SERIO is not set
+# CONFIG_VT is not set
+# CONFIG_LEGACY_PTYS is not set
+CONFIG_HW_RANDOM=y
+CONFIG_HW_RANDOM_TIMERIOMEM=m
+CONFIG_I2C=y
+CONFIG_I2C_CHARDEV=y
+# CONFIG_HWMON is not set
+CONFIG_WATCHDOG=y
+CONFIG_WATCHDOG_NOWAYOUT=y
+# CONFIG_VGA_ARB is not set
+# CONFIG_USB_SUPPORT is not set
+CONFIG_EDAC=y
+CONFIG_EDAC_MM_EDAC=y
+CONFIG_RTC_CLASS=y
+CONFIG_RTC_DRV_TILE=y
+CONFIG_EXT2_FS=y
+CONFIG_EXT2_FS_XATTR=y
+CONFIG_EXT2_FS_POSIX_ACL=y
+CONFIG_EXT2_FS_SECURITY=y
+CONFIG_EXT2_FS_XIP=y
+CONFIG_EXT3_FS=y
+CONFIG_EXT3_FS_POSIX_ACL=y
+CONFIG_EXT3_FS_SECURITY=y
+CONFIG_EXT4_FS=y
+CONFIG_EXT4_FS_POSIX_ACL=y
+CONFIG_EXT4_FS_SECURITY=y
+CONFIG_XFS_FS=y
+CONFIG_XFS_QUOTA=y
+CONFIG_XFS_POSIX_ACL=y
+CONFIG_GFS2_FS=m
+CONFIG_GFS2_FS_LOCKING_DLM=y
+CONFIG_BTRFS_FS=m
+CONFIG_BTRFS_FS_POSIX_ACL=y
+CONFIG_QUOTA=y
+CONFIG_QUOTA_NETLINK_INTERFACE=y
+# CONFIG_PRINT_QUOTA_WARNING is not set
+CONFIG_QFMT_V2=y
+CONFIG_AUTOFS4_FS=m
+CONFIG_FUSE_FS=y
+CONFIG_CUSE=m
+CONFIG_FSCACHE=m
+CONFIG_FSCACHE_STATS=y
+CONFIG_CACHEFILES=m
+CONFIG_ISO9660_FS=m
+CONFIG_JOLIET=y
+CONFIG_ZISOFS=y
+CONFIG_UDF_FS=m
+CONFIG_MSDOS_FS=m
+CONFIG_VFAT_FS=m
+CONFIG_FAT_DEFAULT_IOCHARSET="ascii"
+CONFIG_PROC_KCORE=y
+CONFIG_TMPFS=y
+CONFIG_TMPFS_POSIX_ACL=y
+CONFIG_HUGETLBFS=y
+CONFIG_CONFIGFS_FS=m
+CONFIG_ECRYPT_FS=m
+CONFIG_CRAMFS=m
+CONFIG_SQUASHFS=m
+CONFIG_NFS_FS=m
+CONFIG_NFS_V3_ACL=y
+CONFIG_NFS_V4=m
+CONFIG_NFS_V4_1=y
+CONFIG_NFS_FSCACHE=y
+CONFIG_NFSD=m
+CONFIG_NFSD_V3_ACL=y
+CONFIG_NFSD_V4=y
+CONFIG_CIFS=m
+CONFIG_CIFS_STATS=y
+CONFIG_CIFS_WEAK_PW_HASH=y
+CONFIG_CIFS_UPCALL=y
+CONFIG_CIFS_XATTR=y
+CONFIG_CIFS_POSIX=y
+CONFIG_CIFS_DFS_UPCALL=y
+CONFIG_CIFS_FSCACHE=y
+CONFIG_NLS=y
+CONFIG_NLS_DEFAULT="utf8"
+CONFIG_NLS_CODEPAGE_437=y
+CONFIG_NLS_CODEPAGE_737=m
+CONFIG_NLS_CODEPAGE_775=m
+CONFIG_NLS_CODEPAGE_850=m
+CONFIG_NLS_CODEPAGE_852=m
+CONFIG_NLS_CODEPAGE_855=m
+CONFIG_NLS_CODEPAGE_857=m
+CONFIG_NLS_CODEPAGE_860=m
+CONFIG_NLS_CODEPAGE_861=m
+CONFIG_NLS_CODEPAGE_862=m
+CONFIG_NLS_CODEPAGE_863=m
+CONFIG_NLS_CODEPAGE_864=m
+CONFIG_NLS_CODEPAGE_865=m
+CONFIG_NLS_CODEPAGE_866=m
+CONFIG_NLS_CODEPAGE_869=m
+CONFIG_NLS_CODEPAGE_936=m
+CONFIG_NLS_CODEPAGE_950=m
+CONFIG_NLS_CODEPAGE_932=m
+CONFIG_NLS_CODEPAGE_949=m
+CONFIG_NLS_CODEPAGE_874=m
+CONFIG_NLS_ISO8859_8=m
+CONFIG_NLS_CODEPAGE_1250=m
+CONFIG_NLS_CODEPAGE_1251=m
+CONFIG_NLS_ASCII=y
+CONFIG_NLS_ISO8859_1=m
+CONFIG_NLS_ISO8859_2=m
+CONFIG_NLS_ISO8859_3=m
+CONFIG_NLS_ISO8859_4=m
+CONFIG_NLS_ISO8859_5=m
+CONFIG_NLS_ISO8859_6=m
+CONFIG_NLS_ISO8859_7=m
+CONFIG_NLS_ISO8859_9=m
+CONFIG_NLS_ISO8859_13=m
+CONFIG_NLS_ISO8859_14=m
+CONFIG_NLS_ISO8859_15=m
+CONFIG_NLS_KOI8_R=m
+CONFIG_NLS_KOI8_U=m
+CONFIG_NLS_UTF8=m
+CONFIG_DLM=m
+CONFIG_DLM_DEBUG=y
+CONFIG_DYNAMIC_DEBUG=y
+CONFIG_DEBUG_INFO=y
+CONFIG_DEBUG_INFO_REDUCED=y
+# CONFIG_ENABLE_WARN_DEPRECATED is not set
+CONFIG_FRAME_WARN=2048
+CONFIG_STRIP_ASM_SYMS=y
+CONFIG_DEBUG_FS=y
+CONFIG_HEADERS_CHECK=y
+# CONFIG_FRAME_POINTER is not set
+CONFIG_DEBUG_FORCE_WEAK_PER_CPU=y
+CONFIG_MAGIC_SYSRQ=y
+CONFIG_DEBUG_VM=y
+CONFIG_DEBUG_MEMORY_INIT=y
+CONFIG_DEBUG_STACKOVERFLOW=y
+CONFIG_LOCKUP_DETECTOR=y
+CONFIG_SCHEDSTATS=y
+CONFIG_TIMER_STATS=y
+CONFIG_DEBUG_LIST=y
+CONFIG_DEBUG_CREDENTIALS=y
+CONFIG_RCU_CPU_STALL_TIMEOUT=60
+CONFIG_ASYNC_RAID6_TEST=m
+CONFIG_KEYS_DEBUG_PROC_KEYS=y
+CONFIG_SECURITY=y
+CONFIG_SECURITYFS=y
+CONFIG_SECURITY_NETWORK=y
+CONFIG_SECURITY_NETWORK_XFRM=y
+CONFIG_SECURITY_SELINUX=y
+CONFIG_SECURITY_SELINUX_BOOTPARAM=y
+CONFIG_SECURITY_SELINUX_DISABLE=y
+CONFIG_CRYPTO_PCRYPT=m
+CONFIG_CRYPTO_CRYPTD=m
+CONFIG_CRYPTO_TEST=m
+CONFIG_CRYPTO_CCM=m
+CONFIG_CRYPTO_GCM=m
+CONFIG_CRYPTO_CTS=m
+CONFIG_CRYPTO_LRW=m
+CONFIG_CRYPTO_PCBC=m
+CONFIG_CRYPTO_XTS=m
+CONFIG_CRYPTO_HMAC=y
+CONFIG_CRYPTO_XCBC=m
+CONFIG_CRYPTO_VMAC=m
+CONFIG_CRYPTO_MICHAEL_MIC=m
+CONFIG_CRYPTO_RMD128=m
+CONFIG_CRYPTO_RMD160=m
+CONFIG_CRYPTO_RMD256=m
+CONFIG_CRYPTO_RMD320=m
+CONFIG_CRYPTO_SHA1=y
+CONFIG_CRYPTO_SHA512=m
+CONFIG_CRYPTO_TGR192=m
+CONFIG_CRYPTO_WP512=m
+CONFIG_CRYPTO_ANUBIS=m
+CONFIG_CRYPTO_BLOWFISH=m
+CONFIG_CRYPTO_CAMELLIA=m
+CONFIG_CRYPTO_CAST5=m
+CONFIG_CRYPTO_CAST6=m
+CONFIG_CRYPTO_FCRYPT=m
+CONFIG_CRYPTO_KHAZAD=m
+CONFIG_CRYPTO_SEED=m
+CONFIG_CRYPTO_SERPENT=m
+CONFIG_CRYPTO_TEA=m
+CONFIG_CRYPTO_TWOFISH=m
+CONFIG_CRYPTO_ZLIB=m
+CONFIG_CRYPTO_LZO=m
+CONFIG_CRC_CCITT=m
+CONFIG_CRC7=m
diff --git a/kernel/arch/tile/gxio/Kconfig b/kernel/arch/tile/gxio/Kconfig
new file mode 100644
index 000000000..d4e10d580
--- /dev/null
+++ b/kernel/arch/tile/gxio/Kconfig
@@ -0,0 +1,33 @@
+# Support direct access to TILE-Gx hardware from user space, via the
+# gxio library, or from kernel space, via kernel IORPC support.
+config TILE_GXIO
+	bool
+	depends on TILEGX
+
+# Support direct access to the common I/O DMA facility within the
+# TILE-Gx mPIPE and Trio hardware from kernel space.
+config TILE_GXIO_DMA
+	bool
+	select TILE_GXIO
+
+# Support direct access to the TILE-Gx mPIPE hardware from kernel space.
+config TILE_GXIO_MPIPE
+	bool
+	select TILE_GXIO
+	select TILE_GXIO_DMA
+
+# Support direct access to the TILE-Gx TRIO hardware from kernel space.
+config TILE_GXIO_TRIO
+	bool
+	select TILE_GXIO
+	select TILE_GXIO_DMA
+
+# Support direct access to the TILE-Gx USB hardware from kernel space.
+config TILE_GXIO_USB_HOST
+	bool
+	select TILE_GXIO
+
+# Support direct access to the TILE-Gx UART hardware from kernel space.
+config TILE_GXIO_UART
+	bool
+	select TILE_GXIO
diff --git a/kernel/arch/tile/gxio/Makefile b/kernel/arch/tile/gxio/Makefile
new file mode 100644
index 000000000..26ae2c727
--- /dev/null
+++ b/kernel/arch/tile/gxio/Makefile
@@ -0,0 +1,10 @@
+#
+# Makefile for the Tile-Gx device access support.
+#
+
+obj-$(CONFIG_TILE_GXIO) += iorpc_globals.o kiorpc.o
+obj-$(CONFIG_TILE_GXIO_DMA) += dma_queue.o
+obj-$(CONFIG_TILE_GXIO_MPIPE) += mpipe.o iorpc_mpipe.o iorpc_mpipe_info.o
+obj-$(CONFIG_TILE_GXIO_TRIO) += trio.o iorpc_trio.o
+obj-$(CONFIG_TILE_GXIO_UART) += uart.o iorpc_uart.o
+obj-$(CONFIG_TILE_GXIO_USB_HOST) += usb_host.o iorpc_usb_host.o
diff --git a/kernel/arch/tile/gxio/dma_queue.c b/kernel/arch/tile/gxio/dma_queue.c
new file mode 100644
index 000000000..baa60357f
--- /dev/null
+++ b/kernel/arch/tile/gxio/dma_queue.c
@@ -0,0 +1,176 @@
+/*
+ * Copyright 2012 Tilera Corporation. All Rights Reserved.
+ *
+ *   This program is free software; you can redistribute it and/or
+ *   modify it under the terms of the GNU General Public License
+ *   as published by the Free Software Foundation, version 2.
+ *
+ *   This program is distributed in the hope that it will be useful, but
+ *   WITHOUT ANY WARRANTY; without even the implied warranty of
+ *   MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
+ *   NON INFRINGEMENT.  See the GNU General Public License for
+ *   more details.
+ */
+
+#include <linux/io.h>
+#include <linux/atomic.h>
+#include <linux/module.h>
+#include <gxio/dma_queue.h>
+
+/* Wait for a memory read to complete. */
+#define wait_for_value(val)                             \
+  __asm__ __volatile__("move %0, %0" :: "r"(val))
+
+/* The index is in the low 16. */
+#define DMA_QUEUE_INDEX_MASK ((1 << 16) - 1)
+
+/*
+ * The hardware descriptor-ring type.
+ * This matches the types used by mpipe (MPIPE_EDMA_POST_REGION_VAL_t)
+ * and trio (TRIO_PUSH_DMA_REGION_VAL_t or TRIO_PULL_DMA_REGION_VAL_t).
+ * See those types for more documentation on the individual fields.
+ */
+typedef union {
+	struct {
+#ifndef __BIG_ENDIAN__
+		uint64_t ring_idx:16;
+		uint64_t count:16;
+		uint64_t gen:1;
+		uint64_t __reserved:31;
+#else
+		uint64_t __reserved:31;
+		uint64_t gen:1;
+		uint64_t count:16;
+		uint64_t ring_idx:16;
+#endif
+	};
+	uint64_t word;
+} __gxio_ring_t;
+
+void __gxio_dma_queue_init(__gxio_dma_queue_t *dma_queue,
+			   void *post_region_addr, unsigned int num_entries)
+{
+	/*
+	 * Limit 65536 entry rings to 65535 credits because we only have a
+	 * 16 bit completion counter.
+	 */
+	int64_t credits = (num_entries < 65536) ? num_entries : 65535;
+
+	memset(dma_queue, 0, sizeof(*dma_queue));
+
+	dma_queue->post_region_addr = post_region_addr;
+	dma_queue->hw_complete_count = 0;
+	dma_queue->credits_and_next_index = credits << DMA_QUEUE_CREDIT_SHIFT;
+}
+
+EXPORT_SYMBOL_GPL(__gxio_dma_queue_init);
+
+void __gxio_dma_queue_update_credits(__gxio_dma_queue_t *dma_queue)
+{
+	__gxio_ring_t val;
+	uint64_t count;
+	uint64_t delta;
+	uint64_t new_count;
+
+	/*
+	 * Read the 64-bit completion count without touching the cache, so
+	 * we later avoid having to evict any sharers of this cache line
+	 * when we update it below.
+	 */
+	uint64_t orig_hw_complete_count =
+		cmpxchg(&dma_queue->hw_complete_count,
+			-1, -1);
+
+	/* Make sure the load completes before we access the hardware. */
+	wait_for_value(orig_hw_complete_count);
+
+	/* Read the 16-bit count of how many packets it has completed. */
+	val.word = __gxio_mmio_read(dma_queue->post_region_addr);
+	count = val.count;
+
+	/*
+	 * Calculate the number of completions since we last updated the
+	 * 64-bit counter.  It's safe to ignore the high bits because the
+	 * maximum credit value is 65535.
+	 */
+	delta = (count - orig_hw_complete_count) & 0xffff;
+	if (delta == 0)
+		return;
+
+	/*
+	 * Try to write back the count, advanced by delta.  If we race with
+	 * another thread, this might fail, in which case we return
+	 * immediately on the assumption that some credits are (or at least
+	 * were) available.
+	 */
+	new_count = orig_hw_complete_count + delta;
+	if (cmpxchg(&dma_queue->hw_complete_count,
+		    orig_hw_complete_count,
+		    new_count) != orig_hw_complete_count)
+		return;
+
+	/*
+	 * We succeeded in advancing the completion count; add back the
+	 * corresponding number of egress credits.
+	 */
+	__insn_fetchadd(&dma_queue->credits_and_next_index,
+			(delta << DMA_QUEUE_CREDIT_SHIFT));
+}
+
+EXPORT_SYMBOL_GPL(__gxio_dma_queue_update_credits);
+
+/*
+ * A separate 'blocked' method for put() so that backtraces and
+ * profiles will clearly indicate that we're wasting time spinning on
+ * egress availability rather than actually posting commands.
+ */
+int64_t __gxio_dma_queue_wait_for_credits(__gxio_dma_queue_t *dma_queue,
+					  int64_t modifier)
+{
+	int backoff = 16;
+	int64_t old;
+
+	do {
+		int i;
+		/* Back off to avoid spamming memory networks. */
+		for (i = backoff; i > 0; i--)
+			__insn_mfspr(SPR_PASS);
+
+		/* Check credits again. */
+		__gxio_dma_queue_update_credits(dma_queue);
+		old = __insn_fetchaddgez(&dma_queue->credits_and_next_index,
+					 modifier);
+
+		/* Calculate bounded exponential backoff for next iteration. */
+		if (backoff < 256)
+			backoff *= 2;
+	} while (old + modifier < 0);
+
+	return old;
+}
+
+EXPORT_SYMBOL_GPL(__gxio_dma_queue_wait_for_credits);
+
+int64_t __gxio_dma_queue_reserve_aux(__gxio_dma_queue_t *dma_queue,
+				     unsigned int num, int wait)
+{
+	return __gxio_dma_queue_reserve(dma_queue, num, wait != 0, true);
+}
+
+EXPORT_SYMBOL_GPL(__gxio_dma_queue_reserve_aux);
+
+int __gxio_dma_queue_is_complete(__gxio_dma_queue_t *dma_queue,
+				 int64_t completion_slot, int update)
+{
+	if (update) {
+		if (ACCESS_ONCE(dma_queue->hw_complete_count) >
+		    completion_slot)
+			return 1;
+
+		__gxio_dma_queue_update_credits(dma_queue);
+	}
+
+	return ACCESS_ONCE(dma_queue->hw_complete_count) > completion_slot;
+}
+
+EXPORT_SYMBOL_GPL(__gxio_dma_queue_is_complete);
diff --git a/kernel/arch/tile/gxio/iorpc_globals.c b/kernel/arch/tile/gxio/iorpc_globals.c
new file mode 100644
index 000000000..e178e9080
--- /dev/null
+++ b/kernel/arch/tile/gxio/iorpc_globals.c
@@ -0,0 +1,89 @@
+/*
+ * Copyright 2012 Tilera Corporation. All Rights Reserved.
+ *
+ *   This program is free software; you can redistribute it and/or
+ *   modify it under the terms of the GNU General Public License
+ *   as published by the Free Software Foundation, version 2.
+ *
+ *   This program is distributed in the hope that it will be useful, but
+ *   WITHOUT ANY WARRANTY; without even the implied warranty of
+ *   MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
+ *   NON INFRINGEMENT.  See the GNU General Public License for
+ *   more details.
+ */
+
+/* This file is machine-generated; DO NOT EDIT! */
+#include "gxio/iorpc_globals.h"
+
+struct arm_pollfd_param {
+	union iorpc_pollfd pollfd;
+};
+
+int __iorpc_arm_pollfd(int fd, int pollfd_cookie)
+{
+	struct arm_pollfd_param temp;
+	struct arm_pollfd_param *params = &temp;
+
+	params->pollfd.kernel.cookie = pollfd_cookie;
+
+	return hv_dev_pwrite(fd, 0, (HV_VirtAddr) params, sizeof(*params),
+			     IORPC_OP_ARM_POLLFD);
+}
+
+EXPORT_SYMBOL(__iorpc_arm_pollfd);
+
+struct close_pollfd_param {
+	union iorpc_pollfd pollfd;
+};
+
+int __iorpc_close_pollfd(int fd, int pollfd_cookie)
+{
+	struct close_pollfd_param temp;
+	struct close_pollfd_param *params = &temp;
+
+	params->pollfd.kernel.cookie = pollfd_cookie;
+
+	return hv_dev_pwrite(fd, 0, (HV_VirtAddr) params, sizeof(*params),
+			     IORPC_OP_CLOSE_POLLFD);
+}
+
+EXPORT_SYMBOL(__iorpc_close_pollfd);
+
+struct get_mmio_base_param {
+	HV_PTE base;
+};
+
+int __iorpc_get_mmio_base(int fd, HV_PTE *base)
+{
+	int __result;
+	struct get_mmio_base_param temp;
+	struct get_mmio_base_param *params = &temp;
+
+	__result =
+	    hv_dev_pread(fd, 0, (HV_VirtAddr) params, sizeof(*params),
+			 IORPC_OP_GET_MMIO_BASE);
+	*base = params->base;
+
+	return __result;
+}
+
+EXPORT_SYMBOL(__iorpc_get_mmio_base);
+
+struct check_mmio_offset_param {
+	unsigned long offset;
+	unsigned long size;
+};
+
+int __iorpc_check_mmio_offset(int fd, unsigned long offset, unsigned long size)
+{
+	struct check_mmio_offset_param temp;
+	struct check_mmio_offset_param *params = &temp;
+
+	params->offset = offset;
+	params->size = size;
+
+	return hv_dev_pwrite(fd, 0, (HV_VirtAddr) params, sizeof(*params),
+			     IORPC_OP_CHECK_MMIO_OFFSET);
+}
+
+EXPORT_SYMBOL(__iorpc_check_mmio_offset);
diff --git a/kernel/arch/tile/gxio/iorpc_mpipe.c b/kernel/arch/tile/gxio/iorpc_mpipe.c
new file mode 100644
index 000000000..e19325c4c
--- /dev/null
+++ b/kernel/arch/tile/gxio/iorpc_mpipe.c
@@ -0,0 +1,593 @@
+/*
+ * Copyright 2012 Tilera Corporation. All Rights Reserved.
+ *
+ *   This program is free software; you can redistribute it and/or
+ *   modify it under the terms of the GNU General Public License
+ *   as published by the Free Software Foundation, version 2.
+ *
+ *   This program is distributed in the hope that it will be useful, but
+ *   WITHOUT ANY WARRANTY; without even the implied warranty of
+ *   MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
+ *   NON INFRINGEMENT.  See the GNU General Public License for
+ *   more details.
+ */
+
+/* This file is machine-generated; DO NOT EDIT! */
+#include "gxio/iorpc_mpipe.h"
+
+struct alloc_buffer_stacks_param {
+	unsigned int count;
+	unsigned int first;
+	unsigned int flags;
+};
+
+int gxio_mpipe_alloc_buffer_stacks(gxio_mpipe_context_t *context,
+				   unsigned int count, unsigned int first,
+				   unsigned int flags)
+{
+	struct alloc_buffer_stacks_param temp;
+	struct alloc_buffer_stacks_param *params = &temp;
+
+	params->count = count;
+	params->first = first;
+	params->flags = flags;
+
+	return hv_dev_pwrite(context->fd, 0, (HV_VirtAddr) params,
+			     sizeof(*params),
+			     GXIO_MPIPE_OP_ALLOC_BUFFER_STACKS);
+}
+
+EXPORT_SYMBOL(gxio_mpipe_alloc_buffer_stacks);
+
+struct init_buffer_stack_aux_param {
+	union iorpc_mem_buffer buffer;
+	unsigned int stack;
+	unsigned int buffer_size_enum;
+};
+
+int gxio_mpipe_init_buffer_stack_aux(gxio_mpipe_context_t *context,
+				     void *mem_va, size_t mem_size,
+				     unsigned int mem_flags, unsigned int stack,
+				     unsigned int buffer_size_enum)
+{
+	int __result;
+	unsigned long long __cpa;
+	pte_t __pte;
+	struct init_buffer_stack_aux_param temp;
+	struct init_buffer_stack_aux_param *params = &temp;
+
+	__result = va_to_cpa_and_pte(mem_va, &__cpa, &__pte);
+	if (__result != 0)
+		return __result;
+	params->buffer.kernel.cpa = __cpa;
+	params->buffer.kernel.size = mem_size;
+	params->buffer.kernel.pte = __pte;
+	params->buffer.kernel.flags = mem_flags;
+	params->stack = stack;
+	params->buffer_size_enum = buffer_size_enum;
+
+	return hv_dev_pwrite(context->fd, 0, (HV_VirtAddr) params,
+			     sizeof(*params),
+			     GXIO_MPIPE_OP_INIT_BUFFER_STACK_AUX);
+}
+
+EXPORT_SYMBOL(gxio_mpipe_init_buffer_stack_aux);
+
+
+struct alloc_notif_rings_param {
+	unsigned int count;
+	unsigned int first;
+	unsigned int flags;
+};
+
+int gxio_mpipe_alloc_notif_rings(gxio_mpipe_context_t *context,
+				 unsigned int count, unsigned int first,
+				 unsigned int flags)
+{
+	struct alloc_notif_rings_param temp;
+	struct alloc_notif_rings_param *params = &temp;
+
+	params->count = count;
+	params->first = first;
+	params->flags = flags;
+
+	return hv_dev_pwrite(context->fd, 0, (HV_VirtAddr) params,
+			     sizeof(*params), GXIO_MPIPE_OP_ALLOC_NOTIF_RINGS);
+}
+
+EXPORT_SYMBOL(gxio_mpipe_alloc_notif_rings);
+
+struct init_notif_ring_aux_param {
+	union iorpc_mem_buffer buffer;
+	unsigned int ring;
+};
+
+int gxio_mpipe_init_notif_ring_aux(gxio_mpipe_context_t *context, void *mem_va,
+				   size_t mem_size, unsigned int mem_flags,
+				   unsigned int ring)
+{
+	int __result;
+	unsigned long long __cpa;
+	pte_t __pte;
+	struct init_notif_ring_aux_param temp;
+	struct init_notif_ring_aux_param *params = &temp;
+
+	__result = va_to_cpa_and_pte(mem_va, &__cpa, &__pte);
+	if (__result != 0)
+		return __result;
+	params->buffer.kernel.cpa = __cpa;
+	params->buffer.kernel.size = mem_size;
+	params->buffer.kernel.pte = __pte;
+	params->buffer.kernel.flags = mem_flags;
+	params->ring = ring;
+
+	return hv_dev_pwrite(context->fd, 0, (HV_VirtAddr) params,
+			     sizeof(*params),
+			     GXIO_MPIPE_OP_INIT_NOTIF_RING_AUX);
+}
+
+EXPORT_SYMBOL(gxio_mpipe_init_notif_ring_aux);
+
+struct request_notif_ring_interrupt_param {
+	union iorpc_interrupt interrupt;
+	unsigned int ring;
+};
+
+int gxio_mpipe_request_notif_ring_interrupt(gxio_mpipe_context_t *context,
+					    int inter_x, int inter_y,
+					    int inter_ipi, int inter_event,
+					    unsigned int ring)
+{
+	struct request_notif_ring_interrupt_param temp;
+	struct request_notif_ring_interrupt_param *params = &temp;
+
+	params->interrupt.kernel.x = inter_x;
+	params->interrupt.kernel.y = inter_y;
+	params->interrupt.kernel.ipi = inter_ipi;
+	params->interrupt.kernel.event = inter_event;
+	params->ring = ring;
+
+	return hv_dev_pwrite(context->fd, 0, (HV_VirtAddr) params,
+			     sizeof(*params),
+			     GXIO_MPIPE_OP_REQUEST_NOTIF_RING_INTERRUPT);
+}
+
+EXPORT_SYMBOL(gxio_mpipe_request_notif_ring_interrupt);
+
+struct enable_notif_ring_interrupt_param {
+	unsigned int ring;
+};
+
+int gxio_mpipe_enable_notif_ring_interrupt(gxio_mpipe_context_t *context,
+					   unsigned int ring)
+{
+	struct enable_notif_ring_interrupt_param temp;
+	struct enable_notif_ring_interrupt_param *params = &temp;
+
+	params->ring = ring;
+
+	return hv_dev_pwrite(context->fd, 0, (HV_VirtAddr) params,
+			     sizeof(*params),
+			     GXIO_MPIPE_OP_ENABLE_NOTIF_RING_INTERRUPT);
+}
+
+EXPORT_SYMBOL(gxio_mpipe_enable_notif_ring_interrupt);
+
+struct alloc_notif_groups_param {
+	unsigned int count;
+	unsigned int first;
+	unsigned int flags;
+};
+
+int gxio_mpipe_alloc_notif_groups(gxio_mpipe_context_t *context,
+				  unsigned int count, unsigned int first,
+				  unsigned int flags)
+{
+	struct alloc_notif_groups_param temp;
+	struct alloc_notif_groups_param *params = &temp;
+
+	params->count = count;
+	params->first = first;
+	params->flags = flags;
+
+	return hv_dev_pwrite(context->fd, 0, (HV_VirtAddr) params,
+			     sizeof(*params), GXIO_MPIPE_OP_ALLOC_NOTIF_GROUPS);
+}
+
+EXPORT_SYMBOL(gxio_mpipe_alloc_notif_groups);
+
+struct init_notif_group_param {
+	unsigned int group;
+	gxio_mpipe_notif_group_bits_t bits;
+};
+
+int gxio_mpipe_init_notif_group(gxio_mpipe_context_t *context,
+				unsigned int group,
+				gxio_mpipe_notif_group_bits_t bits)
+{
+	struct init_notif_group_param temp;
+	struct init_notif_group_param *params = &temp;
+
+	params->group = group;
+	params->bits = bits;
+
+	return hv_dev_pwrite(context->fd, 0, (HV_VirtAddr) params,
+			     sizeof(*params), GXIO_MPIPE_OP_INIT_NOTIF_GROUP);
+}
+
+EXPORT_SYMBOL(gxio_mpipe_init_notif_group);
+
+struct alloc_buckets_param {
+	unsigned int count;
+	unsigned int first;
+	unsigned int flags;
+};
+
+int gxio_mpipe_alloc_buckets(gxio_mpipe_context_t *context, unsigned int count,
+			     unsigned int first, unsigned int flags)
+{
+	struct alloc_buckets_param temp;
+	struct alloc_buckets_param *params = &temp;
+
+	params->count = count;
+	params->first = first;
+	params->flags = flags;
+
+	return hv_dev_pwrite(context->fd, 0, (HV_VirtAddr) params,
+			     sizeof(*params), GXIO_MPIPE_OP_ALLOC_BUCKETS);
+}
+
+EXPORT_SYMBOL(gxio_mpipe_alloc_buckets);
+
+struct init_bucket_param {
+	unsigned int bucket;
+	MPIPE_LBL_INIT_DAT_BSTS_TBL_t bucket_info;
+};
+
+int gxio_mpipe_init_bucket(gxio_mpipe_context_t *context, unsigned int bucket,
+			   MPIPE_LBL_INIT_DAT_BSTS_TBL_t bucket_info)
+{
+	struct init_bucket_param temp;
+	struct init_bucket_param *params = &temp;
+
+	params->bucket = bucket;
+	params->bucket_info = bucket_info;
+
+	return hv_dev_pwrite(context->fd, 0, (HV_VirtAddr) params,
+			     sizeof(*params), GXIO_MPIPE_OP_INIT_BUCKET);
+}
+
+EXPORT_SYMBOL(gxio_mpipe_init_bucket);
+
+struct alloc_edma_rings_param {
+	unsigned int count;
+	unsigned int first;
+	unsigned int flags;
+};
+
+int gxio_mpipe_alloc_edma_rings(gxio_mpipe_context_t *context,
+				unsigned int count, unsigned int first,
+				unsigned int flags)
+{
+	struct alloc_edma_rings_param temp;
+	struct alloc_edma_rings_param *params = &temp;
+
+	params->count = count;
+	params->first = first;
+	params->flags = flags;
+
+	return hv_dev_pwrite(context->fd, 0, (HV_VirtAddr) params,
+			     sizeof(*params), GXIO_MPIPE_OP_ALLOC_EDMA_RINGS);
+}
+
+EXPORT_SYMBOL(gxio_mpipe_alloc_edma_rings);
+
+struct init_edma_ring_aux_param {
+	union iorpc_mem_buffer buffer;
+	unsigned int ring;
+	unsigned int channel;
+};
+
+int gxio_mpipe_init_edma_ring_aux(gxio_mpipe_context_t *context, void *mem_va,
+				  size_t mem_size, unsigned int mem_flags,
+				  unsigned int ring, unsigned int channel)
+{
+	int __result;
+	unsigned long long __cpa;
+	pte_t __pte;
+	struct init_edma_ring_aux_param temp;
+	struct init_edma_ring_aux_param *params = &temp;
+
+	__result = va_to_cpa_and_pte(mem_va, &__cpa, &__pte);
+	if (__result != 0)
+		return __result;
+	params->buffer.kernel.cpa = __cpa;
+	params->buffer.kernel.size = mem_size;
+	params->buffer.kernel.pte = __pte;
+	params->buffer.kernel.flags = mem_flags;
+	params->ring = ring;
+	params->channel = channel;
+
+	return hv_dev_pwrite(context->fd, 0, (HV_VirtAddr) params,
+			     sizeof(*params), GXIO_MPIPE_OP_INIT_EDMA_RING_AUX);
+}
+
+EXPORT_SYMBOL(gxio_mpipe_init_edma_ring_aux);
+
+
+int gxio_mpipe_commit_rules(gxio_mpipe_context_t *context, const void *blob,
+			    size_t blob_size)
+{
+	const void *params = blob;
+
+	return hv_dev_pwrite(context->fd, 0, (HV_VirtAddr) params, blob_size,
+			     GXIO_MPIPE_OP_COMMIT_RULES);
+}
+
+EXPORT_SYMBOL(gxio_mpipe_commit_rules);
+
+struct register_client_memory_param {
+	unsigned int iotlb;
+	HV_PTE pte;
+	unsigned int flags;
+};
+
+int gxio_mpipe_register_client_memory(gxio_mpipe_context_t *context,
+				      unsigned int iotlb, HV_PTE pte,
+				      unsigned int flags)
+{
+	struct register_client_memory_param temp;
+	struct register_client_memory_param *params = &temp;
+
+	params->iotlb = iotlb;
+	params->pte = pte;
+	params->flags = flags;
+
+	return hv_dev_pwrite(context->fd, 0, (HV_VirtAddr) params,
+			     sizeof(*params),
+			     GXIO_MPIPE_OP_REGISTER_CLIENT_MEMORY);
+}
+
+EXPORT_SYMBOL(gxio_mpipe_register_client_memory);
+
+struct link_open_aux_param {
+	_gxio_mpipe_link_name_t name;
+	unsigned int flags;
+};
+
+int gxio_mpipe_link_open_aux(gxio_mpipe_context_t *context,
+			     _gxio_mpipe_link_name_t name, unsigned int flags)
+{
+	struct link_open_aux_param temp;
+	struct link_open_aux_param *params = &temp;
+
+	params->name = name;
+	params->flags = flags;
+
+	return hv_dev_pwrite(context->fd, 0, (HV_VirtAddr) params,
+			     sizeof(*params), GXIO_MPIPE_OP_LINK_OPEN_AUX);
+}
+
+EXPORT_SYMBOL(gxio_mpipe_link_open_aux);
+
+struct link_close_aux_param {
+	int mac;
+};
+
+int gxio_mpipe_link_close_aux(gxio_mpipe_context_t *context, int mac)
+{
+	struct link_close_aux_param temp;
+	struct link_close_aux_param *params = &temp;
+
+	params->mac = mac;
+
+	return hv_dev_pwrite(context->fd, 0, (HV_VirtAddr) params,
+			     sizeof(*params), GXIO_MPIPE_OP_LINK_CLOSE_AUX);
+}
+
+EXPORT_SYMBOL(gxio_mpipe_link_close_aux);
+
+struct link_set_attr_aux_param {
+	int mac;
+	uint32_t attr;
+	int64_t val;
+};
+
+int gxio_mpipe_link_set_attr_aux(gxio_mpipe_context_t *context, int mac,
+				 uint32_t attr, int64_t val)
+{
+	struct link_set_attr_aux_param temp;
+	struct link_set_attr_aux_param *params = &temp;
+
+	params->mac = mac;
+	params->attr = attr;
+	params->val = val;
+
+	return hv_dev_pwrite(context->fd, 0, (HV_VirtAddr) params,
+			     sizeof(*params), GXIO_MPIPE_OP_LINK_SET_ATTR_AUX);
+}
+
+EXPORT_SYMBOL(gxio_mpipe_link_set_attr_aux);
+
+struct get_timestamp_aux_param {
+	uint64_t sec;
+	uint64_t nsec;
+	uint64_t cycles;
+};
+
+int gxio_mpipe_get_timestamp_aux(gxio_mpipe_context_t *context, uint64_t *sec,
+				 uint64_t *nsec, uint64_t *cycles)
+{
+	int __result;
+	struct get_timestamp_aux_param temp;
+	struct get_timestamp_aux_param *params = &temp;
+
+	__result =
+	    hv_dev_pread(context->fd, 0, (HV_VirtAddr) params, sizeof(*params),
+			 GXIO_MPIPE_OP_GET_TIMESTAMP_AUX);
+	*sec = params->sec;
+	*nsec = params->nsec;
+	*cycles = params->cycles;
+
+	return __result;
+}
+
+EXPORT_SYMBOL(gxio_mpipe_get_timestamp_aux);
+
+struct set_timestamp_aux_param {
+	uint64_t sec;
+	uint64_t nsec;
+	uint64_t cycles;
+};
+
+int gxio_mpipe_set_timestamp_aux(gxio_mpipe_context_t *context, uint64_t sec,
+				 uint64_t nsec, uint64_t cycles)
+{
+	struct set_timestamp_aux_param temp;
+	struct set_timestamp_aux_param *params = &temp;
+
+	params->sec = sec;
+	params->nsec = nsec;
+	params->cycles = cycles;
+
+	return hv_dev_pwrite(context->fd, 0, (HV_VirtAddr) params,
+			     sizeof(*params), GXIO_MPIPE_OP_SET_TIMESTAMP_AUX);
+}
+
+EXPORT_SYMBOL(gxio_mpipe_set_timestamp_aux);
+
+struct adjust_timestamp_aux_param {
+	int64_t nsec;
+};
+
+int gxio_mpipe_adjust_timestamp_aux(gxio_mpipe_context_t *context, int64_t nsec)
+{
+	struct adjust_timestamp_aux_param temp;
+	struct adjust_timestamp_aux_param *params = &temp;
+
+	params->nsec = nsec;
+
+	return hv_dev_pwrite(context->fd, 0, (HV_VirtAddr) params,
+			     sizeof(*params),
+			     GXIO_MPIPE_OP_ADJUST_TIMESTAMP_AUX);
+}
+
+EXPORT_SYMBOL(gxio_mpipe_adjust_timestamp_aux);
+
+struct config_edma_ring_blks_param {
+	unsigned int ering;
+	unsigned int max_blks;
+	unsigned int min_snf_blks;
+	unsigned int db;
+};
+
+int gxio_mpipe_config_edma_ring_blks(gxio_mpipe_context_t *context,
+				     unsigned int ering, unsigned int max_blks,
+				     unsigned int min_snf_blks, unsigned int db)
+{
+	struct config_edma_ring_blks_param temp;
+	struct config_edma_ring_blks_param *params = &temp;
+
+	params->ering = ering;
+	params->max_blks = max_blks;
+	params->min_snf_blks = min_snf_blks;
+	params->db = db;
+
+	return hv_dev_pwrite(context->fd, 0, (HV_VirtAddr) params,
+			     sizeof(*params),
+			     GXIO_MPIPE_OP_CONFIG_EDMA_RING_BLKS);
+}
+
+EXPORT_SYMBOL(gxio_mpipe_config_edma_ring_blks);
+
+struct adjust_timestamp_freq_param {
+	int32_t ppb;
+};
+
+int gxio_mpipe_adjust_timestamp_freq(gxio_mpipe_context_t *context, int32_t ppb)
+{
+	struct adjust_timestamp_freq_param temp;
+	struct adjust_timestamp_freq_param *params = &temp;
+
+	params->ppb = ppb;
+
+	return hv_dev_pwrite(context->fd, 0, (HV_VirtAddr) params,
+			     sizeof(*params),
+			     GXIO_MPIPE_OP_ADJUST_TIMESTAMP_FREQ);
+}
+
+EXPORT_SYMBOL(gxio_mpipe_adjust_timestamp_freq);
+
+struct arm_pollfd_param {
+	union iorpc_pollfd pollfd;
+};
+
+int gxio_mpipe_arm_pollfd(gxio_mpipe_context_t *context, int pollfd_cookie)
+{
+	struct arm_pollfd_param temp;
+	struct arm_pollfd_param *params = &temp;
+
+	params->pollfd.kernel.cookie = pollfd_cookie;
+
+	return hv_dev_pwrite(context->fd, 0, (HV_VirtAddr) params,
+			     sizeof(*params), GXIO_MPIPE_OP_ARM_POLLFD);
+}
+
+EXPORT_SYMBOL(gxio_mpipe_arm_pollfd);
+
+struct close_pollfd_param {
+	union iorpc_pollfd pollfd;
+};
+
+int gxio_mpipe_close_pollfd(gxio_mpipe_context_t *context, int pollfd_cookie)
+{
+	struct close_pollfd_param temp;
+	struct close_pollfd_param *params = &temp;
+
+	params->pollfd.kernel.cookie = pollfd_cookie;
+
+	return hv_dev_pwrite(context->fd, 0, (HV_VirtAddr) params,
+			     sizeof(*params), GXIO_MPIPE_OP_CLOSE_POLLFD);
+}
+
+EXPORT_SYMBOL(gxio_mpipe_close_pollfd);
+
+struct get_mmio_base_param {
+	HV_PTE base;
+};
+
+int gxio_mpipe_get_mmio_base(gxio_mpipe_context_t *context, HV_PTE *base)
+{
+	int __result;
+	struct get_mmio_base_param temp;
+	struct get_mmio_base_param *params = &temp;
+
+	__result =
+	    hv_dev_pread(context->fd, 0, (HV_VirtAddr) params, sizeof(*params),
+			 GXIO_MPIPE_OP_GET_MMIO_BASE);
+	*base = params->base;
+
+	return __result;
+}
+
+EXPORT_SYMBOL(gxio_mpipe_get_mmio_base);
+
+struct check_mmio_offset_param {
+	unsigned long offset;
+	unsigned long size;
+};
+
+int gxio_mpipe_check_mmio_offset(gxio_mpipe_context_t *context,
+				 unsigned long offset, unsigned long size)
+{
+	struct check_mmio_offset_param temp;
+	struct check_mmio_offset_param *params = &temp;
+
+	params->offset = offset;
+	params->size = size;
+
+	return hv_dev_pwrite(context->fd, 0, (HV_VirtAddr) params,
+			     sizeof(*params), GXIO_MPIPE_OP_CHECK_MMIO_OFFSET);
+}
+
+EXPORT_SYMBOL(gxio_mpipe_check_mmio_offset);
diff --git a/kernel/arch/tile/gxio/iorpc_mpipe_info.c b/kernel/arch/tile/gxio/iorpc_mpipe_info.c
new file mode 100644
index 000000000..77019c6e9
--- /dev/null
+++ b/kernel/arch/tile/gxio/iorpc_mpipe_info.c
@@ -0,0 +1,102 @@
+/*
+ * Copyright 2012 Tilera Corporation. All Rights Reserved.
+ *
+ *   This program is free software; you can redistribute it and/or
+ *   modify it under the terms of the GNU General Public License
+ *   as published by the Free Software Foundation, version 2.
+ *
+ *   This program is distributed in the hope that it will be useful, but
+ *   WITHOUT ANY WARRANTY; without even the implied warranty of
+ *   MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
+ *   NON INFRINGEMENT.  See the GNU General Public License for
+ *   more details.
+ */
+
+/* This file is machine-generated; DO NOT EDIT! */
+#include "gxio/iorpc_mpipe_info.h"
+
+struct instance_aux_param {
+	_gxio_mpipe_link_name_t name;
+};
+
+int gxio_mpipe_info_instance_aux(gxio_mpipe_info_context_t *context,
+				 _gxio_mpipe_link_name_t name)
+{
+	struct instance_aux_param temp;
+	struct instance_aux_param *params = &temp;
+
+	params->name = name;
+
+	return hv_dev_pwrite(context->fd, 0, (HV_VirtAddr) params,
+			     sizeof(*params), GXIO_MPIPE_INFO_OP_INSTANCE_AUX);
+}
+
+EXPORT_SYMBOL(gxio_mpipe_info_instance_aux);
+
+struct enumerate_aux_param {
+	_gxio_mpipe_link_name_t name;
+	_gxio_mpipe_link_mac_t mac;
+};
+
+int gxio_mpipe_info_enumerate_aux(gxio_mpipe_info_context_t *context,
+				  unsigned int idx,
+				  _gxio_mpipe_link_name_t *name,
+				  _gxio_mpipe_link_mac_t *mac)
+{
+	int __result;
+	struct enumerate_aux_param temp;
+	struct enumerate_aux_param *params = &temp;
+
+	__result =
+	    hv_dev_pread(context->fd, 0, (HV_VirtAddr) params, sizeof(*params),
+			 (((uint64_t)idx << 32) |
+			  GXIO_MPIPE_INFO_OP_ENUMERATE_AUX));
+	*name = params->name;
+	*mac = params->mac;
+
+	return __result;
+}
+
+EXPORT_SYMBOL(gxio_mpipe_info_enumerate_aux);
+
+struct get_mmio_base_param {
+	HV_PTE base;
+};
+
+int gxio_mpipe_info_get_mmio_base(gxio_mpipe_info_context_t *context,
+				  HV_PTE *base)
+{
+	int __result;
+	struct get_mmio_base_param temp;
+	struct get_mmio_base_param *params = &temp;
+
+	__result =
+	    hv_dev_pread(context->fd, 0, (HV_VirtAddr) params, sizeof(*params),
+			 GXIO_MPIPE_INFO_OP_GET_MMIO_BASE);
+	*base = params->base;
+
+	return __result;
+}
+
+EXPORT_SYMBOL(gxio_mpipe_info_get_mmio_base);
+
+struct check_mmio_offset_param {
+	unsigned long offset;
+	unsigned long size;
+};
+
+int gxio_mpipe_info_check_mmio_offset(gxio_mpipe_info_context_t *context,
+				      unsigned long offset, unsigned long size)
+{
+	struct check_mmio_offset_param temp;
+	struct check_mmio_offset_param *params = &temp;
+
+	params->offset = offset;
+	params->size = size;
+
+	return hv_dev_pwrite(context->fd, 0, (HV_VirtAddr) params,
+			     sizeof(*params),
+			     GXIO_MPIPE_INFO_OP_CHECK_MMIO_OFFSET);
+}
+
+EXPORT_SYMBOL(gxio_mpipe_info_check_mmio_offset);
diff --git a/kernel/arch/tile/gxio/iorpc_trio.c b/kernel/arch/tile/gxio/iorpc_trio.c
new file mode 100644
index 000000000..1d3cedb9a
--- /dev/null
+++ b/kernel/arch/tile/gxio/iorpc_trio.c
@@ -0,0 +1,350 @@
+/*
+ * Copyright 2012 Tilera Corporation. All Rights Reserved.
+ *
+ *   This program is free software; you can redistribute it and/or
+ *   modify it under the terms of the GNU General Public License
+ *   as published by the Free Software Foundation, version 2.
+ *
+ *   This program is distributed in the hope that it will be useful, but
+ *   WITHOUT ANY WARRANTY; without even the implied warranty of
+ *   MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
+ *   NON INFRINGEMENT.  See the GNU General Public License for
+ *   more details.
+ */
+
+/* This file is machine-generated; DO NOT EDIT! */
+#include "gxio/iorpc_trio.h"
+
+struct alloc_asids_param {
+	unsigned int count;
+	unsigned int first;
+	unsigned int flags;
+};
+
+int gxio_trio_alloc_asids(gxio_trio_context_t *context, unsigned int count,
+			  unsigned int first, unsigned int flags)
+{
+	struct alloc_asids_param temp;
+	struct alloc_asids_param *params = &temp;
+
+	params->count = count;
+	params->first = first;
+	params->flags = flags;
+
+	return hv_dev_pwrite(context->fd, 0, (HV_VirtAddr) params,
+			     sizeof(*params), GXIO_TRIO_OP_ALLOC_ASIDS);
+}
+
+EXPORT_SYMBOL(gxio_trio_alloc_asids);
+
+
+struct alloc_memory_maps_param {
+	unsigned int count;
+	unsigned int first;
+	unsigned int flags;
+};
+
+int gxio_trio_alloc_memory_maps(gxio_trio_context_t *context,
+				unsigned int count, unsigned int first,
+				unsigned int flags)
+{
+	struct alloc_memory_maps_param temp;
+	struct alloc_memory_maps_param *params = &temp;
+
+	params->count = count;
+	params->first = first;
+	params->flags = flags;
+
+	return hv_dev_pwrite(context->fd, 0, (HV_VirtAddr) params,
+			     sizeof(*params), GXIO_TRIO_OP_ALLOC_MEMORY_MAPS);
+}
+
+EXPORT_SYMBOL(gxio_trio_alloc_memory_maps);
+
+struct alloc_scatter_queues_param {
+	unsigned int count;
+	unsigned int first;
+	unsigned int flags;
+};
+
+int gxio_trio_alloc_scatter_queues(gxio_trio_context_t *context,
+				   unsigned int count, unsigned int first,
+				   unsigned int flags)
+{
+	struct alloc_scatter_queues_param temp;
+	struct alloc_scatter_queues_param *params = &temp;
+
+	params->count = count;
+	params->first = first;
+	params->flags = flags;
+
+	return hv_dev_pwrite(context->fd, 0, (HV_VirtAddr) params,
+			     sizeof(*params),
+			     GXIO_TRIO_OP_ALLOC_SCATTER_QUEUES);
+}
+
+EXPORT_SYMBOL(gxio_trio_alloc_scatter_queues);
+
+struct alloc_pio_regions_param {
+	unsigned int count;
+	unsigned int first;
+	unsigned int flags;
+};
+
+int gxio_trio_alloc_pio_regions(gxio_trio_context_t *context,
+				unsigned int count, unsigned int first,
+				unsigned int flags)
+{
+	struct alloc_pio_regions_param temp;
+	struct alloc_pio_regions_param *params = &temp;
+
+	params->count = count;
+	params->first = first;
+	params->flags = flags;
+
+	return hv_dev_pwrite(context->fd, 0, (HV_VirtAddr) params,
+			     sizeof(*params), GXIO_TRIO_OP_ALLOC_PIO_REGIONS);
+}
+
+EXPORT_SYMBOL(gxio_trio_alloc_pio_regions);
+
+struct init_pio_region_aux_param {
+	unsigned int pio_region;
+	unsigned int mac;
+	uint32_t bus_address_hi;
+	unsigned int flags;
+};
+
+int gxio_trio_init_pio_region_aux(gxio_trio_context_t *context,
+				  unsigned int pio_region, unsigned int mac,
+				  uint32_t bus_address_hi, unsigned int flags)
+{
+	struct init_pio_region_aux_param temp;
+	struct init_pio_region_aux_param *params = &temp;
+
+	params->pio_region = pio_region;
+	params->mac = mac;
+	params->bus_address_hi = bus_address_hi;
+	params->flags = flags;
+
+	return hv_dev_pwrite(context->fd, 0, (HV_VirtAddr) params,
+			     sizeof(*params), GXIO_TRIO_OP_INIT_PIO_REGION_AUX);
+}
+
+EXPORT_SYMBOL(gxio_trio_init_pio_region_aux);
+
+
+struct init_memory_map_mmu_aux_param {
+	unsigned int map;
+	unsigned long va;
+	uint64_t size;
+	unsigned int asid;
+	unsigned int mac;
+	uint64_t bus_address;
+	unsigned int node;
+	unsigned int order_mode;
+};
+
+int gxio_trio_init_memory_map_mmu_aux(gxio_trio_context_t *context,
+				      unsigned int map, unsigned long va,
+				      uint64_t size, unsigned int asid,
+				      unsigned int mac, uint64_t bus_address,
+				      unsigned int node,
+				      unsigned int order_mode)
+{
+	struct init_memory_map_mmu_aux_param temp;
+	struct init_memory_map_mmu_aux_param *params = &temp;
+
+	params->map = map;
+	params->va = va;
+	params->size = size;
+	params->asid = asid;
+	params->mac = mac;
+	params->bus_address = bus_address;
+	params->node = node;
+	params->order_mode = order_mode;
+
+	return hv_dev_pwrite(context->fd, 0, (HV_VirtAddr) params,
+			     sizeof(*params),
+			     GXIO_TRIO_OP_INIT_MEMORY_MAP_MMU_AUX);
+}
+
+EXPORT_SYMBOL(gxio_trio_init_memory_map_mmu_aux);
+
+struct get_port_property_param {
+	struct pcie_trio_ports_property trio_ports;
+};
+
+int gxio_trio_get_port_property(gxio_trio_context_t *context,
+				struct pcie_trio_ports_property *trio_ports)
+{
+	int __result;
+	struct get_port_property_param temp;
+	struct get_port_property_param *params = &temp;
+
+	__result =
+	    hv_dev_pread(context->fd, 0, (HV_VirtAddr) params, sizeof(*params),
+			 GXIO_TRIO_OP_GET_PORT_PROPERTY);
+	*trio_ports = params->trio_ports;
+
+	return __result;
+}
+
+EXPORT_SYMBOL(gxio_trio_get_port_property);
+
+struct config_legacy_intr_param {
+	union iorpc_interrupt interrupt;
+	unsigned int mac;
+	unsigned int intx;
+};
+
+int gxio_trio_config_legacy_intr(gxio_trio_context_t *context, int inter_x,
+				 int inter_y, int inter_ipi, int inter_event,
+				 unsigned int mac, unsigned int intx)
+{
+	struct config_legacy_intr_param temp;
+	struct config_legacy_intr_param *params = &temp;
+
+	params->interrupt.kernel.x = inter_x;
+	params->interrupt.kernel.y = inter_y;
+	params->interrupt.kernel.ipi = inter_ipi;
+	params->interrupt.kernel.event = inter_event;
+	params->mac = mac;
+	params->intx = intx;
+
+	return hv_dev_pwrite(context->fd, 0, (HV_VirtAddr) params,
+			     sizeof(*params), GXIO_TRIO_OP_CONFIG_LEGACY_INTR);
+}
+
+EXPORT_SYMBOL(gxio_trio_config_legacy_intr);
+
+struct config_msi_intr_param {
+	union iorpc_interrupt interrupt;
+	unsigned int mac;
+	unsigned int mem_map;
+	uint64_t mem_map_base;
+	uint64_t mem_map_limit;
+	unsigned int asid;
+};
+
+int gxio_trio_config_msi_intr(gxio_trio_context_t *context, int inter_x,
+			      int inter_y, int inter_ipi, int inter_event,
+			      unsigned int mac, unsigned int mem_map,
+			      uint64_t mem_map_base, uint64_t mem_map_limit,
+			      unsigned int asid)
+{
+	struct config_msi_intr_param temp;
+	struct config_msi_intr_param *params = &temp;
+
+	params->interrupt.kernel.x = inter_x;
+	params->interrupt.kernel.y = inter_y;
+	params->interrupt.kernel.ipi = inter_ipi;
+	params->interrupt.kernel.event = inter_event;
+	params->mac = mac;
+	params->mem_map = mem_map;
+	params->mem_map_base = mem_map_base;
+	params->mem_map_limit = mem_map_limit;
+	params->asid = asid;
+
+	return hv_dev_pwrite(context->fd, 0, (HV_VirtAddr) params,
+			     sizeof(*params), GXIO_TRIO_OP_CONFIG_MSI_INTR);
+}
+
+EXPORT_SYMBOL(gxio_trio_config_msi_intr);
+
+
+struct set_mps_mrs_param {
+	uint16_t mps;
+	uint16_t mrs;
+	unsigned int mac;
+};
+
+int gxio_trio_set_mps_mrs(gxio_trio_context_t *context, uint16_t mps,
+			  uint16_t mrs, unsigned int mac)
+{
+	struct set_mps_mrs_param temp;
+	struct set_mps_mrs_param *params = &temp;
+
+	params->mps = mps;
+	params->mrs = mrs;
+	params->mac = mac;
+
+	return hv_dev_pwrite(context->fd, 0, (HV_VirtAddr) params,
+			     sizeof(*params), GXIO_TRIO_OP_SET_MPS_MRS);
+}
+
+EXPORT_SYMBOL(gxio_trio_set_mps_mrs);
+
+struct force_rc_link_up_param {
+	unsigned int mac;
+};
+
+int gxio_trio_force_rc_link_up(gxio_trio_context_t *context, unsigned int mac)
+{
+	struct force_rc_link_up_param temp;
+	struct force_rc_link_up_param *params = &temp;
+
+	params->mac = mac;
+
+	return hv_dev_pwrite(context->fd, 0, (HV_VirtAddr) params,
+			     sizeof(*params), GXIO_TRIO_OP_FORCE_RC_LINK_UP);
+}
+
+EXPORT_SYMBOL(gxio_trio_force_rc_link_up);
+
+struct force_ep_link_up_param {
+	unsigned int mac;
+};
+
+int gxio_trio_force_ep_link_up(gxio_trio_context_t *context, unsigned int mac)
+{
+	struct force_ep_link_up_param temp;
+	struct force_ep_link_up_param *params = &temp;
+
+	params->mac = mac;
+
+	return hv_dev_pwrite(context->fd, 0, (HV_VirtAddr) params,
+			     sizeof(*params), GXIO_TRIO_OP_FORCE_EP_LINK_UP);
+}
+
+EXPORT_SYMBOL(gxio_trio_force_ep_link_up);
+
+struct get_mmio_base_param {
+	HV_PTE base;
+};
+
+int gxio_trio_get_mmio_base(gxio_trio_context_t *context, HV_PTE *base)
+{
+	int __result;
+	struct get_mmio_base_param temp;
+	struct get_mmio_base_param *params = &temp;
+
+	__result =
+	    hv_dev_pread(context->fd, 0, (HV_VirtAddr) params, sizeof(*params),
+			 GXIO_TRIO_OP_GET_MMIO_BASE);
+	*base = params->base;
+
+	return __result;
+}
+
+EXPORT_SYMBOL(gxio_trio_get_mmio_base);
+
+struct check_mmio_offset_param {
+	unsigned long offset;
+	unsigned long size;
+};
+
+int gxio_trio_check_mmio_offset(gxio_trio_context_t *context,
+				unsigned long offset, unsigned long size)
+{
+	struct check_mmio_offset_param temp;
+	struct check_mmio_offset_param *params = &temp;
+
+	params->offset = offset;
+	params->size = size;
+
+	return hv_dev_pwrite(context->fd, 0, (HV_VirtAddr) params,
+			     sizeof(*params), GXIO_TRIO_OP_CHECK_MMIO_OFFSET);
+}
+
+EXPORT_SYMBOL(gxio_trio_check_mmio_offset);
diff --git a/kernel/arch/tile/gxio/iorpc_uart.c b/kernel/arch/tile/gxio/iorpc_uart.c
new file mode 100644
index 000000000..b9a6d6193
--- /dev/null
+++ b/kernel/arch/tile/gxio/iorpc_uart.c
@@ -0,0 +1,77 @@
+/*
+ * Copyright 2013 Tilera Corporation. All Rights Reserved.
+ *
+ *   This program is free software; you can redistribute it and/or
+ *   modify it under the terms of the GNU General Public License
+ *   as published by the Free Software Foundation, version 2.
+ *
+ *   This program is distributed in the hope that it will be useful, but
+ *   WITHOUT ANY WARRANTY; without even the implied warranty of
+ *   MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
+ *   NON INFRINGEMENT.  See the GNU General Public License for
+ *   more details.
+ */
+
+/* This file is machine-generated; DO NOT EDIT! */
+#include "gxio/iorpc_uart.h"
+
+struct cfg_interrupt_param {
+	union iorpc_interrupt interrupt;
+};
+
+int gxio_uart_cfg_interrupt(gxio_uart_context_t *context, int inter_x,
+			    int inter_y, int inter_ipi, int inter_event)
+{
+	struct cfg_interrupt_param temp;
+	struct cfg_interrupt_param *params = &temp;
+
+	params->interrupt.kernel.x = inter_x;
+	params->interrupt.kernel.y = inter_y;
+	params->interrupt.kernel.ipi = inter_ipi;
+	params->interrupt.kernel.event = inter_event;
+
+	return hv_dev_pwrite(context->fd, 0, (HV_VirtAddr) params,
+			     sizeof(*params), GXIO_UART_OP_CFG_INTERRUPT);
+}
+
+EXPORT_SYMBOL(gxio_uart_cfg_interrupt);
+
+struct get_mmio_base_param {
+	HV_PTE base;
+};
+
+int gxio_uart_get_mmio_base(gxio_uart_context_t *context, HV_PTE *base)
+{
+	int __result;
+	struct get_mmio_base_param temp;
+	struct get_mmio_base_param *params = &temp;
+
+	__result =
+	    hv_dev_pread(context->fd, 0, (HV_VirtAddr) params, sizeof(*params),
+			 GXIO_UART_OP_GET_MMIO_BASE);
+	*base = params->base;
+
+	return __result;
+}
+
+EXPORT_SYMBOL(gxio_uart_get_mmio_base);
+
+struct check_mmio_offset_param {
+	unsigned long offset;
+	unsigned long size;
+};
+
+int gxio_uart_check_mmio_offset(gxio_uart_context_t *context,
+				unsigned long offset, unsigned long size)
+{
+	struct check_mmio_offset_param temp;
+	struct check_mmio_offset_param *params = &temp;
+
+	params->offset = offset;
+	params->size = size;
+
+	return hv_dev_pwrite(context->fd, 0, (HV_VirtAddr) params,
+			     sizeof(*params), GXIO_UART_OP_CHECK_MMIO_OFFSET);
+}
+
+EXPORT_SYMBOL(gxio_uart_check_mmio_offset);
diff --git a/kernel/arch/tile/gxio/iorpc_usb_host.c b/kernel/arch/tile/gxio/iorpc_usb_host.c
new file mode 100644
index 000000000..9c820073b
--- /dev/null
+++ b/kernel/arch/tile/gxio/iorpc_usb_host.c
@@ -0,0 +1,99 @@
+/*
+ * Copyright 2012 Tilera Corporation. All Rights Reserved.
+ *
+ *   This program is free software; you can redistribute it and/or
+ *   modify it under the terms of the GNU General Public License
+ *   as published by the Free Software Foundation, version 2.
+ *
+ *   This program is distributed in the hope that it will be useful, but
+ *   WITHOUT ANY WARRANTY; without even the implied warranty of
+ *   MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
+ *   NON INFRINGEMENT.  See the GNU General Public License for
+ *   more details.
+ */
+
+/* This file is machine-generated; DO NOT EDIT! */
+#include "gxio/iorpc_usb_host.h"
+
+struct cfg_interrupt_param {
+	union iorpc_interrupt interrupt;
+};
+
+int gxio_usb_host_cfg_interrupt(gxio_usb_host_context_t *context, int inter_x,
+				int inter_y, int inter_ipi, int inter_event)
+{
+	struct cfg_interrupt_param temp;
+	struct cfg_interrupt_param *params = &temp;
+
+	params->interrupt.kernel.x = inter_x;
+	params->interrupt.kernel.y = inter_y;
+	params->interrupt.kernel.ipi = inter_ipi;
+	params->interrupt.kernel.event = inter_event;
+
+	return hv_dev_pwrite(context->fd, 0, (HV_VirtAddr) params,
+			     sizeof(*params), GXIO_USB_HOST_OP_CFG_INTERRUPT);
+}
+
+EXPORT_SYMBOL(gxio_usb_host_cfg_interrupt);
+
+struct register_client_memory_param {
+	HV_PTE pte;
+	unsigned int flags;
+};
+
+int gxio_usb_host_register_client_memory(gxio_usb_host_context_t *context,
+					 HV_PTE pte, unsigned int flags)
+{
+	struct register_client_memory_param temp;
+	struct register_client_memory_param *params = &temp;
+
+	params->pte = pte;
+	params->flags = flags;
+
+	return hv_dev_pwrite(context->fd, 0, (HV_VirtAddr) params,
+			     sizeof(*params),
+			     GXIO_USB_HOST_OP_REGISTER_CLIENT_MEMORY);
+}
+
+EXPORT_SYMBOL(gxio_usb_host_register_client_memory);
+
+struct get_mmio_base_param {
+	HV_PTE base;
+};
+
+int gxio_usb_host_get_mmio_base(gxio_usb_host_context_t *context, HV_PTE *base)
+{
+	int __result;
+	struct get_mmio_base_param temp;
+	struct get_mmio_base_param *params = &temp;
+
+	__result =
+	    hv_dev_pread(context->fd, 0, (HV_VirtAddr) params, sizeof(*params),
+			 GXIO_USB_HOST_OP_GET_MMIO_BASE);
+	*base = params->base;
+
+	return __result;
+}
+
+EXPORT_SYMBOL(gxio_usb_host_get_mmio_base);
+
+struct check_mmio_offset_param {
+	unsigned long offset;
+	unsigned long size;
+};
+
+int gxio_usb_host_check_mmio_offset(gxio_usb_host_context_t *context,
+				    unsigned long offset, unsigned long size)
+{
+	struct check_mmio_offset_param temp;
+	struct check_mmio_offset_param *params = &temp;
+
+	params->offset = offset;
+	params->size = size;
+
+	return hv_dev_pwrite(context->fd, 0, (HV_VirtAddr) params,
+			     sizeof(*params),
+			     GXIO_USB_HOST_OP_CHECK_MMIO_OFFSET);
+}
+
+EXPORT_SYMBOL(gxio_usb_host_check_mmio_offset);
diff --git a/kernel/arch/tile/gxio/kiorpc.c b/kernel/arch/tile/gxio/kiorpc.c
new file mode 100644
index 000000000..c8096aa5a
--- /dev/null
+++ b/kernel/arch/tile/gxio/kiorpc.c
@@ -0,0 +1,61 @@
+/*
+ * Copyright 2012 Tilera Corporation. All Rights Reserved.
+ *
+ *   This program is free software; you can redistribute it and/or
+ *   modify it under the terms of the GNU General Public License
+ *   as published by the Free Software Foundation, version 2.
+ *
+ *   This program is distributed in the hope that it will be useful, but
+ *   WITHOUT ANY WARRANTY; without even the implied warranty of
+ *   MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
+ *   NON INFRINGEMENT.  See the GNU General Public License for
+ *   more details.
+ *
+ * TILE-Gx IORPC support for kernel I/O drivers.
+ */
+
+#include <linux/mmzone.h>
+#include <linux/module.h>
+#include <linux/io.h>
+#include <gxio/iorpc_globals.h>
+#include <gxio/kiorpc.h>
+
+#ifdef DEBUG_IORPC
+#define TRACE(FMT, ...) pr_info(SIMPLE_MSG_LINE FMT, ## __VA_ARGS__)
+#else
+#define TRACE(...)
+#endif
+
+/* Create kernel-VA-space MMIO mapping for an on-chip IO device. */
+void __iomem *iorpc_ioremap(int hv_fd, resource_size_t offset,
+			    unsigned long size)
+{
+	pgprot_t mmio_base, prot = { 0 };
+	unsigned long pfn;
+	int err;
+
+	/* Look up the shim's lotar and base PA. */
+	err = __iorpc_get_mmio_base(hv_fd, &mmio_base);
+	if (err) {
+		TRACE("get_mmio_base() failure: %d\n", err);
+		return NULL;
+	}
+
+	/* Make sure the HV driver approves of our offset and size. */
+	err = __iorpc_check_mmio_offset(hv_fd, offset, size);
+	if (err) {
+		TRACE("check_mmio_offset() failure: %d\n", err);
+		return NULL;
+	}
+
+	/*
+	 * mmio_base contains a base pfn and homing coordinates.  Turn
+	 * it into an MMIO pgprot and offset pfn.
+	 */
+	prot = hv_pte_set_lotar(prot, hv_pte_get_lotar(mmio_base));
+	pfn = pte_pfn(mmio_base) + PFN_DOWN(offset);
+
+	return ioremap_prot(PFN_PHYS(pfn), size, prot);
+}
+
+EXPORT_SYMBOL(iorpc_ioremap);
diff --git a/kernel/arch/tile/gxio/mpipe.c b/kernel/arch/tile/gxio/mpipe.c
new file mode 100644
index 000000000..ee186e13d
--- /dev/null
+++ b/kernel/arch/tile/gxio/mpipe.c
@@ -0,0 +1,609 @@
+/*
+ * Copyright 2012 Tilera Corporation. All Rights Reserved.
+ *
+ *   This program is free software; you can redistribute it and/or
+ *   modify it under the terms of the GNU General Public License
+ *   as published by the Free Software Foundation, version 2.
+ *
+ *   This program is distributed in the hope that it will be useful, but
+ *   WITHOUT ANY WARRANTY; without even the implied warranty of
+ *   MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
+ *   NON INFRINGEMENT.  See the GNU General Public License for
+ *   more details.
+ */
+
+/*
+ * Implementation of mpipe gxio calls.
+ */
+
+#include <linux/errno.h>
+#include <linux/io.h>
+#include <linux/module.h>
+
+#include <gxio/iorpc_globals.h>
+#include <gxio/iorpc_mpipe.h>
+#include <gxio/iorpc_mpipe_info.h>
+#include <gxio/kiorpc.h>
+#include <gxio/mpipe.h>
+
+/* HACK: Avoid pointless "shadow" warnings. */
+#define link link_shadow
+
+/**
+ * strscpy - Copy a C-string into a sized buffer, but only if it fits
+ * @dest: Where to copy the string to
+ * @src: Where to copy the string from
+ * @size: size of destination buffer
+ *
+ * Use this routine to avoid copying too-long strings.
+ * The routine returns the total number of bytes copied
+ * (including the trailing NUL) or zero if the buffer wasn't
+ * big enough.  To ensure that programmers pay attention
+ * to the return code, the destination has a single NUL
+ * written at the front (if size is non-zero) when the
+ * buffer is not big enough.
+ */
+static size_t strscpy(char *dest, const char *src, size_t size)
+{
+	size_t len = strnlen(src, size) + 1;
+	if (len > size) {
+		if (size)
+			dest[0] = '\0';
+		return 0;
+	}
+	memcpy(dest, src, len);
+	return len;
+}
+
+int gxio_mpipe_init(gxio_mpipe_context_t *context, unsigned int mpipe_index)
+{
+	char file[32];
+
+	int fd;
+	int i;
+
+	if (mpipe_index >= GXIO_MPIPE_INSTANCE_MAX)
+		return -EINVAL;
+
+	snprintf(file, sizeof(file), "mpipe/%d/iorpc", mpipe_index);
+	fd = hv_dev_open((HV_VirtAddr) file, 0);
+
+	context->fd = fd;
+
+	if (fd < 0) {
+		if (fd >= GXIO_ERR_MIN && fd <= GXIO_ERR_MAX)
+			return fd;
+		else
+			return -ENODEV;
+	}
+
+	/* Map in the MMIO space. */
+	context->mmio_cfg_base = (void __force *)
+		iorpc_ioremap(fd, HV_MPIPE_CONFIG_MMIO_OFFSET,
+			      HV_MPIPE_CONFIG_MMIO_SIZE);
+	if (context->mmio_cfg_base == NULL)
+		goto cfg_failed;
+
+	context->mmio_fast_base = (void __force *)
+		iorpc_ioremap(fd, HV_MPIPE_FAST_MMIO_OFFSET,
+			      HV_MPIPE_FAST_MMIO_SIZE);
+	if (context->mmio_fast_base == NULL)
+		goto fast_failed;
+
+	/* Initialize the stacks. */
+	for (i = 0; i < 8; i++)
+		context->__stacks.stacks[i] = 255;
+
+	context->instance = mpipe_index;
+
+	return 0;
+
+      fast_failed:
+	iounmap((void __force __iomem *)(context->mmio_cfg_base));
+      cfg_failed:
+	hv_dev_close(context->fd);
+	context->fd = -1;
+	return -ENODEV;
+}
+
+EXPORT_SYMBOL_GPL(gxio_mpipe_init);
+
+int gxio_mpipe_destroy(gxio_mpipe_context_t *context)
+{
+	iounmap((void __force __iomem *)(context->mmio_cfg_base));
+	iounmap((void __force __iomem *)(context->mmio_fast_base));
+	return hv_dev_close(context->fd);
+}
+
+EXPORT_SYMBOL_GPL(gxio_mpipe_destroy);
+
+static int16_t gxio_mpipe_buffer_sizes[8] =
+	{ 128, 256, 512, 1024, 1664, 4096, 10368, 16384 };
+
+gxio_mpipe_buffer_size_enum_t gxio_mpipe_buffer_size_to_buffer_size_enum(size_t
+									 size)
+{
+	int i;
+	for (i = 0; i < 7; i++)
+		if (size <= gxio_mpipe_buffer_sizes[i])
+			break;
+	return i;
+}
+
+EXPORT_SYMBOL_GPL(gxio_mpipe_buffer_size_to_buffer_size_enum);
+
+size_t gxio_mpipe_buffer_size_enum_to_buffer_size(gxio_mpipe_buffer_size_enum_t
+						  buffer_size_enum)
+{
+	if (buffer_size_enum > 7)
+		buffer_size_enum = 7;
+
+	return gxio_mpipe_buffer_sizes[buffer_size_enum];
+}
+
+EXPORT_SYMBOL_GPL(gxio_mpipe_buffer_size_enum_to_buffer_size);
+
+size_t gxio_mpipe_calc_buffer_stack_bytes(unsigned long buffers)
+{
+	const int BUFFERS_PER_LINE = 12;
+
+	/* Count the number of cachlines. */
+	unsigned long lines =
+		(buffers + BUFFERS_PER_LINE - 1) / BUFFERS_PER_LINE;
+
+	/* Convert to bytes. */
+	return lines * CHIP_L2_LINE_SIZE();
+}
+
+EXPORT_SYMBOL_GPL(gxio_mpipe_calc_buffer_stack_bytes);
+
+int gxio_mpipe_init_buffer_stack(gxio_mpipe_context_t *context,
+				 unsigned int stack,
+				 gxio_mpipe_buffer_size_enum_t
+				 buffer_size_enum, void *mem, size_t mem_size,
+				 unsigned int mem_flags)
+{
+	int result;
+
+	memset(mem, 0, mem_size);
+
+	result = gxio_mpipe_init_buffer_stack_aux(context, mem, mem_size,
+						  mem_flags, stack,
+						  buffer_size_enum);
+	if (result < 0)
+		return result;
+
+	/* Save the stack. */
+	context->__stacks.stacks[buffer_size_enum] = stack;
+
+	return 0;
+}
+
+EXPORT_SYMBOL_GPL(gxio_mpipe_init_buffer_stack);
+
+int gxio_mpipe_init_notif_ring(gxio_mpipe_context_t *context,
+			       unsigned int ring,
+			       void *mem, size_t mem_size,
+			       unsigned int mem_flags)
+{
+	return gxio_mpipe_init_notif_ring_aux(context, mem, mem_size,
+					      mem_flags, ring);
+}
+
+EXPORT_SYMBOL_GPL(gxio_mpipe_init_notif_ring);
+
+int gxio_mpipe_init_notif_group_and_buckets(gxio_mpipe_context_t *context,
+					    unsigned int group,
+					    unsigned int ring,
+					    unsigned int num_rings,
+					    unsigned int bucket,
+					    unsigned int num_buckets,
+					    gxio_mpipe_bucket_mode_t mode)
+{
+	int i;
+	int result;
+
+	gxio_mpipe_bucket_info_t bucket_info = { {
+						  .group = group,
+						  .mode = mode,
+						  }
+	};
+
+	gxio_mpipe_notif_group_bits_t bits = { {0} };
+
+	for (i = 0; i < num_rings; i++)
+		gxio_mpipe_notif_group_add_ring(&bits, ring + i);
+
+	result = gxio_mpipe_init_notif_group(context, group, bits);
+	if (result != 0)
+		return result;
+
+	for (i = 0; i < num_buckets; i++) {
+		bucket_info.notifring = ring + (i % num_rings);
+
+		result = gxio_mpipe_init_bucket(context, bucket + i,
+						bucket_info);
+		if (result != 0)
+			return result;
+	}
+
+	return 0;
+}
+
+EXPORT_SYMBOL_GPL(gxio_mpipe_init_notif_group_and_buckets);
+
+int gxio_mpipe_init_edma_ring(gxio_mpipe_context_t *context,
+			      unsigned int ring, unsigned int channel,
+			      void *mem, size_t mem_size,
+			      unsigned int mem_flags)
+{
+	memset(mem, 0, mem_size);
+
+	return gxio_mpipe_init_edma_ring_aux(context, mem, mem_size, mem_flags,
+					     ring, channel);
+}
+
+EXPORT_SYMBOL_GPL(gxio_mpipe_init_edma_ring);
+
+void gxio_mpipe_rules_init(gxio_mpipe_rules_t *rules,
+			   gxio_mpipe_context_t *context)
+{
+	rules->context = context;
+	memset(&rules->list, 0, sizeof(rules->list));
+}
+
+EXPORT_SYMBOL_GPL(gxio_mpipe_rules_init);
+
+int gxio_mpipe_rules_begin(gxio_mpipe_rules_t *rules,
+			   unsigned int bucket, unsigned int num_buckets,
+			   gxio_mpipe_rules_stacks_t *stacks)
+{
+	int i;
+	int stack = 255;
+
+	gxio_mpipe_rules_list_t *list = &rules->list;
+
+	/* Current rule. */
+	gxio_mpipe_rules_rule_t *rule =
+		(gxio_mpipe_rules_rule_t *) (list->rules + list->head);
+
+	unsigned int head = list->tail;
+
+	/*
+	 * Align next rule properly.
+	 *Note that "dmacs_and_vlans" will also be aligned.
+	 */
+	unsigned int pad = 0;
+	while (((head + pad) % __alignof__(gxio_mpipe_rules_rule_t)) != 0)
+		pad++;
+
+	/*
+	 * Verify room.
+	 * ISSUE: Mark rules as broken on error?
+	 */
+	if (head + pad + sizeof(*rule) >= sizeof(list->rules))
+		return GXIO_MPIPE_ERR_RULES_FULL;
+
+	/* Verify num_buckets is a power of 2. */
+	if (__builtin_popcount(num_buckets) != 1)
+		return GXIO_MPIPE_ERR_RULES_INVALID;
+
+	/* Add padding to previous rule. */
+	rule->size += pad;
+
+	/* Start a new rule. */
+	list->head = head + pad;
+
+	rule = (gxio_mpipe_rules_rule_t *) (list->rules + list->head);
+
+	/* Default some values. */
+	rule->headroom = 2;
+	rule->tailroom = 0;
+	rule->capacity = 16384;
+
+	/* Save the bucket info. */
+	rule->bucket_mask = num_buckets - 1;
+	rule->bucket_first = bucket;
+
+	for (i = 8 - 1; i >= 0; i--) {
+		int maybe =
+			stacks ? stacks->stacks[i] : rules->context->__stacks.
+			stacks[i];
+		if (maybe != 255)
+			stack = maybe;
+		rule->stacks.stacks[i] = stack;
+	}
+
+	if (stack == 255)
+		return GXIO_MPIPE_ERR_RULES_INVALID;
+
+	/* NOTE: Only entries at the end of the array can be 255. */
+	for (i = 8 - 1; i > 0; i--) {
+		if (rule->stacks.stacks[i] == 255) {
+			rule->stacks.stacks[i] = stack;
+			rule->capacity =
+				gxio_mpipe_buffer_size_enum_to_buffer_size(i -
+									   1);
+		}
+	}
+
+	rule->size = sizeof(*rule);
+	list->tail = list->head + rule->size;
+
+	return 0;
+}
+
+EXPORT_SYMBOL_GPL(gxio_mpipe_rules_begin);
+
+int gxio_mpipe_rules_add_channel(gxio_mpipe_rules_t *rules,
+				 unsigned int channel)
+{
+	gxio_mpipe_rules_list_t *list = &rules->list;
+
+	gxio_mpipe_rules_rule_t *rule =
+		(gxio_mpipe_rules_rule_t *) (list->rules + list->head);
+
+	/* Verify channel. */
+	if (channel >= 32)
+		return GXIO_MPIPE_ERR_RULES_INVALID;
+
+	/* Verify begun. */
+	if (list->tail == 0)
+		return GXIO_MPIPE_ERR_RULES_EMPTY;
+
+	rule->channel_bits |= (1UL << channel);
+
+	return 0;
+}
+
+EXPORT_SYMBOL_GPL(gxio_mpipe_rules_add_channel);
+
+int gxio_mpipe_rules_set_headroom(gxio_mpipe_rules_t *rules, uint8_t headroom)
+{
+	gxio_mpipe_rules_list_t *list = &rules->list;
+
+	gxio_mpipe_rules_rule_t *rule =
+		(gxio_mpipe_rules_rule_t *) (list->rules + list->head);
+
+	/* Verify begun. */
+	if (list->tail == 0)
+		return GXIO_MPIPE_ERR_RULES_EMPTY;
+
+	rule->headroom = headroom;
+
+	return 0;
+}
+
+EXPORT_SYMBOL_GPL(gxio_mpipe_rules_set_headroom);
+
+int gxio_mpipe_rules_commit(gxio_mpipe_rules_t *rules)
+{
+	gxio_mpipe_rules_list_t *list = &rules->list;
+	unsigned int size =
+		offsetof(gxio_mpipe_rules_list_t, rules) + list->tail;
+	return gxio_mpipe_commit_rules(rules->context, list, size);
+}
+
+EXPORT_SYMBOL_GPL(gxio_mpipe_rules_commit);
+
+int gxio_mpipe_iqueue_init(gxio_mpipe_iqueue_t *iqueue,
+			   gxio_mpipe_context_t *context,
+			   unsigned int ring,
+			   void *mem, size_t mem_size, unsigned int mem_flags)
+{
+	/* The init call below will verify that "mem_size" is legal. */
+	unsigned int num_entries = mem_size / sizeof(gxio_mpipe_idesc_t);
+
+	iqueue->context = context;
+	iqueue->idescs = (gxio_mpipe_idesc_t *)mem;
+	iqueue->ring = ring;
+	iqueue->num_entries = num_entries;
+	iqueue->mask_num_entries = num_entries - 1;
+	iqueue->log2_num_entries = __builtin_ctz(num_entries);
+	iqueue->head = 1;
+#ifdef __BIG_ENDIAN__
+	iqueue->swapped = 0;
+#endif
+
+	/* Initialize the "tail". */
+	__gxio_mmio_write(mem, iqueue->head);
+
+	return gxio_mpipe_init_notif_ring(context, ring, mem, mem_size,
+					  mem_flags);
+}
+
+EXPORT_SYMBOL_GPL(gxio_mpipe_iqueue_init);
+
+int gxio_mpipe_equeue_init(gxio_mpipe_equeue_t *equeue,
+			   gxio_mpipe_context_t *context,
+			   unsigned int ering,
+			   unsigned int channel,
+			   void *mem, unsigned int mem_size,
+			   unsigned int mem_flags)
+{
+	/* The init call below will verify that "mem_size" is legal. */
+	unsigned int num_entries = mem_size / sizeof(gxio_mpipe_edesc_t);
+
+	/* Offset used to read number of completed commands. */
+	MPIPE_EDMA_POST_REGION_ADDR_t offset;
+
+	int result = gxio_mpipe_init_edma_ring(context, ering, channel,
+					       mem, mem_size, mem_flags);
+	if (result < 0)
+		return result;
+
+	memset(equeue, 0, sizeof(*equeue));
+
+	offset.word = 0;
+	offset.region =
+		MPIPE_MMIO_ADDR__REGION_VAL_EDMA -
+		MPIPE_MMIO_ADDR__REGION_VAL_IDMA;
+	offset.ring = ering;
+
+	__gxio_dma_queue_init(&equeue->dma_queue,
+			      context->mmio_fast_base + offset.word,
+			      num_entries);
+	equeue->edescs = mem;
+	equeue->mask_num_entries = num_entries - 1;
+	equeue->log2_num_entries = __builtin_ctz(num_entries);
+	equeue->context = context;
+	equeue->ering = ering;
+	equeue->channel = channel;
+
+	return 0;
+}
+
+EXPORT_SYMBOL_GPL(gxio_mpipe_equeue_init);
+
+int gxio_mpipe_set_timestamp(gxio_mpipe_context_t *context,
+			     const struct timespec64 *ts)
+{
+	cycles_t cycles = get_cycles();
+	return gxio_mpipe_set_timestamp_aux(context, (uint64_t)ts->tv_sec,
+					    (uint64_t)ts->tv_nsec,
+					    (uint64_t)cycles);
+}
+EXPORT_SYMBOL_GPL(gxio_mpipe_set_timestamp);
+
+int gxio_mpipe_get_timestamp(gxio_mpipe_context_t *context,
+			     struct timespec64 *ts)
+{
+	int ret;
+	cycles_t cycles_prev, cycles_now, clock_rate;
+	cycles_prev = get_cycles();
+	ret = gxio_mpipe_get_timestamp_aux(context, (uint64_t *)&ts->tv_sec,
+					   (uint64_t *)&ts->tv_nsec,
+					   (uint64_t *)&cycles_now);
+	if (ret < 0) {
+		return ret;
+	}
+
+	clock_rate = get_clock_rate();
+	ts->tv_nsec -= (cycles_now - cycles_prev) * 1000000000LL / clock_rate;
+	if (ts->tv_nsec < 0) {
+		ts->tv_nsec += 1000000000LL;
+		ts->tv_sec -= 1;
+	}
+	return ret;
+}
+EXPORT_SYMBOL_GPL(gxio_mpipe_get_timestamp);
+
+int gxio_mpipe_adjust_timestamp(gxio_mpipe_context_t *context, int64_t delta)
+{
+	return gxio_mpipe_adjust_timestamp_aux(context, delta);
+}
+EXPORT_SYMBOL_GPL(gxio_mpipe_adjust_timestamp);
+
+/* Get our internal context used for link name access.  This context is
+ *  special in that it is not associated with an mPIPE service domain.
+ */
+static gxio_mpipe_context_t *_gxio_get_link_context(void)
+{
+	static gxio_mpipe_context_t context;
+	static gxio_mpipe_context_t *contextp;
+	static int tried_open = 0;
+	static DEFINE_MUTEX(mutex);
+
+	mutex_lock(&mutex);
+
+	if (!tried_open) {
+		int i = 0;
+		tried_open = 1;
+
+		/*
+		 * "4" here is the maximum possible number of mPIPE shims; it's
+		 * an exaggeration but we shouldn't ever go beyond 2 anyway.
+		 */
+		for (i = 0; i < 4; i++) {
+			char file[80];
+
+			snprintf(file, sizeof(file), "mpipe/%d/iorpc_info", i);
+			context.fd = hv_dev_open((HV_VirtAddr) file, 0);
+			if (context.fd < 0)
+				continue;
+
+			contextp = &context;
+			break;
+		}
+	}
+
+	mutex_unlock(&mutex);
+
+	return contextp;
+}
+
+int gxio_mpipe_link_instance(const char *link_name)
+{
+	_gxio_mpipe_link_name_t name;
+	gxio_mpipe_context_t *context = _gxio_get_link_context();
+
+	if (!context)
+		return GXIO_ERR_NO_DEVICE;
+
+	if (strscpy(name.name, link_name, sizeof(name.name)) == 0)
+		return GXIO_ERR_NO_DEVICE;
+
+	return gxio_mpipe_info_instance_aux(context, name);
+}
+EXPORT_SYMBOL_GPL(gxio_mpipe_link_instance);
+
+int gxio_mpipe_link_enumerate_mac(int idx, char *link_name, uint8_t *link_mac)
+{
+	int rv;
+	_gxio_mpipe_link_name_t name;
+	_gxio_mpipe_link_mac_t mac;
+
+	gxio_mpipe_context_t *context = _gxio_get_link_context();
+	if (!context)
+		return GXIO_ERR_NO_DEVICE;
+
+	rv = gxio_mpipe_info_enumerate_aux(context, idx, &name, &mac);
+	if (rv >= 0) {
+		if (strscpy(link_name, name.name, sizeof(name.name)) == 0)
+			return GXIO_ERR_INVAL_MEMORY_SIZE;
+		memcpy(link_mac, mac.mac, sizeof(mac.mac));
+	}
+
+	return rv;
+}
+
+EXPORT_SYMBOL_GPL(gxio_mpipe_link_enumerate_mac);
+
+int gxio_mpipe_link_open(gxio_mpipe_link_t *link,
+			 gxio_mpipe_context_t *context, const char *link_name,
+			 unsigned int flags)
+{
+	_gxio_mpipe_link_name_t name;
+	int rv;
+
+	if (strscpy(name.name, link_name, sizeof(name.name)) == 0)
+		return GXIO_ERR_NO_DEVICE;
+
+	rv = gxio_mpipe_link_open_aux(context, name, flags);
+	if (rv < 0)
+		return rv;
+
+	link->context = context;
+	link->channel = rv >> 8;
+	link->mac = rv & 0xFF;
+
+	return 0;
+}
+
+EXPORT_SYMBOL_GPL(gxio_mpipe_link_open);
+
+int gxio_mpipe_link_close(gxio_mpipe_link_t *link)
+{
+	return gxio_mpipe_link_close_aux(link->context, link->mac);
+}
+
+EXPORT_SYMBOL_GPL(gxio_mpipe_link_close);
+
+int gxio_mpipe_link_set_attr(gxio_mpipe_link_t *link, uint32_t attr,
+			     int64_t val)
+{
+	return gxio_mpipe_link_set_attr_aux(link->context, link->mac, attr,
+					    val);
+}
+
+EXPORT_SYMBOL_GPL(gxio_mpipe_link_set_attr);
diff --git a/kernel/arch/tile/gxio/trio.c b/kernel/arch/tile/gxio/trio.c
new file mode 100644
index 000000000..69f0b8df3
--- /dev/null
+++ b/kernel/arch/tile/gxio/trio.c
@@ -0,0 +1,49 @@
+/*
+ * Copyright 2012 Tilera Corporation. All Rights Reserved.
+ *
+ *   This program is free software; you can redistribute it and/or
+ *   modify it under the terms of the GNU General Public License
+ *   as published by the Free Software Foundation, version 2.
+ *
+ *   This program is distributed in the hope that it will be useful, but
+ *   WITHOUT ANY WARRANTY; without even the implied warranty of
+ *   MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
+ *   NON INFRINGEMENT.  See the GNU General Public License for
+ *   more details.
+ */
+
+/*
+ * Implementation of trio gxio calls.
+ */
+
+#include <linux/errno.h>
+#include <linux/io.h>
+#include <linux/module.h>
+
+#include <gxio/trio.h>
+#include <gxio/iorpc_globals.h>
+#include <gxio/iorpc_trio.h>
+#include <gxio/kiorpc.h>
+
+int gxio_trio_init(gxio_trio_context_t *context, unsigned int trio_index)
+{
+	char file[32];
+	int fd;
+
+	snprintf(file, sizeof(file), "trio/%d/iorpc", trio_index);
+	fd = hv_dev_open((HV_VirtAddr) file, 0);
+	if (fd < 0) {
+		context->fd = -1;
+
+		if (fd >= GXIO_ERR_MIN && fd <= GXIO_ERR_MAX)
+			return fd;
+		else
+			return -ENODEV;
+	}
+
+	context->fd = fd;
+
+	return 0;
+}
+
+EXPORT_SYMBOL_GPL(gxio_trio_init);
diff --git a/kernel/arch/tile/gxio/uart.c b/kernel/arch/tile/gxio/uart.c
new file mode 100644
index 000000000..ba585175e
--- /dev/null
+++ b/kernel/arch/tile/gxio/uart.c
@@ -0,0 +1,87 @@
+/*
+ * Copyright 2013 Tilera Corporation. All Rights Reserved.
+ *
+ *   This program is free software; you can redistribute it and/or
+ *   modify it under the terms of the GNU General Public License
+ *   as published by the Free Software Foundation, version 2.
+ *
+ *   This program is distributed in the hope that it will be useful, but
+ *   WITHOUT ANY WARRANTY; without even the implied warranty of
+ *   MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
+ *   NON INFRINGEMENT.  See the GNU General Public License for
+ *   more details.
+ */
+
+/*
+ * Implementation of UART gxio calls.
+ */
+
+#include <linux/io.h>
+#include <linux/errno.h>
+#include <linux/module.h>
+
+#include <gxio/uart.h>
+#include <gxio/iorpc_globals.h>
+#include <gxio/iorpc_uart.h>
+#include <gxio/kiorpc.h>
+
+int gxio_uart_init(gxio_uart_context_t *context, int uart_index)
+{
+	char file[32];
+	int fd;
+
+	snprintf(file, sizeof(file), "uart/%d/iorpc", uart_index);
+	fd = hv_dev_open((HV_VirtAddr) file, 0);
+	if (fd < 0) {
+		if (fd >= GXIO_ERR_MIN && fd <= GXIO_ERR_MAX)
+			return fd;
+		else
+			return -ENODEV;
+	}
+
+	context->fd = fd;
+
+	/* Map in the MMIO space. */
+	context->mmio_base = (void __force *)
+		iorpc_ioremap(fd, HV_UART_MMIO_OFFSET, HV_UART_MMIO_SIZE);
+
+	if (context->mmio_base == NULL) {
+		hv_dev_close(context->fd);
+		context->fd = -1;
+		return -ENODEV;
+	}
+
+	return 0;
+}
+
+EXPORT_SYMBOL_GPL(gxio_uart_init);
+
+int gxio_uart_destroy(gxio_uart_context_t *context)
+{
+	iounmap((void __force __iomem *)(context->mmio_base));
+	hv_dev_close(context->fd);
+
+	context->mmio_base = NULL;
+	context->fd = -1;
+
+	return 0;
+}
+
+EXPORT_SYMBOL_GPL(gxio_uart_destroy);
+
+/* UART register write wrapper. */
+void gxio_uart_write(gxio_uart_context_t *context, uint64_t offset,
+		     uint64_t word)
+{
+	__gxio_mmio_write(context->mmio_base + offset, word);
+}
+
+EXPORT_SYMBOL_GPL(gxio_uart_write);
+
+/* UART register read wrapper. */
+uint64_t gxio_uart_read(gxio_uart_context_t *context, uint64_t offset)
+{
+	return __gxio_mmio_read(context->mmio_base + offset);
+}
+
+EXPORT_SYMBOL_GPL(gxio_uart_read);
diff --git a/kernel/arch/tile/gxio/usb_host.c b/kernel/arch/tile/gxio/usb_host.c
new file mode 100644
index 000000000..785afad79
--- /dev/null
+++ b/kernel/arch/tile/gxio/usb_host.c
@@ -0,0 +1,91 @@
+/*
+ * Copyright 2012 Tilera Corporation. All Rights Reserved.
+ *
+ *   This program is free software; you can redistribute it and/or
+ *   modify it under the terms of the GNU General Public License
+ *   as published by the Free Software Foundation, version 2.
+ *
+ *   This program is distributed in the hope that it will be useful, but
+ *   WITHOUT ANY WARRANTY; without even the implied warranty of
+ *   MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
+ *   NON INFRINGEMENT.  See the GNU General Public License for
+ *   more details.
+ */
+
+/*
+ *
+ * Implementation of USB gxio calls.
+ */
+
+#include <linux/io.h>
+#include <linux/errno.h>
+#include <linux/module.h>
+
+#include <gxio/iorpc_globals.h>
+#include <gxio/iorpc_usb_host.h>
+#include <gxio/kiorpc.h>
+#include <gxio/usb_host.h>
+
+int gxio_usb_host_init(gxio_usb_host_context_t *context, int usb_index,
+		       int is_ehci)
+{
+	char file[32];
+	int fd;
+
+	if (is_ehci)
+		snprintf(file, sizeof(file), "usb_host/%d/iorpc/ehci",
+			 usb_index);
+	else
+		snprintf(file, sizeof(file), "usb_host/%d/iorpc/ohci",
+			 usb_index);
+
+	fd = hv_dev_open((HV_VirtAddr) file, 0);
+	if (fd < 0) {
+		if (fd >= GXIO_ERR_MIN && fd <= GXIO_ERR_MAX)
+			return fd;
+		else
+			return -ENODEV;
+	}
+
+	context->fd = fd;
+
+	// Map in the MMIO space.
+	context->mmio_base =
+		(void __force *)iorpc_ioremap(fd, 0, HV_USB_HOST_MMIO_SIZE);
+
+	if (context->mmio_base == NULL) {
+		hv_dev_close(context->fd);
+		return -ENODEV;
+	}
+
+	return 0;
+}
+
+EXPORT_SYMBOL_GPL(gxio_usb_host_init);
+
+int gxio_usb_host_destroy(gxio_usb_host_context_t *context)
+{
+	iounmap((void __force __iomem *)(context->mmio_base));
+	hv_dev_close(context->fd);
+
+	context->mmio_base = NULL;
+	context->fd = -1;
+
+	return 0;
+}
+
+EXPORT_SYMBOL_GPL(gxio_usb_host_destroy);
+
+void *gxio_usb_host_get_reg_start(gxio_usb_host_context_t *context)
+{
+	return context->mmio_base;
+}
+
+EXPORT_SYMBOL_GPL(gxio_usb_host_get_reg_start);
+
+size_t gxio_usb_host_get_reg_len(gxio_usb_host_context_t *context)
+{
+	return HV_USB_HOST_MMIO_SIZE;
+}
+
+EXPORT_SYMBOL_GPL(gxio_usb_host_get_reg_len);
diff --git a/kernel/arch/tile/include/arch/Kbuild b/kernel/arch/tile/include/arch/Kbuild
new file mode 100644
index 000000000..3751c9fab
--- /dev/null
+++ b/kernel/arch/tile/include/arch/Kbuild
@@ -0,0 +1 @@
+# Tile arch headers
diff --git a/kernel/arch/tile/include/arch/mpipe.h b/kernel/arch/tile/include/arch/mpipe.h
new file mode 100644
index 000000000..904538e75
--- /dev/null
+++ b/kernel/arch/tile/include/arch/mpipe.h
@@ -0,0 +1,371 @@
+/*
+ * Copyright 2012 Tilera Corporation. All Rights Reserved.
+ *
+ *   This program is free software; you can redistribute it and/or
+ *   modify it under the terms of the GNU General Public License
+ *   as published by the Free Software Foundation, version 2.
+ *
+ *   This program is distributed in the hope that it will be useful, but
+ *   WITHOUT ANY WARRANTY; without even the implied warranty of
+ *   MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
+ *   NON INFRINGEMENT.  See the GNU General Public License for
+ *   more details.
+ */
+
+/* Machine-generated file; do not edit. */
+
+#ifndef __ARCH_MPIPE_H__
+#define __ARCH_MPIPE_H__
+
+#include <arch/abi.h>
+#include <arch/mpipe_def.h>
+
+#ifndef __ASSEMBLER__
+
+/*
+ * MMIO Ingress DMA Release Region Address.
+ * This is a description of the physical addresses used to manipulate ingress
+ * credit counters.  Accesses to this address space should use an address of
+ * this form and a value like that specified in IDMA_RELEASE_REGION_VAL.
+ */
+
+__extension__
+typedef union
+{
+  struct
+  {
+#ifndef __BIG_ENDIAN__
+    /* Reserved. */
+    uint_reg_t __reserved_0  : 3;
+    /* NotifRing to be released */
+    uint_reg_t ring          : 8;
+    /* Bucket to be released */
+    uint_reg_t bucket        : 13;
+    /* Enable NotifRing release */
+    uint_reg_t ring_enable   : 1;
+    /* Enable Bucket release */
+    uint_reg_t bucket_enable : 1;
+    /*
+     * This field of the address selects the region (address space) to be
+     * accessed.  For the iDMA release region, this field must be 4.
+     */
+    uint_reg_t region        : 3;
+    /* Reserved. */
+    uint_reg_t __reserved_1  : 6;
+    /* This field of the address indexes the 32 entry service domain table. */
+    uint_reg_t svc_dom       : 5;
+    /* Reserved. */
+    uint_reg_t __reserved_2  : 24;
+#else   /* __BIG_ENDIAN__ */
+    uint_reg_t __reserved_2  : 24;
+    uint_reg_t svc_dom       : 5;
+    uint_reg_t __reserved_1  : 6;
+    uint_reg_t region        : 3;
+    uint_reg_t bucket_enable : 1;
+    uint_reg_t ring_enable   : 1;
+    uint_reg_t bucket        : 13;
+    uint_reg_t ring          : 8;
+    uint_reg_t __reserved_0  : 3;
+#endif
+  };
+
+  uint_reg_t word;
+} MPIPE_IDMA_RELEASE_REGION_ADDR_t;
+
+/*
+ * MMIO Ingress DMA Release Region Value - Release NotifRing and/or Bucket.
+ * Provides release of the associated NotifRing.  The address of the MMIO
+ * operation is described in IDMA_RELEASE_REGION_ADDR.
+ */
+
+__extension__
+typedef union
+{
+  struct
+  {
+#ifndef __BIG_ENDIAN__
+    /*
+     * Number of packets being released.  The load balancer's count of
+     * inflight packets will be decremented by this amount for the associated
+     * Bucket and/or NotifRing
+     */
+    uint_reg_t count      : 16;
+    /* Reserved. */
+    uint_reg_t __reserved : 48;
+#else   /* __BIG_ENDIAN__ */
+    uint_reg_t __reserved : 48;
+    uint_reg_t count      : 16;
+#endif
+  };
+
+  uint_reg_t word;
+} MPIPE_IDMA_RELEASE_REGION_VAL_t;
+
+/*
+ * MMIO Buffer Stack Manager Region Address.
+ * This MMIO region is used for posting or fetching buffers to/from the
+ * buffer stack manager.  On an MMIO load, this pops a buffer descriptor from
+ * the top of stack if one is available.  On an MMIO store, this pushes a
+ * buffer to the stack.  The value read or written is described in
+ * BSM_REGION_VAL.
+ */
+
+__extension__
+typedef union
+{
+  struct
+  {
+#ifndef __BIG_ENDIAN__
+    /* Reserved. */
+    uint_reg_t __reserved_0 : 3;
+    /* BufferStack being accessed. */
+    uint_reg_t stack        : 5;
+    /* Reserved. */
+    uint_reg_t __reserved_1 : 18;
+    /*
+     * This field of the address selects the region (address space) to be
+     * accessed.  For the buffer stack manager region, this field must be 6.
+     */
+    uint_reg_t region       : 3;
+    /* Reserved. */
+    uint_reg_t __reserved_2 : 6;
+    /* This field of the address indexes the 32 entry service domain table. */
+    uint_reg_t svc_dom      : 5;
+    /* Reserved. */
+    uint_reg_t __reserved_3 : 24;
+#else   /* __BIG_ENDIAN__ */
+    uint_reg_t __reserved_3 : 24;
+    uint_reg_t svc_dom      : 5;
+    uint_reg_t __reserved_2 : 6;
+    uint_reg_t region       : 3;
+    uint_reg_t __reserved_1 : 18;
+    uint_reg_t stack        : 5;
+    uint_reg_t __reserved_0 : 3;
+#endif
+  };
+
+  uint_reg_t word;
+} MPIPE_BSM_REGION_ADDR_t;
+
+/*
+ * MMIO Buffer Stack Manager Region Value.
+ * This MMIO region is used for posting or fetching buffers to/from the
+ * buffer stack manager.  On an MMIO load, this pops a buffer descriptor from
+ * the top of stack if one is available. On an MMIO store, this pushes a
+ * buffer to the stack.  The address of the MMIO operation is described in
+ * BSM_REGION_ADDR.
+ */
+
+__extension__
+typedef union
+{
+  struct
+  {
+#ifndef __BIG_ENDIAN__
+    /* Reserved. */
+    uint_reg_t __reserved_0 : 7;
+    /*
+     * Base virtual address of the buffer.  Must be sign extended by consumer.
+     */
+    int_reg_t va           : 35;
+    /* Reserved. */
+    uint_reg_t __reserved_1 : 6;
+    /*
+     * Index of the buffer stack to which this buffer belongs.  Ignored on
+     * writes since the offset bits specify the stack being accessed.
+     */
+    uint_reg_t stack_idx    : 5;
+    /* Reserved. */
+    uint_reg_t __reserved_2 : 3;
+    /*
+     * Instance ID.  For devices that support automatic buffer return between
+     * mPIPE instances, this field indicates the buffer owner.  If the INST
+     * field does not match the mPIPE's instance number when a packet is
+     * egressed, buffers with HWB set will be returned to the other mPIPE
+     * instance.  Note that not all devices support multi-mPIPE buffer
+     * return.  The MPIPE_EDMA_INFO.REMOTE_BUFF_RTN_SUPPORT bit indicates
+     * whether the INST field in the buffer descriptor is populated by iDMA
+     * hardware. This field is ignored on writes.
+     */
+    uint_reg_t inst         : 2;
+    /*
+     * Reads as one to indicate that this is a hardware managed buffer.
+     * Ignored on writes since all buffers on a given stack are the same size.
+     */
+    uint_reg_t hwb          : 1;
+    /*
+     * Encoded size of buffer (ignored on writes):
+     * 0 = 128 bytes
+     * 1 = 256 bytes
+     * 2 = 512 bytes
+     * 3 = 1024 bytes
+     * 4 = 1664 bytes
+     * 5 = 4096 bytes
+     * 6 = 10368 bytes
+     * 7 = 16384 bytes
+     */
+    uint_reg_t size         : 3;
+    /*
+     * Valid indication for the buffer.  Ignored on writes.
+     * 0 : Valid buffer descriptor popped from stack.
+     * 3 : Could not pop a buffer from the stack.  Either the stack is empty,
+     * or the hardware's prefetch buffer is empty for this stack.
+     */
+    uint_reg_t c            : 2;
+#else   /* __BIG_ENDIAN__ */
+    uint_reg_t c            : 2;
+    uint_reg_t size         : 3;
+    uint_reg_t hwb          : 1;
+    uint_reg_t inst         : 2;
+    uint_reg_t __reserved_2 : 3;
+    uint_reg_t stack_idx    : 5;
+    uint_reg_t __reserved_1 : 6;
+    int_reg_t va           : 35;
+    uint_reg_t __reserved_0 : 7;
+#endif
+  };
+
+  uint_reg_t word;
+} MPIPE_BSM_REGION_VAL_t;
+
+/*
+ * MMIO Egress DMA Post Region Address.
+ * Used to post descriptor locations to the eDMA descriptor engine.  The
+ * value to be written is described in EDMA_POST_REGION_VAL
+ */
+
+__extension__
+typedef union
+{
+  struct
+  {
+#ifndef __BIG_ENDIAN__
+    /* Reserved. */
+    uint_reg_t __reserved_0 : 3;
+    /* eDMA ring being accessed */
+    uint_reg_t ring         : 6;
+    /* Reserved. */
+    uint_reg_t __reserved_1 : 17;
+    /*
+     * This field of the address selects the region (address space) to be
+     * accessed.  For the egress DMA post region, this field must be 5.
+     */
+    uint_reg_t region       : 3;
+    /* Reserved. */
+    uint_reg_t __reserved_2 : 6;
+    /* This field of the address indexes the 32 entry service domain table. */
+    uint_reg_t svc_dom      : 5;
+    /* Reserved. */
+    uint_reg_t __reserved_3 : 24;
+#else   /* __BIG_ENDIAN__ */
+    uint_reg_t __reserved_3 : 24;
+    uint_reg_t svc_dom      : 5;
+    uint_reg_t __reserved_2 : 6;
+    uint_reg_t region       : 3;
+    uint_reg_t __reserved_1 : 17;
+    uint_reg_t ring         : 6;
+    uint_reg_t __reserved_0 : 3;
+#endif
+  };
+
+  uint_reg_t word;
+} MPIPE_EDMA_POST_REGION_ADDR_t;
+
+/*
+ * MMIO Egress DMA Post Region Value.
+ * Used to post descriptor locations to the eDMA descriptor engine.  The
+ * address is described in EDMA_POST_REGION_ADDR.
+ */
+
+__extension__
+typedef union
+{
+  struct
+  {
+#ifndef __BIG_ENDIAN__
+    /*
+     * For writes, this specifies the current ring tail pointer prior to any
+     * post.  For example, to post 1 or more descriptors starting at location
+     * 23, this would contain 23 (not 24).  On writes, this index must be
+     * masked based on the ring size.  The new tail pointer after this post
+     * is COUNT+RING_IDX (masked by the ring size).
+     *
+     * For reads, this provides the hardware descriptor fetcher's head
+     * pointer.  The descriptors prior to the head pointer, however, may not
+     * yet have been processed so this indicator is only used to determine
+     * how full the ring is and if software may post more descriptors.
+     */
+    uint_reg_t ring_idx   : 16;
+    /*
+     * For writes, this specifies number of contiguous descriptors that are
+     * being posted.  Software may post up to RingSize descriptors with a
+     * single MMIO store.  A zero in this field on a write will "wake up" an
+     * eDMA ring and cause it fetch descriptors regardless of the hardware's
+     * current view of the state of the tail pointer.
+     *
+     * For reads, this field provides a rolling count of the number of
+     * descriptors that have been completely processed.  This may be used by
+     * software to determine when buffers associated with a descriptor may be
+     * returned or reused.  When the ring's flush bit is cleared by software
+     * (after having been set by HW or SW), the COUNT will be cleared.
+     */
+    uint_reg_t count      : 16;
+    /*
+     * For writes, this specifies the generation number of the tail being
+     * posted. Note that if tail+cnt wraps to the beginning of the ring, the
+     * eDMA hardware assumes that the descriptors posted at the beginning of
+     * the ring are also valid so it is okay to post around the wrap point.
+     *
+     * For reads, this is the current generation number.  Valid descriptors
+     * will have the inverse of this generation number.
+     */
+    uint_reg_t gen        : 1;
+    /* Reserved. */
+    uint_reg_t __reserved : 31;
+#else   /* __BIG_ENDIAN__ */
+    uint_reg_t __reserved : 31;
+    uint_reg_t gen        : 1;
+    uint_reg_t count      : 16;
+    uint_reg_t ring_idx   : 16;
+#endif
+  };
+
+  uint_reg_t word;
+} MPIPE_EDMA_POST_REGION_VAL_t;
+
+/*
+ * Load Balancer Bucket Status Data.
+ * Read/Write data for load balancer Bucket-Status Table. 4160 entries
+ * indexed by LBL_INIT_CTL.IDX when LBL_INIT_CTL.STRUCT_SEL is BSTS_TBL
+ */
+
+__extension__
+typedef union
+{
+  struct
+  {
+#ifndef __BIG_ENDIAN__
+    /* NotifRing currently assigned to this bucket. */
+    uint_reg_t notifring  : 8;
+    /* Current reference count. */
+    uint_reg_t count      : 16;
+    /* Group associated with this bucket. */
+    uint_reg_t group      : 5;
+    /* Mode select for this bucket. */
+    uint_reg_t mode       : 3;
+    /* Reserved. */
+    uint_reg_t __reserved : 32;
+#else   /* __BIG_ENDIAN__ */
+    uint_reg_t __reserved : 32;
+    uint_reg_t mode       : 3;
+    uint_reg_t group      : 5;
+    uint_reg_t count      : 16;
+    uint_reg_t notifring  : 8;
+#endif
+  };
+
+  uint_reg_t word;
+} MPIPE_LBL_INIT_DAT_BSTS_TBL_t;
+#endif /* !defined(__ASSEMBLER__) */
+
+#endif /* !defined(__ARCH_MPIPE_H__) */
diff --git a/kernel/arch/tile/include/arch/mpipe_constants.h b/kernel/arch/tile/include/arch/mpipe_constants.h
new file mode 100644
index 000000000..84022ac5f
--- /dev/null
+++ b/kernel/arch/tile/include/arch/mpipe_constants.h
@@ -0,0 +1,42 @@
+/*
+ * Copyright 2012 Tilera Corporation. All Rights Reserved.
+ *
+ *   This program is free software; you can redistribute it and/or
+ *   modify it under the terms of the GNU General Public License
+ *   as published by the Free Software Foundation, version 2.
+ *
+ *   This program is distributed in the hope that it will be useful, but
+ *   WITHOUT ANY WARRANTY; without even the implied warranty of
+ *   MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
+ *   NON INFRINGEMENT.  See the GNU General Public License for
+ *   more details.
+ */
+
+
+#ifndef __ARCH_MPIPE_CONSTANTS_H__
+#define __ARCH_MPIPE_CONSTANTS_H__
+
+#define MPIPE_NUM_CLASSIFIERS 16
+#define MPIPE_CLS_MHZ 1200
+
+#define MPIPE_NUM_EDMA_RINGS 64
+
+#define MPIPE_NUM_SGMII_MACS 16
+#define MPIPE_NUM_XAUI_MACS 16
+#define MPIPE_NUM_LOOPBACK_CHANNELS 4
+#define MPIPE_NUM_NON_LB_CHANNELS 28
+
+#define MPIPE_NUM_IPKT_BLOCKS 1536
+
+#define MPIPE_NUM_BUCKETS 4160
+
+#define MPIPE_NUM_NOTIF_RINGS 256
+
+#define MPIPE_NUM_NOTIF_GROUPS 32
+
+#define MPIPE_NUM_TLBS_PER_ASID 16
+#define MPIPE_TLB_IDX_WIDTH 4
+
+#define MPIPE_MMIO_NUM_SVC_DOM 32
+
+#endif /* __ARCH_MPIPE_CONSTANTS_H__ */
diff --git a/kernel/arch/tile/include/arch/mpipe_def.h b/kernel/arch/tile/include/arch/mpipe_def.h
new file mode 100644
index 000000000..c3d30217f
--- /dev/null
+++ b/kernel/arch/tile/include/arch/mpipe_def.h
@@ -0,0 +1,39 @@
+/*
+ * Copyright 2012 Tilera Corporation. All Rights Reserved.
+ *
+ *   This program is free software; you can redistribute it and/or
+ *   modify it under the terms of the GNU General Public License
+ *   as published by the Free Software Foundation, version 2.
+ *
+ *   This program is distributed in the hope that it will be useful, but
+ *   WITHOUT ANY WARRANTY; without even the implied warranty of
+ *   MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
+ *   NON INFRINGEMENT.  See the GNU General Public License for
+ *   more details.
+ */
+
+/* Machine-generated file; do not edit. */
+
+#ifndef __ARCH_MPIPE_DEF_H__
+#define __ARCH_MPIPE_DEF_H__
+#define MPIPE_MMIO_ADDR__REGION_SHIFT 26
+#define MPIPE_MMIO_ADDR__REGION_VAL_CFG 0x0
+#define MPIPE_MMIO_ADDR__REGION_VAL_IDMA 0x4
+#define MPIPE_MMIO_ADDR__REGION_VAL_EDMA 0x5
+#define MPIPE_MMIO_ADDR__REGION_VAL_BSM 0x6
+#define MPIPE_BSM_REGION_VAL__VA_SHIFT 7
+#define MPIPE_BSM_INIT_DAT_1__SIZE_VAL_BSZ_128 0x0
+#define MPIPE_BSM_INIT_DAT_1__SIZE_VAL_BSZ_256 0x1
+#define MPIPE_BSM_INIT_DAT_1__SIZE_VAL_BSZ_512 0x2
+#define MPIPE_BSM_INIT_DAT_1__SIZE_VAL_BSZ_1024 0x3
+#define MPIPE_BSM_INIT_DAT_1__SIZE_VAL_BSZ_1664 0x4
+#define MPIPE_BSM_INIT_DAT_1__SIZE_VAL_BSZ_4096 0x5
+#define MPIPE_BSM_INIT_DAT_1__SIZE_VAL_BSZ_10368 0x6
+#define MPIPE_BSM_INIT_DAT_1__SIZE_VAL_BSZ_16384 0x7
+#define MPIPE_LBL_INIT_DAT_BSTS_TBL__MODE_VAL_DFA 0x0
+#define MPIPE_LBL_INIT_DAT_BSTS_TBL__MODE_VAL_FIXED 0x1
+#define MPIPE_LBL_INIT_DAT_BSTS_TBL__MODE_VAL_ALWAYS_PICK 0x2
+#define MPIPE_LBL_INIT_DAT_BSTS_TBL__MODE_VAL_STICKY 0x3
+#define MPIPE_LBL_INIT_DAT_BSTS_TBL__MODE_VAL_STICKY_RAND 0x7
+#define MPIPE_LBL_NR_STATE__FIRST_WORD 0x2138
+#endif /* !defined(__ARCH_MPIPE_DEF_H__) */
diff --git a/kernel/arch/tile/include/arch/mpipe_shm.h b/kernel/arch/tile/include/arch/mpipe_shm.h
new file mode 100644
index 000000000..13b3c4300
--- /dev/null
+++ b/kernel/arch/tile/include/arch/mpipe_shm.h
@@ -0,0 +1,521 @@
+/*
+ * Copyright 2012 Tilera Corporation. All Rights Reserved.
+ *
+ *   This program is free software; you can redistribute it and/or
+ *   modify it under the terms of the GNU General Public License
+ *   as published by the Free Software Foundation, version 2.
+ *
+ *   This program is distributed in the hope that it will be useful, but
+ *   WITHOUT ANY WARRANTY; without even the implied warranty of
+ *   MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
+ *   NON INFRINGEMENT.  See the GNU General Public License for
+ *   more details.
+ */
+
+/* Machine-generated file; do not edit. */
+
+
+#ifndef __ARCH_MPIPE_SHM_H__
+#define __ARCH_MPIPE_SHM_H__
+
+#include <arch/abi.h>
+#include <arch/mpipe_shm_def.h>
+
+#ifndef __ASSEMBLER__
+/**
+ * MPIPE eDMA Descriptor.
+ * The eDMA descriptor is written by software and consumed by hardware.  It
+ * is used to specify the location of egress packet data to be sent out of
+ * the chip via one of the packet interfaces.
+ */
+
+__extension__
+typedef union
+{
+  struct
+  {
+    /* Word 0 */
+
+#ifndef __BIG_ENDIAN__
+    /**
+     * Generation number.  Used to indicate a valid descriptor in ring.  When
+     * a new descriptor is written into the ring, software must toggle this
+     * bit.  The net effect is that the GEN bit being written into new
+     * descriptors toggles each time the ring tail pointer wraps.
+     */
+    uint_reg_t gen        : 1;
+    /**
+     * For devices with EDMA reorder support, this field allows the
+     * descriptor to select the egress FIFO.  The associated DMA ring must
+     * have ALLOW_EFIFO_SEL enabled.
+     */
+    uint_reg_t efifo_sel  : 6;
+    /** Reserved.  Must be zero. */
+    uint_reg_t r0         : 1;
+    /** Checksum generation enabled for this transfer. */
+    uint_reg_t csum       : 1;
+    /**
+     * Nothing to be sent.  Used, for example, when software has dropped a
+     * packet but still wishes to return all of the associated buffers.
+     */
+    uint_reg_t ns         : 1;
+    /**
+     * Notification interrupt will be delivered when packet has been egressed.
+     */
+    uint_reg_t notif      : 1;
+    /**
+     * Boundary indicator.  When 1, this transfer includes the EOP for this
+     * command.  Must be clear on all but the last descriptor for an egress
+     * packet.
+     */
+    uint_reg_t bound      : 1;
+    /** Reserved.  Must be zero. */
+    uint_reg_t r1         : 4;
+    /**
+     * Number of bytes to be sent for this descriptor.  When zero, no data
+     * will be moved and the buffer descriptor will be ignored.  If the
+     * buffer descriptor indicates that it is chained, the low 7 bits of the
+     * VA indicate the offset within the first buffer (e.g. 127 bytes is the
+     * maximum offset into the first buffer).  If the size exceeds a single
+     * buffer, subsequent buffer descriptors will be fetched prior to
+     * processing the next eDMA descriptor in the ring.
+     */
+    uint_reg_t xfer_size  : 14;
+    /** Reserved.  Must be zero. */
+    uint_reg_t r2         : 2;
+    /**
+     * Destination of checksum relative to CSUM_START relative to the first
+     * byte moved by this descriptor.  Must be zero if CSUM=0 in this
+     * descriptor.  Must be less than XFER_SIZE (e.g. the first byte of the
+     * CSUM_DEST must be within the span of this descriptor).
+     */
+    uint_reg_t csum_dest  : 8;
+    /**
+     * Start byte of checksum relative to the first byte moved by this
+     * descriptor.  If this is not the first descriptor for the egress
+     * packet, CSUM_START is still relative to the first byte in this
+     * descriptor.  Must be zero if CSUM=0 in this descriptor.
+     */
+    uint_reg_t csum_start : 8;
+    /**
+     * Initial value for 16-bit 1's compliment checksum if enabled via CSUM.
+     * Specified in network order.  That is, bits[7:0] will be added to the
+     * byte pointed to by CSUM_START and bits[15:8] will be added to the byte
+     * pointed to by CSUM_START+1 (with appropriate 1's compliment carries).
+     * Must be zero if CSUM=0 in this descriptor.
+     */
+    uint_reg_t csum_seed  : 16;
+#else   /* __BIG_ENDIAN__ */
+    uint_reg_t csum_seed  : 16;
+    uint_reg_t csum_start : 8;
+    uint_reg_t csum_dest  : 8;
+    uint_reg_t r2         : 2;
+    uint_reg_t xfer_size  : 14;
+    uint_reg_t r1         : 4;
+    uint_reg_t bound      : 1;
+    uint_reg_t notif      : 1;
+    uint_reg_t ns         : 1;
+    uint_reg_t csum       : 1;
+    uint_reg_t r0         : 1;
+    uint_reg_t efifo_sel  : 6;
+    uint_reg_t gen        : 1;
+#endif
+
+    /* Word 1 */
+
+#ifndef __BIG_ENDIAN__
+    /** Virtual address.  Must be sign extended by consumer. */
+    int_reg_t va           : 42;
+    /** Reserved. */
+    uint_reg_t __reserved_0 : 6;
+    /** Index of the buffer stack to which this buffer belongs. */
+    uint_reg_t stack_idx    : 5;
+    /** Reserved. */
+    uint_reg_t __reserved_1 : 3;
+    /**
+     * Instance ID.  For devices that support automatic buffer return between
+     * mPIPE instances, this field indicates the buffer owner.  If the INST
+     * field does not match the mPIPE's instance number when a packet is
+     * egressed, buffers with HWB set will be returned to the other mPIPE
+     * instance.  Note that not all devices support multi-mPIPE buffer
+     * return.  The MPIPE_EDMA_INFO.REMOTE_BUFF_RTN_SUPPORT bit indicates
+     * whether the INST field in the buffer descriptor is populated by iDMA
+     * hardware.
+     */
+    uint_reg_t inst         : 2;
+    /**
+     * Always set to one by hardware in iDMA packet descriptors.  For eDMA,
+     * indicates whether the buffer will be released to the buffer stack
+     * manager.  When 0, software is responsible for releasing the buffer.
+     */
+    uint_reg_t hwb          : 1;
+    /**
+     * Encoded size of buffer.  Set by the ingress hardware for iDMA packet
+     * descriptors.  For eDMA descriptors, indicates the buffer size if .c
+     * indicates a chained packet.  If an eDMA descriptor is not chained and
+     * the .hwb bit is not set, this field is ignored and the size is
+     * specified by the .xfer_size field.
+     * 0 = 128 bytes
+     * 1 = 256 bytes
+     * 2 = 512 bytes
+     * 3 = 1024 bytes
+     * 4 = 1664 bytes
+     * 5 = 4096 bytes
+     * 6 = 10368 bytes
+     * 7 = 16384 bytes
+     */
+    uint_reg_t size         : 3;
+    /**
+     * Chaining configuration for the buffer.  Indicates that an ingress
+     * packet or egress command is chained across multiple buffers, with each
+     * buffer's size indicated by the .size field.
+     */
+    uint_reg_t c            : 2;
+#else   /* __BIG_ENDIAN__ */
+    uint_reg_t c            : 2;
+    uint_reg_t size         : 3;
+    uint_reg_t hwb          : 1;
+    uint_reg_t inst         : 2;
+    uint_reg_t __reserved_1 : 3;
+    uint_reg_t stack_idx    : 5;
+    uint_reg_t __reserved_0 : 6;
+    int_reg_t va           : 42;
+#endif
+
+  };
+
+  /** Word access */
+  uint_reg_t words[2];
+} MPIPE_EDMA_DESC_t;
+
+/**
+ * MPIPE Packet Descriptor.
+ * The packet descriptor is filled by the mPIPE's classification,
+ * load-balancing, and buffer management services.  Some fields are consumed
+ * by mPIPE hardware, and others are consumed by Tile software.
+ */
+
+__extension__
+typedef union
+{
+  struct
+  {
+    /* Word 0 */
+
+#ifndef __BIG_ENDIAN__
+    /**
+     * Notification ring into which this packet descriptor is written.
+     * Typically written by load balancer, but can be overridden by
+     * classification program if NR is asserted.
+     */
+    uint_reg_t notif_ring   : 8;
+    /** Source channel for this packet.  Written by mPIPE DMA hardware. */
+    uint_reg_t channel      : 5;
+    /** Reserved. */
+    uint_reg_t __reserved_0 : 1;
+    /**
+     * MAC Error.
+     * Generated by the MAC interface.  Asserted if there was an overrun of
+     * the MAC's receive FIFO.  This condition generally only occurs if the
+     * mPIPE clock is running too slowly.
+     */
+    uint_reg_t me           : 1;
+    /**
+     * Truncation Error.
+     * Written by the iDMA hardware.  Asserted if packet was truncated due to
+     * insufficient space in iPkt buffer
+     */
+    uint_reg_t tr           : 1;
+    /**
+     * Written by the iDMA hardware.  Indicates the number of bytes written
+     * to Tile memory.  In general, this is the actual size of the packet as
+     * received from the MAC.  But if the packet is truncated due to running
+     * out of buffers or due to the iPkt buffer filling up, then the L2_SIZE
+     * will be reduced to reflect the actual number of valid bytes written to
+     * Tile memory.
+     */
+    uint_reg_t l2_size      : 14;
+    /**
+     * CRC Error.
+     * Generated by the MAC.  Asserted if MAC indicated an L2 CRC error or
+     * other L2 error (bad length etc.) on the packet.
+     */
+    uint_reg_t ce           : 1;
+    /**
+     * Cut Through.
+     * Written by the iDMA hardware.  Asserted if packet was not completely
+     * received before being sent to classifier.  L2_Size will indicate
+     * number of bytes received so far.
+     */
+    uint_reg_t ct           : 1;
+    /**
+     * Written by the classification program.  Used by the load balancer to
+     * select the ring into which this packet descriptor is written.
+     */
+    uint_reg_t bucket_id    : 13;
+    /** Reserved. */
+    uint_reg_t __reserved_1 : 3;
+    /**
+     * Checksum.
+     * Written by classification program.  When 1, the checksum engine will
+     * perform checksum based on the CSUM_SEED, CSUM_START, and CSUM_BYTES
+     * fields.  The result will be placed in CSUM_VAL.
+     */
+    uint_reg_t cs           : 1;
+    /**
+     * Notification Ring Select.
+     * Written by the classification program.  When 1, the NotifRingIDX is
+     * set by classification program rather than being set by load balancer.
+     */
+    uint_reg_t nr           : 1;
+    /**
+     * Written by classification program.  Indicates whether packet and
+     * descriptor should both be dropped, both be delivered, or only the
+     * descriptor should be delivered.
+     */
+    uint_reg_t dest         : 2;
+    /**
+     * General Purpose Sequence Number Enable.
+     * Written by the classification program.  When 1, the GP_SQN_SEL field
+     * contains the sequence number selector and the GP_SQN field will be
+     * replaced with the associated sequence number.  When clear, the GP_SQN
+     * field is left intact and be used as "Custom" bytes.
+     */
+    uint_reg_t sq           : 1;
+    /**
+     * TimeStamp Enable.
+     * Enable TimeStamp insertion.  When clear, timestamp field may be filled
+     * with custom data by classifier.  When set, hardware inserts the
+     * timestamp when the start of packet is received from the MAC.
+     */
+    uint_reg_t ts           : 1;
+    /**
+     * Packet Sequence Number Enable.
+     * Enable PacketSQN insertion.  When clear, PacketSQN field may be filled
+     * with custom data by classifier.  When set, hardware inserts the packet
+     * sequence number when the packet descriptor is written to a
+     * notification ring.
+     */
+    uint_reg_t ps           : 1;
+    /**
+     * Buffer Error.
+     * Written by the iDMA hardware.  Asserted if iDMA ran out of buffers
+     * while writing the packet. Software must still return any buffer
+     * descriptors whose C field indicates a valid descriptor was consumed.
+     */
+    uint_reg_t be           : 1;
+    /**
+     * Written by  the classification program.  The associated counter is
+     * incremented when the packet is sent.
+     */
+    uint_reg_t ctr0         : 5;
+    /** Reserved. */
+    uint_reg_t __reserved_2 : 3;
+#else   /* __BIG_ENDIAN__ */
+    uint_reg_t __reserved_2 : 3;
+    uint_reg_t ctr0         : 5;
+    uint_reg_t be           : 1;
+    uint_reg_t ps           : 1;
+    uint_reg_t ts           : 1;
+    uint_reg_t sq           : 1;
+    uint_reg_t dest         : 2;
+    uint_reg_t nr           : 1;
+    uint_reg_t cs           : 1;
+    uint_reg_t __reserved_1 : 3;
+    uint_reg_t bucket_id    : 13;
+    uint_reg_t ct           : 1;
+    uint_reg_t ce           : 1;
+    uint_reg_t l2_size      : 14;
+    uint_reg_t tr           : 1;
+    uint_reg_t me           : 1;
+    uint_reg_t __reserved_0 : 1;
+    uint_reg_t channel      : 5;
+    uint_reg_t notif_ring   : 8;
+#endif
+
+    /* Word 1 */
+
+#ifndef __BIG_ENDIAN__
+    /**
+     * Written by  the classification program.  The associated counter is
+     * incremented when the packet is sent.
+     */
+    uint_reg_t ctr1          : 5;
+    /** Reserved. */
+    uint_reg_t __reserved_3  : 3;
+    /**
+     * Written by classification program.  Indicates the start byte for
+     * checksum.  Relative to 1st byte received from MAC.
+     */
+    uint_reg_t csum_start    : 8;
+    /**
+     * Checksum seed written by classification program.  Overwritten with
+     * resultant checksum if CS bit is asserted.  The endianness of the CSUM
+     * value bits when viewed by Tile software match the packet byte order.
+     * That is, bits[7:0] of the resulting checksum value correspond to
+     * earlier (more significant) bytes in the packet.  To avoid classifier
+     * software from having to byte swap the CSUM_SEED, the iDMA checksum
+     * engine byte swaps the classifier's result before seeding the checksum
+     * calculation.  Thus, the CSUM_START byte of packet data is added to
+     * bits[15:8] of the CSUM_SEED field generated by the classifier.  This
+     * byte swap will be visible to Tile software if the CS bit is clear.
+     */
+    uint_reg_t csum_seed_val : 16;
+    /**
+     * Written by  the classification program.  Not interpreted by mPIPE
+     * hardware.
+     */
+    uint_reg_t custom0       : 32;
+#else   /* __BIG_ENDIAN__ */
+    uint_reg_t custom0       : 32;
+    uint_reg_t csum_seed_val : 16;
+    uint_reg_t csum_start    : 8;
+    uint_reg_t __reserved_3  : 3;
+    uint_reg_t ctr1          : 5;
+#endif
+
+    /* Word 2 */
+
+#ifndef __BIG_ENDIAN__
+    /**
+     * Written by  the classification program.  Not interpreted by mPIPE
+     * hardware.
+     */
+    uint_reg_t custom1 : 64;
+#else   /* __BIG_ENDIAN__ */
+    uint_reg_t custom1 : 64;
+#endif
+
+    /* Word 3 */
+
+#ifndef __BIG_ENDIAN__
+    /**
+     * Written by  the classification program.  Not interpreted by mPIPE
+     * hardware.
+     */
+    uint_reg_t custom2 : 64;
+#else   /* __BIG_ENDIAN__ */
+    uint_reg_t custom2 : 64;
+#endif
+
+    /* Word 4 */
+
+#ifndef __BIG_ENDIAN__
+    /**
+     * Written by  the classification program.  Not interpreted by mPIPE
+     * hardware.
+     */
+    uint_reg_t custom3 : 64;
+#else   /* __BIG_ENDIAN__ */
+    uint_reg_t custom3 : 64;
+#endif
+
+    /* Word 5 */
+
+#ifndef __BIG_ENDIAN__
+    /**
+     * Sequence number applied when packet is distributed.   Classifier
+     * selects which sequence number is to be applied by writing the 13-bit
+     * SQN-selector into this field.  For devices that support EXT_SQN (as
+     * indicated in IDMA_INFO.EXT_SQN_SUPPORT), the GP_SQN can be extended to
+     * 32-bits via the IDMA_CTL.EXT_SQN register.  In this case the
+     * PACKET_SQN will be reduced to 32 bits.
+     */
+    uint_reg_t gp_sqn     : 16;
+    /**
+     * Written by notification hardware.  The packet sequence number is
+     * incremented for each packet that wasn't dropped.
+     */
+    uint_reg_t packet_sqn : 48;
+#else   /* __BIG_ENDIAN__ */
+    uint_reg_t packet_sqn : 48;
+    uint_reg_t gp_sqn     : 16;
+#endif
+
+    /* Word 6 */
+
+#ifndef __BIG_ENDIAN__
+    /**
+     * Written by hardware when the start-of-packet is received by the mPIPE
+     * from the MAC.  This is the nanoseconds part of the packet timestamp.
+     */
+    uint_reg_t time_stamp_ns  : 32;
+    /**
+     * Written by hardware when the start-of-packet is received by the mPIPE
+     * from the MAC.  This is the seconds part of the packet timestamp.
+     */
+    uint_reg_t time_stamp_sec : 32;
+#else   /* __BIG_ENDIAN__ */
+    uint_reg_t time_stamp_sec : 32;
+    uint_reg_t time_stamp_ns  : 32;
+#endif
+
+    /* Word 7 */
+
+#ifndef __BIG_ENDIAN__
+    /** Virtual address.  Must be sign extended by consumer. */
+    int_reg_t va           : 42;
+    /** Reserved. */
+    uint_reg_t __reserved_4 : 6;
+    /** Index of the buffer stack to which this buffer belongs. */
+    uint_reg_t stack_idx    : 5;
+    /** Reserved. */
+    uint_reg_t __reserved_5 : 3;
+    /**
+     * Instance ID.  For devices that support automatic buffer return between
+     * mPIPE instances, this field indicates the buffer owner.  If the INST
+     * field does not match the mPIPE's instance number when a packet is
+     * egressed, buffers with HWB set will be returned to the other mPIPE
+     * instance.  Note that not all devices support multi-mPIPE buffer
+     * return.  The MPIPE_EDMA_INFO.REMOTE_BUFF_RTN_SUPPORT bit indicates
+     * whether the INST field in the buffer descriptor is populated by iDMA
+     * hardware.
+     */
+    uint_reg_t inst         : 2;
+    /**
+     * Always set to one by hardware in iDMA packet descriptors.  For eDMA,
+     * indicates whether the buffer will be released to the buffer stack
+     * manager.  When 0, software is responsible for releasing the buffer.
+     */
+    uint_reg_t hwb          : 1;
+    /**
+     * Encoded size of buffer.  Set by the ingress hardware for iDMA packet
+     * descriptors.  For eDMA descriptors, indicates the buffer size if .c
+     * indicates a chained packet.  If an eDMA descriptor is not chained and
+     * the .hwb bit is not set, this field is ignored and the size is
+     * specified by the .xfer_size field.
+     * 0 = 128 bytes
+     * 1 = 256 bytes
+     * 2 = 512 bytes
+     * 3 = 1024 bytes
+     * 4 = 1664 bytes
+     * 5 = 4096 bytes
+     * 6 = 10368 bytes
+     * 7 = 16384 bytes
+     */
+    uint_reg_t size         : 3;
+    /**
+     * Chaining configuration for the buffer.  Indicates that an ingress
+     * packet or egress command is chained across multiple buffers, with each
+     * buffer's size indicated by the .size field.
+     */
+    uint_reg_t c            : 2;
+#else   /* __BIG_ENDIAN__ */
+    uint_reg_t c            : 2;
+    uint_reg_t size         : 3;
+    uint_reg_t hwb          : 1;
+    uint_reg_t inst         : 2;
+    uint_reg_t __reserved_5 : 3;
+    uint_reg_t stack_idx    : 5;
+    uint_reg_t __reserved_4 : 6;
+    int_reg_t va           : 42;
+#endif
+
+  };
+
+  /** Word access */
+  uint_reg_t words[8];
+} MPIPE_PDESC_t;
+#endif /* !defined(__ASSEMBLER__) */
+
+#endif /* !defined(__ARCH_MPIPE_SHM_H__) */
diff --git a/kernel/arch/tile/include/arch/mpipe_shm_def.h b/kernel/arch/tile/include/arch/mpipe_shm_def.h
new file mode 100644
index 000000000..6124d39c8
--- /dev/null
+++ b/kernel/arch/tile/include/arch/mpipe_shm_def.h
@@ -0,0 +1,23 @@
+/*
+ * Copyright 2012 Tilera Corporation. All Rights Reserved.
+ *
+ *   This program is free software; you can redistribute it and/or
+ *   modify it under the terms of the GNU General Public License
+ *   as published by the Free Software Foundation, version 2.
+ *
+ *   This program is distributed in the hope that it will be useful, but
+ *   WITHOUT ANY WARRANTY; without even the implied warranty of
+ *   MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
+ *   NON INFRINGEMENT.  See the GNU General Public License for
+ *   more details.
+ */
+
+/* Machine-generated file; do not edit. */
+
+#ifndef __ARCH_MPIPE_SHM_DEF_H__
+#define __ARCH_MPIPE_SHM_DEF_H__
+#define MPIPE_EDMA_DESC_WORD1__C_VAL_UNCHAINED 0x0
+#define MPIPE_EDMA_DESC_WORD1__C_VAL_CHAINED 0x1
+#define MPIPE_EDMA_DESC_WORD1__C_VAL_NOT_RDY 0x2
+#define MPIPE_EDMA_DESC_WORD1__C_VAL_INVALID 0x3
+#endif /* !defined(__ARCH_MPIPE_SHM_DEF_H__) */
diff --git a/kernel/arch/tile/include/arch/spr_def.h b/kernel/arch/tile/include/arch/spr_def.h
new file mode 100644
index 000000000..2de83e7af
--- /dev/null
+++ b/kernel/arch/tile/include/arch/spr_def.h
@@ -0,0 +1,109 @@
+/*
+ * Copyright 2010 Tilera Corporation. All Rights Reserved.
+ *
+ *   This program is free software; you can redistribute it and/or
+ *   modify it under the terms of the GNU General Public License
+ *   as published by the Free Software Foundation, version 2.
+ *
+ *   This program is distributed in the hope that it will be useful, but
+ *   WITHOUT ANY WARRANTY; without even the implied warranty of
+ *   MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
+ *   NON INFRINGEMENT.  See the GNU General Public License for
+ *   more details.
+ */
+#ifndef __ARCH_SPR_DEF_H__
+#define __ARCH_SPR_DEF_H__
+
+#include <uapi/arch/spr_def.h>
+
+
+/*
+ * In addition to including the proper base SPR definition file, depending
+ * on machine architecture, this file defines several macros which allow
+ * kernel code to use protection-level dependent SPRs without worrying
+ * about which PL it's running at.  In these macros, the PL that the SPR
+ * or interrupt number applies to is replaced by K.
+ */
+
+#if CONFIG_KERNEL_PL != 1 && CONFIG_KERNEL_PL != 2
+#error CONFIG_KERNEL_PL must be 1 or 2
+#endif
+
+/* Concatenate 4 strings. */
+#define __concat4(a, b, c, d) a ## b ## c ## d
+#define _concat4(a, b, c, d)  __concat4(a, b, c, d)
+
+#ifdef __tilegx__
+
+/* TILE-Gx dependent, protection-level dependent SPRs. */
+
+#define SPR_INTERRUPT_MASK_K \
+	_concat4(SPR_INTERRUPT_MASK_, CONFIG_KERNEL_PL,,)
+#define SPR_INTERRUPT_MASK_SET_K \
+	_concat4(SPR_INTERRUPT_MASK_SET_, CONFIG_KERNEL_PL,,)
+#define SPR_INTERRUPT_MASK_RESET_K \
+	_concat4(SPR_INTERRUPT_MASK_RESET_, CONFIG_KERNEL_PL,,)
+#define SPR_INTERRUPT_VECTOR_BASE_K \
+	_concat4(SPR_INTERRUPT_VECTOR_BASE_, CONFIG_KERNEL_PL,,)
+
+#define SPR_IPI_MASK_K \
+	_concat4(SPR_IPI_MASK_, CONFIG_KERNEL_PL,,)
+#define SPR_IPI_MASK_RESET_K \
+	_concat4(SPR_IPI_MASK_RESET_, CONFIG_KERNEL_PL,,)
+#define SPR_IPI_MASK_SET_K \
+	_concat4(SPR_IPI_MASK_SET_, CONFIG_KERNEL_PL,,)
+#define SPR_IPI_EVENT_K \
+	_concat4(SPR_IPI_EVENT_, CONFIG_KERNEL_PL,,)
+#define SPR_IPI_EVENT_RESET_K \
+	_concat4(SPR_IPI_EVENT_RESET_, CONFIG_KERNEL_PL,,)
+#define SPR_IPI_EVENT_SET_K \
+	_concat4(SPR_IPI_EVENT_SET_, CONFIG_KERNEL_PL,,)
+#define INT_IPI_K \
+	_concat4(INT_IPI_, CONFIG_KERNEL_PL,,)
+
+#define SPR_SINGLE_STEP_CONTROL_K \
+	_concat4(SPR_SINGLE_STEP_CONTROL_, CONFIG_KERNEL_PL,,)
+#define SPR_SINGLE_STEP_EN_K_K \
+	_concat4(SPR_SINGLE_STEP_EN_, CONFIG_KERNEL_PL, _, CONFIG_KERNEL_PL)
+#define INT_SINGLE_STEP_K \
+	_concat4(INT_SINGLE_STEP_, CONFIG_KERNEL_PL,,)
+
+#else
+
+/* TILEPro dependent, protection-level dependent SPRs. */
+
+#define SPR_INTERRUPT_MASK_K_0 \
+	_concat4(SPR_INTERRUPT_MASK_, CONFIG_KERNEL_PL, _0,)
+#define SPR_INTERRUPT_MASK_K_1 \
+	_concat4(SPR_INTERRUPT_MASK_, CONFIG_KERNEL_PL, _1,)
+#define SPR_INTERRUPT_MASK_SET_K_0 \
+	_concat4(SPR_INTERRUPT_MASK_SET_, CONFIG_KERNEL_PL, _0,)
+#define SPR_INTERRUPT_MASK_SET_K_1 \
+	_concat4(SPR_INTERRUPT_MASK_SET_, CONFIG_KERNEL_PL, _1,)
+#define SPR_INTERRUPT_MASK_RESET_K_0 \
+	_concat4(SPR_INTERRUPT_MASK_RESET_, CONFIG_KERNEL_PL, _0,)
+#define SPR_INTERRUPT_MASK_RESET_K_1 \
+	_concat4(SPR_INTERRUPT_MASK_RESET_, CONFIG_KERNEL_PL, _1,)
+
+#endif
+
+/* Generic protection-level dependent SPRs. */
+
+#define SPR_SYSTEM_SAVE_K_0 \
+	_concat4(SPR_SYSTEM_SAVE_, CONFIG_KERNEL_PL, _0,)
+#define SPR_SYSTEM_SAVE_K_1 \
+	_concat4(SPR_SYSTEM_SAVE_, CONFIG_KERNEL_PL, _1,)
+#define SPR_SYSTEM_SAVE_K_2 \
+	_concat4(SPR_SYSTEM_SAVE_, CONFIG_KERNEL_PL, _2,)
+#define SPR_SYSTEM_SAVE_K_3 \
+	_concat4(SPR_SYSTEM_SAVE_, CONFIG_KERNEL_PL, _3,)
+#define SPR_EX_CONTEXT_K_0 \
+	_concat4(SPR_EX_CONTEXT_, CONFIG_KERNEL_PL, _0,)
+#define SPR_EX_CONTEXT_K_1 \
+	_concat4(SPR_EX_CONTEXT_, CONFIG_KERNEL_PL, _1,)
+#define SPR_INTCTRL_K_STATUS \
+	_concat4(SPR_INTCTRL_, CONFIG_KERNEL_PL, _STATUS,)
+#define INT_INTCTRL_K \
+	_concat4(INT_INTCTRL_, CONFIG_KERNEL_PL,,)
+
+#endif /* __ARCH_SPR_DEF_H__ */
diff --git a/kernel/arch/tile/include/arch/trio.h b/kernel/arch/tile/include/arch/trio.h
new file mode 100644
index 000000000..c0ddedcae
--- /dev/null
+++ b/kernel/arch/tile/include/arch/trio.h
@@ -0,0 +1,111 @@
+/*
+ * Copyright 2012 Tilera Corporation. All Rights Reserved.
+ *
+ *   This program is free software; you can redistribute it and/or
+ *   modify it under the terms of the GNU General Public License
+ *   as published by the Free Software Foundation, version 2.
+ *
+ *   This program is distributed in the hope that it will be useful, but
+ *   WITHOUT ANY WARRANTY; without even the implied warranty of
+ *   MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
+ *   NON INFRINGEMENT.  See the GNU General Public License for
+ *   more details.
+ */
+
+/* Machine-generated file; do not edit. */
+
+#ifndef __ARCH_TRIO_H__
+#define __ARCH_TRIO_H__
+
+#include <arch/abi.h>
+#include <arch/trio_def.h>
+
+#ifndef __ASSEMBLER__
+
+/*
+ * Map SQ Doorbell Format.
+ * This describes the format of the write-only doorbell register that exists
+ * in the last 8-bytes of the MAP_SQ_BASE/LIM range.  This register is only
+ * writable from PCIe space.  Writes to this register will not be written to
+ * Tile memory space and thus no IO VA translation is required if the last
+ * page of the BASE/LIM range is not otherwise written.
+ */
+
+__extension__
+typedef union
+{
+  struct
+  {
+#ifndef __BIG_ENDIAN__
+    /*
+     * When written with a 1, the associated MAP_SQ region's doorbell
+     * interrupt will be triggered once all previous writes are visible to
+     * Tile software.
+     */
+    uint_reg_t doorbell   : 1;
+    /*
+     * When written with a 1, the descriptor at the head of the associated
+     * MAP_SQ's FIFO will be dequeued.
+     */
+    uint_reg_t pop        : 1;
+    /* Reserved. */
+    uint_reg_t __reserved : 62;
+#else   /* __BIG_ENDIAN__ */
+    uint_reg_t __reserved : 62;
+    uint_reg_t pop        : 1;
+    uint_reg_t doorbell   : 1;
+#endif
+  };
+
+  uint_reg_t word;
+} TRIO_MAP_SQ_DOORBELL_FMT_t;
+
+
+/*
+ * Tile PIO Region Configuration - CFG Address Format.
+ * This register describes the address format for PIO accesses when the
+ * associated region is setup with TYPE=CFG.
+ */
+
+__extension__
+typedef union
+{
+  struct
+  {
+#ifndef __BIG_ENDIAN__
+    /* Register Address (full byte address). */
+    uint_reg_t reg_addr     : 12;
+    /* Function Number */
+    uint_reg_t fn           : 3;
+    /* Device Number */
+    uint_reg_t dev          : 5;
+    /* BUS Number */
+    uint_reg_t bus          : 8;
+    /* Config Type: 0 for access to directly-attached device.  1 otherwise. */
+    uint_reg_t type         : 1;
+    /* Reserved. */
+    uint_reg_t __reserved_0 : 1;
+    /*
+     * MAC select.  This must match the configuration in
+     * TILE_PIO_REGION_SETUP.MAC.
+     */
+    uint_reg_t mac          : 2;
+    /* Reserved. */
+    uint_reg_t __reserved_1 : 32;
+#else   /* __BIG_ENDIAN__ */
+    uint_reg_t __reserved_1 : 32;
+    uint_reg_t mac          : 2;
+    uint_reg_t __reserved_0 : 1;
+    uint_reg_t type         : 1;
+    uint_reg_t bus          : 8;
+    uint_reg_t dev          : 5;
+    uint_reg_t fn           : 3;
+    uint_reg_t reg_addr     : 12;
+#endif
+  };
+
+  uint_reg_t word;
+} TRIO_TILE_PIO_REGION_SETUP_CFG_ADDR_t;
+#endif /* !defined(__ASSEMBLER__) */
+
+#endif /* !defined(__ARCH_TRIO_H__) */
diff --git a/kernel/arch/tile/include/arch/trio_constants.h b/kernel/arch/tile/include/arch/trio_constants.h
new file mode 100644
index 000000000..85647e91a
--- /dev/null
+++ b/kernel/arch/tile/include/arch/trio_constants.h
@@ -0,0 +1,36 @@
+/*
+ * Copyright 2012 Tilera Corporation. All Rights Reserved.
+ *
+ *   This program is free software; you can redistribute it and/or
+ *   modify it under the terms of the GNU General Public License
+ *   as published by the Free Software Foundation, version 2.
+ *
+ *   This program is distributed in the hope that it will be useful, but
+ *   WITHOUT ANY WARRANTY; without even the implied warranty of
+ *   MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
+ *   NON INFRINGEMENT.  See the GNU General Public License for
+ *   more details.
+ */
+
+
+#ifndef __ARCH_TRIO_CONSTANTS_H__
+#define __ARCH_TRIO_CONSTANTS_H__
+
+#define TRIO_NUM_ASIDS 32
+#define TRIO_NUM_TLBS_PER_ASID 16
+
+#define TRIO_NUM_TPIO_REGIONS 8
+#define TRIO_LOG2_NUM_TPIO_REGIONS 3
+
+#define TRIO_NUM_MAP_MEM_REGIONS 32
+#define TRIO_LOG2_NUM_MAP_MEM_REGIONS 5
+#define TRIO_NUM_MAP_SQ_REGIONS 8
+#define TRIO_LOG2_NUM_MAP_SQ_REGIONS 3
+
+#define TRIO_LOG2_NUM_SQ_FIFO_ENTRIES 6
+
+#define TRIO_NUM_PUSH_DMA_RINGS 64
+
+#define TRIO_NUM_PULL_DMA_RINGS 64
+
+#endif /* __ARCH_TRIO_CONSTANTS_H__ */
diff --git a/kernel/arch/tile/include/arch/trio_def.h b/kernel/arch/tile/include/arch/trio_def.h
new file mode 100644
index 000000000..e80500317
--- /dev/null
+++ b/kernel/arch/tile/include/arch/trio_def.h
@@ -0,0 +1,41 @@
+/*
+ * Copyright 2012 Tilera Corporation. All Rights Reserved.
+ *
+ *   This program is free software; you can redistribute it and/or
+ *   modify it under the terms of the GNU General Public License
+ *   as published by the Free Software Foundation, version 2.
+ *
+ *   This program is distributed in the hope that it will be useful, but
+ *   WITHOUT ANY WARRANTY; without even the implied warranty of
+ *   MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
+ *   NON INFRINGEMENT.  See the GNU General Public License for
+ *   more details.
+ */
+
+/* Machine-generated file; do not edit. */
+
+#ifndef __ARCH_TRIO_DEF_H__
+#define __ARCH_TRIO_DEF_H__
+#define TRIO_CFG_REGION_ADDR__REG_SHIFT 0
+#define TRIO_CFG_REGION_ADDR__INTFC_SHIFT 16
+#define TRIO_CFG_REGION_ADDR__INTFC_VAL_TRIO 0x0
+#define TRIO_CFG_REGION_ADDR__INTFC_VAL_MAC_INTERFACE 0x1
+#define TRIO_CFG_REGION_ADDR__INTFC_VAL_MAC_STANDARD 0x2
+#define TRIO_CFG_REGION_ADDR__INTFC_VAL_MAC_PROTECTED 0x3
+#define TRIO_CFG_REGION_ADDR__MAC_SEL_SHIFT 18
+#define TRIO_CFG_REGION_ADDR__PROT_SHIFT 20
+#define TRIO_PIO_REGIONS_ADDR__REGION_SHIFT 32
+#define TRIO_MAP_MEM_REG_INT0 0x1000000000
+#define TRIO_MAP_MEM_REG_INT1 0x1000000008
+#define TRIO_MAP_MEM_REG_INT2 0x1000000010
+#define TRIO_MAP_MEM_REG_INT3 0x1000000018
+#define TRIO_MAP_MEM_REG_INT4 0x1000000020
+#define TRIO_MAP_MEM_REG_INT5 0x1000000028
+#define TRIO_MAP_MEM_REG_INT6 0x1000000030
+#define TRIO_MAP_MEM_REG_INT7 0x1000000038
+#define TRIO_MAP_MEM_LIM__ADDR_SHIFT 12
+#define TRIO_MAP_MEM_SETUP__ORDER_MODE_VAL_UNORDERED 0x0
+#define TRIO_MAP_MEM_SETUP__ORDER_MODE_VAL_STRICT 0x1
+#define TRIO_MAP_MEM_SETUP__ORDER_MODE_VAL_REL_ORD 0x2
+#define TRIO_TILE_PIO_REGION_SETUP_CFG_ADDR__MAC_SHIFT 30
+#endif /* !defined(__ARCH_TRIO_DEF_H__) */
diff --git a/kernel/arch/tile/include/arch/trio_pcie_intfc.h b/kernel/arch/tile/include/arch/trio_pcie_intfc.h
new file mode 100644
index 000000000..0487fdb9d
--- /dev/null
+++ b/kernel/arch/tile/include/arch/trio_pcie_intfc.h
@@ -0,0 +1,229 @@
+/*
+ * Copyright 2012 Tilera Corporation. All Rights Reserved.
+ *
+ *   This program is free software; you can redistribute it and/or
+ *   modify it under the terms of the GNU General Public License
+ *   as published by the Free Software Foundation, version 2.
+ *
+ *   This program is distributed in the hope that it will be useful, but
+ *   WITHOUT ANY WARRANTY; without even the implied warranty of
+ *   MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
+ *   NON INFRINGEMENT.  See the GNU General Public License for
+ *   more details.
+ */
+
+/* Machine-generated file; do not edit. */
+
+#ifndef __ARCH_TRIO_PCIE_INTFC_H__
+#define __ARCH_TRIO_PCIE_INTFC_H__
+
+#include <arch/abi.h>
+#include <arch/trio_pcie_intfc_def.h>
+
+#ifndef __ASSEMBLER__
+
+/*
+ * Port Configuration.
+ * Configuration of the PCIe Port
+ */
+
+__extension__
+typedef union
+{
+  struct
+  {
+#ifndef __BIG_ENDIAN__
+    /* Provides the state of the strapping pins for this port. */
+    uint_reg_t strap_state      : 3;
+    /* Reserved. */
+    uint_reg_t __reserved_0     : 1;
+    /*
+     * When 1, the device type will be overridden using OVD_DEV_TYPE_VAL.
+     * When 0, the device type is determined based on the STRAP_STATE.
+     */
+    uint_reg_t ovd_dev_type     : 1;
+    /* Provides the device type when OVD_DEV_TYPE is 1. */
+    uint_reg_t ovd_dev_type_val : 4;
+    /* Determines how link is trained. */
+    uint_reg_t train_mode       : 2;
+    /* Reserved. */
+    uint_reg_t __reserved_1     : 1;
+    /*
+     * For PCIe, used to flip physical RX lanes that were not properly wired.
+     *  This is not the same as lane reversal which is handled automatically
+     * during link training.  When 0, RX Lane0 must be wired to the link
+     * partner (either to its Lane0 or it's LaneN).  When RX_LANE_FLIP is 1,
+     * the highest numbered lane for this port becomes Lane0 and Lane0 does
+     * NOT have to be wired to the link partner.
+     */
+    uint_reg_t rx_lane_flip     : 1;
+    /*
+     * For PCIe, used to flip physical TX lanes that were not properly wired.
+     *  This is not the same as lane reversal which is handled automatically
+     * during link training.  When 0, TX Lane0 must be wired to the link
+     * partner (either to its Lane0 or it's LaneN).  When TX_LANE_FLIP is 1,
+     * the highest numbered lane for this port becomes Lane0 and Lane0 does
+     * NOT have to be wired to the link partner.
+     */
+    uint_reg_t tx_lane_flip     : 1;
+    /*
+     * For StreamIO port, configures the width of the port when TRAIN_MODE is
+     * not STRAP.
+     */
+    uint_reg_t stream_width     : 2;
+    /*
+     * For StreamIO port, configures the rate of the port when TRAIN_MODE is
+     * not STRAP.
+     */
+    uint_reg_t stream_rate      : 2;
+    /* Reserved. */
+    uint_reg_t __reserved_2     : 46;
+#else   /* __BIG_ENDIAN__ */
+    uint_reg_t __reserved_2     : 46;
+    uint_reg_t stream_rate      : 2;
+    uint_reg_t stream_width     : 2;
+    uint_reg_t tx_lane_flip     : 1;
+    uint_reg_t rx_lane_flip     : 1;
+    uint_reg_t __reserved_1     : 1;
+    uint_reg_t train_mode       : 2;
+    uint_reg_t ovd_dev_type_val : 4;
+    uint_reg_t ovd_dev_type     : 1;
+    uint_reg_t __reserved_0     : 1;
+    uint_reg_t strap_state      : 3;
+#endif
+  };
+
+  uint_reg_t word;
+} TRIO_PCIE_INTFC_PORT_CONFIG_t;
+
+/*
+ * Port Status.
+ * Status of the PCIe Port.  This register applies to the StreamIO port when
+ * StreamIO is enabled.
+ */
+
+__extension__
+typedef union
+{
+  struct
+  {
+#ifndef __BIG_ENDIAN__
+    /*
+     * Indicates the DL state of the port.  When 1, the port is up and ready
+     * to receive traffic.
+     */
+    uint_reg_t dl_up        : 1;
+    /*
+     * Indicates the number of times the link has gone down.  Clears on read.
+     */
+    uint_reg_t dl_down_cnt  : 7;
+    /* Indicates the SERDES PLL has spun up and is providing a valid clock. */
+    uint_reg_t clock_ready  : 1;
+    /* Reserved. */
+    uint_reg_t __reserved_0 : 7;
+    /* Device revision ID. */
+    uint_reg_t device_rev   : 8;
+    /* Link state (PCIe). */
+    uint_reg_t ltssm_state  : 6;
+    /* Link power management state (PCIe). */
+    uint_reg_t pm_state     : 3;
+    /* Reserved. */
+    uint_reg_t __reserved_1 : 31;
+#else   /* __BIG_ENDIAN__ */
+    uint_reg_t __reserved_1 : 31;
+    uint_reg_t pm_state     : 3;
+    uint_reg_t ltssm_state  : 6;
+    uint_reg_t device_rev   : 8;
+    uint_reg_t __reserved_0 : 7;
+    uint_reg_t clock_ready  : 1;
+    uint_reg_t dl_down_cnt  : 7;
+    uint_reg_t dl_up        : 1;
+#endif
+  };
+
+  uint_reg_t word;
+} TRIO_PCIE_INTFC_PORT_STATUS_t;
+
+/*
+ * Transmit FIFO Control.
+ * Contains TX FIFO thresholds.  These registers are for diagnostics purposes
+ * only.  Changing these values causes undefined behavior.
+ */
+
+__extension__
+typedef union
+{
+  struct
+  {
+#ifndef __BIG_ENDIAN__
+    /*
+     * Almost-Empty level for TX0 data.  Typically set to at least
+     * roundup(38.0*M/N) where N=tclk frequency and M=MAC symbol rate in MHz
+     * for a x4 port (250MHz).
+     */
+    uint_reg_t tx0_data_ae_lvl : 7;
+    /* Reserved. */
+    uint_reg_t __reserved_0    : 1;
+    /* Almost-Empty level for TX1 data. */
+    uint_reg_t tx1_data_ae_lvl : 7;
+    /* Reserved. */
+    uint_reg_t __reserved_1    : 1;
+    /* Almost-Full level for TX0 data. */
+    uint_reg_t tx0_data_af_lvl : 7;
+    /* Reserved. */
+    uint_reg_t __reserved_2    : 1;
+    /* Almost-Full level for TX1 data. */
+    uint_reg_t tx1_data_af_lvl : 7;
+    /* Reserved. */
+    uint_reg_t __reserved_3    : 1;
+    /* Almost-Full level for TX0 info. */
+    uint_reg_t tx0_info_af_lvl : 5;
+    /* Reserved. */
+    uint_reg_t __reserved_4    : 3;
+    /* Almost-Full level for TX1 info. */
+    uint_reg_t tx1_info_af_lvl : 5;
+    /* Reserved. */
+    uint_reg_t __reserved_5    : 3;
+    /*
+     * This register provides performance adjustment for high bandwidth
+     * flows.  The MAC will assert almost-full to TRIO if non-posted credits
+     * fall below this level.  Note that setting this larger than the initial
+     * PORT_CREDIT.NPH value will cause READS to never be sent.  If the
+     * initial credit value from the link partner is smaller than this value
+     * when the link comes up, the value will be reset to the initial credit
+     * value to prevent lockup.
+     */
+    uint_reg_t min_np_credits  : 8;
+    /*
+     * This register provides performance adjustment for high bandwidth
+     * flows.  The MAC will assert almost-full to TRIO if posted credits fall
+     * below this level.  Note that setting this larger than the initial
+     * PORT_CREDIT.PH value will cause WRITES to never be sent.  If the
+     * initial credit value from the link partner is smaller than this value
+     * when the link comes up, the value will be reset to the initial credit
+     * value to prevent lockup.
+     */
+    uint_reg_t min_p_credits   : 8;
+#else   /* __BIG_ENDIAN__ */
+    uint_reg_t min_p_credits   : 8;
+    uint_reg_t min_np_credits  : 8;
+    uint_reg_t __reserved_5    : 3;
+    uint_reg_t tx1_info_af_lvl : 5;
+    uint_reg_t __reserved_4    : 3;
+    uint_reg_t tx0_info_af_lvl : 5;
+    uint_reg_t __reserved_3    : 1;
+    uint_reg_t tx1_data_af_lvl : 7;
+    uint_reg_t __reserved_2    : 1;
+    uint_reg_t tx0_data_af_lvl : 7;
+    uint_reg_t __reserved_1    : 1;
+    uint_reg_t tx1_data_ae_lvl : 7;
+    uint_reg_t __reserved_0    : 1;
+    uint_reg_t tx0_data_ae_lvl : 7;
+#endif
+  };
+
+  uint_reg_t word;
+} TRIO_PCIE_INTFC_TX_FIFO_CTL_t;
+#endif /* !defined(__ASSEMBLER__) */
+
+#endif /* !defined(__ARCH_TRIO_PCIE_INTFC_H__) */
diff --git a/kernel/arch/tile/include/arch/trio_pcie_intfc_def.h b/kernel/arch/tile/include/arch/trio_pcie_intfc_def.h
new file mode 100644
index 000000000..d3fd6781f
--- /dev/null
+++ b/kernel/arch/tile/include/arch/trio_pcie_intfc_def.h
@@ -0,0 +1,32 @@
+/*
+ * Copyright 2012 Tilera Corporation. All Rights Reserved.
+ *
+ *   This program is free software; you can redistribute it and/or
+ *   modify it under the terms of the GNU General Public License
+ *   as published by the Free Software Foundation, version 2.
+ *
+ *   This program is distributed in the hope that it will be useful, but
+ *   WITHOUT ANY WARRANTY; without even the implied warranty of
+ *   MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
+ *   NON INFRINGEMENT.  See the GNU General Public License for
+ *   more details.
+ */
+
+/* Machine-generated file; do not edit. */
+
+#ifndef __ARCH_TRIO_PCIE_INTFC_DEF_H__
+#define __ARCH_TRIO_PCIE_INTFC_DEF_H__
+#define TRIO_PCIE_INTFC_MAC_INT_STS 0x0000
+#define TRIO_PCIE_INTFC_MAC_INT_STS__INT_LEVEL_MASK  0xf000
+#define TRIO_PCIE_INTFC_PORT_CONFIG 0x0018
+#define TRIO_PCIE_INTFC_PORT_CONFIG__STRAP_STATE_VAL_DISABLED 0x0
+#define TRIO_PCIE_INTFC_PORT_CONFIG__STRAP_STATE_VAL_AUTO_CONFIG_ENDPOINT 0x1
+#define TRIO_PCIE_INTFC_PORT_CONFIG__STRAP_STATE_VAL_AUTO_CONFIG_RC 0x2
+#define TRIO_PCIE_INTFC_PORT_CONFIG__STRAP_STATE_VAL_AUTO_CONFIG_ENDPOINT_G1 0x3
+#define TRIO_PCIE_INTFC_PORT_CONFIG__STRAP_STATE_VAL_AUTO_CONFIG_RC_G1 0x4
+#define TRIO_PCIE_INTFC_PORT_CONFIG__STRAP_STATE_VAL_AUTO_XLINK 0x5
+#define TRIO_PCIE_INTFC_PORT_CONFIG__STRAP_STATE_VAL_STREAM_X1 0x6
+#define TRIO_PCIE_INTFC_PORT_CONFIG__STRAP_STATE_VAL_STREAM_X4 0x7
+#define TRIO_PCIE_INTFC_PORT_STATUS 0x0020
+#define TRIO_PCIE_INTFC_TX_FIFO_CTL 0x0050
+#endif /* !defined(__ARCH_TRIO_PCIE_INTFC_DEF_H__) */
diff --git a/kernel/arch/tile/include/arch/trio_pcie_rc.h b/kernel/arch/tile/include/arch/trio_pcie_rc.h
new file mode 100644
index 000000000..6a25d0aca
--- /dev/null
+++ b/kernel/arch/tile/include/arch/trio_pcie_rc.h
@@ -0,0 +1,156 @@
+/*
+ * Copyright 2012 Tilera Corporation. All Rights Reserved.
+ *
+ *   This program is free software; you can redistribute it and/or
+ *   modify it under the terms of the GNU General Public License
+ *   as published by the Free Software Foundation, version 2.
+ *
+ *   This program is distributed in the hope that it will be useful, but
+ *   WITHOUT ANY WARRANTY; without even the implied warranty of
+ *   MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
+ *   NON INFRINGEMENT.  See the GNU General Public License for
+ *   more details.
+ */
+
+/* Machine-generated file; do not edit. */
+
+#ifndef __ARCH_TRIO_PCIE_RC_H__
+#define __ARCH_TRIO_PCIE_RC_H__
+
+#include <arch/abi.h>
+#include <arch/trio_pcie_rc_def.h>
+
+#ifndef __ASSEMBLER__
+
+/* Device Capabilities Register. */
+
+__extension__
+typedef union
+{
+  struct
+  {
+#ifndef __BIG_ENDIAN__
+    /*
+     * Max_Payload_Size Supported, writablethrough the MAC_STANDARD interface
+     */
+    uint_reg_t mps_sup                    : 3;
+    /*
+     * This field is writable through the MAC_STANDARD interface.  However,
+     * Phantom Function is not  supported. Therefore, the application must
+     * not write any value other than 0x0 to this  field.
+     */
+    uint_reg_t phantom_function_supported : 2;
+    /* This bit is writable through the MAC_STANDARD interface. */
+    uint_reg_t ext_tag_field_supported    : 1;
+    /* Reserved. */
+    uint_reg_t __reserved_0               : 3;
+    /* Endpoint L1 Acceptable Latency Must be 0x0 for non-Endpoint devices. */
+    uint_reg_t l1_lat                     : 3;
+    /*
+     * Undefined since PCI Express 1.1 (Was Attention Button Present for PCI
+     * Express 1.0a)
+     */
+    uint_reg_t r1                         : 1;
+    /*
+     * Undefined since PCI Express 1.1 (Was Attention Indicator Present for
+     * PCI  Express 1.0a)
+     */
+    uint_reg_t r2                         : 1;
+    /*
+     * Undefined since PCI Express 1.1 (Was Power Indicator Present for PCI
+     * Express 1.0a)
+     */
+    uint_reg_t r3                         : 1;
+    /*
+     * Role-Based Error Reporting, writable through the MAC_STANDARD
+     * interface.  Required to be set for device compliant to 1.1  spec and
+     * later.
+     */
+    uint_reg_t rer                        : 1;
+    /* Reserved. */
+    uint_reg_t __reserved_1               : 2;
+    /* Captured Slot Power Limit Value Upstream port only. */
+    uint_reg_t slot_pwr_lim               : 8;
+    /* Captured Slot Power Limit Scale Upstream port only. */
+    uint_reg_t slot_pwr_scale             : 2;
+    /* Reserved. */
+    uint_reg_t __reserved_2               : 4;
+    /* Endpoint L0s Acceptable LatencyMust be 0x0 for non-Endpoint devices. */
+    uint_reg_t l0s_lat                    : 1;
+    /* Reserved. */
+    uint_reg_t __reserved_3               : 31;
+#else   /* __BIG_ENDIAN__ */
+    uint_reg_t __reserved_3               : 31;
+    uint_reg_t l0s_lat                    : 1;
+    uint_reg_t __reserved_2               : 4;
+    uint_reg_t slot_pwr_scale             : 2;
+    uint_reg_t slot_pwr_lim               : 8;
+    uint_reg_t __reserved_1               : 2;
+    uint_reg_t rer                        : 1;
+    uint_reg_t r3                         : 1;
+    uint_reg_t r2                         : 1;
+    uint_reg_t r1                         : 1;
+    uint_reg_t l1_lat                     : 3;
+    uint_reg_t __reserved_0               : 3;
+    uint_reg_t ext_tag_field_supported    : 1;
+    uint_reg_t phantom_function_supported : 2;
+    uint_reg_t mps_sup                    : 3;
+#endif
+  };
+
+  uint_reg_t word;
+} TRIO_PCIE_RC_DEVICE_CAP_t;
+
+/* Device Control Register. */
+
+__extension__
+typedef union
+{
+  struct
+  {
+#ifndef __BIG_ENDIAN__
+    /* Correctable Error Reporting Enable */
+    uint_reg_t cor_err_ena      : 1;
+    /* Non-Fatal Error Reporting Enable */
+    uint_reg_t nf_err_ena       : 1;
+    /* Fatal Error Reporting Enable */
+    uint_reg_t fatal_err_ena    : 1;
+    /* Unsupported Request Reporting Enable */
+    uint_reg_t ur_ena           : 1;
+    /* Relaxed orderring enable */
+    uint_reg_t ro_ena           : 1;
+    /* Max Payload Size */
+    uint_reg_t max_payload_size : 3;
+    /* Extended Tag Field Enable */
+    uint_reg_t ext_tag          : 1;
+    /* Phantom Function Enable */
+    uint_reg_t ph_fn_ena        : 1;
+    /* AUX Power PM Enable */
+    uint_reg_t aux_pm_ena       : 1;
+    /* Enable NoSnoop */
+    uint_reg_t no_snoop         : 1;
+    /* Max read request size */
+    uint_reg_t max_read_req_sz  : 3;
+    /* Reserved. */
+    uint_reg_t __reserved       : 49;
+#else   /* __BIG_ENDIAN__ */
+    uint_reg_t __reserved       : 49;
+    uint_reg_t max_read_req_sz  : 3;
+    uint_reg_t no_snoop         : 1;
+    uint_reg_t aux_pm_ena       : 1;
+    uint_reg_t ph_fn_ena        : 1;
+    uint_reg_t ext_tag          : 1;
+    uint_reg_t max_payload_size : 3;
+    uint_reg_t ro_ena           : 1;
+    uint_reg_t ur_ena           : 1;
+    uint_reg_t fatal_err_ena    : 1;
+    uint_reg_t nf_err_ena       : 1;
+    uint_reg_t cor_err_ena      : 1;
+#endif
+  };
+
+  uint_reg_t word;
+} TRIO_PCIE_RC_DEVICE_CONTROL_t;
+#endif /* !defined(__ASSEMBLER__) */
+
+#endif /* !defined(__ARCH_TRIO_PCIE_RC_H__) */
diff --git a/kernel/arch/tile/include/arch/trio_pcie_rc_def.h b/kernel/arch/tile/include/arch/trio_pcie_rc_def.h
new file mode 100644
index 000000000..74081a65b
--- /dev/null
+++ b/kernel/arch/tile/include/arch/trio_pcie_rc_def.h
@@ -0,0 +1,24 @@
+/*
+ * Copyright 2012 Tilera Corporation. All Rights Reserved.
+ *
+ *   This program is free software; you can redistribute it and/or
+ *   modify it under the terms of the GNU General Public License
+ *   as published by the Free Software Foundation, version 2.
+ *
+ *   This program is distributed in the hope that it will be useful, but
+ *   WITHOUT ANY WARRANTY; without even the implied warranty of
+ *   MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
+ *   NON INFRINGEMENT.  See the GNU General Public License for
+ *   more details.
+ */
+
+/* Machine-generated file; do not edit. */
+
+#ifndef __ARCH_TRIO_PCIE_RC_DEF_H__
+#define __ARCH_TRIO_PCIE_RC_DEF_H__
+#define TRIO_PCIE_RC_DEVICE_CAP 0x0074
+#define TRIO_PCIE_RC_DEVICE_CONTROL 0x0078
+#define TRIO_PCIE_RC_DEVICE_ID_VEN_ID 0x0000
+#define TRIO_PCIE_RC_DEVICE_ID_VEN_ID__DEV_ID_SHIFT 16
+#define TRIO_PCIE_RC_REVISION_ID 0x0008
+#endif /* !defined(__ARCH_TRIO_PCIE_RC_DEF_H__) */
diff --git a/kernel/arch/tile/include/arch/trio_shm.h b/kernel/arch/tile/include/arch/trio_shm.h
new file mode 100644
index 000000000..3382e3824
--- /dev/null
+++ b/kernel/arch/tile/include/arch/trio_shm.h
@@ -0,0 +1,125 @@
+/*
+ * Copyright 2012 Tilera Corporation. All Rights Reserved.
+ *
+ *   This program is free software; you can redistribute it and/or
+ *   modify it under the terms of the GNU General Public License
+ *   as published by the Free Software Foundation, version 2.
+ *
+ *   This program is distributed in the hope that it will be useful, but
+ *   WITHOUT ANY WARRANTY; without even the implied warranty of
+ *   MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
+ *   NON INFRINGEMENT.  See the GNU General Public License for
+ *   more details.
+ */
+
+/* Machine-generated file; do not edit. */
+
+
+#ifndef __ARCH_TRIO_SHM_H__
+#define __ARCH_TRIO_SHM_H__
+
+#include <arch/abi.h>
+#include <arch/trio_shm_def.h>
+
+#ifndef __ASSEMBLER__
+/**
+ * TRIO DMA Descriptor.
+ * The TRIO DMA descriptor is written by software and consumed by hardware.
+ * It is used to specify the location of transaction data in the IO and Tile
+ * domains.
+ */
+
+__extension__
+typedef union
+{
+  struct
+  {
+    /* Word 0 */
+
+#ifndef __BIG_ENDIAN__
+    /** Tile side virtual address. */
+    int_reg_t va           : 42;
+    /**
+     * Encoded size of buffer used on push DMA when C=1:
+     * 0 = 128 bytes
+     * 1 = 256 bytes
+     * 2 = 512 bytes
+     * 3 = 1024 bytes
+     * 4 = 1664 bytes
+     * 5 = 4096 bytes
+     * 6 = 10368 bytes
+     * 7 = 16384 bytes
+     */
+    uint_reg_t bsz          : 3;
+    /**
+     * Chaining designation.  Always zero for pull DMA
+     * 0 : Unchained buffer pointer
+     * 1 : Chained buffer pointer.  Next buffer descriptor (e.g. VA) stored
+     * in 1st 8-bytes in buffer.  For chained buffers, first 8-bytes of each
+     * buffer contain the next buffer descriptor formatted exactly like a PDE
+     * buffer descriptor.  This allows a chained PDE buffer to be sent using
+     * push DMA.
+     */
+    uint_reg_t c            : 1;
+    /**
+     * Notification interrupt will be delivered when the transaction has
+     * completed (all data has been read from or written to the Tile-side
+     * buffer).
+     */
+    uint_reg_t notif        : 1;
+    /**
+     * When 0, the XSIZE field specifies the total byte count for the
+     * transaction.  When 1, the XSIZE field is encoded as 2^(N+14) for N in
+     * {0..6}:
+     * 0 = 16KB
+     * 1 = 32KB
+     * 2 = 64KB
+     * 3 = 128KB
+     * 4 = 256KB
+     * 5 = 512KB
+     * 6 = 1MB
+     * All other encodings of the XSIZE field are reserved when SMOD=1
+     */
+    uint_reg_t smod         : 1;
+    /**
+     * Total number of bytes to move for this transaction.   When SMOD=1,
+     * this field is encoded - see SMOD description.
+     */
+    uint_reg_t xsize        : 14;
+    /** Reserved. */
+    uint_reg_t __reserved_0 : 1;
+    /**
+     * Generation number.  Used to indicate a valid descriptor in ring.  When
+     * a new descriptor is written into the ring, software must toggle this
+     * bit.  The net effect is that the GEN bit being written into new
+     * descriptors toggles each time the ring tail pointer wraps.
+     */
+    uint_reg_t gen          : 1;
+#else   /* __BIG_ENDIAN__ */
+    uint_reg_t gen          : 1;
+    uint_reg_t __reserved_0 : 1;
+    uint_reg_t xsize        : 14;
+    uint_reg_t smod         : 1;
+    uint_reg_t notif        : 1;
+    uint_reg_t c            : 1;
+    uint_reg_t bsz          : 3;
+    int_reg_t va           : 42;
+#endif
+
+    /* Word 1 */
+
+#ifndef __BIG_ENDIAN__
+    /** IO-side address */
+    uint_reg_t io_address : 64;
+#else   /* __BIG_ENDIAN__ */
+    uint_reg_t io_address : 64;
+#endif
+
+  };
+
+  /** Word access */
+  uint_reg_t words[2];
+} TRIO_DMA_DESC_t;
+#endif /* !defined(__ASSEMBLER__) */
+
+#endif /* !defined(__ARCH_TRIO_SHM_H__) */
diff --git a/kernel/arch/tile/include/arch/trio_shm_def.h b/kernel/arch/tile/include/arch/trio_shm_def.h
new file mode 100644
index 000000000..72a59c88b
--- /dev/null
+++ b/kernel/arch/tile/include/arch/trio_shm_def.h
@@ -0,0 +1,19 @@
+/*
+ * Copyright 2012 Tilera Corporation. All Rights Reserved.
+ *
+ *   This program is free software; you can redistribute it and/or
+ *   modify it under the terms of the GNU General Public License
+ *   as published by the Free Software Foundation, version 2.
+ *
+ *   This program is distributed in the hope that it will be useful, but
+ *   WITHOUT ANY WARRANTY; without even the implied warranty of
+ *   MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
+ *   NON INFRINGEMENT.  See the GNU General Public License for
+ *   more details.
+ */
+
+/* Machine-generated file; do not edit. */
+
+#ifndef __ARCH_TRIO_SHM_DEF_H__
+#define __ARCH_TRIO_SHM_DEF_H__
+#endif /* !defined(__ARCH_TRIO_SHM_DEF_H__) */
diff --git a/kernel/arch/tile/include/arch/uart.h b/kernel/arch/tile/include/arch/uart.h
new file mode 100644
index 000000000..07966970a
--- /dev/null
+++ b/kernel/arch/tile/include/arch/uart.h
@@ -0,0 +1,300 @@
+/*
+ * Copyright 2013 Tilera Corporation. All Rights Reserved.
+ *
+ *   This program is free software; you can redistribute it and/or
+ *   modify it under the terms of the GNU General Public License
+ *   as published by the Free Software Foundation, version 2.
+ *
+ *   This program is distributed in the hope that it will be useful, but
+ *   WITHOUT ANY WARRANTY; without even the implied warranty of
+ *   MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
+ *   NON INFRINGEMENT.  See the GNU General Public License for
+ *   more details.
+ */
+
+/* Machine-generated file; do not edit. */
+
+#ifndef __ARCH_UART_H__
+#define __ARCH_UART_H__
+
+#include <arch/abi.h>
+#include <arch/uart_def.h>
+
+#ifndef __ASSEMBLER__
+
+/* Divisor. */
+
+__extension__
+typedef union
+{
+  struct
+  {
+#ifndef __BIG_ENDIAN__
+    /*
+     * Baud Rate Divisor.  Desired_baud_rate = REF_CLK frequency / (baud *
+     * 16).
+     *                       Note: REF_CLK is always 125 MHz, the default
+     * divisor = 68, baud rate = 125M/(68*16) = 115200 baud.
+     */
+    uint_reg_t divisor    : 12;
+    /* Reserved. */
+    uint_reg_t __reserved : 52;
+#else   /* __BIG_ENDIAN__ */
+    uint_reg_t __reserved : 52;
+    uint_reg_t divisor    : 12;
+#endif
+  };
+
+  uint_reg_t word;
+} UART_DIVISOR_t;
+
+/* FIFO Count. */
+
+__extension__
+typedef union
+{
+  struct
+  {
+#ifndef __BIG_ENDIAN__
+    /*
+     * n: n active entries in the receive FIFO (max is 2**8). Each entry has
+     * 8 bits.
+     * 0: no active entry in the receive FIFO (that is empty).
+     */
+    uint_reg_t rfifo_count  : 9;
+    /* Reserved. */
+    uint_reg_t __reserved_0 : 7;
+    /*
+     * n: n active entries in the transmit FIFO (max is 2**8). Each entry has
+     * 8 bits.
+     * 0: no active entry in the transmit FIFO (that is empty).
+     */
+    uint_reg_t tfifo_count  : 9;
+    /* Reserved. */
+    uint_reg_t __reserved_1 : 7;
+    /*
+     * n: n active entries in the write FIFO (max is 2**2). Each entry has 8
+     * bits.
+     * 0: no active entry in the write FIFO (that is empty).
+     */
+    uint_reg_t wfifo_count  : 3;
+    /* Reserved. */
+    uint_reg_t __reserved_2 : 29;
+#else   /* __BIG_ENDIAN__ */
+    uint_reg_t __reserved_2 : 29;
+    uint_reg_t wfifo_count  : 3;
+    uint_reg_t __reserved_1 : 7;
+    uint_reg_t tfifo_count  : 9;
+    uint_reg_t __reserved_0 : 7;
+    uint_reg_t rfifo_count  : 9;
+#endif
+  };
+
+  uint_reg_t word;
+} UART_FIFO_COUNT_t;
+
+/* FLAG. */
+
+__extension__
+typedef union
+{
+  struct
+  {
+#ifndef __BIG_ENDIAN__
+    /* Reserved. */
+    uint_reg_t __reserved_0 : 1;
+    /* 1: receive FIFO is empty */
+    uint_reg_t rfifo_empty  : 1;
+    /* 1: write FIFO is empty. */
+    uint_reg_t wfifo_empty  : 1;
+    /* 1: transmit FIFO is empty. */
+    uint_reg_t tfifo_empty  : 1;
+    /* 1: receive FIFO is full. */
+    uint_reg_t rfifo_full   : 1;
+    /* 1: write FIFO is full. */
+    uint_reg_t wfifo_full   : 1;
+    /* 1: transmit FIFO is full. */
+    uint_reg_t tfifo_full   : 1;
+    /* Reserved. */
+    uint_reg_t __reserved_1 : 57;
+#else   /* __BIG_ENDIAN__ */
+    uint_reg_t __reserved_1 : 57;
+    uint_reg_t tfifo_full   : 1;
+    uint_reg_t wfifo_full   : 1;
+    uint_reg_t rfifo_full   : 1;
+    uint_reg_t tfifo_empty  : 1;
+    uint_reg_t wfifo_empty  : 1;
+    uint_reg_t rfifo_empty  : 1;
+    uint_reg_t __reserved_0 : 1;
+#endif
+  };
+
+  uint_reg_t word;
+} UART_FLAG_t;
+
+/*
+ * Interrupt Vector Mask.
+ * Each bit in this register corresponds to a specific interrupt. When set,
+ * the associated interrupt will not be dispatched.
+ */
+
+__extension__
+typedef union
+{
+  struct
+  {
+#ifndef __BIG_ENDIAN__
+    /* Read data FIFO read and no data available */
+    uint_reg_t rdat_err       : 1;
+    /* Write FIFO was written but it was full */
+    uint_reg_t wdat_err       : 1;
+    /* Stop bit not found when current data was received */
+    uint_reg_t frame_err      : 1;
+    /* Parity error was detected when current data was received */
+    uint_reg_t parity_err     : 1;
+    /* Data was received but the receive FIFO was full */
+    uint_reg_t rfifo_overflow : 1;
+    /*
+     * An almost full event is reached when data is to be written to the
+     * receive FIFO, and the receive FIFO has more than or equal to
+     * BUFFER_THRESHOLD.RFIFO_AFULL bytes.
+     */
+    uint_reg_t rfifo_afull    : 1;
+    /* Reserved. */
+    uint_reg_t __reserved_0   : 1;
+    /* An entry in the transmit FIFO was popped */
+    uint_reg_t tfifo_re       : 1;
+    /* An entry has been pushed into the receive FIFO */
+    uint_reg_t rfifo_we       : 1;
+    /* An entry of the write FIFO has been popped */
+    uint_reg_t wfifo_re       : 1;
+    /* Rshim read receive FIFO in protocol mode */
+    uint_reg_t rfifo_err      : 1;
+    /*
+     * An almost empty event is reached when data is to be read from the
+     * transmit FIFO, and the transmit FIFO has less than or equal to
+     * BUFFER_THRESHOLD.TFIFO_AEMPTY bytes.
+     */
+    uint_reg_t tfifo_aempty   : 1;
+    /* Reserved. */
+    uint_reg_t __reserved_1   : 52;
+#else   /* __BIG_ENDIAN__ */
+    uint_reg_t __reserved_1   : 52;
+    uint_reg_t tfifo_aempty   : 1;
+    uint_reg_t rfifo_err      : 1;
+    uint_reg_t wfifo_re       : 1;
+    uint_reg_t rfifo_we       : 1;
+    uint_reg_t tfifo_re       : 1;
+    uint_reg_t __reserved_0   : 1;
+    uint_reg_t rfifo_afull    : 1;
+    uint_reg_t rfifo_overflow : 1;
+    uint_reg_t parity_err     : 1;
+    uint_reg_t frame_err      : 1;
+    uint_reg_t wdat_err       : 1;
+    uint_reg_t rdat_err       : 1;
+#endif
+  };
+
+  uint_reg_t word;
+} UART_INTERRUPT_MASK_t;
+
+/*
+ * Interrupt vector, write-one-to-clear.
+ * Each bit in this register corresponds to a specific interrupt. Hardware
+ * sets the bit when the associated condition has occurred. Writing a 1
+ * clears the status bit.
+ */
+
+__extension__
+typedef union
+{
+  struct
+  {
+#ifndef __BIG_ENDIAN__
+    /* Read data FIFO read and no data available */
+    uint_reg_t rdat_err       : 1;
+    /* Write FIFO was written but it was full */
+    uint_reg_t wdat_err       : 1;
+    /* Stop bit not found when current data was received */
+    uint_reg_t frame_err      : 1;
+    /* Parity error was detected when current data was received */
+    uint_reg_t parity_err     : 1;
+    /* Data was received but the receive FIFO was full */
+    uint_reg_t rfifo_overflow : 1;
+    /*
+     * Data was received and the receive FIFO is now almost full (more than
+     * BUFFER_THRESHOLD.RFIFO_AFULL bytes in it)
+     */
+    uint_reg_t rfifo_afull    : 1;
+    /* Reserved. */
+    uint_reg_t __reserved_0   : 1;
+    /* An entry in the transmit FIFO was popped */
+    uint_reg_t tfifo_re       : 1;
+    /* An entry has been pushed into the receive FIFO */
+    uint_reg_t rfifo_we       : 1;
+    /* An entry of the write FIFO has been popped */
+    uint_reg_t wfifo_re       : 1;
+    /* Rshim read receive FIFO in protocol mode */
+    uint_reg_t rfifo_err      : 1;
+    /*
+     * Data was read from the transmit FIFO and now it is almost empty (less
+     * than or equal to BUFFER_THRESHOLD.TFIFO_AEMPTY bytes in it).
+     */
+    uint_reg_t tfifo_aempty   : 1;
+    /* Reserved. */
+    uint_reg_t __reserved_1   : 52;
+#else   /* __BIG_ENDIAN__ */
+    uint_reg_t __reserved_1   : 52;
+    uint_reg_t tfifo_aempty   : 1;
+    uint_reg_t rfifo_err      : 1;
+    uint_reg_t wfifo_re       : 1;
+    uint_reg_t rfifo_we       : 1;
+    uint_reg_t tfifo_re       : 1;
+    uint_reg_t __reserved_0   : 1;
+    uint_reg_t rfifo_afull    : 1;
+    uint_reg_t rfifo_overflow : 1;
+    uint_reg_t parity_err     : 1;
+    uint_reg_t frame_err      : 1;
+    uint_reg_t wdat_err       : 1;
+    uint_reg_t rdat_err       : 1;
+#endif
+  };
+
+  uint_reg_t word;
+} UART_INTERRUPT_STATUS_t;
+
+/* Type. */
+
+__extension__
+typedef union
+{
+  struct
+  {
+#ifndef __BIG_ENDIAN__
+    /* Number of stop bits, rx and tx */
+    uint_reg_t sbits        : 1;
+    /* Reserved. */
+    uint_reg_t __reserved_0 : 1;
+    /* Data word size, rx and tx */
+    uint_reg_t dbits        : 1;
+    /* Reserved. */
+    uint_reg_t __reserved_1 : 1;
+    /* Parity selection, rx and tx */
+    uint_reg_t ptype        : 3;
+    /* Reserved. */
+    uint_reg_t __reserved_2 : 57;
+#else   /* __BIG_ENDIAN__ */
+    uint_reg_t __reserved_2 : 57;
+    uint_reg_t ptype        : 3;
+    uint_reg_t __reserved_1 : 1;
+    uint_reg_t dbits        : 1;
+    uint_reg_t __reserved_0 : 1;
+    uint_reg_t sbits        : 1;
+#endif
+  };
+
+  uint_reg_t word;
+} UART_TYPE_t;
+#endif /* !defined(__ASSEMBLER__) */
+
+#endif /* !defined(__ARCH_UART_H__) */
diff --git a/kernel/arch/tile/include/arch/uart_def.h b/kernel/arch/tile/include/arch/uart_def.h
new file mode 100644
index 000000000..42bcaf535
--- /dev/null
+++ b/kernel/arch/tile/include/arch/uart_def.h
@@ -0,0 +1,120 @@
+/*
+ * Copyright 2013 Tilera Corporation. All Rights Reserved.
+ *
+ *   This program is free software; you can redistribute it and/or
+ *   modify it under the terms of the GNU General Public License
+ *   as published by the Free Software Foundation, version 2.
+ *
+ *   This program is distributed in the hope that it will be useful, but
+ *   WITHOUT ANY WARRANTY; without even the implied warranty of
+ *   MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
+ *   NON INFRINGEMENT.  See the GNU General Public License for
+ *   more details.
+ */
+
+/* Machine-generated file; do not edit. */
+
+#ifndef __ARCH_UART_DEF_H__
+#define __ARCH_UART_DEF_H__
+#define UART_DIVISOR 0x0158
+#define UART_FIFO_COUNT 0x0110
+#define UART_FLAG 0x0108
+#define UART_INTERRUPT_MASK 0x0208
+#define UART_INTERRUPT_MASK__RDAT_ERR_SHIFT 0
+#define UART_INTERRUPT_MASK__RDAT_ERR_WIDTH 1
+#define UART_INTERRUPT_MASK__RDAT_ERR_RESET_VAL 1
+#define UART_INTERRUPT_MASK__RDAT_ERR_RMASK 0x1
+#define UART_INTERRUPT_MASK__RDAT_ERR_MASK  0x1
+#define UART_INTERRUPT_MASK__RDAT_ERR_FIELD 0,0
+#define UART_INTERRUPT_MASK__WDAT_ERR_SHIFT 1
+#define UART_INTERRUPT_MASK__WDAT_ERR_WIDTH 1
+#define UART_INTERRUPT_MASK__WDAT_ERR_RESET_VAL 1
+#define UART_INTERRUPT_MASK__WDAT_ERR_RMASK 0x1
+#define UART_INTERRUPT_MASK__WDAT_ERR_MASK  0x2
+#define UART_INTERRUPT_MASK__WDAT_ERR_FIELD 1,1
+#define UART_INTERRUPT_MASK__FRAME_ERR_SHIFT 2
+#define UART_INTERRUPT_MASK__FRAME_ERR_WIDTH 1
+#define UART_INTERRUPT_MASK__FRAME_ERR_RESET_VAL 1
+#define UART_INTERRUPT_MASK__FRAME_ERR_RMASK 0x1
+#define UART_INTERRUPT_MASK__FRAME_ERR_MASK  0x4
+#define UART_INTERRUPT_MASK__FRAME_ERR_FIELD 2,2
+#define UART_INTERRUPT_MASK__PARITY_ERR_SHIFT 3
+#define UART_INTERRUPT_MASK__PARITY_ERR_WIDTH 1
+#define UART_INTERRUPT_MASK__PARITY_ERR_RESET_VAL 1
+#define UART_INTERRUPT_MASK__PARITY_ERR_RMASK 0x1
+#define UART_INTERRUPT_MASK__PARITY_ERR_MASK  0x8
+#define UART_INTERRUPT_MASK__PARITY_ERR_FIELD 3,3
+#define UART_INTERRUPT_MASK__RFIFO_OVERFLOW_SHIFT 4
+#define UART_INTERRUPT_MASK__RFIFO_OVERFLOW_WIDTH 1
+#define UART_INTERRUPT_MASK__RFIFO_OVERFLOW_RESET_VAL 1
+#define UART_INTERRUPT_MASK__RFIFO_OVERFLOW_RMASK 0x1
+#define UART_INTERRUPT_MASK__RFIFO_OVERFLOW_MASK  0x10
+#define UART_INTERRUPT_MASK__RFIFO_OVERFLOW_FIELD 4,4
+#define UART_INTERRUPT_MASK__RFIFO_AFULL_SHIFT 5
+#define UART_INTERRUPT_MASK__RFIFO_AFULL_WIDTH 1
+#define UART_INTERRUPT_MASK__RFIFO_AFULL_RESET_VAL 1
+#define UART_INTERRUPT_MASK__RFIFO_AFULL_RMASK 0x1
+#define UART_INTERRUPT_MASK__RFIFO_AFULL_MASK  0x20
+#define UART_INTERRUPT_MASK__RFIFO_AFULL_FIELD 5,5
+#define UART_INTERRUPT_MASK__TFIFO_RE_SHIFT 7
+#define UART_INTERRUPT_MASK__TFIFO_RE_WIDTH 1
+#define UART_INTERRUPT_MASK__TFIFO_RE_RESET_VAL 1
+#define UART_INTERRUPT_MASK__TFIFO_RE_RMASK 0x1
+#define UART_INTERRUPT_MASK__TFIFO_RE_MASK  0x80
+#define UART_INTERRUPT_MASK__TFIFO_RE_FIELD 7,7
+#define UART_INTERRUPT_MASK__RFIFO_WE_SHIFT 8
+#define UART_INTERRUPT_MASK__RFIFO_WE_WIDTH 1
+#define UART_INTERRUPT_MASK__RFIFO_WE_RESET_VAL 1
+#define UART_INTERRUPT_MASK__RFIFO_WE_RMASK 0x1
+#define UART_INTERRUPT_MASK__RFIFO_WE_MASK  0x100
+#define UART_INTERRUPT_MASK__RFIFO_WE_FIELD 8,8
+#define UART_INTERRUPT_MASK__WFIFO_RE_SHIFT 9
+#define UART_INTERRUPT_MASK__WFIFO_RE_WIDTH 1
+#define UART_INTERRUPT_MASK__WFIFO_RE_RESET_VAL 1
+#define UART_INTERRUPT_MASK__WFIFO_RE_RMASK 0x1
+#define UART_INTERRUPT_MASK__WFIFO_RE_MASK  0x200
+#define UART_INTERRUPT_MASK__WFIFO_RE_FIELD 9,9
+#define UART_INTERRUPT_MASK__RFIFO_ERR_SHIFT 10
+#define UART_INTERRUPT_MASK__RFIFO_ERR_WIDTH 1
+#define UART_INTERRUPT_MASK__RFIFO_ERR_RESET_VAL 1
+#define UART_INTERRUPT_MASK__RFIFO_ERR_RMASK 0x1
+#define UART_INTERRUPT_MASK__RFIFO_ERR_MASK  0x400
+#define UART_INTERRUPT_MASK__RFIFO_ERR_FIELD 10,10
+#define UART_INTERRUPT_MASK__TFIFO_AEMPTY_SHIFT 11
+#define UART_INTERRUPT_MASK__TFIFO_AEMPTY_WIDTH 1
+#define UART_INTERRUPT_MASK__TFIFO_AEMPTY_RESET_VAL 1
+#define UART_INTERRUPT_MASK__TFIFO_AEMPTY_RMASK 0x1
+#define UART_INTERRUPT_MASK__TFIFO_AEMPTY_MASK  0x800
+#define UART_INTERRUPT_MASK__TFIFO_AEMPTY_FIELD 11,11
+#define UART_INTERRUPT_STATUS 0x0200
+#define UART_RECEIVE_DATA 0x0148
+#define UART_TRANSMIT_DATA 0x0140
+#define UART_TYPE 0x0160
+#define UART_TYPE__SBITS_SHIFT 0
+#define UART_TYPE__SBITS_WIDTH 1
+#define UART_TYPE__SBITS_RESET_VAL 1
+#define UART_TYPE__SBITS_RMASK 0x1
+#define UART_TYPE__SBITS_MASK  0x1
+#define UART_TYPE__SBITS_FIELD 0,0
+#define UART_TYPE__SBITS_VAL_ONE_SBITS 0x0
+#define UART_TYPE__SBITS_VAL_TWO_SBITS 0x1
+#define UART_TYPE__DBITS_SHIFT 2
+#define UART_TYPE__DBITS_WIDTH 1
+#define UART_TYPE__DBITS_RESET_VAL 0
+#define UART_TYPE__DBITS_RMASK 0x1
+#define UART_TYPE__DBITS_MASK  0x4
+#define UART_TYPE__DBITS_FIELD 2,2
+#define UART_TYPE__DBITS_VAL_EIGHT_DBITS 0x0
+#define UART_TYPE__DBITS_VAL_SEVEN_DBITS 0x1
+#define UART_TYPE__PTYPE_SHIFT 4
+#define UART_TYPE__PTYPE_WIDTH 3
+#define UART_TYPE__PTYPE_RESET_VAL 3
+#define UART_TYPE__PTYPE_RMASK 0x7
+#define UART_TYPE__PTYPE_MASK  0x70
+#define UART_TYPE__PTYPE_FIELD 4,6
+#define UART_TYPE__PTYPE_VAL_NONE 0x0
+#define UART_TYPE__PTYPE_VAL_MARK 0x1
+#define UART_TYPE__PTYPE_VAL_SPACE 0x2
+#define UART_TYPE__PTYPE_VAL_EVEN 0x3
+#define UART_TYPE__PTYPE_VAL_ODD 0x4
+#endif /* !defined(__ARCH_UART_DEF_H__) */
diff --git a/kernel/arch/tile/include/arch/usb_host.h b/kernel/arch/tile/include/arch/usb_host.h
new file mode 100644
index 000000000..d09f32683
--- /dev/null
+++ b/kernel/arch/tile/include/arch/usb_host.h
@@ -0,0 +1,26 @@
+/*
+ * Copyright 2012 Tilera Corporation. All Rights Reserved.
+ *
+ *   This program is free software; you can redistribute it and/or
+ *   modify it under the terms of the GNU General Public License
+ *   as published by the Free Software Foundation, version 2.
+ *
+ *   This program is distributed in the hope that it will be useful, but
+ *   WITHOUT ANY WARRANTY; without even the implied warranty of
+ *   MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
+ *   NON INFRINGEMENT.  See the GNU General Public License for
+ *   more details.
+ */
+
+/* Machine-generated file; do not edit. */
+
+#ifndef __ARCH_USB_HOST_H__
+#define __ARCH_USB_HOST_H__
+
+#include <arch/abi.h>
+#include <arch/usb_host_def.h>
+
+#ifndef __ASSEMBLER__
+#endif /* !defined(__ASSEMBLER__) */
+
+#endif /* !defined(__ARCH_USB_HOST_H__) */
diff --git a/kernel/arch/tile/include/arch/usb_host_def.h b/kernel/arch/tile/include/arch/usb_host_def.h
new file mode 100644
index 000000000..aeed7753e
--- /dev/null
+++ b/kernel/arch/tile/include/arch/usb_host_def.h
@@ -0,0 +1,19 @@
+/*
+ * Copyright 2012 Tilera Corporation. All Rights Reserved.
+ *
+ *   This program is free software; you can redistribute it and/or
+ *   modify it under the terms of the GNU General Public License
+ *   as published by the Free Software Foundation, version 2.
+ *
+ *   This program is distributed in the hope that it will be useful, but
+ *   WITHOUT ANY WARRANTY; without even the implied warranty of
+ *   MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
+ *   NON INFRINGEMENT.  See the GNU General Public License for
+ *   more details.
+ */
+
+/* Machine-generated file; do not edit. */
+
+#ifndef __ARCH_USB_HOST_DEF_H__
+#define __ARCH_USB_HOST_DEF_H__
+#endif /* !defined(__ARCH_USB_HOST_DEF_H__) */
diff --git a/kernel/arch/tile/include/asm/Kbuild b/kernel/arch/tile/include/asm/Kbuild
new file mode 100644
index 000000000..f5433e0e3
--- /dev/null
+++ b/kernel/arch/tile/include/asm/Kbuild
@@ -0,0 +1,42 @@
+
+header-y += ../arch/
+
+generic-y += bug.h
+generic-y += bugs.h
+generic-y += clkdev.h
+generic-y += cputime.h
+generic-y += div64.h
+generic-y += emergency-restart.h
+generic-y += errno.h
+generic-y += exec.h
+generic-y += fb.h
+generic-y += fcntl.h
+generic-y += hw_irq.h
+generic-y += ioctl.h
+generic-y += ioctls.h
+generic-y += ipcbuf.h
+generic-y += irq_regs.h
+generic-y += local.h
+generic-y += local64.h
+generic-y += mcs_spinlock.h
+generic-y += msgbuf.h
+generic-y += mutex.h
+generic-y += param.h
+generic-y += parport.h
+generic-y += poll.h
+generic-y += posix_types.h
+generic-y += preempt.h
+generic-y += resource.h
+generic-y += scatterlist.h
+generic-y += sembuf.h
+generic-y += serial.h
+generic-y += shmbuf.h
+generic-y += shmparam.h
+generic-y += socket.h
+generic-y += sockios.h
+generic-y += statfs.h
+generic-y += termbits.h
+generic-y += termios.h
+generic-y += trace_clock.h
+generic-y += types.h
+generic-y += xor.h
diff --git a/kernel/arch/tile/include/asm/asm-offsets.h b/kernel/arch/tile/include/asm/asm-offsets.h
new file mode 100644
index 000000000..d370ee36a
--- /dev/null
+++ b/kernel/arch/tile/include/asm/asm-offsets.h
@@ -0,0 +1 @@
+#include <generated/asm-offsets.h>
diff --git a/kernel/arch/tile/include/asm/atomic.h b/kernel/arch/tile/include/asm/atomic.h
new file mode 100644
index 000000000..709798460
--- /dev/null
+++ b/kernel/arch/tile/include/asm/atomic.h
@@ -0,0 +1,208 @@
+/*
+ * Copyright 2010 Tilera Corporation. All Rights Reserved.
+ *
+ *   This program is free software; you can redistribute it and/or
+ *   modify it under the terms of the GNU General Public License
+ *   as published by the Free Software Foundation, version 2.
+ *
+ *   This program is distributed in the hope that it will be useful, but
+ *   WITHOUT ANY WARRANTY; without even the implied warranty of
+ *   MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
+ *   NON INFRINGEMENT.  See the GNU General Public License for
+ *   more details.
+ *
+ * Atomic primitives.
+ */
+
+#ifndef _ASM_TILE_ATOMIC_H
+#define _ASM_TILE_ATOMIC_H
+
+#include <asm/cmpxchg.h>
+
+#ifndef __ASSEMBLY__
+
+#include <linux/compiler.h>
+#include <linux/types.h>
+
+#define ATOMIC_INIT(i)	{ (i) }
+
+/**
+ * atomic_read - read atomic variable
+ * @v: pointer of type atomic_t
+ *
+ * Atomically reads the value of @v.
+ */
+static inline int atomic_read(const atomic_t *v)
+{
+	return ACCESS_ONCE(v->counter);
+}
+
+/**
+ * atomic_sub_return - subtract integer and return
+ * @v: pointer of type atomic_t
+ * @i: integer value to subtract
+ *
+ * Atomically subtracts @i from @v and returns @v - @i
+ */
+#define atomic_sub_return(i, v)		atomic_add_return((int)(-(i)), (v))
+
+/**
+ * atomic_sub - subtract integer from atomic variable
+ * @i: integer value to subtract
+ * @v: pointer of type atomic_t
+ *
+ * Atomically subtracts @i from @v.
+ */
+#define atomic_sub(i, v)		atomic_add((int)(-(i)), (v))
+
+/**
+ * atomic_sub_and_test - subtract value from variable and test result
+ * @i: integer value to subtract
+ * @v: pointer of type atomic_t
+ *
+ * Atomically subtracts @i from @v and returns true if the result is
+ * zero, or false for all other cases.
+ */
+#define atomic_sub_and_test(i, v)	(atomic_sub_return((i), (v)) == 0)
+
+/**
+ * atomic_inc_return - increment memory and return
+ * @v: pointer of type atomic_t
+ *
+ * Atomically increments @v by 1 and returns the new value.
+ */
+#define atomic_inc_return(v)		atomic_add_return(1, (v))
+
+/**
+ * atomic_dec_return - decrement memory and return
+ * @v: pointer of type atomic_t
+ *
+ * Atomically decrements @v by 1 and returns the new value.
+ */
+#define atomic_dec_return(v)		atomic_sub_return(1, (v))
+
+/**
+ * atomic_inc - increment atomic variable
+ * @v: pointer of type atomic_t
+ *
+ * Atomically increments @v by 1.
+ */
+#define atomic_inc(v)			atomic_add(1, (v))
+
+/**
+ * atomic_dec - decrement atomic variable
+ * @v: pointer of type atomic_t
+ *
+ * Atomically decrements @v by 1.
+ */
+#define atomic_dec(v)			atomic_sub(1, (v))
+
+/**
+ * atomic_dec_and_test - decrement and test
+ * @v: pointer of type atomic_t
+ *
+ * Atomically decrements @v by 1 and returns true if the result is 0.
+ */
+#define atomic_dec_and_test(v)		(atomic_dec_return(v) == 0)
+
+/**
+ * atomic_inc_and_test - increment and test
+ * @v: pointer of type atomic_t
+ *
+ * Atomically increments @v by 1 and returns true if the result is 0.
+ */
+#define atomic_inc_and_test(v)		(atomic_inc_return(v) == 0)
+
+/**
+ * atomic_xchg - atomically exchange contents of memory with a new value
+ * @v: pointer of type atomic_t
+ * @i: integer value to store in memory
+ *
+ * Atomically sets @v to @i and returns old @v
+ */
+static inline int atomic_xchg(atomic_t *v, int n)
+{
+	return xchg(&v->counter, n);
+}
+
+/**
+ * atomic_cmpxchg - atomically exchange contents of memory if it matches
+ * @v: pointer of type atomic_t
+ * @o: old value that memory should have
+ * @n: new value to write to memory if it matches
+ *
+ * Atomically checks if @v holds @o and replaces it with @n if so.
+ * Returns the old value at @v.
+ */
+static inline int atomic_cmpxchg(atomic_t *v, int o, int n)
+{
+	return cmpxchg(&v->counter, o, n);
+}
+
+/**
+ * atomic_add_negative - add and test if negative
+ * @v: pointer of type atomic_t
+ * @i: integer value to add
+ *
+ * Atomically adds @i to @v and returns true if the result is
+ * negative, or false when result is greater than or equal to zero.
+ */
+#define atomic_add_negative(i, v)	(atomic_add_return((i), (v)) < 0)
+
+#endif /* __ASSEMBLY__ */
+
+#ifndef __tilegx__
+#include <asm/atomic_32.h>
+#else
+#include <asm/atomic_64.h>
+#endif
+
+#ifndef __ASSEMBLY__
+
+/**
+ * atomic64_xchg - atomically exchange contents of memory with a new value
+ * @v: pointer of type atomic64_t
+ * @i: integer value to store in memory
+ *
+ * Atomically sets @v to @i and returns old @v
+ */
+static inline long long atomic64_xchg(atomic64_t *v, long long n)
+{
+	return xchg64(&v->counter, n);
+}
+
+/**
+ * atomic64_cmpxchg - atomically exchange contents of memory if it matches
+ * @v: pointer of type atomic64_t
+ * @o: old value that memory should have
+ * @n: new value to write to memory if it matches
+ *
+ * Atomically checks if @v holds @o and replaces it with @n if so.
+ * Returns the old value at @v.
+ */
+static inline long long atomic64_cmpxchg(atomic64_t *v, long long o,
+					long long n)
+{
+	return cmpxchg64(&v->counter, o, n);
+}
+
+static inline long long atomic64_dec_if_positive(atomic64_t *v)
+{
+	long long c, old, dec;
+
+	c = atomic64_read(v);
+	for (;;) {
+		dec = c - 1;
+		if (unlikely(dec < 0))
+			break;
+		old = atomic64_cmpxchg((v), c, dec);
+		if (likely(old == c))
+			break;
+		c = old;
+	}
+	return dec;
+}
+
+#endif /* __ASSEMBLY__ */
+
+#endif /* _ASM_TILE_ATOMIC_H */
diff --git a/kernel/arch/tile/include/asm/atomic_32.h b/kernel/arch/tile/include/asm/atomic_32.h
new file mode 100644
index 000000000..1b109fad9
--- /dev/null
+++ b/kernel/arch/tile/include/asm/atomic_32.h
@@ -0,0 +1,243 @@
+/*
+ * Copyright 2010 Tilera Corporation. All Rights Reserved.
+ *
+ *   This program is free software; you can redistribute it and/or
+ *   modify it under the terms of the GNU General Public License
+ *   as published by the Free Software Foundation, version 2.
+ *
+ *   This program is distributed in the hope that it will be useful, but
+ *   WITHOUT ANY WARRANTY; without even the implied warranty of
+ *   MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
+ *   NON INFRINGEMENT.  See the GNU General Public License for
+ *   more details.
+ *
+ * Do not include directly; use <linux/atomic.h>.
+ */
+
+#ifndef _ASM_TILE_ATOMIC_32_H
+#define _ASM_TILE_ATOMIC_32_H
+
+#include <asm/barrier.h>
+#include <arch/chip.h>
+
+#ifndef __ASSEMBLY__
+
+/**
+ * atomic_add - add integer to atomic variable
+ * @i: integer value to add
+ * @v: pointer of type atomic_t
+ *
+ * Atomically adds @i to @v.
+ */
+static inline void atomic_add(int i, atomic_t *v)
+{
+	_atomic_xchg_add(&v->counter, i);
+}
+
+/**
+ * atomic_add_return - add integer and return
+ * @v: pointer of type atomic_t
+ * @i: integer value to add
+ *
+ * Atomically adds @i to @v and returns @i + @v
+ */
+static inline int atomic_add_return(int i, atomic_t *v)
+{
+	smp_mb();  /* barrier for proper semantics */
+	return _atomic_xchg_add(&v->counter, i) + i;
+}
+
+/**
+ * __atomic_add_unless - add unless the number is already a given value
+ * @v: pointer of type atomic_t
+ * @a: the amount to add to v...
+ * @u: ...unless v is equal to u.
+ *
+ * Atomically adds @a to @v, so long as @v was not already @u.
+ * Returns the old value of @v.
+ */
+static inline int __atomic_add_unless(atomic_t *v, int a, int u)
+{
+	smp_mb();  /* barrier for proper semantics */
+	return _atomic_xchg_add_unless(&v->counter, a, u);
+}
+
+/**
+ * atomic_set - set atomic variable
+ * @v: pointer of type atomic_t
+ * @i: required value
+ *
+ * Atomically sets the value of @v to @i.
+ *
+ * atomic_set() can't be just a raw store, since it would be lost if it
+ * fell between the load and store of one of the other atomic ops.
+ */
+static inline void atomic_set(atomic_t *v, int n)
+{
+	_atomic_xchg(&v->counter, n);
+}
+
+/* A 64bit atomic type */
+
+typedef struct {
+	long long counter;
+} atomic64_t;
+
+#define ATOMIC64_INIT(val) { (val) }
+
+/**
+ * atomic64_read - read atomic variable
+ * @v: pointer of type atomic64_t
+ *
+ * Atomically reads the value of @v.
+ */
+static inline long long atomic64_read(const atomic64_t *v)
+{
+	/*
+	 * Requires an atomic op to read both 32-bit parts consistently.
+	 * Casting away const is safe since the atomic support routines
+	 * do not write to memory if the value has not been modified.
+	 */
+	return _atomic64_xchg_add((long long *)&v->counter, 0);
+}
+
+/**
+ * atomic64_add - add integer to atomic variable
+ * @i: integer value to add
+ * @v: pointer of type atomic64_t
+ *
+ * Atomically adds @i to @v.
+ */
+static inline void atomic64_add(long long i, atomic64_t *v)
+{
+	_atomic64_xchg_add(&v->counter, i);
+}
+
+/**
+ * atomic64_add_return - add integer and return
+ * @v: pointer of type atomic64_t
+ * @i: integer value to add
+ *
+ * Atomically adds @i to @v and returns @i + @v
+ */
+static inline long long atomic64_add_return(long long i, atomic64_t *v)
+{
+	smp_mb();  /* barrier for proper semantics */
+	return _atomic64_xchg_add(&v->counter, i) + i;
+}
+
+/**
+ * atomic64_add_unless - add unless the number is already a given value
+ * @v: pointer of type atomic64_t
+ * @a: the amount to add to v...
+ * @u: ...unless v is equal to u.
+ *
+ * Atomically adds @a to @v, so long as @v was not already @u.
+ * Returns non-zero if @v was not @u, and zero otherwise.
+ */
+static inline long long atomic64_add_unless(atomic64_t *v, long long a,
+					long long u)
+{
+	smp_mb();  /* barrier for proper semantics */
+	return _atomic64_xchg_add_unless(&v->counter, a, u) != u;
+}
+
+/**
+ * atomic64_set - set atomic variable
+ * @v: pointer of type atomic64_t
+ * @i: required value
+ *
+ * Atomically sets the value of @v to @i.
+ *
+ * atomic64_set() can't be just a raw store, since it would be lost if it
+ * fell between the load and store of one of the other atomic ops.
+ */
+static inline void atomic64_set(atomic64_t *v, long long n)
+{
+	_atomic64_xchg(&v->counter, n);
+}
+
+#define atomic64_add_negative(a, v)	(atomic64_add_return((a), (v)) < 0)
+#define atomic64_inc(v)			atomic64_add(1LL, (v))
+#define atomic64_inc_return(v)		atomic64_add_return(1LL, (v))
+#define atomic64_inc_and_test(v)	(atomic64_inc_return(v) == 0)
+#define atomic64_sub_return(i, v)	atomic64_add_return(-(i), (v))
+#define atomic64_sub_and_test(a, v)	(atomic64_sub_return((a), (v)) == 0)
+#define atomic64_sub(i, v)		atomic64_add(-(i), (v))
+#define atomic64_dec(v)			atomic64_sub(1LL, (v))
+#define atomic64_dec_return(v)		atomic64_sub_return(1LL, (v))
+#define atomic64_dec_and_test(v)	(atomic64_dec_return((v)) == 0)
+#define atomic64_inc_not_zero(v)	atomic64_add_unless((v), 1LL, 0LL)
+
+
+#endif /* !__ASSEMBLY__ */
+
+/*
+ * Internal definitions only beyond this point.
+ */
+
+/*
+ * Number of atomic locks in atomic_locks[]. Must be a power of two.
+ * There is no reason for more than PAGE_SIZE / 8 entries, since that
+ * is the maximum number of pointer bits we can use to index this.
+ * And we cannot have more than PAGE_SIZE / 4, since this has to
+ * fit on a single page and each entry takes 4 bytes.
+ */
+#define ATOMIC_HASH_SHIFT (PAGE_SHIFT - 3)
+#define ATOMIC_HASH_SIZE (1 << ATOMIC_HASH_SHIFT)
+
+#ifndef __ASSEMBLY__
+extern int atomic_locks[];
+#endif
+
+/*
+ * All the code that may fault while holding an atomic lock must
+ * place the pointer to the lock in ATOMIC_LOCK_REG so the fault code
+ * can correctly release and reacquire the lock.  Note that we
+ * mention the register number in a comment in "lib/atomic_asm.S" to help
+ * assembly coders from using this register by mistake, so if it
+ * is changed here, change that comment as well.
+ */
+#define ATOMIC_LOCK_REG 20
+#define ATOMIC_LOCK_REG_NAME r20
+
+#ifndef __ASSEMBLY__
+/* Called from setup to initialize a hash table to point to per_cpu locks. */
+void __init_atomic_per_cpu(void);
+
+#ifdef CONFIG_SMP
+/* Support releasing the atomic lock in do_page_fault_ics(). */
+void __atomic_fault_unlock(int *lock_ptr);
+#endif
+
+/* Return a pointer to the lock for the given address. */
+int *__atomic_hashed_lock(volatile void *v);
+
+/* Private helper routines in lib/atomic_asm_32.S */
+struct __get_user {
+	unsigned long val;
+	int err;
+};
+extern struct __get_user __atomic_cmpxchg(volatile int *p,
+					  int *lock, int o, int n);
+extern struct __get_user __atomic_xchg(volatile int *p, int *lock, int n);
+extern struct __get_user __atomic_xchg_add(volatile int *p, int *lock, int n);
+extern struct __get_user __atomic_xchg_add_unless(volatile int *p,
+						  int *lock, int o, int n);
+extern struct __get_user __atomic_or(volatile int *p, int *lock, int n);
+extern struct __get_user __atomic_andn(volatile int *p, int *lock, int n);
+extern struct __get_user __atomic_xor(volatile int *p, int *lock, int n);
+extern long long __atomic64_cmpxchg(volatile long long *p, int *lock,
+					long long o, long long n);
+extern long long __atomic64_xchg(volatile long long *p, int *lock, long long n);
+extern long long __atomic64_xchg_add(volatile long long *p, int *lock,
+					long long n);
+extern long long __atomic64_xchg_add_unless(volatile long long *p,
+					int *lock, long long o, long long n);
+
+/* Return failure from the atomic wrappers. */
+struct __get_user __atomic_bad_address(int __user *addr);
+
+#endif /* !__ASSEMBLY__ */
+
+#endif /* _ASM_TILE_ATOMIC_32_H */
diff --git a/kernel/arch/tile/include/asm/atomic_64.h b/kernel/arch/tile/include/asm/atomic_64.h
new file mode 100644
index 000000000..7b11c5fad
--- /dev/null
+++ b/kernel/arch/tile/include/asm/atomic_64.h
@@ -0,0 +1,113 @@
+/*
+ * Copyright 2011 Tilera Corporation. All Rights Reserved.
+ *
+ *   This program is free software; you can redistribute it and/or
+ *   modify it under the terms of the GNU General Public License
+ *   as published by the Free Software Foundation, version 2.
+ *
+ *   This program is distributed in the hope that it will be useful, but
+ *   WITHOUT ANY WARRANTY; without even the implied warranty of
+ *   MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
+ *   NON INFRINGEMENT.  See the GNU General Public License for
+ *   more details.
+ *
+ * Do not include directly; use <linux/atomic.h>.
+ */
+
+#ifndef _ASM_TILE_ATOMIC_64_H
+#define _ASM_TILE_ATOMIC_64_H
+
+#ifndef __ASSEMBLY__
+
+#include <asm/barrier.h>
+#include <arch/spr_def.h>
+
+/* First, the 32-bit atomic ops that are "real" on our 64-bit platform. */
+
+#define atomic_set(v, i) ((v)->counter = (i))
+
+/*
+ * The smp_mb() operations throughout are to support the fact that
+ * Linux requires memory barriers before and after the operation,
+ * on any routine which updates memory and returns a value.
+ */
+
+static inline void atomic_add(int i, atomic_t *v)
+{
+	__insn_fetchadd4((void *)&v->counter, i);
+}
+
+static inline int atomic_add_return(int i, atomic_t *v)
+{
+	int val;
+	smp_mb();  /* barrier for proper semantics */
+	val = __insn_fetchadd4((void *)&v->counter, i) + i;
+	barrier();  /* the "+ i" above will wait on memory */
+	return val;
+}
+
+static inline int __atomic_add_unless(atomic_t *v, int a, int u)
+{
+	int guess, oldval = v->counter;
+	do {
+		if (oldval == u)
+			break;
+		guess = oldval;
+		oldval = cmpxchg(&v->counter, guess, guess + a);
+	} while (guess != oldval);
+	return oldval;
+}
+
+/* Now the true 64-bit operations. */
+
+#define ATOMIC64_INIT(i)	{ (i) }
+
+#define atomic64_read(v)		((v)->counter)
+#define atomic64_set(v, i) ((v)->counter = (i))
+
+static inline void atomic64_add(long i, atomic64_t *v)
+{
+	__insn_fetchadd((void *)&v->counter, i);
+}
+
+static inline long atomic64_add_return(long i, atomic64_t *v)
+{
+	int val;
+	smp_mb();  /* barrier for proper semantics */
+	val = __insn_fetchadd((void *)&v->counter, i) + i;
+	barrier();  /* the "+ i" above will wait on memory */
+	return val;
+}
+
+static inline long atomic64_add_unless(atomic64_t *v, long a, long u)
+{
+	long guess, oldval = v->counter;
+	do {
+		if (oldval == u)
+			break;
+		guess = oldval;
+		oldval = cmpxchg(&v->counter, guess, guess + a);
+	} while (guess != oldval);
+	return oldval != u;
+}
+
+#define atomic64_sub_return(i, v)	atomic64_add_return(-(i), (v))
+#define atomic64_sub(i, v)		atomic64_add(-(i), (v))
+#define atomic64_inc_return(v)		atomic64_add_return(1, (v))
+#define atomic64_dec_return(v)		atomic64_sub_return(1, (v))
+#define atomic64_inc(v)			atomic64_add(1, (v))
+#define atomic64_dec(v)			atomic64_sub(1, (v))
+
+#define atomic64_inc_and_test(v)	(atomic64_inc_return(v) == 0)
+#define atomic64_dec_and_test(v)	(atomic64_dec_return(v) == 0)
+#define atomic64_sub_and_test(i, v)	(atomic64_sub_return((i), (v)) == 0)
+#define atomic64_add_negative(i, v)	(atomic64_add_return((i), (v)) < 0)
+
+#define atomic64_inc_not_zero(v)	atomic64_add_unless((v), 1, 0)
+
+/* Define this to indicate that cmpxchg is an efficient operation. */
+#define __HAVE_ARCH_CMPXCHG
+
+#endif /* !__ASSEMBLY__ */
+
+#endif /* _ASM_TILE_ATOMIC_64_H */
diff --git a/kernel/arch/tile/include/asm/backtrace.h b/kernel/arch/tile/include/asm/backtrace.h
new file mode 100644
index 000000000..bd5399a69
--- /dev/null
+++ b/kernel/arch/tile/include/asm/backtrace.h
@@ -0,0 +1,162 @@
+/*
+ * Copyright 2010 Tilera Corporation. All Rights Reserved.
+ *
+ *   This program is free software; you can redistribute it and/or
+ *   modify it under the terms of the GNU General Public License
+ *   as published by the Free Software Foundation, version 2.
+ *
+ *   This program is distributed in the hope that it will be useful, but
+ *   WITHOUT ANY WARRANTY; without even the implied warranty of
+ *   MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
+ *   NON INFRINGEMENT.  See the GNU General Public License for
+ *   more details.
+ */
+
+#ifndef _ASM_TILE_BACKTRACE_H
+#define _ASM_TILE_BACKTRACE_H
+
+#include <linux/types.h>
+
+/* Reads 'size' bytes from 'address' and writes the data to 'result'.
+ * Returns true if successful, else false (e.g. memory not readable).
+ */
+typedef bool (*BacktraceMemoryReader)(void *result,
+				      unsigned long address,
+				      unsigned int size,
+				      void *extra);
+
+typedef struct {
+	/* Current PC. */
+	unsigned long pc;
+
+	/* Current stack pointer value. */
+	unsigned long sp;
+
+	/* Current frame pointer value (i.e. caller's stack pointer) */
+	unsigned long fp;
+
+	/* Internal use only: caller's PC for first frame. */
+	unsigned long initial_frame_caller_pc;
+
+	/* Internal use only: callback to read memory. */
+	BacktraceMemoryReader read_memory_func;
+
+	/* Internal use only: arbitrary argument to read_memory_func. */
+	void *read_memory_func_extra;
+
+} BacktraceIterator;
+
+
+typedef enum {
+
+	/* We have no idea what the caller's pc is. */
+	PC_LOC_UNKNOWN,
+
+	/* The caller's pc is currently in lr. */
+	PC_LOC_IN_LR,
+
+	/* The caller's pc can be found by dereferencing the caller's sp. */
+	PC_LOC_ON_STACK
+
+} CallerPCLocation;
+
+
+typedef enum {
+
+	/* We have no idea what the caller's sp is. */
+	SP_LOC_UNKNOWN,
+
+	/* The caller's sp is currently in r52. */
+	SP_LOC_IN_R52,
+
+	/* The caller's sp can be found by adding a certain constant
+	 * to the current value of sp.
+	 */
+	SP_LOC_OFFSET
+
+} CallerSPLocation;
+
+
+/* Bit values ORed into CALLER_* values for info ops. */
+enum {
+	/* Setting the low bit on any of these values means the info op
+	 * applies only to one bundle ago.
+	 */
+	ONE_BUNDLE_AGO_FLAG = 1,
+
+	/* Setting this bit on a CALLER_SP_* value means the PC is in LR.
+	 * If not set, PC is on the stack.
+	 */
+	PC_IN_LR_FLAG = 2,
+
+	/* This many of the low bits of a CALLER_SP_* value are for the
+	 * flag bits above.
+	 */
+	NUM_INFO_OP_FLAGS = 2,
+
+	/* We cannot have one in the memory pipe so this is the maximum. */
+	MAX_INFO_OPS_PER_BUNDLE = 2
+};
+
+
+/* Internal constants used to define 'info' operands. */
+enum {
+	/* 0 and 1 are reserved, as are all negative numbers. */
+
+	CALLER_UNKNOWN_BASE = 2,
+
+	CALLER_SP_IN_R52_BASE = 4,
+
+	CALLER_SP_OFFSET_BASE = 8,
+};
+
+
+/* Current backtracer state describing where it thinks the caller is. */
+typedef struct {
+	/*
+	 * Public fields
+	 */
+
+	/* How do we find the caller's PC? */
+	CallerPCLocation pc_location : 8;
+
+	/* How do we find the caller's SP? */
+	CallerSPLocation sp_location : 8;
+
+	/* If sp_location == SP_LOC_OFFSET, then caller_sp == sp +
+	 * loc->sp_offset. Else this field is undefined.
+	 */
+	uint16_t sp_offset;
+
+	/* In the most recently visited bundle a terminating bundle? */
+	bool at_terminating_bundle;
+
+	/*
+	 * Private fields
+	 */
+
+	/* Will the forward scanner see someone clobbering sp
+	 * (i.e. changing it with something other than addi sp, sp, N?)
+	 */
+	bool sp_clobber_follows;
+
+	/* Operand to next "visible" info op (no more than one bundle past
+	 * the next terminating bundle), or -32768 if none.
+	 */
+	int16_t next_info_operand;
+
+	/* Is the info of in next_info_op in the very next bundle? */
+	bool is_next_info_operand_adjacent;
+
+} CallerLocation;
+
+extern void backtrace_init(BacktraceIterator *state,
+                          BacktraceMemoryReader read_memory_func,
+                          void *read_memory_func_extra,
+                          unsigned long pc, unsigned long lr,
+                          unsigned long sp, unsigned long r52);
+
+
+extern bool backtrace_next(BacktraceIterator *state);
+
+#endif /* _ASM_TILE_BACKTRACE_H */
diff --git a/kernel/arch/tile/include/asm/barrier.h b/kernel/arch/tile/include/asm/barrier.h
new file mode 100644
index 000000000..96a42ae79
--- /dev/null
+++ b/kernel/arch/tile/include/asm/barrier.h
@@ -0,0 +1,92 @@
+/*
+ * Copyright 2010 Tilera Corporation. All Rights Reserved.
+ *
+ *   This program is free software; you can redistribute it and/or
+ *   modify it under the terms of the GNU General Public License
+ *   as published by the Free Software Foundation, version 2.
+ *
+ *   This program is distributed in the hope that it will be useful, but
+ *   WITHOUT ANY WARRANTY; without even the implied warranty of
+ *   MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
+ *   NON INFRINGEMENT.  See the GNU General Public License for
+ *   more details.
+ */
+
+#ifndef _ASM_TILE_BARRIER_H
+#define _ASM_TILE_BARRIER_H
+
+#ifndef __ASSEMBLY__
+
+#include <linux/types.h>
+#include <arch/chip.h>
+#include <arch/spr_def.h>
+#include <asm/timex.h>
+
+#define __sync()	__insn_mf()
+
+#include <hv/syscall_public.h>
+/*
+ * Issue an uncacheable load to each memory controller, then
+ * wait until those loads have completed.
+ */
+static inline void __mb_incoherent(void)
+{
+	long clobber_r10;
+	asm volatile("swint2"
+		     : "=R10" (clobber_r10)
+		     : "R10" (HV_SYS_fence_incoherent)
+		     : "r0", "r1", "r2", "r3", "r4",
+		       "r5", "r6", "r7", "r8", "r9",
+		       "r11", "r12", "r13", "r14",
+		       "r15", "r16", "r17", "r18", "r19",
+		       "r20", "r21", "r22", "r23", "r24",
+		       "r25", "r26", "r27", "r28", "r29");
+}
+
+/* Fence to guarantee visibility of stores to incoherent memory. */
+static inline void
+mb_incoherent(void)
+{
+	__insn_mf();
+
+	{
+#if CHIP_HAS_TILE_WRITE_PENDING()
+		const unsigned long WRITE_TIMEOUT_CYCLES = 400;
+		unsigned long start = get_cycles_low();
+		do {
+			if (__insn_mfspr(SPR_TILE_WRITE_PENDING) == 0)
+				return;
+		} while ((get_cycles_low() - start) < WRITE_TIMEOUT_CYCLES);
+#endif /* CHIP_HAS_TILE_WRITE_PENDING() */
+		(void) __mb_incoherent();
+	}
+}
+
+#define fast_wmb()	__sync()
+#define fast_rmb()	__sync()
+#define fast_mb()	__sync()
+#define fast_iob()	mb_incoherent()
+
+#define wmb()		fast_wmb()
+#define rmb()		fast_rmb()
+#define mb()		fast_mb()
+#define iob()		fast_iob()
+
+#ifndef __tilegx__ /* 32 bit */
+/*
+ * We need to barrier before modifying the word, since the _atomic_xxx()
+ * routines just tns the lock and then read/modify/write of the word.
+ * But after the word is updated, the routine issues an "mf" before returning,
+ * and since it's a function call, we don't even need a compiler barrier.
+ */
+#define smp_mb__before_atomic()	smp_mb()
+#define smp_mb__after_atomic()	do { } while (0)
+#else /* 64 bit */
+#define smp_mb__before_atomic()	smp_mb()
+#define smp_mb__after_atomic()	smp_mb()
+#endif
+
+#include <asm-generic/barrier.h>
+
+#endif /* !__ASSEMBLY__ */
+#endif /* _ASM_TILE_BARRIER_H */
diff --git a/kernel/arch/tile/include/asm/bitops.h b/kernel/arch/tile/include/asm/bitops.h
new file mode 100644
index 000000000..20caa346a
--- /dev/null
+++ b/kernel/arch/tile/include/asm/bitops.h
@@ -0,0 +1,94 @@
+/*
+ * Copyright 1992, Linus Torvalds.
+ * Copyright 2010 Tilera Corporation. All Rights Reserved.
+ *
+ *   This program is free software; you can redistribute it and/or
+ *   modify it under the terms of the GNU General Public License
+ *   as published by the Free Software Foundation, version 2.
+ *
+ *   This program is distributed in the hope that it will be useful, but
+ *   WITHOUT ANY WARRANTY; without even the implied warranty of
+ *   MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
+ *   NON INFRINGEMENT.  See the GNU General Public License for
+ *   more details.
+ */
+
+#ifndef _ASM_TILE_BITOPS_H
+#define _ASM_TILE_BITOPS_H
+
+#include <linux/types.h>
+#include <asm/barrier.h>
+
+#ifndef _LINUX_BITOPS_H
+#error only <linux/bitops.h> can be included directly
+#endif
+
+#ifdef __tilegx__
+#include <asm/bitops_64.h>
+#else
+#include <asm/bitops_32.h>
+#endif
+
+/**
+ * ffz - find first zero bit in word
+ * @word: The word to search
+ *
+ * Undefined if no zero exists, so code should check against ~0UL first.
+ */
+static inline unsigned long ffz(unsigned long word)
+{
+	return __builtin_ctzl(~word);
+}
+
+static inline int fls64(__u64 w)
+{
+	return (sizeof(__u64) * 8) - __builtin_clzll(w);
+}
+
+/**
+ * fls - find last set bit in word
+ * @x: the word to search
+ *
+ * This is defined in a similar way as the libc and compiler builtin
+ * ffs, but returns the position of the most significant set bit.
+ *
+ * fls(value) returns 0 if value is 0 or the position of the last
+ * set bit if value is nonzero. The last (most significant) bit is
+ * at position 32.
+ */
+static inline int fls(int x)
+{
+	return fls64((unsigned int) x);
+}
+
+static inline unsigned int __arch_hweight32(unsigned int w)
+{
+	return __builtin_popcount(w);
+}
+
+static inline unsigned int __arch_hweight16(unsigned int w)
+{
+	return __builtin_popcount(w & 0xffff);
+}
+
+static inline unsigned int __arch_hweight8(unsigned int w)
+{
+	return __builtin_popcount(w & 0xff);
+}
+
+static inline unsigned long __arch_hweight64(__u64 w)
+{
+	return __builtin_popcountll(w);
+}
+
+#include <asm-generic/bitops/builtin-__ffs.h>
+#include <asm-generic/bitops/builtin-__fls.h>
+#include <asm-generic/bitops/builtin-ffs.h>
+#include <asm-generic/bitops/const_hweight.h>
+#include <asm-generic/bitops/lock.h>
+#include <asm-generic/bitops/find.h>
+#include <asm-generic/bitops/sched.h>
+#include <asm-generic/bitops/non-atomic.h>
+#include <asm-generic/bitops/le.h>
+
+#endif /* _ASM_TILE_BITOPS_H */
diff --git a/kernel/arch/tile/include/asm/bitops_32.h b/kernel/arch/tile/include/asm/bitops_32.h
new file mode 100644
index 000000000..bbf7b666f
--- /dev/null
+++ b/kernel/arch/tile/include/asm/bitops_32.h
@@ -0,0 +1,126 @@
+/*
+ * Copyright 2010 Tilera Corporation. All Rights Reserved.
+ *
+ *   This program is free software; you can redistribute it and/or
+ *   modify it under the terms of the GNU General Public License
+ *   as published by the Free Software Foundation, version 2.
+ *
+ *   This program is distributed in the hope that it will be useful, but
+ *   WITHOUT ANY WARRANTY; without even the implied warranty of
+ *   MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
+ *   NON INFRINGEMENT.  See the GNU General Public License for
+ *   more details.
+ */
+
+#ifndef _ASM_TILE_BITOPS_32_H
+#define _ASM_TILE_BITOPS_32_H
+
+#include <linux/compiler.h>
+#include <asm/barrier.h>
+
+/* Tile-specific routines to support <asm/bitops.h>. */
+unsigned long _atomic_or(volatile unsigned long *p, unsigned long mask);
+unsigned long _atomic_andn(volatile unsigned long *p, unsigned long mask);
+unsigned long _atomic_xor(volatile unsigned long *p, unsigned long mask);
+
+/**
+ * set_bit - Atomically set a bit in memory
+ * @nr: the bit to set
+ * @addr: the address to start counting from
+ *
+ * This function is atomic and may not be reordered.
+ * See __set_bit() if you do not require the atomic guarantees.
+ * Note that @nr may be almost arbitrarily large; this function is not
+ * restricted to acting on a single-word quantity.
+ */
+static inline void set_bit(unsigned nr, volatile unsigned long *addr)
+{
+	_atomic_or(addr + BIT_WORD(nr), BIT_MASK(nr));
+}
+
+/**
+ * clear_bit - Clears a bit in memory
+ * @nr: Bit to clear
+ * @addr: Address to start counting from
+ *
+ * clear_bit() is atomic and may not be reordered.
+ * See __clear_bit() if you do not require the atomic guarantees.
+ * Note that @nr may be almost arbitrarily large; this function is not
+ * restricted to acting on a single-word quantity.
+ *
+ * clear_bit() may not contain a memory barrier, so if it is used for
+ * locking purposes, you should call smp_mb__before_atomic() and/or
+ * smp_mb__after_atomic() to ensure changes are visible on other cpus.
+ */
+static inline void clear_bit(unsigned nr, volatile unsigned long *addr)
+{
+	_atomic_andn(addr + BIT_WORD(nr), BIT_MASK(nr));
+}
+
+/**
+ * change_bit - Toggle a bit in memory
+ * @nr: Bit to change
+ * @addr: Address to start counting from
+ *
+ * change_bit() is atomic and may not be reordered.
+ * See __change_bit() if you do not require the atomic guarantees.
+ * Note that @nr may be almost arbitrarily large; this function is not
+ * restricted to acting on a single-word quantity.
+ */
+static inline void change_bit(unsigned nr, volatile unsigned long *addr)
+{
+	_atomic_xor(addr + BIT_WORD(nr), BIT_MASK(nr));
+}
+
+/**
+ * test_and_set_bit - Set a bit and return its old value
+ * @nr: Bit to set
+ * @addr: Address to count from
+ *
+ * This operation is atomic and cannot be reordered.
+ * It also implies a memory barrier.
+ */
+static inline int test_and_set_bit(unsigned nr, volatile unsigned long *addr)
+{
+	unsigned long mask = BIT_MASK(nr);
+	addr += BIT_WORD(nr);
+	smp_mb();  /* barrier for proper semantics */
+	return (_atomic_or(addr, mask) & mask) != 0;
+}
+
+/**
+ * test_and_clear_bit - Clear a bit and return its old value
+ * @nr: Bit to clear
+ * @addr: Address to count from
+ *
+ * This operation is atomic and cannot be reordered.
+ * It also implies a memory barrier.
+ */
+static inline int test_and_clear_bit(unsigned nr, volatile unsigned long *addr)
+{
+	unsigned long mask = BIT_MASK(nr);
+	addr += BIT_WORD(nr);
+	smp_mb();  /* barrier for proper semantics */
+	return (_atomic_andn(addr, mask) & mask) != 0;
+}
+
+/**
+ * test_and_change_bit - Change a bit and return its old value
+ * @nr: Bit to change
+ * @addr: Address to count from
+ *
+ * This operation is atomic and cannot be reordered.
+ * It also implies a memory barrier.
+ */
+static inline int test_and_change_bit(unsigned nr,
+				      volatile unsigned long *addr)
+{
+	unsigned long mask = BIT_MASK(nr);
+	addr += BIT_WORD(nr);
+	smp_mb();  /* barrier for proper semantics */
+	return (_atomic_xor(addr, mask) & mask) != 0;
+}
+
+#include <asm-generic/bitops/ext2-atomic.h>
+
+#endif /* _ASM_TILE_BITOPS_32_H */
diff --git a/kernel/arch/tile/include/asm/bitops_64.h b/kernel/arch/tile/include/asm/bitops_64.h
new file mode 100644
index 000000000..bb1a29221
--- /dev/null
+++ b/kernel/arch/tile/include/asm/bitops_64.h
@@ -0,0 +1,95 @@
+/*
+ * Copyright 2011 Tilera Corporation. All Rights Reserved.
+ *
+ *   This program is free software; you can redistribute it and/or
+ *   modify it under the terms of the GNU General Public License
+ *   as published by the Free Software Foundation, version 2.
+ *
+ *   This program is distributed in the hope that it will be useful, but
+ *   WITHOUT ANY WARRANTY; without even the implied warranty of
+ *   MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
+ *   NON INFRINGEMENT.  See the GNU General Public License for
+ *   more details.
+ */
+
+#ifndef _ASM_TILE_BITOPS_64_H
+#define _ASM_TILE_BITOPS_64_H
+
+#include <linux/compiler.h>
+#include <asm/cmpxchg.h>
+
+/* See <asm/bitops.h> for API comments. */
+
+static inline void set_bit(unsigned nr, volatile unsigned long *addr)
+{
+	unsigned long mask = (1UL << (nr % BITS_PER_LONG));
+	__insn_fetchor((void *)(addr + nr / BITS_PER_LONG), mask);
+}
+
+static inline void clear_bit(unsigned nr, volatile unsigned long *addr)
+{
+	unsigned long mask = (1UL << (nr % BITS_PER_LONG));
+	__insn_fetchand((void *)(addr + nr / BITS_PER_LONG), ~mask);
+}
+
+static inline void change_bit(unsigned nr, volatile unsigned long *addr)
+{
+	unsigned long mask = (1UL << (nr % BITS_PER_LONG));
+	unsigned long guess, oldval;
+	addr += nr / BITS_PER_LONG;
+	oldval = *addr;
+	do {
+		guess = oldval;
+		oldval = cmpxchg(addr, guess, guess ^ mask);
+	} while (guess != oldval);
+}
+
+
+/*
+ * The test_and_xxx_bit() routines require a memory fence before we
+ * start the operation, and after the operation completes.  We use
+ * smp_mb() before, and rely on the "!= 0" comparison, plus a compiler
+ * barrier(), to block until the atomic op is complete.
+ */
+
+static inline int test_and_set_bit(unsigned nr, volatile unsigned long *addr)
+{
+	int val;
+	unsigned long mask = (1UL << (nr % BITS_PER_LONG));
+	smp_mb();  /* barrier for proper semantics */
+	val = (__insn_fetchor((void *)(addr + nr / BITS_PER_LONG), mask)
+	       & mask) != 0;
+	barrier();
+	return val;
+}
+
+
+static inline int test_and_clear_bit(unsigned nr, volatile unsigned long *addr)
+{
+	int val;
+	unsigned long mask = (1UL << (nr % BITS_PER_LONG));
+	smp_mb();  /* barrier for proper semantics */
+	val = (__insn_fetchand((void *)(addr + nr / BITS_PER_LONG), ~mask)
+	       & mask) != 0;
+	barrier();
+	return val;
+}
+
+
+static inline int test_and_change_bit(unsigned nr,
+				      volatile unsigned long *addr)
+{
+	unsigned long mask = (1UL << (nr % BITS_PER_LONG));
+	unsigned long guess, oldval;
+	addr += nr / BITS_PER_LONG;
+	oldval = *addr;
+	do {
+		guess = oldval;
+		oldval = cmpxchg(addr, guess, guess ^ mask);
+	} while (guess != oldval);
+	return (oldval & mask) != 0;
+}
+
+#include <asm-generic/bitops/ext2-atomic-setbit.h>
+
+#endif /* _ASM_TILE_BITOPS_64_H */
diff --git a/kernel/arch/tile/include/asm/cache.h b/kernel/arch/tile/include/asm/cache.h
new file mode 100644
index 000000000..6160761d5
--- /dev/null
+++ b/kernel/arch/tile/include/asm/cache.h
@@ -0,0 +1,64 @@
+/*
+ * Copyright 2010 Tilera Corporation. All Rights Reserved.
+ *
+ *   This program is free software; you can redistribute it and/or
+ *   modify it under the terms of the GNU General Public License
+ *   as published by the Free Software Foundation, version 2.
+ *
+ *   This program is distributed in the hope that it will be useful, but
+ *   WITHOUT ANY WARRANTY; without even the implied warranty of
+ *   MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
+ *   NON INFRINGEMENT.  See the GNU General Public License for
+ *   more details.
+ */
+
+#ifndef _ASM_TILE_CACHE_H
+#define _ASM_TILE_CACHE_H
+
+#include <arch/chip.h>
+
+/* bytes per L1 data cache line */
+#define L1_CACHE_SHIFT		CHIP_L1D_LOG_LINE_SIZE()
+#define L1_CACHE_BYTES		(1 << L1_CACHE_SHIFT)
+
+/* bytes per L2 cache line */
+#define L2_CACHE_SHIFT		CHIP_L2_LOG_LINE_SIZE()
+#define L2_CACHE_BYTES		(1 << L2_CACHE_SHIFT)
+#define L2_CACHE_ALIGN(x)	(((x)+(L2_CACHE_BYTES-1)) & -L2_CACHE_BYTES)
+
+/*
+ * TILEPro I/O is not always coherent (networking typically uses coherent
+ * I/O, but PCI traffic does not) and setting ARCH_DMA_MINALIGN to the
+ * L2 cacheline size helps ensure that kernel heap allocations are aligned.
+ * TILE-Gx I/O is always coherent when used on hash-for-home pages.
+ *
+ * However, it's possible at runtime to request not to use hash-for-home
+ * for the kernel heap, in which case the kernel will use flush-and-inval
+ * to manage coherence.  As a result, we use L2_CACHE_BYTES for the
+ * DMA minimum alignment to avoid false sharing in the kernel heap.
+ */
+#define ARCH_DMA_MINALIGN	L2_CACHE_BYTES
+
+/* use the cache line size for the L2, which is where it counts */
+#define SMP_CACHE_BYTES_SHIFT	L2_CACHE_SHIFT
+#define SMP_CACHE_BYTES		L2_CACHE_BYTES
+#define INTERNODE_CACHE_SHIFT   L2_CACHE_SHIFT
+#define INTERNODE_CACHE_BYTES   L2_CACHE_BYTES
+
+/* Group together read-mostly things to avoid cache false sharing */
+#define __read_mostly __attribute__((__section__(".data..read_mostly")))
+
+/*
+ * Originally we used small TLB pages for kernel data and grouped some
+ * things together as "write once", enforcing the property at the end
+ * of initialization by making those pages read-only and non-coherent.
+ * This allowed better cache utilization since cache inclusion did not
+ * need to be maintained.  However, to do this requires an extra TLB
+ * entry, which on balance is more of a performance hit than the
+ * non-coherence is a performance gain, so we now just make "read
+ * mostly" and "write once" be synonyms.  We keep the attribute
+ * separate in case we change our minds at a future date.
+ */
+#define __write_once __read_mostly
+
+#endif /* _ASM_TILE_CACHE_H */
diff --git a/kernel/arch/tile/include/asm/cacheflush.h b/kernel/arch/tile/include/asm/cacheflush.h
new file mode 100644
index 000000000..92ee4c8a4
--- /dev/null
+++ b/kernel/arch/tile/include/asm/cacheflush.h
@@ -0,0 +1,160 @@
+/*
+ * Copyright 2010 Tilera Corporation. All Rights Reserved.
+ *
+ *   This program is free software; you can redistribute it and/or
+ *   modify it under the terms of the GNU General Public License
+ *   as published by the Free Software Foundation, version 2.
+ *
+ *   This program is distributed in the hope that it will be useful, but
+ *   WITHOUT ANY WARRANTY; without even the implied warranty of
+ *   MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
+ *   NON INFRINGEMENT.  See the GNU General Public License for
+ *   more details.
+ */
+
+#ifndef _ASM_TILE_CACHEFLUSH_H
+#define _ASM_TILE_CACHEFLUSH_H
+
+#include <arch/chip.h>
+
+/* Keep includes the same across arches.  */
+#include <linux/mm.h>
+#include <linux/cache.h>
+#include <arch/icache.h>
+
+/* Caches are physically-indexed and so don't need special treatment */
+#define flush_cache_all()			do { } while (0)
+#define flush_cache_mm(mm)			do { } while (0)
+#define flush_cache_dup_mm(mm)			do { } while (0)
+#define flush_cache_range(vma, start, end)	do { } while (0)
+#define flush_cache_page(vma, vmaddr, pfn)	do { } while (0)
+#define ARCH_IMPLEMENTS_FLUSH_DCACHE_PAGE 0
+#define flush_dcache_page(page)			do { } while (0)
+#define flush_dcache_mmap_lock(mapping)		do { } while (0)
+#define flush_dcache_mmap_unlock(mapping)	do { } while (0)
+#define flush_cache_vmap(start, end)		do { } while (0)
+#define flush_cache_vunmap(start, end)		do { } while (0)
+#define flush_icache_page(vma, pg)		do { } while (0)
+#define flush_icache_user_range(vma, pg, adr, len)	do { } while (0)
+
+/* Flush the icache just on this cpu */
+extern void __flush_icache_range(unsigned long start, unsigned long end);
+
+/* Flush the entire icache on this cpu. */
+#define __flush_icache() __flush_icache_range(0, CHIP_L1I_CACHE_SIZE())
+
+#ifdef CONFIG_SMP
+/*
+ * When the kernel writes to its own text we need to do an SMP
+ * broadcast to make the L1I coherent everywhere.  This includes
+ * module load and single step.
+ */
+extern void flush_icache_range(unsigned long start, unsigned long end);
+#else
+#define flush_icache_range __flush_icache_range
+#endif
+
+/*
+ * An update to an executable user page requires icache flushing.
+ * We could carefully update only tiles that are running this process,
+ * and rely on the fact that we flush the icache on every context
+ * switch to avoid doing extra work here.  But for now, I'll be
+ * conservative and just do a global icache flush.
+ */
+static inline void copy_to_user_page(struct vm_area_struct *vma,
+				     struct page *page, unsigned long vaddr,
+				     void *dst, void *src, int len)
+{
+	memcpy(dst, src, len);
+	if (vma->vm_flags & VM_EXEC) {
+		flush_icache_range((unsigned long) dst,
+				   (unsigned long) dst + len);
+	}
+}
+
+#define copy_from_user_page(vma, page, vaddr, dst, src, len) \
+	memcpy((dst), (src), (len))
+
+/* Flush a VA range; pads to L2 cacheline boundaries. */
+static inline void __flush_buffer(void *buffer, size_t size)
+{
+	char *next = (char *)((long)buffer & -L2_CACHE_BYTES);
+	char *finish = (char *)L2_CACHE_ALIGN((long)buffer + size);
+	while (next < finish) {
+		__insn_flush(next);
+		next += CHIP_FLUSH_STRIDE();
+	}
+}
+
+/* Flush & invalidate a VA range; pads to L2 cacheline boundaries. */
+static inline void __finv_buffer(void *buffer, size_t size)
+{
+	char *next = (char *)((long)buffer & -L2_CACHE_BYTES);
+	char *finish = (char *)L2_CACHE_ALIGN((long)buffer + size);
+	while (next < finish) {
+		__insn_finv(next);
+		next += CHIP_FINV_STRIDE();
+	}
+}
+
+
+/*
+ * Flush a locally-homecached VA range and wait for the evicted
+ * cachelines to hit memory.
+ */
+static inline void flush_buffer_local(void *buffer, size_t size)
+{
+	__flush_buffer(buffer, size);
+	mb_incoherent();
+}
+
+/*
+ * Flush and invalidate a locally-homecached VA range and wait for the
+ * evicted cachelines to hit memory.
+ */
+static inline void finv_buffer_local(void *buffer, size_t size)
+{
+	__finv_buffer(buffer, size);
+	mb_incoherent();
+}
+
+#ifdef __tilepro__
+/* Invalidate a VA range; pads to L2 cacheline boundaries. */
+static inline void __inv_buffer(void *buffer, size_t size)
+{
+	char *next = (char *)((long)buffer & -L2_CACHE_BYTES);
+	char *finish = (char *)L2_CACHE_ALIGN((long)buffer + size);
+	while (next < finish) {
+		__insn_inv(next);
+		next += CHIP_INV_STRIDE();
+	}
+}
+
+/* Invalidate a VA range and wait for it to be complete. */
+static inline void inv_buffer(void *buffer, size_t size)
+{
+	__inv_buffer(buffer, size);
+	mb();
+}
+#endif
+
+/*
+ * Flush and invalidate a VA range that is homed remotely, waiting
+ * until the memory controller holds the flushed values.  If "hfh" is
+ * true, we will do a more expensive flush involving additional loads
+ * to make sure we have touched all the possible home cpus of a buffer
+ * that is homed with "hash for home".
+ */
+void finv_buffer_remote(void *buffer, size_t size, int hfh);
+
+/*
+ * On SMP systems, when the scheduler does migration-cost autodetection,
+ * it needs a way to flush as much of the CPU's caches as possible:
+ *
+ * TODO: fill this in!
+ */
+static inline void sched_cacheflush(void)
+{
+}
+
+#endif /* _ASM_TILE_CACHEFLUSH_H */
diff --git a/kernel/arch/tile/include/asm/checksum.h b/kernel/arch/tile/include/asm/checksum.h
new file mode 100644
index 000000000..b21a2fdec
--- /dev/null
+++ b/kernel/arch/tile/include/asm/checksum.h
@@ -0,0 +1,42 @@
+/*
+ * Copyright 2010 Tilera Corporation. All Rights Reserved.
+ *
+ *   This program is free software; you can redistribute it and/or
+ *   modify it under the terms of the GNU General Public License
+ *   as published by the Free Software Foundation, version 2.
+ *
+ *   This program is distributed in the hope that it will be useful, but
+ *   WITHOUT ANY WARRANTY; without even the implied warranty of
+ *   MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
+ *   NON INFRINGEMENT.  See the GNU General Public License for
+ *   more details.
+ */
+
+#ifndef _ASM_TILE_CHECKSUM_H
+#define _ASM_TILE_CHECKSUM_H
+
+#include <asm-generic/checksum.h>
+
+/* Allow us to provide a more optimized do_csum(). */
+__wsum do_csum(const unsigned char *buff, int len);
+#define do_csum do_csum
+
+/*
+ * Return the sum of all the 16-bit subwords in a long.
+ * This sums two subwords on a 32-bit machine, and four on 64 bits.
+ * The implementation does two vector adds to capture any overflow.
+ */
+static inline unsigned int csum_long(unsigned long x)
+{
+	unsigned long ret;
+#ifdef __tilegx__
+	ret = __insn_v2sadu(x, 0);
+	ret = __insn_v2sadu(ret, 0);
+#else
+	ret = __insn_sadh_u(x, 0);
+	ret = __insn_sadh_u(ret, 0);
+#endif
+	return ret;
+}
+
+#endif /* _ASM_TILE_CHECKSUM_H */
diff --git a/kernel/arch/tile/include/asm/cmpxchg.h b/kernel/arch/tile/include/asm/cmpxchg.h
new file mode 100644
index 000000000..0ccda3c42
--- /dev/null
+++ b/kernel/arch/tile/include/asm/cmpxchg.h
@@ -0,0 +1,134 @@
+/*
+ * cmpxchg.h -- forked from asm/atomic.h with this copyright:
+ *
+ * Copyright 2010 Tilera Corporation. All Rights Reserved.
+ *
+ *   This program is free software; you can redistribute it and/or
+ *   modify it under the terms of the GNU General Public License
+ *   as published by the Free Software Foundation, version 2.
+ *
+ *   This program is distributed in the hope that it will be useful, but
+ *   WITHOUT ANY WARRANTY; without even the implied warranty of
+ *   MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
+ *   NON INFRINGEMENT.  See the GNU General Public License for
+ *   more details.
+ *
+ */
+
+#ifndef _ASM_TILE_CMPXCHG_H
+#define _ASM_TILE_CMPXCHG_H
+
+#ifndef __ASSEMBLY__
+
+#include <asm/barrier.h>
+
+/* Nonexistent functions intended to cause compile errors. */
+extern void __xchg_called_with_bad_pointer(void)
+	__compiletime_error("Bad argument size for xchg");
+extern void __cmpxchg_called_with_bad_pointer(void)
+	__compiletime_error("Bad argument size for cmpxchg");
+
+#ifndef __tilegx__
+
+/* Note the _atomic_xxx() routines include a final mb(). */
+int _atomic_xchg(int *ptr, int n);
+int _atomic_xchg_add(int *v, int i);
+int _atomic_xchg_add_unless(int *v, int a, int u);
+int _atomic_cmpxchg(int *ptr, int o, int n);
+long long _atomic64_xchg(long long *v, long long n);
+long long _atomic64_xchg_add(long long *v, long long i);
+long long _atomic64_xchg_add_unless(long long *v, long long a, long long u);
+long long _atomic64_cmpxchg(long long *v, long long o, long long n);
+
+#define xchg(ptr, n)							\
+	({								\
+		if (sizeof(*(ptr)) != 4)				\
+			__xchg_called_with_bad_pointer();		\
+		smp_mb();						\
+		(typeof(*(ptr)))_atomic_xchg((int *)(ptr), (int)(n));	\
+	})
+
+#define cmpxchg(ptr, o, n)						\
+	({								\
+		if (sizeof(*(ptr)) != 4)				\
+			__cmpxchg_called_with_bad_pointer();		\
+		smp_mb();						\
+		(typeof(*(ptr)))_atomic_cmpxchg((int *)ptr, (int)o,	\
+						(int)n);		\
+	})
+
+#define xchg64(ptr, n)							\
+	({								\
+		if (sizeof(*(ptr)) != 8)				\
+			__xchg_called_with_bad_pointer();		\
+		smp_mb();						\
+		(typeof(*(ptr)))_atomic64_xchg((long long *)(ptr),	\
+						(long long)(n));	\
+	})
+
+#define cmpxchg64(ptr, o, n)						\
+	({								\
+		if (sizeof(*(ptr)) != 8)				\
+			__cmpxchg_called_with_bad_pointer();		\
+		smp_mb();						\
+		(typeof(*(ptr)))_atomic64_cmpxchg((long long *)ptr,	\
+					(long long)o, (long long)n);	\
+	})
+
+#else
+
+#define xchg(ptr, n)							\
+	({								\
+		typeof(*(ptr)) __x;					\
+		smp_mb();						\
+		switch (sizeof(*(ptr))) {				\
+		case 4:							\
+			__x = (typeof(__x))(unsigned long)		\
+				__insn_exch4((ptr),			\
+					(u32)(unsigned long)(n));	\
+			break;						\
+		case 8:							\
+			__x = (typeof(__x))				\
+				__insn_exch((ptr), (unsigned long)(n));	\
+			break;						\
+		default:						\
+			__xchg_called_with_bad_pointer();		\
+			break;						\
+		}							\
+		smp_mb();						\
+		__x;							\
+	})
+
+#define cmpxchg(ptr, o, n)						\
+	({								\
+		typeof(*(ptr)) __x;					\
+		__insn_mtspr(SPR_CMPEXCH_VALUE, (unsigned long)(o));	\
+		smp_mb();						\
+		switch (sizeof(*(ptr))) {				\
+		case 4:							\
+			__x = (typeof(__x))(unsigned long)		\
+				__insn_cmpexch4((ptr),			\
+					(u32)(unsigned long)(n));	\
+			break;						\
+		case 8:							\
+			__x = (typeof(__x))__insn_cmpexch((ptr),	\
+						(long long)(n));	\
+			break;						\
+		default:						\
+			__cmpxchg_called_with_bad_pointer();		\
+			break;						\
+		}							\
+		smp_mb();						\
+		__x;							\
+	})
+
+#define xchg64 xchg
+#define cmpxchg64 cmpxchg
+
+#endif
+
+#define tas(ptr) xchg((ptr), 1)
+
+#endif /* __ASSEMBLY__ */
+
+#endif /* _ASM_TILE_CMPXCHG_H */
diff --git a/kernel/arch/tile/include/asm/compat.h b/kernel/arch/tile/include/asm/compat.h
new file mode 100644
index 000000000..c14e36f00
--- /dev/null
+++ b/kernel/arch/tile/include/asm/compat.h
@@ -0,0 +1,296 @@
+/*
+ * Copyright 2010 Tilera Corporation. All Rights Reserved.
+ *
+ *   This program is free software; you can redistribute it and/or
+ *   modify it under the terms of the GNU General Public License
+ *   as published by the Free Software Foundation, version 2.
+ *
+ *   This program is distributed in the hope that it will be useful, but
+ *   WITHOUT ANY WARRANTY; without even the implied warranty of
+ *   MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
+ *   NON INFRINGEMENT.  See the GNU General Public License for
+ *   more details.
+ */
+
+#ifndef _ASM_TILE_COMPAT_H
+#define _ASM_TILE_COMPAT_H
+
+/*
+ * Architecture specific compatibility types
+ */
+#include <linux/types.h>
+#include <linux/sched.h>
+
+#define COMPAT_USER_HZ	100
+
+/* "long" and pointer-based types are different. */
+typedef s32		compat_long_t;
+typedef u32		compat_ulong_t;
+typedef u32		compat_size_t;
+typedef s32		compat_ssize_t;
+typedef s32		compat_off_t;
+typedef s32		compat_time_t;
+typedef s32		compat_clock_t;
+typedef u32		compat_ino_t;
+typedef u32		compat_caddr_t;
+typedef	u32		compat_uptr_t;
+
+/* Many types are "int" or otherwise the same. */
+typedef __kernel_pid_t compat_pid_t;
+typedef __kernel_uid_t __compat_uid_t;
+typedef __kernel_gid_t __compat_gid_t;
+typedef __kernel_uid32_t __compat_uid32_t;
+typedef __kernel_uid32_t __compat_gid32_t;
+typedef __kernel_mode_t compat_mode_t;
+typedef __kernel_dev_t compat_dev_t;
+typedef __kernel_loff_t compat_loff_t;
+typedef __kernel_ipc_pid_t compat_ipc_pid_t;
+typedef __kernel_daddr_t compat_daddr_t;
+typedef __kernel_fsid_t	compat_fsid_t;
+typedef __kernel_timer_t compat_timer_t;
+typedef __kernel_key_t compat_key_t;
+typedef int compat_int_t;
+typedef s64 compat_s64;
+typedef uint compat_uint_t;
+typedef u64 compat_u64;
+
+/* We use the same register dump format in 32-bit images. */
+typedef unsigned long compat_elf_greg_t;
+#define COMPAT_ELF_NGREG (sizeof(struct pt_regs) / sizeof(compat_elf_greg_t))
+typedef compat_elf_greg_t compat_elf_gregset_t[COMPAT_ELF_NGREG];
+
+struct compat_timespec {
+	compat_time_t	tv_sec;
+	s32		tv_nsec;
+};
+
+struct compat_timeval {
+	compat_time_t	tv_sec;
+	s32		tv_usec;
+};
+
+#define compat_stat stat
+#define compat_statfs statfs
+
+struct compat_sysctl {
+	unsigned int	name;
+	int		nlen;
+	unsigned int	oldval;
+	unsigned int	oldlenp;
+	unsigned int	newval;
+	unsigned int	newlen;
+	unsigned int	__unused[4];
+};
+
+
+struct compat_flock {
+	short		l_type;
+	short		l_whence;
+	compat_off_t	l_start;
+	compat_off_t	l_len;
+	compat_pid_t	l_pid;
+};
+
+#define F_GETLK64	12	/*  using 'struct flock64' */
+#define F_SETLK64	13
+#define F_SETLKW64	14
+
+struct compat_flock64 {
+	short		l_type;
+	short		l_whence;
+	compat_loff_t	l_start;
+	compat_loff_t	l_len;
+	compat_pid_t	l_pid;
+};
+
+#define COMPAT_RLIM_INFINITY		0xffffffff
+
+#define _COMPAT_NSIG		64
+#define _COMPAT_NSIG_BPW	32
+
+typedef u32               compat_sigset_word;
+
+typedef union compat_sigval {
+	compat_int_t	sival_int;
+	compat_uptr_t	sival_ptr;
+} compat_sigval_t;
+
+#define COMPAT_SI_PAD_SIZE	(128/sizeof(int) - 3)
+
+typedef struct compat_siginfo {
+	int si_signo;
+	int si_errno;
+	int si_code;
+
+	union {
+		int _pad[COMPAT_SI_PAD_SIZE];
+
+		/* kill() */
+		struct {
+			unsigned int _pid;	/* sender's pid */
+			unsigned int _uid;	/* sender's uid */
+		} _kill;
+
+		/* POSIX.1b timers */
+		struct {
+			compat_timer_t _tid;	/* timer id */
+			int _overrun;		/* overrun count */
+			compat_sigval_t _sigval;	/* same as below */
+			int _sys_private;	/* not to be passed to user */
+			int _overrun_incr;	/* amount to add to overrun */
+		} _timer;
+
+		/* POSIX.1b signals */
+		struct {
+			unsigned int _pid;	/* sender's pid */
+			unsigned int _uid;	/* sender's uid */
+			compat_sigval_t _sigval;
+		} _rt;
+
+		/* SIGCHLD */
+		struct {
+			unsigned int _pid;	/* which child */
+			unsigned int _uid;	/* sender's uid */
+			int _status;		/* exit code */
+			compat_clock_t _utime;
+			compat_clock_t _stime;
+		} _sigchld;
+
+		/* SIGILL, SIGFPE, SIGSEGV, SIGBUS */
+		struct {
+			unsigned int _addr;	/* faulting insn/memory ref. */
+#ifdef __ARCH_SI_TRAPNO
+			int _trapno;	/* TRAP # which caused the signal */
+#endif
+		} _sigfault;
+
+		/* SIGPOLL */
+		struct {
+			int _band;	/* POLL_IN, POLL_OUT, POLL_MSG */
+			int _fd;
+		} _sigpoll;
+	} _sifields;
+} compat_siginfo_t;
+
+#define COMPAT_OFF_T_MAX	0x7fffffff
+#define COMPAT_LOFF_T_MAX	0x7fffffffffffffffL
+
+struct compat_ipc64_perm {
+	compat_key_t key;
+	__compat_uid32_t uid;
+	__compat_gid32_t gid;
+	__compat_uid32_t cuid;
+	__compat_gid32_t cgid;
+	unsigned short mode;
+	unsigned short __pad1;
+	unsigned short seq;
+	unsigned short __pad2;
+	compat_ulong_t unused1;
+	compat_ulong_t unused2;
+};
+
+struct compat_semid64_ds {
+	struct compat_ipc64_perm sem_perm;
+	compat_time_t  sem_otime;
+	compat_ulong_t __unused1;
+	compat_time_t  sem_ctime;
+	compat_ulong_t __unused2;
+	compat_ulong_t sem_nsems;
+	compat_ulong_t __unused3;
+	compat_ulong_t __unused4;
+};
+
+struct compat_msqid64_ds {
+	struct compat_ipc64_perm msg_perm;
+	compat_time_t  msg_stime;
+	compat_ulong_t __unused1;
+	compat_time_t  msg_rtime;
+	compat_ulong_t __unused2;
+	compat_time_t  msg_ctime;
+	compat_ulong_t __unused3;
+	compat_ulong_t msg_cbytes;
+	compat_ulong_t msg_qnum;
+	compat_ulong_t msg_qbytes;
+	compat_pid_t   msg_lspid;
+	compat_pid_t   msg_lrpid;
+	compat_ulong_t __unused4;
+	compat_ulong_t __unused5;
+};
+
+struct compat_shmid64_ds {
+	struct compat_ipc64_perm shm_perm;
+	compat_size_t  shm_segsz;
+	compat_time_t  shm_atime;
+	compat_ulong_t __unused1;
+	compat_time_t  shm_dtime;
+	compat_ulong_t __unused2;
+	compat_time_t  shm_ctime;
+	compat_ulong_t __unused3;
+	compat_pid_t   shm_cpid;
+	compat_pid_t   shm_lpid;
+	compat_ulong_t shm_nattch;
+	compat_ulong_t __unused4;
+	compat_ulong_t __unused5;
+};
+
+/*
+ * A pointer passed in from user mode. This should not
+ * be used for syscall parameters, just declare them
+ * as pointers because the syscall entry code will have
+ * appropriately converted them already.
+ */
+
+static inline void __user *compat_ptr(compat_uptr_t uptr)
+{
+	return (void __user *)(long)(s32)uptr;
+}
+
+static inline compat_uptr_t ptr_to_compat(void __user *uptr)
+{
+	return (u32)(unsigned long)uptr;
+}
+
+/* Sign-extend when storing a kernel pointer to a user's ptregs. */
+static inline unsigned long ptr_to_compat_reg(void __user *uptr)
+{
+	return (long)(int)(long __force)uptr;
+}
+
+static inline void __user *arch_compat_alloc_user_space(long len)
+{
+	struct pt_regs *regs = task_pt_regs(current);
+	return (void __user *)regs->sp - len;
+}
+
+static inline int is_compat_task(void)
+{
+	return current_thread_info()->status & TS_COMPAT;
+}
+
+extern int compat_setup_rt_frame(struct ksignal *ksig, sigset_t *set,
+				 struct pt_regs *regs);
+
+/* Compat syscalls. */
+struct compat_siginfo;
+struct compat_sigaltstack;
+long compat_sys_rt_sigreturn(void);
+long compat_sys_truncate64(char __user *filename, u32 dummy, u32 low, u32 high);
+long compat_sys_ftruncate64(unsigned int fd, u32 dummy, u32 low, u32 high);
+long compat_sys_pread64(unsigned int fd, char __user *ubuf, size_t count,
+			u32 dummy, u32 low, u32 high);
+long compat_sys_pwrite64(unsigned int fd, char __user *ubuf, size_t count,
+			 u32 dummy, u32 low, u32 high);
+long compat_sys_sync_file_range2(int fd, unsigned int flags,
+				 u32 offset_lo, u32 offset_hi,
+				 u32 nbytes_lo, u32 nbytes_hi);
+long compat_sys_fallocate(int fd, int mode,
+			  u32 offset_lo, u32 offset_hi,
+			  u32 len_lo, u32 len_hi);
+long compat_sys_llseek(unsigned int fd, unsigned int offset_high,
+		       unsigned int offset_low, loff_t __user * result,
+		       unsigned int origin);
+
+/* Assembly trampoline to avoid clobbering r0. */
+long _compat_sys_rt_sigreturn(void);
+
+#endif /* _ASM_TILE_COMPAT_H */
diff --git a/kernel/arch/tile/include/asm/current.h b/kernel/arch/tile/include/asm/current.h
new file mode 100644
index 000000000..da21acf02
--- /dev/null
+++ b/kernel/arch/tile/include/asm/current.h
@@ -0,0 +1,31 @@
+/*
+ * Copyright 2010 Tilera Corporation. All Rights Reserved.
+ *
+ *   This program is free software; you can redistribute it and/or
+ *   modify it under the terms of the GNU General Public License
+ *   as published by the Free Software Foundation, version 2.
+ *
+ *   This program is distributed in the hope that it will be useful, but
+ *   WITHOUT ANY WARRANTY; without even the implied warranty of
+ *   MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
+ *   NON INFRINGEMENT.  See the GNU General Public License for
+ *   more details.
+ */
+
+#ifndef _ASM_TILE_CURRENT_H
+#define _ASM_TILE_CURRENT_H
+
+#include <linux/thread_info.h>
+
+struct task_struct;
+
+static inline struct task_struct *get_current(void)
+{
+	return current_thread_info()->task;
+}
+#define current get_current()
+
+/* Return a usable "task_struct" pointer even if the real one is corrupt. */
+struct task_struct *validate_current(void);
+
+#endif /* _ASM_TILE_CURRENT_H */
diff --git a/kernel/arch/tile/include/asm/delay.h b/kernel/arch/tile/include/asm/delay.h
new file mode 100644
index 000000000..97b0e69e7
--- /dev/null
+++ b/kernel/arch/tile/include/asm/delay.h
@@ -0,0 +1,34 @@
+/*
+ * Copyright 2010 Tilera Corporation. All Rights Reserved.
+ *
+ *   This program is free software; you can redistribute it and/or
+ *   modify it under the terms of the GNU General Public License
+ *   as published by the Free Software Foundation, version 2.
+ *
+ *   This program is distributed in the hope that it will be useful, but
+ *   WITHOUT ANY WARRANTY; without even the implied warranty of
+ *   MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
+ *   NON INFRINGEMENT.  See the GNU General Public License for
+ *   more details.
+ */
+
+#ifndef _ASM_TILE_DELAY_H
+#define _ASM_TILE_DELAY_H
+
+/* Undefined functions to get compile-time errors. */
+extern void __bad_udelay(void);
+extern void __bad_ndelay(void);
+
+extern void __udelay(unsigned long usecs);
+extern void __ndelay(unsigned long nsecs);
+extern void __delay(unsigned long loops);
+
+#define udelay(n) (__builtin_constant_p(n) ? \
+	((n) > 20000 ? __bad_udelay() : __ndelay((n) * 1000)) : \
+	__udelay(n))
+
+#define ndelay(n) (__builtin_constant_p(n) ? \
+	((n) > 20000 ? __bad_ndelay() : __ndelay(n)) : \
+	__ndelay(n))
+
+#endif /* _ASM_TILE_DELAY_H */
diff --git a/kernel/arch/tile/include/asm/device.h b/kernel/arch/tile/include/asm/device.h
new file mode 100644
index 000000000..6ab8bf146
--- /dev/null
+++ b/kernel/arch/tile/include/asm/device.h
@@ -0,0 +1,36 @@
+/*
+ * Copyright 2010 Tilera Corporation. All Rights Reserved.
+ *
+ *   This program is free software; you can redistribute it and/or
+ *   modify it under the terms of the GNU General Public License
+ *   as published by the Free Software Foundation, version 2.
+ *
+ *   This program is distributed in the hope that it will be useful, but
+ *   WITHOUT ANY WARRANTY; without even the implied warranty of
+ *   MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
+ *   NON INFRINGEMENT.  See the GNU General Public License for
+ *   more details.
+ * Arch specific extensions to struct device
+ */
+
+#ifndef _ASM_TILE_DEVICE_H
+#define _ASM_TILE_DEVICE_H
+
+struct dev_archdata {
+	/* DMA operations on that device */
+        struct dma_map_ops	*dma_ops;
+
+	/* Offset of the DMA address from the PA. */
+	dma_addr_t		dma_offset;
+
+	/*
+	 * Highest DMA address that can be generated by devices that
+	 * have limited DMA capability, i.e. non 64-bit capable.
+	 */
+	dma_addr_t		max_direct_dma_addr;
+};
+
+struct pdev_archdata {
+};
+
+#endif /* _ASM_TILE_DEVICE_H */
diff --git a/kernel/arch/tile/include/asm/dma-mapping.h b/kernel/arch/tile/include/asm/dma-mapping.h
new file mode 100644
index 000000000..1eae359d8
--- /dev/null
+++ b/kernel/arch/tile/include/asm/dma-mapping.h
@@ -0,0 +1,158 @@
+/*
+ * Copyright 2010 Tilera Corporation. All Rights Reserved.
+ *
+ *   This program is free software; you can redistribute it and/or
+ *   modify it under the terms of the GNU General Public License
+ *   as published by the Free Software Foundation, version 2.
+ *
+ *   This program is distributed in the hope that it will be useful, but
+ *   WITHOUT ANY WARRANTY; without even the implied warranty of
+ *   MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
+ *   NON INFRINGEMENT.  See the GNU General Public License for
+ *   more details.
+ */
+
+#ifndef _ASM_TILE_DMA_MAPPING_H
+#define _ASM_TILE_DMA_MAPPING_H
+
+#include <linux/mm.h>
+#include <linux/scatterlist.h>
+#include <linux/cache.h>
+#include <linux/io.h>
+
+#ifdef __tilegx__
+#define ARCH_HAS_DMA_GET_REQUIRED_MASK
+#endif
+
+extern struct dma_map_ops *tile_dma_map_ops;
+extern struct dma_map_ops *gx_pci_dma_map_ops;
+extern struct dma_map_ops *gx_legacy_pci_dma_map_ops;
+extern struct dma_map_ops *gx_hybrid_pci_dma_map_ops;
+
+static inline struct dma_map_ops *get_dma_ops(struct device *dev)
+{
+	if (dev && dev->archdata.dma_ops)
+		return dev->archdata.dma_ops;
+	else
+		return tile_dma_map_ops;
+}
+
+static inline dma_addr_t get_dma_offset(struct device *dev)
+{
+	return dev->archdata.dma_offset;
+}
+
+static inline void set_dma_offset(struct device *dev, dma_addr_t off)
+{
+	dev->archdata.dma_offset = off;
+}
+
+static inline dma_addr_t phys_to_dma(struct device *dev, phys_addr_t paddr)
+{
+	return paddr;
+}
+
+static inline phys_addr_t dma_to_phys(struct device *dev, dma_addr_t daddr)
+{
+	return daddr;
+}
+
+static inline void dma_mark_clean(void *addr, size_t size) {}
+
+#include <asm-generic/dma-mapping-common.h>
+
+static inline void set_dma_ops(struct device *dev, struct dma_map_ops *ops)
+{
+	dev->archdata.dma_ops = ops;
+}
+
+static inline bool dma_capable(struct device *dev, dma_addr_t addr, size_t size)
+{
+	if (!dev->dma_mask)
+		return 0;
+
+	return addr + size - 1 <= *dev->dma_mask;
+}
+
+static inline int
+dma_mapping_error(struct device *dev, dma_addr_t dma_addr)
+{
+	debug_dma_mapping_error(dev, dma_addr);
+	return get_dma_ops(dev)->mapping_error(dev, dma_addr);
+}
+
+static inline int
+dma_supported(struct device *dev, u64 mask)
+{
+	return get_dma_ops(dev)->dma_supported(dev, mask);
+}
+
+static inline int
+dma_set_mask(struct device *dev, u64 mask)
+{
+	struct dma_map_ops *dma_ops = get_dma_ops(dev);
+
+	/*
+	 * For PCI devices with 64-bit DMA addressing capability, promote
+	 * the dma_ops to hybrid, with the consistent memory DMA space limited
+	 * to 32-bit. For 32-bit capable devices, limit the streaming DMA
+	 * address range to max_direct_dma_addr.
+	 */
+	if (dma_ops == gx_pci_dma_map_ops ||
+	    dma_ops == gx_hybrid_pci_dma_map_ops ||
+	    dma_ops == gx_legacy_pci_dma_map_ops) {
+		if (mask == DMA_BIT_MASK(64) &&
+		    dma_ops == gx_legacy_pci_dma_map_ops)
+			set_dma_ops(dev, gx_hybrid_pci_dma_map_ops);
+		else if (mask > dev->archdata.max_direct_dma_addr)
+			mask = dev->archdata.max_direct_dma_addr;
+	}
+
+	if (!dev->dma_mask || !dma_supported(dev, mask))
+		return -EIO;
+
+	*dev->dma_mask = mask;
+
+	return 0;
+}
+
+static inline void *dma_alloc_attrs(struct device *dev, size_t size,
+				    dma_addr_t *dma_handle, gfp_t flag,
+				    struct dma_attrs *attrs)
+{
+	struct dma_map_ops *dma_ops = get_dma_ops(dev);
+	void *cpu_addr;
+
+	cpu_addr = dma_ops->alloc(dev, size, dma_handle, flag, attrs);
+
+	debug_dma_alloc_coherent(dev, size, *dma_handle, cpu_addr);
+
+	return cpu_addr;
+}
+
+static inline void dma_free_attrs(struct device *dev, size_t size,
+				  void *cpu_addr, dma_addr_t dma_handle,
+				  struct dma_attrs *attrs)
+{
+	struct dma_map_ops *dma_ops = get_dma_ops(dev);
+
+	debug_dma_free_coherent(dev, size, cpu_addr, dma_handle);
+
+	dma_ops->free(dev, size, cpu_addr, dma_handle, attrs);
+}
+
+#define dma_alloc_coherent(d, s, h, f) dma_alloc_attrs(d, s, h, f, NULL)
+#define dma_alloc_noncoherent(d, s, h, f) dma_alloc_attrs(d, s, h, f, NULL)
+#define dma_free_coherent(d, s, v, h) dma_free_attrs(d, s, v, h, NULL)
+#define dma_free_noncoherent(d, s, v, h) dma_free_attrs(d, s, v, h, NULL)
+
+/*
+ * dma_alloc_noncoherent() is #defined to return coherent memory,
+ * so there's no need to do any flushing here.
+ */
+static inline void dma_cache_sync(struct device *dev, void *vaddr, size_t size,
+				  enum dma_data_direction direction)
+{
+}
+
+#endif /* _ASM_TILE_DMA_MAPPING_H */
diff --git a/kernel/arch/tile/include/asm/dma.h b/kernel/arch/tile/include/asm/dma.h
new file mode 100644
index 000000000..12a7ca16d
--- /dev/null
+++ b/kernel/arch/tile/include/asm/dma.h
@@ -0,0 +1,25 @@
+/*
+ * Copyright 2010 Tilera Corporation. All Rights Reserved.
+ *
+ *   This program is free software; you can redistribute it and/or
+ *   modify it under the terms of the GNU General Public License
+ *   as published by the Free Software Foundation, version 2.
+ *
+ *   This program is distributed in the hope that it will be useful, but
+ *   WITHOUT ANY WARRANTY; without even the implied warranty of
+ *   MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
+ *   NON INFRINGEMENT.  See the GNU General Public License for
+ *   more details.
+ */
+
+#ifndef _ASM_TILE_DMA_H
+#define _ASM_TILE_DMA_H
+
+#include <asm-generic/dma.h>
+
+/* Needed by drivers/pci/quirks.c */
+#ifdef CONFIG_PCI
+extern int isa_dma_bridge_buggy;
+#endif
+
+#endif /* _ASM_TILE_DMA_H */
diff --git a/kernel/arch/tile/include/asm/edac.h b/kernel/arch/tile/include/asm/edac.h
new file mode 100644
index 000000000..87fc83eea
--- /dev/null
+++ b/kernel/arch/tile/include/asm/edac.h
@@ -0,0 +1,29 @@
+/*
+ * Copyright 2011 Tilera Corporation. All Rights Reserved.
+ *
+ *   This program is free software; you can redistribute it and/or
+ *   modify it under the terms of the GNU General Public License
+ *   as published by the Free Software Foundation, version 2.
+ *
+ *   This program is distributed in the hope that it will be useful, but
+ *   WITHOUT ANY WARRANTY; without even the implied warranty of
+ *   MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
+ *   NON INFRINGEMENT.  See the GNU General Public License for
+ *   more details.
+ */
+
+#ifndef _ASM_TILE_EDAC_H
+#define _ASM_TILE_EDAC_H
+
+/* ECC atomic, DMA, SMP and interrupt safe scrub function */
+
+static inline void atomic_scrub(void *va, u32 size)
+{
+	/*
+	 * These is nothing to be done here because CE is
+	 * corrected by the mshim.
+	 */
+	return;
+}
+
+#endif /* _ASM_TILE_EDAC_H */
diff --git a/kernel/arch/tile/include/asm/elf.h b/kernel/arch/tile/include/asm/elf.h
new file mode 100644
index 000000000..41d9878a9
--- /dev/null
+++ b/kernel/arch/tile/include/asm/elf.h
@@ -0,0 +1,183 @@
+/*
+ * Copyright 2010 Tilera Corporation. All Rights Reserved.
+ *
+ *   This program is free software; you can redistribute it and/or
+ *   modify it under the terms of the GNU General Public License
+ *   as published by the Free Software Foundation, version 2.
+ *
+ *   This program is distributed in the hope that it will be useful, but
+ *   WITHOUT ANY WARRANTY; without even the implied warranty of
+ *   MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
+ *   NON INFRINGEMENT.  See the GNU General Public License for
+ *   more details.
+ */
+
+#ifndef _ASM_TILE_ELF_H
+#define _ASM_TILE_ELF_H
+
+/*
+ * ELF register definitions.
+ */
+
+#include <arch/chip.h>
+
+#include <linux/ptrace.h>
+#include <asm/byteorder.h>
+#include <asm/page.h>
+
+typedef unsigned long elf_greg_t;
+
+#define ELF_NGREG (sizeof(struct pt_regs) / sizeof(elf_greg_t))
+typedef elf_greg_t elf_gregset_t[ELF_NGREG];
+
+#define EM_TILEPRO 188
+#define EM_TILEGX  191
+
+/* Provide a nominal data structure. */
+#define ELF_NFPREG	0
+typedef double elf_fpreg_t;
+typedef elf_fpreg_t elf_fpregset_t[ELF_NFPREG];
+
+#ifdef __tilegx__
+#define ELF_CLASS	ELFCLASS64
+#else
+#define ELF_CLASS	ELFCLASS32
+#endif
+#ifdef __BIG_ENDIAN__
+#define ELF_DATA	ELFDATA2MSB
+#else
+#define ELF_DATA	ELFDATA2LSB
+#endif
+
+/*
+ * There seems to be a bug in how compat_binfmt_elf.c works: it
+ * #undefs ELF_ARCH, but it is then used in binfmt_elf.c for fill_note_info().
+ * Hack around this by providing an enum value of ELF_ARCH.
+ */
+enum { ELF_ARCH = CHIP_ELF_TYPE() };
+#define ELF_ARCH ELF_ARCH
+
+/*
+ * This is used to ensure we don't load something for the wrong architecture.
+ */
+#define elf_check_arch(x)  \
+	((x)->e_ident[EI_CLASS] == ELF_CLASS && \
+	 (x)->e_ident[EI_DATA] == ELF_DATA && \
+	 (x)->e_machine == CHIP_ELF_TYPE())
+
+/* The module loader only handles a few relocation types. */
+#ifndef __tilegx__
+#define R_TILE_32                 1
+#define R_TILE_JOFFLONG_X1       15
+#define R_TILE_IMM16_X0_LO       25
+#define R_TILE_IMM16_X1_LO       26
+#define R_TILE_IMM16_X0_HA       29
+#define R_TILE_IMM16_X1_HA       30
+#else
+#define R_TILEGX_64                       1
+#define R_TILEGX_JUMPOFF_X1              21
+#define R_TILEGX_IMM16_X0_HW0            36
+#define R_TILEGX_IMM16_X1_HW0            37
+#define R_TILEGX_IMM16_X0_HW1            38
+#define R_TILEGX_IMM16_X1_HW1            39
+#define R_TILEGX_IMM16_X0_HW2_LAST       48
+#define R_TILEGX_IMM16_X1_HW2_LAST       49
+#endif
+
+/* Use standard page size for core dumps. */
+#define ELF_EXEC_PAGESIZE	PAGE_SIZE
+
+/*
+ * This is the location that an ET_DYN program is loaded if exec'ed.  Typical
+ * use of this is to invoke "./ld.so someprog" to test out a new version of
+ * the loader.  We need to make sure that it is out of the way of the program
+ * that it will "exec", and that there is sufficient room for the brk.
+ */
+#define ELF_ET_DYN_BASE         (TASK_SIZE / 3 * 2)
+
+#define ELF_CORE_COPY_REGS(_dest, _regs)			\
+	memcpy((char *) &_dest, (char *) _regs,			\
+	       sizeof(struct pt_regs));
+
+/* No additional FP registers to copy. */
+#define ELF_CORE_COPY_FPREGS(t, fpu) 0
+
+/*
+ * This yields a mask that user programs can use to figure out what
+ * instruction set this CPU supports.  This could be done in user space,
+ * but it's not easy, and we've already done it here.
+ */
+#define ELF_HWCAP	(0)
+
+/*
+ * This yields a string that ld.so will use to load implementation
+ * specific libraries for optimization.  This is more specific in
+ * intent than poking at uname or /proc/cpuinfo.
+ */
+#define ELF_PLATFORM  (NULL)
+
+extern void elf_plat_init(struct pt_regs *regs, unsigned long load_addr);
+
+#define ELF_PLAT_INIT(_r, load_addr) elf_plat_init(_r, load_addr)
+
+extern int dump_task_regs(struct task_struct *, elf_gregset_t *);
+#define ELF_CORE_COPY_TASK_REGS(tsk, elf_regs) dump_task_regs(tsk, elf_regs)
+
+/* Tilera Linux has no personalities currently, so no need to do anything. */
+#define SET_PERSONALITY(ex) do { } while (0)
+
+#define ARCH_HAS_SETUP_ADDITIONAL_PAGES
+/* Support auto-mapping of the user interrupt vectors. */
+struct linux_binprm;
+extern int arch_setup_additional_pages(struct linux_binprm *bprm,
+				       int executable_stack);
+#define ARCH_DLINFO \
+do { \
+	NEW_AUX_ENT(AT_SYSINFO_EHDR, VDSO_BASE); \
+} while (0)
+
+struct mm_struct;
+extern unsigned long arch_randomize_brk(struct mm_struct *mm);
+#define arch_randomize_brk arch_randomize_brk
+
+#ifdef CONFIG_COMPAT
+
+#define COMPAT_ELF_PLATFORM "tilegx-m32"
+
+/*
+ * "Compat" binaries have the same machine type, but 32-bit class,
+ * since they're not a separate machine type, but just a 32-bit
+ * variant of the standard 64-bit architecture.
+ */
+#define compat_elf_check_arch(x)  \
+	((x)->e_ident[EI_CLASS] == ELFCLASS32 && \
+	 (x)->e_machine == CHIP_ELF_TYPE())
+
+#define compat_start_thread(regs, ip, usp) do { \
+		regs->pc = ptr_to_compat_reg((void *)(ip)); \
+		regs->sp = ptr_to_compat_reg((void *)(usp)); \
+		single_step_execve();	\
+	} while (0)
+
+/*
+ * Use SET_PERSONALITY to indicate compatibility via TS_COMPAT.
+ */
+#undef SET_PERSONALITY
+#define SET_PERSONALITY(ex) \
+do { \
+	set_personality(PER_LINUX | (current->personality & (~PER_MASK))); \
+	current_thread_info()->status &= ~TS_COMPAT; \
+} while (0)
+#define COMPAT_SET_PERSONALITY(ex) \
+do { \
+	set_personality(PER_LINUX | (current->personality & (~PER_MASK))); \
+	current_thread_info()->status |= TS_COMPAT; \
+} while (0)
+
+#define COMPAT_ELF_ET_DYN_BASE (0xffffffff / 3 * 2)
+
+#endif /* CONFIG_COMPAT */
+
+#define CORE_DUMP_USE_REGSET
+
+#endif /* _ASM_TILE_ELF_H */
diff --git a/kernel/arch/tile/include/asm/fixmap.h b/kernel/arch/tile/include/asm/fixmap.h
new file mode 100644
index 000000000..ffe2637ae
--- /dev/null
+++ b/kernel/arch/tile/include/asm/fixmap.h
@@ -0,0 +1,87 @@
+/*
+ * Copyright (C) 1998 Ingo Molnar
+ * Copyright 2010 Tilera Corporation. All Rights Reserved.
+ *
+ *   This program is free software; you can redistribute it and/or
+ *   modify it under the terms of the GNU General Public License
+ *   as published by the Free Software Foundation, version 2.
+ *
+ *   This program is distributed in the hope that it will be useful, but
+ *   WITHOUT ANY WARRANTY; without even the implied warranty of
+ *   MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
+ *   NON INFRINGEMENT.  See the GNU General Public License for
+ *   more details.
+ */
+
+#ifndef _ASM_TILE_FIXMAP_H
+#define _ASM_TILE_FIXMAP_H
+
+#include <asm/page.h>
+
+#ifndef __ASSEMBLY__
+#include <linux/kernel.h>
+#ifdef CONFIG_HIGHMEM
+#include <linux/threads.h>
+#include <asm/kmap_types.h>
+#endif
+
+/*
+ * Here we define all the compile-time 'special' virtual
+ * addresses. The point is to have a constant address at
+ * compile time, but to set the physical address only
+ * in the boot process. We allocate these special addresses
+ * from the end of supervisor virtual memory backwards.
+ * Also this lets us do fail-safe vmalloc(), we
+ * can guarantee that these special addresses and
+ * vmalloc()-ed addresses never overlap.
+ *
+ * these 'compile-time allocated' memory buffers are
+ * fixed-size 4k pages. (or larger if used with an increment
+ * higher than 1) use fixmap_set(idx,phys) to associate
+ * physical memory with fixmap indices.
+ *
+ * TLB entries of such buffers will not be flushed across
+ * task switches.
+ */
+enum fixed_addresses {
+#ifdef __tilegx__
+	/*
+	 * TILEPro has unmapped memory above so the hole isn't needed,
+	 * and in any case the hole pushes us over a single 16MB pmd.
+	 */
+	FIX_HOLE,
+#endif
+#ifdef CONFIG_HIGHMEM
+	FIX_KMAP_BEGIN,	/* reserved pte's for temporary kernel mappings */
+	FIX_KMAP_END = FIX_KMAP_BEGIN+(KM_TYPE_NR*NR_CPUS)-1,
+#endif
+#ifdef __tilegx__  /* see homecache.c */
+	FIX_HOMECACHE_BEGIN,
+	FIX_HOMECACHE_END = FIX_HOMECACHE_BEGIN+(NR_CPUS)-1,
+#endif
+	__end_of_permanent_fixed_addresses,
+
+	/*
+	 * Temporary boot-time mappings, used before ioremap() is functional.
+	 * Not currently needed by the Tile architecture.
+	 */
+#define NR_FIX_BTMAPS	0
+#if NR_FIX_BTMAPS
+	FIX_BTMAP_END = __end_of_permanent_fixed_addresses,
+	FIX_BTMAP_BEGIN = FIX_BTMAP_END + NR_FIX_BTMAPS - 1,
+	__end_of_fixed_addresses
+#else
+	__end_of_fixed_addresses = __end_of_permanent_fixed_addresses
+#endif
+};
+
+#define __FIXADDR_SIZE	(__end_of_permanent_fixed_addresses << PAGE_SHIFT)
+#define __FIXADDR_BOOT_SIZE	(__end_of_fixed_addresses << PAGE_SHIFT)
+#define FIXADDR_START		(FIXADDR_TOP + PAGE_SIZE - __FIXADDR_SIZE)
+#define FIXADDR_BOOT_START	(FIXADDR_TOP + PAGE_SIZE - __FIXADDR_BOOT_SIZE)
+
+#include <asm-generic/fixmap.h>
+
+#endif /* !__ASSEMBLY__ */
+
+#endif /* _ASM_TILE_FIXMAP_H */
diff --git a/kernel/arch/tile/include/asm/ftrace.h b/kernel/arch/tile/include/asm/ftrace.h
new file mode 100644
index 000000000..738d239b7
--- /dev/null
+++ b/kernel/arch/tile/include/asm/ftrace.h
@@ -0,0 +1,42 @@
+/*
+ * Copyright 2010 Tilera Corporation. All Rights Reserved.
+ *
+ *   This program is free software; you can redistribute it and/or
+ *   modify it under the terms of the GNU General Public License
+ *   as published by the Free Software Foundation, version 2.
+ *
+ *   This program is distributed in the hope that it will be useful, but
+ *   WITHOUT ANY WARRANTY; without even the implied warranty of
+ *   MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
+ *   NON INFRINGEMENT.  See the GNU General Public License for
+ *   more details.
+ */
+
+#ifndef _ASM_TILE_FTRACE_H
+#define _ASM_TILE_FTRACE_H
+
+#ifdef CONFIG_FUNCTION_TRACER
+
+#define MCOUNT_ADDR ((unsigned long)(__mcount))
+#define MCOUNT_INSN_SIZE 8		/* sizeof mcount call */
+
+#ifndef __ASSEMBLY__
+extern void __mcount(void);
+
+#define ARCH_SUPPORTS_FTRACE_OPS 1
+
+#ifdef CONFIG_DYNAMIC_FTRACE
+static inline unsigned long ftrace_call_adjust(unsigned long addr)
+{
+	return addr;
+}
+
+struct dyn_arch_ftrace {
+};
+#endif /*  CONFIG_DYNAMIC_FTRACE */
+
+#endif /* __ASSEMBLY__ */
+
+#endif /* CONFIG_FUNCTION_TRACER */
+
+#endif /* _ASM_TILE_FTRACE_H */
diff --git a/kernel/arch/tile/include/asm/futex.h b/kernel/arch/tile/include/asm/futex.h
new file mode 100644
index 000000000..1a6ef1b69
--- /dev/null
+++ b/kernel/arch/tile/include/asm/futex.h
@@ -0,0 +1,196 @@
+/*
+ * Copyright 2010 Tilera Corporation. All Rights Reserved.
+ *
+ *   This program is free software; you can redistribute it and/or
+ *   modify it under the terms of the GNU General Public License
+ *   as published by the Free Software Foundation, version 2.
+ *
+ *   This program is distributed in the hope that it will be useful, but
+ *   WITHOUT ANY WARRANTY; without even the implied warranty of
+ *   MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
+ *   NON INFRINGEMENT.  See the GNU General Public License for
+ *   more details.
+ *
+ * These routines make two important assumptions:
+ *
+ * 1. atomic_t is really an int and can be freely cast back and forth
+ *    (validated in __init_atomic_per_cpu).
+ *
+ * 2. userspace uses sys_cmpxchg() for all atomic operations, thus using
+ *    the same locking convention that all the kernel atomic routines use.
+ */
+
+#ifndef _ASM_TILE_FUTEX_H
+#define _ASM_TILE_FUTEX_H
+
+#ifndef __ASSEMBLY__
+
+#include <linux/futex.h>
+#include <linux/uaccess.h>
+#include <linux/errno.h>
+#include <asm/atomic.h>
+
+/*
+ * Support macros for futex operations.  Do not use these macros directly.
+ * They assume "ret", "val", "oparg", and "uaddr" in the lexical context.
+ * __futex_cmpxchg() additionally assumes "oldval".
+ */
+
+#ifdef __tilegx__
+
+#define __futex_asm(OP) \
+	asm("1: {" #OP " %1, %3, %4; movei %0, 0 }\n"		\
+	    ".pushsection .fixup,\"ax\"\n"			\
+	    "0: { movei %0, %5; j 9f }\n"			\
+	    ".section __ex_table,\"a\"\n"			\
+	    ".align 8\n"					\
+	    ".quad 1b, 0b\n"					\
+	    ".popsection\n"					\
+	    "9:"						\
+	    : "=r" (ret), "=r" (val), "+m" (*(uaddr))		\
+	    : "r" (uaddr), "r" (oparg), "i" (-EFAULT))
+
+#define __futex_set() __futex_asm(exch4)
+#define __futex_add() __futex_asm(fetchadd4)
+#define __futex_or() __futex_asm(fetchor4)
+#define __futex_andn() ({ oparg = ~oparg; __futex_asm(fetchand4); })
+#define __futex_cmpxchg() \
+	({ __insn_mtspr(SPR_CMPEXCH_VALUE, oldval); __futex_asm(cmpexch4); })
+
+#define __futex_xor()						\
+	({							\
+		u32 oldval, n = oparg;				\
+		if ((ret = __get_user(oldval, uaddr)) == 0) {	\
+			do {					\
+				oparg = oldval ^ n;		\
+				__futex_cmpxchg();		\
+			} while (ret == 0 && oldval != val);	\
+		}						\
+	})
+
+/* No need to prefetch, since the atomic ops go to the home cache anyway. */
+#define __futex_prolog()
+
+#else
+
+#define __futex_call(FN)						\
+	{								\
+		struct __get_user gu = FN((u32 __force *)uaddr, lock, oparg); \
+		val = gu.val;						\
+		ret = gu.err;						\
+	}
+
+#define __futex_set() __futex_call(__atomic_xchg)
+#define __futex_add() __futex_call(__atomic_xchg_add)
+#define __futex_or() __futex_call(__atomic_or)
+#define __futex_andn() __futex_call(__atomic_andn)
+#define __futex_xor() __futex_call(__atomic_xor)
+
+#define __futex_cmpxchg()						\
+	{								\
+		struct __get_user gu = __atomic_cmpxchg((u32 __force *)uaddr, \
+							lock, oldval, oparg); \
+		val = gu.val;						\
+		ret = gu.err;						\
+	}
+
+/*
+ * Find the lock pointer for the atomic calls to use, and issue a
+ * prefetch to the user address to bring it into cache.  Similar to
+ * __atomic_setup(), but we can't do a read into the L1 since it might
+ * fault; instead we do a prefetch into the L2.
+ */
+#define __futex_prolog()					\
+	int *lock;						\
+	__insn_prefetch(uaddr);					\
+	lock = __atomic_hashed_lock((int __force *)uaddr)
+#endif
+
+static inline int futex_atomic_op_inuser(int encoded_op, u32 __user *uaddr)
+{
+	int op = (encoded_op >> 28) & 7;
+	int cmp = (encoded_op >> 24) & 15;
+	int oparg = (encoded_op << 8) >> 20;
+	int cmparg = (encoded_op << 20) >> 20;
+	int uninitialized_var(val), ret;
+
+	__futex_prolog();
+
+	/* The 32-bit futex code makes this assumption, so validate it here. */
+	BUILD_BUG_ON(sizeof(atomic_t) != sizeof(int));
+
+	if (encoded_op & (FUTEX_OP_OPARG_SHIFT << 28))
+		oparg = 1 << oparg;
+
+	if (!access_ok(VERIFY_WRITE, uaddr, sizeof(u32)))
+		return -EFAULT;
+
+	pagefault_disable();
+	switch (op) {
+	case FUTEX_OP_SET:
+		__futex_set();
+		break;
+	case FUTEX_OP_ADD:
+		__futex_add();
+		break;
+	case FUTEX_OP_OR:
+		__futex_or();
+		break;
+	case FUTEX_OP_ANDN:
+		__futex_andn();
+		break;
+	case FUTEX_OP_XOR:
+		__futex_xor();
+		break;
+	default:
+		ret = -ENOSYS;
+		break;
+	}
+	pagefault_enable();
+
+	if (!ret) {
+		switch (cmp) {
+		case FUTEX_OP_CMP_EQ:
+			ret = (val == cmparg);
+			break;
+		case FUTEX_OP_CMP_NE:
+			ret = (val != cmparg);
+			break;
+		case FUTEX_OP_CMP_LT:
+			ret = (val < cmparg);
+			break;
+		case FUTEX_OP_CMP_GE:
+			ret = (val >= cmparg);
+			break;
+		case FUTEX_OP_CMP_LE:
+			ret = (val <= cmparg);
+			break;
+		case FUTEX_OP_CMP_GT:
+			ret = (val > cmparg);
+			break;
+		default:
+			ret = -ENOSYS;
+		}
+	}
+	return ret;
+}
+
+static inline int futex_atomic_cmpxchg_inatomic(u32 *uval, u32 __user *uaddr,
+						u32 oldval, u32 oparg)
+{
+	int ret, val;
+
+	__futex_prolog();
+
+	if (!access_ok(VERIFY_WRITE, uaddr, sizeof(u32)))
+		return -EFAULT;
+
+	__futex_cmpxchg();
+
+	*uval = val;
+	return ret;
+}
+
+#endif /* !__ASSEMBLY__ */
+
+#endif /* _ASM_TILE_FUTEX_H */
diff --git a/kernel/arch/tile/include/asm/hardirq.h b/kernel/arch/tile/include/asm/hardirq.h
new file mode 100644
index 000000000..54110af23
--- /dev/null
+++ b/kernel/arch/tile/include/asm/hardirq.h
@@ -0,0 +1,45 @@
+/*
+ * Copyright 2010 Tilera Corporation. All Rights Reserved.
+ *
+ *   This program is free software; you can redistribute it and/or
+ *   modify it under the terms of the GNU General Public License
+ *   as published by the Free Software Foundation, version 2.
+ *
+ *   This program is distributed in the hope that it will be useful, but
+ *   WITHOUT ANY WARRANTY; without even the implied warranty of
+ *   MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
+ *   NON INFRINGEMENT.  See the GNU General Public License for
+ *   more details.
+ */
+
+#ifndef _ASM_TILE_HARDIRQ_H
+#define _ASM_TILE_HARDIRQ_H
+
+#include <linux/threads.h>
+#include <linux/cache.h>
+
+#include <asm/irq.h>
+
+typedef struct {
+	unsigned int __softirq_pending;
+	long idle_timestamp;
+
+	/* Hard interrupt statistics. */
+	unsigned int irq_timer_count;
+	unsigned int irq_syscall_count;
+	unsigned int irq_resched_count;
+	unsigned int irq_hv_flush_count;
+	unsigned int irq_call_count;
+	unsigned int irq_hv_msg_count;
+	unsigned int irq_dev_intr_count;
+
+} ____cacheline_aligned irq_cpustat_t;
+
+DECLARE_PER_CPU(irq_cpustat_t, irq_stat);
+
+#define __ARCH_IRQ_STAT
+#define __IRQ_STAT(cpu, member) (per_cpu(irq_stat, cpu).member)
+
+#include <linux/irq_cpustat.h>	/* Standard mappings for irq_cpustat_t above */
+
+#endif /* _ASM_TILE_HARDIRQ_H */
diff --git a/kernel/arch/tile/include/asm/hardwall.h b/kernel/arch/tile/include/asm/hardwall.h
new file mode 100644
index 000000000..44d2765bd
--- /dev/null
+++ b/kernel/arch/tile/include/asm/hardwall.h
@@ -0,0 +1,30 @@
+/*
+ * Copyright 2010 Tilera Corporation. All Rights Reserved.
+ *
+ *   This program is free software; you can redistribute it and/or
+ *   modify it under the terms of the GNU General Public License
+ *   as published by the Free Software Foundation, version 2.
+ *
+ *   This program is distributed in the hope that it will be useful, but
+ *   WITHOUT ANY WARRANTY; without even the implied warranty of
+ *   MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
+ *   NON INFRINGEMENT.  See the GNU General Public License for
+ *   more details.
+ *
+ * Provide methods for access control of per-cpu resources like
+ * UDN, IDN, or IPI.
+ */
+#ifndef _ASM_TILE_HARDWALL_H
+#define _ASM_TILE_HARDWALL_H
+
+#include <uapi/asm/hardwall.h>
+
+/* /proc hooks for hardwall. */
+struct proc_dir_entry;
+#ifdef CONFIG_HARDWALL
+void proc_tile_hardwall_init(struct proc_dir_entry *root);
+int proc_pid_hardwall(struct seq_file *m, struct pid_namespace *ns, struct pid *pid, struct task_struct *task);
+#else
+static inline void proc_tile_hardwall_init(struct proc_dir_entry *root) {}
+#endif
+#endif /* _ASM_TILE_HARDWALL_H */
diff --git a/kernel/arch/tile/include/asm/highmem.h b/kernel/arch/tile/include/asm/highmem.h
new file mode 100644
index 000000000..fc8429a31
--- /dev/null
+++ b/kernel/arch/tile/include/asm/highmem.h
@@ -0,0 +1,72 @@
+/*
+ * Copyright (C) 1999 Gerhard Wichert, Siemens AG
+ *                   Gerhard.Wichert@pdb.siemens.de
+ * Copyright 2010 Tilera Corporation. All Rights Reserved.
+ *
+ *   This program is free software; you can redistribute it and/or
+ *   modify it under the terms of the GNU General Public License
+ *   as published by the Free Software Foundation, version 2.
+ *
+ *   This program is distributed in the hope that it will be useful, but
+ *   WITHOUT ANY WARRANTY; without even the implied warranty of
+ *   MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
+ *   NON INFRINGEMENT.  See the GNU General Public License for
+ *   more details.
+ *
+ * Used in CONFIG_HIGHMEM systems for memory pages which
+ * are not addressable by direct kernel virtual addresses.
+ *
+ */
+
+#ifndef _ASM_TILE_HIGHMEM_H
+#define _ASM_TILE_HIGHMEM_H
+
+#include <linux/interrupt.h>
+#include <linux/threads.h>
+#include <asm/tlbflush.h>
+#include <asm/homecache.h>
+
+/* declarations for highmem.c */
+extern unsigned long highstart_pfn, highend_pfn;
+
+extern pte_t *pkmap_page_table;
+
+/*
+ * Ordering is:
+ *
+ * FIXADDR_TOP
+ *			fixed_addresses
+ * FIXADDR_START
+ *			temp fixed addresses
+ * FIXADDR_BOOT_START
+ *			Persistent kmap area
+ * PKMAP_BASE
+ * VMALLOC_END
+ *			Vmalloc area
+ * VMALLOC_START
+ * high_memory
+ */
+#define LAST_PKMAP_MASK (LAST_PKMAP-1)
+#define PKMAP_NR(virt)  ((virt-PKMAP_BASE) >> PAGE_SHIFT)
+#define PKMAP_ADDR(nr)  (PKMAP_BASE + ((nr) << PAGE_SHIFT))
+
+void *kmap_high(struct page *page);
+void kunmap_high(struct page *page);
+void *kmap(struct page *page);
+void kunmap(struct page *page);
+void *kmap_fix_kpte(struct page *page, int finished);
+
+/* This macro is used only in map_new_virtual() to map "page". */
+#define kmap_prot page_to_kpgprot(page)
+
+void *kmap_atomic(struct page *page);
+void __kunmap_atomic(void *kvaddr);
+void *kmap_atomic_pfn(unsigned long pfn);
+void *kmap_atomic_prot_pfn(unsigned long pfn, pgprot_t prot);
+struct page *kmap_atomic_to_page(void *ptr);
+void *kmap_atomic_prot(struct page *page, pgprot_t prot);
+void kmap_atomic_fix_kpte(struct page *page, int finished);
+
+#define flush_cache_kmaps()	do { } while (0)
+
+#endif /* _ASM_TILE_HIGHMEM_H */
diff --git a/kernel/arch/tile/include/asm/homecache.h b/kernel/arch/tile/include/asm/homecache.h
new file mode 100644
index 000000000..7ddd1b8d6
--- /dev/null
+++ b/kernel/arch/tile/include/asm/homecache.h
@@ -0,0 +1,123 @@
+/*
+ * Copyright 2010 Tilera Corporation. All Rights Reserved.
+ *
+ *   This program is free software; you can redistribute it and/or
+ *   modify it under the terms of the GNU General Public License
+ *   as published by the Free Software Foundation, version 2.
+ *
+ *   This program is distributed in the hope that it will be useful, but
+ *   WITHOUT ANY WARRANTY; without even the implied warranty of
+ *   MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
+ *   NON INFRINGEMENT.  See the GNU General Public License for
+ *   more details.
+ *
+ * Handle issues around the Tile "home cache" model of coherence.
+ */
+
+#ifndef _ASM_TILE_HOMECACHE_H
+#define _ASM_TILE_HOMECACHE_H
+
+#include <asm/page.h>
+#include <linux/cpumask.h>
+
+struct page;
+struct task_struct;
+struct vm_area_struct;
+struct zone;
+
+/*
+ * Coherence point for the page is its memory controller.
+ * It is not present in any cache (L1 or L2).
+ */
+#define PAGE_HOME_UNCACHED -1
+
+/*
+ * Is this page immutable (unwritable) and thus able to be cached more
+ * widely than would otherwise be possible?  This means we have "nc" set.
+ */
+#define PAGE_HOME_IMMUTABLE -2
+
+/*
+ * Each cpu considers its own cache to be the home for the page,
+ * which makes it incoherent.
+ */
+#define PAGE_HOME_INCOHERENT -3
+
+/* Home for the page is distributed via hash-for-home. */
+#define PAGE_HOME_HASH -4
+
+/* Support wrapper to use instead of explicit hv_flush_remote(). */
+extern void flush_remote(unsigned long cache_pfn, unsigned long cache_length,
+			 const struct cpumask *cache_cpumask,
+			 HV_VirtAddr tlb_va, unsigned long tlb_length,
+			 unsigned long tlb_pgsize,
+			 const struct cpumask *tlb_cpumask,
+			 HV_Remote_ASID *asids, int asidcount);
+
+/* Set homing-related bits in a PTE (can also pass a pgprot_t). */
+extern pte_t pte_set_home(pte_t pte, int home);
+
+/* Do a cache eviction on the specified cpus. */
+extern void homecache_evict(const struct cpumask *mask);
+
+/*
+ * Change a kernel page's homecache.  It must not be mapped in user space.
+ * If !CONFIG_HOMECACHE, only usable on LOWMEM, and can only be called when
+ * no other cpu can reference the page, and causes a full-chip cache/TLB flush.
+ */
+extern void homecache_change_page_home(struct page *, int order, int home);
+
+/*
+ * Flush a page out of whatever cache(s) it is in.
+ * This is more than just finv, since it properly handles waiting
+ * for the data to reach memory, but it can be quite
+ * heavyweight, particularly on incoherent or immutable memory.
+ */
+extern void homecache_finv_page(struct page *);
+
+/*
+ * Flush a page out of the specified home cache.
+ * Note that the specified home need not be the actual home of the page,
+ * as for example might be the case when coordinating with I/O devices.
+ */
+extern void homecache_finv_map_page(struct page *, int home);
+
+/*
+ * Allocate a page with the given GFP flags, home, and optionally
+ * node.  These routines are actually just wrappers around the normal
+ * alloc_pages() / alloc_pages_node() functions, which set and clear
+ * a per-cpu variable to communicate with homecache_new_kernel_page().
+ * If !CONFIG_HOMECACHE, uses homecache_change_page_home().
+ */
+extern struct page *homecache_alloc_pages(gfp_t gfp_mask,
+					  unsigned int order, int home);
+extern struct page *homecache_alloc_pages_node(int nid, gfp_t gfp_mask,
+					       unsigned int order, int home);
+#define homecache_alloc_page(gfp_mask, home) \
+  homecache_alloc_pages(gfp_mask, 0, home)
+
+/*
+ * These routines are just pass-throughs to free_pages() when
+ * we support full homecaching.  If !CONFIG_HOMECACHE, then these
+ * routines use homecache_change_page_home() to reset the home
+ * back to the default before returning the page to the allocator.
+ */
+void __homecache_free_pages(struct page *, unsigned int order);
+void homecache_free_pages(unsigned long addr, unsigned int order);
+#define __homecache_free_page(page) __homecache_free_pages((page), 0)
+#define homecache_free_page(page) homecache_free_pages((page), 0)
+
+
+/*
+ * Report the page home for LOWMEM pages by examining their kernel PTE,
+ * or for highmem pages as the default home.
+ */
+extern int page_home(struct page *);
+
+#define homecache_migrate_kthread() do {} while (0)
+
+#define homecache_kpte_lock() 0
+#define homecache_kpte_unlock(flags) do {} while (0)
+
+
+#endif /* _ASM_TILE_HOMECACHE_H */
diff --git a/kernel/arch/tile/include/asm/hugetlb.h b/kernel/arch/tile/include/asm/hugetlb.h
new file mode 100644
index 000000000..325773300
--- /dev/null
+++ b/kernel/arch/tile/include/asm/hugetlb.h
@@ -0,0 +1,135 @@
+/*
+ * Copyright 2010 Tilera Corporation. All Rights Reserved.
+ *
+ *   This program is free software; you can redistribute it and/or
+ *   modify it under the terms of the GNU General Public License
+ *   as published by the Free Software Foundation, version 2.
+ *
+ *   This program is distributed in the hope that it will be useful, but
+ *   WITHOUT ANY WARRANTY; without even the implied warranty of
+ *   MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
+ *   NON INFRINGEMENT.  See the GNU General Public License for
+ *   more details.
+ */
+
+#ifndef _ASM_TILE_HUGETLB_H
+#define _ASM_TILE_HUGETLB_H
+
+#include <asm/page.h>
+#include <asm-generic/hugetlb.h>
+
+
+static inline int is_hugepage_only_range(struct mm_struct *mm,
+					 unsigned long addr,
+					 unsigned long len) {
+	return 0;
+}
+
+/*
+ * If the arch doesn't supply something else, assume that hugepage
+ * size aligned regions are ok without further preparation.
+ */
+static inline int prepare_hugepage_range(struct file *file,
+					 unsigned long addr, unsigned long len)
+{
+	struct hstate *h = hstate_file(file);
+	if (len & ~huge_page_mask(h))
+		return -EINVAL;
+	if (addr & ~huge_page_mask(h))
+		return -EINVAL;
+	return 0;
+}
+
+static inline void hugetlb_prefault_arch_hook(struct mm_struct *mm)
+{
+}
+
+static inline void hugetlb_free_pgd_range(struct mmu_gather *tlb,
+					  unsigned long addr, unsigned long end,
+					  unsigned long floor,
+					  unsigned long ceiling)
+{
+	free_pgd_range(tlb, addr, end, floor, ceiling);
+}
+
+static inline void set_huge_pte_at(struct mm_struct *mm, unsigned long addr,
+				   pte_t *ptep, pte_t pte)
+{
+	set_pte(ptep, pte);
+}
+
+static inline pte_t huge_ptep_get_and_clear(struct mm_struct *mm,
+					    unsigned long addr, pte_t *ptep)
+{
+	return ptep_get_and_clear(mm, addr, ptep);
+}
+
+static inline void huge_ptep_clear_flush(struct vm_area_struct *vma,
+					 unsigned long addr, pte_t *ptep)
+{
+	ptep_clear_flush(vma, addr, ptep);
+}
+
+static inline int huge_pte_none(pte_t pte)
+{
+	return pte_none(pte);
+}
+
+static inline pte_t huge_pte_wrprotect(pte_t pte)
+{
+	return pte_wrprotect(pte);
+}
+
+static inline void huge_ptep_set_wrprotect(struct mm_struct *mm,
+					   unsigned long addr, pte_t *ptep)
+{
+	ptep_set_wrprotect(mm, addr, ptep);
+}
+
+static inline int huge_ptep_set_access_flags(struct vm_area_struct *vma,
+					     unsigned long addr, pte_t *ptep,
+					     pte_t pte, int dirty)
+{
+	return ptep_set_access_flags(vma, addr, ptep, pte, dirty);
+}
+
+static inline pte_t huge_ptep_get(pte_t *ptep)
+{
+	return *ptep;
+}
+
+static inline int arch_prepare_hugepage(struct page *page)
+{
+	return 0;
+}
+
+static inline void arch_release_hugepage(struct page *page)
+{
+}
+
+static inline void arch_clear_hugepage_flags(struct page *page)
+{
+}
+
+#ifdef CONFIG_HUGETLB_SUPER_PAGES
+static inline pte_t arch_make_huge_pte(pte_t entry, struct vm_area_struct *vma,
+				       struct page *page, int writable)
+{
+	size_t pagesize = huge_page_size(hstate_vma(vma));
+	if (pagesize != PUD_SIZE && pagesize != PMD_SIZE)
+		entry = pte_mksuper(entry);
+	return entry;
+}
+#define arch_make_huge_pte arch_make_huge_pte
+
+/* Sizes to scale up page size for PTEs with HV_PTE_SUPER bit. */
+enum {
+	HUGE_SHIFT_PGDIR = 0,
+	HUGE_SHIFT_PMD = 1,
+	HUGE_SHIFT_PAGE = 2,
+	HUGE_SHIFT_ENTRIES
+};
+extern int huge_shift[HUGE_SHIFT_ENTRIES];
+#endif
+
+#endif /* _ASM_TILE_HUGETLB_H */
diff --git a/kernel/arch/tile/include/asm/hv_driver.h b/kernel/arch/tile/include/asm/hv_driver.h
new file mode 100644
index 000000000..ad614de89
--- /dev/null
+++ b/kernel/arch/tile/include/asm/hv_driver.h
@@ -0,0 +1,60 @@
+/*
+ * Copyright 2010 Tilera Corporation. All Rights Reserved.
+ *
+ *   This program is free software; you can redistribute it and/or
+ *   modify it under the terms of the GNU General Public License
+ *   as published by the Free Software Foundation, version 2.
+ *
+ *   This program is distributed in the hope that it will be useful, but
+ *   WITHOUT ANY WARRANTY; without even the implied warranty of
+ *   MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
+ *   NON INFRINGEMENT.  See the GNU General Public License for
+ *   more details.
+ *
+ * This header defines a wrapper interface for managing hypervisor
+ * device calls that will result in an interrupt at some later time.
+ * In particular, this provides wrappers for hv_preada() and
+ * hv_pwritea().
+ */
+
+#ifndef _ASM_TILE_HV_DRIVER_H
+#define _ASM_TILE_HV_DRIVER_H
+
+#include <hv/hypervisor.h>
+
+struct hv_driver_cb;
+
+/* A callback to be invoked when an operation completes. */
+typedef void hv_driver_callback_t(struct hv_driver_cb *cb, __hv32 result);
+
+/*
+ * A structure to hold information about an outstanding call.
+ * The driver must allocate a separate structure for each call.
+ */
+struct hv_driver_cb {
+	hv_driver_callback_t *callback;  /* Function to call on interrupt. */
+	void *dev;                       /* Driver-specific state variable. */
+};
+
+/* Wrapper for invoking hv_dev_preada(). */
+static inline int
+tile_hv_dev_preada(int devhdl, __hv32 flags, __hv32 sgl_len,
+		   HV_SGL sgl[/* sgl_len */], __hv64 offset,
+		   struct hv_driver_cb *callback)
+{
+	return hv_dev_preada(devhdl, flags, sgl_len, sgl,
+			     offset, (HV_IntArg)callback);
+}
+
+/* Wrapper for invoking hv_dev_pwritea(). */
+static inline int
+tile_hv_dev_pwritea(int devhdl, __hv32 flags, __hv32 sgl_len,
+		    HV_SGL sgl[/* sgl_len */], __hv64 offset,
+		    struct hv_driver_cb *callback)
+{
+	return hv_dev_pwritea(devhdl, flags, sgl_len, sgl,
+			      offset, (HV_IntArg)callback);
+}
+
+
+#endif /* _ASM_TILE_HV_DRIVER_H */
diff --git a/kernel/arch/tile/include/asm/ide.h b/kernel/arch/tile/include/asm/ide.h
new file mode 100644
index 000000000..3c6f2ed89
--- /dev/null
+++ b/kernel/arch/tile/include/asm/ide.h
@@ -0,0 +1,25 @@
+/*
+ * Copyright 2010 Tilera Corporation. All Rights Reserved.
+ *
+ *   This program is free software; you can redistribute it and/or
+ *   modify it under the terms of the GNU General Public License
+ *   as published by the Free Software Foundation, version 2.
+ *
+ *   This program is distributed in the hope that it will be useful, but
+ *   WITHOUT ANY WARRANTY; without even the implied warranty of
+ *   MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
+ *   NON INFRINGEMENT.  See the GNU General Public License for
+ *   more details.
+ */
+
+#ifndef _ASM_TILE_IDE_H
+#define _ASM_TILE_IDE_H
+
+/* For IDE on PCI */
+#define MAX_HWIFS       10
+
+#define ide_default_io_ctl(base)	(0)
+
+#include <asm-generic/ide_iops.h>
+
+#endif /* _ASM_TILE_IDE_H */
diff --git a/kernel/arch/tile/include/asm/io.h b/kernel/arch/tile/include/asm/io.h
new file mode 100644
index 000000000..6ef4ecab1
--- /dev/null
+++ b/kernel/arch/tile/include/asm/io.h
@@ -0,0 +1,508 @@
+/*
+ * Copyright 2010 Tilera Corporation. All Rights Reserved.
+ *
+ *   This program is free software; you can redistribute it and/or
+ *   modify it under the terms of the GNU General Public License
+ *   as published by the Free Software Foundation, version 2.
+ *
+ *   This program is distributed in the hope that it will be useful, but
+ *   WITHOUT ANY WARRANTY; without even the implied warranty of
+ *   MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
+ *   NON INFRINGEMENT.  See the GNU General Public License for
+ *   more details.
+ */
+
+#ifndef _ASM_TILE_IO_H
+#define _ASM_TILE_IO_H
+
+#include <linux/kernel.h>
+#include <linux/bug.h>
+#include <asm/page.h>
+
+/* Maximum PCI I/O space address supported. */
+#define IO_SPACE_LIMIT 0xffffffff
+
+/*
+ * Convert a physical pointer to a virtual kernel pointer for /dev/mem
+ * access.
+ */
+#define xlate_dev_mem_ptr(p)	__va(p)
+
+/*
+ * Convert a virtual cached pointer to an uncached pointer.
+ */
+#define xlate_dev_kmem_ptr(p)	p
+
+/*
+ * Change "struct page" to physical address.
+ */
+#define page_to_phys(page)    ((dma_addr_t)page_to_pfn(page) << PAGE_SHIFT)
+
+/*
+ * Some places try to pass in an loff_t for PHYSADDR (?!), so we cast it to
+ * long before casting it to a pointer to avoid compiler warnings.
+ */
+#if CHIP_HAS_MMIO()
+extern void __iomem *ioremap(resource_size_t offset, unsigned long size);
+extern void __iomem *ioremap_prot(resource_size_t offset, unsigned long size,
+	pgprot_t pgprot);
+extern void iounmap(volatile void __iomem *addr);
+#else
+#define ioremap(physaddr, size)	((void __iomem *)(unsigned long)(physaddr))
+#define iounmap(addr)		((void)0)
+#endif
+
+#define ioremap_nocache(physaddr, size)		ioremap(physaddr, size)
+#define ioremap_wc(physaddr, size)		ioremap(physaddr, size)
+#define ioremap_writethrough(physaddr, size)	ioremap(physaddr, size)
+#define ioremap_fullcache(physaddr, size)	ioremap(physaddr, size)
+
+#define mmiowb()
+
+/* Conversion between virtual and physical mappings.  */
+#define mm_ptov(addr)		((void *)phys_to_virt(addr))
+#define mm_vtop(addr)		((unsigned long)virt_to_phys(addr))
+
+#if CHIP_HAS_MMIO()
+
+/*
+ * We use inline assembly to guarantee that the compiler does not
+ * split an access into multiple byte-sized accesses as it might
+ * sometimes do if a register data structure is marked "packed".
+ * Obviously on tile we can't tolerate such an access being
+ * actually unaligned, but we want to avoid the case where the
+ * compiler conservatively would generate multiple accesses even
+ * for an aligned read or write.
+ */
+
+static inline u8 __raw_readb(const volatile void __iomem *addr)
+{
+	return *(const volatile u8 __force *)addr;
+}
+
+static inline u16 __raw_readw(const volatile void __iomem *addr)
+{
+	u16 ret;
+	asm volatile("ld2u %0, %1" : "=r" (ret) : "r" (addr));
+	barrier();
+	return le16_to_cpu(ret);
+}
+
+static inline u32 __raw_readl(const volatile void __iomem *addr)
+{
+	u32 ret;
+	/* Sign-extend to conform to u32 ABI sign-extension convention. */
+	asm volatile("ld4s %0, %1" : "=r" (ret) : "r" (addr));
+	barrier();
+	return le32_to_cpu(ret);
+}
+
+static inline u64 __raw_readq(const volatile void __iomem *addr)
+{
+	u64 ret;
+	asm volatile("ld %0, %1" : "=r" (ret) : "r" (addr));
+	barrier();
+	return le64_to_cpu(ret);
+}
+
+static inline void __raw_writeb(u8 val, volatile void __iomem *addr)
+{
+	*(volatile u8 __force *)addr = val;
+}
+
+static inline void __raw_writew(u16 val, volatile void __iomem *addr)
+{
+	asm volatile("st2 %0, %1" :: "r" (addr), "r" (cpu_to_le16(val)));
+}
+
+static inline void __raw_writel(u32 val, volatile void __iomem *addr)
+{
+	asm volatile("st4 %0, %1" :: "r" (addr), "r" (cpu_to_le32(val)));
+}
+
+static inline void __raw_writeq(u64 val, volatile void __iomem *addr)
+{
+	asm volatile("st %0, %1" :: "r" (addr), "r" (cpu_to_le64(val)));
+}
+
+/*
+ * The on-chip I/O hardware on tilegx is configured with VA=PA for the
+ * kernel's PA range.  The low-level APIs and field names use "va" and
+ * "void *" nomenclature, to be consistent with the general notion
+ * that the addresses in question are virtualizable, but in the kernel
+ * context we are actually manipulating PA values.  (In other contexts,
+ * e.g. access from user space, we do in fact use real virtual addresses
+ * in the va fields.)  To allow readers of the code to understand what's
+ * happening, we direct their attention to this comment by using the
+ * following two functions that just duplicate __va() and __pa().
+ */
+typedef unsigned long tile_io_addr_t;
+static inline tile_io_addr_t va_to_tile_io_addr(void *va)
+{
+	BUILD_BUG_ON(sizeof(phys_addr_t) != sizeof(tile_io_addr_t));
+	return __pa(va);
+}
+static inline void *tile_io_addr_to_va(tile_io_addr_t tile_io_addr)
+{
+	return __va(tile_io_addr);
+}
+
+#else /* CHIP_HAS_MMIO() */
+
+#ifdef CONFIG_PCI
+
+extern u8 _tile_readb(unsigned long addr);
+extern u16 _tile_readw(unsigned long addr);
+extern u32 _tile_readl(unsigned long addr);
+extern u64 _tile_readq(unsigned long addr);
+extern void _tile_writeb(u8  val, unsigned long addr);
+extern void _tile_writew(u16 val, unsigned long addr);
+extern void _tile_writel(u32 val, unsigned long addr);
+extern void _tile_writeq(u64 val, unsigned long addr);
+
+#define __raw_readb(addr) _tile_readb((unsigned long)addr)
+#define __raw_readw(addr) _tile_readw((unsigned long)addr)
+#define __raw_readl(addr) _tile_readl((unsigned long)addr)
+#define __raw_readq(addr) _tile_readq((unsigned long)addr)
+#define __raw_writeb(val, addr) _tile_writeb(val, (unsigned long)addr)
+#define __raw_writew(val, addr) _tile_writew(val, (unsigned long)addr)
+#define __raw_writel(val, addr) _tile_writel(val, (unsigned long)addr)
+#define __raw_writeq(val, addr) _tile_writeq(val, (unsigned long)addr)
+
+#else /* CONFIG_PCI */
+
+/*
+ * The tilepro architecture does not support IOMEM unless PCI is enabled.
+ * Unfortunately we can't yet simply not declare these methods,
+ * since some generic code that compiles into the kernel, but
+ * we never run, uses them unconditionally.
+ */
+
+static inline int iomem_panic(void)
+{
+	panic("readb/writeb and friends do not exist on tile without PCI");
+	return 0;
+}
+
+static inline u8 readb(unsigned long addr)
+{
+	return iomem_panic();
+}
+
+static inline u16 _readw(unsigned long addr)
+{
+	return iomem_panic();
+}
+
+static inline u32 readl(unsigned long addr)
+{
+	return iomem_panic();
+}
+
+static inline u64 readq(unsigned long addr)
+{
+	return iomem_panic();
+}
+
+static inline void writeb(u8  val, unsigned long addr)
+{
+	iomem_panic();
+}
+
+static inline void writew(u16 val, unsigned long addr)
+{
+	iomem_panic();
+}
+
+static inline void writel(u32 val, unsigned long addr)
+{
+	iomem_panic();
+}
+
+static inline void writeq(u64 val, unsigned long addr)
+{
+	iomem_panic();
+}
+
+#endif /* CONFIG_PCI */
+
+#endif /* CHIP_HAS_MMIO() */
+
+#define readb __raw_readb
+#define readw __raw_readw
+#define readl __raw_readl
+#define readq __raw_readq
+#define writeb __raw_writeb
+#define writew __raw_writew
+#define writel __raw_writel
+#define writeq __raw_writeq
+
+#define readb_relaxed readb
+#define readw_relaxed readw
+#define readl_relaxed readl
+#define readq_relaxed readq
+#define writeb_relaxed writeb
+#define writew_relaxed writew
+#define writel_relaxed writel
+#define writeq_relaxed writeq
+
+#define ioread8 readb
+#define ioread16 readw
+#define ioread32 readl
+#define ioread64 readq
+#define iowrite8 writeb
+#define iowrite16 writew
+#define iowrite32 writel
+#define iowrite64 writeq
+
+#if CHIP_HAS_MMIO() || defined(CONFIG_PCI)
+
+static inline void memset_io(volatile void *dst, int val, size_t len)
+{
+	size_t x;
+	BUG_ON((unsigned long)dst & 0x3);
+	val = (val & 0xff) * 0x01010101;
+	for (x = 0; x < len; x += 4)
+		writel(val, dst + x);
+}
+
+static inline void memcpy_fromio(void *dst, const volatile void __iomem *src,
+				 size_t len)
+{
+	size_t x;
+	BUG_ON((unsigned long)src & 0x3);
+	for (x = 0; x < len; x += 4)
+		*(u32 *)(dst + x) = readl(src + x);
+}
+
+static inline void memcpy_toio(volatile void __iomem *dst, const void *src,
+				size_t len)
+{
+	size_t x;
+	BUG_ON((unsigned long)dst & 0x3);
+	for (x = 0; x < len; x += 4)
+		writel(*(u32 *)(src + x), dst + x);
+}
+
+#endif
+
+#if CHIP_HAS_MMIO() && defined(CONFIG_TILE_PCI_IO)
+
+static inline u8 inb(unsigned long addr)
+{
+	return readb((volatile void __iomem *) addr);
+}
+
+static inline u16 inw(unsigned long addr)
+{
+	return readw((volatile void __iomem *) addr);
+}
+
+static inline u32 inl(unsigned long addr)
+{
+	return readl((volatile void __iomem *) addr);
+}
+
+static inline void outb(u8 b, unsigned long addr)
+{
+	writeb(b, (volatile void __iomem *) addr);
+}
+
+static inline void outw(u16 b, unsigned long addr)
+{
+	writew(b, (volatile void __iomem *) addr);
+}
+
+static inline void outl(u32 b, unsigned long addr)
+{
+	writel(b, (volatile void __iomem *) addr);
+}
+
+static inline void insb(unsigned long addr, void *buffer, int count)
+{
+	if (count) {
+		u8 *buf = buffer;
+		do {
+			u8 x = inb(addr);
+			*buf++ = x;
+		} while (--count);
+	}
+}
+
+static inline void insw(unsigned long addr, void *buffer, int count)
+{
+	if (count) {
+		u16 *buf = buffer;
+		do {
+			u16 x = inw(addr);
+			*buf++ = x;
+		} while (--count);
+	}
+}
+
+static inline void insl(unsigned long addr, void *buffer, int count)
+{
+	if (count) {
+		u32 *buf = buffer;
+		do {
+			u32 x = inl(addr);
+			*buf++ = x;
+		} while (--count);
+	}
+}
+
+static inline void outsb(unsigned long addr, const void *buffer, int count)
+{
+	if (count) {
+		const u8 *buf = buffer;
+		do {
+			outb(*buf++, addr);
+		} while (--count);
+	}
+}
+
+static inline void outsw(unsigned long addr, const void *buffer, int count)
+{
+	if (count) {
+		const u16 *buf = buffer;
+		do {
+			outw(*buf++, addr);
+		} while (--count);
+	}
+}
+
+static inline void outsl(unsigned long addr, const void *buffer, int count)
+{
+	if (count) {
+		const u32 *buf = buffer;
+		do {
+			outl(*buf++, addr);
+		} while (--count);
+	}
+}
+
+extern void __iomem *ioport_map(unsigned long port, unsigned int len);
+extern void ioport_unmap(void __iomem *addr);
+
+#else
+
+/*
+ * The TilePro architecture does not support IOPORT, even with PCI.
+ * Unfortunately we can't yet simply not declare these methods,
+ * since some generic code that compiles into the kernel, but
+ * we never run, uses them unconditionally.
+ */
+
+static inline long ioport_panic(void)
+{
+#ifdef __tilegx__
+	panic("PCI IO space support is disabled. Configure the kernel with CONFIG_TILE_PCI_IO to enable it");
+#else
+	panic("inb/outb and friends do not exist on tile");
+#endif
+	return 0;
+}
+
+static inline void __iomem *ioport_map(unsigned long port, unsigned int len)
+{
+	pr_info("ioport_map: mapping IO resources is unsupported on tile\n");
+	return NULL;
+}
+
+static inline void ioport_unmap(void __iomem *addr)
+{
+	ioport_panic();
+}
+
+static inline u8 inb(unsigned long addr)
+{
+	return ioport_panic();
+}
+
+static inline u16 inw(unsigned long addr)
+{
+	return ioport_panic();
+}
+
+static inline u32 inl(unsigned long addr)
+{
+	return ioport_panic();
+}
+
+static inline void outb(u8 b, unsigned long addr)
+{
+	ioport_panic();
+}
+
+static inline void outw(u16 b, unsigned long addr)
+{
+	ioport_panic();
+}
+
+static inline void outl(u32 b, unsigned long addr)
+{
+	ioport_panic();
+}
+
+static inline void insb(unsigned long addr, void *buffer, int count)
+{
+	ioport_panic();
+}
+
+static inline void insw(unsigned long addr, void *buffer, int count)
+{
+	ioport_panic();
+}
+
+static inline void insl(unsigned long addr, void *buffer, int count)
+{
+	ioport_panic();
+}
+
+static inline void outsb(unsigned long addr, const void *buffer, int count)
+{
+	ioport_panic();
+}
+
+static inline void outsw(unsigned long addr, const void *buffer, int count)
+{
+	ioport_panic();
+}
+
+static inline void outsl(unsigned long addr, const void *buffer, int count)
+{
+	ioport_panic();
+}
+
+#endif /* CHIP_HAS_MMIO() && defined(CONFIG_TILE_PCI_IO) */
+
+#define inb_p(addr)	inb(addr)
+#define inw_p(addr)	inw(addr)
+#define inl_p(addr)	inl(addr)
+#define outb_p(x, addr)	outb((x), (addr))
+#define outw_p(x, addr)	outw((x), (addr))
+#define outl_p(x, addr)	outl((x), (addr))
+
+#define ioread16be(addr)	be16_to_cpu(ioread16(addr))
+#define ioread32be(addr)	be32_to_cpu(ioread32(addr))
+#define iowrite16be(v, addr)	iowrite16(be16_to_cpu(v), (addr))
+#define iowrite32be(v, addr)	iowrite32(be32_to_cpu(v), (addr))
+
+#define ioread8_rep(p, dst, count) \
+	insb((unsigned long) (p), (dst), (count))
+#define ioread16_rep(p, dst, count) \
+	insw((unsigned long) (p), (dst), (count))
+#define ioread32_rep(p, dst, count) \
+	insl((unsigned long) (p), (dst), (count))
+
+#define iowrite8_rep(p, src, count) \
+	outsb((unsigned long) (p), (src), (count))
+#define iowrite16_rep(p, src, count) \
+	outsw((unsigned long) (p), (src), (count))
+#define iowrite32_rep(p, src, count) \
+	outsl((unsigned long) (p), (src), (count))
+
+#define virt_to_bus     virt_to_phys
+#define bus_to_virt     phys_to_virt
+
+#endif /* _ASM_TILE_IO_H */
diff --git a/kernel/arch/tile/include/asm/irq.h b/kernel/arch/tile/include/asm/irq.h
new file mode 100644
index 000000000..1fe869118
--- /dev/null
+++ b/kernel/arch/tile/include/asm/irq.h
@@ -0,0 +1,81 @@
+/*
+ * Copyright 2010 Tilera Corporation. All Rights Reserved.
+ *
+ *   This program is free software; you can redistribute it and/or
+ *   modify it under the terms of the GNU General Public License
+ *   as published by the Free Software Foundation, version 2.
+ *
+ *   This program is distributed in the hope that it will be useful, but
+ *   WITHOUT ANY WARRANTY; without even the implied warranty of
+ *   MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
+ *   NON INFRINGEMENT.  See the GNU General Public License for
+ *   more details.
+ */
+
+#ifndef _ASM_TILE_IRQ_H
+#define _ASM_TILE_IRQ_H
+
+#include <linux/hardirq.h>
+
+/* The hypervisor interface provides 32 IRQs. */
+#define NR_IRQS			32
+
+/* IRQ numbers used for linux IPIs. */
+#define IRQ_RESCHEDULE	0
+/* Interrupts for dynamic allocation start at 1. Let the core allocate irq0 */
+#define NR_IRQS_LEGACY	1
+
+#define irq_canonicalize(irq)   (irq)
+
+void ack_bad_irq(unsigned int irq);
+
+/*
+ * Different ways of handling interrupts.  Tile interrupts are always
+ * per-cpu; there is no global interrupt controller to implement
+ * enable/disable.  Most onboard devices can send their interrupts to
+ * many tiles at the same time, and Tile-specific drivers know how to
+ * deal with this.
+ *
+ * However, generic devices (usually PCIE based, sometimes GPIO)
+ * expect that interrupts will fire on a single core at a time and
+ * that the irq can be enabled or disabled from any core at any time.
+ * We implement this by directing such interrupts to a single core.
+ *
+ * One added wrinkle is that PCI interrupts can be either
+ * hardware-cleared (legacy interrupts) or software cleared (MSI).
+ * Other generic device systems (GPIO) are always software-cleared.
+ *
+ * The enums below are used by drivers for onboard devices, including
+ * the internals of PCI root complex and GPIO.  They allow the driver
+ * to tell the generic irq code what kind of interrupt is mapped to a
+ * particular IRQ number.
+ */
+enum {
+	/* per-cpu interrupt; use enable/disable_percpu_irq() to mask */
+	TILE_IRQ_PERCPU,
+	/* global interrupt, hardware responsible for clearing. */
+	TILE_IRQ_HW_CLEAR,
+	/* global interrupt, software responsible for clearing. */
+	TILE_IRQ_SW_CLEAR,
+};
+
+
+/*
+ * Paravirtualized drivers should call this when they dynamically
+ * allocate a new IRQ or discover an IRQ that was pre-allocated by the
+ * hypervisor for use with their particular device.  This gives the
+ * IRQ subsystem an opportunity to do interrupt-type-specific
+ * initialization.
+ *
+ * ISSUE: We should modify this API so that registering anything
+ * except percpu interrupts also requires providing callback methods
+ * for enabling and disabling the interrupt.  This would allow the
+ * generic IRQ code to proxy enable/disable_irq() calls back into the
+ * PCI subsystem, which in turn could enable or disable the interrupt
+ * at the PCI shim.
+ */
+void tile_irq_activate(unsigned int irq, int tile_irq_type);
+
+void setup_irq_regs(void);
+
+#endif /* _ASM_TILE_IRQ_H */
diff --git a/kernel/arch/tile/include/asm/irq_work.h b/kernel/arch/tile/include/asm/irq_work.h
new file mode 100644
index 000000000..48af33a61
--- /dev/null
+++ b/kernel/arch/tile/include/asm/irq_work.h
@@ -0,0 +1,14 @@
+#ifndef __ASM_IRQ_WORK_H
+#define __ASM_IRQ_WORK_H
+
+static inline bool arch_irq_work_has_interrupt(void)
+{
+#ifdef CONFIG_SMP
+	extern bool self_interrupt_ok;
+	return self_interrupt_ok;
+#else
+	return false;
+#endif
+}
+
+#endif /* __ASM_IRQ_WORK_H */
diff --git a/kernel/arch/tile/include/asm/irqflags.h b/kernel/arch/tile/include/asm/irqflags.h
new file mode 100644
index 000000000..60d62a292
--- /dev/null
+++ b/kernel/arch/tile/include/asm/irqflags.h
@@ -0,0 +1,311 @@
+/*
+ * Copyright 2010 Tilera Corporation. All Rights Reserved.
+ *
+ *   This program is free software; you can redistribute it and/or
+ *   modify it under the terms of the GNU General Public License
+ *   as published by the Free Software Foundation, version 2.
+ *
+ *   This program is distributed in the hope that it will be useful, but
+ *   WITHOUT ANY WARRANTY; without even the implied warranty of
+ *   MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
+ *   NON INFRINGEMENT.  See the GNU General Public License for
+ *   more details.
+ */
+
+#ifndef _ASM_TILE_IRQFLAGS_H
+#define _ASM_TILE_IRQFLAGS_H
+
+#include <arch/interrupts.h>
+#include <arch/chip.h>
+
+/*
+ * The set of interrupts we want to allow when interrupts are nominally
+ * disabled.  The remainder are effectively "NMI" interrupts from
+ * the point of view of the generic Linux code.  Note that synchronous
+ * interrupts (aka "non-queued") are not blocked by the mask in any case.
+ */
+#define LINUX_MASKABLE_INTERRUPTS \
+	(~((_AC(1,ULL) << INT_PERF_COUNT) | (_AC(1,ULL) << INT_AUX_PERF_COUNT)))
+
+#if CHIP_HAS_SPLIT_INTR_MASK()
+/* The same macro, but for the two 32-bit SPRs separately. */
+#define LINUX_MASKABLE_INTERRUPTS_LO (-1)
+#define LINUX_MASKABLE_INTERRUPTS_HI \
+	(~((1 << (INT_PERF_COUNT - 32)) | (1 << (INT_AUX_PERF_COUNT - 32))))
+#endif
+
+#ifndef __ASSEMBLY__
+
+/* NOTE: we can't include <linux/percpu.h> due to #include dependencies. */
+#include <asm/percpu.h>
+#include <arch/spr_def.h>
+
+/*
+ * Set and clear kernel interrupt masks.
+ *
+ * NOTE: __insn_mtspr() is a compiler builtin marked as a memory
+ * clobber.  We rely on it being equivalent to a compiler barrier in
+ * this code since arch_local_irq_save() and friends must act as
+ * compiler barriers.  This compiler semantic is baked into enough
+ * places that the compiler will maintain it going forward.
+ */
+#if CHIP_HAS_SPLIT_INTR_MASK()
+#if INT_PERF_COUNT < 32 || INT_AUX_PERF_COUNT < 32 || INT_MEM_ERROR >= 32
+# error Fix assumptions about which word various interrupts are in
+#endif
+#define interrupt_mask_set(n) do { \
+	int __n = (n); \
+	int __mask = 1 << (__n & 0x1f); \
+	if (__n < 32) \
+		__insn_mtspr(SPR_INTERRUPT_MASK_SET_K_0, __mask); \
+	else \
+		__insn_mtspr(SPR_INTERRUPT_MASK_SET_K_1, __mask); \
+} while (0)
+#define interrupt_mask_reset(n) do { \
+	int __n = (n); \
+	int __mask = 1 << (__n & 0x1f); \
+	if (__n < 32) \
+		__insn_mtspr(SPR_INTERRUPT_MASK_RESET_K_0, __mask); \
+	else \
+		__insn_mtspr(SPR_INTERRUPT_MASK_RESET_K_1, __mask); \
+} while (0)
+#define interrupt_mask_check(n) ({ \
+	int __n = (n); \
+	(((__n < 32) ? \
+	 __insn_mfspr(SPR_INTERRUPT_MASK_K_0) : \
+	 __insn_mfspr(SPR_INTERRUPT_MASK_K_1)) \
+	  >> (__n & 0x1f)) & 1; \
+})
+#define interrupt_mask_set_mask(mask) do { \
+	unsigned long long __m = (mask); \
+	__insn_mtspr(SPR_INTERRUPT_MASK_SET_K_0, (unsigned long)(__m)); \
+	__insn_mtspr(SPR_INTERRUPT_MASK_SET_K_1, (unsigned long)(__m>>32)); \
+} while (0)
+#define interrupt_mask_reset_mask(mask) do { \
+	unsigned long long __m = (mask); \
+	__insn_mtspr(SPR_INTERRUPT_MASK_RESET_K_0, (unsigned long)(__m)); \
+	__insn_mtspr(SPR_INTERRUPT_MASK_RESET_K_1, (unsigned long)(__m>>32)); \
+} while (0)
+#define interrupt_mask_save_mask() \
+	(__insn_mfspr(SPR_INTERRUPT_MASK_SET_K_0) | \
+	 (((unsigned long long)__insn_mfspr(SPR_INTERRUPT_MASK_SET_K_1))<<32))
+#define interrupt_mask_restore_mask(mask) do { \
+	unsigned long long __m = (mask); \
+	__insn_mtspr(SPR_INTERRUPT_MASK_K_0, (unsigned long)(__m)); \
+	__insn_mtspr(SPR_INTERRUPT_MASK_K_1, (unsigned long)(__m>>32)); \
+} while (0)
+#else
+#define interrupt_mask_set(n) \
+	__insn_mtspr(SPR_INTERRUPT_MASK_SET_K, (1UL << (n)))
+#define interrupt_mask_reset(n) \
+	__insn_mtspr(SPR_INTERRUPT_MASK_RESET_K, (1UL << (n)))
+#define interrupt_mask_check(n) \
+	((__insn_mfspr(SPR_INTERRUPT_MASK_K) >> (n)) & 1)
+#define interrupt_mask_set_mask(mask) \
+	__insn_mtspr(SPR_INTERRUPT_MASK_SET_K, (mask))
+#define interrupt_mask_reset_mask(mask) \
+	__insn_mtspr(SPR_INTERRUPT_MASK_RESET_K, (mask))
+#define interrupt_mask_save_mask() \
+	__insn_mfspr(SPR_INTERRUPT_MASK_K)
+#define interrupt_mask_restore_mask(mask) \
+	__insn_mtspr(SPR_INTERRUPT_MASK_K, (mask))
+#endif
+
+/*
+ * The set of interrupts we want active if irqs are enabled.
+ * Note that in particular, the tile timer interrupt comes and goes
+ * from this set, since we have no other way to turn off the timer.
+ * Likewise, INTCTRL_K is removed and re-added during device
+ * interrupts, as is the the hardwall UDN_FIREWALL interrupt.
+ * We use a low bit (MEM_ERROR) as our sentinel value and make sure it
+ * is always claimed as an "active interrupt" so we can query that bit
+ * to know our current state.
+ */
+DECLARE_PER_CPU(unsigned long long, interrupts_enabled_mask);
+#define INITIAL_INTERRUPTS_ENABLED (1ULL << INT_MEM_ERROR)
+
+#ifdef CONFIG_DEBUG_PREEMPT
+/* Due to inclusion issues, we can't rely on <linux/smp.h> here. */
+extern unsigned int debug_smp_processor_id(void);
+# define smp_processor_id() debug_smp_processor_id()
+#endif
+
+/* Disable interrupts. */
+#define arch_local_irq_disable() \
+	interrupt_mask_set_mask(LINUX_MASKABLE_INTERRUPTS)
+
+/* Disable all interrupts, including NMIs. */
+#define arch_local_irq_disable_all() \
+	interrupt_mask_set_mask(-1ULL)
+
+/*
+ * Read the set of maskable interrupts.
+ * We avoid the preemption warning here via raw_cpu_ptr since even
+ * if irqs are already enabled, it's harmless to read the wrong cpu's
+ * enabled mask.
+ */
+#define arch_local_irqs_enabled() \
+	(*raw_cpu_ptr(&interrupts_enabled_mask))
+
+/* Re-enable all maskable interrupts. */
+#define arch_local_irq_enable() \
+	interrupt_mask_reset_mask(arch_local_irqs_enabled())
+
+/* Disable or enable interrupts based on flag argument. */
+#define arch_local_irq_restore(disabled) do { \
+	if (disabled) \
+		arch_local_irq_disable(); \
+	else \
+		arch_local_irq_enable(); \
+} while (0)
+
+/* Return true if "flags" argument means interrupts are disabled. */
+#define arch_irqs_disabled_flags(flags) ((flags) != 0)
+
+/* Return true if interrupts are currently disabled. */
+#define arch_irqs_disabled() interrupt_mask_check(INT_MEM_ERROR)
+
+/* Save whether interrupts are currently disabled. */
+#define arch_local_save_flags() arch_irqs_disabled()
+
+/* Save whether interrupts are currently disabled, then disable them. */
+#define arch_local_irq_save() ({ \
+	unsigned long __flags = arch_local_save_flags(); \
+	arch_local_irq_disable(); \
+	__flags; })
+
+/* Prevent the given interrupt from being enabled next time we enable irqs. */
+#define arch_local_irq_mask(interrupt) \
+	this_cpu_and(interrupts_enabled_mask, ~(1ULL << (interrupt)))
+
+/* Prevent the given interrupt from being enabled immediately. */
+#define arch_local_irq_mask_now(interrupt) do { \
+	arch_local_irq_mask(interrupt); \
+	interrupt_mask_set(interrupt); \
+} while (0)
+
+/* Allow the given interrupt to be enabled next time we enable irqs. */
+#define arch_local_irq_unmask(interrupt) \
+	this_cpu_or(interrupts_enabled_mask, (1ULL << (interrupt)))
+
+/* Allow the given interrupt to be enabled immediately, if !irqs_disabled. */
+#define arch_local_irq_unmask_now(interrupt) do { \
+	arch_local_irq_unmask(interrupt); \
+	if (!irqs_disabled()) \
+		interrupt_mask_reset(interrupt); \
+} while (0)
+
+#else /* __ASSEMBLY__ */
+
+/* We provide a somewhat more restricted set for assembly. */
+
+#ifdef __tilegx__
+
+#if INT_MEM_ERROR != 0
+# error Fix IRQS_DISABLED() macro
+#endif
+
+/* Return 0 or 1 to indicate whether interrupts are currently disabled. */
+#define IRQS_DISABLED(tmp)					\
+	mfspr   tmp, SPR_INTERRUPT_MASK_K;			\
+	andi    tmp, tmp, 1
+
+/* Load up a pointer to &interrupts_enabled_mask. */
+#define GET_INTERRUPTS_ENABLED_MASK_PTR(reg)			\
+	moveli reg, hw2_last(interrupts_enabled_mask);		\
+	shl16insli reg, reg, hw1(interrupts_enabled_mask);	\
+	shl16insli reg, reg, hw0(interrupts_enabled_mask);	\
+	add     reg, reg, tp
+
+/* Disable interrupts. */
+#define IRQ_DISABLE(tmp0, tmp1)					\
+	moveli  tmp0, hw2_last(LINUX_MASKABLE_INTERRUPTS);	\
+	shl16insli tmp0, tmp0, hw1(LINUX_MASKABLE_INTERRUPTS);	\
+	shl16insli tmp0, tmp0, hw0(LINUX_MASKABLE_INTERRUPTS);	\
+	mtspr   SPR_INTERRUPT_MASK_SET_K, tmp0
+
+/* Disable ALL synchronous interrupts (used by NMI entry). */
+#define IRQ_DISABLE_ALL(tmp)					\
+	movei   tmp, -1;					\
+	mtspr   SPR_INTERRUPT_MASK_SET_K, tmp
+
+/* Enable interrupts. */
+#define IRQ_ENABLE_LOAD(tmp0, tmp1)				\
+	GET_INTERRUPTS_ENABLED_MASK_PTR(tmp0);			\
+	ld      tmp0, tmp0
+#define IRQ_ENABLE_APPLY(tmp0, tmp1)				\
+	mtspr   SPR_INTERRUPT_MASK_RESET_K, tmp0
+
+#else /* !__tilegx__ */
+
+/*
+ * Return 0 or 1 to indicate whether interrupts are currently disabled.
+ * Note that it's important that we use a bit from the "low" mask word,
+ * since when we are enabling, that is the word we write first, so if we
+ * are interrupted after only writing half of the mask, the interrupt
+ * handler will correctly observe that we have interrupts enabled, and
+ * will enable interrupts itself on return from the interrupt handler
+ * (making the original code's write of the "high" mask word idempotent).
+ */
+#define IRQS_DISABLED(tmp)					\
+	mfspr   tmp, SPR_INTERRUPT_MASK_K_0;			\
+	shri    tmp, tmp, INT_MEM_ERROR;			\
+	andi    tmp, tmp, 1
+
+/* Load up a pointer to &interrupts_enabled_mask. */
+#define GET_INTERRUPTS_ENABLED_MASK_PTR(reg)			\
+	moveli  reg, lo16(interrupts_enabled_mask);		\
+	auli    reg, reg, ha16(interrupts_enabled_mask);	\
+	add     reg, reg, tp
+
+/* Disable interrupts. */
+#define IRQ_DISABLE(tmp0, tmp1)					\
+	{							\
+	 movei  tmp0, LINUX_MASKABLE_INTERRUPTS_LO;		\
+	 moveli tmp1, lo16(LINUX_MASKABLE_INTERRUPTS_HI)	\
+	};							\
+	{							\
+	 mtspr  SPR_INTERRUPT_MASK_SET_K_0, tmp0;		\
+	 auli   tmp1, tmp1, ha16(LINUX_MASKABLE_INTERRUPTS_HI)	\
+	};							\
+	mtspr   SPR_INTERRUPT_MASK_SET_K_1, tmp1
+
+/* Disable ALL synchronous interrupts (used by NMI entry). */
+#define IRQ_DISABLE_ALL(tmp)					\
+	movei   tmp, -1;					\
+	mtspr   SPR_INTERRUPT_MASK_SET_K_0, tmp;		\
+	mtspr   SPR_INTERRUPT_MASK_SET_K_1, tmp
+
+/* Enable interrupts. */
+#define IRQ_ENABLE_LOAD(tmp0, tmp1)				\
+	GET_INTERRUPTS_ENABLED_MASK_PTR(tmp0);			\
+	{							\
+	 lw     tmp0, tmp0;					\
+	 addi   tmp1, tmp0, 4					\
+	};							\
+	lw      tmp1, tmp1
+#define IRQ_ENABLE_APPLY(tmp0, tmp1)				\
+	mtspr   SPR_INTERRUPT_MASK_RESET_K_0, tmp0;		\
+	mtspr   SPR_INTERRUPT_MASK_RESET_K_1, tmp1
+#endif
+
+#define IRQ_ENABLE(tmp0, tmp1)					\
+	IRQ_ENABLE_LOAD(tmp0, tmp1);				\
+	IRQ_ENABLE_APPLY(tmp0, tmp1)
+
+/*
+ * Do the CPU's IRQ-state tracing from assembly code. We call a
+ * C function, but almost everywhere we do, we don't mind clobbering
+ * all the caller-saved registers.
+ */
+#ifdef CONFIG_TRACE_IRQFLAGS
+# define TRACE_IRQS_ON  jal trace_hardirqs_on
+# define TRACE_IRQS_OFF jal trace_hardirqs_off
+#else
+# define TRACE_IRQS_ON
+# define TRACE_IRQS_OFF
+#endif
+
+#endif /* __ASSEMBLY__ */
+
+#endif /* _ASM_TILE_IRQFLAGS_H */
diff --git a/kernel/arch/tile/include/asm/kdebug.h b/kernel/arch/tile/include/asm/kdebug.h
new file mode 100644
index 000000000..5bbbfa904
--- /dev/null
+++ b/kernel/arch/tile/include/asm/kdebug.h
@@ -0,0 +1,28 @@
+/*
+ * Copyright 2012 Tilera Corporation. All Rights Reserved.
+ *
+ *   This program is free software; you can redistribute it and/or
+ *   modify it under the terms of the GNU General Public License
+ *   as published by the Free Software Foundation, version 2.
+ *
+ *   This program is distributed in the hope that it will be useful, but
+ *   WITHOUT ANY WARRANTY; without even the implied warranty of
+ *   MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
+ *   NON INFRINGEMENT.  See the GNU General Public License for
+ *   more details.
+ */
+
+#ifndef _ASM_TILE_KDEBUG_H
+#define _ASM_TILE_KDEBUG_H
+
+#include <linux/notifier.h>
+
+enum die_val {
+	DIE_OOPS = 1,
+	DIE_BREAK,
+	DIE_SSTEPBP,
+	DIE_PAGE_FAULT,
+	DIE_COMPILED_BPT
+};
+
+#endif /* _ASM_TILE_KDEBUG_H */
diff --git a/kernel/arch/tile/include/asm/kexec.h b/kernel/arch/tile/include/asm/kexec.h
new file mode 100644
index 000000000..fc98ccfc9
--- /dev/null
+++ b/kernel/arch/tile/include/asm/kexec.h
@@ -0,0 +1,65 @@
+/*
+ * Copyright 2010 Tilera Corporation. All Rights Reserved.
+ *
+ *   This program is free software; you can redistribute it and/or
+ *   modify it under the terms of the GNU General Public License
+ *   as published by the Free Software Foundation, version 2.
+ *
+ *   This program is distributed in the hope that it will be useful, but
+ *   WITHOUT ANY WARRANTY; without even the implied warranty of
+ *   MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
+ *   NON INFRINGEMENT.  See the GNU General Public License for
+ *   more details.
+ *
+ * based on kexec.h from other architectures in linux-2.6.18
+ */
+
+#ifndef _ASM_TILE_KEXEC_H
+#define _ASM_TILE_KEXEC_H
+
+#include <asm/page.h>
+
+#ifndef __tilegx__
+/* Maximum physical address we can use pages from. */
+#define KEXEC_SOURCE_MEMORY_LIMIT TASK_SIZE
+/* Maximum address we can reach in physical address mode. */
+#define KEXEC_DESTINATION_MEMORY_LIMIT TASK_SIZE
+/* Maximum address we can use for the control code buffer. */
+#define KEXEC_CONTROL_MEMORY_LIMIT TASK_SIZE
+#else
+/* We need to limit the memory below PGDIR_SIZE since
+ * we only setup page table for [0, PGDIR_SIZE) before final kexec.
+ */
+/* Maximum physical address we can use pages from. */
+#define KEXEC_SOURCE_MEMORY_LIMIT PGDIR_SIZE
+/* Maximum address we can reach in physical address mode. */
+#define KEXEC_DESTINATION_MEMORY_LIMIT PGDIR_SIZE
+/* Maximum address we can use for the control code buffer. */
+#define KEXEC_CONTROL_MEMORY_LIMIT PGDIR_SIZE
+#endif
+
+#define KEXEC_CONTROL_PAGE_SIZE	PAGE_SIZE
+
+/*
+ * We don't bother to provide a unique identifier, since we can only
+ * reboot with a single type of kernel image anyway.
+ */
+#define KEXEC_ARCH KEXEC_ARCH_DEFAULT
+
+/* Use the tile override for the page allocator. */
+struct page *kimage_alloc_pages_arch(gfp_t gfp_mask, unsigned int order);
+#define kimage_alloc_pages_arch kimage_alloc_pages_arch
+
+#define MAX_NOTE_BYTES 1024
+
+/* Defined in arch/tile/kernel/relocate_kernel.S */
+extern const unsigned char relocate_new_kernel[];
+extern const unsigned long relocate_new_kernel_size;
+extern void relocate_new_kernel_end(void);
+
+/* Provide a dummy definition to avoid build failures. */
+static inline void crash_setup_regs(struct pt_regs *n, struct pt_regs *o)
+{
+}
+
+#endif /* _ASM_TILE_KEXEC_H */
diff --git a/kernel/arch/tile/include/asm/kgdb.h b/kernel/arch/tile/include/asm/kgdb.h
new file mode 100644
index 000000000..280c181cf
--- /dev/null
+++ b/kernel/arch/tile/include/asm/kgdb.h
@@ -0,0 +1,71 @@
+/*
+ * Copyright 2013 Tilera Corporation. All Rights Reserved.
+ *
+ *   This program is free software; you can redistribute it and/or
+ *   modify it under the terms of the GNU General Public License
+ *   as published by the Free Software Foundation, version 2.
+ *
+ *   This program is distributed in the hope that it will be useful, but
+ *   WITHOUT ANY WARRANTY; without even the implied warranty of
+ *   MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
+ *   NON INFRINGEMENT.  See the GNU General Public License for
+ *   more details.
+ *
+ * TILE-Gx KGDB support.
+ */
+
+#ifndef __TILE_KGDB_H__
+#define __TILE_KGDB_H__
+
+#include <linux/kdebug.h>
+#include <arch/opcode.h>
+
+#define GDB_SIZEOF_REG		sizeof(unsigned long)
+
+/*
+ * TILE-Gx gdb is expecting the following register layout:
+ * 56 GPRs(R0 - R52, TP, SP, LR), 8 special GPRs(networks and ZERO),
+ * plus the PC and the faultnum.
+ *
+ * Even though kernel not use the 8 special GPRs, they need to be present
+ * in the registers sent for correct processing in the host-side gdb.
+ *
+ */
+#define DBG_MAX_REG_NUM		(56+8+2)
+#define NUMREGBYTES		(DBG_MAX_REG_NUM * GDB_SIZEOF_REG)
+
+/*
+ * BUFMAX defines the maximum number of characters in inbound/outbound
+ * buffers at least NUMREGBYTES*2 are needed for register packets,
+ * Longer buffer is needed to list all threads.
+ */
+#define BUFMAX			2048
+
+#define BREAK_INSTR_SIZE	TILEGX_BUNDLE_SIZE_IN_BYTES
+
+/*
+ * Require cache flush for set/clear a software breakpoint or write memory.
+ */
+#define CACHE_FLUSH_IS_SAFE	1
+
+/*
+ * The compiled-in breakpoint instruction can be used to "break" into
+ * the debugger via magic system request key (sysrq-G).
+ */
+static tile_bundle_bits compiled_bpt = TILEGX_BPT_BUNDLE | DIE_COMPILED_BPT;
+
+enum tilegx_regnum {
+	TILEGX_PC_REGNUM = TREG_LAST_GPR + 9,
+	TILEGX_FAULTNUM_REGNUM,
+};
+
+/*
+ * Generate a breakpoint exception to "break" into the debugger.
+ */
+static inline void arch_kgdb_breakpoint(void)
+{
+	asm volatile (".quad %0\n\t"
+		      ::""(compiled_bpt));
+}
+
+#endif /* __TILE_KGDB_H__ */
diff --git a/kernel/arch/tile/include/asm/kmap_types.h b/kernel/arch/tile/include/asm/kmap_types.h
new file mode 100644
index 000000000..92b28e3e9
--- /dev/null
+++ b/kernel/arch/tile/include/asm/kmap_types.h
@@ -0,0 +1,28 @@
+/*
+ * Copyright 2010 Tilera Corporation. All Rights Reserved.
+ *
+ *   This program is free software; you can redistribute it and/or
+ *   modify it under the terms of the GNU General Public License
+ *   as published by the Free Software Foundation, version 2.
+ *
+ *   This program is distributed in the hope that it will be useful, but
+ *   WITHOUT ANY WARRANTY; without even the implied warranty of
+ *   MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
+ *   NON INFRINGEMENT.  See the GNU General Public License for
+ *   more details.
+ */
+
+#ifndef _ASM_TILE_KMAP_TYPES_H
+#define _ASM_TILE_KMAP_TYPES_H
+
+/*
+ * In 32-bit TILE Linux we have to balance the desire to have a lot of
+ * nested atomic mappings with the fact that large page sizes and many
+ * processors chew up address space quickly.  In a typical
+ * 64-processor, 64KB-page layout build, making KM_TYPE_NR one larger
+ * adds 4MB of required address-space.  For now we leave KM_TYPE_NR
+ * set to depth 8.
+ */
+#define	KM_TYPE_NR 8
+
+#endif /* _ASM_TILE_KMAP_TYPES_H */
diff --git a/kernel/arch/tile/include/asm/kprobes.h b/kernel/arch/tile/include/asm/kprobes.h
new file mode 100644
index 000000000..d8f9a8394
--- /dev/null
+++ b/kernel/arch/tile/include/asm/kprobes.h
@@ -0,0 +1,79 @@
+/*
+ * arch/tile/include/asm/kprobes.h
+ *
+ * Copyright 2012 Tilera Corporation. All Rights Reserved.
+ *
+ *   This program is free software; you can redistribute it and/or
+ *   modify it under the terms of the GNU General Public License
+ *   as published by the Free Software Foundation, version 2.
+ *
+ *   This program is distributed in the hope that it will be useful, but
+ *   WITHOUT ANY WARRANTY; without even the implied warranty of
+ *   MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
+ *   NON INFRINGEMENT.  See the GNU General Public License for
+ *   more details.
+ */
+
+#ifndef _ASM_TILE_KPROBES_H
+#define _ASM_TILE_KPROBES_H
+
+#include <linux/types.h>
+#include <linux/ptrace.h>
+#include <linux/percpu.h>
+
+#include <arch/opcode.h>
+
+#define __ARCH_WANT_KPROBES_INSN_SLOT
+#define MAX_INSN_SIZE			2
+
+#define kretprobe_blacklist_size 0
+
+typedef tile_bundle_bits kprobe_opcode_t;
+
+#define flush_insn_slot(p)						\
+	flush_icache_range((unsigned long)p->addr,			\
+			   (unsigned long)p->addr +			\
+			   (MAX_INSN_SIZE * sizeof(kprobe_opcode_t)))
+
+struct kprobe;
+
+/* Architecture specific copy of original instruction. */
+struct arch_specific_insn {
+	kprobe_opcode_t *insn;
+};
+
+struct prev_kprobe {
+	struct kprobe *kp;
+	unsigned long status;
+	unsigned long saved_pc;
+};
+
+#define MAX_JPROBES_STACK_SIZE 128
+#define MAX_JPROBES_STACK_ADDR \
+	(((unsigned long)current_thread_info()) + THREAD_SIZE - 32 \
+		- sizeof(struct pt_regs))
+
+#define MIN_JPROBES_STACK_SIZE(ADDR)					\
+	((((ADDR) + MAX_JPROBES_STACK_SIZE) > MAX_JPROBES_STACK_ADDR)	\
+		? MAX_JPROBES_STACK_ADDR - (ADDR)			\
+		: MAX_JPROBES_STACK_SIZE)
+
+/* per-cpu kprobe control block. */
+struct kprobe_ctlblk {
+	unsigned long kprobe_status;
+	unsigned long kprobe_saved_pc;
+	unsigned long jprobe_saved_sp;
+	struct prev_kprobe prev_kprobe;
+	struct pt_regs jprobe_saved_regs;
+	char jprobes_stack[MAX_JPROBES_STACK_SIZE];
+};
+
+extern tile_bundle_bits breakpoint2_insn;
+extern tile_bundle_bits breakpoint_insn;
+
+void arch_remove_kprobe(struct kprobe *);
+
+extern int kprobe_exceptions_notify(struct notifier_block *self,
+			     unsigned long val, void *data);
+
+#endif /* _ASM_TILE_KPROBES_H */
diff --git a/kernel/arch/tile/include/asm/linkage.h b/kernel/arch/tile/include/asm/linkage.h
new file mode 100644
index 000000000..e121c3975
--- /dev/null
+++ b/kernel/arch/tile/include/asm/linkage.h
@@ -0,0 +1,51 @@
+/*
+ * Copyright 2010 Tilera Corporation. All Rights Reserved.
+ *
+ *   This program is free software; you can redistribute it and/or
+ *   modify it under the terms of the GNU General Public License
+ *   as published by the Free Software Foundation, version 2.
+ *
+ *   This program is distributed in the hope that it will be useful, but
+ *   WITHOUT ANY WARRANTY; without even the implied warranty of
+ *   MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
+ *   NON INFRINGEMENT.  See the GNU General Public License for
+ *   more details.
+ */
+
+#ifndef _ASM_TILE_LINKAGE_H
+#define _ASM_TILE_LINKAGE_H
+
+#include <feedback.h>
+
+#define __ALIGN .align 8
+
+/*
+ * The STD_ENTRY and STD_ENDPROC macros put the function in a
+ * self-named .text.foo section, and if linker feedback collection
+ * is enabled, add a suitable call to the feedback collection code.
+ * STD_ENTRY_SECTION lets you specify a non-standard section name.
+ */
+
+#define STD_ENTRY(name) \
+  .pushsection .text.##name, "ax"; \
+  ENTRY(name); \
+  FEEDBACK_ENTER(name)
+
+#define STD_ENTRY_SECTION(name, section) \
+  .pushsection section, "ax"; \
+  ENTRY(name); \
+  FEEDBACK_ENTER_EXPLICIT(name, section, .Lend_##name - name)
+
+#define STD_ENDPROC(name) \
+  ENDPROC(name); \
+  .Lend_##name:; \
+  .popsection
+
+/* Create a file-static function entry set up for feedback gathering. */
+#define STD_ENTRY_LOCAL(name) \
+  .pushsection .text.##name, "ax"; \
+  ALIGN; \
+  name:; \
+  FEEDBACK_ENTER(name)
+
+#endif /* _ASM_TILE_LINKAGE_H */
diff --git a/kernel/arch/tile/include/asm/mmu.h b/kernel/arch/tile/include/asm/mmu.h
new file mode 100644
index 000000000..0cab1182b
--- /dev/null
+++ b/kernel/arch/tile/include/asm/mmu.h
@@ -0,0 +1,32 @@
+/*
+ * Copyright 2010 Tilera Corporation. All Rights Reserved.
+ *
+ *   This program is free software; you can redistribute it and/or
+ *   modify it under the terms of the GNU General Public License
+ *   as published by the Free Software Foundation, version 2.
+ *
+ *   This program is distributed in the hope that it will be useful, but
+ *   WITHOUT ANY WARRANTY; without even the implied warranty of
+ *   MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
+ *   NON INFRINGEMENT.  See the GNU General Public License for
+ *   more details.
+ */
+
+#ifndef _ASM_TILE_MMU_H
+#define _ASM_TILE_MMU_H
+
+/* Capture any arch- and mm-specific information. */
+struct mm_context {
+	/*
+	 * Written under the mmap_sem semaphore; read without the
+	 * semaphore but atomically, but it is conservatively set.
+	 */
+	unsigned long priority_cached;
+	unsigned long vdso_base;
+};
+
+typedef struct mm_context mm_context_t;
+
+void leave_mm(int cpu);
+
+#endif /* _ASM_TILE_MMU_H */
diff --git a/kernel/arch/tile/include/asm/mmu_context.h b/kernel/arch/tile/include/asm/mmu_context.h
new file mode 100644
index 000000000..f67753db1
--- /dev/null
+++ b/kernel/arch/tile/include/asm/mmu_context.h
@@ -0,0 +1,135 @@
+/*
+ * Copyright 2010 Tilera Corporation. All Rights Reserved.
+ *
+ *   This program is free software; you can redistribute it and/or
+ *   modify it under the terms of the GNU General Public License
+ *   as published by the Free Software Foundation, version 2.
+ *
+ *   This program is distributed in the hope that it will be useful, but
+ *   WITHOUT ANY WARRANTY; without even the implied warranty of
+ *   MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
+ *   NON INFRINGEMENT.  See the GNU General Public License for
+ *   more details.
+ */
+
+#ifndef _ASM_TILE_MMU_CONTEXT_H
+#define _ASM_TILE_MMU_CONTEXT_H
+
+#include <linux/smp.h>
+#include <asm/setup.h>
+#include <asm/page.h>
+#include <asm/pgalloc.h>
+#include <asm/pgtable.h>
+#include <asm/tlbflush.h>
+#include <asm/homecache.h>
+#include <asm-generic/mm_hooks.h>
+
+static inline int
+init_new_context(struct task_struct *tsk, struct mm_struct *mm)
+{
+	return 0;
+}
+
+/*
+ * Note that arch/tile/kernel/head_NN.S and arch/tile/mm/migrate_NN.S
+ * also call hv_install_context().
+ */
+static inline void __install_page_table(pgd_t *pgdir, int asid, pgprot_t prot)
+{
+	/* FIXME: DIRECTIO should not always be set. FIXME. */
+	int rc = hv_install_context(__pa(pgdir), prot, asid,
+				    HV_CTX_DIRECTIO | CTX_PAGE_FLAG);
+	if (rc < 0)
+		panic("hv_install_context failed: %d", rc);
+}
+
+static inline void install_page_table(pgd_t *pgdir, int asid)
+{
+	pte_t *ptep = virt_to_kpte((unsigned long)pgdir);
+	__install_page_table(pgdir, asid, *ptep);
+}
+
+/*
+ * "Lazy" TLB mode is entered when we are switching to a kernel task,
+ * which borrows the mm of the previous task.  The goal of this
+ * optimization is to avoid having to install a new page table.  On
+ * early x86 machines (where the concept originated) you couldn't do
+ * anything short of a full page table install for invalidation, so
+ * handling a remote TLB invalidate required doing a page table
+ * re-install.  Someone clearly decided that it was silly to keep
+ * doing this while in "lazy" TLB mode, so the optimization involves
+ * installing the swapper page table instead the first time one
+ * occurs, and clearing the cpu out of cpu_vm_mask, so the cpu running
+ * the kernel task doesn't need to take any more interrupts.  At that
+ * point it's then necessary to explicitly reinstall it when context
+ * switching back to the original mm.
+ *
+ * On Tile, we have to do a page-table install whenever DMA is enabled,
+ * so in that case lazy mode doesn't help anyway.  And more generally,
+ * we have efficient per-page TLB shootdown, and don't expect to spend
+ * that much time in kernel tasks in general, so just leaving the
+ * kernel task borrowing the old page table, but handling TLB
+ * shootdowns, is a reasonable thing to do.  And importantly, this
+ * lets us use the hypervisor's internal APIs for TLB shootdown, which
+ * means we don't have to worry about having TLB shootdowns blocked
+ * when Linux is disabling interrupts; see the page migration code for
+ * an example of where it's important for TLB shootdowns to complete
+ * even when interrupts are disabled at the Linux level.
+ */
+static inline void enter_lazy_tlb(struct mm_struct *mm, struct task_struct *t)
+{
+#if CHIP_HAS_TILE_DMA()
+	/*
+	 * We have to do an "identity" page table switch in order to
+	 * clear any pending DMA interrupts.
+	 */
+	if (current->thread.tile_dma_state.enabled)
+		install_page_table(mm->pgd, __this_cpu_read(current_asid));
+#endif
+}
+
+static inline void switch_mm(struct mm_struct *prev, struct mm_struct *next,
+			     struct task_struct *tsk)
+{
+	if (likely(prev != next)) {
+
+		int cpu = smp_processor_id();
+
+		/* Pick new ASID. */
+		int asid = __this_cpu_read(current_asid) + 1;
+		if (asid > max_asid) {
+			asid = min_asid;
+			local_flush_tlb();
+		}
+		__this_cpu_write(current_asid, asid);
+
+		/* Clear cpu from the old mm, and set it in the new one. */
+		cpumask_clear_cpu(cpu, mm_cpumask(prev));
+		cpumask_set_cpu(cpu, mm_cpumask(next));
+
+		/* Re-load page tables */
+		install_page_table(next->pgd, asid);
+
+		/* See how we should set the red/black cache info */
+		check_mm_caching(prev, next);
+
+		/*
+		 * Since we're changing to a new mm, we have to flush
+		 * the icache in case some physical page now being mapped
+		 * has subsequently been repurposed and has new code.
+		 */
+		__flush_icache();
+
+	}
+}
+
+static inline void activate_mm(struct mm_struct *prev_mm,
+			       struct mm_struct *next_mm)
+{
+	switch_mm(prev_mm, next_mm, NULL);
+}
+
+#define destroy_context(mm)		do { } while (0)
+#define deactivate_mm(tsk, mm)          do { } while (0)
+
+#endif /* _ASM_TILE_MMU_CONTEXT_H */
diff --git a/kernel/arch/tile/include/asm/mmzone.h b/kernel/arch/tile/include/asm/mmzone.h
new file mode 100644
index 000000000..804f1098b
--- /dev/null
+++ b/kernel/arch/tile/include/asm/mmzone.h
@@ -0,0 +1,70 @@
+/*
+ * Copyright 2010 Tilera Corporation. All Rights Reserved.
+ *
+ *   This program is free software; you can redistribute it and/or
+ *   modify it under the terms of the GNU General Public License
+ *   as published by the Free Software Foundation, version 2.
+ *
+ *   This program is distributed in the hope that it will be useful, but
+ *   WITHOUT ANY WARRANTY; without even the implied warranty of
+ *   MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
+ *   NON INFRINGEMENT.  See the GNU General Public License for
+ *   more details.
+ */
+
+#ifndef _ASM_TILE_MMZONE_H
+#define _ASM_TILE_MMZONE_H
+
+extern struct pglist_data node_data[];
+#define NODE_DATA(nid)	(&node_data[nid])
+
+extern void get_memcfg_numa(void);
+
+#ifdef CONFIG_DISCONTIGMEM
+
+#include <asm/page.h>
+
+/*
+ * Generally, memory ranges are always doled out by the hypervisor in
+ * fixed-size, power-of-two increments.  That would make computing the node
+ * very easy.  We could just take a couple high bits of the PA, which
+ * denote the memory shim, and we'd be done.  However, when we're doing
+ * memory striping, this may not be true; PAs with different high bit
+ * values might be in the same node.  Thus, we keep a lookup table to
+ * translate the high bits of the PFN to the node number.
+ */
+extern int highbits_to_node[];
+
+static inline int pfn_to_nid(unsigned long pfn)
+{
+	return highbits_to_node[__pfn_to_highbits(pfn)];
+}
+
+#define kern_addr_valid(kaddr)	virt_addr_valid((void *)kaddr)
+
+static inline int pfn_valid(unsigned long pfn)
+{
+	int nid = pfn_to_nid(pfn);
+
+	if (nid >= 0)
+		return (pfn < node_end_pfn(nid));
+	return 0;
+}
+
+/* Information on the NUMA nodes that we compute early */
+extern unsigned long node_start_pfn[];
+extern unsigned long node_end_pfn[];
+extern unsigned long node_memmap_pfn[];
+extern unsigned long node_percpu_pfn[];
+extern unsigned long node_free_pfn[];
+#ifdef CONFIG_HIGHMEM
+extern unsigned long node_lowmem_end_pfn[];
+#endif
+#ifdef CONFIG_PCI
+extern unsigned long pci_reserve_start_pfn;
+extern unsigned long pci_reserve_end_pfn;
+#endif
+
+#endif /* CONFIG_DISCONTIGMEM */
+
+#endif /* _ASM_TILE_MMZONE_H */
diff --git a/kernel/arch/tile/include/asm/module.h b/kernel/arch/tile/include/asm/module.h
new file mode 100644
index 000000000..44ed07ccd
--- /dev/null
+++ b/kernel/arch/tile/include/asm/module.h
@@ -0,0 +1,40 @@
+/*
+ * Copyright 2011 Tilera Corporation. All Rights Reserved.
+ *
+ *   This program is free software; you can redistribute it and/or
+ *   modify it under the terms of the GNU General Public License
+ *   as published by the Free Software Foundation, version 2.
+ *
+ *   This program is distributed in the hope that it will be useful, but
+ *   WITHOUT ANY WARRANTY; without even the implied warranty of
+ *   MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
+ *   NON INFRINGEMENT.  See the GNU General Public License for
+ *   more details.
+ */
+
+#ifndef _ASM_TILE_MODULE_H
+#define _ASM_TILE_MODULE_H
+
+#include <arch/chip.h>
+
+#include <asm-generic/module.h>
+
+/* We can't use modules built with different page sizes. */
+#if defined(CONFIG_PAGE_SIZE_16KB)
+# define MODULE_PGSZ " 16KB"
+#elif defined(CONFIG_PAGE_SIZE_64KB)
+# define MODULE_PGSZ " 64KB"
+#else
+# define MODULE_PGSZ ""
+#endif
+
+/* We don't really support no-SMP so tag if someone tries. */
+#ifdef CONFIG_SMP
+#define MODULE_NOSMP ""
+#else
+#define MODULE_NOSMP " nosmp"
+#endif
+
+#define MODULE_ARCH_VERMAGIC CHIP_ARCH_NAME MODULE_PGSZ MODULE_NOSMP
+
+#endif /* _ASM_TILE_MODULE_H */
diff --git a/kernel/arch/tile/include/asm/page.h b/kernel/arch/tile/include/asm/page.h
new file mode 100644
index 000000000..a213a8d84
--- /dev/null
+++ b/kernel/arch/tile/include/asm/page.h
@@ -0,0 +1,333 @@
+/*
+ * Copyright 2010 Tilera Corporation. All Rights Reserved.
+ *
+ *   This program is free software; you can redistribute it and/or
+ *   modify it under the terms of the GNU General Public License
+ *   as published by the Free Software Foundation, version 2.
+ *
+ *   This program is distributed in the hope that it will be useful, but
+ *   WITHOUT ANY WARRANTY; without even the implied warranty of
+ *   MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
+ *   NON INFRINGEMENT.  See the GNU General Public License for
+ *   more details.
+ */
+
+#ifndef _ASM_TILE_PAGE_H
+#define _ASM_TILE_PAGE_H
+
+#include <linux/const.h>
+#include <hv/hypervisor.h>
+#include <arch/chip.h>
+
+/* PAGE_SHIFT and HPAGE_SHIFT determine the page sizes. */
+#if defined(CONFIG_PAGE_SIZE_16KB)
+#define PAGE_SHIFT	14
+#define CTX_PAGE_FLAG	HV_CTX_PG_SM_16K
+#elif defined(CONFIG_PAGE_SIZE_64KB)
+#define PAGE_SHIFT	16
+#define CTX_PAGE_FLAG	HV_CTX_PG_SM_64K
+#else
+#define PAGE_SHIFT	HV_LOG2_DEFAULT_PAGE_SIZE_SMALL
+#define CTX_PAGE_FLAG	0
+#endif
+#define HPAGE_SHIFT	HV_LOG2_DEFAULT_PAGE_SIZE_LARGE
+
+#define PAGE_SIZE	(_AC(1, UL) << PAGE_SHIFT)
+#define HPAGE_SIZE	(_AC(1, UL) << HPAGE_SHIFT)
+
+#define PAGE_MASK	(~(PAGE_SIZE - 1))
+#define HPAGE_MASK	(~(HPAGE_SIZE - 1))
+
+/*
+ * If the Kconfig doesn't specify, set a maximum zone order that
+ * is enough so that we can create huge pages from small pages given
+ * the respective sizes of the two page types.  See <linux/mmzone.h>.
+ */
+#ifndef CONFIG_FORCE_MAX_ZONEORDER
+#define CONFIG_FORCE_MAX_ZONEORDER (HPAGE_SHIFT - PAGE_SHIFT + 1)
+#endif
+
+#ifndef __ASSEMBLY__
+
+#include <linux/types.h>
+#include <linux/string.h>
+
+struct page;
+
+static inline void clear_page(void *page)
+{
+	memset(page, 0, PAGE_SIZE);
+}
+
+static inline void copy_page(void *to, void *from)
+{
+	memcpy(to, from, PAGE_SIZE);
+}
+
+static inline void clear_user_page(void *page, unsigned long vaddr,
+				struct page *pg)
+{
+	clear_page(page);
+}
+
+static inline void copy_user_page(void *to, void *from, unsigned long vaddr,
+				struct page *topage)
+{
+	copy_page(to, from);
+}
+
+/*
+ * Hypervisor page tables are made of the same basic structure.
+ */
+
+typedef HV_PTE pte_t;
+typedef HV_PTE pgd_t;
+typedef HV_PTE pgprot_t;
+
+/*
+ * User L2 page tables are managed as one L2 page table per page,
+ * because we use the page allocator for them.  This keeps the allocation
+ * simple, but it's also inefficient, since L2 page tables are much smaller
+ * than pages (currently 2KB vs 64KB).  So we should revisit this.
+ */
+typedef struct page *pgtable_t;
+
+/* Must be a macro since it is used to create constants. */
+#define __pgprot(val) hv_pte(val)
+
+/* Rarely-used initializers, typically with a "zero" value. */
+#define __pte(x) hv_pte(x)
+#define __pgd(x) hv_pte(x)
+
+static inline u64 pgprot_val(pgprot_t pgprot)
+{
+	return hv_pte_val(pgprot);
+}
+
+static inline u64 pte_val(pte_t pte)
+{
+	return hv_pte_val(pte);
+}
+
+static inline u64 pgd_val(pgd_t pgd)
+{
+	return hv_pte_val(pgd);
+}
+
+#ifdef __tilegx__
+
+typedef HV_PTE pmd_t;
+
+#define __pmd(x) hv_pte(x)
+
+static inline u64 pmd_val(pmd_t pmd)
+{
+	return hv_pte_val(pmd);
+}
+
+#endif
+
+static inline __attribute_const__ int get_order(unsigned long size)
+{
+	return BITS_PER_LONG - __builtin_clzl((size - 1) >> PAGE_SHIFT);
+}
+
+#endif /* !__ASSEMBLY__ */
+
+#define HUGETLB_PAGE_ORDER	(HPAGE_SHIFT - PAGE_SHIFT)
+
+#define HUGE_MAX_HSTATE		6
+
+#ifdef CONFIG_HUGETLB_PAGE
+#define HAVE_ARCH_HUGETLB_UNMAPPED_AREA
+#endif
+
+/* Allow overriding how much VA or PA the kernel will use. */
+#define MAX_PA_WIDTH CHIP_PA_WIDTH()
+#define MAX_VA_WIDTH CHIP_VA_WIDTH()
+
+/* Each memory controller has PAs distinct in their high bits. */
+#define NR_PA_HIGHBIT_SHIFT (MAX_PA_WIDTH - CHIP_LOG_NUM_MSHIMS())
+#define NR_PA_HIGHBIT_VALUES (1 << CHIP_LOG_NUM_MSHIMS())
+#define __pa_to_highbits(pa) ((phys_addr_t)(pa) >> NR_PA_HIGHBIT_SHIFT)
+#define __pfn_to_highbits(pfn) ((pfn) >> (NR_PA_HIGHBIT_SHIFT - PAGE_SHIFT))
+
+#ifdef __tilegx__
+
+/*
+ * We reserve the lower half of memory for user-space programs, and the
+ * upper half for system code.  We re-map all of physical memory in the
+ * upper half, which takes a quarter of our VA space.  Then we have
+ * the vmalloc regions.  The supervisor code lives at the highest address,
+ * with the hypervisor above that.
+ *
+ * Loadable kernel modules are placed immediately after the static
+ * supervisor code, with each being allocated a 256MB region of
+ * address space, so we don't have to worry about the range of "jal"
+ * and other branch instructions.
+ *
+ * For now we keep life simple and just allocate one pmd (4GB) for vmalloc.
+ * Similarly, for now we don't play any struct page mapping games.
+ */
+
+#if MAX_PA_WIDTH + 2 > MAX_VA_WIDTH
+# error Too much PA to map with the VA available!
+#endif
+
+#define PAGE_OFFSET		(-(_AC(1, UL) << (MAX_VA_WIDTH - 1)))
+#define KERNEL_HIGH_VADDR	_AC(0xfffffff800000000, UL)  /* high 32GB */
+#define FIXADDR_BASE		(KERNEL_HIGH_VADDR - 0x300000000) /* 4 GB */
+#define FIXADDR_TOP		(KERNEL_HIGH_VADDR - 0x200000000) /* 4 GB */
+#define _VMALLOC_START		FIXADDR_TOP
+#define MEM_SV_START		(KERNEL_HIGH_VADDR - 0x100000000) /* 256 MB */
+#define MEM_MODULE_START	(MEM_SV_START + (256*1024*1024)) /* 256 MB */
+#define MEM_MODULE_END		(MEM_MODULE_START + (256*1024*1024))
+
+#else /* !__tilegx__ */
+
+/*
+ * A PAGE_OFFSET of 0xC0000000 means that the kernel has
+ * a virtual address space of one gigabyte, which limits the
+ * amount of physical memory you can use to about 768MB.
+ * If you want more physical memory than this then see the CONFIG_HIGHMEM
+ * option in the kernel configuration.
+ *
+ * The top 16MB chunk in the table below is unavailable to Linux.  Since
+ * the kernel interrupt vectors must live at ether 0xfe000000 or 0xfd000000
+ * (depending on whether the kernel is at PL2 or Pl1), we map all of the
+ * bottom of RAM at this address with a huge page table entry to minimize
+ * its ITLB footprint (as well as at PAGE_OFFSET).  The last architected
+ * requirement is that user interrupt vectors live at 0xfc000000, so we
+ * make that range of memory available to user processes.  The remaining
+ * regions are sized as shown; the first four addresses use the PL 1
+ * values, and after that, we show "typical" values, since the actual
+ * addresses depend on kernel #defines.
+ *
+ * MEM_HV_START                    0xfe000000
+ * MEM_SV_START  (kernel code)     0xfd000000
+ * MEM_USER_INTRPT (user vector)   0xfc000000
+ * FIX_KMAP_xxx                    0xfa000000 (via NR_CPUS * KM_TYPE_NR)
+ * PKMAP_BASE                      0xf9000000 (via LAST_PKMAP)
+ * VMALLOC_START                   0xf7000000 (via VMALLOC_RESERVE)
+ * mapped LOWMEM                   0xc0000000
+ */
+
+#define MEM_USER_INTRPT		_AC(0xfc000000, UL)
+#define MEM_SV_START		_AC(0xfd000000, UL)
+#define MEM_HV_START		_AC(0xfe000000, UL)
+
+#define INTRPT_SIZE		0x4000
+
+/* Tolerate page size larger than the architecture interrupt region size. */
+#if PAGE_SIZE > INTRPT_SIZE
+#undef INTRPT_SIZE
+#define INTRPT_SIZE PAGE_SIZE
+#endif
+
+#define KERNEL_HIGH_VADDR	MEM_USER_INTRPT
+#define FIXADDR_TOP		(KERNEL_HIGH_VADDR - PAGE_SIZE)
+
+#define PAGE_OFFSET		_AC(CONFIG_PAGE_OFFSET, UL)
+
+/* On 32-bit architectures we mix kernel modules in with other vmaps. */
+#define MEM_MODULE_START	VMALLOC_START
+#define MEM_MODULE_END		VMALLOC_END
+
+#endif /* __tilegx__ */
+
+#if !defined(__ASSEMBLY__) && !defined(VDSO_BUILD)
+
+#ifdef CONFIG_HIGHMEM
+
+/* Map kernel virtual addresses to page frames, in HPAGE_SIZE chunks. */
+extern unsigned long pbase_map[];
+extern void *vbase_map[];
+
+static inline unsigned long kaddr_to_pfn(const volatile void *_kaddr)
+{
+	unsigned long kaddr = (unsigned long)_kaddr;
+	return pbase_map[kaddr >> HPAGE_SHIFT] +
+		((kaddr & (HPAGE_SIZE - 1)) >> PAGE_SHIFT);
+}
+
+static inline void *pfn_to_kaddr(unsigned long pfn)
+{
+	return vbase_map[__pfn_to_highbits(pfn)] + (pfn << PAGE_SHIFT);
+}
+
+static inline phys_addr_t virt_to_phys(const volatile void *kaddr)
+{
+	unsigned long pfn = kaddr_to_pfn(kaddr);
+	return ((phys_addr_t)pfn << PAGE_SHIFT) +
+		((unsigned long)kaddr & (PAGE_SIZE-1));
+}
+
+static inline void *phys_to_virt(phys_addr_t paddr)
+{
+	return pfn_to_kaddr(paddr >> PAGE_SHIFT) + (paddr & (PAGE_SIZE-1));
+}
+
+/* With HIGHMEM, we pack PAGE_OFFSET through high_memory with all valid VAs. */
+static inline int virt_addr_valid(const volatile void *kaddr)
+{
+	extern void *high_memory;  /* copied from <linux/mm.h> */
+	return ((unsigned long)kaddr >= PAGE_OFFSET && kaddr < high_memory);
+}
+
+#else /* !CONFIG_HIGHMEM */
+
+static inline unsigned long kaddr_to_pfn(const volatile void *kaddr)
+{
+	return ((unsigned long)kaddr - PAGE_OFFSET) >> PAGE_SHIFT;
+}
+
+static inline void *pfn_to_kaddr(unsigned long pfn)
+{
+	return (void *)((pfn << PAGE_SHIFT) + PAGE_OFFSET);
+}
+
+static inline phys_addr_t virt_to_phys(const volatile void *kaddr)
+{
+	return (phys_addr_t)((unsigned long)kaddr - PAGE_OFFSET);
+}
+
+static inline void *phys_to_virt(phys_addr_t paddr)
+{
+	return (void *)((unsigned long)paddr + PAGE_OFFSET);
+}
+
+/* Check that the given address is within some mapped range of PAs. */
+#define virt_addr_valid(kaddr) pfn_valid(kaddr_to_pfn(kaddr))
+
+#endif /* !CONFIG_HIGHMEM */
+
+/* All callers are not consistent in how they call these functions. */
+#define __pa(kaddr) virt_to_phys((void *)(unsigned long)(kaddr))
+#define __va(paddr) phys_to_virt((phys_addr_t)(paddr))
+
+extern int devmem_is_allowed(unsigned long pagenr);
+
+#ifdef CONFIG_FLATMEM
+static inline int pfn_valid(unsigned long pfn)
+{
+	return pfn < max_mapnr;
+}
+#endif
+
+/* Provide as macros since these require some other headers included. */
+#define page_to_pa(page) ((phys_addr_t)(page_to_pfn(page)) << PAGE_SHIFT)
+#define virt_to_page(kaddr) pfn_to_page(kaddr_to_pfn((void *)(kaddr)))
+#define page_to_virt(page) pfn_to_kaddr(page_to_pfn(page))
+
+struct mm_struct;
+extern pte_t *virt_to_pte(struct mm_struct *mm, unsigned long addr);
+extern pte_t *virt_to_kpte(unsigned long kaddr);
+
+#endif /* !__ASSEMBLY__ */
+
+#define VM_DATA_DEFAULT_FLAGS \
+	(VM_READ | VM_WRITE | VM_MAYREAD | VM_MAYWRITE | VM_MAYEXEC)
+
+#include <asm-generic/memory_model.h>
+
+#endif /* _ASM_TILE_PAGE_H */
diff --git a/kernel/arch/tile/include/asm/pci.h b/kernel/arch/tile/include/asm/pci.h
new file mode 100644
index 000000000..dfedd7ac7
--- /dev/null
+++ b/kernel/arch/tile/include/asm/pci.h
@@ -0,0 +1,232 @@
+/*
+ * Copyright 2010 Tilera Corporation. All Rights Reserved.
+ *
+ *   This program is free software; you can redistribute it and/or
+ *   modify it under the terms of the GNU General Public License
+ *   as published by the Free Software Foundation, version 2.
+ *
+ *   This program is distributed in the hope that it will be useful, but
+ *   WITHOUT ANY WARRANTY; without even the implied warranty of
+ *   MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
+ *   NON INFRINGEMENT.  See the GNU General Public License for
+ *   more details.
+ */
+
+#ifndef _ASM_TILE_PCI_H
+#define _ASM_TILE_PCI_H
+
+#include <linux/dma-mapping.h>
+#include <linux/pci.h>
+#include <asm-generic/pci_iomap.h>
+
+#ifndef __tilegx__
+
+/*
+ * Structure of a PCI controller (host bridge)
+ */
+struct pci_controller {
+	int index;		/* PCI domain number */
+	struct pci_bus *root_bus;
+
+	int last_busno;
+
+	int hv_cfg_fd[2];	/* config{0,1} fds for this PCIe controller */
+	int hv_mem_fd;		/* fd to Hypervisor for MMIO operations */
+
+	struct pci_ops *ops;
+
+	int irq_base;		/* Base IRQ from the Hypervisor	*/
+	int plx_gen1;		/* flag for PLX Gen 1 configuration */
+
+	/* Address ranges that are routed to this controller/bridge. */
+	struct resource mem_resources[3];
+};
+
+/*
+ * This flag tells if the platform is TILEmpower that needs
+ * special configuration for the PLX switch chip.
+ */
+extern int tile_plx_gen1;
+
+static inline void pci_iounmap(struct pci_dev *dev, void __iomem *addr) {}
+
+#define	TILE_NUM_PCIE	2
+
+/*
+ * The hypervisor maps the entirety of CPA-space as bus addresses, so
+ * bus addresses are physical addresses.  The networking and block
+ * device layers use this boolean for bounce buffer decisions.
+ */
+#define PCI_DMA_BUS_IS_PHYS     1
+
+/* generic pci stuff */
+#include <asm-generic/pci.h>
+
+#else
+
+#include <asm/page.h>
+#include <gxio/trio.h>
+
+/**
+ * We reserve the hugepage-size address range at the top of the 64-bit address
+ * space to serve as the PCI window, emulating the BAR0 space of an endpoint
+ * device. This window is used by the chip-to-chip applications running on
+ * the RC node. The reason for carving out this window is that Mem-Maps that
+ * back up this window will not overlap with those that map the real physical
+ * memory.
+ */
+#define PCIE_HOST_BAR0_SIZE		HPAGE_SIZE
+#define PCIE_HOST_BAR0_START		HPAGE_MASK
+
+/**
+ * The first PAGE_SIZE of the above "BAR" window is mapped to the
+ * gxpci_host_regs structure.
+ */
+#define PCIE_HOST_REGS_SIZE		PAGE_SIZE
+
+/*
+ * This is the PCI address where the Mem-Map interrupt regions start.
+ * We use the 2nd to the last huge page of the 64-bit address space.
+ * The last huge page is used for the rootcomplex "bar", for C2C purpose.
+ */
+#define	MEM_MAP_INTR_REGIONS_BASE	(HPAGE_MASK - HPAGE_SIZE)
+
+/*
+ * Each Mem-Map interrupt region occupies 4KB.
+ */
+#define	MEM_MAP_INTR_REGION_SIZE	(1 << TRIO_MAP_MEM_LIM__ADDR_SHIFT)
+
+/*
+ * Allocate the PCI BAR window right below 4GB.
+ */
+#define	TILE_PCI_BAR_WINDOW_TOP		(1ULL << 32)
+
+/*
+ * Allocate 1GB for the PCI BAR window.
+ */
+#define	TILE_PCI_BAR_WINDOW_SIZE	(1 << 30)
+
+/*
+ * This is the highest bus address targeting the host memory that
+ * can be generated by legacy PCI devices with 32-bit or less
+ * DMA capability, dictated by the BAR window size and location.
+ */
+#define	TILE_PCI_MAX_DIRECT_DMA_ADDRESS \
+	(TILE_PCI_BAR_WINDOW_TOP - TILE_PCI_BAR_WINDOW_SIZE - 1)
+
+/*
+ * We shift the PCI bus range for all the physical memory up by the whole PA
+ * range. The corresponding CPA of an incoming PCI request will be the PCI
+ * address minus TILE_PCI_MEM_MAP_BASE_OFFSET. This also implies
+ * that the 64-bit capable devices will be given DMA addresses as
+ * the CPA plus TILE_PCI_MEM_MAP_BASE_OFFSET. To support 32-bit
+ * devices, we create a separate map region that handles the low
+ * 4GB.
+ *
+ * This design lets us avoid the "PCI hole" problem where the host bridge
+ * won't pass DMA traffic with target addresses that happen to fall within the
+ * BAR space. This enables us to use all the physical memory for DMA, instead
+ * of wasting the same amount of physical memory as the BAR window size.
+ */
+#define	TILE_PCI_MEM_MAP_BASE_OFFSET	(1ULL << CHIP_PA_WIDTH())
+
+/*
+ * Start of the PCI memory resource, which starts at the end of the
+ * maximum system physical RAM address.
+ */
+#define	TILE_PCI_MEM_START	(1ULL << CHIP_PA_WIDTH())
+
+/*
+ * Structure of a PCI controller (host bridge) on Gx.
+ */
+struct pci_controller {
+
+	/* Pointer back to the TRIO that this PCIe port is connected to. */
+	gxio_trio_context_t *trio;
+	int mac;		/* PCIe mac index on the TRIO shim */
+	int trio_index;		/* Index of TRIO shim that contains the MAC. */
+
+	int pio_mem_index;	/* PIO region index for memory access */
+
+#ifdef CONFIG_TILE_PCI_IO
+	int pio_io_index;	/* PIO region index for I/O space access */
+#endif
+
+	/*
+	 * Mem-Map regions for all the memory controllers so that Linux can
+	 * map all of its physical memory space to the PCI bus.
+	 */
+	int mem_maps[MAX_NUMNODES];
+
+	int index;		/* PCI domain number */
+	struct pci_bus *root_bus;
+
+	/* PCI I/O space resource for this controller. */
+	struct resource io_space;
+	char io_space_name[32];
+
+	/* PCI memory space resource for this controller. */
+	struct resource mem_space;
+	char mem_space_name[32];
+
+	uint64_t mem_offset;	/* cpu->bus memory mapping offset. */
+
+	int first_busno;
+
+	struct pci_ops *ops;
+
+	/* Table that maps the INTx numbers to Linux irq numbers. */
+	int irq_intx_table[4];
+};
+
+extern struct pci_controller pci_controllers[TILEGX_NUM_TRIO * TILEGX_TRIO_PCIES];
+extern gxio_trio_context_t trio_contexts[TILEGX_NUM_TRIO];
+extern int num_trio_shims;
+
+extern void pci_iounmap(struct pci_dev *dev, void __iomem *);
+
+/*
+ * The PCI address space does not equal the physical memory address
+ * space (we have an IOMMU). The IDE and SCSI device layers use this
+ * boolean for bounce buffer decisions.
+ */
+#define PCI_DMA_BUS_IS_PHYS     0
+
+#endif /* __tilegx__ */
+
+int __init tile_pci_init(void);
+int __init pcibios_init(void);
+
+void pcibios_fixup_bus(struct pci_bus *bus);
+
+#define pci_domain_nr(bus) (((struct pci_controller *)(bus)->sysdata)->index)
+
+/*
+ * This decides whether to display the domain number in /proc.
+ */
+static inline int pci_proc_domain(struct pci_bus *bus)
+{
+	return 1;
+}
+
+/*
+ * pcibios_assign_all_busses() tells whether or not the bus numbers
+ * should be reassigned, in case the BIOS didn't do it correctly, or
+ * in case we don't have a BIOS and we want to let Linux do it.
+ */
+static inline int pcibios_assign_all_busses(void)
+{
+	return 1;
+}
+
+#define PCIBIOS_MIN_MEM		0
+/* Minimum PCI I/O address, starting at the page boundary. */
+#define PCIBIOS_MIN_IO		PAGE_SIZE
+
+/* Use any cpu for PCI. */
+#define cpumask_of_pcibus(bus) cpu_online_mask
+
+/* implement the pci_ DMA API in terms of the generic device dma_ one */
+#include <asm-generic/pci-dma-compat.h>
+
+#endif /* _ASM_TILE_PCI_H */
diff --git a/kernel/arch/tile/include/asm/percpu.h b/kernel/arch/tile/include/asm/percpu.h
new file mode 100644
index 000000000..4f7ae39fa
--- /dev/null
+++ b/kernel/arch/tile/include/asm/percpu.h
@@ -0,0 +1,52 @@
+/*
+ * Copyright 2010 Tilera Corporation. All Rights Reserved.
+ *
+ *   This program is free software; you can redistribute it and/or
+ *   modify it under the terms of the GNU General Public License
+ *   as published by the Free Software Foundation, version 2.
+ *
+ *   This program is distributed in the hope that it will be useful, but
+ *   WITHOUT ANY WARRANTY; without even the implied warranty of
+ *   MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
+ *   NON INFRINGEMENT.  See the GNU General Public License for
+ *   more details.
+ */
+
+#ifndef _ASM_TILE_PERCPU_H
+#define _ASM_TILE_PERCPU_H
+
+register unsigned long my_cpu_offset_reg asm("tp");
+
+#ifdef CONFIG_PREEMPT
+/*
+ * For full preemption, we can't just use the register variable
+ * directly, since we need barrier() to hazard against it, causing the
+ * compiler to reload anything computed from a previous "tp" value.
+ * But we also don't want to use volatile asm, since we'd like the
+ * compiler to be able to cache the value across multiple percpu reads.
+ * So we use a fake stack read as a hazard against barrier().
+ * The 'U' constraint is like 'm' but disallows postincrement.
+ */
+static inline unsigned long __my_cpu_offset(void)
+{
+	unsigned long tp;
+	register unsigned long *sp asm("sp");
+	asm("move %0, tp" : "=r" (tp) : "U" (*sp));
+	return tp;
+}
+#define __my_cpu_offset __my_cpu_offset()
+#else
+/*
+ * We don't need to hazard against barrier() since "tp" doesn't ever
+ * change with PREEMPT_NONE, and with PREEMPT_VOLUNTARY it only
+ * changes at function call points, at which we are already re-reading
+ * the value of "tp" due to "my_cpu_offset_reg" being a global variable.
+ */
+#define __my_cpu_offset my_cpu_offset_reg
+#endif
+
+#define set_my_cpu_offset(tp) (my_cpu_offset_reg = (tp))
+
+#include <asm-generic/percpu.h>
+
+#endif /* _ASM_TILE_PERCPU_H */
diff --git a/kernel/arch/tile/include/asm/perf_event.h b/kernel/arch/tile/include/asm/perf_event.h
new file mode 100644
index 000000000..59c5b164e
--- /dev/null
+++ b/kernel/arch/tile/include/asm/perf_event.h
@@ -0,0 +1,22 @@
+/*
+ * Copyright 2014 Tilera Corporation. All Rights Reserved.
+ *
+ *   This program is free software; you can redistribute it and/or
+ *   modify it under the terms of the GNU General Public License
+ *   as published by the Free Software Foundation, version 2.
+ *
+ *   This program is distributed in the hope that it will be useful, but
+ *   WITHOUT ANY WARRANTY; without even the implied warranty of
+ *   MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
+ *   NON INFRINGEMENT.  See the GNU General Public License for
+ *   more details.
+ */
+
+#ifndef _ASM_TILE_PERF_EVENT_H
+#define _ASM_TILE_PERF_EVENT_H
+
+#include <linux/percpu.h>
+DECLARE_PER_CPU(u64, perf_irqs);
+
+unsigned long handle_syscall_link_address(void);
+#endif /* _ASM_TILE_PERF_EVENT_H */
diff --git a/kernel/arch/tile/include/asm/pgalloc.h b/kernel/arch/tile/include/asm/pgalloc.h
new file mode 100644
index 000000000..1b902508b
--- /dev/null
+++ b/kernel/arch/tile/include/asm/pgalloc.h
@@ -0,0 +1,164 @@
+/*
+ * Copyright 2010 Tilera Corporation. All Rights Reserved.
+ *
+ *   This program is free software; you can redistribute it and/or
+ *   modify it under the terms of the GNU General Public License
+ *   as published by the Free Software Foundation, version 2.
+ *
+ *   This program is distributed in the hope that it will be useful, but
+ *   WITHOUT ANY WARRANTY; without even the implied warranty of
+ *   MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
+ *   NON INFRINGEMENT.  See the GNU General Public License for
+ *   more details.
+ */
+
+#ifndef _ASM_TILE_PGALLOC_H
+#define _ASM_TILE_PGALLOC_H
+
+#include <linux/threads.h>
+#include <linux/mm.h>
+#include <linux/mmzone.h>
+#include <asm/fixmap.h>
+#include <asm/page.h>
+#include <hv/hypervisor.h>
+
+/* Bits for the size of the second-level page table. */
+#define L2_KERNEL_PGTABLE_SHIFT _HV_LOG2_L2_SIZE(HPAGE_SHIFT, PAGE_SHIFT)
+
+/* How big is a kernel L2 page table? */
+#define L2_KERNEL_PGTABLE_SIZE (1UL << L2_KERNEL_PGTABLE_SHIFT)
+
+/* We currently allocate user L2 page tables by page (unlike kernel L2s). */
+#if L2_KERNEL_PGTABLE_SHIFT < PAGE_SHIFT
+#define L2_USER_PGTABLE_SHIFT PAGE_SHIFT
+#else
+#define L2_USER_PGTABLE_SHIFT L2_KERNEL_PGTABLE_SHIFT
+#endif
+
+/* How many pages do we need, as an "order", for a user L2 page table? */
+#define L2_USER_PGTABLE_ORDER (L2_USER_PGTABLE_SHIFT - PAGE_SHIFT)
+
+static inline void set_pmd(pmd_t *pmdp, pmd_t pmd)
+{
+#ifdef CONFIG_64BIT
+	set_pte(pmdp, pmd);
+#else
+	set_pte(&pmdp->pud.pgd, pmd.pud.pgd);
+#endif
+}
+
+static inline void pmd_populate_kernel(struct mm_struct *mm,
+				       pmd_t *pmd, pte_t *ptep)
+{
+	set_pmd(pmd, ptfn_pmd(HV_CPA_TO_PTFN(__pa(ptep)),
+			      __pgprot(_PAGE_PRESENT)));
+}
+
+static inline void pmd_populate(struct mm_struct *mm, pmd_t *pmd,
+				pgtable_t page)
+{
+	set_pmd(pmd, ptfn_pmd(HV_CPA_TO_PTFN(PFN_PHYS(page_to_pfn(page))),
+			      __pgprot(_PAGE_PRESENT)));
+}
+
+/*
+ * Allocate and free page tables.
+ */
+
+extern pgd_t *pgd_alloc(struct mm_struct *mm);
+extern void pgd_free(struct mm_struct *mm, pgd_t *pgd);
+
+extern pgtable_t pgtable_alloc_one(struct mm_struct *mm, unsigned long address,
+				   int order);
+extern void pgtable_free(struct mm_struct *mm, struct page *pte, int order);
+
+static inline pgtable_t pte_alloc_one(struct mm_struct *mm,
+				      unsigned long address)
+{
+	return pgtable_alloc_one(mm, address, L2_USER_PGTABLE_ORDER);
+}
+
+static inline void pte_free(struct mm_struct *mm, struct page *pte)
+{
+	pgtable_free(mm, pte, L2_USER_PGTABLE_ORDER);
+}
+
+#define pmd_pgtable(pmd) pmd_page(pmd)
+
+static inline pte_t *
+pte_alloc_one_kernel(struct mm_struct *mm, unsigned long address)
+{
+	return pfn_to_kaddr(page_to_pfn(pte_alloc_one(mm, address)));
+}
+
+static inline void pte_free_kernel(struct mm_struct *mm, pte_t *pte)
+{
+	BUG_ON((unsigned long)pte & (PAGE_SIZE-1));
+	pte_free(mm, virt_to_page(pte));
+}
+
+extern void __pgtable_free_tlb(struct mmu_gather *tlb, struct page *pte,
+			       unsigned long address, int order);
+static inline void __pte_free_tlb(struct mmu_gather *tlb, struct page *pte,
+				  unsigned long address)
+{
+	__pgtable_free_tlb(tlb, pte, address, L2_USER_PGTABLE_ORDER);
+}
+
+#define check_pgt_cache()	do { } while (0)
+
+/*
+ * Get the small-page pte_t lowmem entry for a given pfn.
+ * This may or may not be in use, depending on whether the initial
+ * huge-page entry for the page has already been shattered.
+ */
+pte_t *get_prealloc_pte(unsigned long pfn);
+
+/* During init, we can shatter kernel huge pages if needed. */
+void shatter_pmd(pmd_t *pmd);
+
+/* After init, a more complex technique is required. */
+void shatter_huge_page(unsigned long addr);
+
+#ifdef __tilegx__
+
+#define pud_populate(mm, pud, pmd) \
+  pmd_populate_kernel((mm), (pmd_t *)(pud), (pte_t *)(pmd))
+
+/* Bits for the size of the L1 (intermediate) page table. */
+#define L1_KERNEL_PGTABLE_SHIFT _HV_LOG2_L1_SIZE(HPAGE_SHIFT)
+
+/* How big is a kernel L2 page table? */
+#define L1_KERNEL_PGTABLE_SIZE (1UL << L1_KERNEL_PGTABLE_SHIFT)
+
+/* We currently allocate L1 page tables by page. */
+#if L1_KERNEL_PGTABLE_SHIFT < PAGE_SHIFT
+#define L1_USER_PGTABLE_SHIFT PAGE_SHIFT
+#else
+#define L1_USER_PGTABLE_SHIFT L1_KERNEL_PGTABLE_SHIFT
+#endif
+
+/* How many pages do we need, as an "order", for an L1 page table? */
+#define L1_USER_PGTABLE_ORDER (L1_USER_PGTABLE_SHIFT - PAGE_SHIFT)
+
+static inline pmd_t *pmd_alloc_one(struct mm_struct *mm, unsigned long address)
+{
+	struct page *p = pgtable_alloc_one(mm, address, L1_USER_PGTABLE_ORDER);
+	return (pmd_t *)page_to_virt(p);
+}
+
+static inline void pmd_free(struct mm_struct *mm, pmd_t *pmdp)
+{
+	pgtable_free(mm, virt_to_page(pmdp), L1_USER_PGTABLE_ORDER);
+}
+
+static inline void __pmd_free_tlb(struct mmu_gather *tlb, pmd_t *pmdp,
+				  unsigned long address)
+{
+	__pgtable_free_tlb(tlb, virt_to_page(pmdp), address,
+			   L1_USER_PGTABLE_ORDER);
+}
+
+#endif /* __tilegx__ */
+
+#endif /* _ASM_TILE_PGALLOC_H */
diff --git a/kernel/arch/tile/include/asm/pgtable.h b/kernel/arch/tile/include/asm/pgtable.h
new file mode 100644
index 000000000..95a4f19d1
--- /dev/null
+++ b/kernel/arch/tile/include/asm/pgtable.h
@@ -0,0 +1,530 @@
+/*
+ * Copyright 2010 Tilera Corporation. All Rights Reserved.
+ *
+ *   This program is free software; you can redistribute it and/or
+ *   modify it under the terms of the GNU General Public License
+ *   as published by the Free Software Foundation, version 2.
+ *
+ *   This program is distributed in the hope that it will be useful, but
+ *   WITHOUT ANY WARRANTY; without even the implied warranty of
+ *   MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
+ *   NON INFRINGEMENT.  See the GNU General Public License for
+ *   more details.
+ *
+ * This file contains the functions and defines necessary to modify and use
+ * the TILE page table tree.
+ */
+
+#ifndef _ASM_TILE_PGTABLE_H
+#define _ASM_TILE_PGTABLE_H
+
+#include <hv/hypervisor.h>
+
+#ifndef __ASSEMBLY__
+
+#include <linux/bitops.h>
+#include <linux/threads.h>
+#include <linux/slab.h>
+#include <linux/list.h>
+#include <linux/spinlock.h>
+#include <linux/pfn.h>
+#include <asm/processor.h>
+#include <asm/fixmap.h>
+#include <asm/page.h>
+
+struct mm_struct;
+struct vm_area_struct;
+
+/*
+ * ZERO_PAGE is a global shared page that is always zero: used
+ * for zero-mapped memory areas etc..
+ */
+extern unsigned long empty_zero_page[PAGE_SIZE/sizeof(unsigned long)];
+#define ZERO_PAGE(vaddr) (virt_to_page(empty_zero_page))
+
+extern pgd_t swapper_pg_dir[];
+extern pgprot_t swapper_pgprot;
+extern struct kmem_cache *pgd_cache;
+extern spinlock_t pgd_lock;
+extern struct list_head pgd_list;
+
+/*
+ * The very last slots in the pgd_t are for addresses unusable by Linux
+ * (pgd_addr_invalid() returns true).  So we use them for the list structure.
+ * The x86 code we are modelled on uses the page->private/index fields
+ * (older 2.6 kernels) or the lru list (newer 2.6 kernels), but since
+ * our pgds are so much smaller than a page, it seems a waste to
+ * spend a whole page on each pgd.
+ */
+#define PGD_LIST_OFFSET \
+  ((PTRS_PER_PGD * sizeof(pgd_t)) - sizeof(struct list_head))
+#define pgd_to_list(pgd) \
+  ((struct list_head *)((char *)(pgd) + PGD_LIST_OFFSET))
+#define list_to_pgd(list) \
+  ((pgd_t *)((char *)(list) - PGD_LIST_OFFSET))
+
+extern void pgtable_cache_init(void);
+extern void paging_init(void);
+extern void set_page_homes(void);
+
+#define FIRST_USER_ADDRESS	0UL
+
+#define _PAGE_PRESENT           HV_PTE_PRESENT
+#define _PAGE_HUGE_PAGE         HV_PTE_PAGE
+#define _PAGE_SUPER_PAGE        HV_PTE_SUPER
+#define _PAGE_READABLE          HV_PTE_READABLE
+#define _PAGE_WRITABLE          HV_PTE_WRITABLE
+#define _PAGE_EXECUTABLE        HV_PTE_EXECUTABLE
+#define _PAGE_ACCESSED          HV_PTE_ACCESSED
+#define _PAGE_DIRTY             HV_PTE_DIRTY
+#define _PAGE_GLOBAL            HV_PTE_GLOBAL
+#define _PAGE_USER              HV_PTE_USER
+
+/*
+ * All the "standard" bits.  Cache-control bits are managed elsewhere.
+ * This is used to test for valid level-2 page table pointers by checking
+ * all the bits, and to mask away the cache control bits for mprotect.
+ */
+#define _PAGE_ALL (\
+  _PAGE_PRESENT | \
+  _PAGE_HUGE_PAGE | \
+  _PAGE_SUPER_PAGE | \
+  _PAGE_READABLE | \
+  _PAGE_WRITABLE | \
+  _PAGE_EXECUTABLE | \
+  _PAGE_ACCESSED | \
+  _PAGE_DIRTY | \
+  _PAGE_GLOBAL | \
+  _PAGE_USER \
+)
+
+#define PAGE_NONE \
+	__pgprot(_PAGE_PRESENT | _PAGE_ACCESSED)
+#define PAGE_SHARED \
+	__pgprot(_PAGE_PRESENT | _PAGE_READABLE | _PAGE_WRITABLE | \
+		 _PAGE_USER | _PAGE_ACCESSED)
+
+#define PAGE_SHARED_EXEC \
+	__pgprot(_PAGE_PRESENT | _PAGE_READABLE | _PAGE_WRITABLE | \
+		 _PAGE_EXECUTABLE | _PAGE_USER | _PAGE_ACCESSED)
+#define PAGE_COPY_NOEXEC \
+	__pgprot(_PAGE_PRESENT | _PAGE_USER | _PAGE_ACCESSED | _PAGE_READABLE)
+#define PAGE_COPY_EXEC \
+	__pgprot(_PAGE_PRESENT | _PAGE_USER | _PAGE_ACCESSED | \
+		 _PAGE_READABLE | _PAGE_EXECUTABLE)
+#define PAGE_COPY \
+	PAGE_COPY_NOEXEC
+#define PAGE_READONLY \
+	__pgprot(_PAGE_PRESENT | _PAGE_USER | _PAGE_ACCESSED | _PAGE_READABLE)
+#define PAGE_READONLY_EXEC \
+	__pgprot(_PAGE_PRESENT | _PAGE_USER | _PAGE_ACCESSED | \
+		 _PAGE_READABLE | _PAGE_EXECUTABLE)
+
+#define _PAGE_KERNEL_RO \
+ (_PAGE_PRESENT | _PAGE_GLOBAL | _PAGE_READABLE | _PAGE_ACCESSED)
+#define _PAGE_KERNEL \
+ (_PAGE_KERNEL_RO | _PAGE_WRITABLE | _PAGE_DIRTY)
+#define _PAGE_KERNEL_EXEC       (_PAGE_KERNEL_RO | _PAGE_EXECUTABLE)
+
+#define PAGE_KERNEL		__pgprot(_PAGE_KERNEL)
+#define PAGE_KERNEL_RO		__pgprot(_PAGE_KERNEL_RO)
+#define PAGE_KERNEL_EXEC	__pgprot(_PAGE_KERNEL_EXEC)
+
+#define page_to_kpgprot(p) PAGE_KERNEL
+
+/*
+ * We could tighten these up, but for now writable or executable
+ * implies readable.
+ */
+#define __P000	PAGE_NONE
+#define __P001	PAGE_READONLY
+#define __P010	PAGE_COPY      /* this is write-only, which we won't support */
+#define __P011	PAGE_COPY
+#define __P100	PAGE_READONLY_EXEC
+#define __P101	PAGE_READONLY_EXEC
+#define __P110	PAGE_COPY_EXEC
+#define __P111	PAGE_COPY_EXEC
+
+#define __S000	PAGE_NONE
+#define __S001	PAGE_READONLY
+#define __S010	PAGE_SHARED
+#define __S011	PAGE_SHARED
+#define __S100	PAGE_READONLY_EXEC
+#define __S101	PAGE_READONLY_EXEC
+#define __S110	PAGE_SHARED_EXEC
+#define __S111	PAGE_SHARED_EXEC
+
+/*
+ * All the normal _PAGE_ALL bits are ignored for PMDs, except PAGE_PRESENT
+ * and PAGE_HUGE_PAGE, which must be one and zero, respectively.
+ * We set the ignored bits to zero.
+ */
+#define _PAGE_TABLE     _PAGE_PRESENT
+
+/* Inherit the caching flags from the old protection bits. */
+#define pgprot_modify(oldprot, newprot) \
+  (pgprot_t) { ((oldprot).val & ~_PAGE_ALL) | (newprot).val }
+
+/* Just setting the PFN to zero suffices. */
+#define pte_pgprot(x) hv_pte_set_pa((x), 0)
+
+/*
+ * For PTEs and PDEs, we must clear the Present bit first when
+ * clearing a page table entry, so clear the bottom half first and
+ * enforce ordering with a barrier.
+ */
+static inline void __pte_clear(pte_t *ptep)
+{
+#ifdef __tilegx__
+	ptep->val = 0;
+#else
+	u32 *tmp = (u32 *)ptep;
+	tmp[0] = 0;
+	barrier();
+	tmp[1] = 0;
+#endif
+}
+#define pte_clear(mm, addr, ptep) __pte_clear(ptep)
+
+/*
+ * The following only work if pte_present() is true.
+ * Undefined behaviour if not..
+ */
+#define pte_present hv_pte_get_present
+#define pte_mknotpresent hv_pte_clear_present
+#define pte_user hv_pte_get_user
+#define pte_read hv_pte_get_readable
+#define pte_dirty hv_pte_get_dirty
+#define pte_young hv_pte_get_accessed
+#define pte_write hv_pte_get_writable
+#define pte_exec hv_pte_get_executable
+#define pte_huge hv_pte_get_page
+#define pte_super hv_pte_get_super
+#define pte_rdprotect hv_pte_clear_readable
+#define pte_exprotect hv_pte_clear_executable
+#define pte_mkclean hv_pte_clear_dirty
+#define pte_mkold hv_pte_clear_accessed
+#define pte_wrprotect hv_pte_clear_writable
+#define pte_mksmall hv_pte_clear_page
+#define pte_mkread hv_pte_set_readable
+#define pte_mkexec hv_pte_set_executable
+#define pte_mkdirty hv_pte_set_dirty
+#define pte_mkyoung hv_pte_set_accessed
+#define pte_mkwrite hv_pte_set_writable
+#define pte_mkhuge hv_pte_set_page
+#define pte_mksuper hv_pte_set_super
+
+#define pte_special(pte) 0
+#define pte_mkspecial(pte) (pte)
+
+/*
+ * Use some spare bits in the PTE for user-caching tags.
+ */
+#define pte_set_forcecache hv_pte_set_client0
+#define pte_get_forcecache hv_pte_get_client0
+#define pte_clear_forcecache hv_pte_clear_client0
+#define pte_set_anyhome hv_pte_set_client1
+#define pte_get_anyhome hv_pte_get_client1
+#define pte_clear_anyhome hv_pte_clear_client1
+
+/*
+ * A migrating PTE has PAGE_PRESENT clear but all the other bits preserved.
+ */
+#define pte_migrating hv_pte_get_migrating
+#define pte_mkmigrate(x) hv_pte_set_migrating(hv_pte_clear_present(x))
+#define pte_donemigrate(x) hv_pte_set_present(hv_pte_clear_migrating(x))
+
+#define pte_ERROR(e) \
+	pr_err("%s:%d: bad pte 0x%016llx\n", __FILE__, __LINE__, pte_val(e))
+#define pgd_ERROR(e) \
+	pr_err("%s:%d: bad pgd 0x%016llx\n", __FILE__, __LINE__, pgd_val(e))
+
+/* Return PA and protection info for a given kernel VA. */
+int va_to_cpa_and_pte(void *va, phys_addr_t *cpa, pte_t *pte);
+
+/*
+ * __set_pte() ensures we write the 64-bit PTE with 32-bit words in
+ * the right order on 32-bit platforms and also allows us to write
+ * hooks to check valid PTEs, etc., if we want.
+ */
+void __set_pte(pte_t *ptep, pte_t pte);
+
+/*
+ * set_pte() sets the given PTE and also sanity-checks the
+ * requested PTE against the page homecaching.  Unspecified parts
+ * of the PTE are filled in when it is written to memory, i.e. all
+ * caching attributes if "!forcecache", or the home cpu if "anyhome".
+ */
+extern void set_pte(pte_t *ptep, pte_t pte);
+#define set_pte_at(mm, addr, ptep, pteval) set_pte(ptep, pteval)
+#define set_pte_atomic(pteptr, pteval) set_pte(pteptr, pteval)
+
+#define pte_page(x)		pfn_to_page(pte_pfn(x))
+
+static inline int pte_none(pte_t pte)
+{
+	return !pte.val;
+}
+
+static inline unsigned long pte_pfn(pte_t pte)
+{
+	return PFN_DOWN(hv_pte_get_pa(pte));
+}
+
+/* Set or get the remote cache cpu in a pgprot with remote caching. */
+extern pgprot_t set_remote_cache_cpu(pgprot_t prot, int cpu);
+extern int get_remote_cache_cpu(pgprot_t prot);
+
+static inline pte_t pfn_pte(unsigned long pfn, pgprot_t prot)
+{
+	return hv_pte_set_pa(prot, PFN_PHYS(pfn));
+}
+
+/* Support for priority mappings. */
+extern void start_mm_caching(struct mm_struct *mm);
+extern void check_mm_caching(struct mm_struct *prev, struct mm_struct *next);
+
+/*
+ * Encode and de-code a swap entry (see <linux/swapops.h>).
+ * We put the swap file type+offset in the 32 high bits;
+ * I believe we can just leave the low bits clear.
+ */
+#define __swp_type(swp)		((swp).val & 0x1f)
+#define __swp_offset(swp)	((swp).val >> 5)
+#define __swp_entry(type, off)	((swp_entry_t) { (type) | ((off) << 5) })
+#define __pte_to_swp_entry(pte)	((swp_entry_t) { (pte).val >> 32 })
+#define __swp_entry_to_pte(swp)	((pte_t) { (((long long) ((swp).val)) << 32) })
+
+/*
+ * Conversion functions: convert a page and protection to a page entry,
+ * and a page entry and page directory to the page they refer to.
+ */
+
+#define mk_pte(page, pgprot)	pfn_pte(page_to_pfn(page), (pgprot))
+
+/*
+ * If we are doing an mprotect(), just accept the new vma->vm_page_prot
+ * value and combine it with the PFN from the old PTE to get a new PTE.
+ */
+static inline pte_t pte_modify(pte_t pte, pgprot_t newprot)
+{
+	return pfn_pte(pte_pfn(pte), newprot);
+}
+
+/*
+ * The pgd page can be thought of an array like this: pgd_t[PTRS_PER_PGD]
+ *
+ * This macro returns the index of the entry in the pgd page which would
+ * control the given virtual address.
+ */
+#define pgd_index(address) (((address) >> PGDIR_SHIFT) & (PTRS_PER_PGD - 1))
+
+/*
+ * pgd_offset() returns a (pgd_t *)
+ * pgd_index() is used get the offset into the pgd page's array of pgd_t's.
+ */
+#define pgd_offset(mm, address) ((mm)->pgd + pgd_index(address))
+
+/*
+ * A shortcut which implies the use of the kernel's pgd, instead
+ * of a process's.
+ */
+#define pgd_offset_k(address) pgd_offset(&init_mm, address)
+
+#define pte_offset_map(dir, address) pte_offset_kernel(dir, address)
+#define pte_unmap(pte) do { } while (0)
+
+/* Clear a non-executable kernel PTE and flush it from the TLB. */
+#define kpte_clear_flush(ptep, vaddr)		\
+do {						\
+	pte_clear(&init_mm, (vaddr), (ptep));	\
+	local_flush_tlb_page(FLUSH_NONEXEC, (vaddr), PAGE_SIZE); \
+} while (0)
+
+/*
+ * The kernel page tables contain what we need, and we flush when we
+ * change specific page table entries.
+ */
+#define update_mmu_cache(vma, address, pte) do { } while (0)
+
+#ifdef CONFIG_FLATMEM
+#define kern_addr_valid(addr)	(1)
+#endif /* CONFIG_FLATMEM */
+
+extern void vmalloc_sync_all(void);
+
+#endif /* !__ASSEMBLY__ */
+
+#ifdef __tilegx__
+#include <asm/pgtable_64.h>
+#else
+#include <asm/pgtable_32.h>
+#endif
+
+#ifndef __ASSEMBLY__
+
+static inline int pmd_none(pmd_t pmd)
+{
+	/*
+	 * Only check low word on 32-bit platforms, since it might be
+	 * out of sync with upper half.
+	 */
+	return (unsigned long)pmd_val(pmd) == 0;
+}
+
+static inline int pmd_present(pmd_t pmd)
+{
+	return pmd_val(pmd) & _PAGE_PRESENT;
+}
+
+static inline int pmd_bad(pmd_t pmd)
+{
+	return ((pmd_val(pmd) & _PAGE_ALL) != _PAGE_TABLE);
+}
+
+static inline unsigned long pages_to_mb(unsigned long npg)
+{
+	return npg >> (20 - PAGE_SHIFT);
+}
+
+/*
+ * The pmd can be thought of an array like this: pmd_t[PTRS_PER_PMD]
+ *
+ * This function returns the index of the entry in the pmd which would
+ * control the given virtual address.
+ */
+static inline unsigned long pmd_index(unsigned long address)
+{
+	return (address >> PMD_SHIFT) & (PTRS_PER_PMD - 1);
+}
+
+#define __HAVE_ARCH_PMDP_TEST_AND_CLEAR_YOUNG
+static inline int pmdp_test_and_clear_young(struct vm_area_struct *vma,
+					    unsigned long address,
+					    pmd_t *pmdp)
+{
+	return ptep_test_and_clear_young(vma, address, pmdp_ptep(pmdp));
+}
+
+#define __HAVE_ARCH_PMDP_SET_WRPROTECT
+static inline void pmdp_set_wrprotect(struct mm_struct *mm,
+				      unsigned long address, pmd_t *pmdp)
+{
+	ptep_set_wrprotect(mm, address, pmdp_ptep(pmdp));
+}
+
+
+#define __HAVE_ARCH_PMDP_GET_AND_CLEAR
+static inline pmd_t pmdp_get_and_clear(struct mm_struct *mm,
+				       unsigned long address,
+				       pmd_t *pmdp)
+{
+	return pte_pmd(ptep_get_and_clear(mm, address, pmdp_ptep(pmdp)));
+}
+
+static inline void __set_pmd(pmd_t *pmdp, pmd_t pmdval)
+{
+	set_pte(pmdp_ptep(pmdp), pmd_pte(pmdval));
+}
+
+#define set_pmd_at(mm, addr, pmdp, pmdval) __set_pmd(pmdp, pmdval)
+
+/* Create a pmd from a PTFN. */
+static inline pmd_t ptfn_pmd(unsigned long ptfn, pgprot_t prot)
+{
+	return pte_pmd(hv_pte_set_ptfn(prot, ptfn));
+}
+
+/* Return the page-table frame number (ptfn) that a pmd_t points at. */
+#define pmd_ptfn(pmd) hv_pte_get_ptfn(pmd_pte(pmd))
+
+/*
+ * A given kernel pmd_t maps to a specific virtual address (either a
+ * kernel huge page or a kernel pte_t table).  Since kernel pte_t
+ * tables can be aligned at sub-page granularity, this function can
+ * return non-page-aligned pointers, despite its name.
+ */
+static inline unsigned long pmd_page_vaddr(pmd_t pmd)
+{
+	phys_addr_t pa =
+		(phys_addr_t)pmd_ptfn(pmd) << HV_LOG2_PAGE_TABLE_ALIGN;
+	return (unsigned long)__va(pa);
+}
+
+/*
+ * A pmd_t points to the base of a huge page or to a pte_t array.
+ * If a pte_t array, since we can have multiple per page, we don't
+ * have a one-to-one mapping of pmd_t's to pages.  However, this is
+ * OK for pte_lockptr(), since we just end up with potentially one
+ * lock being used for several pte_t arrays.
+ */
+#define pmd_page(pmd) pfn_to_page(PFN_DOWN(HV_PTFN_TO_CPA(pmd_ptfn(pmd))))
+
+static inline void pmd_clear(pmd_t *pmdp)
+{
+	__pte_clear(pmdp_ptep(pmdp));
+}
+
+#define pmd_mknotpresent(pmd)	pte_pmd(pte_mknotpresent(pmd_pte(pmd)))
+#define pmd_young(pmd)		pte_young(pmd_pte(pmd))
+#define pmd_mkyoung(pmd)	pte_pmd(pte_mkyoung(pmd_pte(pmd)))
+#define pmd_mkold(pmd)		pte_pmd(pte_mkold(pmd_pte(pmd)))
+#define pmd_mkwrite(pmd)	pte_pmd(pte_mkwrite(pmd_pte(pmd)))
+#define pmd_write(pmd)		pte_write(pmd_pte(pmd))
+#define pmd_wrprotect(pmd)	pte_pmd(pte_wrprotect(pmd_pte(pmd)))
+#define pmd_mkdirty(pmd)	pte_pmd(pte_mkdirty(pmd_pte(pmd)))
+#define pmd_huge_page(pmd)	pte_huge(pmd_pte(pmd))
+#define pmd_mkhuge(pmd)		pte_pmd(pte_mkhuge(pmd_pte(pmd)))
+#define __HAVE_ARCH_PMD_WRITE
+
+#define pfn_pmd(pfn, pgprot)	pte_pmd(pfn_pte((pfn), (pgprot)))
+#define pmd_pfn(pmd)		pte_pfn(pmd_pte(pmd))
+#define mk_pmd(page, pgprot)	pfn_pmd(page_to_pfn(page), (pgprot))
+
+static inline pmd_t pmd_modify(pmd_t pmd, pgprot_t newprot)
+{
+	return pfn_pmd(pmd_pfn(pmd), newprot);
+}
+
+#ifdef CONFIG_TRANSPARENT_HUGEPAGE
+#define has_transparent_hugepage() 1
+#define pmd_trans_huge pmd_huge_page
+
+static inline pmd_t pmd_mksplitting(pmd_t pmd)
+{
+	return pte_pmd(hv_pte_set_client2(pmd_pte(pmd)));
+}
+
+static inline int pmd_trans_splitting(pmd_t pmd)
+{
+	return hv_pte_get_client2(pmd_pte(pmd));
+}
+#endif /* CONFIG_TRANSPARENT_HUGEPAGE */
+
+/*
+ * The pte page can be thought of an array like this: pte_t[PTRS_PER_PTE]
+ *
+ * This macro returns the index of the entry in the pte page which would
+ * control the given virtual address.
+ */
+static inline unsigned long pte_index(unsigned long address)
+{
+	return (address >> PAGE_SHIFT) & (PTRS_PER_PTE - 1);
+}
+
+static inline pte_t *pte_offset_kernel(pmd_t *pmd, unsigned long address)
+{
+       return (pte_t *)pmd_page_vaddr(*pmd) + pte_index(address);
+}
+
+#include <asm-generic/pgtable.h>
+
+/* Support /proc/NN/pgtable API. */
+struct seq_file;
+int arch_proc_pgtable_show(struct seq_file *m, struct mm_struct *mm,
+			   unsigned long vaddr, unsigned long pagesize,
+			   pte_t *ptep, void **datap);
+
+#endif /* !__ASSEMBLY__ */
+
+#endif /* _ASM_TILE_PGTABLE_H */
diff --git a/kernel/arch/tile/include/asm/pgtable_32.h b/kernel/arch/tile/include/asm/pgtable_32.h
new file mode 100644
index 000000000..d26a42279
--- /dev/null
+++ b/kernel/arch/tile/include/asm/pgtable_32.h
@@ -0,0 +1,121 @@
+/*
+ * Copyright 2010 Tilera Corporation. All Rights Reserved.
+ *
+ *   This program is free software; you can redistribute it and/or
+ *   modify it under the terms of the GNU General Public License
+ *   as published by the Free Software Foundation, version 2.
+ *
+ *   This program is distributed in the hope that it will be useful, but
+ *   WITHOUT ANY WARRANTY; without even the implied warranty of
+ *   MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
+ *   NON INFRINGEMENT.  See the GNU General Public License for
+ *   more details.
+ *
+ */
+
+#ifndef _ASM_TILE_PGTABLE_32_H
+#define _ASM_TILE_PGTABLE_32_H
+
+/*
+ * The level-1 index is defined by the huge page size.  A PGD is composed
+ * of PTRS_PER_PGD pgd_t's and is the top level of the page table.
+ */
+#define PGDIR_SHIFT	HPAGE_SHIFT
+#define PGDIR_SIZE	HPAGE_SIZE
+#define PGDIR_MASK	(~(PGDIR_SIZE-1))
+#define PTRS_PER_PGD	_HV_L1_ENTRIES(HPAGE_SHIFT)
+#define PGD_INDEX(va)	_HV_L1_INDEX(va, HPAGE_SHIFT)
+#define SIZEOF_PGD	_HV_L1_SIZE(HPAGE_SHIFT)
+
+/*
+ * The level-2 index is defined by the difference between the huge
+ * page size and the normal page size.  A PTE is composed of
+ * PTRS_PER_PTE pte_t's and is the bottom level of the page table.
+ * Note that the hypervisor docs use PTE for what we call pte_t, so
+ * this nomenclature is somewhat confusing.
+ */
+#define PTRS_PER_PTE	_HV_L2_ENTRIES(HPAGE_SHIFT, PAGE_SHIFT)
+#define PTE_INDEX(va)	_HV_L2_INDEX(va, HPAGE_SHIFT, PAGE_SHIFT)
+#define SIZEOF_PTE	_HV_L2_SIZE(HPAGE_SHIFT, PAGE_SHIFT)
+
+#ifndef __ASSEMBLY__
+
+/*
+ * Right now we initialize only a single pte table. It can be extended
+ * easily, subsequent pte tables have to be allocated in one physical
+ * chunk of RAM.
+ *
+ * HOWEVER, if we are using an allocation scheme with slop after the
+ * end of the page table (e.g. where our L2 page tables are 2KB but
+ * our pages are 64KB and we are allocating via the page allocator)
+ * we can't extend it easily.
+ */
+#define LAST_PKMAP PTRS_PER_PTE
+
+#define PKMAP_BASE   ((FIXADDR_BOOT_START - PAGE_SIZE*LAST_PKMAP) & PGDIR_MASK)
+
+#ifdef CONFIG_HIGHMEM
+# define _VMALLOC_END	(PKMAP_BASE & ~(HPAGE_SIZE-1))
+#else
+# define _VMALLOC_END	(FIXADDR_START & ~(HPAGE_SIZE-1))
+#endif
+
+/*
+ * Align the vmalloc area to an L2 page table, and leave a guard page
+ * at the beginning and end.  The vmalloc code also puts in an internal
+ * guard page between each allocation.
+ */
+#define VMALLOC_END	(_VMALLOC_END - PAGE_SIZE)
+extern unsigned long VMALLOC_RESERVE /* = CONFIG_VMALLOC_RESERVE */;
+#define _VMALLOC_START	(_VMALLOC_END - VMALLOC_RESERVE)
+#define VMALLOC_START	(_VMALLOC_START + PAGE_SIZE)
+
+/* This is the maximum possible amount of lowmem. */
+#define MAXMEM		(_VMALLOC_START - PAGE_OFFSET)
+
+/* We have no pmd or pud since we are strictly a two-level page table */
+#include <asm-generic/pgtable-nopmd.h>
+
+static inline int pud_huge_page(pud_t pud)	{ return 0; }
+
+/* We don't define any pgds for these addresses. */
+static inline int pgd_addr_invalid(unsigned long addr)
+{
+	return addr >= MEM_HV_START;
+}
+
+/*
+ * Provide versions of these routines that can be used safely when
+ * the hypervisor may be asynchronously modifying dirty/accessed bits.
+ * ptep_get_and_clear() matches the generic one but we provide it to
+ * be parallel with the 64-bit code.
+ */
+#define __HAVE_ARCH_PTEP_TEST_AND_CLEAR_YOUNG
+#define __HAVE_ARCH_PTEP_SET_WRPROTECT
+
+extern int ptep_test_and_clear_young(struct vm_area_struct *,
+				     unsigned long addr, pte_t *);
+extern void ptep_set_wrprotect(struct mm_struct *,
+			       unsigned long addr, pte_t *);
+
+#define __HAVE_ARCH_PTEP_GET_AND_CLEAR
+static inline pte_t ptep_get_and_clear(struct mm_struct *mm,
+				       unsigned long addr, pte_t *ptep)
+{
+	pte_t pte = *ptep;
+	pte_clear(_mm, addr, ptep);
+	return pte;
+}
+
+/*
+ * pmds are wrappers around pgds, which are the same as ptes.
+ * It's often convenient to "cast" back and forth and use the pte methods,
+ * which are the methods supplied by the hypervisor.
+ */
+#define pmd_pte(pmd) ((pmd).pud.pgd)
+#define pmdp_ptep(pmdp) (&(pmdp)->pud.pgd)
+#define pte_pmd(pte) ((pmd_t){ { (pte) } })
+
+#endif /* __ASSEMBLY__ */
+
+#endif /* _ASM_TILE_PGTABLE_32_H */
diff --git a/kernel/arch/tile/include/asm/pgtable_64.h b/kernel/arch/tile/include/asm/pgtable_64.h
new file mode 100644
index 000000000..e96cec52f
--- /dev/null
+++ b/kernel/arch/tile/include/asm/pgtable_64.h
@@ -0,0 +1,171 @@
+/*
+ * Copyright 2011 Tilera Corporation. All Rights Reserved.
+ *
+ *   This program is free software; you can redistribute it and/or
+ *   modify it under the terms of the GNU General Public License
+ *   as published by the Free Software Foundation, version 2.
+ *
+ *   This program is distributed in the hope that it will be useful, but
+ *   WITHOUT ANY WARRANTY; without even the implied warranty of
+ *   MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
+ *   NON INFRINGEMENT.  See the GNU General Public License for
+ *   more details.
+ *
+ */
+
+#ifndef _ASM_TILE_PGTABLE_64_H
+#define _ASM_TILE_PGTABLE_64_H
+
+/* The level-0 page table breaks the address space into 32-bit chunks. */
+#define PGDIR_SHIFT	HV_LOG2_L1_SPAN
+#define PGDIR_SIZE	HV_L1_SPAN
+#define PGDIR_MASK	(~(PGDIR_SIZE-1))
+#define PTRS_PER_PGD	HV_L0_ENTRIES
+#define PGD_INDEX(va)	HV_L0_INDEX(va)
+#define SIZEOF_PGD	HV_L0_SIZE
+
+/*
+ * The level-1 index is defined by the huge page size.  A PMD is composed
+ * of PTRS_PER_PMD pgd_t's and is the middle level of the page table.
+ */
+#define PMD_SHIFT	HPAGE_SHIFT
+#define PMD_SIZE	HPAGE_SIZE
+#define PMD_MASK	(~(PMD_SIZE-1))
+#define PTRS_PER_PMD	_HV_L1_ENTRIES(HPAGE_SHIFT)
+#define PMD_INDEX(va)	_HV_L1_INDEX(va, HPAGE_SHIFT)
+#define SIZEOF_PMD	_HV_L1_SIZE(HPAGE_SHIFT)
+
+/*
+ * The level-2 index is defined by the difference between the huge
+ * page size and the normal page size.  A PTE is composed of
+ * PTRS_PER_PTE pte_t's and is the bottom level of the page table.
+ * Note that the hypervisor docs use PTE for what we call pte_t, so
+ * this nomenclature is somewhat confusing.
+ */
+#define PTRS_PER_PTE	_HV_L2_ENTRIES(HPAGE_SHIFT, PAGE_SHIFT)
+#define PTE_INDEX(va)	_HV_L2_INDEX(va, HPAGE_SHIFT, PAGE_SHIFT)
+#define SIZEOF_PTE	_HV_L2_SIZE(HPAGE_SHIFT, PAGE_SHIFT)
+
+/*
+ * Align the vmalloc area to an L2 page table.  Omit guard pages at
+ * the beginning and end for simplicity (particularly in the per-cpu
+ * memory allocation code).  The vmalloc code puts in an internal
+ * guard page between each allocation.
+ */
+#define _VMALLOC_END	MEM_SV_START
+#define VMALLOC_END	_VMALLOC_END
+#define VMALLOC_START	_VMALLOC_START
+
+#ifndef __ASSEMBLY__
+
+/* We have no pud since we are a three-level page table. */
+#include <asm-generic/pgtable-nopud.h>
+
+/*
+ * pmds are the same as pgds and ptes, so converting is a no-op.
+ */
+#define pmd_pte(pmd) (pmd)
+#define pmdp_ptep(pmdp) (pmdp)
+#define pte_pmd(pte) (pte)
+
+#define pud_pte(pud) ((pud).pgd)
+
+static inline int pud_none(pud_t pud)
+{
+	return pud_val(pud) == 0;
+}
+
+static inline int pud_present(pud_t pud)
+{
+	return pud_val(pud) & _PAGE_PRESENT;
+}
+
+static inline int pud_huge_page(pud_t pud)
+{
+	return pud_val(pud) & _PAGE_HUGE_PAGE;
+}
+
+#define pmd_ERROR(e) \
+	pr_err("%s:%d: bad pmd 0x%016llx\n", __FILE__, __LINE__, pmd_val(e))
+
+static inline void pud_clear(pud_t *pudp)
+{
+	__pte_clear(&pudp->pgd);
+}
+
+static inline int pud_bad(pud_t pud)
+{
+	return ((pud_val(pud) & _PAGE_ALL) != _PAGE_TABLE);
+}
+
+/* Return the page-table frame number (ptfn) that a pud_t points at. */
+#define pud_ptfn(pud) hv_pte_get_ptfn((pud).pgd)
+
+/* Return the page frame number (pfn) that a pud_t points at. */
+#define pud_pfn(pud) pte_pfn(pud_pte(pud))
+
+/*
+ * A given kernel pud_t maps to a kernel pmd_t table at a specific
+ * virtual address.  Since kernel pmd_t tables can be aligned at
+ * sub-page granularity, this macro can return non-page-aligned
+ * pointers, despite its name.
+ */
+#define pud_page_vaddr(pud) \
+	(__va((phys_addr_t)pud_ptfn(pud) << HV_LOG2_PAGE_TABLE_ALIGN))
+
+/*
+ * A pud_t points to a pmd_t array.  Since we can have multiple per
+ * page, we don't have a one-to-one mapping of pud_t's to pages.
+ */
+#define pud_page(pud) pfn_to_page(PFN_DOWN(HV_PTFN_TO_CPA(pud_ptfn(pud))))
+
+static inline unsigned long pud_index(unsigned long address)
+{
+	return (address >> PUD_SHIFT) & (PTRS_PER_PUD - 1);
+}
+
+#define pmd_offset(pud, address) \
+	((pmd_t *)pud_page_vaddr(*(pud)) + pmd_index(address))
+
+/* Normalize an address to having the correct high bits set. */
+#define pgd_addr_normalize pgd_addr_normalize
+static inline unsigned long pgd_addr_normalize(unsigned long addr)
+{
+	return ((long)addr << (CHIP_WORD_SIZE() - CHIP_VA_WIDTH())) >>
+		(CHIP_WORD_SIZE() - CHIP_VA_WIDTH());
+}
+
+/* We don't define any pgds for these addresses. */
+static inline int pgd_addr_invalid(unsigned long addr)
+{
+	return addr >= KERNEL_HIGH_VADDR || addr != pgd_addr_normalize(addr);
+}
+
+/*
+ * Use atomic instructions to provide atomicity against the hypervisor.
+ */
+#define __HAVE_ARCH_PTEP_TEST_AND_CLEAR_YOUNG
+static inline int ptep_test_and_clear_young(struct vm_area_struct *vma,
+					    unsigned long addr, pte_t *ptep)
+{
+	return (__insn_fetchand(&ptep->val, ~HV_PTE_ACCESSED) >>
+		HV_PTE_INDEX_ACCESSED) & 0x1;
+}
+
+#define __HAVE_ARCH_PTEP_SET_WRPROTECT
+static inline void ptep_set_wrprotect(struct mm_struct *mm,
+				      unsigned long addr, pte_t *ptep)
+{
+	__insn_fetchand(&ptep->val, ~HV_PTE_WRITABLE);
+}
+
+#define __HAVE_ARCH_PTEP_GET_AND_CLEAR
+static inline pte_t ptep_get_and_clear(struct mm_struct *mm,
+				       unsigned long addr, pte_t *ptep)
+{
+	return hv_pte(__insn_exch(&ptep->val, 0UL));
+}
+
+#endif /* __ASSEMBLY__ */
+
+#endif /* _ASM_TILE_PGTABLE_64_H */
diff --git a/kernel/arch/tile/include/asm/pmc.h b/kernel/arch/tile/include/asm/pmc.h
new file mode 100644
index 000000000..7ae3956d9
--- /dev/null
+++ b/kernel/arch/tile/include/asm/pmc.h
@@ -0,0 +1,64 @@
+/*
+ * Copyright 2014 Tilera Corporation. All Rights Reserved.
+ *
+ *   This program is free software; you can redistribute it and/or
+ *   modify it under the terms of the GNU General Public License
+ *   as published by the Free Software Foundation, version 2.
+ *
+ *   This program is distributed in the hope that it will be useful, but
+ *   WITHOUT ANY WARRANTY; without even the implied warranty of
+ *   MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
+ *   NON INFRINGEMENT.  See the GNU General Public License for
+ *   more details.
+ */
+
+#ifndef _ASM_TILE_PMC_H
+#define _ASM_TILE_PMC_H
+
+#include <linux/ptrace.h>
+
+#define TILE_BASE_COUNTERS	2
+
+/* Bitfields below are derived from SPR PERF_COUNT_CTL*/
+#ifndef __tilegx__
+/* PERF_COUNT_CTL on TILEPro */
+#define TILE_CTL_EXCL_USER	(1 << 7) /* exclude user level */
+#define TILE_CTL_EXCL_KERNEL	(1 << 8) /* exclude kernel level */
+#define TILE_CTL_EXCL_HV	(1 << 9) /* exclude hypervisor level */
+
+#define TILE_SEL_MASK		0x7f	/* 7 bits for event SEL,
+					COUNT_0_SEL */
+#define TILE_PLM_MASK		0x780	/* 4 bits priv level msks,
+					COUNT_0_MASK*/
+#define TILE_EVENT_MASK	(TILE_SEL_MASK | TILE_PLM_MASK)
+
+#else /* __tilegx__*/
+/* PERF_COUNT_CTL on TILEGx*/
+#define TILE_CTL_EXCL_USER	(1 << 10) /* exclude user level */
+#define TILE_CTL_EXCL_KERNEL	(1 << 11) /* exclude kernel level */
+#define TILE_CTL_EXCL_HV	(1 << 12) /* exclude hypervisor level */
+
+#define TILE_SEL_MASK		0x3f	/* 6 bits for event SEL,
+					COUNT_0_SEL*/
+#define TILE_BOX_MASK		0x1c0	/* 3 bits box msks,
+					COUNT_0_BOX */
+#define TILE_PLM_MASK		0x3c00	/* 4 bits priv level msks,
+					COUNT_0_MASK */
+#define TILE_EVENT_MASK	(TILE_SEL_MASK | TILE_BOX_MASK | TILE_PLM_MASK)
+#endif /* __tilegx__*/
+
+/* Takes register and fault number.  Returns error to disable the interrupt. */
+typedef int (*perf_irq_t)(struct pt_regs *, int);
+
+int userspace_perf_handler(struct pt_regs *regs, int fault);
+
+perf_irq_t reserve_pmc_hardware(perf_irq_t new_perf_irq);
+void release_pmc_hardware(void);
+
+unsigned long pmc_get_overflow(void);
+void pmc_ack_overflow(unsigned long status);
+
+void unmask_pmc_interrupts(void);
+void mask_pmc_interrupts(void);
+
+#endif /* _ASM_TILE_PMC_H */
diff --git a/kernel/arch/tile/include/asm/processor.h b/kernel/arch/tile/include/asm/processor.h
new file mode 100644
index 000000000..dd4f9f17e
--- /dev/null
+++ b/kernel/arch/tile/include/asm/processor.h
@@ -0,0 +1,379 @@
+/*
+ * Copyright 2010 Tilera Corporation. All Rights Reserved.
+ *
+ *   This program is free software; you can redistribute it and/or
+ *   modify it under the terms of the GNU General Public License
+ *   as published by the Free Software Foundation, version 2.
+ *
+ *   This program is distributed in the hope that it will be useful, but
+ *   WITHOUT ANY WARRANTY; without even the implied warranty of
+ *   MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
+ *   NON INFRINGEMENT.  See the GNU General Public License for
+ *   more details.
+ */
+
+#ifndef _ASM_TILE_PROCESSOR_H
+#define _ASM_TILE_PROCESSOR_H
+
+#include <arch/chip.h>
+
+#ifndef __ASSEMBLY__
+
+/*
+ * NOTE: we don't include <linux/ptrace.h> or <linux/percpu.h> as one
+ * normally would, due to #include dependencies.
+ */
+#include <linux/types.h>
+#include <asm/ptrace.h>
+#include <asm/percpu.h>
+
+#include <arch/spr_def.h>
+
+struct task_struct;
+struct thread_struct;
+
+typedef struct {
+	unsigned long seg;
+} mm_segment_t;
+
+/*
+ * Default implementation of macro that returns current
+ * instruction pointer ("program counter").
+ */
+void *current_text_addr(void);
+
+#if CHIP_HAS_TILE_DMA()
+/* Capture the state of a suspended DMA. */
+struct tile_dma_state {
+	int enabled;
+	unsigned long src;
+	unsigned long dest;
+	unsigned long strides;
+	unsigned long chunk_size;
+	unsigned long src_chunk;
+	unsigned long dest_chunk;
+	unsigned long byte;
+	unsigned long status;
+};
+
+/*
+ * A mask of the DMA status register for selecting only the 'running'
+ * and 'done' bits.
+ */
+#define DMA_STATUS_MASK \
+  (SPR_DMA_STATUS__RUNNING_MASK | SPR_DMA_STATUS__DONE_MASK)
+#endif
+
+/*
+ * Track asynchronous TLB events (faults and access violations)
+ * that occur while we are in kernel mode from DMA or the SN processor.
+ */
+struct async_tlb {
+	short fault_num;         /* original fault number; 0 if none */
+	char is_fault;           /* was it a fault (vs an access violation) */
+	char is_write;           /* for fault: was it caused by a write? */
+	unsigned long address;   /* what address faulted? */
+};
+
+#ifdef CONFIG_HARDWALL
+struct hardwall_info;
+struct hardwall_task {
+	/* Which hardwall is this task tied to? (or NULL if none) */
+	struct hardwall_info *info;
+	/* Chains this task into the list at info->task_head. */
+	struct list_head list;
+};
+#ifdef __tilepro__
+#define HARDWALL_TYPES 1   /* udn */
+#else
+#define HARDWALL_TYPES 3   /* udn, idn, and ipi */
+#endif
+#endif
+
+struct thread_struct {
+	/* kernel stack pointer */
+	unsigned long  ksp;
+	/* kernel PC */
+	unsigned long  pc;
+	/* starting user stack pointer (for page migration) */
+	unsigned long  usp0;
+	/* pid of process that created this one */
+	pid_t creator_pid;
+#if CHIP_HAS_TILE_DMA()
+	/* DMA info for suspended threads (byte == 0 means no DMA state) */
+	struct tile_dma_state tile_dma_state;
+#endif
+	/* User EX_CONTEXT registers */
+	unsigned long ex_context[2];
+	/* User SYSTEM_SAVE registers */
+	unsigned long system_save[4];
+	/* User interrupt mask */
+	unsigned long long interrupt_mask;
+	/* User interrupt-control 0 state */
+	unsigned long intctrl_0;
+	/* Is this task currently doing a backtrace? */
+	bool in_backtrace;
+	/* Any other miscellaneous processor state bits */
+	unsigned long proc_status;
+#if !CHIP_HAS_FIXED_INTVEC_BASE()
+	/* Interrupt base for PL0 interrupts */
+	unsigned long interrupt_vector_base;
+#endif
+	/* Tile cache retry fifo high-water mark */
+	unsigned long tile_rtf_hwm;
+#if CHIP_HAS_DSTREAM_PF()
+	/* Data stream prefetch control */
+	unsigned long dstream_pf;
+#endif
+#ifdef CONFIG_HARDWALL
+	/* Hardwall information for various resources. */
+	struct hardwall_task hardwall[HARDWALL_TYPES];
+#endif
+#if CHIP_HAS_TILE_DMA()
+	/* Async DMA TLB fault information */
+	struct async_tlb dma_async_tlb;
+#endif
+};
+
+#endif /* !__ASSEMBLY__ */
+
+/*
+ * Start with "sp" this many bytes below the top of the kernel stack.
+ * This allows us to be cache-aware when handling the initial save
+ * of the pt_regs value to the stack.
+ */
+#define STACK_TOP_DELTA 64
+
+/*
+ * When entering the kernel via a fault, start with the top of the
+ * pt_regs structure this many bytes below the top of the page.
+ * This aligns the pt_regs structure optimally for cache-line access.
+ */
+#ifdef __tilegx__
+#define KSTK_PTREGS_GAP  48
+#else
+#define KSTK_PTREGS_GAP  56
+#endif
+
+#ifndef __ASSEMBLY__
+
+#ifdef __tilegx__
+#define TASK_SIZE_MAX		(_AC(1, UL) << (MAX_VA_WIDTH - 1))
+#else
+#define TASK_SIZE_MAX		PAGE_OFFSET
+#endif
+
+/* TASK_SIZE and related variables are always checked in "current" context. */
+#ifdef CONFIG_COMPAT
+#define COMPAT_TASK_SIZE	(1UL << 31)
+#define TASK_SIZE		((current_thread_info()->status & TS_COMPAT) ?\
+				 COMPAT_TASK_SIZE : TASK_SIZE_MAX)
+#else
+#define TASK_SIZE		TASK_SIZE_MAX
+#endif
+
+#define VDSO_BASE	((unsigned long)current->active_mm->context.vdso_base)
+#define VDSO_SYM(x)	(VDSO_BASE + (unsigned long)(x))
+
+#define STACK_TOP		TASK_SIZE
+
+/* STACK_TOP_MAX is used temporarily in execve and should not check COMPAT. */
+#define STACK_TOP_MAX		TASK_SIZE_MAX
+
+/*
+ * This decides where the kernel will search for a free chunk of vm
+ * space during mmap's, if it is using bottom-up mapping.
+ */
+#define TASK_UNMAPPED_BASE	(PAGE_ALIGN(TASK_SIZE / 3))
+
+#define HAVE_ARCH_PICK_MMAP_LAYOUT
+
+#define INIT_THREAD {                                                   \
+	.ksp = (unsigned long)init_stack + THREAD_SIZE - STACK_TOP_DELTA, \
+	.interrupt_mask = -1ULL                                         \
+}
+
+/* Kernel stack top for the task that first boots on this cpu. */
+DECLARE_PER_CPU(unsigned long, boot_sp);
+
+/* PC to boot from on this cpu. */
+DECLARE_PER_CPU(unsigned long, boot_pc);
+
+/* Do necessary setup to start up a newly executed thread. */
+static inline void start_thread(struct pt_regs *regs,
+				unsigned long pc, unsigned long usp)
+{
+	regs->pc = pc;
+	regs->sp = usp;
+	single_step_execve();
+}
+
+/* Free all resources held by a thread. */
+static inline void release_thread(struct task_struct *dead_task)
+{
+	/* Nothing for now */
+}
+
+extern int do_work_pending(struct pt_regs *regs, u32 flags);
+
+
+/*
+ * Return saved (kernel) PC of a blocked thread.
+ * Only used in a printk() in kernel/sched/core.c, so don't work too hard.
+ */
+#define thread_saved_pc(t)   ((t)->thread.pc)
+
+unsigned long get_wchan(struct task_struct *p);
+
+/* Return initial ksp value for given task. */
+#define task_ksp0(task) \
+	((unsigned long)(task)->stack + THREAD_SIZE - STACK_TOP_DELTA)
+
+/* Return some info about the user process TASK. */
+#define task_pt_regs(task) \
+	((struct pt_regs *)(task_ksp0(task) - KSTK_PTREGS_GAP) - 1)
+#define current_pt_regs()                                   \
+	((struct pt_regs *)((stack_pointer | (THREAD_SIZE - 1)) - \
+			    STACK_TOP_DELTA - (KSTK_PTREGS_GAP - 1)) - 1)
+#define task_sp(task)	(task_pt_regs(task)->sp)
+#define task_pc(task)	(task_pt_regs(task)->pc)
+/* Aliases for pc and sp (used in fs/proc/array.c) */
+#define KSTK_EIP(task)	task_pc(task)
+#define KSTK_ESP(task)	task_sp(task)
+
+/* Fine-grained unaligned JIT support */
+#define GET_UNALIGN_CTL(tsk, adr)	get_unalign_ctl((tsk), (adr))
+#define SET_UNALIGN_CTL(tsk, val)	set_unalign_ctl((tsk), (val))
+
+extern int get_unalign_ctl(struct task_struct *tsk, unsigned long adr);
+extern int set_unalign_ctl(struct task_struct *tsk, unsigned int val);
+
+/* Standard format for printing registers and other word-size data. */
+#ifdef __tilegx__
+# define REGFMT "0x%016lx"
+#else
+# define REGFMT "0x%08lx"
+#endif
+
+/*
+ * Do some slow action (e.g. read a slow SPR).
+ * Note that this must also have compiler-barrier semantics since
+ * it may be used in a busy loop reading memory.
+ */
+static inline void cpu_relax(void)
+{
+	__insn_mfspr(SPR_PASS);
+	barrier();
+}
+
+#define cpu_relax_lowlatency() cpu_relax()
+
+/* Info on this processor (see fs/proc/cpuinfo.c) */
+struct seq_operations;
+extern const struct seq_operations cpuinfo_op;
+
+/* Provide information about the chip model. */
+extern char chip_model[64];
+
+/* Data on which physical memory controller corresponds to which NUMA node. */
+extern int node_controller[];
+
+/* Does the heap allocator return hash-for-home pages by default? */
+extern int hash_default;
+
+/* Should kernel stack pages be hash-for-home? */
+extern int kstack_hash;
+
+/* Does MAP_ANONYMOUS return hash-for-home pages by default? */
+#define uheap_hash hash_default
+
+
+/* Are we using huge pages in the TLB for kernel data? */
+extern int kdata_huge;
+
+/* Support standard Linux prefetching. */
+#define ARCH_HAS_PREFETCH
+#define prefetch(x) __builtin_prefetch(x)
+#define PREFETCH_STRIDE CHIP_L2_LINE_SIZE()
+
+/* Bring a value into the L1D, faulting the TLB if necessary. */
+#ifdef __tilegx__
+#define prefetch_L1(x) __insn_prefetch_l1_fault((void *)(x))
+#else
+#define prefetch_L1(x) __insn_prefetch_L1((void *)(x))
+#endif
+
+#else /* __ASSEMBLY__ */
+
+/* Do some slow action (e.g. read a slow SPR). */
+#define CPU_RELAX       mfspr zero, SPR_PASS
+
+#endif /* !__ASSEMBLY__ */
+
+/* Assembly code assumes that the PL is in the low bits. */
+#if SPR_EX_CONTEXT_1_1__PL_SHIFT != 0
+# error Fix assembly assumptions about PL
+#endif
+
+/* We sometimes use these macros for EX_CONTEXT_0_1 as well. */
+#if SPR_EX_CONTEXT_1_1__PL_SHIFT != SPR_EX_CONTEXT_0_1__PL_SHIFT || \
+    SPR_EX_CONTEXT_1_1__PL_RMASK != SPR_EX_CONTEXT_0_1__PL_RMASK || \
+    SPR_EX_CONTEXT_1_1__ICS_SHIFT != SPR_EX_CONTEXT_0_1__ICS_SHIFT || \
+    SPR_EX_CONTEXT_1_1__ICS_RMASK != SPR_EX_CONTEXT_0_1__ICS_RMASK
+# error Fix assumptions that EX1 macros work for both PL0 and PL1
+#endif
+
+/* Allow pulling apart and recombining the PL and ICS bits in EX_CONTEXT. */
+#define EX1_PL(ex1) \
+  (((ex1) >> SPR_EX_CONTEXT_1_1__PL_SHIFT) & SPR_EX_CONTEXT_1_1__PL_RMASK)
+#define EX1_ICS(ex1) \
+  (((ex1) >> SPR_EX_CONTEXT_1_1__ICS_SHIFT) & SPR_EX_CONTEXT_1_1__ICS_RMASK)
+#define PL_ICS_EX1(pl, ics) \
+  (((pl) << SPR_EX_CONTEXT_1_1__PL_SHIFT) | \
+   ((ics) << SPR_EX_CONTEXT_1_1__ICS_SHIFT))
+
+/*
+ * Provide symbolic constants for PLs.
+ */
+#define USER_PL 0
+#if CONFIG_KERNEL_PL == 2
+#define GUEST_PL 1
+#endif
+#define KERNEL_PL CONFIG_KERNEL_PL
+
+/* SYSTEM_SAVE_K_0 holds the current cpu number ORed with ksp0. */
+#ifdef __tilegx__
+#define CPU_SHIFT 48
+#if CHIP_VA_WIDTH() > CPU_SHIFT
+# error Too many VA bits!
+#endif
+#define MAX_CPU_ID ((1 << (64 - CPU_SHIFT)) - 1)
+#define raw_smp_processor_id() \
+	((int)(__insn_mfspr(SPR_SYSTEM_SAVE_K_0) >> CPU_SHIFT))
+#define get_current_ksp0() \
+	((unsigned long)(((long)__insn_mfspr(SPR_SYSTEM_SAVE_K_0) << \
+			  (64 - CPU_SHIFT)) >> (64 - CPU_SHIFT)))
+#define next_current_ksp0(task) ({ \
+	unsigned long __ksp0 = task_ksp0(task) & ((1UL << CPU_SHIFT) - 1); \
+	unsigned long __cpu = (long)raw_smp_processor_id() << CPU_SHIFT; \
+	__ksp0 | __cpu; \
+})
+#else
+#define LOG2_NR_CPU_IDS 6
+#define MAX_CPU_ID ((1 << LOG2_NR_CPU_IDS) - 1)
+#define raw_smp_processor_id() \
+	((int)__insn_mfspr(SPR_SYSTEM_SAVE_K_0) & MAX_CPU_ID)
+#define get_current_ksp0() \
+	(__insn_mfspr(SPR_SYSTEM_SAVE_K_0) & ~MAX_CPU_ID)
+#define next_current_ksp0(task) ({ \
+	unsigned long __ksp0 = task_ksp0(task); \
+	int __cpu = raw_smp_processor_id(); \
+	BUG_ON(__ksp0 & MAX_CPU_ID); \
+	__ksp0 | __cpu; \
+})
+#endif
+#if CONFIG_NR_CPUS > (MAX_CPU_ID + 1)
+# error Too many cpus!
+#endif
+
+#endif /* _ASM_TILE_PROCESSOR_H */
diff --git a/kernel/arch/tile/include/asm/ptrace.h b/kernel/arch/tile/include/asm/ptrace.h
new file mode 100644
index 000000000..b9620c077
--- /dev/null
+++ b/kernel/arch/tile/include/asm/ptrace.h
@@ -0,0 +1,97 @@
+/*
+ * Copyright 2010 Tilera Corporation. All Rights Reserved.
+ *
+ *   This program is free software; you can redistribute it and/or
+ *   modify it under the terms of the GNU General Public License
+ *   as published by the Free Software Foundation, version 2.
+ *
+ *   This program is distributed in the hope that it will be useful, but
+ *   WITHOUT ANY WARRANTY; without even the implied warranty of
+ *   MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
+ *   NON INFRINGEMENT.  See the GNU General Public License for
+ *   more details.
+ */
+#ifndef _ASM_TILE_PTRACE_H
+#define _ASM_TILE_PTRACE_H
+
+#include <linux/compiler.h>
+
+#ifndef __ASSEMBLY__
+/* Benefit from consistent use of "long" on all chips. */
+typedef unsigned long pt_reg_t;
+#endif
+
+#include <uapi/asm/ptrace.h>
+
+#define PTRACE_O_MASK_TILE	(PTRACE_O_TRACEMIGRATE)
+#define PT_TRACE_MIGRATE	PT_EVENT_FLAG(PTRACE_EVENT_MIGRATE)
+
+/* Flag bits in pt_regs.flags */
+#define PT_FLAGS_DISABLE_IRQ    1  /* on return to kernel, disable irqs */
+#define PT_FLAGS_CALLER_SAVES   2  /* caller-save registers are valid */
+#define PT_FLAGS_RESTORE_REGS   4  /* restore callee-save regs on return */
+
+#ifndef __ASSEMBLY__
+
+#define regs_return_value(regs) ((regs)->regs[0])
+#define instruction_pointer(regs) ((regs)->pc)
+#define profile_pc(regs) instruction_pointer(regs)
+#define user_stack_pointer(regs) ((regs)->sp)
+
+/* Does the process account for user or for system time? */
+#define user_mode(regs) (EX1_PL((regs)->ex1) < KERNEL_PL)
+
+/* Fill in a struct pt_regs with the current kernel registers. */
+struct pt_regs *get_pt_regs(struct pt_regs *);
+
+/* Trace the current syscall. */
+extern int do_syscall_trace_enter(struct pt_regs *regs);
+extern void do_syscall_trace_exit(struct pt_regs *regs);
+
+#define arch_has_single_step()	(1)
+
+/*
+ * A structure for all single-stepper state.
+ *
+ * Also update defines in assembler section if it changes
+ */
+struct single_step_state {
+	/* the page to which we will write hacked-up bundles */
+	void __user *buffer;
+
+	union {
+		int flags;
+		struct {
+			unsigned long is_enabled:1, update:1, update_reg:6;
+		};
+	};
+
+	unsigned long orig_pc;		/* the original PC */
+	unsigned long next_pc;		/* return PC if no branch (PC + 1) */
+	unsigned long branch_next_pc;	/* return PC if we did branch/jump */
+	unsigned long update_value;	/* value to restore to update_target */
+};
+
+/* Single-step the instruction at regs->pc */
+extern void single_step_once(struct pt_regs *regs);
+
+/* Clean up after execve(). */
+extern void single_step_execve(void);
+
+struct task_struct;
+
+extern void send_sigtrap(struct task_struct *tsk, struct pt_regs *regs);
+
+#ifdef __tilegx__
+/* We need this since sigval_t has a user pointer in it, for GETSIGINFO etc. */
+#define __ARCH_WANT_COMPAT_SYS_PTRACE
+#endif
+
+#endif /* !__ASSEMBLY__ */
+
+#define SINGLESTEP_STATE_MASK_IS_ENABLED      0x1
+#define SINGLESTEP_STATE_MASK_UPDATE          0x2
+#define SINGLESTEP_STATE_TARGET_LB              2
+#define SINGLESTEP_STATE_TARGET_UB              7
+
+#endif /* _ASM_TILE_PTRACE_H */
diff --git a/kernel/arch/tile/include/asm/sections.h b/kernel/arch/tile/include/asm/sections.h
new file mode 100644
index 000000000..86a746243
--- /dev/null
+++ b/kernel/arch/tile/include/asm/sections.h
@@ -0,0 +1,47 @@
+/*
+ * Copyright 2010 Tilera Corporation. All Rights Reserved.
+ *
+ *   This program is free software; you can redistribute it and/or
+ *   modify it under the terms of the GNU General Public License
+ *   as published by the Free Software Foundation, version 2.
+ *
+ *   This program is distributed in the hope that it will be useful, but
+ *   WITHOUT ANY WARRANTY; without even the implied warranty of
+ *   MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
+ *   NON INFRINGEMENT.  See the GNU General Public License for
+ *   more details.
+ */
+
+#ifndef _ASM_TILE_SECTIONS_H
+#define _ASM_TILE_SECTIONS_H
+
+#define arch_is_kernel_data arch_is_kernel_data
+
+#include <asm-generic/sections.h>
+
+/* Write-once data is writable only till the end of initialization. */
+extern char __w1data_begin[], __w1data_end[];
+
+extern char vdso_start[], vdso_end[];
+#ifdef CONFIG_COMPAT
+extern char vdso32_start[], vdso32_end[];
+#endif
+
+/* Not exactly sections, but PC comparison points in the code. */
+extern char __rt_sigreturn[], __rt_sigreturn_end[];
+#ifdef __tilegx__
+extern char __start_unalign_asm_code[], __end_unalign_asm_code[];
+#else
+extern char sys_cmpxchg[], __sys_cmpxchg_end[];
+extern char __sys_cmpxchg_grab_lock[];
+extern char __start_atomic_asm_code[], __end_atomic_asm_code[];
+#endif
+
+/* Handle the discontiguity between _sdata and _text. */
+static inline int arch_is_kernel_data(unsigned long addr)
+{
+	return addr >= (unsigned long)_sdata &&
+		addr < (unsigned long)_end;
+}
+
+#endif /* _ASM_TILE_SECTIONS_H */
diff --git a/kernel/arch/tile/include/asm/setup.h b/kernel/arch/tile/include/asm/setup.h
new file mode 100644
index 000000000..e98909033
--- /dev/null
+++ b/kernel/arch/tile/include/asm/setup.h
@@ -0,0 +1,52 @@
+/*
+ * Copyright 2010 Tilera Corporation. All Rights Reserved.
+ *
+ *   This program is free software; you can redistribute it and/or
+ *   modify it under the terms of the GNU General Public License
+ *   as published by the Free Software Foundation, version 2.
+ *
+ *   This program is distributed in the hope that it will be useful, but
+ *   WITHOUT ANY WARRANTY; without even the implied warranty of
+ *   MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
+ *   NON INFRINGEMENT.  See the GNU General Public License for
+ *   more details.
+ */
+#ifndef _ASM_TILE_SETUP_H
+#define _ASM_TILE_SETUP_H
+
+
+#include <linux/pfn.h>
+#include <linux/init.h>
+#include <uapi/asm/setup.h>
+
+/*
+ * Reserved space for vmalloc and iomap - defined in asm/page.h
+ */
+#define MAXMEM_PFN	PFN_DOWN(MAXMEM)
+
+int tile_console_write(const char *buf, int count);
+void early_panic(const char *fmt, ...);
+
+/* Init-time routine to do tile-specific per-cpu setup. */
+void setup_cpu(int boot);
+
+/* User-level DMA management functions */
+void grant_dma_mpls(void);
+void restrict_dma_mpls(void);
+
+#ifdef CONFIG_HARDWALL
+/* User-level network management functions */
+void reset_network_state(void);
+struct task_struct;
+void hardwall_switch_tasks(struct task_struct *prev, struct task_struct *next);
+void hardwall_deactivate_all(struct task_struct *task);
+int hardwall_ipi_valid(int cpu);
+
+/* Hook hardwall code into changes in affinity. */
+#define arch_set_cpus_allowed(p, new_mask) do { \
+	if (!cpumask_equal(&p->cpus_allowed, new_mask)) \
+		hardwall_deactivate_all(p); \
+} while (0)
+#endif
+
+#endif /* _ASM_TILE_SETUP_H */
diff --git a/kernel/arch/tile/include/asm/sigframe.h b/kernel/arch/tile/include/asm/sigframe.h
new file mode 100644
index 000000000..994d3d302
--- /dev/null
+++ b/kernel/arch/tile/include/asm/sigframe.h
@@ -0,0 +1,33 @@
+/*
+ * Copyright 2010 Tilera Corporation. All Rights Reserved.
+ *
+ *   This program is free software; you can redistribute it and/or
+ *   modify it under the terms of the GNU General Public License
+ *   as published by the Free Software Foundation, version 2.
+ *
+ *   This program is distributed in the hope that it will be useful, but
+ *   WITHOUT ANY WARRANTY; without even the implied warranty of
+ *   MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
+ *   NON INFRINGEMENT.  See the GNU General Public License for
+ *   more details.
+ */
+
+#ifndef _ASM_TILE_SIGFRAME_H
+#define _ASM_TILE_SIGFRAME_H
+
+/* Indicate that syscall return should not examine r0 */
+#define INT_SWINT_1_SIGRETURN (~0)
+
+#ifndef __ASSEMBLY__
+
+#include <arch/abi.h>
+
+struct rt_sigframe {
+	unsigned char save_area[C_ABI_SAVE_AREA_SIZE]; /* caller save area */
+	struct siginfo info;
+	struct ucontext uc;
+};
+
+#endif /* !__ASSEMBLY__ */
+
+#endif /* _ASM_TILE_SIGFRAME_H */
diff --git a/kernel/arch/tile/include/asm/signal.h b/kernel/arch/tile/include/asm/signal.h
new file mode 100644
index 000000000..10e183de9
--- /dev/null
+++ b/kernel/arch/tile/include/asm/signal.h
@@ -0,0 +1,29 @@
+/*
+ * Copyright 2010 Tilera Corporation. All Rights Reserved.
+ *
+ *   This program is free software; you can redistribute it and/or
+ *   modify it under the terms of the GNU General Public License
+ *   as published by the Free Software Foundation, version 2.
+ *
+ *   This program is distributed in the hope that it will be useful, but
+ *   WITHOUT ANY WARRANTY; without even the implied warranty of
+ *   MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
+ *   NON INFRINGEMENT.  See the GNU General Public License for
+ *   more details.
+ */
+#ifndef _ASM_TILE_SIGNAL_H
+#define _ASM_TILE_SIGNAL_H
+
+#include <uapi/asm/signal.h>
+
+#if !defined(__ASSEMBLY__)
+struct pt_regs;
+int restore_sigcontext(struct pt_regs *, struct sigcontext __user *);
+int setup_sigcontext(struct sigcontext __user *, struct pt_regs *);
+void do_signal(struct pt_regs *regs);
+void signal_fault(const char *type, struct pt_regs *,
+		  void __user *frame, int sig);
+void trace_unhandled_signal(const char *type, struct pt_regs *regs,
+			    unsigned long address, int signo);
+#endif
+#endif /* _ASM_TILE_SIGNAL_H */
diff --git a/kernel/arch/tile/include/asm/smp.h b/kernel/arch/tile/include/asm/smp.h
new file mode 100644
index 000000000..735e7f144
--- /dev/null
+++ b/kernel/arch/tile/include/asm/smp.h
@@ -0,0 +1,139 @@
+/*
+ * Copyright 2010 Tilera Corporation. All Rights Reserved.
+ *
+ *   This program is free software; you can redistribute it and/or
+ *   modify it under the terms of the GNU General Public License
+ *   as published by the Free Software Foundation, version 2.
+ *
+ *   This program is distributed in the hope that it will be useful, but
+ *   WITHOUT ANY WARRANTY; without even the implied warranty of
+ *   MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
+ *   NON INFRINGEMENT.  See the GNU General Public License for
+ *   more details.
+ */
+
+#ifndef _ASM_TILE_SMP_H
+#define _ASM_TILE_SMP_H
+
+#ifdef CONFIG_SMP
+
+#include <asm/processor.h>
+#include <linux/cpumask.h>
+#include <linux/irqreturn.h>
+#include <hv/hypervisor.h>
+
+/* Set up this tile to support receiving hypervisor messages */
+void init_messaging(void);
+
+/* Set up this tile to support receiving device interrupts and IPIs. */
+void init_per_tile_IRQs(void);
+
+/* Send a message to processors specified in mask */
+void send_IPI_many(const struct cpumask *mask, int tag);
+
+/* Send a message to all but the sending processor */
+void send_IPI_allbutself(int tag);
+
+/* Send a message to a specific processor */
+void send_IPI_single(int dest, int tag);
+
+/* Process an IPI message */
+void evaluate_message(int tag);
+
+/* Boot a secondary cpu */
+void online_secondary(void);
+
+/* Topology of the supervisor tile grid, and coordinates of boot processor */
+extern HV_Topology smp_topology;
+
+/* Accessors for grid size */
+#define smp_height		(smp_topology.height)
+#define smp_width		(smp_topology.width)
+
+/* Convenience functions for converting cpu <-> coords. */
+static inline int cpu_x(int cpu)
+{
+	return cpu % smp_width;
+}
+static inline int cpu_y(int cpu)
+{
+	return cpu / smp_width;
+}
+static inline int xy_to_cpu(int x, int y)
+{
+	return y * smp_width + x;
+}
+
+/* Hypervisor message tags sent via the tile send_IPI*() routines. */
+#define MSG_TAG_START_CPU		1
+#define MSG_TAG_STOP_CPU		2
+#define MSG_TAG_CALL_FUNCTION_MANY	3
+#define MSG_TAG_CALL_FUNCTION_SINGLE	4
+#define MSG_TAG_IRQ_WORK		5
+
+/* Hook for the generic smp_call_function_many() routine. */
+static inline void arch_send_call_function_ipi_mask(struct cpumask *mask)
+{
+	send_IPI_many(mask, MSG_TAG_CALL_FUNCTION_MANY);
+}
+
+/* Hook for the generic smp_call_function_single() routine. */
+static inline void arch_send_call_function_single_ipi(int cpu)
+{
+	send_IPI_single(cpu, MSG_TAG_CALL_FUNCTION_SINGLE);
+}
+
+/* Print out the boot string describing which cpus were disabled. */
+void print_disabled_cpus(void);
+
+#else /* !CONFIG_SMP */
+
+#define smp_master_cpu		0
+#define smp_height		1
+#define smp_width		1
+#define cpu_x(cpu)		0
+#define cpu_y(cpu)		0
+#define xy_to_cpu(x, y)		0
+
+#endif /* !CONFIG_SMP */
+
+
+/* Which cpus may be used as the lotar in a page table entry. */
+extern struct cpumask cpu_lotar_map;
+#define cpu_is_valid_lotar(cpu) cpumask_test_cpu((cpu), &cpu_lotar_map)
+
+/* Which processors are used for hash-for-home mapping */
+extern struct cpumask hash_for_home_map;
+
+/* Which cpus can have their cache flushed by hv_flush_remote(). */
+extern struct cpumask cpu_cacheable_map;
+#define cpu_cacheable(cpu) cpumask_test_cpu((cpu), &cpu_cacheable_map)
+
+/* Convert an HV_LOTAR value into a cpu. */
+static inline int hv_lotar_to_cpu(HV_LOTAR lotar)
+{
+	return HV_LOTAR_X(lotar) + (HV_LOTAR_Y(lotar) * smp_width);
+}
+
+/*
+ * Extension of <linux/cpumask.h> functionality when you just want
+ * to express a mask or suppression or inclusion region without
+ * being too concerned about exactly which cpus are valid in that region.
+ */
+int bitmap_parselist_crop(const char *bp, unsigned long *maskp, int nmaskbits);
+
+#define cpulist_parse_crop(buf, dst) \
+			__cpulist_parse_crop((buf), (dst), NR_CPUS)
+static inline int __cpulist_parse_crop(const char *buf, struct cpumask *dstp,
+					int nbits)
+{
+	return bitmap_parselist_crop(buf, cpumask_bits(dstp), nbits);
+}
+
+/* Initialize the IPI subsystem. */
+void ipi_init(void);
+
+/* Function for start-cpu message to cause us to jump to. */
+extern unsigned long start_cpu_function_addr;
+
+#endif /* _ASM_TILE_SMP_H */
diff --git a/kernel/arch/tile/include/asm/spinlock.h b/kernel/arch/tile/include/asm/spinlock.h
new file mode 100644
index 000000000..1a8bd4740
--- /dev/null
+++ b/kernel/arch/tile/include/asm/spinlock.h
@@ -0,0 +1,24 @@
+/*
+ * Copyright 2010 Tilera Corporation. All Rights Reserved.
+ *
+ *   This program is free software; you can redistribute it and/or
+ *   modify it under the terms of the GNU General Public License
+ *   as published by the Free Software Foundation, version 2.
+ *
+ *   This program is distributed in the hope that it will be useful, but
+ *   WITHOUT ANY WARRANTY; without even the implied warranty of
+ *   MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
+ *   NON INFRINGEMENT.  See the GNU General Public License for
+ *   more details.
+ */
+
+#ifndef _ASM_TILE_SPINLOCK_H
+#define _ASM_TILE_SPINLOCK_H
+
+#ifdef __tilegx__
+#include <asm/spinlock_64.h>
+#else
+#include <asm/spinlock_32.h>
+#endif
+
+#endif /* _ASM_TILE_SPINLOCK_H */
diff --git a/kernel/arch/tile/include/asm/spinlock_32.h b/kernel/arch/tile/include/asm/spinlock_32.h
new file mode 100644
index 000000000..c0a77b38d
--- /dev/null
+++ b/kernel/arch/tile/include/asm/spinlock_32.h
@@ -0,0 +1,129 @@
+/*
+ * Copyright 2010 Tilera Corporation. All Rights Reserved.
+ *
+ *   This program is free software; you can redistribute it and/or
+ *   modify it under the terms of the GNU General Public License
+ *   as published by the Free Software Foundation, version 2.
+ *
+ *   This program is distributed in the hope that it will be useful, but
+ *   WITHOUT ANY WARRANTY; without even the implied warranty of
+ *   MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
+ *   NON INFRINGEMENT.  See the GNU General Public License for
+ *   more details.
+ *
+ * 32-bit SMP spinlocks.
+ */
+
+#ifndef _ASM_TILE_SPINLOCK_32_H
+#define _ASM_TILE_SPINLOCK_32_H
+
+#include <linux/atomic.h>
+#include <asm/page.h>
+#include <linux/compiler.h>
+
+/*
+ * We only use even ticket numbers so the '1' inserted by a tns is
+ * an unambiguous "ticket is busy" flag.
+ */
+#define TICKET_QUANTUM 2
+
+
+/*
+ * SMP ticket spinlocks, allowing only a single CPU anywhere
+ *
+ * (the type definitions are in asm/spinlock_types.h)
+ */
+static inline int arch_spin_is_locked(arch_spinlock_t *lock)
+{
+	/*
+	 * Note that even if a new ticket is in the process of being
+	 * acquired, so lock->next_ticket is 1, it's still reasonable
+	 * to claim the lock is held, since it will be momentarily
+	 * if not already.  There's no need to wait for a "valid"
+	 * lock->next_ticket to become available.
+	 */
+	return lock->next_ticket != lock->current_ticket;
+}
+
+void arch_spin_lock(arch_spinlock_t *lock);
+
+/* We cannot take an interrupt after getting a ticket, so don't enable them. */
+#define arch_spin_lock_flags(lock, flags) arch_spin_lock(lock)
+
+int arch_spin_trylock(arch_spinlock_t *lock);
+
+static inline void arch_spin_unlock(arch_spinlock_t *lock)
+{
+	/* For efficiency, overlap fetching the old ticket with the wmb(). */
+	int old_ticket = lock->current_ticket;
+	wmb();  /* guarantee anything modified under the lock is visible */
+	lock->current_ticket = old_ticket + TICKET_QUANTUM;
+}
+
+void arch_spin_unlock_wait(arch_spinlock_t *lock);
+
+/*
+ * Read-write spinlocks, allowing multiple readers
+ * but only one writer.
+ *
+ * We use a "tns/store-back" technique on a single word to manage
+ * the lock state, looping around to retry if the tns returns 1.
+ */
+
+/* Internal layout of the word; do not use. */
+#define _WR_NEXT_SHIFT	8
+#define _WR_CURR_SHIFT  16
+#define _WR_WIDTH       8
+#define _RD_COUNT_SHIFT 24
+#define _RD_COUNT_WIDTH 8
+
+/**
+ * arch_read_can_lock() - would read_trylock() succeed?
+ */
+static inline int arch_read_can_lock(arch_rwlock_t *rwlock)
+{
+	return (rwlock->lock << _RD_COUNT_WIDTH) == 0;
+}
+
+/**
+ * arch_write_can_lock() - would write_trylock() succeed?
+ */
+static inline int arch_write_can_lock(arch_rwlock_t *rwlock)
+{
+	return rwlock->lock == 0;
+}
+
+/**
+ * arch_read_lock() - acquire a read lock.
+ */
+void arch_read_lock(arch_rwlock_t *rwlock);
+
+/**
+ * arch_write_lock() - acquire a write lock.
+ */
+void arch_write_lock(arch_rwlock_t *rwlock);
+
+/**
+ * arch_read_trylock() - try to acquire a read lock.
+ */
+int arch_read_trylock(arch_rwlock_t *rwlock);
+
+/**
+ * arch_write_trylock() - try to acquire a write lock.
+ */
+int arch_write_trylock(arch_rwlock_t *rwlock);
+
+/**
+ * arch_read_unlock() - release a read lock.
+ */
+void arch_read_unlock(arch_rwlock_t *rwlock);
+
+/**
+ * arch_write_unlock() - release a write lock.
+ */
+void arch_write_unlock(arch_rwlock_t *rwlock);
+
+#define arch_read_lock_flags(lock, flags) arch_read_lock(lock)
+#define arch_write_lock_flags(lock, flags) arch_write_lock(lock)
+
+#endif /* _ASM_TILE_SPINLOCK_32_H */
diff --git a/kernel/arch/tile/include/asm/spinlock_64.h b/kernel/arch/tile/include/asm/spinlock_64.h
new file mode 100644
index 000000000..9a12b9c7e
--- /dev/null
+++ b/kernel/arch/tile/include/asm/spinlock_64.h
@@ -0,0 +1,161 @@
+/*
+ * Copyright 2011 Tilera Corporation. All Rights Reserved.
+ *
+ *   This program is free software; you can redistribute it and/or
+ *   modify it under the terms of the GNU General Public License
+ *   as published by the Free Software Foundation, version 2.
+ *
+ *   This program is distributed in the hope that it will be useful, but
+ *   WITHOUT ANY WARRANTY; without even the implied warranty of
+ *   MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
+ *   NON INFRINGEMENT.  See the GNU General Public License for
+ *   more details.
+ *
+ * 64-bit SMP ticket spinlocks, allowing only a single CPU anywhere
+ * (the type definitions are in asm/spinlock_types.h)
+ */
+
+#ifndef _ASM_TILE_SPINLOCK_64_H
+#define _ASM_TILE_SPINLOCK_64_H
+
+/* Shifts and masks for the various fields in "lock". */
+#define __ARCH_SPIN_CURRENT_SHIFT	17
+#define __ARCH_SPIN_NEXT_MASK		0x7fff
+#define __ARCH_SPIN_NEXT_OVERFLOW	0x8000
+
+/*
+ * Return the "current" portion of a ticket lock value,
+ * i.e. the number that currently owns the lock.
+ */
+static inline u32 arch_spin_current(u32 val)
+{
+	return val >> __ARCH_SPIN_CURRENT_SHIFT;
+}
+
+/*
+ * Return the "next" portion of a ticket lock value,
+ * i.e. the number that the next task to try to acquire the lock will get.
+ */
+static inline u32 arch_spin_next(u32 val)
+{
+	return val & __ARCH_SPIN_NEXT_MASK;
+}
+
+/* The lock is locked if a task would have to wait to get it. */
+static inline int arch_spin_is_locked(arch_spinlock_t *lock)
+{
+	u32 val = lock->lock;
+	return arch_spin_current(val) != arch_spin_next(val);
+}
+
+/* Bump the current ticket so the next task owns the lock. */
+static inline void arch_spin_unlock(arch_spinlock_t *lock)
+{
+	wmb();  /* guarantee anything modified under the lock is visible */
+	__insn_fetchadd4(&lock->lock, 1U << __ARCH_SPIN_CURRENT_SHIFT);
+}
+
+void arch_spin_unlock_wait(arch_spinlock_t *lock);
+
+void arch_spin_lock_slow(arch_spinlock_t *lock, u32 val);
+
+/* Grab the "next" ticket number and bump it atomically.
+ * If the current ticket is not ours, go to the slow path.
+ * We also take the slow path if the "next" value overflows.
+ */
+static inline void arch_spin_lock(arch_spinlock_t *lock)
+{
+	u32 val = __insn_fetchadd4(&lock->lock, 1);
+	u32 ticket = val & (__ARCH_SPIN_NEXT_MASK | __ARCH_SPIN_NEXT_OVERFLOW);
+	if (unlikely(arch_spin_current(val) != ticket))
+		arch_spin_lock_slow(lock, ticket);
+}
+
+/* Try to get the lock, and return whether we succeeded. */
+int arch_spin_trylock(arch_spinlock_t *lock);
+
+/* We cannot take an interrupt after getting a ticket, so don't enable them. */
+#define arch_spin_lock_flags(lock, flags) arch_spin_lock(lock)
+
+/*
+ * Read-write spinlocks, allowing multiple readers
+ * but only one writer.
+ *
+ * We use fetchadd() for readers, and fetchor() with the sign bit
+ * for writers.
+ */
+
+#define __WRITE_LOCK_BIT (1 << 31)
+
+static inline int arch_write_val_locked(int val)
+{
+	return val < 0;  /* Optimize "val & __WRITE_LOCK_BIT". */
+}
+
+/**
+ * read_can_lock - would read_trylock() succeed?
+ * @lock: the rwlock in question.
+ */
+static inline int arch_read_can_lock(arch_rwlock_t *rw)
+{
+	return !arch_write_val_locked(rw->lock);
+}
+
+/**
+ * write_can_lock - would write_trylock() succeed?
+ * @lock: the rwlock in question.
+ */
+static inline int arch_write_can_lock(arch_rwlock_t *rw)
+{
+	return rw->lock == 0;
+}
+
+extern void __read_lock_failed(arch_rwlock_t *rw);
+
+static inline void arch_read_lock(arch_rwlock_t *rw)
+{
+	u32 val = __insn_fetchaddgez4(&rw->lock, 1);
+	if (unlikely(arch_write_val_locked(val)))
+		__read_lock_failed(rw);
+}
+
+extern void __write_lock_failed(arch_rwlock_t *rw, u32 val);
+
+static inline void arch_write_lock(arch_rwlock_t *rw)
+{
+	u32 val = __insn_fetchor4(&rw->lock, __WRITE_LOCK_BIT);
+	if (unlikely(val != 0))
+		__write_lock_failed(rw, val);
+}
+
+static inline void arch_read_unlock(arch_rwlock_t *rw)
+{
+	__insn_mf();
+	__insn_fetchadd4(&rw->lock, -1);
+}
+
+static inline void arch_write_unlock(arch_rwlock_t *rw)
+{
+	__insn_mf();
+	__insn_exch4(&rw->lock, 0);  /* Avoid waiting in the write buffer. */
+}
+
+static inline int arch_read_trylock(arch_rwlock_t *rw)
+{
+	return !arch_write_val_locked(__insn_fetchaddgez4(&rw->lock, 1));
+}
+
+static inline int arch_write_trylock(arch_rwlock_t *rw)
+{
+	u32 val = __insn_fetchor4(&rw->lock, __WRITE_LOCK_BIT);
+	if (likely(val == 0))
+		return 1;
+	if (!arch_write_val_locked(val))
+		__insn_fetchand4(&rw->lock, ~__WRITE_LOCK_BIT);
+	return 0;
+}
+
+#define arch_read_lock_flags(lock, flags) arch_read_lock(lock)
+#define arch_write_lock_flags(lock, flags) arch_write_lock(lock)
+
+#endif /* _ASM_TILE_SPINLOCK_64_H */
diff --git a/kernel/arch/tile/include/asm/spinlock_types.h b/kernel/arch/tile/include/asm/spinlock_types.h
new file mode 100644
index 000000000..a71f59b49
--- /dev/null
+++ b/kernel/arch/tile/include/asm/spinlock_types.h
@@ -0,0 +1,60 @@
+/*
+ * Copyright 2010 Tilera Corporation. All Rights Reserved.
+ *
+ *   This program is free software; you can redistribute it and/or
+ *   modify it under the terms of the GNU General Public License
+ *   as published by the Free Software Foundation, version 2.
+ *
+ *   This program is distributed in the hope that it will be useful, but
+ *   WITHOUT ANY WARRANTY; without even the implied warranty of
+ *   MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
+ *   NON INFRINGEMENT.  See the GNU General Public License for
+ *   more details.
+ */
+
+#ifndef _ASM_TILE_SPINLOCK_TYPES_H
+#define _ASM_TILE_SPINLOCK_TYPES_H
+
+#ifndef __LINUX_SPINLOCK_TYPES_H
+# error "please don't include this file directly"
+#endif
+
+#ifdef __tilegx__
+
+/* Low 15 bits are "next"; high 15 bits are "current". */
+typedef struct arch_spinlock {
+	unsigned int lock;
+} arch_spinlock_t;
+
+#define __ARCH_SPIN_LOCK_UNLOCKED	{ 0 }
+
+/* High bit is "writer owns"; low 31 bits are a count of readers. */
+typedef struct arch_rwlock {
+	unsigned int lock;
+} arch_rwlock_t;
+
+#define __ARCH_RW_LOCK_UNLOCKED		{ 0 }
+
+#else
+
+typedef struct arch_spinlock {
+	/* Next ticket number to hand out. */
+	int next_ticket;
+	/* The ticket number that currently owns this lock. */
+	int current_ticket;
+} arch_spinlock_t;
+
+#define __ARCH_SPIN_LOCK_UNLOCKED	{ 0, 0 }
+
+/*
+ * Byte 0 for tns (only the low bit is used), byte 1 for ticket-lock "next",
+ * byte 2 for ticket-lock "current", byte 3 for reader count.
+ */
+typedef struct arch_rwlock {
+	unsigned int lock;
+} arch_rwlock_t;
+
+#define __ARCH_RW_LOCK_UNLOCKED		{ 0 }
+
+#endif
+#endif /* _ASM_TILE_SPINLOCK_TYPES_H */
diff --git a/kernel/arch/tile/include/asm/stack.h b/kernel/arch/tile/include/asm/stack.h
new file mode 100644
index 000000000..0e9d382a2
--- /dev/null
+++ b/kernel/arch/tile/include/asm/stack.h
@@ -0,0 +1,74 @@
+/*
+ * Copyright 2010 Tilera Corporation. All Rights Reserved.
+ *
+ *   This program is free software; you can redistribute it and/or
+ *   modify it under the terms of the GNU General Public License
+ *   as published by the Free Software Foundation, version 2.
+ *
+ *   This program is distributed in the hope that it will be useful, but
+ *   WITHOUT ANY WARRANTY; without even the implied warranty of
+ *   MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
+ *   NON INFRINGEMENT.  See the GNU General Public License for
+ *   more details.
+ */
+
+#ifndef _ASM_TILE_STACK_H
+#define _ASM_TILE_STACK_H
+
+#include <linux/types.h>
+#include <linux/sched.h>
+#include <asm/backtrace.h>
+#include <asm/page.h>
+#include <hv/hypervisor.h>
+
+/* Everything we need to keep track of a backtrace iteration */
+struct KBacktraceIterator {
+	BacktraceIterator it;
+	struct task_struct *task;     /* task we are backtracing */
+	int end;		      /* iteration complete. */
+	int new_context;              /* new context is starting */
+	int profile;                  /* profiling, so stop on async intrpt */
+	int verbose;		      /* printk extra info (don't want to
+				       * do this for profiling) */
+	int is_current;               /* backtracing current task */
+};
+
+/* Iteration methods for kernel backtraces */
+
+/*
+ * Initialize a KBacktraceIterator from a task_struct, and optionally from
+ * a set of registers.  If the registers are omitted, the process is
+ * assumed to be descheduled, and registers are read from the process's
+ * thread_struct and stack.  "verbose" means to printk some additional
+ * information about fault handlers as we pass them on the stack.
+ */
+extern void KBacktraceIterator_init(struct KBacktraceIterator *kbt,
+				    struct task_struct *, struct pt_regs *);
+
+/* Initialize iterator based on current stack. */
+extern void KBacktraceIterator_init_current(struct KBacktraceIterator *kbt);
+
+/* Helper method for above. */
+extern void _KBacktraceIterator_init_current(struct KBacktraceIterator *kbt,
+				ulong pc, ulong lr, ulong sp, ulong r52);
+
+/* No more frames? */
+extern int KBacktraceIterator_end(struct KBacktraceIterator *kbt);
+
+/* Advance to the next frame. */
+extern void KBacktraceIterator_next(struct KBacktraceIterator *kbt);
+
+/*
+ * Dump stack given complete register info. Use only from the
+ * architecture-specific code; show_stack()
+ * and dump_stack() (in entry.S) are architecture-independent entry points.
+ */
+extern void tile_show_stack(struct KBacktraceIterator *, int headers);
+
+/* Dump stack of current process, with registers to seed the backtrace. */
+extern void dump_stack_regs(struct pt_regs *);
+
+/* Helper method for assembly dump_stack(). */
+extern void _dump_stack(int dummy, ulong pc, ulong lr, ulong sp, ulong r52);
+
+#endif /* _ASM_TILE_STACK_H */
diff --git a/kernel/arch/tile/include/asm/string.h b/kernel/arch/tile/include/asm/string.h
new file mode 100644
index 000000000..92b271bd9
--- /dev/null
+++ b/kernel/arch/tile/include/asm/string.h
@@ -0,0 +1,34 @@
+/*
+ * Copyright 2010 Tilera Corporation. All Rights Reserved.
+ *
+ *   This program is free software; you can redistribute it and/or
+ *   modify it under the terms of the GNU General Public License
+ *   as published by the Free Software Foundation, version 2.
+ *
+ *   This program is distributed in the hope that it will be useful, but
+ *   WITHOUT ANY WARRANTY; without even the implied warranty of
+ *   MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
+ *   NON INFRINGEMENT.  See the GNU General Public License for
+ *   more details.
+ */
+
+#ifndef _ASM_TILE_STRING_H
+#define _ASM_TILE_STRING_H
+
+#define __HAVE_ARCH_MEMCHR
+#define __HAVE_ARCH_MEMSET
+#define __HAVE_ARCH_MEMCPY
+#define __HAVE_ARCH_MEMMOVE
+#define __HAVE_ARCH_STRCHR
+#define __HAVE_ARCH_STRLEN
+#define __HAVE_ARCH_STRNLEN
+
+extern __kernel_size_t strlen(const char *);
+extern __kernel_size_t strnlen(const char *, __kernel_size_t);
+extern char *strchr(const char *s, int c);
+extern void *memchr(const void *s, int c, size_t n);
+extern void *memset(void *, int, __kernel_size_t);
+extern void *memcpy(void *, const void *, __kernel_size_t);
+extern void *memmove(void *, const void *, __kernel_size_t);
+
+#endif /* _ASM_TILE_STRING_H */
diff --git a/kernel/arch/tile/include/asm/switch_to.h b/kernel/arch/tile/include/asm/switch_to.h
new file mode 100644
index 000000000..b8f888cbe
--- /dev/null
+++ b/kernel/arch/tile/include/asm/switch_to.h
@@ -0,0 +1,79 @@
+/*
+ * Copyright 2010 Tilera Corporation. All Rights Reserved.
+ *
+ *   This program is free software; you can redistribute it and/or
+ *   modify it under the terms of the GNU General Public License
+ *   as published by the Free Software Foundation, version 2.
+ *
+ *   This program is distributed in the hope that it will be useful, but
+ *   WITHOUT ANY WARRANTY; without even the implied warranty of
+ *   MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
+ *   NON INFRINGEMENT.  See the GNU General Public License for
+ *   more details.
+ */
+
+#ifndef _ASM_TILE_SWITCH_TO_H
+#define _ASM_TILE_SWITCH_TO_H
+
+#include <arch/sim_def.h>
+
+/*
+ * switch_to(n) should switch tasks to task nr n, first
+ * checking that n isn't the current task, in which case it does nothing.
+ * The number of callee-saved registers saved on the kernel stack
+ * is defined here for use in copy_thread() and must agree with __switch_to().
+ */
+#define CALLEE_SAVED_FIRST_REG 30
+#define CALLEE_SAVED_REGS_COUNT 24   /* r30 to r52, plus an empty to align */
+
+#ifndef __ASSEMBLY__
+
+struct task_struct;
+
+/*
+ * Pause the DMA engine and static network before task switching.
+ */
+#define prepare_arch_switch(next) _prepare_arch_switch(next)
+void _prepare_arch_switch(struct task_struct *next);
+
+struct task_struct;
+#define switch_to(prev, next, last) ((last) = _switch_to((prev), (next)))
+extern struct task_struct *_switch_to(struct task_struct *prev,
+				      struct task_struct *next);
+
+/* Helper function for _switch_to(). */
+extern struct task_struct *__switch_to(struct task_struct *prev,
+				       struct task_struct *next,
+				       unsigned long new_system_save_k_0);
+
+/* Address that switched-away from tasks are at. */
+extern unsigned long get_switch_to_pc(void);
+
+/*
+ * Kernel threads can check to see if they need to migrate their
+ * stack whenever they return from a context switch; for user
+ * threads, we defer until they are returning to user-space.
+ */
+#define finish_arch_switch(prev) do {                                     \
+	if (unlikely((prev)->state == TASK_DEAD))                         \
+		__insn_mtspr(SPR_SIM_CONTROL, SIM_CONTROL_OS_EXIT |       \
+			((prev)->pid << _SIM_CONTROL_OPERATOR_BITS));     \
+	__insn_mtspr(SPR_SIM_CONTROL, SIM_CONTROL_OS_SWITCH |             \
+		(current->pid << _SIM_CONTROL_OPERATOR_BITS));            \
+	if (current->mm == NULL && !kstack_hash &&                        \
+	    current_thread_info()->homecache_cpu != smp_processor_id())   \
+		homecache_migrate_kthread();                              \
+} while (0)
+
+/* Support function for forking a new task. */
+void ret_from_fork(void);
+
+/* Support function for forking a new kernel thread. */
+void ret_from_kernel_thread(void *fn, void *arg);
+
+/* Called from ret_from_xxx() when a new process starts up. */
+struct task_struct *sim_notify_fork(struct task_struct *prev);
+
+#endif /* !__ASSEMBLY__ */
+
+#endif /* _ASM_TILE_SWITCH_TO_H */
diff --git a/kernel/arch/tile/include/asm/syscall.h b/kernel/arch/tile/include/asm/syscall.h
new file mode 100644
index 000000000..9644b88f1
--- /dev/null
+++ b/kernel/arch/tile/include/asm/syscall.h
@@ -0,0 +1,85 @@
+/*
+ * Copyright (C) 2008-2009 Red Hat, Inc.  All rights reserved.
+ * Copyright 2010 Tilera Corporation. All Rights Reserved.
+ *
+ *   This program is free software; you can redistribute it and/or
+ *   modify it under the terms of the GNU General Public License
+ *   as published by the Free Software Foundation, version 2.
+ *
+ *   This program is distributed in the hope that it will be useful, but
+ *   WITHOUT ANY WARRANTY; without even the implied warranty of
+ *   MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
+ *   NON INFRINGEMENT.  See the GNU General Public License for
+ *   more details.
+ *
+ * See asm-generic/syscall.h for descriptions of what we must do here.
+ */
+
+#ifndef _ASM_TILE_SYSCALL_H
+#define _ASM_TILE_SYSCALL_H
+
+#include <linux/sched.h>
+#include <linux/err.h>
+#include <arch/abi.h>
+
+/* The array of function pointers for syscalls. */
+extern void *sys_call_table[];
+#ifdef CONFIG_COMPAT
+extern void *compat_sys_call_table[];
+#endif
+
+/*
+ * Only the low 32 bits of orig_r0 are meaningful, so we return int.
+ * This importantly ignores the high bits on 64-bit, so comparisons
+ * sign-extend the low 32 bits.
+ */
+static inline int syscall_get_nr(struct task_struct *t, struct pt_regs *regs)
+{
+	return regs->regs[TREG_SYSCALL_NR];
+}
+
+static inline void syscall_rollback(struct task_struct *task,
+				    struct pt_regs *regs)
+{
+	regs->regs[0] = regs->orig_r0;
+}
+
+static inline long syscall_get_error(struct task_struct *task,
+				     struct pt_regs *regs)
+{
+	unsigned long error = regs->regs[0];
+	return IS_ERR_VALUE(error) ? error : 0;
+}
+
+static inline long syscall_get_return_value(struct task_struct *task,
+					    struct pt_regs *regs)
+{
+	return regs->regs[0];
+}
+
+static inline void syscall_set_return_value(struct task_struct *task,
+					    struct pt_regs *regs,
+					    int error, long val)
+{
+	regs->regs[0] = (long) error ?: val;
+}
+
+static inline void syscall_get_arguments(struct task_struct *task,
+					 struct pt_regs *regs,
+					 unsigned int i, unsigned int n,
+					 unsigned long *args)
+{
+	BUG_ON(i + n > 6);
+	memcpy(args, &regs[i], n * sizeof(args[0]));
+}
+
+static inline void syscall_set_arguments(struct task_struct *task,
+					 struct pt_regs *regs,
+					 unsigned int i, unsigned int n,
+					 const unsigned long *args)
+{
+	BUG_ON(i + n > 6);
+	memcpy(&regs[i], args, n * sizeof(args[0]));
+}
+
+#endif	/* _ASM_TILE_SYSCALL_H */
diff --git a/kernel/arch/tile/include/asm/syscalls.h b/kernel/arch/tile/include/asm/syscalls.h
new file mode 100644
index 000000000..07b298450
--- /dev/null
+++ b/kernel/arch/tile/include/asm/syscalls.h
@@ -0,0 +1,70 @@
+/*
+ * syscalls.h - Linux syscall interfaces (arch-specific)
+ *
+ * Copyright (c) 2008 Jaswinder Singh Rajput
+ * Copyright 2010 Tilera Corporation. All Rights Reserved.
+ *
+ *   This program is free software; you can redistribute it and/or
+ *   modify it under the terms of the GNU General Public License
+ *   as published by the Free Software Foundation, version 2.
+ *
+ *   This program is distributed in the hope that it will be useful, but
+ *   WITHOUT ANY WARRANTY; without even the implied warranty of
+ *   MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
+ *   NON INFRINGEMENT.  See the GNU General Public License for
+ *   more details.
+ */
+
+#ifndef _ASM_TILE_SYSCALLS_H
+#define _ASM_TILE_SYSCALLS_H
+
+#include <linux/compiler.h>
+#include <linux/linkage.h>
+#include <linux/signal.h>
+#include <linux/types.h>
+#include <linux/compat.h>
+
+/*
+ * Note that by convention, any syscall which requires the current
+ * register set takes an additional "struct pt_regs *" pointer; a
+ * _sys_xxx() trampoline in intvec*.S just sets up the pointer and
+ * jumps to sys_xxx().
+ */
+
+/* kernel/sys.c */
+ssize_t sys32_readahead(int fd, u32 offset_lo, u32 offset_hi, u32 count);
+long sys32_fadvise64(int fd, u32 offset_lo, u32 offset_hi,
+		     u32 len, int advice);
+int sys32_fadvise64_64(int fd, u32 offset_lo, u32 offset_hi,
+		       u32 len_lo, u32 len_hi, int advice);
+long sys_cacheflush(unsigned long addr, unsigned long len,
+		    unsigned long flags);
+#ifndef __tilegx__  /* No mmap() in the 32-bit kernel. */
+#define sys_mmap sys_mmap
+#endif
+
+#ifndef __tilegx__
+/* mm/fault.c */
+long sys_cmpxchg_badaddr(unsigned long address);
+#endif
+
+#ifdef CONFIG_COMPAT
+/* These four are not defined for 64-bit, but serve as "compat" syscalls. */
+long sys_fcntl64(unsigned int fd, unsigned int cmd, unsigned long arg);
+long sys_fstat64(unsigned long fd, struct stat64 __user *statbuf);
+long sys_truncate64(const char __user *path, loff_t length);
+long sys_ftruncate64(unsigned int fd, loff_t length);
+#endif
+
+/* Provide versions of standard syscalls that use current_pt_regs(). */
+long sys_rt_sigreturn(void);
+#define sys_rt_sigreturn sys_rt_sigreturn
+
+/* These are the intvec*.S trampolines. */
+long _sys_rt_sigreturn(void);
+long _sys_clone(unsigned long clone_flags, unsigned long newsp,
+		void __user *parent_tid, void __user *child_tid);
+
+#include <asm-generic/syscalls.h>
+
+#endif /* _ASM_TILE_SYSCALLS_H */
diff --git a/kernel/arch/tile/include/asm/thread_info.h b/kernel/arch/tile/include/asm/thread_info.h
new file mode 100644
index 000000000..f804c39a5
--- /dev/null
+++ b/kernel/arch/tile/include/asm/thread_info.h
@@ -0,0 +1,192 @@
+/*
+ * Copyright (C) 2002  David Howells (dhowells@redhat.com)
+ * Copyright 2010 Tilera Corporation. All Rights Reserved.
+ *
+ *   This program is free software; you can redistribute it and/or
+ *   modify it under the terms of the GNU General Public License
+ *   as published by the Free Software Foundation, version 2.
+ *
+ *   This program is distributed in the hope that it will be useful, but
+ *   WITHOUT ANY WARRANTY; without even the implied warranty of
+ *   MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
+ *   NON INFRINGEMENT.  See the GNU General Public License for
+ *   more details.
+ */
+
+#ifndef _ASM_TILE_THREAD_INFO_H
+#define _ASM_TILE_THREAD_INFO_H
+
+#include <asm/processor.h>
+#include <asm/page.h>
+#ifndef __ASSEMBLY__
+
+/*
+ * Low level task data that assembly code needs immediate access to.
+ * The structure is placed at the bottom of the supervisor stack.
+ */
+struct thread_info {
+	struct task_struct	*task;		/* main task structure */
+	unsigned long		flags;		/* low level flags */
+	unsigned long		status;		/* thread-synchronous flags */
+	__u32			homecache_cpu;	/* CPU we are homecached on */
+	__u32			cpu;		/* current CPU */
+	int			preempt_count;	/* 0 => preemptable,
+						   <0 => BUG */
+
+	mm_segment_t		addr_limit;	/* thread address space
+						   (KERNEL_DS or USER_DS) */
+	struct single_step_state *step_state;	/* single step state
+						   (if non-zero) */
+	int			align_ctl;	/* controls unaligned access */
+#ifdef __tilegx__
+	unsigned long		unalign_jit_tmp[4]; /* temp r0..r3 storage */
+	void __user		*unalign_jit_base; /* unalign fixup JIT base */
+#endif
+};
+
+/*
+ * macros/functions for gaining access to the thread information structure.
+ */
+#define INIT_THREAD_INFO(tsk)			\
+{						\
+	.task		= &tsk,			\
+	.flags		= 0,			\
+	.cpu		= 0,			\
+	.preempt_count	= INIT_PREEMPT_COUNT,	\
+	.addr_limit	= KERNEL_DS,		\
+	.step_state	= NULL,			\
+	.align_ctl	= 0,			\
+}
+
+#define init_thread_info	(init_thread_union.thread_info)
+#define init_stack		(init_thread_union.stack)
+
+#endif /* !__ASSEMBLY__ */
+
+#if PAGE_SIZE < 8192
+#define THREAD_SIZE_ORDER (13 - PAGE_SHIFT)
+#else
+#define THREAD_SIZE_ORDER (0)
+#endif
+#define THREAD_SIZE_PAGES (1 << THREAD_SIZE_ORDER)
+
+#define THREAD_SIZE (PAGE_SIZE << THREAD_SIZE_ORDER)
+#define LOG2_THREAD_SIZE (PAGE_SHIFT + THREAD_SIZE_ORDER)
+
+#define STACK_WARN             (THREAD_SIZE/8)
+
+#ifndef __ASSEMBLY__
+
+void arch_release_thread_info(struct thread_info *info);
+
+/* How to get the thread information struct from C. */
+register unsigned long stack_pointer __asm__("sp");
+
+#define current_thread_info() \
+  ((struct thread_info *)(stack_pointer & -THREAD_SIZE))
+
+/* Sit on a nap instruction until interrupted. */
+extern void smp_nap(void);
+
+/* Enable interrupts racelessly and nap forever: helper for arch_cpu_idle(). */
+extern void _cpu_idle(void);
+
+#else /* __ASSEMBLY__ */
+
+/*
+ * How to get the thread information struct from assembly.
+ * Note that we use different macros since different architectures
+ * have different semantics in their "mm" instruction and we would
+ * like to guarantee that the macro expands to exactly one instruction.
+ */
+#ifdef __tilegx__
+#define EXTRACT_THREAD_INFO(reg) mm reg, zero, LOG2_THREAD_SIZE, 63
+#else
+#define GET_THREAD_INFO(reg) mm reg, sp, zero, LOG2_THREAD_SIZE, 31
+#endif
+
+#endif /* !__ASSEMBLY__ */
+
+/*
+ * Thread information flags that various assembly files may need to access.
+ * Keep flags accessed frequently in low bits, particular since it makes
+ * it easier to build constants in assembly.
+ */
+#define TIF_SIGPENDING		0	/* signal pending */
+#define TIF_NEED_RESCHED	1	/* rescheduling necessary */
+#define TIF_SINGLESTEP		2	/* restore singlestep on return to
+					   user mode */
+#define TIF_ASYNC_TLB		3	/* got an async TLB fault in kernel */
+#define TIF_SYSCALL_TRACE	4	/* syscall trace active */
+#define TIF_SYSCALL_AUDIT	5	/* syscall auditing active */
+#define TIF_SECCOMP		6	/* secure computing */
+#define TIF_MEMDIE		7	/* OOM killer at work */
+#define TIF_NOTIFY_RESUME	8	/* callback before returning to user */
+#define TIF_SYSCALL_TRACEPOINT	9	/* syscall tracepoint instrumentation */
+#define TIF_POLLING_NRFLAG	10	/* idle is polling for TIF_NEED_RESCHED */
+#define TIF_NOHZ		11	/* in adaptive nohz mode */
+
+#define _TIF_SIGPENDING		(1<<TIF_SIGPENDING)
+#define _TIF_NEED_RESCHED	(1<<TIF_NEED_RESCHED)
+#define _TIF_SINGLESTEP		(1<<TIF_SINGLESTEP)
+#define _TIF_ASYNC_TLB		(1<<TIF_ASYNC_TLB)
+#define _TIF_SYSCALL_TRACE	(1<<TIF_SYSCALL_TRACE)
+#define _TIF_SYSCALL_AUDIT	(1<<TIF_SYSCALL_AUDIT)
+#define _TIF_SECCOMP		(1<<TIF_SECCOMP)
+#define _TIF_MEMDIE		(1<<TIF_MEMDIE)
+#define _TIF_NOTIFY_RESUME	(1<<TIF_NOTIFY_RESUME)
+#define _TIF_SYSCALL_TRACEPOINT	(1<<TIF_SYSCALL_TRACEPOINT)
+#define _TIF_POLLING_NRFLAG	(1<<TIF_POLLING_NRFLAG)
+#define _TIF_NOHZ		(1<<TIF_NOHZ)
+
+/* Work to do on any return to user space. */
+#define _TIF_ALLWORK_MASK \
+	(_TIF_SIGPENDING | _TIF_NEED_RESCHED | _TIF_SINGLESTEP | \
+	 _TIF_ASYNC_TLB | _TIF_NOTIFY_RESUME | _TIF_NOHZ)
+
+/* Work to do at syscall entry. */
+#define _TIF_SYSCALL_ENTRY_WORK \
+	(_TIF_SYSCALL_TRACE | _TIF_SYSCALL_TRACEPOINT | _TIF_NOHZ)
+
+/* Work to do at syscall exit. */
+#define _TIF_SYSCALL_EXIT_WORK (_TIF_SYSCALL_TRACE | _TIF_SYSCALL_TRACEPOINT)
+
+/*
+ * Thread-synchronous status.
+ *
+ * This is different from the flags in that nobody else
+ * ever touches our thread-synchronous status, so we don't
+ * have to worry about atomic accesses.
+ */
+#ifdef __tilegx__
+#define TS_COMPAT		0x0001	/* 32-bit compatibility mode */
+#endif
+#define TS_RESTORE_SIGMASK	0x0008	/* restore signal mask in do_signal */
+
+#ifndef __ASSEMBLY__
+#define HAVE_SET_RESTORE_SIGMASK	1
+static inline void set_restore_sigmask(void)
+{
+	struct thread_info *ti = current_thread_info();
+	ti->status |= TS_RESTORE_SIGMASK;
+	WARN_ON(!test_bit(TIF_SIGPENDING, &ti->flags));
+}
+static inline void clear_restore_sigmask(void)
+{
+	current_thread_info()->status &= ~TS_RESTORE_SIGMASK;
+}
+static inline bool test_restore_sigmask(void)
+{
+	return current_thread_info()->status & TS_RESTORE_SIGMASK;
+}
+static inline bool test_and_clear_restore_sigmask(void)
+{
+	struct thread_info *ti = current_thread_info();
+	if (!(ti->status & TS_RESTORE_SIGMASK))
+		return false;
+	ti->status &= ~TS_RESTORE_SIGMASK;
+	return true;
+}
+#endif	/* !__ASSEMBLY__ */
+
+#endif /* _ASM_TILE_THREAD_INFO_H */
diff --git a/kernel/arch/tile/include/asm/tile-desc.h b/kernel/arch/tile/include/asm/tile-desc.h
new file mode 100644
index 000000000..43849bf79
--- /dev/null
+++ b/kernel/arch/tile/include/asm/tile-desc.h
@@ -0,0 +1,19 @@
+/*
+ * Copyright 2011 Tilera Corporation. All Rights Reserved.
+ *
+ *   This program is free software; you can redistribute it and/or
+ *   modify it under the terms of the GNU General Public License
+ *   as published by the Free Software Foundation, version 2.
+ *
+ *   This program is distributed in the hope that it will be useful, but
+ *   WITHOUT ANY WARRANTY; without even the implied warranty of
+ *   MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
+ *   NON INFRINGEMENT.  See the GNU General Public License for
+ *   more details.
+ */
+
+#ifndef __tilegx__
+#include <asm/tile-desc_32.h>
+#else
+#include <asm/tile-desc_64.h>
+#endif
diff --git a/kernel/arch/tile/include/asm/tile-desc_32.h b/kernel/arch/tile/include/asm/tile-desc_32.h
new file mode 100644
index 000000000..f09c5c43b
--- /dev/null
+++ b/kernel/arch/tile/include/asm/tile-desc_32.h
@@ -0,0 +1,553 @@
+/* TILEPro opcode information.
+ *
+ * Copyright 2011 Tilera Corporation. All Rights Reserved.
+ *
+ *   This program is free software; you can redistribute it and/or
+ *   modify it under the terms of the GNU General Public License
+ *   as published by the Free Software Foundation, version 2.
+ *
+ *   This program is distributed in the hope that it will be useful, but
+ *   WITHOUT ANY WARRANTY; without even the implied warranty of
+ *   MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
+ *   NON INFRINGEMENT.  See the GNU General Public License for
+ *   more details.
+ *
+ *
+ *
+ *
+ *
+ */
+
+#ifndef opcode_tilepro_h
+#define opcode_tilepro_h
+
+#include <arch/opcode.h>
+
+
+enum
+{
+  TILEPRO_MAX_OPERANDS = 5 /* mm */
+};
+
+typedef enum
+{
+  TILEPRO_OPC_BPT,
+  TILEPRO_OPC_INFO,
+  TILEPRO_OPC_INFOL,
+  TILEPRO_OPC_J,
+  TILEPRO_OPC_JAL,
+  TILEPRO_OPC_MOVE,
+  TILEPRO_OPC_MOVE_SN,
+  TILEPRO_OPC_MOVEI,
+  TILEPRO_OPC_MOVEI_SN,
+  TILEPRO_OPC_MOVELI,
+  TILEPRO_OPC_MOVELI_SN,
+  TILEPRO_OPC_MOVELIS,
+  TILEPRO_OPC_PREFETCH,
+  TILEPRO_OPC_RAISE,
+  TILEPRO_OPC_ADD,
+  TILEPRO_OPC_ADD_SN,
+  TILEPRO_OPC_ADDB,
+  TILEPRO_OPC_ADDB_SN,
+  TILEPRO_OPC_ADDBS_U,
+  TILEPRO_OPC_ADDBS_U_SN,
+  TILEPRO_OPC_ADDH,
+  TILEPRO_OPC_ADDH_SN,
+  TILEPRO_OPC_ADDHS,
+  TILEPRO_OPC_ADDHS_SN,
+  TILEPRO_OPC_ADDI,
+  TILEPRO_OPC_ADDI_SN,
+  TILEPRO_OPC_ADDIB,
+  TILEPRO_OPC_ADDIB_SN,
+  TILEPRO_OPC_ADDIH,
+  TILEPRO_OPC_ADDIH_SN,
+  TILEPRO_OPC_ADDLI,
+  TILEPRO_OPC_ADDLI_SN,
+  TILEPRO_OPC_ADDLIS,
+  TILEPRO_OPC_ADDS,
+  TILEPRO_OPC_ADDS_SN,
+  TILEPRO_OPC_ADIFFB_U,
+  TILEPRO_OPC_ADIFFB_U_SN,
+  TILEPRO_OPC_ADIFFH,
+  TILEPRO_OPC_ADIFFH_SN,
+  TILEPRO_OPC_AND,
+  TILEPRO_OPC_AND_SN,
+  TILEPRO_OPC_ANDI,
+  TILEPRO_OPC_ANDI_SN,
+  TILEPRO_OPC_AULI,
+  TILEPRO_OPC_AVGB_U,
+  TILEPRO_OPC_AVGB_U_SN,
+  TILEPRO_OPC_AVGH,
+  TILEPRO_OPC_AVGH_SN,
+  TILEPRO_OPC_BBNS,
+  TILEPRO_OPC_BBNS_SN,
+  TILEPRO_OPC_BBNST,
+  TILEPRO_OPC_BBNST_SN,
+  TILEPRO_OPC_BBS,
+  TILEPRO_OPC_BBS_SN,
+  TILEPRO_OPC_BBST,
+  TILEPRO_OPC_BBST_SN,
+  TILEPRO_OPC_BGEZ,
+  TILEPRO_OPC_BGEZ_SN,
+  TILEPRO_OPC_BGEZT,
+  TILEPRO_OPC_BGEZT_SN,
+  TILEPRO_OPC_BGZ,
+  TILEPRO_OPC_BGZ_SN,
+  TILEPRO_OPC_BGZT,
+  TILEPRO_OPC_BGZT_SN,
+  TILEPRO_OPC_BITX,
+  TILEPRO_OPC_BITX_SN,
+  TILEPRO_OPC_BLEZ,
+  TILEPRO_OPC_BLEZ_SN,
+  TILEPRO_OPC_BLEZT,
+  TILEPRO_OPC_BLEZT_SN,
+  TILEPRO_OPC_BLZ,
+  TILEPRO_OPC_BLZ_SN,
+  TILEPRO_OPC_BLZT,
+  TILEPRO_OPC_BLZT_SN,
+  TILEPRO_OPC_BNZ,
+  TILEPRO_OPC_BNZ_SN,
+  TILEPRO_OPC_BNZT,
+  TILEPRO_OPC_BNZT_SN,
+  TILEPRO_OPC_BYTEX,
+  TILEPRO_OPC_BYTEX_SN,
+  TILEPRO_OPC_BZ,
+  TILEPRO_OPC_BZ_SN,
+  TILEPRO_OPC_BZT,
+  TILEPRO_OPC_BZT_SN,
+  TILEPRO_OPC_CLZ,
+  TILEPRO_OPC_CLZ_SN,
+  TILEPRO_OPC_CRC32_32,
+  TILEPRO_OPC_CRC32_32_SN,
+  TILEPRO_OPC_CRC32_8,
+  TILEPRO_OPC_CRC32_8_SN,
+  TILEPRO_OPC_CTZ,
+  TILEPRO_OPC_CTZ_SN,
+  TILEPRO_OPC_DRAIN,
+  TILEPRO_OPC_DTLBPR,
+  TILEPRO_OPC_DWORD_ALIGN,
+  TILEPRO_OPC_DWORD_ALIGN_SN,
+  TILEPRO_OPC_FINV,
+  TILEPRO_OPC_FLUSH,
+  TILEPRO_OPC_FNOP,
+  TILEPRO_OPC_ICOH,
+  TILEPRO_OPC_ILL,
+  TILEPRO_OPC_INTHB,
+  TILEPRO_OPC_INTHB_SN,
+  TILEPRO_OPC_INTHH,
+  TILEPRO_OPC_INTHH_SN,
+  TILEPRO_OPC_INTLB,
+  TILEPRO_OPC_INTLB_SN,
+  TILEPRO_OPC_INTLH,
+  TILEPRO_OPC_INTLH_SN,
+  TILEPRO_OPC_INV,
+  TILEPRO_OPC_IRET,
+  TILEPRO_OPC_JALB,
+  TILEPRO_OPC_JALF,
+  TILEPRO_OPC_JALR,
+  TILEPRO_OPC_JALRP,
+  TILEPRO_OPC_JB,
+  TILEPRO_OPC_JF,
+  TILEPRO_OPC_JR,
+  TILEPRO_OPC_JRP,
+  TILEPRO_OPC_LB,
+  TILEPRO_OPC_LB_SN,
+  TILEPRO_OPC_LB_U,
+  TILEPRO_OPC_LB_U_SN,
+  TILEPRO_OPC_LBADD,
+  TILEPRO_OPC_LBADD_SN,
+  TILEPRO_OPC_LBADD_U,
+  TILEPRO_OPC_LBADD_U_SN,
+  TILEPRO_OPC_LH,
+  TILEPRO_OPC_LH_SN,
+  TILEPRO_OPC_LH_U,
+  TILEPRO_OPC_LH_U_SN,
+  TILEPRO_OPC_LHADD,
+  TILEPRO_OPC_LHADD_SN,
+  TILEPRO_OPC_LHADD_U,
+  TILEPRO_OPC_LHADD_U_SN,
+  TILEPRO_OPC_LNK,
+  TILEPRO_OPC_LNK_SN,
+  TILEPRO_OPC_LW,
+  TILEPRO_OPC_LW_SN,
+  TILEPRO_OPC_LW_NA,
+  TILEPRO_OPC_LW_NA_SN,
+  TILEPRO_OPC_LWADD,
+  TILEPRO_OPC_LWADD_SN,
+  TILEPRO_OPC_LWADD_NA,
+  TILEPRO_OPC_LWADD_NA_SN,
+  TILEPRO_OPC_MAXB_U,
+  TILEPRO_OPC_MAXB_U_SN,
+  TILEPRO_OPC_MAXH,
+  TILEPRO_OPC_MAXH_SN,
+  TILEPRO_OPC_MAXIB_U,
+  TILEPRO_OPC_MAXIB_U_SN,
+  TILEPRO_OPC_MAXIH,
+  TILEPRO_OPC_MAXIH_SN,
+  TILEPRO_OPC_MF,
+  TILEPRO_OPC_MFSPR,
+  TILEPRO_OPC_MINB_U,
+  TILEPRO_OPC_MINB_U_SN,
+  TILEPRO_OPC_MINH,
+  TILEPRO_OPC_MINH_SN,
+  TILEPRO_OPC_MINIB_U,
+  TILEPRO_OPC_MINIB_U_SN,
+  TILEPRO_OPC_MINIH,
+  TILEPRO_OPC_MINIH_SN,
+  TILEPRO_OPC_MM,
+  TILEPRO_OPC_MNZ,
+  TILEPRO_OPC_MNZ_SN,
+  TILEPRO_OPC_MNZB,
+  TILEPRO_OPC_MNZB_SN,
+  TILEPRO_OPC_MNZH,
+  TILEPRO_OPC_MNZH_SN,
+  TILEPRO_OPC_MTSPR,
+  TILEPRO_OPC_MULHH_SS,
+  TILEPRO_OPC_MULHH_SS_SN,
+  TILEPRO_OPC_MULHH_SU,
+  TILEPRO_OPC_MULHH_SU_SN,
+  TILEPRO_OPC_MULHH_UU,
+  TILEPRO_OPC_MULHH_UU_SN,
+  TILEPRO_OPC_MULHHA_SS,
+  TILEPRO_OPC_MULHHA_SS_SN,
+  TILEPRO_OPC_MULHHA_SU,
+  TILEPRO_OPC_MULHHA_SU_SN,
+  TILEPRO_OPC_MULHHA_UU,
+  TILEPRO_OPC_MULHHA_UU_SN,
+  TILEPRO_OPC_MULHHSA_UU,
+  TILEPRO_OPC_MULHHSA_UU_SN,
+  TILEPRO_OPC_MULHL_SS,
+  TILEPRO_OPC_MULHL_SS_SN,
+  TILEPRO_OPC_MULHL_SU,
+  TILEPRO_OPC_MULHL_SU_SN,
+  TILEPRO_OPC_MULHL_US,
+  TILEPRO_OPC_MULHL_US_SN,
+  TILEPRO_OPC_MULHL_UU,
+  TILEPRO_OPC_MULHL_UU_SN,
+  TILEPRO_OPC_MULHLA_SS,
+  TILEPRO_OPC_MULHLA_SS_SN,
+  TILEPRO_OPC_MULHLA_SU,
+  TILEPRO_OPC_MULHLA_SU_SN,
+  TILEPRO_OPC_MULHLA_US,
+  TILEPRO_OPC_MULHLA_US_SN,
+  TILEPRO_OPC_MULHLA_UU,
+  TILEPRO_OPC_MULHLA_UU_SN,
+  TILEPRO_OPC_MULHLSA_UU,
+  TILEPRO_OPC_MULHLSA_UU_SN,
+  TILEPRO_OPC_MULLL_SS,
+  TILEPRO_OPC_MULLL_SS_SN,
+  TILEPRO_OPC_MULLL_SU,
+  TILEPRO_OPC_MULLL_SU_SN,
+  TILEPRO_OPC_MULLL_UU,
+  TILEPRO_OPC_MULLL_UU_SN,
+  TILEPRO_OPC_MULLLA_SS,
+  TILEPRO_OPC_MULLLA_SS_SN,
+  TILEPRO_OPC_MULLLA_SU,
+  TILEPRO_OPC_MULLLA_SU_SN,
+  TILEPRO_OPC_MULLLA_UU,
+  TILEPRO_OPC_MULLLA_UU_SN,
+  TILEPRO_OPC_MULLLSA_UU,
+  TILEPRO_OPC_MULLLSA_UU_SN,
+  TILEPRO_OPC_MVNZ,
+  TILEPRO_OPC_MVNZ_SN,
+  TILEPRO_OPC_MVZ,
+  TILEPRO_OPC_MVZ_SN,
+  TILEPRO_OPC_MZ,
+  TILEPRO_OPC_MZ_SN,
+  TILEPRO_OPC_MZB,
+  TILEPRO_OPC_MZB_SN,
+  TILEPRO_OPC_MZH,
+  TILEPRO_OPC_MZH_SN,
+  TILEPRO_OPC_NAP,
+  TILEPRO_OPC_NOP,
+  TILEPRO_OPC_NOR,
+  TILEPRO_OPC_NOR_SN,
+  TILEPRO_OPC_OR,
+  TILEPRO_OPC_OR_SN,
+  TILEPRO_OPC_ORI,
+  TILEPRO_OPC_ORI_SN,
+  TILEPRO_OPC_PACKBS_U,
+  TILEPRO_OPC_PACKBS_U_SN,
+  TILEPRO_OPC_PACKHB,
+  TILEPRO_OPC_PACKHB_SN,
+  TILEPRO_OPC_PACKHS,
+  TILEPRO_OPC_PACKHS_SN,
+  TILEPRO_OPC_PACKLB,
+  TILEPRO_OPC_PACKLB_SN,
+  TILEPRO_OPC_PCNT,
+  TILEPRO_OPC_PCNT_SN,
+  TILEPRO_OPC_RL,
+  TILEPRO_OPC_RL_SN,
+  TILEPRO_OPC_RLI,
+  TILEPRO_OPC_RLI_SN,
+  TILEPRO_OPC_S1A,
+  TILEPRO_OPC_S1A_SN,
+  TILEPRO_OPC_S2A,
+  TILEPRO_OPC_S2A_SN,
+  TILEPRO_OPC_S3A,
+  TILEPRO_OPC_S3A_SN,
+  TILEPRO_OPC_SADAB_U,
+  TILEPRO_OPC_SADAB_U_SN,
+  TILEPRO_OPC_SADAH,
+  TILEPRO_OPC_SADAH_SN,
+  TILEPRO_OPC_SADAH_U,
+  TILEPRO_OPC_SADAH_U_SN,
+  TILEPRO_OPC_SADB_U,
+  TILEPRO_OPC_SADB_U_SN,
+  TILEPRO_OPC_SADH,
+  TILEPRO_OPC_SADH_SN,
+  TILEPRO_OPC_SADH_U,
+  TILEPRO_OPC_SADH_U_SN,
+  TILEPRO_OPC_SB,
+  TILEPRO_OPC_SBADD,
+  TILEPRO_OPC_SEQ,
+  TILEPRO_OPC_SEQ_SN,
+  TILEPRO_OPC_SEQB,
+  TILEPRO_OPC_SEQB_SN,
+  TILEPRO_OPC_SEQH,
+  TILEPRO_OPC_SEQH_SN,
+  TILEPRO_OPC_SEQI,
+  TILEPRO_OPC_SEQI_SN,
+  TILEPRO_OPC_SEQIB,
+  TILEPRO_OPC_SEQIB_SN,
+  TILEPRO_OPC_SEQIH,
+  TILEPRO_OPC_SEQIH_SN,
+  TILEPRO_OPC_SH,
+  TILEPRO_OPC_SHADD,
+  TILEPRO_OPC_SHL,
+  TILEPRO_OPC_SHL_SN,
+  TILEPRO_OPC_SHLB,
+  TILEPRO_OPC_SHLB_SN,
+  TILEPRO_OPC_SHLH,
+  TILEPRO_OPC_SHLH_SN,
+  TILEPRO_OPC_SHLI,
+  TILEPRO_OPC_SHLI_SN,
+  TILEPRO_OPC_SHLIB,
+  TILEPRO_OPC_SHLIB_SN,
+  TILEPRO_OPC_SHLIH,
+  TILEPRO_OPC_SHLIH_SN,
+  TILEPRO_OPC_SHR,
+  TILEPRO_OPC_SHR_SN,
+  TILEPRO_OPC_SHRB,
+  TILEPRO_OPC_SHRB_SN,
+  TILEPRO_OPC_SHRH,
+  TILEPRO_OPC_SHRH_SN,
+  TILEPRO_OPC_SHRI,
+  TILEPRO_OPC_SHRI_SN,
+  TILEPRO_OPC_SHRIB,
+  TILEPRO_OPC_SHRIB_SN,
+  TILEPRO_OPC_SHRIH,
+  TILEPRO_OPC_SHRIH_SN,
+  TILEPRO_OPC_SLT,
+  TILEPRO_OPC_SLT_SN,
+  TILEPRO_OPC_SLT_U,
+  TILEPRO_OPC_SLT_U_SN,
+  TILEPRO_OPC_SLTB,
+  TILEPRO_OPC_SLTB_SN,
+  TILEPRO_OPC_SLTB_U,
+  TILEPRO_OPC_SLTB_U_SN,
+  TILEPRO_OPC_SLTE,
+  TILEPRO_OPC_SLTE_SN,
+  TILEPRO_OPC_SLTE_U,
+  TILEPRO_OPC_SLTE_U_SN,
+  TILEPRO_OPC_SLTEB,
+  TILEPRO_OPC_SLTEB_SN,
+  TILEPRO_OPC_SLTEB_U,
+  TILEPRO_OPC_SLTEB_U_SN,
+  TILEPRO_OPC_SLTEH,
+  TILEPRO_OPC_SLTEH_SN,
+  TILEPRO_OPC_SLTEH_U,
+  TILEPRO_OPC_SLTEH_U_SN,
+  TILEPRO_OPC_SLTH,
+  TILEPRO_OPC_SLTH_SN,
+  TILEPRO_OPC_SLTH_U,
+  TILEPRO_OPC_SLTH_U_SN,
+  TILEPRO_OPC_SLTI,
+  TILEPRO_OPC_SLTI_SN,
+  TILEPRO_OPC_SLTI_U,
+  TILEPRO_OPC_SLTI_U_SN,
+  TILEPRO_OPC_SLTIB,
+  TILEPRO_OPC_SLTIB_SN,
+  TILEPRO_OPC_SLTIB_U,
+  TILEPRO_OPC_SLTIB_U_SN,
+  TILEPRO_OPC_SLTIH,
+  TILEPRO_OPC_SLTIH_SN,
+  TILEPRO_OPC_SLTIH_U,
+  TILEPRO_OPC_SLTIH_U_SN,
+  TILEPRO_OPC_SNE,
+  TILEPRO_OPC_SNE_SN,
+  TILEPRO_OPC_SNEB,
+  TILEPRO_OPC_SNEB_SN,
+  TILEPRO_OPC_SNEH,
+  TILEPRO_OPC_SNEH_SN,
+  TILEPRO_OPC_SRA,
+  TILEPRO_OPC_SRA_SN,
+  TILEPRO_OPC_SRAB,
+  TILEPRO_OPC_SRAB_SN,
+  TILEPRO_OPC_SRAH,
+  TILEPRO_OPC_SRAH_SN,
+  TILEPRO_OPC_SRAI,
+  TILEPRO_OPC_SRAI_SN,
+  TILEPRO_OPC_SRAIB,
+  TILEPRO_OPC_SRAIB_SN,
+  TILEPRO_OPC_SRAIH,
+  TILEPRO_OPC_SRAIH_SN,
+  TILEPRO_OPC_SUB,
+  TILEPRO_OPC_SUB_SN,
+  TILEPRO_OPC_SUBB,
+  TILEPRO_OPC_SUBB_SN,
+  TILEPRO_OPC_SUBBS_U,
+  TILEPRO_OPC_SUBBS_U_SN,
+  TILEPRO_OPC_SUBH,
+  TILEPRO_OPC_SUBH_SN,
+  TILEPRO_OPC_SUBHS,
+  TILEPRO_OPC_SUBHS_SN,
+  TILEPRO_OPC_SUBS,
+  TILEPRO_OPC_SUBS_SN,
+  TILEPRO_OPC_SW,
+  TILEPRO_OPC_SWADD,
+  TILEPRO_OPC_SWINT0,
+  TILEPRO_OPC_SWINT1,
+  TILEPRO_OPC_SWINT2,
+  TILEPRO_OPC_SWINT3,
+  TILEPRO_OPC_TBLIDXB0,
+  TILEPRO_OPC_TBLIDXB0_SN,
+  TILEPRO_OPC_TBLIDXB1,
+  TILEPRO_OPC_TBLIDXB1_SN,
+  TILEPRO_OPC_TBLIDXB2,
+  TILEPRO_OPC_TBLIDXB2_SN,
+  TILEPRO_OPC_TBLIDXB3,
+  TILEPRO_OPC_TBLIDXB3_SN,
+  TILEPRO_OPC_TNS,
+  TILEPRO_OPC_TNS_SN,
+  TILEPRO_OPC_WH64,
+  TILEPRO_OPC_XOR,
+  TILEPRO_OPC_XOR_SN,
+  TILEPRO_OPC_XORI,
+  TILEPRO_OPC_XORI_SN,
+  TILEPRO_OPC_NONE
+} tilepro_mnemonic;
+
+
+
+
+typedef enum
+{
+  TILEPRO_PIPELINE_X0,
+  TILEPRO_PIPELINE_X1,
+  TILEPRO_PIPELINE_Y0,
+  TILEPRO_PIPELINE_Y1,
+  TILEPRO_PIPELINE_Y2,
+} tilepro_pipeline;
+
+#define tilepro_is_x_pipeline(p) ((int)(p) <= (int)TILEPRO_PIPELINE_X1)
+
+typedef enum
+{
+  TILEPRO_OP_TYPE_REGISTER,
+  TILEPRO_OP_TYPE_IMMEDIATE,
+  TILEPRO_OP_TYPE_ADDRESS,
+  TILEPRO_OP_TYPE_SPR
+} tilepro_operand_type;
+
+struct tilepro_operand
+{
+  /* Is this operand a register, immediate or address? */
+  tilepro_operand_type type;
+
+  /* The default relocation type for this operand.  */
+  signed int default_reloc : 16;
+
+  /* How many bits is this value? (used for range checking) */
+  unsigned int num_bits : 5;
+
+  /* Is the value signed? (used for range checking) */
+  unsigned int is_signed : 1;
+
+  /* Is this operand a source register? */
+  unsigned int is_src_reg : 1;
+
+  /* Is this operand written? (i.e. is it a destination register) */
+  unsigned int is_dest_reg : 1;
+
+  /* Is this operand PC-relative? */
+  unsigned int is_pc_relative : 1;
+
+  /* By how many bits do we right shift the value before inserting? */
+  unsigned int rightshift : 2;
+
+  /* Return the bits for this operand to be ORed into an existing bundle. */
+  tilepro_bundle_bits (*insert) (int op);
+
+  /* Extract this operand and return it. */
+  unsigned int (*extract) (tilepro_bundle_bits bundle);
+};
+
+
+extern const struct tilepro_operand tilepro_operands[];
+
+/* One finite-state machine per pipe for rapid instruction decoding. */
+extern const unsigned short * const
+tilepro_bundle_decoder_fsms[TILEPRO_NUM_PIPELINE_ENCODINGS];
+
+
+struct tilepro_opcode
+{
+  /* The opcode mnemonic, e.g. "add" */
+  const char *name;
+
+  /* The enum value for this mnemonic. */
+  tilepro_mnemonic mnemonic;
+
+  /* A bit mask of which of the five pipes this instruction
+     is compatible with:
+     X0  0x01
+     X1  0x02
+     Y0  0x04
+     Y1  0x08
+     Y2  0x10 */
+  unsigned char pipes;
+
+  /* How many operands are there? */
+  unsigned char num_operands;
+
+  /* Which register does this write implicitly, or TREG_ZERO if none? */
+  unsigned char implicitly_written_register;
+
+  /* Can this be bundled with other instructions (almost always true). */
+  unsigned char can_bundle;
+
+  /* The description of the operands. Each of these is an
+   * index into the tilepro_operands[] table. */
+  unsigned char operands[TILEPRO_NUM_PIPELINE_ENCODINGS][TILEPRO_MAX_OPERANDS];
+
+};
+
+extern const struct tilepro_opcode tilepro_opcodes[];
+
+
+/* Used for non-textual disassembly into structs. */
+struct tilepro_decoded_instruction
+{
+  const struct tilepro_opcode *opcode;
+  const struct tilepro_operand *operands[TILEPRO_MAX_OPERANDS];
+  int operand_values[TILEPRO_MAX_OPERANDS];
+};
+
+
+/* Disassemble a bundle into a struct for machine processing. */
+extern int parse_insn_tilepro(tilepro_bundle_bits bits,
+                              unsigned int pc,
+                              struct tilepro_decoded_instruction
+                              decoded[TILEPRO_MAX_INSTRUCTIONS_PER_BUNDLE]);
+
+
+/* Given a set of bundle bits and a specific pipe, returns which
+ * instruction the bundle contains in that pipe.
+ */
+extern const struct tilepro_opcode *
+find_opcode(tilepro_bundle_bits bits, tilepro_pipeline pipe);
+
+
+
+#endif /* opcode_tilepro_h */
diff --git a/kernel/arch/tile/include/asm/tile-desc_64.h b/kernel/arch/tile/include/asm/tile-desc_64.h
new file mode 100644
index 000000000..1819efcba
--- /dev/null
+++ b/kernel/arch/tile/include/asm/tile-desc_64.h
@@ -0,0 +1,483 @@
+/* TILE-Gx opcode information.
+ *
+ * Copyright 2011 Tilera Corporation. All Rights Reserved.
+ *
+ *   This program is free software; you can redistribute it and/or
+ *   modify it under the terms of the GNU General Public License
+ *   as published by the Free Software Foundation, version 2.
+ *
+ *   This program is distributed in the hope that it will be useful, but
+ *   WITHOUT ANY WARRANTY; without even the implied warranty of
+ *   MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
+ *   NON INFRINGEMENT.  See the GNU General Public License for
+ *   more details.
+ *
+ *
+ *
+ *
+ *
+ */
+
+#ifndef opcode_tile_h
+#define opcode_tile_h
+
+#include <arch/opcode.h>
+
+
+enum
+{
+  TILEGX_MAX_OPERANDS = 4 /* bfexts */
+};
+
+typedef enum
+{
+  TILEGX_OPC_BPT,
+  TILEGX_OPC_INFO,
+  TILEGX_OPC_INFOL,
+  TILEGX_OPC_MOVE,
+  TILEGX_OPC_MOVEI,
+  TILEGX_OPC_MOVELI,
+  TILEGX_OPC_PREFETCH,
+  TILEGX_OPC_PREFETCH_ADD_L1,
+  TILEGX_OPC_PREFETCH_ADD_L1_FAULT,
+  TILEGX_OPC_PREFETCH_ADD_L2,
+  TILEGX_OPC_PREFETCH_ADD_L2_FAULT,
+  TILEGX_OPC_PREFETCH_ADD_L3,
+  TILEGX_OPC_PREFETCH_ADD_L3_FAULT,
+  TILEGX_OPC_PREFETCH_L1,
+  TILEGX_OPC_PREFETCH_L1_FAULT,
+  TILEGX_OPC_PREFETCH_L2,
+  TILEGX_OPC_PREFETCH_L2_FAULT,
+  TILEGX_OPC_PREFETCH_L3,
+  TILEGX_OPC_PREFETCH_L3_FAULT,
+  TILEGX_OPC_RAISE,
+  TILEGX_OPC_ADD,
+  TILEGX_OPC_ADDI,
+  TILEGX_OPC_ADDLI,
+  TILEGX_OPC_ADDX,
+  TILEGX_OPC_ADDXI,
+  TILEGX_OPC_ADDXLI,
+  TILEGX_OPC_ADDXSC,
+  TILEGX_OPC_AND,
+  TILEGX_OPC_ANDI,
+  TILEGX_OPC_BEQZ,
+  TILEGX_OPC_BEQZT,
+  TILEGX_OPC_BFEXTS,
+  TILEGX_OPC_BFEXTU,
+  TILEGX_OPC_BFINS,
+  TILEGX_OPC_BGEZ,
+  TILEGX_OPC_BGEZT,
+  TILEGX_OPC_BGTZ,
+  TILEGX_OPC_BGTZT,
+  TILEGX_OPC_BLBC,
+  TILEGX_OPC_BLBCT,
+  TILEGX_OPC_BLBS,
+  TILEGX_OPC_BLBST,
+  TILEGX_OPC_BLEZ,
+  TILEGX_OPC_BLEZT,
+  TILEGX_OPC_BLTZ,
+  TILEGX_OPC_BLTZT,
+  TILEGX_OPC_BNEZ,
+  TILEGX_OPC_BNEZT,
+  TILEGX_OPC_CLZ,
+  TILEGX_OPC_CMOVEQZ,
+  TILEGX_OPC_CMOVNEZ,
+  TILEGX_OPC_CMPEQ,
+  TILEGX_OPC_CMPEQI,
+  TILEGX_OPC_CMPEXCH,
+  TILEGX_OPC_CMPEXCH4,
+  TILEGX_OPC_CMPLES,
+  TILEGX_OPC_CMPLEU,
+  TILEGX_OPC_CMPLTS,
+  TILEGX_OPC_CMPLTSI,
+  TILEGX_OPC_CMPLTU,
+  TILEGX_OPC_CMPLTUI,
+  TILEGX_OPC_CMPNE,
+  TILEGX_OPC_CMUL,
+  TILEGX_OPC_CMULA,
+  TILEGX_OPC_CMULAF,
+  TILEGX_OPC_CMULF,
+  TILEGX_OPC_CMULFR,
+  TILEGX_OPC_CMULH,
+  TILEGX_OPC_CMULHR,
+  TILEGX_OPC_CRC32_32,
+  TILEGX_OPC_CRC32_8,
+  TILEGX_OPC_CTZ,
+  TILEGX_OPC_DBLALIGN,
+  TILEGX_OPC_DBLALIGN2,
+  TILEGX_OPC_DBLALIGN4,
+  TILEGX_OPC_DBLALIGN6,
+  TILEGX_OPC_DRAIN,
+  TILEGX_OPC_DTLBPR,
+  TILEGX_OPC_EXCH,
+  TILEGX_OPC_EXCH4,
+  TILEGX_OPC_FDOUBLE_ADD_FLAGS,
+  TILEGX_OPC_FDOUBLE_ADDSUB,
+  TILEGX_OPC_FDOUBLE_MUL_FLAGS,
+  TILEGX_OPC_FDOUBLE_PACK1,
+  TILEGX_OPC_FDOUBLE_PACK2,
+  TILEGX_OPC_FDOUBLE_SUB_FLAGS,
+  TILEGX_OPC_FDOUBLE_UNPACK_MAX,
+  TILEGX_OPC_FDOUBLE_UNPACK_MIN,
+  TILEGX_OPC_FETCHADD,
+  TILEGX_OPC_FETCHADD4,
+  TILEGX_OPC_FETCHADDGEZ,
+  TILEGX_OPC_FETCHADDGEZ4,
+  TILEGX_OPC_FETCHAND,
+  TILEGX_OPC_FETCHAND4,
+  TILEGX_OPC_FETCHOR,
+  TILEGX_OPC_FETCHOR4,
+  TILEGX_OPC_FINV,
+  TILEGX_OPC_FLUSH,
+  TILEGX_OPC_FLUSHWB,
+  TILEGX_OPC_FNOP,
+  TILEGX_OPC_FSINGLE_ADD1,
+  TILEGX_OPC_FSINGLE_ADDSUB2,
+  TILEGX_OPC_FSINGLE_MUL1,
+  TILEGX_OPC_FSINGLE_MUL2,
+  TILEGX_OPC_FSINGLE_PACK1,
+  TILEGX_OPC_FSINGLE_PACK2,
+  TILEGX_OPC_FSINGLE_SUB1,
+  TILEGX_OPC_ICOH,
+  TILEGX_OPC_ILL,
+  TILEGX_OPC_INV,
+  TILEGX_OPC_IRET,
+  TILEGX_OPC_J,
+  TILEGX_OPC_JAL,
+  TILEGX_OPC_JALR,
+  TILEGX_OPC_JALRP,
+  TILEGX_OPC_JR,
+  TILEGX_OPC_JRP,
+  TILEGX_OPC_LD,
+  TILEGX_OPC_LD1S,
+  TILEGX_OPC_LD1S_ADD,
+  TILEGX_OPC_LD1U,
+  TILEGX_OPC_LD1U_ADD,
+  TILEGX_OPC_LD2S,
+  TILEGX_OPC_LD2S_ADD,
+  TILEGX_OPC_LD2U,
+  TILEGX_OPC_LD2U_ADD,
+  TILEGX_OPC_LD4S,
+  TILEGX_OPC_LD4S_ADD,
+  TILEGX_OPC_LD4U,
+  TILEGX_OPC_LD4U_ADD,
+  TILEGX_OPC_LD_ADD,
+  TILEGX_OPC_LDNA,
+  TILEGX_OPC_LDNA_ADD,
+  TILEGX_OPC_LDNT,
+  TILEGX_OPC_LDNT1S,
+  TILEGX_OPC_LDNT1S_ADD,
+  TILEGX_OPC_LDNT1U,
+  TILEGX_OPC_LDNT1U_ADD,
+  TILEGX_OPC_LDNT2S,
+  TILEGX_OPC_LDNT2S_ADD,
+  TILEGX_OPC_LDNT2U,
+  TILEGX_OPC_LDNT2U_ADD,
+  TILEGX_OPC_LDNT4S,
+  TILEGX_OPC_LDNT4S_ADD,
+  TILEGX_OPC_LDNT4U,
+  TILEGX_OPC_LDNT4U_ADD,
+  TILEGX_OPC_LDNT_ADD,
+  TILEGX_OPC_LNK,
+  TILEGX_OPC_MF,
+  TILEGX_OPC_MFSPR,
+  TILEGX_OPC_MM,
+  TILEGX_OPC_MNZ,
+  TILEGX_OPC_MTSPR,
+  TILEGX_OPC_MUL_HS_HS,
+  TILEGX_OPC_MUL_HS_HU,
+  TILEGX_OPC_MUL_HS_LS,
+  TILEGX_OPC_MUL_HS_LU,
+  TILEGX_OPC_MUL_HU_HU,
+  TILEGX_OPC_MUL_HU_LS,
+  TILEGX_OPC_MUL_HU_LU,
+  TILEGX_OPC_MUL_LS_LS,
+  TILEGX_OPC_MUL_LS_LU,
+  TILEGX_OPC_MUL_LU_LU,
+  TILEGX_OPC_MULA_HS_HS,
+  TILEGX_OPC_MULA_HS_HU,
+  TILEGX_OPC_MULA_HS_LS,
+  TILEGX_OPC_MULA_HS_LU,
+  TILEGX_OPC_MULA_HU_HU,
+  TILEGX_OPC_MULA_HU_LS,
+  TILEGX_OPC_MULA_HU_LU,
+  TILEGX_OPC_MULA_LS_LS,
+  TILEGX_OPC_MULA_LS_LU,
+  TILEGX_OPC_MULA_LU_LU,
+  TILEGX_OPC_MULAX,
+  TILEGX_OPC_MULX,
+  TILEGX_OPC_MZ,
+  TILEGX_OPC_NAP,
+  TILEGX_OPC_NOP,
+  TILEGX_OPC_NOR,
+  TILEGX_OPC_OR,
+  TILEGX_OPC_ORI,
+  TILEGX_OPC_PCNT,
+  TILEGX_OPC_REVBITS,
+  TILEGX_OPC_REVBYTES,
+  TILEGX_OPC_ROTL,
+  TILEGX_OPC_ROTLI,
+  TILEGX_OPC_SHL,
+  TILEGX_OPC_SHL16INSLI,
+  TILEGX_OPC_SHL1ADD,
+  TILEGX_OPC_SHL1ADDX,
+  TILEGX_OPC_SHL2ADD,
+  TILEGX_OPC_SHL2ADDX,
+  TILEGX_OPC_SHL3ADD,
+  TILEGX_OPC_SHL3ADDX,
+  TILEGX_OPC_SHLI,
+  TILEGX_OPC_SHLX,
+  TILEGX_OPC_SHLXI,
+  TILEGX_OPC_SHRS,
+  TILEGX_OPC_SHRSI,
+  TILEGX_OPC_SHRU,
+  TILEGX_OPC_SHRUI,
+  TILEGX_OPC_SHRUX,
+  TILEGX_OPC_SHRUXI,
+  TILEGX_OPC_SHUFFLEBYTES,
+  TILEGX_OPC_ST,
+  TILEGX_OPC_ST1,
+  TILEGX_OPC_ST1_ADD,
+  TILEGX_OPC_ST2,
+  TILEGX_OPC_ST2_ADD,
+  TILEGX_OPC_ST4,
+  TILEGX_OPC_ST4_ADD,
+  TILEGX_OPC_ST_ADD,
+  TILEGX_OPC_STNT,
+  TILEGX_OPC_STNT1,
+  TILEGX_OPC_STNT1_ADD,
+  TILEGX_OPC_STNT2,
+  TILEGX_OPC_STNT2_ADD,
+  TILEGX_OPC_STNT4,
+  TILEGX_OPC_STNT4_ADD,
+  TILEGX_OPC_STNT_ADD,
+  TILEGX_OPC_SUB,
+  TILEGX_OPC_SUBX,
+  TILEGX_OPC_SUBXSC,
+  TILEGX_OPC_SWINT0,
+  TILEGX_OPC_SWINT1,
+  TILEGX_OPC_SWINT2,
+  TILEGX_OPC_SWINT3,
+  TILEGX_OPC_TBLIDXB0,
+  TILEGX_OPC_TBLIDXB1,
+  TILEGX_OPC_TBLIDXB2,
+  TILEGX_OPC_TBLIDXB3,
+  TILEGX_OPC_V1ADD,
+  TILEGX_OPC_V1ADDI,
+  TILEGX_OPC_V1ADDUC,
+  TILEGX_OPC_V1ADIFFU,
+  TILEGX_OPC_V1AVGU,
+  TILEGX_OPC_V1CMPEQ,
+  TILEGX_OPC_V1CMPEQI,
+  TILEGX_OPC_V1CMPLES,
+  TILEGX_OPC_V1CMPLEU,
+  TILEGX_OPC_V1CMPLTS,
+  TILEGX_OPC_V1CMPLTSI,
+  TILEGX_OPC_V1CMPLTU,
+  TILEGX_OPC_V1CMPLTUI,
+  TILEGX_OPC_V1CMPNE,
+  TILEGX_OPC_V1DDOTPU,
+  TILEGX_OPC_V1DDOTPUA,
+  TILEGX_OPC_V1DDOTPUS,
+  TILEGX_OPC_V1DDOTPUSA,
+  TILEGX_OPC_V1DOTP,
+  TILEGX_OPC_V1DOTPA,
+  TILEGX_OPC_V1DOTPU,
+  TILEGX_OPC_V1DOTPUA,
+  TILEGX_OPC_V1DOTPUS,
+  TILEGX_OPC_V1DOTPUSA,
+  TILEGX_OPC_V1INT_H,
+  TILEGX_OPC_V1INT_L,
+  TILEGX_OPC_V1MAXU,
+  TILEGX_OPC_V1MAXUI,
+  TILEGX_OPC_V1MINU,
+  TILEGX_OPC_V1MINUI,
+  TILEGX_OPC_V1MNZ,
+  TILEGX_OPC_V1MULTU,
+  TILEGX_OPC_V1MULU,
+  TILEGX_OPC_V1MULUS,
+  TILEGX_OPC_V1MZ,
+  TILEGX_OPC_V1SADAU,
+  TILEGX_OPC_V1SADU,
+  TILEGX_OPC_V1SHL,
+  TILEGX_OPC_V1SHLI,
+  TILEGX_OPC_V1SHRS,
+  TILEGX_OPC_V1SHRSI,
+  TILEGX_OPC_V1SHRU,
+  TILEGX_OPC_V1SHRUI,
+  TILEGX_OPC_V1SUB,
+  TILEGX_OPC_V1SUBUC,
+  TILEGX_OPC_V2ADD,
+  TILEGX_OPC_V2ADDI,
+  TILEGX_OPC_V2ADDSC,
+  TILEGX_OPC_V2ADIFFS,
+  TILEGX_OPC_V2AVGS,
+  TILEGX_OPC_V2CMPEQ,
+  TILEGX_OPC_V2CMPEQI,
+  TILEGX_OPC_V2CMPLES,
+  TILEGX_OPC_V2CMPLEU,
+  TILEGX_OPC_V2CMPLTS,
+  TILEGX_OPC_V2CMPLTSI,
+  TILEGX_OPC_V2CMPLTU,
+  TILEGX_OPC_V2CMPLTUI,
+  TILEGX_OPC_V2CMPNE,
+  TILEGX_OPC_V2DOTP,
+  TILEGX_OPC_V2DOTPA,
+  TILEGX_OPC_V2INT_H,
+  TILEGX_OPC_V2INT_L,
+  TILEGX_OPC_V2MAXS,
+  TILEGX_OPC_V2MAXSI,
+  TILEGX_OPC_V2MINS,
+  TILEGX_OPC_V2MINSI,
+  TILEGX_OPC_V2MNZ,
+  TILEGX_OPC_V2MULFSC,
+  TILEGX_OPC_V2MULS,
+  TILEGX_OPC_V2MULTS,
+  TILEGX_OPC_V2MZ,
+  TILEGX_OPC_V2PACKH,
+  TILEGX_OPC_V2PACKL,
+  TILEGX_OPC_V2PACKUC,
+  TILEGX_OPC_V2SADAS,
+  TILEGX_OPC_V2SADAU,
+  TILEGX_OPC_V2SADS,
+  TILEGX_OPC_V2SADU,
+  TILEGX_OPC_V2SHL,
+  TILEGX_OPC_V2SHLI,
+  TILEGX_OPC_V2SHLSC,
+  TILEGX_OPC_V2SHRS,
+  TILEGX_OPC_V2SHRSI,
+  TILEGX_OPC_V2SHRU,
+  TILEGX_OPC_V2SHRUI,
+  TILEGX_OPC_V2SUB,
+  TILEGX_OPC_V2SUBSC,
+  TILEGX_OPC_V4ADD,
+  TILEGX_OPC_V4ADDSC,
+  TILEGX_OPC_V4INT_H,
+  TILEGX_OPC_V4INT_L,
+  TILEGX_OPC_V4PACKSC,
+  TILEGX_OPC_V4SHL,
+  TILEGX_OPC_V4SHLSC,
+  TILEGX_OPC_V4SHRS,
+  TILEGX_OPC_V4SHRU,
+  TILEGX_OPC_V4SUB,
+  TILEGX_OPC_V4SUBSC,
+  TILEGX_OPC_WH64,
+  TILEGX_OPC_XOR,
+  TILEGX_OPC_XORI,
+  TILEGX_OPC_NONE
+} tilegx_mnemonic;
+
+
+
+typedef enum
+{
+  TILEGX_PIPELINE_X0,
+  TILEGX_PIPELINE_X1,
+  TILEGX_PIPELINE_Y0,
+  TILEGX_PIPELINE_Y1,
+  TILEGX_PIPELINE_Y2,
+} tilegx_pipeline;
+
+#define tilegx_is_x_pipeline(p) ((int)(p) <= (int)TILEGX_PIPELINE_X1)
+
+typedef enum
+{
+  TILEGX_OP_TYPE_REGISTER,
+  TILEGX_OP_TYPE_IMMEDIATE,
+  TILEGX_OP_TYPE_ADDRESS,
+  TILEGX_OP_TYPE_SPR
+} tilegx_operand_type;
+
+struct tilegx_operand
+{
+  /* Is this operand a register, immediate or address? */
+  tilegx_operand_type type;
+
+  /* The default relocation type for this operand.  */
+  signed int default_reloc : 16;
+
+  /* How many bits is this value? (used for range checking) */
+  unsigned int num_bits : 5;
+
+  /* Is the value signed? (used for range checking) */
+  unsigned int is_signed : 1;
+
+  /* Is this operand a source register? */
+  unsigned int is_src_reg : 1;
+
+  /* Is this operand written? (i.e. is it a destination register) */
+  unsigned int is_dest_reg : 1;
+
+  /* Is this operand PC-relative? */
+  unsigned int is_pc_relative : 1;
+
+  /* By how many bits do we right shift the value before inserting? */
+  unsigned int rightshift : 2;
+
+  /* Return the bits for this operand to be ORed into an existing bundle. */
+  tilegx_bundle_bits (*insert) (int op);
+
+  /* Extract this operand and return it. */
+  unsigned int (*extract) (tilegx_bundle_bits bundle);
+};
+
+
+extern const struct tilegx_operand tilegx_operands[];
+
+/* One finite-state machine per pipe for rapid instruction decoding. */
+extern const unsigned short * const
+tilegx_bundle_decoder_fsms[TILEGX_NUM_PIPELINE_ENCODINGS];
+
+
+struct tilegx_opcode
+{
+  /* The opcode mnemonic, e.g. "add" */
+  const char *name;
+
+  /* The enum value for this mnemonic. */
+  tilegx_mnemonic mnemonic;
+
+  /* A bit mask of which of the five pipes this instruction
+     is compatible with:
+     X0  0x01
+     X1  0x02
+     Y0  0x04
+     Y1  0x08
+     Y2  0x10 */
+  unsigned char pipes;
+
+  /* How many operands are there? */
+  unsigned char num_operands;
+
+  /* Which register does this write implicitly, or TREG_ZERO if none? */
+  unsigned char implicitly_written_register;
+
+  /* Can this be bundled with other instructions (almost always true). */
+  unsigned char can_bundle;
+
+  /* The description of the operands. Each of these is an
+   * index into the tilegx_operands[] table. */
+  unsigned char operands[TILEGX_NUM_PIPELINE_ENCODINGS][TILEGX_MAX_OPERANDS];
+
+};
+
+extern const struct tilegx_opcode tilegx_opcodes[];
+
+/* Used for non-textual disassembly into structs. */
+struct tilegx_decoded_instruction
+{
+  const struct tilegx_opcode *opcode;
+  const struct tilegx_operand *operands[TILEGX_MAX_OPERANDS];
+  long long operand_values[TILEGX_MAX_OPERANDS];
+};
+
+
+/* Disassemble a bundle into a struct for machine processing. */
+extern int parse_insn_tilegx(tilegx_bundle_bits bits,
+                             unsigned long long pc,
+                             struct tilegx_decoded_instruction
+                             decoded[TILEGX_MAX_INSTRUCTIONS_PER_BUNDLE]);
+
+
+
+#endif /* opcode_tilegx_h */
diff --git a/kernel/arch/tile/include/asm/timex.h b/kernel/arch/tile/include/asm/timex.h
new file mode 100644
index 000000000..dc987d53e
--- /dev/null
+++ b/kernel/arch/tile/include/asm/timex.h
@@ -0,0 +1,52 @@
+/*
+ * Copyright 2010 Tilera Corporation. All Rights Reserved.
+ *
+ *   This program is free software; you can redistribute it and/or
+ *   modify it under the terms of the GNU General Public License
+ *   as published by the Free Software Foundation, version 2.
+ *
+ *   This program is distributed in the hope that it will be useful, but
+ *   WITHOUT ANY WARRANTY; without even the implied warranty of
+ *   MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
+ *   NON INFRINGEMENT.  See the GNU General Public License for
+ *   more details.
+ */
+
+#ifndef _ASM_TILE_TIMEX_H
+#define _ASM_TILE_TIMEX_H
+
+/*
+ * This rate should be a multiple of the possible HZ values (100, 250, 1000)
+ * and a fraction of the possible hardware timer frequencies.  Our timer
+ * frequency is highly tunable but also quite precise, so for the primary use
+ * of this value (setting ACT_HZ from HZ) we just pick a value that causes
+ * ACT_HZ to be set to HZ.  We make the value somewhat large just to be
+ * more robust in case someone tries out a new value of HZ.
+ */
+#define CLOCK_TICK_RATE	1000000
+
+typedef unsigned long long cycles_t;
+
+#if CHIP_HAS_SPLIT_CYCLE()
+cycles_t get_cycles(void);
+#define get_cycles_low() __insn_mfspr(SPR_CYCLE_LOW)
+#else
+static inline cycles_t get_cycles(void)
+{
+	return __insn_mfspr(SPR_CYCLE);
+}
+#define get_cycles_low() __insn_mfspr(SPR_CYCLE)   /* just get all 64 bits */
+#endif
+
+cycles_t get_clock_rate(void);
+
+/* Convert nanoseconds to core clock cycles. */
+cycles_t ns2cycles(unsigned long nsecs);
+
+/* Called at cpu initialization to set some low-level constants. */
+void setup_clock(void);
+
+/* Called at cpu initialization to start the tile-timer clock device. */
+void setup_tile_timer(void);
+
+#endif /* _ASM_TILE_TIMEX_H */
diff --git a/kernel/arch/tile/include/asm/tlb.h b/kernel/arch/tile/include/asm/tlb.h
new file mode 100644
index 000000000..4a891a1a8
--- /dev/null
+++ b/kernel/arch/tile/include/asm/tlb.h
@@ -0,0 +1,25 @@
+/*
+ * Copyright 2010 Tilera Corporation. All Rights Reserved.
+ *
+ *   This program is free software; you can redistribute it and/or
+ *   modify it under the terms of the GNU General Public License
+ *   as published by the Free Software Foundation, version 2.
+ *
+ *   This program is distributed in the hope that it will be useful, but
+ *   WITHOUT ANY WARRANTY; without even the implied warranty of
+ *   MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
+ *   NON INFRINGEMENT.  See the GNU General Public License for
+ *   more details.
+ */
+
+#ifndef _ASM_TILE_TLB_H
+#define _ASM_TILE_TLB_H
+
+#define tlb_start_vma(tlb, vma) do { } while (0)
+#define tlb_end_vma(tlb, vma) do { } while (0)
+#define __tlb_remove_tlb_entry(tlb, ptep, address) do { } while (0)
+#define tlb_flush(tlb) flush_tlb_mm((tlb)->mm)
+
+#include <asm-generic/tlb.h>
+
+#endif /* _ASM_TILE_TLB_H */
diff --git a/kernel/arch/tile/include/asm/tlbflush.h b/kernel/arch/tile/include/asm/tlbflush.h
new file mode 100644
index 000000000..dcf91b25a
--- /dev/null
+++ b/kernel/arch/tile/include/asm/tlbflush.h
@@ -0,0 +1,123 @@
+/*
+ * Copyright 2010 Tilera Corporation. All Rights Reserved.
+ *
+ *   This program is free software; you can redistribute it and/or
+ *   modify it under the terms of the GNU General Public License
+ *   as published by the Free Software Foundation, version 2.
+ *
+ *   This program is distributed in the hope that it will be useful, but
+ *   WITHOUT ANY WARRANTY; without even the implied warranty of
+ *   MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
+ *   NON INFRINGEMENT.  See the GNU General Public License for
+ *   more details.
+ */
+
+#ifndef _ASM_TILE_TLBFLUSH_H
+#define _ASM_TILE_TLBFLUSH_H
+
+#include <linux/mm.h>
+#include <linux/sched.h>
+#include <linux/smp.h>
+#include <asm/cacheflush.h>
+#include <asm/page.h>
+#include <hv/hypervisor.h>
+
+/*
+ * Rather than associating each mm with its own ASID, we just use
+ * ASIDs to allow us to lazily flush the TLB when we switch mms.
+ * This way we only have to do an actual TLB flush on mm switch
+ * every time we wrap ASIDs, not every single time we switch.
+ *
+ * FIXME: We might improve performance by keeping ASIDs around
+ * properly, though since the hypervisor direct-maps VAs to TSB
+ * entries, we're likely to have lost at least the executable page
+ * mappings by the time we switch back to the original mm.
+ */
+DECLARE_PER_CPU(int, current_asid);
+
+/* The hypervisor tells us what ASIDs are available to us. */
+extern int min_asid, max_asid;
+
+/* Pass as vma pointer for non-executable mapping, if no vma available. */
+#define FLUSH_NONEXEC ((struct vm_area_struct *)-1UL)
+
+/* Flush a single user page on this cpu. */
+static inline void local_flush_tlb_page(struct vm_area_struct *vma,
+					unsigned long addr,
+					unsigned long page_size)
+{
+	int rc = hv_flush_page(addr, page_size);
+	if (rc < 0)
+		panic("hv_flush_page(%#lx,%#lx) failed: %d",
+		      addr, page_size, rc);
+	if (!vma || (vma != FLUSH_NONEXEC && (vma->vm_flags & VM_EXEC)))
+		__flush_icache();
+}
+
+/* Flush range of user pages on this cpu. */
+static inline void local_flush_tlb_pages(struct vm_area_struct *vma,
+					 unsigned long addr,
+					 unsigned long page_size,
+					 unsigned long len)
+{
+	int rc = hv_flush_pages(addr, page_size, len);
+	if (rc < 0)
+		panic("hv_flush_pages(%#lx,%#lx,%#lx) failed: %d",
+		      addr, page_size, len, rc);
+	if (!vma || (vma != FLUSH_NONEXEC && (vma->vm_flags & VM_EXEC)))
+		__flush_icache();
+}
+
+/* Flush all user pages on this cpu. */
+static inline void local_flush_tlb(void)
+{
+	int rc = hv_flush_all(1);   /* preserve global mappings */
+	if (rc < 0)
+		panic("hv_flush_all(1) failed: %d", rc);
+	__flush_icache();
+}
+
+/*
+ * Global pages have to be flushed a bit differently. Not a real
+ * performance problem because this does not happen often.
+ */
+static inline void local_flush_tlb_all(void)
+{
+	int i;
+	for (i = 0; ; ++i) {
+		HV_VirtAddrRange r = hv_inquire_virtual(i);
+		if (r.size == 0)
+			break;
+		local_flush_tlb_pages(NULL, r.start, PAGE_SIZE, r.size);
+		local_flush_tlb_pages(NULL, r.start, HPAGE_SIZE, r.size);
+	}
+}
+
+/*
+ * TLB flushing:
+ *
+ *  - flush_tlb() flushes the current mm struct TLBs
+ *  - flush_tlb_all() flushes all processes TLBs
+ *  - flush_tlb_mm(mm) flushes the specified mm context TLB's
+ *  - flush_tlb_page(vma, vmaddr) flushes one page
+ *  - flush_tlb_range(vma, start, end) flushes a range of pages
+ *  - flush_tlb_kernel_range(start, end) flushes a range of kernel pages
+ *  - flush_tlb_others(cpumask, mm, va) flushes TLBs on other cpus
+ *
+ * Here (as in vm_area_struct), "end" means the first byte after
+ * our end address.
+ */
+
+extern void flush_tlb_all(void);
+extern void flush_tlb_kernel_range(unsigned long start, unsigned long end);
+extern void flush_tlb_current_task(void);
+extern void flush_tlb_mm(struct mm_struct *);
+extern void flush_tlb_page(struct vm_area_struct *, unsigned long);
+extern void flush_tlb_page_mm(struct vm_area_struct *,
+			      struct mm_struct *, unsigned long);
+extern void flush_tlb_range(struct vm_area_struct *,
+			    unsigned long start, unsigned long end);
+
+#define flush_tlb()     flush_tlb_current_task()
+
+#endif /* _ASM_TILE_TLBFLUSH_H */
diff --git a/kernel/arch/tile/include/asm/topology.h b/kernel/arch/tile/include/asm/topology.h
new file mode 100644
index 000000000..938311844
--- /dev/null
+++ b/kernel/arch/tile/include/asm/topology.h
@@ -0,0 +1,61 @@
+/*
+ * Copyright 2010 Tilera Corporation. All Rights Reserved.
+ *
+ *   This program is free software; you can redistribute it and/or
+ *   modify it under the terms of the GNU General Public License
+ *   as published by the Free Software Foundation, version 2.
+ *
+ *   This program is distributed in the hope that it will be useful, but
+ *   WITHOUT ANY WARRANTY; without even the implied warranty of
+ *   MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
+ *   NON INFRINGEMENT.  See the GNU General Public License for
+ *   more details.
+ */
+
+#ifndef _ASM_TILE_TOPOLOGY_H
+#define _ASM_TILE_TOPOLOGY_H
+
+#ifdef CONFIG_NUMA
+
+#include <linux/cpumask.h>
+
+/* Mappings between logical cpu number and node number. */
+extern struct cpumask node_2_cpu_mask[];
+extern char cpu_2_node[];
+
+/* Returns the number of the node containing CPU 'cpu'. */
+static inline int cpu_to_node(int cpu)
+{
+	return cpu_2_node[cpu];
+}
+
+/*
+ * Returns the number of the node containing Node 'node'.
+ * This architecture is flat, so it is a pretty simple function!
+ */
+#define parent_node(node) (node)
+
+/* Returns a bitmask of CPUs on Node 'node'. */
+static inline const struct cpumask *cpumask_of_node(int node)
+{
+	return &node_2_cpu_mask[node];
+}
+
+/* For now, use numa node -1 for global allocation. */
+#define pcibus_to_node(bus)		((void)(bus), -1)
+
+/* By definition, we create nodes based on online memory. */
+#define node_has_online_mem(nid) 1
+
+#endif /* CONFIG_NUMA */
+
+#include <asm-generic/topology.h>
+
+#ifdef CONFIG_SMP
+#define topology_physical_package_id(cpu)       ((void)(cpu), 0)
+#define topology_core_id(cpu)                   (cpu)
+#define topology_core_cpumask(cpu)              ((void)(cpu), cpu_online_mask)
+#define topology_thread_cpumask(cpu)            cpumask_of(cpu)
+#endif
+
+#endif /* _ASM_TILE_TOPOLOGY_H */
diff --git a/kernel/arch/tile/include/asm/traps.h b/kernel/arch/tile/include/asm/traps.h
new file mode 100644
index 000000000..4b99a1c3a
--- /dev/null
+++ b/kernel/arch/tile/include/asm/traps.h
@@ -0,0 +1,85 @@
+/*
+ * Copyright 2010 Tilera Corporation. All Rights Reserved.
+ *
+ *   This program is free software; you can redistribute it and/or
+ *   modify it under the terms of the GNU General Public License
+ *   as published by the Free Software Foundation, version 2.
+ *
+ *   This program is distributed in the hope that it will be useful, but
+ *   WITHOUT ANY WARRANTY; without even the implied warranty of
+ *   MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
+ *   NON INFRINGEMENT.  See the GNU General Public License for
+ *   more details.
+ */
+
+#ifndef _ASM_TILE_TRAPS_H
+#define _ASM_TILE_TRAPS_H
+
+#ifndef __ASSEMBLY__
+#include <arch/chip.h>
+
+/* mm/fault.c */
+void do_page_fault(struct pt_regs *, int fault_num,
+		   unsigned long address, unsigned long write);
+#if CHIP_HAS_TILE_DMA()
+void do_async_page_fault(struct pt_regs *);
+#endif
+
+#ifndef __tilegx__
+/*
+ * We return this structure in registers to avoid having to write
+ * additional save/restore code in the intvec.S caller.
+ */
+struct intvec_state {
+	void *handler;
+	unsigned long vecnum;
+	unsigned long fault_num;
+	unsigned long info;
+	unsigned long retval;
+};
+struct intvec_state do_page_fault_ics(struct pt_regs *regs, int fault_num,
+				      unsigned long address,
+				      unsigned long info);
+#endif
+
+/* kernel/traps.c */
+void do_trap(struct pt_regs *, int fault_num, unsigned long reason);
+void kernel_double_fault(int dummy, ulong pc, ulong lr, ulong sp, ulong r52);
+
+/* kernel/time.c */
+void do_timer_interrupt(struct pt_regs *, int fault_num);
+
+/* kernel/messaging.c */
+void hv_message_intr(struct pt_regs *, int intnum);
+
+/* kernel/irq.c */
+void tile_dev_intr(struct pt_regs *, int intnum);
+
+#ifdef CONFIG_HARDWALL
+/* kernel/hardwall.c */
+void do_hardwall_trap(struct pt_regs *, int fault_num);
+#endif
+
+/* kernel/ptrace.c */
+void do_breakpoint(struct pt_regs *, int fault_num);
+
+
+#ifdef __tilegx__
+/* kernel/single_step.c */
+void gx_singlestep_handle(struct pt_regs *, int fault_num);
+
+/* kernel/intvec_64.S */
+void fill_ra_stack(void);
+
+/* Handle unalign data fixup. */
+extern void do_unaligned(struct pt_regs *regs, int vecnum);
+#endif
+
+#endif /* __ASSEMBLY__ */
+
+#ifdef __tilegx__
+/* 128 byte JIT per unalign fixup. */
+#define UNALIGN_JIT_SHIFT    7
+#endif
+
+#endif /* _ASM_TILE_TRAPS_H */
diff --git a/kernel/arch/tile/include/asm/uaccess.h b/kernel/arch/tile/include/asm/uaccess.h
new file mode 100644
index 000000000..a33276bf5
--- /dev/null
+++ b/kernel/arch/tile/include/asm/uaccess.h
@@ -0,0 +1,611 @@
+/*
+ * Copyright 2010 Tilera Corporation. All Rights Reserved.
+ *
+ *   This program is free software; you can redistribute it and/or
+ *   modify it under the terms of the GNU General Public License
+ *   as published by the Free Software Foundation, version 2.
+ *
+ *   This program is distributed in the hope that it will be useful, but
+ *   WITHOUT ANY WARRANTY; without even the implied warranty of
+ *   MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
+ *   NON INFRINGEMENT.  See the GNU General Public License for
+ *   more details.
+ */
+
+#ifndef _ASM_TILE_UACCESS_H
+#define _ASM_TILE_UACCESS_H
+
+/*
+ * User space memory access functions
+ */
+#include <linux/sched.h>
+#include <linux/mm.h>
+#include <asm-generic/uaccess-unaligned.h>
+#include <asm/processor.h>
+#include <asm/page.h>
+
+#define VERIFY_READ	0
+#define VERIFY_WRITE	1
+
+/*
+ * The fs value determines whether argument validity checking should be
+ * performed or not.  If get_fs() == USER_DS, checking is performed, with
+ * get_fs() == KERNEL_DS, checking is bypassed.
+ *
+ * For historical reasons, these macros are grossly misnamed.
+ */
+#define MAKE_MM_SEG(a)  ((mm_segment_t) { (a) })
+
+#define KERNEL_DS	MAKE_MM_SEG(-1UL)
+#define USER_DS		MAKE_MM_SEG(PAGE_OFFSET)
+
+#define get_ds()	(KERNEL_DS)
+#define get_fs()	(current_thread_info()->addr_limit)
+#define set_fs(x)	(current_thread_info()->addr_limit = (x))
+
+#define segment_eq(a, b) ((a).seg == (b).seg)
+
+#ifndef __tilegx__
+/*
+ * We could allow mapping all 16 MB at 0xfc000000, but we set up a
+ * special hack in arch_setup_additional_pages() to auto-create a mapping
+ * for the first 16 KB, and it would seem strange to have different
+ * user-accessible semantics for memory at 0xfc000000 and above 0xfc004000.
+ */
+static inline int is_arch_mappable_range(unsigned long addr,
+					 unsigned long size)
+{
+	return (addr >= MEM_USER_INTRPT &&
+		addr < (MEM_USER_INTRPT + INTRPT_SIZE) &&
+		size <= (MEM_USER_INTRPT + INTRPT_SIZE) - addr);
+}
+#define is_arch_mappable_range is_arch_mappable_range
+#else
+#define is_arch_mappable_range(addr, size) 0
+#endif
+
+/*
+ * Test whether a block of memory is a valid user space address.
+ * Returns 0 if the range is valid, nonzero otherwise.
+ */
+int __range_ok(unsigned long addr, unsigned long size);
+
+/**
+ * access_ok: - Checks if a user space pointer is valid
+ * @type: Type of access: %VERIFY_READ or %VERIFY_WRITE.  Note that
+ *        %VERIFY_WRITE is a superset of %VERIFY_READ - if it is safe
+ *        to write to a block, it is always safe to read from it.
+ * @addr: User space pointer to start of block to check
+ * @size: Size of block to check
+ *
+ * Context: User context only. This function may sleep if pagefaults are
+ *          enabled.
+ *
+ * Checks if a pointer to a block of memory in user space is valid.
+ *
+ * Returns true (nonzero) if the memory block may be valid, false (zero)
+ * if it is definitely invalid.
+ *
+ * Note that, depending on architecture, this function probably just
+ * checks that the pointer is in the user space range - after calling
+ * this function, memory access functions may still return -EFAULT.
+ */
+#define access_ok(type, addr, size) ({ \
+	__chk_user_ptr(addr); \
+	likely(__range_ok((unsigned long)(addr), (size)) == 0);	\
+})
+
+/*
+ * The exception table consists of pairs of addresses: the first is the
+ * address of an instruction that is allowed to fault, and the second is
+ * the address at which the program should continue.  No registers are
+ * modified, so it is entirely up to the continuation code to figure out
+ * what to do.
+ *
+ * All the routines below use bits of fixup code that are out of line
+ * with the main instruction path.  This means when everything is well,
+ * we don't even have to jump over them.  Further, they do not intrude
+ * on our cache or tlb entries.
+ */
+
+struct exception_table_entry {
+	unsigned long insn, fixup;
+};
+
+extern int fixup_exception(struct pt_regs *regs);
+
+/*
+ * This is a type: either unsigned long, if the argument fits into
+ * that type, or otherwise unsigned long long.
+ */
+#define __inttype(x) \
+	__typeof__(__builtin_choose_expr(sizeof(x) > sizeof(0UL), 0ULL, 0UL))
+
+/*
+ * Support macros for __get_user().
+ * Note that __get_user() and __put_user() assume proper alignment.
+ */
+
+#ifdef __LP64__
+#define _ASM_PTR	".quad"
+#define _ASM_ALIGN	".align 8"
+#else
+#define _ASM_PTR	".long"
+#define _ASM_ALIGN	".align 4"
+#endif
+
+#define __get_user_asm(OP, x, ptr, ret)					\
+	asm volatile("1: {" #OP " %1, %2; movei %0, 0 }\n"		\
+		     ".pushsection .fixup,\"ax\"\n"			\
+		     "0: { movei %1, 0; movei %0, %3 }\n"		\
+		     "j 9f\n"						\
+		     ".section __ex_table,\"a\"\n"			\
+		     _ASM_ALIGN "\n"					\
+		     _ASM_PTR " 1b, 0b\n"				\
+		     ".popsection\n"					\
+		     "9:"						\
+		     : "=r" (ret), "=r" (x)				\
+		     : "r" (ptr), "i" (-EFAULT))
+
+#ifdef __tilegx__
+#define __get_user_1(x, ptr, ret) __get_user_asm(ld1u, x, ptr, ret)
+#define __get_user_2(x, ptr, ret) __get_user_asm(ld2u, x, ptr, ret)
+#define __get_user_4(x, ptr, ret) __get_user_asm(ld4s, x, ptr, ret)
+#define __get_user_8(x, ptr, ret) __get_user_asm(ld, x, ptr, ret)
+#else
+#define __get_user_1(x, ptr, ret) __get_user_asm(lb_u, x, ptr, ret)
+#define __get_user_2(x, ptr, ret) __get_user_asm(lh_u, x, ptr, ret)
+#define __get_user_4(x, ptr, ret) __get_user_asm(lw, x, ptr, ret)
+#ifdef __LITTLE_ENDIAN
+#define __lo32(a, b) a
+#define __hi32(a, b) b
+#else
+#define __lo32(a, b) b
+#define __hi32(a, b) a
+#endif
+#define __get_user_8(x, ptr, ret)					\
+	({								\
+		unsigned int __a, __b;					\
+		asm volatile("1: { lw %1, %3; addi %2, %3, 4 }\n"	\
+			     "2: { lw %2, %2; movei %0, 0 }\n"		\
+			     ".pushsection .fixup,\"ax\"\n"		\
+			     "0: { movei %1, 0; movei %2, 0 }\n"	\
+			     "{ movei %0, %4; j 9f }\n"			\
+			     ".section __ex_table,\"a\"\n"		\
+			     ".align 4\n"				\
+			     ".word 1b, 0b\n"				\
+			     ".word 2b, 0b\n"				\
+			     ".popsection\n"				\
+			     "9:"					\
+			     : "=r" (ret), "=r" (__a), "=&r" (__b)	\
+			     : "r" (ptr), "i" (-EFAULT));		\
+		(x) = (__force __typeof(x))(__inttype(x))		\
+			(((u64)__hi32(__a, __b) << 32) |		\
+			 __lo32(__a, __b));				\
+	})
+#endif
+
+extern int __get_user_bad(void)
+  __attribute__((warning("sizeof __get_user argument not 1, 2, 4 or 8")));
+
+/**
+ * __get_user: - Get a simple variable from user space, with less checking.
+ * @x:   Variable to store result.
+ * @ptr: Source address, in user space.
+ *
+ * Context: User context only. This function may sleep if pagefaults are
+ *          enabled.
+ *
+ * This macro copies a single simple variable from user space to kernel
+ * space.  It supports simple types like char and int, but not larger
+ * data types like structures or arrays.
+ *
+ * @ptr must have pointer-to-simple-variable type, and the result of
+ * dereferencing @ptr must be assignable to @x without a cast.
+ *
+ * Returns zero on success, or -EFAULT on error.
+ * On error, the variable @x is set to zero.
+ *
+ * Caller must check the pointer with access_ok() before calling this
+ * function.
+ */
+#define __get_user(x, ptr)						\
+	({								\
+		int __ret;						\
+		typeof(x) _x;						\
+		__chk_user_ptr(ptr);					\
+		switch (sizeof(*(ptr))) {				\
+		case 1: __get_user_1(_x, ptr, __ret); break;		\
+		case 2: __get_user_2(_x, ptr, __ret); break;		\
+		case 4: __get_user_4(_x, ptr, __ret); break;		\
+		case 8: __get_user_8(_x, ptr, __ret); break;		\
+		default: __ret = __get_user_bad(); break;		\
+		}							\
+		(x) = (typeof(*(ptr))) _x;				\
+		__ret;							\
+	})
+
+/* Support macros for __put_user(). */
+
+#define __put_user_asm(OP, x, ptr, ret)			\
+	asm volatile("1: {" #OP " %1, %2; movei %0, 0 }\n"		\
+		     ".pushsection .fixup,\"ax\"\n"			\
+		     "0: { movei %0, %3; j 9f }\n"			\
+		     ".section __ex_table,\"a\"\n"			\
+		     _ASM_ALIGN "\n"					\
+		     _ASM_PTR " 1b, 0b\n"				\
+		     ".popsection\n"					\
+		     "9:"						\
+		     : "=r" (ret)					\
+		     : "r" (ptr), "r" (x), "i" (-EFAULT))
+
+#ifdef __tilegx__
+#define __put_user_1(x, ptr, ret) __put_user_asm(st1, x, ptr, ret)
+#define __put_user_2(x, ptr, ret) __put_user_asm(st2, x, ptr, ret)
+#define __put_user_4(x, ptr, ret) __put_user_asm(st4, x, ptr, ret)
+#define __put_user_8(x, ptr, ret) __put_user_asm(st, x, ptr, ret)
+#else
+#define __put_user_1(x, ptr, ret) __put_user_asm(sb, x, ptr, ret)
+#define __put_user_2(x, ptr, ret) __put_user_asm(sh, x, ptr, ret)
+#define __put_user_4(x, ptr, ret) __put_user_asm(sw, x, ptr, ret)
+#define __put_user_8(x, ptr, ret)					\
+	({								\
+		u64 __x = (__force __inttype(x))(x);			\
+		int __lo = (int) __x, __hi = (int) (__x >> 32);		\
+		asm volatile("1: { sw %1, %2; addi %0, %1, 4 }\n"	\
+			     "2: { sw %0, %3; movei %0, 0 }\n"		\
+			     ".pushsection .fixup,\"ax\"\n"		\
+			     "0: { movei %0, %4; j 9f }\n"		\
+			     ".section __ex_table,\"a\"\n"		\
+			     ".align 4\n"				\
+			     ".word 1b, 0b\n"				\
+			     ".word 2b, 0b\n"				\
+			     ".popsection\n"				\
+			     "9:"					\
+			     : "=&r" (ret)				\
+			     : "r" (ptr), "r" (__lo32(__lo, __hi)),	\
+			     "r" (__hi32(__lo, __hi)), "i" (-EFAULT));	\
+	})
+#endif
+
+extern int __put_user_bad(void)
+  __attribute__((warning("sizeof __put_user argument not 1, 2, 4 or 8")));
+
+/**
+ * __put_user: - Write a simple value into user space, with less checking.
+ * @x:   Value to copy to user space.
+ * @ptr: Destination address, in user space.
+ *
+ * Context: User context only. This function may sleep if pagefaults are
+ *          enabled.
+ *
+ * This macro copies a single simple value from kernel space to user
+ * space.  It supports simple types like char and int, but not larger
+ * data types like structures or arrays.
+ *
+ * @ptr must have pointer-to-simple-variable type, and @x must be assignable
+ * to the result of dereferencing @ptr.
+ *
+ * Caller must check the pointer with access_ok() before calling this
+ * function.
+ *
+ * Returns zero on success, or -EFAULT on error.
+ */
+#define __put_user(x, ptr)						\
+({									\
+	int __ret;							\
+	typeof(*(ptr)) _x = (x);					\
+	__chk_user_ptr(ptr);						\
+	switch (sizeof(*(ptr))) {					\
+	case 1: __put_user_1(_x, ptr, __ret); break;			\
+	case 2: __put_user_2(_x, ptr, __ret); break;			\
+	case 4: __put_user_4(_x, ptr, __ret); break;			\
+	case 8: __put_user_8(_x, ptr, __ret); break;			\
+	default: __ret = __put_user_bad(); break;			\
+	}								\
+	__ret;								\
+})
+
+/*
+ * The versions of get_user and put_user without initial underscores
+ * check the address of their arguments to make sure they are not
+ * in kernel space.
+ */
+#define put_user(x, ptr)						\
+({									\
+	__typeof__(*(ptr)) __user *__Pu_addr = (ptr);			\
+	access_ok(VERIFY_WRITE, (__Pu_addr), sizeof(*(__Pu_addr))) ?	\
+		__put_user((x), (__Pu_addr)) :				\
+		-EFAULT;						\
+})
+
+#define get_user(x, ptr)						\
+({									\
+	__typeof__(*(ptr)) const __user *__Gu_addr = (ptr);		\
+	access_ok(VERIFY_READ, (__Gu_addr), sizeof(*(__Gu_addr))) ?	\
+		__get_user((x), (__Gu_addr)) :				\
+		((x) = 0, -EFAULT);					\
+})
+
+/**
+ * __copy_to_user() - copy data into user space, with less checking.
+ * @to:   Destination address, in user space.
+ * @from: Source address, in kernel space.
+ * @n:    Number of bytes to copy.
+ *
+ * Context: User context only. This function may sleep if pagefaults are
+ *          enabled.
+ *
+ * Copy data from kernel space to user space.  Caller must check
+ * the specified block with access_ok() before calling this function.
+ *
+ * Returns number of bytes that could not be copied.
+ * On success, this will be zero.
+ *
+ * An alternate version - __copy_to_user_inatomic() - is designed
+ * to be called from atomic context, typically bracketed by calls
+ * to pagefault_disable() and pagefault_enable().
+ */
+extern unsigned long __must_check __copy_to_user_inatomic(
+	void __user *to, const void *from, unsigned long n);
+
+static inline unsigned long __must_check
+__copy_to_user(void __user *to, const void *from, unsigned long n)
+{
+	might_fault();
+	return __copy_to_user_inatomic(to, from, n);
+}
+
+static inline unsigned long __must_check
+copy_to_user(void __user *to, const void *from, unsigned long n)
+{
+	if (access_ok(VERIFY_WRITE, to, n))
+		n = __copy_to_user(to, from, n);
+	return n;
+}
+
+/**
+ * __copy_from_user() - copy data from user space, with less checking.
+ * @to:   Destination address, in kernel space.
+ * @from: Source address, in user space.
+ * @n:    Number of bytes to copy.
+ *
+ * Context: User context only. This function may sleep if pagefaults are
+ *          enabled.
+ *
+ * Copy data from user space to kernel space.  Caller must check
+ * the specified block with access_ok() before calling this function.
+ *
+ * Returns number of bytes that could not be copied.
+ * On success, this will be zero.
+ *
+ * If some data could not be copied, this function will pad the copied
+ * data to the requested size using zero bytes.
+ *
+ * An alternate version - __copy_from_user_inatomic() - is designed
+ * to be called from atomic context, typically bracketed by calls
+ * to pagefault_disable() and pagefault_enable().  This version
+ * does *NOT* pad with zeros.
+ */
+extern unsigned long __must_check __copy_from_user_inatomic(
+	void *to, const void __user *from, unsigned long n);
+extern unsigned long __must_check __copy_from_user_zeroing(
+	void *to, const void __user *from, unsigned long n);
+
+static inline unsigned long __must_check
+__copy_from_user(void *to, const void __user *from, unsigned long n)
+{
+       might_fault();
+       return __copy_from_user_zeroing(to, from, n);
+}
+
+static inline unsigned long __must_check
+_copy_from_user(void *to, const void __user *from, unsigned long n)
+{
+	if (access_ok(VERIFY_READ, from, n))
+		n = __copy_from_user(to, from, n);
+	else
+		memset(to, 0, n);
+	return n;
+}
+
+#ifdef CONFIG_DEBUG_STRICT_USER_COPY_CHECKS
+/*
+ * There are still unprovable places in the generic code as of 2.6.34, so this
+ * option is not really compatible with -Werror, which is more useful in
+ * general.
+ */
+extern void copy_from_user_overflow(void)
+	__compiletime_warning("copy_from_user() size is not provably correct");
+
+static inline unsigned long __must_check copy_from_user(void *to,
+					  const void __user *from,
+					  unsigned long n)
+{
+	int sz = __compiletime_object_size(to);
+
+	if (likely(sz == -1 || sz >= n))
+		n = _copy_from_user(to, from, n);
+	else
+		copy_from_user_overflow();
+
+	return n;
+}
+#else
+#define copy_from_user _copy_from_user
+#endif
+
+#ifdef __tilegx__
+/**
+ * __copy_in_user() - copy data within user space, with less checking.
+ * @to:   Destination address, in user space.
+ * @from: Source address, in user space.
+ * @n:    Number of bytes to copy.
+ *
+ * Context: User context only. This function may sleep if pagefaults are
+ *          enabled.
+ *
+ * Copy data from user space to user space.  Caller must check
+ * the specified blocks with access_ok() before calling this function.
+ *
+ * Returns number of bytes that could not be copied.
+ * On success, this will be zero.
+ */
+extern unsigned long __copy_in_user_inatomic(
+	void __user *to, const void __user *from, unsigned long n);
+
+static inline unsigned long __must_check
+__copy_in_user(void __user *to, const void __user *from, unsigned long n)
+{
+	might_fault();
+	return __copy_in_user_inatomic(to, from, n);
+}
+
+static inline unsigned long __must_check
+copy_in_user(void __user *to, const void __user *from, unsigned long n)
+{
+	if (access_ok(VERIFY_WRITE, to, n) && access_ok(VERIFY_READ, from, n))
+		n = __copy_in_user(to, from, n);
+	return n;
+}
+#endif
+
+
+/**
+ * strlen_user: - Get the size of a string in user space.
+ * @str: The string to measure.
+ *
+ * Context: User context only.  This function may sleep.
+ *
+ * Get the size of a NUL-terminated string in user space.
+ *
+ * Returns the size of the string INCLUDING the terminating NUL.
+ * On exception, returns 0.
+ *
+ * If there is a limit on the length of a valid string, you may wish to
+ * consider using strnlen_user() instead.
+ */
+extern long strnlen_user_asm(const char __user *str, long n);
+static inline long __must_check strnlen_user(const char __user *str, long n)
+{
+	might_fault();
+	return strnlen_user_asm(str, n);
+}
+#define strlen_user(str) strnlen_user(str, LONG_MAX)
+
+/**
+ * strncpy_from_user: - Copy a NUL terminated string from userspace, with less checking.
+ * @dst:   Destination address, in kernel space.  This buffer must be at
+ *         least @count bytes long.
+ * @src:   Source address, in user space.
+ * @count: Maximum number of bytes to copy, including the trailing NUL.
+ *
+ * Copies a NUL-terminated string from userspace to kernel space.
+ * Caller must check the specified block with access_ok() before calling
+ * this function.
+ *
+ * On success, returns the length of the string (not including the trailing
+ * NUL).
+ *
+ * If access to userspace fails, returns -EFAULT (some data may have been
+ * copied).
+ *
+ * If @count is smaller than the length of the string, copies @count bytes
+ * and returns @count.
+ */
+extern long strncpy_from_user_asm(char *dst, const char __user *src, long);
+static inline long __must_check __strncpy_from_user(
+	char *dst, const char __user *src, long count)
+{
+	might_fault();
+	return strncpy_from_user_asm(dst, src, count);
+}
+static inline long __must_check strncpy_from_user(
+	char *dst, const char __user *src, long count)
+{
+	if (access_ok(VERIFY_READ, src, 1))
+		return __strncpy_from_user(dst, src, count);
+	return -EFAULT;
+}
+
+/**
+ * clear_user: - Zero a block of memory in user space.
+ * @mem:   Destination address, in user space.
+ * @len:   Number of bytes to zero.
+ *
+ * Zero a block of memory in user space.
+ *
+ * Returns number of bytes that could not be cleared.
+ * On success, this will be zero.
+ */
+extern unsigned long clear_user_asm(void __user *mem, unsigned long len);
+static inline unsigned long __must_check __clear_user(
+	void __user *mem, unsigned long len)
+{
+	might_fault();
+	return clear_user_asm(mem, len);
+}
+static inline unsigned long __must_check clear_user(
+	void __user *mem, unsigned long len)
+{
+	if (access_ok(VERIFY_WRITE, mem, len))
+		return __clear_user(mem, len);
+	return len;
+}
+
+/**
+ * flush_user: - Flush a block of memory in user space from cache.
+ * @mem:   Destination address, in user space.
+ * @len:   Number of bytes to flush.
+ *
+ * Returns number of bytes that could not be flushed.
+ * On success, this will be zero.
+ */
+extern unsigned long flush_user_asm(void __user *mem, unsigned long len);
+static inline unsigned long __must_check __flush_user(
+	void __user *mem, unsigned long len)
+{
+	int retval;
+
+	might_fault();
+	retval = flush_user_asm(mem, len);
+	mb_incoherent();
+	return retval;
+}
+
+static inline unsigned long __must_check flush_user(
+	void __user *mem, unsigned long len)
+{
+	if (access_ok(VERIFY_WRITE, mem, len))
+		return __flush_user(mem, len);
+	return len;
+}
+
+/**
+ * finv_user: - Flush-inval a block of memory in user space from cache.
+ * @mem:   Destination address, in user space.
+ * @len:   Number of bytes to invalidate.
+ *
+ * Returns number of bytes that could not be flush-invalidated.
+ * On success, this will be zero.
+ */
+extern unsigned long finv_user_asm(void __user *mem, unsigned long len);
+static inline unsigned long __must_check __finv_user(
+	void __user *mem, unsigned long len)
+{
+	int retval;
+
+	might_fault();
+	retval = finv_user_asm(mem, len);
+	mb_incoherent();
+	return retval;
+}
+static inline unsigned long __must_check finv_user(
+	void __user *mem, unsigned long len)
+{
+	if (access_ok(VERIFY_WRITE, mem, len))
+		return __finv_user(mem, len);
+	return len;
+}
+
+#endif /* _ASM_TILE_UACCESS_H */
diff --git a/kernel/arch/tile/include/asm/unaligned.h b/kernel/arch/tile/include/asm/unaligned.h
new file mode 100644
index 000000000..5a58a0d11
--- /dev/null
+++ b/kernel/arch/tile/include/asm/unaligned.h
@@ -0,0 +1,43 @@
+/*
+ * Copyright 2010 Tilera Corporation. All Rights Reserved.
+ *
+ *   This program is free software; you can redistribute it and/or
+ *   modify it under the terms of the GNU General Public License
+ *   as published by the Free Software Foundation, version 2.
+ *
+ *   This program is distributed in the hope that it will be useful, but
+ *   WITHOUT ANY WARRANTY; without even the implied warranty of
+ *   MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
+ *   NON INFRINGEMENT.  See the GNU General Public License for
+ *   more details.
+ */
+
+#ifndef _ASM_TILE_UNALIGNED_H
+#define _ASM_TILE_UNALIGNED_H
+
+/*
+ * We could implement faster get_unaligned_[be/le]64 using the ldna
+ * instruction on tilegx; however, we need to either copy all of the
+ * other generic functions to here (which is pretty ugly) or else
+ * modify both the generic code and other arch code to allow arch
+ * specific unaligned data access functions.  Given these functions
+ * are not often called, we'll stick with the generic version.
+ */
+#include <asm-generic/unaligned.h>
+
+/*
+ * Is the kernel doing fixups of unaligned accesses?  If <0, no kernel
+ * intervention occurs and SIGBUS is delivered with no data address
+ * info.  If 0, the kernel single-steps the instruction to discover
+ * the data address to provide with the SIGBUS.  If 1, the kernel does
+ * a fixup.
+ */
+extern int unaligned_fixup;
+
+/* Is the kernel printing on each unaligned fixup? */
+extern int unaligned_printk;
+
+/* Number of unaligned fixups performed */
+extern unsigned int unaligned_fixup_count;
+
+#endif /* _ASM_TILE_UNALIGNED_H */
diff --git a/kernel/arch/tile/include/asm/unistd.h b/kernel/arch/tile/include/asm/unistd.h
new file mode 100644
index 000000000..940831fe9
--- /dev/null
+++ b/kernel/arch/tile/include/asm/unistd.h
@@ -0,0 +1,20 @@
+/*
+ * Copyright 2010 Tilera Corporation. All Rights Reserved.
+ *
+ *   This program is free software; you can redistribute it and/or
+ *   modify it under the terms of the GNU General Public License
+ *   as published by the Free Software Foundation, version 2.
+ *
+ *   This program is distributed in the hope that it will be useful, but
+ *   WITHOUT ANY WARRANTY; without even the implied warranty of
+ *   MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
+ *   NON INFRINGEMENT.  See the GNU General Public License for
+ *   more details.
+ */
+/* In compat mode, we use sys_llseek() for compat_sys_llseek(). */
+#ifdef CONFIG_COMPAT
+#define __ARCH_WANT_SYS_LLSEEK
+#endif
+#define __ARCH_WANT_SYS_NEWFSTATAT
+#define __ARCH_WANT_SYS_CLONE
+#include <uapi/asm/unistd.h>
diff --git a/kernel/arch/tile/include/asm/user.h b/kernel/arch/tile/include/asm/user.h
new file mode 100644
index 000000000..cbc8b4d5a
--- /dev/null
+++ b/kernel/arch/tile/include/asm/user.h
@@ -0,0 +1,21 @@
+/*
+ * Copyright 2010 Tilera Corporation. All Rights Reserved.
+ *
+ *   This program is free software; you can redistribute it and/or
+ *   modify it under the terms of the GNU General Public License
+ *   as published by the Free Software Foundation, version 2.
+ *
+ *   This program is distributed in the hope that it will be useful, but
+ *   WITHOUT ANY WARRANTY; without even the implied warranty of
+ *   MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
+ *   NON INFRINGEMENT.  See the GNU General Public License for
+ *   more details.
+ *
+ */
+
+#ifndef _ASM_TILE_USER_H
+#define _ASM_TILE_USER_H
+
+/* This header is for a.out file formats, which TILE does not support. */
+
+#endif /* _ASM_TILE_USER_H */
diff --git a/kernel/arch/tile/include/asm/vdso.h b/kernel/arch/tile/include/asm/vdso.h
new file mode 100644
index 000000000..9b0696921
--- /dev/null
+++ b/kernel/arch/tile/include/asm/vdso.h
@@ -0,0 +1,55 @@
+/*
+ * Copyright 2012 Tilera Corporation. All Rights Reserved.
+ *
+ *   This program is free software; you can redistribute it and/or
+ *   modify it under the terms of the GNU General Public License
+ *   as published by the Free Software Foundation, version 2.
+ *
+ *   This program is distributed in the hope that it will be useful, but
+ *   WITHOUT ANY WARRANTY; without even the implied warranty of
+ *   MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
+ *   NON INFRINGEMENT.  See the GNU General Public License for
+ *   more details.
+ */
+
+#ifndef __TILE_VDSO_H__
+#define __TILE_VDSO_H__
+
+#include <linux/seqlock.h>
+#include <linux/types.h>
+
+/*
+ * Note about the vdso_data structure:
+ *
+ * NEVER USE THEM IN USERSPACE CODE DIRECTLY. The layout of the
+ * structure is supposed to be known only to the function in the vdso
+ * itself and may change without notice.
+ */
+
+struct vdso_data {
+	seqcount_t tz_seq;	/* Timezone seqlock                   */
+	seqcount_t tb_seq;	/* Timebase seqlock                   */
+	__u64 cycle_last;       /* TOD clock for xtime                */
+	__u64 mask;             /* Cycle mask                         */
+	__u32 mult;             /* Cycle to nanosecond multiplier     */
+	__u32 shift;            /* Cycle to nanosecond divisor (power of two) */
+	__u64 wall_time_sec;
+	__u64 wall_time_snsec;
+	__u64 monotonic_time_sec;
+	__u64 monotonic_time_snsec;
+	__u64 wall_time_coarse_sec;
+	__u64 wall_time_coarse_nsec;
+	__u64 monotonic_time_coarse_sec;
+	__u64 monotonic_time_coarse_nsec;
+	__u32 tz_minuteswest;   /* Minutes west of Greenwich          */
+	__u32 tz_dsttime;       /* Type of dst correction             */
+};
+
+extern struct vdso_data *vdso_data;
+
+/* __vdso_rt_sigreturn is defined with the addresses in the vdso page. */
+extern void __vdso_rt_sigreturn(void);
+
+extern int setup_vdso_pages(void);
+
+#endif /* __TILE_VDSO_H__ */
diff --git a/kernel/arch/tile/include/asm/vga.h b/kernel/arch/tile/include/asm/vga.h
new file mode 100644
index 000000000..7b46e754d
--- /dev/null
+++ b/kernel/arch/tile/include/asm/vga.h
@@ -0,0 +1,39 @@
+/*
+ * Copyright 2010 Tilera Corporation. All Rights Reserved.
+ *
+ *   This program is free software; you can redistribute it and/or
+ *   modify it under the terms of the GNU General Public License
+ *   as published by the Free Software Foundation, version 2.
+ *
+ *   This program is distributed in the hope that it will be useful, but
+ *   WITHOUT ANY WARRANTY; without even the implied warranty of
+ *   MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
+ *   NON INFRINGEMENT.  See the GNU General Public License for
+ *   more details.
+ *
+ * Access to VGA videoram.
+ */
+
+#ifndef _ASM_TILE_VGA_H
+#define _ASM_TILE_VGA_H
+
+#include <asm/io.h>
+
+#define VT_BUF_HAVE_RW
+
+static inline void scr_writew(u16 val, volatile u16 *addr)
+{
+	__raw_writew(val, (volatile u16 __iomem *) addr);
+}
+
+static inline u16 scr_readw(volatile const u16 *addr)
+{
+	return __raw_readw((volatile const u16 __iomem *) addr);
+}
+
+#define vga_readb(a)	readb((u8 __iomem *)(a))
+#define vga_writeb(v,a)	writeb(v, (u8 __iomem *)(a))
+
+#define VGA_MAP_MEM(x,s)	((unsigned long) ioremap(x, s))
+
+#endif
diff --git a/kernel/arch/tile/include/gxio/common.h b/kernel/arch/tile/include/gxio/common.h
new file mode 100644
index 000000000..724595a24
--- /dev/null
+++ b/kernel/arch/tile/include/gxio/common.h
@@ -0,0 +1,40 @@
+/*
+ * Copyright 2012 Tilera Corporation. All Rights Reserved.
+ *
+ *   This program is free software; you can redistribute it and/or
+ *   modify it under the terms of the GNU General Public License
+ *   as published by the Free Software Foundation, version 2.
+ *
+ *   This program is distributed in the hope that it will be useful, but
+ *   WITHOUT ANY WARRANTY; without even the implied warranty of
+ *   MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
+ *   NON INFRINGEMENT.  See the GNU General Public License for
+ *   more details.
+ */
+
+#ifndef _GXIO_COMMON_H_
+#define _GXIO_COMMON_H_
+
+/*
+ * Routines shared between the various GXIO device components.
+ */
+
+#include <hv/iorpc.h>
+
+#include <linux/types.h>
+#include <linux/compiler.h>
+#include <linux/io.h>
+
+/* Define the standard gxio MMIO functions using kernel functions. */
+#define __gxio_mmio_read8(addr)		readb(addr)
+#define __gxio_mmio_read16(addr)	readw(addr)
+#define __gxio_mmio_read32(addr)	readl(addr)
+#define __gxio_mmio_read64(addr)	readq(addr)
+#define __gxio_mmio_write8(addr, val)	writeb((val), (addr))
+#define __gxio_mmio_write16(addr, val)	writew((val), (addr))
+#define __gxio_mmio_write32(addr, val)	writel((val), (addr))
+#define __gxio_mmio_write64(addr, val)	writeq((val), (addr))
+#define __gxio_mmio_read(addr)		__gxio_mmio_read64(addr)
+#define __gxio_mmio_write(addr, val)	__gxio_mmio_write64((addr), (val))
+
+#endif /* !_GXIO_COMMON_H_ */
diff --git a/kernel/arch/tile/include/gxio/dma_queue.h b/kernel/arch/tile/include/gxio/dma_queue.h
new file mode 100644
index 000000000..b9e45e376
--- /dev/null
+++ b/kernel/arch/tile/include/gxio/dma_queue.h
@@ -0,0 +1,161 @@
+/*
+ * Copyright 2012 Tilera Corporation. All Rights Reserved.
+ *
+ *   This program is free software; you can redistribute it and/or
+ *   modify it under the terms of the GNU General Public License
+ *   as published by the Free Software Foundation, version 2.
+ *
+ *   This program is distributed in the hope that it will be useful, but
+ *   WITHOUT ANY WARRANTY; without even the implied warranty of
+ *   MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
+ *   NON INFRINGEMENT.  See the GNU General Public License for
+ *   more details.
+ */
+
+#ifndef _GXIO_DMA_QUEUE_H_
+#define _GXIO_DMA_QUEUE_H_
+
+/*
+ * DMA queue management APIs shared between TRIO and mPIPE.
+ */
+
+#include <gxio/common.h>
+
+/* The credit counter lives in the high 32 bits. */
+#define DMA_QUEUE_CREDIT_SHIFT 32
+
+/*
+ * State object that tracks a DMA queue's head and tail indices, as
+ * well as the number of commands posted and completed.  The
+ * structure is accessed via a thread-safe, lock-free algorithm.
+ */
+typedef struct {
+	/*
+	 * Address of a MPIPE_EDMA_POST_REGION_VAL_t,
+	 * TRIO_PUSH_DMA_REGION_VAL_t, or TRIO_PULL_DMA_REGION_VAL_t
+	 * register.  These register have identical encodings and provide
+	 * information about how many commands have been processed.
+	 */
+	void *post_region_addr;
+
+	/*
+	 * A lazily-updated count of how many edescs the hardware has
+	 * completed.
+	 */
+	uint64_t hw_complete_count __attribute__ ((aligned(64)));
+
+	/*
+	 * High 32 bits are a count of available egress command credits,
+	 * low 24 bits are the next egress "slot".
+	 */
+	int64_t credits_and_next_index;
+
+} __gxio_dma_queue_t;
+
+/* Initialize a dma queue. */
+extern void __gxio_dma_queue_init(__gxio_dma_queue_t *dma_queue,
+				  void *post_region_addr,
+				  unsigned int num_entries);
+
+/*
+ * Update the "credits_and_next_index" and "hw_complete_count" fields
+ * based on pending hardware completions.  Note that some other thread
+ * may have already done this and, importantly, may still be in the
+ * process of updating "credits_and_next_index".
+ */
+extern void __gxio_dma_queue_update_credits(__gxio_dma_queue_t *dma_queue);
+
+/* Wait for credits to become available. */
+extern int64_t __gxio_dma_queue_wait_for_credits(__gxio_dma_queue_t *dma_queue,
+						 int64_t modifier);
+
+/* Reserve slots in the queue, optionally waiting for slots to become
+ * available, and optionally returning a "completion_slot" suitable for
+ * direct comparison to "hw_complete_count".
+ */
+static inline int64_t __gxio_dma_queue_reserve(__gxio_dma_queue_t *dma_queue,
+					       unsigned int num, bool wait,
+					       bool completion)
+{
+	uint64_t slot;
+
+	/*
+	 * Try to reserve 'num' egress command slots.  We do this by
+	 * constructing a constant that subtracts N credits and adds N to
+	 * the index, and using fetchaddgez to only apply it if the credits
+	 * count doesn't go negative.
+	 */
+	int64_t modifier = (((int64_t)(-num)) << DMA_QUEUE_CREDIT_SHIFT) | num;
+	int64_t old =
+		__insn_fetchaddgez(&dma_queue->credits_and_next_index,
+				   modifier);
+
+	if (unlikely(old + modifier < 0)) {
+		/*
+		 * We're out of credits.  Try once to get more by checking for
+		 * completed egress commands.  If that fails, wait or fail.
+		 */
+		__gxio_dma_queue_update_credits(dma_queue);
+		old = __insn_fetchaddgez(&dma_queue->credits_and_next_index,
+					 modifier);
+		if (old + modifier < 0) {
+			if (wait)
+				old = __gxio_dma_queue_wait_for_credits
+					(dma_queue, modifier);
+			else
+				return GXIO_ERR_DMA_CREDITS;
+		}
+	}
+
+	/* The bottom 24 bits of old encode the "slot". */
+	slot = (old & 0xffffff);
+
+	if (completion) {
+		/*
+		 * A "completion_slot" is a "slot" which can be compared to
+		 * "hw_complete_count" at any time in the future.  To convert
+		 * "slot" into a "completion_slot", we access "hw_complete_count"
+		 * once (knowing that we have reserved a slot, and thus, it will
+		 * be "basically" accurate), and combine its high 40 bits with
+		 * the 24 bit "slot", and handle "wrapping" by adding "1 << 24"
+		 * if the result is LESS than "hw_complete_count".
+		 */
+		uint64_t complete;
+		complete = ACCESS_ONCE(dma_queue->hw_complete_count);
+		slot |= (complete & 0xffffffffff000000);
+		if (slot < complete)
+			slot += 0x1000000;
+	}
+
+	/*
+	 * If any of our slots mod 256 were equivalent to 0, go ahead and
+	 * collect some egress credits, and update "hw_complete_count", and
+	 * make sure the index doesn't overflow into the credits.
+	 */
+	if (unlikely(((old + num) & 0xff) < num)) {
+		__gxio_dma_queue_update_credits(dma_queue);
+
+		/* Make sure the index doesn't overflow into the credits. */
+#ifdef __BIG_ENDIAN__
+		*(((uint8_t *)&dma_queue->credits_and_next_index) + 4) = 0;
+#else
+		*(((uint8_t *)&dma_queue->credits_and_next_index) + 3) = 0;
+#endif
+	}
+
+	return slot;
+}
+
+/* Non-inlinable "__gxio_dma_queue_reserve(..., true)". */
+extern int64_t __gxio_dma_queue_reserve_aux(__gxio_dma_queue_t *dma_queue,
+					    unsigned int num, int wait);
+
+/* Check whether a particular "completion slot" has completed.
+ *
+ * Note that this function requires a "completion slot", and thus
+ * cannot be used with the result of any "reserve_fast" function.
+ */
+extern int __gxio_dma_queue_is_complete(__gxio_dma_queue_t *dma_queue,
+					int64_t completion_slot, int update);
+
+#endif /* !_GXIO_DMA_QUEUE_H_ */
diff --git a/kernel/arch/tile/include/gxio/iorpc_globals.h b/kernel/arch/tile/include/gxio/iorpc_globals.h
new file mode 100644
index 000000000..52c721f8d
--- /dev/null
+++ b/kernel/arch/tile/include/gxio/iorpc_globals.h
@@ -0,0 +1,38 @@
+/*
+ * Copyright 2012 Tilera Corporation. All Rights Reserved.
+ *
+ *   This program is free software; you can redistribute it and/or
+ *   modify it under the terms of the GNU General Public License
+ *   as published by the Free Software Foundation, version 2.
+ *
+ *   This program is distributed in the hope that it will be useful, but
+ *   WITHOUT ANY WARRANTY; without even the implied warranty of
+ *   MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
+ *   NON INFRINGEMENT.  See the GNU General Public License for
+ *   more details.
+ */
+
+/* This file is machine-generated; DO NOT EDIT! */
+#ifndef __IORPC_LINUX_RPC_H__
+#define __IORPC_LINUX_RPC_H__
+
+#include <hv/iorpc.h>
+
+#include <linux/string.h>
+#include <linux/module.h>
+#include <asm/pgtable.h>
+
+#define IORPC_OP_ARM_POLLFD            IORPC_OPCODE(IORPC_FORMAT_KERNEL_POLLFD, 0x9000)
+#define IORPC_OP_CLOSE_POLLFD          IORPC_OPCODE(IORPC_FORMAT_KERNEL_POLLFD, 0x9001)
+#define IORPC_OP_GET_MMIO_BASE         IORPC_OPCODE(IORPC_FORMAT_NONE_NOUSER, 0x8000)
+#define IORPC_OP_CHECK_MMIO_OFFSET     IORPC_OPCODE(IORPC_FORMAT_NONE_NOUSER, 0x8001)
+
+int __iorpc_arm_pollfd(int fd, int pollfd_cookie);
+
+int __iorpc_close_pollfd(int fd, int pollfd_cookie);
+
+int __iorpc_get_mmio_base(int fd, HV_PTE *base);
+
+int __iorpc_check_mmio_offset(int fd, unsigned long offset, unsigned long size);
+
+#endif /* !__IORPC_LINUX_RPC_H__ */
diff --git a/kernel/arch/tile/include/gxio/iorpc_mpipe.h b/kernel/arch/tile/include/gxio/iorpc_mpipe.h
new file mode 100644
index 000000000..4cda03de7
--- /dev/null
+++ b/kernel/arch/tile/include/gxio/iorpc_mpipe.h
@@ -0,0 +1,144 @@
+/*
+ * Copyright 2012 Tilera Corporation. All Rights Reserved.
+ *
+ *   This program is free software; you can redistribute it and/or
+ *   modify it under the terms of the GNU General Public License
+ *   as published by the Free Software Foundation, version 2.
+ *
+ *   This program is distributed in the hope that it will be useful, but
+ *   WITHOUT ANY WARRANTY; without even the implied warranty of
+ *   MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
+ *   NON INFRINGEMENT.  See the GNU General Public License for
+ *   more details.
+ */
+
+/* This file is machine-generated; DO NOT EDIT! */
+#ifndef __GXIO_MPIPE_LINUX_RPC_H__
+#define __GXIO_MPIPE_LINUX_RPC_H__
+
+#include <hv/iorpc.h>
+
+#include <hv/drv_mpipe_intf.h>
+#include <asm/page.h>
+#include <gxio/kiorpc.h>
+#include <gxio/mpipe.h>
+#include <linux/string.h>
+#include <linux/module.h>
+#include <asm/pgtable.h>
+
+#define GXIO_MPIPE_OP_ALLOC_BUFFER_STACKS IORPC_OPCODE(IORPC_FORMAT_NONE, 0x1200)
+#define GXIO_MPIPE_OP_INIT_BUFFER_STACK_AUX IORPC_OPCODE(IORPC_FORMAT_KERNEL_MEM, 0x1201)
+
+#define GXIO_MPIPE_OP_ALLOC_NOTIF_RINGS IORPC_OPCODE(IORPC_FORMAT_NONE, 0x1203)
+#define GXIO_MPIPE_OP_INIT_NOTIF_RING_AUX IORPC_OPCODE(IORPC_FORMAT_KERNEL_MEM, 0x1204)
+#define GXIO_MPIPE_OP_REQUEST_NOTIF_RING_INTERRUPT IORPC_OPCODE(IORPC_FORMAT_KERNEL_INTERRUPT, 0x1205)
+#define GXIO_MPIPE_OP_ENABLE_NOTIF_RING_INTERRUPT IORPC_OPCODE(IORPC_FORMAT_NONE, 0x1206)
+#define GXIO_MPIPE_OP_ALLOC_NOTIF_GROUPS IORPC_OPCODE(IORPC_FORMAT_NONE, 0x1207)
+#define GXIO_MPIPE_OP_INIT_NOTIF_GROUP IORPC_OPCODE(IORPC_FORMAT_NONE, 0x1208)
+#define GXIO_MPIPE_OP_ALLOC_BUCKETS    IORPC_OPCODE(IORPC_FORMAT_NONE, 0x1209)
+#define GXIO_MPIPE_OP_INIT_BUCKET      IORPC_OPCODE(IORPC_FORMAT_NONE, 0x120a)
+#define GXIO_MPIPE_OP_ALLOC_EDMA_RINGS IORPC_OPCODE(IORPC_FORMAT_NONE, 0x120b)
+#define GXIO_MPIPE_OP_INIT_EDMA_RING_AUX IORPC_OPCODE(IORPC_FORMAT_KERNEL_MEM, 0x120c)
+
+#define GXIO_MPIPE_OP_COMMIT_RULES     IORPC_OPCODE(IORPC_FORMAT_NONE, 0x120f)
+#define GXIO_MPIPE_OP_REGISTER_CLIENT_MEMORY IORPC_OPCODE(IORPC_FORMAT_NONE_NOUSER, 0x1210)
+#define GXIO_MPIPE_OP_LINK_OPEN_AUX    IORPC_OPCODE(IORPC_FORMAT_NONE, 0x1211)
+#define GXIO_MPIPE_OP_LINK_CLOSE_AUX   IORPC_OPCODE(IORPC_FORMAT_NONE, 0x1212)
+#define GXIO_MPIPE_OP_LINK_SET_ATTR_AUX IORPC_OPCODE(IORPC_FORMAT_NONE, 0x1213)
+
+#define GXIO_MPIPE_OP_GET_TIMESTAMP_AUX IORPC_OPCODE(IORPC_FORMAT_NONE, 0x121e)
+#define GXIO_MPIPE_OP_SET_TIMESTAMP_AUX IORPC_OPCODE(IORPC_FORMAT_NONE, 0x121f)
+#define GXIO_MPIPE_OP_ADJUST_TIMESTAMP_AUX IORPC_OPCODE(IORPC_FORMAT_NONE, 0x1220)
+#define GXIO_MPIPE_OP_CONFIG_EDMA_RING_BLKS IORPC_OPCODE(IORPC_FORMAT_NONE, 0x1221)
+#define GXIO_MPIPE_OP_ADJUST_TIMESTAMP_FREQ IORPC_OPCODE(IORPC_FORMAT_NONE, 0x1222)
+#define GXIO_MPIPE_OP_ARM_POLLFD       IORPC_OPCODE(IORPC_FORMAT_KERNEL_POLLFD, 0x9000)
+#define GXIO_MPIPE_OP_CLOSE_POLLFD     IORPC_OPCODE(IORPC_FORMAT_KERNEL_POLLFD, 0x9001)
+#define GXIO_MPIPE_OP_GET_MMIO_BASE    IORPC_OPCODE(IORPC_FORMAT_NONE_NOUSER, 0x8000)
+#define GXIO_MPIPE_OP_CHECK_MMIO_OFFSET IORPC_OPCODE(IORPC_FORMAT_NONE_NOUSER, 0x8001)
+
+int gxio_mpipe_alloc_buffer_stacks(gxio_mpipe_context_t *context,
+				   unsigned int count, unsigned int first,
+				   unsigned int flags);
+
+int gxio_mpipe_init_buffer_stack_aux(gxio_mpipe_context_t *context,
+				     void *mem_va, size_t mem_size,
+				     unsigned int mem_flags, unsigned int stack,
+				     unsigned int buffer_size_enum);
+
+
+int gxio_mpipe_alloc_notif_rings(gxio_mpipe_context_t *context,
+				 unsigned int count, unsigned int first,
+				 unsigned int flags);
+
+int gxio_mpipe_init_notif_ring_aux(gxio_mpipe_context_t *context, void *mem_va,
+				   size_t mem_size, unsigned int mem_flags,
+				   unsigned int ring);
+
+int gxio_mpipe_request_notif_ring_interrupt(gxio_mpipe_context_t *context,
+					    int inter_x, int inter_y,
+					    int inter_ipi, int inter_event,
+					    unsigned int ring);
+
+int gxio_mpipe_enable_notif_ring_interrupt(gxio_mpipe_context_t *context,
+					   unsigned int ring);
+
+int gxio_mpipe_alloc_notif_groups(gxio_mpipe_context_t *context,
+				  unsigned int count, unsigned int first,
+				  unsigned int flags);
+
+int gxio_mpipe_init_notif_group(gxio_mpipe_context_t *context,
+				unsigned int group,
+				gxio_mpipe_notif_group_bits_t bits);
+
+int gxio_mpipe_alloc_buckets(gxio_mpipe_context_t *context, unsigned int count,
+			     unsigned int first, unsigned int flags);
+
+int gxio_mpipe_init_bucket(gxio_mpipe_context_t *context, unsigned int bucket,
+			   MPIPE_LBL_INIT_DAT_BSTS_TBL_t bucket_info);
+
+int gxio_mpipe_alloc_edma_rings(gxio_mpipe_context_t *context,
+				unsigned int count, unsigned int first,
+				unsigned int flags);
+
+int gxio_mpipe_init_edma_ring_aux(gxio_mpipe_context_t *context, void *mem_va,
+				  size_t mem_size, unsigned int mem_flags,
+				  unsigned int ring, unsigned int channel);
+
+
+int gxio_mpipe_commit_rules(gxio_mpipe_context_t *context, const void *blob,
+			    size_t blob_size);
+
+int gxio_mpipe_register_client_memory(gxio_mpipe_context_t *context,
+				      unsigned int iotlb, HV_PTE pte,
+				      unsigned int flags);
+
+int gxio_mpipe_link_open_aux(gxio_mpipe_context_t *context,
+			     _gxio_mpipe_link_name_t name, unsigned int flags);
+
+int gxio_mpipe_link_close_aux(gxio_mpipe_context_t *context, int mac);
+
+int gxio_mpipe_link_set_attr_aux(gxio_mpipe_context_t *context, int mac,
+				 uint32_t attr, int64_t val);
+
+int gxio_mpipe_get_timestamp_aux(gxio_mpipe_context_t *context, uint64_t *sec,
+				 uint64_t *nsec, uint64_t *cycles);
+
+int gxio_mpipe_set_timestamp_aux(gxio_mpipe_context_t *context, uint64_t sec,
+				 uint64_t nsec, uint64_t cycles);
+
+int gxio_mpipe_adjust_timestamp_aux(gxio_mpipe_context_t *context,
+				    int64_t nsec);
+
+int gxio_mpipe_adjust_timestamp_freq(gxio_mpipe_context_t *context,
+				     int32_t ppb);
+
+int gxio_mpipe_arm_pollfd(gxio_mpipe_context_t *context, int pollfd_cookie);
+
+int gxio_mpipe_close_pollfd(gxio_mpipe_context_t *context, int pollfd_cookie);
+
+int gxio_mpipe_get_mmio_base(gxio_mpipe_context_t *context, HV_PTE *base);
+
+int gxio_mpipe_check_mmio_offset(gxio_mpipe_context_t *context,
+				 unsigned long offset, unsigned long size);
+
+#endif /* !__GXIO_MPIPE_LINUX_RPC_H__ */
diff --git a/kernel/arch/tile/include/gxio/iorpc_mpipe_info.h b/kernel/arch/tile/include/gxio/iorpc_mpipe_info.h
new file mode 100644
index 000000000..f0b042844
--- /dev/null
+++ b/kernel/arch/tile/include/gxio/iorpc_mpipe_info.h
@@ -0,0 +1,50 @@
+/*
+ * Copyright 2012 Tilera Corporation. All Rights Reserved.
+ *
+ *   This program is free software; you can redistribute it and/or
+ *   modify it under the terms of the GNU General Public License
+ *   as published by the Free Software Foundation, version 2.
+ *
+ *   This program is distributed in the hope that it will be useful, but
+ *   WITHOUT ANY WARRANTY; without even the implied warranty of
+ *   MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
+ *   NON INFRINGEMENT.  See the GNU General Public License for
+ *   more details.
+ */
+
+/* This file is machine-generated; DO NOT EDIT! */
+#ifndef __GXIO_MPIPE_INFO_LINUX_RPC_H__
+#define __GXIO_MPIPE_INFO_LINUX_RPC_H__
+
+#include <hv/iorpc.h>
+
+#include <hv/drv_mpipe_intf.h>
+#include <asm/page.h>
+#include <gxio/kiorpc.h>
+#include <gxio/mpipe.h>
+#include <linux/string.h>
+#include <linux/module.h>
+#include <asm/pgtable.h>
+
+
+#define GXIO_MPIPE_INFO_OP_INSTANCE_AUX IORPC_OPCODE(IORPC_FORMAT_NONE, 0x1250)
+#define GXIO_MPIPE_INFO_OP_ENUMERATE_AUX IORPC_OPCODE(IORPC_FORMAT_NONE, 0x1251)
+#define GXIO_MPIPE_INFO_OP_GET_MMIO_BASE IORPC_OPCODE(IORPC_FORMAT_NONE_NOUSER, 0x8000)
+#define GXIO_MPIPE_INFO_OP_CHECK_MMIO_OFFSET IORPC_OPCODE(IORPC_FORMAT_NONE_NOUSER, 0x8001)
+
+
+int gxio_mpipe_info_instance_aux(gxio_mpipe_info_context_t *context,
+				 _gxio_mpipe_link_name_t name);
+
+int gxio_mpipe_info_enumerate_aux(gxio_mpipe_info_context_t *context,
+				  unsigned int idx,
+				  _gxio_mpipe_link_name_t *name,
+				  _gxio_mpipe_link_mac_t *mac);
+
+int gxio_mpipe_info_get_mmio_base(gxio_mpipe_info_context_t *context,
+				  HV_PTE *base);
+
+int gxio_mpipe_info_check_mmio_offset(gxio_mpipe_info_context_t *context,
+				      unsigned long offset, unsigned long size);
+
+#endif /* !__GXIO_MPIPE_INFO_LINUX_RPC_H__ */
diff --git a/kernel/arch/tile/include/gxio/iorpc_trio.h b/kernel/arch/tile/include/gxio/iorpc_trio.h
new file mode 100644
index 000000000..376a4f771
--- /dev/null
+++ b/kernel/arch/tile/include/gxio/iorpc_trio.h
@@ -0,0 +1,104 @@
+/*
+ * Copyright 2012 Tilera Corporation. All Rights Reserved.
+ *
+ *   This program is free software; you can redistribute it and/or
+ *   modify it under the terms of the GNU General Public License
+ *   as published by the Free Software Foundation, version 2.
+ *
+ *   This program is distributed in the hope that it will be useful, but
+ *   WITHOUT ANY WARRANTY; without even the implied warranty of
+ *   MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
+ *   NON INFRINGEMENT.  See the GNU General Public License for
+ *   more details.
+ */
+
+/* This file is machine-generated; DO NOT EDIT! */
+#ifndef __GXIO_TRIO_LINUX_RPC_H__
+#define __GXIO_TRIO_LINUX_RPC_H__
+
+#include <hv/iorpc.h>
+
+#include <hv/drv_trio_intf.h>
+#include <gxio/trio.h>
+#include <gxio/kiorpc.h>
+#include <linux/string.h>
+#include <linux/module.h>
+#include <asm/pgtable.h>
+
+#define GXIO_TRIO_OP_DEALLOC_ASID      IORPC_OPCODE(IORPC_FORMAT_NONE, 0x1400)
+#define GXIO_TRIO_OP_ALLOC_ASIDS       IORPC_OPCODE(IORPC_FORMAT_NONE, 0x1401)
+
+#define GXIO_TRIO_OP_ALLOC_MEMORY_MAPS IORPC_OPCODE(IORPC_FORMAT_NONE, 0x1404)
+
+#define GXIO_TRIO_OP_ALLOC_SCATTER_QUEUES IORPC_OPCODE(IORPC_FORMAT_NONE, 0x140e)
+#define GXIO_TRIO_OP_ALLOC_PIO_REGIONS IORPC_OPCODE(IORPC_FORMAT_NONE, 0x1412)
+
+#define GXIO_TRIO_OP_INIT_PIO_REGION_AUX IORPC_OPCODE(IORPC_FORMAT_NONE, 0x1414)
+
+#define GXIO_TRIO_OP_INIT_MEMORY_MAP_MMU_AUX IORPC_OPCODE(IORPC_FORMAT_NONE_NOUSER, 0x141e)
+#define GXIO_TRIO_OP_GET_PORT_PROPERTY IORPC_OPCODE(IORPC_FORMAT_NONE_NOUSER, 0x141f)
+#define GXIO_TRIO_OP_CONFIG_LEGACY_INTR IORPC_OPCODE(IORPC_FORMAT_KERNEL_INTERRUPT, 0x1420)
+#define GXIO_TRIO_OP_CONFIG_MSI_INTR   IORPC_OPCODE(IORPC_FORMAT_KERNEL_INTERRUPT, 0x1421)
+
+#define GXIO_TRIO_OP_SET_MPS_MRS       IORPC_OPCODE(IORPC_FORMAT_NONE_NOUSER, 0x1423)
+#define GXIO_TRIO_OP_FORCE_RC_LINK_UP  IORPC_OPCODE(IORPC_FORMAT_NONE_NOUSER, 0x1424)
+#define GXIO_TRIO_OP_FORCE_EP_LINK_UP  IORPC_OPCODE(IORPC_FORMAT_NONE_NOUSER, 0x1425)
+#define GXIO_TRIO_OP_GET_MMIO_BASE     IORPC_OPCODE(IORPC_FORMAT_NONE_NOUSER, 0x8000)
+#define GXIO_TRIO_OP_CHECK_MMIO_OFFSET IORPC_OPCODE(IORPC_FORMAT_NONE_NOUSER, 0x8001)
+
+int gxio_trio_alloc_asids(gxio_trio_context_t *context, unsigned int count,
+			  unsigned int first, unsigned int flags);
+
+
+int gxio_trio_alloc_memory_maps(gxio_trio_context_t *context,
+				unsigned int count, unsigned int first,
+				unsigned int flags);
+
+
+int gxio_trio_alloc_scatter_queues(gxio_trio_context_t *context,
+				   unsigned int count, unsigned int first,
+				   unsigned int flags);
+
+int gxio_trio_alloc_pio_regions(gxio_trio_context_t *context,
+				unsigned int count, unsigned int first,
+				unsigned int flags);
+
+int gxio_trio_init_pio_region_aux(gxio_trio_context_t *context,
+				  unsigned int pio_region, unsigned int mac,
+				  uint32_t bus_address_hi, unsigned int flags);
+
+
+int gxio_trio_init_memory_map_mmu_aux(gxio_trio_context_t *context,
+				      unsigned int map, unsigned long va,
+				      uint64_t size, unsigned int asid,
+				      unsigned int mac, uint64_t bus_address,
+				      unsigned int node,
+				      unsigned int order_mode);
+
+int gxio_trio_get_port_property(gxio_trio_context_t *context,
+				struct pcie_trio_ports_property *trio_ports);
+
+int gxio_trio_config_legacy_intr(gxio_trio_context_t *context, int inter_x,
+				 int inter_y, int inter_ipi, int inter_event,
+				 unsigned int mac, unsigned int intx);
+
+int gxio_trio_config_msi_intr(gxio_trio_context_t *context, int inter_x,
+			      int inter_y, int inter_ipi, int inter_event,
+			      unsigned int mac, unsigned int mem_map,
+			      uint64_t mem_map_base, uint64_t mem_map_limit,
+			      unsigned int asid);
+
+
+int gxio_trio_set_mps_mrs(gxio_trio_context_t *context, uint16_t mps,
+			  uint16_t mrs, unsigned int mac);
+
+int gxio_trio_force_rc_link_up(gxio_trio_context_t *context, unsigned int mac);
+
+int gxio_trio_force_ep_link_up(gxio_trio_context_t *context, unsigned int mac);
+
+int gxio_trio_get_mmio_base(gxio_trio_context_t *context, HV_PTE *base);
+
+int gxio_trio_check_mmio_offset(gxio_trio_context_t *context,
+				unsigned long offset, unsigned long size);
+
+#endif /* !__GXIO_TRIO_LINUX_RPC_H__ */
diff --git a/kernel/arch/tile/include/gxio/iorpc_uart.h b/kernel/arch/tile/include/gxio/iorpc_uart.h
new file mode 100644
index 000000000..55429d48e
--- /dev/null
+++ b/kernel/arch/tile/include/gxio/iorpc_uart.h
@@ -0,0 +1,40 @@
+/*
+ * Copyright 2013 Tilera Corporation. All Rights Reserved.
+ *
+ *   This program is free software; you can redistribute it and/or
+ *   modify it under the terms of the GNU General Public License
+ *   as published by the Free Software Foundation, version 2.
+ *
+ *   This program is distributed in the hope that it will be useful, but
+ *   WITHOUT ANY WARRANTY; without even the implied warranty of
+ *   MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
+ *   NON INFRINGEMENT.  See the GNU General Public License for
+ *   more details.
+ */
+
+/* This file is machine-generated; DO NOT EDIT! */
+#ifndef __GXIO_UART_LINUX_RPC_H__
+#define __GXIO_UART_LINUX_RPC_H__
+
+#include <hv/iorpc.h>
+
+#include <hv/drv_uart_intf.h>
+#include <gxio/uart.h>
+#include <gxio/kiorpc.h>
+#include <linux/string.h>
+#include <linux/module.h>
+#include <asm/pgtable.h>
+
+#define GXIO_UART_OP_CFG_INTERRUPT     IORPC_OPCODE(IORPC_FORMAT_KERNEL_INTERRUPT, 0x1900)
+#define GXIO_UART_OP_GET_MMIO_BASE     IORPC_OPCODE(IORPC_FORMAT_NONE_NOUSER, 0x8000)
+#define GXIO_UART_OP_CHECK_MMIO_OFFSET IORPC_OPCODE(IORPC_FORMAT_NONE_NOUSER, 0x8001)
+
+int gxio_uart_cfg_interrupt(gxio_uart_context_t *context, int inter_x,
+			    int inter_y, int inter_ipi, int inter_event);
+
+int gxio_uart_get_mmio_base(gxio_uart_context_t *context, HV_PTE *base);
+
+int gxio_uart_check_mmio_offset(gxio_uart_context_t *context,
+				unsigned long offset, unsigned long size);
+
+#endif /* !__GXIO_UART_LINUX_RPC_H__ */
diff --git a/kernel/arch/tile/include/gxio/iorpc_usb_host.h b/kernel/arch/tile/include/gxio/iorpc_usb_host.h
new file mode 100644
index 000000000..79962a97d
--- /dev/null
+++ b/kernel/arch/tile/include/gxio/iorpc_usb_host.h
@@ -0,0 +1,46 @@
+/*
+ * Copyright 2012 Tilera Corporation. All Rights Reserved.
+ *
+ *   This program is free software; you can redistribute it and/or
+ *   modify it under the terms of the GNU General Public License
+ *   as published by the Free Software Foundation, version 2.
+ *
+ *   This program is distributed in the hope that it will be useful, but
+ *   WITHOUT ANY WARRANTY; without even the implied warranty of
+ *   MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
+ *   NON INFRINGEMENT.  See the GNU General Public License for
+ *   more details.
+ */
+
+/* This file is machine-generated; DO NOT EDIT! */
+#ifndef __GXIO_USB_HOST_LINUX_RPC_H__
+#define __GXIO_USB_HOST_LINUX_RPC_H__
+
+#include <hv/iorpc.h>
+
+#include <hv/drv_usb_host_intf.h>
+#include <asm/page.h>
+#include <gxio/kiorpc.h>
+#include <gxio/usb_host.h>
+#include <linux/string.h>
+#include <linux/module.h>
+#include <asm/pgtable.h>
+
+#define GXIO_USB_HOST_OP_CFG_INTERRUPT IORPC_OPCODE(IORPC_FORMAT_KERNEL_INTERRUPT, 0x1800)
+#define GXIO_USB_HOST_OP_REGISTER_CLIENT_MEMORY IORPC_OPCODE(IORPC_FORMAT_NONE_NOUSER, 0x1801)
+#define GXIO_USB_HOST_OP_GET_MMIO_BASE IORPC_OPCODE(IORPC_FORMAT_NONE_NOUSER, 0x8000)
+#define GXIO_USB_HOST_OP_CHECK_MMIO_OFFSET IORPC_OPCODE(IORPC_FORMAT_NONE_NOUSER, 0x8001)
+
+int gxio_usb_host_cfg_interrupt(gxio_usb_host_context_t *context, int inter_x,
+				int inter_y, int inter_ipi, int inter_event);
+
+int gxio_usb_host_register_client_memory(gxio_usb_host_context_t *context,
+					 HV_PTE pte, unsigned int flags);
+
+int gxio_usb_host_get_mmio_base(gxio_usb_host_context_t *context,
+				HV_PTE *base);
+
+int gxio_usb_host_check_mmio_offset(gxio_usb_host_context_t *context,
+				    unsigned long offset, unsigned long size);
+
+#endif /* !__GXIO_USB_HOST_LINUX_RPC_H__ */
diff --git a/kernel/arch/tile/include/gxio/kiorpc.h b/kernel/arch/tile/include/gxio/kiorpc.h
new file mode 100644
index 000000000..ee5820979
--- /dev/null
+++ b/kernel/arch/tile/include/gxio/kiorpc.h
@@ -0,0 +1,29 @@
+/*
+ * Copyright 2012 Tilera Corporation. All Rights Reserved.
+ *
+ *   This program is free software; you can redistribute it and/or
+ *   modify it under the terms of the GNU General Public License
+ *   as published by the Free Software Foundation, version 2.
+ *
+ *   This program is distributed in the hope that it will be useful, but
+ *   WITHOUT ANY WARRANTY; without even the implied warranty of
+ *   MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
+ *   NON INFRINGEMENT.  See the GNU General Public License for
+ *   more details.
+ *
+ * Support routines for kernel IORPC drivers.
+ */
+
+#ifndef _GXIO_KIORPC_H
+#define _GXIO_KIORPC_H
+
+#include <linux/types.h>
+#include <asm/page.h>
+#include <arch/chip.h>
+
+#if CHIP_HAS_MMIO()
+void __iomem *iorpc_ioremap(int hv_fd, resource_size_t offset,
+			    unsigned long size);
+#endif
+
+#endif /* _GXIO_KIORPC_H */
diff --git a/kernel/arch/tile/include/gxio/mpipe.h b/kernel/arch/tile/include/gxio/mpipe.h
new file mode 100644
index 000000000..73e83a187
--- /dev/null
+++ b/kernel/arch/tile/include/gxio/mpipe.h
@@ -0,0 +1,1871 @@
+/*
+ * Copyright 2012 Tilera Corporation. All Rights Reserved.
+ *
+ *   This program is free software; you can redistribute it and/or
+ *   modify it under the terms of the GNU General Public License
+ *   as published by the Free Software Foundation, version 2.
+ *
+ *   This program is distributed in the hope that it will be useful, but
+ *   WITHOUT ANY WARRANTY; without even the implied warranty of
+ *   MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
+ *   NON INFRINGEMENT.  See the GNU General Public License for
+ *   more details.
+ */
+
+#ifndef _GXIO_MPIPE_H_
+#define _GXIO_MPIPE_H_
+
+/*
+ *
+ * An API for allocating, configuring, and manipulating mPIPE hardware
+ * resources.
+ */
+
+#include <gxio/common.h>
+#include <gxio/dma_queue.h>
+
+#include <linux/time.h>
+
+#include <arch/mpipe_def.h>
+#include <arch/mpipe_shm.h>
+
+#include <hv/drv_mpipe_intf.h>
+#include <hv/iorpc.h>
+
+/*
+ *
+ * The TILE-Gx mPIPE&tm; shim provides Ethernet connectivity, packet
+ * classification, and packet load balancing services.  The
+ * gxio_mpipe_ API, declared in <gxio/mpipe.h>, allows applications to
+ * allocate mPIPE IO channels, configure packet distribution
+ * parameters, and send and receive Ethernet packets.  The API is
+ * designed to be a minimal wrapper around the mPIPE hardware, making
+ * system calls only where necessary to preserve inter-process
+ * protection guarantees.
+ *
+ * The APIs described below allow the programmer to allocate and
+ * configure mPIPE resources.  As described below, the mPIPE is a
+ * single shared hardware device that provides partitionable resources
+ * that are shared between all applications in the system.  The
+ * gxio_mpipe_ API allows userspace code to make resource request
+ * calls to the hypervisor, which in turns keeps track of the
+ * resources in use by all applications, maintains protection
+ * guarantees, and resets resources upon application shutdown.
+ *
+ * We strongly recommend reading the mPIPE section of the IO Device
+ * Guide (UG404) before working with this API.  Most functions in the
+ * gxio_mpipe_ API are directly analogous to hardware interfaces and
+ * the documentation assumes that the reader understands those
+ * hardware interfaces.
+ *
+ * @section mpipe__ingress mPIPE Ingress Hardware Resources
+ *
+ * The mPIPE ingress hardware provides extensive hardware offload for
+ * tasks like packet header parsing, load balancing, and memory
+ * management.  This section provides a brief introduction to the
+ * hardware components and the gxio_mpipe_ calls used to manage them;
+ * see the IO Device Guide for a much more detailed description of the
+ * mPIPE's capabilities.
+ *
+ * When a packet arrives at one of the mPIPE's Ethernet MACs, it is
+ * assigned a channel number indicating which MAC received it.  It
+ * then proceeds through the following hardware pipeline:
+ *
+ * @subsection mpipe__classification Classification
+ *
+ * A set of classification processors run header parsing code on each
+ * incoming packet, extracting information including the destination
+ * MAC address, VLAN, Ethernet type, and five-tuple hash.  Some of
+ * this information is then used to choose which buffer stack will be
+ * used to hold the packet, and which bucket will be used by the load
+ * balancer to determine which application will receive the packet.
+ *
+ * The rules by which the buffer stack and bucket are chosen can be
+ * configured via the @ref gxio_mpipe_classifier API.  A given app can
+ * specify multiple rules, each one specifying a bucket range, and a
+ * set of buffer stacks, to be used for packets matching the rule.
+ * Each rule can optionally specify a restricted set of channels,
+ * VLANs, and/or dMACs, in which it is interested.  By default, a
+ * given rule starts out matching all channels associated with the
+ * mPIPE context's set of open links; all VLANs; and all dMACs.
+ * Subsequent restrictions can then be added.
+ *
+ * @subsection mpipe__load_balancing Load Balancing
+ *
+ * The mPIPE load balancer is responsible for choosing the NotifRing
+ * to which the packet will be delivered.  This decision is based on
+ * the bucket number indicated by the classification program.  In
+ * general, the bucket number is based on some number of low bits of
+ * the packet's flow hash (applications that aren't interested in flow
+ * hashing use a single bucket).  Each load balancer bucket keeps a
+ * record of the NotifRing to which packets directed to that bucket
+ * are currently being delivered.  Based on the bucket's load
+ * balancing mode (@ref gxio_mpipe_bucket_mode_t), the load balancer
+ * either forwards the packet to the previously assigned NotifRing or
+ * decides to choose a new NotifRing.  If a new NotifRing is required,
+ * the load balancer chooses the least loaded ring in the NotifGroup
+ * associated with the bucket.
+ *
+ * The load balancer is a shared resource.  Each application needs to
+ * explicitly allocate NotifRings, NotifGroups, and buckets, using
+ * gxio_mpipe_alloc_notif_rings(), gxio_mpipe_alloc_notif_groups(),
+ * and gxio_mpipe_alloc_buckets().  Then the application needs to
+ * configure them using gxio_mpipe_init_notif_ring() and
+ * gxio_mpipe_init_notif_group_and_buckets().
+ *
+ * @subsection mpipe__buffers Buffer Selection and Packet Delivery
+ *
+ * Once the load balancer has chosen the destination NotifRing, the
+ * mPIPE DMA engine pops at least one buffer off of the 'buffer stack'
+ * chosen by the classification program and DMAs the packet data into
+ * that buffer.  Each buffer stack provides a hardware-accelerated
+ * stack of data buffers with the same size.  If the packet data is
+ * larger than the buffers provided by the chosen buffer stack, the
+ * mPIPE hardware pops off multiple buffers and chains the packet data
+ * through a multi-buffer linked list.  Once the packet data is
+ * delivered to the buffer(s), the mPIPE hardware writes the
+ * ::gxio_mpipe_idesc_t metadata object (calculated by the classifier)
+ * into the NotifRing and increments the number of packets delivered
+ * to that ring.
+ *
+ * Applications can push buffers onto a buffer stack by calling
+ * gxio_mpipe_push_buffer() or by egressing a packet with the
+ * ::gxio_mpipe_edesc_t::hwb bit set, indicating that the egressed
+ * buffers should be returned to the stack.
+ *
+ * Applications can allocate and initialize buffer stacks with the
+ * gxio_mpipe_alloc_buffer_stacks() and gxio_mpipe_init_buffer_stack()
+ * APIs.
+ *
+ * The application must also register the memory pages that will hold
+ * packets.  This requires calling gxio_mpipe_register_page() for each
+ * memory page that will hold packets allocated by the application for
+ * a given buffer stack.  Since each buffer stack is limited to 16
+ * registered pages, it may be necessary to use huge pages, or even
+ * extremely huge pages, to hold all the buffers.
+ *
+ * @subsection mpipe__iqueue NotifRings
+ *
+ * Each NotifRing is a region of shared memory, allocated by the
+ * application, to which the mPIPE delivers packet descriptors
+ * (::gxio_mpipe_idesc_t).  The application can allocate them via
+ * gxio_mpipe_alloc_notif_rings().  The application can then either
+ * explicitly initialize them with gxio_mpipe_init_notif_ring() and
+ * then read from them manually, or can make use of the convenience
+ * wrappers provided by @ref gxio_mpipe_wrappers.
+ *
+ * @section mpipe__egress mPIPE Egress Hardware
+ *
+ * Applications use eDMA rings to queue packets for egress.  The
+ * application can allocate them via gxio_mpipe_alloc_edma_rings().
+ * The application can then either explicitly initialize them with
+ * gxio_mpipe_init_edma_ring() and then write to them manually, or
+ * can make use of the convenience wrappers provided by
+ * @ref gxio_mpipe_wrappers.
+ *
+ * @section gxio__shortcomings Plans for Future API Revisions
+ *
+ * The API defined here is only an initial version of the mPIPE API.
+ * Future plans include:
+ *
+ * - Higher level wrapper functions to provide common initialization
+ * patterns.  This should help users start writing mPIPE programs
+ * without having to learn the details of the hardware.
+ *
+ * - Support for reset and deallocation of resources, including
+ * cleanup upon application shutdown.
+ *
+ * - Support for calling these APIs in the BME.
+ *
+ * - Support for IO interrupts.
+ *
+ * - Clearer definitions of thread safety guarantees.
+ *
+ * @section gxio__mpipe_examples Examples
+ *
+ * See the following mPIPE example programs for more information about
+ * allocating mPIPE resources and using them in real applications:
+ *
+ * - @ref mpipe/ingress/app.c : Receiving packets.
+ *
+ * - @ref mpipe/forward/app.c : Forwarding packets.
+ *
+ * Note that there are several more examples.
+ */
+
+/* Flags that can be passed to resource allocation functions. */
+enum gxio_mpipe_alloc_flags_e {
+	/* Require an allocation to start at a specified resource index. */
+	GXIO_MPIPE_ALLOC_FIXED = HV_MPIPE_ALLOC_FIXED,
+};
+
+/* Flags that can be passed to memory registration functions. */
+enum gxio_mpipe_mem_flags_e {
+	/* Do not fill L3 when writing, and invalidate lines upon egress. */
+	GXIO_MPIPE_MEM_FLAG_NT_HINT = IORPC_MEM_BUFFER_FLAG_NT_HINT,
+
+	/* L3 cache fills should only populate IO cache ways. */
+	GXIO_MPIPE_MEM_FLAG_IO_PIN = IORPC_MEM_BUFFER_FLAG_IO_PIN,
+};
+
+/* An ingress packet descriptor.  When a packet arrives, the mPIPE
+ * hardware generates this structure and writes it into a NotifRing.
+ */
+typedef MPIPE_PDESC_t gxio_mpipe_idesc_t;
+
+/* An egress command descriptor.  Applications write this structure
+ * into eDMA rings and the hardware performs the indicated operation
+ * (normally involving egressing some bytes).  Note that egressing a
+ * single packet may involve multiple egress command descriptors.
+ */
+typedef MPIPE_EDMA_DESC_t gxio_mpipe_edesc_t;
+
+/*
+ * Max # of mpipe instances. 2 currently.
+ */
+#define GXIO_MPIPE_INSTANCE_MAX  HV_MPIPE_INSTANCE_MAX
+
+#define NR_MPIPE_MAX   GXIO_MPIPE_INSTANCE_MAX
+
+/* Get the "va" field from an "idesc".
+ *
+ * This is the address at which the ingress hardware copied the first
+ * byte of the packet.
+ *
+ * If the classifier detected a custom header, then this will point to
+ * the custom header, and gxio_mpipe_idesc_get_l2_start() will point
+ * to the actual L2 header.
+ *
+ * Note that this value may be misleading if "idesc->be" is set.
+ *
+ * @param idesc An ingress packet descriptor.
+ */
+static inline unsigned char *gxio_mpipe_idesc_get_va(gxio_mpipe_idesc_t *idesc)
+{
+	return (unsigned char *)(long)idesc->va;
+}
+
+/* Get the "xfer_size" from an "idesc".
+ *
+ * This is the actual number of packet bytes transferred into memory
+ * by the hardware.
+ *
+ * Note that this value may be misleading if "idesc->be" is set.
+ *
+ * @param idesc An ingress packet descriptor.
+ *
+ * ISSUE: Is this the best name for this?
+ * FIXME: Add more docs about chaining, clipping, etc.
+ */
+static inline unsigned int gxio_mpipe_idesc_get_xfer_size(gxio_mpipe_idesc_t
+							  *idesc)
+{
+	return idesc->l2_size;
+}
+
+/* Get the "l2_offset" from an "idesc".
+ *
+ * Extremely customized classifiers might not support this function.
+ *
+ * This is the number of bytes between the "va" and the L2 header.
+ *
+ * The L2 header consists of a destination mac address, a source mac
+ * address, and an initial ethertype.  Various initial ethertypes
+ * allow encoding extra information in the L2 header, often including
+ * a vlan, and/or a new ethertype.
+ *
+ * Note that the "l2_offset" will be non-zero if (and only if) the
+ * classifier processed a custom header for the packet.
+ *
+ * @param idesc An ingress packet descriptor.
+ */
+static inline uint8_t gxio_mpipe_idesc_get_l2_offset(gxio_mpipe_idesc_t *idesc)
+{
+	return (idesc->custom1 >> 32) & 0xFF;
+}
+
+/* Get the "l2_start" from an "idesc".
+ *
+ * This is simply gxio_mpipe_idesc_get_va() plus
+ * gxio_mpipe_idesc_get_l2_offset().
+ *
+ * @param idesc An ingress packet descriptor.
+ */
+static inline unsigned char *gxio_mpipe_idesc_get_l2_start(gxio_mpipe_idesc_t
+							   *idesc)
+{
+	unsigned char *va = gxio_mpipe_idesc_get_va(idesc);
+	return va + gxio_mpipe_idesc_get_l2_offset(idesc);
+}
+
+/* Get the "l2_length" from an "idesc".
+ *
+ * This is simply gxio_mpipe_idesc_get_xfer_size() minus
+ * gxio_mpipe_idesc_get_l2_offset().
+ *
+ * @param idesc An ingress packet descriptor.
+ */
+static inline unsigned int gxio_mpipe_idesc_get_l2_length(gxio_mpipe_idesc_t
+							  *idesc)
+{
+	unsigned int xfer_size = idesc->l2_size;
+	return xfer_size - gxio_mpipe_idesc_get_l2_offset(idesc);
+}
+
+/* A context object used to manage mPIPE hardware resources. */
+typedef struct {
+
+	/* File descriptor for calling up to Linux (and thus the HV). */
+	int fd;
+
+	/* Corresponding mpipe instance #. */
+	int instance;
+
+	/* The VA at which configuration registers are mapped. */
+	char *mmio_cfg_base;
+
+	/* The VA at which IDMA, EDMA, and buffer manager are mapped. */
+	char *mmio_fast_base;
+
+	/* The "initialized" buffer stacks. */
+	gxio_mpipe_rules_stacks_t __stacks;
+
+} gxio_mpipe_context_t;
+
+/* This is only used internally, but it's most easily made visible here. */
+typedef gxio_mpipe_context_t gxio_mpipe_info_context_t;
+
+/* Initialize an mPIPE context.
+ *
+ * This function allocates an mPIPE "service domain" and maps the MMIO
+ * registers into the caller's VA space.
+ *
+ * @param context Context object to be initialized.
+ * @param mpipe_instance Instance number of mPIPE shim to be controlled via
+ *  context.
+ */
+extern int gxio_mpipe_init(gxio_mpipe_context_t *context,
+			   unsigned int mpipe_instance);
+
+/* Destroy an mPIPE context.
+ *
+ * This function frees the mPIPE "service domain" and unmaps the MMIO
+ * registers from the caller's VA space.
+ *
+ * If a user process exits without calling this routine, the kernel
+ * will destroy the mPIPE context as part of process teardown.
+ *
+ * @param context Context object to be destroyed.
+ */
+extern int gxio_mpipe_destroy(gxio_mpipe_context_t *context);
+
+/*****************************************************************
+ *                         Buffer Stacks                          *
+ ******************************************************************/
+
+/* Allocate a set of buffer stacks.
+ *
+ * The return value is NOT interesting if count is zero.
+ *
+ * @param context An initialized mPIPE context.
+ * @param count Number of stacks required.
+ * @param first Index of first stack if ::GXIO_MPIPE_ALLOC_FIXED flag is set,
+ *   otherwise ignored.
+ * @param flags Flag bits from ::gxio_mpipe_alloc_flags_e.
+ * @return Index of first allocated buffer stack, or
+ * ::GXIO_MPIPE_ERR_NO_BUFFER_STACK if allocation failed.
+ */
+extern int gxio_mpipe_alloc_buffer_stacks(gxio_mpipe_context_t *context,
+					  unsigned int count,
+					  unsigned int first,
+					  unsigned int flags);
+
+/* Enum codes for buffer sizes supported by mPIPE. */
+typedef enum {
+	/* 128 byte packet data buffer. */
+	GXIO_MPIPE_BUFFER_SIZE_128 = MPIPE_BSM_INIT_DAT_1__SIZE_VAL_BSZ_128,
+	/* 256 byte packet data buffer. */
+	GXIO_MPIPE_BUFFER_SIZE_256 = MPIPE_BSM_INIT_DAT_1__SIZE_VAL_BSZ_256,
+	/* 512 byte packet data buffer. */
+	GXIO_MPIPE_BUFFER_SIZE_512 = MPIPE_BSM_INIT_DAT_1__SIZE_VAL_BSZ_512,
+	/* 1024 byte packet data buffer. */
+	GXIO_MPIPE_BUFFER_SIZE_1024 = MPIPE_BSM_INIT_DAT_1__SIZE_VAL_BSZ_1024,
+	/* 1664 byte packet data buffer. */
+	GXIO_MPIPE_BUFFER_SIZE_1664 = MPIPE_BSM_INIT_DAT_1__SIZE_VAL_BSZ_1664,
+	/* 4096 byte packet data buffer. */
+	GXIO_MPIPE_BUFFER_SIZE_4096 = MPIPE_BSM_INIT_DAT_1__SIZE_VAL_BSZ_4096,
+	/* 10368 byte packet data buffer. */
+	GXIO_MPIPE_BUFFER_SIZE_10368 =
+		MPIPE_BSM_INIT_DAT_1__SIZE_VAL_BSZ_10368,
+	/* 16384 byte packet data buffer. */
+	GXIO_MPIPE_BUFFER_SIZE_16384 = MPIPE_BSM_INIT_DAT_1__SIZE_VAL_BSZ_16384
+} gxio_mpipe_buffer_size_enum_t;
+
+/* Convert a buffer size in bytes into a buffer size enum. */
+extern gxio_mpipe_buffer_size_enum_t
+gxio_mpipe_buffer_size_to_buffer_size_enum(size_t size);
+
+/* Convert a buffer size enum into a buffer size in bytes. */
+extern size_t
+gxio_mpipe_buffer_size_enum_to_buffer_size(gxio_mpipe_buffer_size_enum_t
+					   buffer_size_enum);
+
+/* Calculate the number of bytes required to store a given number of
+ * buffers in the memory registered with a buffer stack via
+ * gxio_mpipe_init_buffer_stack().
+ */
+extern size_t gxio_mpipe_calc_buffer_stack_bytes(unsigned long buffers);
+
+/* Initialize a buffer stack.  This function binds a region of memory
+ * to be used by the hardware for storing buffer addresses pushed via
+ * gxio_mpipe_push_buffer() or as the result of sending a buffer out
+ * the egress with the 'push to stack when done' bit set.  Once this
+ * function returns, the memory region's contents may be arbitrarily
+ * modified by the hardware at any time and software should not access
+ * the memory region again.
+ *
+ * @param context An initialized mPIPE context.
+ * @param stack The buffer stack index.
+ * @param buffer_size_enum The size of each buffer in the buffer stack,
+ * as an enum.
+ * @param mem The address of the buffer stack.  This memory must be
+ * physically contiguous and aligned to a 64kB boundary.
+ * @param mem_size The size of the buffer stack, in bytes.
+ * @param mem_flags ::gxio_mpipe_mem_flags_e memory flags.
+ * @return Zero on success, ::GXIO_MPIPE_ERR_INVAL_BUFFER_SIZE if
+ * buffer_size_enum is invalid, ::GXIO_MPIPE_ERR_BAD_BUFFER_STACK if
+ * stack has not been allocated.
+ */
+extern int gxio_mpipe_init_buffer_stack(gxio_mpipe_context_t *context,
+					unsigned int stack,
+					gxio_mpipe_buffer_size_enum_t
+					buffer_size_enum, void *mem,
+					size_t mem_size,
+					unsigned int mem_flags);
+
+/* Push a buffer onto a previously initialized buffer stack.
+ *
+ * The size of the buffer being pushed must match the size that was
+ * registered with gxio_mpipe_init_buffer_stack().  All packet buffer
+ * addresses are 128-byte aligned; the low 7 bits of the specified
+ * buffer address will be ignored.
+ *
+ * @param context An initialized mPIPE context.
+ * @param stack The buffer stack index.
+ * @param buffer The buffer (the low seven bits are ignored).
+ */
+static inline void gxio_mpipe_push_buffer(gxio_mpipe_context_t *context,
+					  unsigned int stack, void *buffer)
+{
+	MPIPE_BSM_REGION_ADDR_t offset = { {0} };
+	MPIPE_BSM_REGION_VAL_t val = { {0} };
+
+	/*
+	 * The mmio_fast_base region starts at the IDMA region, so subtract
+	 * off that initial offset.
+	 */
+	offset.region =
+		MPIPE_MMIO_ADDR__REGION_VAL_BSM -
+		MPIPE_MMIO_ADDR__REGION_VAL_IDMA;
+	offset.stack = stack;
+
+#if __SIZEOF_POINTER__ == 4
+	val.va = ((ulong) buffer) >> MPIPE_BSM_REGION_VAL__VA_SHIFT;
+#else
+	val.va = ((long)buffer) >> MPIPE_BSM_REGION_VAL__VA_SHIFT;
+#endif
+
+	__gxio_mmio_write(context->mmio_fast_base + offset.word, val.word);
+}
+
+/* Pop a buffer off of a previously initialized buffer stack.
+ *
+ * @param context An initialized mPIPE context.
+ * @param stack The buffer stack index.
+ * @return The buffer, or NULL if the stack is empty.
+ */
+static inline void *gxio_mpipe_pop_buffer(gxio_mpipe_context_t *context,
+					  unsigned int stack)
+{
+	MPIPE_BSM_REGION_ADDR_t offset = { {0} };
+
+	/*
+	 * The mmio_fast_base region starts at the IDMA region, so subtract
+	 * off that initial offset.
+	 */
+	offset.region =
+		MPIPE_MMIO_ADDR__REGION_VAL_BSM -
+		MPIPE_MMIO_ADDR__REGION_VAL_IDMA;
+	offset.stack = stack;
+
+	while (1) {
+		/*
+		 * Case 1: val.c == ..._UNCHAINED, va is non-zero.
+		 * Case 2: val.c == ..._INVALID, va is zero.
+		 * Case 3: val.c == ..._NOT_RDY, va is zero.
+		 */
+		MPIPE_BSM_REGION_VAL_t val;
+		val.word =
+			__gxio_mmio_read(context->mmio_fast_base +
+					 offset.word);
+
+		/*
+		 * Handle case 1 and 2 by returning the buffer (or NULL).
+		 * Handle case 3 by waiting for the prefetch buffer to refill.
+		 */
+		if (val.c != MPIPE_EDMA_DESC_WORD1__C_VAL_NOT_RDY)
+			return (void *)((unsigned long)val.
+					va << MPIPE_BSM_REGION_VAL__VA_SHIFT);
+	}
+}
+
+/*****************************************************************
+ *                          NotifRings                            *
+ ******************************************************************/
+
+/* Allocate a set of NotifRings.
+ *
+ * The return value is NOT interesting if count is zero.
+ *
+ * Note that NotifRings are allocated in chunks, so allocating one at
+ * a time is much less efficient than allocating several at once.
+ *
+ * @param context An initialized mPIPE context.
+ * @param count Number of NotifRings required.
+ * @param first Index of first NotifRing if ::GXIO_MPIPE_ALLOC_FIXED flag
+ *   is set, otherwise ignored.
+ * @param flags Flag bits from ::gxio_mpipe_alloc_flags_e.
+ * @return Index of first allocated buffer NotifRing, or
+ * ::GXIO_MPIPE_ERR_NO_NOTIF_RING if allocation failed.
+ */
+extern int gxio_mpipe_alloc_notif_rings(gxio_mpipe_context_t *context,
+					unsigned int count, unsigned int first,
+					unsigned int flags);
+
+/* Initialize a NotifRing, using the given memory and size.
+ *
+ * @param context An initialized mPIPE context.
+ * @param ring The NotifRing index.
+ * @param mem A physically contiguous region of memory to be filled
+ * with a ring of ::gxio_mpipe_idesc_t structures.
+ * @param mem_size Number of bytes in the ring.  Must be 128, 512,
+ * 2048, or 65536 * sizeof(gxio_mpipe_idesc_t).
+ * @param mem_flags ::gxio_mpipe_mem_flags_e memory flags.
+ *
+ * @return 0 on success, ::GXIO_MPIPE_ERR_BAD_NOTIF_RING or
+ * ::GXIO_ERR_INVAL_MEMORY_SIZE on failure.
+ */
+extern int gxio_mpipe_init_notif_ring(gxio_mpipe_context_t *context,
+				      unsigned int ring,
+				      void *mem, size_t mem_size,
+				      unsigned int mem_flags);
+
+/* Configure an interrupt to be sent to a tile on incoming NotifRing
+ *  traffic.  Once an interrupt is sent for a particular ring, no more
+ *  will be sent until gxio_mica_enable_notif_ring_interrupt() is called.
+ *
+ * @param context An initialized mPIPE context.
+ * @param x X coordinate of interrupt target tile.
+ * @param y Y coordinate of interrupt target tile.
+ * @param i Index of the IPI register which will receive the interrupt.
+ * @param e Specific event which will be set in the target IPI register when
+ * the interrupt occurs.
+ * @param ring The NotifRing index.
+ * @return Zero on success, GXIO_ERR_INVAL if params are out of range.
+ */
+extern int gxio_mpipe_request_notif_ring_interrupt(gxio_mpipe_context_t
+						   *context, int x, int y,
+						   int i, int e,
+						   unsigned int ring);
+
+/* Enable an interrupt on incoming NotifRing traffic.
+ *
+ * @param context An initialized mPIPE context.
+ * @param ring The NotifRing index.
+ * @return Zero on success, GXIO_ERR_INVAL if params are out of range.
+ */
+extern int gxio_mpipe_enable_notif_ring_interrupt(gxio_mpipe_context_t
+						  *context, unsigned int ring);
+
+/* Map all of a client's memory via the given IOTLB.
+ * @param context An initialized mPIPE context.
+ * @param iotlb IOTLB index.
+ * @param pte Page table entry.
+ * @param flags Flags.
+ * @return Zero on success, or a negative error code.
+ */
+extern int gxio_mpipe_register_client_memory(gxio_mpipe_context_t *context,
+					     unsigned int iotlb, HV_PTE pte,
+					     unsigned int flags);
+
+/*****************************************************************
+ *                        Notif Groups                            *
+ ******************************************************************/
+
+/* Allocate a set of NotifGroups.
+ *
+ * The return value is NOT interesting if count is zero.
+ *
+ * @param context An initialized mPIPE context.
+ * @param count Number of NotifGroups required.
+ * @param first Index of first NotifGroup if ::GXIO_MPIPE_ALLOC_FIXED flag
+ *   is set, otherwise ignored.
+ * @param flags Flag bits from ::gxio_mpipe_alloc_flags_e.
+ * @return Index of first allocated buffer NotifGroup, or
+ * ::GXIO_MPIPE_ERR_NO_NOTIF_GROUP if allocation failed.
+ */
+extern int gxio_mpipe_alloc_notif_groups(gxio_mpipe_context_t *context,
+					 unsigned int count,
+					 unsigned int first,
+					 unsigned int flags);
+
+/* Add a NotifRing to a NotifGroup.  This only sets a bit in the
+ * application's 'group' object; the hardware NotifGroup can be
+ * initialized by passing 'group' to gxio_mpipe_init_notif_group() or
+ * gxio_mpipe_init_notif_group_and_buckets().
+ */
+static inline void
+gxio_mpipe_notif_group_add_ring(gxio_mpipe_notif_group_bits_t *bits, int ring)
+{
+	bits->ring_mask[ring / 64] |= (1ull << (ring % 64));
+}
+
+/* Set a particular NotifGroup bitmask.  Since the load balancer
+ * makes decisions based on both bucket and NotifGroup state, most
+ * applications should use gxio_mpipe_init_notif_group_and_buckets()
+ * rather than using this function to configure just a NotifGroup.
+ */
+extern int gxio_mpipe_init_notif_group(gxio_mpipe_context_t *context,
+				       unsigned int group,
+				       gxio_mpipe_notif_group_bits_t bits);
+
+/*****************************************************************
+ *                         Load Balancer                          *
+ ******************************************************************/
+
+/* Allocate a set of load balancer buckets.
+ *
+ * The return value is NOT interesting if count is zero.
+ *
+ * Note that buckets are allocated in chunks, so allocating one at
+ * a time is much less efficient than allocating several at once.
+ *
+ * Note that the buckets are actually divided into two sub-ranges, of
+ * different sizes, and different chunk sizes, and the range you get
+ * by default is determined by the size of the request.  Allocations
+ * cannot span the two sub-ranges.
+ *
+ * @param context An initialized mPIPE context.
+ * @param count Number of buckets required.
+ * @param first Index of first bucket if ::GXIO_MPIPE_ALLOC_FIXED flag is set,
+ *   otherwise ignored.
+ * @param flags Flag bits from ::gxio_mpipe_alloc_flags_e.
+ * @return Index of first allocated buffer bucket, or
+ * ::GXIO_MPIPE_ERR_NO_BUCKET if allocation failed.
+ */
+extern int gxio_mpipe_alloc_buckets(gxio_mpipe_context_t *context,
+				    unsigned int count, unsigned int first,
+				    unsigned int flags);
+
+/* The legal modes for gxio_mpipe_bucket_info_t and
+ * gxio_mpipe_init_notif_group_and_buckets().
+ *
+ * All modes except ::GXIO_MPIPE_BUCKET_ROUND_ROBIN expect that the user
+ * will allocate a power-of-two number of buckets and initialize them
+ * to the same mode.  The classifier program then uses the appropriate
+ * number of low bits from the incoming packet's flow hash to choose a
+ * load balancer bucket.  Based on that bucket's load balancing mode,
+ * reference count, and currently active NotifRing, the load balancer
+ * chooses the NotifRing to which the packet will be delivered.
+ */
+typedef enum {
+	/* All packets for a bucket go to the same NotifRing unless the
+	 * NotifRing gets full, in which case packets will be dropped.  If
+	 * the bucket reference count ever reaches zero, a new NotifRing may
+	 * be chosen.
+	 */
+	GXIO_MPIPE_BUCKET_DYNAMIC_FLOW_AFFINITY =
+		MPIPE_LBL_INIT_DAT_BSTS_TBL__MODE_VAL_DFA,
+
+	/* All packets for a bucket always go to the same NotifRing.
+	 */
+	GXIO_MPIPE_BUCKET_STATIC_FLOW_AFFINITY =
+		MPIPE_LBL_INIT_DAT_BSTS_TBL__MODE_VAL_FIXED,
+
+	/* All packets for a bucket go to the least full NotifRing in the
+	 * group, providing load balancing round robin behavior.
+	 */
+	GXIO_MPIPE_BUCKET_ROUND_ROBIN =
+		MPIPE_LBL_INIT_DAT_BSTS_TBL__MODE_VAL_ALWAYS_PICK,
+
+	/* All packets for a bucket go to the same NotifRing unless the
+	 * NotifRing gets full, at which point the bucket starts using the
+	 * least full NotifRing in the group.  If all NotifRings in the
+	 * group are full, packets will be dropped.
+	 */
+	GXIO_MPIPE_BUCKET_STICKY_FLOW_LOCALITY =
+		MPIPE_LBL_INIT_DAT_BSTS_TBL__MODE_VAL_STICKY,
+
+	/* All packets for a bucket go to the same NotifRing unless the
+	 * NotifRing gets full, or a random timer fires, at which point the
+	 * bucket starts using the least full NotifRing in the group.  If
+	 * all NotifRings in the group are full, packets will be dropped.
+	 * WARNING: This mode is BROKEN on chips with fewer than 64 tiles.
+	 */
+	GXIO_MPIPE_BUCKET_PREFER_FLOW_LOCALITY =
+		MPIPE_LBL_INIT_DAT_BSTS_TBL__MODE_VAL_STICKY_RAND,
+
+} gxio_mpipe_bucket_mode_t;
+
+/* Copy a set of bucket initialization values into the mPIPE
+ * hardware.  Since the load balancer makes decisions based on both
+ * bucket and NotifGroup state, most applications should use
+ * gxio_mpipe_init_notif_group_and_buckets() rather than using this
+ * function to configure a single bucket.
+ *
+ * @param context An initialized mPIPE context.
+ * @param bucket Bucket index to be initialized.
+ * @param bucket_info Initial reference count, NotifRing index, and mode.
+ * @return 0 on success, ::GXIO_MPIPE_ERR_BAD_BUCKET on failure.
+ */
+extern int gxio_mpipe_init_bucket(gxio_mpipe_context_t *context,
+				  unsigned int bucket,
+				  gxio_mpipe_bucket_info_t bucket_info);
+
+/* Initializes a group and range of buckets and range of rings such
+ * that the load balancer runs a particular load balancing function.
+ *
+ * First, the group is initialized with the given rings.
+ *
+ * Second, each bucket is initialized with the mode and group, and a
+ * ring chosen round-robin from the given rings.
+ *
+ * Normally, the classifier picks a bucket, and then the load balancer
+ * picks a ring, based on the bucket's mode, group, and current ring,
+ * possibly updating the bucket's ring.
+ *
+ * @param context An initialized mPIPE context.
+ * @param group The group.
+ * @param ring The first ring.
+ * @param num_rings The number of rings.
+ * @param bucket The first bucket.
+ * @param num_buckets The number of buckets.
+ * @param mode The load balancing mode.
+ *
+ * @return 0 on success, ::GXIO_MPIPE_ERR_BAD_BUCKET,
+ * ::GXIO_MPIPE_ERR_BAD_NOTIF_GROUP, or
+ * ::GXIO_MPIPE_ERR_BAD_NOTIF_RING on failure.
+ */
+extern int gxio_mpipe_init_notif_group_and_buckets(gxio_mpipe_context_t
+						   *context,
+						   unsigned int group,
+						   unsigned int ring,
+						   unsigned int num_rings,
+						   unsigned int bucket,
+						   unsigned int num_buckets,
+						   gxio_mpipe_bucket_mode_t
+						   mode);
+
+/* Return credits to a NotifRing and/or bucket.
+ *
+ * @param context An initialized mPIPE context.
+ * @param ring The NotifRing index, or -1.
+ * @param bucket The bucket, or -1.
+ * @param count The number of credits to return.
+ */
+static inline void gxio_mpipe_credit(gxio_mpipe_context_t *context,
+				     int ring, int bucket, unsigned int count)
+{
+	/* NOTE: Fancy struct initialization would break "C89" header test. */
+
+	MPIPE_IDMA_RELEASE_REGION_ADDR_t offset = { {0} };
+	MPIPE_IDMA_RELEASE_REGION_VAL_t val = { {0} };
+
+	/*
+	 * The mmio_fast_base region starts at the IDMA region, so subtract
+	 * off that initial offset.
+	 */
+	offset.region =
+		MPIPE_MMIO_ADDR__REGION_VAL_IDMA -
+		MPIPE_MMIO_ADDR__REGION_VAL_IDMA;
+	offset.ring = ring;
+	offset.bucket = bucket;
+	offset.ring_enable = (ring >= 0);
+	offset.bucket_enable = (bucket >= 0);
+	val.count = count;
+
+	__gxio_mmio_write(context->mmio_fast_base + offset.word, val.word);
+}
+
+/*****************************************************************
+ *                         Egress Rings                           *
+ ******************************************************************/
+
+/* Allocate a set of eDMA rings.
+ *
+ * The return value is NOT interesting if count is zero.
+ *
+ * @param context An initialized mPIPE context.
+ * @param count Number of eDMA rings required.
+ * @param first Index of first eDMA ring if ::GXIO_MPIPE_ALLOC_FIXED flag
+ *   is set, otherwise ignored.
+ * @param flags Flag bits from ::gxio_mpipe_alloc_flags_e.
+ * @return Index of first allocated buffer eDMA ring, or
+ * ::GXIO_MPIPE_ERR_NO_EDMA_RING if allocation failed.
+ */
+extern int gxio_mpipe_alloc_edma_rings(gxio_mpipe_context_t *context,
+				       unsigned int count, unsigned int first,
+				       unsigned int flags);
+
+/* Initialize an eDMA ring, using the given memory and size.
+ *
+ * @param context An initialized mPIPE context.
+ * @param ering The eDMA ring index.
+ * @param channel The channel to use.  This must be one of the channels
+ * associated with the context's set of open links.
+ * @param mem A physically contiguous region of memory to be filled
+ * with a ring of ::gxio_mpipe_edesc_t structures.
+ * @param mem_size Number of bytes in the ring.  Must be 512, 2048,
+ * 8192 or 65536, times 16 (i.e. sizeof(gxio_mpipe_edesc_t)).
+ * @param mem_flags ::gxio_mpipe_mem_flags_e memory flags.
+ *
+ * @return 0 on success, ::GXIO_MPIPE_ERR_BAD_EDMA_RING or
+ * ::GXIO_ERR_INVAL_MEMORY_SIZE on failure.
+ */
+extern int gxio_mpipe_init_edma_ring(gxio_mpipe_context_t *context,
+				     unsigned int ering, unsigned int channel,
+				     void *mem, size_t mem_size,
+				     unsigned int mem_flags);
+
+/* Set the "max_blks", "min_snf_blks", and "db" fields of
+ * ::MPIPE_EDMA_RG_INIT_DAT_THRESH_t for a given edma ring.
+ *
+ * The global pool of dynamic blocks will be automatically adjusted.
+ *
+ * This function should not be called after any egress has been done
+ * on the edma ring.
+ *
+ * Most applications should just use gxio_mpipe_equeue_set_snf_size().
+ *
+ * @param context An initialized mPIPE context.
+ * @param ering The eDMA ring index.
+ * @param max_blks The number of blocks to dedicate to the ring
+ * (normally min_snf_blks + 1).  Must be greater than min_snf_blocks.
+ * @param min_snf_blks The number of blocks which must be stored
+ * prior to starting to send the packet (normally 12).
+ * @param db Whether to allow use of dynamic blocks by the ring
+ * (normally 1).
+ *
+ * @return 0 on success, negative on error.
+ */
+extern int gxio_mpipe_config_edma_ring_blks(gxio_mpipe_context_t *context,
+					    unsigned int ering,
+					    unsigned int max_blks,
+					    unsigned int min_snf_blks,
+					    unsigned int db);
+
+/*****************************************************************
+ *                      Classifier Program                        *
+ ******************************************************************/
+
+/*
+ *
+ * Functions for loading or configuring the mPIPE classifier program.
+ *
+ * The mPIPE classification processors all run a special "classifier"
+ * program which, for each incoming packet, parses the packet headers,
+ * encodes some packet metadata in the "idesc", and either drops the
+ * packet, or picks a notif ring to handle the packet, and a buffer
+ * stack to contain the packet, usually based on the channel, VLAN,
+ * dMAC, flow hash, and packet size, under the guidance of the "rules"
+ * API described below.
+ *
+ * @section gxio_mpipe_classifier_default Default Classifier
+ *
+ * The MDE provides a simple "default" classifier program.  It is
+ * shipped as source in "$TILERA_ROOT/src/sys/mpipe/classifier.c",
+ * which serves as its official documentation.  It is shipped as a
+ * binary program in "$TILERA_ROOT/tile/boot/classifier", which is
+ * automatically included in bootroms created by "tile-monitor", and
+ * is automatically loaded by the hypervisor at boot time.
+ *
+ * The L2 analysis handles LLC packets, SNAP packets, and "VLAN
+ * wrappers" (keeping the outer VLAN).
+ *
+ * The L3 analysis handles IPv4 and IPv6, dropping packets with bad
+ * IPv4 header checksums, requesting computation of a TCP/UDP checksum
+ * if appropriate, and hashing the dest and src IP addresses, plus the
+ * ports for TCP/UDP packets, into the flow hash.  No special analysis
+ * is done for "fragmented" packets or "tunneling" protocols.  Thus,
+ * the first fragment of a fragmented TCP/UDP packet is hashed using
+ * src/dest IP address and ports and all subsequent fragments are only
+ * hashed according to src/dest IP address.
+ *
+ * The L3 analysis handles other packets too, hashing the dMAC
+ * smac into a flow hash.
+ *
+ * The channel, VLAN, and dMAC used to pick a "rule" (see the
+ * "rules" APIs below), which in turn is used to pick a buffer stack
+ * (based on the packet size) and a bucket (based on the flow hash).
+ *
+ * To receive traffic matching a particular (channel/VLAN/dMAC
+ * pattern, an application should allocate its own buffer stacks and
+ * load balancer buckets, and map traffic to those stacks and buckets,
+ * as decribed by the "rules" API below.
+ *
+ * Various packet metadata is encoded in the idesc.  The flow hash is
+ * four bytes at 0x0C.  The VLAN is two bytes at 0x10.  The ethtype is
+ * two bytes at 0x12.  The l3 start is one byte at 0x14.  The l4 start
+ * is one byte at 0x15 for IPv4 and IPv6 packets, and otherwise zero.
+ * The protocol is one byte at 0x16 for IPv4 and IPv6 packets, and
+ * otherwise zero.
+ *
+ * @section gxio_mpipe_classifier_custom Custom Classifiers.
+ *
+ * A custom classifier may be created using "tile-mpipe-cc" with a
+ * customized version of the default classifier sources.
+ *
+ * The custom classifier may be included in bootroms using the
+ * "--classifier" option to "tile-monitor", or loaded dynamically
+ * using gxio_mpipe_classifier_load_from_file().
+ *
+ * Be aware that "extreme" customizations may break the assumptions of
+ * the "rules" APIs described below, but simple customizations, such
+ * as adding new packet metadata, should be fine.
+ */
+
+/* A set of classifier rules, plus a context. */
+typedef struct {
+
+	/* The context. */
+	gxio_mpipe_context_t *context;
+
+	/* The actual rules. */
+	gxio_mpipe_rules_list_t list;
+
+} gxio_mpipe_rules_t;
+
+/* Initialize a classifier program rules list.
+ *
+ * This function can be called on a previously initialized rules list
+ * to discard any previously added rules.
+ *
+ * @param rules Rules list to initialize.
+ * @param context An initialized mPIPE context.
+ */
+extern void gxio_mpipe_rules_init(gxio_mpipe_rules_t *rules,
+				  gxio_mpipe_context_t *context);
+
+/* Begin a new rule on the indicated rules list.
+ *
+ * Note that an empty rule matches all packets, but an empty rule list
+ * matches no packets.
+ *
+ * @param rules Rules list to which new rule is appended.
+ * @param bucket First load balancer bucket to which packets will be
+ * delivered.
+ * @param num_buckets Number of buckets (must be a power of two) across
+ * which packets will be distributed based on the "flow hash".
+ * @param stacks Either NULL, to assign each packet to the smallest
+ * initialized buffer stack which does not induce chaining (and to
+ * drop packets which exceed the largest initialized buffer stack
+ * buffer size), or an array, with each entry indicating which buffer
+ * stack should be used for packets up to that size (with 255
+ * indicating that those packets should be dropped).
+ * @return 0 on success, or a negative error code on failure.
+ */
+extern int gxio_mpipe_rules_begin(gxio_mpipe_rules_t *rules,
+				  unsigned int bucket,
+				  unsigned int num_buckets,
+				  gxio_mpipe_rules_stacks_t *stacks);
+
+/* Set the headroom of the current rule.
+ *
+ * @param rules Rules list whose current rule will be modified.
+ * @param headroom The headroom.
+ * @return 0 on success, or a negative error code on failure.
+ */
+extern int gxio_mpipe_rules_set_headroom(gxio_mpipe_rules_t *rules,
+					 uint8_t headroom);
+
+/* Indicate that packets from a particular channel can be delivered
+ * to the buckets and buffer stacks associated with the current rule.
+ *
+ * Channels added must be associated with links opened by the mPIPE context
+ * used in gxio_mpipe_rules_init().  A rule with no channels is equivalent
+ * to a rule naming all such associated channels.
+ *
+ * @param rules Rules list whose current rule will be modified.
+ * @param channel The channel to add.
+ * @return 0 on success, or a negative error code on failure.
+ */
+extern int gxio_mpipe_rules_add_channel(gxio_mpipe_rules_t *rules,
+					unsigned int channel);
+
+/* Commit rules.
+ *
+ * The rules are sent to the hypervisor, where they are combined with
+ * the rules from other apps, and used to program the hardware classifier.
+ *
+ * Note that if this function returns an error, then the rules will NOT
+ * have been committed, even if the error is due to interactions with
+ * rules from another app.
+ *
+ * @param rules Rules list to commit.
+ * @return 0 on success, or a negative error code on failure.
+ */
+extern int gxio_mpipe_rules_commit(gxio_mpipe_rules_t *rules);
+
+/*****************************************************************
+ *                     Ingress Queue Wrapper                      *
+ ******************************************************************/
+
+/*
+ *
+ * Convenience functions for receiving packets from a NotifRing and
+ * sending packets via an eDMA ring.
+ *
+ * The mpipe ingress and egress hardware uses shared memory packet
+ * descriptors to describe packets that have arrived on ingress or
+ * are destined for egress.  These descriptors are stored in shared
+ * memory ring buffers and written or read by hardware as necessary.
+ * The gxio library provides wrapper functions that manage the head and
+ * tail pointers for these rings, allowing the user to easily read or
+ * write packet descriptors.
+ *
+ * The initialization interface for ingress and egress rings is quite
+ * similar.  For example, to create an ingress queue, the user passes
+ * a ::gxio_mpipe_iqueue_t state object, a ring number from
+ * gxio_mpipe_alloc_notif_rings(), and the address of memory to hold a
+ * ring buffer to the gxio_mpipe_iqueue_init() function.  The function
+ * returns success when the state object has been initialized and the
+ * hardware configured to deliver packets to the specified ring
+ * buffer.  Similarly, gxio_mpipe_equeue_init() takes a
+ * ::gxio_mpipe_equeue_t state object, a ring number from
+ * gxio_mpipe_alloc_edma_rings(), and a shared memory buffer.
+ *
+ * @section gxio_mpipe_iqueue Working with Ingress Queues
+ *
+ * Once initialized, the gxio_mpipe_iqueue_t API provides two flows
+ * for getting the ::gxio_mpipe_idesc_t packet descriptor associated
+ * with incoming packets.  The simplest is to call
+ * gxio_mpipe_iqueue_get() or gxio_mpipe_iqueue_try_get().  These
+ * functions copy the oldest packet descriptor out of the NotifRing and
+ * into a descriptor provided by the caller.  They also immediately
+ * inform the hardware that a descriptor has been processed.
+ *
+ * For applications with stringent performance requirements, higher
+ * efficiency can be achieved by avoiding the packet descriptor copy
+ * and processing multiple descriptors at once.  The
+ * gxio_mpipe_iqueue_peek() and gxio_mpipe_iqueue_try_peek() functions
+ * allow such optimizations.  These functions provide a pointer to the
+ * next valid ingress descriptor in the NotifRing's shared memory ring
+ * buffer, and a count of how many contiguous descriptors are ready to
+ * be processed.  The application can then process any number of those
+ * descriptors in place, calling gxio_mpipe_iqueue_consume() to inform
+ * the hardware after each one has been processed.
+ *
+ * @section gxio_mpipe_equeue Working with Egress Queues
+ *
+ * Similarly, the egress queue API provides a high-performance
+ * interface plus a simple wrapper for use in posting
+ * ::gxio_mpipe_edesc_t egress packet descriptors.  The simple
+ * version, gxio_mpipe_equeue_put(), allows the programmer to wait for
+ * an eDMA ring slot to become available and write a single descriptor
+ * into the ring.
+ *
+ * Alternatively, you can reserve slots in the eDMA ring using
+ * gxio_mpipe_equeue_reserve() or gxio_mpipe_equeue_try_reserve(), and
+ * then fill in each slot using gxio_mpipe_equeue_put_at().  This
+ * capability can be used to amortize the cost of reserving slots
+ * across several packets.  It also allows gather operations to be
+ * performed on a shared equeue, by ensuring that the edescs for all
+ * the fragments are all contiguous in the eDMA ring.
+ *
+ * The gxio_mpipe_equeue_reserve() and gxio_mpipe_equeue_try_reserve()
+ * functions return a 63-bit "completion slot", which is actually a
+ * sequence number, the low bits of which indicate the ring buffer
+ * index and the high bits the number of times the application has
+ * gone around the egress ring buffer.  The extra bits allow an
+ * application to check for egress completion by calling
+ * gxio_mpipe_equeue_is_complete() to see whether a particular 'slot'
+ * number has finished.  Given the maximum packet rates of the Gx
+ * processor, the 63-bit slot number will never wrap.
+ *
+ * In practice, most applications use the ::gxio_mpipe_edesc_t::hwb
+ * bit to indicate that the buffers containing egress packet data
+ * should be pushed onto a buffer stack when egress is complete.  Such
+ * applications generally do not need to know when an egress operation
+ * completes (since there is no need to free a buffer post-egress),
+ * and thus can use the optimized gxio_mpipe_equeue_reserve_fast() or
+ * gxio_mpipe_equeue_try_reserve_fast() functions, which return a 24
+ * bit "slot", instead of a 63-bit "completion slot".
+ *
+ * Once a slot has been "reserved", it MUST be filled.  If the
+ * application reserves a slot and then decides that it does not
+ * actually need it, it can set the ::gxio_mpipe_edesc_t::ns (no send)
+ * bit on the descriptor passed to gxio_mpipe_equeue_put_at() to
+ * indicate that no data should be sent.  This technique can also be
+ * used to drop an incoming packet, instead of forwarding it, since
+ * any buffer will still be pushed onto the buffer stack when the
+ * egress descriptor is processed.
+ */
+
+/* A convenient interface to a NotifRing, for use by a single thread.
+ */
+typedef struct {
+
+	/* The context. */
+	gxio_mpipe_context_t *context;
+
+	/* The actual NotifRing. */
+	gxio_mpipe_idesc_t *idescs;
+
+	/* The number of entries. */
+	unsigned long num_entries;
+
+	/* The number of entries minus one. */
+	unsigned long mask_num_entries;
+
+	/* The log2() of the number of entries. */
+	unsigned long log2_num_entries;
+
+	/* The next entry. */
+	unsigned int head;
+
+	/* The NotifRing id. */
+	unsigned int ring;
+
+#ifdef __BIG_ENDIAN__
+	/* The number of byteswapped entries. */
+	unsigned int swapped;
+#endif
+
+} gxio_mpipe_iqueue_t;
+
+/* Initialize an "iqueue".
+ *
+ * Takes the iqueue plus the same args as gxio_mpipe_init_notif_ring().
+ */
+extern int gxio_mpipe_iqueue_init(gxio_mpipe_iqueue_t *iqueue,
+				  gxio_mpipe_context_t *context,
+				  unsigned int ring,
+				  void *mem, size_t mem_size,
+				  unsigned int mem_flags);
+
+/* Advance over some old entries in an iqueue.
+ *
+ * Please see the documentation for gxio_mpipe_iqueue_consume().
+ *
+ * @param iqueue An ingress queue initialized via gxio_mpipe_iqueue_init().
+ * @param count The number of entries to advance over.
+ */
+static inline void gxio_mpipe_iqueue_advance(gxio_mpipe_iqueue_t *iqueue,
+					     int count)
+{
+	/* Advance with proper wrap. */
+	int head = iqueue->head + count;
+	iqueue->head =
+		(head & iqueue->mask_num_entries) +
+		(head >> iqueue->log2_num_entries);
+
+#ifdef __BIG_ENDIAN__
+	/* HACK: Track swapped entries. */
+	iqueue->swapped -= count;
+#endif
+}
+
+/* Release the ring and bucket for an old entry in an iqueue.
+ *
+ * Releasing the ring allows more packets to be delivered to the ring.
+ *
+ * Releasing the bucket allows flows using the bucket to be moved to a
+ * new ring when using GXIO_MPIPE_BUCKET_DYNAMIC_FLOW_AFFINITY.
+ *
+ * This function is shorthand for "gxio_mpipe_credit(iqueue->context,
+ * iqueue->ring, idesc->bucket_id, 1)", and it may be more convenient
+ * to make that underlying call, using those values, instead of
+ * tracking the entire "idesc".
+ *
+ * If packet processing is deferred, optimal performance requires that
+ * the releasing be deferred as well.
+ *
+ * Please see the documentation for gxio_mpipe_iqueue_consume().
+ *
+ * @param iqueue An ingress queue initialized via gxio_mpipe_iqueue_init().
+ * @param idesc The descriptor which was processed.
+ */
+static inline void gxio_mpipe_iqueue_release(gxio_mpipe_iqueue_t *iqueue,
+					     gxio_mpipe_idesc_t *idesc)
+{
+	gxio_mpipe_credit(iqueue->context, iqueue->ring, idesc->bucket_id, 1);
+}
+
+/* Consume a packet from an "iqueue".
+ *
+ * After processing packets peeked at via gxio_mpipe_iqueue_peek()
+ * or gxio_mpipe_iqueue_try_peek(), you must call this function, or
+ * gxio_mpipe_iqueue_advance() plus gxio_mpipe_iqueue_release(), to
+ * advance over those entries, and release their rings and buckets.
+ *
+ * You may call this function as each packet is processed, or you can
+ * wait until several packets have been processed.
+ *
+ * Note that if you are using a single bucket, and you are handling
+ * batches of N packets, then you can replace several calls to this
+ * function with calls to "gxio_mpipe_iqueue_advance(iqueue, N)" and
+ * "gxio_mpipe_credit(iqueue->context, iqueue->ring, bucket, N)".
+ *
+ * Note that if your classifier sets "idesc->nr", then you should
+ * explicitly call "gxio_mpipe_iqueue_advance(iqueue, idesc)" plus
+ * "gxio_mpipe_credit(iqueue->context, iqueue->ring, -1, 1)", to
+ * avoid incorrectly crediting the (unused) bucket.
+ *
+ * @param iqueue An ingress queue initialized via gxio_mpipe_iqueue_init().
+ * @param idesc The descriptor which was processed.
+ */
+static inline void gxio_mpipe_iqueue_consume(gxio_mpipe_iqueue_t *iqueue,
+					     gxio_mpipe_idesc_t *idesc)
+{
+	gxio_mpipe_iqueue_advance(iqueue, 1);
+	gxio_mpipe_iqueue_release(iqueue, idesc);
+}
+
+/* Peek at the next packet(s) in an "iqueue", without waiting.
+ *
+ * If no packets are available, fills idesc_ref with NULL, and then
+ * returns ::GXIO_MPIPE_ERR_IQUEUE_EMPTY.  Otherwise, fills idesc_ref
+ * with the address of the next valid packet descriptor, and returns
+ * the maximum number of valid descriptors which can be processed.
+ * You may process fewer descriptors if desired.
+ *
+ * Call gxio_mpipe_iqueue_consume() on each packet once it has been
+ * processed (or dropped), to allow more packets to be delivered.
+ *
+ * @param iqueue An ingress queue initialized via gxio_mpipe_iqueue_init().
+ * @param idesc_ref A pointer to a packet descriptor pointer.
+ * @return The (positive) number of packets which can be processed,
+ * or ::GXIO_MPIPE_ERR_IQUEUE_EMPTY if no packets are available.
+ */
+static inline int gxio_mpipe_iqueue_try_peek(gxio_mpipe_iqueue_t *iqueue,
+					     gxio_mpipe_idesc_t **idesc_ref)
+{
+	gxio_mpipe_idesc_t *next;
+
+	uint64_t head = iqueue->head;
+	uint64_t tail = __gxio_mmio_read(iqueue->idescs);
+
+	/* Available entries. */
+	uint64_t avail =
+		(tail >= head) ? (tail - head) : (iqueue->num_entries - head);
+
+	if (avail == 0) {
+		*idesc_ref = NULL;
+		return GXIO_MPIPE_ERR_IQUEUE_EMPTY;
+	}
+
+	next = &iqueue->idescs[head];
+
+	/* ISSUE: Is this helpful? */
+	__insn_prefetch(next);
+
+#ifdef __BIG_ENDIAN__
+	/* HACK: Swap new entries directly in memory. */
+	{
+		int i, j;
+		for (i = iqueue->swapped; i < avail; i++) {
+			for (j = 0; j < 8; j++)
+				next[i].words[j] =
+					__builtin_bswap64(next[i].words[j]);
+		}
+		iqueue->swapped = avail;
+	}
+#endif
+
+	*idesc_ref = next;
+
+	return avail;
+}
+
+/* Drop a packet by pushing its buffer (if appropriate).
+ *
+ * NOTE: The caller must still call gxio_mpipe_iqueue_consume() if idesc
+ * came from gxio_mpipe_iqueue_try_peek() or gxio_mpipe_iqueue_peek().
+ *
+ * @param iqueue An ingress queue initialized via gxio_mpipe_iqueue_init().
+ * @param idesc A packet descriptor.
+ */
+static inline void gxio_mpipe_iqueue_drop(gxio_mpipe_iqueue_t *iqueue,
+					  gxio_mpipe_idesc_t *idesc)
+{
+	/* FIXME: Handle "chaining" properly. */
+
+	if (!idesc->be) {
+		unsigned char *va = gxio_mpipe_idesc_get_va(idesc);
+		gxio_mpipe_push_buffer(iqueue->context, idesc->stack_idx, va);
+	}
+}
+
+/*****************************************************************
+ *                      Egress Queue Wrapper                      *
+ ******************************************************************/
+
+/* A convenient, thread-safe interface to an eDMA ring. */
+typedef struct {
+
+	/* State object for tracking head and tail pointers. */
+	__gxio_dma_queue_t dma_queue;
+
+	/* The ring entries. */
+	gxio_mpipe_edesc_t *edescs;
+
+	/* The number of entries minus one. */
+	unsigned long mask_num_entries;
+
+	/* The log2() of the number of entries. */
+	unsigned long log2_num_entries;
+
+	/* The context. */
+	gxio_mpipe_context_t *context;
+
+	/* The ering. */
+	unsigned int ering;
+
+	/* The channel. */
+	unsigned int channel;
+
+} gxio_mpipe_equeue_t;
+
+/* Initialize an "equeue".
+ *
+ * This function uses gxio_mpipe_init_edma_ring() to initialize the
+ * underlying edma_ring using the provided arguments.
+ *
+ * @param equeue An egress queue to be initialized.
+ * @param context An initialized mPIPE context.
+ * @param ering The eDMA ring index.
+ * @param channel The channel to use.  This must be one of the channels
+ * associated with the context's set of open links.
+ * @param mem A physically contiguous region of memory to be filled
+ * with a ring of ::gxio_mpipe_edesc_t structures.
+ * @param mem_size Number of bytes in the ring.  Must be 512, 2048,
+ * 8192 or 65536, times 16 (i.e. sizeof(gxio_mpipe_edesc_t)).
+ * @param mem_flags ::gxio_mpipe_mem_flags_e memory flags.
+ *
+ * @return 0 on success, ::GXIO_MPIPE_ERR_BAD_EDMA_RING or
+ * ::GXIO_ERR_INVAL_MEMORY_SIZE on failure.
+ */
+extern int gxio_mpipe_equeue_init(gxio_mpipe_equeue_t *equeue,
+				  gxio_mpipe_context_t *context,
+				  unsigned int ering,
+				  unsigned int channel,
+				  void *mem, unsigned int mem_size,
+				  unsigned int mem_flags);
+
+/* Reserve completion slots for edescs.
+ *
+ * Use gxio_mpipe_equeue_put_at() to actually populate the slots.
+ *
+ * This function is slower than gxio_mpipe_equeue_reserve_fast(), but
+ * returns a full 64 bit completion slot, which can be used with
+ * gxio_mpipe_equeue_is_complete().
+ *
+ * @param equeue An egress queue initialized via gxio_mpipe_equeue_init().
+ * @param num Number of slots to reserve (must be non-zero).
+ * @return The first reserved completion slot, or a negative error code.
+ */
+static inline int64_t gxio_mpipe_equeue_reserve(gxio_mpipe_equeue_t *equeue,
+						unsigned int num)
+{
+	return __gxio_dma_queue_reserve_aux(&equeue->dma_queue, num, true);
+}
+
+/* Reserve completion slots for edescs, if possible.
+ *
+ * Use gxio_mpipe_equeue_put_at() to actually populate the slots.
+ *
+ * This function is slower than gxio_mpipe_equeue_try_reserve_fast(),
+ * but returns a full 64 bit completion slot, which can be used with
+ * gxio_mpipe_equeue_is_complete().
+ *
+ * @param equeue An egress queue initialized via gxio_mpipe_equeue_init().
+ * @param num Number of slots to reserve (must be non-zero).
+ * @return The first reserved completion slot, or a negative error code.
+ */
+static inline int64_t gxio_mpipe_equeue_try_reserve(gxio_mpipe_equeue_t
+						    *equeue, unsigned int num)
+{
+	return __gxio_dma_queue_reserve_aux(&equeue->dma_queue, num, false);
+}
+
+/* Reserve slots for edescs.
+ *
+ * Use gxio_mpipe_equeue_put_at() to actually populate the slots.
+ *
+ * This function is faster than gxio_mpipe_equeue_reserve(), but
+ * returns a 24 bit slot (instead of a 64 bit completion slot), which
+ * thus cannot be used with gxio_mpipe_equeue_is_complete().
+ *
+ * @param equeue An egress queue initialized via gxio_mpipe_equeue_init().
+ * @param num Number of slots to reserve (should be non-zero).
+ * @return The first reserved slot, or a negative error code.
+ */
+static inline int64_t gxio_mpipe_equeue_reserve_fast(gxio_mpipe_equeue_t
+						     *equeue, unsigned int num)
+{
+	return __gxio_dma_queue_reserve(&equeue->dma_queue, num, true, false);
+}
+
+/* Reserve slots for edescs, if possible.
+ *
+ * Use gxio_mpipe_equeue_put_at() to actually populate the slots.
+ *
+ * This function is faster than gxio_mpipe_equeue_try_reserve(), but
+ * returns a 24 bit slot (instead of a 64 bit completion slot), which
+ * thus cannot be used with gxio_mpipe_equeue_is_complete().
+ *
+ * @param equeue An egress queue initialized via gxio_mpipe_equeue_init().
+ * @param num Number of slots to reserve (should be non-zero).
+ * @return The first reserved slot, or a negative error code.
+ */
+static inline int64_t gxio_mpipe_equeue_try_reserve_fast(gxio_mpipe_equeue_t
+							 *equeue,
+							 unsigned int num)
+{
+	return __gxio_dma_queue_reserve(&equeue->dma_queue, num, false, false);
+}
+
+/*
+ * HACK: This helper function tricks gcc 4.6 into avoiding saving
+ * a copy of "edesc->words[0]" on the stack for no obvious reason.
+ */
+
+static inline void gxio_mpipe_equeue_put_at_aux(gxio_mpipe_equeue_t *equeue,
+						uint_reg_t ew[2],
+						unsigned long slot)
+{
+	unsigned long edma_slot = slot & equeue->mask_num_entries;
+	gxio_mpipe_edesc_t *edesc_p = &equeue->edescs[edma_slot];
+
+	/*
+	 * ISSUE: Could set eDMA ring to be on generation 1 at start, which
+	 * would avoid the negation here, perhaps allowing "__insn_bfins()".
+	 */
+	ew[0] |= !((slot >> equeue->log2_num_entries) & 1);
+
+	/*
+	 * NOTE: We use "__gxio_mpipe_write()", plus the fact that the eDMA
+	 * queue alignment restrictions ensure that these two words are on
+	 * the same cacheline, to force proper ordering between the stores.
+	 */
+	__gxio_mmio_write64(&edesc_p->words[1], ew[1]);
+	__gxio_mmio_write64(&edesc_p->words[0], ew[0]);
+}
+
+/* Post an edesc to a given slot in an equeue.
+ *
+ * This function copies the supplied edesc into entry "slot mod N" in
+ * the underlying ring, setting the "gen" bit to the appropriate value
+ * based on "(slot mod N*2)", where "N" is the size of the ring.  Note
+ * that the higher bits of slot are unused, and thus, this function
+ * can handle "slots" as well as "completion slots".
+ *
+ * Normally this function is used to fill in slots reserved by
+ * gxio_mpipe_equeue_try_reserve(), gxio_mpipe_equeue_reserve(),
+ * gxio_mpipe_equeue_try_reserve_fast(), or
+ * gxio_mpipe_equeue_reserve_fast(),
+ *
+ * This function can also be used without "reserving" slots, if the
+ * application KNOWS that the ring can never overflow, for example, by
+ * pushing fewer buffers into the buffer stacks than there are total
+ * slots in the equeue, but this is NOT recommended.
+ *
+ * @param equeue An egress queue initialized via gxio_mpipe_equeue_init().
+ * @param edesc The egress descriptor to be posted.
+ * @param slot An egress slot (only the low bits are actually used).
+ */
+static inline void gxio_mpipe_equeue_put_at(gxio_mpipe_equeue_t *equeue,
+					    gxio_mpipe_edesc_t edesc,
+					    unsigned long slot)
+{
+	gxio_mpipe_equeue_put_at_aux(equeue, edesc.words, slot);
+}
+
+/* Post an edesc to the next slot in an equeue.
+ *
+ * This is a convenience wrapper around
+ * gxio_mpipe_equeue_reserve_fast() and gxio_mpipe_equeue_put_at().
+ *
+ * @param equeue An egress queue initialized via gxio_mpipe_equeue_init().
+ * @param edesc The egress descriptor to be posted.
+ * @return 0 on success.
+ */
+static inline int gxio_mpipe_equeue_put(gxio_mpipe_equeue_t *equeue,
+					gxio_mpipe_edesc_t edesc)
+{
+	int64_t slot = gxio_mpipe_equeue_reserve_fast(equeue, 1);
+	if (slot < 0)
+		return (int)slot;
+
+	gxio_mpipe_equeue_put_at(equeue, edesc, slot);
+
+	return 0;
+}
+
+/* Ask the mPIPE hardware to egress outstanding packets immediately.
+ *
+ * This call is not necessary, but may slightly reduce overall latency.
+ *
+ * Technically, you should flush all gxio_mpipe_equeue_put_at() writes
+ * to memory before calling this function, to ensure the descriptors
+ * are visible in memory before the mPIPE hardware actually looks for
+ * them.  But this should be very rare, and the only side effect would
+ * be increased latency, so it is up to the caller to decide whether
+ * or not to flush memory.
+ *
+ * @param equeue An egress queue initialized via gxio_mpipe_equeue_init().
+ */
+static inline void gxio_mpipe_equeue_flush(gxio_mpipe_equeue_t *equeue)
+{
+	/* Use "ring_idx = 0" and "count = 0" to "wake up" the eDMA ring. */
+	MPIPE_EDMA_POST_REGION_VAL_t val = { {0} };
+	/* Flush the write buffers. */
+	__insn_flushwb();
+	__gxio_mmio_write(equeue->dma_queue.post_region_addr, val.word);
+}
+
+/* Determine if a given edesc has been completed.
+ *
+ * Note that this function requires a "completion slot", and thus may
+ * NOT be used with a "slot" from gxio_mpipe_equeue_reserve_fast() or
+ * gxio_mpipe_equeue_try_reserve_fast().
+ *
+ * @param equeue An egress queue initialized via gxio_mpipe_equeue_init().
+ * @param completion_slot The completion slot used by the edesc.
+ * @param update If true, and the desc does not appear to have completed
+ * yet, then update any software cache of the hardware completion counter,
+ * and check again.  This should normally be true.
+ * @return True iff the given edesc has been completed.
+ */
+static inline int gxio_mpipe_equeue_is_complete(gxio_mpipe_equeue_t *equeue,
+						int64_t completion_slot,
+						int update)
+{
+	return __gxio_dma_queue_is_complete(&equeue->dma_queue,
+					    completion_slot, update);
+}
+
+/* Set the snf (store and forward) size for an equeue.
+ *
+ * The snf size for an equeue defaults to 1536, and encodes the size
+ * of the largest packet for which egress is guaranteed to avoid
+ * transmission underruns and/or corrupt checksums under heavy load.
+ *
+ * The snf size affects a global resource pool which cannot support,
+ * for example, all 24 equeues each requesting an snf size of 8K.
+ *
+ * To ensure that jumbo packets can be egressed properly, the snf size
+ * should be set to the size of the largest possible packet, which
+ * will usually be limited by the size of the app's largest buffer.
+ *
+ * This is a convenience wrapper around
+ * gxio_mpipe_config_edma_ring_blks().
+ *
+ * This function should not be called after any egress has been done
+ * on the equeue.
+ *
+ * @param equeue An egress queue initialized via gxio_mpipe_equeue_init().
+ * @param size The snf size, in bytes.
+ * @return Zero on success, negative error otherwise.
+ */
+static inline int gxio_mpipe_equeue_set_snf_size(gxio_mpipe_equeue_t *equeue,
+						 size_t size)
+{
+	int blks = (size + 127) / 128;
+	return gxio_mpipe_config_edma_ring_blks(equeue->context, equeue->ering,
+						blks + 1, blks, 1);
+}
+
+/*****************************************************************
+ *                        Link Management                         *
+ ******************************************************************/
+
+/*
+ *
+ * Functions for manipulating and sensing the state and configuration
+ * of physical network links.
+ *
+ * @section gxio_mpipe_link_perm Link Permissions
+ *
+ * Opening a link (with gxio_mpipe_link_open()) requests a set of link
+ * permissions, which control what may be done with the link, and potentially
+ * what permissions may be granted to other processes.
+ *
+ * Data permission allows the process to receive packets from the link by
+ * specifying the link's channel number in mPIPE packet distribution rules,
+ * and to send packets to the link by using the link's channel number as
+ * the target for an eDMA ring.
+ *
+ * Stats permission allows the process to retrieve link attributes (such as
+ * the speeds it is capable of running at, or whether it is currently up), and
+ * to read and write certain statistics-related registers in the link's MAC.
+ *
+ * Control permission allows the process to retrieve and modify link attributes
+ * (so that it may, for example, bring the link up and take it down), and
+ * read and write many registers in the link's MAC and PHY.
+ *
+ * Any permission may be requested as shared, which allows other processes
+ * to also request shared permission, or exclusive, which prevents other
+ * processes from requesting it.  In keeping with GXIO's typical usage in
+ * an embedded environment, the defaults for all permissions are shared.
+ *
+ * Permissions are granted on a first-come, first-served basis, so if two
+ * applications request an exclusive permission on the same link, the one
+ * to run first will win.  Note, however, that some system components, like
+ * the kernel Ethernet driver, may get an opportunity to open links before
+ * any applications run.
+ *
+ * @section gxio_mpipe_link_names Link Names
+ *
+ * Link names are of the form gbe<em>number</em> (for Gigabit Ethernet),
+ * xgbe<em>number</em> (for 10 Gigabit Ethernet), loop<em>number</em> (for
+ * internal mPIPE loopback), or ilk<em>number</em>/<em>channel</em>
+ * (for Interlaken links); for instance, gbe0, xgbe1, loop3, and
+ * ilk0/12 are all possible link names.  The correspondence between
+ * the link name and an mPIPE instance number or mPIPE channel number is
+ * system-dependent; all links will not exist on all systems, and the set
+ * of numbers used for a particular link type may not start at zero and may
+ * not be contiguous.  Use gxio_mpipe_link_enumerate() to retrieve the set of
+ * links which exist on a system, and always use gxio_mpipe_link_instance()
+ * to determine which mPIPE controls a particular link.
+ *
+ * Note that in some cases, links may share hardware, such as PHYs, or
+ * internal mPIPE buffers; in these cases, only one of the links may be
+ * opened at a time.  This is especially common with xgbe and gbe ports,
+ * since each xgbe port uses 4 SERDES lanes, each of which may also be
+ * configured as one gbe port.
+ *
+ * @section gxio_mpipe_link_states Link States
+ *
+ * The mPIPE link management model revolves around three different states,
+ * which are maintained for each link:
+ *
+ * 1. The <em>current</em> link state: is the link up now, and if so, at
+ *    what speed?
+ *
+ * 2. The <em>desired</em> link state: what do we want the link state to be?
+ *    The system is always working to make this state the current state;
+ *    thus, if the desired state is up, and the link is down, we'll be
+ *    constantly trying to bring it up, automatically.
+ *
+ * 3. The <em>possible</em> link state: what speeds are valid for this
+ *    particular link?  Or, in other words, what are the capabilities of
+ *    the link hardware?
+ *
+ * These link states are not, strictly speaking, related to application
+ * state; they may be manipulated at any time, whether or not the link
+ * is currently being used for data transfer.  However, for convenience,
+ * gxio_mpipe_link_open() and gxio_mpipe_link_close() (or application exit)
+ * can affect the link state.  These implicit link management operations
+ * may be modified or disabled by the use of link open flags.
+ *
+ * From an application, you can use gxio_mpipe_link_get_attr()
+ * and gxio_mpipe_link_set_attr() to manipulate the link states.
+ * gxio_mpipe_link_get_attr() with ::GXIO_MPIPE_LINK_POSSIBLE_STATE
+ * gets you the possible link state.  gxio_mpipe_link_get_attr() with
+ * ::GXIO_MPIPE_LINK_CURRENT_STATE gets you the current link state.
+ * Finally, gxio_mpipe_link_set_attr() and gxio_mpipe_link_get_attr()
+ * with ::GXIO_MPIPE_LINK_DESIRED_STATE allow you to modify or retrieve
+ * the desired link state.
+ *
+ * If you want to manage a link from a part of your application which isn't
+ * involved in packet processing, you can use the ::GXIO_MPIPE_LINK_NO_DATA
+ * flags on a gxio_mpipe_link_open() call.  This opens the link, but does
+ * not request data permission, so it does not conflict with any exclusive
+ * permissions which may be held by other processes.  You can then can use
+ * gxio_mpipe_link_get_attr() and gxio_mpipe_link_set_attr() on this link
+ * object to bring up or take down the link.
+ *
+ * Some links support link state bits which support various loopback
+ * modes. ::GXIO_MPIPE_LINK_LOOP_MAC tests datapaths within the Tile
+ * Processor itself; ::GXIO_MPIPE_LINK_LOOP_PHY tests the datapath between
+ * the Tile Processor and the external physical layer interface chip; and
+ * ::GXIO_MPIPE_LINK_LOOP_EXT tests the entire network datapath with the
+ * aid of an external loopback connector.  In addition to enabling hardware
+ * testing, such configuration can be useful for software testing, as well.
+ *
+ * When LOOP_MAC or LOOP_PHY is enabled, packets transmitted on a channel
+ * will be received by that channel, instead of being emitted on the
+ * physical link, and packets received on the physical link will be ignored.
+ * Other than that, all standard GXIO operations work as you might expect.
+ * Note that loopback operation requires that the link be brought up using
+ * one or more of the GXIO_MPIPE_LINK_SPEED_xxx link state bits.
+ *
+ * Those familiar with previous versions of the MDE on TILEPro hardware
+ * will notice significant similarities between the NetIO link management
+ * model and the mPIPE link management model.  However, the NetIO model
+ * was developed in stages, and some of its features -- for instance,
+ * the default setting of certain flags -- were shaped by the need to be
+ * compatible with previous versions of NetIO.  Since the features provided
+ * by the mPIPE hardware and the mPIPE GXIO library are significantly
+ * different than those provided by NetIO, in some cases, we have made
+ * different choices in the mPIPE link management API.  Thus, please read
+ * this documentation carefully before assuming that mPIPE link management
+ * operations are exactly equivalent to their NetIO counterparts.
+ */
+
+/* An object used to manage mPIPE link state and resources. */
+typedef struct {
+	/* The overall mPIPE context. */
+	gxio_mpipe_context_t *context;
+
+	/* The channel number used by this link. */
+	uint8_t channel;
+
+	/* The MAC index used by this link. */
+	uint8_t mac;
+} gxio_mpipe_link_t;
+
+/* Translate a link name to the instance number of the mPIPE shim which is
+ *  connected to that link.  This call does not verify whether the link is
+ *  currently available, and does not reserve any link resources;
+ *  gxio_mpipe_link_open() must be called to perform those functions.
+ *
+ *  Typically applications will call this function to translate a link name
+ *  to an mPIPE instance number; call gxio_mpipe_init(), passing it that
+ *  instance number, to initialize the mPIPE shim; and then call
+ *  gxio_mpipe_link_open(), passing it the same link name plus the mPIPE
+ *  context, to configure the link.
+ *
+ * @param link_name Name of the link; see @ref gxio_mpipe_link_names.
+ * @return The mPIPE instance number which is associated with the named
+ *  link, or a negative error code (::GXIO_ERR_NO_DEVICE) if the link does
+ *  not exist.
+ */
+extern int gxio_mpipe_link_instance(const char *link_name);
+
+/* Retrieve one of this system's legal link names, and its MAC address.
+ *
+ * @param index Link name index.  If a system supports N legal link names,
+ *  then indices between 0 and N - 1, inclusive, each correspond to one of
+ *  those names.  Thus, to retrieve all of a system's legal link names,
+ *  call this function in a loop, starting with an index of zero, and
+ *  incrementing it once per iteration until -1 is returned.
+ * @param link_name Pointer to the buffer which will receive the retrieved
+ *  link name.  The buffer should contain space for at least
+ *  ::GXIO_MPIPE_LINK_NAME_LEN bytes; the returned name, including the
+ *  terminating null byte, will be no longer than that.
+ * @param link_name Pointer to the buffer which will receive the retrieved
+ *  MAC address.  The buffer should contain space for at least 6 bytes.
+ * @return Zero if a link name was successfully retrieved; -1 if one was
+ *  not.
+ */
+extern int gxio_mpipe_link_enumerate_mac(int index, char *link_name,
+					 uint8_t *mac_addr);
+
+/* Open an mPIPE link.
+ *
+ *  A link must be opened before it may be used to send or receive packets,
+ *  and before its state may be examined or changed.  Depending up on the
+ *  link's intended use, one or more link permissions may be requested via
+ *  the flags parameter; see @ref gxio_mpipe_link_perm.  In addition, flags
+ *  may request that the link's state be modified at open time.  See @ref
+ *  gxio_mpipe_link_states and @ref gxio_mpipe_link_open_flags for more detail.
+ *
+ * @param link A link state object, which will be initialized if this
+ *  function completes successfully.
+ * @param context An initialized mPIPE context.
+ * @param link_name Name of the link.
+ * @param flags Zero or more @ref gxio_mpipe_link_open_flags, ORed together.
+ * @return 0 if the link was successfully opened, or a negative error code.
+ *
+ */
+extern int gxio_mpipe_link_open(gxio_mpipe_link_t *link,
+				gxio_mpipe_context_t *context,
+				const char *link_name, unsigned int flags);
+
+/* Close an mPIPE link.
+ *
+ *  Closing a link makes it available for use by other processes.  Once
+ *  a link has been closed, packets may no longer be sent on or received
+ *  from the link, and its state may not be examined or changed.
+ *
+ * @param link A link state object, which will no longer be initialized
+ *  if this function completes successfully.
+ * @return 0 if the link was successfully closed, or a negative error code.
+ *
+ */
+extern int gxio_mpipe_link_close(gxio_mpipe_link_t *link);
+
+/* Return a link's channel number.
+ *
+ * @param link A properly initialized link state object.
+ * @return The channel number for the link.
+ */
+static inline int gxio_mpipe_link_channel(gxio_mpipe_link_t *link)
+{
+	return link->channel;
+}
+
+/* Set a link attribute.
+ *
+ * @param link A properly initialized link state object.
+ * @param attr An attribute from the set of @ref gxio_mpipe_link_attrs.
+ * @param val New value of the attribute.
+ * @return 0 if the attribute was successfully set, or a negative error
+ *  code.
+ */
+extern int gxio_mpipe_link_set_attr(gxio_mpipe_link_t *link, uint32_t attr,
+				    int64_t val);
+
+///////////////////////////////////////////////////////////////////
+//                             Timestamp                         //
+///////////////////////////////////////////////////////////////////
+
+/* Get the timestamp of mPIPE when this routine is called.
+ *
+ * @param context An initialized mPIPE context.
+ * @param ts A timespec structure to store the current clock.
+ * @return If the call was successful, zero; otherwise, a negative error
+ *  code.
+ */
+extern int gxio_mpipe_get_timestamp(gxio_mpipe_context_t *context,
+				    struct timespec64 *ts);
+
+/* Set the timestamp of mPIPE.
+ *
+ * @param context An initialized mPIPE context.
+ * @param ts A timespec structure to store the requested clock.
+ * @return If the call was successful, zero; otherwise, a negative error
+ *  code.
+ */
+extern int gxio_mpipe_set_timestamp(gxio_mpipe_context_t *context,
+				    const struct timespec64 *ts);
+
+/* Adjust the timestamp of mPIPE.
+ *
+ * @param context An initialized mPIPE context.
+ * @param delta A signed time offset to adjust, in nanoseconds.
+ * The absolute value of this parameter must be less than or
+ * equal to 1000000000.
+ * @return If the call was successful, zero; otherwise, a negative error
+ *  code.
+ */
+extern int gxio_mpipe_adjust_timestamp(gxio_mpipe_context_t *context,
+				       int64_t delta);
+
+/** Adjust the mPIPE timestamp clock frequency.
+ *
+ * @param context An initialized mPIPE context.
+ * @param ppb A 32-bit signed PPB (Parts Per Billion) value to adjust.
+ * The absolute value of ppb must be less than or equal to 1000000000.
+ * Values less than about 30000 will generally cause a GXIO_ERR_INVAL
+ * return due to the granularity of the hardware that converts reference
+ * clock cycles into seconds and nanoseconds.
+ * @return If the call was successful, zero; otherwise, a negative error
+ *  code.
+ */
+extern int gxio_mpipe_adjust_timestamp_freq(gxio_mpipe_context_t* context,
+                                            int32_t ppb);
+
+#endif /* !_GXIO_MPIPE_H_ */
diff --git a/kernel/arch/tile/include/gxio/trio.h b/kernel/arch/tile/include/gxio/trio.h
new file mode 100644
index 000000000..df10a662c
--- /dev/null
+++ b/kernel/arch/tile/include/gxio/trio.h
@@ -0,0 +1,298 @@
+/*
+ * Copyright 2012 Tilera Corporation. All Rights Reserved.
+ *
+ *   This program is free software; you can redistribute it and/or
+ *   modify it under the terms of the GNU General Public License
+ *   as published by the Free Software Foundation, version 2.
+ *
+ *   This program is distributed in the hope that it will be useful, but
+ *   WITHOUT ANY WARRANTY; without even the implied warranty of
+ *   MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
+ *   NON INFRINGEMENT.  See the GNU General Public License for
+ *   more details.
+ */
+
+/*
+ *
+ * An API for allocating, configuring, and manipulating TRIO hardware
+ * resources
+ */
+
+/*
+ *
+ * The TILE-Gx TRIO shim provides connections to external devices via
+ * PCIe or other transaction IO standards.  The gxio_trio_ API,
+ * declared in <gxio/trio.h>, allows applications to allocate and
+ * configure TRIO IO resources like DMA command rings, memory map
+ * windows, and device interrupts.  The following sections introduce
+ * the various components of the API.  We strongly recommend reading
+ * the TRIO section of the IO Device Guide (UG404) before working with
+ * this API.
+ *
+ * @section trio__ingress TRIO Ingress Hardware Resources
+ *
+ * The TRIO ingress hardware is responsible for examining incoming
+ * PCIe or StreamIO packets and choosing a processing mechanism based
+ * on the packets' bus address.  The gxio_trio_ API can be used to
+ * configure different handlers for different ranges of bus address
+ * space.  The user can configure "mapped memory" and "scatter queue"
+ * regions to match incoming packets within 4kB-aligned ranges of bus
+ * addresses.  Each range specifies a different set of mapping
+ * parameters to be applied when handling the ingress packet.  The
+ * following sections describe how to work with MapMem and scatter
+ * queue regions.
+ *
+ * @subsection trio__mapmem TRIO MapMem Regions
+ *
+ * TRIO mapped memory (or MapMem) regions allow the user to map
+ * incoming read and write requests directly to the application's
+ * memory space.  MapMem regions are allocated via
+ * gxio_trio_alloc_memory_maps().  Given an integer MapMem number,
+ * applications can use gxio_trio_init_memory_map() to specify the
+ * range of bus addresses that will match the region and the range of
+ * virtual addresses to which those packets will be applied.
+ *
+ * As with many other gxio APIs, the programmer must be sure to
+ * register memory pages that will be used with MapMem regions.  Pages
+ * can be registered with TRIO by allocating an ASID (address space
+ * identifier) and then using gxio_trio_register_page() to register up to
+ * 16 pages with the hardware.  The initialization functions for
+ * resources that require registered memory (MapMem, scatter queues,
+ * push DMA, and pull DMA) then take an 'asid' parameter in order to
+ * configure which set of registered pages is used by each resource.
+ *
+ * @subsection trio__scatter_queue TRIO Scatter Queues
+ *
+ * The TRIO shim's scatter queue regions allow users to dynamically
+ * map buffers from a large address space into a small range of bus
+ * addresses.  This is particularly helpful for PCIe endpoint devices,
+ * where the host generally limits the size of BARs to tens of
+ * megabytes.
+ *
+ * Each scatter queue consists of a memory map region, a queue of
+ * tile-side buffer VAs to be mapped to that region, and a bus-mapped
+ * "doorbell" register that the remote endpoint can write to trigger a
+ * dequeue of the current buffer VA, thus swapping in a new buffer.
+ * The VAs pushed onto a scatter queue must be 4kB aligned, so
+ * applications may need to use higher-level protocols to inform
+ * remote entities that they should apply some additional, sub-4kB
+ * offset when reading or writing the scatter queue region.  For more
+ * information, see the IO Device Guide (UG404).
+ *
+ * @section trio__egress TRIO Egress Hardware Resources
+ *
+ * The TRIO shim supports two mechanisms for egress packet generation:
+ * programmed IO (PIO) and push/pull DMA.  PIO allows applications to
+ * create MMIO mappings for PCIe or StreamIO address space, such that
+ * the application can generate word-sized read or write transactions
+ * by issuing load or store instructions.  Push and pull DMA are tuned
+ * for larger transactions; they use specialized hardware engines to
+ * transfer large blocks of data at line rate.
+ *
+ * @subsection trio__pio TRIO Programmed IO
+ *
+ * Programmed IO allows applications to create MMIO mappings for PCIe
+ * or StreamIO address space.  The hardware PIO regions support access
+ * to PCIe configuration, IO, and memory space, but the gxio_trio API
+ * only supports memory space accesses.  PIO regions are allocated
+ * with gxio_trio_alloc_pio_regions() and initialized via
+ * gxio_trio_init_pio_region().  Once a region is bound to a range of
+ * bus address via the initialization function, the application can
+ * use gxio_trio_map_pio_region() to create MMIO mappings from its VA
+ * space onto the range of bus addresses supported by the PIO region.
+ *
+ * @subsection trio_dma TRIO Push and Pull DMA
+ *
+ * The TRIO push and pull DMA engines allow users to copy blocks of
+ * data between application memory and the bus.  Push DMA generates
+ * write packets that copy from application memory to the bus and pull
+ * DMA generates read packets that copy from the bus into application
+ * memory.  The DMA engines are managed via an API that is very
+ * similar to the mPIPE eDMA interface.  For a detailed explanation of
+ * the eDMA queue API, see @ref gxio_mpipe_wrappers.
+ *
+ * Push and pull DMA queues are allocated via
+ * gxio_trio_alloc_push_dma_ring() / gxio_trio_alloc_pull_dma_ring().
+ * Once allocated, users generally use a ::gxio_trio_dma_queue_t
+ * object to manage the queue, providing easy wrappers for reserving
+ * command slots in the DMA command ring, filling those slots, and
+ * waiting for commands to complete.  DMA queues can be initialized
+ * via gxio_trio_init_push_dma_queue() or
+ * gxio_trio_init_pull_dma_queue().
+ *
+ * See @ref trio/push_dma/app.c for an example of how to use push DMA.
+ *
+ * @section trio_shortcomings Plans for Future API Revisions
+ *
+ * The simulation framework is incomplete.  Future features include:
+ *
+ * - Support for reset and deallocation of resources.
+ *
+ * - Support for pull DMA.
+ *
+ * - Support for interrupt regions and user-space interrupt delivery.
+ *
+ * - Support for getting BAR mappings and reserving regions of BAR
+ *   address space.
+ */
+#ifndef _GXIO_TRIO_H_
+#define _GXIO_TRIO_H_
+
+#include <linux/types.h>
+
+#include <gxio/common.h>
+#include <gxio/dma_queue.h>
+
+#include <arch/trio_constants.h>
+#include <arch/trio.h>
+#include <arch/trio_pcie_intfc.h>
+#include <arch/trio_pcie_rc.h>
+#include <arch/trio_shm.h>
+#include <hv/drv_trio_intf.h>
+#include <hv/iorpc.h>
+
+/* A context object used to manage TRIO hardware resources. */
+typedef struct {
+
+	/* File descriptor for calling up to Linux (and thus the HV). */
+	int fd;
+
+	/* The VA at which the MAC MMIO registers are mapped. */
+	char *mmio_base_mac;
+
+	/* The VA at which the PIO config space are mapped for each PCIe MAC.
+	   Gx36 has max 3 PCIe MACs per TRIO shim. */
+	char *mmio_base_pio_cfg[TILEGX_TRIO_PCIES];
+
+#ifdef USE_SHARED_PCIE_CONFIG_REGION
+	/* Index of the shared PIO region for PCI config access. */
+	int pio_cfg_index;
+#else
+	/* Index of the PIO region for PCI config access per MAC. */
+	int pio_cfg_index[TILEGX_TRIO_PCIES];
+#endif
+
+	/*  The VA at which the push DMA MMIO registers are mapped. */
+	char *mmio_push_dma[TRIO_NUM_PUSH_DMA_RINGS];
+
+	/*  The VA at which the pull DMA MMIO registers are mapped. */
+	char *mmio_pull_dma[TRIO_NUM_PUSH_DMA_RINGS];
+
+	/* Application space ID. */
+	unsigned int asid;
+
+} gxio_trio_context_t;
+
+/* Command descriptor for push or pull DMA. */
+typedef TRIO_DMA_DESC_t gxio_trio_dma_desc_t;
+
+/* A convenient, thread-safe interface to an eDMA ring. */
+typedef struct {
+
+	/* State object for tracking head and tail pointers. */
+	__gxio_dma_queue_t dma_queue;
+
+	/* The ring entries. */
+	gxio_trio_dma_desc_t *dma_descs;
+
+	/* The number of entries minus one. */
+	unsigned long mask_num_entries;
+
+	/* The log2() of the number of entries. */
+	unsigned int log2_num_entries;
+
+} gxio_trio_dma_queue_t;
+
+/* Initialize a TRIO context.
+ *
+ * This function allocates a TRIO "service domain" and maps the MMIO
+ * registers into the the caller's VA space.
+ *
+ * @param trio_index Which TRIO shim; Gx36 must pass 0.
+ * @param context Context object to be initialized.
+ */
+extern int gxio_trio_init(gxio_trio_context_t *context,
+			  unsigned int trio_index);
+
+/* This indicates that an ASID hasn't been allocated. */
+#define GXIO_ASID_NULL -1
+
+/* Ordering modes for map memory regions and scatter queue regions. */
+typedef enum gxio_trio_order_mode_e {
+	/* Writes are not ordered.  Reads always wait for previous writes. */
+	GXIO_TRIO_ORDER_MODE_UNORDERED =
+		TRIO_MAP_MEM_SETUP__ORDER_MODE_VAL_UNORDERED,
+	/* Both writes and reads wait for previous transactions to complete. */
+	GXIO_TRIO_ORDER_MODE_STRICT =
+		TRIO_MAP_MEM_SETUP__ORDER_MODE_VAL_STRICT,
+	/* Writes are ordered unless the incoming packet has the
+	   relaxed-ordering attributes set. */
+	GXIO_TRIO_ORDER_MODE_OBEY_PACKET =
+		TRIO_MAP_MEM_SETUP__ORDER_MODE_VAL_REL_ORD
+} gxio_trio_order_mode_t;
+
+/* Initialize a memory mapping region.
+ *
+ * @param context An initialized TRIO context.
+ * @param map A Memory map region allocated by gxio_trio_alloc_memory_map().
+ * @param target_mem VA of backing memory, should be registered via
+ *   gxio_trio_register_page() and aligned to 4kB.
+ * @param target_size Length of the memory mapping, must be a multiple
+ * of 4kB.
+ * @param asid ASID to be used for Tile-side address translation.
+ * @param mac MAC number.
+ * @param bus_address Bus address at which the mapping starts.
+ * @param order_mode Memory ordering mode for this mapping.
+ * @return Zero on success, else ::GXIO_TRIO_ERR_BAD_MEMORY_MAP,
+ * GXIO_TRIO_ERR_BAD_ASID, or ::GXIO_TRIO_ERR_BAD_BUS_RANGE.
+ */
+extern int gxio_trio_init_memory_map(gxio_trio_context_t *context,
+				     unsigned int map, void *target_mem,
+				     size_t target_size, unsigned int asid,
+				     unsigned int mac, uint64_t bus_address,
+				     gxio_trio_order_mode_t order_mode);
+
+/* Flags that can be passed to resource allocation functions. */
+enum gxio_trio_alloc_flags_e {
+	GXIO_TRIO_ALLOC_FIXED = HV_TRIO_ALLOC_FIXED,
+};
+
+/* Flags that can be passed to memory registration functions. */
+enum gxio_trio_mem_flags_e {
+	/* Do not fill L3 when writing, and invalidate lines upon egress. */
+	GXIO_TRIO_MEM_FLAG_NT_HINT = IORPC_MEM_BUFFER_FLAG_NT_HINT,
+
+	/* L3 cache fills should only populate IO cache ways. */
+	GXIO_TRIO_MEM_FLAG_IO_PIN = IORPC_MEM_BUFFER_FLAG_IO_PIN,
+};
+
+/* Flag indicating a request generator uses a special traffic
+    class. */
+#define GXIO_TRIO_FLAG_TRAFFIC_CLASS(N) HV_TRIO_FLAG_TC(N)
+
+/* Flag indicating a request generator uses a virtual function
+    number. */
+#define GXIO_TRIO_FLAG_VFUNC(N) HV_TRIO_FLAG_VFUNC(N)
+
+/*****************************************************************
+ *                       Memory Registration                      *
+ ******************************************************************/
+
+/* Allocate Application Space Identifiers (ASIDs).  Each ASID can
+ * register up to 16 page translations.  ASIDs are used by memory map
+ * regions, scatter queues, and DMA queues to translate application
+ * VAs into memory system PAs.
+ *
+ * @param context An initialized TRIO context.
+ * @param count Number of ASIDs required.
+ * @param first Index of first ASID if ::GXIO_TRIO_ALLOC_FIXED flag
+ *   is set, otherwise ignored.
+ * @param flags Flag bits, including bits from ::gxio_trio_alloc_flags_e.
+ * @return Index of first ASID, or ::GXIO_TRIO_ERR_NO_ASID if allocation
+ *   failed.
+ */
+extern int gxio_trio_alloc_asids(gxio_trio_context_t *context,
+				 unsigned int count, unsigned int first,
+				 unsigned int flags);
+
+#endif /* ! _GXIO_TRIO_H_ */
diff --git a/kernel/arch/tile/include/gxio/uart.h b/kernel/arch/tile/include/gxio/uart.h
new file mode 100644
index 000000000..438ee7e46
--- /dev/null
+++ b/kernel/arch/tile/include/gxio/uart.h
@@ -0,0 +1,105 @@
+/*
+ * Copyright 2013 Tilera Corporation. All Rights Reserved.
+ *
+ *   This program is free software; you can redistribute it and/or
+ *   modify it under the terms of the GNU General Public License
+ *   as published by the Free Software Foundation, version 2.
+ *
+ *   This program is distributed in the hope that it will be useful, but
+ *   WITHOUT ANY WARRANTY; without even the implied warranty of
+ *   MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
+ *   NON INFRINGEMENT.  See the GNU General Public License for
+ *   more details.
+ */
+
+#ifndef _GXIO_UART_H_
+#define _GXIO_UART_H_
+
+#include "common.h"
+
+#include <hv/drv_uart_intf.h>
+#include <hv/iorpc.h>
+
+/*
+ *
+ * An API for manipulating UART interface.
+ */
+
+/*
+ *
+ * The Rshim allows access to the processor's UART interface.
+ */
+
+/* A context object used to manage UART resources. */
+typedef struct {
+
+	/* File descriptor for calling up to the hypervisor. */
+	int fd;
+
+	/* The VA at which our MMIO registers are mapped. */
+	char *mmio_base;
+
+} gxio_uart_context_t;
+
+/* Request UART interrupts.
+ *
+ *  Request that interrupts be delivered to a tile when the UART's
+ *  Receive FIFO is written, or the Write FIFO is read.
+ *
+ * @param context Pointer to a properly initialized gxio_uart_context_t.
+ * @param bind_cpu_x X coordinate of CPU to which interrupt will be delivered.
+ * @param bind_cpu_y Y coordinate of CPU to which interrupt will be delivered.
+ * @param bind_interrupt IPI interrupt number.
+ * @param bind_event Sub-interrupt event bit number; a negative value can
+ *  disable the interrupt.
+ * @return Zero if all of the requested UART events were successfully
+ *  configured to interrupt.
+ */
+extern int gxio_uart_cfg_interrupt(gxio_uart_context_t *context,
+				   int bind_cpu_x,
+				   int bind_cpu_y,
+				   int bind_interrupt, int bind_event);
+
+/* Initialize a UART context.
+ *
+ *  A properly initialized context must be obtained before any of the other
+ *  gxio_uart routines may be used.
+ *
+ * @param context Pointer to a gxio_uart_context_t, which will be initialized
+ *  by this routine, if it succeeds.
+ * @param uart_index Index of the UART to use.
+ * @return Zero if the context was successfully initialized, else a
+ *  GXIO_ERR_xxx error code.
+ */
+extern int gxio_uart_init(gxio_uart_context_t *context, int uart_index);
+
+/* Destroy a UART context.
+ *
+ *  Once destroyed, a context may not be used with any gxio_uart routines
+ *  other than gxio_uart_init().  After this routine returns, no further
+ *  interrupts requested on this context will be delivered.  The state and
+ *  configuration of the pins which had been attached to this context are
+ *  unchanged by this operation.
+ *
+ * @param context Pointer to a gxio_uart_context_t.
+ * @return Zero if the context was successfully destroyed, else a
+ *  GXIO_ERR_xxx error code.
+ */
+extern int gxio_uart_destroy(gxio_uart_context_t *context);
+
+/* Write UART register.
+ * @param context Pointer to a gxio_uart_context_t.
+ * @param offset UART register offset.
+ * @param word Data will be wrote to UART reigister.
+ */
+extern void gxio_uart_write(gxio_uart_context_t *context, uint64_t offset,
+			    uint64_t word);
+
+/* Read UART register.
+ * @param context Pointer to a gxio_uart_context_t.
+ * @param offset UART register offset.
+ * @return Data read from UART register.
+ */
+extern uint64_t gxio_uart_read(gxio_uart_context_t *context, uint64_t offset);
+
+#endif /* _GXIO_UART_H_ */
diff --git a/kernel/arch/tile/include/gxio/usb_host.h b/kernel/arch/tile/include/gxio/usb_host.h
new file mode 100644
index 000000000..93c9636d2
--- /dev/null
+++ b/kernel/arch/tile/include/gxio/usb_host.h
@@ -0,0 +1,87 @@
+/*
+ * Copyright 2012 Tilera Corporation. All Rights Reserved.
+ *
+ *   This program is free software; you can redistribute it and/or
+ *   modify it under the terms of the GNU General Public License
+ *   as published by the Free Software Foundation, version 2.
+ *
+ *   This program is distributed in the hope that it will be useful, but
+ *   WITHOUT ANY WARRANTY; without even the implied warranty of
+ *   MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
+ *   NON INFRINGEMENT.  See the GNU General Public License for
+ *   more details.
+ */
+#ifndef _GXIO_USB_H_
+#define _GXIO_USB_H_
+
+#include <gxio/common.h>
+
+#include <hv/drv_usb_host_intf.h>
+#include <hv/iorpc.h>
+
+/*
+ *
+ * An API for manipulating general-purpose I/O pins.
+ */
+
+/*
+ *
+ * The USB shim allows access to the processor's Universal Serial Bus
+ * connections.
+ */
+
+/* A context object used to manage USB hardware resources. */
+typedef struct {
+
+	/* File descriptor for calling up to the hypervisor. */
+	int fd;
+
+	/* The VA at which our MMIO registers are mapped. */
+	char *mmio_base;
+} gxio_usb_host_context_t;
+
+/* Initialize a USB context.
+ *
+ *  A properly initialized context must be obtained before any of the other
+ *  gxio_usb_host routines may be used.
+ *
+ * @param context Pointer to a gxio_usb_host_context_t, which will be
+ *  initialized by this routine, if it succeeds.
+ * @param usb_index Index of the USB shim to use.
+ * @param is_ehci Nonzero to use the EHCI interface; zero to use the OHCI
+ *  intereface.
+ * @return Zero if the context was successfully initialized, else a
+ *  GXIO_ERR_xxx error code.
+ */
+extern int gxio_usb_host_init(gxio_usb_host_context_t *context, int usb_index,
+			      int is_ehci);
+
+/* Destroy a USB context.
+ *
+ *  Once destroyed, a context may not be used with any gxio_usb_host routines
+ *  other than gxio_usb_host_init().  After this routine returns, no further
+ *  interrupts or signals requested on this context will be delivered.  The
+ *  state and configuration of the pins which had been attached to this
+ *  context are unchanged by this operation.
+ *
+ * @param context Pointer to a gxio_usb_host_context_t.
+ * @return Zero if the context was successfully destroyed, else a
+ *  GXIO_ERR_xxx error code.
+ */
+extern int gxio_usb_host_destroy(gxio_usb_host_context_t *context);
+
+/* Retrieve the address of the shim's MMIO registers.
+ *
+ * @param context Pointer to a properly initialized gxio_usb_host_context_t.
+ * @return The address of the shim's MMIO registers.
+ */
+extern void *gxio_usb_host_get_reg_start(gxio_usb_host_context_t *context);
+
+/* Retrieve the length of the shim's MMIO registers.
+ *
+ * @param context Pointer to a properly initialized gxio_usb_host_context_t.
+ * @return The length of the shim's MMIO registers.
+ */
+extern size_t gxio_usb_host_get_reg_len(gxio_usb_host_context_t *context);
+
+#endif /* _GXIO_USB_H_ */
diff --git a/kernel/arch/tile/include/hv/drv_mpipe_intf.h b/kernel/arch/tile/include/hv/drv_mpipe_intf.h
new file mode 100644
index 000000000..c97e416dd
--- /dev/null
+++ b/kernel/arch/tile/include/hv/drv_mpipe_intf.h
@@ -0,0 +1,605 @@
+/*
+ * Copyright 2011 Tilera Corporation. All Rights Reserved.
+ *
+ *   This program is free software; you can redistribute it and/or
+ *   modify it under the terms of the GNU General Public License
+ *   as published by the Free Software Foundation, version 2.
+ *
+ *   This program is distributed in the hope that it will be useful, but
+ *   WITHOUT ANY WARRANTY; without even the implied warranty of
+ *   MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
+ *   NON INFRINGEMENT.  See the GNU General Public License for
+ *   more details.
+ */
+
+/**
+ * Interface definitions for the mpipe driver.
+ */
+
+#ifndef _SYS_HV_DRV_MPIPE_INTF_H
+#define _SYS_HV_DRV_MPIPE_INTF_H
+
+#include <arch/mpipe.h>
+#include <arch/mpipe_constants.h>
+
+
+/** Number of mPIPE instances supported */
+#define HV_MPIPE_INSTANCE_MAX   (2)
+
+/** Number of buffer stacks (32). */
+#define HV_MPIPE_NUM_BUFFER_STACKS \
+  (MPIPE_MMIO_INIT_DAT_GX36_1__BUFFER_STACK_MASK_WIDTH)
+
+/** Number of NotifRings (256). */
+#define HV_MPIPE_NUM_NOTIF_RINGS (MPIPE_NUM_NOTIF_RINGS)
+
+/** Number of NotifGroups (32). */
+#define HV_MPIPE_NUM_NOTIF_GROUPS (MPIPE_NUM_NOTIF_GROUPS)
+
+/** Number of buckets (4160). */
+#define HV_MPIPE_NUM_BUCKETS (MPIPE_NUM_BUCKETS)
+
+/** Number of "lo" buckets (4096). */
+#define HV_MPIPE_NUM_LO_BUCKETS 4096
+
+/** Number of "hi" buckets (64). */
+#define HV_MPIPE_NUM_HI_BUCKETS \
+  (HV_MPIPE_NUM_BUCKETS - HV_MPIPE_NUM_LO_BUCKETS)
+
+/** Number of edma rings (24). */
+#define HV_MPIPE_NUM_EDMA_RINGS \
+  (MPIPE_MMIO_INIT_DAT_GX36_1__EDMA_POST_MASK_WIDTH)
+
+
+
+
+/** A flag bit indicating a fixed resource allocation. */
+#define HV_MPIPE_ALLOC_FIXED 0x01
+
+/** Offset for the config register MMIO region. */
+#define HV_MPIPE_CONFIG_MMIO_OFFSET \
+  (MPIPE_MMIO_ADDR__REGION_VAL_CFG << MPIPE_MMIO_ADDR__REGION_SHIFT)
+
+/** Size of the config register MMIO region. */
+#define HV_MPIPE_CONFIG_MMIO_SIZE (64 * 1024)
+
+/** Offset for the config register MMIO region. */
+#define HV_MPIPE_FAST_MMIO_OFFSET \
+  (MPIPE_MMIO_ADDR__REGION_VAL_IDMA << MPIPE_MMIO_ADDR__REGION_SHIFT)
+
+/** Size of the fast register MMIO region (IDMA, EDMA, buffer stack). */
+#define HV_MPIPE_FAST_MMIO_SIZE \
+  ((MPIPE_MMIO_ADDR__REGION_VAL_BSM + 1 - MPIPE_MMIO_ADDR__REGION_VAL_IDMA) \
+   << MPIPE_MMIO_ADDR__REGION_SHIFT)
+
+
+/*
+ * Each type of resource allocation comes in quantized chunks, where
+ * XXX_BITS is the number of chunks, and XXX_RES_PER_BIT is the number
+ * of resources in each chunk.
+ */
+
+/** Number of buffer stack chunks available (32). */
+#define HV_MPIPE_ALLOC_BUFFER_STACKS_BITS \
+  MPIPE_MMIO_INIT_DAT_GX36_1__BUFFER_STACK_MASK_WIDTH
+
+/** Granularity of buffer stack allocation (1). */
+#define HV_MPIPE_ALLOC_BUFFER_STACKS_RES_PER_BIT \
+  (HV_MPIPE_NUM_BUFFER_STACKS / HV_MPIPE_ALLOC_BUFFER_STACKS_BITS)
+
+/** Number of NotifRing chunks available (32). */
+#define HV_MPIPE_ALLOC_NOTIF_RINGS_BITS \
+  MPIPE_MMIO_INIT_DAT_GX36_0__NOTIF_RING_MASK_WIDTH
+
+/** Granularity of NotifRing allocation (8). */
+#define HV_MPIPE_ALLOC_NOTIF_RINGS_RES_PER_BIT \
+  (HV_MPIPE_NUM_NOTIF_RINGS / HV_MPIPE_ALLOC_NOTIF_RINGS_BITS)
+
+/** Number of NotifGroup chunks available (32). */
+#define HV_MPIPE_ALLOC_NOTIF_GROUPS_BITS \
+  HV_MPIPE_NUM_NOTIF_GROUPS
+
+/** Granularity of NotifGroup allocation (1). */
+#define HV_MPIPE_ALLOC_NOTIF_GROUPS_RES_PER_BIT \
+  (HV_MPIPE_NUM_NOTIF_GROUPS / HV_MPIPE_ALLOC_NOTIF_GROUPS_BITS)
+
+/** Number of lo bucket chunks available (16). */
+#define HV_MPIPE_ALLOC_LO_BUCKETS_BITS \
+  MPIPE_MMIO_INIT_DAT_GX36_0__BUCKET_RELEASE_MASK_LO_WIDTH
+
+/** Granularity of lo bucket allocation (256). */
+#define HV_MPIPE_ALLOC_LO_BUCKETS_RES_PER_BIT \
+  (HV_MPIPE_NUM_LO_BUCKETS / HV_MPIPE_ALLOC_LO_BUCKETS_BITS)
+
+/** Number of hi bucket chunks available (16). */
+#define HV_MPIPE_ALLOC_HI_BUCKETS_BITS \
+  MPIPE_MMIO_INIT_DAT_GX36_0__BUCKET_RELEASE_MASK_HI_WIDTH
+
+/** Granularity of hi bucket allocation (4). */
+#define HV_MPIPE_ALLOC_HI_BUCKETS_RES_PER_BIT \
+  (HV_MPIPE_NUM_HI_BUCKETS / HV_MPIPE_ALLOC_HI_BUCKETS_BITS)
+
+/** Number of eDMA ring chunks available (24). */
+#define HV_MPIPE_ALLOC_EDMA_RINGS_BITS \
+  MPIPE_MMIO_INIT_DAT_GX36_1__EDMA_POST_MASK_WIDTH
+
+/** Granularity of eDMA ring allocation (1). */
+#define HV_MPIPE_ALLOC_EDMA_RINGS_RES_PER_BIT \
+  (HV_MPIPE_NUM_EDMA_RINGS / HV_MPIPE_ALLOC_EDMA_RINGS_BITS)
+
+
+
+
+/** Bit vector encoding which NotifRings are in a NotifGroup. */
+typedef struct
+{
+  /** The actual bits. */
+  uint64_t ring_mask[4];
+
+} gxio_mpipe_notif_group_bits_t;
+
+
+/** Another name for MPIPE_LBL_INIT_DAT_BSTS_TBL_t. */
+typedef MPIPE_LBL_INIT_DAT_BSTS_TBL_t gxio_mpipe_bucket_info_t;
+
+
+
+/** Eight buffer stack ids. */
+typedef struct
+{
+  /** The stacks. */
+  uint8_t stacks[8];
+
+} gxio_mpipe_rules_stacks_t;
+
+
+/** A destination mac address. */
+typedef struct
+{
+  /** The octets. */
+  uint8_t octets[6];
+
+} gxio_mpipe_rules_dmac_t;
+
+
+/** A vlan. */
+typedef uint16_t gxio_mpipe_rules_vlan_t;
+
+
+
+/** Maximum number of characters in a link name. */
+#define GXIO_MPIPE_LINK_NAME_LEN  32
+
+
+/** Structure holding a link name.  Only needed, and only typedef'ed,
+ *  because the IORPC stub generator only handles types which are single
+ *  words coming before the parameter name. */
+typedef struct
+{
+  /** The name itself. */
+  char name[GXIO_MPIPE_LINK_NAME_LEN];
+}
+_gxio_mpipe_link_name_t;
+
+/** Maximum number of characters in a symbol name. */
+#define GXIO_MPIPE_SYMBOL_NAME_LEN  128
+
+
+/** Structure holding a symbol name.  Only needed, and only typedef'ed,
+ *  because the IORPC stub generator only handles types which are single
+ *  words coming before the parameter name. */
+typedef struct
+{
+  /** The name itself. */
+  char name[GXIO_MPIPE_SYMBOL_NAME_LEN];
+}
+_gxio_mpipe_symbol_name_t;
+
+
+/** Structure holding a MAC address. */
+typedef struct
+{
+  /** The address. */
+  uint8_t mac[6];
+}
+_gxio_mpipe_link_mac_t;
+
+
+
+/** Request shared data permission -- that is, the ability to send and
+ *  receive packets -- on the specified link.  Other processes may also
+ *  request shared data permission on the same link.
+ *
+ *  No more than one of ::GXIO_MPIPE_LINK_DATA, ::GXIO_MPIPE_LINK_NO_DATA,
+ *  or ::GXIO_MPIPE_LINK_EXCL_DATA may be specifed in a gxio_mpipe_link_open()
+ *  call.  If none are specified, ::GXIO_MPIPE_LINK_DATA is assumed.
+ */
+#define GXIO_MPIPE_LINK_DATA               0x00000001UL
+
+/** Do not request data permission on the specified link.
+ *
+ *  No more than one of ::GXIO_MPIPE_LINK_DATA, ::GXIO_MPIPE_LINK_NO_DATA,
+ *  or ::GXIO_MPIPE_LINK_EXCL_DATA may be specifed in a gxio_mpipe_link_open()
+ *  call.  If none are specified, ::GXIO_MPIPE_LINK_DATA is assumed.
+ */
+#define GXIO_MPIPE_LINK_NO_DATA            0x00000002UL
+
+/** Request exclusive data permission -- that is, the ability to send and
+ *  receive packets -- on the specified link.  No other processes may
+ *  request data permission on this link, and if any process already has
+ *  data permission on it, this open will fail.
+ *
+ *  No more than one of ::GXIO_MPIPE_LINK_DATA, ::GXIO_MPIPE_LINK_NO_DATA,
+ *  or ::GXIO_MPIPE_LINK_EXCL_DATA may be specifed in a gxio_mpipe_link_open()
+ *  call.  If none are specified, ::GXIO_MPIPE_LINK_DATA is assumed.
+ */
+#define GXIO_MPIPE_LINK_EXCL_DATA          0x00000004UL
+
+/** Request shared stats permission -- that is, the ability to read and write
+ *  registers which contain link statistics, and to get link attributes --
+ *  on the specified link.  Other processes may also request shared stats
+ *  permission on the same link.
+ *
+ *  No more than one of ::GXIO_MPIPE_LINK_STATS, ::GXIO_MPIPE_LINK_NO_STATS,
+ *  or ::GXIO_MPIPE_LINK_EXCL_STATS may be specifed in a gxio_mpipe_link_open()
+ *  call.  If none are specified, ::GXIO_MPIPE_LINK_STATS is assumed.
+ */
+#define GXIO_MPIPE_LINK_STATS              0x00000008UL
+
+/** Do not request stats permission on the specified link.
+ *
+ *  No more than one of ::GXIO_MPIPE_LINK_STATS, ::GXIO_MPIPE_LINK_NO_STATS,
+ *  or ::GXIO_MPIPE_LINK_EXCL_STATS may be specifed in a gxio_mpipe_link_open()
+ *  call.  If none are specified, ::GXIO_MPIPE_LINK_STATS is assumed.
+ */
+#define GXIO_MPIPE_LINK_NO_STATS           0x00000010UL
+
+/** Request exclusive stats permission -- that is, the ability to read and
+ *  write registers which contain link statistics, and to get link
+ *  attributes -- on the specified link.  No other processes may request
+ *  stats permission on this link, and if any process already
+ *  has stats permission on it, this open will fail.
+ *
+ *  Requesting exclusive stats permission is normally a very bad idea, since
+ *  it prevents programs like mpipe-stat from providing information on this
+ *  link.  Applications should only do this if they use MAC statistics
+ *  registers, and cannot tolerate any of the clear-on-read registers being
+ *  reset by other statistics programs.
+ *
+ *  No more than one of ::GXIO_MPIPE_LINK_STATS, ::GXIO_MPIPE_LINK_NO_STATS,
+ *  or ::GXIO_MPIPE_LINK_EXCL_STATS may be specifed in a gxio_mpipe_link_open()
+ *  call.  If none are specified, ::GXIO_MPIPE_LINK_STATS is assumed.
+ */
+#define GXIO_MPIPE_LINK_EXCL_STATS         0x00000020UL
+
+/** Request shared control permission -- that is, the ability to modify link
+ *  attributes, and read and write MAC and MDIO registers -- on the
+ *  specified link.  Other processes may also request shared control
+ *  permission on the same link.
+ *
+ *  No more than one of ::GXIO_MPIPE_LINK_CTL, ::GXIO_MPIPE_LINK_NO_CTL,
+ *  or ::GXIO_MPIPE_LINK_EXCL_CTL may be specifed in a gxio_mpipe_link_open()
+ *  call.  If none are specified, ::GXIO_MPIPE_LINK_CTL is assumed.
+ */
+#define GXIO_MPIPE_LINK_CTL                0x00000040UL
+
+/** Do not request control permission on the specified link.
+ *
+ *  No more than one of ::GXIO_MPIPE_LINK_CTL, ::GXIO_MPIPE_LINK_NO_CTL,
+ *  or ::GXIO_MPIPE_LINK_EXCL_CTL may be specifed in a gxio_mpipe_link_open()
+ *  call.  If none are specified, ::GXIO_MPIPE_LINK_CTL is assumed.
+ */
+#define GXIO_MPIPE_LINK_NO_CTL             0x00000080UL
+
+/** Request exclusive control permission -- that is, the ability to modify
+ *  link attributes, and read and write MAC and MDIO registers -- on the
+ *  specified link.  No other processes may request control permission on
+ *  this link, and if any process already has control permission on it,
+ *  this open will fail.
+ *
+ *  Requesting exclusive control permission is not always a good idea, since
+ *  it prevents programs like mpipe-link from configuring the link.
+ *
+ *  No more than one of ::GXIO_MPIPE_LINK_CTL, ::GXIO_MPIPE_LINK_NO_CTL,
+ *  or ::GXIO_MPIPE_LINK_EXCL_CTL may be specifed in a gxio_mpipe_link_open()
+ *  call.  If none are specified, ::GXIO_MPIPE_LINK_CTL is assumed.
+ */
+#define GXIO_MPIPE_LINK_EXCL_CTL           0x00000100UL
+
+/** Set the desired state of the link to up, allowing any speeds which are
+ *  supported by the link hardware, as part of this open operation; do not
+ *  change the desired state of the link when it is closed or the process
+ *  exits.  No more than one of ::GXIO_MPIPE_LINK_AUTO_UP,
+ *  ::GXIO_MPIPE_LINK_AUTO_UPDOWN, ::GXIO_MPIPE_LINK_AUTO_DOWN, or
+ *  ::GXIO_MPIPE_LINK_AUTO_NONE may be specifed in a gxio_mpipe_link_open()
+ *  call.  If none are specified, ::GXIO_MPIPE_LINK_AUTO_UPDOWN is assumed.
+ */
+#define GXIO_MPIPE_LINK_AUTO_UP            0x00000200UL
+
+/** Set the desired state of the link to up, allowing any speeds which are
+ *  supported by the link hardware, as part of this open operation; when the
+ *  link is closed or this process exits, if no other process has the link
+ *  open, set the desired state of the link to down.  No more than one of
+ *  ::GXIO_MPIPE_LINK_AUTO_UP, ::GXIO_MPIPE_LINK_AUTO_UPDOWN,
+ *  ::GXIO_MPIPE_LINK_AUTO_DOWN, or ::GXIO_MPIPE_LINK_AUTO_NONE may be
+ *  specifed in a gxio_mpipe_link_open() call.  If none are specified,
+ *  ::GXIO_MPIPE_LINK_AUTO_UPDOWN is assumed.
+ */
+#define GXIO_MPIPE_LINK_AUTO_UPDOWN        0x00000400UL
+
+/** Do not change the desired state of the link as part of the open
+ *  operation; when the link is closed or this process exits, if no other
+ *  process has the link open, set the desired state of the link to down.
+ *  No more than one of ::GXIO_MPIPE_LINK_AUTO_UP,
+ *  ::GXIO_MPIPE_LINK_AUTO_UPDOWN, ::GXIO_MPIPE_LINK_AUTO_DOWN, or
+ *  ::GXIO_MPIPE_LINK_AUTO_NONE may be specifed in a gxio_mpipe_link_open()
+ *  call.  If none are specified, ::GXIO_MPIPE_LINK_AUTO_UPDOWN is assumed.
+ */
+#define GXIO_MPIPE_LINK_AUTO_DOWN          0x00000800UL
+
+/** Do not change the desired state of the link as part of the open
+ *  operation; do not change the desired state of the link when it is
+ *  closed or the process exits.  No more than one of
+ *  ::GXIO_MPIPE_LINK_AUTO_UP, ::GXIO_MPIPE_LINK_AUTO_UPDOWN,
+ *  ::GXIO_MPIPE_LINK_AUTO_DOWN, or ::GXIO_MPIPE_LINK_AUTO_NONE may be
+ *  specifed in a gxio_mpipe_link_open() call.  If none are specified,
+ *  ::GXIO_MPIPE_LINK_AUTO_UPDOWN is assumed.
+ */
+#define GXIO_MPIPE_LINK_AUTO_NONE          0x00001000UL
+
+/** Request that this open call not complete until the network link is up.
+ *  The process will wait as long as necessary for this to happen;
+ *  applications which wish to abandon waiting for the link after a
+ *  specific time period should not specify this flag when opening a link,
+ *  but should instead call gxio_mpipe_link_wait() afterward.  The link
+ *  must be opened with stats permission.  Note that this flag by itself
+ *  does not change the desired link state; if other open flags or previous
+ *  link state changes have not requested a desired state of up, the open
+ *  call will never complete.  This flag is not available to kernel
+ *  clients.
+ */
+#define GXIO_MPIPE_LINK_WAIT               0x00002000UL
+
+
+/*
+ * Note: link attributes must fit in 24 bits, since we use the top 8 bits
+ * of the IORPC offset word for the channel number.
+ */
+
+/** Determine whether jumbo frames may be received.  If this attribute's
+ *  value value is nonzero, the MAC will accept frames of up to 10240 bytes.
+ *  If the value is zero, the MAC will only accept frames of up to 1544
+ *  bytes.  The default value is zero. */
+#define GXIO_MPIPE_LINK_RECEIVE_JUMBO      0x010000
+
+/** Determine whether to send pause frames on this link if the mPIPE packet
+ *  FIFO is nearly full.  If the value is zero, pause frames are not sent.
+ *  If the value is nonzero, it is the delay value which will be sent in any
+ *  pause frames which are output, in units of 512 bit times.
+ *
+ *  Bear in mind that in almost all circumstances, the mPIPE packet FIFO
+ *  will never fill up, since mPIPE will empty it as fast as or faster than
+ *  the incoming data rate, by either delivering or dropping packets.  The
+ *  only situation in which this is not true is if the memory and cache
+ *  subsystem is extremely heavily loaded, and mPIPE cannot perform DMA of
+ *  packet data to memory in a timely fashion.  In particular, pause frames
+ *  will <em>not</em> be sent if packets cannot be delivered because
+ *  NotifRings are full, buckets are full, or buffers are not available in
+ *  a buffer stack. */
+#define GXIO_MPIPE_LINK_SEND_PAUSE         0x020000
+
+/** Determine whether to suspend output on the receipt of pause frames.
+ *  If the value is nonzero, mPIPE shim will suspend output on the link's
+ *  channel when a pause frame is received.  If the value is zero, pause
+ *  frames will be ignored.  The default value is zero. */
+#define GXIO_MPIPE_LINK_RECEIVE_PAUSE      0x030000
+
+/** Interface MAC address.  The value is a 6-byte MAC address, in the least
+ *  significant 48 bits of the value; in other words, an address which would
+ *  be printed as '12:34:56:78:90:AB' in IEEE 802 canonical format would
+ *  be returned as 0x12345678ab.
+ *
+ *  Depending upon the overall system design, a MAC address may or may not
+ *  be available for each interface.  Note that the interface's MAC address
+ *  does not limit the packets received on its channel, although the
+ *  classifier's rules could be configured to do that.  Similarly, the MAC
+ *  address is not used when transmitting packets, although applications
+ *  could certainly decide to use the assigned address as a source MAC
+ *  address when doing so.  This attribute may only be retrieved with
+ *  gxio_mpipe_link_get_attr(); it may not be modified.
+ */
+#define GXIO_MPIPE_LINK_MAC                0x040000
+
+/** Determine whether to discard egress packets on link down. If this value
+ *  is nonzero, packets sent on this link while the link is down will be
+ *  discarded.  If this value is zero, no packets will be sent on this link
+ *  while it is down.  The default value is one. */
+#define GXIO_MPIPE_LINK_DISCARD_IF_DOWN    0x050000
+
+/** Possible link state.  The value is a combination of link state flags,
+ *  ORed together, that indicate link modes which are actually supported by
+ *  the hardware.  This attribute may only be retrieved with
+ *  gxio_mpipe_link_get_attr(); it may not be modified. */
+#define GXIO_MPIPE_LINK_POSSIBLE_STATE     0x060000
+
+/** Current link state.  The value is a combination of link state flags,
+ *  ORed together, that indicate the current state of the hardware.  If the
+ *  link is down, the value ANDed with ::GXIO_MPIPE_LINK_SPEED will be zero;
+ *  if the link is up, the value ANDed with ::GXIO_MPIPE_LINK_SPEED will
+ *  result in exactly one of the speed values, indicating the current speed.
+ *  This attribute may only be retrieved with gxio_mpipe_link_get_attr(); it
+ *  may not be modified. */
+#define GXIO_MPIPE_LINK_CURRENT_STATE      0x070000
+
+/** Desired link state. The value is a conbination of flags, which specify
+ *  the desired state for the link.  With gxio_mpipe_link_set_attr(), this
+ *  will, in the background, attempt to bring up the link using whichever of
+ *  the requested flags are reasonable, or take down the link if the flags
+ *  are zero.  The actual link up or down operation may happen after this
+ *  call completes.  If the link state changes in the future, the system
+ *  will continue to try to get back to the desired link state; for
+ *  instance, if the link is brought up successfully, and then the network
+ *  cable is disconnected, the link will go down.  However, the desired
+ *  state of the link is still up, so if the cable is reconnected, the link
+ *  will be brought up again.
+ *
+ *  With gxio_mpipe_link_set_attr(), this will indicate the desired state
+ *  for the link, as set with a previous gxio_mpipe_link_set_attr() call,
+ *  or implicitly by a gxio_mpipe_link_open() or link close operation.
+ *  This may not reflect the current state of the link; to get that, use
+ *  ::GXIO_MPIPE_LINK_CURRENT_STATE.
+ */
+#define GXIO_MPIPE_LINK_DESIRED_STATE      0x080000
+
+
+
+/** Link can run, should run, or is running at 10 Mbps. */
+#define GXIO_MPIPE_LINK_10M        0x0000000000000001UL
+
+/** Link can run, should run, or is running at 100 Mbps. */
+#define GXIO_MPIPE_LINK_100M       0x0000000000000002UL
+
+/** Link can run, should run, or is running at 1 Gbps. */
+#define GXIO_MPIPE_LINK_1G         0x0000000000000004UL
+
+/** Link can run, should run, or is running at 10 Gbps. */
+#define GXIO_MPIPE_LINK_10G        0x0000000000000008UL
+
+/** Link can run, should run, or is running at 20 Gbps. */
+#define GXIO_MPIPE_LINK_20G        0x0000000000000010UL
+
+/** Link can run, should run, or is running at 25 Gbps. */
+#define GXIO_MPIPE_LINK_25G        0x0000000000000020UL
+
+/** Link can run, should run, or is running at 50 Gbps. */
+#define GXIO_MPIPE_LINK_50G        0x0000000000000040UL
+
+/** Link should run at the highest speed supported by the link and by
+ *  the device connected to the link.  Only usable as a value for
+ *  the link's desired state; never returned as a value for the current
+ *  or possible states. */
+#define GXIO_MPIPE_LINK_ANYSPEED   0x0000000000000800UL
+
+/** All legal link speeds.  This value is provided for use in extracting
+ *  the speed-related subset of the link state flags; it is not intended
+ *  to be set directly as a value for one of the GXIO_MPIPE_LINK_xxx_STATE
+ *  attributes.  A link is up or is requested to be up if its current or
+ *  desired state, respectively, ANDED with this value, is nonzero. */
+#define GXIO_MPIPE_LINK_SPEED_MASK 0x0000000000000FFFUL
+
+/** Link can run, should run, or is running in MAC loopback mode.  This
+ *  loops transmitted packets back to the receiver, inside the Tile
+ *  Processor. */
+#define GXIO_MPIPE_LINK_LOOP_MAC   0x0000000000001000UL
+
+/** Link can run, should run, or is running in PHY loopback mode.  This
+ *  loops transmitted packets back to the receiver, inside the external
+ *  PHY chip. */
+#define GXIO_MPIPE_LINK_LOOP_PHY   0x0000000000002000UL
+
+/** Link can run, should run, or is running in external loopback mode.
+ *  This requires that an external loopback plug be installed on the
+ *  Ethernet port.  Note that only some links require that this be
+ *  configured via the gxio_mpipe_link routines; other links can do
+ *  external loopack with the plug and no special configuration. */
+#define GXIO_MPIPE_LINK_LOOP_EXT   0x0000000000004000UL
+
+/** All legal loopback types. */
+#define GXIO_MPIPE_LINK_LOOP_MASK  0x000000000000F000UL
+
+/** Link can run, should run, or is running in full-duplex mode.
+ *  If neither ::GXIO_MPIPE_LINK_FDX nor ::GXIO_MPIPE_LINK_HDX are
+ *  specified in a set of desired state flags, both are assumed. */
+#define GXIO_MPIPE_LINK_FDX        0x0000000000010000UL
+
+/** Link can run, should run, or is running in half-duplex mode.
+ *  If neither ::GXIO_MPIPE_LINK_FDX nor ::GXIO_MPIPE_LINK_HDX are
+ *  specified in a set of desired state flags, both are assumed. */
+#define GXIO_MPIPE_LINK_HDX        0x0000000000020000UL
+
+
+/** An individual rule. */
+typedef struct
+{
+  /** The total size. */
+  uint16_t size;
+
+  /** The priority. */
+  int16_t priority;
+
+  /** The "headroom" in each buffer. */
+  uint8_t headroom;
+
+  /** The "tailroom" in each buffer. */
+  uint8_t tailroom;
+
+  /** The "capacity" of the largest buffer. */
+  uint16_t capacity;
+
+  /** The mask for converting a flow hash into a bucket. */
+  uint16_t bucket_mask;
+
+  /** The offset for converting a flow hash into a bucket. */
+  uint16_t bucket_first;
+
+  /** The buffer stack ids. */
+  gxio_mpipe_rules_stacks_t stacks;
+
+  /** The actual channels. */
+  uint32_t channel_bits;
+
+  /** The number of dmacs. */
+  uint16_t num_dmacs;
+
+  /** The number of vlans. */
+  uint16_t num_vlans;
+
+  /** The actual dmacs and vlans. */
+  uint8_t dmacs_and_vlans[];
+
+} gxio_mpipe_rules_rule_t;
+
+
+/** A list of classifier rules. */
+typedef struct
+{
+  /** The offset to the end of the current rule. */
+  uint16_t tail;
+
+  /** The offset to the start of the current rule. */
+  uint16_t head;
+
+  /** The actual rules. */
+  uint8_t rules[4096 - 4];
+
+} gxio_mpipe_rules_list_t;
+
+
+
+
+/** mPIPE statistics structure. These counters include all relevant
+ *  events occurring on all links within the mPIPE shim. */
+typedef struct
+{
+  /** Number of ingress packets dropped for any reason. */
+  uint64_t ingress_drops;
+  /** Number of ingress packets dropped because a buffer stack was empty. */
+  uint64_t ingress_drops_no_buf;
+  /** Number of ingress packets dropped or truncated due to lack of space in
+   *  the iPkt buffer. */
+  uint64_t ingress_drops_ipkt;
+  /** Number of ingress packets dropped by the classifier or load balancer */
+  uint64_t ingress_drops_cls_lb;
+  /** Total number of ingress packets. */
+  uint64_t ingress_packets;
+  /** Total number of egress packets. */
+  uint64_t egress_packets;
+  /** Total number of ingress bytes. */
+  uint64_t ingress_bytes;
+  /** Total number of egress bytes. */
+  uint64_t egress_bytes;
+}
+gxio_mpipe_stats_t;
+
+
+#endif /* _SYS_HV_DRV_MPIPE_INTF_H */
diff --git a/kernel/arch/tile/include/hv/drv_mshim_intf.h b/kernel/arch/tile/include/hv/drv_mshim_intf.h
new file mode 100644
index 000000000..c6ef3bdc5
--- /dev/null
+++ b/kernel/arch/tile/include/hv/drv_mshim_intf.h
@@ -0,0 +1,50 @@
+/*
+ * Copyright 2011 Tilera Corporation. All Rights Reserved.
+ *
+ *   This program is free software; you can redistribute it and/or
+ *   modify it under the terms of the GNU General Public License
+ *   as published by the Free Software Foundation, version 2.
+ *
+ *   This program is distributed in the hope that it will be useful, but
+ *   WITHOUT ANY WARRANTY; without even the implied warranty of
+ *   MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
+ *   NON INFRINGEMENT.  See the GNU General Public License for
+ *   more details.
+ */
+
+/**
+ * @file drv_mshim_intf.h
+ * Interface definitions for the Linux EDAC memory controller driver.
+ */
+
+#ifndef _SYS_HV_INCLUDE_DRV_MSHIM_INTF_H
+#define _SYS_HV_INCLUDE_DRV_MSHIM_INTF_H
+
+/** Number of memory controllers in the public API. */
+#define TILE_MAX_MSHIMS 4
+
+/** Memory info under each memory controller. */
+struct mshim_mem_info
+{
+  uint64_t mem_size;     /**< Total memory size in bytes. */
+  uint8_t mem_type;      /**< Memory type, DDR2 or DDR3. */
+  uint8_t mem_ecc;       /**< Memory supports ECC. */
+};
+
+/**
+ * DIMM error structure.
+ * For now, only correctable errors are counted and the mshim doesn't record
+ * the error PA. HV takes panic upon uncorrectable errors.
+ */
+struct mshim_mem_error
+{
+  uint32_t sbe_count;     /**< Number of single-bit errors. */
+};
+
+/** Read this offset to get the memory info per mshim. */
+#define MSHIM_MEM_INFO_OFF 0x100
+
+/** Read this offset to check DIMM error. */
+#define MSHIM_MEM_ERROR_OFF 0x200
+
+#endif /* _SYS_HV_INCLUDE_DRV_MSHIM_INTF_H */
diff --git a/kernel/arch/tile/include/hv/drv_pcie_rc_intf.h b/kernel/arch/tile/include/hv/drv_pcie_rc_intf.h
new file mode 100644
index 000000000..9bd2243be
--- /dev/null
+++ b/kernel/arch/tile/include/hv/drv_pcie_rc_intf.h
@@ -0,0 +1,38 @@
+/*
+ * Copyright 2010 Tilera Corporation. All Rights Reserved.
+ *
+ *   This program is free software; you can redistribute it and/or
+ *   modify it under the terms of the GNU General Public License
+ *   as published by the Free Software Foundation, version 2.
+ *
+ *   This program is distributed in the hope that it will be useful, but
+ *   WITHOUT ANY WARRANTY; without even the implied warranty of
+ *   MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
+ *   NON INFRINGEMENT.  See the GNU General Public License for
+ *   more details.
+ */
+
+/**
+ * @file drv_pcie_rc_intf.h
+ * Interface definitions for the PCIE Root Complex.
+ */
+
+#ifndef _SYS_HV_DRV_PCIE_RC_INTF_H
+#define _SYS_HV_DRV_PCIE_RC_INTF_H
+
+/** File offset for reading the interrupt base number used for PCIE legacy
+    interrupts and PLX Gen 1 requirement flag */
+#define PCIE_RC_CONFIG_MASK_OFF 0
+
+
+/**
+ * Structure used for obtaining PCIe config information, read from the PCIE
+ * subsystem /ctl file at initialization
+ */
+typedef struct pcie_rc_config
+{
+  int intr;                     /**< interrupt number used for downcall */
+  int plx_gen1;                 /**< flag for PLX Gen 1 configuration */
+} pcie_rc_config_t;
+
+#endif  /* _SYS_HV_DRV_PCIE_RC_INTF_H */
diff --git a/kernel/arch/tile/include/hv/drv_srom_intf.h b/kernel/arch/tile/include/hv/drv_srom_intf.h
new file mode 100644
index 000000000..6395faa6d
--- /dev/null
+++ b/kernel/arch/tile/include/hv/drv_srom_intf.h
@@ -0,0 +1,41 @@
+/*
+ * Copyright 2011 Tilera Corporation. All Rights Reserved.
+ *
+ *   This program is free software; you can redistribute it and/or
+ *   modify it under the terms of the GNU General Public License
+ *   as published by the Free Software Foundation, version 2.
+ *
+ *   This program is distributed in the hope that it will be useful, but
+ *   WITHOUT ANY WARRANTY; without even the implied warranty of
+ *   MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
+ *   NON INFRINGEMENT.  See the GNU General Public License for
+ *   more details.
+ */
+
+/**
+ * @file drv_srom_intf.h
+ * Interface definitions for the SPI Flash ROM driver.
+ */
+
+#ifndef _SYS_HV_INCLUDE_DRV_SROM_INTF_H
+#define _SYS_HV_INCLUDE_DRV_SROM_INTF_H
+
+/** Read this offset to get the total device size. */
+#define SROM_TOTAL_SIZE_OFF   0xF0000000
+
+/** Read this offset to get the device sector size. */
+#define SROM_SECTOR_SIZE_OFF  0xF0000004
+
+/** Read this offset to get the device page size. */
+#define SROM_PAGE_SIZE_OFF    0xF0000008
+
+/** Write this offset to flush any pending writes. */
+#define SROM_FLUSH_OFF        0xF1000000
+
+/** Write this offset, plus the byte offset of the start of a sector, to
+ *  erase a sector.  Any write data is ignored, but there must be at least
+ *  one byte of write data.  Only applies when the driver is in MTD mode.
+ */
+#define SROM_ERASE_OFF        0xF2000000
+
+#endif /* _SYS_HV_INCLUDE_DRV_SROM_INTF_H */
diff --git a/kernel/arch/tile/include/hv/drv_trio_intf.h b/kernel/arch/tile/include/hv/drv_trio_intf.h
new file mode 100644
index 000000000..237e04dee
--- /dev/null
+++ b/kernel/arch/tile/include/hv/drv_trio_intf.h
@@ -0,0 +1,199 @@
+/*
+ * Copyright 2012 Tilera Corporation. All Rights Reserved.
+ *
+ *   This program is free software; you can redistribute it and/or
+ *   modify it under the terms of the GNU General Public License
+ *   as published by the Free Software Foundation, version 2.
+ *
+ *   This program is distributed in the hope that it will be useful, but
+ *   WITHOUT ANY WARRANTY; without even the implied warranty of
+ *   MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
+ *   NON INFRINGEMENT.  See the GNU General Public License for
+ *   more details.
+ */
+
+/**
+ * Interface definitions for the trio driver.
+ */
+
+#ifndef _SYS_HV_DRV_TRIO_INTF_H
+#define _SYS_HV_DRV_TRIO_INTF_H
+
+#include <arch/trio.h>
+
+/** The vendor ID for all Tilera processors. */
+#define TILERA_VENDOR_ID 0x1a41
+
+/** The device ID for the Gx36 processor. */
+#define TILERA_GX36_DEV_ID 0x0200
+
+/** Device ID for our internal bridge when running as RC. */
+#define TILERA_GX36_RC_DEV_ID 0x2000
+
+/** Maximum number of TRIO interfaces. */
+#define TILEGX_NUM_TRIO         2
+
+/** Gx36 has max 3 PCIe MACs per TRIO interface. */
+#define TILEGX_TRIO_PCIES       3
+
+/** Specify port properties for a PCIe MAC. */
+struct pcie_port_property
+{
+  /** If true, the link can be configured in PCIe root complex mode. */
+  uint8_t allow_rc: 1;
+
+  /** If true, the link can be configured in PCIe endpoint mode. */
+  uint8_t allow_ep: 1;
+
+  /** If true, the link can be configured in StreamIO mode. */
+  uint8_t allow_sio: 1;
+
+  /** If true, the link is allowed to support 1-lane operation. Software
+   *  will not consider it an error if the link comes up as a x1 link. */
+  uint8_t allow_x1: 1;
+
+  /** If true, the link is allowed to support 2-lane operation. Software
+   *  will not consider it an error if the link comes up as a x2 link. */
+  uint8_t allow_x2: 1;
+
+  /** If true, the link is allowed to support 4-lane operation. Software
+   *  will not consider it an error if the link comes up as a x4 link. */
+  uint8_t allow_x4: 1;
+
+  /** If true, the link is allowed to support 8-lane operation. Software
+   *  will not consider it an error if the link comes up as a x8 link. */
+  uint8_t allow_x8: 1;
+
+  /** If true, this link is connected to a device which may or may not
+   *  be present. */
+  uint8_t removable: 1;
+
+};
+
+/** Configurations can be issued to configure a char stream interrupt. */
+typedef enum pcie_stream_intr_config_sel_e
+{
+  /** Interrupt configuration for memory map regions. */
+  MEM_MAP_SEL,
+
+  /** Interrupt configuration for push DMAs. */
+  PUSH_DMA_SEL,
+
+  /** Interrupt configuration for pull DMAs. */
+  PULL_DMA_SEL,
+}
+pcie_stream_intr_config_sel_t;
+
+
+/** The mmap file offset (PA) of the TRIO config region. */
+#define HV_TRIO_CONFIG_OFFSET                                        \
+  ((unsigned long long)TRIO_MMIO_ADDRESS_SPACE__REGION_VAL_CFG <<   \
+    TRIO_MMIO_ADDRESS_SPACE__REGION_SHIFT)
+
+/** The maximum size of the TRIO config region. */
+#define HV_TRIO_CONFIG_SIZE                                 \
+  (1ULL << TRIO_CFG_REGION_ADDR__REGION_SHIFT)
+
+/** Size of the config region mapped into client. We can't use
+ *  TRIO_MMIO_ADDRESS_SPACE__OFFSET_WIDTH because it
+ *  will require the kernel to allocate 4GB VA space
+ *  from the VMALLOC region which has a total range
+ *  of 4GB.
+ */
+#define HV_TRIO_CONFIG_IOREMAP_SIZE                            \
+  ((uint64_t) 1 << TRIO_CFG_REGION_ADDR__PROT_SHIFT)
+
+/** The mmap file offset (PA) of a scatter queue region. */
+#define HV_TRIO_SQ_OFFSET(queue)                                        \
+  (((unsigned long long)TRIO_MMIO_ADDRESS_SPACE__REGION_VAL_MAP_SQ <<   \
+    TRIO_MMIO_ADDRESS_SPACE__REGION_SHIFT) |                            \
+   ((queue) << TRIO_MAP_SQ_REGION_ADDR__SQ_SEL_SHIFT))
+
+/** The maximum size of a scatter queue region. */
+#define HV_TRIO_SQ_SIZE                                 \
+  (1ULL << TRIO_MAP_SQ_REGION_ADDR__SQ_SEL_SHIFT)
+
+
+/** The "hardware MMIO region" of the first PIO region. */
+#define HV_TRIO_FIRST_PIO_REGION 8
+
+/** The mmap file offset (PA) of a PIO region. */
+#define HV_TRIO_PIO_OFFSET(region)                           \
+  (((unsigned long long)(region) + HV_TRIO_FIRST_PIO_REGION) \
+   << TRIO_PIO_REGIONS_ADDR__REGION_SHIFT)
+
+/** The maximum size of a PIO region. */
+#define HV_TRIO_PIO_SIZE (1ULL << TRIO_PIO_REGIONS_ADDR__ADDR_WIDTH)
+
+
+/** The mmap file offset (PA) of a push DMA region. */
+#define HV_TRIO_PUSH_DMA_OFFSET(ring)                                   \
+  (((unsigned long long)TRIO_MMIO_ADDRESS_SPACE__REGION_VAL_PUSH_DMA << \
+    TRIO_MMIO_ADDRESS_SPACE__REGION_SHIFT) |                            \
+   ((ring) << TRIO_PUSH_DMA_REGION_ADDR__RING_SEL_SHIFT))
+
+/** The mmap file offset (PA) of a pull DMA region. */
+#define HV_TRIO_PULL_DMA_OFFSET(ring)                                   \
+  (((unsigned long long)TRIO_MMIO_ADDRESS_SPACE__REGION_VAL_PULL_DMA << \
+    TRIO_MMIO_ADDRESS_SPACE__REGION_SHIFT) |                            \
+   ((ring) << TRIO_PULL_DMA_REGION_ADDR__RING_SEL_SHIFT))
+
+/** The maximum size of a DMA region. */
+#define HV_TRIO_DMA_REGION_SIZE                         \
+  (1ULL << TRIO_PUSH_DMA_REGION_ADDR__RING_SEL_SHIFT)
+
+
+/** The mmap file offset (PA) of a Mem-Map interrupt region. */
+#define HV_TRIO_MEM_MAP_INTR_OFFSET(map)                                 \
+  (((unsigned long long)TRIO_MMIO_ADDRESS_SPACE__REGION_VAL_MAP_MEM <<   \
+    TRIO_MMIO_ADDRESS_SPACE__REGION_SHIFT) |                            \
+   ((map) << TRIO_MAP_MEM_REGION_ADDR__MAP_SEL_SHIFT))
+
+/** The maximum size of a Mem-Map interrupt region. */
+#define HV_TRIO_MEM_MAP_INTR_SIZE                                 \
+  (1ULL << TRIO_MAP_MEM_REGION_ADDR__MAP_SEL_SHIFT)
+
+
+/** A flag bit indicating a fixed resource allocation. */
+#define HV_TRIO_ALLOC_FIXED 0x01
+
+/** TRIO requires that all mappings have 4kB aligned start addresses. */
+#define HV_TRIO_PAGE_SHIFT 12
+
+/** TRIO requires that all mappings have 4kB aligned start addresses. */
+#define HV_TRIO_PAGE_SIZE (1ull << HV_TRIO_PAGE_SHIFT)
+
+
+/* Specify all PCIe port properties for a TRIO. */
+struct pcie_trio_ports_property
+{
+  struct pcie_port_property ports[TILEGX_TRIO_PCIES];
+
+  /** Set if this TRIO belongs to a Gx72 device. */
+  uint8_t is_gx72;
+};
+
+/* Flags indicating traffic class. */
+#define HV_TRIO_FLAG_TC_SHIFT 4
+#define HV_TRIO_FLAG_TC_RMASK 0xf
+#define HV_TRIO_FLAG_TC(N) \
+  ((((N) & HV_TRIO_FLAG_TC_RMASK) + 1) << HV_TRIO_FLAG_TC_SHIFT)
+
+/* Flags indicating virtual functions. */
+#define HV_TRIO_FLAG_VFUNC_SHIFT 8
+#define HV_TRIO_FLAG_VFUNC_RMASK 0xff
+#define HV_TRIO_FLAG_VFUNC(N) \
+  ((((N) & HV_TRIO_FLAG_VFUNC_RMASK) + 1) << HV_TRIO_FLAG_VFUNC_SHIFT)
+
+
+/* Flag indicating an ordered PIO region. */
+#define HV_TRIO_PIO_FLAG_ORDERED (1 << 16)
+
+/* Flags indicating special types of PIO regions. */
+#define HV_TRIO_PIO_FLAG_SPACE_SHIFT 17
+#define HV_TRIO_PIO_FLAG_SPACE_MASK (0x3 << HV_TRIO_PIO_FLAG_SPACE_SHIFT)
+#define HV_TRIO_PIO_FLAG_CONFIG_SPACE (0x1 << HV_TRIO_PIO_FLAG_SPACE_SHIFT)
+#define HV_TRIO_PIO_FLAG_IO_SPACE (0x2 << HV_TRIO_PIO_FLAG_SPACE_SHIFT)
+
+
+#endif /* _SYS_HV_DRV_TRIO_INTF_H */
diff --git a/kernel/arch/tile/include/hv/drv_uart_intf.h b/kernel/arch/tile/include/hv/drv_uart_intf.h
new file mode 100644
index 000000000..f5379e240
--- /dev/null
+++ b/kernel/arch/tile/include/hv/drv_uart_intf.h
@@ -0,0 +1,33 @@
+/*
+ * Copyright 2013 Tilera Corporation. All Rights Reserved.
+ *
+ *   This program is free software; you can redistribute it and/or
+ *   modify it under the terms of the GNU General Public License
+ *   as published by the Free Software Foundation, version 2.
+ *
+ *   This program is distributed in the hope that it will be useful, but
+ *   WITHOUT ANY WARRANTY; without even the implied warranty of
+ *   MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
+ *   NON INFRINGEMENT.  See the GNU General Public License for
+ *   more details.
+ */
+
+/**
+ * Interface definitions for the UART driver.
+ */
+
+#ifndef _SYS_HV_DRV_UART_INTF_H
+#define _SYS_HV_DRV_UART_INTF_H
+
+#include <arch/uart.h>
+
+/** Number of UART ports supported. */
+#define TILEGX_UART_NR        2
+
+/** The mmap file offset (PA) of the UART MMIO region. */
+#define HV_UART_MMIO_OFFSET   0
+
+/** The maximum size of the UARTs MMIO region (64K Bytes). */
+#define HV_UART_MMIO_SIZE     (1UL << 16)
+
+#endif /* _SYS_HV_DRV_UART_INTF_H */
diff --git a/kernel/arch/tile/include/hv/drv_usb_host_intf.h b/kernel/arch/tile/include/hv/drv_usb_host_intf.h
new file mode 100644
index 000000000..24ce774a3
--- /dev/null
+++ b/kernel/arch/tile/include/hv/drv_usb_host_intf.h
@@ -0,0 +1,39 @@
+/*
+ * Copyright 2012 Tilera Corporation. All Rights Reserved.
+ *
+ *   This program is free software; you can redistribute it and/or
+ *   modify it under the terms of the GNU General Public License
+ *   as published by the Free Software Foundation, version 2.
+ *
+ *   This program is distributed in the hope that it will be useful, but
+ *   WITHOUT ANY WARRANTY; without even the implied warranty of
+ *   MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
+ *   NON INFRINGEMENT.  See the GNU General Public License for
+ *   more details.
+ */
+
+/**
+ * Interface definitions for the USB host driver.
+ */
+
+#ifndef _SYS_HV_DRV_USB_HOST_INTF_H
+#define _SYS_HV_DRV_USB_HOST_INTF_H
+
+#include <arch/usb_host.h>
+
+
+/** Offset for the EHCI register MMIO region. */
+#define HV_USB_HOST_MMIO_OFFSET_EHCI ((uint64_t) USB_HOST_HCCAPBASE_REG)
+
+/** Offset for the OHCI register MMIO region. */
+#define HV_USB_HOST_MMIO_OFFSET_OHCI ((uint64_t) USB_HOST_OHCD_HC_REVISION_REG)
+
+/** Size of the register MMIO region.  This turns out to be the same for
+ *  both EHCI and OHCI. */
+#define HV_USB_HOST_MMIO_SIZE ((uint64_t) 0x1000)
+
+/** The number of service domains supported by the USB host shim. */
+#define HV_USB_HOST_NUM_SVC_DOM 1
+
+
+#endif /* _SYS_HV_DRV_USB_HOST_INTF_H */
diff --git a/kernel/arch/tile/include/hv/drv_xgbe_impl.h b/kernel/arch/tile/include/hv/drv_xgbe_impl.h
new file mode 100644
index 000000000..3a73b2b44
--- /dev/null
+++ b/kernel/arch/tile/include/hv/drv_xgbe_impl.h
@@ -0,0 +1,300 @@
+/*
+ * Copyright 2010 Tilera Corporation. All Rights Reserved.
+ *
+ *   This program is free software; you can redistribute it and/or
+ *   modify it under the terms of the GNU General Public License
+ *   as published by the Free Software Foundation, version 2.
+ *
+ *   This program is distributed in the hope that it will be useful, but
+ *   WITHOUT ANY WARRANTY; without even the implied warranty of
+ *   MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
+ *   NON INFRINGEMENT.  See the GNU General Public License for
+ *   more details.
+ */
+
+/**
+ * @file drivers/xgbe/impl.h
+ * Implementation details for the NetIO library.
+ */
+
+#ifndef __DRV_XGBE_IMPL_H__
+#define __DRV_XGBE_IMPL_H__
+
+#include <hv/netio_errors.h>
+#include <hv/netio_intf.h>
+#include <hv/drv_xgbe_intf.h>
+
+
+/** How many groups we have (log2). */
+#define LOG2_NUM_GROUPS (12)
+/** How many groups we have. */
+#define NUM_GROUPS (1 << LOG2_NUM_GROUPS)
+
+/** Number of output requests we'll buffer per tile. */
+#define EPP_REQS_PER_TILE (32)
+
+/** Words used in an eDMA command without checksum acceleration. */
+#define EDMA_WDS_NO_CSUM      8
+/** Words used in an eDMA command with checksum acceleration. */
+#define EDMA_WDS_CSUM        10
+/** Total available words in the eDMA command FIFO. */
+#define EDMA_WDS_TOTAL      128
+
+
+/*
+ * FIXME: These definitions are internal and should have underscores!
+ * NOTE: The actual numeric values here are intentional and allow us to
+ * optimize the concept "if small ... else if large ... else ...", by
+ * checking for the low bit being set, and then for non-zero.
+ * These are used as array indices, so they must have the values (0, 1, 2)
+ * in some order.
+ */
+#define SIZE_SMALL (1)       /**< Small packet queue. */
+#define SIZE_LARGE (2)       /**< Large packet queue. */
+#define SIZE_JUMBO (0)       /**< Jumbo packet queue. */
+
+/** The number of "SIZE_xxx" values. */
+#define NETIO_NUM_SIZES 3
+
+
+/*
+ * Default numbers of packets for IPP drivers.  These values are chosen
+ * such that CIPP1 will not overflow its L2 cache.
+ */
+
+/** The default number of small packets. */
+#define NETIO_DEFAULT_SMALL_PACKETS 2750
+/** The default number of large packets. */
+#define NETIO_DEFAULT_LARGE_PACKETS 2500
+/** The default number of jumbo packets. */
+#define NETIO_DEFAULT_JUMBO_PACKETS 250
+
+
+/** Log2 of the size of a memory arena. */
+#define NETIO_ARENA_SHIFT      24      /* 16 MB */
+/** Size of a memory arena. */
+#define NETIO_ARENA_SIZE       (1 << NETIO_ARENA_SHIFT)
+
+
+/** A queue of packets.
+ *
+ * This structure partially defines a queue of packets waiting to be
+ * processed.  The queue as a whole is written to by an interrupt handler and
+ * read by non-interrupt code; this data structure is what's touched by the
+ * interrupt handler.  The other part of the queue state, the read offset, is
+ * kept in user space, not in hypervisor space, so it is in a separate data
+ * structure.
+ *
+ * The read offset (__packet_receive_read in the user part of the queue
+ * structure) points to the next packet to be read. When the read offset is
+ * equal to the write offset, the queue is empty; therefore the queue must
+ * contain one more slot than the required maximum queue size.
+ *
+ * Here's an example of all 3 state variables and what they mean.  All
+ * pointers move left to right.
+ *
+ * @code
+ *   I   I   V   V   V   V   I   I   I   I
+ *   0   1   2   3   4   5   6   7   8   9  10
+ *           ^       ^       ^               ^
+ *           |               |               |
+ *           |               |               __last_packet_plus_one
+ *           |               __buffer_write
+ *           __packet_receive_read
+ * @endcode
+ *
+ * This queue has 10 slots, and thus can hold 9 packets (_last_packet_plus_one
+ * = 10).  The read pointer is at 2, and the write pointer is at 6; thus,
+ * there are valid, unread packets in slots 2, 3, 4, and 5.  The remaining
+ * slots are invalid (do not contain a packet).
+ */
+typedef struct {
+  /** Byte offset of the next notify packet to be written: zero for the first
+   *  packet on the queue, sizeof (netio_pkt_t) for the second packet on the
+   *  queue, etc. */
+  volatile uint32_t __packet_write;
+
+  /** Offset of the packet after the last valid packet (i.e., when any
+   *  pointer is incremented to this value, it wraps back to zero). */
+  uint32_t __last_packet_plus_one;
+}
+__netio_packet_queue_t;
+
+
+/** A queue of buffers.
+ *
+ * This structure partially defines a queue of empty buffers which have been
+ * obtained via requests to the IPP.  (The elements of the queue are packet
+ * handles, which are transformed into a full netio_pkt_t when the buffer is
+ * retrieved.)  The queue as a whole is written to by an interrupt handler and
+ * read by non-interrupt code; this data structure is what's touched by the
+ * interrupt handler.  The other parts of the queue state, the read offset and
+ * requested write offset, are kept in user space, not in hypervisor space, so
+ * they are in a separate data structure.
+ *
+ * The read offset (__buffer_read in the user part of the queue structure)
+ * points to the next buffer to be read. When the read offset is equal to the
+ * write offset, the queue is empty; therefore the queue must contain one more
+ * slot than the required maximum queue size.
+ *
+ * The requested write offset (__buffer_requested_write in the user part of
+ * the queue structure) points to the slot which will hold the next buffer we
+ * request from the IPP, once we get around to sending such a request.  When
+ * the requested write offset is equal to the write offset, no requests for
+ * new buffers are outstanding; when the requested write offset is one greater
+ * than the read offset, no more requests may be sent.
+ *
+ * Note that, unlike the packet_queue, the buffer_queue places incoming
+ * buffers at decreasing addresses.  This makes the check for "is it time to
+ * wrap the buffer pointer" cheaper in the assembly code which receives new
+ * buffers, and means that the value which defines the queue size,
+ * __last_buffer, is different than in the packet queue.  Also, the offset
+ * used in the packet_queue is already scaled by the size of a packet; here we
+ * use unscaled slot indices for the offsets.  (These differences are
+ * historical, and in the future it's possible that the packet_queue will look
+ * more like this queue.)
+ *
+ * @code
+ * Here's an example of all 4 state variables and what they mean.  Remember:
+ * all pointers move right to left.
+ *
+ *   V   V   V   I   I   R   R   V   V   V
+ *   0   1   2   3   4   5   6   7   8   9
+ *           ^       ^       ^           ^
+ *           |       |       |           |
+ *           |       |       |           __last_buffer
+ *           |       |       __buffer_write
+ *           |       __buffer_requested_write
+ *           __buffer_read
+ * @endcode
+ *
+ * This queue has 10 slots, and thus can hold 9 buffers (_last_buffer = 9).
+ * The read pointer is at 2, and the write pointer is at 6; thus, there are
+ * valid, unread buffers in slots 2, 1, 0, 9, 8, and 7.  The requested write
+ * pointer is at 4; thus, requests have been made to the IPP for buffers which
+ * will be placed in slots 6 and 5 when they arrive.  Finally, the remaining
+ * slots are invalid (do not contain a buffer).
+ */
+typedef struct
+{
+  /** Ordinal number of the next buffer to be written: 0 for the first slot in
+   *  the queue, 1 for the second slot in the queue, etc. */
+  volatile uint32_t __buffer_write;
+
+  /** Ordinal number of the last buffer (i.e., when any pointer is decremented
+   *  below zero, it is reloaded with this value). */
+  uint32_t __last_buffer;
+}
+__netio_buffer_queue_t;
+
+
+/**
+ * An object for providing Ethernet packets to a process.
+ */
+typedef struct __netio_queue_impl_t
+{
+  /** The queue of packets waiting to be received. */
+  __netio_packet_queue_t __packet_receive_queue;
+  /** The intr bit mask that IDs this device. */
+  unsigned int __intr_id;
+  /** Offset to queues of empty buffers, one per size. */
+  uint32_t __buffer_queue[NETIO_NUM_SIZES];
+  /** The address of the first EPP tile, or -1 if no EPP. */
+  /* ISSUE: Actually this is always "0" or "~0". */
+  uint32_t __epp_location;
+  /** The queue ID that this queue represents. */
+  unsigned int __queue_id;
+  /** Number of acknowledgements received. */
+  volatile uint32_t __acks_received;
+  /** Last completion number received for packet_sendv. */
+  volatile uint32_t __last_completion_rcv;
+  /** Number of packets allowed to be outstanding. */
+  uint32_t __max_outstanding;
+  /** First VA available for packets. */
+  void* __va_0;
+  /** First VA in second range available for packets. */
+  void* __va_1;
+  /** Padding to align the "__packets" field to the size of a netio_pkt_t. */
+  uint32_t __padding[3];
+  /** The packets themselves. */
+  netio_pkt_t __packets[0];
+}
+netio_queue_impl_t;
+
+
+/**
+ * An object for managing the user end of a NetIO queue.
+ */
+typedef struct __netio_queue_user_impl_t
+{
+  /** The next incoming packet to be read. */
+  uint32_t __packet_receive_read;
+  /** The next empty buffers to be read, one index per size. */
+  uint8_t __buffer_read[NETIO_NUM_SIZES];
+  /** Where the empty buffer we next request from the IPP will go, one index
+   * per size. */
+  uint8_t __buffer_requested_write[NETIO_NUM_SIZES];
+  /** PCIe interface flag. */
+  uint8_t __pcie;
+  /** Number of packets left to be received before we send a credit update. */
+  uint32_t __receive_credit_remaining;
+  /** Value placed in __receive_credit_remaining when it reaches zero. */
+  uint32_t __receive_credit_interval;
+  /** First fast I/O routine index. */
+  uint32_t __fastio_index;
+  /** Number of acknowledgements expected. */
+  uint32_t __acks_outstanding;
+  /** Last completion number requested. */
+  uint32_t __last_completion_req;
+  /** File descriptor for driver. */
+  int __fd;
+}
+netio_queue_user_impl_t;
+
+
+#define NETIO_GROUP_CHUNK_SIZE   64   /**< Max # groups in one IPP request */
+#define NETIO_BUCKET_CHUNK_SIZE  64   /**< Max # buckets in one IPP request */
+
+
+/** Internal structure used to convey packet send information to the
+ * hypervisor.  FIXME: Actually, it's not used for that anymore, but
+ * netio_packet_send() still uses it internally.
+ */
+typedef struct
+{
+  uint16_t flags;              /**< Packet flags (__NETIO_SEND_FLG_xxx) */
+  uint16_t transfer_size;      /**< Size of packet */
+  uint32_t va;                 /**< VA of start of packet */
+  __netio_pkt_handle_t handle; /**< Packet handle */
+  uint32_t csum0;              /**< First checksum word */
+  uint32_t csum1;              /**< Second checksum word */
+}
+__netio_send_cmd_t;
+
+
+/** Flags used in two contexts:
+ *  - As the "flags" member in the __netio_send_cmd_t, above; used only
+ *    for netio_pkt_send_{prepare,commit}.
+ *  - As part of the flags passed to the various send packet fast I/O calls.
+ */
+
+/** Need acknowledgement on this packet.  Note that some code in the
+ *  normal send_pkt fast I/O handler assumes that this is equal to 1. */
+#define __NETIO_SEND_FLG_ACK    0x1
+
+/** Do checksum on this packet.  (Only used with the __netio_send_cmd_t;
+ *  normal packet sends use a special fast I/O index to denote checksumming,
+ *  and multi-segment sends test the checksum descriptor.) */
+#define __NETIO_SEND_FLG_CSUM   0x2
+
+/** Get a completion on this packet.  Only used with multi-segment sends.  */
+#define __NETIO_SEND_FLG_COMPLETION 0x4
+
+/** Position of the number-of-extra-segments value in the flags word.
+    Only used with multi-segment sends. */
+#define __NETIO_SEND_FLG_XSEG_SHIFT 3
+
+/** Width of the number-of-extra-segments value in the flags word. */
+#define __NETIO_SEND_FLG_XSEG_WIDTH 2
+
+#endif /* __DRV_XGBE_IMPL_H__ */
diff --git a/kernel/arch/tile/include/hv/drv_xgbe_intf.h b/kernel/arch/tile/include/hv/drv_xgbe_intf.h
new file mode 100644
index 000000000..2a20b266d
--- /dev/null
+++ b/kernel/arch/tile/include/hv/drv_xgbe_intf.h
@@ -0,0 +1,615 @@
+/*
+ * Copyright 2010 Tilera Corporation. All Rights Reserved.
+ *
+ *   This program is free software; you can redistribute it and/or
+ *   modify it under the terms of the GNU General Public License
+ *   as published by the Free Software Foundation, version 2.
+ *
+ *   This program is distributed in the hope that it will be useful, but
+ *   WITHOUT ANY WARRANTY; without even the implied warranty of
+ *   MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
+ *   NON INFRINGEMENT.  See the GNU General Public License for
+ *   more details.
+ */
+
+/**
+ * @file drv_xgbe_intf.h
+ * Interface to the hypervisor XGBE driver.
+ */
+
+#ifndef __DRV_XGBE_INTF_H__
+#define __DRV_XGBE_INTF_H__
+
+/**
+ * An object for forwarding VAs and PAs to the hypervisor.
+ * @ingroup types
+ *
+ * This allows the supervisor to specify a number of areas of memory to
+ * store packet buffers.
+ */
+typedef struct
+{
+  /** The physical address of the memory. */
+  HV_PhysAddr pa;
+  /** Page table entry for the memory.  This is only used to derive the
+   *  memory's caching mode; the PA bits are ignored. */
+  HV_PTE pte;
+  /** The virtual address of the memory. */
+  HV_VirtAddr va;
+  /** Size (in bytes) of the memory area. */
+  int size;
+
+}
+netio_ipp_address_t;
+
+/** The various pread/pwrite offsets into the hypervisor-level driver.
+ * @ingroup types
+ */
+typedef enum
+{
+  /** Inform the Linux driver of the address of the NetIO arena memory.
+   *  This offset is actually only used to convey information from netio
+   *  to the Linux driver; it never makes it from there to the hypervisor.
+   *  Write-only; takes a uint32_t specifying the VA address. */
+  NETIO_FIXED_ADDR               = 0x5000000000000000ULL,
+
+  /** Inform the Linux driver of the size of the NetIO arena memory.
+   *  This offset is actually only used to convey information from netio
+   *  to the Linux driver; it never makes it from there to the hypervisor.
+   *  Write-only; takes a uint32_t specifying the VA size. */
+  NETIO_FIXED_SIZE               = 0x5100000000000000ULL,
+
+  /** Register current tile with IPP.  Write then read: write, takes a
+   *  netio_input_config_t, read returns a pointer to a netio_queue_impl_t. */
+  NETIO_IPP_INPUT_REGISTER_OFF   = 0x6000000000000000ULL,
+
+  /** Unregister current tile from IPP.  Write-only, takes a dummy argument. */
+  NETIO_IPP_INPUT_UNREGISTER_OFF = 0x6100000000000000ULL,
+
+  /** Start packets flowing.  Write-only, takes a dummy argument. */
+  NETIO_IPP_INPUT_INIT_OFF       = 0x6200000000000000ULL,
+
+  /** Stop packets flowing.  Write-only, takes a dummy argument. */
+  NETIO_IPP_INPUT_UNINIT_OFF     = 0x6300000000000000ULL,
+
+  /** Configure group (typically we group on VLAN).  Write-only: takes an
+   *  array of netio_group_t's, low 24 bits of the offset is the base group
+   *  number times the size of a netio_group_t. */
+  NETIO_IPP_INPUT_GROUP_CFG_OFF  = 0x6400000000000000ULL,
+
+  /** Configure bucket.  Write-only: takes an array of netio_bucket_t's, low
+   *  24 bits of the offset is the base bucket number times the size of a
+   *  netio_bucket_t. */
+  NETIO_IPP_INPUT_BUCKET_CFG_OFF = 0x6500000000000000ULL,
+
+  /** Get/set a parameter.  Read or write: read or write data is the parameter
+   *  value, low 32 bits of the offset is a __netio_getset_offset_t. */
+  NETIO_IPP_PARAM_OFF            = 0x6600000000000000ULL,
+
+  /** Get fast I/O index.  Read-only; returns a 4-byte base index value. */
+  NETIO_IPP_GET_FASTIO_OFF       = 0x6700000000000000ULL,
+
+  /** Configure hijack IP address.  Packets with this IPv4 dest address
+   *  go to bucket NETIO_NUM_BUCKETS - 1.  Write-only: takes an IP address
+   *  in some standard form.  FIXME: Define the form! */
+  NETIO_IPP_INPUT_HIJACK_CFG_OFF  = 0x6800000000000000ULL,
+
+  /**
+   * Offsets beyond this point are reserved for the supervisor (although that
+   * enforcement must be done by the supervisor driver itself).
+   */
+  NETIO_IPP_USER_MAX_OFF         = 0x6FFFFFFFFFFFFFFFULL,
+
+  /** Register I/O memory.  Write-only, takes a netio_ipp_address_t. */
+  NETIO_IPP_IOMEM_REGISTER_OFF   = 0x7000000000000000ULL,
+
+  /** Unregister I/O memory.  Write-only, takes a netio_ipp_address_t. */
+  NETIO_IPP_IOMEM_UNREGISTER_OFF = 0x7100000000000000ULL,
+
+  /* Offsets greater than 0x7FFFFFFF can't be used directly from Linux
+   * userspace code due to limitations in the pread/pwrite syscalls. */
+
+  /** Drain LIPP buffers. */
+  NETIO_IPP_DRAIN_OFF              = 0xFA00000000000000ULL,
+
+  /** Supply a netio_ipp_address_t to be used as shared memory for the
+   *  LEPP command queue. */
+  NETIO_EPP_SHM_OFF              = 0xFB00000000000000ULL,
+
+  /* 0xFC... is currently unused. */
+
+  /** Stop IPP/EPP tiles.  Write-only, takes a dummy argument.  */
+  NETIO_IPP_STOP_SHIM_OFF        = 0xFD00000000000000ULL,
+
+  /** Start IPP/EPP tiles.  Write-only, takes a dummy argument.  */
+  NETIO_IPP_START_SHIM_OFF       = 0xFE00000000000000ULL,
+
+  /** Supply packet arena.  Write-only, takes an array of
+    * netio_ipp_address_t values. */
+  NETIO_IPP_ADDRESS_OFF          = 0xFF00000000000000ULL,
+} netio_hv_offset_t;
+
+/** Extract the base offset from an offset */
+#define NETIO_BASE_OFFSET(off)    ((off) & 0xFF00000000000000ULL)
+/** Extract the local offset from an offset */
+#define NETIO_LOCAL_OFFSET(off)   ((off) & 0x00FFFFFFFFFFFFFFULL)
+
+
+/**
+ * Get/set offset.
+ */
+typedef union
+{
+  struct
+  {
+    uint64_t addr:48;        /**< Class-specific address */
+    unsigned int class:8;    /**< Class (e.g., NETIO_PARAM) */
+    unsigned int opcode:8;   /**< High 8 bits of NETIO_IPP_PARAM_OFF */
+  }
+  bits;                      /**< Bitfields */
+  uint64_t word;             /**< Aggregated value to use as the offset */
+}
+__netio_getset_offset_t;
+
+/**
+ * Fast I/O index offsets (must be contiguous).
+ */
+typedef enum
+{
+  NETIO_FASTIO_ALLOCATE         = 0, /**< Get empty packet buffer */
+  NETIO_FASTIO_FREE_BUFFER      = 1, /**< Give buffer back to IPP */
+  NETIO_FASTIO_RETURN_CREDITS   = 2, /**< Give credits to IPP */
+  NETIO_FASTIO_SEND_PKT_NOCK    = 3, /**< Send a packet, no checksum */
+  NETIO_FASTIO_SEND_PKT_CK      = 4, /**< Send a packet, with checksum */
+  NETIO_FASTIO_SEND_PKT_VEC     = 5, /**< Send a vector of packets */
+  NETIO_FASTIO_SENDV_PKT        = 6, /**< Sendv one packet */
+  NETIO_FASTIO_NUM_INDEX        = 7, /**< Total number of fast I/O indices */
+} netio_fastio_index_t;
+
+/** 3-word return type for Fast I/O call. */
+typedef struct
+{
+  int err;            /**< Error code. */
+  uint32_t val0;      /**< Value.  Meaning depends upon the specific call. */
+  uint32_t val1;      /**< Value.  Meaning depends upon the specific call. */
+} netio_fastio_rv3_t;
+
+/** 0-argument fast I/O call */
+int __netio_fastio0(uint32_t fastio_index);
+/** 1-argument fast I/O call */
+int __netio_fastio1(uint32_t fastio_index, uint32_t arg0);
+/** 3-argument fast I/O call, 2-word return value */
+netio_fastio_rv3_t __netio_fastio3_rv3(uint32_t fastio_index, uint32_t arg0,
+                                       uint32_t arg1, uint32_t arg2);
+/** 4-argument fast I/O call */
+int __netio_fastio4(uint32_t fastio_index, uint32_t arg0, uint32_t arg1,
+                    uint32_t arg2, uint32_t arg3);
+/** 6-argument fast I/O call */
+int __netio_fastio6(uint32_t fastio_index, uint32_t arg0, uint32_t arg1,
+                    uint32_t arg2, uint32_t arg3, uint32_t arg4, uint32_t arg5);
+/** 9-argument fast I/O call */
+int __netio_fastio9(uint32_t fastio_index, uint32_t arg0, uint32_t arg1,
+                    uint32_t arg2, uint32_t arg3, uint32_t arg4, uint32_t arg5,
+                    uint32_t arg6, uint32_t arg7, uint32_t arg8);
+
+/** Allocate an empty packet.
+ * @param fastio_index Fast I/O index.
+ * @param size Size of the packet to allocate.
+ */
+#define __netio_fastio_allocate(fastio_index, size) \
+  __netio_fastio1((fastio_index) + NETIO_FASTIO_ALLOCATE, size)
+
+/** Free a buffer.
+ * @param fastio_index Fast I/O index.
+ * @param handle Handle for the packet to free.
+ */
+#define __netio_fastio_free_buffer(fastio_index, handle) \
+  __netio_fastio1((fastio_index) + NETIO_FASTIO_FREE_BUFFER, handle)
+
+/** Increment our receive credits.
+ * @param fastio_index Fast I/O index.
+ * @param credits Number of credits to add.
+ */
+#define __netio_fastio_return_credits(fastio_index, credits) \
+  __netio_fastio1((fastio_index) + NETIO_FASTIO_RETURN_CREDITS, credits)
+
+/** Send packet, no checksum.
+ * @param fastio_index Fast I/O index.
+ * @param ackflag Nonzero if we want an ack.
+ * @param size Size of the packet.
+ * @param va Virtual address of start of packet.
+ * @param handle Packet handle.
+ */
+#define __netio_fastio_send_pkt_nock(fastio_index, ackflag, size, va, handle) \
+  __netio_fastio4((fastio_index) + NETIO_FASTIO_SEND_PKT_NOCK, ackflag, \
+                  size, va, handle)
+
+/** Send packet, calculate checksum.
+ * @param fastio_index Fast I/O index.
+ * @param ackflag Nonzero if we want an ack.
+ * @param size Size of the packet.
+ * @param va Virtual address of start of packet.
+ * @param handle Packet handle.
+ * @param csum0 Shim checksum header.
+ * @param csum1 Checksum seed.
+ */
+#define __netio_fastio_send_pkt_ck(fastio_index, ackflag, size, va, handle, \
+                                   csum0, csum1) \
+  __netio_fastio6((fastio_index) + NETIO_FASTIO_SEND_PKT_CK, ackflag, \
+                  size, va, handle, csum0, csum1)
+
+
+/** Format for the "csum0" argument to the __netio_fastio_send routines
+ * and LEPP.  Note that this is currently exactly identical to the
+ * ShimProtocolOffloadHeader.
+ */
+typedef union
+{
+  struct
+  {
+    unsigned int start_byte:7;       /**< The first byte to be checksummed */
+    unsigned int count:14;           /**< Number of bytes to be checksummed. */
+    unsigned int destination_byte:7; /**< The byte to write the checksum to. */
+    unsigned int reserved:4;         /**< Reserved. */
+  } bits;                            /**< Decomposed method of access. */
+  unsigned int word;                 /**< To send out the IDN. */
+} __netio_checksum_header_t;
+
+
+/** Sendv packet with 1 or 2 segments.
+ * @param fastio_index Fast I/O index.
+ * @param flags Ack/csum/notify flags in low 3 bits; number of segments minus
+ *        1 in next 2 bits; expected checksum in high 16 bits.
+ * @param confno Confirmation number to request, if notify flag set.
+ * @param csum0 Checksum descriptor; if zero, no checksum.
+ * @param va_F Virtual address of first segment.
+ * @param va_L Virtual address of last segment, if 2 segments.
+ * @param len_F_L Length of first segment in low 16 bits; length of last
+ *        segment, if 2 segments, in high 16 bits.
+ */
+#define __netio_fastio_sendv_pkt_1_2(fastio_index, flags, confno, csum0, \
+                                     va_F, va_L, len_F_L) \
+  __netio_fastio6((fastio_index) + NETIO_FASTIO_SENDV_PKT, flags, confno, \
+                  csum0, va_F, va_L, len_F_L)
+
+/** Send packet on PCIe interface.
+ * @param fastio_index Fast I/O index.
+ * @param flags Ack/csum/notify flags in low 3 bits.
+ * @param confno Confirmation number to request, if notify flag set.
+ * @param csum0 Checksum descriptor; Hard wired 0, not needed for PCIe.
+ * @param va_F Virtual address of the packet buffer.
+ * @param va_L Virtual address of last segment, if 2 segments. Hard wired 0.
+ * @param len_F_L Length of the packet buffer in low 16 bits.
+ */
+#define __netio_fastio_send_pcie_pkt(fastio_index, flags, confno, csum0, \
+                                     va_F, va_L, len_F_L) \
+  __netio_fastio6((fastio_index) + PCIE_FASTIO_SENDV_PKT, flags, confno, \
+                  csum0, va_F, va_L, len_F_L)
+
+/** Sendv packet with 3 or 4 segments.
+ * @param fastio_index Fast I/O index.
+ * @param flags Ack/csum/notify flags in low 3 bits; number of segments minus
+ *        1 in next 2 bits; expected checksum in high 16 bits.
+ * @param confno Confirmation number to request, if notify flag set.
+ * @param csum0 Checksum descriptor; if zero, no checksum.
+ * @param va_F Virtual address of first segment.
+ * @param va_L Virtual address of last segment (third segment if 3 segments,
+ *        fourth segment if 4 segments).
+ * @param len_F_L Length of first segment in low 16 bits; length of last
+ *        segment in high 16 bits.
+ * @param va_M0 Virtual address of "middle 0" segment; this segment is sent
+ *        second when there are three segments, and third if there are four.
+ * @param va_M1 Virtual address of "middle 1" segment; this segment is sent
+ *        second when there are four segments.
+ * @param len_M0_M1 Length of middle 0 segment in low 16 bits; length of middle
+ *        1 segment, if 4 segments, in high 16 bits.
+ */
+#define __netio_fastio_sendv_pkt_3_4(fastio_index, flags, confno, csum0, va_F, \
+                                     va_L, len_F_L, va_M0, va_M1, len_M0_M1) \
+  __netio_fastio9((fastio_index) + NETIO_FASTIO_SENDV_PKT, flags, confno, \
+                  csum0, va_F, va_L, len_F_L, va_M0, va_M1, len_M0_M1)
+
+/** Send vector of packets.
+ * @param fastio_index Fast I/O index.
+ * @param seqno Number of packets transmitted so far on this interface;
+ *        used to decide which packets should be acknowledged.
+ * @param nentries Number of entries in vector.
+ * @param va Virtual address of start of vector entry array.
+ * @return 3-word netio_fastio_rv3_t structure.  The structure's err member
+ *         is an error code, or zero if no error.  The val0 member is the
+ *         updated value of seqno; it has been incremented by 1 for each
+ *         packet sent.  That increment may be less than nentries if an
+ *         error occurred, or if some of the entries in the vector contain
+ *         handles equal to NETIO_PKT_HANDLE_NONE.  The val1 member is the
+ *         updated value of nentries; it has been decremented by 1 for each
+ *         vector entry processed.  Again, that decrement may be less than
+ *         nentries (leaving the returned value positive) if an error
+ *         occurred.
+ */
+#define __netio_fastio_send_pkt_vec(fastio_index, seqno, nentries, va) \
+  __netio_fastio3_rv3((fastio_index) + NETIO_FASTIO_SEND_PKT_VEC, seqno, \
+                      nentries, va)
+
+
+/** An egress DMA command for LEPP. */
+typedef struct
+{
+  /** Is this a TSO transfer?
+   *
+   * NOTE: This field is always 0, to distinguish it from
+   * lepp_tso_cmd_t.  It must come first!
+   */
+  uint8_t tso               : 1;
+
+  /** Unused padding bits. */
+  uint8_t _unused           : 3;
+
+  /** Should this packet be sent directly from caches instead of DRAM,
+   * using hash-for-home to locate the packet data?
+   */
+  uint8_t hash_for_home     : 1;
+
+  /** Should we compute a checksum? */
+  uint8_t compute_checksum  : 1;
+
+  /** Is this the final buffer for this packet?
+   *
+   * A single packet can be split over several input buffers (a "gather"
+   * operation).  This flag indicates that this is the last buffer
+   * in a packet.
+   */
+  uint8_t end_of_packet     : 1;
+
+  /** Should LEPP advance 'comp_busy' when this DMA is fully finished? */
+  uint8_t send_completion   : 1;
+
+  /** High bits of Client Physical Address of the start of the buffer
+   *  to be egressed.
+   *
+   *  NOTE: Only 6 bits are actually needed here, as CPAs are
+   *  currently 38 bits.  So two bits could be scavenged from this.
+   */
+  uint8_t cpa_hi;
+
+  /** The number of bytes to be egressed. */
+  uint16_t length;
+
+  /** Low 32 bits of Client Physical Address of the start of the buffer
+   *  to be egressed.
+   */
+  uint32_t cpa_lo;
+
+  /** Checksum information (only used if 'compute_checksum'). */
+  __netio_checksum_header_t checksum_data;
+
+} lepp_cmd_t;
+
+
+/** A chunk of physical memory for a TSO egress. */
+typedef struct
+{
+  /** The low bits of the CPA. */
+  uint32_t cpa_lo;
+  /** The high bits of the CPA. */
+  uint16_t cpa_hi		: 15;
+  /** Should this packet be sent directly from caches instead of DRAM,
+   *  using hash-for-home to locate the packet data?
+   */
+  uint16_t hash_for_home	: 1;
+  /** The length in bytes. */
+  uint16_t length;
+} lepp_frag_t;
+
+
+/** An LEPP command that handles TSO. */
+typedef struct
+{
+  /** Is this a TSO transfer?
+   *
+   *  NOTE: This field is always 1, to distinguish it from
+   *  lepp_cmd_t.  It must come first!
+   */
+  uint8_t tso             : 1;
+
+  /** Unused padding bits. */
+  uint8_t _unused         : 7;
+
+  /** Size of the header[] array in bytes.  It must be in the range
+   *  [40, 127], which are the smallest header for a TCP packet over
+   *  Ethernet and the maximum possible prepend size supported by
+   *  hardware, respectively.  Note that the array storage must be
+   *  padded out to a multiple of four bytes so that the following
+   *  LEPP command is aligned properly.
+   */
+  uint8_t header_size;
+
+  /** Byte offset of the IP header in header[]. */
+  uint8_t ip_offset;
+
+  /** Byte offset of the TCP header in header[]. */
+  uint8_t tcp_offset;
+
+  /** The number of bytes to use for the payload of each packet,
+   *  except of course the last one, which may not have enough bytes.
+   *  This means that each Ethernet packet except the last will have a
+   *  size of header_size + payload_size.
+   */
+  uint16_t payload_size;
+
+  /** The length of the 'frags' array that follows this struct. */
+  uint16_t num_frags;
+
+  /** The actual frags. */
+  lepp_frag_t frags[0 /* Variable-sized; num_frags entries. */];
+
+  /*
+   * The packet header template logically follows frags[],
+   * but you can't declare that in C.
+   *
+   * uint32_t header[header_size_in_words_rounded_up];
+   */
+
+} lepp_tso_cmd_t;
+
+
+/** An LEPP completion ring entry. */
+typedef void* lepp_comp_t;
+
+
+/** Maximum number of frags for one TSO command.  This is adapted from
+ *  linux's "MAX_SKB_FRAGS", and presumably over-estimates by one, for
+ *  our page size of exactly 65536.  We add one for a "body" fragment.
+ */
+#define LEPP_MAX_FRAGS (65536 / HV_DEFAULT_PAGE_SIZE_SMALL + 2 + 1)
+
+/** Total number of bytes needed for an lepp_tso_cmd_t. */
+#define LEPP_TSO_CMD_SIZE(num_frags, header_size) \
+  (sizeof(lepp_tso_cmd_t) + \
+   (num_frags) * sizeof(lepp_frag_t) + \
+   (((header_size) + 3) & -4))
+
+/** The size of the lepp "cmd" queue. */
+#define LEPP_CMD_QUEUE_BYTES \
+ (((CHIP_L2_CACHE_SIZE() - 2 * CHIP_L2_LINE_SIZE()) / \
+  (sizeof(lepp_cmd_t) + sizeof(lepp_comp_t))) * sizeof(lepp_cmd_t))
+
+/** The largest possible command that can go in lepp_queue_t::cmds[]. */
+#define LEPP_MAX_CMD_SIZE LEPP_TSO_CMD_SIZE(LEPP_MAX_FRAGS, 128)
+
+/** The largest possible value of lepp_queue_t::cmd_{head, tail} (inclusive).
+ */
+#define LEPP_CMD_LIMIT \
+  (LEPP_CMD_QUEUE_BYTES - LEPP_MAX_CMD_SIZE)
+
+/** The maximum number of completions in an LEPP queue. */
+#define LEPP_COMP_QUEUE_SIZE \
+  ((LEPP_CMD_LIMIT + sizeof(lepp_cmd_t) - 1) / sizeof(lepp_cmd_t))
+
+/** Increment an index modulo the queue size. */
+#define LEPP_QINC(var) \
+  (var = __insn_mnz(var - (LEPP_COMP_QUEUE_SIZE - 1), var + 1))
+
+/** A queue used to convey egress commands from the client to LEPP. */
+typedef struct
+{
+  /** Index of first completion not yet processed by user code.
+   *  If this is equal to comp_busy, there are no such completions.
+   *
+   *  NOTE: This is only read/written by the user.
+   */
+  unsigned int comp_head;
+
+  /** Index of first completion record not yet completed.
+   *  If this is equal to comp_tail, there are no such completions.
+   *  This index gets advanced (modulo LEPP_QUEUE_SIZE) whenever
+   *  a command with the 'completion' bit set is finished.
+   *
+   *  NOTE: This is only written by LEPP, only read by the user.
+   */
+  volatile unsigned int comp_busy;
+
+  /** Index of the first empty slot in the completion ring.
+   *  Entries from this up to but not including comp_head (in ring order)
+   *  can be filled in with completion data.
+   *
+   *  NOTE: This is only read/written by the user.
+   */
+  unsigned int comp_tail;
+
+  /** Byte index of first command enqueued for LEPP but not yet processed.
+   *
+   *  This is always divisible by sizeof(void*) and always <= LEPP_CMD_LIMIT.
+   *
+   *  NOTE: LEPP advances this counter as soon as it no longer needs
+   *  the cmds[] storage for this entry, but the transfer is not actually
+   *  complete (i.e. the buffer pointed to by the command is no longer
+   *  needed) until comp_busy advances.
+   *
+   *  If this is equal to cmd_tail, the ring is empty.
+   *
+   *  NOTE: This is only written by LEPP, only read by the user.
+   */
+  volatile unsigned int cmd_head;
+
+  /** Byte index of first empty slot in the command ring.  This field can
+   *  be incremented up to but not equal to cmd_head (because that would
+   *  mean the ring is empty).
+   *
+   *  This is always divisible by sizeof(void*) and always <= LEPP_CMD_LIMIT.
+   *
+   *  NOTE: This is read/written by the user, only read by LEPP.
+   */
+  volatile unsigned int cmd_tail;
+
+  /** A ring of variable-sized egress DMA commands.
+   *
+   *  NOTE: Only written by the user, only read by LEPP.
+   */
+  char cmds[LEPP_CMD_QUEUE_BYTES]
+    __attribute__((aligned(CHIP_L2_LINE_SIZE())));
+
+  /** A ring of user completion data.
+   *  NOTE: Only read/written by the user.
+   */
+  lepp_comp_t comps[LEPP_COMP_QUEUE_SIZE]
+    __attribute__((aligned(CHIP_L2_LINE_SIZE())));
+} lepp_queue_t;
+
+
+/** An internal helper function for determining the number of entries
+ *  available in a ring buffer, given that there is one sentinel.
+ */
+static inline unsigned int
+_lepp_num_free_slots(unsigned int head, unsigned int tail)
+{
+  /*
+   * One entry is reserved for use as a sentinel, to distinguish
+   * "empty" from "full".  So we compute
+   * (head - tail - 1) % LEPP_QUEUE_SIZE, but without using a slow % operation.
+   */
+  return (head - tail - 1) + ((head <= tail) ? LEPP_COMP_QUEUE_SIZE : 0);
+}
+
+
+/** Returns how many new comp entries can be enqueued. */
+static inline unsigned int
+lepp_num_free_comp_slots(const lepp_queue_t* q)
+{
+  return _lepp_num_free_slots(q->comp_head, q->comp_tail);
+}
+
+static inline int
+lepp_qsub(int v1, int v2)
+{
+  int delta = v1 - v2;
+  return delta + ((delta >> 31) & LEPP_COMP_QUEUE_SIZE);
+}
+
+
+/** FIXME: Check this from linux, via a new "pwrite()" call. */
+#define LIPP_VERSION 1
+
+
+/** We use exactly two bytes of alignment padding. */
+#define LIPP_PACKET_PADDING 2
+
+/** The minimum size of a "small" buffer (including the padding). */
+#define LIPP_SMALL_PACKET_SIZE 128
+
+/*
+ * NOTE: The following two values should total to less than around
+ * 13582, to keep the total size used for "lipp_state_t" below 64K.
+ */
+
+/** The maximum number of "small" buffers.
+ *  This is enough for 53 network cpus with 128 credits.  Note that
+ *  if these are exhausted, we will fall back to using large buffers.
+ */
+#define LIPP_SMALL_BUFFERS 6785
+
+/** The maximum number of "large" buffers.
+ *  This is enough for 53 network cpus with 128 credits.
+ */
+#define LIPP_LARGE_BUFFERS 6785
+
+#endif /* __DRV_XGBE_INTF_H__ */
diff --git a/kernel/arch/tile/include/hv/hypervisor.h b/kernel/arch/tile/include/hv/hypervisor.h
new file mode 100644
index 000000000..e0e6af4e7
--- /dev/null
+++ b/kernel/arch/tile/include/hv/hypervisor.h
@@ -0,0 +1,2598 @@
+/*
+ * Copyright 2010 Tilera Corporation. All Rights Reserved.
+ *
+ *   This program is free software; you can redistribute it and/or
+ *   modify it under the terms of the GNU General Public License
+ *   as published by the Free Software Foundation, version 2.
+ *
+ *   This program is distributed in the hope that it will be useful, but
+ *   WITHOUT ANY WARRANTY; without even the implied warranty of
+ *   MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
+ *   NON INFRINGEMENT.  See the GNU General Public License for
+ *   more details.
+ */
+
+/**
+ * @file hypervisor.h
+ * The hypervisor's public API.
+ */
+
+#ifndef _HV_HV_H
+#define _HV_HV_H
+
+#include <arch/chip.h>
+
+/* Linux builds want unsigned long constants, but assembler wants numbers */
+#ifdef __ASSEMBLER__
+/** One, for assembler */
+#define __HV_SIZE_ONE 1
+#elif !defined(__tile__) && CHIP_VA_WIDTH() > 32
+/** One, for 64-bit on host */
+#define __HV_SIZE_ONE 1ULL
+#else
+/** One, for Linux */
+#define __HV_SIZE_ONE 1UL
+#endif
+
+/** The log2 of the span of a level-1 page table, in bytes.
+ */
+#define HV_LOG2_L1_SPAN 32
+
+/** The span of a level-1 page table, in bytes.
+ */
+#define HV_L1_SPAN (__HV_SIZE_ONE << HV_LOG2_L1_SPAN)
+
+/** The log2 of the initial size of small pages, in bytes.
+ * See HV_DEFAULT_PAGE_SIZE_SMALL.
+ */
+#define HV_LOG2_DEFAULT_PAGE_SIZE_SMALL 16
+
+/** The initial size of small pages, in bytes. This value should be verified
+ * at runtime by calling hv_sysconf(HV_SYSCONF_PAGE_SIZE_SMALL).
+ * It may also be modified when installing a new context.
+ */
+#define HV_DEFAULT_PAGE_SIZE_SMALL \
+  (__HV_SIZE_ONE << HV_LOG2_DEFAULT_PAGE_SIZE_SMALL)
+
+/** The log2 of the initial size of large pages, in bytes.
+ * See HV_DEFAULT_PAGE_SIZE_LARGE.
+ */
+#define HV_LOG2_DEFAULT_PAGE_SIZE_LARGE 24
+
+/** The initial size of large pages, in bytes. This value should be verified
+ * at runtime by calling hv_sysconf(HV_SYSCONF_PAGE_SIZE_LARGE).
+ * It may also be modified when installing a new context.
+ */
+#define HV_DEFAULT_PAGE_SIZE_LARGE \
+  (__HV_SIZE_ONE << HV_LOG2_DEFAULT_PAGE_SIZE_LARGE)
+
+#if CHIP_VA_WIDTH() > 32
+
+/** The log2 of the initial size of jumbo pages, in bytes.
+ * See HV_DEFAULT_PAGE_SIZE_JUMBO.
+ */
+#define HV_LOG2_DEFAULT_PAGE_SIZE_JUMBO 32
+
+/** The initial size of jumbo pages, in bytes. This value should
+ * be verified at runtime by calling hv_sysconf(HV_SYSCONF_PAGE_SIZE_JUMBO).
+ * It may also be modified when installing a new context.
+ */
+#define HV_DEFAULT_PAGE_SIZE_JUMBO \
+  (__HV_SIZE_ONE << HV_LOG2_DEFAULT_PAGE_SIZE_JUMBO)
+
+#endif
+
+/** The log2 of the granularity at which page tables must be aligned;
+ *  in other words, the CPA for a page table must have this many zero
+ *  bits at the bottom of the address.
+ */
+#define HV_LOG2_PAGE_TABLE_ALIGN 11
+
+/** The granularity at which page tables must be aligned.
+ */
+#define HV_PAGE_TABLE_ALIGN (__HV_SIZE_ONE << HV_LOG2_PAGE_TABLE_ALIGN)
+
+/** Normal start of hypervisor glue in client physical memory. */
+#define HV_GLUE_START_CPA 0x10000
+
+/** This much space is reserved at HV_GLUE_START_CPA
+ * for the hypervisor glue. The client program must start at
+ * some address higher than this, and in particular the address of
+ * its text section should be equal to zero modulo HV_PAGE_SIZE_LARGE
+ * so that relative offsets to the HV glue are correct.
+ */
+#define HV_GLUE_RESERVED_SIZE 0x10000
+
+/** Each entry in the hv dispatch array takes this many bytes. */
+#define HV_DISPATCH_ENTRY_SIZE 32
+
+/** Version of the hypervisor interface defined by this file */
+#define _HV_VERSION 13
+
+/** Last version of the hypervisor interface with old hv_init() ABI.
+ *
+ * The change from version 12 to version 13 corresponds to launching
+ * the client by default at PL2 instead of PL1 (corresponding to the
+ * hv itself running at PL3 instead of PL2).  To make this explicit,
+ * the hv_init() API was also extended so the client can report its
+ * desired PL, resulting in a more helpful failure diagnostic.  If you
+ * call hv_init() with _HV_VERSION_OLD_HV_INIT and omit the client_pl
+ * argument, the hypervisor will assume client_pl = 1.
+ *
+ * Note that this is a deprecated solution and we do not expect to
+ * support clients of the Tilera hypervisor running at PL1 indefinitely.
+ */
+#define _HV_VERSION_OLD_HV_INIT 12
+
+/* Index into hypervisor interface dispatch code blocks.
+ *
+ * Hypervisor calls are invoked from user space by calling code
+ * at an address HV_BASE_ADDRESS + (index) * HV_DISPATCH_ENTRY_SIZE,
+ * where index is one of these enum values.
+ *
+ * Normally a supervisor is expected to produce a set of symbols
+ * starting at HV_BASE_ADDRESS that obey this convention, but a user
+ * program could call directly through function pointers if desired.
+ *
+ * These numbers are part of the binary API and will not be changed
+ * without updating HV_VERSION, which should be a rare event.
+ */
+
+/** reserved. */
+#define _HV_DISPATCH_RESERVED                     0
+
+/** hv_init  */
+#define HV_DISPATCH_INIT                          1
+
+/** hv_install_context */
+#define HV_DISPATCH_INSTALL_CONTEXT               2
+
+/** hv_sysconf */
+#define HV_DISPATCH_SYSCONF                       3
+
+/** hv_get_rtc */
+#define HV_DISPATCH_GET_RTC                       4
+
+/** hv_set_rtc */
+#define HV_DISPATCH_SET_RTC                       5
+
+/** hv_flush_asid */
+#define HV_DISPATCH_FLUSH_ASID                    6
+
+/** hv_flush_page */
+#define HV_DISPATCH_FLUSH_PAGE                    7
+
+/** hv_flush_pages */
+#define HV_DISPATCH_FLUSH_PAGES                   8
+
+/** hv_restart */
+#define HV_DISPATCH_RESTART                       9
+
+/** hv_halt */
+#define HV_DISPATCH_HALT                          10
+
+/** hv_power_off */
+#define HV_DISPATCH_POWER_OFF                     11
+
+/** hv_inquire_physical */
+#define HV_DISPATCH_INQUIRE_PHYSICAL              12
+
+/** hv_inquire_memory_controller */
+#define HV_DISPATCH_INQUIRE_MEMORY_CONTROLLER     13
+
+/** hv_inquire_virtual */
+#define HV_DISPATCH_INQUIRE_VIRTUAL               14
+
+/** hv_inquire_asid */
+#define HV_DISPATCH_INQUIRE_ASID                  15
+
+/** hv_nanosleep */
+#define HV_DISPATCH_NANOSLEEP                     16
+
+/** hv_console_read_if_ready */
+#define HV_DISPATCH_CONSOLE_READ_IF_READY         17
+
+/** hv_console_write */
+#define HV_DISPATCH_CONSOLE_WRITE                 18
+
+/** hv_downcall_dispatch */
+#define HV_DISPATCH_DOWNCALL_DISPATCH             19
+
+/** hv_inquire_topology */
+#define HV_DISPATCH_INQUIRE_TOPOLOGY              20
+
+/** hv_fs_findfile */
+#define HV_DISPATCH_FS_FINDFILE                   21
+
+/** hv_fs_fstat */
+#define HV_DISPATCH_FS_FSTAT                      22
+
+/** hv_fs_pread */
+#define HV_DISPATCH_FS_PREAD                      23
+
+/** hv_physaddr_read64 */
+#define HV_DISPATCH_PHYSADDR_READ64               24
+
+/** hv_physaddr_write64 */
+#define HV_DISPATCH_PHYSADDR_WRITE64              25
+
+/** hv_get_command_line */
+#define HV_DISPATCH_GET_COMMAND_LINE              26
+
+/** hv_set_caching */
+#define HV_DISPATCH_SET_CACHING                   27
+
+/** hv_bzero_page */
+#define HV_DISPATCH_BZERO_PAGE                    28
+
+/** hv_register_message_state */
+#define HV_DISPATCH_REGISTER_MESSAGE_STATE        29
+
+/** hv_send_message */
+#define HV_DISPATCH_SEND_MESSAGE                  30
+
+/** hv_receive_message */
+#define HV_DISPATCH_RECEIVE_MESSAGE               31
+
+/** hv_inquire_context */
+#define HV_DISPATCH_INQUIRE_CONTEXT               32
+
+/** hv_start_all_tiles */
+#define HV_DISPATCH_START_ALL_TILES               33
+
+/** hv_dev_open */
+#define HV_DISPATCH_DEV_OPEN                      34
+
+/** hv_dev_close */
+#define HV_DISPATCH_DEV_CLOSE                     35
+
+/** hv_dev_pread */
+#define HV_DISPATCH_DEV_PREAD                     36
+
+/** hv_dev_pwrite */
+#define HV_DISPATCH_DEV_PWRITE                    37
+
+/** hv_dev_poll */
+#define HV_DISPATCH_DEV_POLL                      38
+
+/** hv_dev_poll_cancel */
+#define HV_DISPATCH_DEV_POLL_CANCEL               39
+
+/** hv_dev_preada */
+#define HV_DISPATCH_DEV_PREADA                    40
+
+/** hv_dev_pwritea */
+#define HV_DISPATCH_DEV_PWRITEA                   41
+
+/** hv_flush_remote */
+#define HV_DISPATCH_FLUSH_REMOTE                  42
+
+/** hv_console_putc */
+#define HV_DISPATCH_CONSOLE_PUTC                  43
+
+/** hv_inquire_tiles */
+#define HV_DISPATCH_INQUIRE_TILES                 44
+
+/** hv_confstr */
+#define HV_DISPATCH_CONFSTR                       45
+
+/** hv_reexec */
+#define HV_DISPATCH_REEXEC                        46
+
+/** hv_set_command_line */
+#define HV_DISPATCH_SET_COMMAND_LINE              47
+
+#if !CHIP_HAS_IPI()
+
+/** hv_clear_intr */
+#define HV_DISPATCH_CLEAR_INTR                    48
+
+/** hv_enable_intr */
+#define HV_DISPATCH_ENABLE_INTR                   49
+
+/** hv_disable_intr */
+#define HV_DISPATCH_DISABLE_INTR                  50
+
+/** hv_raise_intr */
+#define HV_DISPATCH_RAISE_INTR                    51
+
+/** hv_trigger_ipi */
+#define HV_DISPATCH_TRIGGER_IPI                   52
+
+#endif /* !CHIP_HAS_IPI() */
+
+/** hv_store_mapping */
+#define HV_DISPATCH_STORE_MAPPING                 53
+
+/** hv_inquire_realpa */
+#define HV_DISPATCH_INQUIRE_REALPA                54
+
+/** hv_flush_all */
+#define HV_DISPATCH_FLUSH_ALL                     55
+
+#if CHIP_HAS_IPI()
+/** hv_get_ipi_pte */
+#define HV_DISPATCH_GET_IPI_PTE                   56
+#endif
+
+/** hv_set_pte_super_shift */
+#define HV_DISPATCH_SET_PTE_SUPER_SHIFT           57
+
+/** hv_console_set_ipi */
+#define HV_DISPATCH_CONSOLE_SET_IPI               63
+
+/** One more than the largest dispatch value */
+#define _HV_DISPATCH_END                          64
+
+
+#ifndef __ASSEMBLER__
+
+#ifdef __KERNEL__
+#include <asm/types.h>
+typedef u32 __hv32;        /**< 32-bit value */
+typedef u64 __hv64;        /**< 64-bit value */
+#else
+#include <stdint.h>
+typedef uint32_t __hv32;   /**< 32-bit value */
+typedef uint64_t __hv64;   /**< 64-bit value */
+#endif
+
+
+/** Hypervisor physical address. */
+typedef __hv64 HV_PhysAddr;
+
+#if CHIP_VA_WIDTH() > 32
+/** Hypervisor virtual address. */
+typedef __hv64 HV_VirtAddr;
+#else
+/** Hypervisor virtual address. */
+typedef __hv32 HV_VirtAddr;
+#endif /* CHIP_VA_WIDTH() > 32 */
+
+/** Hypervisor ASID. */
+typedef unsigned int HV_ASID;
+
+/** Hypervisor tile location for a memory access
+ * ("location overridden target").
+ */
+typedef unsigned int HV_LOTAR;
+
+/** Hypervisor size of a page. */
+typedef unsigned long HV_PageSize;
+
+/** A page table entry.
+ */
+typedef struct
+{
+  __hv64 val;                /**< Value of PTE */
+} HV_PTE;
+
+/** Hypervisor error code. */
+typedef int HV_Errno;
+
+#endif /* !__ASSEMBLER__ */
+
+#define HV_OK           0    /**< No error */
+#define HV_EINVAL      -801  /**< Invalid argument */
+#define HV_ENODEV      -802  /**< No such device */
+#define HV_ENOENT      -803  /**< No such file or directory */
+#define HV_EBADF       -804  /**< Bad file number */
+#define HV_EFAULT      -805  /**< Bad address */
+#define HV_ERECIP      -806  /**< Bad recipients */
+#define HV_E2BIG       -807  /**< Message too big */
+#define HV_ENOTSUP     -808  /**< Service not supported */
+#define HV_EBUSY       -809  /**< Device busy */
+#define HV_ENOSYS      -810  /**< Invalid syscall */
+#define HV_EPERM       -811  /**< No permission */
+#define HV_ENOTREADY   -812  /**< Device not ready */
+#define HV_EIO         -813  /**< I/O error */
+#define HV_ENOMEM      -814  /**< Out of memory */
+#define HV_EAGAIN      -815  /**< Try again */
+
+#define HV_ERR_MAX     -801  /**< Largest HV error code */
+#define HV_ERR_MIN     -815  /**< Smallest HV error code */
+
+#ifndef __ASSEMBLER__
+
+/** Pass HV_VERSION to hv_init to request this version of the interface. */
+typedef enum {
+  HV_VERSION = _HV_VERSION,
+  HV_VERSION_OLD_HV_INIT = _HV_VERSION_OLD_HV_INIT,
+
+} HV_VersionNumber;
+
+/** Initializes the hypervisor.
+ *
+ * @param interface_version_number The version of the hypervisor interface
+ * that this program expects, typically HV_VERSION.
+ * @param chip_num Architecture number of the chip the client was built for.
+ * @param chip_rev_num Revision number of the chip the client was built for.
+ * @param client_pl Privilege level the client is built for
+ *   (not required if interface_version_number == HV_VERSION_OLD_HV_INIT).
+ */
+void hv_init(HV_VersionNumber interface_version_number,
+             int chip_num, int chip_rev_num, int client_pl);
+
+
+/** Queries we can make for hv_sysconf().
+ *
+ * These numbers are part of the binary API and guaranteed not to change.
+ */
+typedef enum {
+  /** An invalid value; do not use. */
+  _HV_SYSCONF_RESERVED       = 0,
+
+  /** The length of the glue section containing the hv_ procs, in bytes. */
+  HV_SYSCONF_GLUE_SIZE       = 1,
+
+  /** The size of small pages, in bytes. */
+  HV_SYSCONF_PAGE_SIZE_SMALL = 2,
+
+  /** The size of large pages, in bytes. */
+  HV_SYSCONF_PAGE_SIZE_LARGE = 3,
+
+  /** Processor clock speed, in hertz. */
+  HV_SYSCONF_CPU_SPEED       = 4,
+
+  /** Processor temperature, in degrees Kelvin.  The value
+   *  HV_SYSCONF_TEMP_KTOC may be subtracted from this to get degrees
+   *  Celsius.  If that Celsius value is HV_SYSCONF_OVERTEMP, this indicates
+   *  that the temperature has hit an upper limit and is no longer being
+   *  accurately tracked.
+   */
+  HV_SYSCONF_CPU_TEMP        = 5,
+
+  /** Board temperature, in degrees Kelvin.  The value
+   *  HV_SYSCONF_TEMP_KTOC may be subtracted from this to get degrees
+   *  Celsius.  If that Celsius value is HV_SYSCONF_OVERTEMP, this indicates
+   *  that the temperature has hit an upper limit and is no longer being
+   *  accurately tracked.
+   */
+  HV_SYSCONF_BOARD_TEMP      = 6,
+
+  /** Legal page size bitmask for hv_install_context().
+   * For example, if 16KB and 64KB small pages are supported,
+   * it would return "HV_CTX_PG_SM_16K | HV_CTX_PG_SM_64K".
+   */
+  HV_SYSCONF_VALID_PAGE_SIZES = 7,
+
+  /** The size of jumbo pages, in bytes.
+   * If no jumbo pages are available, zero will be returned.
+   */
+  HV_SYSCONF_PAGE_SIZE_JUMBO = 8,
+
+} HV_SysconfQuery;
+
+/** Offset to subtract from returned Kelvin temperature to get degrees
+    Celsius. */
+#define HV_SYSCONF_TEMP_KTOC 273
+
+/** Pseudo-temperature value indicating that the temperature has
+ *  pegged at its upper limit and is no longer accurate; note that this is
+ *  the value after subtracting HV_SYSCONF_TEMP_KTOC. */
+#define HV_SYSCONF_OVERTEMP 999
+
+/** Query a configuration value from the hypervisor.
+ * @param query Which value is requested (HV_SYSCONF_xxx).
+ * @return The requested value, or -1 the requested value is illegal or
+ *         unavailable.
+ */
+long hv_sysconf(HV_SysconfQuery query);
+
+
+/** Queries we can make for hv_confstr().
+ *
+ * These numbers are part of the binary API and guaranteed not to change.
+ */
+typedef enum {
+  /** An invalid value; do not use. */
+  _HV_CONFSTR_RESERVED        = 0,
+
+  /** Board part number. */
+  HV_CONFSTR_BOARD_PART_NUM   = 1,
+
+  /** Board serial number. */
+  HV_CONFSTR_BOARD_SERIAL_NUM = 2,
+
+  /** Chip serial number. */
+  HV_CONFSTR_CHIP_SERIAL_NUM  = 3,
+
+  /** Board revision level. */
+  HV_CONFSTR_BOARD_REV        = 4,
+
+  /** Hypervisor software version. */
+  HV_CONFSTR_HV_SW_VER        = 5,
+
+  /** The name for this chip model. */
+  HV_CONFSTR_CHIP_MODEL       = 6,
+
+  /** Human-readable board description. */
+  HV_CONFSTR_BOARD_DESC       = 7,
+
+  /** Human-readable description of the hypervisor configuration. */
+  HV_CONFSTR_HV_CONFIG        = 8,
+
+  /** Human-readable version string for the boot image (for instance,
+   *  who built it and when, what configuration file was used). */
+  HV_CONFSTR_HV_CONFIG_VER    = 9,
+
+  /** Mezzanine part number. */
+  HV_CONFSTR_MEZZ_PART_NUM   = 10,
+
+  /** Mezzanine serial number. */
+  HV_CONFSTR_MEZZ_SERIAL_NUM = 11,
+
+  /** Mezzanine revision level. */
+  HV_CONFSTR_MEZZ_REV        = 12,
+
+  /** Human-readable mezzanine description. */
+  HV_CONFSTR_MEZZ_DESC       = 13,
+
+  /** Control path for the onboard network switch. */
+  HV_CONFSTR_SWITCH_CONTROL  = 14,
+
+  /** Chip revision level. */
+  HV_CONFSTR_CHIP_REV        = 15,
+
+  /** CPU module part number. */
+  HV_CONFSTR_CPUMOD_PART_NUM = 16,
+
+  /** CPU module serial number. */
+  HV_CONFSTR_CPUMOD_SERIAL_NUM = 17,
+
+  /** CPU module revision level. */
+  HV_CONFSTR_CPUMOD_REV      = 18,
+
+  /** Human-readable CPU module description. */
+  HV_CONFSTR_CPUMOD_DESC     = 19,
+
+  /** Per-tile hypervisor statistics.  When this identifier is specified,
+   *  the hv_confstr call takes two extra arguments.  The first is the
+   *  HV_XY_TO_LOTAR of the target tile's coordinates.  The second is
+   *  a flag word.  The only current flag is the lowest bit, which means
+   *  "zero out the stats instead of retrieving them"; in this case the
+   *  buffer and buffer length are ignored. */
+  HV_CONFSTR_HV_STATS        = 20
+
+} HV_ConfstrQuery;
+
+/** Query a configuration string from the hypervisor.
+ *
+ * @param query Identifier for the specific string to be retrieved
+ *        (HV_CONFSTR_xxx).  Some strings may require or permit extra
+ *        arguments to be appended which select specific objects to be
+ *        described; see the string descriptions above.
+ * @param buf Buffer in which to place the string.
+ * @param len Length of the buffer.
+ * @return If query is valid, then the length of the corresponding string,
+ *        including the trailing null; if this is greater than len, the string
+ *        was truncated.  If query is invalid, HV_EINVAL.  If the specified
+ *        buffer is not writable by the client, HV_EFAULT.
+ */
+int hv_confstr(HV_ConfstrQuery query, HV_VirtAddr buf, int len, ...);
+
+/** Tile coordinate */
+typedef struct
+{
+  /** X coordinate, relative to supervisor's top-left coordinate */
+  int x;
+
+  /** Y coordinate, relative to supervisor's top-left coordinate */
+  int y;
+} HV_Coord;
+
+
+#if CHIP_HAS_IPI()
+
+/** Get the PTE for sending an IPI to a particular tile.
+ *
+ * @param tile Tile which will receive the IPI.
+ * @param pl Indicates which IPI registers: 0 = IPI_0, 1 = IPI_1.
+ * @param pte Filled with resulting PTE.
+ * @result Zero if no error, non-zero for invalid parameters.
+ */
+int hv_get_ipi_pte(HV_Coord tile, int pl, HV_PTE* pte);
+
+/** Configure the console interrupt.
+ *
+ * When the console client interrupt is enabled, the hypervisor will
+ * deliver the specified IPI to the client in the following situations:
+ *
+ * - The console has at least one character available for input.
+ *
+ * - The console can accept new characters for output, and the last call
+ *   to hv_console_write() did not write all of the characters requested
+ *   by the client.
+ *
+ * Note that in some system configurations, console interrupt will not
+ * be available; clients should be prepared for this routine to fail and
+ * to fall back to periodic console polling in that case.
+ *
+ * @param ipi Index of the IPI register which will receive the interrupt.
+ * @param event IPI event number for console interrupt. If less than 0,
+ *        disable the console IPI interrupt.
+ * @param coord Tile to be targeted for console interrupt.
+ * @return 0 on success, otherwise, HV_EINVAL if illegal parameter,
+ *         HV_ENOTSUP if console interrupt are not available.
+ */
+int hv_console_set_ipi(int ipi, int event, HV_Coord coord);
+
+#else /* !CHIP_HAS_IPI() */
+
+/** A set of interrupts. */
+typedef __hv32 HV_IntrMask;
+
+/** The low interrupt numbers are reserved for use by the client in
+ *  delivering IPIs.  Any interrupt numbers higher than this value are
+ *  reserved for use by HV device drivers. */
+#define HV_MAX_IPI_INTERRUPT 7
+
+/** Enable a set of device interrupts.
+ *
+ * @param enab_mask Bitmap of interrupts to enable.
+ */
+void hv_enable_intr(HV_IntrMask enab_mask);
+
+/** Disable a set of device interrupts.
+ *
+ * @param disab_mask Bitmap of interrupts to disable.
+ */
+void hv_disable_intr(HV_IntrMask disab_mask);
+
+/** Clear a set of device interrupts.
+ *
+ * @param clear_mask Bitmap of interrupts to clear.
+ */
+void hv_clear_intr(HV_IntrMask clear_mask);
+
+/** Raise a set of device interrupts.
+ *
+ * @param raise_mask Bitmap of interrupts to raise.
+ */
+void hv_raise_intr(HV_IntrMask raise_mask);
+
+/** Trigger a one-shot interrupt on some tile
+ *
+ * @param tile Which tile to interrupt.
+ * @param interrupt Interrupt number to trigger; must be between 0 and
+ *        HV_MAX_IPI_INTERRUPT.
+ * @return HV_OK on success, or a hypervisor error code.
+ */
+HV_Errno hv_trigger_ipi(HV_Coord tile, int interrupt);
+
+#endif /* !CHIP_HAS_IPI() */
+
+/** Store memory mapping in debug memory so that external debugger can read it.
+ * A maximum of 16 entries can be stored.
+ *
+ * @param va VA of memory that is mapped.
+ * @param len Length of mapped memory.
+ * @param pa PA of memory that is mapped.
+ * @return 0 on success, -1 if the maximum number of mappings is exceeded.
+ */
+int hv_store_mapping(HV_VirtAddr va, unsigned int len, HV_PhysAddr pa);
+
+/** Given a client PA and a length, return its real (HV) PA.
+ *
+ * @param cpa Client physical address.
+ * @param len Length of mapped memory.
+ * @return physical address, or -1 if cpa or len is not valid.
+ */
+HV_PhysAddr hv_inquire_realpa(HV_PhysAddr cpa, unsigned int len);
+
+/** RTC return flag for no RTC chip present.
+ */
+#define HV_RTC_NO_CHIP     0x1
+
+/** RTC return flag for low-voltage condition, indicating that battery had
+ * died and time read is unreliable.
+ */
+#define HV_RTC_LOW_VOLTAGE 0x2
+
+/** Date/Time of day */
+typedef struct {
+#if CHIP_WORD_SIZE() > 32
+  __hv64 tm_sec;   /**< Seconds, 0-59 */
+  __hv64 tm_min;   /**< Minutes, 0-59 */
+  __hv64 tm_hour;  /**< Hours, 0-23 */
+  __hv64 tm_mday;  /**< Day of month, 0-30 */
+  __hv64 tm_mon;   /**< Month, 0-11 */
+  __hv64 tm_year;  /**< Years since 1900, 0-199 */
+  __hv64 flags;    /**< Return flags, 0 if no error */
+#else
+  __hv32 tm_sec;   /**< Seconds, 0-59 */
+  __hv32 tm_min;   /**< Minutes, 0-59 */
+  __hv32 tm_hour;  /**< Hours, 0-23 */
+  __hv32 tm_mday;  /**< Day of month, 0-30 */
+  __hv32 tm_mon;   /**< Month, 0-11 */
+  __hv32 tm_year;  /**< Years since 1900, 0-199 */
+  __hv32 flags;    /**< Return flags, 0 if no error */
+#endif
+} HV_RTCTime;
+
+/** Read the current time-of-day clock.
+ * @return HV_RTCTime of current time (GMT).
+ */
+HV_RTCTime hv_get_rtc(void);
+
+
+/** Set the current time-of-day clock.
+ * @param time time to reset time-of-day to (GMT).
+ */
+void hv_set_rtc(HV_RTCTime time);
+
+/** Installs a context, comprising a page table and other attributes.
+ *
+ *  Once this service completes, page_table will be used to translate
+ *  subsequent virtual address references to physical memory.
+ *
+ *  Installing a context does not cause an implicit TLB flush.  Before
+ *  reusing an ASID value for a different address space, the client is
+ *  expected to flush old references from the TLB with hv_flush_asid().
+ *  (Alternately, hv_flush_all() may be used to flush many ASIDs at once.)
+ *  After invalidating a page table entry, changing its attributes, or
+ *  changing its target CPA, the client is expected to flush old references
+ *  from the TLB with hv_flush_page() or hv_flush_pages(). Making a
+ *  previously invalid page valid does not require a flush.
+ *
+ *  Specifying an invalid ASID, or an invalid CPA (client physical address)
+ *  (either as page_table_pointer, or within the referenced table),
+ *  or another page table data item documented as above as illegal may
+ *  lead to client termination; since the validation of the table is
+ *  done as needed, this may happen before the service returns, or at
+ *  some later time, or never, depending upon the client's pattern of
+ *  memory references.  Page table entries which supply translations for
+ *  invalid virtual addresses may result in client termination, or may
+ *  be silently ignored.  "Invalid" in this context means a value which
+ *  was not provided to the client via the appropriate hv_inquire_* routine.
+ *
+ *  To support changing the instruction VAs at the same time as
+ *  installing the new page table, this call explicitly supports
+ *  setting the "lr" register to a different address and then jumping
+ *  directly to the hv_install_context() routine.  In this case, the
+ *  new page table does not need to contain any mapping for the
+ *  hv_install_context address itself.
+ *
+ *  At most one HV_CTX_PG_SM_* flag may be specified in "flags";
+ *  if multiple flags are specified, HV_EINVAL is returned.
+ *  Specifying none of the flags results in using the default page size.
+ *  All cores participating in a given client must request the same
+ *  page size, or the results are undefined.
+ *
+ * @param page_table Root of the page table.
+ * @param access PTE providing info on how to read the page table.  This
+ *   value must be consistent between multiple tiles sharing a page table,
+ *   and must also be consistent with any virtual mappings the client
+ *   may be using to access the page table.
+ * @param asid HV_ASID the page table is to be used for.
+ * @param flags Context flags, denoting attributes or privileges of the
+ *   current context (HV_CTX_xxx).
+ * @return Zero on success, or a hypervisor error code on failure.
+ */
+int hv_install_context(HV_PhysAddr page_table, HV_PTE access, HV_ASID asid,
+                       __hv32 flags);
+
+#endif /* !__ASSEMBLER__ */
+
+#define HV_CTX_DIRECTIO     0x1   /**< Direct I/O requests are accepted from
+                                       PL0. */
+
+#define HV_CTX_PG_SM_4K     0x10  /**< Use 4K small pages, if available. */
+#define HV_CTX_PG_SM_16K    0x20  /**< Use 16K small pages, if available. */
+#define HV_CTX_PG_SM_64K    0x40  /**< Use 64K small pages, if available. */
+#define HV_CTX_PG_SM_MASK   0xf0  /**< Mask of all possible small pages. */
+
+#ifndef __ASSEMBLER__
+
+
+/** Set the number of pages ganged together by HV_PTE_SUPER at a
+ * particular level of the page table.
+ *
+ * The current TILE-Gx hardware only supports powers of four
+ * (i.e. log2_count must be a multiple of two), and the requested
+ * "super" page size must be less than the span of the next level in
+ * the page table.  The largest size that can be requested is 64GB.
+ *
+ * The shift value is initially "0" for all page table levels,
+ * indicating that the HV_PTE_SUPER bit is effectively ignored.
+ *
+ * If you change the count from one non-zero value to another, the
+ * hypervisor will flush the entire TLB and TSB to avoid confusion.
+ *
+ * @param level Page table level (0, 1, or 2)
+ * @param log2_count Base-2 log of the number of pages to gang together,
+ * i.e. how much to shift left the base page size for the super page size.
+ * @return Zero on success, or a hypervisor error code on failure.
+ */
+int hv_set_pte_super_shift(int level, int log2_count);
+
+
+/** Value returned from hv_inquire_context(). */
+typedef struct
+{
+  /** Physical address of page table */
+  HV_PhysAddr page_table;
+
+  /** PTE which defines access method for top of page table */
+  HV_PTE access;
+
+  /** ASID associated with this page table */
+  HV_ASID asid;
+
+  /** Context flags */
+  __hv32 flags;
+} HV_Context;
+
+/** Retrieve information about the currently installed context.
+ * @return The data passed to the last successful hv_install_context call.
+ */
+HV_Context hv_inquire_context(void);
+
+
+/** Flushes all translations associated with the named address space
+ *  identifier from the TLB and any other hypervisor data structures.
+ *  Translations installed with the "global" bit are not flushed.
+ *
+ *  Specifying an invalid ASID may lead to client termination.  "Invalid"
+ *  in this context means a value which was not provided to the client
+ *  via <tt>hv_inquire_asid()</tt>.
+ *
+ * @param asid HV_ASID whose entries are to be flushed.
+ * @return Zero on success, or a hypervisor error code on failure.
+*/
+int hv_flush_asid(HV_ASID asid);
+
+
+/** Flushes all translations associated with the named virtual address
+ *  and page size from the TLB and other hypervisor data structures. Only
+ *  pages visible to the current ASID are affected; note that this includes
+ *  global pages in addition to pages specific to the current ASID.
+ *
+ *  The supplied VA need not be aligned; it may be anywhere in the
+ *  subject page.
+ *
+ *  Specifying an invalid virtual address may lead to client termination,
+ *  or may silently succeed.  "Invalid" in this context means a value
+ *  which was not provided to the client via hv_inquire_virtual.
+ *
+ * @param address Address of the page to flush.
+ * @param page_size Size of pages to assume.
+ * @return Zero on success, or a hypervisor error code on failure.
+ */
+int hv_flush_page(HV_VirtAddr address, HV_PageSize page_size);
+
+
+/** Flushes all translations associated with the named virtual address range
+ *  and page size from the TLB and other hypervisor data structures. Only
+ *  pages visible to the current ASID are affected; note that this includes
+ *  global pages in addition to pages specific to the current ASID.
+ *
+ *  The supplied VA need not be aligned; it may be anywhere in the
+ *  subject page.
+ *
+ *  Specifying an invalid virtual address may lead to client termination,
+ *  or may silently succeed.  "Invalid" in this context means a value
+ *  which was not provided to the client via hv_inquire_virtual.
+ *
+ * @param start Address to flush.
+ * @param page_size Size of pages to assume.
+ * @param size The number of bytes to flush. Any page in the range
+ *        [start, start + size) will be flushed from the TLB.
+ * @return Zero on success, or a hypervisor error code on failure.
+ */
+int hv_flush_pages(HV_VirtAddr start, HV_PageSize page_size,
+                   unsigned long size);
+
+
+/** Flushes all non-global translations (if preserve_global is true),
+ *  or absolutely all translations (if preserve_global is false).
+ *
+ * @param preserve_global Non-zero if we want to preserve "global" mappings.
+ * @return Zero on success, or a hypervisor error code on failure.
+*/
+int hv_flush_all(int preserve_global);
+
+
+/** Restart machine with optional restart command and optional args.
+ * @param cmd Const pointer to command to restart with, or NULL
+ * @param args Const pointer to argument string to restart with, or NULL
+ */
+void hv_restart(HV_VirtAddr cmd, HV_VirtAddr args);
+
+
+/** Halt machine. */
+void hv_halt(void);
+
+
+/** Power off machine. */
+void hv_power_off(void);
+
+
+/** Re-enter virtual-is-physical memory translation mode and restart
+ *  execution at a given address.
+ * @param entry Client physical address at which to begin execution.
+ * @return A hypervisor error code on failure; if the operation is
+ *         successful the call does not return.
+ */
+int hv_reexec(HV_PhysAddr entry);
+
+
+/** Chip topology */
+typedef struct
+{
+  /** Relative coordinates of the querying tile */
+  HV_Coord coord;
+
+  /** Width of the querying supervisor's tile rectangle. */
+  int width;
+
+  /** Height of the querying supervisor's tile rectangle. */
+  int height;
+
+} HV_Topology;
+
+/** Returns information about the tile coordinate system.
+ *
+ * Each supervisor is given a rectangle of tiles it potentially controls.
+ * These tiles are labeled using a relative coordinate system with (0,0) as
+ * the upper left tile regardless of their physical location on the chip.
+ *
+ * This call returns both the size of that rectangle and the position
+ * within that rectangle of the querying tile.
+ *
+ * Not all tiles within that rectangle may be available to the supervisor;
+ * to get the precise set of available tiles, you must also call
+ * hv_inquire_tiles(HV_INQ_TILES_AVAIL, ...).
+ **/
+HV_Topology hv_inquire_topology(void);
+
+/** Sets of tiles we can retrieve with hv_inquire_tiles().
+ *
+ * These numbers are part of the binary API and guaranteed not to change.
+ */
+typedef enum {
+  /** An invalid value; do not use. */
+  _HV_INQ_TILES_RESERVED       = 0,
+
+  /** All available tiles within the supervisor's tile rectangle. */
+  HV_INQ_TILES_AVAIL           = 1,
+
+  /** The set of tiles used for hash-for-home caching. */
+  HV_INQ_TILES_HFH_CACHE       = 2,
+
+  /** The set of tiles that can be legally used as a LOTAR for a PTE. */
+  HV_INQ_TILES_LOTAR           = 3,
+
+  /** The set of "shared" driver tiles that the hypervisor may
+   *  periodically interrupt. */
+  HV_INQ_TILES_SHARED          = 4
+} HV_InqTileSet;
+
+/** Returns specific information about various sets of tiles within the
+ *  supervisor's tile rectangle.
+ *
+ * @param set Which set of tiles to retrieve.
+ * @param cpumask Pointer to a returned bitmask (in row-major order,
+ *        supervisor-relative) of tiles.  The low bit of the first word
+ *        corresponds to the tile at the upper left-hand corner of the
+ *        supervisor's rectangle.  In order for the supervisor to know the
+ *        buffer length to supply, it should first call hv_inquire_topology.
+ * @param length Number of bytes available for the returned bitmask.
+ **/
+HV_Errno hv_inquire_tiles(HV_InqTileSet set, HV_VirtAddr cpumask, int length);
+
+
+/** An identifier for a memory controller. Multiple memory controllers
+ * may be connected to one chip, and this uniquely identifies each one.
+ */
+typedef int HV_MemoryController;
+
+/** A range of physical memory. */
+typedef struct
+{
+  HV_PhysAddr start;   /**< Starting address. */
+  __hv64 size;         /**< Size in bytes. */
+  HV_MemoryController controller;  /**< Which memory controller owns this. */
+} HV_PhysAddrRange;
+
+/** Returns information about a range of physical memory.
+ *
+ * hv_inquire_physical() returns one of the ranges of client
+ * physical addresses which are available to this client.
+ *
+ * The first range is retrieved by specifying an idx of 0, and
+ * successive ranges are returned with subsequent idx values.  Ranges
+ * are ordered by increasing start address (i.e., as idx increases,
+ * so does start), do not overlap, and do not touch (i.e., the
+ * available memory is described with the fewest possible ranges).
+ *
+ * If an out-of-range idx value is specified, the returned size will be zero.
+ * A client can count the number of ranges by increasing idx until the
+ * returned size is zero. There will always be at least one valid range.
+ *
+ * Some clients might not be prepared to deal with more than one
+ * physical address range; they still ought to call this routine and
+ * issue a warning message if they're given more than one range, on the
+ * theory that whoever configured the hypervisor to provide that memory
+ * should know that it's being wasted.
+ */
+HV_PhysAddrRange hv_inquire_physical(int idx);
+
+/** Possible DIMM types. */
+typedef enum
+{
+  NO_DIMM                    = 0,  /**< No DIMM */
+  DDR2                       = 1,  /**< DDR2 */
+  DDR3                       = 2   /**< DDR3 */
+} HV_DIMM_Type;
+
+#ifdef __tilegx__
+
+/** Log2 of minimum DIMM bytes supported by the memory controller. */
+#define HV_MSH_MIN_DIMM_SIZE_SHIFT 29
+
+/** Max number of DIMMs contained by one memory controller. */
+#define HV_MSH_MAX_DIMMS 8
+
+#else
+
+/** Log2 of minimum DIMM bytes supported by the memory controller. */
+#define HV_MSH_MIN_DIMM_SIZE_SHIFT 26
+
+/** Max number of DIMMs contained by one memory controller. */
+#define HV_MSH_MAX_DIMMS 2
+
+#endif
+
+/** Number of bits to right-shift to get the DIMM type. */
+#define HV_DIMM_TYPE_SHIFT 0
+
+/** Bits to mask to get the DIMM type. */
+#define HV_DIMM_TYPE_MASK 0xf
+
+/** Number of bits to right-shift to get the DIMM size. */
+#define HV_DIMM_SIZE_SHIFT 4
+
+/** Bits to mask to get the DIMM size. */
+#define HV_DIMM_SIZE_MASK 0xf
+
+/** Memory controller information. */
+typedef struct
+{
+  HV_Coord coord;   /**< Relative tile coordinates of the port used by a
+                         specified tile to communicate with this controller. */
+  __hv64 speed;     /**< Speed of this controller in bytes per second. */
+} HV_MemoryControllerInfo;
+
+/** Returns information about a particular memory controller.
+ *
+ *  hv_inquire_memory_controller(coord,idx) returns information about a
+ *  particular controller.  Two pieces of information are returned:
+ *  - The relative coordinates of the port on the controller that the specified
+ *    tile would use to contact it.  The relative coordinates may lie
+ *    outside the supervisor's rectangle, i.e. the controller may not
+ *    be attached to a node managed by the querying node's supervisor.
+ *    In particular note that x or y may be negative.
+ *  - The speed of the memory controller.  (This is a not-to-exceed value
+ *    based on the raw hardware data rate, and may not be achievable in
+ *    practice; it is provided to give clients information on the relative
+ *    performance of the available controllers.)
+ *
+ *  Clients should avoid calling this interface with invalid values.
+ *  A client who does may be terminated.
+ * @param coord Tile for which to calculate the relative port position.
+ * @param controller Index of the controller; identical to value returned
+ *        from other routines like hv_inquire_physical.
+ * @return Information about the controller.
+ */
+HV_MemoryControllerInfo hv_inquire_memory_controller(HV_Coord coord,
+                                                     int controller);
+
+
+/** A range of virtual memory. */
+typedef struct
+{
+  HV_VirtAddr start;   /**< Starting address. */
+  __hv64 size;         /**< Size in bytes. */
+} HV_VirtAddrRange;
+
+/** Returns information about a range of virtual memory.
+ *
+ * hv_inquire_virtual() returns one of the ranges of client
+ * virtual addresses which are available to this client.
+ *
+ * The first range is retrieved by specifying an idx of 0, and
+ * successive ranges are returned with subsequent idx values.  Ranges
+ * are ordered by increasing start address (i.e., as idx increases,
+ * so does start), do not overlap, and do not touch (i.e., the
+ * available memory is described with the fewest possible ranges).
+ *
+ * If an out-of-range idx value is specified, the returned size will be zero.
+ * A client can count the number of ranges by increasing idx until the
+ * returned size is zero. There will always be at least one valid range.
+ *
+ * Some clients may well have various virtual addresses hardwired
+ * into themselves; for instance, their instruction stream may
+ * have been compiled expecting to live at a particular address.
+ * Such clients should use this interface to verify they've been
+ * given the virtual address space they expect, and issue a (potentially
+ * fatal) warning message otherwise.
+ *
+ * Note that the returned size is a __hv64, not a __hv32, so it is
+ * possible to express a single range spanning the entire 32-bit
+ * address space.
+ */
+HV_VirtAddrRange hv_inquire_virtual(int idx);
+
+
+/** A range of ASID values. */
+typedef struct
+{
+  HV_ASID start;        /**< First ASID in the range. */
+  unsigned int size;    /**< Number of ASIDs. Zero for an invalid range. */
+} HV_ASIDRange;
+
+/** Returns information about a range of ASIDs.
+ *
+ * hv_inquire_asid() returns one of the ranges of address
+ * space identifiers which are available to this client.
+ *
+ * The first range is retrieved by specifying an idx of 0, and
+ * successive ranges are returned with subsequent idx values.  Ranges
+ * are ordered by increasing start value (i.e., as idx increases,
+ * so does start), do not overlap, and do not touch (i.e., the
+ * available ASIDs are described with the fewest possible ranges).
+ *
+ * If an out-of-range idx value is specified, the returned size will be zero.
+ * A client can count the number of ranges by increasing idx until the
+ * returned size is zero. There will always be at least one valid range.
+ */
+HV_ASIDRange hv_inquire_asid(int idx);
+
+
+/** Waits for at least the specified number of nanoseconds then returns.
+ *
+ * NOTE: this deprecated function currently assumes a 750 MHz clock,
+ * and is thus not generally suitable for use.  New code should call
+ * hv_sysconf(HV_SYSCONF_CPU_SPEED), compute a cycle count to wait for,
+ * and delay by looping while checking the cycle counter SPR.
+ *
+ * @param nanosecs The number of nanoseconds to sleep.
+ */
+void hv_nanosleep(int nanosecs);
+
+
+/** Reads a character from the console without blocking.
+ *
+ * @return A value from 0-255 indicates the value successfully read.
+ * A negative value means no value was ready.
+ */
+int hv_console_read_if_ready(void);
+
+
+/** Writes a character to the console, blocking if the console is busy.
+ *
+ *  This call cannot fail. If the console is broken for some reason,
+ *  output will simply vanish.
+ * @param byte Character to write.
+ */
+void hv_console_putc(int byte);
+
+
+/** Writes a string to the console, blocking if the console is busy.
+ * @param bytes Pointer to characters to write.
+ * @param len Number of characters to write.
+ * @return Number of characters written, or HV_EFAULT if the buffer is invalid.
+ */
+int hv_console_write(HV_VirtAddr bytes, int len);
+
+
+/** Dispatch the next interrupt from the client downcall mechanism.
+ *
+ *  The hypervisor uses downcalls to notify the client of asynchronous
+ *  events.  Some of these events are hypervisor-created (like incoming
+ *  messages).  Some are regular interrupts which initially occur in
+ *  the hypervisor, and are normally handled directly by the client;
+ *  when these occur in a client's interrupt critical section, they must
+ *  be delivered through the downcall mechanism.
+ *
+ *  A downcall is initially delivered to the client as an INTCTRL_CL
+ *  interrupt, where CL is the client's PL.  Upon entry to the INTCTRL_CL
+ *  vector, the client must immediately invoke the hv_downcall_dispatch
+ *  service.  This service will not return; instead it will cause one of
+ *  the client's actual downcall-handling interrupt vectors to be entered.
+ *  The EX_CONTEXT registers in the client will be set so that when the
+ *  client irets, it will return to the code which was interrupted by the
+ *  INTCTRL_CL interrupt.
+ *
+ *  Under some circumstances, the firing of INTCTRL_CL can race with
+ *  the lowering of a device interrupt.  In such a case, the
+ *  hv_downcall_dispatch service may issue an iret instruction instead
+ *  of entering one of the client's actual downcall-handling interrupt
+ *  vectors.  This will return execution to the location that was
+ *  interrupted by INTCTRL_CL.
+ *
+ *  Any saving of registers should be done by the actual handling
+ *  vectors; no registers should be changed by the INTCTRL_CL handler.
+ *  In particular, the client should not use a jal instruction to invoke
+ *  the hv_downcall_dispatch service, as that would overwrite the client's
+ *  lr register.  Note that the hv_downcall_dispatch service may overwrite
+ *  one or more of the client's system save registers.
+ *
+ *  The client must not modify the INTCTRL_CL_STATUS SPR.  The hypervisor
+ *  will set this register to cause a downcall to happen, and will clear
+ *  it when no further downcalls are pending.
+ *
+ *  When a downcall vector is entered, the INTCTRL_CL interrupt will be
+ *  masked.  When the client is done processing a downcall, and is ready
+ *  to accept another, it must unmask this interrupt; if more downcalls
+ *  are pending, this will cause the INTCTRL_CL vector to be reentered.
+ *  Currently the following interrupt vectors can be entered through a
+ *  downcall:
+ *
+ *  INT_MESSAGE_RCV_DWNCL   (hypervisor message available)
+ *  INT_DEV_INTR_DWNCL      (device interrupt)
+ *  INT_DMATLB_MISS_DWNCL   (DMA TLB miss)
+ *  INT_SNITLB_MISS_DWNCL   (SNI TLB miss)
+ *  INT_DMATLB_ACCESS_DWNCL (DMA TLB access violation)
+ */
+void hv_downcall_dispatch(void);
+
+#endif /* !__ASSEMBLER__ */
+
+/** We use actual interrupt vectors which never occur (they're only there
+ *  to allow setting MPLs for related SPRs) for our downcall vectors.
+ */
+/** Message receive downcall interrupt vector */
+#define INT_MESSAGE_RCV_DWNCL    INT_BOOT_ACCESS
+/** DMA TLB miss downcall interrupt vector */
+#define INT_DMATLB_MISS_DWNCL    INT_DMA_ASID
+/** Static nework processor instruction TLB miss interrupt vector */
+#define INT_SNITLB_MISS_DWNCL    INT_SNI_ASID
+/** DMA TLB access violation downcall interrupt vector */
+#define INT_DMATLB_ACCESS_DWNCL  INT_DMA_CPL
+/** Device interrupt downcall interrupt vector */
+#define INT_DEV_INTR_DWNCL       INT_WORLD_ACCESS
+
+#ifndef __ASSEMBLER__
+
+/** Requests the inode for a specific full pathname.
+ *
+ * Performs a lookup in the hypervisor filesystem for a given filename.
+ * Multiple calls with the same filename will always return the same inode.
+ * If there is no such filename, HV_ENOENT is returned.
+ * A bad filename pointer may result in HV_EFAULT instead.
+ *
+ * @param filename Constant pointer to name of requested file
+ * @return Inode of requested file
+ */
+int hv_fs_findfile(HV_VirtAddr filename);
+
+
+/** Data returned from an fstat request.
+ * Note that this structure should be no more than 40 bytes in size so
+ * that it can always be returned completely in registers.
+ */
+typedef struct
+{
+  int size;             /**< Size of file (or HV_Errno on error) */
+  unsigned int flags;   /**< Flags (see HV_FS_FSTAT_FLAGS) */
+} HV_FS_StatInfo;
+
+/** Bitmask flags for fstat request */
+typedef enum
+{
+  HV_FS_ISDIR    = 0x0001   /**< Is the entry a directory? */
+} HV_FS_FSTAT_FLAGS;
+
+/** Get stat information on a given file inode.
+ *
+ * Return information on the file with the given inode.
+ *
+ * IF the HV_FS_ISDIR bit is set, the "file" is a directory.  Reading
+ * it will return NUL-separated filenames (no directory part) relative
+ * to the path to the inode of the directory "file".  These can be
+ * appended to the path to the directory "file" after a forward slash
+ * to create additional filenames.  Note that it is not required
+ * that all valid paths be decomposable into valid parent directories;
+ * a filesystem may validly have just a few files, none of which have
+ * HV_FS_ISDIR set.  However, if clients may wish to enumerate the
+ * files in the filesystem, it is recommended to include all the
+ * appropriate parent directory "files" to give a consistent view.
+ *
+ * An invalid file inode will cause an HV_EBADF error to be returned.
+ *
+ * @param inode The inode number of the query
+ * @return An HV_FS_StatInfo structure
+ */
+HV_FS_StatInfo hv_fs_fstat(int inode);
+
+
+/** Read data from a specific hypervisor file.
+ * On error, may return HV_EBADF for a bad inode or HV_EFAULT for a bad buf.
+ * Reads near the end of the file will return fewer bytes than requested.
+ * Reads at or beyond the end of a file will return zero.
+ *
+ * @param inode the hypervisor file to read
+ * @param buf the buffer to read data into
+ * @param length the number of bytes of data to read
+ * @param offset the offset into the file to read the data from
+ * @return number of bytes successfully read, or an HV_Errno code
+ */
+int hv_fs_pread(int inode, HV_VirtAddr buf, int length, int offset);
+
+
+/** Read a 64-bit word from the specified physical address.
+ * The address must be 8-byte aligned.
+ * Specifying an invalid physical address will lead to client termination.
+ * @param addr The physical address to read
+ * @param access The PTE describing how to read the memory
+ * @return The 64-bit value read from the given address
+ */
+unsigned long long hv_physaddr_read64(HV_PhysAddr addr, HV_PTE access);
+
+
+/** Write a 64-bit word to the specified physical address.
+ * The address must be 8-byte aligned.
+ * Specifying an invalid physical address will lead to client termination.
+ * @param addr The physical address to write
+ * @param access The PTE that says how to write the memory
+ * @param val The 64-bit value to write to the given address
+ */
+void hv_physaddr_write64(HV_PhysAddr addr, HV_PTE access,
+                         unsigned long long val);
+
+
+/** Get the value of the command-line for the supervisor, if any.
+ * This will not include the filename of the booted supervisor, but may
+ * include configured-in boot arguments or the hv_restart() arguments.
+ * If the buffer is not long enough the hypervisor will NUL the first
+ * character of the buffer but not write any other data.
+ * @param buf The virtual address to write the command-line string to.
+ * @param length The length of buf, in characters.
+ * @return The actual length of the command line, including the trailing NUL
+ *         (may be larger than "length").
+ */
+int hv_get_command_line(HV_VirtAddr buf, int length);
+
+
+/** Set a new value for the command-line for the supervisor, which will
+ *  be returned from subsequent invocations of hv_get_command_line() on
+ *  this tile.
+ * @param buf The virtual address to read the command-line string from.
+ * @param length The length of buf, in characters; must be no more than
+ *        HV_COMMAND_LINE_LEN.
+ * @return Zero if successful, or a hypervisor error code.
+ */
+HV_Errno hv_set_command_line(HV_VirtAddr buf, int length);
+
+/** Maximum size of a command line passed to hv_set_command_line(); note
+ *  that a line returned from hv_get_command_line() could be larger than
+ *  this.*/
+#define HV_COMMAND_LINE_LEN  256
+
+/** Tell the hypervisor how to cache non-priority pages
+ * (its own as well as pages explicitly represented in page tables).
+ * Normally these will be represented as red/black pages, but
+ * when the supervisor starts to allocate "priority" pages in the PTE
+ * the hypervisor will need to start marking those pages as (e.g.) "red"
+ * and non-priority pages as either "black" (if they cache-alias
+ * with the existing priority pages) or "red/black" (if they don't).
+ * The bitmask provides information on which parts of the cache
+ * have been used for pinned pages so far on this tile; if (1 << N)
+ * appears in the bitmask, that indicates that a 4KB region of the
+ * cache starting at (N * 4KB) is in use by a "priority" page.
+ * The portion of cache used by a particular page can be computed
+ * by taking the page's PA, modulo CHIP_L2_CACHE_SIZE(), and setting
+ * all the "4KB" bits corresponding to the actual page size.
+ * @param bitmask A bitmap of priority page set values
+ */
+void hv_set_caching(unsigned long bitmask);
+
+
+/** Zero out a specified number of pages.
+ * The va and size must both be multiples of 4096.
+ * Caches are bypassed and memory is directly set to zero.
+ * This API is implemented only in the magic hypervisor and is intended
+ * to provide a performance boost to the minimal supervisor by
+ * giving it a fast way to zero memory pages when allocating them.
+ * @param va Virtual address where the page has been mapped
+ * @param size Number of bytes (must be a page size multiple)
+ */
+void hv_bzero_page(HV_VirtAddr va, unsigned int size);
+
+
+/** State object for the hypervisor messaging subsystem. */
+typedef struct
+{
+#if CHIP_VA_WIDTH() > 32
+  __hv64 opaque[2]; /**< No user-serviceable parts inside */
+#else
+  __hv32 opaque[2]; /**< No user-serviceable parts inside */
+#endif
+}
+HV_MsgState;
+
+/** Register to receive incoming messages.
+ *
+ *  This routine configures the current tile so that it can receive
+ *  incoming messages.  It must be called before the client can receive
+ *  messages with the hv_receive_message routine, and must be called on
+ *  each tile which will receive messages.
+ *
+ *  msgstate is the virtual address of a state object of type HV_MsgState.
+ *  Once the state is registered, the client must not read or write the
+ *  state object; doing so will cause undefined results.
+ *
+ *  If this routine is called with msgstate set to 0, the client's message
+ *  state will be freed and it will no longer be able to receive messages.
+ *  Note that this may cause the loss of any as-yet-undelivered messages
+ *  for the client.
+ *
+ *  If another client attempts to send a message to a client which has
+ *  not yet called hv_register_message_state, or which has freed its
+ *  message state, the message will not be delivered, as if the client
+ *  had insufficient buffering.
+ *
+ *  This routine returns HV_OK if the registration was successful, and
+ *  HV_EINVAL if the supplied state object is unsuitable.  Note that some
+ *  errors may not be detected during this routine, but might be detected
+ *  during a subsequent message delivery.
+ * @param msgstate State object.
+ **/
+HV_Errno hv_register_message_state(HV_MsgState* msgstate);
+
+/** Possible message recipient states. */
+typedef enum
+{
+  HV_TO_BE_SENT,    /**< Not sent (not attempted, or recipient not ready) */
+  HV_SENT,          /**< Successfully sent */
+  HV_BAD_RECIP      /**< Bad recipient coordinates (permanent error) */
+} HV_Recip_State;
+
+/** Message recipient. */
+typedef struct
+{
+  /** X coordinate, relative to supervisor's top-left coordinate */
+  unsigned int x:11;
+
+  /** Y coordinate, relative to supervisor's top-left coordinate */
+  unsigned int y:11;
+
+  /** Status of this recipient */
+  HV_Recip_State state:10;
+} HV_Recipient;
+
+/** Send a message to a set of recipients.
+ *
+ *  This routine sends a message to a set of recipients.
+ *
+ *  recips is an array of HV_Recipient structures.  Each specifies a tile,
+ *  and a message state; initially, it is expected that the state will
+ *  be set to HV_TO_BE_SENT.  nrecip specifies the number of recipients
+ *  in the recips array.
+ *
+ *  For each recipient whose state is HV_TO_BE_SENT, the hypervisor attempts
+ *  to send that tile the specified message.  In order to successfully
+ *  receive the message, the receiver must be a valid tile to which the
+ *  sender has access, must not be the sending tile itself, and must have
+ *  sufficient free buffer space.  (The hypervisor guarantees that each
+ *  tile which has called hv_register_message_state() will be able to
+ *  buffer one message from every other tile which can legally send to it;
+ *  more space may be provided but is not guaranteed.)  If an invalid tile
+ *  is specified, the recipient's state is set to HV_BAD_RECIP; this is a
+ *  permanent delivery error.  If the message is successfully delivered
+ *  to the recipient's buffer, the recipient's state is set to HV_SENT.
+ *  Otherwise, the recipient's state is unchanged.  Message delivery is
+ *  synchronous; all attempts to send messages are completed before this
+ *  routine returns.
+ *
+ *  If no permanent delivery errors were encountered, the routine returns
+ *  the number of messages successfully sent: that is, the number of
+ *  recipients whose states changed from HV_TO_BE_SENT to HV_SENT during
+ *  this operation.  If any permanent delivery errors were encountered,
+ *  the routine returns HV_ERECIP.  In the event of permanent delivery
+ *  errors, it may be the case that delivery was not attempted to all
+ *  recipients; if any messages were successfully delivered, however,
+ *  recipients' state values will be updated appropriately.
+ *
+ *  It is explicitly legal to specify a recipient structure whose state
+ *  is not HV_TO_BE_SENT; such a recipient is ignored.  One suggested way
+ *  of using hv_send_message to send a message to multiple tiles is to set
+ *  up a list of recipients, and then call the routine repeatedly with the
+ *  same list, each time accumulating the number of messages successfully
+ *  sent, until all messages are sent, a permanent error is encountered,
+ *  or the desired number of attempts have been made.  When used in this
+ *  way, the routine will deliver each message no more than once to each
+ *  recipient.
+ *
+ *  Note that a message being successfully delivered to the recipient's
+ *  buffer space does not guarantee that it is received by the recipient,
+ *  either immediately or at any time in the future; the recipient might
+ *  never call hv_receive_message, or could register a different state
+ *  buffer, losing the message.
+ *
+ *  Specifying the same recipient more than once in the recipient list
+ *  is an error, which will not result in an error return but which may
+ *  or may not result in more than one message being delivered to the
+ *  recipient tile.
+ *
+ *  buf and buflen specify the message to be sent.  buf is a virtual address
+ *  which must be currently mapped in the client's page table; if not, the
+ *  routine returns HV_EFAULT.  buflen must be greater than zero and less
+ *  than or equal to HV_MAX_MESSAGE_SIZE, and nrecip must be less than the
+ *  number of tiles to which the sender has access; if not, the routine
+ *  returns HV_EINVAL.
+ * @param recips List of recipients.
+ * @param nrecip Number of recipients.
+ * @param buf Address of message data.
+ * @param buflen Length of message data.
+ **/
+int hv_send_message(HV_Recipient *recips, int nrecip,
+                    HV_VirtAddr buf, int buflen);
+
+/** Maximum hypervisor message size, in bytes */
+#define HV_MAX_MESSAGE_SIZE 28
+
+
+/** Return value from hv_receive_message() */
+typedef struct
+{
+  int msglen;     /**< Message length in bytes, or an error code */
+  __hv32 source;  /**< Code identifying message sender (HV_MSG_xxx) */
+} HV_RcvMsgInfo;
+
+#define HV_MSG_TILE 0x0         /**< Message source is another tile */
+#define HV_MSG_INTR 0x1         /**< Message source is a driver interrupt */
+
+/** Receive a message.
+ *
+ * This routine retrieves a message from the client's incoming message
+ * buffer.
+ *
+ * Multiple messages sent from a particular sending tile to a particular
+ * receiving tile are received in the order that they were sent; however,
+ * no ordering is guaranteed between messages sent by different tiles.
+ *
+ * Whenever the a client's message buffer is empty, the first message
+ * subsequently received will cause the client's MESSAGE_RCV_DWNCL
+ * interrupt vector to be invoked through the interrupt downcall mechanism
+ * (see the description of the hv_downcall_dispatch() routine for details
+ * on downcalls).
+ *
+ * Another message-available downcall will not occur until a call to
+ * this routine is made when the message buffer is empty, and a message
+ * subsequently arrives.  Note that such a downcall could occur while
+ * this routine is executing.  If the calling code does not wish this
+ * to happen, it is recommended that this routine be called with the
+ * INTCTRL_1 interrupt masked, or inside an interrupt critical section.
+ *
+ * msgstate is the value previously passed to hv_register_message_state().
+ * buf is the virtual address of the buffer into which the message will
+ * be written; buflen is the length of the buffer.
+ *
+ * This routine returns an HV_RcvMsgInfo structure.  The msglen member
+ * of that structure is the length of the message received, zero if no
+ * message is available, or HV_E2BIG if the message is too large for the
+ * specified buffer.  If the message is too large, it is not consumed,
+ * and may be retrieved by a subsequent call to this routine specifying
+ * a sufficiently large buffer.  A buffer which is HV_MAX_MESSAGE_SIZE
+ * bytes long is guaranteed to be able to receive any possible message.
+ *
+ * The source member of the HV_RcvMsgInfo structure describes the sender
+ * of the message.  For messages sent by another client tile via an
+ * hv_send_message() call, this value is HV_MSG_TILE; for messages sent
+ * as a result of a device interrupt, this value is HV_MSG_INTR.
+ */
+
+HV_RcvMsgInfo hv_receive_message(HV_MsgState msgstate, HV_VirtAddr buf,
+                                 int buflen);
+
+
+/** Start remaining tiles owned by this supervisor.  Initially, only one tile
+ *  executes the client program; after it calls this service, the other tiles
+ *  are started.  This allows the initial tile to do one-time configuration
+ *  of shared data structures without having to lock them against simultaneous
+ *  access.
+ */
+void hv_start_all_tiles(void);
+
+
+/** Open a hypervisor device.
+ *
+ *  This service initializes an I/O device and its hypervisor driver software,
+ *  and makes it available for use.  The open operation is per-device per-chip;
+ *  once it has been performed, the device handle returned may be used in other
+ *  device services calls made by any tile.
+ *
+ * @param name Name of the device.  A base device name is just a text string
+ *        (say, "pcie").  If there is more than one instance of a device, the
+ *        base name is followed by a slash and a device number (say, "pcie/0").
+ *        Some devices may support further structure beneath those components;
+ *        most notably, devices which require control operations do so by
+ *        supporting reads and/or writes to a control device whose name
+ *        includes a trailing "/ctl" (say, "pcie/0/ctl").
+ * @param flags Flags (HV_DEV_xxx).
+ * @return A positive integer device handle, or a negative error code.
+ */
+int hv_dev_open(HV_VirtAddr name, __hv32 flags);
+
+
+/** Close a hypervisor device.
+ *
+ *  This service uninitializes an I/O device and its hypervisor driver
+ *  software, and makes it unavailable for use.  The close operation is
+ *  per-device per-chip; once it has been performed, the device is no longer
+ *  available.  Normally there is no need to ever call the close service.
+ *
+ * @param devhdl Device handle of the device to be closed.
+ * @return Zero if the close is successful, otherwise, a negative error code.
+ */
+int hv_dev_close(int devhdl);
+
+
+/** Read data from a hypervisor device synchronously.
+ *
+ *  This service transfers data from a hypervisor device to a memory buffer.
+ *  When the service returns, the data has been written from the memory buffer,
+ *  and the buffer will not be further modified by the driver.
+ *
+ *  No ordering is guaranteed between requests issued from different tiles.
+ *
+ *  Devices may choose to support both the synchronous and asynchronous read
+ *  operations, only one of them, or neither of them.
+ *
+ * @param devhdl Device handle of the device to be read from.
+ * @param flags Flags (HV_DEV_xxx).
+ * @param va Virtual address of the target data buffer.  This buffer must
+ *        be mapped in the currently installed page table; if not, HV_EFAULT
+ *        may be returned.
+ * @param len Number of bytes to be transferred.
+ * @param offset Driver-dependent offset.  For a random-access device, this is
+ *        often a byte offset from the beginning of the device; in other cases,
+ *        like on a control device, it may have a different meaning.
+ * @return A non-negative value if the read was at least partially successful;
+ *         otherwise, a negative error code.  The precise interpretation of
+ *         the return value is driver-dependent, but many drivers will return
+ *         the number of bytes successfully transferred.
+ */
+int hv_dev_pread(int devhdl, __hv32 flags, HV_VirtAddr va, __hv32 len,
+                 __hv64 offset);
+
+#define HV_DEV_NB_EMPTY     0x1   /**< Don't block when no bytes of data can
+                                       be transferred. */
+#define HV_DEV_NB_PARTIAL   0x2   /**< Don't block when some bytes, but not all
+                                       of the requested bytes, can be
+                                       transferred. */
+#define HV_DEV_NOCACHE      0x4   /**< The caller warrants that none of the
+                                       cache lines which might contain data
+                                       from the requested buffer are valid.
+                                       Useful with asynchronous operations
+                                       only. */
+
+#define HV_DEV_ALLFLAGS     (HV_DEV_NB_EMPTY | HV_DEV_NB_PARTIAL | \
+                             HV_DEV_NOCACHE)   /**< All HV_DEV_xxx flags */
+
+/** Write data to a hypervisor device synchronously.
+ *
+ *  This service transfers data from a memory buffer to a hypervisor device.
+ *  When the service returns, the data has been read from the memory buffer,
+ *  and the buffer may be overwritten by the client; the data may not
+ *  necessarily have been conveyed to the actual hardware I/O interface.
+ *
+ *  No ordering is guaranteed between requests issued from different tiles.
+ *
+ *  Devices may choose to support both the synchronous and asynchronous write
+ *  operations, only one of them, or neither of them.
+ *
+ * @param devhdl Device handle of the device to be written to.
+ * @param flags Flags (HV_DEV_xxx).
+ * @param va Virtual address of the source data buffer.  This buffer must
+ *        be mapped in the currently installed page table; if not, HV_EFAULT
+ *        may be returned.
+ * @param len Number of bytes to be transferred.
+ * @param offset Driver-dependent offset.  For a random-access device, this is
+ *        often a byte offset from the beginning of the device; in other cases,
+ *        like on a control device, it may have a different meaning.
+ * @return A non-negative value if the write was at least partially successful;
+ *         otherwise, a negative error code.  The precise interpretation of
+ *         the return value is driver-dependent, but many drivers will return
+ *         the number of bytes successfully transferred.
+ */
+int hv_dev_pwrite(int devhdl, __hv32 flags, HV_VirtAddr va, __hv32 len,
+                  __hv64 offset);
+
+
+/** Interrupt arguments, used in the asynchronous I/O interfaces. */
+#if CHIP_VA_WIDTH() > 32
+typedef __hv64 HV_IntArg;
+#else
+typedef __hv32 HV_IntArg;
+#endif
+
+/** Interrupt messages are delivered via the mechanism as normal messages,
+ *  but have a message source of HV_DEV_INTR.  The message is formatted
+ *  as an HV_IntrMsg structure.
+ */
+
+typedef struct
+{
+  HV_IntArg intarg;  /**< Interrupt argument, passed to the poll/preada/pwritea
+                          services */
+  HV_IntArg intdata; /**< Interrupt-specific interrupt data */
+} HV_IntrMsg;
+
+/** Request an interrupt message when a device condition is satisfied.
+ *
+ *  This service requests that an interrupt message be delivered to the
+ *  requesting tile when a device becomes readable or writable, or when any
+ *  data queued to the device via previous write operations from this tile
+ *  has been actually sent out on the hardware I/O interface.  Devices may
+ *  choose to support any, all, or none of the available conditions.
+ *
+ *  If multiple conditions are specified, only one message will be
+ *  delivered.  If the event mask delivered to that interrupt handler
+ *  indicates that some of the conditions have not yet occurred, the
+ *  client must issue another poll() call if it wishes to wait for those
+ *  conditions.
+ *
+ *  Only one poll may be outstanding per device handle per tile.  If more than
+ *  one tile is polling on the same device and condition, they will all be
+ *  notified when it happens.  Because of this, clients may not assume that
+ *  the condition signaled is necessarily still true when they request a
+ *  subsequent service; for instance, the readable data which caused the
+ *  poll call to interrupt may have been read by another tile in the interim.
+ *
+ *  The notification interrupt message could come directly, or via the
+ *  downcall (intctrl1) method, depending on what the tile is doing
+ *  when the condition is satisfied.  Note that it is possible for the
+ *  requested interrupt to be delivered after this service is called but
+ *  before it returns.
+ *
+ * @param devhdl Device handle of the device to be polled.
+ * @param events Flags denoting the events which will cause the interrupt to
+ *        be delivered (HV_DEVPOLL_xxx).
+ * @param intarg Value which will be delivered as the intarg member of the
+ *        eventual interrupt message; the intdata member will be set to a
+ *        mask of HV_DEVPOLL_xxx values indicating which conditions have been
+ *        satisifed.
+ * @return Zero if the interrupt was successfully scheduled; otherwise, a
+ *         negative error code.
+ */
+int hv_dev_poll(int devhdl, __hv32 events, HV_IntArg intarg);
+
+#define HV_DEVPOLL_READ     0x1   /**< Test device for readability */
+#define HV_DEVPOLL_WRITE    0x2   /**< Test device for writability */
+#define HV_DEVPOLL_FLUSH    0x4   /**< Test device for output drained */
+
+
+/** Cancel a request for an interrupt when a device event occurs.
+ *
+ *  This service requests that no interrupt be delivered when the events
+ *  noted in the last-issued poll() call happen.  Once this service returns,
+ *  the interrupt has been canceled; however, it is possible for the interrupt
+ *  to be delivered after this service is called but before it returns.
+ *
+ * @param devhdl Device handle of the device on which to cancel polling.
+ * @return Zero if the poll was successfully canceled; otherwise, a negative
+ *         error code.
+ */
+int hv_dev_poll_cancel(int devhdl);
+
+
+/** Scatter-gather list for preada/pwritea calls. */
+typedef struct
+#if CHIP_VA_WIDTH() <= 32
+__attribute__ ((packed, aligned(4)))
+#endif
+{
+  HV_PhysAddr pa;  /**< Client physical address of the buffer segment. */
+  HV_PTE pte;      /**< Page table entry describing the caching and location
+                        override characteristics of the buffer segment.  Some
+                        drivers ignore this element and will require that
+                        the NOCACHE flag be set on their requests. */
+  __hv32 len;      /**< Length of the buffer segment. */
+} HV_SGL;
+
+#define HV_SGL_MAXLEN 16  /**< Maximum number of entries in a scatter-gather
+                               list */
+
+/** Read data from a hypervisor device asynchronously.
+ *
+ *  This service transfers data from a hypervisor device to a memory buffer.
+ *  When the service returns, the read has been scheduled.  When the read
+ *  completes, an interrupt message will be delivered, and the buffer will
+ *  not be further modified by the driver.
+ *
+ *  The number of possible outstanding asynchronous requests is defined by
+ *  each driver, but it is recommended that it be at least two requests
+ *  per tile per device.
+ *
+ *  No ordering is guaranteed between synchronous and asynchronous requests,
+ *  even those issued on the same tile.
+ *
+ *  The completion interrupt message could come directly, or via the downcall
+ *  (intctrl1) method, depending on what the tile is doing when the read
+ *  completes.  Interrupts do not coalesce; one is delivered for each
+ *  asynchronous I/O request.  Note that it is possible for the requested
+ *  interrupt to be delivered after this service is called but before it
+ *  returns.
+ *
+ *  Devices may choose to support both the synchronous and asynchronous read
+ *  operations, only one of them, or neither of them.
+ *
+ * @param devhdl Device handle of the device to be read from.
+ * @param flags Flags (HV_DEV_xxx).
+ * @param sgl_len Number of elements in the scatter-gather list.
+ * @param sgl Scatter-gather list describing the memory to which data will be
+ *        written.
+ * @param offset Driver-dependent offset.  For a random-access device, this is
+ *        often a byte offset from the beginning of the device; in other cases,
+ *        like on a control device, it may have a different meaning.
+ * @param intarg Value which will be delivered as the intarg member of the
+ *        eventual interrupt message; the intdata member will be set to the
+ *        normal return value from the read request.
+ * @return Zero if the read was successfully scheduled; otherwise, a negative
+ *         error code.  Note that some drivers may choose to pre-validate
+ *         their arguments, and may thus detect certain device error
+ *         conditions at this time rather than when the completion notification
+ *         occurs, but this is not required.
+ */
+int hv_dev_preada(int devhdl, __hv32 flags, __hv32 sgl_len,
+                  HV_SGL sgl[/* sgl_len */], __hv64 offset, HV_IntArg intarg);
+
+
+/** Write data to a hypervisor device asynchronously.
+ *
+ *  This service transfers data from a memory buffer to a hypervisor
+ *  device.  When the service returns, the write has been scheduled.
+ *  When the write completes, an interrupt message will be delivered,
+ *  and the buffer may be overwritten by the client; the data may not
+ *  necessarily have been conveyed to the actual hardware I/O interface.
+ *
+ *  The number of possible outstanding asynchronous requests is defined by
+ *  each driver, but it is recommended that it be at least two requests
+ *  per tile per device.
+ *
+ *  No ordering is guaranteed between synchronous and asynchronous requests,
+ *  even those issued on the same tile.
+ *
+ *  The completion interrupt message could come directly, or via the downcall
+ *  (intctrl1) method, depending on what the tile is doing when the read
+ *  completes.  Interrupts do not coalesce; one is delivered for each
+ *  asynchronous I/O request.  Note that it is possible for the requested
+ *  interrupt to be delivered after this service is called but before it
+ *  returns.
+ *
+ *  Devices may choose to support both the synchronous and asynchronous write
+ *  operations, only one of them, or neither of them.
+ *
+ * @param devhdl Device handle of the device to be read from.
+ * @param flags Flags (HV_DEV_xxx).
+ * @param sgl_len Number of elements in the scatter-gather list.
+ * @param sgl Scatter-gather list describing the memory from which data will be
+ *        read.
+ * @param offset Driver-dependent offset.  For a random-access device, this is
+ *        often a byte offset from the beginning of the device; in other cases,
+ *        like on a control device, it may have a different meaning.
+ * @param intarg Value which will be delivered as the intarg member of the
+ *        eventual interrupt message; the intdata member will be set to the
+ *        normal return value from the write request.
+ * @return Zero if the write was successfully scheduled; otherwise, a negative
+ *         error code.  Note that some drivers may choose to pre-validate
+ *         their arguments, and may thus detect certain device error
+ *         conditions at this time rather than when the completion notification
+ *         occurs, but this is not required.
+ */
+int hv_dev_pwritea(int devhdl, __hv32 flags, __hv32 sgl_len,
+                   HV_SGL sgl[/* sgl_len */], __hv64 offset, HV_IntArg intarg);
+
+
+/** Define a pair of tile and ASID to identify a user process context. */
+typedef struct
+{
+  /** X coordinate, relative to supervisor's top-left coordinate */
+  unsigned int x:11;
+
+  /** Y coordinate, relative to supervisor's top-left coordinate */
+  unsigned int y:11;
+
+  /** ASID of the process on this x,y tile */
+  HV_ASID asid:10;
+} HV_Remote_ASID;
+
+/** Flush cache and/or TLB state on remote tiles.
+ *
+ * @param cache_pa Client physical address to flush from cache (ignored if
+ *        the length encoded in cache_control is zero, or if
+ *        HV_FLUSH_EVICT_L2 is set, or if cache_cpumask is NULL).
+ * @param cache_control This argument allows you to specify a length of
+ *        physical address space to flush (maximum HV_FLUSH_MAX_CACHE_LEN).
+ *        You can "or" in HV_FLUSH_EVICT_L2 to flush the whole L2 cache.
+ *        You can "or" in HV_FLUSH_EVICT_L1I to flush the whole L1I cache.
+ *        HV_FLUSH_ALL flushes all caches.
+ * @param cache_cpumask Bitmask (in row-major order, supervisor-relative) of
+ *        tile indices to perform cache flush on.  The low bit of the first
+ *        word corresponds to the tile at the upper left-hand corner of the
+ *        supervisor's rectangle.  If passed as a NULL pointer, equivalent
+ *        to an empty bitmask.  On chips which support hash-for-home caching,
+ *        if passed as -1, equivalent to a mask containing tiles which could
+ *        be doing hash-for-home caching.
+ * @param tlb_va Virtual address to flush from TLB (ignored if
+ *        tlb_length is zero or tlb_cpumask is NULL).
+ * @param tlb_length Number of bytes of data to flush from the TLB.
+ * @param tlb_pgsize Page size to use for TLB flushes.
+ *        tlb_va and tlb_length need not be aligned to this size.
+ * @param tlb_cpumask Bitmask for tlb flush, like cache_cpumask.
+ *        If passed as a NULL pointer, equivalent to an empty bitmask.
+ * @param asids Pointer to an HV_Remote_ASID array of tile/ASID pairs to flush.
+ * @param asidcount Number of HV_Remote_ASID entries in asids[].
+ * @return Zero for success, or else HV_EINVAL or HV_EFAULT for errors that
+ *        are detected while parsing the arguments.
+ */
+int hv_flush_remote(HV_PhysAddr cache_pa, unsigned long cache_control,
+                    unsigned long* cache_cpumask,
+                    HV_VirtAddr tlb_va, unsigned long tlb_length,
+                    unsigned long tlb_pgsize, unsigned long* tlb_cpumask,
+                    HV_Remote_ASID* asids, int asidcount);
+
+/** Include in cache_control to ensure a flush of the entire L2. */
+#define HV_FLUSH_EVICT_L2 (1UL << 31)
+
+/** Include in cache_control to ensure a flush of the entire L1I. */
+#define HV_FLUSH_EVICT_L1I (1UL << 30)
+
+/** Maximum legal size to use for the "length" component of cache_control. */
+#define HV_FLUSH_MAX_CACHE_LEN ((1UL << 30) - 1)
+
+/** Use for cache_control to ensure a flush of all caches. */
+#define HV_FLUSH_ALL -1UL
+
+#else   /* __ASSEMBLER__ */
+
+/** Include in cache_control to ensure a flush of the entire L2. */
+#define HV_FLUSH_EVICT_L2 (1 << 31)
+
+/** Include in cache_control to ensure a flush of the entire L1I. */
+#define HV_FLUSH_EVICT_L1I (1 << 30)
+
+/** Maximum legal size to use for the "length" component of cache_control. */
+#define HV_FLUSH_MAX_CACHE_LEN ((1 << 30) - 1)
+
+/** Use for cache_control to ensure a flush of all caches. */
+#define HV_FLUSH_ALL -1
+
+#endif  /* __ASSEMBLER__ */
+
+#ifndef __ASSEMBLER__
+
+/** Return a 64-bit value corresponding to the PTE if needed */
+#define hv_pte_val(pte) ((pte).val)
+
+/** Cast a 64-bit value to an HV_PTE */
+#define hv_pte(val) ((HV_PTE) { val })
+
+#endif  /* !__ASSEMBLER__ */
+
+
+/** Bits in the size of an HV_PTE */
+#define HV_LOG2_PTE_SIZE 3
+
+/** Size of an HV_PTE */
+#define HV_PTE_SIZE (1 << HV_LOG2_PTE_SIZE)
+
+
+/* Bits in HV_PTE's low word. */
+#define HV_PTE_INDEX_PRESENT          0  /**< PTE is valid */
+#define HV_PTE_INDEX_MIGRATING        1  /**< Page is migrating */
+#define HV_PTE_INDEX_CLIENT0          2  /**< Page client state 0 */
+#define HV_PTE_INDEX_CLIENT1          3  /**< Page client state 1 */
+#define HV_PTE_INDEX_NC               4  /**< L1$/L2$ incoherent with L3$ */
+#define HV_PTE_INDEX_NO_ALLOC_L1      5  /**< Page is uncached in local L1$ */
+#define HV_PTE_INDEX_NO_ALLOC_L2      6  /**< Page is uncached in local L2$ */
+#define HV_PTE_INDEX_CACHED_PRIORITY  7  /**< Page is priority cached */
+#define HV_PTE_INDEX_PAGE             8  /**< PTE describes a page */
+#define HV_PTE_INDEX_GLOBAL           9  /**< Page is global */
+#define HV_PTE_INDEX_USER            10  /**< Page is user-accessible */
+#define HV_PTE_INDEX_ACCESSED        11  /**< Page has been accessed */
+#define HV_PTE_INDEX_DIRTY           12  /**< Page has been written */
+                                         /*   Bits 13-14 are reserved for
+                                              future use. */
+#define HV_PTE_INDEX_SUPER           15  /**< Pages ganged together for TLB */
+#define HV_PTE_INDEX_MODE            16  /**< Page mode; see HV_PTE_MODE_xxx */
+#define HV_PTE_MODE_BITS              3  /**< Number of bits in mode */
+#define HV_PTE_INDEX_CLIENT2         19  /**< Page client state 2 */
+#define HV_PTE_INDEX_LOTAR           20  /**< Page's LOTAR; must be high bits
+                                              of word */
+#define HV_PTE_LOTAR_BITS            12  /**< Number of bits in a LOTAR */
+
+/* Bits in HV_PTE's high word. */
+#define HV_PTE_INDEX_READABLE        32  /**< Page is readable */
+#define HV_PTE_INDEX_WRITABLE        33  /**< Page is writable */
+#define HV_PTE_INDEX_EXECUTABLE      34  /**< Page is executable */
+#define HV_PTE_INDEX_PTFN            35  /**< Page's PTFN; must be high bits
+                                              of word */
+#define HV_PTE_PTFN_BITS             29  /**< Number of bits in a PTFN */
+
+/*
+ * Legal values for the PTE's mode field
+ */
+/** Data is not resident in any caches; loads and stores access memory
+ *  directly.
+ */
+#define HV_PTE_MODE_UNCACHED          1
+
+/** Data is resident in the tile's local L1 and/or L2 caches; if a load
+ *  or store misses there, it goes to memory.
+ *
+ *  The copy in the local L1$/L2$ is not invalidated when the copy in
+ *  memory is changed.
+ */
+#define HV_PTE_MODE_CACHE_NO_L3       2
+
+/** Data is resident in the tile's local L1 and/or L2 caches.  If a load
+ *  or store misses there, it goes to an L3 cache in a designated tile;
+ *  if it misses there, it goes to memory.
+ *
+ *  If the NC bit is not set, the copy in the local L1$/L2$ is invalidated
+ *  when the copy in the remote L3$ is changed.  Otherwise, such
+ *  invalidation will not occur.
+ *
+ *  Chips for which CHIP_HAS_COHERENT_LOCAL_CACHE() is 0 do not support
+ *  invalidation from an L3$ to another tile's L1$/L2$.  If the NC bit is
+ *  clear on such a chip, no copy is kept in the local L1$/L2$ in this mode.
+ */
+#define HV_PTE_MODE_CACHE_TILE_L3     3
+
+/** Data is resident in the tile's local L1 and/or L2 caches.  If a load
+ *  or store misses there, it goes to an L3 cache in one of a set of
+ *  designated tiles; if it misses there, it goes to memory.  Which tile
+ *  is chosen from the set depends upon a hash function applied to the
+ *  physical address.  This mode is not supported on chips for which
+ *  CHIP_HAS_CBOX_HOME_MAP() is 0.
+ *
+ *  If the NC bit is not set, the copy in the local L1$/L2$ is invalidated
+ *  when the copy in the remote L3$ is changed.  Otherwise, such
+ *  invalidation will not occur.
+ *
+ *  Chips for which CHIP_HAS_COHERENT_LOCAL_CACHE() is 0 do not support
+ *  invalidation from an L3$ to another tile's L1$/L2$.  If the NC bit is
+ *  clear on such a chip, no copy is kept in the local L1$/L2$ in this mode.
+ */
+#define HV_PTE_MODE_CACHE_HASH_L3     4
+
+/** Data is not resident in memory; accesses are instead made to an I/O
+ *  device, whose tile coordinates are given by the PTE's LOTAR field.
+ *  This mode is only supported on chips for which CHIP_HAS_MMIO() is 1.
+ *  The EXECUTABLE bit may not be set in an MMIO PTE.
+ */
+#define HV_PTE_MODE_MMIO              5
+
+
+/* C wants 1ULL so it is typed as __hv64, but the assembler needs just numbers.
+ * The assembler can't handle shifts greater than 31, but treats them
+ * as shifts mod 32, so assembler code must be aware of which word
+ * the bit belongs in when using these macros.
+ */
+#ifdef __ASSEMBLER__
+#define __HV_PTE_ONE 1        /**< One, for assembler */
+#else
+#define __HV_PTE_ONE 1ULL     /**< One, for C */
+#endif
+
+/** Is this PTE present?
+ *
+ * If this bit is set, this PTE represents a valid translation or level-2
+ * page table pointer.  Otherwise, the page table does not contain a
+ * translation for the subject virtual pages.
+ *
+ * If this bit is not set, the other bits in the PTE are not
+ * interpreted by the hypervisor, and may contain any value.
+ */
+#define HV_PTE_PRESENT               (__HV_PTE_ONE << HV_PTE_INDEX_PRESENT)
+
+/** Does this PTE map a page?
+ *
+ * If this bit is set in a level-0 page table, the entry should be
+ * interpreted as a level-2 page table entry mapping a jumbo page.
+ *
+ * If this bit is set in a level-1 page table, the entry should be
+ * interpreted as a level-2 page table entry mapping a large page.
+ *
+ * This bit should not be modified by the client while PRESENT is set, as
+ * doing so may race with the hypervisor's update of ACCESSED and DIRTY bits.
+ *
+ * In a level-2 page table, this bit is ignored and must be zero.
+ */
+#define HV_PTE_PAGE                  (__HV_PTE_ONE << HV_PTE_INDEX_PAGE)
+
+/** Does this PTE implicitly reference multiple pages?
+ *
+ * If this bit is set in the page table (either in the level-2 page table,
+ * or in a higher level page table in conjunction with the PAGE bit)
+ * then the PTE specifies a range of contiguous pages, not a single page.
+ * The hv_set_pte_super_shift() allows you to specify the count for
+ * each level of the page table.
+ *
+ * Note: this bit is not supported on TILEPro systems.
+ */
+#define HV_PTE_SUPER                 (__HV_PTE_ONE << HV_PTE_INDEX_SUPER)
+
+/** Is this a global (non-ASID) mapping?
+ *
+ * If this bit is set, the translations established by this PTE will
+ * not be flushed from the TLB by the hv_flush_asid() service; they
+ * will be flushed by the hv_flush_page() or hv_flush_pages() services.
+ *
+ * Setting this bit for translations which are identical in all page
+ * tables (for instance, code and data belonging to a client OS) can
+ * be very beneficial, as it will reduce the number of TLB misses.
+ * Note that, while it is not an error which will be detected by the
+ * hypervisor, it is an extremely bad idea to set this bit for
+ * translations which are _not_ identical in all page tables.
+ *
+ * This bit should not be modified by the client while PRESENT is set, as
+ * doing so may race with the hypervisor's update of ACCESSED and DIRTY bits.
+ *
+ * This bit is ignored in level-1 PTEs unless the Page bit is set.
+ */
+#define HV_PTE_GLOBAL                (__HV_PTE_ONE << HV_PTE_INDEX_GLOBAL)
+
+/** Is this mapping accessible to users?
+ *
+ * If this bit is set, code running at any PL will be permitted to
+ * access the virtual addresses mapped by this PTE.  Otherwise, only
+ * code running at PL 1 or above will be allowed to do so.
+ *
+ * This bit should not be modified by the client while PRESENT is set, as
+ * doing so may race with the hypervisor's update of ACCESSED and DIRTY bits.
+ *
+ * This bit is ignored in level-1 PTEs unless the Page bit is set.
+ */
+#define HV_PTE_USER                  (__HV_PTE_ONE << HV_PTE_INDEX_USER)
+
+/** Has this mapping been accessed?
+ *
+ * This bit is set by the hypervisor when the memory described by the
+ * translation is accessed for the first time.  It is never cleared by
+ * the hypervisor, but may be cleared by the client.  After the bit
+ * has been cleared, subsequent references are not guaranteed to set
+ * it again until the translation has been flushed from the TLB.
+ *
+ * This bit is ignored in level-1 PTEs unless the Page bit is set.
+ */
+#define HV_PTE_ACCESSED              (__HV_PTE_ONE << HV_PTE_INDEX_ACCESSED)
+
+/** Is this mapping dirty?
+ *
+ * This bit is set by the hypervisor when the memory described by the
+ * translation is written for the first time.  It is never cleared by
+ * the hypervisor, but may be cleared by the client.  After the bit
+ * has been cleared, subsequent references are not guaranteed to set
+ * it again until the translation has been flushed from the TLB.
+ *
+ * This bit is ignored in level-1 PTEs unless the Page bit is set.
+ */
+#define HV_PTE_DIRTY                 (__HV_PTE_ONE << HV_PTE_INDEX_DIRTY)
+
+/** Migrating bit in PTE.
+ *
+ * This bit is guaranteed not to be inspected or modified by the
+ * hypervisor.  The name is indicative of the suggested use by the client
+ * to tag pages whose L3 cache is being migrated from one cpu to another.
+ */
+#define HV_PTE_MIGRATING             (__HV_PTE_ONE << HV_PTE_INDEX_MIGRATING)
+
+/** Client-private bit in PTE.
+ *
+ * This bit is guaranteed not to be inspected or modified by the
+ * hypervisor.
+ */
+#define HV_PTE_CLIENT0               (__HV_PTE_ONE << HV_PTE_INDEX_CLIENT0)
+
+/** Client-private bit in PTE.
+ *
+ * This bit is guaranteed not to be inspected or modified by the
+ * hypervisor.
+ */
+#define HV_PTE_CLIENT1               (__HV_PTE_ONE << HV_PTE_INDEX_CLIENT1)
+
+/** Client-private bit in PTE.
+ *
+ * This bit is guaranteed not to be inspected or modified by the
+ * hypervisor.
+ */
+#define HV_PTE_CLIENT2               (__HV_PTE_ONE << HV_PTE_INDEX_CLIENT2)
+
+/** Non-coherent (NC) bit in PTE.
+ *
+ * If this bit is set, the mapping that is set up will be non-coherent
+ * (also known as non-inclusive).  This means that changes to the L3
+ * cache will not cause a local copy to be invalidated.  It is generally
+ * recommended only for read-only mappings.
+ *
+ * In level-1 PTEs, if the Page bit is clear, this bit determines how the
+ * level-2 page table is accessed.
+ */
+#define HV_PTE_NC                    (__HV_PTE_ONE << HV_PTE_INDEX_NC)
+
+/** Is this page prevented from filling the L1$?
+ *
+ * If this bit is set, the page described by the PTE will not be cached
+ * the local cpu's L1 cache.
+ *
+ * If CHIP_HAS_NC_AND_NOALLOC_BITS() is not true in <chip.h> for this chip,
+ * it is illegal to use this attribute, and may cause client termination.
+ *
+ * In level-1 PTEs, if the Page bit is clear, this bit
+ * determines how the level-2 page table is accessed.
+ */
+#define HV_PTE_NO_ALLOC_L1           (__HV_PTE_ONE << HV_PTE_INDEX_NO_ALLOC_L1)
+
+/** Is this page prevented from filling the L2$?
+ *
+ * If this bit is set, the page described by the PTE will not be cached
+ * the local cpu's L2 cache.
+ *
+ * If CHIP_HAS_NC_AND_NOALLOC_BITS() is not true in <chip.h> for this chip,
+ * it is illegal to use this attribute, and may cause client termination.
+ *
+ * In level-1 PTEs, if the Page bit is clear, this bit determines how the
+ * level-2 page table is accessed.
+ */
+#define HV_PTE_NO_ALLOC_L2           (__HV_PTE_ONE << HV_PTE_INDEX_NO_ALLOC_L2)
+
+/** Is this a priority page?
+ *
+ * If this bit is set, the page described by the PTE will be given
+ * priority in the cache.  Normally this translates into allowing the
+ * page to use only the "red" half of the cache.  The client may wish to
+ * then use the hv_set_caching service to specify that other pages which
+ * alias this page will use only the "black" half of the cache.
+ *
+ * If the Cached Priority bit is clear, the hypervisor uses the
+ * current hv_set_caching() value to choose how to cache the page.
+ *
+ * It is illegal to set the Cached Priority bit if the Non-Cached bit
+ * is set and the Cached Remotely bit is clear, i.e. if requests to
+ * the page map directly to memory.
+ *
+ * This bit is ignored in level-1 PTEs unless the Page bit is set.
+ */
+#define HV_PTE_CACHED_PRIORITY       (__HV_PTE_ONE << \
+                                      HV_PTE_INDEX_CACHED_PRIORITY)
+
+/** Is this a readable mapping?
+ *
+ * If this bit is set, code will be permitted to read from (e.g.,
+ * issue load instructions against) the virtual addresses mapped by
+ * this PTE.
+ *
+ * It is illegal for this bit to be clear if the Writable bit is set.
+ *
+ * This bit is ignored in level-1 PTEs unless the Page bit is set.
+ */
+#define HV_PTE_READABLE              (__HV_PTE_ONE << HV_PTE_INDEX_READABLE)
+
+/** Is this a writable mapping?
+ *
+ * If this bit is set, code will be permitted to write to (e.g., issue
+ * store instructions against) the virtual addresses mapped by this
+ * PTE.
+ *
+ * This bit is ignored in level-1 PTEs unless the Page bit is set.
+ */
+#define HV_PTE_WRITABLE              (__HV_PTE_ONE << HV_PTE_INDEX_WRITABLE)
+
+/** Is this an executable mapping?
+ *
+ * If this bit is set, code will be permitted to execute from
+ * (e.g., jump to) the virtual addresses mapped by this PTE.
+ *
+ * This bit applies to any processor on the tile, if there are more
+ * than one.
+ *
+ * This bit is ignored in level-1 PTEs unless the Page bit is set.
+ */
+#define HV_PTE_EXECUTABLE            (__HV_PTE_ONE << HV_PTE_INDEX_EXECUTABLE)
+
+/** The width of a LOTAR's x or y bitfield. */
+#define HV_LOTAR_WIDTH 11
+
+/** Converts an x,y pair to a LOTAR value. */
+#define HV_XY_TO_LOTAR(x, y) ((HV_LOTAR)(((x) << HV_LOTAR_WIDTH) | (y)))
+
+/** Extracts the X component of a lotar. */
+#define HV_LOTAR_X(lotar) ((lotar) >> HV_LOTAR_WIDTH)
+
+/** Extracts the Y component of a lotar. */
+#define HV_LOTAR_Y(lotar) ((lotar) & ((1 << HV_LOTAR_WIDTH) - 1))
+
+#ifndef __ASSEMBLER__
+
+/** Define accessor functions for a PTE bit. */
+#define _HV_BIT(name, bit)                                      \
+static __inline int                                             \
+hv_pte_get_##name(HV_PTE pte)                                   \
+{                                                               \
+  return (pte.val >> HV_PTE_INDEX_##bit) & 1;                   \
+}                                                               \
+                                                                \
+static __inline HV_PTE                                          \
+hv_pte_set_##name(HV_PTE pte)                                   \
+{                                                               \
+  pte.val |= 1ULL << HV_PTE_INDEX_##bit;                        \
+  return pte;                                                   \
+}                                                               \
+                                                                \
+static __inline HV_PTE                                          \
+hv_pte_clear_##name(HV_PTE pte)                                 \
+{                                                               \
+  pte.val &= ~(1ULL << HV_PTE_INDEX_##bit);                     \
+  return pte;                                                   \
+}
+
+/* Generate accessors to get, set, and clear various PTE flags.
+ */
+_HV_BIT(present,         PRESENT)
+_HV_BIT(page,            PAGE)
+_HV_BIT(super,           SUPER)
+_HV_BIT(client0,         CLIENT0)
+_HV_BIT(client1,         CLIENT1)
+_HV_BIT(client2,         CLIENT2)
+_HV_BIT(migrating,       MIGRATING)
+_HV_BIT(nc,              NC)
+_HV_BIT(readable,        READABLE)
+_HV_BIT(writable,        WRITABLE)
+_HV_BIT(executable,      EXECUTABLE)
+_HV_BIT(accessed,        ACCESSED)
+_HV_BIT(dirty,           DIRTY)
+_HV_BIT(no_alloc_l1,     NO_ALLOC_L1)
+_HV_BIT(no_alloc_l2,     NO_ALLOC_L2)
+_HV_BIT(cached_priority, CACHED_PRIORITY)
+_HV_BIT(global,          GLOBAL)
+_HV_BIT(user,            USER)
+
+#undef _HV_BIT
+
+/** Get the page mode from the PTE.
+ *
+ * This field generally determines whether and how accesses to the page
+ * are cached; the HV_PTE_MODE_xxx symbols define the legal values for the
+ * page mode.  The NC, NO_ALLOC_L1, and NO_ALLOC_L2 bits modify this
+ * general policy.
+ */
+static __inline unsigned int
+hv_pte_get_mode(const HV_PTE pte)
+{
+  return (((__hv32) pte.val) >> HV_PTE_INDEX_MODE) &
+         ((1 << HV_PTE_MODE_BITS) - 1);
+}
+
+/** Set the page mode into a PTE.  See hv_pte_get_mode. */
+static __inline HV_PTE
+hv_pte_set_mode(HV_PTE pte, unsigned int val)
+{
+  pte.val &= ~(((1ULL << HV_PTE_MODE_BITS) - 1) << HV_PTE_INDEX_MODE);
+  pte.val |= val << HV_PTE_INDEX_MODE;
+  return pte;
+}
+
+/** Get the page frame number from the PTE.
+ *
+ * This field contains the upper bits of the CPA (client physical
+ * address) of the target page; the complete CPA is this field with
+ * HV_LOG2_PAGE_TABLE_ALIGN zero bits appended to it.
+ *
+ * For all PTEs in the lowest-level page table, and for all PTEs with
+ * the Page bit set in all page tables, the CPA must be aligned modulo
+ * the relevant page size.
+ */
+static __inline unsigned long
+hv_pte_get_ptfn(const HV_PTE pte)
+{
+  return pte.val >> HV_PTE_INDEX_PTFN;
+}
+
+/** Set the page table frame number into a PTE.  See hv_pte_get_ptfn. */
+static __inline HV_PTE
+hv_pte_set_ptfn(HV_PTE pte, unsigned long val)
+{
+  pte.val &= ~(((1ULL << HV_PTE_PTFN_BITS)-1) << HV_PTE_INDEX_PTFN);
+  pte.val |= (__hv64) val << HV_PTE_INDEX_PTFN;
+  return pte;
+}
+
+/** Get the client physical address from the PTE.  See hv_pte_set_ptfn. */
+static __inline HV_PhysAddr
+hv_pte_get_pa(const HV_PTE pte)
+{
+  return (__hv64) hv_pte_get_ptfn(pte) << HV_LOG2_PAGE_TABLE_ALIGN;
+}
+
+/** Set the client physical address into a PTE.  See hv_pte_get_ptfn. */
+static __inline HV_PTE
+hv_pte_set_pa(HV_PTE pte, HV_PhysAddr pa)
+{
+  return hv_pte_set_ptfn(pte, pa >> HV_LOG2_PAGE_TABLE_ALIGN);
+}
+
+
+/** Get the remote tile caching this page.
+ *
+ * Specifies the remote tile which is providing the L3 cache for this page.
+ *
+ * This field is ignored unless the page mode is HV_PTE_MODE_CACHE_TILE_L3.
+ *
+ * In level-1 PTEs, if the Page bit is clear, this field determines how the
+ * level-2 page table is accessed.
+ */
+static __inline unsigned int
+hv_pte_get_lotar(const HV_PTE pte)
+{
+  unsigned int lotar = ((__hv32) pte.val) >> HV_PTE_INDEX_LOTAR;
+
+  return HV_XY_TO_LOTAR( (lotar >> (HV_PTE_LOTAR_BITS / 2)),
+                         (lotar & ((1 << (HV_PTE_LOTAR_BITS / 2)) - 1)) );
+}
+
+
+/** Set the remote tile caching a page into a PTE.  See hv_pte_get_lotar. */
+static __inline HV_PTE
+hv_pte_set_lotar(HV_PTE pte, unsigned int val)
+{
+  unsigned int x = HV_LOTAR_X(val);
+  unsigned int y = HV_LOTAR_Y(val);
+
+  pte.val &= ~(((1ULL << HV_PTE_LOTAR_BITS)-1) << HV_PTE_INDEX_LOTAR);
+  pte.val |= (x << (HV_PTE_INDEX_LOTAR + HV_PTE_LOTAR_BITS / 2)) |
+             (y << HV_PTE_INDEX_LOTAR);
+  return pte;
+}
+
+#endif  /* !__ASSEMBLER__ */
+
+/** Converts a client physical address to a ptfn. */
+#define HV_CPA_TO_PTFN(p) ((p) >> HV_LOG2_PAGE_TABLE_ALIGN)
+
+/** Converts a ptfn to a client physical address. */
+#define HV_PTFN_TO_CPA(p) (((HV_PhysAddr)(p)) << HV_LOG2_PAGE_TABLE_ALIGN)
+
+#if CHIP_VA_WIDTH() > 32
+
+/*
+ * Note that we currently do not allow customizing the page size
+ * of the L0 pages, but fix them at 4GB, so we do not use the
+ * "_HV_xxx" nomenclature for the L0 macros.
+ */
+
+/** Log number of HV_PTE entries in L0 page table */
+#define HV_LOG2_L0_ENTRIES (CHIP_VA_WIDTH() - HV_LOG2_L1_SPAN)
+
+/** Number of HV_PTE entries in L0 page table */
+#define HV_L0_ENTRIES (1 << HV_LOG2_L0_ENTRIES)
+
+/** Log size of L0 page table in bytes */
+#define HV_LOG2_L0_SIZE (HV_LOG2_PTE_SIZE + HV_LOG2_L0_ENTRIES)
+
+/** Size of L0 page table in bytes */
+#define HV_L0_SIZE (1 << HV_LOG2_L0_SIZE)
+
+#ifdef __ASSEMBLER__
+
+/** Index in L0 for a specific VA */
+#define HV_L0_INDEX(va) \
+  (((va) >> HV_LOG2_L1_SPAN) & (HV_L0_ENTRIES - 1))
+
+#else
+
+/** Index in L1 for a specific VA */
+#define HV_L0_INDEX(va) \
+  (((HV_VirtAddr)(va) >> HV_LOG2_L1_SPAN) & (HV_L0_ENTRIES - 1))
+
+#endif
+
+#endif /* CHIP_VA_WIDTH() > 32 */
+
+/** Log number of HV_PTE entries in L1 page table */
+#define _HV_LOG2_L1_ENTRIES(log2_page_size_large) \
+  (HV_LOG2_L1_SPAN - log2_page_size_large)
+
+/** Number of HV_PTE entries in L1 page table */
+#define _HV_L1_ENTRIES(log2_page_size_large) \
+  (1 << _HV_LOG2_L1_ENTRIES(log2_page_size_large))
+
+/** Log size of L1 page table in bytes */
+#define _HV_LOG2_L1_SIZE(log2_page_size_large) \
+  (HV_LOG2_PTE_SIZE + _HV_LOG2_L1_ENTRIES(log2_page_size_large))
+
+/** Size of L1 page table in bytes */
+#define _HV_L1_SIZE(log2_page_size_large) \
+  (1 << _HV_LOG2_L1_SIZE(log2_page_size_large))
+
+/** Log number of HV_PTE entries in level-2 page table */
+#define _HV_LOG2_L2_ENTRIES(log2_page_size_large, log2_page_size_small) \
+  (log2_page_size_large - log2_page_size_small)
+
+/** Number of HV_PTE entries in level-2 page table */
+#define _HV_L2_ENTRIES(log2_page_size_large, log2_page_size_small) \
+  (1 << _HV_LOG2_L2_ENTRIES(log2_page_size_large, log2_page_size_small))
+
+/** Log size of level-2 page table in bytes */
+#define _HV_LOG2_L2_SIZE(log2_page_size_large, log2_page_size_small) \
+  (HV_LOG2_PTE_SIZE + \
+   _HV_LOG2_L2_ENTRIES(log2_page_size_large, log2_page_size_small))
+
+/** Size of level-2 page table in bytes */
+#define _HV_L2_SIZE(log2_page_size_large, log2_page_size_small) \
+  (1 << _HV_LOG2_L2_SIZE(log2_page_size_large, log2_page_size_small))
+
+#ifdef __ASSEMBLER__
+
+#if CHIP_VA_WIDTH() > 32
+
+/** Index in L1 for a specific VA */
+#define _HV_L1_INDEX(va, log2_page_size_large) \
+  (((va) >> log2_page_size_large) & (_HV_L1_ENTRIES(log2_page_size_large) - 1))
+
+#else /* CHIP_VA_WIDTH() > 32 */
+
+/** Index in L1 for a specific VA */
+#define _HV_L1_INDEX(va, log2_page_size_large) \
+  (((va) >> log2_page_size_large))
+
+#endif /* CHIP_VA_WIDTH() > 32 */
+
+/** Index in level-2 page table for a specific VA */
+#define _HV_L2_INDEX(va, log2_page_size_large, log2_page_size_small) \
+  (((va) >> log2_page_size_small) & \
+   (_HV_L2_ENTRIES(log2_page_size_large, log2_page_size_small) - 1))
+
+#else /* __ASSEMBLER __ */
+
+#if CHIP_VA_WIDTH() > 32
+
+/** Index in L1 for a specific VA */
+#define _HV_L1_INDEX(va, log2_page_size_large) \
+  (((HV_VirtAddr)(va) >> log2_page_size_large) & \
+   (_HV_L1_ENTRIES(log2_page_size_large) - 1))
+
+#else /* CHIP_VA_WIDTH() > 32 */
+
+/** Index in L1 for a specific VA */
+#define _HV_L1_INDEX(va, log2_page_size_large) \
+  (((HV_VirtAddr)(va) >> log2_page_size_large))
+
+#endif /* CHIP_VA_WIDTH() > 32 */
+
+/** Index in level-2 page table for a specific VA */
+#define _HV_L2_INDEX(va, log2_page_size_large, log2_page_size_small) \
+  (((HV_VirtAddr)(va) >> log2_page_size_small) & \
+   (_HV_L2_ENTRIES(log2_page_size_large, log2_page_size_small) - 1))
+
+#endif /* __ASSEMBLER __ */
+
+/** Position of the PFN field within the PTE (subset of the PTFN). */
+#define _HV_PTE_INDEX_PFN(log2_page_size) \
+  (HV_PTE_INDEX_PTFN + (log2_page_size - HV_LOG2_PAGE_TABLE_ALIGN))
+
+/** Length of the PFN field within the PTE (subset of the PTFN). */
+#define _HV_PTE_INDEX_PFN_BITS(log2_page_size) \
+  (HV_PTE_INDEX_PTFN_BITS - (log2_page_size - HV_LOG2_PAGE_TABLE_ALIGN))
+
+/** Converts a client physical address to a pfn. */
+#define _HV_CPA_TO_PFN(p, log2_page_size) ((p) >> log2_page_size)
+
+/** Converts a pfn to a client physical address. */
+#define _HV_PFN_TO_CPA(p, log2_page_size) \
+  (((HV_PhysAddr)(p)) << log2_page_size)
+
+/** Converts a ptfn to a pfn. */
+#define _HV_PTFN_TO_PFN(p, log2_page_size) \
+  ((p) >> (log2_page_size - HV_LOG2_PAGE_TABLE_ALIGN))
+
+/** Converts a pfn to a ptfn. */
+#define _HV_PFN_TO_PTFN(p, log2_page_size) \
+  ((p) << (log2_page_size - HV_LOG2_PAGE_TABLE_ALIGN))
+
+#endif /* _HV_HV_H */
diff --git a/kernel/arch/tile/include/hv/iorpc.h b/kernel/arch/tile/include/hv/iorpc.h
new file mode 100644
index 000000000..ddf160448
--- /dev/null
+++ b/kernel/arch/tile/include/hv/iorpc.h
@@ -0,0 +1,714 @@
+/*
+ * Copyright 2012 Tilera Corporation. All Rights Reserved.
+ *
+ *   This program is free software; you can redistribute it and/or
+ *   modify it under the terms of the GNU General Public License
+ *   as published by the Free Software Foundation, version 2.
+ *
+ *   This program is distributed in the hope that it will be useful, but
+ *   WITHOUT ANY WARRANTY; without even the implied warranty of
+ *   MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
+ *   NON INFRINGEMENT.  See the GNU General Public License for
+ *   more details.
+ */
+#ifndef _HV_IORPC_H_
+#define _HV_IORPC_H_
+
+/**
+ *
+ * Error codes and struct definitions for the IO RPC library.
+ *
+ * The hypervisor's IO RPC component provides a convenient way for
+ * driver authors to proxy system calls between user space, linux, and
+ * the hypervisor driver.  The core of the system is a set of Python
+ * files that take ".idl" files as input and generates the following
+ * source code:
+ *
+ * - _rpc_call() routines for use in userspace IO libraries.  These
+ * routines take an argument list specified in the .idl file, pack the
+ * arguments in to a buffer, and read or write that buffer via the
+ * Linux iorpc driver.
+ *
+ * - dispatch_read() and dispatch_write() routines that hypervisor
+ * drivers can use to implement most of their dev_pread() and
+ * dev_pwrite() methods.  These routines decode the incoming parameter
+ * blob, permission check and translate parameters where appropriate,
+ * and then invoke a callback routine for whichever RPC call has
+ * arrived.  The driver simply implements the set of callback
+ * routines.
+ *
+ * The IO RPC system also includes the Linux 'iorpc' driver, which
+ * proxies calls between the userspace library and the hypervisor
+ * driver.  The Linux driver is almost entirely device agnostic; it
+ * watches for special flags indicating cases where a memory buffer
+ * address might need to be translated, etc.  As a result, driver
+ * writers can avoid many of the problem cases related to registering
+ * hardware resources like memory pages or interrupts.  However, the
+ * drivers must be careful to obey the conventions documented below in
+ * order to work properly with the generic Linux iorpc driver.
+ *
+ * @section iorpc_domains Service Domains
+ *
+ * All iorpc-based drivers must support a notion of service domains.
+ * A service domain is basically an application context - state
+ * indicating resources that are allocated to that particular app
+ * which it may access and (perhaps) other applications may not
+ * access.  Drivers can support any number of service domains they
+ * choose.  In some cases the design is limited by a number of service
+ * domains supported by the IO hardware; in other cases the service
+ * domains are a purely software concept and the driver chooses a
+ * maximum number of domains based on how much state memory it is
+ * willing to preallocate.
+ *
+ * For example, the mPIPE driver only supports as many service domains
+ * as are supported by the mPIPE hardware.  This limitation is
+ * required because the hardware implements its own MMIO protection
+ * scheme to allow large MMIO mappings while still protecting small
+ * register ranges within the page that should only be accessed by the
+ * hypervisor.
+ *
+ * In contrast, drivers with no hardware service domain limitations
+ * (for instance the TRIO shim) can implement an arbitrary number of
+ * service domains.  In these cases, each service domain is limited to
+ * a carefully restricted set of legal MMIO addresses if necessary to
+ * keep one application from corrupting another application's state.
+ *
+ * @section iorpc_conventions System Call Conventions
+ *
+ * The driver's open routine is responsible for allocating a new
+ * service domain for each hv_dev_open() call.  By convention, the
+ * return value from open() should be the service domain number on
+ * success, or GXIO_ERR_NO_SVC_DOM if no more service domains are
+ * available.
+ *
+ * The implementations of hv_dev_pread() and hv_dev_pwrite() are
+ * responsible for validating the devhdl value passed up by the
+ * client.  Since the device handle returned by hv_dev_open() should
+ * embed the positive service domain number, drivers should make sure
+ * that DRV_HDL2BITS(devhdl) is a legal service domain.  If the client
+ * passes an illegal service domain number, the routine should return
+ * GXIO_ERR_INVAL_SVC_DOM.  Once the service domain number has been
+ * validated, the driver can copy to/from the client buffer and call
+ * the dispatch_read() or dispatch_write() methods created by the RPC
+ * generator.
+ *
+ * The hv_dev_close() implementation should reset all service domain
+ * state and put the service domain back on a free list for
+ * reallocation by a future application.  In most cases, this will
+ * require executing a hardware reset or drain flow and denying any
+ * MMIO regions that were created for the service domain.
+ *
+ * @section iorpc_data Special Data Types
+ *
+ * The .idl file syntax allows the creation of syscalls with special
+ * parameters that require permission checks or translations as part
+ * of the system call path.  Because of limitations in the code
+ * generator, APIs are generally limited to just one of these special
+ * parameters per system call, and they are sometimes required to be
+ * the first or last parameter to the call.  Special parameters
+ * include:
+ *
+ * @subsection iorpc_mem_buffer MEM_BUFFER
+ *
+ * The MEM_BUFFER() datatype allows user space to "register" memory
+ * buffers with a device.  Registering memory accomplishes two tasks:
+ * Linux keeps track of all buffers that might be modified by a
+ * hardware device, and the hardware device drivers bind registered
+ * buffers to particular hardware resources like ingress NotifRings.
+ * The MEM_BUFFER() idl syntax can take extra flags like ALIGN_64KB,
+ * ALIGN_SELF_SIZE, and FLAGS indicating that memory buffers must have
+ * certain alignment or that the user should be able to pass a "memory
+ * flags" word specifying attributes like nt_hint or IO cache pinning.
+ * The parser will accept multiple MEM_BUFFER() flags.
+ *
+ * Implementations must obey the following conventions when
+ * registering memory buffers via the iorpc flow.  These rules are a
+ * result of the Linux driver implementation, which needs to keep
+ * track of how many times a particular page has been registered with
+ * the hardware so that it can release the page when all those
+ * registrations are cleared.
+ *
+ * - Memory registrations that refer to a resource which has already
+ * been bound must return GXIO_ERR_ALREADY_INIT.  Thus, it is an
+ * error to register memory twice without resetting (i.e. closing) the
+ * resource in between.  This convention keeps the Linux driver from
+ * having to track which particular devices a page is bound to.
+ *
+ * - At present, a memory registration is only cleared when the
+ * service domain is reset.  In this case, the Linux driver simply
+ * closes the HV device file handle and then decrements the reference
+ * counts of all pages that were previously registered with the
+ * device.
+ *
+ * - In the future, we may add a mechanism for unregistering memory.
+ * One possible implementation would require that the user specify
+ * which buffer is currently registered.  The HV would then verify
+ * that that page was actually the one currently mapped and return
+ * success or failure to Linux, which would then only decrement the
+ * page reference count if the addresses were mapped.  Another scheme
+ * might allow Linux to pass a token to the HV to be returned when the
+ * resource is unmapped.
+ *
+ * @subsection iorpc_interrupt INTERRUPT
+ *
+ * The INTERRUPT .idl datatype allows the client to bind hardware
+ * interrupts to a particular combination of IPI parameters - CPU, IPI
+ * PL, and event bit number.  This data is passed via a special
+ * datatype so that the Linux driver can validate the CPU and PL and
+ * the HV generic iorpc code can translate client CPUs to real CPUs.
+ *
+ * @subsection iorpc_pollfd_setup POLLFD_SETUP
+ *
+ * The POLLFD_SETUP .idl datatype allows the client to set up hardware
+ * interrupt bindings which are received by Linux but which are made
+ * visible to user processes as state transitions on a file descriptor;
+ * this allows user processes to use Linux primitives, such as poll(), to
+ * await particular hardware events.  This data is passed via a special
+ * datatype so that the Linux driver may recognize the pollable file
+ * descriptor and translate it to a set of interrupt target information,
+ * and so that the HV generic iorpc code can translate client CPUs to real
+ * CPUs.
+ *
+ * @subsection iorpc_pollfd POLLFD
+ *
+ * The POLLFD .idl datatype allows manipulation of hardware interrupt
+ * bindings set up via the POLLFD_SETUP datatype; common operations are
+ * resetting the state of the requested interrupt events, and unbinding any
+ * bound interrupts.  This data is passed via a special datatype so that
+ * the Linux driver may recognize the pollable file descriptor and
+ * translate it to an interrupt identifier previously supplied by the
+ * hypervisor as the result of an earlier pollfd_setup operation.
+ *
+ * @subsection iorpc_blob BLOB
+ *
+ * The BLOB .idl datatype allows the client to write an arbitrary
+ * length string of bytes up to the hypervisor driver.  This can be
+ * useful for passing up large, arbitrarily structured data like
+ * classifier programs.  The iorpc stack takes care of validating the
+ * buffer VA and CPA as the data passes up to the hypervisor.  Unlike
+ * MEM_BUFFER(), the buffer is not registered - Linux does not bump
+ * page refcounts and the HV driver should not reuse the buffer once
+ * the system call is complete.
+ *
+ * @section iorpc_translation Translating User Space Calls
+ *
+ * The ::iorpc_offset structure describes the formatting of the offset
+ * that is passed to pread() or pwrite() as part of the generated RPC code.
+ * When the user calls up to Linux, the rpc code fills in all the fields of
+ * the offset, including a 16-bit opcode, a 16 bit format indicator, and 32
+ * bits of user-specified "sub-offset".  The opcode indicates which syscall
+ * is being requested.  The format indicates whether there is a "prefix
+ * struct" at the start of the memory buffer passed to pwrite(), and if so
+ * what data is in that prefix struct.  These prefix structs are used to
+ * implement special datatypes like MEM_BUFFER() and INTERRUPT - we arrange
+ * to put data that needs translation and permission checks at the start of
+ * the buffer so that the Linux driver and generic portions of the HV iorpc
+ * code can easily access the data.  The 32 bits of user-specified
+ * "sub-offset" are most useful for pread() calls where the user needs to
+ * also pass in a few bits indicating which register to read, etc.
+ *
+ * The Linux iorpc driver watches for system calls that contain prefix
+ * structs so that it can translate parameters and bump reference
+ * counts as appropriate.  It does not (currently) have any knowledge
+ * of the per-device opcodes - it doesn't care what operation you're
+ * doing to mPIPE, so long as it can do all the generic book-keeping.
+ * The hv/iorpc.h header file defines all of the generic encoding bits
+ * needed to translate iorpc calls without knowing which particular
+ * opcode is being issued.
+ *
+ * @section iorpc_globals Global iorpc Calls
+ *
+ * Implementing mmap() required adding some special iorpc syscalls
+ * that are only called by the Linux driver, never by userspace.
+ * These include get_mmio_base() and check_mmio_offset().  These
+ * routines are described in globals.idl and must be included in every
+ * iorpc driver.  By providing these routines in every driver, Linux's
+ * mmap implementation can easily get the PTE bits it needs and
+ * validate the PA offset without needing to know the per-device
+ * opcodes to perform those tasks.
+ *
+ * @section iorpc_kernel Supporting gxio APIs in the Kernel
+ *
+ * The iorpc code generator also supports generation of kernel code
+ * implementing the gxio APIs.  This capability is currently used by
+ * the mPIPE network driver, and will likely be used by the TRIO root
+ * complex and endpoint drivers and perhaps an in-kernel crypto
+ * driver.  Each driver that wants to instantiate iorpc calls in the
+ * kernel needs to generate a kernel version of the generate rpc code
+ * and (probably) copy any related gxio source files into the kernel.
+ * The mPIPE driver provides a good example of this pattern.
+ */
+
+#ifdef __KERNEL__
+#include <linux/stddef.h>
+#else
+#include <stddef.h>
+#endif
+
+#if defined(__HV__)
+#include <hv/hypervisor.h>
+#elif defined(__KERNEL__)
+#include <hv/hypervisor.h>
+#include <linux/types.h>
+#else
+#include <stdint.h>
+#endif
+
+
+/** Code indicating translation services required within the RPC path.
+ * These indicate whether there is a translatable struct at the start
+ * of the RPC buffer and what information that struct contains.
+ */
+enum iorpc_format_e
+{
+  /** No translation required, no prefix struct. */
+  IORPC_FORMAT_NONE,
+
+  /** No translation required, no prefix struct, no access to this
+   *  operation from user space. */
+  IORPC_FORMAT_NONE_NOUSER,
+
+  /** Prefix struct contains user VA and size. */
+  IORPC_FORMAT_USER_MEM,
+
+  /** Prefix struct contains CPA, size, and homing bits. */
+  IORPC_FORMAT_KERNEL_MEM,
+
+  /** Prefix struct contains interrupt. */
+  IORPC_FORMAT_KERNEL_INTERRUPT,
+
+  /** Prefix struct contains user-level interrupt. */
+  IORPC_FORMAT_USER_INTERRUPT,
+
+  /** Prefix struct contains pollfd_setup (interrupt information). */
+  IORPC_FORMAT_KERNEL_POLLFD_SETUP,
+
+  /** Prefix struct contains user-level pollfd_setup (file descriptor). */
+  IORPC_FORMAT_USER_POLLFD_SETUP,
+
+  /** Prefix struct contains pollfd (interrupt cookie). */
+  IORPC_FORMAT_KERNEL_POLLFD,
+
+  /** Prefix struct contains user-level pollfd (file descriptor). */
+  IORPC_FORMAT_USER_POLLFD,
+};
+
+
+/** Generate an opcode given format and code. */
+#define IORPC_OPCODE(FORMAT, CODE) (((FORMAT) << 16) | (CODE))
+
+/** The offset passed through the read() and write() system calls
+    combines an opcode with 32 bits of user-specified offset. */
+union iorpc_offset
+{
+#ifndef __BIG_ENDIAN__
+  uint64_t offset;              /**< All bits. */
+
+  struct
+  {
+    uint16_t code;              /**< RPC code. */
+    uint16_t format;            /**< iorpc_format_e */
+    uint32_t sub_offset;        /**< caller-specified offset. */
+  };
+
+  uint32_t opcode;              /**< Opcode combines code & format. */
+#else
+  uint64_t offset;              /**< All bits. */
+
+  struct
+  {
+    uint32_t sub_offset;        /**< caller-specified offset. */
+    uint16_t format;            /**< iorpc_format_e */
+    uint16_t code;              /**< RPC code. */
+  };
+
+  struct
+  {
+    uint32_t padding;
+    uint32_t opcode;              /**< Opcode combines code & format. */
+  };
+#endif
+};
+
+
+/** Homing and cache hinting bits that can be used by IO devices. */
+struct iorpc_mem_attr
+{
+  unsigned int lotar_x:4;       /**< lotar X bits (or Gx page_mask). */
+  unsigned int lotar_y:4;       /**< lotar Y bits (or Gx page_offset). */
+  unsigned int hfh:1;           /**< Uses hash-for-home. */
+  unsigned int nt_hint:1;       /**< Non-temporal hint. */
+  unsigned int io_pin:1;        /**< Only fill 'IO' cache ways. */
+};
+
+/** Set the nt_hint bit. */
+#define IORPC_MEM_BUFFER_FLAG_NT_HINT (1 << 0)
+
+/** Set the IO pin bit. */
+#define IORPC_MEM_BUFFER_FLAG_IO_PIN (1 << 1)
+
+
+/** A structure used to describe memory registration.  Different
+    protection levels describe memory differently, so this union
+    contains all the different possible descriptions.  As a request
+    moves up the call chain, each layer translates from one
+    description format to the next.  In particular, the Linux iorpc
+    driver translates user VAs into CPAs and homing parameters. */
+union iorpc_mem_buffer
+{
+  struct
+  {
+    uint64_t va;                /**< User virtual address. */
+    uint64_t size;              /**< Buffer size. */
+    unsigned int flags;         /**< nt_hint, IO pin. */
+  }
+  user;                         /**< Buffer as described by user apps. */
+
+  struct
+  {
+    unsigned long long cpa;     /**< Client physical address. */
+#if defined(__KERNEL__) || defined(__HV__)
+    size_t size;                /**< Buffer size. */
+    HV_PTE pte;                 /**< PTE describing memory homing. */
+#else
+    uint64_t size;
+    uint64_t pte;
+#endif
+    unsigned int flags;         /**< nt_hint, IO pin. */
+  }
+  kernel;                       /**< Buffer as described by kernel. */
+
+  struct
+  {
+    unsigned long long pa;      /**< Physical address. */
+    size_t size;                /**< Buffer size. */
+    struct iorpc_mem_attr attr;      /**< Homing and locality hint bits. */
+  }
+  hv;                           /**< Buffer parameters for HV driver. */
+};
+
+
+/** A structure used to describe interrupts.  The format differs slightly
+ *  for user and kernel interrupts.  As with the mem_buffer_t, translation
+ *  between the formats is done at each level. */
+union iorpc_interrupt
+{
+  struct
+  {
+    int cpu;   /**< CPU. */
+    int event; /**< evt_num */
+  }
+  user;        /**< Interrupt as described by user applications. */
+
+  struct
+  {
+    int x;     /**< X coord. */
+    int y;     /**< Y coord. */
+    int ipi;   /**< int_num */
+    int event; /**< evt_num */
+  }
+  kernel;      /**< Interrupt as described by the kernel. */
+
+};
+
+
+/** A structure used to describe interrupts used with poll().  The format
+ *  differs significantly for requests from user to kernel, and kernel to
+ *  hypervisor.  As with the mem_buffer_t, translation between the formats
+ *  is done at each level. */
+union iorpc_pollfd_setup
+{
+  struct
+  {
+    int fd;    /**< Pollable file descriptor. */
+  }
+  user;        /**< pollfd_setup as described by user applications. */
+
+  struct
+  {
+    int x;     /**< X coord. */
+    int y;     /**< Y coord. */
+    int ipi;   /**< int_num */
+    int event; /**< evt_num */
+  }
+  kernel;      /**< pollfd_setup as described by the kernel. */
+
+};
+
+
+/** A structure used to describe previously set up interrupts used with
+ *  poll().  The format differs significantly for requests from user to
+ *  kernel, and kernel to hypervisor.  As with the mem_buffer_t, translation
+ *  between the formats is done at each level. */
+union iorpc_pollfd
+{
+  struct
+  {
+    int fd;    /**< Pollable file descriptor. */
+  }
+  user;        /**< pollfd as described by user applications. */
+
+  struct
+  {
+    int cookie; /**< hv cookie returned by the pollfd_setup operation. */
+  }
+  kernel;      /**< pollfd as described by the kernel. */
+
+};
+
+
+/** The various iorpc devices use error codes from -1100 to -1299.
+ *
+ * This range is distinct from netio (-700 to -799), the hypervisor
+ * (-800 to -899), tilepci (-900 to -999), ilib (-1000 to -1099),
+ * gxcr (-1300 to -1399) and gxpci (-1400 to -1499).
+ */
+enum gxio_err_e {
+
+  /** Largest iorpc error number. */
+  GXIO_ERR_MAX = -1101,
+
+
+  /********************************************************/
+  /*                   Generic Error Codes                */
+  /********************************************************/
+
+  /** Bad RPC opcode - possible version incompatibility. */
+  GXIO_ERR_OPCODE = -1101,
+
+  /** Invalid parameter. */
+  GXIO_ERR_INVAL = -1102,
+
+  /** Memory buffer did not meet alignment requirements. */
+  GXIO_ERR_ALIGNMENT = -1103,
+
+  /** Memory buffers must be coherent and cacheable. */
+  GXIO_ERR_COHERENCE = -1104,
+
+  /** Resource already initialized. */
+  GXIO_ERR_ALREADY_INIT = -1105,
+
+  /** No service domains available. */
+  GXIO_ERR_NO_SVC_DOM = -1106,
+
+  /** Illegal service domain number. */
+  GXIO_ERR_INVAL_SVC_DOM = -1107,
+
+  /** Illegal MMIO address. */
+  GXIO_ERR_MMIO_ADDRESS = -1108,
+
+  /** Illegal interrupt binding. */
+  GXIO_ERR_INTERRUPT = -1109,
+
+  /** Unreasonable client memory. */
+  GXIO_ERR_CLIENT_MEMORY = -1110,
+
+  /** No more IOTLB entries. */
+  GXIO_ERR_IOTLB_ENTRY = -1111,
+
+  /** Invalid memory size. */
+  GXIO_ERR_INVAL_MEMORY_SIZE = -1112,
+
+  /** Unsupported operation. */
+  GXIO_ERR_UNSUPPORTED_OP = -1113,
+
+  /** Insufficient DMA credits. */
+  GXIO_ERR_DMA_CREDITS = -1114,
+
+  /** Operation timed out. */
+  GXIO_ERR_TIMEOUT = -1115,
+
+  /** No such device or object. */
+  GXIO_ERR_NO_DEVICE = -1116,
+
+  /** Device or resource busy. */
+  GXIO_ERR_BUSY = -1117,
+
+  /** I/O error. */
+  GXIO_ERR_IO = -1118,
+
+  /** Permissions error. */
+  GXIO_ERR_PERM = -1119,
+
+
+
+  /********************************************************/
+  /*                 Test Device Error Codes              */
+  /********************************************************/
+
+  /** Illegal register number. */
+  GXIO_TEST_ERR_REG_NUMBER = -1120,
+
+  /** Illegal buffer slot. */
+  GXIO_TEST_ERR_BUFFER_SLOT = -1121,
+
+
+  /********************************************************/
+  /*                    MPIPE Error Codes                 */
+  /********************************************************/
+
+
+  /** Invalid buffer size. */
+  GXIO_MPIPE_ERR_INVAL_BUFFER_SIZE = -1131,
+
+  /** Cannot allocate buffer stack. */
+  GXIO_MPIPE_ERR_NO_BUFFER_STACK = -1140,
+
+  /** Invalid buffer stack number. */
+  GXIO_MPIPE_ERR_BAD_BUFFER_STACK = -1141,
+
+  /** Cannot allocate NotifRing. */
+  GXIO_MPIPE_ERR_NO_NOTIF_RING = -1142,
+
+  /** Invalid NotifRing number. */
+  GXIO_MPIPE_ERR_BAD_NOTIF_RING = -1143,
+
+  /** Cannot allocate NotifGroup. */
+  GXIO_MPIPE_ERR_NO_NOTIF_GROUP = -1144,
+
+  /** Invalid NotifGroup number. */
+  GXIO_MPIPE_ERR_BAD_NOTIF_GROUP = -1145,
+
+  /** Cannot allocate bucket. */
+  GXIO_MPIPE_ERR_NO_BUCKET = -1146,
+
+  /** Invalid bucket number. */
+  GXIO_MPIPE_ERR_BAD_BUCKET = -1147,
+
+  /** Cannot allocate eDMA ring. */
+  GXIO_MPIPE_ERR_NO_EDMA_RING = -1148,
+
+  /** Invalid eDMA ring number. */
+  GXIO_MPIPE_ERR_BAD_EDMA_RING = -1149,
+
+  /** Invalid channel number. */
+  GXIO_MPIPE_ERR_BAD_CHANNEL = -1150,
+
+  /** Bad configuration. */
+  GXIO_MPIPE_ERR_BAD_CONFIG = -1151,
+
+  /** Empty iqueue. */
+  GXIO_MPIPE_ERR_IQUEUE_EMPTY = -1152,
+
+  /** Empty rules. */
+  GXIO_MPIPE_ERR_RULES_EMPTY = -1160,
+
+  /** Full rules. */
+  GXIO_MPIPE_ERR_RULES_FULL = -1161,
+
+  /** Corrupt rules. */
+  GXIO_MPIPE_ERR_RULES_CORRUPT = -1162,
+
+  /** Invalid rules. */
+  GXIO_MPIPE_ERR_RULES_INVALID = -1163,
+
+  /** Classifier is too big. */
+  GXIO_MPIPE_ERR_CLASSIFIER_TOO_BIG = -1170,
+
+  /** Classifier is too complex. */
+  GXIO_MPIPE_ERR_CLASSIFIER_TOO_COMPLEX = -1171,
+
+  /** Classifier has bad header. */
+  GXIO_MPIPE_ERR_CLASSIFIER_BAD_HEADER = -1172,
+
+  /** Classifier has bad contents. */
+  GXIO_MPIPE_ERR_CLASSIFIER_BAD_CONTENTS = -1173,
+
+  /** Classifier encountered invalid symbol. */
+  GXIO_MPIPE_ERR_CLASSIFIER_INVAL_SYMBOL = -1174,
+
+  /** Classifier encountered invalid bounds. */
+  GXIO_MPIPE_ERR_CLASSIFIER_INVAL_BOUNDS = -1175,
+
+  /** Classifier encountered invalid relocation. */
+  GXIO_MPIPE_ERR_CLASSIFIER_INVAL_RELOCATION = -1176,
+
+  /** Classifier encountered undefined symbol. */
+  GXIO_MPIPE_ERR_CLASSIFIER_UNDEF_SYMBOL = -1177,
+
+
+  /********************************************************/
+  /*                    TRIO  Error Codes                 */
+  /********************************************************/
+
+  /** Cannot allocate memory map region. */
+  GXIO_TRIO_ERR_NO_MEMORY_MAP = -1180,
+
+  /** Invalid memory map region number. */
+  GXIO_TRIO_ERR_BAD_MEMORY_MAP = -1181,
+
+  /** Cannot allocate scatter queue. */
+  GXIO_TRIO_ERR_NO_SCATTER_QUEUE = -1182,
+
+  /** Invalid scatter queue number. */
+  GXIO_TRIO_ERR_BAD_SCATTER_QUEUE = -1183,
+
+  /** Cannot allocate push DMA ring. */
+  GXIO_TRIO_ERR_NO_PUSH_DMA_RING = -1184,
+
+  /** Invalid push DMA ring index. */
+  GXIO_TRIO_ERR_BAD_PUSH_DMA_RING = -1185,
+
+  /** Cannot allocate pull DMA ring. */
+  GXIO_TRIO_ERR_NO_PULL_DMA_RING = -1186,
+
+  /** Invalid pull DMA ring index. */
+  GXIO_TRIO_ERR_BAD_PULL_DMA_RING = -1187,
+
+  /** Cannot allocate PIO region. */
+  GXIO_TRIO_ERR_NO_PIO = -1188,
+
+  /** Invalid PIO region index. */
+  GXIO_TRIO_ERR_BAD_PIO = -1189,
+
+  /** Cannot allocate ASID. */
+  GXIO_TRIO_ERR_NO_ASID = -1190,
+
+  /** Invalid ASID. */
+  GXIO_TRIO_ERR_BAD_ASID = -1191,
+
+
+  /********************************************************/
+  /*                    MICA Error Codes                  */
+  /********************************************************/
+
+  /** No such accelerator type. */
+  GXIO_MICA_ERR_BAD_ACCEL_TYPE = -1220,
+
+  /** Cannot allocate context. */
+  GXIO_MICA_ERR_NO_CONTEXT = -1221,
+
+  /** PKA command queue is full, can't add another command. */
+  GXIO_MICA_ERR_PKA_CMD_QUEUE_FULL = -1222,
+
+  /** PKA result queue is empty, can't get a result from the queue. */
+  GXIO_MICA_ERR_PKA_RESULT_QUEUE_EMPTY = -1223,
+
+  /********************************************************/
+  /*                    GPIO Error Codes                  */
+  /********************************************************/
+
+  /** Pin not available.  Either the physical pin does not exist, or
+   *  it is reserved by the hypervisor for system usage. */
+  GXIO_GPIO_ERR_PIN_UNAVAILABLE = -1240,
+
+  /** Pin busy.  The pin exists, and is available for use via GXIO, but
+   *  it has been attached by some other process or driver. */
+  GXIO_GPIO_ERR_PIN_BUSY = -1241,
+
+  /** Cannot access unattached pin.  One or more of the pins being
+   *  manipulated by this call are not attached to the requesting
+   *  context. */
+  GXIO_GPIO_ERR_PIN_UNATTACHED = -1242,
+
+  /** Invalid I/O mode for pin.  The wiring of the pin in the system
+   *  is such that the I/O mode or electrical control parameters
+   *  requested could cause damage. */
+  GXIO_GPIO_ERR_PIN_INVALID_MODE = -1243,
+
+  /** Smallest iorpc error number. */
+  GXIO_ERR_MIN = -1299
+};
+
+
+#endif /* !_HV_IORPC_H_ */
diff --git a/kernel/arch/tile/include/hv/netio_errors.h b/kernel/arch/tile/include/hv/netio_errors.h
new file mode 100644
index 000000000..e1591bff6
--- /dev/null
+++ b/kernel/arch/tile/include/hv/netio_errors.h
@@ -0,0 +1,122 @@
+/*
+ * Copyright 2010 Tilera Corporation. All Rights Reserved.
+ *
+ *   This program is free software; you can redistribute it and/or
+ *   modify it under the terms of the GNU General Public License
+ *   as published by the Free Software Foundation, version 2.
+ *
+ *   This program is distributed in the hope that it will be useful, but
+ *   WITHOUT ANY WARRANTY; without even the implied warranty of
+ *   MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
+ *   NON INFRINGEMENT.  See the GNU General Public License for
+ *   more details.
+ */
+
+/**
+ * Error codes returned from NetIO routines.
+ */
+
+#ifndef __NETIO_ERRORS_H__
+#define __NETIO_ERRORS_H__
+
+/**
+ * @addtogroup error
+ *
+ * @brief The error codes returned by NetIO functions.
+ *
+ * NetIO functions return 0 (defined as ::NETIO_NO_ERROR) on success, and
+ * a negative value if an error occurs.
+ *
+ * In cases where a NetIO function failed due to a error reported by
+ * system libraries, the error code will be the negation of the
+ * system errno at the time of failure.  The @ref netio_strerror()
+ * function will deliver error strings for both NetIO and system error
+ * codes.
+ *
+ * @{
+ */
+
+/** The set of all NetIO errors. */
+typedef enum
+{
+  /** Operation successfully completed. */
+  NETIO_NO_ERROR        = 0,
+
+  /** A packet was successfully retrieved from an input queue. */
+  NETIO_PKT             = 0,
+
+  /** Largest NetIO error number. */
+  NETIO_ERR_MAX         = -701,
+
+  /** The tile is not registered with the IPP. */
+  NETIO_NOT_REGISTERED  = -701,
+
+  /** No packet was available to retrieve from the input queue. */
+  NETIO_NOPKT           = -702,
+
+  /** The requested function is not implemented. */
+  NETIO_NOT_IMPLEMENTED = -703,
+
+  /** On a registration operation, the target queue already has the maximum
+   *  number of tiles registered for it, and no more may be added.  On a
+   *  packet send operation, the output queue is full and nothing more can
+   *  be queued until some of the queued packets are actually transmitted. */
+  NETIO_QUEUE_FULL      = -704,
+
+  /** The calling process or thread is not bound to exactly one CPU. */
+  NETIO_BAD_AFFINITY    = -705,
+
+  /** Cannot allocate memory on requested controllers. */
+  NETIO_CANNOT_HOME     = -706,
+
+  /** On a registration operation, the IPP specified is not configured
+   *  to support the options requested; for instance, the application
+   *  wants a specific type of tagged headers which the configured IPP
+   *  doesn't support.  Or, the supplied configuration information is
+   *  not self-consistent, or is out of range; for instance, specifying
+   *  both NETIO_RECV and NETIO_NO_RECV, or asking for more than
+   *  NETIO_MAX_SEND_BUFFERS to be preallocated.  On a VLAN or bucket
+   *  configure operation, the number of items, or the base item, was
+   *  out of range.
+   */
+  NETIO_BAD_CONFIG      = -707,
+
+  /** Too many tiles have registered to transmit packets. */
+  NETIO_TOOMANY_XMIT    = -708,
+
+  /** Packet transmission was attempted on a queue which was registered
+      with transmit disabled. */
+  NETIO_UNREG_XMIT      = -709,
+
+  /** This tile is already registered with the IPP. */
+  NETIO_ALREADY_REGISTERED = -710,
+
+  /** The Ethernet link is down. The application should try again later. */
+  NETIO_LINK_DOWN       = -711,
+
+  /** An invalid memory buffer has been specified.  This may be an unmapped
+   * virtual address, or one which does not meet alignment requirements.
+   * For netio_input_register(), this error may be returned when multiple
+   * processes specify different memory regions to be used for NetIO
+   * buffers.  That can happen if these processes specify explicit memory
+   * regions with the ::NETIO_FIXED_BUFFER_VA flag, or if tmc_cmem_init()
+   * has not been called by a common ancestor of the processes.
+   */
+  NETIO_FAULT           = -712,
+
+  /** Cannot combine user-managed shared memory and cache coherence. */
+  NETIO_BAD_CACHE_CONFIG = -713,
+
+  /** Smallest NetIO error number. */
+  NETIO_ERR_MIN         = -713,
+
+#ifndef __DOXYGEN__
+  /** Used internally to mean that no response is needed; never returned to
+   *  an application. */
+  NETIO_NO_RESPONSE     = 1
+#endif
+} netio_error_t;
+
+/** @} */
+
+#endif /* __NETIO_ERRORS_H__ */
diff --git a/kernel/arch/tile/include/hv/netio_intf.h b/kernel/arch/tile/include/hv/netio_intf.h
new file mode 100644
index 000000000..8d20972ab
--- /dev/null
+++ b/kernel/arch/tile/include/hv/netio_intf.h
@@ -0,0 +1,2975 @@
+/*
+ * Copyright 2010 Tilera Corporation. All Rights Reserved.
+ *
+ *   This program is free software; you can redistribute it and/or
+ *   modify it under the terms of the GNU General Public License
+ *   as published by the Free Software Foundation, version 2.
+ *
+ *   This program is distributed in the hope that it will be useful, but
+ *   WITHOUT ANY WARRANTY; without even the implied warranty of
+ *   MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
+ *   NON INFRINGEMENT.  See the GNU General Public License for
+ *   more details.
+ */
+
+/**
+ * NetIO interface structures and macros.
+ */
+
+#ifndef __NETIO_INTF_H__
+#define __NETIO_INTF_H__
+
+#include <hv/netio_errors.h>
+
+#ifdef __KERNEL__
+#include <linux/types.h>
+#else
+#include <stdint.h>
+#endif
+
+#if !defined(__HV__) && !defined(__BOGUX__) && !defined(__KERNEL__)
+#include <assert.h>
+#define netio_assert assert  /**< Enable assertions from macros */
+#else
+#define netio_assert(...) ((void)(0))  /**< Disable assertions from macros */
+#endif
+
+/*
+ * If none of these symbols are defined, we're building libnetio in an
+ * environment where we have pthreads, so we'll enable locking.
+ */
+#if !defined(__HV__) && !defined(__BOGUX__) && !defined(__KERNEL__) && \
+    !defined(__NEWLIB__)
+#define _NETIO_PTHREAD       /**< Include a mutex in netio_queue_t below */
+
+/*
+ * If NETIO_UNLOCKED is defined, we don't do use per-cpu locks on
+ * per-packet NetIO operations.  We still do pthread locking on things
+ * like netio_input_register, though.  This is used for building
+ * libnetio_unlocked.
+ */
+#ifndef NETIO_UNLOCKED
+
+/* Avoid PLT overhead by using our own inlined per-cpu lock. */
+#include <sched.h>
+typedef int _netio_percpu_mutex_t;
+
+static __inline int
+_netio_percpu_mutex_init(_netio_percpu_mutex_t* lock)
+{
+  *lock = 0;
+  return 0;
+}
+
+static __inline int
+_netio_percpu_mutex_lock(_netio_percpu_mutex_t* lock)
+{
+  while (__builtin_expect(__insn_tns(lock), 0))
+    sched_yield();
+  return 0;
+}
+
+static __inline int
+_netio_percpu_mutex_unlock(_netio_percpu_mutex_t* lock)
+{
+  *lock = 0;
+  return 0;
+}
+
+#else /* NETIO_UNLOCKED */
+
+/* Don't do any locking for per-packet NetIO operations. */
+typedef int _netio_percpu_mutex_t;
+#define _netio_percpu_mutex_init(L)
+#define _netio_percpu_mutex_lock(L)
+#define _netio_percpu_mutex_unlock(L)
+
+#endif /* NETIO_UNLOCKED */
+#endif /* !__HV__, !__BOGUX, !__KERNEL__, !__NEWLIB__ */
+
+/** How many tiles can register for a given queue.
+ *  @ingroup setup */
+#define NETIO_MAX_TILES_PER_QUEUE  64
+
+
+/** Largest permissible queue identifier.
+ *  @ingroup setup  */
+#define NETIO_MAX_QUEUE_ID        255
+
+
+#ifndef __DOXYGEN__
+
+/* Metadata packet checksum/ethertype flags. */
+
+/** The L4 checksum has not been calculated. */
+#define _NETIO_PKT_NO_L4_CSUM_SHIFT           0
+#define _NETIO_PKT_NO_L4_CSUM_RMASK           1
+#define _NETIO_PKT_NO_L4_CSUM_MASK \
+         (_NETIO_PKT_NO_L4_CSUM_RMASK << _NETIO_PKT_NO_L4_CSUM_SHIFT)
+
+/** The L3 checksum has not been calculated. */
+#define _NETIO_PKT_NO_L3_CSUM_SHIFT           1
+#define _NETIO_PKT_NO_L3_CSUM_RMASK           1
+#define _NETIO_PKT_NO_L3_CSUM_MASK \
+         (_NETIO_PKT_NO_L3_CSUM_RMASK << _NETIO_PKT_NO_L3_CSUM_SHIFT)
+
+/** The L3 checksum is incorrect (or perhaps has not been calculated). */
+#define _NETIO_PKT_BAD_L3_CSUM_SHIFT          2
+#define _NETIO_PKT_BAD_L3_CSUM_RMASK          1
+#define _NETIO_PKT_BAD_L3_CSUM_MASK \
+         (_NETIO_PKT_BAD_L3_CSUM_RMASK << _NETIO_PKT_BAD_L3_CSUM_SHIFT)
+
+/** The Ethernet packet type is unrecognized. */
+#define _NETIO_PKT_TYPE_UNRECOGNIZED_SHIFT    3
+#define _NETIO_PKT_TYPE_UNRECOGNIZED_RMASK    1
+#define _NETIO_PKT_TYPE_UNRECOGNIZED_MASK \
+         (_NETIO_PKT_TYPE_UNRECOGNIZED_RMASK << \
+          _NETIO_PKT_TYPE_UNRECOGNIZED_SHIFT)
+
+/* Metadata packet type flags. */
+
+/** Where the packet type bits are; this field is the index into
+ *  _netio_pkt_info. */
+#define _NETIO_PKT_TYPE_SHIFT        4
+#define _NETIO_PKT_TYPE_RMASK        0x3F
+
+/** How many VLAN tags the packet has, and, if we have two, which one we
+ *  actually grouped on.  A VLAN within a proprietary (Marvell or Broadcom)
+ *  tag is counted here. */
+#define _NETIO_PKT_VLAN_SHIFT        4
+#define _NETIO_PKT_VLAN_RMASK        0x3
+#define _NETIO_PKT_VLAN_MASK \
+         (_NETIO_PKT_VLAN_RMASK << _NETIO_PKT_VLAN_SHIFT)
+#define _NETIO_PKT_VLAN_NONE         0   /* No VLAN tag. */
+#define _NETIO_PKT_VLAN_ONE          1   /* One VLAN tag. */
+#define _NETIO_PKT_VLAN_TWO_OUTER    2   /* Two VLAN tags, outer one used. */
+#define _NETIO_PKT_VLAN_TWO_INNER    3   /* Two VLAN tags, inner one used. */
+
+/** Which proprietary tags the packet has. */
+#define _NETIO_PKT_TAG_SHIFT         6
+#define _NETIO_PKT_TAG_RMASK         0x3
+#define _NETIO_PKT_TAG_MASK \
+          (_NETIO_PKT_TAG_RMASK << _NETIO_PKT_TAG_SHIFT)
+#define _NETIO_PKT_TAG_NONE          0   /* No proprietary tags. */
+#define _NETIO_PKT_TAG_MRVL          1   /* Marvell HyperG.Stack tags. */
+#define _NETIO_PKT_TAG_MRVL_EXT      2   /* HyperG.Stack extended tags. */
+#define _NETIO_PKT_TAG_BRCM          3   /* Broadcom HiGig tags. */
+
+/** Whether a packet has an LLC + SNAP header. */
+#define _NETIO_PKT_SNAP_SHIFT        8
+#define _NETIO_PKT_SNAP_RMASK        0x1
+#define _NETIO_PKT_SNAP_MASK \
+          (_NETIO_PKT_SNAP_RMASK << _NETIO_PKT_SNAP_SHIFT)
+
+/* NOTE: Bits 9 and 10 are unused. */
+
+/** Length of any custom data before the L2 header, in words. */
+#define _NETIO_PKT_CUSTOM_LEN_SHIFT  11
+#define _NETIO_PKT_CUSTOM_LEN_RMASK  0x1F
+#define _NETIO_PKT_CUSTOM_LEN_MASK \
+          (_NETIO_PKT_CUSTOM_LEN_RMASK << _NETIO_PKT_CUSTOM_LEN_SHIFT)
+
+/** The L4 checksum is incorrect (or perhaps has not been calculated). */
+#define _NETIO_PKT_BAD_L4_CSUM_SHIFT 16
+#define _NETIO_PKT_BAD_L4_CSUM_RMASK 0x1
+#define _NETIO_PKT_BAD_L4_CSUM_MASK \
+          (_NETIO_PKT_BAD_L4_CSUM_RMASK << _NETIO_PKT_BAD_L4_CSUM_SHIFT)
+
+/** Length of the L2 header, in words. */
+#define _NETIO_PKT_L2_LEN_SHIFT  17
+#define _NETIO_PKT_L2_LEN_RMASK  0x1F
+#define _NETIO_PKT_L2_LEN_MASK \
+          (_NETIO_PKT_L2_LEN_RMASK << _NETIO_PKT_L2_LEN_SHIFT)
+
+
+/* Flags in minimal packet metadata. */
+
+/** We need an eDMA checksum on this packet. */
+#define _NETIO_PKT_NEED_EDMA_CSUM_SHIFT            0
+#define _NETIO_PKT_NEED_EDMA_CSUM_RMASK            1
+#define _NETIO_PKT_NEED_EDMA_CSUM_MASK \
+         (_NETIO_PKT_NEED_EDMA_CSUM_RMASK << _NETIO_PKT_NEED_EDMA_CSUM_SHIFT)
+
+/* Data within the packet information table. */
+
+/* Note that, for efficiency, code which uses these fields assumes that none
+ * of the shift values below are zero.  See uses below for an explanation. */
+
+/** Offset within the L2 header of the innermost ethertype (in halfwords). */
+#define _NETIO_PKT_INFO_ETYPE_SHIFT       6
+#define _NETIO_PKT_INFO_ETYPE_RMASK    0x1F
+
+/** Offset within the L2 header of the VLAN tag (in halfwords). */
+#define _NETIO_PKT_INFO_VLAN_SHIFT       11
+#define _NETIO_PKT_INFO_VLAN_RMASK     0x1F
+
+#endif
+
+
+/** The size of a memory buffer representing a small packet.
+ *  @ingroup egress */
+#define SMALL_PACKET_SIZE 256
+
+/** The size of a memory buffer representing a large packet.
+ *  @ingroup egress */
+#define LARGE_PACKET_SIZE 2048
+
+/** The size of a memory buffer representing a jumbo packet.
+ *  @ingroup egress */
+#define JUMBO_PACKET_SIZE (12 * 1024)
+
+
+/* Common ethertypes.
+ * @ingroup ingress */
+/** @{ */
+/** The ethertype of IPv4. */
+#define ETHERTYPE_IPv4 (0x0800)
+/** The ethertype of ARP. */
+#define ETHERTYPE_ARP (0x0806)
+/** The ethertype of VLANs. */
+#define ETHERTYPE_VLAN (0x8100)
+/** The ethertype of a Q-in-Q header. */
+#define ETHERTYPE_Q_IN_Q (0x9100)
+/** The ethertype of IPv6. */
+#define ETHERTYPE_IPv6 (0x86DD)
+/** The ethertype of MPLS. */
+#define ETHERTYPE_MPLS (0x8847)
+/** @} */
+
+
+/** The possible return values of NETIO_PKT_STATUS.
+ * @ingroup ingress
+ */
+typedef enum
+{
+  /** No problems were detected with this packet. */
+  NETIO_PKT_STATUS_OK,
+  /** The packet is undersized; this is expected behavior if the packet's
+    * ethertype is unrecognized, but otherwise the packet is likely corrupt. */
+  NETIO_PKT_STATUS_UNDERSIZE,
+  /** The packet is oversized and some trailing bytes have been discarded.
+      This is expected behavior for short packets, since it's impossible to
+      precisely determine the amount of padding which may have been added to
+      them to make them meet the minimum Ethernet packet size. */
+  NETIO_PKT_STATUS_OVERSIZE,
+  /** The packet was judged to be corrupt by hardware (for instance, it had
+      a bad CRC, or part of it was discarded due to lack of buffer space in
+      the I/O shim) and should be discarded. */
+  NETIO_PKT_STATUS_BAD
+} netio_pkt_status_t;
+
+
+/** Log2 of how many buckets we have. */
+#define NETIO_LOG2_NUM_BUCKETS (10)
+
+/** How many buckets we have.
+ * @ingroup ingress */
+#define NETIO_NUM_BUCKETS (1 << NETIO_LOG2_NUM_BUCKETS)
+
+
+/**
+ * @brief A group-to-bucket identifier.
+ *
+ * @ingroup setup
+ *
+ * This tells us what to do with a given group.
+ */
+typedef union {
+  /** The header broken down into bits. */
+  struct {
+    /** Whether we should balance on L4, if available */
+    unsigned int __balance_on_l4:1;
+    /** Whether we should balance on L3, if available */
+    unsigned int __balance_on_l3:1;
+    /** Whether we should balance on L2, if available */
+    unsigned int __balance_on_l2:1;
+    /** Reserved for future use */
+    unsigned int __reserved:1;
+    /** The base bucket to use to send traffic */
+    unsigned int __bucket_base:NETIO_LOG2_NUM_BUCKETS;
+    /** The mask to apply to the balancing value. This must be one less
+     * than a power of two, e.g. 0x3 or 0xFF.
+     */
+    unsigned int __bucket_mask:NETIO_LOG2_NUM_BUCKETS;
+    /** Pad to 32 bits */
+    unsigned int __padding:(32 - 4 - 2 * NETIO_LOG2_NUM_BUCKETS);
+  } bits;
+  /** To send out the IDN. */
+  unsigned int word;
+}
+netio_group_t;
+
+
+/**
+ * @brief A VLAN-to-bucket identifier.
+ *
+ * @ingroup setup
+ *
+ * This tells us what to do with a given VLAN.
+ */
+typedef netio_group_t netio_vlan_t;
+
+
+/**
+ * A bucket-to-queue mapping.
+ * @ingroup setup
+ */
+typedef unsigned char netio_bucket_t;
+
+
+/**
+ * A packet size can always fit in a netio_size_t.
+ * @ingroup setup
+ */
+typedef unsigned int netio_size_t;
+
+
+/**
+ * @brief Ethernet standard (ingress) packet metadata.
+ *
+ * @ingroup ingress
+ *
+ * This is additional data associated with each packet.
+ * This structure is opaque and accessed through the @ref ingress.
+ *
+ * Also, the buffer population operation currently assumes that standard
+ * metadata is at least as large as minimal metadata, and will need to be
+ * modified if that is no longer the case.
+ */
+typedef struct
+{
+#ifdef __DOXYGEN__
+  /** This structure is opaque. */
+  unsigned char opaque[24];
+#else
+  /** The overall ordinal of the packet */
+  unsigned int __packet_ordinal;
+  /** The ordinal of the packet within the group */
+  unsigned int __group_ordinal;
+  /** The best flow hash IPP could compute. */
+  unsigned int __flow_hash;
+  /** Flags pertaining to checksum calculation, packet type, etc. */
+  unsigned int __flags;
+  /** The first word of "user data". */
+  unsigned int __user_data_0;
+  /** The second word of "user data". */
+  unsigned int __user_data_1;
+#endif
+}
+netio_pkt_metadata_t;
+
+
+/** To ensure that the L3 header is aligned mod 4, the L2 header should be
+ * aligned mod 4 plus 2, since every supported L2 header is 4n + 2 bytes
+ * long.  The standard way to do this is to simply add 2 bytes of padding
+ * before the L2 header.
+ */
+#define NETIO_PACKET_PADDING 2
+
+
+
+/**
+ * @brief Ethernet minimal (egress) packet metadata.
+ *
+ * @ingroup egress
+ *
+ * This structure represents information about packets which have
+ * been processed by @ref netio_populate_buffer() or
+ * @ref netio_populate_prepend_buffer().  This structure is opaque
+ * and accessed through the @ref egress.
+ *
+ * @internal This structure is actually copied into the memory used by
+ * standard metadata, which is assumed to be large enough.
+ */
+typedef struct
+{
+#ifdef __DOXYGEN__
+  /** This structure is opaque. */
+  unsigned char opaque[14];
+#else
+  /** The offset of the L2 header from the start of the packet data. */
+  unsigned short l2_offset;
+  /** The offset of the L3 header from the start of the packet data. */
+  unsigned short l3_offset;
+  /** Where to write the checksum. */
+  unsigned char csum_location;
+  /** Where to start checksumming from. */
+  unsigned char csum_start;
+  /** Flags pertaining to checksum calculation etc. */
+  unsigned short flags;
+  /** The L2 length of the packet. */
+  unsigned short l2_length;
+  /** The checksum with which to seed the checksum generator. */
+  unsigned short csum_seed;
+  /** How much to checksum. */
+  unsigned short csum_length;
+#endif
+}
+netio_pkt_minimal_metadata_t;
+
+
+#ifndef __DOXYGEN__
+
+/**
+ * @brief An I/O notification header.
+ *
+ * This is the first word of data received from an I/O shim in a notification
+ * packet. It contains framing and status information.
+ */
+typedef union
+{
+  unsigned int word; /**< The whole word. */
+  /** The various fields. */
+  struct
+  {
+    unsigned int __channel:7;    /**< Resource channel. */
+    unsigned int __type:4;       /**< Type. */
+    unsigned int __ack:1;        /**< Whether an acknowledgement is needed. */
+    unsigned int __reserved:1;   /**< Reserved. */
+    unsigned int __protocol:1;   /**< A protocol-specific word is added. */
+    unsigned int __status:2;     /**< Status of the transfer. */
+    unsigned int __framing:2;    /**< Framing of the transfer. */
+    unsigned int __transfer_size:14; /**< Transfer size in bytes (total). */
+  } bits;
+}
+__netio_pkt_notif_t;
+
+
+/**
+ * Returns the base address of the packet.
+ */
+#define _NETIO_PKT_HANDLE_BASE(p) \
+  ((unsigned char*)((p).word & 0xFFFFFFC0))
+
+/**
+ * Returns the base address of the packet.
+ */
+#define _NETIO_PKT_BASE(p) \
+  _NETIO_PKT_HANDLE_BASE(p->__packet)
+
+/**
+ * @brief An I/O notification packet (second word)
+ *
+ * This is the second word of data received from an I/O shim in a notification
+ * packet.  This is the virtual address of the packet buffer, plus some flag
+ * bits.  (The virtual address of the packet is always 256-byte aligned so we
+ * have room for 8 bits' worth of flags in the low 8 bits.)
+ *
+ * @internal
+ * NOTE: The low two bits must contain "__queue", so the "packet size"
+ * (SIZE_SMALL, SIZE_LARGE, or SIZE_JUMBO) can be determined quickly.
+ *
+ * If __addr or __offset are moved, _NETIO_PKT_BASE
+ * (defined right below this) must be changed.
+ */
+typedef union
+{
+  unsigned int word; /**< The whole word. */
+  /** The various fields. */
+  struct
+  {
+    /** Which queue the packet will be returned to once it is sent back to
+        the IPP.  This is one of the SIZE_xxx values. */
+    unsigned int __queue:2;
+
+    /** The IPP handle of the sending IPP. */
+    unsigned int __ipp_handle:2;
+
+    /** Reserved for future use. */
+    unsigned int __reserved:1;
+
+    /** If 1, this packet has minimal (egress) metadata; otherwise, it
+        has standard (ingress) metadata. */
+    unsigned int __minimal:1;
+
+    /** Offset of the metadata within the packet.  This value is multiplied
+     *  by 64 and added to the base packet address to get the metadata
+     *  address.  Note that this field is aligned within the word such that
+     *  you can easily extract the metadata address with a 26-bit mask. */
+    unsigned int __offset:2;
+
+    /** The top 24 bits of the packet's virtual address. */
+    unsigned int __addr:24;
+  } bits;
+}
+__netio_pkt_handle_t;
+
+#endif /* !__DOXYGEN__ */
+
+
+/**
+ * @brief A handle for an I/O packet's storage.
+ * @ingroup ingress
+ *
+ * netio_pkt_handle_t encodes the concept of a ::netio_pkt_t with its
+ * packet metadata removed.  It is a much smaller type that exists to
+ * facilitate applications where the full ::netio_pkt_t type is too
+ * large, such as those that cache enormous numbers of packets or wish
+ * to transmit packet descriptors over the UDN.
+ *
+ * Because there is no metadata, most ::netio_pkt_t operations cannot be
+ * performed on a netio_pkt_handle_t.  It supports only
+ * netio_free_handle() (to free the buffer) and
+ * NETIO_PKT_CUSTOM_DATA_H() (to access a pointer to its contents).
+ * The application must acquire any additional metadata it wants from the
+ * original ::netio_pkt_t and record it separately.
+ *
+ * A netio_pkt_handle_t can be extracted from a ::netio_pkt_t by calling
+ * NETIO_PKT_HANDLE().  An invalid handle (analogous to NULL) can be
+ * created by assigning the value ::NETIO_PKT_HANDLE_NONE. A handle can
+ * be tested for validity with NETIO_PKT_HANDLE_IS_VALID().
+ */
+typedef struct
+{
+  unsigned int word; /**< Opaque bits. */
+} netio_pkt_handle_t;
+
+/**
+ * @brief A packet descriptor.
+ *
+ * @ingroup ingress
+ * @ingroup egress
+ *
+ * This data structure represents a packet.  The structure is manipulated
+ * through the @ref ingress and the @ref egress.
+ *
+ * While the contents of a netio_pkt_t are opaque, the structure itself is
+ * portable.  This means that it may be shared between all tiles which have
+ * done a netio_input_register() call for the interface on which the pkt_t
+ * was initially received (via netio_get_packet()) or retrieved (via
+ * netio_get_buffer()).  The contents of a netio_pkt_t can be transmitted to
+ * another tile via shared memory, or via a UDN message, or by other means.
+ * The destination tile may then use the pkt_t as if it had originally been
+ * received locally; it may read or write the packet's data, read its
+ * metadata, free the packet, send the packet, transfer the netio_pkt_t to
+ * yet another tile, and so forth.
+ *
+ * Once a netio_pkt_t has been transferred to a second tile, the first tile
+ * should not reference the original copy; in particular, if more than one
+ * tile frees or sends the same netio_pkt_t, the IPP's packet free lists will
+ * become corrupted.  Note also that each tile which reads or modifies
+ * packet data must obey the memory coherency rules outlined in @ref input.
+ */
+typedef struct
+{
+#ifdef __DOXYGEN__
+  /** This structure is opaque. */
+  unsigned char opaque[32];
+#else
+  /** For an ingress packet (one with standard metadata), this is the
+   *  notification header we got from the I/O shim.  For an egress packet
+   *  (one with minimal metadata), this word is zero if the packet has not
+   *  been populated, and nonzero if it has. */
+  __netio_pkt_notif_t __notif_header;
+
+  /** Virtual address of the packet buffer, plus state flags. */
+  __netio_pkt_handle_t __packet;
+
+  /** Metadata associated with the packet. */
+  netio_pkt_metadata_t __metadata;
+#endif
+}
+netio_pkt_t;
+
+
+#ifndef __DOXYGEN__
+
+#define __NETIO_PKT_NOTIF_HEADER(pkt) ((pkt)->__notif_header)
+#define __NETIO_PKT_IPP_HANDLE(pkt) ((pkt)->__packet.bits.__ipp_handle)
+#define __NETIO_PKT_QUEUE(pkt) ((pkt)->__packet.bits.__queue)
+#define __NETIO_PKT_NOTIF_HEADER_M(mda, pkt) ((pkt)->__notif_header)
+#define __NETIO_PKT_IPP_HANDLE_M(mda, pkt) ((pkt)->__packet.bits.__ipp_handle)
+#define __NETIO_PKT_MINIMAL(pkt) ((pkt)->__packet.bits.__minimal)
+#define __NETIO_PKT_QUEUE_M(mda, pkt) ((pkt)->__packet.bits.__queue)
+#define __NETIO_PKT_FLAGS_M(mda, pkt) ((mda)->__flags)
+
+/* Packet information table, used by the attribute access functions below. */
+extern const uint16_t _netio_pkt_info[];
+
+#endif /* __DOXYGEN__ */
+
+
+#ifndef __DOXYGEN__
+/* These macros are deprecated and will disappear in a future MDE release. */
+#define NETIO_PKT_GOOD_CHECKSUM(pkt) \
+  NETIO_PKT_L4_CSUM_CORRECT(pkt)
+#define NETIO_PKT_GOOD_CHECKSUM_M(mda, pkt) \
+  NETIO_PKT_L4_CSUM_CORRECT_M(mda, pkt)
+#endif /* __DOXYGEN__ */
+
+
+/* Packet attribute access functions. */
+
+/** Return a pointer to the metadata for a packet.
+ * @ingroup ingress
+ *
+ * Calling this function once and passing the result to other retrieval
+ * functions with a "_M" suffix usually improves performance.  This
+ * function must be called on an 'ingress' packet (i.e. one retrieved
+ * by @ref netio_get_packet(), on which @ref netio_populate_buffer() or
+ * @ref netio_populate_prepend_buffer have not been called). Use of this
+ * function on an 'egress' packet will cause an assertion failure.
+ *
+ * @param[in] pkt Packet on which to operate.
+ * @return A pointer to the packet's standard metadata.
+ */
+static __inline netio_pkt_metadata_t*
+NETIO_PKT_METADATA(netio_pkt_t* pkt)
+{
+  netio_assert(!pkt->__packet.bits.__minimal);
+  return &pkt->__metadata;
+}
+
+
+/** Return a pointer to the minimal metadata for a packet.
+ * @ingroup egress
+ *
+ * Calling this function once and passing the result to other retrieval
+ * functions with a "_MM" suffix usually improves performance.  This
+ * function must be called on an 'egress' packet (i.e. one on which
+ * @ref netio_populate_buffer() or @ref netio_populate_prepend_buffer()
+ * have been called, or one retrieved by @ref netio_get_buffer()). Use of
+ * this function on an 'ingress' packet will cause an assertion failure.
+ *
+ * @param[in] pkt Packet on which to operate.
+ * @return A pointer to the packet's standard metadata.
+ */
+static __inline netio_pkt_minimal_metadata_t*
+NETIO_PKT_MINIMAL_METADATA(netio_pkt_t* pkt)
+{
+  netio_assert(pkt->__packet.bits.__minimal);
+  return (netio_pkt_minimal_metadata_t*) &pkt->__metadata;
+}
+
+
+/** Determine whether a packet has 'minimal' metadata.
+ * @ingroup pktfuncs
+ *
+ * This function will return nonzero if the packet is an 'egress'
+ * packet (i.e. one on which @ref netio_populate_buffer() or
+ * @ref netio_populate_prepend_buffer() have been called, or one
+ * retrieved by @ref netio_get_buffer()), and zero if the packet
+ * is an 'ingress' packet (i.e. one retrieved by @ref netio_get_packet(),
+ * which has not been converted into an 'egress' packet).
+ *
+ * @param[in] pkt Packet on which to operate.
+ * @return Nonzero if the packet has minimal metadata.
+ */
+static __inline unsigned int
+NETIO_PKT_IS_MINIMAL(netio_pkt_t* pkt)
+{
+  return pkt->__packet.bits.__minimal;
+}
+
+
+/** Return a handle for a packet's storage.
+ * @ingroup pktfuncs
+ *
+ * @param[in] pkt Packet on which to operate.
+ * @return A handle for the packet's storage.
+ */
+static __inline netio_pkt_handle_t
+NETIO_PKT_HANDLE(netio_pkt_t* pkt)
+{
+  netio_pkt_handle_t h;
+  h.word = pkt->__packet.word;
+  return h;
+}
+
+
+/** A special reserved value indicating the absence of a packet handle.
+ *
+ * @ingroup pktfuncs
+ */
+#define NETIO_PKT_HANDLE_NONE ((netio_pkt_handle_t) { 0 })
+
+
+/** Test whether a packet handle is valid.
+ *
+ * Applications may wish to use the reserved value NETIO_PKT_HANDLE_NONE
+ * to indicate no packet at all.  This function tests to see if a packet
+ * handle is a real handle, not this special reserved value.
+ *
+ * @ingroup pktfuncs
+ *
+ * @param[in] handle Handle on which to operate.
+ * @return One if the packet handle is valid, else zero.
+ */
+static __inline unsigned int
+NETIO_PKT_HANDLE_IS_VALID(netio_pkt_handle_t handle)
+{
+  return handle.word != 0;
+}
+
+
+
+/** Return a pointer to the start of the packet's custom header.
+ *  A custom header may or may not be present, depending upon the IPP; its
+ *  contents and alignment are also IPP-dependent.  Currently, none of the
+ *  standard IPPs supplied by Tilera produce a custom header.  If present,
+ *  the custom header precedes the L2 header in the packet buffer.
+ * @ingroup ingress
+ *
+ * @param[in] handle Handle on which to operate.
+ * @return A pointer to start of the packet.
+ */
+static __inline unsigned char*
+NETIO_PKT_CUSTOM_DATA_H(netio_pkt_handle_t handle)
+{
+  return _NETIO_PKT_HANDLE_BASE(handle) + NETIO_PACKET_PADDING;
+}
+
+
+/** Return the length of the packet's custom header.
+ *  A custom header may or may not be present, depending upon the IPP; its
+ *  contents and alignment are also IPP-dependent.  Currently, none of the
+ *  standard IPPs supplied by Tilera produce a custom header.  If present,
+ *  the custom header precedes the L2 header in the packet buffer.
+ *
+ * @ingroup ingress
+ *
+ * @param[in] mda Pointer to packet's standard metadata.
+ * @param[in] pkt Packet on which to operate.
+ * @return The length of the packet's custom header, in bytes.
+ */
+static __inline netio_size_t
+NETIO_PKT_CUSTOM_HEADER_LENGTH_M(netio_pkt_metadata_t* mda, netio_pkt_t* pkt)
+{
+  /*
+   * Note that we effectively need to extract a quantity from the flags word
+   * which is measured in words, and then turn it into bytes by shifting
+   * it left by 2.  We do this all at once by just shifting right two less
+   * bits, and shifting the mask up two bits.
+   */
+  return ((mda->__flags >> (_NETIO_PKT_CUSTOM_LEN_SHIFT - 2)) &
+          (_NETIO_PKT_CUSTOM_LEN_RMASK << 2));
+}
+
+
+/** Return the length of the packet, starting with the custom header.
+ *  A custom header may or may not be present, depending upon the IPP; its
+ *  contents and alignment are also IPP-dependent.  Currently, none of the
+ *  standard IPPs supplied by Tilera produce a custom header.  If present,
+ *  the custom header precedes the L2 header in the packet buffer.
+ * @ingroup ingress
+ *
+ * @param[in] mda Pointer to packet's standard metadata.
+ * @param[in] pkt Packet on which to operate.
+ * @return The length of the packet, in bytes.
+ */
+static __inline netio_size_t
+NETIO_PKT_CUSTOM_LENGTH_M(netio_pkt_metadata_t* mda, netio_pkt_t* pkt)
+{
+  return (__NETIO_PKT_NOTIF_HEADER(pkt).bits.__transfer_size -
+          NETIO_PACKET_PADDING);
+}
+
+
+/** Return a pointer to the start of the packet's custom header.
+ *  A custom header may or may not be present, depending upon the IPP; its
+ *  contents and alignment are also IPP-dependent.  Currently, none of the
+ *  standard IPPs supplied by Tilera produce a custom header.  If present,
+ *  the custom header precedes the L2 header in the packet buffer.
+ * @ingroup ingress
+ *
+ * @param[in] mda Pointer to packet's standard metadata.
+ * @param[in] pkt Packet on which to operate.
+ * @return A pointer to start of the packet.
+ */
+static __inline unsigned char*
+NETIO_PKT_CUSTOM_DATA_M(netio_pkt_metadata_t* mda, netio_pkt_t* pkt)
+{
+  return NETIO_PKT_CUSTOM_DATA_H(NETIO_PKT_HANDLE(pkt));
+}
+
+
+/** Return the length of the packet's L2 (Ethernet plus VLAN or SNAP) header.
+ * @ingroup ingress
+ *
+ * @param[in] mda Pointer to packet's standard metadata.
+ * @param[in] pkt Packet on which to operate.
+ * @return The length of the packet's L2 header, in bytes.
+ */
+static __inline netio_size_t
+NETIO_PKT_L2_HEADER_LENGTH_M(netio_pkt_metadata_t* mda, netio_pkt_t* pkt)
+{
+  /*
+   * Note that we effectively need to extract a quantity from the flags word
+   * which is measured in words, and then turn it into bytes by shifting
+   * it left by 2.  We do this all at once by just shifting right two less
+   * bits, and shifting the mask up two bits.  We then add two bytes.
+   */
+  return ((mda->__flags >> (_NETIO_PKT_L2_LEN_SHIFT - 2)) &
+          (_NETIO_PKT_L2_LEN_RMASK << 2)) + 2;
+}
+
+
+/** Return the length of the packet, starting with the L2 (Ethernet) header.
+ * @ingroup ingress
+ *
+ * @param[in] mda Pointer to packet's standard metadata.
+ * @param[in] pkt Packet on which to operate.
+ * @return The length of the packet, in bytes.
+ */
+static __inline netio_size_t
+NETIO_PKT_L2_LENGTH_M(netio_pkt_metadata_t* mda, netio_pkt_t* pkt)
+{
+  return (NETIO_PKT_CUSTOM_LENGTH_M(mda, pkt) -
+          NETIO_PKT_CUSTOM_HEADER_LENGTH_M(mda,pkt));
+}
+
+
+/** Return a pointer to the start of the packet's L2 (Ethernet) header.
+ * @ingroup ingress
+ *
+ * @param[in] mda Pointer to packet's standard metadata.
+ * @param[in] pkt Packet on which to operate.
+ * @return A pointer to start of the packet.
+ */
+static __inline unsigned char*
+NETIO_PKT_L2_DATA_M(netio_pkt_metadata_t* mda, netio_pkt_t* pkt)
+{
+  return (NETIO_PKT_CUSTOM_DATA_M(mda, pkt) +
+          NETIO_PKT_CUSTOM_HEADER_LENGTH_M(mda, pkt));
+}
+
+
+/** Retrieve the length of the packet, starting with the L3 (generally,
+ *  the IP) header.
+ * @ingroup ingress
+ *
+ * @param[in] mda Pointer to packet's standard metadata.
+ * @param[in] pkt Packet on which to operate.
+ * @return Length of the packet's L3 header and data, in bytes.
+ */
+static __inline netio_size_t
+NETIO_PKT_L3_LENGTH_M(netio_pkt_metadata_t* mda, netio_pkt_t* pkt)
+{
+  return (NETIO_PKT_L2_LENGTH_M(mda, pkt) -
+          NETIO_PKT_L2_HEADER_LENGTH_M(mda,pkt));
+}
+
+
+/** Return a pointer to the packet's L3 (generally, the IP) header.
+ * @ingroup ingress
+ *
+ * Note that we guarantee word alignment of the L3 header.
+ *
+ * @param[in] mda Pointer to packet's standard metadata.
+ * @param[in] pkt Packet on which to operate.
+ * @return A pointer to the packet's L3 header.
+ */
+static __inline unsigned char*
+NETIO_PKT_L3_DATA_M(netio_pkt_metadata_t* mda, netio_pkt_t* pkt)
+{
+  return (NETIO_PKT_L2_DATA_M(mda, pkt) +
+          NETIO_PKT_L2_HEADER_LENGTH_M(mda, pkt));
+}
+
+
+/** Return the ordinal of the packet.
+ * @ingroup ingress
+ *
+ * Each packet is given an ordinal number when it is delivered by the IPP.
+ * In the medium term, the ordinal is unique and monotonically increasing,
+ * being incremented by 1 for each packet; the ordinal of the first packet
+ * delivered after the IPP starts is zero.  (Since the ordinal is of finite
+ * size, given enough input packets, it will eventually wrap around to zero;
+ * in the long term, therefore, ordinals are not unique.)  The ordinals
+ * handed out by different IPPs are not disjoint, so two packets from
+ * different IPPs may have identical ordinals.  Packets dropped by the
+ * IPP or by the I/O shim are not assigned ordinals.
+ *
+ * @param[in] mda Pointer to packet's standard metadata.
+ * @param[in] pkt Packet on which to operate.
+ * @return The packet's per-IPP packet ordinal.
+ */
+static __inline unsigned int
+NETIO_PKT_ORDINAL_M(netio_pkt_metadata_t* mda, netio_pkt_t* pkt)
+{
+  return mda->__packet_ordinal;
+}
+
+
+/** Return the per-group ordinal of the packet.
+ * @ingroup ingress
+ *
+ * Each packet is given a per-group ordinal number when it is
+ * delivered by the IPP. By default, the group is the packet's VLAN,
+ * although IPP can be recompiled to use different values.  In
+ * the medium term, the ordinal is unique and monotonically
+ * increasing, being incremented by 1 for each packet; the ordinal of
+ * the first packet distributed to a particular group is zero.
+ * (Since the ordinal is of finite size, given enough input packets,
+ * it will eventually wrap around to zero; in the long term,
+ * therefore, ordinals are not unique.)  The ordinals handed out by
+ * different IPPs are not disjoint, so two packets from different IPPs
+ * may have identical ordinals; similarly, packets distributed to
+ * different groups may have identical ordinals.  Packets dropped by
+ * the IPP or by the I/O shim are not assigned ordinals.
+ *
+ * @param[in] mda Pointer to packet's standard metadata.
+ * @param[in] pkt Packet on which to operate.
+ * @return The packet's per-IPP, per-group ordinal.
+ */
+static __inline unsigned int
+NETIO_PKT_GROUP_ORDINAL_M(netio_pkt_metadata_t* mda, netio_pkt_t* pkt)
+{
+  return mda->__group_ordinal;
+}
+
+
+/** Return the VLAN ID assigned to the packet.
+ * @ingroup ingress
+ *
+ * This value is usually contained within the packet header.
+ *
+ * This value will be zero if the packet does not have a VLAN tag, or if
+ * this value was not extracted from the packet.
+ *
+ * @param[in] mda Pointer to packet's standard metadata.
+ * @param[in] pkt Packet on which to operate.
+ * @return The packet's VLAN ID.
+ */
+static __inline unsigned short
+NETIO_PKT_VLAN_ID_M(netio_pkt_metadata_t* mda, netio_pkt_t* pkt)
+{
+  int vl = (mda->__flags >> _NETIO_PKT_VLAN_SHIFT) & _NETIO_PKT_VLAN_RMASK;
+  unsigned short* pkt_p;
+  int index;
+  unsigned short val;
+
+  if (vl == _NETIO_PKT_VLAN_NONE)
+    return 0;
+
+  pkt_p = (unsigned short*) NETIO_PKT_L2_DATA_M(mda, pkt);
+  index = (mda->__flags >> _NETIO_PKT_TYPE_SHIFT) & _NETIO_PKT_TYPE_RMASK;
+
+  val = pkt_p[(_netio_pkt_info[index] >> _NETIO_PKT_INFO_VLAN_SHIFT) &
+              _NETIO_PKT_INFO_VLAN_RMASK];
+
+#ifdef __TILECC__
+  return (__insn_bytex(val) >> 16) & 0xFFF;
+#else
+  return (__builtin_bswap32(val) >> 16) & 0xFFF;
+#endif
+}
+
+
+/** Return the ethertype of the packet.
+ * @ingroup ingress
+ *
+ * This value is usually contained within the packet header.
+ *
+ * This value is reliable if @ref NETIO_PKT_ETHERTYPE_RECOGNIZED_M()
+ * returns true, and otherwise, may not be well defined.
+ *
+ * @param[in] mda Pointer to packet's standard metadata.
+ * @param[in] pkt Packet on which to operate.
+ * @return The packet's ethertype.
+ */
+static __inline unsigned short
+NETIO_PKT_ETHERTYPE_M(netio_pkt_metadata_t* mda, netio_pkt_t* pkt)
+{
+  unsigned short* pkt_p = (unsigned short*) NETIO_PKT_L2_DATA_M(mda, pkt);
+  int index = (mda->__flags >> _NETIO_PKT_TYPE_SHIFT) & _NETIO_PKT_TYPE_RMASK;
+
+  unsigned short val =
+    pkt_p[(_netio_pkt_info[index] >> _NETIO_PKT_INFO_ETYPE_SHIFT) &
+          _NETIO_PKT_INFO_ETYPE_RMASK];
+
+  return __builtin_bswap32(val) >> 16;
+}
+
+
+/** Return the flow hash computed on the packet.
+ * @ingroup ingress
+ *
+ * For TCP and UDP packets, this hash is calculated by hashing together
+ * the "5-tuple" values, specifically the source IP address, destination
+ * IP address, protocol type, source port and destination port.
+ * The hash value is intended to be helpful for millions of distinct
+ * flows.
+ *
+ * For IPv4 or IPv6 packets which are neither TCP nor UDP, the flow hash is
+ * derived by hashing together the source and destination IP addresses.
+ *
+ * For MPLS-encapsulated packets, the flow hash is derived by hashing
+ * the first MPLS label.
+ *
+ * For all other packets the flow hash is computed from the source
+ * and destination Ethernet addresses.
+ *
+ * The hash is symmetric, meaning it produces the same value if the
+ * source and destination are swapped. The only exceptions are
+ * tunneling protocols 0x04 (IP in IP Encapsulation), 0x29 (Simple
+ * Internet Protocol), 0x2F (General Routing Encapsulation) and 0x32
+ * (Encap Security Payload), which use only the destination address
+ * since the source address is not meaningful.
+ *
+ * @param[in] mda Pointer to packet's standard metadata.
+ * @param[in] pkt Packet on which to operate.
+ * @return The packet's 32-bit flow hash.
+ */
+static __inline unsigned int
+NETIO_PKT_FLOW_HASH_M(netio_pkt_metadata_t* mda, netio_pkt_t* pkt)
+{
+  return mda->__flow_hash;
+}
+
+
+/** Return the first word of "user data" for the packet.
+ *
+ * The contents of the user data words depend on the IPP.
+ *
+ * When using the standard ipp1, ipp2, or ipp4 sub-drivers, the first
+ * word of user data contains the least significant bits of the 64-bit
+ * arrival cycle count (see @c get_cycle_count_low()).
+ *
+ * See the <em>System Programmer's Guide</em> for details.
+ *
+ * @ingroup ingress
+ *
+ * @param[in] mda Pointer to packet's standard metadata.
+ * @param[in] pkt Packet on which to operate.
+ * @return The packet's first word of "user data".
+ */
+static __inline unsigned int
+NETIO_PKT_USER_DATA_0_M(netio_pkt_metadata_t* mda, netio_pkt_t* pkt)
+{
+  return mda->__user_data_0;
+}
+
+
+/** Return the second word of "user data" for the packet.
+ *
+ * The contents of the user data words depend on the IPP.
+ *
+ * When using the standard ipp1, ipp2, or ipp4 sub-drivers, the second
+ * word of user data contains the most significant bits of the 64-bit
+ * arrival cycle count (see @c get_cycle_count_high()).
+ *
+ * See the <em>System Programmer's Guide</em> for details.
+ *
+ * @ingroup ingress
+ *
+ * @param[in] mda Pointer to packet's standard metadata.
+ * @param[in] pkt Packet on which to operate.
+ * @return The packet's second word of "user data".
+ */
+static __inline unsigned int
+NETIO_PKT_USER_DATA_1_M(netio_pkt_metadata_t* mda, netio_pkt_t* pkt)
+{
+  return mda->__user_data_1;
+}
+
+
+/** Determine whether the L4 (TCP/UDP) checksum was calculated.
+ * @ingroup ingress
+ *
+ * @param[in] mda Pointer to packet's standard metadata.
+ * @param[in] pkt Packet on which to operate.
+ * @return Nonzero if the L4 checksum was calculated.
+ */
+static __inline unsigned int
+NETIO_PKT_L4_CSUM_CALCULATED_M(netio_pkt_metadata_t* mda, netio_pkt_t* pkt)
+{
+  return !(mda->__flags & _NETIO_PKT_NO_L4_CSUM_MASK);
+}
+
+
+/** Determine whether the L4 (TCP/UDP) checksum was calculated and found to
+ *  be correct.
+ * @ingroup ingress
+ *
+ * @param[in] mda Pointer to packet's standard metadata.
+ * @param[in] pkt Packet on which to operate.
+ * @return Nonzero if the checksum was calculated and is correct.
+ */
+static __inline unsigned int
+NETIO_PKT_L4_CSUM_CORRECT_M(netio_pkt_metadata_t* mda, netio_pkt_t* pkt)
+{
+  return !(mda->__flags &
+           (_NETIO_PKT_BAD_L4_CSUM_MASK | _NETIO_PKT_NO_L4_CSUM_MASK));
+}
+
+
+/** Determine whether the L3 (IP) checksum was calculated.
+ * @ingroup ingress
+ *
+ * @param[in] mda Pointer to packet's standard metadata.
+ * @param[in] pkt Packet on which to operate.
+ * @return Nonzero if the L3 (IP) checksum was calculated.
+*/
+static __inline unsigned int
+NETIO_PKT_L3_CSUM_CALCULATED_M(netio_pkt_metadata_t* mda, netio_pkt_t* pkt)
+{
+  return !(mda->__flags & _NETIO_PKT_NO_L3_CSUM_MASK);
+}
+
+
+/** Determine whether the L3 (IP) checksum was calculated and found to be
+ *  correct.
+ * @ingroup ingress
+ *
+ * @param[in] mda Pointer to packet's standard metadata.
+ * @param[in] pkt Packet on which to operate.
+ * @return Nonzero if the checksum was calculated and is correct.
+ */
+static __inline unsigned int
+NETIO_PKT_L3_CSUM_CORRECT_M(netio_pkt_metadata_t* mda, netio_pkt_t* pkt)
+{
+  return !(mda->__flags &
+           (_NETIO_PKT_BAD_L3_CSUM_MASK | _NETIO_PKT_NO_L3_CSUM_MASK));
+}
+
+
+/** Determine whether the ethertype was recognized and L3 packet data was
+ *  processed.
+ * @ingroup ingress
+ *
+ * @param[in] mda Pointer to packet's standard metadata.
+ * @param[in] pkt Packet on which to operate.
+ * @return Nonzero if the ethertype was recognized and L3 packet data was
+ *   processed.
+ */
+static __inline unsigned int
+NETIO_PKT_ETHERTYPE_RECOGNIZED_M(netio_pkt_metadata_t* mda, netio_pkt_t* pkt)
+{
+  return !(mda->__flags & _NETIO_PKT_TYPE_UNRECOGNIZED_MASK);
+}
+
+
+/** Retrieve the status of a packet and any errors that may have occurred
+ * during ingress processing (length mismatches, CRC errors, etc.).
+ * @ingroup ingress
+ *
+ * Note that packets for which @ref NETIO_PKT_ETHERTYPE_RECOGNIZED()
+ * returns zero are always reported as underlength, as there is no a priori
+ * means to determine their length.  Normally, applications should use
+ * @ref NETIO_PKT_BAD_M() instead of explicitly checking status with this
+ * function.
+ *
+ * @param[in] mda Pointer to packet's standard metadata.
+ * @param[in] pkt Packet on which to operate.
+ * @return The packet's status.
+ */
+static __inline netio_pkt_status_t
+NETIO_PKT_STATUS_M(netio_pkt_metadata_t* mda, netio_pkt_t* pkt)
+{
+  return (netio_pkt_status_t) __NETIO_PKT_NOTIF_HEADER(pkt).bits.__status;
+}
+
+
+/** Report whether a packet is bad (i.e., was shorter than expected based on
+ *  its headers, or had a bad CRC).
+ * @ingroup ingress
+ *
+ * Note that this function does not verify L3 or L4 checksums.
+ *
+ * @param[in] mda Pointer to packet's standard metadata.
+ * @param[in] pkt Packet on which to operate.
+ * @return Nonzero if the packet is bad and should be discarded.
+ */
+static __inline unsigned int
+NETIO_PKT_BAD_M(netio_pkt_metadata_t* mda, netio_pkt_t* pkt)
+{
+  return ((NETIO_PKT_STATUS_M(mda, pkt) & 1) &&
+          (NETIO_PKT_ETHERTYPE_RECOGNIZED_M(mda, pkt) ||
+           NETIO_PKT_STATUS_M(mda, pkt) == NETIO_PKT_STATUS_BAD));
+}
+
+
+/** Return the length of the packet, starting with the L2 (Ethernet) header.
+ * @ingroup egress
+ *
+ * @param[in] mmd Pointer to packet's minimal metadata.
+ * @param[in] pkt Packet on which to operate.
+ * @return The length of the packet, in bytes.
+ */
+static __inline netio_size_t
+NETIO_PKT_L2_LENGTH_MM(netio_pkt_minimal_metadata_t* mmd, netio_pkt_t* pkt)
+{
+  return mmd->l2_length;
+}
+
+
+/** Return the length of the L2 (Ethernet) header.
+ * @ingroup egress
+ *
+ * @param[in] mmd Pointer to packet's minimal metadata.
+ * @param[in] pkt Packet on which to operate.
+ * @return The length of the packet's L2 header, in bytes.
+ */
+static __inline netio_size_t
+NETIO_PKT_L2_HEADER_LENGTH_MM(netio_pkt_minimal_metadata_t* mmd,
+                              netio_pkt_t* pkt)
+{
+  return mmd->l3_offset - mmd->l2_offset;
+}
+
+
+/** Return the length of the packet, starting with the L3 (IP) header.
+ * @ingroup egress
+ *
+ * @param[in] mmd Pointer to packet's minimal metadata.
+ * @param[in] pkt Packet on which to operate.
+ * @return Length of the packet's L3 header and data, in bytes.
+ */
+static __inline netio_size_t
+NETIO_PKT_L3_LENGTH_MM(netio_pkt_minimal_metadata_t* mmd, netio_pkt_t* pkt)
+{
+  return (NETIO_PKT_L2_LENGTH_MM(mmd, pkt) -
+          NETIO_PKT_L2_HEADER_LENGTH_MM(mmd, pkt));
+}
+
+
+/** Return a pointer to the packet's L3 (generally, the IP) header.
+ * @ingroup egress
+ *
+ * Note that we guarantee word alignment of the L3 header.
+ *
+ * @param[in] mmd Pointer to packet's minimal metadata.
+ * @param[in] pkt Packet on which to operate.
+ * @return A pointer to the packet's L3 header.
+ */
+static __inline unsigned char*
+NETIO_PKT_L3_DATA_MM(netio_pkt_minimal_metadata_t* mmd, netio_pkt_t* pkt)
+{
+  return _NETIO_PKT_BASE(pkt) + mmd->l3_offset;
+}
+
+
+/** Return a pointer to the packet's L2 (Ethernet) header.
+ * @ingroup egress
+ *
+ * @param[in] mmd Pointer to packet's minimal metadata.
+ * @param[in] pkt Packet on which to operate.
+ * @return A pointer to start of the packet.
+ */
+static __inline unsigned char*
+NETIO_PKT_L2_DATA_MM(netio_pkt_minimal_metadata_t* mmd, netio_pkt_t* pkt)
+{
+  return _NETIO_PKT_BASE(pkt) + mmd->l2_offset;
+}
+
+
+/** Retrieve the status of a packet and any errors that may have occurred
+ * during ingress processing (length mismatches, CRC errors, etc.).
+ * @ingroup ingress
+ *
+ * Note that packets for which @ref NETIO_PKT_ETHERTYPE_RECOGNIZED()
+ * returns zero are always reported as underlength, as there is no a priori
+ * means to determine their length.  Normally, applications should use
+ * @ref NETIO_PKT_BAD() instead of explicitly checking status with this
+ * function.
+ *
+ * @param[in] pkt Packet on which to operate.
+ * @return The packet's status.
+ */
+static __inline netio_pkt_status_t
+NETIO_PKT_STATUS(netio_pkt_t* pkt)
+{
+  netio_assert(!pkt->__packet.bits.__minimal);
+
+  return (netio_pkt_status_t) __NETIO_PKT_NOTIF_HEADER(pkt).bits.__status;
+}
+
+
+/** Report whether a packet is bad (i.e., was shorter than expected based on
+ *  its headers, or had a bad CRC).
+ * @ingroup ingress
+ *
+ * Note that this function does not verify L3 or L4 checksums.
+ *
+ * @param[in] pkt Packet on which to operate.
+ * @return Nonzero if the packet is bad and should be discarded.
+ */
+static __inline unsigned int
+NETIO_PKT_BAD(netio_pkt_t* pkt)
+{
+  netio_pkt_metadata_t* mda = NETIO_PKT_METADATA(pkt);
+
+  return NETIO_PKT_BAD_M(mda, pkt);
+}
+
+
+/** Return the length of the packet's custom header.
+ *  A custom header may or may not be present, depending upon the IPP; its
+ *  contents and alignment are also IPP-dependent.  Currently, none of the
+ *  standard IPPs supplied by Tilera produce a custom header.  If present,
+ *  the custom header precedes the L2 header in the packet buffer.
+ * @ingroup pktfuncs
+ *
+ * @param[in] pkt Packet on which to operate.
+ * @return The length of the packet's custom header, in bytes.
+ */
+static __inline netio_size_t
+NETIO_PKT_CUSTOM_HEADER_LENGTH(netio_pkt_t* pkt)
+{
+  netio_pkt_metadata_t* mda = NETIO_PKT_METADATA(pkt);
+
+  return NETIO_PKT_CUSTOM_HEADER_LENGTH_M(mda, pkt);
+}
+
+
+/** Return the length of the packet, starting with the custom header.
+ *  A custom header may or may not be present, depending upon the IPP; its
+ *  contents and alignment are also IPP-dependent.  Currently, none of the
+ *  standard IPPs supplied by Tilera produce a custom header.  If present,
+ *  the custom header precedes the L2 header in the packet buffer.
+ * @ingroup pktfuncs
+ *
+ * @param[in] pkt Packet on which to operate.
+ * @return  The length of the packet, in bytes.
+ */
+static __inline netio_size_t
+NETIO_PKT_CUSTOM_LENGTH(netio_pkt_t* pkt)
+{
+  netio_pkt_metadata_t* mda = NETIO_PKT_METADATA(pkt);
+
+  return NETIO_PKT_CUSTOM_LENGTH_M(mda, pkt);
+}
+
+
+/** Return a pointer to the packet's custom header.
+ *  A custom header may or may not be present, depending upon the IPP; its
+ *  contents and alignment are also IPP-dependent.  Currently, none of the
+ *  standard IPPs supplied by Tilera produce a custom header.  If present,
+ *  the custom header precedes the L2 header in the packet buffer.
+ * @ingroup pktfuncs
+ *
+ * @param[in] pkt Packet on which to operate.
+ * @return A pointer to start of the packet.
+ */
+static __inline unsigned char*
+NETIO_PKT_CUSTOM_DATA(netio_pkt_t* pkt)
+{
+  netio_pkt_metadata_t* mda = NETIO_PKT_METADATA(pkt);
+
+  return NETIO_PKT_CUSTOM_DATA_M(mda, pkt);
+}
+
+
+/** Return the length of the packet's L2 (Ethernet plus VLAN or SNAP) header.
+ * @ingroup pktfuncs
+ *
+ * @param[in] pkt Packet on which to operate.
+ * @return The length of the packet's L2 header, in bytes.
+ */
+static __inline netio_size_t
+NETIO_PKT_L2_HEADER_LENGTH(netio_pkt_t* pkt)
+{
+  if (NETIO_PKT_IS_MINIMAL(pkt))
+  {
+    netio_pkt_minimal_metadata_t* mmd = NETIO_PKT_MINIMAL_METADATA(pkt);
+
+    return NETIO_PKT_L2_HEADER_LENGTH_MM(mmd, pkt);
+  }
+  else
+  {
+    netio_pkt_metadata_t* mda = NETIO_PKT_METADATA(pkt);
+
+    return NETIO_PKT_L2_HEADER_LENGTH_M(mda, pkt);
+  }
+}
+
+
+/** Return the length of the packet, starting with the L2 (Ethernet) header.
+ * @ingroup pktfuncs
+ *
+ * @param[in] pkt Packet on which to operate.
+ * @return  The length of the packet, in bytes.
+ */
+static __inline netio_size_t
+NETIO_PKT_L2_LENGTH(netio_pkt_t* pkt)
+{
+  if (NETIO_PKT_IS_MINIMAL(pkt))
+  {
+    netio_pkt_minimal_metadata_t* mmd = NETIO_PKT_MINIMAL_METADATA(pkt);
+
+    return NETIO_PKT_L2_LENGTH_MM(mmd, pkt);
+  }
+  else
+  {
+    netio_pkt_metadata_t* mda = NETIO_PKT_METADATA(pkt);
+
+    return NETIO_PKT_L2_LENGTH_M(mda, pkt);
+  }
+}
+
+
+/** Return a pointer to the packet's L2 (Ethernet) header.
+ * @ingroup pktfuncs
+ *
+ * @param[in] pkt Packet on which to operate.
+ * @return A pointer to start of the packet.
+ */
+static __inline unsigned char*
+NETIO_PKT_L2_DATA(netio_pkt_t* pkt)
+{
+  if (NETIO_PKT_IS_MINIMAL(pkt))
+  {
+    netio_pkt_minimal_metadata_t* mmd = NETIO_PKT_MINIMAL_METADATA(pkt);
+
+    return NETIO_PKT_L2_DATA_MM(mmd, pkt);
+  }
+  else
+  {
+    netio_pkt_metadata_t* mda = NETIO_PKT_METADATA(pkt);
+
+    return NETIO_PKT_L2_DATA_M(mda, pkt);
+  }
+}
+
+
+/** Retrieve the length of the packet, starting with the L3 (generally, the IP)
+ * header.
+ * @ingroup pktfuncs
+ *
+ * @param[in] pkt Packet on which to operate.
+ * @return Length of the packet's L3 header and data, in bytes.
+ */
+static __inline netio_size_t
+NETIO_PKT_L3_LENGTH(netio_pkt_t* pkt)
+{
+  if (NETIO_PKT_IS_MINIMAL(pkt))
+  {
+    netio_pkt_minimal_metadata_t* mmd = NETIO_PKT_MINIMAL_METADATA(pkt);
+
+    return NETIO_PKT_L3_LENGTH_MM(mmd, pkt);
+  }
+  else
+  {
+    netio_pkt_metadata_t* mda = NETIO_PKT_METADATA(pkt);
+
+    return NETIO_PKT_L3_LENGTH_M(mda, pkt);
+  }
+}
+
+
+/** Return a pointer to the packet's L3 (generally, the IP) header.
+ * @ingroup pktfuncs
+ *
+ * Note that we guarantee word alignment of the L3 header.
+ *
+ * @param[in] pkt Packet on which to operate.
+ * @return A pointer to the packet's L3 header.
+ */
+static __inline unsigned char*
+NETIO_PKT_L3_DATA(netio_pkt_t* pkt)
+{
+  if (NETIO_PKT_IS_MINIMAL(pkt))
+  {
+    netio_pkt_minimal_metadata_t* mmd = NETIO_PKT_MINIMAL_METADATA(pkt);
+
+    return NETIO_PKT_L3_DATA_MM(mmd, pkt);
+  }
+  else
+  {
+    netio_pkt_metadata_t* mda = NETIO_PKT_METADATA(pkt);
+
+    return NETIO_PKT_L3_DATA_M(mda, pkt);
+  }
+}
+
+
+/** Return the ordinal of the packet.
+ * @ingroup ingress
+ *
+ * Each packet is given an ordinal number when it is delivered by the IPP.
+ * In the medium term, the ordinal is unique and monotonically increasing,
+ * being incremented by 1 for each packet; the ordinal of the first packet
+ * delivered after the IPP starts is zero.  (Since the ordinal is of finite
+ * size, given enough input packets, it will eventually wrap around to zero;
+ * in the long term, therefore, ordinals are not unique.)  The ordinals
+ * handed out by different IPPs are not disjoint, so two packets from
+ * different IPPs may have identical ordinals.  Packets dropped by the
+ * IPP or by the I/O shim are not assigned ordinals.
+ *
+ *
+ * @param[in] pkt Packet on which to operate.
+ * @return The packet's per-IPP packet ordinal.
+ */
+static __inline unsigned int
+NETIO_PKT_ORDINAL(netio_pkt_t* pkt)
+{
+  netio_pkt_metadata_t* mda = NETIO_PKT_METADATA(pkt);
+
+  return NETIO_PKT_ORDINAL_M(mda, pkt);
+}
+
+
+/** Return the per-group ordinal of the packet.
+ * @ingroup ingress
+ *
+ * Each packet is given a per-group ordinal number when it is
+ * delivered by the IPP. By default, the group is the packet's VLAN,
+ * although IPP can be recompiled to use different values.  In
+ * the medium term, the ordinal is unique and monotonically
+ * increasing, being incremented by 1 for each packet; the ordinal of
+ * the first packet distributed to a particular group is zero.
+ * (Since the ordinal is of finite size, given enough input packets,
+ * it will eventually wrap around to zero; in the long term,
+ * therefore, ordinals are not unique.)  The ordinals handed out by
+ * different IPPs are not disjoint, so two packets from different IPPs
+ * may have identical ordinals; similarly, packets distributed to
+ * different groups may have identical ordinals.  Packets dropped by
+ * the IPP or by the I/O shim are not assigned ordinals.
+ *
+ * @param[in] pkt Packet on which to operate.
+ * @return The packet's per-IPP, per-group ordinal.
+ */
+static __inline unsigned int
+NETIO_PKT_GROUP_ORDINAL(netio_pkt_t* pkt)
+{
+  netio_pkt_metadata_t* mda = NETIO_PKT_METADATA(pkt);
+
+  return NETIO_PKT_GROUP_ORDINAL_M(mda, pkt);
+}
+
+
+/** Return the VLAN ID assigned to the packet.
+ * @ingroup ingress
+ *
+ * This is usually also contained within the packet header.  If the packet
+ * does not have a VLAN tag, the VLAN ID returned by this function is zero.
+ *
+ * @param[in] pkt Packet on which to operate.
+ * @return The packet's VLAN ID.
+ */
+static __inline unsigned short
+NETIO_PKT_VLAN_ID(netio_pkt_t* pkt)
+{
+  netio_pkt_metadata_t* mda = NETIO_PKT_METADATA(pkt);
+
+  return NETIO_PKT_VLAN_ID_M(mda, pkt);
+}
+
+
+/** Return the ethertype of the packet.
+ * @ingroup ingress
+ *
+ * This value is reliable if @ref NETIO_PKT_ETHERTYPE_RECOGNIZED()
+ * returns true, and otherwise, may not be well defined.
+ *
+ * @param[in] pkt Packet on which to operate.
+ * @return The packet's ethertype.
+ */
+static __inline unsigned short
+NETIO_PKT_ETHERTYPE(netio_pkt_t* pkt)
+{
+  netio_pkt_metadata_t* mda = NETIO_PKT_METADATA(pkt);
+
+  return NETIO_PKT_ETHERTYPE_M(mda, pkt);
+}
+
+
+/** Return the flow hash computed on the packet.
+ * @ingroup ingress
+ *
+ * For TCP and UDP packets, this hash is calculated by hashing together
+ * the "5-tuple" values, specifically the source IP address, destination
+ * IP address, protocol type, source port and destination port.
+ * The hash value is intended to be helpful for millions of distinct
+ * flows.
+ *
+ * For IPv4 or IPv6 packets which are neither TCP nor UDP, the flow hash is
+ * derived by hashing together the source and destination IP addresses.
+ *
+ * For MPLS-encapsulated packets, the flow hash is derived by hashing
+ * the first MPLS label.
+ *
+ * For all other packets the flow hash is computed from the source
+ * and destination Ethernet addresses.
+ *
+ * The hash is symmetric, meaning it produces the same value if the
+ * source and destination are swapped. The only exceptions are
+ * tunneling protocols 0x04 (IP in IP Encapsulation), 0x29 (Simple
+ * Internet Protocol), 0x2F (General Routing Encapsulation) and 0x32
+ * (Encap Security Payload), which use only the destination address
+ * since the source address is not meaningful.
+ *
+ * @param[in] pkt Packet on which to operate.
+ * @return The packet's 32-bit flow hash.
+ */
+static __inline unsigned int
+NETIO_PKT_FLOW_HASH(netio_pkt_t* pkt)
+{
+  netio_pkt_metadata_t* mda = NETIO_PKT_METADATA(pkt);
+
+  return NETIO_PKT_FLOW_HASH_M(mda, pkt);
+}
+
+
+/** Return the first word of "user data" for the packet.
+ *
+ * The contents of the user data words depend on the IPP.
+ *
+ * When using the standard ipp1, ipp2, or ipp4 sub-drivers, the first
+ * word of user data contains the least significant bits of the 64-bit
+ * arrival cycle count (see @c get_cycle_count_low()).
+ *
+ * See the <em>System Programmer's Guide</em> for details.
+ *
+ * @ingroup ingress
+ *
+ * @param[in] pkt Packet on which to operate.
+ * @return The packet's first word of "user data".
+ */
+static __inline unsigned int
+NETIO_PKT_USER_DATA_0(netio_pkt_t* pkt)
+{
+  netio_pkt_metadata_t* mda = NETIO_PKT_METADATA(pkt);
+
+  return NETIO_PKT_USER_DATA_0_M(mda, pkt);
+}
+
+
+/** Return the second word of "user data" for the packet.
+ *
+ * The contents of the user data words depend on the IPP.
+ *
+ * When using the standard ipp1, ipp2, or ipp4 sub-drivers, the second
+ * word of user data contains the most significant bits of the 64-bit
+ * arrival cycle count (see @c get_cycle_count_high()).
+ *
+ * See the <em>System Programmer's Guide</em> for details.
+ *
+ * @ingroup ingress
+ *
+ * @param[in] pkt Packet on which to operate.
+ * @return The packet's second word of "user data".
+ */
+static __inline unsigned int
+NETIO_PKT_USER_DATA_1(netio_pkt_t* pkt)
+{
+  netio_pkt_metadata_t* mda = NETIO_PKT_METADATA(pkt);
+
+  return NETIO_PKT_USER_DATA_1_M(mda, pkt);
+}
+
+
+/** Determine whether the L4 (TCP/UDP) checksum was calculated.
+ * @ingroup ingress
+ *
+ * @param[in] pkt Packet on which to operate.
+ * @return Nonzero if the L4 checksum was calculated.
+ */
+static __inline unsigned int
+NETIO_PKT_L4_CSUM_CALCULATED(netio_pkt_t* pkt)
+{
+  netio_pkt_metadata_t* mda = NETIO_PKT_METADATA(pkt);
+
+  return NETIO_PKT_L4_CSUM_CALCULATED_M(mda, pkt);
+}
+
+
+/** Determine whether the L4 (TCP/UDP) checksum was calculated and found to
+ *  be correct.
+ * @ingroup ingress
+ *
+ * @param[in] pkt Packet on which to operate.
+ * @return Nonzero if the checksum was calculated and is correct.
+ */
+static __inline unsigned int
+NETIO_PKT_L4_CSUM_CORRECT(netio_pkt_t* pkt)
+{
+  netio_pkt_metadata_t* mda = NETIO_PKT_METADATA(pkt);
+
+  return NETIO_PKT_L4_CSUM_CORRECT_M(mda, pkt);
+}
+
+
+/** Determine whether the L3 (IP) checksum was calculated.
+ * @ingroup ingress
+ *
+ * @param[in] pkt Packet on which to operate.
+ * @return Nonzero if the L3 (IP) checksum was calculated.
+*/
+static __inline unsigned int
+NETIO_PKT_L3_CSUM_CALCULATED(netio_pkt_t* pkt)
+{
+  netio_pkt_metadata_t* mda = NETIO_PKT_METADATA(pkt);
+
+  return NETIO_PKT_L3_CSUM_CALCULATED_M(mda, pkt);
+}
+
+
+/** Determine whether the L3 (IP) checksum was calculated and found to be
+ *  correct.
+ * @ingroup ingress
+ *
+ * @param[in] pkt Packet on which to operate.
+ * @return Nonzero if the checksum was calculated and is correct.
+ */
+static __inline unsigned int
+NETIO_PKT_L3_CSUM_CORRECT(netio_pkt_t* pkt)
+{
+  netio_pkt_metadata_t* mda = NETIO_PKT_METADATA(pkt);
+
+  return NETIO_PKT_L3_CSUM_CORRECT_M(mda, pkt);
+}
+
+
+/** Determine whether the Ethertype was recognized and L3 packet data was
+ *  processed.
+ * @ingroup ingress
+ *
+ * @param[in] pkt Packet on which to operate.
+ * @return Nonzero if the Ethertype was recognized and L3 packet data was
+ *   processed.
+ */
+static __inline unsigned int
+NETIO_PKT_ETHERTYPE_RECOGNIZED(netio_pkt_t* pkt)
+{
+  netio_pkt_metadata_t* mda = NETIO_PKT_METADATA(pkt);
+
+  return NETIO_PKT_ETHERTYPE_RECOGNIZED_M(mda, pkt);
+}
+
+
+/** Set an egress packet's L2 length, using a metadata pointer to speed the
+ * computation.
+ * @ingroup egress
+ *
+ * @param[in,out] mmd Pointer to packet's minimal metadata.
+ * @param[in] pkt Packet on which to operate.
+ * @param[in] len Packet L2 length, in bytes.
+ */
+static __inline void
+NETIO_PKT_SET_L2_LENGTH_MM(netio_pkt_minimal_metadata_t* mmd, netio_pkt_t* pkt,
+                           int len)
+{
+  mmd->l2_length = len;
+}
+
+
+/** Set an egress packet's L2 length.
+ * @ingroup egress
+ *
+ * @param[in,out] pkt Packet on which to operate.
+ * @param[in] len Packet L2 length, in bytes.
+ */
+static __inline void
+NETIO_PKT_SET_L2_LENGTH(netio_pkt_t* pkt, int len)
+{
+  netio_pkt_minimal_metadata_t* mmd = NETIO_PKT_MINIMAL_METADATA(pkt);
+
+  NETIO_PKT_SET_L2_LENGTH_MM(mmd, pkt, len);
+}
+
+
+/** Set an egress packet's L2 header length, using a metadata pointer to
+ *  speed the computation.
+ * @ingroup egress
+ *
+ * It is not normally necessary to call this routine; only the L2 length,
+ * not the header length, is needed to transmit a packet.  It may be useful if
+ * the egress packet will later be processed by code which expects to use
+ * functions like @ref NETIO_PKT_L3_DATA() to get a pointer to the L3 payload.
+ *
+ * @param[in,out] mmd Pointer to packet's minimal metadata.
+ * @param[in] pkt Packet on which to operate.
+ * @param[in] len Packet L2 header length, in bytes.
+ */
+static __inline void
+NETIO_PKT_SET_L2_HEADER_LENGTH_MM(netio_pkt_minimal_metadata_t* mmd,
+                                  netio_pkt_t* pkt, int len)
+{
+  mmd->l3_offset = mmd->l2_offset + len;
+}
+
+
+/** Set an egress packet's L2 header length.
+ * @ingroup egress
+ *
+ * It is not normally necessary to call this routine; only the L2 length,
+ * not the header length, is needed to transmit a packet.  It may be useful if
+ * the egress packet will later be processed by code which expects to use
+ * functions like @ref NETIO_PKT_L3_DATA() to get a pointer to the L3 payload.
+ *
+ * @param[in,out] pkt Packet on which to operate.
+ * @param[in] len Packet L2 header length, in bytes.
+ */
+static __inline void
+NETIO_PKT_SET_L2_HEADER_LENGTH(netio_pkt_t* pkt, int len)
+{
+  netio_pkt_minimal_metadata_t* mmd = NETIO_PKT_MINIMAL_METADATA(pkt);
+
+  NETIO_PKT_SET_L2_HEADER_LENGTH_MM(mmd, pkt, len);
+}
+
+
+/** Set up an egress packet for hardware checksum computation, using a
+ *  metadata pointer to speed the operation.
+ * @ingroup egress
+ *
+ *  NetIO provides the ability to automatically calculate a standard
+ *  16-bit Internet checksum on transmitted packets.  The application
+ *  may specify the point in the packet where the checksum starts, the
+ *  number of bytes to be checksummed, and the two bytes in the packet
+ *  which will be replaced with the completed checksum.  (If the range
+ *  of bytes to be checksummed includes the bytes to be replaced, the
+ *  initial values of those bytes will be included in the checksum.)
+ *
+ *  For some protocols, the packet checksum covers data which is not present
+ *  in the packet, or is at least not contiguous to the main data payload.
+ *  For instance, the TCP checksum includes a "pseudo-header" which includes
+ *  the source and destination IP addresses of the packet.  To accommodate
+ *  this, the checksum engine may be "seeded" with an initial value, which
+ *  the application would need to compute based on the specific protocol's
+ *  requirements.  Note that the seed is given in host byte order (little-
+ *  endian), not network byte order (big-endian); code written to compute a
+ *  pseudo-header checksum in network byte order will need to byte-swap it
+ *  before use as the seed.
+ *
+ *  Note that the checksum is computed as part of the transmission process,
+ *  so it will not be present in the packet upon completion of this routine.
+ *
+ * @param[in,out] mmd Pointer to packet's minimal metadata.
+ * @param[in] pkt Packet on which to operate.
+ * @param[in] start Offset within L2 packet of the first byte to include in
+ *   the checksum.
+ * @param[in] length Number of bytes to include in the checksum.
+ *   the checksum.
+ * @param[in] location Offset within L2 packet of the first of the two bytes
+ *   to be replaced with the calculated checksum.
+ * @param[in] seed Initial value of the running checksum before any of the
+ *   packet data is added.
+ */
+static __inline void
+NETIO_PKT_DO_EGRESS_CSUM_MM(netio_pkt_minimal_metadata_t* mmd,
+                            netio_pkt_t* pkt, int start, int length,
+                            int location, uint16_t seed)
+{
+  mmd->csum_start = start;
+  mmd->csum_length = length;
+  mmd->csum_location = location;
+  mmd->csum_seed = seed;
+  mmd->flags |= _NETIO_PKT_NEED_EDMA_CSUM_MASK;
+}
+
+
+/** Set up an egress packet for hardware checksum computation.
+ * @ingroup egress
+ *
+ *  NetIO provides the ability to automatically calculate a standard
+ *  16-bit Internet checksum on transmitted packets.  The application
+ *  may specify the point in the packet where the checksum starts, the
+ *  number of bytes to be checksummed, and the two bytes in the packet
+ *  which will be replaced with the completed checksum.  (If the range
+ *  of bytes to be checksummed includes the bytes to be replaced, the
+ *  initial values of those bytes will be included in the checksum.)
+ *
+ *  For some protocols, the packet checksum covers data which is not present
+ *  in the packet, or is at least not contiguous to the main data payload.
+ *  For instance, the TCP checksum includes a "pseudo-header" which includes
+ *  the source and destination IP addresses of the packet.  To accommodate
+ *  this, the checksum engine may be "seeded" with an initial value, which
+ *  the application would need to compute based on the specific protocol's
+ *  requirements.  Note that the seed is given in host byte order (little-
+ *  endian), not network byte order (big-endian); code written to compute a
+ *  pseudo-header checksum in network byte order will need to byte-swap it
+ *  before use as the seed.
+ *
+ *  Note that the checksum is computed as part of the transmission process,
+ *  so it will not be present in the packet upon completion of this routine.
+ *
+ * @param[in,out] pkt Packet on which to operate.
+ * @param[in] start Offset within L2 packet of the first byte to include in
+ *   the checksum.
+ * @param[in] length Number of bytes to include in the checksum.
+ *   the checksum.
+ * @param[in] location Offset within L2 packet of the first of the two bytes
+ *   to be replaced with the calculated checksum.
+ * @param[in] seed Initial value of the running checksum before any of the
+ *   packet data is added.
+ */
+static __inline void
+NETIO_PKT_DO_EGRESS_CSUM(netio_pkt_t* pkt, int start, int length,
+                         int location, uint16_t seed)
+{
+  netio_pkt_minimal_metadata_t* mmd = NETIO_PKT_MINIMAL_METADATA(pkt);
+
+  NETIO_PKT_DO_EGRESS_CSUM_MM(mmd, pkt, start, length, location, seed);
+}
+
+
+/** Return the number of bytes which could be prepended to a packet, using a
+ *  metadata pointer to speed the operation.
+ *  See @ref netio_populate_prepend_buffer() to get a full description of
+ *  prepending.
+ *
+ * @param[in,out] mda Pointer to packet's standard metadata.
+ * @param[in] pkt Packet on which to operate.
+ */
+static __inline int
+NETIO_PKT_PREPEND_AVAIL_M(netio_pkt_metadata_t* mda, netio_pkt_t* pkt)
+{
+  return (pkt->__packet.bits.__offset << 6) +
+         NETIO_PKT_CUSTOM_HEADER_LENGTH_M(mda, pkt);
+}
+
+
+/** Return the number of bytes which could be prepended to a packet, using a
+ *  metadata pointer to speed the operation.
+ *  See @ref netio_populate_prepend_buffer() to get a full description of
+ *  prepending.
+ * @ingroup egress
+ *
+ * @param[in,out] mmd Pointer to packet's minimal metadata.
+ * @param[in] pkt Packet on which to operate.
+ */
+static __inline int
+NETIO_PKT_PREPEND_AVAIL_MM(netio_pkt_minimal_metadata_t* mmd, netio_pkt_t* pkt)
+{
+  return (pkt->__packet.bits.__offset << 6) + mmd->l2_offset;
+}
+
+
+/** Return the number of bytes which could be prepended to a packet.
+ *  See @ref netio_populate_prepend_buffer() to get a full description of
+ *  prepending.
+ * @ingroup egress
+ *
+ * @param[in] pkt Packet on which to operate.
+ */
+static __inline int
+NETIO_PKT_PREPEND_AVAIL(netio_pkt_t* pkt)
+{
+  if (NETIO_PKT_IS_MINIMAL(pkt))
+  {
+    netio_pkt_minimal_metadata_t* mmd = NETIO_PKT_MINIMAL_METADATA(pkt);
+
+    return NETIO_PKT_PREPEND_AVAIL_MM(mmd, pkt);
+  }
+  else
+  {
+    netio_pkt_metadata_t* mda = NETIO_PKT_METADATA(pkt);
+
+    return NETIO_PKT_PREPEND_AVAIL_M(mda, pkt);
+  }
+}
+
+
+/** Flush a packet's minimal metadata from the cache, using a metadata pointer
+ *  to speed the operation.
+ * @ingroup egress
+ *
+ * @param[in] mmd Pointer to packet's minimal metadata.
+ * @param[in] pkt Packet on which to operate.
+ */
+static __inline void
+NETIO_PKT_FLUSH_MINIMAL_METADATA_MM(netio_pkt_minimal_metadata_t* mmd,
+                                    netio_pkt_t* pkt)
+{
+}
+
+
+/** Invalidate a packet's minimal metadata from the cache, using a metadata
+ *  pointer to speed the operation.
+ * @ingroup egress
+ *
+ * @param[in] mmd Pointer to packet's minimal metadata.
+ * @param[in] pkt Packet on which to operate.
+ */
+static __inline void
+NETIO_PKT_INV_MINIMAL_METADATA_MM(netio_pkt_minimal_metadata_t* mmd,
+                                  netio_pkt_t* pkt)
+{
+}
+
+
+/** Flush and then invalidate a packet's minimal metadata from the cache,
+ *  using a metadata pointer to speed the operation.
+ * @ingroup egress
+ *
+ * @param[in] mmd Pointer to packet's minimal metadata.
+ * @param[in] pkt Packet on which to operate.
+ */
+static __inline void
+NETIO_PKT_FLUSH_INV_MINIMAL_METADATA_MM(netio_pkt_minimal_metadata_t* mmd,
+                                        netio_pkt_t* pkt)
+{
+}
+
+
+/** Flush a packet's metadata from the cache, using a metadata pointer
+ *  to speed the operation.
+ * @ingroup ingress
+ *
+ * @param[in] mda Pointer to packet's minimal metadata.
+ * @param[in] pkt Packet on which to operate.
+ */
+static __inline void
+NETIO_PKT_FLUSH_METADATA_M(netio_pkt_metadata_t* mda, netio_pkt_t* pkt)
+{
+}
+
+
+/** Invalidate a packet's metadata from the cache, using a metadata
+ *  pointer to speed the operation.
+ * @ingroup ingress
+ *
+ * @param[in] mda Pointer to packet's metadata.
+ * @param[in] pkt Packet on which to operate.
+ */
+static __inline void
+NETIO_PKT_INV_METADATA_M(netio_pkt_metadata_t* mda, netio_pkt_t* pkt)
+{
+}
+
+
+/** Flush and then invalidate a packet's metadata from the cache,
+ *  using a metadata pointer to speed the operation.
+ * @ingroup ingress
+ *
+ * @param[in] mda Pointer to packet's metadata.
+ * @param[in] pkt Packet on which to operate.
+ */
+static __inline void
+NETIO_PKT_FLUSH_INV_METADATA_M(netio_pkt_metadata_t* mda, netio_pkt_t* pkt)
+{
+}
+
+
+/** Flush a packet's minimal metadata from the cache.
+ * @ingroup egress
+ *
+ * @param[in] pkt Packet on which to operate.
+ */
+static __inline void
+NETIO_PKT_FLUSH_MINIMAL_METADATA(netio_pkt_t* pkt)
+{
+}
+
+
+/** Invalidate a packet's minimal metadata from the cache.
+ * @ingroup egress
+ *
+ * @param[in] pkt Packet on which to operate.
+ */
+static __inline void
+NETIO_PKT_INV_MINIMAL_METADATA(netio_pkt_t* pkt)
+{
+}
+
+
+/** Flush and then invalidate a packet's minimal metadata from the cache.
+ * @ingroup egress
+ *
+ * @param[in] pkt Packet on which to operate.
+ */
+static __inline void
+NETIO_PKT_FLUSH_INV_MINIMAL_METADATA(netio_pkt_t* pkt)
+{
+}
+
+
+/** Flush a packet's metadata from the cache.
+ * @ingroup ingress
+ *
+ * @param[in] pkt Packet on which to operate.
+ */
+static __inline void
+NETIO_PKT_FLUSH_METADATA(netio_pkt_t* pkt)
+{
+}
+
+
+/** Invalidate a packet's metadata from the cache.
+ * @ingroup ingress
+ *
+ * @param[in] pkt Packet on which to operate.
+ */
+static __inline void
+NETIO_PKT_INV_METADATA(netio_pkt_t* pkt)
+{
+}
+
+
+/** Flush and then invalidate a packet's metadata from the cache.
+ * @ingroup ingress
+ *
+ * @param[in] pkt Packet on which to operate.
+ */
+static __inline void
+NETIO_PKT_FLUSH_INV_METADATA(netio_pkt_t* pkt)
+{
+}
+
+/** Number of NUMA nodes we can distribute buffers to.
+ * @ingroup setup */
+#define NETIO_NUM_NODE_WEIGHTS  16
+
+/**
+ * @brief An object for specifying the characteristics of NetIO communication
+ * endpoint.
+ *
+ * @ingroup setup
+ *
+ * The @ref netio_input_register() function uses this structure to define
+ * how an application tile will communicate with an IPP.
+ *
+ *
+ * Future updates to NetIO may add new members to this structure,
+ * which can affect the success of the registration operation.  Thus,
+ * if dynamically initializing the structure, applications are urged to
+ * zero it out first, for example:
+ *
+ * @code
+ * netio_input_config_t config;
+ * memset(&config, 0, sizeof (config));
+ * config.flags = NETIO_RECV | NETIO_XMIT_CSUM | NETIO_TAG_NONE;
+ * config.num_receive_packets = NETIO_MAX_RECEIVE_PKTS;
+ * config.queue_id = 0;
+ *     .
+ *     .
+ *     .
+ * @endcode
+ *
+ * since that guarantees that any unused structure members, including
+ * members which did not exist when the application was first developed,
+ * will not have unexpected values.
+ *
+ * If statically initializing the structure, we strongly recommend use of
+ * C99-style named initializers, for example:
+ *
+ * @code
+ * netio_input_config_t config = {
+ *    .flags = NETIO_RECV | NETIO_XMIT_CSUM | NETIO_TAG_NONE,
+ *    .num_receive_packets = NETIO_MAX_RECEIVE_PKTS,
+ *    .queue_id = 0,
+ * },
+ * @endcode
+ *
+ * instead of the old-style structure initialization:
+ *
+ * @code
+ * // Bad example! Currently equivalent to the above, but don't do this.
+ * netio_input_config_t config = {
+ *    NETIO_RECV | NETIO_XMIT_CSUM | NETIO_TAG_NONE, NETIO_MAX_RECEIVE_PKTS, 0
+ * },
+ * @endcode
+ *
+ * since the C99 style requires no changes to the code if elements of the
+ * config structure are rearranged.  (It also makes the initialization much
+ * easier to understand.)
+ *
+ * Except for items which address a particular tile's transmit or receive
+ * characteristics, such as the ::NETIO_RECV flag, applications are advised
+ * to specify the same set of configuration data on all registrations.
+ * This prevents differing results if multiple tiles happen to do their
+ * registration operations in a different order on different invocations of
+ * the application.  This is particularly important for things like link
+ * management flags, and buffer size and homing specifications.
+ *
+ * Unless the ::NETIO_FIXED_BUFFER_VA flag is specified in flags, the NetIO
+ * buffer pool is automatically created and mapped into the application's
+ * virtual address space at an address chosen by the operating system,
+ * using the common memory (cmem) facility in the Tilera Multicore
+ * Components library.  The cmem facility allows multiple processes to gain
+ * access to shared memory which is mapped into each process at an
+ * identical virtual address.  In order for this to work, the processes
+ * must have a common ancestor, which must create the common memory using
+ * tmc_cmem_init().
+ *
+ * In programs using the iLib process creation API, or in programs which use
+ * only one process (which include programs using the pthreads library),
+ * tmc_cmem_init() is called automatically.  All other applications
+ * must call it explicitly, before any child processes which might call
+ * netio_input_register() are created.
+ */
+typedef struct
+{
+  /** Registration characteristics.
+
+      This value determines several characteristics of the registration;
+      flags for different types of behavior are ORed together to make the
+      final flag value.  Generally applications should specify exactly
+      one flag from each of the following categories:
+
+      - Whether the application will be receiving packets on this queue
+        (::NETIO_RECV or ::NETIO_NO_RECV).
+
+      - Whether the application will be transmitting packets on this queue,
+        and if so, whether it will request egress checksum calculation
+        (::NETIO_XMIT, ::NETIO_XMIT_CSUM, or ::NETIO_NO_XMIT).  It is
+        legal to call netio_get_buffer() without one of the XMIT flags,
+        as long as ::NETIO_RECV is specified; in this case, the retrieved
+        buffers must be passed to another tile for transmission.
+
+      - Whether the application expects any vendor-specific tags in
+        its packets' L2 headers (::NETIO_TAG_NONE, ::NETIO_TAG_BRCM,
+        or ::NETIO_TAG_MRVL).  This must match the configuration of the
+        target IPP.
+
+      To accommodate applications written to previous versions of the NetIO
+      interface, none of the flags above are currently required; if omitted,
+      NetIO behaves more or less as if ::NETIO_RECV | ::NETIO_XMIT_CSUM |
+      ::NETIO_TAG_NONE were used.  However, explicit specification of
+      the relevant flags allows NetIO to do a better job of resource
+      allocation, allows earlier detection of certain configuration errors,
+      and may enable advanced features or higher performance in the future,
+      so their use is strongly recommended.
+
+      Note that specifying ::NETIO_NO_RECV along with ::NETIO_NO_XMIT
+      is a special case, intended primarily for use by programs which
+      retrieve network statistics or do link management operations.
+      When these flags are both specified, the resulting queue may not
+      be used with NetIO routines other than netio_get(), netio_set(),
+      and netio_input_unregister().  See @ref link for more information
+      on link management.
+
+      Other flags are optional; their use is described below.
+  */
+  int flags;
+
+  /** Interface name.  This is a string which identifies the specific
+      Ethernet controller hardware to be used.  The format of the string
+      is a device type and a device index, separated by a slash; so,
+      the first 10 Gigabit Ethernet controller is named "xgbe/0", while
+      the second 10/100/1000 Megabit Ethernet controller is named "gbe/1".
+   */
+  const char* interface;
+
+  /** Receive packet queue size.  This specifies the maximum number
+      of ingress packets that can be received on this queue without
+      being retrieved by @ref netio_get_packet().  If the IPP's distribution
+      algorithm calls for a packet to be sent to this queue, and this
+      number of packets are already pending there, the new packet
+      will either be discarded, or sent to another tile registered
+      for the same queue_id (see @ref drops).  This value must
+      be at least ::NETIO_MIN_RECEIVE_PKTS, can always be at least
+      ::NETIO_MAX_RECEIVE_PKTS, and may be larger than that on certain
+      interfaces.
+   */
+  int num_receive_packets;
+
+  /** The queue ID being requested.  Legal values for this range from 0
+      to ::NETIO_MAX_QUEUE_ID, inclusive.  ::NETIO_MAX_QUEUE_ID is always
+      greater than or equal to the number of tiles; this allows one queue
+      for each tile, plus at least one additional queue.  Some applications
+      may wish to use the additional queue as a destination for unwanted
+      packets, since packets delivered to queues for which no tiles have
+      registered are discarded.
+   */
+  unsigned int queue_id;
+
+  /** Maximum number of small send buffers to be held in the local empty
+      buffer cache.  This specifies the size of the area which holds
+      empty small egress buffers requested from the IPP but not yet
+      retrieved via @ref netio_get_buffer().  This value must be greater
+      than zero if the application will ever use @ref netio_get_buffer()
+      to allocate empty small egress buffers; it may be no larger than
+      ::NETIO_MAX_SEND_BUFFERS.  See @ref epp for more details on empty
+      buffer caching.
+   */
+  int num_send_buffers_small_total;
+
+  /** Number of small send buffers to be preallocated at registration.
+      If this value is nonzero, the specified number of empty small egress
+      buffers will be requested from the IPP during the netio_input_register
+      operation; this may speed the execution of @ref netio_get_buffer().
+      This may be no larger than @ref num_send_buffers_small_total.  See @ref
+      epp for more details on empty buffer caching.
+   */
+  int num_send_buffers_small_prealloc;
+
+  /** Maximum number of large send buffers to be held in the local empty
+      buffer cache.  This specifies the size of the area which holds empty
+      large egress buffers requested from the IPP but not yet retrieved via
+      @ref netio_get_buffer().  This value must be greater than zero if the
+      application will ever use @ref netio_get_buffer() to allocate empty
+      large egress buffers; it may be no larger than ::NETIO_MAX_SEND_BUFFERS.
+      See @ref epp for more details on empty buffer caching.
+   */
+  int num_send_buffers_large_total;
+
+  /** Number of large send buffers to be preallocated at registration.
+      If this value is nonzero, the specified number of empty large egress
+      buffers will be requested from the IPP during the netio_input_register
+      operation; this may speed the execution of @ref netio_get_buffer().
+      This may be no larger than @ref num_send_buffers_large_total.  See @ref
+      epp for more details on empty buffer caching.
+   */
+  int num_send_buffers_large_prealloc;
+
+  /** Maximum number of jumbo send buffers to be held in the local empty
+      buffer cache.  This specifies the size of the area which holds empty
+      jumbo egress buffers requested from the IPP but not yet retrieved via
+      @ref netio_get_buffer().  This value must be greater than zero if the
+      application will ever use @ref netio_get_buffer() to allocate empty
+      jumbo egress buffers; it may be no larger than ::NETIO_MAX_SEND_BUFFERS.
+      See @ref epp for more details on empty buffer caching.
+   */
+  int num_send_buffers_jumbo_total;
+
+  /** Number of jumbo send buffers to be preallocated at registration.
+      If this value is nonzero, the specified number of empty jumbo egress
+      buffers will be requested from the IPP during the netio_input_register
+      operation; this may speed the execution of @ref netio_get_buffer().
+      This may be no larger than @ref num_send_buffers_jumbo_total.  See @ref
+      epp for more details on empty buffer caching.
+   */
+  int num_send_buffers_jumbo_prealloc;
+
+  /** Total packet buffer size.  This determines the total size, in bytes,
+      of the NetIO buffer pool.  Note that the maximum number of available
+      buffers of each size is determined during hypervisor configuration
+      (see the <em>System Programmer's Guide</em> for details); this just
+      influences how much host memory is allocated for those buffers.
+
+      The buffer pool is allocated from common memory, which will be
+      automatically initialized if needed.  If your buffer pool is larger
+      than 240 MB, you might need to explicitly call @c tmc_cmem_init(),
+      as described in the Application Libraries Reference Manual (UG227).
+
+      Packet buffers are currently allocated in chunks of 16 MB; this
+      value will be rounded up to the next larger multiple of 16 MB.
+      If this value is zero, a default of 32 MB will be used; this was
+      the value used by previous versions of NetIO.  Note that taking this
+      default also affects the placement of buffers on Linux NUMA nodes.
+      See @ref buffer_node_weights for an explanation of buffer placement.
+
+      In order to successfully allocate packet buffers, Linux must have
+      available huge pages on the relevant Linux NUMA nodes.  See the
+      <em>System Programmer's Guide</em> for information on configuring
+      huge page support in Linux.
+   */
+  uint64_t total_buffer_size;
+
+  /** Buffer placement weighting factors.
+
+      This array specifies the relative amount of buffering to place
+      on each of the available Linux NUMA nodes.  This array is
+      indexed by the NUMA node, and the values in the array are
+      proportional to the amount of buffer space to allocate on that
+      node.
+
+      If memory striping is enabled in the Hypervisor, then there is
+      only one logical NUMA node (node 0). In that case, NetIO will by
+      default ignore the suggested buffer node weights, and buffers
+      will be striped across the physical memory controllers. See
+      UG209 System Programmer's Guide for a description of the
+      hypervisor option that controls memory striping.
+
+      If memory striping is disabled, then there are up to four NUMA
+      nodes, corresponding to the four DDRAM controllers in the TILE
+      processor architecture.  See UG100 Tile Processor Architecture
+      Overview for a diagram showing the location of each of the DDRAM
+      controllers relative to the tile array.
+
+      For instance, if memory striping is disabled, the following
+      configuration strucure:
+
+      @code
+      netio_input_config_t config = {
+            .
+            .
+            .
+        .total_buffer_size = 4 * 16 * 1024 * 1024;
+        .buffer_node_weights = { 1, 0, 1, 0 },
+      },
+      @endcode
+
+      would result in 32 MB of buffers being placed on controller 0, and
+      32 MB on controller 2.  (Since buffers are allocated in units of
+      16 MB, some sets of weights will not be able to be matched exactly.)
+
+      For the weights to be effective, @ref total_buffer_size must be
+      nonzero.  If @ref total_buffer_size is zero, causing the default
+      32 MB of buffer space to be used, then any specified weights will
+      be ignored, and buffers will positioned as they were in previous
+      versions of NetIO:
+
+      - For xgbe/0 and gbe/0, 16 MB of buffers will be placed on controller 1,
+        and the other 16 MB will be placed on controller 2.
+
+      - For xgbe/1 and gbe/1, 16 MB of buffers will be placed on controller 2,
+        and the other 16 MB will be placed on controller 3.
+
+      If @ref total_buffer_size is nonzero, but all weights are zero,
+      then all buffer space will be allocated on Linux NUMA node zero.
+
+      By default, the specified buffer placement is treated as a hint;
+      if sufficient free memory is not available on the specified
+      controllers, the buffers will be allocated elsewhere.  However,
+      if the ::NETIO_STRICT_HOMING flag is specified in @ref flags, then a
+      failure to allocate buffer space exactly as requested will cause the
+      registration operation to fail with an error of ::NETIO_CANNOT_HOME.
+
+      Note that maximal network performance cannot be achieved with
+      only one memory controller.
+   */
+  uint8_t buffer_node_weights[NETIO_NUM_NODE_WEIGHTS];
+
+  /** Fixed virtual address for packet buffers.  Only valid when
+      ::NETIO_FIXED_BUFFER_VA is specified in @ref flags; see the
+      description of that flag for details.
+   */
+  void* fixed_buffer_va;
+
+  /**
+      Maximum number of outstanding send packet requests.  This value is
+      only relevant when an EPP is in use; it determines the number of
+      slots in the EPP's outgoing packet queue which this tile is allowed
+      to consume, and thus the number of packets which may be sent before
+      the sending tile must wait for an acknowledgment from the EPP.
+      Modifying this value is generally only helpful when using @ref
+      netio_send_packet_vector(), where it can help improve performance by
+      allowing a single vector send operation to process more packets.
+      Typically it is not specified, and the default, which divides the
+      outgoing packet slots evenly between all tiles on the chip, is used.
+
+      If a registration asks for more outgoing packet queue slots than are
+      available, ::NETIO_TOOMANY_XMIT will be returned.  The total number
+      of packet queue slots which are available for all tiles for each EPP
+      is subject to change, but is currently ::NETIO_TOTAL_SENDS_OUTSTANDING.
+
+
+      This value is ignored if ::NETIO_XMIT is not specified in flags.
+      If you want to specify a large value here for a specific tile, you are
+      advised to specify NETIO_NO_XMIT on other, non-transmitting tiles so
+      that they do not consume a default number of packet slots.  Any tile
+      transmitting is required to have at least ::NETIO_MIN_SENDS_OUTSTANDING
+      slots allocated to it; values less than that will be silently
+      increased by the NetIO library.
+   */
+  int num_sends_outstanding;
+}
+netio_input_config_t;
+
+
+/** Registration flags; used in the @ref netio_input_config_t structure.
+ * @addtogroup setup
+ */
+/** @{ */
+
+/** Fail a registration request if we can't put packet buffers
+    on the specified memory controllers. */
+#define NETIO_STRICT_HOMING   0x00000002
+
+/** This application expects no tags on its L2 headers. */
+#define NETIO_TAG_NONE        0x00000004
+
+/** This application expects Marvell extended tags on its L2 headers. */
+#define NETIO_TAG_MRVL        0x00000008
+
+/** This application expects Broadcom tags on its L2 headers. */
+#define NETIO_TAG_BRCM        0x00000010
+
+/** This registration may call routines which receive packets. */
+#define NETIO_RECV            0x00000020
+
+/** This registration may not call routines which receive packets. */
+#define NETIO_NO_RECV         0x00000040
+
+/** This registration may call routines which transmit packets. */
+#define NETIO_XMIT            0x00000080
+
+/** This registration may call routines which transmit packets with
+    checksum acceleration. */
+#define NETIO_XMIT_CSUM       0x00000100
+
+/** This registration may not call routines which transmit packets. */
+#define NETIO_NO_XMIT         0x00000200
+
+/** This registration wants NetIO buffers mapped at an application-specified
+    virtual address.
+
+    NetIO buffers are by default created by the TMC common memory facility,
+    which must be configured by a common ancestor of all processes sharing
+    a network interface.  When this flag is specified, NetIO buffers are
+    instead mapped at an address chosen by the application (and specified
+    in @ref netio_input_config_t::fixed_buffer_va).  This allows multiple
+    unrelated but cooperating processes to share a NetIO interface.
+    All processes sharing the same interface must specify this flag,
+    and all must specify the same fixed virtual address.
+
+    @ref netio_input_config_t::fixed_buffer_va must be a
+    multiple of 16 MB, and the packet buffers will occupy @ref
+    netio_input_config_t::total_buffer_size bytes of virtual address
+    space, beginning at that address.  If any of those virtual addresses
+    are currently occupied by other memory objects, like application or
+    shared library code or data, @ref netio_input_register() will return
+    ::NETIO_FAULT.  While it is impossible to provide a fixed_buffer_va
+    which will work for all applications, a good first guess might be to
+    use 0xb0000000 minus @ref netio_input_config_t::total_buffer_size.
+    If that fails, it might be helpful to consult the running application's
+    virtual address description file (/proc/<em>pid</em>/maps) to see
+    which regions of virtual address space are available.
+ */
+#define NETIO_FIXED_BUFFER_VA 0x00000400
+
+/** This registration call will not complete unless the network link
+    is up.  The process will wait several seconds for this to happen (the
+    precise interval is link-dependent), but if the link does not come up,
+    ::NETIO_LINK_DOWN will be returned.  This flag is the default if
+    ::NETIO_NOREQUIRE_LINK_UP is not specified.  Note that this flag by
+    itself does not request that the link be brought up; that can be done
+    with the ::NETIO_AUTO_LINK_UPDN or ::NETIO_AUTO_LINK_UP flags (the
+    latter is the default if no NETIO_AUTO_LINK_xxx flags are specified),
+    or by explicitly setting the link's desired state via netio_set().
+    If the link is not brought up by one of those methods, and this flag
+    is specified, the registration operation will return ::NETIO_LINK_DOWN.
+    This flag is ignored if it is specified along with ::NETIO_NO_XMIT and
+    ::NETIO_NO_RECV.  See @ref link for more information on link
+    management.
+ */
+#define NETIO_REQUIRE_LINK_UP    0x00000800
+
+/** This registration call will complete even if the network link is not up.
+    Whenever the link is not up, packets will not be sent or received:
+    netio_get_packet() will return ::NETIO_NOPKT once all queued packets
+    have been drained, and netio_send_packet() and similar routines will
+    return NETIO_QUEUE_FULL once the outgoing packet queue in the EPP
+    or the I/O shim is full.  See @ref link for more information on link
+    management.
+ */
+#define NETIO_NOREQUIRE_LINK_UP  0x00001000
+
+#ifndef __DOXYGEN__
+/*
+ * These are part of the implementation of the NETIO_AUTO_LINK_xxx flags,
+ * but should not be used directly by applications, and are thus not
+ * documented.
+ */
+#define _NETIO_AUTO_UP        0x00002000
+#define _NETIO_AUTO_DN        0x00004000
+#define _NETIO_AUTO_PRESENT   0x00008000
+#endif
+
+/** Set the desired state of the link to up, allowing any speeds which are
+    supported by the link hardware, as part of this registration operation.
+    Do not take down the link automatically.  This is the default if
+    no other NETIO_AUTO_LINK_xxx flags are specified.  This flag is ignored
+    if it is specified along with ::NETIO_NO_XMIT and ::NETIO_NO_RECV.
+    See @ref link for more information on link management.
+ */
+#define NETIO_AUTO_LINK_UP     (_NETIO_AUTO_PRESENT | _NETIO_AUTO_UP)
+
+/** Set the desired state of the link to up, allowing any speeds which are
+    supported by the link hardware, as part of this registration operation.
+    Set the desired state of the link to down the next time no tiles are
+    registered for packet reception or transmission.  This flag is ignored
+    if it is specified along with ::NETIO_NO_XMIT and ::NETIO_NO_RECV.
+    See @ref link for more information on link management.
+ */
+#define NETIO_AUTO_LINK_UPDN   (_NETIO_AUTO_PRESENT | _NETIO_AUTO_UP | \
+                                _NETIO_AUTO_DN)
+
+/** Set the desired state of the link to down the next time no tiles are
+    registered for packet reception or transmission.  This flag is ignored
+    if it is specified along with ::NETIO_NO_XMIT and ::NETIO_NO_RECV.
+    See @ref link for more information on link management.
+ */
+#define NETIO_AUTO_LINK_DN     (_NETIO_AUTO_PRESENT | _NETIO_AUTO_DN)
+
+/** Do not bring up the link automatically as part of this registration
+    operation.  Do not take down the link automatically.  This flag
+    is ignored if it is specified along with ::NETIO_NO_XMIT and
+    ::NETIO_NO_RECV.  See @ref link for more information on link management.
+  */
+#define NETIO_AUTO_LINK_NONE   _NETIO_AUTO_PRESENT
+
+
+/** Minimum number of receive packets. */
+#define NETIO_MIN_RECEIVE_PKTS            16
+
+/** Lower bound on the maximum number of receive packets; may be higher
+    than this on some interfaces. */
+#define NETIO_MAX_RECEIVE_PKTS           128
+
+/** Maximum number of send buffers, per packet size. */
+#define NETIO_MAX_SEND_BUFFERS            16
+
+/** Number of EPP queue slots, and thus outstanding sends, per EPP. */
+#define NETIO_TOTAL_SENDS_OUTSTANDING   2015
+
+/** Minimum number of EPP queue slots, and thus outstanding sends, per
+ *  transmitting tile. */
+#define NETIO_MIN_SENDS_OUTSTANDING       16
+
+
+/**@}*/
+
+#ifndef __DOXYGEN__
+
+/**
+ * An object for providing Ethernet packets to a process.
+ */
+struct __netio_queue_impl_t;
+
+/**
+ * An object for managing the user end of a NetIO queue.
+ */
+struct __netio_queue_user_impl_t;
+
+#endif /* !__DOXYGEN__ */
+
+
+/** A netio_queue_t describes a NetIO communications endpoint.
+ * @ingroup setup
+ */
+typedef struct
+{
+#ifdef __DOXYGEN__
+  uint8_t opaque[8];                 /**< This is an opaque structure. */
+#else
+  struct __netio_queue_impl_t* __system_part;    /**< The system part. */
+  struct __netio_queue_user_impl_t* __user_part; /**< The user part. */
+#ifdef _NETIO_PTHREAD
+  _netio_percpu_mutex_t lock;                    /**< Queue lock. */
+#endif
+#endif
+}
+netio_queue_t;
+
+
+/**
+ * @brief Packet send context.
+ *
+ * @ingroup egress
+ *
+ * Packet send context for use with netio_send_packet_prepare and _commit.
+ */
+typedef struct
+{
+#ifdef __DOXYGEN__
+  uint8_t opaque[44];   /**< This is an opaque structure. */
+#else
+  uint8_t flags;        /**< Defined below */
+  uint8_t datalen;      /**< Number of valid words pointed to by data. */
+  uint32_t request[9];  /**< Request to be sent to the EPP or shim.  Note
+                             that this is smaller than the 11-word maximum
+                             request size, since some constant values are
+                             not saved in the context. */
+  uint32_t *data;       /**< Data to be sent to the EPP or shim via IDN. */
+#endif
+}
+netio_send_pkt_context_t;
+
+
+#ifndef __DOXYGEN__
+#define SEND_PKT_CTX_USE_EPP   1  /**< We're sending to an EPP. */
+#define SEND_PKT_CTX_SEND_CSUM 2  /**< Request includes a checksum. */
+#endif
+
+/**
+ * @brief Packet vector entry.
+ *
+ * @ingroup egress
+ *
+ * This data structure is used with netio_send_packet_vector() to send multiple
+ * packets with one NetIO call.  The structure should be initialized by
+ * calling netio_pkt_vector_set(), rather than by setting the fields
+ * directly.
+ *
+ * This structure is guaranteed to be a power of two in size, no
+ * bigger than one L2 cache line, and to be aligned modulo its size.
+ */
+typedef struct
+#ifndef __DOXYGEN__
+__attribute__((aligned(8)))
+#endif
+{
+  /** Reserved for use by the user application.  When initialized with
+   *  the netio_set_pkt_vector_entry() function, this field is guaranteed
+   *  to be visible to readers only after all other fields are already
+   *  visible.  This way it can be used as a valid flag or generation
+   *  counter. */
+  uint8_t user_data;
+
+  /* Structure members below this point should not be accessed directly by
+   * applications, as they may change in the future. */
+
+  /** Low 8 bits of the packet address to send.  The high bits are
+   *  acquired from the 'handle' field. */
+  uint8_t buffer_address_low;
+
+  /** Number of bytes to transmit. */
+  uint16_t size;
+
+  /** The raw handle from a netio_pkt_t.  If this is NETIO_PKT_HANDLE_NONE,
+   *  this vector entry will be skipped and no packet will be transmitted. */
+  netio_pkt_handle_t handle;
+}
+netio_pkt_vector_entry_t;
+
+
+/**
+ * @brief Initialize fields in a packet vector entry.
+ *
+ * @ingroup egress
+ *
+ * @param[out] v Pointer to the vector entry to be initialized.
+ * @param[in] pkt Packet to be transmitted when the vector entry is passed to
+ *        netio_send_packet_vector().  Note that the packet's attributes
+ *        (e.g., its L2 offset and length) are captured at the time this
+ *        routine is called; subsequent changes in those attributes will not
+ *        be reflected in the packet which is actually transmitted.
+ *        Changes in the packet's contents, however, will be so reflected.
+ *        If this is NULL, no packet will be transmitted.
+ * @param[in] user_data User data to be set in the vector entry.
+ *        This function guarantees that the "user_data" field will become
+ *        visible to a reader only after all other fields have become visible.
+ *        This allows a structure in a ring buffer to be written and read
+ *        by a polling reader without any locks or other synchronization.
+ */
+static __inline void
+netio_pkt_vector_set(volatile netio_pkt_vector_entry_t* v, netio_pkt_t* pkt,
+                     uint8_t user_data)
+{
+  if (pkt)
+  {
+    if (NETIO_PKT_IS_MINIMAL(pkt))
+    {
+      netio_pkt_minimal_metadata_t* mmd =
+        (netio_pkt_minimal_metadata_t*) &pkt->__metadata;
+      v->buffer_address_low = (uintptr_t) NETIO_PKT_L2_DATA_MM(mmd, pkt) & 0xFF;
+      v->size = NETIO_PKT_L2_LENGTH_MM(mmd, pkt);
+    }
+    else
+    {
+      netio_pkt_metadata_t* mda = &pkt->__metadata;
+      v->buffer_address_low = (uintptr_t) NETIO_PKT_L2_DATA_M(mda, pkt) & 0xFF;
+      v->size = NETIO_PKT_L2_LENGTH_M(mda, pkt);
+    }
+    v->handle.word = pkt->__packet.word;
+  }
+  else
+  {
+    v->handle.word = 0;   /* Set handle to NETIO_PKT_HANDLE_NONE. */
+  }
+
+  __asm__("" : : : "memory");
+
+  v->user_data = user_data;
+}
+
+
+/**
+ * Flags and structures for @ref netio_get() and @ref netio_set().
+ * @ingroup config
+ */
+
+/** @{ */
+/** Parameter class; addr is a NETIO_PARAM_xxx value. */
+#define NETIO_PARAM       0
+/** Interface MAC address. This address is only valid with @ref netio_get().
+ *  The value is a 6-byte MAC address.  Depending upon the overall system
+ *  design, a MAC address may or may not be available for each interface. */
+#define NETIO_PARAM_MAC        0
+
+/** Determine whether to suspend output on the receipt of pause frames.
+ *  If the value is nonzero, the I/O shim will suspend output when a pause
+ *  frame is received.  If the value is zero, pause frames will be ignored. */
+#define NETIO_PARAM_PAUSE_IN   1
+
+/** Determine whether to send pause frames if the I/O shim packet FIFOs are
+ *  nearly full.  If the value is zero, pause frames are not sent.  If
+ *  the value is nonzero, it is the delay value which will be sent in any
+ *  pause frames which are output, in units of 512 bit times. */
+#define NETIO_PARAM_PAUSE_OUT  2
+
+/** Jumbo frame support.  The value is a 4-byte integer.  If the value is
+ *  nonzero, the MAC will accept frames of up to 10240 bytes.  If the value
+ *  is zero, the MAC will only accept frames of up to 1544 bytes. */
+#define NETIO_PARAM_JUMBO      3
+
+/** I/O shim's overflow statistics register.  The value is two 16-bit integers.
+ *  The first 16-bit value (or the low 16 bits, if the value is treated as a
+ *  32-bit number) is the count of packets which were completely dropped and
+ *  not delivered by the shim.  The second 16-bit value (or the high 16 bits,
+ *  if the value is treated as a 32-bit number) is the count of packets
+ *  which were truncated and thus only partially delivered by the shim.  This
+ *  register is automatically reset to zero after it has been read.
+ */
+#define NETIO_PARAM_OVERFLOW   4
+
+/** IPP statistics.  This address is only valid with @ref netio_get().  The
+ *  value is a netio_stat_t structure.  Unlike the I/O shim statistics, the
+ *  IPP statistics are not all reset to zero on read; see the description
+ *  of the netio_stat_t for details. */
+#define NETIO_PARAM_STAT 5
+
+/** Possible link state.  The value is a combination of "NETIO_LINK_xxx"
+ *  flags.  With @ref netio_get(), this will indicate which flags are
+ *  actually supported by the hardware.
+ *
+ *  For historical reasons, specifying this value to netio_set() will have
+ *  the same behavior as using ::NETIO_PARAM_LINK_CONFIG, but this usage is
+ *  discouraged.
+ */
+#define NETIO_PARAM_LINK_POSSIBLE_STATE 6
+
+/** Link configuration. The value is a combination of "NETIO_LINK_xxx" flags.
+ *  With @ref netio_set(), this will attempt to immediately bring up the
+ *  link using whichever of the requested flags are supported by the
+ *  hardware, or take down the link if the flags are zero; if this is
+ *  not possible, an error will be returned.  Many programs will want
+ *  to use ::NETIO_PARAM_LINK_DESIRED_STATE instead.
+ *
+ *  For historical reasons, specifying this value to netio_get() will
+ *  have the same behavior as using ::NETIO_PARAM_LINK_POSSIBLE_STATE,
+ *  but this usage is discouraged.
+ */
+#define NETIO_PARAM_LINK_CONFIG NETIO_PARAM_LINK_POSSIBLE_STATE
+
+/** Current link state. This address is only valid with @ref netio_get().
+ *  The value is zero or more of the "NETIO_LINK_xxx" flags, ORed together.
+ *  If the link is down, the value ANDed with NETIO_LINK_SPEED will be
+ *  zero; if the link is up, the value ANDed with NETIO_LINK_SPEED will
+ *  result in exactly one of the NETIO_LINK_xxx values, indicating the
+ *  current speed. */
+#define NETIO_PARAM_LINK_CURRENT_STATE 7
+
+/** Variant symbol for current state, retained for compatibility with
+ *  pre-MDE-2.1 programs. */
+#define NETIO_PARAM_LINK_STATUS NETIO_PARAM_LINK_CURRENT_STATE
+
+/** Packet Coherence protocol. This address is only valid with @ref netio_get().
+ *  The value is nonzero if the interface is configured for cache-coherent DMA.
+ */
+#define NETIO_PARAM_COHERENT 8
+
+/** Desired link state. The value is a conbination of "NETIO_LINK_xxx"
+ *  flags, which specify the desired state for the link.  With @ref
+ *  netio_set(), this will, in the background, attempt to bring up the link
+ *  using whichever of the requested flags are reasonable, or take down the
+ *  link if the flags are zero.  The actual link up or down operation may
+ *  happen after this call completes.  If the link state changes in the
+ *  future, the system will continue to try to get back to the desired link
+ *  state; for instance, if the link is brought up successfully, and then
+ *  the network cable is disconnected, the link will go down.  However, the
+ *  desired state of the link is still up, so if the cable is reconnected,
+ *  the link will be brought up again.
+ *
+ *  With @ref netio_get(), this will indicate the desired state for the
+ *  link, as set with a previous netio_set() call, or implicitly by a
+ *  netio_input_register() or netio_input_unregister() operation.  This may
+ *  not reflect the current state of the link; to get that, use
+ *  ::NETIO_PARAM_LINK_CURRENT_STATE. */
+#define NETIO_PARAM_LINK_DESIRED_STATE 9
+
+/** NetIO statistics structure.  Retrieved using the ::NETIO_PARAM_STAT
+ *  address passed to @ref netio_get(). */
+typedef struct
+{
+  /** Number of packets which have been received by the IPP and forwarded
+   *  to a tile's receive queue for processing.  This value wraps at its
+   *  maximum, and is not cleared upon read. */
+  uint32_t packets_received;
+
+  /** Number of packets which have been dropped by the IPP, because they could
+   *  not be received, or could not be forwarded to a tile.  The former happens
+   *  when the IPP does not have a free packet buffer of suitable size for an
+   *  incoming frame.  The latter happens when all potential destination tiles
+   *  for a packet, as defined by the group, bucket, and queue configuration,
+   *  have full receive queues.   This value wraps at its maximum, and is not
+   *  cleared upon read. */
+  uint32_t packets_dropped;
+
+  /*
+   * Note: the #defines after each of the following four one-byte values
+   * denote their location within the third word of the netio_stat_t.  They
+   * are intended for use only by the IPP implementation and are thus omitted
+   * from the Doxygen output.
+   */
+
+  /** Number of packets dropped because no worker was able to accept a new
+   *  packet.  This value saturates at its maximum, and is cleared upon
+   *  read. */
+  uint8_t drops_no_worker;
+#ifndef __DOXYGEN__
+#define NETIO_STAT_DROPS_NO_WORKER   0
+#endif
+
+  /** Number of packets dropped because no small buffers were available.
+   *  This value saturates at its maximum, and is cleared upon read. */
+  uint8_t drops_no_smallbuf;
+#ifndef __DOXYGEN__
+#define NETIO_STAT_DROPS_NO_SMALLBUF 1
+#endif
+
+  /** Number of packets dropped because no large buffers were available.
+   *  This value saturates at its maximum, and is cleared upon read. */
+  uint8_t drops_no_largebuf;
+#ifndef __DOXYGEN__
+#define NETIO_STAT_DROPS_NO_LARGEBUF 2
+#endif
+
+  /** Number of packets dropped because no jumbo buffers were available.
+   *  This value saturates at its maximum, and is cleared upon read. */
+  uint8_t drops_no_jumbobuf;
+#ifndef __DOXYGEN__
+#define NETIO_STAT_DROPS_NO_JUMBOBUF 3
+#endif
+}
+netio_stat_t;
+
+
+/** Link can run, should run, or is running at 10 Mbps. */
+#define NETIO_LINK_10M         0x01
+
+/** Link can run, should run, or is running at 100 Mbps. */
+#define NETIO_LINK_100M        0x02
+
+/** Link can run, should run, or is running at 1 Gbps. */
+#define NETIO_LINK_1G          0x04
+
+/** Link can run, should run, or is running at 10 Gbps. */
+#define NETIO_LINK_10G         0x08
+
+/** Link should run at the highest speed supported by the link and by
+ *  the device connected to the link.  Only usable as a value for
+ *  the link's desired state; never returned as a value for the current
+ *  or possible states. */
+#define NETIO_LINK_ANYSPEED    0x10
+
+/** All legal link speeds. */
+#define NETIO_LINK_SPEED  (NETIO_LINK_10M  | \
+                           NETIO_LINK_100M | \
+                           NETIO_LINK_1G   | \
+                           NETIO_LINK_10G  | \
+                           NETIO_LINK_ANYSPEED)
+
+
+/** MAC register class.  Addr is a register offset within the MAC.
+ *  Registers within the XGbE and GbE MACs are documented in the Tile
+ *  Processor I/O Device Guide (UG104). MAC registers start at address
+ *  0x4000, and do not include the MAC_INTERFACE registers. */
+#define NETIO_MAC             1
+
+/** MDIO register class (IEEE 802.3 clause 22 format).  Addr is the "addr"
+ *  member of a netio_mdio_addr_t structure. */
+#define NETIO_MDIO            2
+
+/** MDIO register class (IEEE 802.3 clause 45 format).  Addr is the "addr"
+ *  member of a netio_mdio_addr_t structure. */
+#define NETIO_MDIO_CLAUSE45   3
+
+/** NetIO MDIO address type.  Retrieved or provided using the ::NETIO_MDIO
+ *  address passed to @ref netio_get() or @ref netio_set(). */
+typedef union
+{
+  struct
+  {
+    unsigned int reg:16;  /**< MDIO register offset.  For clause 22 access,
+                               must be less than 32. */
+    unsigned int phy:5;   /**< Which MDIO PHY to access. */
+    unsigned int dev:5;   /**< Which MDIO device to access within that PHY.
+                               Applicable for clause 45 access only; ignored
+                               for clause 22 access. */
+  }
+  bits;                   /**< Container for bitfields. */
+  uint64_t addr;          /**< Value to pass to @ref netio_get() or
+                           *   @ref netio_set(). */
+}
+netio_mdio_addr_t;
+
+/** @} */
+
+#endif /* __NETIO_INTF_H__ */
diff --git a/kernel/arch/tile/include/hv/syscall_public.h b/kernel/arch/tile/include/hv/syscall_public.h
new file mode 100644
index 000000000..9cc0837e6
--- /dev/null
+++ b/kernel/arch/tile/include/hv/syscall_public.h
@@ -0,0 +1,42 @@
+/*
+ * Copyright 2010 Tilera Corporation. All Rights Reserved.
+ *
+ *   This program is free software; you can redistribute it and/or
+ *   modify it under the terms of the GNU General Public License
+ *   as published by the Free Software Foundation, version 2.
+ *
+ *   This program is distributed in the hope that it will be useful, but
+ *   WITHOUT ANY WARRANTY; without even the implied warranty of
+ *   MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
+ *   NON INFRINGEMENT.  See the GNU General Public License for
+ *   more details.
+ */
+
+/**
+ * @file syscall.h
+ * Indices for the hypervisor system calls that are intended to be called
+ * directly, rather than only through hypervisor-generated "glue" code.
+ */
+
+#ifndef _SYS_HV_INCLUDE_SYSCALL_PUBLIC_H
+#define _SYS_HV_INCLUDE_SYSCALL_PUBLIC_H
+
+/** Fast syscall flag bit location.  When this bit is set, the hypervisor
+ *  handles the syscall specially.
+ */
+#define HV_SYS_FAST_SHIFT                 14
+
+/** Fast syscall flag bit mask. */
+#define HV_SYS_FAST_MASK                  (1 << HV_SYS_FAST_SHIFT)
+
+/** Bit location for flagging fast syscalls that can be called from PL0. */
+#define HV_SYS_FAST_PLO_SHIFT             13
+
+/** Fast syscall allowing PL0 bit mask. */
+#define HV_SYS_FAST_PL0_MASK              (1 << HV_SYS_FAST_PLO_SHIFT)
+
+/** Perform an MF that waits for all victims to reach DRAM. */
+#define HV_SYS_fence_incoherent         (51 | HV_SYS_FAST_MASK \
+                                       | HV_SYS_FAST_PL0_MASK)
+
+#endif /* !_SYS_HV_INCLUDE_SYSCALL_PUBLIC_H */
diff --git a/kernel/arch/tile/include/uapi/arch/Kbuild b/kernel/arch/tile/include/uapi/arch/Kbuild
new file mode 100644
index 000000000..97dfbecec
--- /dev/null
+++ b/kernel/arch/tile/include/uapi/arch/Kbuild
@@ -0,0 +1,17 @@
+# UAPI Header export list
+header-y += abi.h
+header-y += chip.h
+header-y += chip_tilegx.h
+header-y += chip_tilepro.h
+header-y += icache.h
+header-y += interrupts.h
+header-y += interrupts_32.h
+header-y += interrupts_64.h
+header-y += opcode.h
+header-y += opcode_tilegx.h
+header-y += opcode_tilepro.h
+header-y += sim.h
+header-y += sim_def.h
+header-y += spr_def.h
+header-y += spr_def_32.h
+header-y += spr_def_64.h
diff --git a/kernel/arch/tile/include/uapi/arch/abi.h b/kernel/arch/tile/include/uapi/arch/abi.h
new file mode 100644
index 000000000..c55a3d432
--- /dev/null
+++ b/kernel/arch/tile/include/uapi/arch/abi.h
@@ -0,0 +1,141 @@
+/*
+ * Copyright 2010 Tilera Corporation. All Rights Reserved.
+ *
+ *   This program is free software; you can redistribute it and/or
+ *   modify it under the terms of the GNU General Public License
+ *   as published by the Free Software Foundation, version 2.
+ *
+ *   This program is distributed in the hope that it will be useful, but
+ *   WITHOUT ANY WARRANTY; without even the implied warranty of
+ *   MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
+ *   NON INFRINGEMENT.  See the GNU General Public License for
+ *   more details.
+ */
+
+/**
+ * @file
+ *
+ * ABI-related register definitions.
+ */
+
+#ifndef __ARCH_ABI_H__
+
+#if !defined __need_int_reg_t && !defined __DOXYGEN__
+# define __ARCH_ABI_H__
+# include <arch/chip.h>
+#endif
+
+/* Provide the basic machine types. */
+#ifndef __INT_REG_BITS
+
+/** Number of bits in a register. */
+#if defined __tilegx__
+# define __INT_REG_BITS 64
+#elif defined __tilepro__
+# define __INT_REG_BITS 32
+#elif !defined __need_int_reg_t
+# include <arch/chip.h>
+# define __INT_REG_BITS CHIP_WORD_SIZE()
+#else
+# error Unrecognized architecture with __need_int_reg_t
+#endif
+
+#if __INT_REG_BITS == 64
+
+#ifndef __ASSEMBLER__
+/** Unsigned type that can hold a register. */
+typedef unsigned long long __uint_reg_t;
+
+/** Signed type that can hold a register. */
+typedef long long __int_reg_t;
+#endif
+
+/** String prefix to use for printf(). */
+#define __INT_REG_FMT "ll"
+
+#else
+
+#ifndef __ASSEMBLER__
+/** Unsigned type that can hold a register. */
+typedef unsigned long __uint_reg_t;
+
+/** Signed type that can hold a register. */
+typedef long __int_reg_t;
+#endif
+
+/** String prefix to use for printf(). */
+#define __INT_REG_FMT "l"
+
+#endif
+#endif /* __INT_REG_BITS */
+
+
+#ifndef __need_int_reg_t
+
+
+#ifndef __ASSEMBLER__
+/** Unsigned type that can hold a register. */
+typedef __uint_reg_t uint_reg_t;
+
+/** Signed type that can hold a register. */
+typedef __int_reg_t int_reg_t;
+#endif
+
+/** String prefix to use for printf(). */
+#define INT_REG_FMT __INT_REG_FMT
+
+/** Number of bits in a register. */
+#define INT_REG_BITS __INT_REG_BITS
+
+
+/* Registers 0 - 55 are "normal", but some perform special roles. */
+
+#define TREG_FP       52   /**< Frame pointer. */
+#define TREG_TP       53   /**< Thread pointer. */
+#define TREG_SP       54   /**< Stack pointer. */
+#define TREG_LR       55   /**< Link to calling function PC. */
+
+/** Index of last normal general-purpose register. */
+#define TREG_LAST_GPR 55
+
+/* Registers 56 - 62 are "special" network registers. */
+
+#define TREG_SN       56   /**< Static network access. */
+#define TREG_IDN0     57   /**< IDN demux 0 access. */
+#define TREG_IDN1     58   /**< IDN demux 1 access. */
+#define TREG_UDN0     59   /**< UDN demux 0 access. */
+#define TREG_UDN1     60   /**< UDN demux 1 access. */
+#define TREG_UDN2     61   /**< UDN demux 2 access. */
+#define TREG_UDN3     62   /**< UDN demux 3 access. */
+
+/* Register 63 is the "special" zero register. */
+
+#define TREG_ZERO     63   /**< "Zero" register; always reads as "0". */
+
+
+/** By convention, this register is used to hold the syscall number. */
+#define TREG_SYSCALL_NR      10
+
+/** Name of register that holds the syscall number, for use in assembly. */
+#define TREG_SYSCALL_NR_NAME r10
+
+
+/**
+ * The ABI requires callers to allocate a caller state save area of
+ * this many bytes at the bottom of each stack frame.
+ */
+#define C_ABI_SAVE_AREA_SIZE (2 * (INT_REG_BITS / 8))
+
+/**
+ * The operand to an 'info' opcode directing the backtracer to not
+ * try to find the calling frame.
+ */
+#define INFO_OP_CANNOT_BACKTRACE 2
+
+
+#endif /* !__need_int_reg_t */
+
+/* Make sure we later can get all the definitions and declarations.  */
+#undef __need_int_reg_t
+
+#endif /* !__ARCH_ABI_H__ */
diff --git a/kernel/arch/tile/include/uapi/arch/chip.h b/kernel/arch/tile/include/uapi/arch/chip.h
new file mode 100644
index 000000000..4c91f90b9
--- /dev/null
+++ b/kernel/arch/tile/include/uapi/arch/chip.h
@@ -0,0 +1,21 @@
+/*
+ * Copyright 2010 Tilera Corporation. All Rights Reserved.
+ *
+ *   This program is free software; you can redistribute it and/or
+ *   modify it under the terms of the GNU General Public License
+ *   as published by the Free Software Foundation, version 2.
+ *
+ *   This program is distributed in the hope that it will be useful, but
+ *   WITHOUT ANY WARRANTY; without even the implied warranty of
+ *   MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
+ *   NON INFRINGEMENT.  See the GNU General Public License for
+ *   more details.
+ */
+
+#if __tile_chip__ == 1
+#include <arch/chip_tilepro.h>
+#elif defined(__tilegx__)
+#include <arch/chip_tilegx.h>
+#else
+#error Unexpected Tilera chip type
+#endif
diff --git a/kernel/arch/tile/include/uapi/arch/chip_tilegx.h b/kernel/arch/tile/include/uapi/arch/chip_tilegx.h
new file mode 100644
index 000000000..ea8e4f2c9
--- /dev/null
+++ b/kernel/arch/tile/include/uapi/arch/chip_tilegx.h
@@ -0,0 +1,258 @@
+/*
+ * Copyright 2011 Tilera Corporation. All Rights Reserved.
+ *
+ *   This program is free software; you can redistribute it and/or
+ *   modify it under the terms of the GNU General Public License
+ *   as published by the Free Software Foundation, version 2.
+ *
+ *   This program is distributed in the hope that it will be useful, but
+ *   WITHOUT ANY WARRANTY; without even the implied warranty of
+ *   MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
+ *   NON INFRINGEMENT.  See the GNU General Public License for
+ *   more details.
+ */
+
+/*
+ * @file
+ * Global header file.
+ * This header file specifies defines for TILE-Gx.
+ */
+
+#ifndef __ARCH_CHIP_H__
+#define __ARCH_CHIP_H__
+
+/** Specify chip version.
+ * When possible, prefer the CHIP_xxx symbols below for future-proofing.
+ * This is intended for cross-compiling; native compilation should
+ * use the predefined __tile_chip__ symbol.
+ */
+#define TILE_CHIP 10
+
+/** Specify chip revision.
+ * This provides for the case of a respin of a particular chip type;
+ * the normal value for this symbol is "0".
+ * This is intended for cross-compiling; native compilation should
+ * use the predefined __tile_chip_rev__ symbol.
+ */
+#define TILE_CHIP_REV 0
+
+/** The name of this architecture. */
+#define CHIP_ARCH_NAME "tilegx"
+
+/** The ELF e_machine type for binaries for this chip. */
+#define CHIP_ELF_TYPE() EM_TILEGX
+
+/** The alternate ELF e_machine type for binaries for this chip. */
+#define CHIP_COMPAT_ELF_TYPE() 0x2597
+
+/** What is the native word size of the machine? */
+#define CHIP_WORD_SIZE() 64
+
+/** How many bits of a virtual address are used. Extra bits must be
+ * the sign extension of the low bits.
+ */
+#define CHIP_VA_WIDTH() 42
+
+/** How many bits are in a physical address? */
+#define CHIP_PA_WIDTH() 40
+
+/** Size of the L2 cache, in bytes. */
+#define CHIP_L2_CACHE_SIZE() 262144
+
+/** Log size of an L2 cache line in bytes. */
+#define CHIP_L2_LOG_LINE_SIZE() 6
+
+/** Size of an L2 cache line, in bytes. */
+#define CHIP_L2_LINE_SIZE() (1 << CHIP_L2_LOG_LINE_SIZE())
+
+/** Associativity of the L2 cache. */
+#define CHIP_L2_ASSOC() 8
+
+/** Size of the L1 data cache, in bytes. */
+#define CHIP_L1D_CACHE_SIZE() 32768
+
+/** Log size of an L1 data cache line in bytes. */
+#define CHIP_L1D_LOG_LINE_SIZE() 6
+
+/** Size of an L1 data cache line, in bytes. */
+#define CHIP_L1D_LINE_SIZE() (1 << CHIP_L1D_LOG_LINE_SIZE())
+
+/** Associativity of the L1 data cache. */
+#define CHIP_L1D_ASSOC() 2
+
+/** Size of the L1 instruction cache, in bytes. */
+#define CHIP_L1I_CACHE_SIZE() 32768
+
+/** Log size of an L1 instruction cache line in bytes. */
+#define CHIP_L1I_LOG_LINE_SIZE() 6
+
+/** Size of an L1 instruction cache line, in bytes. */
+#define CHIP_L1I_LINE_SIZE() (1 << CHIP_L1I_LOG_LINE_SIZE())
+
+/** Associativity of the L1 instruction cache. */
+#define CHIP_L1I_ASSOC() 2
+
+/** Stride with which flush instructions must be issued. */
+#define CHIP_FLUSH_STRIDE() CHIP_L2_LINE_SIZE()
+
+/** Stride with which inv instructions must be issued. */
+#define CHIP_INV_STRIDE() CHIP_L2_LINE_SIZE()
+
+/** Stride with which finv instructions must be issued. */
+#define CHIP_FINV_STRIDE() CHIP_L2_LINE_SIZE()
+
+/** Can the local cache coherently cache data that is homed elsewhere? */
+#define CHIP_HAS_COHERENT_LOCAL_CACHE() 1
+
+/** How many simultaneous outstanding victims can the L2 cache have? */
+#define CHIP_MAX_OUTSTANDING_VICTIMS() 128
+
+/** Does the TLB support the NC and NOALLOC bits? */
+#define CHIP_HAS_NC_AND_NOALLOC_BITS() 1
+
+/** Does the chip support hash-for-home caching? */
+#define CHIP_HAS_CBOX_HOME_MAP() 1
+
+/** Number of entries in the chip's home map tables. */
+#define CHIP_CBOX_HOME_MAP_SIZE() 128
+
+/** Do uncacheable requests miss in the cache regardless of whether
+ * there is matching data? */
+#define CHIP_HAS_ENFORCED_UNCACHEABLE_REQUESTS() 1
+
+/** Does the mf instruction wait for victims? */
+#define CHIP_HAS_MF_WAITS_FOR_VICTIMS() 0
+
+/** Does the chip have an "inv" instruction that doesn't also flush? */
+#define CHIP_HAS_INV() 1
+
+/** Does the chip have a "wh64" instruction? */
+#define CHIP_HAS_WH64() 1
+
+/** Does this chip have a 'dword_align' instruction? */
+#define CHIP_HAS_DWORD_ALIGN() 0
+
+/** Number of performance counters. */
+#define CHIP_PERFORMANCE_COUNTERS() 4
+
+/** Does this chip have auxiliary performance counters? */
+#define CHIP_HAS_AUX_PERF_COUNTERS() 1
+
+/** Is the CBOX_MSR1 SPR supported? */
+#define CHIP_HAS_CBOX_MSR1() 0
+
+/** Is the TILE_RTF_HWM SPR supported? */
+#define CHIP_HAS_TILE_RTF_HWM() 1
+
+/** Is the TILE_WRITE_PENDING SPR supported? */
+#define CHIP_HAS_TILE_WRITE_PENDING() 0
+
+/** Is the PROC_STATUS SPR supported? */
+#define CHIP_HAS_PROC_STATUS_SPR() 1
+
+/** Is the DSTREAM_PF SPR supported? */
+#define CHIP_HAS_DSTREAM_PF() 1
+
+/** Log of the number of mshims we have. */
+#define CHIP_LOG_NUM_MSHIMS() 2
+
+/** Are the bases of the interrupt vector areas fixed? */
+#define CHIP_HAS_FIXED_INTVEC_BASE() 0
+
+/** Are the interrupt masks split up into 2 SPRs? */
+#define CHIP_HAS_SPLIT_INTR_MASK() 0
+
+/** Is the cycle count split up into 2 SPRs? */
+#define CHIP_HAS_SPLIT_CYCLE() 0
+
+/** Does the chip have a static network? */
+#define CHIP_HAS_SN() 0
+
+/** Does the chip have a static network processor? */
+#define CHIP_HAS_SN_PROC() 0
+
+/** Size of the L1 static network processor instruction cache, in bytes. */
+/* #define CHIP_L1SNI_CACHE_SIZE() -- does not apply to chip 10 */
+
+/** Does the chip have DMA support in each tile? */
+#define CHIP_HAS_TILE_DMA() 0
+
+/** Does the chip have the second revision of the directly accessible
+ *  dynamic networks?  This encapsulates a number of characteristics,
+ *  including the absence of the catch-all, the absence of inline message
+ *  tags, the absence of support for network context-switching, and so on.
+ */
+#define CHIP_HAS_REV1_XDN() 1
+
+/** Does the chip have cmpexch and similar (fetchadd, exch, etc.)? */
+#define CHIP_HAS_CMPEXCH() 1
+
+/** Does the chip have memory-mapped I/O support? */
+#define CHIP_HAS_MMIO() 1
+
+/** Does the chip have post-completion interrupts? */
+#define CHIP_HAS_POST_COMPLETION_INTERRUPTS() 1
+
+/** Does the chip have native single step support? */
+#define CHIP_HAS_SINGLE_STEP() 1
+
+#ifndef __OPEN_SOURCE__  /* features only relevant to hypervisor-level code */
+
+/** How many entries are present in the instruction TLB? */
+#define CHIP_ITLB_ENTRIES() 16
+
+/** How many entries are present in the data TLB? */
+#define CHIP_DTLB_ENTRIES() 32
+
+/** How many MAF entries does the XAUI shim have? */
+#define CHIP_XAUI_MAF_ENTRIES() 32
+
+/** Does the memory shim have a source-id table? */
+#define CHIP_HAS_MSHIM_SRCID_TABLE() 0
+
+/** Does the L1 instruction cache clear on reset? */
+#define CHIP_HAS_L1I_CLEAR_ON_RESET() 1
+
+/** Does the chip come out of reset with valid coordinates on all tiles?
+ * Note that if defined, this also implies that the upper left is 1,1.
+ */
+#define CHIP_HAS_VALID_TILE_COORD_RESET() 1
+
+/** Does the chip have unified packet formats? */
+#define CHIP_HAS_UNIFIED_PACKET_FORMATS() 1
+
+/** Does the chip support write reordering? */
+#define CHIP_HAS_WRITE_REORDERING() 1
+
+/** Does the chip support Y-X routing as well as X-Y? */
+#define CHIP_HAS_Y_X_ROUTING() 1
+
+/** Is INTCTRL_3 managed with the correct MPL? */
+#define CHIP_HAS_INTCTRL_3_STATUS_FIX() 1
+
+/** Is it possible to configure the chip to be big-endian? */
+#define CHIP_HAS_BIG_ENDIAN_CONFIG() 1
+
+/** Is the CACHE_RED_WAY_OVERRIDDEN SPR supported? */
+#define CHIP_HAS_CACHE_RED_WAY_OVERRIDDEN() 0
+
+/** Is the DIAG_TRACE_WAY SPR supported? */
+#define CHIP_HAS_DIAG_TRACE_WAY() 0
+
+/** Is the MEM_STRIPE_CONFIG SPR supported? */
+#define CHIP_HAS_MEM_STRIPE_CONFIG() 1
+
+/** Are the TLB_PERF SPRs supported? */
+#define CHIP_HAS_TLB_PERF() 1
+
+/** Is the VDN_SNOOP_SHIM_CTL SPR supported? */
+#define CHIP_HAS_VDN_SNOOP_SHIM_CTL() 0
+
+/** Does the chip support rev1 DMA packets? */
+#define CHIP_HAS_REV1_DMA_PACKETS() 1
+
+/** Does the chip have an IPI shim? */
+#define CHIP_HAS_IPI() 1
+
+#endif /* !__OPEN_SOURCE__ */
+#endif /* __ARCH_CHIP_H__ */
diff --git a/kernel/arch/tile/include/uapi/arch/chip_tilepro.h b/kernel/arch/tile/include/uapi/arch/chip_tilepro.h
new file mode 100644
index 000000000..70017699a
--- /dev/null
+++ b/kernel/arch/tile/include/uapi/arch/chip_tilepro.h
@@ -0,0 +1,258 @@
+/*
+ * Copyright 2010 Tilera Corporation. All Rights Reserved.
+ *
+ *   This program is free software; you can redistribute it and/or
+ *   modify it under the terms of the GNU General Public License
+ *   as published by the Free Software Foundation, version 2.
+ *
+ *   This program is distributed in the hope that it will be useful, but
+ *   WITHOUT ANY WARRANTY; without even the implied warranty of
+ *   MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
+ *   NON INFRINGEMENT.  See the GNU General Public License for
+ *   more details.
+ */
+
+/*
+ * @file
+ * Global header file.
+ * This header file specifies defines for TILEPro.
+ */
+
+#ifndef __ARCH_CHIP_H__
+#define __ARCH_CHIP_H__
+
+/** Specify chip version.
+ * When possible, prefer the CHIP_xxx symbols below for future-proofing.
+ * This is intended for cross-compiling; native compilation should
+ * use the predefined __tile_chip__ symbol.
+ */
+#define TILE_CHIP 1
+
+/** Specify chip revision.
+ * This provides for the case of a respin of a particular chip type;
+ * the normal value for this symbol is "0".
+ * This is intended for cross-compiling; native compilation should
+ * use the predefined __tile_chip_rev__ symbol.
+ */
+#define TILE_CHIP_REV 0
+
+/** The name of this architecture. */
+#define CHIP_ARCH_NAME "tilepro"
+
+/** The ELF e_machine type for binaries for this chip. */
+#define CHIP_ELF_TYPE() EM_TILEPRO
+
+/** The alternate ELF e_machine type for binaries for this chip. */
+#define CHIP_COMPAT_ELF_TYPE() 0x2507
+
+/** What is the native word size of the machine? */
+#define CHIP_WORD_SIZE() 32
+
+/** How many bits of a virtual address are used. Extra bits must be
+ * the sign extension of the low bits.
+ */
+#define CHIP_VA_WIDTH() 32
+
+/** How many bits are in a physical address? */
+#define CHIP_PA_WIDTH() 36
+
+/** Size of the L2 cache, in bytes. */
+#define CHIP_L2_CACHE_SIZE() 65536
+
+/** Log size of an L2 cache line in bytes. */
+#define CHIP_L2_LOG_LINE_SIZE() 6
+
+/** Size of an L2 cache line, in bytes. */
+#define CHIP_L2_LINE_SIZE() (1 << CHIP_L2_LOG_LINE_SIZE())
+
+/** Associativity of the L2 cache. */
+#define CHIP_L2_ASSOC() 4
+
+/** Size of the L1 data cache, in bytes. */
+#define CHIP_L1D_CACHE_SIZE() 8192
+
+/** Log size of an L1 data cache line in bytes. */
+#define CHIP_L1D_LOG_LINE_SIZE() 4
+
+/** Size of an L1 data cache line, in bytes. */
+#define CHIP_L1D_LINE_SIZE() (1 << CHIP_L1D_LOG_LINE_SIZE())
+
+/** Associativity of the L1 data cache. */
+#define CHIP_L1D_ASSOC() 2
+
+/** Size of the L1 instruction cache, in bytes. */
+#define CHIP_L1I_CACHE_SIZE() 16384
+
+/** Log size of an L1 instruction cache line in bytes. */
+#define CHIP_L1I_LOG_LINE_SIZE() 6
+
+/** Size of an L1 instruction cache line, in bytes. */
+#define CHIP_L1I_LINE_SIZE() (1 << CHIP_L1I_LOG_LINE_SIZE())
+
+/** Associativity of the L1 instruction cache. */
+#define CHIP_L1I_ASSOC() 1
+
+/** Stride with which flush instructions must be issued. */
+#define CHIP_FLUSH_STRIDE() CHIP_L2_LINE_SIZE()
+
+/** Stride with which inv instructions must be issued. */
+#define CHIP_INV_STRIDE() CHIP_L2_LINE_SIZE()
+
+/** Stride with which finv instructions must be issued. */
+#define CHIP_FINV_STRIDE() CHIP_L2_LINE_SIZE()
+
+/** Can the local cache coherently cache data that is homed elsewhere? */
+#define CHIP_HAS_COHERENT_LOCAL_CACHE() 1
+
+/** How many simultaneous outstanding victims can the L2 cache have? */
+#define CHIP_MAX_OUTSTANDING_VICTIMS() 4
+
+/** Does the TLB support the NC and NOALLOC bits? */
+#define CHIP_HAS_NC_AND_NOALLOC_BITS() 1
+
+/** Does the chip support hash-for-home caching? */
+#define CHIP_HAS_CBOX_HOME_MAP() 1
+
+/** Number of entries in the chip's home map tables. */
+#define CHIP_CBOX_HOME_MAP_SIZE() 64
+
+/** Do uncacheable requests miss in the cache regardless of whether
+ * there is matching data? */
+#define CHIP_HAS_ENFORCED_UNCACHEABLE_REQUESTS() 1
+
+/** Does the mf instruction wait for victims? */
+#define CHIP_HAS_MF_WAITS_FOR_VICTIMS() 0
+
+/** Does the chip have an "inv" instruction that doesn't also flush? */
+#define CHIP_HAS_INV() 1
+
+/** Does the chip have a "wh64" instruction? */
+#define CHIP_HAS_WH64() 1
+
+/** Does this chip have a 'dword_align' instruction? */
+#define CHIP_HAS_DWORD_ALIGN() 1
+
+/** Number of performance counters. */
+#define CHIP_PERFORMANCE_COUNTERS() 4
+
+/** Does this chip have auxiliary performance counters? */
+#define CHIP_HAS_AUX_PERF_COUNTERS() 1
+
+/** Is the CBOX_MSR1 SPR supported? */
+#define CHIP_HAS_CBOX_MSR1() 1
+
+/** Is the TILE_RTF_HWM SPR supported? */
+#define CHIP_HAS_TILE_RTF_HWM() 1
+
+/** Is the TILE_WRITE_PENDING SPR supported? */
+#define CHIP_HAS_TILE_WRITE_PENDING() 1
+
+/** Is the PROC_STATUS SPR supported? */
+#define CHIP_HAS_PROC_STATUS_SPR() 1
+
+/** Is the DSTREAM_PF SPR supported? */
+#define CHIP_HAS_DSTREAM_PF() 0
+
+/** Log of the number of mshims we have. */
+#define CHIP_LOG_NUM_MSHIMS() 2
+
+/** Are the bases of the interrupt vector areas fixed? */
+#define CHIP_HAS_FIXED_INTVEC_BASE() 1
+
+/** Are the interrupt masks split up into 2 SPRs? */
+#define CHIP_HAS_SPLIT_INTR_MASK() 1
+
+/** Is the cycle count split up into 2 SPRs? */
+#define CHIP_HAS_SPLIT_CYCLE() 1
+
+/** Does the chip have a static network? */
+#define CHIP_HAS_SN() 1
+
+/** Does the chip have a static network processor? */
+#define CHIP_HAS_SN_PROC() 0
+
+/** Size of the L1 static network processor instruction cache, in bytes. */
+/* #define CHIP_L1SNI_CACHE_SIZE() -- does not apply to chip 1 */
+
+/** Does the chip have DMA support in each tile? */
+#define CHIP_HAS_TILE_DMA() 1
+
+/** Does the chip have the second revision of the directly accessible
+ *  dynamic networks?  This encapsulates a number of characteristics,
+ *  including the absence of the catch-all, the absence of inline message
+ *  tags, the absence of support for network context-switching, and so on.
+ */
+#define CHIP_HAS_REV1_XDN() 0
+
+/** Does the chip have cmpexch and similar (fetchadd, exch, etc.)? */
+#define CHIP_HAS_CMPEXCH() 0
+
+/** Does the chip have memory-mapped I/O support? */
+#define CHIP_HAS_MMIO() 0
+
+/** Does the chip have post-completion interrupts? */
+#define CHIP_HAS_POST_COMPLETION_INTERRUPTS() 0
+
+/** Does the chip have native single step support? */
+#define CHIP_HAS_SINGLE_STEP() 0
+
+#ifndef __OPEN_SOURCE__  /* features only relevant to hypervisor-level code */
+
+/** How many entries are present in the instruction TLB? */
+#define CHIP_ITLB_ENTRIES() 16
+
+/** How many entries are present in the data TLB? */
+#define CHIP_DTLB_ENTRIES() 16
+
+/** How many MAF entries does the XAUI shim have? */
+#define CHIP_XAUI_MAF_ENTRIES() 32
+
+/** Does the memory shim have a source-id table? */
+#define CHIP_HAS_MSHIM_SRCID_TABLE() 0
+
+/** Does the L1 instruction cache clear on reset? */
+#define CHIP_HAS_L1I_CLEAR_ON_RESET() 1
+
+/** Does the chip come out of reset with valid coordinates on all tiles?
+ * Note that if defined, this also implies that the upper left is 1,1.
+ */
+#define CHIP_HAS_VALID_TILE_COORD_RESET() 1
+
+/** Does the chip have unified packet formats? */
+#define CHIP_HAS_UNIFIED_PACKET_FORMATS() 1
+
+/** Does the chip support write reordering? */
+#define CHIP_HAS_WRITE_REORDERING() 1
+
+/** Does the chip support Y-X routing as well as X-Y? */
+#define CHIP_HAS_Y_X_ROUTING() 1
+
+/** Is INTCTRL_3 managed with the correct MPL? */
+#define CHIP_HAS_INTCTRL_3_STATUS_FIX() 1
+
+/** Is it possible to configure the chip to be big-endian? */
+#define CHIP_HAS_BIG_ENDIAN_CONFIG() 1
+
+/** Is the CACHE_RED_WAY_OVERRIDDEN SPR supported? */
+#define CHIP_HAS_CACHE_RED_WAY_OVERRIDDEN() 1
+
+/** Is the DIAG_TRACE_WAY SPR supported? */
+#define CHIP_HAS_DIAG_TRACE_WAY() 1
+
+/** Is the MEM_STRIPE_CONFIG SPR supported? */
+#define CHIP_HAS_MEM_STRIPE_CONFIG() 1
+
+/** Are the TLB_PERF SPRs supported? */
+#define CHIP_HAS_TLB_PERF() 1
+
+/** Is the VDN_SNOOP_SHIM_CTL SPR supported? */
+#define CHIP_HAS_VDN_SNOOP_SHIM_CTL() 1
+
+/** Does the chip support rev1 DMA packets? */
+#define CHIP_HAS_REV1_DMA_PACKETS() 1
+
+/** Does the chip have an IPI shim? */
+#define CHIP_HAS_IPI() 0
+
+#endif /* !__OPEN_SOURCE__ */
+#endif /* __ARCH_CHIP_H__ */
diff --git a/kernel/arch/tile/include/uapi/arch/icache.h b/kernel/arch/tile/include/uapi/arch/icache.h
new file mode 100644
index 000000000..762eafa8a
--- /dev/null
+++ b/kernel/arch/tile/include/uapi/arch/icache.h
@@ -0,0 +1,93 @@
+/*
+ * Copyright 2010 Tilera Corporation. All Rights Reserved.
+ *
+ *   This program is free software; you can redistribute it and/or
+ *   modify it under the terms of the GNU General Public License
+ *   as published by the Free Software Foundation, version 2.
+ *
+ *   This program is distributed in the hope that it will be useful, but
+ *   WITHOUT ANY WARRANTY; without even the implied warranty of
+ *   MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
+ *   NON INFRINGEMENT.  See the GNU General Public License for
+ *   more details.
+ *
+ */
+
+/**
+ * @file
+ *
+ * Support for invalidating bytes in the instruction cache.
+ */
+
+#ifndef __ARCH_ICACHE_H__
+#define __ARCH_ICACHE_H__
+
+#include <arch/chip.h>
+
+
+/**
+ * Invalidate the instruction cache for the given range of memory.
+ *
+ * @param addr The start of memory to be invalidated.
+ * @param size The number of bytes to be invalidated.
+ * @param page_size The system's page size, e.g. getpagesize() in userspace.
+ * This value must be a power of two no larger than the page containing
+ * the code to be invalidated. If the value is smaller than the actual page
+ * size, this function will still work, but may run slower than necessary.
+ */
+static __inline void
+invalidate_icache(const void* addr, unsigned long size,
+                  unsigned long page_size)
+{
+  const unsigned long cache_way_size =
+    CHIP_L1I_CACHE_SIZE() / CHIP_L1I_ASSOC();
+  unsigned long max_useful_size;
+  const char* start, *end;
+  long num_passes;
+
+  if (__builtin_expect(size == 0, 0))
+    return;
+
+#ifdef __tilegx__
+  /* Limit the number of bytes visited to avoid redundant iterations. */
+  max_useful_size = (page_size < cache_way_size) ? page_size : cache_way_size;
+
+  /* No PA aliasing is possible, so one pass always suffices. */
+  num_passes = 1;
+#else
+  /* Limit the number of bytes visited to avoid redundant iterations. */
+  max_useful_size = cache_way_size;
+
+  /*
+   * Compute how many passes we need (we'll treat 0 as if it were 1).
+   * This works because we know the page size is a power of two.
+   */
+  num_passes = cache_way_size >> __builtin_ctzl(page_size);
+#endif
+
+  if (__builtin_expect(size > max_useful_size, 0))
+    size = max_useful_size;
+
+  /* Locate the first and last bytes to be invalidated. */
+  start = (const char *)((unsigned long)addr & -CHIP_L1I_LINE_SIZE());
+  end = (const char*)addr + size - 1;
+
+  __insn_mf();
+
+  do
+  {
+    const char* p;
+
+    for (p = start; p <= end; p += CHIP_L1I_LINE_SIZE())
+      __insn_icoh(p);
+
+    start += page_size;
+    end += page_size;
+  }
+  while (--num_passes > 0);
+
+  __insn_drain();
+}
+
+
+#endif /* __ARCH_ICACHE_H__ */
diff --git a/kernel/arch/tile/include/uapi/arch/interrupts.h b/kernel/arch/tile/include/uapi/arch/interrupts.h
new file mode 100644
index 000000000..20f8f07d2
--- /dev/null
+++ b/kernel/arch/tile/include/uapi/arch/interrupts.h
@@ -0,0 +1,19 @@
+/*
+ * Copyright 2010 Tilera Corporation. All Rights Reserved.
+ *
+ *   This program is free software; you can redistribute it and/or
+ *   modify it under the terms of the GNU General Public License
+ *   as published by the Free Software Foundation, version 2.
+ *
+ *   This program is distributed in the hope that it will be useful, but
+ *   WITHOUT ANY WARRANTY; without even the implied warranty of
+ *   MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
+ *   NON INFRINGEMENT.  See the GNU General Public License for
+ *   more details.
+ */
+
+#ifdef __tilegx__
+#include <arch/interrupts_64.h>
+#else
+#include <arch/interrupts_32.h>
+#endif
diff --git a/kernel/arch/tile/include/uapi/arch/interrupts_32.h b/kernel/arch/tile/include/uapi/arch/interrupts_32.h
new file mode 100644
index 000000000..2efe3f68b
--- /dev/null
+++ b/kernel/arch/tile/include/uapi/arch/interrupts_32.h
@@ -0,0 +1,309 @@
+/*
+ * Copyright 2010 Tilera Corporation. All Rights Reserved.
+ *
+ *   This program is free software; you can redistribute it and/or
+ *   modify it under the terms of the GNU General Public License
+ *   as published by the Free Software Foundation, version 2.
+ *
+ *   This program is distributed in the hope that it will be useful, but
+ *   WITHOUT ANY WARRANTY; without even the implied warranty of
+ *   MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
+ *   NON INFRINGEMENT.  See the GNU General Public License for
+ *   more details.
+ */
+
+#ifndef __ARCH_INTERRUPTS_H__
+#define __ARCH_INTERRUPTS_H__
+
+#ifndef __KERNEL__
+/** Mask for an interrupt. */
+/* Note: must handle breaking interrupts into high and low words manually. */
+#define INT_MASK_LO(intno) (1 << (intno))
+#define INT_MASK_HI(intno) (1 << ((intno) - 32))
+
+#ifndef __ASSEMBLER__
+#define INT_MASK(intno) (1ULL << (intno))
+#endif
+#endif
+
+
+/** Where a given interrupt executes */
+#define INTERRUPT_VECTOR(i, pl) (0xFC000000 + ((pl) << 24) + ((i) << 8))
+
+/** Where to store a vector for a given interrupt. */
+#define USER_INTERRUPT_VECTOR(i) INTERRUPT_VECTOR(i, 0)
+
+/** The base address of user-level interrupts. */
+#define USER_INTERRUPT_VECTOR_BASE INTERRUPT_VECTOR(0, 0)
+
+
+/** Additional synthetic interrupt. */
+#define INT_BREAKPOINT (63)
+
+#define INT_ITLB_MISS    0
+#define INT_MEM_ERROR    1
+#define INT_ILL    2
+#define INT_GPV    3
+#define INT_SN_ACCESS    4
+#define INT_IDN_ACCESS    5
+#define INT_UDN_ACCESS    6
+#define INT_IDN_REFILL    7
+#define INT_UDN_REFILL    8
+#define INT_IDN_COMPLETE    9
+#define INT_UDN_COMPLETE   10
+#define INT_SWINT_3   11
+#define INT_SWINT_2   12
+#define INT_SWINT_1   13
+#define INT_SWINT_0   14
+#define INT_UNALIGN_DATA   15
+#define INT_DTLB_MISS   16
+#define INT_DTLB_ACCESS   17
+#define INT_DMATLB_MISS   18
+#define INT_DMATLB_ACCESS   19
+#define INT_SNITLB_MISS   20
+#define INT_SN_NOTIFY   21
+#define INT_SN_FIREWALL   22
+#define INT_IDN_FIREWALL   23
+#define INT_UDN_FIREWALL   24
+#define INT_TILE_TIMER   25
+#define INT_IDN_TIMER   26
+#define INT_UDN_TIMER   27
+#define INT_DMA_NOTIFY   28
+#define INT_IDN_CA   29
+#define INT_UDN_CA   30
+#define INT_IDN_AVAIL   31
+#define INT_UDN_AVAIL   32
+#define INT_PERF_COUNT   33
+#define INT_INTCTRL_3   34
+#define INT_INTCTRL_2   35
+#define INT_INTCTRL_1   36
+#define INT_INTCTRL_0   37
+#define INT_BOOT_ACCESS   38
+#define INT_WORLD_ACCESS   39
+#define INT_I_ASID   40
+#define INT_D_ASID   41
+#define INT_DMA_ASID   42
+#define INT_SNI_ASID   43
+#define INT_DMA_CPL   44
+#define INT_SN_CPL   45
+#define INT_DOUBLE_FAULT   46
+#define INT_SN_STATIC_ACCESS   47
+#define INT_AUX_PERF_COUNT   48
+
+#define NUM_INTERRUPTS 49
+
+#ifndef __ASSEMBLER__
+#define QUEUED_INTERRUPTS ( \
+    (1ULL << INT_MEM_ERROR) | \
+    (1ULL << INT_DMATLB_MISS) | \
+    (1ULL << INT_DMATLB_ACCESS) | \
+    (1ULL << INT_SNITLB_MISS) | \
+    (1ULL << INT_SN_NOTIFY) | \
+    (1ULL << INT_SN_FIREWALL) | \
+    (1ULL << INT_IDN_FIREWALL) | \
+    (1ULL << INT_UDN_FIREWALL) | \
+    (1ULL << INT_TILE_TIMER) | \
+    (1ULL << INT_IDN_TIMER) | \
+    (1ULL << INT_UDN_TIMER) | \
+    (1ULL << INT_DMA_NOTIFY) | \
+    (1ULL << INT_IDN_CA) | \
+    (1ULL << INT_UDN_CA) | \
+    (1ULL << INT_IDN_AVAIL) | \
+    (1ULL << INT_UDN_AVAIL) | \
+    (1ULL << INT_PERF_COUNT) | \
+    (1ULL << INT_INTCTRL_3) | \
+    (1ULL << INT_INTCTRL_2) | \
+    (1ULL << INT_INTCTRL_1) | \
+    (1ULL << INT_INTCTRL_0) | \
+    (1ULL << INT_BOOT_ACCESS) | \
+    (1ULL << INT_WORLD_ACCESS) | \
+    (1ULL << INT_I_ASID) | \
+    (1ULL << INT_D_ASID) | \
+    (1ULL << INT_DMA_ASID) | \
+    (1ULL << INT_SNI_ASID) | \
+    (1ULL << INT_DMA_CPL) | \
+    (1ULL << INT_SN_CPL) | \
+    (1ULL << INT_DOUBLE_FAULT) | \
+    (1ULL << INT_AUX_PERF_COUNT) | \
+    0)
+#define NONQUEUED_INTERRUPTS ( \
+    (1ULL << INT_ITLB_MISS) | \
+    (1ULL << INT_ILL) | \
+    (1ULL << INT_GPV) | \
+    (1ULL << INT_SN_ACCESS) | \
+    (1ULL << INT_IDN_ACCESS) | \
+    (1ULL << INT_UDN_ACCESS) | \
+    (1ULL << INT_IDN_REFILL) | \
+    (1ULL << INT_UDN_REFILL) | \
+    (1ULL << INT_IDN_COMPLETE) | \
+    (1ULL << INT_UDN_COMPLETE) | \
+    (1ULL << INT_SWINT_3) | \
+    (1ULL << INT_SWINT_2) | \
+    (1ULL << INT_SWINT_1) | \
+    (1ULL << INT_SWINT_0) | \
+    (1ULL << INT_UNALIGN_DATA) | \
+    (1ULL << INT_DTLB_MISS) | \
+    (1ULL << INT_DTLB_ACCESS) | \
+    (1ULL << INT_SN_STATIC_ACCESS) | \
+    0)
+#define CRITICAL_MASKED_INTERRUPTS ( \
+    (1ULL << INT_MEM_ERROR) | \
+    (1ULL << INT_DMATLB_MISS) | \
+    (1ULL << INT_DMATLB_ACCESS) | \
+    (1ULL << INT_SNITLB_MISS) | \
+    (1ULL << INT_SN_NOTIFY) | \
+    (1ULL << INT_SN_FIREWALL) | \
+    (1ULL << INT_IDN_FIREWALL) | \
+    (1ULL << INT_UDN_FIREWALL) | \
+    (1ULL << INT_TILE_TIMER) | \
+    (1ULL << INT_IDN_TIMER) | \
+    (1ULL << INT_UDN_TIMER) | \
+    (1ULL << INT_DMA_NOTIFY) | \
+    (1ULL << INT_IDN_CA) | \
+    (1ULL << INT_UDN_CA) | \
+    (1ULL << INT_IDN_AVAIL) | \
+    (1ULL << INT_UDN_AVAIL) | \
+    (1ULL << INT_PERF_COUNT) | \
+    (1ULL << INT_INTCTRL_3) | \
+    (1ULL << INT_INTCTRL_2) | \
+    (1ULL << INT_INTCTRL_1) | \
+    (1ULL << INT_INTCTRL_0) | \
+    (1ULL << INT_AUX_PERF_COUNT) | \
+    0)
+#define CRITICAL_UNMASKED_INTERRUPTS ( \
+    (1ULL << INT_ITLB_MISS) | \
+    (1ULL << INT_ILL) | \
+    (1ULL << INT_GPV) | \
+    (1ULL << INT_SN_ACCESS) | \
+    (1ULL << INT_IDN_ACCESS) | \
+    (1ULL << INT_UDN_ACCESS) | \
+    (1ULL << INT_IDN_REFILL) | \
+    (1ULL << INT_UDN_REFILL) | \
+    (1ULL << INT_IDN_COMPLETE) | \
+    (1ULL << INT_UDN_COMPLETE) | \
+    (1ULL << INT_SWINT_3) | \
+    (1ULL << INT_SWINT_2) | \
+    (1ULL << INT_SWINT_1) | \
+    (1ULL << INT_SWINT_0) | \
+    (1ULL << INT_UNALIGN_DATA) | \
+    (1ULL << INT_DTLB_MISS) | \
+    (1ULL << INT_DTLB_ACCESS) | \
+    (1ULL << INT_BOOT_ACCESS) | \
+    (1ULL << INT_WORLD_ACCESS) | \
+    (1ULL << INT_I_ASID) | \
+    (1ULL << INT_D_ASID) | \
+    (1ULL << INT_DMA_ASID) | \
+    (1ULL << INT_SNI_ASID) | \
+    (1ULL << INT_DMA_CPL) | \
+    (1ULL << INT_SN_CPL) | \
+    (1ULL << INT_DOUBLE_FAULT) | \
+    (1ULL << INT_SN_STATIC_ACCESS) | \
+    0)
+#define MASKABLE_INTERRUPTS ( \
+    (1ULL << INT_MEM_ERROR) | \
+    (1ULL << INT_IDN_REFILL) | \
+    (1ULL << INT_UDN_REFILL) | \
+    (1ULL << INT_IDN_COMPLETE) | \
+    (1ULL << INT_UDN_COMPLETE) | \
+    (1ULL << INT_DMATLB_MISS) | \
+    (1ULL << INT_DMATLB_ACCESS) | \
+    (1ULL << INT_SNITLB_MISS) | \
+    (1ULL << INT_SN_NOTIFY) | \
+    (1ULL << INT_SN_FIREWALL) | \
+    (1ULL << INT_IDN_FIREWALL) | \
+    (1ULL << INT_UDN_FIREWALL) | \
+    (1ULL << INT_TILE_TIMER) | \
+    (1ULL << INT_IDN_TIMER) | \
+    (1ULL << INT_UDN_TIMER) | \
+    (1ULL << INT_DMA_NOTIFY) | \
+    (1ULL << INT_IDN_CA) | \
+    (1ULL << INT_UDN_CA) | \
+    (1ULL << INT_IDN_AVAIL) | \
+    (1ULL << INT_UDN_AVAIL) | \
+    (1ULL << INT_PERF_COUNT) | \
+    (1ULL << INT_INTCTRL_3) | \
+    (1ULL << INT_INTCTRL_2) | \
+    (1ULL << INT_INTCTRL_1) | \
+    (1ULL << INT_INTCTRL_0) | \
+    (1ULL << INT_AUX_PERF_COUNT) | \
+    0)
+#define UNMASKABLE_INTERRUPTS ( \
+    (1ULL << INT_ITLB_MISS) | \
+    (1ULL << INT_ILL) | \
+    (1ULL << INT_GPV) | \
+    (1ULL << INT_SN_ACCESS) | \
+    (1ULL << INT_IDN_ACCESS) | \
+    (1ULL << INT_UDN_ACCESS) | \
+    (1ULL << INT_SWINT_3) | \
+    (1ULL << INT_SWINT_2) | \
+    (1ULL << INT_SWINT_1) | \
+    (1ULL << INT_SWINT_0) | \
+    (1ULL << INT_UNALIGN_DATA) | \
+    (1ULL << INT_DTLB_MISS) | \
+    (1ULL << INT_DTLB_ACCESS) | \
+    (1ULL << INT_BOOT_ACCESS) | \
+    (1ULL << INT_WORLD_ACCESS) | \
+    (1ULL << INT_I_ASID) | \
+    (1ULL << INT_D_ASID) | \
+    (1ULL << INT_DMA_ASID) | \
+    (1ULL << INT_SNI_ASID) | \
+    (1ULL << INT_DMA_CPL) | \
+    (1ULL << INT_SN_CPL) | \
+    (1ULL << INT_DOUBLE_FAULT) | \
+    (1ULL << INT_SN_STATIC_ACCESS) | \
+    0)
+#define SYNC_INTERRUPTS ( \
+    (1ULL << INT_ITLB_MISS) | \
+    (1ULL << INT_ILL) | \
+    (1ULL << INT_GPV) | \
+    (1ULL << INT_SN_ACCESS) | \
+    (1ULL << INT_IDN_ACCESS) | \
+    (1ULL << INT_UDN_ACCESS) | \
+    (1ULL << INT_IDN_REFILL) | \
+    (1ULL << INT_UDN_REFILL) | \
+    (1ULL << INT_IDN_COMPLETE) | \
+    (1ULL << INT_UDN_COMPLETE) | \
+    (1ULL << INT_SWINT_3) | \
+    (1ULL << INT_SWINT_2) | \
+    (1ULL << INT_SWINT_1) | \
+    (1ULL << INT_SWINT_0) | \
+    (1ULL << INT_UNALIGN_DATA) | \
+    (1ULL << INT_DTLB_MISS) | \
+    (1ULL << INT_DTLB_ACCESS) | \
+    (1ULL << INT_SN_STATIC_ACCESS) | \
+    0)
+#define NON_SYNC_INTERRUPTS ( \
+    (1ULL << INT_MEM_ERROR) | \
+    (1ULL << INT_DMATLB_MISS) | \
+    (1ULL << INT_DMATLB_ACCESS) | \
+    (1ULL << INT_SNITLB_MISS) | \
+    (1ULL << INT_SN_NOTIFY) | \
+    (1ULL << INT_SN_FIREWALL) | \
+    (1ULL << INT_IDN_FIREWALL) | \
+    (1ULL << INT_UDN_FIREWALL) | \
+    (1ULL << INT_TILE_TIMER) | \
+    (1ULL << INT_IDN_TIMER) | \
+    (1ULL << INT_UDN_TIMER) | \
+    (1ULL << INT_DMA_NOTIFY) | \
+    (1ULL << INT_IDN_CA) | \
+    (1ULL << INT_UDN_CA) | \
+    (1ULL << INT_IDN_AVAIL) | \
+    (1ULL << INT_UDN_AVAIL) | \
+    (1ULL << INT_PERF_COUNT) | \
+    (1ULL << INT_INTCTRL_3) | \
+    (1ULL << INT_INTCTRL_2) | \
+    (1ULL << INT_INTCTRL_1) | \
+    (1ULL << INT_INTCTRL_0) | \
+    (1ULL << INT_BOOT_ACCESS) | \
+    (1ULL << INT_WORLD_ACCESS) | \
+    (1ULL << INT_I_ASID) | \
+    (1ULL << INT_D_ASID) | \
+    (1ULL << INT_DMA_ASID) | \
+    (1ULL << INT_SNI_ASID) | \
+    (1ULL << INT_DMA_CPL) | \
+    (1ULL << INT_SN_CPL) | \
+    (1ULL << INT_DOUBLE_FAULT) | \
+    (1ULL << INT_AUX_PERF_COUNT) | \
+    0)
+#endif /* !__ASSEMBLER__ */
+#endif /* !__ARCH_INTERRUPTS_H__ */
diff --git a/kernel/arch/tile/include/uapi/arch/interrupts_64.h b/kernel/arch/tile/include/uapi/arch/interrupts_64.h
new file mode 100644
index 000000000..13c9f9182
--- /dev/null
+++ b/kernel/arch/tile/include/uapi/arch/interrupts_64.h
@@ -0,0 +1,278 @@
+/*
+ * Copyright 2011 Tilera Corporation. All Rights Reserved.
+ *
+ *   This program is free software; you can redistribute it and/or
+ *   modify it under the terms of the GNU General Public License
+ *   as published by the Free Software Foundation, version 2.
+ *
+ *   This program is distributed in the hope that it will be useful, but
+ *   WITHOUT ANY WARRANTY; without even the implied warranty of
+ *   MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
+ *   NON INFRINGEMENT.  See the GNU General Public License for
+ *   more details.
+ */
+
+#ifndef __ARCH_INTERRUPTS_H__
+#define __ARCH_INTERRUPTS_H__
+
+#ifndef __KERNEL__
+/** Mask for an interrupt. */
+#ifdef __ASSEMBLER__
+/* Note: must handle breaking interrupts into high and low words manually. */
+#define INT_MASK(intno) (1 << (intno))
+#else
+#define INT_MASK(intno) (1ULL << (intno))
+#endif
+#endif
+
+
+/** Where a given interrupt executes */
+#define INTERRUPT_VECTOR(i, pl) (0xFC000000 + ((pl) << 24) + ((i) << 8))
+
+/** Where to store a vector for a given interrupt. */
+#define USER_INTERRUPT_VECTOR(i) INTERRUPT_VECTOR(i, 0)
+
+/** The base address of user-level interrupts. */
+#define USER_INTERRUPT_VECTOR_BASE INTERRUPT_VECTOR(0, 0)
+
+
+/** Additional synthetic interrupt. */
+#define INT_BREAKPOINT (63)
+
+#define INT_MEM_ERROR    0
+#define INT_SINGLE_STEP_3    1
+#define INT_SINGLE_STEP_2    2
+#define INT_SINGLE_STEP_1    3
+#define INT_SINGLE_STEP_0    4
+#define INT_IDN_COMPLETE    5
+#define INT_UDN_COMPLETE    6
+#define INT_ITLB_MISS    7
+#define INT_ILL    8
+#define INT_GPV    9
+#define INT_IDN_ACCESS   10
+#define INT_UDN_ACCESS   11
+#define INT_SWINT_3   12
+#define INT_SWINT_2   13
+#define INT_SWINT_1   14
+#define INT_SWINT_0   15
+#define INT_ILL_TRANS   16
+#define INT_UNALIGN_DATA   17
+#define INT_DTLB_MISS   18
+#define INT_DTLB_ACCESS   19
+#define INT_IDN_FIREWALL   20
+#define INT_UDN_FIREWALL   21
+#define INT_TILE_TIMER   22
+#define INT_AUX_TILE_TIMER   23
+#define INT_IDN_TIMER   24
+#define INT_UDN_TIMER   25
+#define INT_IDN_AVAIL   26
+#define INT_UDN_AVAIL   27
+#define INT_IPI_3   28
+#define INT_IPI_2   29
+#define INT_IPI_1   30
+#define INT_IPI_0   31
+#define INT_PERF_COUNT   32
+#define INT_AUX_PERF_COUNT   33
+#define INT_INTCTRL_3   34
+#define INT_INTCTRL_2   35
+#define INT_INTCTRL_1   36
+#define INT_INTCTRL_0   37
+#define INT_BOOT_ACCESS   38
+#define INT_WORLD_ACCESS   39
+#define INT_I_ASID   40
+#define INT_D_ASID   41
+#define INT_DOUBLE_FAULT   42
+
+#define NUM_INTERRUPTS 43
+
+#ifndef __ASSEMBLER__
+#define QUEUED_INTERRUPTS ( \
+    (1ULL << INT_MEM_ERROR) | \
+    (1ULL << INT_IDN_COMPLETE) | \
+    (1ULL << INT_UDN_COMPLETE) | \
+    (1ULL << INT_IDN_FIREWALL) | \
+    (1ULL << INT_UDN_FIREWALL) | \
+    (1ULL << INT_TILE_TIMER) | \
+    (1ULL << INT_AUX_TILE_TIMER) | \
+    (1ULL << INT_IDN_TIMER) | \
+    (1ULL << INT_UDN_TIMER) | \
+    (1ULL << INT_IDN_AVAIL) | \
+    (1ULL << INT_UDN_AVAIL) | \
+    (1ULL << INT_IPI_3) | \
+    (1ULL << INT_IPI_2) | \
+    (1ULL << INT_IPI_1) | \
+    (1ULL << INT_IPI_0) | \
+    (1ULL << INT_PERF_COUNT) | \
+    (1ULL << INT_AUX_PERF_COUNT) | \
+    (1ULL << INT_INTCTRL_3) | \
+    (1ULL << INT_INTCTRL_2) | \
+    (1ULL << INT_INTCTRL_1) | \
+    (1ULL << INT_INTCTRL_0) | \
+    (1ULL << INT_BOOT_ACCESS) | \
+    (1ULL << INT_WORLD_ACCESS) | \
+    (1ULL << INT_I_ASID) | \
+    (1ULL << INT_D_ASID) | \
+    (1ULL << INT_DOUBLE_FAULT) | \
+    0)
+#define NONQUEUED_INTERRUPTS ( \
+    (1ULL << INT_SINGLE_STEP_3) | \
+    (1ULL << INT_SINGLE_STEP_2) | \
+    (1ULL << INT_SINGLE_STEP_1) | \
+    (1ULL << INT_SINGLE_STEP_0) | \
+    (1ULL << INT_ITLB_MISS) | \
+    (1ULL << INT_ILL) | \
+    (1ULL << INT_GPV) | \
+    (1ULL << INT_IDN_ACCESS) | \
+    (1ULL << INT_UDN_ACCESS) | \
+    (1ULL << INT_SWINT_3) | \
+    (1ULL << INT_SWINT_2) | \
+    (1ULL << INT_SWINT_1) | \
+    (1ULL << INT_SWINT_0) | \
+    (1ULL << INT_ILL_TRANS) | \
+    (1ULL << INT_UNALIGN_DATA) | \
+    (1ULL << INT_DTLB_MISS) | \
+    (1ULL << INT_DTLB_ACCESS) | \
+    0)
+#define CRITICAL_MASKED_INTERRUPTS ( \
+    (1ULL << INT_MEM_ERROR) | \
+    (1ULL << INT_SINGLE_STEP_3) | \
+    (1ULL << INT_SINGLE_STEP_2) | \
+    (1ULL << INT_SINGLE_STEP_1) | \
+    (1ULL << INT_SINGLE_STEP_0) | \
+    (1ULL << INT_IDN_COMPLETE) | \
+    (1ULL << INT_UDN_COMPLETE) | \
+    (1ULL << INT_IDN_FIREWALL) | \
+    (1ULL << INT_UDN_FIREWALL) | \
+    (1ULL << INT_TILE_TIMER) | \
+    (1ULL << INT_AUX_TILE_TIMER) | \
+    (1ULL << INT_IDN_TIMER) | \
+    (1ULL << INT_UDN_TIMER) | \
+    (1ULL << INT_IDN_AVAIL) | \
+    (1ULL << INT_UDN_AVAIL) | \
+    (1ULL << INT_IPI_3) | \
+    (1ULL << INT_IPI_2) | \
+    (1ULL << INT_IPI_1) | \
+    (1ULL << INT_IPI_0) | \
+    (1ULL << INT_PERF_COUNT) | \
+    (1ULL << INT_AUX_PERF_COUNT) | \
+    (1ULL << INT_INTCTRL_3) | \
+    (1ULL << INT_INTCTRL_2) | \
+    (1ULL << INT_INTCTRL_1) | \
+    (1ULL << INT_INTCTRL_0) | \
+    0)
+#define CRITICAL_UNMASKED_INTERRUPTS ( \
+    (1ULL << INT_ITLB_MISS) | \
+    (1ULL << INT_ILL) | \
+    (1ULL << INT_GPV) | \
+    (1ULL << INT_IDN_ACCESS) | \
+    (1ULL << INT_UDN_ACCESS) | \
+    (1ULL << INT_SWINT_3) | \
+    (1ULL << INT_SWINT_2) | \
+    (1ULL << INT_SWINT_1) | \
+    (1ULL << INT_SWINT_0) | \
+    (1ULL << INT_ILL_TRANS) | \
+    (1ULL << INT_UNALIGN_DATA) | \
+    (1ULL << INT_DTLB_MISS) | \
+    (1ULL << INT_DTLB_ACCESS) | \
+    (1ULL << INT_BOOT_ACCESS) | \
+    (1ULL << INT_WORLD_ACCESS) | \
+    (1ULL << INT_I_ASID) | \
+    (1ULL << INT_D_ASID) | \
+    (1ULL << INT_DOUBLE_FAULT) | \
+    0)
+#define MASKABLE_INTERRUPTS ( \
+    (1ULL << INT_MEM_ERROR) | \
+    (1ULL << INT_SINGLE_STEP_3) | \
+    (1ULL << INT_SINGLE_STEP_2) | \
+    (1ULL << INT_SINGLE_STEP_1) | \
+    (1ULL << INT_SINGLE_STEP_0) | \
+    (1ULL << INT_IDN_COMPLETE) | \
+    (1ULL << INT_UDN_COMPLETE) | \
+    (1ULL << INT_IDN_FIREWALL) | \
+    (1ULL << INT_UDN_FIREWALL) | \
+    (1ULL << INT_TILE_TIMER) | \
+    (1ULL << INT_AUX_TILE_TIMER) | \
+    (1ULL << INT_IDN_TIMER) | \
+    (1ULL << INT_UDN_TIMER) | \
+    (1ULL << INT_IDN_AVAIL) | \
+    (1ULL << INT_UDN_AVAIL) | \
+    (1ULL << INT_IPI_3) | \
+    (1ULL << INT_IPI_2) | \
+    (1ULL << INT_IPI_1) | \
+    (1ULL << INT_IPI_0) | \
+    (1ULL << INT_PERF_COUNT) | \
+    (1ULL << INT_AUX_PERF_COUNT) | \
+    (1ULL << INT_INTCTRL_3) | \
+    (1ULL << INT_INTCTRL_2) | \
+    (1ULL << INT_INTCTRL_1) | \
+    (1ULL << INT_INTCTRL_0) | \
+    0)
+#define UNMASKABLE_INTERRUPTS ( \
+    (1ULL << INT_ITLB_MISS) | \
+    (1ULL << INT_ILL) | \
+    (1ULL << INT_GPV) | \
+    (1ULL << INT_IDN_ACCESS) | \
+    (1ULL << INT_UDN_ACCESS) | \
+    (1ULL << INT_SWINT_3) | \
+    (1ULL << INT_SWINT_2) | \
+    (1ULL << INT_SWINT_1) | \
+    (1ULL << INT_SWINT_0) | \
+    (1ULL << INT_ILL_TRANS) | \
+    (1ULL << INT_UNALIGN_DATA) | \
+    (1ULL << INT_DTLB_MISS) | \
+    (1ULL << INT_DTLB_ACCESS) | \
+    (1ULL << INT_BOOT_ACCESS) | \
+    (1ULL << INT_WORLD_ACCESS) | \
+    (1ULL << INT_I_ASID) | \
+    (1ULL << INT_D_ASID) | \
+    (1ULL << INT_DOUBLE_FAULT) | \
+    0)
+#define SYNC_INTERRUPTS ( \
+    (1ULL << INT_SINGLE_STEP_3) | \
+    (1ULL << INT_SINGLE_STEP_2) | \
+    (1ULL << INT_SINGLE_STEP_1) | \
+    (1ULL << INT_SINGLE_STEP_0) | \
+    (1ULL << INT_IDN_COMPLETE) | \
+    (1ULL << INT_UDN_COMPLETE) | \
+    (1ULL << INT_ITLB_MISS) | \
+    (1ULL << INT_ILL) | \
+    (1ULL << INT_GPV) | \
+    (1ULL << INT_IDN_ACCESS) | \
+    (1ULL << INT_UDN_ACCESS) | \
+    (1ULL << INT_SWINT_3) | \
+    (1ULL << INT_SWINT_2) | \
+    (1ULL << INT_SWINT_1) | \
+    (1ULL << INT_SWINT_0) | \
+    (1ULL << INT_ILL_TRANS) | \
+    (1ULL << INT_UNALIGN_DATA) | \
+    (1ULL << INT_DTLB_MISS) | \
+    (1ULL << INT_DTLB_ACCESS) | \
+    0)
+#define NON_SYNC_INTERRUPTS ( \
+    (1ULL << INT_MEM_ERROR) | \
+    (1ULL << INT_IDN_FIREWALL) | \
+    (1ULL << INT_UDN_FIREWALL) | \
+    (1ULL << INT_TILE_TIMER) | \
+    (1ULL << INT_AUX_TILE_TIMER) | \
+    (1ULL << INT_IDN_TIMER) | \
+    (1ULL << INT_UDN_TIMER) | \
+    (1ULL << INT_IDN_AVAIL) | \
+    (1ULL << INT_UDN_AVAIL) | \
+    (1ULL << INT_IPI_3) | \
+    (1ULL << INT_IPI_2) | \
+    (1ULL << INT_IPI_1) | \
+    (1ULL << INT_IPI_0) | \
+    (1ULL << INT_PERF_COUNT) | \
+    (1ULL << INT_AUX_PERF_COUNT) | \
+    (1ULL << INT_INTCTRL_3) | \
+    (1ULL << INT_INTCTRL_2) | \
+    (1ULL << INT_INTCTRL_1) | \
+    (1ULL << INT_INTCTRL_0) | \
+    (1ULL << INT_BOOT_ACCESS) | \
+    (1ULL << INT_WORLD_ACCESS) | \
+    (1ULL << INT_I_ASID) | \
+    (1ULL << INT_D_ASID) | \
+    (1ULL << INT_DOUBLE_FAULT) | \
+    0)
+#endif /* !__ASSEMBLER__ */
+#endif /* !__ARCH_INTERRUPTS_H__ */
diff --git a/kernel/arch/tile/include/uapi/arch/opcode.h b/kernel/arch/tile/include/uapi/arch/opcode.h
new file mode 100644
index 000000000..92d15229e
--- /dev/null
+++ b/kernel/arch/tile/include/uapi/arch/opcode.h
@@ -0,0 +1,21 @@
+/*
+ * Copyright 2011 Tilera Corporation. All Rights Reserved.
+ *
+ *   This program is free software; you can redistribute it and/or
+ *   modify it under the terms of the GNU General Public License
+ *   as published by the Free Software Foundation, version 2.
+ *
+ *   This program is distributed in the hope that it will be useful, but
+ *   WITHOUT ANY WARRANTY; without even the implied warranty of
+ *   MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
+ *   NON INFRINGEMENT.  See the GNU General Public License for
+ *   more details.
+ */
+
+#if defined(__tilepro__)
+#include <arch/opcode_tilepro.h>
+#elif defined(__tilegx__)
+#include <arch/opcode_tilegx.h>
+#else
+#error Unexpected Tilera chip type
+#endif
diff --git a/kernel/arch/tile/include/uapi/arch/opcode_tilegx.h b/kernel/arch/tile/include/uapi/arch/opcode_tilegx.h
new file mode 100644
index 000000000..d76ff2db7
--- /dev/null
+++ b/kernel/arch/tile/include/uapi/arch/opcode_tilegx.h
@@ -0,0 +1,1406 @@
+/* TILE-Gx opcode information.
+ *
+ * Copyright 2011 Tilera Corporation. All Rights Reserved.
+ *
+ *   This program is free software; you can redistribute it and/or
+ *   modify it under the terms of the GNU General Public License
+ *   as published by the Free Software Foundation, version 2.
+ *
+ *   This program is distributed in the hope that it will be useful, but
+ *   WITHOUT ANY WARRANTY; without even the implied warranty of
+ *   MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
+ *   NON INFRINGEMENT.  See the GNU General Public License for
+ *   more details.
+ *
+ *
+ *
+ *
+ *
+ */
+
+#ifndef __ARCH_OPCODE_H__
+#define __ARCH_OPCODE_H__
+
+#ifndef __ASSEMBLER__
+
+typedef unsigned long long tilegx_bundle_bits;
+
+/* These are the bits that determine if a bundle is in the X encoding. */
+#define TILEGX_BUNDLE_MODE_MASK ((tilegx_bundle_bits)3 << 62)
+
+enum
+{
+  /* Maximum number of instructions in a bundle (2 for X, 3 for Y). */
+  TILEGX_MAX_INSTRUCTIONS_PER_BUNDLE = 3,
+
+  /* How many different pipeline encodings are there? X0, X1, Y0, Y1, Y2. */
+  TILEGX_NUM_PIPELINE_ENCODINGS = 5,
+
+  /* Log base 2 of TILEGX_BUNDLE_SIZE_IN_BYTES. */
+  TILEGX_LOG2_BUNDLE_SIZE_IN_BYTES = 3,
+
+  /* Instructions take this many bytes. */
+  TILEGX_BUNDLE_SIZE_IN_BYTES = 1 << TILEGX_LOG2_BUNDLE_SIZE_IN_BYTES,
+
+  /* Log base 2 of TILEGX_BUNDLE_ALIGNMENT_IN_BYTES. */
+  TILEGX_LOG2_BUNDLE_ALIGNMENT_IN_BYTES = 3,
+
+  /* Bundles should be aligned modulo this number of bytes. */
+  TILEGX_BUNDLE_ALIGNMENT_IN_BYTES =
+    (1 << TILEGX_LOG2_BUNDLE_ALIGNMENT_IN_BYTES),
+
+  /* Number of registers (some are magic, such as network I/O). */
+  TILEGX_NUM_REGISTERS = 64,
+};
+
+/* Make a few "tile_" variables to simplify common code between
+   architectures.  */
+
+typedef tilegx_bundle_bits tile_bundle_bits;
+#define TILE_BUNDLE_SIZE_IN_BYTES TILEGX_BUNDLE_SIZE_IN_BYTES
+#define TILE_BUNDLE_ALIGNMENT_IN_BYTES TILEGX_BUNDLE_ALIGNMENT_IN_BYTES
+#define TILE_LOG2_BUNDLE_ALIGNMENT_IN_BYTES \
+  TILEGX_LOG2_BUNDLE_ALIGNMENT_IN_BYTES
+#define TILE_BPT_BUNDLE TILEGX_BPT_BUNDLE
+
+/* 64-bit pattern for a { bpt ; nop } bundle. */
+#define TILEGX_BPT_BUNDLE 0x286a44ae51485000ULL
+
+static __inline unsigned int
+get_BFEnd_X0(tilegx_bundle_bits num)
+{
+  const unsigned int n = (unsigned int)num;
+  return (((n >> 12)) & 0x3f);
+}
+
+static __inline unsigned int
+get_BFOpcodeExtension_X0(tilegx_bundle_bits num)
+{
+  const unsigned int n = (unsigned int)num;
+  return (((n >> 24)) & 0xf);
+}
+
+static __inline unsigned int
+get_BFStart_X0(tilegx_bundle_bits num)
+{
+  const unsigned int n = (unsigned int)num;
+  return (((n >> 18)) & 0x3f);
+}
+
+static __inline unsigned int
+get_BrOff_X1(tilegx_bundle_bits n)
+{
+  return (((unsigned int)(n >> 31)) & 0x0000003f) |
+         (((unsigned int)(n >> 37)) & 0x0001ffc0);
+}
+
+static __inline unsigned int
+get_BrType_X1(tilegx_bundle_bits n)
+{
+  return (((unsigned int)(n >> 54)) & 0x1f);
+}
+
+static __inline unsigned int
+get_Dest_Imm8_X1(tilegx_bundle_bits n)
+{
+  return (((unsigned int)(n >> 31)) & 0x0000003f) |
+         (((unsigned int)(n >> 43)) & 0x000000c0);
+}
+
+static __inline unsigned int
+get_Dest_X0(tilegx_bundle_bits num)
+{
+  const unsigned int n = (unsigned int)num;
+  return (((n >> 0)) & 0x3f);
+}
+
+static __inline unsigned int
+get_Dest_X1(tilegx_bundle_bits n)
+{
+  return (((unsigned int)(n >> 31)) & 0x3f);
+}
+
+static __inline unsigned int
+get_Dest_Y0(tilegx_bundle_bits num)
+{
+  const unsigned int n = (unsigned int)num;
+  return (((n >> 0)) & 0x3f);
+}
+
+static __inline unsigned int
+get_Dest_Y1(tilegx_bundle_bits n)
+{
+  return (((unsigned int)(n >> 31)) & 0x3f);
+}
+
+static __inline unsigned int
+get_Imm16_X0(tilegx_bundle_bits num)
+{
+  const unsigned int n = (unsigned int)num;
+  return (((n >> 12)) & 0xffff);
+}
+
+static __inline unsigned int
+get_Imm16_X1(tilegx_bundle_bits n)
+{
+  return (((unsigned int)(n >> 43)) & 0xffff);
+}
+
+static __inline unsigned int
+get_Imm8OpcodeExtension_X0(tilegx_bundle_bits num)
+{
+  const unsigned int n = (unsigned int)num;
+  return (((n >> 20)) & 0xff);
+}
+
+static __inline unsigned int
+get_Imm8OpcodeExtension_X1(tilegx_bundle_bits n)
+{
+  return (((unsigned int)(n >> 51)) & 0xff);
+}
+
+static __inline unsigned int
+get_Imm8_X0(tilegx_bundle_bits num)
+{
+  const unsigned int n = (unsigned int)num;
+  return (((n >> 12)) & 0xff);
+}
+
+static __inline unsigned int
+get_Imm8_X1(tilegx_bundle_bits n)
+{
+  return (((unsigned int)(n >> 43)) & 0xff);
+}
+
+static __inline unsigned int
+get_Imm8_Y0(tilegx_bundle_bits num)
+{
+  const unsigned int n = (unsigned int)num;
+  return (((n >> 12)) & 0xff);
+}
+
+static __inline unsigned int
+get_Imm8_Y1(tilegx_bundle_bits n)
+{
+  return (((unsigned int)(n >> 43)) & 0xff);
+}
+
+static __inline unsigned int
+get_JumpOff_X1(tilegx_bundle_bits n)
+{
+  return (((unsigned int)(n >> 31)) & 0x7ffffff);
+}
+
+static __inline unsigned int
+get_JumpOpcodeExtension_X1(tilegx_bundle_bits n)
+{
+  return (((unsigned int)(n >> 58)) & 0x1);
+}
+
+static __inline unsigned int
+get_MF_Imm14_X1(tilegx_bundle_bits n)
+{
+  return (((unsigned int)(n >> 37)) & 0x3fff);
+}
+
+static __inline unsigned int
+get_MT_Imm14_X1(tilegx_bundle_bits n)
+{
+  return (((unsigned int)(n >> 31)) & 0x0000003f) |
+         (((unsigned int)(n >> 37)) & 0x00003fc0);
+}
+
+static __inline unsigned int
+get_Mode(tilegx_bundle_bits n)
+{
+  return (((unsigned int)(n >> 62)) & 0x3);
+}
+
+static __inline unsigned int
+get_Opcode_X0(tilegx_bundle_bits num)
+{
+  const unsigned int n = (unsigned int)num;
+  return (((n >> 28)) & 0x7);
+}
+
+static __inline unsigned int
+get_Opcode_X1(tilegx_bundle_bits n)
+{
+  return (((unsigned int)(n >> 59)) & 0x7);
+}
+
+static __inline unsigned int
+get_Opcode_Y0(tilegx_bundle_bits num)
+{
+  const unsigned int n = (unsigned int)num;
+  return (((n >> 27)) & 0xf);
+}
+
+static __inline unsigned int
+get_Opcode_Y1(tilegx_bundle_bits n)
+{
+  return (((unsigned int)(n >> 58)) & 0xf);
+}
+
+static __inline unsigned int
+get_Opcode_Y2(tilegx_bundle_bits n)
+{
+  return (((n >> 26)) & 0x00000001) |
+         (((unsigned int)(n >> 56)) & 0x00000002);
+}
+
+static __inline unsigned int
+get_RRROpcodeExtension_X0(tilegx_bundle_bits num)
+{
+  const unsigned int n = (unsigned int)num;
+  return (((n >> 18)) & 0x3ff);
+}
+
+static __inline unsigned int
+get_RRROpcodeExtension_X1(tilegx_bundle_bits n)
+{
+  return (((unsigned int)(n >> 49)) & 0x3ff);
+}
+
+static __inline unsigned int
+get_RRROpcodeExtension_Y0(tilegx_bundle_bits num)
+{
+  const unsigned int n = (unsigned int)num;
+  return (((n >> 18)) & 0x3);
+}
+
+static __inline unsigned int
+get_RRROpcodeExtension_Y1(tilegx_bundle_bits n)
+{
+  return (((unsigned int)(n >> 49)) & 0x3);
+}
+
+static __inline unsigned int
+get_ShAmt_X0(tilegx_bundle_bits num)
+{
+  const unsigned int n = (unsigned int)num;
+  return (((n >> 12)) & 0x3f);
+}
+
+static __inline unsigned int
+get_ShAmt_X1(tilegx_bundle_bits n)
+{
+  return (((unsigned int)(n >> 43)) & 0x3f);
+}
+
+static __inline unsigned int
+get_ShAmt_Y0(tilegx_bundle_bits num)
+{
+  const unsigned int n = (unsigned int)num;
+  return (((n >> 12)) & 0x3f);
+}
+
+static __inline unsigned int
+get_ShAmt_Y1(tilegx_bundle_bits n)
+{
+  return (((unsigned int)(n >> 43)) & 0x3f);
+}
+
+static __inline unsigned int
+get_ShiftOpcodeExtension_X0(tilegx_bundle_bits num)
+{
+  const unsigned int n = (unsigned int)num;
+  return (((n >> 18)) & 0x3ff);
+}
+
+static __inline unsigned int
+get_ShiftOpcodeExtension_X1(tilegx_bundle_bits n)
+{
+  return (((unsigned int)(n >> 49)) & 0x3ff);
+}
+
+static __inline unsigned int
+get_ShiftOpcodeExtension_Y0(tilegx_bundle_bits num)
+{
+  const unsigned int n = (unsigned int)num;
+  return (((n >> 18)) & 0x3);
+}
+
+static __inline unsigned int
+get_ShiftOpcodeExtension_Y1(tilegx_bundle_bits n)
+{
+  return (((unsigned int)(n >> 49)) & 0x3);
+}
+
+static __inline unsigned int
+get_SrcA_X0(tilegx_bundle_bits num)
+{
+  const unsigned int n = (unsigned int)num;
+  return (((n >> 6)) & 0x3f);
+}
+
+static __inline unsigned int
+get_SrcA_X1(tilegx_bundle_bits n)
+{
+  return (((unsigned int)(n >> 37)) & 0x3f);
+}
+
+static __inline unsigned int
+get_SrcA_Y0(tilegx_bundle_bits num)
+{
+  const unsigned int n = (unsigned int)num;
+  return (((n >> 6)) & 0x3f);
+}
+
+static __inline unsigned int
+get_SrcA_Y1(tilegx_bundle_bits n)
+{
+  return (((unsigned int)(n >> 37)) & 0x3f);
+}
+
+static __inline unsigned int
+get_SrcA_Y2(tilegx_bundle_bits num)
+{
+  const unsigned int n = (unsigned int)num;
+  return (((n >> 20)) & 0x3f);
+}
+
+static __inline unsigned int
+get_SrcBDest_Y2(tilegx_bundle_bits n)
+{
+  return (((unsigned int)(n >> 51)) & 0x3f);
+}
+
+static __inline unsigned int
+get_SrcB_X0(tilegx_bundle_bits num)
+{
+  const unsigned int n = (unsigned int)num;
+  return (((n >> 12)) & 0x3f);
+}
+
+static __inline unsigned int
+get_SrcB_X1(tilegx_bundle_bits n)
+{
+  return (((unsigned int)(n >> 43)) & 0x3f);
+}
+
+static __inline unsigned int
+get_SrcB_Y0(tilegx_bundle_bits num)
+{
+  const unsigned int n = (unsigned int)num;
+  return (((n >> 12)) & 0x3f);
+}
+
+static __inline unsigned int
+get_SrcB_Y1(tilegx_bundle_bits n)
+{
+  return (((unsigned int)(n >> 43)) & 0x3f);
+}
+
+static __inline unsigned int
+get_UnaryOpcodeExtension_X0(tilegx_bundle_bits num)
+{
+  const unsigned int n = (unsigned int)num;
+  return (((n >> 12)) & 0x3f);
+}
+
+static __inline unsigned int
+get_UnaryOpcodeExtension_X1(tilegx_bundle_bits n)
+{
+  return (((unsigned int)(n >> 43)) & 0x3f);
+}
+
+static __inline unsigned int
+get_UnaryOpcodeExtension_Y0(tilegx_bundle_bits num)
+{
+  const unsigned int n = (unsigned int)num;
+  return (((n >> 12)) & 0x3f);
+}
+
+static __inline unsigned int
+get_UnaryOpcodeExtension_Y1(tilegx_bundle_bits n)
+{
+  return (((unsigned int)(n >> 43)) & 0x3f);
+}
+
+
+static __inline int
+sign_extend(int n, int num_bits)
+{
+  int shift = (int)(sizeof(int) * 8 - num_bits);
+  return (n << shift) >> shift;
+}
+
+
+
+static __inline tilegx_bundle_bits
+create_BFEnd_X0(int num)
+{
+  const unsigned int n = (unsigned int)num;
+  return ((n & 0x3f) << 12);
+}
+
+static __inline tilegx_bundle_bits
+create_BFOpcodeExtension_X0(int num)
+{
+  const unsigned int n = (unsigned int)num;
+  return ((n & 0xf) << 24);
+}
+
+static __inline tilegx_bundle_bits
+create_BFStart_X0(int num)
+{
+  const unsigned int n = (unsigned int)num;
+  return ((n & 0x3f) << 18);
+}
+
+static __inline tilegx_bundle_bits
+create_BrOff_X1(int num)
+{
+  const unsigned int n = (unsigned int)num;
+  return (((tilegx_bundle_bits)(n & 0x0000003f)) << 31) |
+         (((tilegx_bundle_bits)(n & 0x0001ffc0)) << 37);
+}
+
+static __inline tilegx_bundle_bits
+create_BrType_X1(int num)
+{
+  const unsigned int n = (unsigned int)num;
+  return (((tilegx_bundle_bits)(n & 0x1f)) << 54);
+}
+
+static __inline tilegx_bundle_bits
+create_Dest_Imm8_X1(int num)
+{
+  const unsigned int n = (unsigned int)num;
+  return (((tilegx_bundle_bits)(n & 0x0000003f)) << 31) |
+         (((tilegx_bundle_bits)(n & 0x000000c0)) << 43);
+}
+
+static __inline tilegx_bundle_bits
+create_Dest_X0(int num)
+{
+  const unsigned int n = (unsigned int)num;
+  return ((n & 0x3f) << 0);
+}
+
+static __inline tilegx_bundle_bits
+create_Dest_X1(int num)
+{
+  const unsigned int n = (unsigned int)num;
+  return (((tilegx_bundle_bits)(n & 0x3f)) << 31);
+}
+
+static __inline tilegx_bundle_bits
+create_Dest_Y0(int num)
+{
+  const unsigned int n = (unsigned int)num;
+  return ((n & 0x3f) << 0);
+}
+
+static __inline tilegx_bundle_bits
+create_Dest_Y1(int num)
+{
+  const unsigned int n = (unsigned int)num;
+  return (((tilegx_bundle_bits)(n & 0x3f)) << 31);
+}
+
+static __inline tilegx_bundle_bits
+create_Imm16_X0(int num)
+{
+  const unsigned int n = (unsigned int)num;
+  return ((n & 0xffff) << 12);
+}
+
+static __inline tilegx_bundle_bits
+create_Imm16_X1(int num)
+{
+  const unsigned int n = (unsigned int)num;
+  return (((tilegx_bundle_bits)(n & 0xffff)) << 43);
+}
+
+static __inline tilegx_bundle_bits
+create_Imm8OpcodeExtension_X0(int num)
+{
+  const unsigned int n = (unsigned int)num;
+  return ((n & 0xff) << 20);
+}
+
+static __inline tilegx_bundle_bits
+create_Imm8OpcodeExtension_X1(int num)
+{
+  const unsigned int n = (unsigned int)num;
+  return (((tilegx_bundle_bits)(n & 0xff)) << 51);
+}
+
+static __inline tilegx_bundle_bits
+create_Imm8_X0(int num)
+{
+  const unsigned int n = (unsigned int)num;
+  return ((n & 0xff) << 12);
+}
+
+static __inline tilegx_bundle_bits
+create_Imm8_X1(int num)
+{
+  const unsigned int n = (unsigned int)num;
+  return (((tilegx_bundle_bits)(n & 0xff)) << 43);
+}
+
+static __inline tilegx_bundle_bits
+create_Imm8_Y0(int num)
+{
+  const unsigned int n = (unsigned int)num;
+  return ((n & 0xff) << 12);
+}
+
+static __inline tilegx_bundle_bits
+create_Imm8_Y1(int num)
+{
+  const unsigned int n = (unsigned int)num;
+  return (((tilegx_bundle_bits)(n & 0xff)) << 43);
+}
+
+static __inline tilegx_bundle_bits
+create_JumpOff_X1(int num)
+{
+  const unsigned int n = (unsigned int)num;
+  return (((tilegx_bundle_bits)(n & 0x7ffffff)) << 31);
+}
+
+static __inline tilegx_bundle_bits
+create_JumpOpcodeExtension_X1(int num)
+{
+  const unsigned int n = (unsigned int)num;
+  return (((tilegx_bundle_bits)(n & 0x1)) << 58);
+}
+
+static __inline tilegx_bundle_bits
+create_MF_Imm14_X1(int num)
+{
+  const unsigned int n = (unsigned int)num;
+  return (((tilegx_bundle_bits)(n & 0x3fff)) << 37);
+}
+
+static __inline tilegx_bundle_bits
+create_MT_Imm14_X1(int num)
+{
+  const unsigned int n = (unsigned int)num;
+  return (((tilegx_bundle_bits)(n & 0x0000003f)) << 31) |
+         (((tilegx_bundle_bits)(n & 0x00003fc0)) << 37);
+}
+
+static __inline tilegx_bundle_bits
+create_Mode(int num)
+{
+  const unsigned int n = (unsigned int)num;
+  return (((tilegx_bundle_bits)(n & 0x3)) << 62);
+}
+
+static __inline tilegx_bundle_bits
+create_Opcode_X0(int num)
+{
+  const unsigned int n = (unsigned int)num;
+  return ((n & 0x7) << 28);
+}
+
+static __inline tilegx_bundle_bits
+create_Opcode_X1(int num)
+{
+  const unsigned int n = (unsigned int)num;
+  return (((tilegx_bundle_bits)(n & 0x7)) << 59);
+}
+
+static __inline tilegx_bundle_bits
+create_Opcode_Y0(int num)
+{
+  const unsigned int n = (unsigned int)num;
+  return ((n & 0xf) << 27);
+}
+
+static __inline tilegx_bundle_bits
+create_Opcode_Y1(int num)
+{
+  const unsigned int n = (unsigned int)num;
+  return (((tilegx_bundle_bits)(n & 0xf)) << 58);
+}
+
+static __inline tilegx_bundle_bits
+create_Opcode_Y2(int num)
+{
+  const unsigned int n = (unsigned int)num;
+  return ((n & 0x00000001) << 26) |
+         (((tilegx_bundle_bits)(n & 0x00000002)) << 56);
+}
+
+static __inline tilegx_bundle_bits
+create_RRROpcodeExtension_X0(int num)
+{
+  const unsigned int n = (unsigned int)num;
+  return ((n & 0x3ff) << 18);
+}
+
+static __inline tilegx_bundle_bits
+create_RRROpcodeExtension_X1(int num)
+{
+  const unsigned int n = (unsigned int)num;
+  return (((tilegx_bundle_bits)(n & 0x3ff)) << 49);
+}
+
+static __inline tilegx_bundle_bits
+create_RRROpcodeExtension_Y0(int num)
+{
+  const unsigned int n = (unsigned int)num;
+  return ((n & 0x3) << 18);
+}
+
+static __inline tilegx_bundle_bits
+create_RRROpcodeExtension_Y1(int num)
+{
+  const unsigned int n = (unsigned int)num;
+  return (((tilegx_bundle_bits)(n & 0x3)) << 49);
+}
+
+static __inline tilegx_bundle_bits
+create_ShAmt_X0(int num)
+{
+  const unsigned int n = (unsigned int)num;
+  return ((n & 0x3f) << 12);
+}
+
+static __inline tilegx_bundle_bits
+create_ShAmt_X1(int num)
+{
+  const unsigned int n = (unsigned int)num;
+  return (((tilegx_bundle_bits)(n & 0x3f)) << 43);
+}
+
+static __inline tilegx_bundle_bits
+create_ShAmt_Y0(int num)
+{
+  const unsigned int n = (unsigned int)num;
+  return ((n & 0x3f) << 12);
+}
+
+static __inline tilegx_bundle_bits
+create_ShAmt_Y1(int num)
+{
+  const unsigned int n = (unsigned int)num;
+  return (((tilegx_bundle_bits)(n & 0x3f)) << 43);
+}
+
+static __inline tilegx_bundle_bits
+create_ShiftOpcodeExtension_X0(int num)
+{
+  const unsigned int n = (unsigned int)num;
+  return ((n & 0x3ff) << 18);
+}
+
+static __inline tilegx_bundle_bits
+create_ShiftOpcodeExtension_X1(int num)
+{
+  const unsigned int n = (unsigned int)num;
+  return (((tilegx_bundle_bits)(n & 0x3ff)) << 49);
+}
+
+static __inline tilegx_bundle_bits
+create_ShiftOpcodeExtension_Y0(int num)
+{
+  const unsigned int n = (unsigned int)num;
+  return ((n & 0x3) << 18);
+}
+
+static __inline tilegx_bundle_bits
+create_ShiftOpcodeExtension_Y1(int num)
+{
+  const unsigned int n = (unsigned int)num;
+  return (((tilegx_bundle_bits)(n & 0x3)) << 49);
+}
+
+static __inline tilegx_bundle_bits
+create_SrcA_X0(int num)
+{
+  const unsigned int n = (unsigned int)num;
+  return ((n & 0x3f) << 6);
+}
+
+static __inline tilegx_bundle_bits
+create_SrcA_X1(int num)
+{
+  const unsigned int n = (unsigned int)num;
+  return (((tilegx_bundle_bits)(n & 0x3f)) << 37);
+}
+
+static __inline tilegx_bundle_bits
+create_SrcA_Y0(int num)
+{
+  const unsigned int n = (unsigned int)num;
+  return ((n & 0x3f) << 6);
+}
+
+static __inline tilegx_bundle_bits
+create_SrcA_Y1(int num)
+{
+  const unsigned int n = (unsigned int)num;
+  return (((tilegx_bundle_bits)(n & 0x3f)) << 37);
+}
+
+static __inline tilegx_bundle_bits
+create_SrcA_Y2(int num)
+{
+  const unsigned int n = (unsigned int)num;
+  return ((n & 0x3f) << 20);
+}
+
+static __inline tilegx_bundle_bits
+create_SrcBDest_Y2(int num)
+{
+  const unsigned int n = (unsigned int)num;
+  return (((tilegx_bundle_bits)(n & 0x3f)) << 51);
+}
+
+static __inline tilegx_bundle_bits
+create_SrcB_X0(int num)
+{
+  const unsigned int n = (unsigned int)num;
+  return ((n & 0x3f) << 12);
+}
+
+static __inline tilegx_bundle_bits
+create_SrcB_X1(int num)
+{
+  const unsigned int n = (unsigned int)num;
+  return (((tilegx_bundle_bits)(n & 0x3f)) << 43);
+}
+
+static __inline tilegx_bundle_bits
+create_SrcB_Y0(int num)
+{
+  const unsigned int n = (unsigned int)num;
+  return ((n & 0x3f) << 12);
+}
+
+static __inline tilegx_bundle_bits
+create_SrcB_Y1(int num)
+{
+  const unsigned int n = (unsigned int)num;
+  return (((tilegx_bundle_bits)(n & 0x3f)) << 43);
+}
+
+static __inline tilegx_bundle_bits
+create_UnaryOpcodeExtension_X0(int num)
+{
+  const unsigned int n = (unsigned int)num;
+  return ((n & 0x3f) << 12);
+}
+
+static __inline tilegx_bundle_bits
+create_UnaryOpcodeExtension_X1(int num)
+{
+  const unsigned int n = (unsigned int)num;
+  return (((tilegx_bundle_bits)(n & 0x3f)) << 43);
+}
+
+static __inline tilegx_bundle_bits
+create_UnaryOpcodeExtension_Y0(int num)
+{
+  const unsigned int n = (unsigned int)num;
+  return ((n & 0x3f) << 12);
+}
+
+static __inline tilegx_bundle_bits
+create_UnaryOpcodeExtension_Y1(int num)
+{
+  const unsigned int n = (unsigned int)num;
+  return (((tilegx_bundle_bits)(n & 0x3f)) << 43);
+}
+
+
+enum
+{
+  ADDI_IMM8_OPCODE_X0 = 1,
+  ADDI_IMM8_OPCODE_X1 = 1,
+  ADDI_OPCODE_Y0 = 0,
+  ADDI_OPCODE_Y1 = 1,
+  ADDLI_OPCODE_X0 = 1,
+  ADDLI_OPCODE_X1 = 0,
+  ADDXI_IMM8_OPCODE_X0 = 2,
+  ADDXI_IMM8_OPCODE_X1 = 2,
+  ADDXI_OPCODE_Y0 = 1,
+  ADDXI_OPCODE_Y1 = 2,
+  ADDXLI_OPCODE_X0 = 2,
+  ADDXLI_OPCODE_X1 = 1,
+  ADDXSC_RRR_0_OPCODE_X0 = 1,
+  ADDXSC_RRR_0_OPCODE_X1 = 1,
+  ADDX_RRR_0_OPCODE_X0 = 2,
+  ADDX_RRR_0_OPCODE_X1 = 2,
+  ADDX_RRR_0_OPCODE_Y0 = 0,
+  ADDX_SPECIAL_0_OPCODE_Y1 = 0,
+  ADD_RRR_0_OPCODE_X0 = 3,
+  ADD_RRR_0_OPCODE_X1 = 3,
+  ADD_RRR_0_OPCODE_Y0 = 1,
+  ADD_SPECIAL_0_OPCODE_Y1 = 1,
+  ANDI_IMM8_OPCODE_X0 = 3,
+  ANDI_IMM8_OPCODE_X1 = 3,
+  ANDI_OPCODE_Y0 = 2,
+  ANDI_OPCODE_Y1 = 3,
+  AND_RRR_0_OPCODE_X0 = 4,
+  AND_RRR_0_OPCODE_X1 = 4,
+  AND_RRR_5_OPCODE_Y0 = 0,
+  AND_RRR_5_OPCODE_Y1 = 0,
+  BEQZT_BRANCH_OPCODE_X1 = 16,
+  BEQZ_BRANCH_OPCODE_X1 = 17,
+  BFEXTS_BF_OPCODE_X0 = 4,
+  BFEXTU_BF_OPCODE_X0 = 5,
+  BFINS_BF_OPCODE_X0 = 6,
+  BF_OPCODE_X0 = 3,
+  BGEZT_BRANCH_OPCODE_X1 = 18,
+  BGEZ_BRANCH_OPCODE_X1 = 19,
+  BGTZT_BRANCH_OPCODE_X1 = 20,
+  BGTZ_BRANCH_OPCODE_X1 = 21,
+  BLBCT_BRANCH_OPCODE_X1 = 22,
+  BLBC_BRANCH_OPCODE_X1 = 23,
+  BLBST_BRANCH_OPCODE_X1 = 24,
+  BLBS_BRANCH_OPCODE_X1 = 25,
+  BLEZT_BRANCH_OPCODE_X1 = 26,
+  BLEZ_BRANCH_OPCODE_X1 = 27,
+  BLTZT_BRANCH_OPCODE_X1 = 28,
+  BLTZ_BRANCH_OPCODE_X1 = 29,
+  BNEZT_BRANCH_OPCODE_X1 = 30,
+  BNEZ_BRANCH_OPCODE_X1 = 31,
+  BRANCH_OPCODE_X1 = 2,
+  CMOVEQZ_RRR_0_OPCODE_X0 = 5,
+  CMOVEQZ_RRR_4_OPCODE_Y0 = 0,
+  CMOVNEZ_RRR_0_OPCODE_X0 = 6,
+  CMOVNEZ_RRR_4_OPCODE_Y0 = 1,
+  CMPEQI_IMM8_OPCODE_X0 = 4,
+  CMPEQI_IMM8_OPCODE_X1 = 4,
+  CMPEQI_OPCODE_Y0 = 3,
+  CMPEQI_OPCODE_Y1 = 4,
+  CMPEQ_RRR_0_OPCODE_X0 = 7,
+  CMPEQ_RRR_0_OPCODE_X1 = 5,
+  CMPEQ_RRR_3_OPCODE_Y0 = 0,
+  CMPEQ_RRR_3_OPCODE_Y1 = 2,
+  CMPEXCH4_RRR_0_OPCODE_X1 = 6,
+  CMPEXCH_RRR_0_OPCODE_X1 = 7,
+  CMPLES_RRR_0_OPCODE_X0 = 8,
+  CMPLES_RRR_0_OPCODE_X1 = 8,
+  CMPLES_RRR_2_OPCODE_Y0 = 0,
+  CMPLES_RRR_2_OPCODE_Y1 = 0,
+  CMPLEU_RRR_0_OPCODE_X0 = 9,
+  CMPLEU_RRR_0_OPCODE_X1 = 9,
+  CMPLEU_RRR_2_OPCODE_Y0 = 1,
+  CMPLEU_RRR_2_OPCODE_Y1 = 1,
+  CMPLTSI_IMM8_OPCODE_X0 = 5,
+  CMPLTSI_IMM8_OPCODE_X1 = 5,
+  CMPLTSI_OPCODE_Y0 = 4,
+  CMPLTSI_OPCODE_Y1 = 5,
+  CMPLTS_RRR_0_OPCODE_X0 = 10,
+  CMPLTS_RRR_0_OPCODE_X1 = 10,
+  CMPLTS_RRR_2_OPCODE_Y0 = 2,
+  CMPLTS_RRR_2_OPCODE_Y1 = 2,
+  CMPLTUI_IMM8_OPCODE_X0 = 6,
+  CMPLTUI_IMM8_OPCODE_X1 = 6,
+  CMPLTU_RRR_0_OPCODE_X0 = 11,
+  CMPLTU_RRR_0_OPCODE_X1 = 11,
+  CMPLTU_RRR_2_OPCODE_Y0 = 3,
+  CMPLTU_RRR_2_OPCODE_Y1 = 3,
+  CMPNE_RRR_0_OPCODE_X0 = 12,
+  CMPNE_RRR_0_OPCODE_X1 = 12,
+  CMPNE_RRR_3_OPCODE_Y0 = 1,
+  CMPNE_RRR_3_OPCODE_Y1 = 3,
+  CMULAF_RRR_0_OPCODE_X0 = 13,
+  CMULA_RRR_0_OPCODE_X0 = 14,
+  CMULFR_RRR_0_OPCODE_X0 = 15,
+  CMULF_RRR_0_OPCODE_X0 = 16,
+  CMULHR_RRR_0_OPCODE_X0 = 17,
+  CMULH_RRR_0_OPCODE_X0 = 18,
+  CMUL_RRR_0_OPCODE_X0 = 19,
+  CNTLZ_UNARY_OPCODE_X0 = 1,
+  CNTLZ_UNARY_OPCODE_Y0 = 1,
+  CNTTZ_UNARY_OPCODE_X0 = 2,
+  CNTTZ_UNARY_OPCODE_Y0 = 2,
+  CRC32_32_RRR_0_OPCODE_X0 = 20,
+  CRC32_8_RRR_0_OPCODE_X0 = 21,
+  DBLALIGN2_RRR_0_OPCODE_X0 = 22,
+  DBLALIGN2_RRR_0_OPCODE_X1 = 13,
+  DBLALIGN4_RRR_0_OPCODE_X0 = 23,
+  DBLALIGN4_RRR_0_OPCODE_X1 = 14,
+  DBLALIGN6_RRR_0_OPCODE_X0 = 24,
+  DBLALIGN6_RRR_0_OPCODE_X1 = 15,
+  DBLALIGN_RRR_0_OPCODE_X0 = 25,
+  DRAIN_UNARY_OPCODE_X1 = 1,
+  DTLBPR_UNARY_OPCODE_X1 = 2,
+  EXCH4_RRR_0_OPCODE_X1 = 16,
+  EXCH_RRR_0_OPCODE_X1 = 17,
+  FDOUBLE_ADDSUB_RRR_0_OPCODE_X0 = 26,
+  FDOUBLE_ADD_FLAGS_RRR_0_OPCODE_X0 = 27,
+  FDOUBLE_MUL_FLAGS_RRR_0_OPCODE_X0 = 28,
+  FDOUBLE_PACK1_RRR_0_OPCODE_X0 = 29,
+  FDOUBLE_PACK2_RRR_0_OPCODE_X0 = 30,
+  FDOUBLE_SUB_FLAGS_RRR_0_OPCODE_X0 = 31,
+  FDOUBLE_UNPACK_MAX_RRR_0_OPCODE_X0 = 32,
+  FDOUBLE_UNPACK_MIN_RRR_0_OPCODE_X0 = 33,
+  FETCHADD4_RRR_0_OPCODE_X1 = 18,
+  FETCHADDGEZ4_RRR_0_OPCODE_X1 = 19,
+  FETCHADDGEZ_RRR_0_OPCODE_X1 = 20,
+  FETCHADD_RRR_0_OPCODE_X1 = 21,
+  FETCHAND4_RRR_0_OPCODE_X1 = 22,
+  FETCHAND_RRR_0_OPCODE_X1 = 23,
+  FETCHOR4_RRR_0_OPCODE_X1 = 24,
+  FETCHOR_RRR_0_OPCODE_X1 = 25,
+  FINV_UNARY_OPCODE_X1 = 3,
+  FLUSHWB_UNARY_OPCODE_X1 = 4,
+  FLUSH_UNARY_OPCODE_X1 = 5,
+  FNOP_UNARY_OPCODE_X0 = 3,
+  FNOP_UNARY_OPCODE_X1 = 6,
+  FNOP_UNARY_OPCODE_Y0 = 3,
+  FNOP_UNARY_OPCODE_Y1 = 8,
+  FSINGLE_ADD1_RRR_0_OPCODE_X0 = 34,
+  FSINGLE_ADDSUB2_RRR_0_OPCODE_X0 = 35,
+  FSINGLE_MUL1_RRR_0_OPCODE_X0 = 36,
+  FSINGLE_MUL2_RRR_0_OPCODE_X0 = 37,
+  FSINGLE_PACK1_UNARY_OPCODE_X0 = 4,
+  FSINGLE_PACK1_UNARY_OPCODE_Y0 = 4,
+  FSINGLE_PACK2_RRR_0_OPCODE_X0 = 38,
+  FSINGLE_SUB1_RRR_0_OPCODE_X0 = 39,
+  ICOH_UNARY_OPCODE_X1 = 7,
+  ILL_UNARY_OPCODE_X1 = 8,
+  ILL_UNARY_OPCODE_Y1 = 9,
+  IMM8_OPCODE_X0 = 4,
+  IMM8_OPCODE_X1 = 3,
+  INV_UNARY_OPCODE_X1 = 9,
+  IRET_UNARY_OPCODE_X1 = 10,
+  JALRP_UNARY_OPCODE_X1 = 11,
+  JALRP_UNARY_OPCODE_Y1 = 10,
+  JALR_UNARY_OPCODE_X1 = 12,
+  JALR_UNARY_OPCODE_Y1 = 11,
+  JAL_JUMP_OPCODE_X1 = 0,
+  JRP_UNARY_OPCODE_X1 = 13,
+  JRP_UNARY_OPCODE_Y1 = 12,
+  JR_UNARY_OPCODE_X1 = 14,
+  JR_UNARY_OPCODE_Y1 = 13,
+  JUMP_OPCODE_X1 = 4,
+  J_JUMP_OPCODE_X1 = 1,
+  LD1S_ADD_IMM8_OPCODE_X1 = 7,
+  LD1S_OPCODE_Y2 = 0,
+  LD1S_UNARY_OPCODE_X1 = 15,
+  LD1U_ADD_IMM8_OPCODE_X1 = 8,
+  LD1U_OPCODE_Y2 = 1,
+  LD1U_UNARY_OPCODE_X1 = 16,
+  LD2S_ADD_IMM8_OPCODE_X1 = 9,
+  LD2S_OPCODE_Y2 = 2,
+  LD2S_UNARY_OPCODE_X1 = 17,
+  LD2U_ADD_IMM8_OPCODE_X1 = 10,
+  LD2U_OPCODE_Y2 = 3,
+  LD2U_UNARY_OPCODE_X1 = 18,
+  LD4S_ADD_IMM8_OPCODE_X1 = 11,
+  LD4S_OPCODE_Y2 = 1,
+  LD4S_UNARY_OPCODE_X1 = 19,
+  LD4U_ADD_IMM8_OPCODE_X1 = 12,
+  LD4U_OPCODE_Y2 = 2,
+  LD4U_UNARY_OPCODE_X1 = 20,
+  LDNA_UNARY_OPCODE_X1 = 21,
+  LDNT1S_ADD_IMM8_OPCODE_X1 = 13,
+  LDNT1S_UNARY_OPCODE_X1 = 22,
+  LDNT1U_ADD_IMM8_OPCODE_X1 = 14,
+  LDNT1U_UNARY_OPCODE_X1 = 23,
+  LDNT2S_ADD_IMM8_OPCODE_X1 = 15,
+  LDNT2S_UNARY_OPCODE_X1 = 24,
+  LDNT2U_ADD_IMM8_OPCODE_X1 = 16,
+  LDNT2U_UNARY_OPCODE_X1 = 25,
+  LDNT4S_ADD_IMM8_OPCODE_X1 = 17,
+  LDNT4S_UNARY_OPCODE_X1 = 26,
+  LDNT4U_ADD_IMM8_OPCODE_X1 = 18,
+  LDNT4U_UNARY_OPCODE_X1 = 27,
+  LDNT_ADD_IMM8_OPCODE_X1 = 19,
+  LDNT_UNARY_OPCODE_X1 = 28,
+  LD_ADD_IMM8_OPCODE_X1 = 20,
+  LD_OPCODE_Y2 = 3,
+  LD_UNARY_OPCODE_X1 = 29,
+  LNK_UNARY_OPCODE_X1 = 30,
+  LNK_UNARY_OPCODE_Y1 = 14,
+  LWNA_ADD_IMM8_OPCODE_X1 = 21,
+  MFSPR_IMM8_OPCODE_X1 = 22,
+  MF_UNARY_OPCODE_X1 = 31,
+  MM_BF_OPCODE_X0 = 7,
+  MNZ_RRR_0_OPCODE_X0 = 40,
+  MNZ_RRR_0_OPCODE_X1 = 26,
+  MNZ_RRR_4_OPCODE_Y0 = 2,
+  MNZ_RRR_4_OPCODE_Y1 = 2,
+  MODE_OPCODE_YA2 = 1,
+  MODE_OPCODE_YB2 = 2,
+  MODE_OPCODE_YC2 = 3,
+  MTSPR_IMM8_OPCODE_X1 = 23,
+  MULAX_RRR_0_OPCODE_X0 = 41,
+  MULAX_RRR_3_OPCODE_Y0 = 2,
+  MULA_HS_HS_RRR_0_OPCODE_X0 = 42,
+  MULA_HS_HS_RRR_9_OPCODE_Y0 = 0,
+  MULA_HS_HU_RRR_0_OPCODE_X0 = 43,
+  MULA_HS_LS_RRR_0_OPCODE_X0 = 44,
+  MULA_HS_LU_RRR_0_OPCODE_X0 = 45,
+  MULA_HU_HU_RRR_0_OPCODE_X0 = 46,
+  MULA_HU_HU_RRR_9_OPCODE_Y0 = 1,
+  MULA_HU_LS_RRR_0_OPCODE_X0 = 47,
+  MULA_HU_LU_RRR_0_OPCODE_X0 = 48,
+  MULA_LS_LS_RRR_0_OPCODE_X0 = 49,
+  MULA_LS_LS_RRR_9_OPCODE_Y0 = 2,
+  MULA_LS_LU_RRR_0_OPCODE_X0 = 50,
+  MULA_LU_LU_RRR_0_OPCODE_X0 = 51,
+  MULA_LU_LU_RRR_9_OPCODE_Y0 = 3,
+  MULX_RRR_0_OPCODE_X0 = 52,
+  MULX_RRR_3_OPCODE_Y0 = 3,
+  MUL_HS_HS_RRR_0_OPCODE_X0 = 53,
+  MUL_HS_HS_RRR_8_OPCODE_Y0 = 0,
+  MUL_HS_HU_RRR_0_OPCODE_X0 = 54,
+  MUL_HS_LS_RRR_0_OPCODE_X0 = 55,
+  MUL_HS_LU_RRR_0_OPCODE_X0 = 56,
+  MUL_HU_HU_RRR_0_OPCODE_X0 = 57,
+  MUL_HU_HU_RRR_8_OPCODE_Y0 = 1,
+  MUL_HU_LS_RRR_0_OPCODE_X0 = 58,
+  MUL_HU_LU_RRR_0_OPCODE_X0 = 59,
+  MUL_LS_LS_RRR_0_OPCODE_X0 = 60,
+  MUL_LS_LS_RRR_8_OPCODE_Y0 = 2,
+  MUL_LS_LU_RRR_0_OPCODE_X0 = 61,
+  MUL_LU_LU_RRR_0_OPCODE_X0 = 62,
+  MUL_LU_LU_RRR_8_OPCODE_Y0 = 3,
+  MZ_RRR_0_OPCODE_X0 = 63,
+  MZ_RRR_0_OPCODE_X1 = 27,
+  MZ_RRR_4_OPCODE_Y0 = 3,
+  MZ_RRR_4_OPCODE_Y1 = 3,
+  NAP_UNARY_OPCODE_X1 = 32,
+  NOP_UNARY_OPCODE_X0 = 5,
+  NOP_UNARY_OPCODE_X1 = 33,
+  NOP_UNARY_OPCODE_Y0 = 5,
+  NOP_UNARY_OPCODE_Y1 = 15,
+  NOR_RRR_0_OPCODE_X0 = 64,
+  NOR_RRR_0_OPCODE_X1 = 28,
+  NOR_RRR_5_OPCODE_Y0 = 1,
+  NOR_RRR_5_OPCODE_Y1 = 1,
+  ORI_IMM8_OPCODE_X0 = 7,
+  ORI_IMM8_OPCODE_X1 = 24,
+  OR_RRR_0_OPCODE_X0 = 65,
+  OR_RRR_0_OPCODE_X1 = 29,
+  OR_RRR_5_OPCODE_Y0 = 2,
+  OR_RRR_5_OPCODE_Y1 = 2,
+  PCNT_UNARY_OPCODE_X0 = 6,
+  PCNT_UNARY_OPCODE_Y0 = 6,
+  REVBITS_UNARY_OPCODE_X0 = 7,
+  REVBITS_UNARY_OPCODE_Y0 = 7,
+  REVBYTES_UNARY_OPCODE_X0 = 8,
+  REVBYTES_UNARY_OPCODE_Y0 = 8,
+  ROTLI_SHIFT_OPCODE_X0 = 1,
+  ROTLI_SHIFT_OPCODE_X1 = 1,
+  ROTLI_SHIFT_OPCODE_Y0 = 0,
+  ROTLI_SHIFT_OPCODE_Y1 = 0,
+  ROTL_RRR_0_OPCODE_X0 = 66,
+  ROTL_RRR_0_OPCODE_X1 = 30,
+  ROTL_RRR_6_OPCODE_Y0 = 0,
+  ROTL_RRR_6_OPCODE_Y1 = 0,
+  RRR_0_OPCODE_X0 = 5,
+  RRR_0_OPCODE_X1 = 5,
+  RRR_0_OPCODE_Y0 = 5,
+  RRR_0_OPCODE_Y1 = 6,
+  RRR_1_OPCODE_Y0 = 6,
+  RRR_1_OPCODE_Y1 = 7,
+  RRR_2_OPCODE_Y0 = 7,
+  RRR_2_OPCODE_Y1 = 8,
+  RRR_3_OPCODE_Y0 = 8,
+  RRR_3_OPCODE_Y1 = 9,
+  RRR_4_OPCODE_Y0 = 9,
+  RRR_4_OPCODE_Y1 = 10,
+  RRR_5_OPCODE_Y0 = 10,
+  RRR_5_OPCODE_Y1 = 11,
+  RRR_6_OPCODE_Y0 = 11,
+  RRR_6_OPCODE_Y1 = 12,
+  RRR_7_OPCODE_Y0 = 12,
+  RRR_7_OPCODE_Y1 = 13,
+  RRR_8_OPCODE_Y0 = 13,
+  RRR_9_OPCODE_Y0 = 14,
+  SHIFT_OPCODE_X0 = 6,
+  SHIFT_OPCODE_X1 = 6,
+  SHIFT_OPCODE_Y0 = 15,
+  SHIFT_OPCODE_Y1 = 14,
+  SHL16INSLI_OPCODE_X0 = 7,
+  SHL16INSLI_OPCODE_X1 = 7,
+  SHL1ADDX_RRR_0_OPCODE_X0 = 67,
+  SHL1ADDX_RRR_0_OPCODE_X1 = 31,
+  SHL1ADDX_RRR_7_OPCODE_Y0 = 1,
+  SHL1ADDX_RRR_7_OPCODE_Y1 = 1,
+  SHL1ADD_RRR_0_OPCODE_X0 = 68,
+  SHL1ADD_RRR_0_OPCODE_X1 = 32,
+  SHL1ADD_RRR_1_OPCODE_Y0 = 0,
+  SHL1ADD_RRR_1_OPCODE_Y1 = 0,
+  SHL2ADDX_RRR_0_OPCODE_X0 = 69,
+  SHL2ADDX_RRR_0_OPCODE_X1 = 33,
+  SHL2ADDX_RRR_7_OPCODE_Y0 = 2,
+  SHL2ADDX_RRR_7_OPCODE_Y1 = 2,
+  SHL2ADD_RRR_0_OPCODE_X0 = 70,
+  SHL2ADD_RRR_0_OPCODE_X1 = 34,
+  SHL2ADD_RRR_1_OPCODE_Y0 = 1,
+  SHL2ADD_RRR_1_OPCODE_Y1 = 1,
+  SHL3ADDX_RRR_0_OPCODE_X0 = 71,
+  SHL3ADDX_RRR_0_OPCODE_X1 = 35,
+  SHL3ADDX_RRR_7_OPCODE_Y0 = 3,
+  SHL3ADDX_RRR_7_OPCODE_Y1 = 3,
+  SHL3ADD_RRR_0_OPCODE_X0 = 72,
+  SHL3ADD_RRR_0_OPCODE_X1 = 36,
+  SHL3ADD_RRR_1_OPCODE_Y0 = 2,
+  SHL3ADD_RRR_1_OPCODE_Y1 = 2,
+  SHLI_SHIFT_OPCODE_X0 = 2,
+  SHLI_SHIFT_OPCODE_X1 = 2,
+  SHLI_SHIFT_OPCODE_Y0 = 1,
+  SHLI_SHIFT_OPCODE_Y1 = 1,
+  SHLXI_SHIFT_OPCODE_X0 = 3,
+  SHLXI_SHIFT_OPCODE_X1 = 3,
+  SHLX_RRR_0_OPCODE_X0 = 73,
+  SHLX_RRR_0_OPCODE_X1 = 37,
+  SHL_RRR_0_OPCODE_X0 = 74,
+  SHL_RRR_0_OPCODE_X1 = 38,
+  SHL_RRR_6_OPCODE_Y0 = 1,
+  SHL_RRR_6_OPCODE_Y1 = 1,
+  SHRSI_SHIFT_OPCODE_X0 = 4,
+  SHRSI_SHIFT_OPCODE_X1 = 4,
+  SHRSI_SHIFT_OPCODE_Y0 = 2,
+  SHRSI_SHIFT_OPCODE_Y1 = 2,
+  SHRS_RRR_0_OPCODE_X0 = 75,
+  SHRS_RRR_0_OPCODE_X1 = 39,
+  SHRS_RRR_6_OPCODE_Y0 = 2,
+  SHRS_RRR_6_OPCODE_Y1 = 2,
+  SHRUI_SHIFT_OPCODE_X0 = 5,
+  SHRUI_SHIFT_OPCODE_X1 = 5,
+  SHRUI_SHIFT_OPCODE_Y0 = 3,
+  SHRUI_SHIFT_OPCODE_Y1 = 3,
+  SHRUXI_SHIFT_OPCODE_X0 = 6,
+  SHRUXI_SHIFT_OPCODE_X1 = 6,
+  SHRUX_RRR_0_OPCODE_X0 = 76,
+  SHRUX_RRR_0_OPCODE_X1 = 40,
+  SHRU_RRR_0_OPCODE_X0 = 77,
+  SHRU_RRR_0_OPCODE_X1 = 41,
+  SHRU_RRR_6_OPCODE_Y0 = 3,
+  SHRU_RRR_6_OPCODE_Y1 = 3,
+  SHUFFLEBYTES_RRR_0_OPCODE_X0 = 78,
+  ST1_ADD_IMM8_OPCODE_X1 = 25,
+  ST1_OPCODE_Y2 = 0,
+  ST1_RRR_0_OPCODE_X1 = 42,
+  ST2_ADD_IMM8_OPCODE_X1 = 26,
+  ST2_OPCODE_Y2 = 1,
+  ST2_RRR_0_OPCODE_X1 = 43,
+  ST4_ADD_IMM8_OPCODE_X1 = 27,
+  ST4_OPCODE_Y2 = 2,
+  ST4_RRR_0_OPCODE_X1 = 44,
+  STNT1_ADD_IMM8_OPCODE_X1 = 28,
+  STNT1_RRR_0_OPCODE_X1 = 45,
+  STNT2_ADD_IMM8_OPCODE_X1 = 29,
+  STNT2_RRR_0_OPCODE_X1 = 46,
+  STNT4_ADD_IMM8_OPCODE_X1 = 30,
+  STNT4_RRR_0_OPCODE_X1 = 47,
+  STNT_ADD_IMM8_OPCODE_X1 = 31,
+  STNT_RRR_0_OPCODE_X1 = 48,
+  ST_ADD_IMM8_OPCODE_X1 = 32,
+  ST_OPCODE_Y2 = 3,
+  ST_RRR_0_OPCODE_X1 = 49,
+  SUBXSC_RRR_0_OPCODE_X0 = 79,
+  SUBXSC_RRR_0_OPCODE_X1 = 50,
+  SUBX_RRR_0_OPCODE_X0 = 80,
+  SUBX_RRR_0_OPCODE_X1 = 51,
+  SUBX_RRR_0_OPCODE_Y0 = 2,
+  SUBX_RRR_0_OPCODE_Y1 = 2,
+  SUB_RRR_0_OPCODE_X0 = 81,
+  SUB_RRR_0_OPCODE_X1 = 52,
+  SUB_RRR_0_OPCODE_Y0 = 3,
+  SUB_RRR_0_OPCODE_Y1 = 3,
+  SWINT0_UNARY_OPCODE_X1 = 34,
+  SWINT1_UNARY_OPCODE_X1 = 35,
+  SWINT2_UNARY_OPCODE_X1 = 36,
+  SWINT3_UNARY_OPCODE_X1 = 37,
+  TBLIDXB0_UNARY_OPCODE_X0 = 9,
+  TBLIDXB0_UNARY_OPCODE_Y0 = 9,
+  TBLIDXB1_UNARY_OPCODE_X0 = 10,
+  TBLIDXB1_UNARY_OPCODE_Y0 = 10,
+  TBLIDXB2_UNARY_OPCODE_X0 = 11,
+  TBLIDXB2_UNARY_OPCODE_Y0 = 11,
+  TBLIDXB3_UNARY_OPCODE_X0 = 12,
+  TBLIDXB3_UNARY_OPCODE_Y0 = 12,
+  UNARY_RRR_0_OPCODE_X0 = 82,
+  UNARY_RRR_0_OPCODE_X1 = 53,
+  UNARY_RRR_1_OPCODE_Y0 = 3,
+  UNARY_RRR_1_OPCODE_Y1 = 3,
+  V1ADDI_IMM8_OPCODE_X0 = 8,
+  V1ADDI_IMM8_OPCODE_X1 = 33,
+  V1ADDUC_RRR_0_OPCODE_X0 = 83,
+  V1ADDUC_RRR_0_OPCODE_X1 = 54,
+  V1ADD_RRR_0_OPCODE_X0 = 84,
+  V1ADD_RRR_0_OPCODE_X1 = 55,
+  V1ADIFFU_RRR_0_OPCODE_X0 = 85,
+  V1AVGU_RRR_0_OPCODE_X0 = 86,
+  V1CMPEQI_IMM8_OPCODE_X0 = 9,
+  V1CMPEQI_IMM8_OPCODE_X1 = 34,
+  V1CMPEQ_RRR_0_OPCODE_X0 = 87,
+  V1CMPEQ_RRR_0_OPCODE_X1 = 56,
+  V1CMPLES_RRR_0_OPCODE_X0 = 88,
+  V1CMPLES_RRR_0_OPCODE_X1 = 57,
+  V1CMPLEU_RRR_0_OPCODE_X0 = 89,
+  V1CMPLEU_RRR_0_OPCODE_X1 = 58,
+  V1CMPLTSI_IMM8_OPCODE_X0 = 10,
+  V1CMPLTSI_IMM8_OPCODE_X1 = 35,
+  V1CMPLTS_RRR_0_OPCODE_X0 = 90,
+  V1CMPLTS_RRR_0_OPCODE_X1 = 59,
+  V1CMPLTUI_IMM8_OPCODE_X0 = 11,
+  V1CMPLTUI_IMM8_OPCODE_X1 = 36,
+  V1CMPLTU_RRR_0_OPCODE_X0 = 91,
+  V1CMPLTU_RRR_0_OPCODE_X1 = 60,
+  V1CMPNE_RRR_0_OPCODE_X0 = 92,
+  V1CMPNE_RRR_0_OPCODE_X1 = 61,
+  V1DDOTPUA_RRR_0_OPCODE_X0 = 161,
+  V1DDOTPUSA_RRR_0_OPCODE_X0 = 93,
+  V1DDOTPUS_RRR_0_OPCODE_X0 = 94,
+  V1DDOTPU_RRR_0_OPCODE_X0 = 162,
+  V1DOTPA_RRR_0_OPCODE_X0 = 95,
+  V1DOTPUA_RRR_0_OPCODE_X0 = 163,
+  V1DOTPUSA_RRR_0_OPCODE_X0 = 96,
+  V1DOTPUS_RRR_0_OPCODE_X0 = 97,
+  V1DOTPU_RRR_0_OPCODE_X0 = 164,
+  V1DOTP_RRR_0_OPCODE_X0 = 98,
+  V1INT_H_RRR_0_OPCODE_X0 = 99,
+  V1INT_H_RRR_0_OPCODE_X1 = 62,
+  V1INT_L_RRR_0_OPCODE_X0 = 100,
+  V1INT_L_RRR_0_OPCODE_X1 = 63,
+  V1MAXUI_IMM8_OPCODE_X0 = 12,
+  V1MAXUI_IMM8_OPCODE_X1 = 37,
+  V1MAXU_RRR_0_OPCODE_X0 = 101,
+  V1MAXU_RRR_0_OPCODE_X1 = 64,
+  V1MINUI_IMM8_OPCODE_X0 = 13,
+  V1MINUI_IMM8_OPCODE_X1 = 38,
+  V1MINU_RRR_0_OPCODE_X0 = 102,
+  V1MINU_RRR_0_OPCODE_X1 = 65,
+  V1MNZ_RRR_0_OPCODE_X0 = 103,
+  V1MNZ_RRR_0_OPCODE_X1 = 66,
+  V1MULTU_RRR_0_OPCODE_X0 = 104,
+  V1MULUS_RRR_0_OPCODE_X0 = 105,
+  V1MULU_RRR_0_OPCODE_X0 = 106,
+  V1MZ_RRR_0_OPCODE_X0 = 107,
+  V1MZ_RRR_0_OPCODE_X1 = 67,
+  V1SADAU_RRR_0_OPCODE_X0 = 108,
+  V1SADU_RRR_0_OPCODE_X0 = 109,
+  V1SHLI_SHIFT_OPCODE_X0 = 7,
+  V1SHLI_SHIFT_OPCODE_X1 = 7,
+  V1SHL_RRR_0_OPCODE_X0 = 110,
+  V1SHL_RRR_0_OPCODE_X1 = 68,
+  V1SHRSI_SHIFT_OPCODE_X0 = 8,
+  V1SHRSI_SHIFT_OPCODE_X1 = 8,
+  V1SHRS_RRR_0_OPCODE_X0 = 111,
+  V1SHRS_RRR_0_OPCODE_X1 = 69,
+  V1SHRUI_SHIFT_OPCODE_X0 = 9,
+  V1SHRUI_SHIFT_OPCODE_X1 = 9,
+  V1SHRU_RRR_0_OPCODE_X0 = 112,
+  V1SHRU_RRR_0_OPCODE_X1 = 70,
+  V1SUBUC_RRR_0_OPCODE_X0 = 113,
+  V1SUBUC_RRR_0_OPCODE_X1 = 71,
+  V1SUB_RRR_0_OPCODE_X0 = 114,
+  V1SUB_RRR_0_OPCODE_X1 = 72,
+  V2ADDI_IMM8_OPCODE_X0 = 14,
+  V2ADDI_IMM8_OPCODE_X1 = 39,
+  V2ADDSC_RRR_0_OPCODE_X0 = 115,
+  V2ADDSC_RRR_0_OPCODE_X1 = 73,
+  V2ADD_RRR_0_OPCODE_X0 = 116,
+  V2ADD_RRR_0_OPCODE_X1 = 74,
+  V2ADIFFS_RRR_0_OPCODE_X0 = 117,
+  V2AVGS_RRR_0_OPCODE_X0 = 118,
+  V2CMPEQI_IMM8_OPCODE_X0 = 15,
+  V2CMPEQI_IMM8_OPCODE_X1 = 40,
+  V2CMPEQ_RRR_0_OPCODE_X0 = 119,
+  V2CMPEQ_RRR_0_OPCODE_X1 = 75,
+  V2CMPLES_RRR_0_OPCODE_X0 = 120,
+  V2CMPLES_RRR_0_OPCODE_X1 = 76,
+  V2CMPLEU_RRR_0_OPCODE_X0 = 121,
+  V2CMPLEU_RRR_0_OPCODE_X1 = 77,
+  V2CMPLTSI_IMM8_OPCODE_X0 = 16,
+  V2CMPLTSI_IMM8_OPCODE_X1 = 41,
+  V2CMPLTS_RRR_0_OPCODE_X0 = 122,
+  V2CMPLTS_RRR_0_OPCODE_X1 = 78,
+  V2CMPLTUI_IMM8_OPCODE_X0 = 17,
+  V2CMPLTUI_IMM8_OPCODE_X1 = 42,
+  V2CMPLTU_RRR_0_OPCODE_X0 = 123,
+  V2CMPLTU_RRR_0_OPCODE_X1 = 79,
+  V2CMPNE_RRR_0_OPCODE_X0 = 124,
+  V2CMPNE_RRR_0_OPCODE_X1 = 80,
+  V2DOTPA_RRR_0_OPCODE_X0 = 125,
+  V2DOTP_RRR_0_OPCODE_X0 = 126,
+  V2INT_H_RRR_0_OPCODE_X0 = 127,
+  V2INT_H_RRR_0_OPCODE_X1 = 81,
+  V2INT_L_RRR_0_OPCODE_X0 = 128,
+  V2INT_L_RRR_0_OPCODE_X1 = 82,
+  V2MAXSI_IMM8_OPCODE_X0 = 18,
+  V2MAXSI_IMM8_OPCODE_X1 = 43,
+  V2MAXS_RRR_0_OPCODE_X0 = 129,
+  V2MAXS_RRR_0_OPCODE_X1 = 83,
+  V2MINSI_IMM8_OPCODE_X0 = 19,
+  V2MINSI_IMM8_OPCODE_X1 = 44,
+  V2MINS_RRR_0_OPCODE_X0 = 130,
+  V2MINS_RRR_0_OPCODE_X1 = 84,
+  V2MNZ_RRR_0_OPCODE_X0 = 131,
+  V2MNZ_RRR_0_OPCODE_X1 = 85,
+  V2MULFSC_RRR_0_OPCODE_X0 = 132,
+  V2MULS_RRR_0_OPCODE_X0 = 133,
+  V2MULTS_RRR_0_OPCODE_X0 = 134,
+  V2MZ_RRR_0_OPCODE_X0 = 135,
+  V2MZ_RRR_0_OPCODE_X1 = 86,
+  V2PACKH_RRR_0_OPCODE_X0 = 136,
+  V2PACKH_RRR_0_OPCODE_X1 = 87,
+  V2PACKL_RRR_0_OPCODE_X0 = 137,
+  V2PACKL_RRR_0_OPCODE_X1 = 88,
+  V2PACKUC_RRR_0_OPCODE_X0 = 138,
+  V2PACKUC_RRR_0_OPCODE_X1 = 89,
+  V2SADAS_RRR_0_OPCODE_X0 = 139,
+  V2SADAU_RRR_0_OPCODE_X0 = 140,
+  V2SADS_RRR_0_OPCODE_X0 = 141,
+  V2SADU_RRR_0_OPCODE_X0 = 142,
+  V2SHLI_SHIFT_OPCODE_X0 = 10,
+  V2SHLI_SHIFT_OPCODE_X1 = 10,
+  V2SHLSC_RRR_0_OPCODE_X0 = 143,
+  V2SHLSC_RRR_0_OPCODE_X1 = 90,
+  V2SHL_RRR_0_OPCODE_X0 = 144,
+  V2SHL_RRR_0_OPCODE_X1 = 91,
+  V2SHRSI_SHIFT_OPCODE_X0 = 11,
+  V2SHRSI_SHIFT_OPCODE_X1 = 11,
+  V2SHRS_RRR_0_OPCODE_X0 = 145,
+  V2SHRS_RRR_0_OPCODE_X1 = 92,
+  V2SHRUI_SHIFT_OPCODE_X0 = 12,
+  V2SHRUI_SHIFT_OPCODE_X1 = 12,
+  V2SHRU_RRR_0_OPCODE_X0 = 146,
+  V2SHRU_RRR_0_OPCODE_X1 = 93,
+  V2SUBSC_RRR_0_OPCODE_X0 = 147,
+  V2SUBSC_RRR_0_OPCODE_X1 = 94,
+  V2SUB_RRR_0_OPCODE_X0 = 148,
+  V2SUB_RRR_0_OPCODE_X1 = 95,
+  V4ADDSC_RRR_0_OPCODE_X0 = 149,
+  V4ADDSC_RRR_0_OPCODE_X1 = 96,
+  V4ADD_RRR_0_OPCODE_X0 = 150,
+  V4ADD_RRR_0_OPCODE_X1 = 97,
+  V4INT_H_RRR_0_OPCODE_X0 = 151,
+  V4INT_H_RRR_0_OPCODE_X1 = 98,
+  V4INT_L_RRR_0_OPCODE_X0 = 152,
+  V4INT_L_RRR_0_OPCODE_X1 = 99,
+  V4PACKSC_RRR_0_OPCODE_X0 = 153,
+  V4PACKSC_RRR_0_OPCODE_X1 = 100,
+  V4SHLSC_RRR_0_OPCODE_X0 = 154,
+  V4SHLSC_RRR_0_OPCODE_X1 = 101,
+  V4SHL_RRR_0_OPCODE_X0 = 155,
+  V4SHL_RRR_0_OPCODE_X1 = 102,
+  V4SHRS_RRR_0_OPCODE_X0 = 156,
+  V4SHRS_RRR_0_OPCODE_X1 = 103,
+  V4SHRU_RRR_0_OPCODE_X0 = 157,
+  V4SHRU_RRR_0_OPCODE_X1 = 104,
+  V4SUBSC_RRR_0_OPCODE_X0 = 158,
+  V4SUBSC_RRR_0_OPCODE_X1 = 105,
+  V4SUB_RRR_0_OPCODE_X0 = 159,
+  V4SUB_RRR_0_OPCODE_X1 = 106,
+  WH64_UNARY_OPCODE_X1 = 38,
+  XORI_IMM8_OPCODE_X0 = 20,
+  XORI_IMM8_OPCODE_X1 = 45,
+  XOR_RRR_0_OPCODE_X0 = 160,
+  XOR_RRR_0_OPCODE_X1 = 107,
+  XOR_RRR_5_OPCODE_Y0 = 3,
+  XOR_RRR_5_OPCODE_Y1 = 3
+};
+
+
+#endif /* __ASSEMBLER__ */
+
+#endif /* __ARCH_OPCODE_H__ */
diff --git a/kernel/arch/tile/include/uapi/arch/opcode_tilepro.h b/kernel/arch/tile/include/uapi/arch/opcode_tilepro.h
new file mode 100644
index 000000000..4451cff1a
--- /dev/null
+++ b/kernel/arch/tile/include/uapi/arch/opcode_tilepro.h
@@ -0,0 +1,1472 @@
+/* TILEPro opcode information.
+ *
+ * Copyright 2011 Tilera Corporation. All Rights Reserved.
+ *
+ *   This program is free software; you can redistribute it and/or
+ *   modify it under the terms of the GNU General Public License
+ *   as published by the Free Software Foundation, version 2.
+ *
+ *   This program is distributed in the hope that it will be useful, but
+ *   WITHOUT ANY WARRANTY; without even the implied warranty of
+ *   MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
+ *   NON INFRINGEMENT.  See the GNU General Public License for
+ *   more details.
+ *
+ *
+ *
+ *
+ *
+ */
+
+#ifndef __ARCH_OPCODE_H__
+#define __ARCH_OPCODE_H__
+
+#ifndef __ASSEMBLER__
+
+typedef unsigned long long tilepro_bundle_bits;
+
+/* This is the bit that determines if a bundle is in the Y encoding. */
+#define TILEPRO_BUNDLE_Y_ENCODING_MASK ((tilepro_bundle_bits)1 << 63)
+
+enum
+{
+  /* Maximum number of instructions in a bundle (2 for X, 3 for Y). */
+  TILEPRO_MAX_INSTRUCTIONS_PER_BUNDLE = 3,
+
+  /* How many different pipeline encodings are there? X0, X1, Y0, Y1, Y2. */
+  TILEPRO_NUM_PIPELINE_ENCODINGS = 5,
+
+  /* Log base 2 of TILEPRO_BUNDLE_SIZE_IN_BYTES. */
+  TILEPRO_LOG2_BUNDLE_SIZE_IN_BYTES = 3,
+
+  /* Instructions take this many bytes. */
+  TILEPRO_BUNDLE_SIZE_IN_BYTES = 1 << TILEPRO_LOG2_BUNDLE_SIZE_IN_BYTES,
+
+  /* Log base 2 of TILEPRO_BUNDLE_ALIGNMENT_IN_BYTES. */
+  TILEPRO_LOG2_BUNDLE_ALIGNMENT_IN_BYTES = 3,
+
+  /* Bundles should be aligned modulo this number of bytes. */
+  TILEPRO_BUNDLE_ALIGNMENT_IN_BYTES =
+    (1 << TILEPRO_LOG2_BUNDLE_ALIGNMENT_IN_BYTES),
+
+  /* Log base 2 of TILEPRO_SN_INSTRUCTION_SIZE_IN_BYTES. */
+  TILEPRO_LOG2_SN_INSTRUCTION_SIZE_IN_BYTES = 1,
+
+  /* Static network instructions take this many bytes. */
+  TILEPRO_SN_INSTRUCTION_SIZE_IN_BYTES =
+    (1 << TILEPRO_LOG2_SN_INSTRUCTION_SIZE_IN_BYTES),
+
+  /* Number of registers (some are magic, such as network I/O). */
+  TILEPRO_NUM_REGISTERS = 64,
+
+  /* Number of static network registers. */
+  TILEPRO_NUM_SN_REGISTERS = 4
+};
+
+/* Make a few "tile_" variables to simplify common code between
+   architectures.  */
+
+typedef tilepro_bundle_bits tile_bundle_bits;
+#define TILE_BUNDLE_SIZE_IN_BYTES TILEPRO_BUNDLE_SIZE_IN_BYTES
+#define TILE_BUNDLE_ALIGNMENT_IN_BYTES TILEPRO_BUNDLE_ALIGNMENT_IN_BYTES
+#define TILE_LOG2_BUNDLE_ALIGNMENT_IN_BYTES \
+  TILEPRO_LOG2_BUNDLE_ALIGNMENT_IN_BYTES
+#define TILE_BPT_BUNDLE TILEPRO_BPT_BUNDLE
+
+/* 64-bit pattern for a { bpt ; nop } bundle. */
+#define TILEPRO_BPT_BUNDLE 0x400b3cae70166000ULL
+
+static __inline unsigned int
+get_BrOff_SN(tilepro_bundle_bits num)
+{
+  const unsigned int n = (unsigned int)num;
+  return (((n >> 0)) & 0x3ff);
+}
+
+static __inline unsigned int
+get_BrOff_X1(tilepro_bundle_bits n)
+{
+  return (((unsigned int)(n >> 43)) & 0x00007fff) |
+         (((unsigned int)(n >> 20)) & 0x00018000);
+}
+
+static __inline unsigned int
+get_BrType_X1(tilepro_bundle_bits n)
+{
+  return (((unsigned int)(n >> 31)) & 0xf);
+}
+
+static __inline unsigned int
+get_Dest_Imm8_X1(tilepro_bundle_bits n)
+{
+  return (((unsigned int)(n >> 31)) & 0x0000003f) |
+         (((unsigned int)(n >> 43)) & 0x000000c0);
+}
+
+static __inline unsigned int
+get_Dest_SN(tilepro_bundle_bits num)
+{
+  const unsigned int n = (unsigned int)num;
+  return (((n >> 2)) & 0x3);
+}
+
+static __inline unsigned int
+get_Dest_X0(tilepro_bundle_bits num)
+{
+  const unsigned int n = (unsigned int)num;
+  return (((n >> 0)) & 0x3f);
+}
+
+static __inline unsigned int
+get_Dest_X1(tilepro_bundle_bits n)
+{
+  return (((unsigned int)(n >> 31)) & 0x3f);
+}
+
+static __inline unsigned int
+get_Dest_Y0(tilepro_bundle_bits num)
+{
+  const unsigned int n = (unsigned int)num;
+  return (((n >> 0)) & 0x3f);
+}
+
+static __inline unsigned int
+get_Dest_Y1(tilepro_bundle_bits n)
+{
+  return (((unsigned int)(n >> 31)) & 0x3f);
+}
+
+static __inline unsigned int
+get_Imm16_X0(tilepro_bundle_bits num)
+{
+  const unsigned int n = (unsigned int)num;
+  return (((n >> 12)) & 0xffff);
+}
+
+static __inline unsigned int
+get_Imm16_X1(tilepro_bundle_bits n)
+{
+  return (((unsigned int)(n >> 43)) & 0xffff);
+}
+
+static __inline unsigned int
+get_Imm8_SN(tilepro_bundle_bits num)
+{
+  const unsigned int n = (unsigned int)num;
+  return (((n >> 0)) & 0xff);
+}
+
+static __inline unsigned int
+get_Imm8_X0(tilepro_bundle_bits num)
+{
+  const unsigned int n = (unsigned int)num;
+  return (((n >> 12)) & 0xff);
+}
+
+static __inline unsigned int
+get_Imm8_X1(tilepro_bundle_bits n)
+{
+  return (((unsigned int)(n >> 43)) & 0xff);
+}
+
+static __inline unsigned int
+get_Imm8_Y0(tilepro_bundle_bits num)
+{
+  const unsigned int n = (unsigned int)num;
+  return (((n >> 12)) & 0xff);
+}
+
+static __inline unsigned int
+get_Imm8_Y1(tilepro_bundle_bits n)
+{
+  return (((unsigned int)(n >> 43)) & 0xff);
+}
+
+static __inline unsigned int
+get_ImmOpcodeExtension_X0(tilepro_bundle_bits num)
+{
+  const unsigned int n = (unsigned int)num;
+  return (((n >> 20)) & 0x7f);
+}
+
+static __inline unsigned int
+get_ImmOpcodeExtension_X1(tilepro_bundle_bits n)
+{
+  return (((unsigned int)(n >> 51)) & 0x7f);
+}
+
+static __inline unsigned int
+get_ImmRROpcodeExtension_SN(tilepro_bundle_bits num)
+{
+  const unsigned int n = (unsigned int)num;
+  return (((n >> 8)) & 0x3);
+}
+
+static __inline unsigned int
+get_JOffLong_X1(tilepro_bundle_bits n)
+{
+  return (((unsigned int)(n >> 43)) & 0x00007fff) |
+         (((unsigned int)(n >> 20)) & 0x00018000) |
+         (((unsigned int)(n >> 14)) & 0x001e0000) |
+         (((unsigned int)(n >> 16)) & 0x07e00000) |
+         (((unsigned int)(n >> 31)) & 0x18000000);
+}
+
+static __inline unsigned int
+get_JOff_X1(tilepro_bundle_bits n)
+{
+  return (((unsigned int)(n >> 43)) & 0x00007fff) |
+         (((unsigned int)(n >> 20)) & 0x00018000) |
+         (((unsigned int)(n >> 14)) & 0x001e0000) |
+         (((unsigned int)(n >> 16)) & 0x07e00000) |
+         (((unsigned int)(n >> 31)) & 0x08000000);
+}
+
+static __inline unsigned int
+get_MF_Imm15_X1(tilepro_bundle_bits n)
+{
+  return (((unsigned int)(n >> 37)) & 0x00003fff) |
+         (((unsigned int)(n >> 44)) & 0x00004000);
+}
+
+static __inline unsigned int
+get_MMEnd_X0(tilepro_bundle_bits num)
+{
+  const unsigned int n = (unsigned int)num;
+  return (((n >> 18)) & 0x1f);
+}
+
+static __inline unsigned int
+get_MMEnd_X1(tilepro_bundle_bits n)
+{
+  return (((unsigned int)(n >> 49)) & 0x1f);
+}
+
+static __inline unsigned int
+get_MMStart_X0(tilepro_bundle_bits num)
+{
+  const unsigned int n = (unsigned int)num;
+  return (((n >> 23)) & 0x1f);
+}
+
+static __inline unsigned int
+get_MMStart_X1(tilepro_bundle_bits n)
+{
+  return (((unsigned int)(n >> 54)) & 0x1f);
+}
+
+static __inline unsigned int
+get_MT_Imm15_X1(tilepro_bundle_bits n)
+{
+  return (((unsigned int)(n >> 31)) & 0x0000003f) |
+         (((unsigned int)(n >> 37)) & 0x00003fc0) |
+         (((unsigned int)(n >> 44)) & 0x00004000);
+}
+
+static __inline unsigned int
+get_Mode(tilepro_bundle_bits n)
+{
+  return (((unsigned int)(n >> 63)) & 0x1);
+}
+
+static __inline unsigned int
+get_NoRegOpcodeExtension_SN(tilepro_bundle_bits num)
+{
+  const unsigned int n = (unsigned int)num;
+  return (((n >> 0)) & 0xf);
+}
+
+static __inline unsigned int
+get_Opcode_SN(tilepro_bundle_bits num)
+{
+  const unsigned int n = (unsigned int)num;
+  return (((n >> 10)) & 0x3f);
+}
+
+static __inline unsigned int
+get_Opcode_X0(tilepro_bundle_bits num)
+{
+  const unsigned int n = (unsigned int)num;
+  return (((n >> 28)) & 0x7);
+}
+
+static __inline unsigned int
+get_Opcode_X1(tilepro_bundle_bits n)
+{
+  return (((unsigned int)(n >> 59)) & 0xf);
+}
+
+static __inline unsigned int
+get_Opcode_Y0(tilepro_bundle_bits num)
+{
+  const unsigned int n = (unsigned int)num;
+  return (((n >> 27)) & 0xf);
+}
+
+static __inline unsigned int
+get_Opcode_Y1(tilepro_bundle_bits n)
+{
+  return (((unsigned int)(n >> 59)) & 0xf);
+}
+
+static __inline unsigned int
+get_Opcode_Y2(tilepro_bundle_bits n)
+{
+  return (((unsigned int)(n >> 56)) & 0x7);
+}
+
+static __inline unsigned int
+get_RROpcodeExtension_SN(tilepro_bundle_bits num)
+{
+  const unsigned int n = (unsigned int)num;
+  return (((n >> 4)) & 0xf);
+}
+
+static __inline unsigned int
+get_RRROpcodeExtension_X0(tilepro_bundle_bits num)
+{
+  const unsigned int n = (unsigned int)num;
+  return (((n >> 18)) & 0x1ff);
+}
+
+static __inline unsigned int
+get_RRROpcodeExtension_X1(tilepro_bundle_bits n)
+{
+  return (((unsigned int)(n >> 49)) & 0x1ff);
+}
+
+static __inline unsigned int
+get_RRROpcodeExtension_Y0(tilepro_bundle_bits num)
+{
+  const unsigned int n = (unsigned int)num;
+  return (((n >> 18)) & 0x3);
+}
+
+static __inline unsigned int
+get_RRROpcodeExtension_Y1(tilepro_bundle_bits n)
+{
+  return (((unsigned int)(n >> 49)) & 0x3);
+}
+
+static __inline unsigned int
+get_RouteOpcodeExtension_SN(tilepro_bundle_bits num)
+{
+  const unsigned int n = (unsigned int)num;
+  return (((n >> 0)) & 0x3ff);
+}
+
+static __inline unsigned int
+get_S_X0(tilepro_bundle_bits num)
+{
+  const unsigned int n = (unsigned int)num;
+  return (((n >> 27)) & 0x1);
+}
+
+static __inline unsigned int
+get_S_X1(tilepro_bundle_bits n)
+{
+  return (((unsigned int)(n >> 58)) & 0x1);
+}
+
+static __inline unsigned int
+get_ShAmt_X0(tilepro_bundle_bits num)
+{
+  const unsigned int n = (unsigned int)num;
+  return (((n >> 12)) & 0x1f);
+}
+
+static __inline unsigned int
+get_ShAmt_X1(tilepro_bundle_bits n)
+{
+  return (((unsigned int)(n >> 43)) & 0x1f);
+}
+
+static __inline unsigned int
+get_ShAmt_Y0(tilepro_bundle_bits num)
+{
+  const unsigned int n = (unsigned int)num;
+  return (((n >> 12)) & 0x1f);
+}
+
+static __inline unsigned int
+get_ShAmt_Y1(tilepro_bundle_bits n)
+{
+  return (((unsigned int)(n >> 43)) & 0x1f);
+}
+
+static __inline unsigned int
+get_SrcA_X0(tilepro_bundle_bits num)
+{
+  const unsigned int n = (unsigned int)num;
+  return (((n >> 6)) & 0x3f);
+}
+
+static __inline unsigned int
+get_SrcA_X1(tilepro_bundle_bits n)
+{
+  return (((unsigned int)(n >> 37)) & 0x3f);
+}
+
+static __inline unsigned int
+get_SrcA_Y0(tilepro_bundle_bits num)
+{
+  const unsigned int n = (unsigned int)num;
+  return (((n >> 6)) & 0x3f);
+}
+
+static __inline unsigned int
+get_SrcA_Y1(tilepro_bundle_bits n)
+{
+  return (((unsigned int)(n >> 37)) & 0x3f);
+}
+
+static __inline unsigned int
+get_SrcA_Y2(tilepro_bundle_bits n)
+{
+  return (((n >> 26)) & 0x00000001) |
+         (((unsigned int)(n >> 50)) & 0x0000003e);
+}
+
+static __inline unsigned int
+get_SrcBDest_Y2(tilepro_bundle_bits num)
+{
+  const unsigned int n = (unsigned int)num;
+  return (((n >> 20)) & 0x3f);
+}
+
+static __inline unsigned int
+get_SrcB_X0(tilepro_bundle_bits num)
+{
+  const unsigned int n = (unsigned int)num;
+  return (((n >> 12)) & 0x3f);
+}
+
+static __inline unsigned int
+get_SrcB_X1(tilepro_bundle_bits n)
+{
+  return (((unsigned int)(n >> 43)) & 0x3f);
+}
+
+static __inline unsigned int
+get_SrcB_Y0(tilepro_bundle_bits num)
+{
+  const unsigned int n = (unsigned int)num;
+  return (((n >> 12)) & 0x3f);
+}
+
+static __inline unsigned int
+get_SrcB_Y1(tilepro_bundle_bits n)
+{
+  return (((unsigned int)(n >> 43)) & 0x3f);
+}
+
+static __inline unsigned int
+get_Src_SN(tilepro_bundle_bits num)
+{
+  const unsigned int n = (unsigned int)num;
+  return (((n >> 0)) & 0x3);
+}
+
+static __inline unsigned int
+get_UnOpcodeExtension_X0(tilepro_bundle_bits num)
+{
+  const unsigned int n = (unsigned int)num;
+  return (((n >> 12)) & 0x1f);
+}
+
+static __inline unsigned int
+get_UnOpcodeExtension_X1(tilepro_bundle_bits n)
+{
+  return (((unsigned int)(n >> 43)) & 0x1f);
+}
+
+static __inline unsigned int
+get_UnOpcodeExtension_Y0(tilepro_bundle_bits num)
+{
+  const unsigned int n = (unsigned int)num;
+  return (((n >> 12)) & 0x1f);
+}
+
+static __inline unsigned int
+get_UnOpcodeExtension_Y1(tilepro_bundle_bits n)
+{
+  return (((unsigned int)(n >> 43)) & 0x1f);
+}
+
+static __inline unsigned int
+get_UnShOpcodeExtension_X0(tilepro_bundle_bits num)
+{
+  const unsigned int n = (unsigned int)num;
+  return (((n >> 17)) & 0x3ff);
+}
+
+static __inline unsigned int
+get_UnShOpcodeExtension_X1(tilepro_bundle_bits n)
+{
+  return (((unsigned int)(n >> 48)) & 0x3ff);
+}
+
+static __inline unsigned int
+get_UnShOpcodeExtension_Y0(tilepro_bundle_bits num)
+{
+  const unsigned int n = (unsigned int)num;
+  return (((n >> 17)) & 0x7);
+}
+
+static __inline unsigned int
+get_UnShOpcodeExtension_Y1(tilepro_bundle_bits n)
+{
+  return (((unsigned int)(n >> 48)) & 0x7);
+}
+
+
+static __inline int
+sign_extend(int n, int num_bits)
+{
+  int shift = (int)(sizeof(int) * 8 - num_bits);
+  return (n << shift) >> shift;
+}
+
+
+
+static __inline tilepro_bundle_bits
+create_BrOff_SN(int num)
+{
+  const unsigned int n = (unsigned int)num;
+  return ((n & 0x3ff) << 0);
+}
+
+static __inline tilepro_bundle_bits
+create_BrOff_X1(int num)
+{
+  const unsigned int n = (unsigned int)num;
+  return (((tilepro_bundle_bits)(n & 0x00007fff)) << 43) |
+         (((tilepro_bundle_bits)(n & 0x00018000)) << 20);
+}
+
+static __inline tilepro_bundle_bits
+create_BrType_X1(int num)
+{
+  const unsigned int n = (unsigned int)num;
+  return (((tilepro_bundle_bits)(n & 0xf)) << 31);
+}
+
+static __inline tilepro_bundle_bits
+create_Dest_Imm8_X1(int num)
+{
+  const unsigned int n = (unsigned int)num;
+  return (((tilepro_bundle_bits)(n & 0x0000003f)) << 31) |
+         (((tilepro_bundle_bits)(n & 0x000000c0)) << 43);
+}
+
+static __inline tilepro_bundle_bits
+create_Dest_SN(int num)
+{
+  const unsigned int n = (unsigned int)num;
+  return ((n & 0x3) << 2);
+}
+
+static __inline tilepro_bundle_bits
+create_Dest_X0(int num)
+{
+  const unsigned int n = (unsigned int)num;
+  return ((n & 0x3f) << 0);
+}
+
+static __inline tilepro_bundle_bits
+create_Dest_X1(int num)
+{
+  const unsigned int n = (unsigned int)num;
+  return (((tilepro_bundle_bits)(n & 0x3f)) << 31);
+}
+
+static __inline tilepro_bundle_bits
+create_Dest_Y0(int num)
+{
+  const unsigned int n = (unsigned int)num;
+  return ((n & 0x3f) << 0);
+}
+
+static __inline tilepro_bundle_bits
+create_Dest_Y1(int num)
+{
+  const unsigned int n = (unsigned int)num;
+  return (((tilepro_bundle_bits)(n & 0x3f)) << 31);
+}
+
+static __inline tilepro_bundle_bits
+create_Imm16_X0(int num)
+{
+  const unsigned int n = (unsigned int)num;
+  return ((n & 0xffff) << 12);
+}
+
+static __inline tilepro_bundle_bits
+create_Imm16_X1(int num)
+{
+  const unsigned int n = (unsigned int)num;
+  return (((tilepro_bundle_bits)(n & 0xffff)) << 43);
+}
+
+static __inline tilepro_bundle_bits
+create_Imm8_SN(int num)
+{
+  const unsigned int n = (unsigned int)num;
+  return ((n & 0xff) << 0);
+}
+
+static __inline tilepro_bundle_bits
+create_Imm8_X0(int num)
+{
+  const unsigned int n = (unsigned int)num;
+  return ((n & 0xff) << 12);
+}
+
+static __inline tilepro_bundle_bits
+create_Imm8_X1(int num)
+{
+  const unsigned int n = (unsigned int)num;
+  return (((tilepro_bundle_bits)(n & 0xff)) << 43);
+}
+
+static __inline tilepro_bundle_bits
+create_Imm8_Y0(int num)
+{
+  const unsigned int n = (unsigned int)num;
+  return ((n & 0xff) << 12);
+}
+
+static __inline tilepro_bundle_bits
+create_Imm8_Y1(int num)
+{
+  const unsigned int n = (unsigned int)num;
+  return (((tilepro_bundle_bits)(n & 0xff)) << 43);
+}
+
+static __inline tilepro_bundle_bits
+create_ImmOpcodeExtension_X0(int num)
+{
+  const unsigned int n = (unsigned int)num;
+  return ((n & 0x7f) << 20);
+}
+
+static __inline tilepro_bundle_bits
+create_ImmOpcodeExtension_X1(int num)
+{
+  const unsigned int n = (unsigned int)num;
+  return (((tilepro_bundle_bits)(n & 0x7f)) << 51);
+}
+
+static __inline tilepro_bundle_bits
+create_ImmRROpcodeExtension_SN(int num)
+{
+  const unsigned int n = (unsigned int)num;
+  return ((n & 0x3) << 8);
+}
+
+static __inline tilepro_bundle_bits
+create_JOffLong_X1(int num)
+{
+  const unsigned int n = (unsigned int)num;
+  return (((tilepro_bundle_bits)(n & 0x00007fff)) << 43) |
+         (((tilepro_bundle_bits)(n & 0x00018000)) << 20) |
+         (((tilepro_bundle_bits)(n & 0x001e0000)) << 14) |
+         (((tilepro_bundle_bits)(n & 0x07e00000)) << 16) |
+         (((tilepro_bundle_bits)(n & 0x18000000)) << 31);
+}
+
+static __inline tilepro_bundle_bits
+create_JOff_X1(int num)
+{
+  const unsigned int n = (unsigned int)num;
+  return (((tilepro_bundle_bits)(n & 0x00007fff)) << 43) |
+         (((tilepro_bundle_bits)(n & 0x00018000)) << 20) |
+         (((tilepro_bundle_bits)(n & 0x001e0000)) << 14) |
+         (((tilepro_bundle_bits)(n & 0x07e00000)) << 16) |
+         (((tilepro_bundle_bits)(n & 0x08000000)) << 31);
+}
+
+static __inline tilepro_bundle_bits
+create_MF_Imm15_X1(int num)
+{
+  const unsigned int n = (unsigned int)num;
+  return (((tilepro_bundle_bits)(n & 0x00003fff)) << 37) |
+         (((tilepro_bundle_bits)(n & 0x00004000)) << 44);
+}
+
+static __inline tilepro_bundle_bits
+create_MMEnd_X0(int num)
+{
+  const unsigned int n = (unsigned int)num;
+  return ((n & 0x1f) << 18);
+}
+
+static __inline tilepro_bundle_bits
+create_MMEnd_X1(int num)
+{
+  const unsigned int n = (unsigned int)num;
+  return (((tilepro_bundle_bits)(n & 0x1f)) << 49);
+}
+
+static __inline tilepro_bundle_bits
+create_MMStart_X0(int num)
+{
+  const unsigned int n = (unsigned int)num;
+  return ((n & 0x1f) << 23);
+}
+
+static __inline tilepro_bundle_bits
+create_MMStart_X1(int num)
+{
+  const unsigned int n = (unsigned int)num;
+  return (((tilepro_bundle_bits)(n & 0x1f)) << 54);
+}
+
+static __inline tilepro_bundle_bits
+create_MT_Imm15_X1(int num)
+{
+  const unsigned int n = (unsigned int)num;
+  return (((tilepro_bundle_bits)(n & 0x0000003f)) << 31) |
+         (((tilepro_bundle_bits)(n & 0x00003fc0)) << 37) |
+         (((tilepro_bundle_bits)(n & 0x00004000)) << 44);
+}
+
+static __inline tilepro_bundle_bits
+create_Mode(int num)
+{
+  const unsigned int n = (unsigned int)num;
+  return (((tilepro_bundle_bits)(n & 0x1)) << 63);
+}
+
+static __inline tilepro_bundle_bits
+create_NoRegOpcodeExtension_SN(int num)
+{
+  const unsigned int n = (unsigned int)num;
+  return ((n & 0xf) << 0);
+}
+
+static __inline tilepro_bundle_bits
+create_Opcode_SN(int num)
+{
+  const unsigned int n = (unsigned int)num;
+  return ((n & 0x3f) << 10);
+}
+
+static __inline tilepro_bundle_bits
+create_Opcode_X0(int num)
+{
+  const unsigned int n = (unsigned int)num;
+  return ((n & 0x7) << 28);
+}
+
+static __inline tilepro_bundle_bits
+create_Opcode_X1(int num)
+{
+  const unsigned int n = (unsigned int)num;
+  return (((tilepro_bundle_bits)(n & 0xf)) << 59);
+}
+
+static __inline tilepro_bundle_bits
+create_Opcode_Y0(int num)
+{
+  const unsigned int n = (unsigned int)num;
+  return ((n & 0xf) << 27);
+}
+
+static __inline tilepro_bundle_bits
+create_Opcode_Y1(int num)
+{
+  const unsigned int n = (unsigned int)num;
+  return (((tilepro_bundle_bits)(n & 0xf)) << 59);
+}
+
+static __inline tilepro_bundle_bits
+create_Opcode_Y2(int num)
+{
+  const unsigned int n = (unsigned int)num;
+  return (((tilepro_bundle_bits)(n & 0x7)) << 56);
+}
+
+static __inline tilepro_bundle_bits
+create_RROpcodeExtension_SN(int num)
+{
+  const unsigned int n = (unsigned int)num;
+  return ((n & 0xf) << 4);
+}
+
+static __inline tilepro_bundle_bits
+create_RRROpcodeExtension_X0(int num)
+{
+  const unsigned int n = (unsigned int)num;
+  return ((n & 0x1ff) << 18);
+}
+
+static __inline tilepro_bundle_bits
+create_RRROpcodeExtension_X1(int num)
+{
+  const unsigned int n = (unsigned int)num;
+  return (((tilepro_bundle_bits)(n & 0x1ff)) << 49);
+}
+
+static __inline tilepro_bundle_bits
+create_RRROpcodeExtension_Y0(int num)
+{
+  const unsigned int n = (unsigned int)num;
+  return ((n & 0x3) << 18);
+}
+
+static __inline tilepro_bundle_bits
+create_RRROpcodeExtension_Y1(int num)
+{
+  const unsigned int n = (unsigned int)num;
+  return (((tilepro_bundle_bits)(n & 0x3)) << 49);
+}
+
+static __inline tilepro_bundle_bits
+create_RouteOpcodeExtension_SN(int num)
+{
+  const unsigned int n = (unsigned int)num;
+  return ((n & 0x3ff) << 0);
+}
+
+static __inline tilepro_bundle_bits
+create_S_X0(int num)
+{
+  const unsigned int n = (unsigned int)num;
+  return ((n & 0x1) << 27);
+}
+
+static __inline tilepro_bundle_bits
+create_S_X1(int num)
+{
+  const unsigned int n = (unsigned int)num;
+  return (((tilepro_bundle_bits)(n & 0x1)) << 58);
+}
+
+static __inline tilepro_bundle_bits
+create_ShAmt_X0(int num)
+{
+  const unsigned int n = (unsigned int)num;
+  return ((n & 0x1f) << 12);
+}
+
+static __inline tilepro_bundle_bits
+create_ShAmt_X1(int num)
+{
+  const unsigned int n = (unsigned int)num;
+  return (((tilepro_bundle_bits)(n & 0x1f)) << 43);
+}
+
+static __inline tilepro_bundle_bits
+create_ShAmt_Y0(int num)
+{
+  const unsigned int n = (unsigned int)num;
+  return ((n & 0x1f) << 12);
+}
+
+static __inline tilepro_bundle_bits
+create_ShAmt_Y1(int num)
+{
+  const unsigned int n = (unsigned int)num;
+  return (((tilepro_bundle_bits)(n & 0x1f)) << 43);
+}
+
+static __inline tilepro_bundle_bits
+create_SrcA_X0(int num)
+{
+  const unsigned int n = (unsigned int)num;
+  return ((n & 0x3f) << 6);
+}
+
+static __inline tilepro_bundle_bits
+create_SrcA_X1(int num)
+{
+  const unsigned int n = (unsigned int)num;
+  return (((tilepro_bundle_bits)(n & 0x3f)) << 37);
+}
+
+static __inline tilepro_bundle_bits
+create_SrcA_Y0(int num)
+{
+  const unsigned int n = (unsigned int)num;
+  return ((n & 0x3f) << 6);
+}
+
+static __inline tilepro_bundle_bits
+create_SrcA_Y1(int num)
+{
+  const unsigned int n = (unsigned int)num;
+  return (((tilepro_bundle_bits)(n & 0x3f)) << 37);
+}
+
+static __inline tilepro_bundle_bits
+create_SrcA_Y2(int num)
+{
+  const unsigned int n = (unsigned int)num;
+  return ((n & 0x00000001) << 26) |
+         (((tilepro_bundle_bits)(n & 0x0000003e)) << 50);
+}
+
+static __inline tilepro_bundle_bits
+create_SrcBDest_Y2(int num)
+{
+  const unsigned int n = (unsigned int)num;
+  return ((n & 0x3f) << 20);
+}
+
+static __inline tilepro_bundle_bits
+create_SrcB_X0(int num)
+{
+  const unsigned int n = (unsigned int)num;
+  return ((n & 0x3f) << 12);
+}
+
+static __inline tilepro_bundle_bits
+create_SrcB_X1(int num)
+{
+  const unsigned int n = (unsigned int)num;
+  return (((tilepro_bundle_bits)(n & 0x3f)) << 43);
+}
+
+static __inline tilepro_bundle_bits
+create_SrcB_Y0(int num)
+{
+  const unsigned int n = (unsigned int)num;
+  return ((n & 0x3f) << 12);
+}
+
+static __inline tilepro_bundle_bits
+create_SrcB_Y1(int num)
+{
+  const unsigned int n = (unsigned int)num;
+  return (((tilepro_bundle_bits)(n & 0x3f)) << 43);
+}
+
+static __inline tilepro_bundle_bits
+create_Src_SN(int num)
+{
+  const unsigned int n = (unsigned int)num;
+  return ((n & 0x3) << 0);
+}
+
+static __inline tilepro_bundle_bits
+create_UnOpcodeExtension_X0(int num)
+{
+  const unsigned int n = (unsigned int)num;
+  return ((n & 0x1f) << 12);
+}
+
+static __inline tilepro_bundle_bits
+create_UnOpcodeExtension_X1(int num)
+{
+  const unsigned int n = (unsigned int)num;
+  return (((tilepro_bundle_bits)(n & 0x1f)) << 43);
+}
+
+static __inline tilepro_bundle_bits
+create_UnOpcodeExtension_Y0(int num)
+{
+  const unsigned int n = (unsigned int)num;
+  return ((n & 0x1f) << 12);
+}
+
+static __inline tilepro_bundle_bits
+create_UnOpcodeExtension_Y1(int num)
+{
+  const unsigned int n = (unsigned int)num;
+  return (((tilepro_bundle_bits)(n & 0x1f)) << 43);
+}
+
+static __inline tilepro_bundle_bits
+create_UnShOpcodeExtension_X0(int num)
+{
+  const unsigned int n = (unsigned int)num;
+  return ((n & 0x3ff) << 17);
+}
+
+static __inline tilepro_bundle_bits
+create_UnShOpcodeExtension_X1(int num)
+{
+  const unsigned int n = (unsigned int)num;
+  return (((tilepro_bundle_bits)(n & 0x3ff)) << 48);
+}
+
+static __inline tilepro_bundle_bits
+create_UnShOpcodeExtension_Y0(int num)
+{
+  const unsigned int n = (unsigned int)num;
+  return ((n & 0x7) << 17);
+}
+
+static __inline tilepro_bundle_bits
+create_UnShOpcodeExtension_Y1(int num)
+{
+  const unsigned int n = (unsigned int)num;
+  return (((tilepro_bundle_bits)(n & 0x7)) << 48);
+}
+
+
+enum
+{
+  ADDBS_U_SPECIAL_0_OPCODE_X0 = 98,
+  ADDBS_U_SPECIAL_0_OPCODE_X1 = 68,
+  ADDB_SPECIAL_0_OPCODE_X0 = 1,
+  ADDB_SPECIAL_0_OPCODE_X1 = 1,
+  ADDHS_SPECIAL_0_OPCODE_X0 = 99,
+  ADDHS_SPECIAL_0_OPCODE_X1 = 69,
+  ADDH_SPECIAL_0_OPCODE_X0 = 2,
+  ADDH_SPECIAL_0_OPCODE_X1 = 2,
+  ADDIB_IMM_0_OPCODE_X0 = 1,
+  ADDIB_IMM_0_OPCODE_X1 = 1,
+  ADDIH_IMM_0_OPCODE_X0 = 2,
+  ADDIH_IMM_0_OPCODE_X1 = 2,
+  ADDI_IMM_0_OPCODE_X0 = 3,
+  ADDI_IMM_0_OPCODE_X1 = 3,
+  ADDI_IMM_1_OPCODE_SN = 1,
+  ADDI_OPCODE_Y0 = 9,
+  ADDI_OPCODE_Y1 = 7,
+  ADDLIS_OPCODE_X0 = 1,
+  ADDLIS_OPCODE_X1 = 2,
+  ADDLI_OPCODE_X0 = 2,
+  ADDLI_OPCODE_X1 = 3,
+  ADDS_SPECIAL_0_OPCODE_X0 = 96,
+  ADDS_SPECIAL_0_OPCODE_X1 = 66,
+  ADD_SPECIAL_0_OPCODE_X0 = 3,
+  ADD_SPECIAL_0_OPCODE_X1 = 3,
+  ADD_SPECIAL_0_OPCODE_Y0 = 0,
+  ADD_SPECIAL_0_OPCODE_Y1 = 0,
+  ADIFFB_U_SPECIAL_0_OPCODE_X0 = 4,
+  ADIFFH_SPECIAL_0_OPCODE_X0 = 5,
+  ANDI_IMM_0_OPCODE_X0 = 1,
+  ANDI_IMM_0_OPCODE_X1 = 4,
+  ANDI_OPCODE_Y0 = 10,
+  ANDI_OPCODE_Y1 = 8,
+  AND_SPECIAL_0_OPCODE_X0 = 6,
+  AND_SPECIAL_0_OPCODE_X1 = 4,
+  AND_SPECIAL_2_OPCODE_Y0 = 0,
+  AND_SPECIAL_2_OPCODE_Y1 = 0,
+  AULI_OPCODE_X0 = 3,
+  AULI_OPCODE_X1 = 4,
+  AVGB_U_SPECIAL_0_OPCODE_X0 = 7,
+  AVGH_SPECIAL_0_OPCODE_X0 = 8,
+  BBNST_BRANCH_OPCODE_X1 = 15,
+  BBNS_BRANCH_OPCODE_X1 = 14,
+  BBNS_OPCODE_SN = 63,
+  BBST_BRANCH_OPCODE_X1 = 13,
+  BBS_BRANCH_OPCODE_X1 = 12,
+  BBS_OPCODE_SN = 62,
+  BGEZT_BRANCH_OPCODE_X1 = 7,
+  BGEZ_BRANCH_OPCODE_X1 = 6,
+  BGEZ_OPCODE_SN = 61,
+  BGZT_BRANCH_OPCODE_X1 = 5,
+  BGZ_BRANCH_OPCODE_X1 = 4,
+  BGZ_OPCODE_SN = 58,
+  BITX_UN_0_SHUN_0_OPCODE_X0 = 1,
+  BITX_UN_0_SHUN_0_OPCODE_Y0 = 1,
+  BLEZT_BRANCH_OPCODE_X1 = 11,
+  BLEZ_BRANCH_OPCODE_X1 = 10,
+  BLEZ_OPCODE_SN = 59,
+  BLZT_BRANCH_OPCODE_X1 = 9,
+  BLZ_BRANCH_OPCODE_X1 = 8,
+  BLZ_OPCODE_SN = 60,
+  BNZT_BRANCH_OPCODE_X1 = 3,
+  BNZ_BRANCH_OPCODE_X1 = 2,
+  BNZ_OPCODE_SN = 57,
+  BPT_NOREG_RR_IMM_0_OPCODE_SN = 1,
+  BRANCH_OPCODE_X1 = 5,
+  BYTEX_UN_0_SHUN_0_OPCODE_X0 = 2,
+  BYTEX_UN_0_SHUN_0_OPCODE_Y0 = 2,
+  BZT_BRANCH_OPCODE_X1 = 1,
+  BZ_BRANCH_OPCODE_X1 = 0,
+  BZ_OPCODE_SN = 56,
+  CLZ_UN_0_SHUN_0_OPCODE_X0 = 3,
+  CLZ_UN_0_SHUN_0_OPCODE_Y0 = 3,
+  CRC32_32_SPECIAL_0_OPCODE_X0 = 9,
+  CRC32_8_SPECIAL_0_OPCODE_X0 = 10,
+  CTZ_UN_0_SHUN_0_OPCODE_X0 = 4,
+  CTZ_UN_0_SHUN_0_OPCODE_Y0 = 4,
+  DRAIN_UN_0_SHUN_0_OPCODE_X1 = 1,
+  DTLBPR_UN_0_SHUN_0_OPCODE_X1 = 2,
+  DWORD_ALIGN_SPECIAL_0_OPCODE_X0 = 95,
+  FINV_UN_0_SHUN_0_OPCODE_X1 = 3,
+  FLUSH_UN_0_SHUN_0_OPCODE_X1 = 4,
+  FNOP_NOREG_RR_IMM_0_OPCODE_SN = 3,
+  FNOP_UN_0_SHUN_0_OPCODE_X0 = 5,
+  FNOP_UN_0_SHUN_0_OPCODE_X1 = 5,
+  FNOP_UN_0_SHUN_0_OPCODE_Y0 = 5,
+  FNOP_UN_0_SHUN_0_OPCODE_Y1 = 1,
+  HALT_NOREG_RR_IMM_0_OPCODE_SN = 0,
+  ICOH_UN_0_SHUN_0_OPCODE_X1 = 6,
+  ILL_UN_0_SHUN_0_OPCODE_X1 = 7,
+  ILL_UN_0_SHUN_0_OPCODE_Y1 = 2,
+  IMM_0_OPCODE_SN = 0,
+  IMM_0_OPCODE_X0 = 4,
+  IMM_0_OPCODE_X1 = 6,
+  IMM_1_OPCODE_SN = 1,
+  IMM_OPCODE_0_X0 = 5,
+  INTHB_SPECIAL_0_OPCODE_X0 = 11,
+  INTHB_SPECIAL_0_OPCODE_X1 = 5,
+  INTHH_SPECIAL_0_OPCODE_X0 = 12,
+  INTHH_SPECIAL_0_OPCODE_X1 = 6,
+  INTLB_SPECIAL_0_OPCODE_X0 = 13,
+  INTLB_SPECIAL_0_OPCODE_X1 = 7,
+  INTLH_SPECIAL_0_OPCODE_X0 = 14,
+  INTLH_SPECIAL_0_OPCODE_X1 = 8,
+  INV_UN_0_SHUN_0_OPCODE_X1 = 8,
+  IRET_UN_0_SHUN_0_OPCODE_X1 = 9,
+  JALB_OPCODE_X1 = 13,
+  JALF_OPCODE_X1 = 12,
+  JALRP_SPECIAL_0_OPCODE_X1 = 9,
+  JALRR_IMM_1_OPCODE_SN = 3,
+  JALR_RR_IMM_0_OPCODE_SN = 5,
+  JALR_SPECIAL_0_OPCODE_X1 = 10,
+  JB_OPCODE_X1 = 11,
+  JF_OPCODE_X1 = 10,
+  JRP_SPECIAL_0_OPCODE_X1 = 11,
+  JRR_IMM_1_OPCODE_SN = 2,
+  JR_RR_IMM_0_OPCODE_SN = 4,
+  JR_SPECIAL_0_OPCODE_X1 = 12,
+  LBADD_IMM_0_OPCODE_X1 = 22,
+  LBADD_U_IMM_0_OPCODE_X1 = 23,
+  LB_OPCODE_Y2 = 0,
+  LB_UN_0_SHUN_0_OPCODE_X1 = 10,
+  LB_U_OPCODE_Y2 = 1,
+  LB_U_UN_0_SHUN_0_OPCODE_X1 = 11,
+  LHADD_IMM_0_OPCODE_X1 = 24,
+  LHADD_U_IMM_0_OPCODE_X1 = 25,
+  LH_OPCODE_Y2 = 2,
+  LH_UN_0_SHUN_0_OPCODE_X1 = 12,
+  LH_U_OPCODE_Y2 = 3,
+  LH_U_UN_0_SHUN_0_OPCODE_X1 = 13,
+  LNK_SPECIAL_0_OPCODE_X1 = 13,
+  LWADD_IMM_0_OPCODE_X1 = 26,
+  LWADD_NA_IMM_0_OPCODE_X1 = 27,
+  LW_NA_UN_0_SHUN_0_OPCODE_X1 = 24,
+  LW_OPCODE_Y2 = 4,
+  LW_UN_0_SHUN_0_OPCODE_X1 = 14,
+  MAXB_U_SPECIAL_0_OPCODE_X0 = 15,
+  MAXB_U_SPECIAL_0_OPCODE_X1 = 14,
+  MAXH_SPECIAL_0_OPCODE_X0 = 16,
+  MAXH_SPECIAL_0_OPCODE_X1 = 15,
+  MAXIB_U_IMM_0_OPCODE_X0 = 4,
+  MAXIB_U_IMM_0_OPCODE_X1 = 5,
+  MAXIH_IMM_0_OPCODE_X0 = 5,
+  MAXIH_IMM_0_OPCODE_X1 = 6,
+  MFSPR_IMM_0_OPCODE_X1 = 7,
+  MF_UN_0_SHUN_0_OPCODE_X1 = 15,
+  MINB_U_SPECIAL_0_OPCODE_X0 = 17,
+  MINB_U_SPECIAL_0_OPCODE_X1 = 16,
+  MINH_SPECIAL_0_OPCODE_X0 = 18,
+  MINH_SPECIAL_0_OPCODE_X1 = 17,
+  MINIB_U_IMM_0_OPCODE_X0 = 6,
+  MINIB_U_IMM_0_OPCODE_X1 = 8,
+  MINIH_IMM_0_OPCODE_X0 = 7,
+  MINIH_IMM_0_OPCODE_X1 = 9,
+  MM_OPCODE_X0 = 6,
+  MM_OPCODE_X1 = 7,
+  MNZB_SPECIAL_0_OPCODE_X0 = 19,
+  MNZB_SPECIAL_0_OPCODE_X1 = 18,
+  MNZH_SPECIAL_0_OPCODE_X0 = 20,
+  MNZH_SPECIAL_0_OPCODE_X1 = 19,
+  MNZ_SPECIAL_0_OPCODE_X0 = 21,
+  MNZ_SPECIAL_0_OPCODE_X1 = 20,
+  MNZ_SPECIAL_1_OPCODE_Y0 = 0,
+  MNZ_SPECIAL_1_OPCODE_Y1 = 1,
+  MOVEI_IMM_1_OPCODE_SN = 0,
+  MOVE_RR_IMM_0_OPCODE_SN = 8,
+  MTSPR_IMM_0_OPCODE_X1 = 10,
+  MULHHA_SS_SPECIAL_0_OPCODE_X0 = 22,
+  MULHHA_SS_SPECIAL_7_OPCODE_Y0 = 0,
+  MULHHA_SU_SPECIAL_0_OPCODE_X0 = 23,
+  MULHHA_UU_SPECIAL_0_OPCODE_X0 = 24,
+  MULHHA_UU_SPECIAL_7_OPCODE_Y0 = 1,
+  MULHHSA_UU_SPECIAL_0_OPCODE_X0 = 25,
+  MULHH_SS_SPECIAL_0_OPCODE_X0 = 26,
+  MULHH_SS_SPECIAL_6_OPCODE_Y0 = 0,
+  MULHH_SU_SPECIAL_0_OPCODE_X0 = 27,
+  MULHH_UU_SPECIAL_0_OPCODE_X0 = 28,
+  MULHH_UU_SPECIAL_6_OPCODE_Y0 = 1,
+  MULHLA_SS_SPECIAL_0_OPCODE_X0 = 29,
+  MULHLA_SU_SPECIAL_0_OPCODE_X0 = 30,
+  MULHLA_US_SPECIAL_0_OPCODE_X0 = 31,
+  MULHLA_UU_SPECIAL_0_OPCODE_X0 = 32,
+  MULHLSA_UU_SPECIAL_0_OPCODE_X0 = 33,
+  MULHLSA_UU_SPECIAL_5_OPCODE_Y0 = 0,
+  MULHL_SS_SPECIAL_0_OPCODE_X0 = 34,
+  MULHL_SU_SPECIAL_0_OPCODE_X0 = 35,
+  MULHL_US_SPECIAL_0_OPCODE_X0 = 36,
+  MULHL_UU_SPECIAL_0_OPCODE_X0 = 37,
+  MULLLA_SS_SPECIAL_0_OPCODE_X0 = 38,
+  MULLLA_SS_SPECIAL_7_OPCODE_Y0 = 2,
+  MULLLA_SU_SPECIAL_0_OPCODE_X0 = 39,
+  MULLLA_UU_SPECIAL_0_OPCODE_X0 = 40,
+  MULLLA_UU_SPECIAL_7_OPCODE_Y0 = 3,
+  MULLLSA_UU_SPECIAL_0_OPCODE_X0 = 41,
+  MULLL_SS_SPECIAL_0_OPCODE_X0 = 42,
+  MULLL_SS_SPECIAL_6_OPCODE_Y0 = 2,
+  MULLL_SU_SPECIAL_0_OPCODE_X0 = 43,
+  MULLL_UU_SPECIAL_0_OPCODE_X0 = 44,
+  MULLL_UU_SPECIAL_6_OPCODE_Y0 = 3,
+  MVNZ_SPECIAL_0_OPCODE_X0 = 45,
+  MVNZ_SPECIAL_1_OPCODE_Y0 = 1,
+  MVZ_SPECIAL_0_OPCODE_X0 = 46,
+  MVZ_SPECIAL_1_OPCODE_Y0 = 2,
+  MZB_SPECIAL_0_OPCODE_X0 = 47,
+  MZB_SPECIAL_0_OPCODE_X1 = 21,
+  MZH_SPECIAL_0_OPCODE_X0 = 48,
+  MZH_SPECIAL_0_OPCODE_X1 = 22,
+  MZ_SPECIAL_0_OPCODE_X0 = 49,
+  MZ_SPECIAL_0_OPCODE_X1 = 23,
+  MZ_SPECIAL_1_OPCODE_Y0 = 3,
+  MZ_SPECIAL_1_OPCODE_Y1 = 2,
+  NAP_UN_0_SHUN_0_OPCODE_X1 = 16,
+  NOP_NOREG_RR_IMM_0_OPCODE_SN = 2,
+  NOP_UN_0_SHUN_0_OPCODE_X0 = 6,
+  NOP_UN_0_SHUN_0_OPCODE_X1 = 17,
+  NOP_UN_0_SHUN_0_OPCODE_Y0 = 6,
+  NOP_UN_0_SHUN_0_OPCODE_Y1 = 3,
+  NOREG_RR_IMM_0_OPCODE_SN = 0,
+  NOR_SPECIAL_0_OPCODE_X0 = 50,
+  NOR_SPECIAL_0_OPCODE_X1 = 24,
+  NOR_SPECIAL_2_OPCODE_Y0 = 1,
+  NOR_SPECIAL_2_OPCODE_Y1 = 1,
+  ORI_IMM_0_OPCODE_X0 = 8,
+  ORI_IMM_0_OPCODE_X1 = 11,
+  ORI_OPCODE_Y0 = 11,
+  ORI_OPCODE_Y1 = 9,
+  OR_SPECIAL_0_OPCODE_X0 = 51,
+  OR_SPECIAL_0_OPCODE_X1 = 25,
+  OR_SPECIAL_2_OPCODE_Y0 = 2,
+  OR_SPECIAL_2_OPCODE_Y1 = 2,
+  PACKBS_U_SPECIAL_0_OPCODE_X0 = 103,
+  PACKBS_U_SPECIAL_0_OPCODE_X1 = 73,
+  PACKHB_SPECIAL_0_OPCODE_X0 = 52,
+  PACKHB_SPECIAL_0_OPCODE_X1 = 26,
+  PACKHS_SPECIAL_0_OPCODE_X0 = 102,
+  PACKHS_SPECIAL_0_OPCODE_X1 = 72,
+  PACKLB_SPECIAL_0_OPCODE_X0 = 53,
+  PACKLB_SPECIAL_0_OPCODE_X1 = 27,
+  PCNT_UN_0_SHUN_0_OPCODE_X0 = 7,
+  PCNT_UN_0_SHUN_0_OPCODE_Y0 = 7,
+  RLI_SHUN_0_OPCODE_X0 = 1,
+  RLI_SHUN_0_OPCODE_X1 = 1,
+  RLI_SHUN_0_OPCODE_Y0 = 1,
+  RLI_SHUN_0_OPCODE_Y1 = 1,
+  RL_SPECIAL_0_OPCODE_X0 = 54,
+  RL_SPECIAL_0_OPCODE_X1 = 28,
+  RL_SPECIAL_3_OPCODE_Y0 = 0,
+  RL_SPECIAL_3_OPCODE_Y1 = 0,
+  RR_IMM_0_OPCODE_SN = 0,
+  S1A_SPECIAL_0_OPCODE_X0 = 55,
+  S1A_SPECIAL_0_OPCODE_X1 = 29,
+  S1A_SPECIAL_0_OPCODE_Y0 = 1,
+  S1A_SPECIAL_0_OPCODE_Y1 = 1,
+  S2A_SPECIAL_0_OPCODE_X0 = 56,
+  S2A_SPECIAL_0_OPCODE_X1 = 30,
+  S2A_SPECIAL_0_OPCODE_Y0 = 2,
+  S2A_SPECIAL_0_OPCODE_Y1 = 2,
+  S3A_SPECIAL_0_OPCODE_X0 = 57,
+  S3A_SPECIAL_0_OPCODE_X1 = 31,
+  S3A_SPECIAL_5_OPCODE_Y0 = 1,
+  S3A_SPECIAL_5_OPCODE_Y1 = 1,
+  SADAB_U_SPECIAL_0_OPCODE_X0 = 58,
+  SADAH_SPECIAL_0_OPCODE_X0 = 59,
+  SADAH_U_SPECIAL_0_OPCODE_X0 = 60,
+  SADB_U_SPECIAL_0_OPCODE_X0 = 61,
+  SADH_SPECIAL_0_OPCODE_X0 = 62,
+  SADH_U_SPECIAL_0_OPCODE_X0 = 63,
+  SBADD_IMM_0_OPCODE_X1 = 28,
+  SB_OPCODE_Y2 = 5,
+  SB_SPECIAL_0_OPCODE_X1 = 32,
+  SEQB_SPECIAL_0_OPCODE_X0 = 64,
+  SEQB_SPECIAL_0_OPCODE_X1 = 33,
+  SEQH_SPECIAL_0_OPCODE_X0 = 65,
+  SEQH_SPECIAL_0_OPCODE_X1 = 34,
+  SEQIB_IMM_0_OPCODE_X0 = 9,
+  SEQIB_IMM_0_OPCODE_X1 = 12,
+  SEQIH_IMM_0_OPCODE_X0 = 10,
+  SEQIH_IMM_0_OPCODE_X1 = 13,
+  SEQI_IMM_0_OPCODE_X0 = 11,
+  SEQI_IMM_0_OPCODE_X1 = 14,
+  SEQI_OPCODE_Y0 = 12,
+  SEQI_OPCODE_Y1 = 10,
+  SEQ_SPECIAL_0_OPCODE_X0 = 66,
+  SEQ_SPECIAL_0_OPCODE_X1 = 35,
+  SEQ_SPECIAL_5_OPCODE_Y0 = 2,
+  SEQ_SPECIAL_5_OPCODE_Y1 = 2,
+  SHADD_IMM_0_OPCODE_X1 = 29,
+  SHL8II_IMM_0_OPCODE_SN = 3,
+  SHLB_SPECIAL_0_OPCODE_X0 = 67,
+  SHLB_SPECIAL_0_OPCODE_X1 = 36,
+  SHLH_SPECIAL_0_OPCODE_X0 = 68,
+  SHLH_SPECIAL_0_OPCODE_X1 = 37,
+  SHLIB_SHUN_0_OPCODE_X0 = 2,
+  SHLIB_SHUN_0_OPCODE_X1 = 2,
+  SHLIH_SHUN_0_OPCODE_X0 = 3,
+  SHLIH_SHUN_0_OPCODE_X1 = 3,
+  SHLI_SHUN_0_OPCODE_X0 = 4,
+  SHLI_SHUN_0_OPCODE_X1 = 4,
+  SHLI_SHUN_0_OPCODE_Y0 = 2,
+  SHLI_SHUN_0_OPCODE_Y1 = 2,
+  SHL_SPECIAL_0_OPCODE_X0 = 69,
+  SHL_SPECIAL_0_OPCODE_X1 = 38,
+  SHL_SPECIAL_3_OPCODE_Y0 = 1,
+  SHL_SPECIAL_3_OPCODE_Y1 = 1,
+  SHR1_RR_IMM_0_OPCODE_SN = 9,
+  SHRB_SPECIAL_0_OPCODE_X0 = 70,
+  SHRB_SPECIAL_0_OPCODE_X1 = 39,
+  SHRH_SPECIAL_0_OPCODE_X0 = 71,
+  SHRH_SPECIAL_0_OPCODE_X1 = 40,
+  SHRIB_SHUN_0_OPCODE_X0 = 5,
+  SHRIB_SHUN_0_OPCODE_X1 = 5,
+  SHRIH_SHUN_0_OPCODE_X0 = 6,
+  SHRIH_SHUN_0_OPCODE_X1 = 6,
+  SHRI_SHUN_0_OPCODE_X0 = 7,
+  SHRI_SHUN_0_OPCODE_X1 = 7,
+  SHRI_SHUN_0_OPCODE_Y0 = 3,
+  SHRI_SHUN_0_OPCODE_Y1 = 3,
+  SHR_SPECIAL_0_OPCODE_X0 = 72,
+  SHR_SPECIAL_0_OPCODE_X1 = 41,
+  SHR_SPECIAL_3_OPCODE_Y0 = 2,
+  SHR_SPECIAL_3_OPCODE_Y1 = 2,
+  SHUN_0_OPCODE_X0 = 7,
+  SHUN_0_OPCODE_X1 = 8,
+  SHUN_0_OPCODE_Y0 = 13,
+  SHUN_0_OPCODE_Y1 = 11,
+  SH_OPCODE_Y2 = 6,
+  SH_SPECIAL_0_OPCODE_X1 = 42,
+  SLTB_SPECIAL_0_OPCODE_X0 = 73,
+  SLTB_SPECIAL_0_OPCODE_X1 = 43,
+  SLTB_U_SPECIAL_0_OPCODE_X0 = 74,
+  SLTB_U_SPECIAL_0_OPCODE_X1 = 44,
+  SLTEB_SPECIAL_0_OPCODE_X0 = 75,
+  SLTEB_SPECIAL_0_OPCODE_X1 = 45,
+  SLTEB_U_SPECIAL_0_OPCODE_X0 = 76,
+  SLTEB_U_SPECIAL_0_OPCODE_X1 = 46,
+  SLTEH_SPECIAL_0_OPCODE_X0 = 77,
+  SLTEH_SPECIAL_0_OPCODE_X1 = 47,
+  SLTEH_U_SPECIAL_0_OPCODE_X0 = 78,
+  SLTEH_U_SPECIAL_0_OPCODE_X1 = 48,
+  SLTE_SPECIAL_0_OPCODE_X0 = 79,
+  SLTE_SPECIAL_0_OPCODE_X1 = 49,
+  SLTE_SPECIAL_4_OPCODE_Y0 = 0,
+  SLTE_SPECIAL_4_OPCODE_Y1 = 0,
+  SLTE_U_SPECIAL_0_OPCODE_X0 = 80,
+  SLTE_U_SPECIAL_0_OPCODE_X1 = 50,
+  SLTE_U_SPECIAL_4_OPCODE_Y0 = 1,
+  SLTE_U_SPECIAL_4_OPCODE_Y1 = 1,
+  SLTH_SPECIAL_0_OPCODE_X0 = 81,
+  SLTH_SPECIAL_0_OPCODE_X1 = 51,
+  SLTH_U_SPECIAL_0_OPCODE_X0 = 82,
+  SLTH_U_SPECIAL_0_OPCODE_X1 = 52,
+  SLTIB_IMM_0_OPCODE_X0 = 12,
+  SLTIB_IMM_0_OPCODE_X1 = 15,
+  SLTIB_U_IMM_0_OPCODE_X0 = 13,
+  SLTIB_U_IMM_0_OPCODE_X1 = 16,
+  SLTIH_IMM_0_OPCODE_X0 = 14,
+  SLTIH_IMM_0_OPCODE_X1 = 17,
+  SLTIH_U_IMM_0_OPCODE_X0 = 15,
+  SLTIH_U_IMM_0_OPCODE_X1 = 18,
+  SLTI_IMM_0_OPCODE_X0 = 16,
+  SLTI_IMM_0_OPCODE_X1 = 19,
+  SLTI_OPCODE_Y0 = 14,
+  SLTI_OPCODE_Y1 = 12,
+  SLTI_U_IMM_0_OPCODE_X0 = 17,
+  SLTI_U_IMM_0_OPCODE_X1 = 20,
+  SLTI_U_OPCODE_Y0 = 15,
+  SLTI_U_OPCODE_Y1 = 13,
+  SLT_SPECIAL_0_OPCODE_X0 = 83,
+  SLT_SPECIAL_0_OPCODE_X1 = 53,
+  SLT_SPECIAL_4_OPCODE_Y0 = 2,
+  SLT_SPECIAL_4_OPCODE_Y1 = 2,
+  SLT_U_SPECIAL_0_OPCODE_X0 = 84,
+  SLT_U_SPECIAL_0_OPCODE_X1 = 54,
+  SLT_U_SPECIAL_4_OPCODE_Y0 = 3,
+  SLT_U_SPECIAL_4_OPCODE_Y1 = 3,
+  SNEB_SPECIAL_0_OPCODE_X0 = 85,
+  SNEB_SPECIAL_0_OPCODE_X1 = 55,
+  SNEH_SPECIAL_0_OPCODE_X0 = 86,
+  SNEH_SPECIAL_0_OPCODE_X1 = 56,
+  SNE_SPECIAL_0_OPCODE_X0 = 87,
+  SNE_SPECIAL_0_OPCODE_X1 = 57,
+  SNE_SPECIAL_5_OPCODE_Y0 = 3,
+  SNE_SPECIAL_5_OPCODE_Y1 = 3,
+  SPECIAL_0_OPCODE_X0 = 0,
+  SPECIAL_0_OPCODE_X1 = 1,
+  SPECIAL_0_OPCODE_Y0 = 1,
+  SPECIAL_0_OPCODE_Y1 = 1,
+  SPECIAL_1_OPCODE_Y0 = 2,
+  SPECIAL_1_OPCODE_Y1 = 2,
+  SPECIAL_2_OPCODE_Y0 = 3,
+  SPECIAL_2_OPCODE_Y1 = 3,
+  SPECIAL_3_OPCODE_Y0 = 4,
+  SPECIAL_3_OPCODE_Y1 = 4,
+  SPECIAL_4_OPCODE_Y0 = 5,
+  SPECIAL_4_OPCODE_Y1 = 5,
+  SPECIAL_5_OPCODE_Y0 = 6,
+  SPECIAL_5_OPCODE_Y1 = 6,
+  SPECIAL_6_OPCODE_Y0 = 7,
+  SPECIAL_7_OPCODE_Y0 = 8,
+  SRAB_SPECIAL_0_OPCODE_X0 = 88,
+  SRAB_SPECIAL_0_OPCODE_X1 = 58,
+  SRAH_SPECIAL_0_OPCODE_X0 = 89,
+  SRAH_SPECIAL_0_OPCODE_X1 = 59,
+  SRAIB_SHUN_0_OPCODE_X0 = 8,
+  SRAIB_SHUN_0_OPCODE_X1 = 8,
+  SRAIH_SHUN_0_OPCODE_X0 = 9,
+  SRAIH_SHUN_0_OPCODE_X1 = 9,
+  SRAI_SHUN_0_OPCODE_X0 = 10,
+  SRAI_SHUN_0_OPCODE_X1 = 10,
+  SRAI_SHUN_0_OPCODE_Y0 = 4,
+  SRAI_SHUN_0_OPCODE_Y1 = 4,
+  SRA_SPECIAL_0_OPCODE_X0 = 90,
+  SRA_SPECIAL_0_OPCODE_X1 = 60,
+  SRA_SPECIAL_3_OPCODE_Y0 = 3,
+  SRA_SPECIAL_3_OPCODE_Y1 = 3,
+  SUBBS_U_SPECIAL_0_OPCODE_X0 = 100,
+  SUBBS_U_SPECIAL_0_OPCODE_X1 = 70,
+  SUBB_SPECIAL_0_OPCODE_X0 = 91,
+  SUBB_SPECIAL_0_OPCODE_X1 = 61,
+  SUBHS_SPECIAL_0_OPCODE_X0 = 101,
+  SUBHS_SPECIAL_0_OPCODE_X1 = 71,
+  SUBH_SPECIAL_0_OPCODE_X0 = 92,
+  SUBH_SPECIAL_0_OPCODE_X1 = 62,
+  SUBS_SPECIAL_0_OPCODE_X0 = 97,
+  SUBS_SPECIAL_0_OPCODE_X1 = 67,
+  SUB_SPECIAL_0_OPCODE_X0 = 93,
+  SUB_SPECIAL_0_OPCODE_X1 = 63,
+  SUB_SPECIAL_0_OPCODE_Y0 = 3,
+  SUB_SPECIAL_0_OPCODE_Y1 = 3,
+  SWADD_IMM_0_OPCODE_X1 = 30,
+  SWINT0_UN_0_SHUN_0_OPCODE_X1 = 18,
+  SWINT1_UN_0_SHUN_0_OPCODE_X1 = 19,
+  SWINT2_UN_0_SHUN_0_OPCODE_X1 = 20,
+  SWINT3_UN_0_SHUN_0_OPCODE_X1 = 21,
+  SW_OPCODE_Y2 = 7,
+  SW_SPECIAL_0_OPCODE_X1 = 64,
+  TBLIDXB0_UN_0_SHUN_0_OPCODE_X0 = 8,
+  TBLIDXB0_UN_0_SHUN_0_OPCODE_Y0 = 8,
+  TBLIDXB1_UN_0_SHUN_0_OPCODE_X0 = 9,
+  TBLIDXB1_UN_0_SHUN_0_OPCODE_Y0 = 9,
+  TBLIDXB2_UN_0_SHUN_0_OPCODE_X0 = 10,
+  TBLIDXB2_UN_0_SHUN_0_OPCODE_Y0 = 10,
+  TBLIDXB3_UN_0_SHUN_0_OPCODE_X0 = 11,
+  TBLIDXB3_UN_0_SHUN_0_OPCODE_Y0 = 11,
+  TNS_UN_0_SHUN_0_OPCODE_X1 = 22,
+  UN_0_SHUN_0_OPCODE_X0 = 11,
+  UN_0_SHUN_0_OPCODE_X1 = 11,
+  UN_0_SHUN_0_OPCODE_Y0 = 5,
+  UN_0_SHUN_0_OPCODE_Y1 = 5,
+  WH64_UN_0_SHUN_0_OPCODE_X1 = 23,
+  XORI_IMM_0_OPCODE_X0 = 2,
+  XORI_IMM_0_OPCODE_X1 = 21,
+  XOR_SPECIAL_0_OPCODE_X0 = 94,
+  XOR_SPECIAL_0_OPCODE_X1 = 65,
+  XOR_SPECIAL_2_OPCODE_Y0 = 3,
+  XOR_SPECIAL_2_OPCODE_Y1 = 3
+};
+
+
+#endif /* __ASSEMBLER__ */
+
+#endif /* __ARCH_OPCODE_H__ */
diff --git a/kernel/arch/tile/include/uapi/arch/sim.h b/kernel/arch/tile/include/uapi/arch/sim.h
new file mode 100644
index 000000000..e54b7b052
--- /dev/null
+++ b/kernel/arch/tile/include/uapi/arch/sim.h
@@ -0,0 +1,643 @@
+/*
+ * Copyright 2010 Tilera Corporation. All Rights Reserved.
+ *
+ *   This program is free software; you can redistribute it and/or
+ *   modify it under the terms of the GNU General Public License
+ *   as published by the Free Software Foundation, version 2.
+ *
+ *   This program is distributed in the hope that it will be useful, but
+ *   WITHOUT ANY WARRANTY; without even the implied warranty of
+ *   MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
+ *   NON INFRINGEMENT.  See the GNU General Public License for
+ *   more details.
+ */
+
+/**
+ * @file
+ *
+ * Provides an API for controlling the simulator at runtime.
+ */
+
+/**
+ * @addtogroup arch_sim
+ * @{
+ *
+ * An API for controlling the simulator at runtime.
+ *
+ * The simulator's behavior can be modified while it is running.
+ * For example, human-readable trace output can be enabled and disabled
+ * around code of interest.
+ *
+ * There are two ways to modify simulator behavior:
+ * programmatically, by calling various sim_* functions, and
+ * interactively, by entering commands like "sim set functional true"
+ * at the tile-monitor prompt.  Typing "sim help" at that prompt provides
+ * a list of interactive commands.
+ *
+ * All interactive commands can also be executed programmatically by
+ * passing a string to the sim_command function.
+ */
+
+#ifndef __ARCH_SIM_H__
+#define __ARCH_SIM_H__
+
+#include <arch/sim_def.h>
+#include <arch/abi.h>
+
+#ifndef __ASSEMBLER__
+
+#include <arch/spr_def.h>
+
+
+/**
+ * Return true if the current program is running under a simulator,
+ * rather than on real hardware.  If running on hardware, other "sim_xxx()"
+ * calls have no useful effect.
+ */
+static inline int
+sim_is_simulator(void)
+{
+  return __insn_mfspr(SPR_SIM_CONTROL) != 0;
+}
+
+
+/**
+ * Checkpoint the simulator state to a checkpoint file.
+ *
+ * The checkpoint file name is either the default or the name specified
+ * on the command line with "--checkpoint-file".
+ */
+static __inline void
+sim_checkpoint(void)
+{
+  __insn_mtspr(SPR_SIM_CONTROL, SIM_CONTROL_CHECKPOINT);
+}
+
+
+/**
+ * Report whether or not various kinds of simulator tracing are enabled.
+ *
+ * @return The bitwise OR of these values:
+ *
+ * SIM_TRACE_CYCLES (--trace-cycles),
+ * SIM_TRACE_ROUTER (--trace-router),
+ * SIM_TRACE_REGISTER_WRITES (--trace-register-writes),
+ * SIM_TRACE_DISASM (--trace-disasm),
+ * SIM_TRACE_STALL_INFO (--trace-stall-info)
+ * SIM_TRACE_MEMORY_CONTROLLER (--trace-memory-controller)
+ * SIM_TRACE_L2_CACHE (--trace-l2)
+ * SIM_TRACE_LINES (--trace-lines)
+ */
+static __inline unsigned int
+sim_get_tracing(void)
+{
+  return __insn_mfspr(SPR_SIM_CONTROL) & SIM_TRACE_FLAG_MASK;
+}
+
+
+/**
+ * Turn on or off different kinds of simulator tracing.
+ *
+ * @param mask Either one of these special values:
+ *
+ * SIM_TRACE_NONE (turns off tracing),
+ * SIM_TRACE_ALL (turns on all possible tracing).
+ *
+ * or the bitwise OR of these values:
+ *
+ * SIM_TRACE_CYCLES (--trace-cycles),
+ * SIM_TRACE_ROUTER (--trace-router),
+ * SIM_TRACE_REGISTER_WRITES (--trace-register-writes),
+ * SIM_TRACE_DISASM (--trace-disasm),
+ * SIM_TRACE_STALL_INFO (--trace-stall-info)
+ * SIM_TRACE_MEMORY_CONTROLLER (--trace-memory-controller)
+ * SIM_TRACE_L2_CACHE (--trace-l2)
+ * SIM_TRACE_LINES (--trace-lines)
+ */
+static __inline void
+sim_set_tracing(unsigned int mask)
+{
+  __insn_mtspr(SPR_SIM_CONTROL, SIM_TRACE_SPR_ARG(mask));
+}
+
+
+/**
+ * Request dumping of different kinds of simulator state.
+ *
+ * @param mask Either this special value:
+ *
+ * SIM_DUMP_ALL (dump all known state)
+ *
+ * or the bitwise OR of these values:
+ *
+ * SIM_DUMP_REGS (the register file),
+ * SIM_DUMP_SPRS (the SPRs),
+ * SIM_DUMP_ITLB (the iTLB),
+ * SIM_DUMP_DTLB (the dTLB),
+ * SIM_DUMP_L1I (the L1 I-cache),
+ * SIM_DUMP_L1D (the L1 D-cache),
+ * SIM_DUMP_L2 (the L2 cache),
+ * SIM_DUMP_SNREGS (the switch register file),
+ * SIM_DUMP_SNITLB (the switch iTLB),
+ * SIM_DUMP_SNL1I (the switch L1 I-cache),
+ * SIM_DUMP_BACKTRACE (the current backtrace)
+ */
+static __inline void
+sim_dump(unsigned int mask)
+{
+  __insn_mtspr(SPR_SIM_CONTROL, SIM_DUMP_SPR_ARG(mask));
+}
+
+
+/**
+ * Print a string to the simulator stdout.
+ *
+ * @param str The string to be written.
+ */
+static __inline void
+sim_print(const char* str)
+{
+  for ( ; *str != '\0'; str++)
+  {
+    __insn_mtspr(SPR_SIM_CONTROL, SIM_CONTROL_PUTC |
+                 (*str << _SIM_CONTROL_OPERATOR_BITS));
+  }
+  __insn_mtspr(SPR_SIM_CONTROL, SIM_CONTROL_PUTC |
+               (SIM_PUTC_FLUSH_BINARY << _SIM_CONTROL_OPERATOR_BITS));
+}
+
+
+/**
+ * Print a string to the simulator stdout.
+ *
+ * @param str The string to be written (a newline is automatically added).
+ */
+static __inline void
+sim_print_string(const char* str)
+{
+  for ( ; *str != '\0'; str++)
+  {
+    __insn_mtspr(SPR_SIM_CONTROL, SIM_CONTROL_PUTC |
+                 (*str << _SIM_CONTROL_OPERATOR_BITS));
+  }
+  __insn_mtspr(SPR_SIM_CONTROL, SIM_CONTROL_PUTC |
+               (SIM_PUTC_FLUSH_STRING << _SIM_CONTROL_OPERATOR_BITS));
+}
+
+
+/**
+ * Execute a simulator command string.
+ *
+ * Type 'sim help' at the tile-monitor prompt to learn what commands
+ * are available.  Note the use of the tile-monitor "sim" command to
+ * pass commands to the simulator.
+ *
+ * The argument to sim_command() does not include the leading "sim"
+ * prefix used at the tile-monitor prompt; for example, you might call
+ * sim_command("trace disasm").
+ */
+static __inline void
+sim_command(const char* str)
+{
+  int c;
+  do
+  {
+    c = *str++;
+    __insn_mtspr(SPR_SIM_CONTROL, SIM_CONTROL_COMMAND |
+                 (c << _SIM_CONTROL_OPERATOR_BITS));
+  }
+  while (c);
+}
+
+
+
+#ifndef __DOXYGEN__
+
+/**
+ * The underlying implementation of "_sim_syscall()".
+ *
+ * We use extra "and" instructions to ensure that all the values
+ * we are passing to the simulator are actually valid in the registers
+ * (i.e. returned from memory) prior to the SIM_CONTROL spr.
+ */
+static __inline long _sim_syscall0(int val)
+{
+  long result;
+  __asm__ __volatile__ ("mtspr SIM_CONTROL, r0"
+                        : "=R00" (result) : "R00" (val));
+  return result;
+}
+
+static __inline long _sim_syscall1(int val, long arg1)
+{
+  long result;
+  __asm__ __volatile__ ("{ and zero, r1, r1; mtspr SIM_CONTROL, r0 }"
+                        : "=R00" (result) : "R00" (val), "R01" (arg1));
+  return result;
+}
+
+static __inline long _sim_syscall2(int val, long arg1, long arg2)
+{
+  long result;
+  __asm__ __volatile__ ("{ and zero, r1, r2; mtspr SIM_CONTROL, r0 }"
+                        : "=R00" (result)
+                        : "R00" (val), "R01" (arg1), "R02" (arg2));
+  return result;
+}
+
+/* Note that _sim_syscall3() and higher are technically at risk of
+   receiving an interrupt right before the mtspr bundle, in which case
+   the register values for arguments 3 and up may still be in flight
+   to the core from a stack frame reload. */
+
+static __inline long _sim_syscall3(int val, long arg1, long arg2, long arg3)
+{
+  long result;
+  __asm__ __volatile__ ("{ and zero, r3, r3 };"
+                        "{ and zero, r1, r2; mtspr SIM_CONTROL, r0 }"
+                        : "=R00" (result)
+                        : "R00" (val), "R01" (arg1), "R02" (arg2),
+                          "R03" (arg3));
+  return result;
+}
+
+static __inline long _sim_syscall4(int val, long arg1, long arg2, long arg3,
+                                  long arg4)
+{
+  long result;
+  __asm__ __volatile__ ("{ and zero, r3, r4 };"
+                        "{ and zero, r1, r2; mtspr SIM_CONTROL, r0 }"
+                        : "=R00" (result)
+                        : "R00" (val), "R01" (arg1), "R02" (arg2),
+                          "R03" (arg3), "R04" (arg4));
+  return result;
+}
+
+static __inline long _sim_syscall5(int val, long arg1, long arg2, long arg3,
+                                  long arg4, long arg5)
+{
+  long result;
+  __asm__ __volatile__ ("{ and zero, r3, r4; and zero, r5, r5 };"
+                        "{ and zero, r1, r2; mtspr SIM_CONTROL, r0 }"
+                        : "=R00" (result)
+                        : "R00" (val), "R01" (arg1), "R02" (arg2),
+                          "R03" (arg3), "R04" (arg4), "R05" (arg5));
+  return result;
+}
+
+/**
+ * Make a special syscall to the simulator itself, if running under
+ * simulation. This is used as the implementation of other functions
+ * and should not be used outside this file.
+ *
+ * @param syscall_num The simulator syscall number.
+ * @param nr The number of additional arguments provided.
+ *
+ * @return Varies by syscall.
+ */
+#define _sim_syscall(syscall_num, nr, args...) \
+  _sim_syscall##nr( \
+    ((syscall_num) << _SIM_CONTROL_OPERATOR_BITS) | SIM_CONTROL_SYSCALL, \
+    ##args)
+
+
+/* Values for the "access_mask" parameters below. */
+#define SIM_WATCHPOINT_READ    1
+#define SIM_WATCHPOINT_WRITE   2
+#define SIM_WATCHPOINT_EXECUTE 4
+
+
+static __inline int
+sim_add_watchpoint(unsigned int process_id,
+                   unsigned long address,
+                   unsigned long size,
+                   unsigned int access_mask,
+                   unsigned long user_data)
+{
+  return _sim_syscall(SIM_SYSCALL_ADD_WATCHPOINT, 5, process_id,
+                     address, size, access_mask, user_data);
+}
+
+
+static __inline int
+sim_remove_watchpoint(unsigned int process_id,
+                      unsigned long address,
+                      unsigned long size,
+                      unsigned int access_mask,
+                      unsigned long user_data)
+{
+  return _sim_syscall(SIM_SYSCALL_REMOVE_WATCHPOINT, 5, process_id,
+                     address, size, access_mask, user_data);
+}
+
+
+/**
+ * Return value from sim_query_watchpoint.
+ */
+struct SimQueryWatchpointStatus
+{
+  /**
+   * 0 if a watchpoint fired, 1 if no watchpoint fired, or -1 for
+   * error (meaning a bad process_id).
+   */
+  int syscall_status;
+
+  /**
+   * The address of the watchpoint that fired (this is the address
+   * passed to sim_add_watchpoint, not an address within that range
+   * that actually triggered the watchpoint).
+   */
+  unsigned long address;
+
+  /** The arbitrary user_data installed by sim_add_watchpoint. */
+  unsigned long user_data;
+};
+
+
+static __inline struct SimQueryWatchpointStatus
+sim_query_watchpoint(unsigned int process_id)
+{
+  struct SimQueryWatchpointStatus status;
+  long val = SIM_CONTROL_SYSCALL |
+    (SIM_SYSCALL_QUERY_WATCHPOINT << _SIM_CONTROL_OPERATOR_BITS);
+  __asm__ __volatile__ ("{ and zero, r1, r1; mtspr SIM_CONTROL, r0 }"
+                        : "=R00" (status.syscall_status),
+                          "=R01" (status.address),
+                          "=R02" (status.user_data)
+                        : "R00" (val), "R01" (process_id));
+  return status;
+}
+
+
+/* On the simulator, confirm lines have been evicted everywhere. */
+static __inline void
+sim_validate_lines_evicted(unsigned long long pa, unsigned long length)
+{
+#ifdef __LP64__
+  _sim_syscall(SIM_SYSCALL_VALIDATE_LINES_EVICTED, 2, pa, length);
+#else
+  _sim_syscall(SIM_SYSCALL_VALIDATE_LINES_EVICTED, 4,
+               0 /* dummy */, (long)(pa), (long)(pa >> 32), length);
+#endif
+}
+
+
+/* Return the current CPU speed in cycles per second. */
+static __inline long
+sim_query_cpu_speed(void)
+{
+  return _sim_syscall(SIM_SYSCALL_QUERY_CPU_SPEED, 0);
+}
+
+#endif /* !__DOXYGEN__ */
+
+
+
+
+/**
+ * Modify the shaping parameters of a shim.
+ *
+ * @param shim The shim to modify. One of:
+ *   SIM_CONTROL_SHAPING_GBE_0
+ *   SIM_CONTROL_SHAPING_GBE_1
+ *   SIM_CONTROL_SHAPING_GBE_2
+ *   SIM_CONTROL_SHAPING_GBE_3
+ *   SIM_CONTROL_SHAPING_XGBE_0
+ *   SIM_CONTROL_SHAPING_XGBE_1
+ *
+ * @param type The type of shaping. This should be the same type of
+ * shaping that is already in place on the shim. One of:
+ *   SIM_CONTROL_SHAPING_MULTIPLIER
+ *   SIM_CONTROL_SHAPING_PPS
+ *   SIM_CONTROL_SHAPING_BPS
+ *
+ * @param units The magnitude of the rate. One of:
+ *   SIM_CONTROL_SHAPING_UNITS_SINGLE
+ *   SIM_CONTROL_SHAPING_UNITS_KILO
+ *   SIM_CONTROL_SHAPING_UNITS_MEGA
+ *   SIM_CONTROL_SHAPING_UNITS_GIGA
+ *
+ * @param rate The rate to which to change it. This must fit in
+ * SIM_CONTROL_SHAPING_RATE_BITS bits or a warning is issued and
+ * the shaping is not changed.
+ *
+ * @return 0 if no problems were detected in the arguments to sim_set_shaping
+ * or 1 if problems were detected (for example, rate does not fit in 17 bits).
+ */
+static __inline int
+sim_set_shaping(unsigned shim,
+                unsigned type,
+                unsigned units,
+                unsigned rate)
+{
+  if ((rate & ~((1 << SIM_CONTROL_SHAPING_RATE_BITS) - 1)) != 0)
+    return 1;
+
+  __insn_mtspr(SPR_SIM_CONTROL, SIM_SHAPING_SPR_ARG(shim, type, units, rate));
+  return 0;
+}
+
+#ifdef __tilegx__
+
+/** Enable a set of mPIPE links.  Pass a -1 link_mask to enable all links. */
+static __inline void
+sim_enable_mpipe_links(unsigned mpipe, unsigned long link_mask)
+{
+  __insn_mtspr(SPR_SIM_CONTROL,
+               (SIM_CONTROL_ENABLE_MPIPE_LINK_MAGIC_BYTE |
+                (mpipe << 8) | (1 << 16) | ((uint_reg_t)link_mask << 32)));
+}
+
+/** Disable a set of mPIPE links.  Pass a -1 link_mask to disable all links. */
+static __inline void
+sim_disable_mpipe_links(unsigned mpipe, unsigned long link_mask)
+{
+  __insn_mtspr(SPR_SIM_CONTROL,
+               (SIM_CONTROL_ENABLE_MPIPE_LINK_MAGIC_BYTE |
+                (mpipe << 8) | (0 << 16) | ((uint_reg_t)link_mask << 32)));
+}
+
+#endif /* __tilegx__ */
+
+
+/*
+ * An API for changing "functional" mode.
+ */
+
+#ifndef __DOXYGEN__
+
+#define sim_enable_functional() \
+  __insn_mtspr(SPR_SIM_CONTROL, SIM_CONTROL_ENABLE_FUNCTIONAL)
+
+#define sim_disable_functional() \
+  __insn_mtspr(SPR_SIM_CONTROL, SIM_CONTROL_DISABLE_FUNCTIONAL)
+
+#endif /* __DOXYGEN__ */
+
+
+/*
+ * Profiler support.
+ */
+
+/**
+ * Turn profiling on for the current task.
+ *
+ * Note that this has no effect if run in an environment without
+ * profiling support (thus, the proper flags to the simulator must
+ * be supplied).
+ */
+static __inline void
+sim_profiler_enable(void)
+{
+  __insn_mtspr(SPR_SIM_CONTROL, SIM_CONTROL_PROFILER_ENABLE);
+}
+
+
+/** Turn profiling off for the current task. */
+static __inline void
+sim_profiler_disable(void)
+{
+  __insn_mtspr(SPR_SIM_CONTROL, SIM_CONTROL_PROFILER_DISABLE);
+}
+
+
+/**
+ * Turn profiling on or off for the current task.
+ *
+ * @param enabled If true, turns on profiling. If false, turns it off.
+ *
+ * Note that this has no effect if run in an environment without
+ * profiling support (thus, the proper flags to the simulator must
+ * be supplied).
+ */
+static __inline void
+sim_profiler_set_enabled(int enabled)
+{
+  int val =
+    enabled ? SIM_CONTROL_PROFILER_ENABLE : SIM_CONTROL_PROFILER_DISABLE;
+  __insn_mtspr(SPR_SIM_CONTROL, val);
+}
+
+
+/**
+ * Return true if and only if profiling is currently enabled
+ * for the current task.
+ *
+ * This returns false even if sim_profiler_enable() was called
+ * if the current execution environment does not support profiling.
+ */
+static __inline int
+sim_profiler_is_enabled(void)
+{
+  return ((__insn_mfspr(SPR_SIM_CONTROL) & SIM_PROFILER_ENABLED_MASK) != 0);
+}
+
+
+/**
+ * Reset profiling counters to zero for the current task.
+ *
+ * Resetting can be done while profiling is enabled.  It does not affect
+ * the chip-wide profiling counters.
+ */
+static __inline void
+sim_profiler_clear(void)
+{
+  __insn_mtspr(SPR_SIM_CONTROL, SIM_CONTROL_PROFILER_CLEAR);
+}
+
+
+/**
+ * Enable specified chip-level profiling counters.
+ *
+ * Does not affect the per-task profiling counters.
+ *
+ * @param mask Either this special value:
+ *
+ * SIM_CHIP_ALL (enables all chip-level components).
+ *
+ * or the bitwise OR of these values:
+ *
+ * SIM_CHIP_MEMCTL (enable all memory controllers)
+ * SIM_CHIP_XAUI (enable all XAUI controllers)
+ * SIM_CHIP_MPIPE (enable all MPIPE controllers)
+ */
+static __inline void
+sim_profiler_chip_enable(unsigned int mask)
+{
+  __insn_mtspr(SPR_SIM_CONTROL, SIM_PROFILER_CHIP_ENABLE_SPR_ARG(mask));
+}
+
+
+/**
+ * Disable specified chip-level profiling counters.
+ *
+ * Does not affect the per-task profiling counters.
+ *
+ * @param mask Either this special value:
+ *
+ * SIM_CHIP_ALL (disables all chip-level components).
+ *
+ * or the bitwise OR of these values:
+ *
+ * SIM_CHIP_MEMCTL (disable all memory controllers)
+ * SIM_CHIP_XAUI (disable all XAUI controllers)
+ * SIM_CHIP_MPIPE (disable all MPIPE controllers)
+ */
+static __inline void
+sim_profiler_chip_disable(unsigned int mask)
+{
+  __insn_mtspr(SPR_SIM_CONTROL, SIM_PROFILER_CHIP_DISABLE_SPR_ARG(mask));
+}
+
+
+/**
+ * Reset specified chip-level profiling counters to zero.
+ *
+ * Does not affect the per-task profiling counters.
+ *
+ * @param mask Either this special value:
+ *
+ * SIM_CHIP_ALL (clears all chip-level components).
+ *
+ * or the bitwise OR of these values:
+ *
+ * SIM_CHIP_MEMCTL (clear all memory controllers)
+ * SIM_CHIP_XAUI (clear all XAUI controllers)
+ * SIM_CHIP_MPIPE (clear all MPIPE controllers)
+ */
+static __inline void
+sim_profiler_chip_clear(unsigned int mask)
+{
+  __insn_mtspr(SPR_SIM_CONTROL, SIM_PROFILER_CHIP_CLEAR_SPR_ARG(mask));
+}
+
+
+/*
+ * Event support.
+ */
+
+#ifndef __DOXYGEN__
+
+static __inline void
+sim_event_begin(unsigned int x)
+{
+#if defined(__tile__) && !defined(__NO_EVENT_SPR__)
+  __insn_mtspr(SPR_EVENT_BEGIN, x);
+#endif
+}
+
+static __inline void
+sim_event_end(unsigned int x)
+{
+#if defined(__tile__) && !defined(__NO_EVENT_SPR__)
+  __insn_mtspr(SPR_EVENT_END, x);
+#endif
+}
+
+#endif /* !__DOXYGEN__ */
+
+#endif /* !__ASSEMBLER__ */
+
+#endif /* !__ARCH_SIM_H__ */
+
+/** @} */
diff --git a/kernel/arch/tile/include/uapi/arch/sim_def.h b/kernel/arch/tile/include/uapi/arch/sim_def.h
new file mode 100644
index 000000000..1c069537a
--- /dev/null
+++ b/kernel/arch/tile/include/uapi/arch/sim_def.h
@@ -0,0 +1,505 @@
+/*
+ * Copyright 2010 Tilera Corporation. All Rights Reserved.
+ *
+ *   This program is free software; you can redistribute it and/or
+ *   modify it under the terms of the GNU General Public License
+ *   as published by the Free Software Foundation, version 2.
+ *
+ *   This program is distributed in the hope that it will be useful, but
+ *   WITHOUT ANY WARRANTY; without even the implied warranty of
+ *   MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
+ *   NON INFRINGEMENT.  See the GNU General Public License for
+ *   more details.
+ */
+
+/**
+ * @file
+ *
+ * Some low-level simulator definitions.
+ */
+
+#ifndef __ARCH_SIM_DEF_H__
+#define __ARCH_SIM_DEF_H__
+
+
+/**
+ * Internal: the low bits of the SIM_CONTROL_* SPR values specify
+ * the operation to perform, and the remaining bits are
+ * an operation-specific parameter (often unused).
+ */
+#define _SIM_CONTROL_OPERATOR_BITS 8
+
+
+/*
+ * Values which can be written to SPR_SIM_CONTROL.
+ */
+
+/** If written to SPR_SIM_CONTROL, stops profiling. */
+#define SIM_CONTROL_PROFILER_DISABLE 0
+
+/** If written to SPR_SIM_CONTROL, starts profiling. */
+#define SIM_CONTROL_PROFILER_ENABLE 1
+
+/** If written to SPR_SIM_CONTROL, clears profiling counters. */
+#define SIM_CONTROL_PROFILER_CLEAR 2
+
+/** If written to SPR_SIM_CONTROL, checkpoints the simulator. */
+#define SIM_CONTROL_CHECKPOINT 3
+
+/**
+ * If written to SPR_SIM_CONTROL, combined with a mask (shifted by 8),
+ * sets the tracing mask to the given mask. See "sim_set_tracing()".
+ */
+#define SIM_CONTROL_SET_TRACING 4
+
+/**
+ * If written to SPR_SIM_CONTROL, combined with a mask (shifted by 8),
+ * dumps the requested items of machine state to the log.
+ */
+#define SIM_CONTROL_DUMP 5
+
+/** If written to SPR_SIM_CONTROL, clears chip-level profiling counters. */
+#define SIM_CONTROL_PROFILER_CHIP_CLEAR 6
+
+/** If written to SPR_SIM_CONTROL, disables chip-level profiling. */
+#define SIM_CONTROL_PROFILER_CHIP_DISABLE 7
+
+/** If written to SPR_SIM_CONTROL, enables chip-level profiling. */
+#define SIM_CONTROL_PROFILER_CHIP_ENABLE 8
+
+/** If written to SPR_SIM_CONTROL, enables chip-level functional mode */
+#define SIM_CONTROL_ENABLE_FUNCTIONAL 9
+
+/** If written to SPR_SIM_CONTROL, disables chip-level functional mode. */
+#define SIM_CONTROL_DISABLE_FUNCTIONAL 10
+
+/**
+ * If written to SPR_SIM_CONTROL, enables chip-level functional mode.
+ * All tiles must perform this write for functional mode to be enabled.
+ * Ignored in naked boot mode unless --functional is specified.
+ * WARNING: Only the hypervisor startup code should use this!
+ */
+#define SIM_CONTROL_ENABLE_FUNCTIONAL_BARRIER 11
+
+/**
+ * If written to SPR_SIM_CONTROL, combined with a character (shifted by 8),
+ * writes a string directly to the simulator output.  Written to once for
+ * each character in the string, plus a final NUL.  Instead of NUL,
+ * you can also use "SIM_PUTC_FLUSH_STRING" or "SIM_PUTC_FLUSH_BINARY".
+ */
+/* ISSUE: Document the meaning of "newline", and the handling of NUL. */
+#define SIM_CONTROL_PUTC 12
+
+/**
+ * If written to SPR_SIM_CONTROL, clears the --grind-coherence state for
+ * this core.  This is intended to be used before a loop that will
+ * invalidate the cache by loading new data and evicting all current data.
+ * Generally speaking, this API should only be used by system code.
+ */
+#define SIM_CONTROL_GRINDER_CLEAR 13
+
+/** If written to SPR_SIM_CONTROL, shuts down the simulator. */
+#define SIM_CONTROL_SHUTDOWN 14
+
+/**
+ * If written to SPR_SIM_CONTROL, combined with a pid (shifted by 8),
+ * indicates that a fork syscall just created the given process.
+ */
+#define SIM_CONTROL_OS_FORK 15
+
+/**
+ * If written to SPR_SIM_CONTROL, combined with a pid (shifted by 8),
+ * indicates that an exit syscall was just executed by the given process.
+ */
+#define SIM_CONTROL_OS_EXIT 16
+
+/**
+ * If written to SPR_SIM_CONTROL, combined with a pid (shifted by 8),
+ * indicates that the OS just switched to the given process.
+ */
+#define SIM_CONTROL_OS_SWITCH 17
+
+/**
+ * If written to SPR_SIM_CONTROL, combined with a character (shifted by 8),
+ * indicates that an exec syscall was just executed. Written to once for
+ * each character in the executable name, plus a final NUL.
+ */
+#define SIM_CONTROL_OS_EXEC 18
+
+/**
+ * If written to SPR_SIM_CONTROL, combined with a character (shifted by 8),
+ * indicates that an interpreter (PT_INTERP) was loaded.  Written to once
+ * for each character in "ADDR:PATH", plus a final NUL, where "ADDR" is a
+ * hex load address starting with "0x", and "PATH" is the executable name.
+ */
+#define SIM_CONTROL_OS_INTERP 19
+
+/**
+ * If written to SPR_SIM_CONTROL, combined with a character (shifted by 8),
+ * indicates that a dll was loaded.  Written to once for each character
+ * in "ADDR:PATH", plus a final NUL, where "ADDR" is a hexadecimal load
+ * address starting with "0x", and "PATH" is the executable name.
+ */
+#define SIM_CONTROL_DLOPEN 20
+
+/**
+ * If written to SPR_SIM_CONTROL, combined with a character (shifted by 8),
+ * indicates that a dll was unloaded.  Written to once for each character
+ * in "ADDR", plus a final NUL, where "ADDR" is a hexadecimal load
+ * address starting with "0x".
+ */
+#define SIM_CONTROL_DLCLOSE 21
+
+/**
+ * If written to SPR_SIM_CONTROL, combined with a flag (shifted by 8),
+ * indicates whether to allow data reads to remotely-cached
+ * dirty cache lines to be cached locally without grinder warnings or
+ * assertions (used by Linux kernel fast memcpy).
+ */
+#define SIM_CONTROL_ALLOW_MULTIPLE_CACHING 22
+
+/** If written to SPR_SIM_CONTROL, enables memory tracing. */
+#define SIM_CONTROL_ENABLE_MEM_LOGGING 23
+
+/** If written to SPR_SIM_CONTROL, disables memory tracing. */
+#define SIM_CONTROL_DISABLE_MEM_LOGGING 24
+
+/**
+ * If written to SPR_SIM_CONTROL, changes the shaping parameters of one of
+ * the gbe or xgbe shims. Must specify the shim id, the type, the units, and
+ * the rate, as defined in SIM_SHAPING_SPR_ARG.
+ */
+#define SIM_CONTROL_SHAPING 25
+
+/**
+ * If written to SPR_SIM_CONTROL, combined with character (shifted by 8),
+ * requests that a simulator command be executed.  Written to once for each
+ * character in the command, plus a final NUL.
+ */
+#define SIM_CONTROL_COMMAND 26
+
+/**
+ * If written to SPR_SIM_CONTROL, indicates that the simulated system
+ * is panicking, to allow debugging via --debug-on-panic.
+ */
+#define SIM_CONTROL_PANIC 27
+
+/**
+ * If written to SPR_SIM_CONTROL, triggers a simulator syscall.
+ * See "sim_syscall()" for more info.
+ */
+#define SIM_CONTROL_SYSCALL 32
+
+/**
+ * If written to SPR_SIM_CONTROL, combined with a pid (shifted by 8),
+ * provides the pid that subsequent SIM_CONTROL_OS_FORK writes should
+ * use as the pid, rather than the default previous SIM_CONTROL_OS_SWITCH.
+ */
+#define SIM_CONTROL_OS_FORK_PARENT 33
+
+/**
+ * If written to SPR_SIM_CONTROL, combined with a mPIPE shim number
+ * (shifted by 8), clears the pending magic data section.  The cleared
+ * pending magic data section and any subsequently appended magic bytes
+ * will only take effect when the classifier blast programmer is run.
+ */
+#define SIM_CONTROL_CLEAR_MPIPE_MAGIC_BYTES 34
+
+/**
+ * If written to SPR_SIM_CONTROL, combined with a mPIPE shim number
+ * (shifted by 8) and a byte of data (shifted by 16), appends that byte
+ * to the shim's pending magic data section.  The pending magic data
+ * section takes effect when the classifier blast programmer is run.
+ */
+#define SIM_CONTROL_APPEND_MPIPE_MAGIC_BYTE 35
+
+/**
+ * If written to SPR_SIM_CONTROL, combined with a mPIPE shim number
+ * (shifted by 8), an enable=1/disable=0 bit (shifted by 16), and a
+ * mask of links (shifted by 32), enable or disable the corresponding
+ * mPIPE links.
+ */
+#define SIM_CONTROL_ENABLE_MPIPE_LINK_MAGIC_BYTE 36
+
+
+/*
+ * Syscall numbers for use with "sim_syscall()".
+ */
+
+/** Syscall number for sim_add_watchpoint(). */
+#define SIM_SYSCALL_ADD_WATCHPOINT 2
+
+/** Syscall number for sim_remove_watchpoint(). */
+#define SIM_SYSCALL_REMOVE_WATCHPOINT 3
+
+/** Syscall number for sim_query_watchpoint(). */
+#define SIM_SYSCALL_QUERY_WATCHPOINT 4
+
+/**
+ * Syscall number that asserts that the cache lines whose 64-bit PA
+ * is passed as the second argument to sim_syscall(), and over a
+ * range passed as the third argument, are no longer in cache.
+ * The simulator raises an error if this is not the case.
+ */
+#define SIM_SYSCALL_VALIDATE_LINES_EVICTED 5
+
+/** Syscall number for sim_query_cpu_speed(). */
+#define SIM_SYSCALL_QUERY_CPU_SPEED 6
+
+
+/*
+ * Bit masks which can be shifted by 8, combined with
+ * SIM_CONTROL_SET_TRACING, and written to SPR_SIM_CONTROL.
+ */
+
+/**
+ * @addtogroup arch_sim
+ * @{
+ */
+
+/** Enable --trace-cycle when passed to simulator_set_tracing(). */
+#define SIM_TRACE_CYCLES          0x01
+
+/** Enable --trace-router when passed to simulator_set_tracing(). */
+#define SIM_TRACE_ROUTER          0x02
+
+/** Enable --trace-register-writes when passed to simulator_set_tracing(). */
+#define SIM_TRACE_REGISTER_WRITES 0x04
+
+/** Enable --trace-disasm when passed to simulator_set_tracing(). */
+#define SIM_TRACE_DISASM          0x08
+
+/** Enable --trace-stall-info when passed to simulator_set_tracing(). */
+#define SIM_TRACE_STALL_INFO      0x10
+
+/** Enable --trace-memory-controller when passed to simulator_set_tracing(). */
+#define SIM_TRACE_MEMORY_CONTROLLER 0x20
+
+/** Enable --trace-l2 when passed to simulator_set_tracing(). */
+#define SIM_TRACE_L2_CACHE 0x40
+
+/** Enable --trace-lines when passed to simulator_set_tracing(). */
+#define SIM_TRACE_LINES 0x80
+
+/** Turn off all tracing when passed to simulator_set_tracing(). */
+#define SIM_TRACE_NONE 0
+
+/** Turn on all tracing when passed to simulator_set_tracing(). */
+#define SIM_TRACE_ALL (-1)
+
+/** @} */
+
+/** Computes the value to write to SPR_SIM_CONTROL to set tracing flags. */
+#define SIM_TRACE_SPR_ARG(mask) \
+  (SIM_CONTROL_SET_TRACING | ((mask) << _SIM_CONTROL_OPERATOR_BITS))
+
+
+/*
+ * Bit masks which can be shifted by 8, combined with
+ * SIM_CONTROL_DUMP, and written to SPR_SIM_CONTROL.
+ */
+
+/**
+ * @addtogroup arch_sim
+ * @{
+ */
+
+/** Dump the general-purpose registers. */
+#define SIM_DUMP_REGS          0x001
+
+/** Dump the SPRs. */
+#define SIM_DUMP_SPRS          0x002
+
+/** Dump the ITLB. */
+#define SIM_DUMP_ITLB          0x004
+
+/** Dump the DTLB. */
+#define SIM_DUMP_DTLB          0x008
+
+/** Dump the L1 I-cache. */
+#define SIM_DUMP_L1I           0x010
+
+/** Dump the L1 D-cache. */
+#define SIM_DUMP_L1D           0x020
+
+/** Dump the L2 cache. */
+#define SIM_DUMP_L2            0x040
+
+/** Dump the switch registers. */
+#define SIM_DUMP_SNREGS        0x080
+
+/** Dump the switch ITLB. */
+#define SIM_DUMP_SNITLB        0x100
+
+/** Dump the switch L1 I-cache. */
+#define SIM_DUMP_SNL1I         0x200
+
+/** Dump the current backtrace. */
+#define SIM_DUMP_BACKTRACE     0x400
+
+/** Only dump valid lines in caches. */
+#define SIM_DUMP_VALID_LINES   0x800
+
+/** Dump everything that is dumpable. */
+#define SIM_DUMP_ALL (-1 & ~SIM_DUMP_VALID_LINES)
+
+/** @} */
+
+/** Computes the value to write to SPR_SIM_CONTROL to dump machine state. */
+#define SIM_DUMP_SPR_ARG(mask) \
+  (SIM_CONTROL_DUMP | ((mask) << _SIM_CONTROL_OPERATOR_BITS))
+
+
+/*
+ * Bit masks which can be shifted by 8, combined with
+ * SIM_CONTROL_PROFILER_CHIP_xxx, and written to SPR_SIM_CONTROL.
+ */
+
+/**
+ * @addtogroup arch_sim
+ * @{
+ */
+
+/** Use with SIM_PROFILER_CHIP_xxx to control the memory controllers. */
+#define SIM_CHIP_MEMCTL        0x001
+
+/** Use with SIM_PROFILER_CHIP_xxx to control the XAUI interface. */
+#define SIM_CHIP_XAUI          0x002
+
+/** Use with SIM_PROFILER_CHIP_xxx to control the PCIe interface. */
+#define SIM_CHIP_PCIE          0x004
+
+/** Use with SIM_PROFILER_CHIP_xxx to control the MPIPE interface. */
+#define SIM_CHIP_MPIPE         0x008
+
+/** Use with SIM_PROFILER_CHIP_xxx to control the TRIO interface. */
+#define SIM_CHIP_TRIO          0x010
+
+/** Reference all chip devices. */
+#define SIM_CHIP_ALL (-1)
+
+/** @} */
+
+/** Computes the value to write to SPR_SIM_CONTROL to clear chip statistics. */
+#define SIM_PROFILER_CHIP_CLEAR_SPR_ARG(mask) \
+  (SIM_CONTROL_PROFILER_CHIP_CLEAR | ((mask) << _SIM_CONTROL_OPERATOR_BITS))
+
+/** Computes the value to write to SPR_SIM_CONTROL to disable chip statistics.*/
+#define SIM_PROFILER_CHIP_DISABLE_SPR_ARG(mask) \
+  (SIM_CONTROL_PROFILER_CHIP_DISABLE | ((mask) << _SIM_CONTROL_OPERATOR_BITS))
+
+/** Computes the value to write to SPR_SIM_CONTROL to enable chip statistics. */
+#define SIM_PROFILER_CHIP_ENABLE_SPR_ARG(mask) \
+  (SIM_CONTROL_PROFILER_CHIP_ENABLE | ((mask) << _SIM_CONTROL_OPERATOR_BITS))
+
+
+
+/* Shim bitrate controls. */
+
+/** The number of bits used to store the shim id. */
+#define SIM_CONTROL_SHAPING_SHIM_ID_BITS 3
+
+/**
+ * @addtogroup arch_sim
+ * @{
+ */
+
+/** Change the gbe 0 bitrate. */
+#define SIM_CONTROL_SHAPING_GBE_0 0x0
+
+/** Change the gbe 1 bitrate. */
+#define SIM_CONTROL_SHAPING_GBE_1 0x1
+
+/** Change the gbe 2 bitrate. */
+#define SIM_CONTROL_SHAPING_GBE_2 0x2
+
+/** Change the gbe 3 bitrate. */
+#define SIM_CONTROL_SHAPING_GBE_3 0x3
+
+/** Change the xgbe 0 bitrate. */
+#define SIM_CONTROL_SHAPING_XGBE_0 0x4
+
+/** Change the xgbe 1 bitrate. */
+#define SIM_CONTROL_SHAPING_XGBE_1 0x5
+
+/** The type of shaping to do. */
+#define SIM_CONTROL_SHAPING_TYPE_BITS 2
+
+/** Control the multiplier. */
+#define SIM_CONTROL_SHAPING_MULTIPLIER 0
+
+/** Control the PPS. */
+#define SIM_CONTROL_SHAPING_PPS 1
+
+/** Control the BPS. */
+#define SIM_CONTROL_SHAPING_BPS 2
+
+/** The number of bits for the units for the shaping parameter. */
+#define SIM_CONTROL_SHAPING_UNITS_BITS 2
+
+/** Provide a number in single units. */
+#define SIM_CONTROL_SHAPING_UNITS_SINGLE 0
+
+/** Provide a number in kilo units. */
+#define SIM_CONTROL_SHAPING_UNITS_KILO 1
+
+/** Provide a number in mega units. */
+#define SIM_CONTROL_SHAPING_UNITS_MEGA 2
+
+/** Provide a number in giga units. */
+#define SIM_CONTROL_SHAPING_UNITS_GIGA 3
+
+/** @} */
+
+/** How many bits are available for the rate. */
+#define SIM_CONTROL_SHAPING_RATE_BITS \
+  (32 - (_SIM_CONTROL_OPERATOR_BITS + \
+         SIM_CONTROL_SHAPING_SHIM_ID_BITS + \
+         SIM_CONTROL_SHAPING_TYPE_BITS + \
+         SIM_CONTROL_SHAPING_UNITS_BITS))
+
+/** Computes the value to write to SPR_SIM_CONTROL to change a bitrate. */
+#define SIM_SHAPING_SPR_ARG(shim, type, units, rate) \
+  (SIM_CONTROL_SHAPING | \
+   ((shim) | \
+   ((type) << (SIM_CONTROL_SHAPING_SHIM_ID_BITS)) | \
+   ((units) << (SIM_CONTROL_SHAPING_SHIM_ID_BITS + \
+                SIM_CONTROL_SHAPING_TYPE_BITS)) | \
+   ((rate) << (SIM_CONTROL_SHAPING_SHIM_ID_BITS + \
+               SIM_CONTROL_SHAPING_TYPE_BITS + \
+               SIM_CONTROL_SHAPING_UNITS_BITS))) << _SIM_CONTROL_OPERATOR_BITS)
+
+
+/*
+ * Values returned when reading SPR_SIM_CONTROL.
+ * ISSUE: These names should share a longer common prefix.
+ */
+
+/**
+ * When reading SPR_SIM_CONTROL, the mask of simulator tracing bits
+ * (SIM_TRACE_xxx values).
+ */
+#define SIM_TRACE_FLAG_MASK 0xFFFF
+
+/** When reading SPR_SIM_CONTROL, the mask for whether profiling is enabled. */
+#define SIM_PROFILER_ENABLED_MASK 0x10000
+
+
+/*
+ * Special arguments for "SIM_CONTROL_PUTC".
+ */
+
+/**
+ * Flag value for forcing a PUTC string-flush, including
+ * coordinate/cycle prefix and newline.
+ */
+#define SIM_PUTC_FLUSH_STRING 0x100
+
+/**
+ * Flag value for forcing a PUTC binary-data-flush, which skips the
+ * prefix and does not append a newline.
+ */
+#define SIM_PUTC_FLUSH_BINARY 0x101
+
+
+#endif /* __ARCH_SIM_DEF_H__ */
diff --git a/kernel/arch/tile/include/uapi/arch/spr_def.h b/kernel/arch/tile/include/uapi/arch/spr_def.h
new file mode 100644
index 000000000..c250c5adb
--- /dev/null
+++ b/kernel/arch/tile/include/uapi/arch/spr_def.h
@@ -0,0 +1,26 @@
+/*
+ * Copyright 2010 Tilera Corporation. All Rights Reserved.
+ *
+ *   This program is free software; you can redistribute it and/or
+ *   modify it under the terms of the GNU General Public License
+ *   as published by the Free Software Foundation, version 2.
+ *
+ *   This program is distributed in the hope that it will be useful, but
+ *   WITHOUT ANY WARRANTY; without even the implied warranty of
+ *   MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
+ *   NON INFRINGEMENT.  See the GNU General Public License for
+ *   more details.
+ */
+
+#ifndef _UAPI__ARCH_SPR_DEF_H__
+#define _UAPI__ARCH_SPR_DEF_H__
+
+/* Include the proper base SPR definition file. */
+#ifdef __tilegx__
+#include <arch/spr_def_64.h>
+#else
+#include <arch/spr_def_32.h>
+#endif
+
+
+#endif /* _UAPI__ARCH_SPR_DEF_H__ */
diff --git a/kernel/arch/tile/include/uapi/arch/spr_def_32.h b/kernel/arch/tile/include/uapi/arch/spr_def_32.h
new file mode 100644
index 000000000..78daa3146
--- /dev/null
+++ b/kernel/arch/tile/include/uapi/arch/spr_def_32.h
@@ -0,0 +1,255 @@
+/*
+ * Copyright 2010 Tilera Corporation. All Rights Reserved.
+ *
+ *   This program is free software; you can redistribute it and/or
+ *   modify it under the terms of the GNU General Public License
+ *   as published by the Free Software Foundation, version 2.
+ *
+ *   This program is distributed in the hope that it will be useful, but
+ *   WITHOUT ANY WARRANTY; without even the implied warranty of
+ *   MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
+ *   NON INFRINGEMENT.  See the GNU General Public License for
+ *   more details.
+ */
+
+#ifndef __DOXYGEN__
+
+#ifndef __ARCH_SPR_DEF_32_H__
+#define __ARCH_SPR_DEF_32_H__
+
+#define SPR_AUX_PERF_COUNT_0 0x6005
+#define SPR_AUX_PERF_COUNT_1 0x6006
+#define SPR_AUX_PERF_COUNT_CTL 0x6007
+#define SPR_AUX_PERF_COUNT_STS 0x6008
+#define SPR_CYCLE_HIGH 0x4e06
+#define SPR_CYCLE_LOW 0x4e07
+#define SPR_DMA_BYTE 0x3900
+#define SPR_DMA_CHUNK_SIZE 0x3901
+#define SPR_DMA_CTR 0x3902
+#define SPR_DMA_CTR__REQUEST_MASK  0x1
+#define SPR_DMA_CTR__SUSPEND_MASK  0x2
+#define SPR_DMA_DST_ADDR 0x3903
+#define SPR_DMA_DST_CHUNK_ADDR 0x3904
+#define SPR_DMA_SRC_ADDR 0x3905
+#define SPR_DMA_SRC_CHUNK_ADDR 0x3906
+#define SPR_DMA_STATUS__DONE_MASK  0x1
+#define SPR_DMA_STATUS__BUSY_MASK  0x2
+#define SPR_DMA_STATUS__RUNNING_MASK  0x10
+#define SPR_DMA_STRIDE 0x3907
+#define SPR_DMA_USER_STATUS 0x3908
+#define SPR_DONE 0x4e08
+#define SPR_EVENT_BEGIN 0x4e0d
+#define SPR_EVENT_END 0x4e0e
+#define SPR_EX_CONTEXT_0_0 0x4a05
+#define SPR_EX_CONTEXT_0_1 0x4a06
+#define SPR_EX_CONTEXT_0_1__PL_SHIFT 0
+#define SPR_EX_CONTEXT_0_1__PL_RMASK 0x3
+#define SPR_EX_CONTEXT_0_1__PL_MASK  0x3
+#define SPR_EX_CONTEXT_0_1__ICS_SHIFT 2
+#define SPR_EX_CONTEXT_0_1__ICS_RMASK 0x1
+#define SPR_EX_CONTEXT_0_1__ICS_MASK  0x4
+#define SPR_EX_CONTEXT_1_0 0x4805
+#define SPR_EX_CONTEXT_1_1 0x4806
+#define SPR_EX_CONTEXT_1_1__PL_SHIFT 0
+#define SPR_EX_CONTEXT_1_1__PL_RMASK 0x3
+#define SPR_EX_CONTEXT_1_1__PL_MASK  0x3
+#define SPR_EX_CONTEXT_1_1__ICS_SHIFT 2
+#define SPR_EX_CONTEXT_1_1__ICS_RMASK 0x1
+#define SPR_EX_CONTEXT_1_1__ICS_MASK  0x4
+#define SPR_EX_CONTEXT_2_0 0x4605
+#define SPR_EX_CONTEXT_2_1 0x4606
+#define SPR_EX_CONTEXT_2_1__PL_SHIFT 0
+#define SPR_EX_CONTEXT_2_1__PL_RMASK 0x3
+#define SPR_EX_CONTEXT_2_1__PL_MASK  0x3
+#define SPR_EX_CONTEXT_2_1__ICS_SHIFT 2
+#define SPR_EX_CONTEXT_2_1__ICS_RMASK 0x1
+#define SPR_EX_CONTEXT_2_1__ICS_MASK  0x4
+#define SPR_FAIL 0x4e09
+#define SPR_IDN_AVAIL_EN 0x3e05
+#define SPR_IDN_CA_DATA 0x0b00
+#define SPR_IDN_DATA_AVAIL 0x0b03
+#define SPR_IDN_DEADLOCK_TIMEOUT 0x3406
+#define SPR_IDN_DEMUX_CA_COUNT 0x0a05
+#define SPR_IDN_DEMUX_COUNT_0 0x0a06
+#define SPR_IDN_DEMUX_COUNT_1 0x0a07
+#define SPR_IDN_DEMUX_CTL 0x0a08
+#define SPR_IDN_DEMUX_QUEUE_SEL 0x0a0a
+#define SPR_IDN_DEMUX_STATUS 0x0a0b
+#define SPR_IDN_DEMUX_WRITE_FIFO 0x0a0c
+#define SPR_IDN_DIRECTION_PROTECT 0x2e05
+#define SPR_IDN_PENDING 0x0a0e
+#define SPR_IDN_REFILL_EN 0x0e05
+#define SPR_IDN_SP_FIFO_DATA 0x0a0f
+#define SPR_IDN_SP_FIFO_SEL 0x0a10
+#define SPR_IDN_SP_FREEZE 0x0a11
+#define SPR_IDN_SP_FREEZE__SP_FRZ_MASK  0x1
+#define SPR_IDN_SP_FREEZE__DEMUX_FRZ_MASK  0x2
+#define SPR_IDN_SP_FREEZE__NON_DEST_EXT_MASK  0x4
+#define SPR_IDN_SP_STATE 0x0a12
+#define SPR_IDN_TAG_0 0x0a13
+#define SPR_IDN_TAG_1 0x0a14
+#define SPR_IDN_TAG_VALID 0x0a15
+#define SPR_IDN_TILE_COORD 0x0a16
+#define SPR_INTCTRL_0_STATUS 0x4a07
+#define SPR_INTCTRL_1_STATUS 0x4807
+#define SPR_INTCTRL_2_STATUS 0x4607
+#define SPR_INTERRUPT_CRITICAL_SECTION 0x4e0a
+#define SPR_INTERRUPT_MASK_0_0 0x4a08
+#define SPR_INTERRUPT_MASK_0_1 0x4a09
+#define SPR_INTERRUPT_MASK_1_0 0x4809
+#define SPR_INTERRUPT_MASK_1_1 0x480a
+#define SPR_INTERRUPT_MASK_2_0 0x4608
+#define SPR_INTERRUPT_MASK_2_1 0x4609
+#define SPR_INTERRUPT_MASK_RESET_0_0 0x4a0a
+#define SPR_INTERRUPT_MASK_RESET_0_1 0x4a0b
+#define SPR_INTERRUPT_MASK_RESET_1_0 0x480b
+#define SPR_INTERRUPT_MASK_RESET_1_1 0x480c
+#define SPR_INTERRUPT_MASK_RESET_2_0 0x460a
+#define SPR_INTERRUPT_MASK_RESET_2_1 0x460b
+#define SPR_INTERRUPT_MASK_SET_0_0 0x4a0c
+#define SPR_INTERRUPT_MASK_SET_0_1 0x4a0d
+#define SPR_INTERRUPT_MASK_SET_1_0 0x480d
+#define SPR_INTERRUPT_MASK_SET_1_1 0x480e
+#define SPR_INTERRUPT_MASK_SET_2_0 0x460c
+#define SPR_INTERRUPT_MASK_SET_2_1 0x460d
+#define SPR_MPL_AUX_PERF_COUNT_SET_0 0x6000
+#define SPR_MPL_AUX_PERF_COUNT_SET_1 0x6001
+#define SPR_MPL_AUX_PERF_COUNT_SET_2 0x6002
+#define SPR_MPL_DMA_CPL_SET_0 0x5800
+#define SPR_MPL_DMA_CPL_SET_1 0x5801
+#define SPR_MPL_DMA_CPL_SET_2 0x5802
+#define SPR_MPL_DMA_NOTIFY_SET_0 0x3800
+#define SPR_MPL_DMA_NOTIFY_SET_1 0x3801
+#define SPR_MPL_DMA_NOTIFY_SET_2 0x3802
+#define SPR_MPL_IDN_ACCESS_SET_0 0x0a00
+#define SPR_MPL_IDN_ACCESS_SET_1 0x0a01
+#define SPR_MPL_IDN_ACCESS_SET_2 0x0a02
+#define SPR_MPL_IDN_AVAIL_SET_0 0x3e00
+#define SPR_MPL_IDN_AVAIL_SET_1 0x3e01
+#define SPR_MPL_IDN_AVAIL_SET_2 0x3e02
+#define SPR_MPL_IDN_CA_SET_0 0x3a00
+#define SPR_MPL_IDN_CA_SET_1 0x3a01
+#define SPR_MPL_IDN_CA_SET_2 0x3a02
+#define SPR_MPL_IDN_COMPLETE_SET_0 0x1200
+#define SPR_MPL_IDN_COMPLETE_SET_1 0x1201
+#define SPR_MPL_IDN_COMPLETE_SET_2 0x1202
+#define SPR_MPL_IDN_FIREWALL_SET_0 0x2e00
+#define SPR_MPL_IDN_FIREWALL_SET_1 0x2e01
+#define SPR_MPL_IDN_FIREWALL_SET_2 0x2e02
+#define SPR_MPL_IDN_REFILL_SET_0 0x0e00
+#define SPR_MPL_IDN_REFILL_SET_1 0x0e01
+#define SPR_MPL_IDN_REFILL_SET_2 0x0e02
+#define SPR_MPL_IDN_TIMER_SET_0 0x3400
+#define SPR_MPL_IDN_TIMER_SET_1 0x3401
+#define SPR_MPL_IDN_TIMER_SET_2 0x3402
+#define SPR_MPL_INTCTRL_0_SET_0 0x4a00
+#define SPR_MPL_INTCTRL_0_SET_1 0x4a01
+#define SPR_MPL_INTCTRL_0_SET_2 0x4a02
+#define SPR_MPL_INTCTRL_1_SET_0 0x4800
+#define SPR_MPL_INTCTRL_1_SET_1 0x4801
+#define SPR_MPL_INTCTRL_1_SET_2 0x4802
+#define SPR_MPL_INTCTRL_2_SET_0 0x4600
+#define SPR_MPL_INTCTRL_2_SET_1 0x4601
+#define SPR_MPL_INTCTRL_2_SET_2 0x4602
+#define SPR_MPL_PERF_COUNT_SET_0 0x4200
+#define SPR_MPL_PERF_COUNT_SET_1 0x4201
+#define SPR_MPL_PERF_COUNT_SET_2 0x4202
+#define SPR_MPL_SN_ACCESS_SET_0 0x0800
+#define SPR_MPL_SN_ACCESS_SET_1 0x0801
+#define SPR_MPL_SN_ACCESS_SET_2 0x0802
+#define SPR_MPL_SN_CPL_SET_0 0x5a00
+#define SPR_MPL_SN_CPL_SET_1 0x5a01
+#define SPR_MPL_SN_CPL_SET_2 0x5a02
+#define SPR_MPL_SN_FIREWALL_SET_0 0x2c00
+#define SPR_MPL_SN_FIREWALL_SET_1 0x2c01
+#define SPR_MPL_SN_FIREWALL_SET_2 0x2c02
+#define SPR_MPL_SN_NOTIFY_SET_0 0x2a00
+#define SPR_MPL_SN_NOTIFY_SET_1 0x2a01
+#define SPR_MPL_SN_NOTIFY_SET_2 0x2a02
+#define SPR_MPL_UDN_ACCESS_SET_0 0x0c00
+#define SPR_MPL_UDN_ACCESS_SET_1 0x0c01
+#define SPR_MPL_UDN_ACCESS_SET_2 0x0c02
+#define SPR_MPL_UDN_AVAIL_SET_0 0x4000
+#define SPR_MPL_UDN_AVAIL_SET_1 0x4001
+#define SPR_MPL_UDN_AVAIL_SET_2 0x4002
+#define SPR_MPL_UDN_CA_SET_0 0x3c00
+#define SPR_MPL_UDN_CA_SET_1 0x3c01
+#define SPR_MPL_UDN_CA_SET_2 0x3c02
+#define SPR_MPL_UDN_COMPLETE_SET_0 0x1400
+#define SPR_MPL_UDN_COMPLETE_SET_1 0x1401
+#define SPR_MPL_UDN_COMPLETE_SET_2 0x1402
+#define SPR_MPL_UDN_FIREWALL_SET_0 0x3000
+#define SPR_MPL_UDN_FIREWALL_SET_1 0x3001
+#define SPR_MPL_UDN_FIREWALL_SET_2 0x3002
+#define SPR_MPL_UDN_REFILL_SET_0 0x1000
+#define SPR_MPL_UDN_REFILL_SET_1 0x1001
+#define SPR_MPL_UDN_REFILL_SET_2 0x1002
+#define SPR_MPL_UDN_TIMER_SET_0 0x3600
+#define SPR_MPL_UDN_TIMER_SET_1 0x3601
+#define SPR_MPL_UDN_TIMER_SET_2 0x3602
+#define SPR_MPL_WORLD_ACCESS_SET_0 0x4e00
+#define SPR_MPL_WORLD_ACCESS_SET_1 0x4e01
+#define SPR_MPL_WORLD_ACCESS_SET_2 0x4e02
+#define SPR_PASS 0x4e0b
+#define SPR_PERF_COUNT_0 0x4205
+#define SPR_PERF_COUNT_1 0x4206
+#define SPR_PERF_COUNT_CTL 0x4207
+#define SPR_PERF_COUNT_DN_CTL 0x4210
+#define SPR_PERF_COUNT_STS 0x4208
+#define SPR_PROC_STATUS 0x4f00
+#define SPR_SIM_CONTROL 0x4e0c
+#define SPR_SNCTL 0x0805
+#define SPR_SNCTL__FRZFABRIC_MASK  0x1
+#define SPR_SNSTATIC 0x080c
+#define SPR_SYSTEM_SAVE_0_0 0x4b00
+#define SPR_SYSTEM_SAVE_0_1 0x4b01
+#define SPR_SYSTEM_SAVE_0_2 0x4b02
+#define SPR_SYSTEM_SAVE_0_3 0x4b03
+#define SPR_SYSTEM_SAVE_1_0 0x4900
+#define SPR_SYSTEM_SAVE_1_1 0x4901
+#define SPR_SYSTEM_SAVE_1_2 0x4902
+#define SPR_SYSTEM_SAVE_1_3 0x4903
+#define SPR_SYSTEM_SAVE_2_0 0x4700
+#define SPR_SYSTEM_SAVE_2_1 0x4701
+#define SPR_SYSTEM_SAVE_2_2 0x4702
+#define SPR_SYSTEM_SAVE_2_3 0x4703
+#define SPR_TILE_COORD 0x4c17
+#define SPR_TILE_RTF_HWM 0x4e10
+#define SPR_TILE_TIMER_CONTROL 0x3205
+#define SPR_TILE_WRITE_PENDING 0x4e0f
+#define SPR_UDN_AVAIL_EN 0x4005
+#define SPR_UDN_CA_DATA 0x0d00
+#define SPR_UDN_DATA_AVAIL 0x0d03
+#define SPR_UDN_DEADLOCK_TIMEOUT 0x3606
+#define SPR_UDN_DEMUX_CA_COUNT 0x0c05
+#define SPR_UDN_DEMUX_COUNT_0 0x0c06
+#define SPR_UDN_DEMUX_COUNT_1 0x0c07
+#define SPR_UDN_DEMUX_COUNT_2 0x0c08
+#define SPR_UDN_DEMUX_COUNT_3 0x0c09
+#define SPR_UDN_DEMUX_CTL 0x0c0a
+#define SPR_UDN_DEMUX_QUEUE_SEL 0x0c0c
+#define SPR_UDN_DEMUX_STATUS 0x0c0d
+#define SPR_UDN_DEMUX_WRITE_FIFO 0x0c0e
+#define SPR_UDN_DIRECTION_PROTECT 0x3005
+#define SPR_UDN_PENDING 0x0c10
+#define SPR_UDN_REFILL_EN 0x1005
+#define SPR_UDN_SP_FIFO_DATA 0x0c11
+#define SPR_UDN_SP_FIFO_SEL 0x0c12
+#define SPR_UDN_SP_FREEZE 0x0c13
+#define SPR_UDN_SP_FREEZE__SP_FRZ_MASK  0x1
+#define SPR_UDN_SP_FREEZE__DEMUX_FRZ_MASK  0x2
+#define SPR_UDN_SP_FREEZE__NON_DEST_EXT_MASK  0x4
+#define SPR_UDN_SP_STATE 0x0c14
+#define SPR_UDN_TAG_0 0x0c15
+#define SPR_UDN_TAG_1 0x0c16
+#define SPR_UDN_TAG_2 0x0c17
+#define SPR_UDN_TAG_3 0x0c18
+#define SPR_UDN_TAG_VALID 0x0c19
+#define SPR_UDN_TILE_COORD 0x0c1a
+#define SPR_WATCH_CTL 0x4209
+#define SPR_WATCH_MASK 0x420a
+#define SPR_WATCH_VAL 0x420b
+
+#endif /* !defined(__ARCH_SPR_DEF_32_H__) */
+
+#endif /* !defined(__DOXYGEN__) */
diff --git a/kernel/arch/tile/include/uapi/arch/spr_def_64.h b/kernel/arch/tile/include/uapi/arch/spr_def_64.h
new file mode 100644
index 000000000..67a6c1751
--- /dev/null
+++ b/kernel/arch/tile/include/uapi/arch/spr_def_64.h
@@ -0,0 +1,216 @@
+/*
+ * Copyright 2011 Tilera Corporation. All Rights Reserved.
+ *
+ *   This program is free software; you can redistribute it and/or
+ *   modify it under the terms of the GNU General Public License
+ *   as published by the Free Software Foundation, version 2.
+ *
+ *   This program is distributed in the hope that it will be useful, but
+ *   WITHOUT ANY WARRANTY; without even the implied warranty of
+ *   MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
+ *   NON INFRINGEMENT.  See the GNU General Public License for
+ *   more details.
+ */
+
+#ifndef __DOXYGEN__
+
+#ifndef __ARCH_SPR_DEF_64_H__
+#define __ARCH_SPR_DEF_64_H__
+
+#define SPR_AUX_PERF_COUNT_0 0x2105
+#define SPR_AUX_PERF_COUNT_1 0x2106
+#define SPR_AUX_PERF_COUNT_CTL 0x2107
+#define SPR_AUX_PERF_COUNT_STS 0x2108
+#define SPR_CMPEXCH_VALUE 0x2780
+#define SPR_CYCLE 0x2781
+#define SPR_DONE 0x2705
+#define SPR_DSTREAM_PF 0x2706
+#define SPR_EVENT_BEGIN 0x2782
+#define SPR_EVENT_END 0x2783
+#define SPR_EX_CONTEXT_0_0 0x2580
+#define SPR_EX_CONTEXT_0_1 0x2581
+#define SPR_EX_CONTEXT_0_1__PL_SHIFT 0
+#define SPR_EX_CONTEXT_0_1__PL_RMASK 0x3
+#define SPR_EX_CONTEXT_0_1__PL_MASK  0x3
+#define SPR_EX_CONTEXT_0_1__ICS_SHIFT 2
+#define SPR_EX_CONTEXT_0_1__ICS_RMASK 0x1
+#define SPR_EX_CONTEXT_0_1__ICS_MASK  0x4
+#define SPR_EX_CONTEXT_1_0 0x2480
+#define SPR_EX_CONTEXT_1_1 0x2481
+#define SPR_EX_CONTEXT_1_1__PL_SHIFT 0
+#define SPR_EX_CONTEXT_1_1__PL_RMASK 0x3
+#define SPR_EX_CONTEXT_1_1__PL_MASK  0x3
+#define SPR_EX_CONTEXT_1_1__ICS_SHIFT 2
+#define SPR_EX_CONTEXT_1_1__ICS_RMASK 0x1
+#define SPR_EX_CONTEXT_1_1__ICS_MASK  0x4
+#define SPR_EX_CONTEXT_2_0 0x2380
+#define SPR_EX_CONTEXT_2_1 0x2381
+#define SPR_EX_CONTEXT_2_1__PL_SHIFT 0
+#define SPR_EX_CONTEXT_2_1__PL_RMASK 0x3
+#define SPR_EX_CONTEXT_2_1__PL_MASK  0x3
+#define SPR_EX_CONTEXT_2_1__ICS_SHIFT 2
+#define SPR_EX_CONTEXT_2_1__ICS_RMASK 0x1
+#define SPR_EX_CONTEXT_2_1__ICS_MASK  0x4
+#define SPR_FAIL 0x2707
+#define SPR_IDN_AVAIL_EN 0x1a05
+#define SPR_IDN_DATA_AVAIL 0x0a80
+#define SPR_IDN_DEADLOCK_TIMEOUT 0x1806
+#define SPR_IDN_DEMUX_COUNT_0 0x0a05
+#define SPR_IDN_DEMUX_COUNT_1 0x0a06
+#define SPR_IDN_DIRECTION_PROTECT 0x1405
+#define SPR_IDN_PENDING 0x0a08
+#define SPR_ILL_TRANS_REASON__I_STREAM_VA_RMASK 0x1
+#define SPR_INTCTRL_0_STATUS 0x2505
+#define SPR_INTCTRL_1_STATUS 0x2405
+#define SPR_INTCTRL_2_STATUS 0x2305
+#define SPR_INTERRUPT_CRITICAL_SECTION 0x2708
+#define SPR_INTERRUPT_MASK_0 0x2506
+#define SPR_INTERRUPT_MASK_1 0x2406
+#define SPR_INTERRUPT_MASK_2 0x2306
+#define SPR_INTERRUPT_MASK_RESET_0 0x2507
+#define SPR_INTERRUPT_MASK_RESET_1 0x2407
+#define SPR_INTERRUPT_MASK_RESET_2 0x2307
+#define SPR_INTERRUPT_MASK_SET_0 0x2508
+#define SPR_INTERRUPT_MASK_SET_1 0x2408
+#define SPR_INTERRUPT_MASK_SET_2 0x2308
+#define SPR_INTERRUPT_VECTOR_BASE_0 0x2509
+#define SPR_INTERRUPT_VECTOR_BASE_1 0x2409
+#define SPR_INTERRUPT_VECTOR_BASE_2 0x2309
+#define SPR_INTERRUPT_VECTOR_BASE_3 0x2209
+#define SPR_IPI_EVENT_0 0x1f05
+#define SPR_IPI_EVENT_1 0x1e05
+#define SPR_IPI_EVENT_2 0x1d05
+#define SPR_IPI_EVENT_RESET_0 0x1f06
+#define SPR_IPI_EVENT_RESET_1 0x1e06
+#define SPR_IPI_EVENT_RESET_2 0x1d06
+#define SPR_IPI_EVENT_SET_0 0x1f07
+#define SPR_IPI_EVENT_SET_1 0x1e07
+#define SPR_IPI_EVENT_SET_2 0x1d07
+#define SPR_IPI_MASK_0 0x1f08
+#define SPR_IPI_MASK_1 0x1e08
+#define SPR_IPI_MASK_2 0x1d08
+#define SPR_IPI_MASK_RESET_0 0x1f09
+#define SPR_IPI_MASK_RESET_1 0x1e09
+#define SPR_IPI_MASK_RESET_2 0x1d09
+#define SPR_IPI_MASK_SET_0 0x1f0a
+#define SPR_IPI_MASK_SET_1 0x1e0a
+#define SPR_IPI_MASK_SET_2 0x1d0a
+#define SPR_MPL_AUX_PERF_COUNT_SET_0 0x2100
+#define SPR_MPL_AUX_PERF_COUNT_SET_1 0x2101
+#define SPR_MPL_AUX_PERF_COUNT_SET_2 0x2102
+#define SPR_MPL_AUX_TILE_TIMER_SET_0 0x1700
+#define SPR_MPL_AUX_TILE_TIMER_SET_1 0x1701
+#define SPR_MPL_AUX_TILE_TIMER_SET_2 0x1702
+#define SPR_MPL_IDN_ACCESS_SET_0 0x0a00
+#define SPR_MPL_IDN_ACCESS_SET_1 0x0a01
+#define SPR_MPL_IDN_ACCESS_SET_2 0x0a02
+#define SPR_MPL_IDN_AVAIL_SET_0 0x1a00
+#define SPR_MPL_IDN_AVAIL_SET_1 0x1a01
+#define SPR_MPL_IDN_AVAIL_SET_2 0x1a02
+#define SPR_MPL_IDN_COMPLETE_SET_0 0x0500
+#define SPR_MPL_IDN_COMPLETE_SET_1 0x0501
+#define SPR_MPL_IDN_COMPLETE_SET_2 0x0502
+#define SPR_MPL_IDN_FIREWALL_SET_0 0x1400
+#define SPR_MPL_IDN_FIREWALL_SET_1 0x1401
+#define SPR_MPL_IDN_FIREWALL_SET_2 0x1402
+#define SPR_MPL_IDN_TIMER_SET_0 0x1800
+#define SPR_MPL_IDN_TIMER_SET_1 0x1801
+#define SPR_MPL_IDN_TIMER_SET_2 0x1802
+#define SPR_MPL_INTCTRL_0_SET_0 0x2500
+#define SPR_MPL_INTCTRL_0_SET_1 0x2501
+#define SPR_MPL_INTCTRL_0_SET_2 0x2502
+#define SPR_MPL_INTCTRL_1_SET_0 0x2400
+#define SPR_MPL_INTCTRL_1_SET_1 0x2401
+#define SPR_MPL_INTCTRL_1_SET_2 0x2402
+#define SPR_MPL_INTCTRL_2_SET_0 0x2300
+#define SPR_MPL_INTCTRL_2_SET_1 0x2301
+#define SPR_MPL_INTCTRL_2_SET_2 0x2302
+#define SPR_MPL_IPI_0 0x1f04
+#define SPR_MPL_IPI_0_SET_0 0x1f00
+#define SPR_MPL_IPI_0_SET_1 0x1f01
+#define SPR_MPL_IPI_0_SET_2 0x1f02
+#define SPR_MPL_IPI_1 0x1e04
+#define SPR_MPL_IPI_1_SET_0 0x1e00
+#define SPR_MPL_IPI_1_SET_1 0x1e01
+#define SPR_MPL_IPI_1_SET_2 0x1e02
+#define SPR_MPL_IPI_2 0x1d04
+#define SPR_MPL_IPI_2_SET_0 0x1d00
+#define SPR_MPL_IPI_2_SET_1 0x1d01
+#define SPR_MPL_IPI_2_SET_2 0x1d02
+#define SPR_MPL_PERF_COUNT_SET_0 0x2000
+#define SPR_MPL_PERF_COUNT_SET_1 0x2001
+#define SPR_MPL_PERF_COUNT_SET_2 0x2002
+#define SPR_MPL_UDN_ACCESS_SET_0 0x0b00
+#define SPR_MPL_UDN_ACCESS_SET_1 0x0b01
+#define SPR_MPL_UDN_ACCESS_SET_2 0x0b02
+#define SPR_MPL_UDN_AVAIL_SET_0 0x1b00
+#define SPR_MPL_UDN_AVAIL_SET_1 0x1b01
+#define SPR_MPL_UDN_AVAIL_SET_2 0x1b02
+#define SPR_MPL_UDN_COMPLETE_SET_0 0x0600
+#define SPR_MPL_UDN_COMPLETE_SET_1 0x0601
+#define SPR_MPL_UDN_COMPLETE_SET_2 0x0602
+#define SPR_MPL_UDN_FIREWALL_SET_0 0x1500
+#define SPR_MPL_UDN_FIREWALL_SET_1 0x1501
+#define SPR_MPL_UDN_FIREWALL_SET_2 0x1502
+#define SPR_MPL_UDN_TIMER_SET_0 0x1900
+#define SPR_MPL_UDN_TIMER_SET_1 0x1901
+#define SPR_MPL_UDN_TIMER_SET_2 0x1902
+#define SPR_MPL_WORLD_ACCESS_SET_0 0x2700
+#define SPR_MPL_WORLD_ACCESS_SET_1 0x2701
+#define SPR_MPL_WORLD_ACCESS_SET_2 0x2702
+#define SPR_PASS 0x2709
+#define SPR_PERF_COUNT_0 0x2005
+#define SPR_PERF_COUNT_1 0x2006
+#define SPR_PERF_COUNT_CTL 0x2007
+#define SPR_PERF_COUNT_DN_CTL 0x2008
+#define SPR_PERF_COUNT_STS 0x2009
+#define SPR_PROC_STATUS 0x2784
+#define SPR_SIM_CONTROL 0x2785
+#define SPR_SINGLE_STEP_CONTROL_0 0x0405
+#define SPR_SINGLE_STEP_CONTROL_0__CANCELED_MASK  0x1
+#define SPR_SINGLE_STEP_CONTROL_0__INHIBIT_MASK  0x2
+#define SPR_SINGLE_STEP_CONTROL_1 0x0305
+#define SPR_SINGLE_STEP_CONTROL_1__CANCELED_MASK  0x1
+#define SPR_SINGLE_STEP_CONTROL_1__INHIBIT_MASK  0x2
+#define SPR_SINGLE_STEP_CONTROL_2 0x0205
+#define SPR_SINGLE_STEP_CONTROL_2__CANCELED_MASK  0x1
+#define SPR_SINGLE_STEP_CONTROL_2__INHIBIT_MASK  0x2
+#define SPR_SINGLE_STEP_EN_0_0 0x250a
+#define SPR_SINGLE_STEP_EN_0_1 0x240a
+#define SPR_SINGLE_STEP_EN_0_2 0x230a
+#define SPR_SINGLE_STEP_EN_1_0 0x250b
+#define SPR_SINGLE_STEP_EN_1_1 0x240b
+#define SPR_SINGLE_STEP_EN_1_2 0x230b
+#define SPR_SINGLE_STEP_EN_2_0 0x250c
+#define SPR_SINGLE_STEP_EN_2_1 0x240c
+#define SPR_SINGLE_STEP_EN_2_2 0x230c
+#define SPR_SYSTEM_SAVE_0_0 0x2582
+#define SPR_SYSTEM_SAVE_0_1 0x2583
+#define SPR_SYSTEM_SAVE_0_2 0x2584
+#define SPR_SYSTEM_SAVE_0_3 0x2585
+#define SPR_SYSTEM_SAVE_1_0 0x2482
+#define SPR_SYSTEM_SAVE_1_1 0x2483
+#define SPR_SYSTEM_SAVE_1_2 0x2484
+#define SPR_SYSTEM_SAVE_1_3 0x2485
+#define SPR_SYSTEM_SAVE_2_0 0x2382
+#define SPR_SYSTEM_SAVE_2_1 0x2383
+#define SPR_SYSTEM_SAVE_2_2 0x2384
+#define SPR_SYSTEM_SAVE_2_3 0x2385
+#define SPR_TILE_COORD 0x270b
+#define SPR_TILE_RTF_HWM 0x270c
+#define SPR_TILE_TIMER_CONTROL 0x1605
+#define SPR_UDN_AVAIL_EN 0x1b05
+#define SPR_UDN_DATA_AVAIL 0x0b80
+#define SPR_UDN_DEADLOCK_TIMEOUT 0x1906
+#define SPR_UDN_DEMUX_COUNT_0 0x0b05
+#define SPR_UDN_DEMUX_COUNT_1 0x0b06
+#define SPR_UDN_DEMUX_COUNT_2 0x0b07
+#define SPR_UDN_DEMUX_COUNT_3 0x0b08
+#define SPR_UDN_DIRECTION_PROTECT 0x1505
+#define SPR_UDN_PENDING 0x0b0a
+#define SPR_WATCH_MASK 0x200a
+#define SPR_WATCH_VAL 0x200b
+
+#endif /* !defined(__ARCH_SPR_DEF_64_H__) */
+
+#endif /* !defined(__DOXYGEN__) */
diff --git a/kernel/arch/tile/include/uapi/asm/Kbuild b/kernel/arch/tile/include/uapi/asm/Kbuild
new file mode 100644
index 000000000..c20db8e42
--- /dev/null
+++ b/kernel/arch/tile/include/uapi/asm/Kbuild
@@ -0,0 +1,21 @@
+# UAPI Header export list
+include include/uapi/asm-generic/Kbuild.asm
+
+header-y += auxvec.h
+header-y += bitsperlong.h
+header-y += byteorder.h
+header-y += cachectl.h
+header-y += hardwall.h
+header-y += kvm_para.h
+header-y += mman.h
+header-y += ptrace.h
+header-y += setup.h
+header-y += sigcontext.h
+header-y += siginfo.h
+header-y += signal.h
+header-y += stat.h
+header-y += swab.h
+header-y += ucontext.h
+header-y += unistd.h
+
+generic-y += ucontext.h
diff --git a/kernel/arch/tile/include/uapi/asm/auxvec.h b/kernel/arch/tile/include/uapi/asm/auxvec.h
new file mode 100644
index 000000000..c93e92709
--- /dev/null
+++ b/kernel/arch/tile/include/uapi/asm/auxvec.h
@@ -0,0 +1,21 @@
+/*
+ * Copyright 2010 Tilera Corporation. All Rights Reserved.
+ *
+ *   This program is free software; you can redistribute it and/or
+ *   modify it under the terms of the GNU General Public License
+ *   as published by the Free Software Foundation, version 2.
+ *
+ *   This program is distributed in the hope that it will be useful, but
+ *   WITHOUT ANY WARRANTY; without even the implied warranty of
+ *   MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
+ *   NON INFRINGEMENT.  See the GNU General Public License for
+ *   more details.
+ */
+
+#ifndef _ASM_TILE_AUXVEC_H
+#define _ASM_TILE_AUXVEC_H
+
+/* The vDSO location. */
+#define AT_SYSINFO_EHDR         33
+
+#endif /* _ASM_TILE_AUXVEC_H */
diff --git a/kernel/arch/tile/include/uapi/asm/bitsperlong.h b/kernel/arch/tile/include/uapi/asm/bitsperlong.h
new file mode 100644
index 000000000..58c771f2a
--- /dev/null
+++ b/kernel/arch/tile/include/uapi/asm/bitsperlong.h
@@ -0,0 +1,26 @@
+/*
+ * Copyright 2010 Tilera Corporation. All Rights Reserved.
+ *
+ *   This program is free software; you can redistribute it and/or
+ *   modify it under the terms of the GNU General Public License
+ *   as published by the Free Software Foundation, version 2.
+ *
+ *   This program is distributed in the hope that it will be useful, but
+ *   WITHOUT ANY WARRANTY; without even the implied warranty of
+ *   MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
+ *   NON INFRINGEMENT.  See the GNU General Public License for
+ *   more details.
+ */
+
+#ifndef _ASM_TILE_BITSPERLONG_H
+#define _ASM_TILE_BITSPERLONG_H
+
+#ifdef __LP64__
+# define __BITS_PER_LONG 64
+#else
+# define __BITS_PER_LONG 32
+#endif
+
+#include <asm-generic/bitsperlong.h>
+
+#endif /* _ASM_TILE_BITSPERLONG_H */
diff --git a/kernel/arch/tile/include/uapi/asm/byteorder.h b/kernel/arch/tile/include/uapi/asm/byteorder.h
new file mode 100644
index 000000000..6b8fa2e1c
--- /dev/null
+++ b/kernel/arch/tile/include/uapi/asm/byteorder.h
@@ -0,0 +1,19 @@
+/*
+ * Copyright 2011 Tilera Corporation. All Rights Reserved.
+ *
+ *   This program is free software; you can redistribute it and/or
+ *   modify it under the terms of the GNU General Public License
+ *   as published by the Free Software Foundation, version 2.
+ *
+ *   This program is distributed in the hope that it will be useful, but
+ *   WITHOUT ANY WARRANTY; without even the implied warranty of
+ *   MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
+ *   NON INFRINGEMENT.  See the GNU General Public License for
+ *   more details.
+ */
+
+#if defined (__BIG_ENDIAN__)
+#include <linux/byteorder/big_endian.h>
+#else
+#include <linux/byteorder/little_endian.h>
+#endif
diff --git a/kernel/arch/tile/include/uapi/asm/cachectl.h b/kernel/arch/tile/include/uapi/asm/cachectl.h
new file mode 100644
index 000000000..572ddcad2
--- /dev/null
+++ b/kernel/arch/tile/include/uapi/asm/cachectl.h
@@ -0,0 +1,42 @@
+/*
+ * Copyright 2011 Tilera Corporation. All Rights Reserved.
+ *
+ *   This program is free software; you can redistribute it and/or
+ *   modify it under the terms of the GNU General Public License
+ *   as published by the Free Software Foundation, version 2.
+ *
+ *   This program is distributed in the hope that it will be useful, but
+ *   WITHOUT ANY WARRANTY; without even the implied warranty of
+ *   MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
+ *   NON INFRINGEMENT.  See the GNU General Public License for
+ *   more details.
+ */
+
+#ifndef _ASM_TILE_CACHECTL_H
+#define _ASM_TILE_CACHECTL_H
+
+/*
+ * Options for cacheflush system call.
+ *
+ * The ICACHE flush is performed on all cores currently running the
+ * current process's address space.  The intent is for user
+ * applications to be able to modify code, invoke the system call,
+ * then allow arbitrary other threads in the same address space to see
+ * the newly-modified code.  Passing a length of CHIP_L1I_CACHE_SIZE()
+ * or more invalidates the entire icache on all cores in the address
+ * spaces.  (Note: currently this option invalidates the entire icache
+ * regardless of the requested address and length, but we may choose
+ * to honor the arguments at some point.)
+ *
+ * Flush and invalidation of memory can normally be performed with the
+ * __insn_flush() and __insn_finv() instructions from userspace.
+ * The DCACHE option to the system call allows userspace
+ * to flush the entire L1+L2 data cache from the core.  In this case,
+ * the address and length arguments are not used.  The DCACHE flush is
+ * restricted to the current core, not all cores in the address space.
+ */
+#define	ICACHE	(1<<0)		/* invalidate L1 instruction cache */
+#define	DCACHE	(1<<1)		/* flush and invalidate data cache */
+#define	BCACHE	(ICACHE|DCACHE)	/* flush both caches               */
+
+#endif	/* _ASM_TILE_CACHECTL_H */
diff --git a/kernel/arch/tile/include/uapi/asm/hardwall.h b/kernel/arch/tile/include/uapi/asm/hardwall.h
new file mode 100644
index 000000000..c2169d4f4
--- /dev/null
+++ b/kernel/arch/tile/include/uapi/asm/hardwall.h
@@ -0,0 +1,51 @@
+/*
+ * Copyright 2010 Tilera Corporation. All Rights Reserved.
+ *
+ *   This program is free software; you can redistribute it and/or
+ *   modify it under the terms of the GNU General Public License
+ *   as published by the Free Software Foundation, version 2.
+ *
+ *   This program is distributed in the hope that it will be useful, but
+ *   WITHOUT ANY WARRANTY; without even the implied warranty of
+ *   MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
+ *   NON INFRINGEMENT.  See the GNU General Public License for
+ *   more details.
+ *
+ * Provide methods for access control of per-cpu resources like
+ * UDN, IDN, or IPI.
+ */
+
+#ifndef _UAPI_ASM_TILE_HARDWALL_H
+#define _UAPI_ASM_TILE_HARDWALL_H
+
+#include <arch/chip.h>
+#include <linux/ioctl.h>
+
+#define HARDWALL_IOCTL_BASE 0xa2
+
+/*
+ * The HARDWALL_CREATE() ioctl is a macro with a "size" argument.
+ * The resulting ioctl value is passed to the kernel in conjunction
+ * with a pointer to a standard kernel bitmask of cpus.
+ * For network resources (UDN or IDN) the bitmask must physically
+ * represent a rectangular configuration on the chip.
+ * The "size" is the number of bytes of cpu mask data.
+ */
+#define _HARDWALL_CREATE 1
+#define HARDWALL_CREATE(size) \
+  _IOC(_IOC_READ, HARDWALL_IOCTL_BASE, _HARDWALL_CREATE, (size))
+
+#define _HARDWALL_ACTIVATE 2
+#define HARDWALL_ACTIVATE \
+  _IO(HARDWALL_IOCTL_BASE, _HARDWALL_ACTIVATE)
+
+#define _HARDWALL_DEACTIVATE 3
+#define HARDWALL_DEACTIVATE \
+ _IO(HARDWALL_IOCTL_BASE, _HARDWALL_DEACTIVATE)
+
+#define _HARDWALL_GET_ID 4
+#define HARDWALL_GET_ID \
+ _IO(HARDWALL_IOCTL_BASE, _HARDWALL_GET_ID)
+
+
+#endif /* _UAPI_ASM_TILE_HARDWALL_H */
diff --git a/kernel/arch/tile/include/uapi/asm/kvm_para.h b/kernel/arch/tile/include/uapi/asm/kvm_para.h
new file mode 100644
index 000000000..14fab8f0b
--- /dev/null
+++ b/kernel/arch/tile/include/uapi/asm/kvm_para.h
@@ -0,0 +1 @@
+#include <asm-generic/kvm_para.h>
diff --git a/kernel/arch/tile/include/uapi/asm/mman.h b/kernel/arch/tile/include/uapi/asm/mman.h
new file mode 100644
index 000000000..81b8fc348
--- /dev/null
+++ b/kernel/arch/tile/include/uapi/asm/mman.h
@@ -0,0 +1,41 @@
+/*
+ * Copyright 2010 Tilera Corporation. All Rights Reserved.
+ *
+ *   This program is free software; you can redistribute it and/or
+ *   modify it under the terms of the GNU General Public License
+ *   as published by the Free Software Foundation, version 2.
+ *
+ *   This program is distributed in the hope that it will be useful, but
+ *   WITHOUT ANY WARRANTY; without even the implied warranty of
+ *   MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
+ *   NON INFRINGEMENT.  See the GNU General Public License for
+ *   more details.
+ */
+
+#ifndef _ASM_TILE_MMAN_H
+#define _ASM_TILE_MMAN_H
+
+#include <asm-generic/mman-common.h>
+#include <arch/chip.h>
+
+/* Standard Linux flags */
+
+#define MAP_POPULATE	0x0040		/* populate (prefault) pagetables */
+#define MAP_NONBLOCK	0x0080		/* do not block on IO */
+#define MAP_GROWSDOWN	0x0100		/* stack-like segment */
+#define MAP_STACK	MAP_GROWSDOWN	/* provide convenience alias */
+#define MAP_LOCKED	0x0200		/* pages are locked */
+#define MAP_NORESERVE	0x0400		/* don't check for reservations */
+#define MAP_DENYWRITE	0x0800		/* ETXTBSY */
+#define MAP_EXECUTABLE	0x1000		/* mark it as an executable */
+#define MAP_HUGETLB	0x4000		/* create a huge page mapping */
+
+
+/*
+ * Flags for mlockall
+ */
+#define MCL_CURRENT	1		/* lock all current mappings */
+#define MCL_FUTURE	2		/* lock all future mappings */
+
+
+#endif /* _ASM_TILE_MMAN_H */
diff --git a/kernel/arch/tile/include/uapi/asm/ptrace.h b/kernel/arch/tile/include/uapi/asm/ptrace.h
new file mode 100644
index 000000000..d03b82985
--- /dev/null
+++ b/kernel/arch/tile/include/uapi/asm/ptrace.h
@@ -0,0 +1,98 @@
+/*
+ * Copyright 2010 Tilera Corporation. All Rights Reserved.
+ *
+ *   This program is free software; you can redistribute it and/or
+ *   modify it under the terms of the GNU General Public License
+ *   as published by the Free Software Foundation, version 2.
+ *
+ *   This program is distributed in the hope that it will be useful, but
+ *   WITHOUT ANY WARRANTY; without even the implied warranty of
+ *   MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
+ *   NON INFRINGEMENT.  See the GNU General Public License for
+ *   more details.
+ */
+
+#ifndef _UAPI_ASM_TILE_PTRACE_H
+#define _UAPI_ASM_TILE_PTRACE_H
+
+#include <arch/chip.h>
+#include <arch/abi.h>
+
+/* These must match struct pt_regs, below. */
+#if CHIP_WORD_SIZE() == 32
+#define PTREGS_OFFSET_REG(n)    ((n)*4)
+#else
+#define PTREGS_OFFSET_REG(n)    ((n)*8)
+#endif
+#define PTREGS_OFFSET_BASE      0
+#define PTREGS_OFFSET_TP        PTREGS_OFFSET_REG(53)
+#define PTREGS_OFFSET_SP        PTREGS_OFFSET_REG(54)
+#define PTREGS_OFFSET_LR        PTREGS_OFFSET_REG(55)
+#define PTREGS_NR_GPRS          56
+#define PTREGS_OFFSET_PC        PTREGS_OFFSET_REG(56)
+#define PTREGS_OFFSET_EX1       PTREGS_OFFSET_REG(57)
+#define PTREGS_OFFSET_FAULTNUM  PTREGS_OFFSET_REG(58)
+#define PTREGS_OFFSET_ORIG_R0   PTREGS_OFFSET_REG(59)
+#define PTREGS_OFFSET_FLAGS     PTREGS_OFFSET_REG(60)
+#if CHIP_HAS_CMPEXCH()
+#define PTREGS_OFFSET_CMPEXCH   PTREGS_OFFSET_REG(61)
+#endif
+#define PTREGS_SIZE             PTREGS_OFFSET_REG(64)
+
+
+#ifndef __ASSEMBLY__
+
+#ifndef __KERNEL__
+/* Provide appropriate length type to userspace regardless of -m32/-m64. */
+typedef uint_reg_t pt_reg_t;
+#endif
+
+/*
+ * This struct defines the way the registers are stored on the stack during a
+ * system call or exception.  "struct sigcontext" has the same shape.
+ */
+struct pt_regs {
+	union {
+		/* Saved main processor registers; 56..63 are special. */
+		pt_reg_t regs[56];
+		struct {
+			pt_reg_t __regs[53];
+			pt_reg_t tp;		/* aliases regs[TREG_TP] */
+			pt_reg_t sp;		/* aliases regs[TREG_SP] */
+			pt_reg_t lr;		/* aliases regs[TREG_LR] */
+		};
+	};
+
+	/* Saved special registers. */
+	pt_reg_t pc;		/* stored in EX_CONTEXT_K_0 */
+	pt_reg_t ex1;		/* stored in EX_CONTEXT_K_1 (PL and ICS bit) */
+	pt_reg_t faultnum;	/* fault number (INT_SWINT_1 for syscall) */
+	pt_reg_t orig_r0;	/* r0 at syscall entry, else zero */
+	pt_reg_t flags;		/* flags (see below) */
+#if !CHIP_HAS_CMPEXCH()
+	pt_reg_t pad[3];
+#else
+	pt_reg_t cmpexch;	/* value of CMPEXCH_VALUE SPR at interrupt */
+	pt_reg_t pad[2];
+#endif
+};
+
+#endif /* __ASSEMBLY__ */
+
+#define PTRACE_GETREGS		12
+#define PTRACE_SETREGS		13
+#define PTRACE_GETFPREGS	14
+#define PTRACE_SETFPREGS	15
+
+/* Support TILE-specific ptrace options, with events starting at 16. */
+#define PTRACE_EVENT_MIGRATE	16
+#define PTRACE_O_TRACEMIGRATE	(1 << PTRACE_EVENT_MIGRATE)
+
+/*
+ * Flag bits in pt_regs.flags that are part of the ptrace API.
+ * We start our numbering higher up to avoid confusion with the
+ * non-ABI kernel-internal values that use the low 16 bits.
+ */
+#define PT_FLAGS_COMPAT		0x10000  /* process is an -m32 compat process */
+
+#endif /* _UAPI_ASM_TILE_PTRACE_H */
diff --git a/kernel/arch/tile/include/uapi/asm/setup.h b/kernel/arch/tile/include/uapi/asm/setup.h
new file mode 100644
index 000000000..e6f7da265
--- /dev/null
+++ b/kernel/arch/tile/include/uapi/asm/setup.h
@@ -0,0 +1,21 @@
+/*
+ * Copyright 2010 Tilera Corporation. All Rights Reserved.
+ *
+ *   This program is free software; you can redistribute it and/or
+ *   modify it under the terms of the GNU General Public License
+ *   as published by the Free Software Foundation, version 2.
+ *
+ *   This program is distributed in the hope that it will be useful, but
+ *   WITHOUT ANY WARRANTY; without even the implied warranty of
+ *   MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
+ *   NON INFRINGEMENT.  See the GNU General Public License for
+ *   more details.
+ */
+
+#ifndef _UAPI_ASM_TILE_SETUP_H
+#define _UAPI_ASM_TILE_SETUP_H
+
+#define COMMAND_LINE_SIZE	2048
+
+
+#endif /* _UAPI_ASM_TILE_SETUP_H */
diff --git a/kernel/arch/tile/include/uapi/asm/sigcontext.h b/kernel/arch/tile/include/uapi/asm/sigcontext.h
new file mode 100644
index 000000000..39ff5d1a2
--- /dev/null
+++ b/kernel/arch/tile/include/uapi/asm/sigcontext.h
@@ -0,0 +1,43 @@
+/*
+ * Copyright 2010 Tilera Corporation. All Rights Reserved.
+ *
+ *   This program is free software; you can redistribute it and/or
+ *   modify it under the terms of the GNU General Public License
+ *   as published by the Free Software Foundation, version 2.
+ *
+ *   This program is distributed in the hope that it will be useful, but
+ *   WITHOUT ANY WARRANTY; without even the implied warranty of
+ *   MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
+ *   NON INFRINGEMENT.  See the GNU General Public License for
+ *   more details.
+ */
+
+#ifndef _ASM_TILE_SIGCONTEXT_H
+#define _ASM_TILE_SIGCONTEXT_H
+
+/* Don't pollute the namespace since <signal.h> includes this file. */
+#define __need_int_reg_t
+#include <arch/abi.h>
+
+/*
+ * struct sigcontext has the same shape as struct pt_regs,
+ * but is simplified since we know the fault is from userspace.
+ */
+struct sigcontext {
+	__extension__ union {
+		/* General-purpose registers.  */
+		__uint_reg_t gregs[56];
+		__extension__ struct {
+			__uint_reg_t __gregs[53];
+			__uint_reg_t tp;	/* Aliases gregs[TREG_TP].  */
+			__uint_reg_t sp;	/* Aliases gregs[TREG_SP].  */
+			__uint_reg_t lr;	/* Aliases gregs[TREG_LR].  */
+		};
+	};
+	__uint_reg_t pc;	/* Program counter.  */
+	__uint_reg_t ics;	/* In Interrupt Critical Section?  */
+	__uint_reg_t faultnum;	/* Fault number.  */
+	__uint_reg_t pad[5];
+};
+
+#endif /* _ASM_TILE_SIGCONTEXT_H */
diff --git a/kernel/arch/tile/include/uapi/asm/siginfo.h b/kernel/arch/tile/include/uapi/asm/siginfo.h
new file mode 100644
index 000000000..56d661bb0
--- /dev/null
+++ b/kernel/arch/tile/include/uapi/asm/siginfo.h
@@ -0,0 +1,34 @@
+/*
+ * Copyright 2010 Tilera Corporation. All Rights Reserved.
+ *
+ *   This program is free software; you can redistribute it and/or
+ *   modify it under the terms of the GNU General Public License
+ *   as published by the Free Software Foundation, version 2.
+ *
+ *   This program is distributed in the hope that it will be useful, but
+ *   WITHOUT ANY WARRANTY; without even the implied warranty of
+ *   MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
+ *   NON INFRINGEMENT.  See the GNU General Public License for
+ *   more details.
+ */
+
+#ifndef _ASM_TILE_SIGINFO_H
+#define _ASM_TILE_SIGINFO_H
+
+#define __ARCH_SI_TRAPNO
+
+#ifdef __LP64__
+# define __ARCH_SI_PREAMBLE_SIZE	(4 * sizeof(int))
+#endif
+
+#include <asm-generic/siginfo.h>
+
+/*
+ * Additional Tile-specific SIGILL si_codes
+ */
+#define ILL_DBLFLT	(__SI_FAULT|9)	/* double fault */
+#define ILL_HARDWALL	(__SI_FAULT|10)	/* user networks hardwall violation */
+#undef NSIGILL
+#define NSIGILL		10
+
+#endif /* _ASM_TILE_SIGINFO_H */
diff --git a/kernel/arch/tile/include/uapi/asm/signal.h b/kernel/arch/tile/include/uapi/asm/signal.h
new file mode 100644
index 000000000..ef0d32d84
--- /dev/null
+++ b/kernel/arch/tile/include/uapi/asm/signal.h
@@ -0,0 +1,27 @@
+/*
+ * Copyright 2010 Tilera Corporation. All Rights Reserved.
+ *
+ *   This program is free software; you can redistribute it and/or
+ *   modify it under the terms of the GNU General Public License
+ *   as published by the Free Software Foundation, version 2.
+ *
+ *   This program is distributed in the hope that it will be useful, but
+ *   WITHOUT ANY WARRANTY; without even the implied warranty of
+ *   MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
+ *   NON INFRINGEMENT.  See the GNU General Public License for
+ *   more details.
+ */
+
+#ifndef _UAPI_ASM_TILE_SIGNAL_H
+#define _UAPI_ASM_TILE_SIGNAL_H
+
+/* Do not notify a ptracer when this signal is handled. */
+#define SA_NOPTRACE 0x02000000u
+
+/* Used in earlier Tilera releases, so keeping for binary compatibility. */
+#define SA_RESTORER 0x04000000u
+
+#include <asm-generic/signal.h>
+
+
+#endif /* _UAPI_ASM_TILE_SIGNAL_H */
diff --git a/kernel/arch/tile/include/uapi/asm/stat.h b/kernel/arch/tile/include/uapi/asm/stat.h
new file mode 100644
index 000000000..c0db34d56
--- /dev/null
+++ b/kernel/arch/tile/include/uapi/asm/stat.h
@@ -0,0 +1,4 @@
+#if defined(__KERNEL__) && defined(CONFIG_COMPAT)
+#define __ARCH_WANT_STAT64	/* Used for compat_sys_stat64() etc. */
+#endif
+#include <asm-generic/stat.h>
diff --git a/kernel/arch/tile/include/uapi/asm/swab.h b/kernel/arch/tile/include/uapi/asm/swab.h
new file mode 100644
index 000000000..7c37b38f6
--- /dev/null
+++ b/kernel/arch/tile/include/uapi/asm/swab.h
@@ -0,0 +1,23 @@
+/*
+ * Copyright 2010 Tilera Corporation. All Rights Reserved.
+ *
+ *   This program is free software; you can redistribute it and/or
+ *   modify it under the terms of the GNU General Public License
+ *   as published by the Free Software Foundation, version 2.
+ *
+ *   This program is distributed in the hope that it will be useful, but
+ *   WITHOUT ANY WARRANTY; without even the implied warranty of
+ *   MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
+ *   NON INFRINGEMENT.  See the GNU General Public License for
+ *   more details.
+ */
+
+#ifndef _ASM_TILE_SWAB_H
+#define _ASM_TILE_SWAB_H
+
+/* Tile gcc is always >= 4.3.0, so we use __builtin_bswap. */
+#define __arch_swab32(x) __builtin_bswap32(x)
+#define __arch_swab64(x) __builtin_bswap64(x)
+#define __arch_swab16(x) (__builtin_bswap32(x) >> 16)
+
+#endif /* _ASM_TILE_SWAB_H */
diff --git a/kernel/arch/tile/include/uapi/asm/unistd.h b/kernel/arch/tile/include/uapi/asm/unistd.h
new file mode 100644
index 000000000..3866397aa
--- /dev/null
+++ b/kernel/arch/tile/include/uapi/asm/unistd.h
@@ -0,0 +1,36 @@
+/*
+ * Copyright 2010 Tilera Corporation. All Rights Reserved.
+ *
+ *   This program is free software; you can redistribute it and/or
+ *   modify it under the terms of the GNU General Public License
+ *   as published by the Free Software Foundation, version 2.
+ *
+ *   This program is distributed in the hope that it will be useful, but
+ *   WITHOUT ANY WARRANTY; without even the implied warranty of
+ *   MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
+ *   NON INFRINGEMENT.  See the GNU General Public License for
+ *   more details.
+ */
+
+#if !defined(__LP64__) || defined(__SYSCALL_COMPAT)
+/* Use the flavor of this syscall that matches the 32-bit API better. */
+#define __ARCH_WANT_SYNC_FILE_RANGE2
+#endif
+
+/* Use the standard ABI for syscalls. */
+#include <asm-generic/unistd.h>
+
+#define NR_syscalls __NR_syscalls
+
+/* Additional Tilera-specific syscalls. */
+#define __NR_cacheflush	(__NR_arch_specific_syscall + 1)
+__SYSCALL(__NR_cacheflush, sys_cacheflush)
+
+#ifndef __tilegx__
+/* "Fast" syscalls provide atomic support for 32-bit chips. */
+#define __NR_FAST_cmpxchg	-1
+#define __NR_FAST_atomic_update	-2
+#define __NR_FAST_cmpxchg64	-3
+#define __NR_cmpxchg_badaddr	(__NR_arch_specific_syscall + 0)
+__SYSCALL(__NR_cmpxchg_badaddr, sys_cmpxchg_badaddr)
+#endif
diff --git a/kernel/arch/tile/kernel/Makefile b/kernel/arch/tile/kernel/Makefile
new file mode 100644
index 000000000..21f77bf68
--- /dev/null
+++ b/kernel/arch/tile/kernel/Makefile
@@ -0,0 +1,36 @@
+#
+# Makefile for the Linux/TILE kernel.
+#
+
+extra-y := vmlinux.lds head_$(BITS).o
+obj-y := backtrace.o entry.o hvglue.o irq.o messaging.o \
+	pci-dma.o proc.o process.o ptrace.o reboot.o \
+	setup.o signal.o single_step.o stack.o sys.o \
+	sysfs.o time.o traps.o unaligned.o vdso.o \
+	intvec_$(BITS).o regs_$(BITS).o tile-desc_$(BITS).o
+
+ifdef CONFIG_FUNCTION_TRACER
+CFLAGS_REMOVE_ftrace.o = -pg
+CFLAGS_REMOVE_early_printk.o = -pg
+endif
+
+obj-$(CONFIG_HARDWALL)		+= hardwall.o
+obj-$(CONFIG_COMPAT)		+= compat.o compat_signal.o
+obj-$(CONFIG_SMP)		+= smpboot.o smp.o tlb.o
+obj-$(CONFIG_MODULES)		+= module.o
+obj-$(CONFIG_EARLY_PRINTK)	+= early_printk.o
+obj-$(CONFIG_KEXEC)		+= machine_kexec.o relocate_kernel_$(BITS).o
+ifdef CONFIG_TILEGX
+obj-$(CONFIG_PCI)		+= pci_gx.o
+else
+obj-$(CONFIG_PCI)		+= pci.o
+endif
+obj-$(CONFIG_PERF_EVENTS)	+= perf_event.o
+obj-$(CONFIG_USE_PMC)		+= pmc.o
+obj-$(CONFIG_TILE_USB)		+= usb.o
+obj-$(CONFIG_TILE_HVGLUE_TRACE)	+= hvglue_trace.o
+obj-$(CONFIG_FUNCTION_TRACER)	+= ftrace.o mcount_64.o
+obj-$(CONFIG_KPROBES)		+= kprobes.o
+obj-$(CONFIG_KGDB)		+= kgdb.o
+
+obj-y				+= vdso/
diff --git a/kernel/arch/tile/kernel/asm-offsets.c b/kernel/arch/tile/kernel/asm-offsets.c
new file mode 100644
index 000000000..375e7c321
--- /dev/null
+++ b/kernel/arch/tile/kernel/asm-offsets.c
@@ -0,0 +1,84 @@
+/*
+ * Copyright 2010 Tilera Corporation. All Rights Reserved.
+ *
+ *   This program is free software; you can redistribute it and/or
+ *   modify it under the terms of the GNU General Public License
+ *   as published by the Free Software Foundation, version 2.
+ *
+ *   This program is distributed in the hope that it will be useful, but
+ *   WITHOUT ANY WARRANTY; without even the implied warranty of
+ *   MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
+ *   NON INFRINGEMENT.  See the GNU General Public License for
+ *   more details.
+ *
+ * Generates definitions from c-type structures used by assembly sources.
+ */
+
+/* Check for compatible compiler early in the build. */
+#ifdef CONFIG_TILEGX
+# ifndef __tilegx__
+#  error Can only build TILE-Gx configurations with tilegx compiler
+# endif
+# ifndef __LP64__
+#  error Must not specify -m32 when building the TILE-Gx kernel
+# endif
+#else
+# ifdef __tilegx__
+#  error Can not build TILEPro configurations with tilegx compiler
+# endif
+#endif
+
+#include <linux/kbuild.h>
+#include <linux/thread_info.h>
+#include <linux/sched.h>
+#include <linux/hardirq.h>
+#include <linux/ptrace.h>
+#include <hv/hypervisor.h>
+
+void foo(void)
+{
+	DEFINE(SINGLESTEP_STATE_BUFFER_OFFSET,
+	       offsetof(struct single_step_state, buffer));
+	DEFINE(SINGLESTEP_STATE_FLAGS_OFFSET,
+	       offsetof(struct single_step_state, flags));
+	DEFINE(SINGLESTEP_STATE_ORIG_PC_OFFSET,
+	       offsetof(struct single_step_state, orig_pc));
+	DEFINE(SINGLESTEP_STATE_NEXT_PC_OFFSET,
+	       offsetof(struct single_step_state, next_pc));
+	DEFINE(SINGLESTEP_STATE_BRANCH_NEXT_PC_OFFSET,
+	       offsetof(struct single_step_state, branch_next_pc));
+	DEFINE(SINGLESTEP_STATE_UPDATE_VALUE_OFFSET,
+	       offsetof(struct single_step_state, update_value));
+
+	DEFINE(THREAD_INFO_TASK_OFFSET,
+	       offsetof(struct thread_info, task));
+	DEFINE(THREAD_INFO_FLAGS_OFFSET,
+	       offsetof(struct thread_info, flags));
+	DEFINE(THREAD_INFO_STATUS_OFFSET,
+	       offsetof(struct thread_info, status));
+	DEFINE(THREAD_INFO_HOMECACHE_CPU_OFFSET,
+	       offsetof(struct thread_info, homecache_cpu));
+	DEFINE(THREAD_INFO_PREEMPT_COUNT_OFFSET,
+	       offsetof(struct thread_info, preempt_count));
+	DEFINE(THREAD_INFO_STEP_STATE_OFFSET,
+	       offsetof(struct thread_info, step_state));
+#ifdef __tilegx__
+	DEFINE(THREAD_INFO_UNALIGN_JIT_BASE_OFFSET,
+	       offsetof(struct thread_info, unalign_jit_base));
+	DEFINE(THREAD_INFO_UNALIGN_JIT_TMP_OFFSET,
+	       offsetof(struct thread_info, unalign_jit_tmp));
+#endif
+
+	DEFINE(TASK_STRUCT_THREAD_KSP_OFFSET,
+	       offsetof(struct task_struct, thread.ksp));
+	DEFINE(TASK_STRUCT_THREAD_PC_OFFSET,
+	       offsetof(struct task_struct, thread.pc));
+
+	DEFINE(HV_TOPOLOGY_WIDTH_OFFSET,
+	       offsetof(HV_Topology, width));
+	DEFINE(HV_TOPOLOGY_HEIGHT_OFFSET,
+	       offsetof(HV_Topology, height));
+
+	DEFINE(IRQ_CPUSTAT_SYSCALL_COUNT_OFFSET,
+	       offsetof(irq_cpustat_t, irq_syscall_count));
+}
diff --git a/kernel/arch/tile/kernel/backtrace.c b/kernel/arch/tile/kernel/backtrace.c
new file mode 100644
index 000000000..f8b74ca83
--- /dev/null
+++ b/kernel/arch/tile/kernel/backtrace.c
@@ -0,0 +1,683 @@
+/*
+ * Copyright 2011 Tilera Corporation. All Rights Reserved.
+ *
+ *   This program is free software; you can redistribute it and/or
+ *   modify it under the terms of the GNU General Public License
+ *   as published by the Free Software Foundation, version 2.
+ *
+ *   This program is distributed in the hope that it will be useful, but
+ *   WITHOUT ANY WARRANTY; without even the implied warranty of
+ *   MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
+ *   NON INFRINGEMENT.  See the GNU General Public License for
+ *   more details.
+ */
+
+#include <linux/kernel.h>
+#include <linux/string.h>
+#include <asm/byteorder.h>
+#include <asm/backtrace.h>
+#include <asm/tile-desc.h>
+#include <arch/abi.h>
+
+#ifdef __tilegx__
+#define TILE_MAX_INSTRUCTIONS_PER_BUNDLE TILEGX_MAX_INSTRUCTIONS_PER_BUNDLE
+#define tile_decoded_instruction tilegx_decoded_instruction
+#define tile_mnemonic tilegx_mnemonic
+#define parse_insn_tile parse_insn_tilegx
+#define TILE_OPC_IRET TILEGX_OPC_IRET
+#define TILE_OPC_ADDI TILEGX_OPC_ADDI
+#define TILE_OPC_ADDLI TILEGX_OPC_ADDLI
+#define TILE_OPC_INFO TILEGX_OPC_INFO
+#define TILE_OPC_INFOL TILEGX_OPC_INFOL
+#define TILE_OPC_JRP TILEGX_OPC_JRP
+#define TILE_OPC_MOVE TILEGX_OPC_MOVE
+#define OPCODE_STORE TILEGX_OPC_ST
+typedef long long bt_int_reg_t;
+#else
+#define TILE_MAX_INSTRUCTIONS_PER_BUNDLE TILEPRO_MAX_INSTRUCTIONS_PER_BUNDLE
+#define tile_decoded_instruction tilepro_decoded_instruction
+#define tile_mnemonic tilepro_mnemonic
+#define parse_insn_tile parse_insn_tilepro
+#define TILE_OPC_IRET TILEPRO_OPC_IRET
+#define TILE_OPC_ADDI TILEPRO_OPC_ADDI
+#define TILE_OPC_ADDLI TILEPRO_OPC_ADDLI
+#define TILE_OPC_INFO TILEPRO_OPC_INFO
+#define TILE_OPC_INFOL TILEPRO_OPC_INFOL
+#define TILE_OPC_JRP TILEPRO_OPC_JRP
+#define TILE_OPC_MOVE TILEPRO_OPC_MOVE
+#define OPCODE_STORE TILEPRO_OPC_SW
+typedef int bt_int_reg_t;
+#endif
+
+/* A decoded bundle used for backtracer analysis. */
+struct BacktraceBundle {
+	tile_bundle_bits bits;
+	int num_insns;
+	struct tile_decoded_instruction
+	insns[TILE_MAX_INSTRUCTIONS_PER_BUNDLE];
+};
+
+
+/* Locates an instruction inside the given bundle that
+ * has the specified mnemonic, and whose first 'num_operands_to_match'
+ * operands exactly match those in 'operand_values'.
+ */
+static const struct tile_decoded_instruction *find_matching_insn(
+	const struct BacktraceBundle *bundle,
+	tile_mnemonic mnemonic,
+	const int *operand_values,
+	int num_operands_to_match)
+{
+	int i, j;
+	bool match;
+
+	for (i = 0; i < bundle->num_insns; i++) {
+		const struct tile_decoded_instruction *insn =
+			&bundle->insns[i];
+
+		if (insn->opcode->mnemonic != mnemonic)
+			continue;
+
+		match = true;
+		for (j = 0; j < num_operands_to_match; j++) {
+			if (operand_values[j] != insn->operand_values[j]) {
+				match = false;
+				break;
+			}
+		}
+
+		if (match)
+			return insn;
+	}
+
+	return NULL;
+}
+
+/* Does this bundle contain an 'iret' instruction? */
+static inline bool bt_has_iret(const struct BacktraceBundle *bundle)
+{
+	return find_matching_insn(bundle, TILE_OPC_IRET, NULL, 0) != NULL;
+}
+
+/* Does this bundle contain an 'addi sp, sp, OFFSET' or
+ * 'addli sp, sp, OFFSET' instruction, and if so, what is OFFSET?
+ */
+static bool bt_has_addi_sp(const struct BacktraceBundle *bundle, int *adjust)
+{
+	static const int vals[2] = { TREG_SP, TREG_SP };
+
+	const struct tile_decoded_instruction *insn =
+		find_matching_insn(bundle, TILE_OPC_ADDI, vals, 2);
+	if (insn == NULL)
+		insn = find_matching_insn(bundle, TILE_OPC_ADDLI, vals, 2);
+#ifdef __tilegx__
+	if (insn == NULL)
+		insn = find_matching_insn(bundle, TILEGX_OPC_ADDXLI, vals, 2);
+	if (insn == NULL)
+		insn = find_matching_insn(bundle, TILEGX_OPC_ADDXI, vals, 2);
+#endif
+	if (insn == NULL)
+		return false;
+
+	*adjust = insn->operand_values[2];
+	return true;
+}
+
+/* Does this bundle contain any 'info OP' or 'infol OP'
+ * instruction, and if so, what are their OP?  Note that OP is interpreted
+ * as an unsigned value by this code since that's what the caller wants.
+ * Returns the number of info ops found.
+ */
+static int bt_get_info_ops(const struct BacktraceBundle *bundle,
+		int operands[MAX_INFO_OPS_PER_BUNDLE])
+{
+	int num_ops = 0;
+	int i;
+
+	for (i = 0; i < bundle->num_insns; i++) {
+		const struct tile_decoded_instruction *insn =
+			&bundle->insns[i];
+
+		if (insn->opcode->mnemonic == TILE_OPC_INFO ||
+		    insn->opcode->mnemonic == TILE_OPC_INFOL) {
+			operands[num_ops++] = insn->operand_values[0];
+		}
+	}
+
+	return num_ops;
+}
+
+/* Does this bundle contain a jrp instruction, and if so, to which
+ * register is it jumping?
+ */
+static bool bt_has_jrp(const struct BacktraceBundle *bundle, int *target_reg)
+{
+	const struct tile_decoded_instruction *insn =
+		find_matching_insn(bundle, TILE_OPC_JRP, NULL, 0);
+	if (insn == NULL)
+		return false;
+
+	*target_reg = insn->operand_values[0];
+	return true;
+}
+
+/* Does this bundle modify the specified register in any way? */
+static bool bt_modifies_reg(const struct BacktraceBundle *bundle, int reg)
+{
+	int i, j;
+	for (i = 0; i < bundle->num_insns; i++) {
+		const struct tile_decoded_instruction *insn =
+			&bundle->insns[i];
+
+		if (insn->opcode->implicitly_written_register == reg)
+			return true;
+
+		for (j = 0; j < insn->opcode->num_operands; j++)
+			if (insn->operands[j]->is_dest_reg &&
+			    insn->operand_values[j] == reg)
+				return true;
+	}
+
+	return false;
+}
+
+/* Does this bundle modify sp? */
+static inline bool bt_modifies_sp(const struct BacktraceBundle *bundle)
+{
+	return bt_modifies_reg(bundle, TREG_SP);
+}
+
+/* Does this bundle modify lr? */
+static inline bool bt_modifies_lr(const struct BacktraceBundle *bundle)
+{
+	return bt_modifies_reg(bundle, TREG_LR);
+}
+
+/* Does this bundle contain the instruction 'move fp, sp'? */
+static inline bool bt_has_move_r52_sp(const struct BacktraceBundle *bundle)
+{
+	static const int vals[2] = { 52, TREG_SP };
+	return find_matching_insn(bundle, TILE_OPC_MOVE, vals, 2) != NULL;
+}
+
+/* Does this bundle contain a store of lr to sp? */
+static inline bool bt_has_sw_sp_lr(const struct BacktraceBundle *bundle)
+{
+	static const int vals[2] = { TREG_SP, TREG_LR };
+	return find_matching_insn(bundle, OPCODE_STORE, vals, 2) != NULL;
+}
+
+#ifdef __tilegx__
+/* Track moveli values placed into registers. */
+static inline void bt_update_moveli(const struct BacktraceBundle *bundle,
+				    int moveli_args[])
+{
+	int i;
+	for (i = 0; i < bundle->num_insns; i++) {
+		const struct tile_decoded_instruction *insn =
+			&bundle->insns[i];
+
+		if (insn->opcode->mnemonic == TILEGX_OPC_MOVELI) {
+			int reg = insn->operand_values[0];
+			moveli_args[reg] = insn->operand_values[1];
+		}
+	}
+}
+
+/* Does this bundle contain an 'add sp, sp, reg' instruction
+ * from a register that we saw a moveli into, and if so, what
+ * is the value in the register?
+ */
+static bool bt_has_add_sp(const struct BacktraceBundle *bundle, int *adjust,
+			  int moveli_args[])
+{
+	static const int vals[2] = { TREG_SP, TREG_SP };
+
+	const struct tile_decoded_instruction *insn =
+		find_matching_insn(bundle, TILEGX_OPC_ADDX, vals, 2);
+	if (insn) {
+		int reg = insn->operand_values[2];
+		if (moveli_args[reg]) {
+			*adjust = moveli_args[reg];
+			return true;
+		}
+	}
+	return false;
+}
+#endif
+
+/* Locates the caller's PC and SP for a program starting at the
+ * given address.
+ */
+static void find_caller_pc_and_caller_sp(CallerLocation *location,
+					 const unsigned long start_pc,
+					 BacktraceMemoryReader read_memory_func,
+					 void *read_memory_func_extra)
+{
+	/* Have we explicitly decided what the sp is,
+	 * rather than just the default?
+	 */
+	bool sp_determined = false;
+
+	/* Has any bundle seen so far modified lr? */
+	bool lr_modified = false;
+
+	/* Have we seen a move from sp to fp? */
+	bool sp_moved_to_r52 = false;
+
+	/* Have we seen a terminating bundle? */
+	bool seen_terminating_bundle = false;
+
+	/* Cut down on round-trip reading overhead by reading several
+	 * bundles at a time.
+	 */
+	tile_bundle_bits prefetched_bundles[32];
+	int num_bundles_prefetched = 0;
+	int next_bundle = 0;
+	unsigned long pc;
+
+#ifdef __tilegx__
+	/* Naively try to track moveli values to support addx for -m32. */
+	int moveli_args[TILEGX_NUM_REGISTERS] = { 0 };
+#endif
+
+	/* Default to assuming that the caller's sp is the current sp.
+	 * This is necessary to handle the case where we start backtracing
+	 * right at the end of the epilog.
+	 */
+	location->sp_location = SP_LOC_OFFSET;
+	location->sp_offset = 0;
+
+	/* Default to having no idea where the caller PC is. */
+	location->pc_location = PC_LOC_UNKNOWN;
+
+	/* Don't even try if the PC is not aligned. */
+	if (start_pc % TILE_BUNDLE_ALIGNMENT_IN_BYTES != 0)
+		return;
+
+	for (pc = start_pc;; pc += sizeof(tile_bundle_bits)) {
+
+		struct BacktraceBundle bundle;
+		int num_info_ops, info_operands[MAX_INFO_OPS_PER_BUNDLE];
+		int one_ago, jrp_reg;
+		bool has_jrp;
+
+		if (next_bundle >= num_bundles_prefetched) {
+			/* Prefetch some bytes, but don't cross a page
+			 * boundary since that might cause a read failure we
+			 * don't care about if we only need the first few
+			 * bytes. Note: we don't care what the actual page
+			 * size is; using the minimum possible page size will
+			 * prevent any problems.
+			 */
+			unsigned int bytes_to_prefetch = 4096 - (pc & 4095);
+			if (bytes_to_prefetch > sizeof prefetched_bundles)
+				bytes_to_prefetch = sizeof prefetched_bundles;
+
+			if (!read_memory_func(prefetched_bundles, pc,
+					      bytes_to_prefetch,
+					      read_memory_func_extra)) {
+				if (pc == start_pc) {
+					/* The program probably called a bad
+					 * address, such as a NULL pointer.
+					 * So treat this as if we are at the
+					 * start of the function prolog so the
+					 * backtrace will show how we got here.
+					 */
+					location->pc_location = PC_LOC_IN_LR;
+					return;
+				}
+
+				/* Unreadable address. Give up. */
+				break;
+			}
+
+			next_bundle = 0;
+			num_bundles_prefetched =
+				bytes_to_prefetch / sizeof(tile_bundle_bits);
+		}
+
+		/*
+		 * Decode the next bundle.
+		 * TILE always stores instruction bundles in little-endian
+		 * mode, even when the chip is running in big-endian mode.
+		 */
+		bundle.bits = le64_to_cpu(prefetched_bundles[next_bundle++]);
+		bundle.num_insns =
+			parse_insn_tile(bundle.bits, pc, bundle.insns);
+		num_info_ops = bt_get_info_ops(&bundle, info_operands);
+
+		/* First look at any one_ago info ops if they are interesting,
+		 * since they should shadow any non-one-ago info ops.
+		 */
+		for (one_ago = (pc != start_pc) ? 1 : 0;
+		     one_ago >= 0; one_ago--) {
+			int i;
+			for (i = 0; i < num_info_ops; i++) {
+				int info_operand = info_operands[i];
+				if (info_operand < CALLER_UNKNOWN_BASE)	{
+					/* Weird; reserved value, ignore it. */
+					continue;
+				}
+
+				/* Skip info ops which are not in the
+				 * "one_ago" mode we want right now.
+				 */
+				if (((info_operand & ONE_BUNDLE_AGO_FLAG) != 0)
+				    != (one_ago != 0))
+					continue;
+
+				/* Clear the flag to make later checking
+				 * easier. */
+				info_operand &= ~ONE_BUNDLE_AGO_FLAG;
+
+				/* Default to looking at PC_IN_LR_FLAG. */
+				if (info_operand & PC_IN_LR_FLAG)
+					location->pc_location =
+						PC_LOC_IN_LR;
+				else
+					location->pc_location =
+						PC_LOC_ON_STACK;
+
+				switch (info_operand) {
+				case CALLER_UNKNOWN_BASE:
+					location->pc_location = PC_LOC_UNKNOWN;
+					location->sp_location = SP_LOC_UNKNOWN;
+					return;
+
+				case CALLER_SP_IN_R52_BASE:
+				case CALLER_SP_IN_R52_BASE | PC_IN_LR_FLAG:
+					location->sp_location = SP_LOC_IN_R52;
+					return;
+
+				default:
+				{
+					const unsigned int val = info_operand
+						- CALLER_SP_OFFSET_BASE;
+					const unsigned int sp_offset =
+						(val >> NUM_INFO_OP_FLAGS) * 8;
+					if (sp_offset < 32768) {
+						/* This is a properly encoded
+						 * SP offset. */
+						location->sp_location =
+							SP_LOC_OFFSET;
+						location->sp_offset =
+							sp_offset;
+						return;
+					} else {
+						/* This looked like an SP
+						 * offset, but it's outside
+						 * the legal range, so this
+						 * must be an unrecognized
+						 * info operand.  Ignore it.
+						 */
+					}
+				}
+				break;
+				}
+			}
+		}
+
+		if (seen_terminating_bundle) {
+			/* We saw a terminating bundle during the previous
+			 * iteration, so we were only looking for an info op.
+			 */
+			break;
+		}
+
+		if (bundle.bits == 0) {
+			/* Wacky terminating bundle. Stop looping, and hope
+			 * we've already seen enough to find the caller.
+			 */
+			break;
+		}
+
+		/*
+		 * Try to determine caller's SP.
+		 */
+
+		if (!sp_determined) {
+			int adjust;
+			if (bt_has_addi_sp(&bundle, &adjust)
+#ifdef __tilegx__
+			    || bt_has_add_sp(&bundle, &adjust, moveli_args)
+#endif
+				) {
+				location->sp_location = SP_LOC_OFFSET;
+
+				if (adjust <= 0) {
+					/* We are in prolog about to adjust
+					 * SP. */
+					location->sp_offset = 0;
+				} else {
+					/* We are in epilog restoring SP. */
+					location->sp_offset = adjust;
+				}
+
+				sp_determined = true;
+			} else {
+				if (bt_has_move_r52_sp(&bundle)) {
+					/* Maybe in prolog, creating an
+					 * alloca-style frame.  But maybe in
+					 * the middle of a fixed-size frame
+					 * clobbering r52 with SP.
+					 */
+					sp_moved_to_r52 = true;
+				}
+
+				if (bt_modifies_sp(&bundle)) {
+					if (sp_moved_to_r52) {
+						/* We saw SP get saved into
+						 * r52 earlier (or now), which
+						 * must have been in the
+						 * prolog, so we now know that
+						 * SP is still holding the
+						 * caller's sp value.
+						 */
+						location->sp_location =
+							SP_LOC_OFFSET;
+						location->sp_offset = 0;
+					} else {
+						/* Someone must have saved
+						 * aside the caller's SP value
+						 * into r52, so r52 holds the
+						 * current value.
+						 */
+						location->sp_location =
+							SP_LOC_IN_R52;
+					}
+					sp_determined = true;
+				}
+			}
+
+#ifdef __tilegx__
+			/* Track moveli arguments for -m32 mode. */
+			bt_update_moveli(&bundle, moveli_args);
+#endif
+		}
+
+		if (bt_has_iret(&bundle)) {
+			/* This is a terminating bundle. */
+			seen_terminating_bundle = true;
+			continue;
+		}
+
+		/*
+		 * Try to determine caller's PC.
+		 */
+
+		jrp_reg = -1;
+		has_jrp = bt_has_jrp(&bundle, &jrp_reg);
+		if (has_jrp)
+			seen_terminating_bundle = true;
+
+		if (location->pc_location == PC_LOC_UNKNOWN) {
+			if (has_jrp) {
+				if (jrp_reg == TREG_LR && !lr_modified) {
+					/* Looks like a leaf function, or else
+					 * lr is already restored. */
+					location->pc_location =
+						PC_LOC_IN_LR;
+				} else {
+					location->pc_location =
+						PC_LOC_ON_STACK;
+				}
+			} else if (bt_has_sw_sp_lr(&bundle)) {
+				/* In prolog, spilling initial lr to stack. */
+				location->pc_location = PC_LOC_IN_LR;
+			} else if (bt_modifies_lr(&bundle)) {
+				lr_modified = true;
+			}
+		}
+	}
+}
+
+/* Initializes a backtracer to start from the given location.
+ *
+ * If the frame pointer cannot be determined it is set to -1.
+ *
+ * state: The state to be filled in.
+ * read_memory_func: A callback that reads memory.
+ * read_memory_func_extra: An arbitrary argument to read_memory_func.
+ * pc: The current PC.
+ * lr: The current value of the 'lr' register.
+ * sp: The current value of the 'sp' register.
+ * r52: The current value of the 'r52' register.
+ */
+void backtrace_init(BacktraceIterator *state,
+		    BacktraceMemoryReader read_memory_func,
+		    void *read_memory_func_extra,
+		    unsigned long pc, unsigned long lr,
+		    unsigned long sp, unsigned long r52)
+{
+	CallerLocation location;
+	unsigned long fp, initial_frame_caller_pc;
+
+	/* Find out where we are in the initial frame. */
+	find_caller_pc_and_caller_sp(&location, pc,
+				     read_memory_func, read_memory_func_extra);
+
+	switch (location.sp_location) {
+	case SP_LOC_UNKNOWN:
+		/* Give up. */
+		fp = -1;
+		break;
+
+	case SP_LOC_IN_R52:
+		fp = r52;
+		break;
+
+	case SP_LOC_OFFSET:
+		fp = sp + location.sp_offset;
+		break;
+
+	default:
+		/* Give up. */
+		fp = -1;
+		break;
+	}
+
+	/* If the frame pointer is not aligned to the basic word size
+	 * something terrible happened and we should mark it as invalid.
+	 */
+	if (fp % sizeof(bt_int_reg_t) != 0)
+		fp = -1;
+
+	/* -1 means "don't know initial_frame_caller_pc". */
+	initial_frame_caller_pc = -1;
+
+	switch (location.pc_location) {
+	case PC_LOC_UNKNOWN:
+		/* Give up. */
+		fp = -1;
+		break;
+
+	case PC_LOC_IN_LR:
+		if (lr == 0 || lr % TILE_BUNDLE_ALIGNMENT_IN_BYTES != 0) {
+			/* Give up. */
+			fp = -1;
+		} else {
+			initial_frame_caller_pc = lr;
+		}
+		break;
+
+	case PC_LOC_ON_STACK:
+		/* Leave initial_frame_caller_pc as -1,
+		 * meaning check the stack.
+		 */
+		break;
+
+	default:
+		/* Give up. */
+		fp = -1;
+		break;
+	}
+
+	state->pc = pc;
+	state->sp = sp;
+	state->fp = fp;
+	state->initial_frame_caller_pc = initial_frame_caller_pc;
+	state->read_memory_func = read_memory_func;
+	state->read_memory_func_extra = read_memory_func_extra;
+}
+
+/* Handle the case where the register holds more bits than the VA. */
+static bool valid_addr_reg(bt_int_reg_t reg)
+{
+	return ((unsigned long)reg == reg);
+}
+
+/* Advances the backtracing state to the calling frame, returning
+ * true iff successful.
+ */
+bool backtrace_next(BacktraceIterator *state)
+{
+	unsigned long next_fp, next_pc;
+	bt_int_reg_t next_frame[2];
+
+	if (state->fp == -1) {
+		/* No parent frame. */
+		return false;
+	}
+
+	/* Try to read the frame linkage data chaining to the next function. */
+	if (!state->read_memory_func(&next_frame, state->fp, sizeof next_frame,
+				     state->read_memory_func_extra)) {
+		return false;
+	}
+
+	next_fp = next_frame[1];
+	if (!valid_addr_reg(next_frame[1]) ||
+	    next_fp % sizeof(bt_int_reg_t) != 0) {
+		/* Caller's frame pointer is suspect, so give up. */
+		return false;
+	}
+
+	if (state->initial_frame_caller_pc != -1) {
+		/* We must be in the initial stack frame and already know the
+		 * caller PC.
+		 */
+		next_pc = state->initial_frame_caller_pc;
+
+		/* Force reading stack next time, in case we were in the
+		 * initial frame.  We don't do this above just to paranoidly
+		 * avoid changing the struct at all when we return false.
+		 */
+		state->initial_frame_caller_pc = -1;
+	} else {
+		/* Get the caller PC from the frame linkage area. */
+		next_pc = next_frame[0];
+		if (!valid_addr_reg(next_frame[0]) || next_pc == 0 ||
+		    next_pc % TILE_BUNDLE_ALIGNMENT_IN_BYTES != 0) {
+			/* The PC is suspect, so give up. */
+			return false;
+		}
+	}
+
+	/* Update state to become the caller's stack frame. */
+	state->pc = next_pc;
+	state->sp = state->fp;
+	state->fp = next_fp;
+
+	return true;
+}
diff --git a/kernel/arch/tile/kernel/compat.c b/kernel/arch/tile/kernel/compat.c
new file mode 100644
index 000000000..49120843f
--- /dev/null
+++ b/kernel/arch/tile/kernel/compat.c
@@ -0,0 +1,108 @@
+/*
+ * Copyright 2010 Tilera Corporation. All Rights Reserved.
+ *
+ *   This program is free software; you can redistribute it and/or
+ *   modify it under the terms of the GNU General Public License
+ *   as published by the Free Software Foundation, version 2.
+ *
+ *   This program is distributed in the hope that it will be useful, but
+ *   WITHOUT ANY WARRANTY; without even the implied warranty of
+ *   MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
+ *   NON INFRINGEMENT.  See the GNU General Public License for
+ *   more details.
+ */
+
+/* Adjust unistd.h to provide 32-bit numbers and functions. */
+#define __SYSCALL_COMPAT
+
+#include <linux/compat.h>
+#include <linux/syscalls.h>
+#include <linux/kdev_t.h>
+#include <linux/fs.h>
+#include <linux/fcntl.h>
+#include <linux/uaccess.h>
+#include <linux/signal.h>
+#include <asm/syscalls.h>
+
+/*
+ * Syscalls that take 64-bit numbers traditionally take them in 32-bit
+ * "high" and "low" value parts on 32-bit architectures.
+ * In principle, one could imagine passing some register arguments as
+ * fully 64-bit on TILE-Gx in 32-bit mode, but it seems easier to
+ * adapt the usual convention.
+ */
+
+COMPAT_SYSCALL_DEFINE4(truncate64, char __user *, filename, u32, dummy,
+                       u32, low, u32, high)
+{
+	return sys_truncate(filename, ((loff_t)high << 32) | low);
+}
+
+COMPAT_SYSCALL_DEFINE4(ftruncate64, unsigned int, fd, u32, dummy,
+                       u32, low, u32, high)
+{
+	return sys_ftruncate(fd, ((loff_t)high << 32) | low);
+}
+
+COMPAT_SYSCALL_DEFINE6(pread64, unsigned int, fd, char __user *, ubuf,
+                       size_t, count, u32, dummy, u32, low, u32, high)
+{
+	return sys_pread64(fd, ubuf, count, ((loff_t)high << 32) | low);
+}
+
+COMPAT_SYSCALL_DEFINE6(pwrite64, unsigned int, fd, char __user *, ubuf,
+                       size_t, count, u32, dummy, u32, low, u32, high)
+{
+	return sys_pwrite64(fd, ubuf, count, ((loff_t)high << 32) | low);
+}
+
+COMPAT_SYSCALL_DEFINE6(sync_file_range2, int, fd, unsigned int, flags,
+                       u32, offset_lo, u32, offset_hi,
+                       u32, nbytes_lo, u32, nbytes_hi)
+{
+	return sys_sync_file_range(fd, ((loff_t)offset_hi << 32) | offset_lo,
+				   ((loff_t)nbytes_hi << 32) | nbytes_lo,
+				   flags);
+}
+
+COMPAT_SYSCALL_DEFINE6(fallocate, int, fd, int, mode,
+                       u32, offset_lo, u32, offset_hi,
+                       u32, len_lo, u32, len_hi)
+{
+	return sys_fallocate(fd, mode, ((loff_t)offset_hi << 32) | offset_lo,
+			     ((loff_t)len_hi << 32) | len_lo);
+}
+
+/*
+ * Avoid bug in generic sys_llseek() that specifies offset_high and
+ * offset_low as "unsigned long", thus making it possible to pass
+ * a sign-extended high 32 bits in offset_low.
+ */
+COMPAT_SYSCALL_DEFINE5(llseek, unsigned int, fd, unsigned int, offset_high,
+		       unsigned int, offset_low, loff_t __user *, result,
+		       unsigned int, origin)
+{
+	return sys_llseek(fd, offset_high, offset_low, result, origin);
+}
+
+/* Provide the compat syscall number to call mapping. */
+#undef __SYSCALL
+#define __SYSCALL(nr, call) [nr] = (call),
+
+/* See comments in sys.c */
+#define compat_sys_fadvise64_64 sys32_fadvise64_64
+#define compat_sys_readahead sys32_readahead
+#define sys_llseek compat_sys_llseek
+
+/* Call the assembly trampolines where necessary. */
+#define compat_sys_rt_sigreturn _compat_sys_rt_sigreturn
+#define sys_clone _sys_clone
+
+/*
+ * Note that we can't include <linux/unistd.h> here since the header
+ * guard will defeat us; <asm/unistd.h> checks for __SYSCALL as well.
+ */
+void *compat_sys_call_table[__NR_syscalls] = {
+	[0 ... __NR_syscalls-1] = sys_ni_syscall,
+#include <asm/unistd.h>
+};
diff --git a/kernel/arch/tile/kernel/compat_signal.c b/kernel/arch/tile/kernel/compat_signal.c
new file mode 100644
index 000000000..e8c2c0414
--- /dev/null
+++ b/kernel/arch/tile/kernel/compat_signal.c
@@ -0,0 +1,248 @@
+/*
+ * Copyright 2010 Tilera Corporation. All Rights Reserved.
+ *
+ *   This program is free software; you can redistribute it and/or
+ *   modify it under the terms of the GNU General Public License
+ *   as published by the Free Software Foundation, version 2.
+ *
+ *   This program is distributed in the hope that it will be useful, but
+ *   WITHOUT ANY WARRANTY; without even the implied warranty of
+ *   MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
+ *   NON INFRINGEMENT.  See the GNU General Public License for
+ *   more details.
+ */
+
+#include <linux/sched.h>
+#include <linux/mm.h>
+#include <linux/smp.h>
+#include <linux/kernel.h>
+#include <linux/signal.h>
+#include <linux/errno.h>
+#include <linux/wait.h>
+#include <linux/unistd.h>
+#include <linux/stddef.h>
+#include <linux/personality.h>
+#include <linux/suspend.h>
+#include <linux/ptrace.h>
+#include <linux/elf.h>
+#include <linux/compat.h>
+#include <linux/syscalls.h>
+#include <linux/uaccess.h>
+#include <asm/processor.h>
+#include <asm/ucontext.h>
+#include <asm/sigframe.h>
+#include <asm/syscalls.h>
+#include <asm/vdso.h>
+#include <arch/interrupts.h>
+
+struct compat_ucontext {
+	compat_ulong_t	  uc_flags;
+	compat_uptr_t     uc_link;
+	struct compat_sigaltstack	  uc_stack;
+	struct sigcontext uc_mcontext;
+	sigset_t	  uc_sigmask;	/* mask last for extensibility */
+};
+
+struct compat_rt_sigframe {
+	unsigned char save_area[C_ABI_SAVE_AREA_SIZE]; /* caller save area */
+	struct compat_siginfo info;
+	struct compat_ucontext uc;
+};
+
+int copy_siginfo_to_user32(struct compat_siginfo __user *to, const siginfo_t *from)
+{
+	int err;
+
+	if (!access_ok(VERIFY_WRITE, to, sizeof(struct compat_siginfo)))
+		return -EFAULT;
+
+	/* If you change siginfo_t structure, please make sure that
+	   this code is fixed accordingly.
+	   It should never copy any pad contained in the structure
+	   to avoid security leaks, but must copy the generic
+	   3 ints plus the relevant union member.  */
+	err = __put_user(from->si_signo, &to->si_signo);
+	err |= __put_user(from->si_errno, &to->si_errno);
+	err |= __put_user((short)from->si_code, &to->si_code);
+
+	if (from->si_code < 0) {
+		err |= __put_user(from->si_pid, &to->si_pid);
+		err |= __put_user(from->si_uid, &to->si_uid);
+		err |= __put_user(from->si_int, &to->si_int);
+	} else {
+		/*
+		 * First 32bits of unions are always present:
+		 * si_pid === si_band === si_tid === si_addr(LS half)
+		 */
+		err |= __put_user(from->_sifields._pad[0],
+				  &to->_sifields._pad[0]);
+		switch (from->si_code >> 16) {
+		case __SI_FAULT >> 16:
+			break;
+		case __SI_CHLD >> 16:
+			err |= __put_user(from->si_utime, &to->si_utime);
+			err |= __put_user(from->si_stime, &to->si_stime);
+			err |= __put_user(from->si_status, &to->si_status);
+			/* FALL THROUGH */
+		default:
+		case __SI_KILL >> 16:
+			err |= __put_user(from->si_uid, &to->si_uid);
+			break;
+		case __SI_POLL >> 16:
+			err |= __put_user(from->si_fd, &to->si_fd);
+			break;
+		case __SI_TIMER >> 16:
+			err |= __put_user(from->si_overrun, &to->si_overrun);
+			err |= __put_user(from->si_int, &to->si_int);
+			break;
+			 /* This is not generated by the kernel as of now.  */
+		case __SI_RT >> 16:
+		case __SI_MESGQ >> 16:
+			err |= __put_user(from->si_uid, &to->si_uid);
+			err |= __put_user(from->si_int, &to->si_int);
+			break;
+		}
+	}
+	return err;
+}
+
+int copy_siginfo_from_user32(siginfo_t *to, struct compat_siginfo __user *from)
+{
+	int err;
+
+	if (!access_ok(VERIFY_READ, from, sizeof(struct compat_siginfo)))
+		return -EFAULT;
+
+	memset(to, 0, sizeof(*to));
+
+	err = __get_user(to->si_signo, &from->si_signo);
+	err |= __get_user(to->si_errno, &from->si_errno);
+	err |= __get_user(to->si_code, &from->si_code);
+
+	err |= __get_user(to->si_pid, &from->si_pid);
+	err |= __get_user(to->si_uid, &from->si_uid);
+	err |= __get_user(to->si_int, &from->si_int);
+
+	return err;
+}
+
+/* The assembly shim for this function arranges to ignore the return value. */
+long compat_sys_rt_sigreturn(void)
+{
+	struct pt_regs *regs = current_pt_regs();
+	struct compat_rt_sigframe __user *frame =
+		(struct compat_rt_sigframe __user *) compat_ptr(regs->sp);
+	sigset_t set;
+
+	if (!access_ok(VERIFY_READ, frame, sizeof(*frame)))
+		goto badframe;
+	if (__copy_from_user(&set, &frame->uc.uc_sigmask, sizeof(set)))
+		goto badframe;
+
+	set_current_blocked(&set);
+
+	if (restore_sigcontext(regs, &frame->uc.uc_mcontext))
+		goto badframe;
+
+	if (compat_restore_altstack(&frame->uc.uc_stack))
+		goto badframe;
+
+	return 0;
+
+badframe:
+	signal_fault("bad sigreturn frame", regs, frame, 0);
+	return 0;
+}
+
+/*
+ * Determine which stack to use..
+ */
+static inline void __user *compat_get_sigframe(struct k_sigaction *ka,
+					       struct pt_regs *regs,
+					       size_t frame_size)
+{
+	unsigned long sp;
+
+	/* Default to using normal stack */
+	sp = (unsigned long)compat_ptr(regs->sp);
+
+	/*
+	 * If we are on the alternate signal stack and would overflow
+	 * it, don't.  Return an always-bogus address instead so we
+	 * will die with SIGSEGV.
+	 */
+	if (on_sig_stack(sp) && !likely(on_sig_stack(sp - frame_size)))
+		return (void __user __force *)-1UL;
+
+	/* This is the X/Open sanctioned signal stack switching.  */
+	if (ka->sa.sa_flags & SA_ONSTACK) {
+		if (sas_ss_flags(sp) == 0)
+			sp = current->sas_ss_sp + current->sas_ss_size;
+	}
+
+	sp -= frame_size;
+	/*
+	 * Align the stack pointer according to the TILE ABI,
+	 * i.e. so that on function entry (sp & 15) == 0.
+	 */
+	sp &= -16UL;
+	return (void __user *) sp;
+}
+
+int compat_setup_rt_frame(struct ksignal *ksig, sigset_t *set,
+			  struct pt_regs *regs)
+{
+	unsigned long restorer;
+	struct compat_rt_sigframe __user *frame;
+	int err = 0, sig = ksig->sig;
+
+	frame = compat_get_sigframe(&ksig->ka, regs, sizeof(*frame));
+
+	if (!access_ok(VERIFY_WRITE, frame, sizeof(*frame)))
+		goto err;
+
+	/* Always write at least the signal number for the stack backtracer. */
+	if (ksig->ka.sa.sa_flags & SA_SIGINFO) {
+		/* At sigreturn time, restore the callee-save registers too. */
+		err |= copy_siginfo_to_user32(&frame->info, &ksig->info);
+		regs->flags |= PT_FLAGS_RESTORE_REGS;
+	} else {
+		err |= __put_user(ksig->info.si_signo, &frame->info.si_signo);
+	}
+
+	/* Create the ucontext.  */
+	err |= __clear_user(&frame->save_area, sizeof(frame->save_area));
+	err |= __put_user(0, &frame->uc.uc_flags);
+	err |= __put_user(0, &frame->uc.uc_link);
+	err |= __compat_save_altstack(&frame->uc.uc_stack, regs->sp);
+	err |= setup_sigcontext(&frame->uc.uc_mcontext, regs);
+	err |= __copy_to_user(&frame->uc.uc_sigmask, set, sizeof(*set));
+	if (err)
+		goto err;
+
+	restorer = VDSO_SYM(&__vdso_rt_sigreturn);
+	if (ksig->ka.sa.sa_flags & SA_RESTORER)
+		restorer = ptr_to_compat_reg(ksig->ka.sa.sa_restorer);
+
+	/*
+	 * Set up registers for signal handler.
+	 * Registers that we don't modify keep the value they had from
+	 * user-space at the time we took the signal.
+	 * We always pass siginfo and mcontext, regardless of SA_SIGINFO,
+	 * since some things rely on this (e.g. glibc's debug/segfault.c).
+	 */
+	regs->pc = ptr_to_compat_reg(ksig->ka.sa.sa_handler);
+	regs->ex1 = PL_ICS_EX1(USER_PL, 1); /* set crit sec in handler */
+	regs->sp = ptr_to_compat_reg(frame);
+	regs->lr = restorer;
+	regs->regs[0] = (unsigned long) sig;
+	regs->regs[1] = ptr_to_compat_reg(&frame->info);
+	regs->regs[2] = ptr_to_compat_reg(&frame->uc);
+	regs->flags |= PT_FLAGS_CALLER_SAVES;
+	return 0;
+
+err:
+	trace_unhandled_signal("bad sigreturn frame", regs,
+			      (unsigned long)frame, SIGSEGV);
+	return -EFAULT;
+}
diff --git a/kernel/arch/tile/kernel/early_printk.c b/kernel/arch/tile/kernel/early_printk.c
new file mode 100644
index 000000000..aefb2c086
--- /dev/null
+++ b/kernel/arch/tile/kernel/early_printk.c
@@ -0,0 +1,75 @@
+/*
+ * Copyright 2010 Tilera Corporation. All Rights Reserved.
+ *
+ *   This program is free software; you can redistribute it and/or
+ *   modify it under the terms of the GNU General Public License
+ *   as published by the Free Software Foundation, version 2.
+ *
+ *   This program is distributed in the hope that it will be useful, but
+ *   WITHOUT ANY WARRANTY; without even the implied warranty of
+ *   MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
+ *   NON INFRINGEMENT.  See the GNU General Public License for
+ *   more details.
+ */
+
+#include <linux/console.h>
+#include <linux/kernel.h>
+#include <linux/init.h>
+#include <linux/string.h>
+#include <linux/irqflags.h>
+#include <linux/printk.h>
+#include <asm/setup.h>
+#include <hv/hypervisor.h>
+
+static void early_hv_write(struct console *con, const char *s, unsigned n)
+{
+	tile_console_write(s, n);
+
+	/*
+	 * Convert NL to NLCR (close enough to CRNL) during early boot.
+	 * We assume newlines are at the ends of strings, which turns out
+	 * to be good enough for early boot console output.
+	 */
+	if (n && s[n-1] == '\n')
+		tile_console_write("\r", 1);
+}
+
+static struct console early_hv_console = {
+	.name =		"earlyhv",
+	.write =	early_hv_write,
+	.flags =	CON_PRINTBUFFER | CON_BOOT,
+	.index =	-1,
+};
+
+void early_panic(const char *fmt, ...)
+{
+	struct va_format vaf;
+	va_list args;
+
+	arch_local_irq_disable_all();
+
+	va_start(args, fmt);
+
+	vaf.fmt = fmt;
+	vaf.va = &args;
+
+	early_printk("Kernel panic - not syncing: %pV", &vaf);
+
+	va_end(args);
+
+	dump_stack();
+	hv_halt();
+}
+
+static int __init setup_early_printk(char *str)
+{
+	if (early_console)
+		return 1;
+
+	early_console = &early_hv_console;
+	register_console(early_console);
+
+	return 0;
+}
+
+early_param("earlyprintk", setup_early_printk);
diff --git a/kernel/arch/tile/kernel/entry.S b/kernel/arch/tile/kernel/entry.S
new file mode 100644
index 000000000..3d9175992
--- /dev/null
+++ b/kernel/arch/tile/kernel/entry.S
@@ -0,0 +1,71 @@
+/*
+ * Copyright 2010 Tilera Corporation. All Rights Reserved.
+ *
+ *   This program is free software; you can redistribute it and/or
+ *   modify it under the terms of the GNU General Public License
+ *   as published by the Free Software Foundation, version 2.
+ *
+ *   This program is distributed in the hope that it will be useful, but
+ *   WITHOUT ANY WARRANTY; without even the implied warranty of
+ *   MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
+ *   NON INFRINGEMENT.  See the GNU General Public License for
+ *   more details.
+ */
+
+#include <linux/linkage.h>
+#include <linux/unistd.h>
+#include <asm/irqflags.h>
+#include <asm/processor.h>
+#include <arch/abi.h>
+#include <arch/spr_def.h>
+
+#ifdef __tilegx__
+#define bnzt bnezt
+#endif
+
+STD_ENTRY(current_text_addr)
+	{ move r0, lr; jrp lr }
+	STD_ENDPROC(current_text_addr)
+
+STD_ENTRY(dump_stack)
+	{ move r2, lr; lnk r1 }
+	{ move r4, r52; addli r1, r1, dump_stack - . }
+	{ move r3, sp; j _dump_stack }
+	jrp lr   /* keep backtracer happy */
+	STD_ENDPROC(dump_stack)
+
+STD_ENTRY(KBacktraceIterator_init_current)
+	{ move r2, lr; lnk r1 }
+	{ move r4, r52; addli r1, r1, KBacktraceIterator_init_current - . }
+	{ move r3, sp; j _KBacktraceIterator_init_current }
+	jrp lr   /* keep backtracer happy */
+	STD_ENDPROC(KBacktraceIterator_init_current)
+
+/* Loop forever on a nap during SMP boot. */
+STD_ENTRY(smp_nap)
+	nap
+	nop       /* avoid provoking the icache prefetch with a jump */
+	j smp_nap /* we are not architecturally guaranteed not to exit nap */
+	jrp lr    /* clue in the backtracer */
+	STD_ENDPROC(smp_nap)
+
+/*
+ * Enable interrupts racelessly and then nap until interrupted.
+ * Architecturally, we are guaranteed that enabling interrupts via
+ * mtspr to INTERRUPT_CRITICAL_SECTION only interrupts at the next PC.
+ * This function's _cpu_idle_nap address is special; see intvec.S.
+ * When interrupted at _cpu_idle_nap, we bump the PC forward 8, and
+ * as a result return to the function that called _cpu_idle().
+ */
+STD_ENTRY(_cpu_idle)
+	movei r1, 1
+	IRQ_ENABLE_LOAD(r2, r3)
+	mtspr INTERRUPT_CRITICAL_SECTION, r1
+	IRQ_ENABLE_APPLY(r2, r3)       /* unmask, but still with ICS set */
+	mtspr INTERRUPT_CRITICAL_SECTION, zero
+	.global _cpu_idle_nap
+_cpu_idle_nap:
+	nap
+	nop       /* avoid provoking the icache prefetch with a jump */
+	jrp lr
+	STD_ENDPROC(_cpu_idle)
diff --git a/kernel/arch/tile/kernel/ftrace.c b/kernel/arch/tile/kernel/ftrace.c
new file mode 100644
index 000000000..0c0996175
--- /dev/null
+++ b/kernel/arch/tile/kernel/ftrace.c
@@ -0,0 +1,248 @@
+/*
+ * Copyright 2012 Tilera Corporation. All Rights Reserved.
+ *
+ *   This program is free software; you can redistribute it and/or
+ *   modify it under the terms of the GNU General Public License
+ *   as published by the Free Software Foundation, version 2.
+ *
+ *   This program is distributed in the hope that it will be useful, but
+ *   WITHOUT ANY WARRANTY; without even the implied warranty of
+ *   MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
+ *   NON INFRINGEMENT.  See the GNU General Public License for
+ *   more details.
+ *
+ * TILE-Gx specific ftrace support
+ */
+
+#include <linux/ftrace.h>
+#include <linux/uaccess.h>
+
+#include <asm/cacheflush.h>
+#include <asm/ftrace.h>
+#include <asm/sections.h>
+
+#include <arch/opcode.h>
+
+#ifdef CONFIG_DYNAMIC_FTRACE
+
+static inline tilegx_bundle_bits NOP(void)
+{
+	return create_UnaryOpcodeExtension_X0(FNOP_UNARY_OPCODE_X0) |
+		create_RRROpcodeExtension_X0(UNARY_RRR_0_OPCODE_X0) |
+		create_Opcode_X0(RRR_0_OPCODE_X0) |
+		create_UnaryOpcodeExtension_X1(NOP_UNARY_OPCODE_X1) |
+		create_RRROpcodeExtension_X1(UNARY_RRR_0_OPCODE_X1) |
+		create_Opcode_X1(RRR_0_OPCODE_X1);
+}
+
+static int machine_stopped __read_mostly;
+
+int ftrace_arch_code_modify_prepare(void)
+{
+	machine_stopped = 1;
+	return 0;
+}
+
+int ftrace_arch_code_modify_post_process(void)
+{
+	flush_icache_range(0, CHIP_L1I_CACHE_SIZE());
+	machine_stopped = 0;
+	return 0;
+}
+
+/*
+ * Put { move r10, lr; jal ftrace_caller } in a bundle, this lets dynamic
+ * tracer just add one cycle overhead to every kernel function when disabled.
+ */
+static unsigned long ftrace_gen_branch(unsigned long pc, unsigned long addr,
+				       bool link)
+{
+	tilegx_bundle_bits opcode_x0, opcode_x1;
+	long pcrel_by_instr = (addr - pc) >> TILEGX_LOG2_BUNDLE_SIZE_IN_BYTES;
+
+	if (link) {
+		/* opcode: jal addr */
+		opcode_x1 =
+			create_Opcode_X1(JUMP_OPCODE_X1) |
+			create_JumpOpcodeExtension_X1(JAL_JUMP_OPCODE_X1) |
+			create_JumpOff_X1(pcrel_by_instr);
+	} else {
+		/* opcode: j addr */
+		opcode_x1 =
+			create_Opcode_X1(JUMP_OPCODE_X1) |
+			create_JumpOpcodeExtension_X1(J_JUMP_OPCODE_X1) |
+			create_JumpOff_X1(pcrel_by_instr);
+	}
+
+	/*
+	 * Also put { move r10, lr; jal ftrace_stub } in a bundle, which
+	 * is used to replace the instruction in address ftrace_call.
+	 */
+	if (addr == FTRACE_ADDR || addr == (unsigned long)ftrace_stub) {
+		/* opcode: or r10, lr, zero */
+		opcode_x0 =
+			create_Dest_X0(10) |
+			create_SrcA_X0(TREG_LR) |
+			create_SrcB_X0(TREG_ZERO) |
+			create_RRROpcodeExtension_X0(OR_RRR_0_OPCODE_X0) |
+			create_Opcode_X0(RRR_0_OPCODE_X0);
+	} else {
+		/* opcode: fnop */
+		opcode_x0 =
+			create_UnaryOpcodeExtension_X0(FNOP_UNARY_OPCODE_X0) |
+			create_RRROpcodeExtension_X0(UNARY_RRR_0_OPCODE_X0) |
+			create_Opcode_X0(RRR_0_OPCODE_X0);
+	}
+
+	return opcode_x1 | opcode_x0;
+}
+
+static unsigned long ftrace_nop_replace(struct dyn_ftrace *rec)
+{
+	return NOP();
+}
+
+static unsigned long ftrace_call_replace(unsigned long pc, unsigned long addr)
+{
+	return ftrace_gen_branch(pc, addr, true);
+}
+
+static int ftrace_modify_code(unsigned long pc, unsigned long old,
+			      unsigned long new)
+{
+	unsigned long pc_wr;
+
+	/* Check if the address is in kernel text space and module space. */
+	if (!kernel_text_address(pc))
+		return -EINVAL;
+
+	/* Operate on writable kernel text mapping. */
+	pc_wr = pc - MEM_SV_START + PAGE_OFFSET;
+
+	if (probe_kernel_write((void *)pc_wr, &new, MCOUNT_INSN_SIZE))
+		return -EPERM;
+
+	smp_wmb();
+
+	if (!machine_stopped && num_online_cpus() > 1)
+		flush_icache_range(pc, pc + MCOUNT_INSN_SIZE);
+
+	return 0;
+}
+
+int ftrace_update_ftrace_func(ftrace_func_t func)
+{
+	unsigned long pc, old;
+	unsigned long new;
+	int ret;
+
+	pc = (unsigned long)&ftrace_call;
+	memcpy(&old, &ftrace_call, MCOUNT_INSN_SIZE);
+	new = ftrace_call_replace(pc, (unsigned long)func);
+
+	ret = ftrace_modify_code(pc, old, new);
+
+	return ret;
+}
+
+int ftrace_make_call(struct dyn_ftrace *rec, unsigned long addr)
+{
+	unsigned long new, old;
+	unsigned long ip = rec->ip;
+
+	old = ftrace_nop_replace(rec);
+	new = ftrace_call_replace(ip, addr);
+
+	return ftrace_modify_code(rec->ip, old, new);
+}
+
+int ftrace_make_nop(struct module *mod,
+		    struct dyn_ftrace *rec, unsigned long addr)
+{
+	unsigned long ip = rec->ip;
+	unsigned long old;
+	unsigned long new;
+	int ret;
+
+	old = ftrace_call_replace(ip, addr);
+	new = ftrace_nop_replace(rec);
+	ret = ftrace_modify_code(ip, old, new);
+
+	return ret;
+}
+
+int __init ftrace_dyn_arch_init(void)
+{
+	return 0;
+}
+#endif /* CONFIG_DYNAMIC_FTRACE */
+
+#ifdef CONFIG_FUNCTION_GRAPH_TRACER
+void prepare_ftrace_return(unsigned long *parent, unsigned long self_addr,
+			   unsigned long frame_pointer)
+{
+	unsigned long return_hooker = (unsigned long) &return_to_handler;
+	struct ftrace_graph_ent trace;
+	unsigned long old;
+	int err;
+
+	if (unlikely(atomic_read(&current->tracing_graph_pause)))
+		return;
+
+	old = *parent;
+	*parent = return_hooker;
+
+	err = ftrace_push_return_trace(old, self_addr, &trace.depth,
+				       frame_pointer);
+	if (err == -EBUSY) {
+		*parent = old;
+		return;
+	}
+
+	trace.func = self_addr;
+
+	/* Only trace if the calling function expects to */
+	if (!ftrace_graph_entry(&trace)) {
+		current->curr_ret_stack--;
+		*parent = old;
+	}
+}
+
+#ifdef CONFIG_DYNAMIC_FTRACE
+extern unsigned long ftrace_graph_call;
+
+static int __ftrace_modify_caller(unsigned long *callsite,
+				  void (*func) (void), bool enable)
+{
+	unsigned long caller_fn = (unsigned long) func;
+	unsigned long pc = (unsigned long) callsite;
+	unsigned long branch = ftrace_gen_branch(pc, caller_fn, false);
+	unsigned long nop = NOP();
+	unsigned long old = enable ? nop : branch;
+	unsigned long new = enable ? branch : nop;
+
+	return ftrace_modify_code(pc, old, new);
+}
+
+static int ftrace_modify_graph_caller(bool enable)
+{
+	int ret;
+
+	ret = __ftrace_modify_caller(&ftrace_graph_call,
+				     ftrace_graph_caller,
+				     enable);
+
+	return ret;
+}
+
+int ftrace_enable_ftrace_graph_caller(void)
+{
+	return ftrace_modify_graph_caller(true);
+}
+
+int ftrace_disable_ftrace_graph_caller(void)
+{
+	return ftrace_modify_graph_caller(false);
+}
+#endif /* CONFIG_DYNAMIC_FTRACE */
+#endif /* CONFIG_FUNCTION_GRAPH_TRACER */
diff --git a/kernel/arch/tile/kernel/hardwall.c b/kernel/arch/tile/kernel/hardwall.c
new file mode 100644
index 000000000..2fd1694ac
--- /dev/null
+++ b/kernel/arch/tile/kernel/hardwall.c
@@ -0,0 +1,1096 @@
+/*
+ * Copyright 2010 Tilera Corporation. All Rights Reserved.
+ *
+ *   This program is free software; you can redistribute it and/or
+ *   modify it under the terms of the GNU General Public License
+ *   as published by the Free Software Foundation, version 2.
+ *
+ *   This program is distributed in the hope that it will be useful, but
+ *   WITHOUT ANY WARRANTY; without even the implied warranty of
+ *   MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
+ *   NON INFRINGEMENT.  See the GNU General Public License for
+ *   more details.
+ */
+
+#include <linux/fs.h>
+#include <linux/proc_fs.h>
+#include <linux/seq_file.h>
+#include <linux/rwsem.h>
+#include <linux/kprobes.h>
+#include <linux/sched.h>
+#include <linux/hardirq.h>
+#include <linux/uaccess.h>
+#include <linux/smp.h>
+#include <linux/cdev.h>
+#include <linux/compat.h>
+#include <asm/hardwall.h>
+#include <asm/traps.h>
+#include <asm/siginfo.h>
+#include <asm/irq_regs.h>
+
+#include <arch/interrupts.h>
+#include <arch/spr_def.h>
+
+
+/*
+ * Implement a per-cpu "hardwall" resource class such as UDN or IPI.
+ * We use "hardwall" nomenclature throughout for historical reasons.
+ * The lock here controls access to the list data structure as well as
+ * to the items on the list.
+ */
+struct hardwall_type {
+	int index;
+	int is_xdn;
+	int is_idn;
+	int disabled;
+	const char *name;
+	struct list_head list;
+	spinlock_t lock;
+	struct proc_dir_entry *proc_dir;
+};
+
+enum hardwall_index {
+	HARDWALL_UDN = 0,
+#ifndef __tilepro__
+	HARDWALL_IDN = 1,
+	HARDWALL_IPI = 2,
+#endif
+	_HARDWALL_TYPES
+};
+
+static struct hardwall_type hardwall_types[] = {
+	{  /* user-space access to UDN */
+		0,
+		1,
+		0,
+		0,
+		"udn",
+		LIST_HEAD_INIT(hardwall_types[HARDWALL_UDN].list),
+		__SPIN_LOCK_UNLOCKED(hardwall_types[HARDWALL_UDN].lock),
+		NULL
+	},
+#ifndef __tilepro__
+	{  /* user-space access to IDN */
+		1,
+		1,
+		1,
+		1,  /* disabled pending hypervisor support */
+		"idn",
+		LIST_HEAD_INIT(hardwall_types[HARDWALL_IDN].list),
+		__SPIN_LOCK_UNLOCKED(hardwall_types[HARDWALL_IDN].lock),
+		NULL
+	},
+	{  /* access to user-space IPI */
+		2,
+		0,
+		0,
+		0,
+		"ipi",
+		LIST_HEAD_INIT(hardwall_types[HARDWALL_IPI].list),
+		__SPIN_LOCK_UNLOCKED(hardwall_types[HARDWALL_IPI].lock),
+		NULL
+	},
+#endif
+};
+
+/*
+ * This data structure tracks the cpu data, etc., associated
+ * one-to-one with a "struct file *" from opening a hardwall device file.
+ * Note that the file's private data points back to this structure.
+ */
+struct hardwall_info {
+	struct list_head list;             /* for hardwall_types.list */
+	struct list_head task_head;        /* head of tasks in this hardwall */
+	struct hardwall_type *type;        /* type of this resource */
+	struct cpumask cpumask;            /* cpus reserved */
+	int id;                            /* integer id for this hardwall */
+	int teardown_in_progress;          /* are we tearing this one down? */
+
+	/* Remaining fields only valid for user-network resources. */
+	int ulhc_x;                        /* upper left hand corner x coord */
+	int ulhc_y;                        /* upper left hand corner y coord */
+	int width;                         /* rectangle width */
+	int height;                        /* rectangle height */
+#if CHIP_HAS_REV1_XDN()
+	atomic_t xdn_pending_count;        /* cores in phase 1 of drain */
+#endif
+};
+
+
+/* /proc/tile/hardwall */
+static struct proc_dir_entry *hardwall_proc_dir;
+
+/* Functions to manage files in /proc/tile/hardwall. */
+static void hardwall_add_proc(struct hardwall_info *);
+static void hardwall_remove_proc(struct hardwall_info *);
+
+/* Allow disabling UDN access. */
+static int __init noudn(char *str)
+{
+	pr_info("User-space UDN access is disabled\n");
+	hardwall_types[HARDWALL_UDN].disabled = 1;
+	return 0;
+}
+early_param("noudn", noudn);
+
+#ifndef __tilepro__
+/* Allow disabling IDN access. */
+static int __init noidn(char *str)
+{
+	pr_info("User-space IDN access is disabled\n");
+	hardwall_types[HARDWALL_IDN].disabled = 1;
+	return 0;
+}
+early_param("noidn", noidn);
+
+/* Allow disabling IPI access. */
+static int __init noipi(char *str)
+{
+	pr_info("User-space IPI access is disabled\n");
+	hardwall_types[HARDWALL_IPI].disabled = 1;
+	return 0;
+}
+early_param("noipi", noipi);
+#endif
+
+
+/*
+ * Low-level primitives for UDN/IDN
+ */
+
+#ifdef __tilepro__
+#define mtspr_XDN(hwt, name, val) \
+	do { (void)(hwt); __insn_mtspr(SPR_UDN_##name, (val)); } while (0)
+#define mtspr_MPL_XDN(hwt, name, val) \
+	do { (void)(hwt); __insn_mtspr(SPR_MPL_UDN_##name, (val)); } while (0)
+#define mfspr_XDN(hwt, name) \
+	((void)(hwt), __insn_mfspr(SPR_UDN_##name))
+#else
+#define mtspr_XDN(hwt, name, val)					\
+	do {								\
+		if ((hwt)->is_idn)					\
+			__insn_mtspr(SPR_IDN_##name, (val));		\
+		else							\
+			__insn_mtspr(SPR_UDN_##name, (val));		\
+	} while (0)
+#define mtspr_MPL_XDN(hwt, name, val)					\
+	do {								\
+		if ((hwt)->is_idn)					\
+			__insn_mtspr(SPR_MPL_IDN_##name, (val));	\
+		else							\
+			__insn_mtspr(SPR_MPL_UDN_##name, (val));	\
+	} while (0)
+#define mfspr_XDN(hwt, name) \
+  ((hwt)->is_idn ? __insn_mfspr(SPR_IDN_##name) : __insn_mfspr(SPR_UDN_##name))
+#endif
+
+/* Set a CPU bit if the CPU is online. */
+#define cpu_online_set(cpu, dst) do { \
+	if (cpu_online(cpu))          \
+		cpumask_set_cpu(cpu, dst);    \
+} while (0)
+
+
+/* Does the given rectangle contain the given x,y coordinate? */
+static int contains(struct hardwall_info *r, int x, int y)
+{
+	return (x >= r->ulhc_x && x < r->ulhc_x + r->width) &&
+		(y >= r->ulhc_y && y < r->ulhc_y + r->height);
+}
+
+/* Compute the rectangle parameters and validate the cpumask. */
+static int check_rectangle(struct hardwall_info *r, struct cpumask *mask)
+{
+	int x, y, cpu, ulhc, lrhc;
+
+	/* The first cpu is the ULHC, the last the LRHC. */
+	ulhc = find_first_bit(cpumask_bits(mask), nr_cpumask_bits);
+	lrhc = find_last_bit(cpumask_bits(mask), nr_cpumask_bits);
+
+	/* Compute the rectangle attributes from the cpus. */
+	r->ulhc_x = cpu_x(ulhc);
+	r->ulhc_y = cpu_y(ulhc);
+	r->width = cpu_x(lrhc) - r->ulhc_x + 1;
+	r->height = cpu_y(lrhc) - r->ulhc_y + 1;
+
+	/* Width and height must be positive */
+	if (r->width <= 0 || r->height <= 0)
+		return -EINVAL;
+
+	/* Confirm that the cpumask is exactly the rectangle. */
+	for (y = 0, cpu = 0; y < smp_height; ++y)
+		for (x = 0; x < smp_width; ++x, ++cpu)
+			if (cpumask_test_cpu(cpu, mask) != contains(r, x, y))
+				return -EINVAL;
+
+	/*
+	 * Note that offline cpus can't be drained when this user network
+	 * rectangle eventually closes.  We used to detect this
+	 * situation and print a warning, but it annoyed users and
+	 * they ignored it anyway, so now we just return without a
+	 * warning.
+	 */
+	return 0;
+}
+
+/*
+ * Hardware management of hardwall setup, teardown, trapping,
+ * and enabling/disabling PL0 access to the networks.
+ */
+
+/* Bit field values to mask together for writes to SPR_XDN_DIRECTION_PROTECT */
+enum direction_protect {
+	N_PROTECT = (1 << 0),
+	E_PROTECT = (1 << 1),
+	S_PROTECT = (1 << 2),
+	W_PROTECT = (1 << 3),
+	C_PROTECT = (1 << 4),
+};
+
+static inline int xdn_which_interrupt(struct hardwall_type *hwt)
+{
+#ifndef __tilepro__
+	if (hwt->is_idn)
+		return INT_IDN_FIREWALL;
+#endif
+	return INT_UDN_FIREWALL;
+}
+
+static void enable_firewall_interrupts(struct hardwall_type *hwt)
+{
+	arch_local_irq_unmask_now(xdn_which_interrupt(hwt));
+}
+
+static void disable_firewall_interrupts(struct hardwall_type *hwt)
+{
+	arch_local_irq_mask_now(xdn_which_interrupt(hwt));
+}
+
+/* Set up hardwall on this cpu based on the passed hardwall_info. */
+static void hardwall_setup_func(void *info)
+{
+	struct hardwall_info *r = info;
+	struct hardwall_type *hwt = r->type;
+
+	int cpu = smp_processor_id();  /* on_each_cpu disables preemption */
+	int x = cpu_x(cpu);
+	int y = cpu_y(cpu);
+	int bits = 0;
+	if (x == r->ulhc_x)
+		bits |= W_PROTECT;
+	if (x == r->ulhc_x + r->width - 1)
+		bits |= E_PROTECT;
+	if (y == r->ulhc_y)
+		bits |= N_PROTECT;
+	if (y == r->ulhc_y + r->height - 1)
+		bits |= S_PROTECT;
+	BUG_ON(bits == 0);
+	mtspr_XDN(hwt, DIRECTION_PROTECT, bits);
+	enable_firewall_interrupts(hwt);
+}
+
+/* Set up all cpus on edge of rectangle to enable/disable hardwall SPRs. */
+static void hardwall_protect_rectangle(struct hardwall_info *r)
+{
+	int x, y, cpu, delta;
+	struct cpumask rect_cpus;
+
+	cpumask_clear(&rect_cpus);
+
+	/* First include the top and bottom edges */
+	cpu = r->ulhc_y * smp_width + r->ulhc_x;
+	delta = (r->height - 1) * smp_width;
+	for (x = 0; x < r->width; ++x, ++cpu) {
+		cpu_online_set(cpu, &rect_cpus);
+		cpu_online_set(cpu + delta, &rect_cpus);
+	}
+
+	/* Then the left and right edges */
+	cpu -= r->width;
+	delta = r->width - 1;
+	for (y = 0; y < r->height; ++y, cpu += smp_width) {
+		cpu_online_set(cpu, &rect_cpus);
+		cpu_online_set(cpu + delta, &rect_cpus);
+	}
+
+	/* Then tell all the cpus to set up their protection SPR */
+	on_each_cpu_mask(&rect_cpus, hardwall_setup_func, r, 1);
+}
+
+/* Entered from INT_xDN_FIREWALL interrupt vector with irqs disabled. */
+void __kprobes do_hardwall_trap(struct pt_regs* regs, int fault_num)
+{
+	struct hardwall_info *rect;
+	struct hardwall_type *hwt;
+	struct task_struct *p;
+	struct siginfo info;
+	int cpu = smp_processor_id();
+	int found_processes;
+	struct pt_regs *old_regs = set_irq_regs(regs);
+
+	irq_enter();
+
+	/* Figure out which network trapped. */
+	switch (fault_num) {
+#ifndef __tilepro__
+	case INT_IDN_FIREWALL:
+		hwt = &hardwall_types[HARDWALL_IDN];
+		break;
+#endif
+	case INT_UDN_FIREWALL:
+		hwt = &hardwall_types[HARDWALL_UDN];
+		break;
+	default:
+		BUG();
+	}
+	BUG_ON(hwt->disabled);
+
+	/* This tile trapped a network access; find the rectangle. */
+	spin_lock(&hwt->lock);
+	list_for_each_entry(rect, &hwt->list, list) {
+		if (cpumask_test_cpu(cpu, &rect->cpumask))
+			break;
+	}
+
+	/*
+	 * It shouldn't be possible not to find this cpu on the
+	 * rectangle list, since only cpus in rectangles get hardwalled.
+	 * The hardwall is only removed after the user network is drained.
+	 */
+	BUG_ON(&rect->list == &hwt->list);
+
+	/*
+	 * If we already started teardown on this hardwall, don't worry;
+	 * the abort signal has been sent and we are just waiting for things
+	 * to quiesce.
+	 */
+	if (rect->teardown_in_progress) {
+		pr_notice("cpu %d: detected %s hardwall violation %#lx while teardown already in progress\n",
+			  cpu, hwt->name,
+			  (long)mfspr_XDN(hwt, DIRECTION_PROTECT));
+		goto done;
+	}
+
+	/*
+	 * Kill off any process that is activated in this rectangle.
+	 * We bypass security to deliver the signal, since it must be
+	 * one of the activated processes that generated the user network
+	 * message that caused this trap, and all the activated
+	 * processes shared a single open file so are pretty tightly
+	 * bound together from a security point of view to begin with.
+	 */
+	rect->teardown_in_progress = 1;
+	wmb(); /* Ensure visibility of rectangle before notifying processes. */
+	pr_notice("cpu %d: detected %s hardwall violation %#lx...\n",
+		  cpu, hwt->name, (long)mfspr_XDN(hwt, DIRECTION_PROTECT));
+	info.si_signo = SIGILL;
+	info.si_errno = 0;
+	info.si_code = ILL_HARDWALL;
+	found_processes = 0;
+	list_for_each_entry(p, &rect->task_head,
+			    thread.hardwall[hwt->index].list) {
+		BUG_ON(p->thread.hardwall[hwt->index].info != rect);
+		if (!(p->flags & PF_EXITING)) {
+			found_processes = 1;
+			pr_notice("hardwall: killing %d\n", p->pid);
+			do_send_sig_info(info.si_signo, &info, p, false);
+		}
+	}
+	if (!found_processes)
+		pr_notice("hardwall: no associated processes!\n");
+
+ done:
+	spin_unlock(&hwt->lock);
+
+	/*
+	 * We have to disable firewall interrupts now, or else when we
+	 * return from this handler, we will simply re-interrupt back to
+	 * it.  However, we can't clear the protection bits, since we
+	 * haven't yet drained the network, and that would allow packets
+	 * to cross out of the hardwall region.
+	 */
+	disable_firewall_interrupts(hwt);
+
+	irq_exit();
+	set_irq_regs(old_regs);
+}
+
+/* Allow access from user space to the user network. */
+void grant_hardwall_mpls(struct hardwall_type *hwt)
+{
+#ifndef __tilepro__
+	if (!hwt->is_xdn) {
+		__insn_mtspr(SPR_MPL_IPI_0_SET_0, 1);
+		return;
+	}
+#endif
+	mtspr_MPL_XDN(hwt, ACCESS_SET_0, 1);
+	mtspr_MPL_XDN(hwt, AVAIL_SET_0, 1);
+	mtspr_MPL_XDN(hwt, COMPLETE_SET_0, 1);
+	mtspr_MPL_XDN(hwt, TIMER_SET_0, 1);
+#if !CHIP_HAS_REV1_XDN()
+	mtspr_MPL_XDN(hwt, REFILL_SET_0, 1);
+	mtspr_MPL_XDN(hwt, CA_SET_0, 1);
+#endif
+}
+
+/* Deny access from user space to the user network. */
+void restrict_hardwall_mpls(struct hardwall_type *hwt)
+{
+#ifndef __tilepro__
+	if (!hwt->is_xdn) {
+		__insn_mtspr(SPR_MPL_IPI_0_SET_1, 1);
+		return;
+	}
+#endif
+	mtspr_MPL_XDN(hwt, ACCESS_SET_1, 1);
+	mtspr_MPL_XDN(hwt, AVAIL_SET_1, 1);
+	mtspr_MPL_XDN(hwt, COMPLETE_SET_1, 1);
+	mtspr_MPL_XDN(hwt, TIMER_SET_1, 1);
+#if !CHIP_HAS_REV1_XDN()
+	mtspr_MPL_XDN(hwt, REFILL_SET_1, 1);
+	mtspr_MPL_XDN(hwt, CA_SET_1, 1);
+#endif
+}
+
+/* Restrict or deny as necessary for the task we're switching to. */
+void hardwall_switch_tasks(struct task_struct *prev,
+			   struct task_struct *next)
+{
+	int i;
+	for (i = 0; i < HARDWALL_TYPES; ++i) {
+		if (prev->thread.hardwall[i].info != NULL) {
+			if (next->thread.hardwall[i].info == NULL)
+				restrict_hardwall_mpls(&hardwall_types[i]);
+		} else if (next->thread.hardwall[i].info != NULL) {
+			grant_hardwall_mpls(&hardwall_types[i]);
+		}
+	}
+}
+
+/* Does this task have the right to IPI the given cpu? */
+int hardwall_ipi_valid(int cpu)
+{
+#ifdef __tilegx__
+	struct hardwall_info *info =
+		current->thread.hardwall[HARDWALL_IPI].info;
+	return info && cpumask_test_cpu(cpu, &info->cpumask);
+#else
+	return 0;
+#endif
+}
+
+/*
+ * Code to create, activate, deactivate, and destroy hardwall resources.
+ */
+
+/* Create a hardwall for the given resource */
+static struct hardwall_info *hardwall_create(struct hardwall_type *hwt,
+					     size_t size,
+					     const unsigned char __user *bits)
+{
+	struct hardwall_info *iter, *info;
+	struct cpumask mask;
+	unsigned long flags;
+	int rc;
+
+	/* Reject crazy sizes out of hand, a la sys_mbind(). */
+	if (size > PAGE_SIZE)
+		return ERR_PTR(-EINVAL);
+
+	/* Copy whatever fits into a cpumask. */
+	if (copy_from_user(&mask, bits, min(sizeof(struct cpumask), size)))
+		return ERR_PTR(-EFAULT);
+
+	/*
+	 * If the size was short, clear the rest of the mask;
+	 * otherwise validate that the rest of the user mask was zero
+	 * (we don't try hard to be efficient when validating huge masks).
+	 */
+	if (size < sizeof(struct cpumask)) {
+		memset((char *)&mask + size, 0, sizeof(struct cpumask) - size);
+	} else if (size > sizeof(struct cpumask)) {
+		size_t i;
+		for (i = sizeof(struct cpumask); i < size; ++i) {
+			char c;
+			if (get_user(c, &bits[i]))
+				return ERR_PTR(-EFAULT);
+			if (c)
+				return ERR_PTR(-EINVAL);
+		}
+	}
+
+	/* Allocate a new hardwall_info optimistically. */
+	info = kmalloc(sizeof(struct hardwall_info),
+			GFP_KERNEL | __GFP_ZERO);
+	if (info == NULL)
+		return ERR_PTR(-ENOMEM);
+	INIT_LIST_HEAD(&info->task_head);
+	info->type = hwt;
+
+	/* Compute the rectangle size and validate that it's plausible. */
+	cpumask_copy(&info->cpumask, &mask);
+	info->id = find_first_bit(cpumask_bits(&mask), nr_cpumask_bits);
+	if (hwt->is_xdn) {
+		rc = check_rectangle(info, &mask);
+		if (rc != 0) {
+			kfree(info);
+			return ERR_PTR(rc);
+		}
+	}
+
+	/*
+	 * Eliminate cpus that are not part of this Linux client.
+	 * Note that this allows for configurations that we might not want to
+	 * support, such as one client on every even cpu, another client on
+	 * every odd cpu.
+	 */
+	cpumask_and(&info->cpumask, &info->cpumask, cpu_online_mask);
+
+	/* Confirm it doesn't overlap and add it to the list. */
+	spin_lock_irqsave(&hwt->lock, flags);
+	list_for_each_entry(iter, &hwt->list, list) {
+		if (cpumask_intersects(&iter->cpumask, &info->cpumask)) {
+			spin_unlock_irqrestore(&hwt->lock, flags);
+			kfree(info);
+			return ERR_PTR(-EBUSY);
+		}
+	}
+	list_add_tail(&info->list, &hwt->list);
+	spin_unlock_irqrestore(&hwt->lock, flags);
+
+	/* Set up appropriate hardwalling on all affected cpus. */
+	if (hwt->is_xdn)
+		hardwall_protect_rectangle(info);
+
+	/* Create a /proc/tile/hardwall entry. */
+	hardwall_add_proc(info);
+
+	return info;
+}
+
+/* Activate a given hardwall on this cpu for this process. */
+static int hardwall_activate(struct hardwall_info *info)
+{
+	int cpu;
+	unsigned long flags;
+	struct task_struct *p = current;
+	struct thread_struct *ts = &p->thread;
+	struct hardwall_type *hwt;
+
+	/* Require a hardwall. */
+	if (info == NULL)
+		return -ENODATA;
+
+	/* Not allowed to activate a hardwall that is being torn down. */
+	if (info->teardown_in_progress)
+		return -EINVAL;
+
+	/*
+	 * Get our affinity; if we're not bound to this tile uniquely,
+	 * we can't access the network registers.
+	 */
+	if (cpumask_weight(&p->cpus_allowed) != 1)
+		return -EPERM;
+
+	/* Make sure we are bound to a cpu assigned to this resource. */
+	cpu = smp_processor_id();
+	BUG_ON(cpumask_first(&p->cpus_allowed) != cpu);
+	if (!cpumask_test_cpu(cpu, &info->cpumask))
+		return -EINVAL;
+
+	/* If we are already bound to this hardwall, it's a no-op. */
+	hwt = info->type;
+	if (ts->hardwall[hwt->index].info) {
+		BUG_ON(ts->hardwall[hwt->index].info != info);
+		return 0;
+	}
+
+	/* Success!  This process gets to use the resource on this cpu. */
+	ts->hardwall[hwt->index].info = info;
+	spin_lock_irqsave(&hwt->lock, flags);
+	list_add(&ts->hardwall[hwt->index].list, &info->task_head);
+	spin_unlock_irqrestore(&hwt->lock, flags);
+	grant_hardwall_mpls(hwt);
+	printk(KERN_DEBUG "Pid %d (%s) activated for %s hardwall: cpu %d\n",
+	       p->pid, p->comm, hwt->name, cpu);
+	return 0;
+}
+
+/*
+ * Deactivate a task's hardwall.  Must hold lock for hardwall_type.
+ * This method may be called from exit_thread(), so we don't want to
+ * rely on too many fields of struct task_struct still being valid.
+ * We assume the cpus_allowed, pid, and comm fields are still valid.
+ */
+static void _hardwall_deactivate(struct hardwall_type *hwt,
+				 struct task_struct *task)
+{
+	struct thread_struct *ts = &task->thread;
+
+	if (cpumask_weight(&task->cpus_allowed) != 1) {
+		pr_err("pid %d (%s) releasing %s hardwall with an affinity mask containing %d cpus!\n",
+		       task->pid, task->comm, hwt->name,
+		       cpumask_weight(&task->cpus_allowed));
+		BUG();
+	}
+
+	BUG_ON(ts->hardwall[hwt->index].info == NULL);
+	ts->hardwall[hwt->index].info = NULL;
+	list_del(&ts->hardwall[hwt->index].list);
+	if (task == current)
+		restrict_hardwall_mpls(hwt);
+}
+
+/* Deactivate a task's hardwall. */
+static int hardwall_deactivate(struct hardwall_type *hwt,
+			       struct task_struct *task)
+{
+	unsigned long flags;
+	int activated;
+
+	spin_lock_irqsave(&hwt->lock, flags);
+	activated = (task->thread.hardwall[hwt->index].info != NULL);
+	if (activated)
+		_hardwall_deactivate(hwt, task);
+	spin_unlock_irqrestore(&hwt->lock, flags);
+
+	if (!activated)
+		return -EINVAL;
+
+	printk(KERN_DEBUG "Pid %d (%s) deactivated for %s hardwall: cpu %d\n",
+	       task->pid, task->comm, hwt->name, raw_smp_processor_id());
+	return 0;
+}
+
+void hardwall_deactivate_all(struct task_struct *task)
+{
+	int i;
+	for (i = 0; i < HARDWALL_TYPES; ++i)
+		if (task->thread.hardwall[i].info)
+			hardwall_deactivate(&hardwall_types[i], task);
+}
+
+/* Stop the switch before draining the network. */
+static void stop_xdn_switch(void *arg)
+{
+#if !CHIP_HAS_REV1_XDN()
+	/* Freeze the switch and the demux. */
+	__insn_mtspr(SPR_UDN_SP_FREEZE,
+		     SPR_UDN_SP_FREEZE__SP_FRZ_MASK |
+		     SPR_UDN_SP_FREEZE__DEMUX_FRZ_MASK |
+		     SPR_UDN_SP_FREEZE__NON_DEST_EXT_MASK);
+#else
+	/*
+	 * Drop all packets bound for the core or off the edge.
+	 * We rely on the normal hardwall protection setup code
+	 * to have set the low four bits to trigger firewall interrupts,
+	 * and shift those bits up to trigger "drop on send" semantics,
+	 * plus adding "drop on send to core" for all switches.
+	 * In practice it seems the switches latch the DIRECTION_PROTECT
+	 * SPR so they won't start dropping if they're already
+	 * delivering the last message to the core, but it doesn't
+	 * hurt to enable it here.
+	 */
+	struct hardwall_type *hwt = arg;
+	unsigned long protect = mfspr_XDN(hwt, DIRECTION_PROTECT);
+	mtspr_XDN(hwt, DIRECTION_PROTECT, (protect | C_PROTECT) << 5);
+#endif
+}
+
+static void empty_xdn_demuxes(struct hardwall_type *hwt)
+{
+#ifndef __tilepro__
+	if (hwt->is_idn) {
+		while (__insn_mfspr(SPR_IDN_DATA_AVAIL) & (1 << 0))
+			(void) __tile_idn0_receive();
+		while (__insn_mfspr(SPR_IDN_DATA_AVAIL) & (1 << 1))
+			(void) __tile_idn1_receive();
+		return;
+	}
+#endif
+	while (__insn_mfspr(SPR_UDN_DATA_AVAIL) & (1 << 0))
+		(void) __tile_udn0_receive();
+	while (__insn_mfspr(SPR_UDN_DATA_AVAIL) & (1 << 1))
+		(void) __tile_udn1_receive();
+	while (__insn_mfspr(SPR_UDN_DATA_AVAIL) & (1 << 2))
+		(void) __tile_udn2_receive();
+	while (__insn_mfspr(SPR_UDN_DATA_AVAIL) & (1 << 3))
+		(void) __tile_udn3_receive();
+}
+
+/* Drain all the state from a stopped switch. */
+static void drain_xdn_switch(void *arg)
+{
+	struct hardwall_info *info = arg;
+	struct hardwall_type *hwt = info->type;
+
+#if CHIP_HAS_REV1_XDN()
+	/*
+	 * The switches have been configured to drop any messages
+	 * destined for cores (or off the edge of the rectangle).
+	 * But the current message may continue to be delivered,
+	 * so we wait until all the cores have finished any pending
+	 * messages before we stop draining.
+	 */
+	int pending = mfspr_XDN(hwt, PENDING);
+	while (pending--) {
+		empty_xdn_demuxes(hwt);
+		if (hwt->is_idn)
+			__tile_idn_send(0);
+		else
+			__tile_udn_send(0);
+	}
+	atomic_dec(&info->xdn_pending_count);
+	while (atomic_read(&info->xdn_pending_count))
+		empty_xdn_demuxes(hwt);
+#else
+	int i;
+	int from_tile_words, ca_count;
+
+	/* Empty out the 5 switch point fifos. */
+	for (i = 0; i < 5; i++) {
+		int words, j;
+		__insn_mtspr(SPR_UDN_SP_FIFO_SEL, i);
+		words = __insn_mfspr(SPR_UDN_SP_STATE) & 0xF;
+		for (j = 0; j < words; j++)
+			(void) __insn_mfspr(SPR_UDN_SP_FIFO_DATA);
+		BUG_ON((__insn_mfspr(SPR_UDN_SP_STATE) & 0xF) != 0);
+	}
+
+	/* Dump out the 3 word fifo at top. */
+	from_tile_words = (__insn_mfspr(SPR_UDN_DEMUX_STATUS) >> 10) & 0x3;
+	for (i = 0; i < from_tile_words; i++)
+		(void) __insn_mfspr(SPR_UDN_DEMUX_WRITE_FIFO);
+
+	/* Empty out demuxes. */
+	empty_xdn_demuxes(hwt);
+
+	/* Empty out catch all. */
+	ca_count = __insn_mfspr(SPR_UDN_DEMUX_CA_COUNT);
+	for (i = 0; i < ca_count; i++)
+		(void) __insn_mfspr(SPR_UDN_CA_DATA);
+	BUG_ON(__insn_mfspr(SPR_UDN_DEMUX_CA_COUNT) != 0);
+
+	/* Clear demux logic. */
+	__insn_mtspr(SPR_UDN_DEMUX_CTL, 1);
+
+	/*
+	 * Write switch state; experimentation indicates that 0xc3000
+	 * is an idle switch point.
+	 */
+	for (i = 0; i < 5; i++) {
+		__insn_mtspr(SPR_UDN_SP_FIFO_SEL, i);
+		__insn_mtspr(SPR_UDN_SP_STATE, 0xc3000);
+	}
+#endif
+}
+
+/* Reset random XDN state registers at boot up and during hardwall teardown. */
+static void reset_xdn_network_state(struct hardwall_type *hwt)
+{
+	if (hwt->disabled)
+		return;
+
+	/* Clear out other random registers so we have a clean slate. */
+	mtspr_XDN(hwt, DIRECTION_PROTECT, 0);
+	mtspr_XDN(hwt, AVAIL_EN, 0);
+	mtspr_XDN(hwt, DEADLOCK_TIMEOUT, 0);
+
+#if !CHIP_HAS_REV1_XDN()
+	/* Reset UDN coordinates to their standard value */
+	{
+		unsigned int cpu = smp_processor_id();
+		unsigned int x = cpu_x(cpu);
+		unsigned int y = cpu_y(cpu);
+		__insn_mtspr(SPR_UDN_TILE_COORD, (x << 18) | (y << 7));
+	}
+
+	/* Set demux tags to predefined values and enable them. */
+	__insn_mtspr(SPR_UDN_TAG_VALID, 0xf);
+	__insn_mtspr(SPR_UDN_TAG_0, (1 << 0));
+	__insn_mtspr(SPR_UDN_TAG_1, (1 << 1));
+	__insn_mtspr(SPR_UDN_TAG_2, (1 << 2));
+	__insn_mtspr(SPR_UDN_TAG_3, (1 << 3));
+
+	/* Set other rev0 random registers to a clean state. */
+	__insn_mtspr(SPR_UDN_REFILL_EN, 0);
+	__insn_mtspr(SPR_UDN_DEMUX_QUEUE_SEL, 0);
+	__insn_mtspr(SPR_UDN_SP_FIFO_SEL, 0);
+
+	/* Start the switch and demux. */
+	__insn_mtspr(SPR_UDN_SP_FREEZE, 0);
+#endif
+}
+
+void reset_network_state(void)
+{
+	reset_xdn_network_state(&hardwall_types[HARDWALL_UDN]);
+#ifndef __tilepro__
+	reset_xdn_network_state(&hardwall_types[HARDWALL_IDN]);
+#endif
+}
+
+/* Restart an XDN switch after draining. */
+static void restart_xdn_switch(void *arg)
+{
+	struct hardwall_type *hwt = arg;
+
+#if CHIP_HAS_REV1_XDN()
+	/* One last drain step to avoid races with injection and draining. */
+	empty_xdn_demuxes(hwt);
+#endif
+
+	reset_xdn_network_state(hwt);
+
+	/* Disable firewall interrupts. */
+	disable_firewall_interrupts(hwt);
+}
+
+/* Last reference to a hardwall is gone, so clear the network. */
+static void hardwall_destroy(struct hardwall_info *info)
+{
+	struct task_struct *task;
+	struct hardwall_type *hwt;
+	unsigned long flags;
+
+	/* Make sure this file actually represents a hardwall. */
+	if (info == NULL)
+		return;
+
+	/*
+	 * Deactivate any remaining tasks.  It's possible to race with
+	 * some other thread that is exiting and hasn't yet called
+	 * deactivate (when freeing its thread_info), so we carefully
+	 * deactivate any remaining tasks before freeing the
+	 * hardwall_info object itself.
+	 */
+	hwt = info->type;
+	info->teardown_in_progress = 1;
+	spin_lock_irqsave(&hwt->lock, flags);
+	list_for_each_entry(task, &info->task_head,
+			    thread.hardwall[hwt->index].list)
+		_hardwall_deactivate(hwt, task);
+	spin_unlock_irqrestore(&hwt->lock, flags);
+
+	if (hwt->is_xdn) {
+		/* Configure the switches for draining the user network. */
+		printk(KERN_DEBUG
+		       "Clearing %s hardwall rectangle %dx%d %d,%d\n",
+		       hwt->name, info->width, info->height,
+		       info->ulhc_x, info->ulhc_y);
+		on_each_cpu_mask(&info->cpumask, stop_xdn_switch, hwt, 1);
+
+		/* Drain the network. */
+#if CHIP_HAS_REV1_XDN()
+		atomic_set(&info->xdn_pending_count,
+			   cpumask_weight(&info->cpumask));
+		on_each_cpu_mask(&info->cpumask, drain_xdn_switch, info, 0);
+#else
+		on_each_cpu_mask(&info->cpumask, drain_xdn_switch, info, 1);
+#endif
+
+		/* Restart switch and disable firewall. */
+		on_each_cpu_mask(&info->cpumask, restart_xdn_switch, hwt, 1);
+	}
+
+	/* Remove the /proc/tile/hardwall entry. */
+	hardwall_remove_proc(info);
+
+	/* Now free the hardwall from the list. */
+	spin_lock_irqsave(&hwt->lock, flags);
+	BUG_ON(!list_empty(&info->task_head));
+	list_del(&info->list);
+	spin_unlock_irqrestore(&hwt->lock, flags);
+	kfree(info);
+}
+
+
+static int hardwall_proc_show(struct seq_file *sf, void *v)
+{
+	struct hardwall_info *info = sf->private;
+
+	seq_printf(sf, "%*pbl\n", cpumask_pr_args(&info->cpumask));
+	return 0;
+}
+
+static int hardwall_proc_open(struct inode *inode,
+			      struct file *file)
+{
+	return single_open(file, hardwall_proc_show, PDE_DATA(inode));
+}
+
+static const struct file_operations hardwall_proc_fops = {
+	.open		= hardwall_proc_open,
+	.read		= seq_read,
+	.llseek		= seq_lseek,
+	.release	= single_release,
+};
+
+static void hardwall_add_proc(struct hardwall_info *info)
+{
+	char buf[64];
+	snprintf(buf, sizeof(buf), "%d", info->id);
+	proc_create_data(buf, 0444, info->type->proc_dir,
+			 &hardwall_proc_fops, info);
+}
+
+static void hardwall_remove_proc(struct hardwall_info *info)
+{
+	char buf[64];
+	snprintf(buf, sizeof(buf), "%d", info->id);
+	remove_proc_entry(buf, info->type->proc_dir);
+}
+
+int proc_pid_hardwall(struct seq_file *m, struct pid_namespace *ns,
+		      struct pid *pid, struct task_struct *task)
+{
+	int i;
+	int n = 0;
+	for (i = 0; i < HARDWALL_TYPES; ++i) {
+		struct hardwall_info *info = task->thread.hardwall[i].info;
+		if (info)
+			seq_printf(m, "%s: %d\n", info->type->name, info->id);
+	}
+	return n;
+}
+
+void proc_tile_hardwall_init(struct proc_dir_entry *root)
+{
+	int i;
+	for (i = 0; i < HARDWALL_TYPES; ++i) {
+		struct hardwall_type *hwt = &hardwall_types[i];
+		if (hwt->disabled)
+			continue;
+		if (hardwall_proc_dir == NULL)
+			hardwall_proc_dir = proc_mkdir("hardwall", root);
+		hwt->proc_dir = proc_mkdir(hwt->name, hardwall_proc_dir);
+	}
+}
+
+
+/*
+ * Character device support via ioctl/close.
+ */
+
+static long hardwall_ioctl(struct file *file, unsigned int a, unsigned long b)
+{
+	struct hardwall_info *info = file->private_data;
+	int minor = iminor(file->f_mapping->host);
+	struct hardwall_type* hwt;
+
+	if (_IOC_TYPE(a) != HARDWALL_IOCTL_BASE)
+		return -EINVAL;
+
+	BUILD_BUG_ON(HARDWALL_TYPES != _HARDWALL_TYPES);
+	BUILD_BUG_ON(HARDWALL_TYPES !=
+		     sizeof(hardwall_types)/sizeof(hardwall_types[0]));
+
+	if (minor < 0 || minor >= HARDWALL_TYPES)
+		return -EINVAL;
+	hwt = &hardwall_types[minor];
+	WARN_ON(info && hwt != info->type);
+
+	switch (_IOC_NR(a)) {
+	case _HARDWALL_CREATE:
+		if (hwt->disabled)
+			return -ENOSYS;
+		if (info != NULL)
+			return -EALREADY;
+		info = hardwall_create(hwt, _IOC_SIZE(a),
+				       (const unsigned char __user *)b);
+		if (IS_ERR(info))
+			return PTR_ERR(info);
+		file->private_data = info;
+		return 0;
+
+	case _HARDWALL_ACTIVATE:
+		return hardwall_activate(info);
+
+	case _HARDWALL_DEACTIVATE:
+		if (current->thread.hardwall[hwt->index].info != info)
+			return -EINVAL;
+		return hardwall_deactivate(hwt, current);
+
+	case _HARDWALL_GET_ID:
+		return info ? info->id : -EINVAL;
+
+	default:
+		return -EINVAL;
+	}
+}
+
+#ifdef CONFIG_COMPAT
+static long hardwall_compat_ioctl(struct file *file,
+				  unsigned int a, unsigned long b)
+{
+	/* Sign-extend the argument so it can be used as a pointer. */
+	return hardwall_ioctl(file, a, (unsigned long)compat_ptr(b));
+}
+#endif
+
+/* The user process closed the file; revoke access to user networks. */
+static int hardwall_flush(struct file *file, fl_owner_t owner)
+{
+	struct hardwall_info *info = file->private_data;
+	struct task_struct *task, *tmp;
+	unsigned long flags;
+
+	if (info) {
+		/*
+		 * NOTE: if multiple threads are activated on this hardwall
+		 * file, the other threads will continue having access to the
+		 * user network until they are context-switched out and back
+		 * in again.
+		 *
+		 * NOTE: A NULL files pointer means the task is being torn
+		 * down, so in that case we also deactivate it.
+		 */
+		struct hardwall_type *hwt = info->type;
+		spin_lock_irqsave(&hwt->lock, flags);
+		list_for_each_entry_safe(task, tmp, &info->task_head,
+					 thread.hardwall[hwt->index].list) {
+			if (task->files == owner || task->files == NULL)
+				_hardwall_deactivate(hwt, task);
+		}
+		spin_unlock_irqrestore(&hwt->lock, flags);
+	}
+
+	return 0;
+}
+
+/* This hardwall is gone, so destroy it. */
+static int hardwall_release(struct inode *inode, struct file *file)
+{
+	hardwall_destroy(file->private_data);
+	return 0;
+}
+
+static const struct file_operations dev_hardwall_fops = {
+	.open           = nonseekable_open,
+	.unlocked_ioctl = hardwall_ioctl,
+#ifdef CONFIG_COMPAT
+	.compat_ioctl   = hardwall_compat_ioctl,
+#endif
+	.flush          = hardwall_flush,
+	.release        = hardwall_release,
+};
+
+static struct cdev hardwall_dev;
+
+static int __init dev_hardwall_init(void)
+{
+	int rc;
+	dev_t dev;
+
+	rc = alloc_chrdev_region(&dev, 0, HARDWALL_TYPES, "hardwall");
+	if (rc < 0)
+		return rc;
+	cdev_init(&hardwall_dev, &dev_hardwall_fops);
+	rc = cdev_add(&hardwall_dev, dev, HARDWALL_TYPES);
+	if (rc < 0)
+		return rc;
+
+	return 0;
+}
+late_initcall(dev_hardwall_init);
diff --git a/kernel/arch/tile/kernel/head_32.S b/kernel/arch/tile/kernel/head_32.S
new file mode 100644
index 000000000..8d5b40ff2
--- /dev/null
+++ b/kernel/arch/tile/kernel/head_32.S
@@ -0,0 +1,183 @@
+/*
+ * Copyright 2010 Tilera Corporation. All Rights Reserved.
+ *
+ *   This program is free software; you can redistribute it and/or
+ *   modify it under the terms of the GNU General Public License
+ *   as published by the Free Software Foundation, version 2.
+ *
+ *   This program is distributed in the hope that it will be useful, but
+ *   WITHOUT ANY WARRANTY; without even the implied warranty of
+ *   MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
+ *   NON INFRINGEMENT.  See the GNU General Public License for
+ *   more details.
+ *
+ * TILE startup code.
+ */
+
+#include <linux/linkage.h>
+#include <linux/init.h>
+#include <asm/page.h>
+#include <asm/pgtable.h>
+#include <asm/thread_info.h>
+#include <asm/processor.h>
+#include <asm/asm-offsets.h>
+#include <hv/hypervisor.h>
+#include <arch/chip.h>
+#include <arch/spr_def.h>
+
+/*
+ * This module contains the entry code for kernel images. It performs the
+ * minimal setup needed to call the generic C routines.
+ */
+
+	__HEAD
+ENTRY(_start)
+	/* Notify the hypervisor of what version of the API we want */
+	{
+	  movei r1, TILE_CHIP
+	  movei r2, TILE_CHIP_REV
+	}
+	{
+	  moveli r0, _HV_VERSION_OLD_HV_INIT
+	  jal _hv_init
+	}
+	/* Get a reasonable default ASID in r0 */
+	{
+	  move r0, zero
+	  jal _hv_inquire_asid
+	}
+	/* Install the default page table */
+	{
+	  moveli r6, lo16(swapper_pgprot - PAGE_OFFSET)
+	  move r4, r0     /* use starting ASID of range for this page table */
+	}
+	{
+	  moveli r0, lo16(swapper_pg_dir - PAGE_OFFSET)
+	  auli r6, r6, ha16(swapper_pgprot - PAGE_OFFSET)
+	}
+	{
+	  lw r2, r6
+	  addi r6, r6, 4
+	}
+	{
+	  lw r3, r6
+	  auli r0, r0, ha16(swapper_pg_dir - PAGE_OFFSET)
+	}
+	{
+	  finv r6
+	  move r1, zero   /* high 32 bits of CPA is zero */
+	}
+	{
+	  moveli lr, lo16(1f)
+	  moveli r5, CTX_PAGE_FLAG
+	}
+	{
+	  auli lr, lr, ha16(1f)
+	  j _hv_install_context
+	}
+1:
+
+	/* Get our processor number and save it away in SAVE_K_0. */
+	jal _hv_inquire_topology
+	mulll_uu r4, r1, r2        /* r1 == y, r2 == width */
+	add r4, r4, r0             /* r0 == x, so r4 == cpu == y*width + x */
+
+#ifdef CONFIG_SMP
+	/*
+	 * Load up our per-cpu offset.  When the first (master) tile
+	 * boots, this value is still zero, so we will load boot_pc
+	 * with start_kernel, and boot_sp at the top of init_stack.
+	 * The master tile initializes the per-cpu offset array, so that
+	 * when subsequent (secondary) tiles boot, they will instead load
+	 * from their per-cpu versions of boot_sp and boot_pc.
+	 */
+	moveli r5, lo16(__per_cpu_offset)
+	auli r5, r5, ha16(__per_cpu_offset)
+	s2a r5, r4, r5
+	lw r5, r5
+	bnz r5, 1f
+
+	/*
+	 * Save the width and height to the smp_topology variable
+	 * for later use.
+	 */
+	moveli r0, lo16(smp_topology + HV_TOPOLOGY_WIDTH_OFFSET)
+	auli r0, r0, ha16(smp_topology + HV_TOPOLOGY_WIDTH_OFFSET)
+	{
+	  sw r0, r2
+	  addi r0, r0, (HV_TOPOLOGY_HEIGHT_OFFSET - HV_TOPOLOGY_WIDTH_OFFSET)
+	}
+	sw r0, r3
+1:
+#else
+	move r5, zero
+#endif
+
+	/* Load and go with the correct pc and sp. */
+	{
+	  addli r1, r5, lo16(boot_sp)
+	  addli r0, r5, lo16(boot_pc)
+	}
+	{
+	  auli r1, r1, ha16(boot_sp)
+	  auli r0, r0, ha16(boot_pc)
+	}
+	lw r0, r0
+	lw sp, r1
+	or r4, sp, r4
+	mtspr SPR_SYSTEM_SAVE_K_0, r4  /* save ksp0 + cpu */
+	{
+	  move lr, zero   /* stop backtraces in the called function */
+	  jr r0
+	}
+	ENDPROC(_start)
+
+__PAGE_ALIGNED_BSS
+	.align PAGE_SIZE
+ENTRY(empty_zero_page)
+	.fill PAGE_SIZE,1,0
+	END(empty_zero_page)
+
+	.macro PTE va, cpa, bits1, no_org=0
+	.ifeq \no_org
+	.org swapper_pg_dir + PGD_INDEX(\va) * HV_PTE_SIZE
+	.endif
+	.word HV_PTE_PAGE | HV_PTE_DIRTY | HV_PTE_PRESENT | HV_PTE_ACCESSED | \
+	      (HV_PTE_MODE_CACHE_NO_L3 << HV_PTE_INDEX_MODE)
+	.word (\bits1) | (HV_CPA_TO_PTFN(\cpa) << (HV_PTE_INDEX_PTFN - 32))
+	.endm
+
+__PAGE_ALIGNED_DATA
+	.align PAGE_SIZE
+ENTRY(swapper_pg_dir)
+	/*
+	 * All data pages from PAGE_OFFSET to MEM_USER_INTRPT are mapped as
+	 * VA = PA + PAGE_OFFSET.  We remap things with more precise access
+	 * permissions and more respect for size of RAM later.
+	 */
+	.set addr, 0
+	.rept (MEM_USER_INTRPT - PAGE_OFFSET) >> PGDIR_SHIFT
+	PTE addr + PAGE_OFFSET, addr, (1 << (HV_PTE_INDEX_READABLE - 32)) | \
+				      (1 << (HV_PTE_INDEX_WRITABLE - 32))
+	.set addr, addr + PGDIR_SIZE
+	.endr
+
+	/* The true text VAs are mapped as VA = PA + MEM_SV_START */
+	PTE MEM_SV_START, 0, (1 << (HV_PTE_INDEX_READABLE - 32)) | \
+			      (1 << (HV_PTE_INDEX_EXECUTABLE - 32))
+	.org swapper_pg_dir + PGDIR_SIZE
+	END(swapper_pg_dir)
+
+	/*
+	 * Isolate swapper_pgprot to its own cache line, since each cpu
+	 * starting up will read it using VA-is-PA and local homing.
+	 * This would otherwise likely conflict with other data on the cache
+	 * line, once we have set its permanent home in the page tables.
+	 */
+	__INITDATA
+	.align CHIP_L2_LINE_SIZE()
+ENTRY(swapper_pgprot)
+	PTE	0, 0, (1 << (HV_PTE_INDEX_READABLE - 32)) | \
+		      (1 << (HV_PTE_INDEX_WRITABLE - 32)), 1
+	.align CHIP_L2_LINE_SIZE()
+	END(swapper_pgprot)
diff --git a/kernel/arch/tile/kernel/head_64.S b/kernel/arch/tile/kernel/head_64.S
new file mode 100644
index 000000000..bd0e12f28
--- /dev/null
+++ b/kernel/arch/tile/kernel/head_64.S
@@ -0,0 +1,279 @@
+/*
+ * Copyright 2011 Tilera Corporation. All Rights Reserved.
+ *
+ *   This program is free software; you can redistribute it and/or
+ *   modify it under the terms of the GNU General Public License
+ *   as published by the Free Software Foundation, version 2.
+ *
+ *   This program is distributed in the hope that it will be useful, but
+ *   WITHOUT ANY WARRANTY; without even the implied warranty of
+ *   MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
+ *   NON INFRINGEMENT.  See the GNU General Public License for
+ *   more details.
+ *
+ * TILE startup code.
+ */
+
+#include <linux/linkage.h>
+#include <linux/init.h>
+#include <asm/page.h>
+#include <asm/pgtable.h>
+#include <asm/thread_info.h>
+#include <asm/processor.h>
+#include <asm/asm-offsets.h>
+#include <hv/hypervisor.h>
+#include <arch/chip.h>
+#include <arch/spr_def.h>
+
+/* Extract two 32-bit bit values that were read into one register. */
+#ifdef __BIG_ENDIAN__
+#define GET_FIRST_INT(rd, rs) shrsi rd, rs, 32
+#define GET_SECOND_INT(rd, rs) addxi rd, rs, 0
+#else
+#define GET_FIRST_INT(rd, rs) addxi rd, rs, 0
+#define GET_SECOND_INT(rd, rs) shrsi rd, rs, 32
+#endif
+
+/*
+ * This module contains the entry code for kernel images. It performs the
+ * minimal setup needed to call the generic C routines.
+ */
+
+	__HEAD
+ENTRY(_start)
+	/* Notify the hypervisor of what version of the API we want */
+	{
+#if KERNEL_PL == 1 && _HV_VERSION == 13
+	  /* Support older hypervisors by asking for API version 12. */
+	  movei r0, _HV_VERSION_OLD_HV_INIT
+#else
+	  movei r0, _HV_VERSION
+#endif
+	  movei r1, TILE_CHIP
+	}
+	{
+	  movei r2, TILE_CHIP_REV
+	  movei r3, KERNEL_PL
+	}
+	jal _hv_init
+	/* Get a reasonable default ASID in r0 */
+	{
+	  move r0, zero
+	  jal _hv_inquire_asid
+	}
+
+	/*
+	 * Install the default page table.  The relocation required to
+	 * statically define the table is a bit too complex, so we have
+	 * to plug in the pointer from the L0 to the L1 table by hand.
+	 * We only do this on the first cpu to boot, though, since the
+	 * other CPUs should see a properly-constructed page table.
+	 */
+	{
+	  GET_FIRST_INT(r2, r0)    /* ASID for hv_install_context */
+	  moveli r4, hw1_last(swapper_pgprot - PAGE_OFFSET)
+	}
+	{
+	  shl16insli r4, r4, hw0(swapper_pgprot - PAGE_OFFSET)
+	}
+	{
+	  ld r1, r4               /* access_pte for hv_install_context */
+	}
+	{
+	  moveli r0, hw1_last(.Lsv_data_pmd - PAGE_OFFSET)
+	  moveli r6, hw1_last(temp_data_pmd - PAGE_OFFSET)
+	}
+	{
+	  /* After initializing swapper_pgprot, HV_PTE_GLOBAL is set. */
+	  bfextu r7, r1, HV_PTE_INDEX_GLOBAL, HV_PTE_INDEX_GLOBAL
+	  finv r4
+	}
+	bnez r7, .Lno_write
+	{
+	  shl16insli r0, r0, hw0(.Lsv_data_pmd - PAGE_OFFSET)
+	  shl16insli r6, r6, hw0(temp_data_pmd - PAGE_OFFSET)
+	}
+	{
+	  /* Cut off the low bits of the PT address. */
+	  shrui r6, r6, HV_LOG2_PAGE_TABLE_ALIGN
+	  /* Start with our access pte. */
+	  move r5, r1
+	}
+	{
+	  /* Stuff the address into the page table pointer slot of the PTE. */
+	  bfins r5, r6, HV_PTE_INDEX_PTFN, \
+			HV_PTE_INDEX_PTFN + HV_PTE_PTFN_BITS - 1
+	}
+	{
+	  /* Store the L0 data PTE. */
+	  st r0, r5
+	  addli r6, r6, (temp_code_pmd - temp_data_pmd) >> \
+			HV_LOG2_PAGE_TABLE_ALIGN
+	}
+	{
+	  addli r0, r0, .Lsv_code_pmd - .Lsv_data_pmd
+	  bfins r5, r6, HV_PTE_INDEX_PTFN, \
+			HV_PTE_INDEX_PTFN + HV_PTE_PTFN_BITS - 1
+	}
+	/* Store the L0 code PTE. */
+	st r0, r5
+
+.Lno_write:
+	moveli lr, hw2_last(1f)
+	{
+	  shl16insli lr, lr, hw1(1f)
+	  moveli r0, hw1_last(swapper_pg_dir - PAGE_OFFSET)
+	}
+	{
+	  shl16insli lr, lr, hw0(1f)
+	  shl16insli r0, r0, hw0(swapper_pg_dir - PAGE_OFFSET)
+	}
+	{
+	  moveli r3, CTX_PAGE_FLAG
+	  j _hv_install_context
+	}
+1:
+
+	/* Install the interrupt base. */
+	moveli r0, hw2_last(intrpt_start)
+	shl16insli r0, r0, hw1(intrpt_start)
+	shl16insli r0, r0, hw0(intrpt_start)
+	mtspr SPR_INTERRUPT_VECTOR_BASE_K, r0
+
+	/* Get our processor number and save it away in SAVE_K_0. */
+	jal _hv_inquire_topology
+	{
+	  GET_FIRST_INT(r5, r1)   /* r5 = width */
+	  GET_SECOND_INT(r4, r0)  /* r4 = y */
+	}
+	{
+	  GET_FIRST_INT(r6, r0)   /* r6 = x */
+	  mul_lu_lu r4, r4, r5
+	}
+	{
+	  add r4, r4, r6          /* r4 == cpu == y*width + x */
+	}
+
+#ifdef CONFIG_SMP
+	/*
+	 * Load up our per-cpu offset.  When the first (master) tile
+	 * boots, this value is still zero, so we will load boot_pc
+	 * with start_kernel, and boot_sp with at the top of init_stack.
+	 * The master tile initializes the per-cpu offset array, so that
+	 * when subsequent (secondary) tiles boot, they will instead load
+	 * from their per-cpu versions of boot_sp and boot_pc.
+	 */
+	moveli r5, hw2_last(__per_cpu_offset)
+	shl16insli r5, r5, hw1(__per_cpu_offset)
+	shl16insli r5, r5, hw0(__per_cpu_offset)
+	shl3add r5, r4, r5
+	ld r5, r5
+	bnez r5, 1f
+
+	/*
+	 * Save the width and height to the smp_topology variable
+	 * for later use.
+	 */
+	moveli r0, hw2_last(smp_topology + HV_TOPOLOGY_WIDTH_OFFSET)
+	shl16insli r0, r0, hw1(smp_topology + HV_TOPOLOGY_WIDTH_OFFSET)
+	shl16insli r0, r0, hw0(smp_topology + HV_TOPOLOGY_WIDTH_OFFSET)
+	st r0, r1
+1:
+#else
+	move r5, zero
+#endif
+
+	/* Load and go with the correct pc and sp. */
+	{
+	  moveli r1, hw2_last(boot_sp)
+	  moveli r0, hw2_last(boot_pc)
+	}
+	{
+	  shl16insli r1, r1, hw1(boot_sp)
+	  shl16insli r0, r0, hw1(boot_pc)
+	}
+	{
+	  shl16insli r1, r1, hw0(boot_sp)
+	  shl16insli r0, r0, hw0(boot_pc)
+	}
+	{
+	  add r1, r1, r5
+	  add r0, r0, r5
+	}
+	ld r0, r0
+	ld sp, r1
+	shli r4, r4, CPU_SHIFT
+	bfins r4, sp, 0, CPU_SHIFT-1
+	mtspr SPR_SYSTEM_SAVE_K_0, r4  /* save ksp0 + cpu */
+	{
+	  move lr, zero   /* stop backtraces in the called function */
+	  jr r0
+	}
+	ENDPROC(_start)
+
+__PAGE_ALIGNED_BSS
+	.align PAGE_SIZE
+ENTRY(empty_zero_page)
+	.fill PAGE_SIZE,1,0
+	END(empty_zero_page)
+
+	.macro PTE cpa, bits1
+	.quad HV_PTE_PAGE | HV_PTE_DIRTY | HV_PTE_PRESENT | HV_PTE_ACCESSED |\
+	      HV_PTE_GLOBAL | (HV_PTE_MODE_CACHE_NO_L3 << HV_PTE_INDEX_MODE) |\
+	      (\bits1) | (HV_CPA_TO_PTFN(\cpa) << HV_PTE_INDEX_PTFN)
+	.endm
+
+__PAGE_ALIGNED_DATA
+	.align PAGE_SIZE
+ENTRY(swapper_pg_dir)
+	.org swapper_pg_dir + PGD_INDEX(PAGE_OFFSET) * HV_PTE_SIZE
+.Lsv_data_pmd:
+	.quad 0  /* PTE temp_data_pmd - PAGE_OFFSET, 0 */
+	.org swapper_pg_dir + PGD_INDEX(MEM_SV_START) * HV_PTE_SIZE
+.Lsv_code_pmd:
+	.quad 0  /* PTE temp_code_pmd - PAGE_OFFSET, 0 */
+	.org swapper_pg_dir + SIZEOF_PGD
+	END(swapper_pg_dir)
+
+	.align HV_PAGE_TABLE_ALIGN
+ENTRY(temp_data_pmd)
+	/*
+	 * We fill the PAGE_OFFSET pmd with huge pages with
+	 * VA = PA + PAGE_OFFSET.  We remap things with more precise access
+	 * permissions later.
+	 */
+	.set addr, 0
+	.rept PTRS_PER_PMD
+	PTE addr, HV_PTE_READABLE | HV_PTE_WRITABLE
+	.set addr, addr + HPAGE_SIZE
+	.endr
+	.org temp_data_pmd + SIZEOF_PMD
+	END(temp_data_pmd)
+
+	.align HV_PAGE_TABLE_ALIGN
+ENTRY(temp_code_pmd)
+	/*
+	 * We fill the MEM_SV_START pmd with huge pages with
+	 * VA = PA + PAGE_OFFSET.  We remap things with more precise access
+	 * permissions later.
+	 */
+	.set addr, 0
+	.rept PTRS_PER_PMD
+	PTE addr, HV_PTE_READABLE | HV_PTE_EXECUTABLE
+	.set addr, addr + HPAGE_SIZE
+	.endr
+	.org temp_code_pmd + SIZEOF_PMD
+	END(temp_code_pmd)
+
+	/*
+	 * Isolate swapper_pgprot to its own cache line, since each cpu
+	 * starting up will read it using VA-is-PA and local homing.
+	 * This would otherwise likely conflict with other data on the cache
+	 * line, once we have set its permanent home in the page tables.
+	 */
+	__INITDATA
+	.align CHIP_L2_LINE_SIZE()
+ENTRY(swapper_pgprot)
+	.quad HV_PTE_PRESENT | (HV_PTE_MODE_CACHE_NO_L3 << HV_PTE_INDEX_MODE)
+	.align CHIP_L2_LINE_SIZE()
+	END(swapper_pgprot)
diff --git a/kernel/arch/tile/kernel/hvglue.S b/kernel/arch/tile/kernel/hvglue.S
new file mode 100644
index 000000000..2ab456622
--- /dev/null
+++ b/kernel/arch/tile/kernel/hvglue.S
@@ -0,0 +1,74 @@
+/* Hypervisor call vector addresses; see <hv/hypervisor.h> */
+.macro gensym sym, val, size
+.org \val
+.global _\sym
+.type _\sym,function
+_\sym:
+.size _\sym,\size
+#ifndef CONFIG_TILE_HVGLUE_TRACE
+.globl \sym
+.set \sym,_\sym
+#endif
+.endm
+
+.section .hvglue,"x",@nobits
+.align 8
+gensym hv_init, 0x20, 32
+gensym hv_install_context, 0x40, 32
+gensym hv_sysconf, 0x60, 32
+gensym hv_get_rtc, 0x80, 32
+gensym hv_set_rtc, 0xa0, 32
+gensym hv_flush_asid, 0xc0, 32
+gensym hv_flush_page, 0xe0, 32
+gensym hv_flush_pages, 0x100, 32
+gensym hv_restart, 0x120, 32
+gensym hv_halt, 0x140, 32
+gensym hv_power_off, 0x160, 32
+gensym hv_inquire_physical, 0x180, 32
+gensym hv_inquire_memory_controller, 0x1a0, 32
+gensym hv_inquire_virtual, 0x1c0, 32
+gensym hv_inquire_asid, 0x1e0, 32
+gensym hv_nanosleep, 0x200, 32
+gensym hv_console_read_if_ready, 0x220, 32
+gensym hv_console_write, 0x240, 32
+gensym hv_downcall_dispatch, 0x260, 32
+gensym hv_inquire_topology, 0x280, 32
+gensym hv_fs_findfile, 0x2a0, 32
+gensym hv_fs_fstat, 0x2c0, 32
+gensym hv_fs_pread, 0x2e0, 32
+gensym hv_physaddr_read64, 0x300, 32
+gensym hv_physaddr_write64, 0x320, 32
+gensym hv_get_command_line, 0x340, 32
+gensym hv_set_caching, 0x360, 32
+gensym hv_bzero_page, 0x380, 32
+gensym hv_register_message_state, 0x3a0, 32
+gensym hv_send_message, 0x3c0, 32
+gensym hv_receive_message, 0x3e0, 32
+gensym hv_inquire_context, 0x400, 32
+gensym hv_start_all_tiles, 0x420, 32
+gensym hv_dev_open, 0x440, 32
+gensym hv_dev_close, 0x460, 32
+gensym hv_dev_pread, 0x480, 32
+gensym hv_dev_pwrite, 0x4a0, 32
+gensym hv_dev_poll, 0x4c0, 32
+gensym hv_dev_poll_cancel, 0x4e0, 32
+gensym hv_dev_preada, 0x500, 32
+gensym hv_dev_pwritea, 0x520, 32
+gensym hv_flush_remote, 0x540, 32
+gensym hv_console_putc, 0x560, 32
+gensym hv_inquire_tiles, 0x580, 32
+gensym hv_confstr, 0x5a0, 32
+gensym hv_reexec, 0x5c0, 32
+gensym hv_set_command_line, 0x5e0, 32
+gensym hv_clear_intr, 0x600, 32
+gensym hv_enable_intr, 0x620, 32
+gensym hv_disable_intr, 0x640, 32
+gensym hv_raise_intr, 0x660, 32
+gensym hv_trigger_ipi, 0x680, 32
+gensym hv_store_mapping, 0x6a0, 32
+gensym hv_inquire_realpa, 0x6c0, 32
+gensym hv_flush_all, 0x6e0, 32
+gensym hv_get_ipi_pte, 0x700, 32
+gensym hv_set_pte_super_shift, 0x720, 32
+gensym hv_console_set_ipi, 0x7e0, 32
+gensym hv_glue_internals, 0x800, 30720
diff --git a/kernel/arch/tile/kernel/hvglue_trace.c b/kernel/arch/tile/kernel/hvglue_trace.c
new file mode 100644
index 000000000..85c74ad29
--- /dev/null
+++ b/kernel/arch/tile/kernel/hvglue_trace.c
@@ -0,0 +1,266 @@
+/*
+ * Copyright 2013 Tilera Corporation. All Rights Reserved.
+ *
+ *   This program is free software; you can redistribute it and/or
+ *   modify it under the terms of the GNU General Public License
+ *   as published by the Free Software Foundation, version 2.
+ *
+ *   This program is distributed in the hope that it will be useful, but
+ *   WITHOUT ANY WARRANTY; without even the implied warranty of
+ *   MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
+ *   NON INFRINGEMENT.  See the GNU General Public License for
+ *   more details.
+ */
+
+/*
+ * Pull in the hypervisor header so we declare all the ABI functions
+ * with the underscore versions, then undef the names so that we can
+ * provide our own wrapper versions.
+ */
+#define hv_init _hv_init
+#define hv_install_context _hv_install_context
+#define hv_sysconf _hv_sysconf
+#define hv_get_rtc _hv_get_rtc
+#define hv_set_rtc _hv_set_rtc
+#define hv_flush_asid _hv_flush_asid
+#define hv_flush_page _hv_flush_page
+#define hv_flush_pages _hv_flush_pages
+#define hv_restart _hv_restart
+#define hv_halt _hv_halt
+#define hv_power_off _hv_power_off
+#define hv_inquire_physical _hv_inquire_physical
+#define hv_inquire_memory_controller _hv_inquire_memory_controller
+#define hv_inquire_virtual _hv_inquire_virtual
+#define hv_inquire_asid _hv_inquire_asid
+#define hv_nanosleep _hv_nanosleep
+#define hv_console_read_if_ready _hv_console_read_if_ready
+#define hv_console_write _hv_console_write
+#define hv_downcall_dispatch _hv_downcall_dispatch
+#define hv_inquire_topology _hv_inquire_topology
+#define hv_fs_findfile _hv_fs_findfile
+#define hv_fs_fstat _hv_fs_fstat
+#define hv_fs_pread _hv_fs_pread
+#define hv_physaddr_read64 _hv_physaddr_read64
+#define hv_physaddr_write64 _hv_physaddr_write64
+#define hv_get_command_line _hv_get_command_line
+#define hv_set_caching _hv_set_caching
+#define hv_bzero_page _hv_bzero_page
+#define hv_register_message_state _hv_register_message_state
+#define hv_send_message _hv_send_message
+#define hv_receive_message _hv_receive_message
+#define hv_inquire_context _hv_inquire_context
+#define hv_start_all_tiles _hv_start_all_tiles
+#define hv_dev_open _hv_dev_open
+#define hv_dev_close _hv_dev_close
+#define hv_dev_pread _hv_dev_pread
+#define hv_dev_pwrite _hv_dev_pwrite
+#define hv_dev_poll _hv_dev_poll
+#define hv_dev_poll_cancel _hv_dev_poll_cancel
+#define hv_dev_preada _hv_dev_preada
+#define hv_dev_pwritea _hv_dev_pwritea
+#define hv_flush_remote _hv_flush_remote
+#define hv_console_putc _hv_console_putc
+#define hv_inquire_tiles _hv_inquire_tiles
+#define hv_confstr _hv_confstr
+#define hv_reexec _hv_reexec
+#define hv_set_command_line _hv_set_command_line
+#define hv_clear_intr _hv_clear_intr
+#define hv_enable_intr _hv_enable_intr
+#define hv_disable_intr _hv_disable_intr
+#define hv_raise_intr _hv_raise_intr
+#define hv_trigger_ipi _hv_trigger_ipi
+#define hv_store_mapping _hv_store_mapping
+#define hv_inquire_realpa _hv_inquire_realpa
+#define hv_flush_all _hv_flush_all
+#define hv_get_ipi_pte _hv_get_ipi_pte
+#define hv_set_pte_super_shift _hv_set_pte_super_shift
+#define hv_console_set_ipi _hv_console_set_ipi
+#include <hv/hypervisor.h>
+#undef hv_init
+#undef hv_install_context
+#undef hv_sysconf
+#undef hv_get_rtc
+#undef hv_set_rtc
+#undef hv_flush_asid
+#undef hv_flush_page
+#undef hv_flush_pages
+#undef hv_restart
+#undef hv_halt
+#undef hv_power_off
+#undef hv_inquire_physical
+#undef hv_inquire_memory_controller
+#undef hv_inquire_virtual
+#undef hv_inquire_asid
+#undef hv_nanosleep
+#undef hv_console_read_if_ready
+#undef hv_console_write
+#undef hv_downcall_dispatch
+#undef hv_inquire_topology
+#undef hv_fs_findfile
+#undef hv_fs_fstat
+#undef hv_fs_pread
+#undef hv_physaddr_read64
+#undef hv_physaddr_write64
+#undef hv_get_command_line
+#undef hv_set_caching
+#undef hv_bzero_page
+#undef hv_register_message_state
+#undef hv_send_message
+#undef hv_receive_message
+#undef hv_inquire_context
+#undef hv_start_all_tiles
+#undef hv_dev_open
+#undef hv_dev_close
+#undef hv_dev_pread
+#undef hv_dev_pwrite
+#undef hv_dev_poll
+#undef hv_dev_poll_cancel
+#undef hv_dev_preada
+#undef hv_dev_pwritea
+#undef hv_flush_remote
+#undef hv_console_putc
+#undef hv_inquire_tiles
+#undef hv_confstr
+#undef hv_reexec
+#undef hv_set_command_line
+#undef hv_clear_intr
+#undef hv_enable_intr
+#undef hv_disable_intr
+#undef hv_raise_intr
+#undef hv_trigger_ipi
+#undef hv_store_mapping
+#undef hv_inquire_realpa
+#undef hv_flush_all
+#undef hv_get_ipi_pte
+#undef hv_set_pte_super_shift
+#undef hv_console_set_ipi
+
+/*
+ * Provide macros based on <linux/syscalls.h> to provide a wrapper
+ * function that invokes the same function with an underscore prefix.
+ * We can't use the existing __SC_xxx macros because we need to
+ * support up to nine arguments rather than up to six, and also this
+ * way the file stands alone from possible changes in the
+ * implementation of <linux/syscalls.h>.
+ */
+#define HV_WRAP0(type, name)					\
+	type name(void);					\
+	type name(void)						\
+	{							\
+		return _##name();				\
+	}
+#define __HV_DECL1(t1, a1)	t1 a1
+#define __HV_DECL2(t2, a2, ...) t2 a2, __HV_DECL1(__VA_ARGS__)
+#define __HV_DECL3(t3, a3, ...) t3 a3, __HV_DECL2(__VA_ARGS__)
+#define __HV_DECL4(t4, a4, ...) t4 a4, __HV_DECL3(__VA_ARGS__)
+#define __HV_DECL5(t5, a5, ...) t5 a5, __HV_DECL4(__VA_ARGS__)
+#define __HV_DECL6(t6, a6, ...) t6 a6, __HV_DECL5(__VA_ARGS__)
+#define __HV_DECL7(t7, a7, ...) t7 a7, __HV_DECL6(__VA_ARGS__)
+#define __HV_DECL8(t8, a8, ...) t8 a8, __HV_DECL7(__VA_ARGS__)
+#define __HV_DECL9(t9, a9, ...) t9 a9, __HV_DECL8(__VA_ARGS__)
+#define __HV_PASS1(t1, a1)	a1
+#define __HV_PASS2(t2, a2, ...) a2, __HV_PASS1(__VA_ARGS__)
+#define __HV_PASS3(t3, a3, ...) a3, __HV_PASS2(__VA_ARGS__)
+#define __HV_PASS4(t4, a4, ...) a4, __HV_PASS3(__VA_ARGS__)
+#define __HV_PASS5(t5, a5, ...) a5, __HV_PASS4(__VA_ARGS__)
+#define __HV_PASS6(t6, a6, ...) a6, __HV_PASS5(__VA_ARGS__)
+#define __HV_PASS7(t7, a7, ...) a7, __HV_PASS6(__VA_ARGS__)
+#define __HV_PASS8(t8, a8, ...) a8, __HV_PASS7(__VA_ARGS__)
+#define __HV_PASS9(t9, a9, ...) a9, __HV_PASS8(__VA_ARGS__)
+#define HV_WRAPx(x, type, name, ...)				\
+	type name(__HV_DECL##x(__VA_ARGS__));			\
+	type name(__HV_DECL##x(__VA_ARGS__))			\
+	{							\
+		return _##name(__HV_PASS##x(__VA_ARGS__));	\
+	}
+#define HV_WRAP1(type, name, ...) HV_WRAPx(1, type, name, __VA_ARGS__)
+#define HV_WRAP2(type, name, ...) HV_WRAPx(2, type, name, __VA_ARGS__)
+#define HV_WRAP3(type, name, ...) HV_WRAPx(3, type, name, __VA_ARGS__)
+#define HV_WRAP4(type, name, ...) HV_WRAPx(4, type, name, __VA_ARGS__)
+#define HV_WRAP5(type, name, ...) HV_WRAPx(5, type, name, __VA_ARGS__)
+#define HV_WRAP6(type, name, ...) HV_WRAPx(6, type, name, __VA_ARGS__)
+#define HV_WRAP7(type, name, ...) HV_WRAPx(7, type, name, __VA_ARGS__)
+#define HV_WRAP8(type, name, ...) HV_WRAPx(8, type, name, __VA_ARGS__)
+#define HV_WRAP9(type, name, ...) HV_WRAPx(9, type, name, __VA_ARGS__)
+
+/* List all the hypervisor API functions. */
+HV_WRAP4(void, hv_init, HV_VersionNumber, interface_version_number,
+	 int, chip_num, int, chip_rev_num, int, client_pl)
+HV_WRAP1(long, hv_sysconf, HV_SysconfQuery, query)
+HV_WRAP3(int, hv_confstr, HV_ConfstrQuery, query, HV_VirtAddr, buf, int, len)
+#if CHIP_HAS_IPI()
+HV_WRAP3(int, hv_get_ipi_pte, HV_Coord, tile, int, pl, HV_PTE*, pte)
+HV_WRAP3(int, hv_console_set_ipi, int, ipi, int, event, HV_Coord, coord);
+#else
+HV_WRAP1(void, hv_enable_intr, HV_IntrMask, enab_mask)
+HV_WRAP1(void, hv_disable_intr, HV_IntrMask, disab_mask)
+HV_WRAP1(void, hv_clear_intr, HV_IntrMask, clear_mask)
+HV_WRAP1(void, hv_raise_intr, HV_IntrMask, raise_mask)
+HV_WRAP2(HV_Errno, hv_trigger_ipi, HV_Coord, tile, int, interrupt)
+#endif /* !CHIP_HAS_IPI() */
+HV_WRAP3(int, hv_store_mapping, HV_VirtAddr, va, unsigned int, len,
+	 HV_PhysAddr, pa)
+HV_WRAP2(HV_PhysAddr, hv_inquire_realpa, HV_PhysAddr, cpa, unsigned int, len)
+HV_WRAP0(HV_RTCTime, hv_get_rtc)
+HV_WRAP1(void, hv_set_rtc, HV_RTCTime, time)
+HV_WRAP4(int, hv_install_context, HV_PhysAddr, page_table, HV_PTE, access,
+	 HV_ASID, asid, __hv32, flags)
+HV_WRAP2(int, hv_set_pte_super_shift, int, level, int, log2_count)
+HV_WRAP0(HV_Context, hv_inquire_context)
+HV_WRAP1(int, hv_flush_asid, HV_ASID, asid)
+HV_WRAP2(int, hv_flush_page, HV_VirtAddr, address, HV_PageSize, page_size)
+HV_WRAP3(int, hv_flush_pages, HV_VirtAddr, start, HV_PageSize, page_size,
+	 unsigned long, size)
+HV_WRAP1(int, hv_flush_all, int, preserve_global)
+HV_WRAP2(void, hv_restart, HV_VirtAddr, cmd, HV_VirtAddr, args)
+HV_WRAP0(void, hv_halt)
+HV_WRAP0(void, hv_power_off)
+HV_WRAP1(int, hv_reexec, HV_PhysAddr, entry)
+HV_WRAP0(HV_Topology, hv_inquire_topology)
+HV_WRAP3(HV_Errno, hv_inquire_tiles, HV_InqTileSet, set, HV_VirtAddr, cpumask,
+	 int, length)
+HV_WRAP1(HV_PhysAddrRange, hv_inquire_physical, int, idx)
+HV_WRAP2(HV_MemoryControllerInfo, hv_inquire_memory_controller, HV_Coord, coord,
+	 int, controller)
+HV_WRAP1(HV_VirtAddrRange, hv_inquire_virtual, int, idx)
+HV_WRAP1(HV_ASIDRange, hv_inquire_asid, int, idx)
+HV_WRAP1(void, hv_nanosleep, int, nanosecs)
+HV_WRAP0(int, hv_console_read_if_ready)
+HV_WRAP1(void, hv_console_putc, int, byte)
+HV_WRAP2(int, hv_console_write, HV_VirtAddr, bytes, int, len)
+HV_WRAP0(void, hv_downcall_dispatch)
+HV_WRAP1(int, hv_fs_findfile, HV_VirtAddr, filename)
+HV_WRAP1(HV_FS_StatInfo, hv_fs_fstat, int, inode)
+HV_WRAP4(int, hv_fs_pread, int, inode, HV_VirtAddr, buf,
+	 int, length, int, offset)
+HV_WRAP2(unsigned long long, hv_physaddr_read64, HV_PhysAddr, addr,
+	 HV_PTE, access)
+HV_WRAP3(void, hv_physaddr_write64, HV_PhysAddr, addr, HV_PTE, access,
+	 unsigned long long, val)
+HV_WRAP2(int, hv_get_command_line, HV_VirtAddr, buf, int, length)
+HV_WRAP2(HV_Errno, hv_set_command_line, HV_VirtAddr, buf, int, length)
+HV_WRAP1(void, hv_set_caching, unsigned long, bitmask)
+HV_WRAP2(void, hv_bzero_page, HV_VirtAddr, va, unsigned int, size)
+HV_WRAP1(HV_Errno, hv_register_message_state, HV_MsgState*, msgstate)
+HV_WRAP4(int, hv_send_message, HV_Recipient *, recips, int, nrecip,
+	 HV_VirtAddr, buf, int, buflen)
+HV_WRAP3(HV_RcvMsgInfo, hv_receive_message, HV_MsgState, msgstate,
+	 HV_VirtAddr, buf, int, buflen)
+HV_WRAP0(void, hv_start_all_tiles)
+HV_WRAP2(int, hv_dev_open, HV_VirtAddr, name, __hv32, flags)
+HV_WRAP1(int, hv_dev_close, int, devhdl)
+HV_WRAP5(int, hv_dev_pread, int, devhdl, __hv32, flags, HV_VirtAddr, va,
+	 __hv32, len, __hv64, offset)
+HV_WRAP5(int, hv_dev_pwrite, int, devhdl, __hv32, flags, HV_VirtAddr, va,
+	 __hv32, len, __hv64, offset)
+HV_WRAP3(int, hv_dev_poll, int, devhdl, __hv32, events, HV_IntArg, intarg)
+HV_WRAP1(int, hv_dev_poll_cancel, int, devhdl)
+HV_WRAP6(int, hv_dev_preada, int, devhdl, __hv32, flags, __hv32, sgl_len,
+	 HV_SGL *, sglp, __hv64, offset, HV_IntArg, intarg)
+HV_WRAP6(int, hv_dev_pwritea, int, devhdl, __hv32, flags, __hv32, sgl_len,
+	 HV_SGL *, sglp, __hv64, offset, HV_IntArg, intarg)
+HV_WRAP9(int, hv_flush_remote, HV_PhysAddr, cache_pa,
+	 unsigned long, cache_control, unsigned long*, cache_cpumask,
+	 HV_VirtAddr, tlb_va, unsigned long, tlb_length,
+	 unsigned long, tlb_pgsize, unsigned long*, tlb_cpumask,
+	 HV_Remote_ASID*, asids, int, asidcount)
diff --git a/kernel/arch/tile/kernel/intvec_32.S b/kernel/arch/tile/kernel/intvec_32.S
new file mode 100644
index 000000000..cdbda45a4
--- /dev/null
+++ b/kernel/arch/tile/kernel/intvec_32.S
@@ -0,0 +1,1920 @@
+/*
+ * Copyright 2010 Tilera Corporation. All Rights Reserved.
+ *
+ *   This program is free software; you can redistribute it and/or
+ *   modify it under the terms of the GNU General Public License
+ *   as published by the Free Software Foundation, version 2.
+ *
+ *   This program is distributed in the hope that it will be useful, but
+ *   WITHOUT ANY WARRANTY; without even the implied warranty of
+ *   MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
+ *   NON INFRINGEMENT.  See the GNU General Public License for
+ *   more details.
+ *
+ * Linux interrupt vectors.
+ */
+
+#include <linux/linkage.h>
+#include <linux/errno.h>
+#include <linux/init.h>
+#include <linux/unistd.h>
+#include <asm/ptrace.h>
+#include <asm/thread_info.h>
+#include <asm/irqflags.h>
+#include <asm/atomic_32.h>
+#include <asm/asm-offsets.h>
+#include <hv/hypervisor.h>
+#include <arch/abi.h>
+#include <arch/interrupts.h>
+#include <arch/spr_def.h>
+
+#define PTREGS_PTR(reg, ptreg) addli reg, sp, C_ABI_SAVE_AREA_SIZE + (ptreg)
+
+#define PTREGS_OFFSET_SYSCALL PTREGS_OFFSET_REG(TREG_SYSCALL_NR)
+
+	.macro  push_reg reg, ptr=sp, delta=-4
+	{
+	 sw     \ptr, \reg
+	 addli  \ptr, \ptr, \delta
+	}
+	.endm
+
+	.macro  pop_reg reg, ptr=sp, delta=4
+	{
+	 lw     \reg, \ptr
+	 addli  \ptr, \ptr, \delta
+	}
+	.endm
+
+	.macro  pop_reg_zero reg, zreg, ptr=sp, delta=4
+	{
+	 move   \zreg, zero
+	 lw     \reg, \ptr
+	 addi   \ptr, \ptr, \delta
+	}
+	.endm
+
+	.macro  push_extra_callee_saves reg
+	PTREGS_PTR(\reg, PTREGS_OFFSET_REG(51))
+	push_reg r51, \reg
+	push_reg r50, \reg
+	push_reg r49, \reg
+	push_reg r48, \reg
+	push_reg r47, \reg
+	push_reg r46, \reg
+	push_reg r45, \reg
+	push_reg r44, \reg
+	push_reg r43, \reg
+	push_reg r42, \reg
+	push_reg r41, \reg
+	push_reg r40, \reg
+	push_reg r39, \reg
+	push_reg r38, \reg
+	push_reg r37, \reg
+	push_reg r36, \reg
+	push_reg r35, \reg
+	push_reg r34, \reg, PTREGS_OFFSET_BASE - PTREGS_OFFSET_REG(34)
+	.endm
+
+	.macro  panic str
+	.pushsection .rodata, "a"
+1:
+	.asciz  "\str"
+	.popsection
+	{
+	 moveli r0, lo16(1b)
+	}
+	{
+	 auli   r0, r0, ha16(1b)
+	 jal    panic
+	}
+	.endm
+
+#ifdef __COLLECT_LINKER_FEEDBACK__
+	.pushsection .text.intvec_feedback,"ax"
+intvec_feedback:
+	.popsection
+#endif
+
+	/*
+	 * Default interrupt handler.
+	 *
+	 * vecnum is where we'll put this code.
+	 * c_routine is the C routine we'll call.
+	 *
+	 * The C routine is passed two arguments:
+	 * - A pointer to the pt_regs state.
+	 * - The interrupt vector number.
+	 *
+	 * The "processing" argument specifies the code for processing
+	 * the interrupt. Defaults to "handle_interrupt".
+	 */
+	.macro  int_hand vecnum, vecname, c_routine, processing=handle_interrupt
+	.org    (\vecnum << 8)
+intvec_\vecname:
+	.ifc    \vecnum, INT_SWINT_1
+	blz     TREG_SYSCALL_NR_NAME, sys_cmpxchg
+	.endif
+
+	/* Temporarily save a register so we have somewhere to work. */
+
+	mtspr   SPR_SYSTEM_SAVE_K_1, r0
+	mfspr   r0, SPR_EX_CONTEXT_K_1
+
+	/* The cmpxchg code clears sp to force us to reset it here on fault. */
+	{
+	 bz     sp, 2f
+	 andi   r0, r0, SPR_EX_CONTEXT_1_1__PL_MASK  /* mask off ICS */
+	}
+
+	.ifc    \vecnum, INT_DOUBLE_FAULT
+	/*
+	 * For double-faults from user-space, fall through to the normal
+	 * register save and stack setup path.  Otherwise, it's the
+	 * hypervisor giving us one last chance to dump diagnostics, and we
+	 * branch to the kernel_double_fault routine to do so.
+	 */
+	bz      r0, 1f
+	j       _kernel_double_fault
+1:
+	.else
+	/*
+	 * If we're coming from user-space, then set sp to the top of
+	 * the kernel stack.  Otherwise, assume sp is already valid.
+	 */
+	{
+	 bnz    r0, 0f
+	 move   r0, sp
+	}
+	.endif
+
+	.ifc    \c_routine, do_page_fault
+	/*
+	 * The page_fault handler may be downcalled directly by the
+	 * hypervisor even when Linux is running and has ICS set.
+	 *
+	 * In this case the contents of EX_CONTEXT_K_1 reflect the
+	 * previous fault and can't be relied on to choose whether or
+	 * not to reinitialize the stack pointer.  So we add a test
+	 * to see whether SYSTEM_SAVE_K_2 has the high bit set,
+	 * and if so we don't reinitialize sp, since we must be coming
+	 * from Linux.  (In fact the precise case is !(val & ~1),
+	 * but any Linux PC has to have the high bit set.)
+	 *
+	 * Note that the hypervisor *always* sets SYSTEM_SAVE_K_2 for
+	 * any path that turns into a downcall to one of our TLB handlers.
+	 */
+	mfspr   r0, SPR_SYSTEM_SAVE_K_2
+	{
+	 blz    r0, 0f    /* high bit in S_S_1_2 is for a PC to use */
+	 move   r0, sp
+	}
+	.endif
+
+2:
+	/*
+	 * SYSTEM_SAVE_K_0 holds the cpu number in the low bits, and
+	 * the current stack top in the higher bits.  So we recover
+	 * our stack top by just masking off the low bits, then
+	 * point sp at the top aligned address on the actual stack page.
+	 */
+	mfspr   r0, SPR_SYSTEM_SAVE_K_0
+	mm      r0, r0, zero, LOG2_NR_CPU_IDS, 31
+
+0:
+	/*
+	 * Align the stack mod 64 so we can properly predict what
+	 * cache lines we need to write-hint to reduce memory fetch
+	 * latency as we enter the kernel.  The layout of memory is
+	 * as follows, with cache line 0 at the lowest VA, and cache
+	 * line 4 just below the r0 value this "andi" computes.
+	 * Note that we never write to cache line 4, and we skip
+	 * cache line 1 for syscalls.
+	 *
+	 *    cache line 4: ptregs padding (two words)
+	 *    cache line 3: r46...lr, pc, ex1, faultnum, orig_r0, flags, pad
+	 *    cache line 2: r30...r45
+	 *    cache line 1: r14...r29
+	 *    cache line 0: 2 x frame, r0..r13
+	 */
+#if STACK_TOP_DELTA != 64
+#error STACK_TOP_DELTA must be 64 for assumptions here and in task_pt_regs()
+#endif
+	andi    r0, r0, -64
+
+	/*
+	 * Push the first four registers on the stack, so that we can set
+	 * them to vector-unique values before we jump to the common code.
+	 *
+	 * Registers are pushed on the stack as a struct pt_regs,
+	 * with the sp initially just above the struct, and when we're
+	 * done, sp points to the base of the struct, minus
+	 * C_ABI_SAVE_AREA_SIZE, so we can directly jal to C code.
+	 *
+	 * This routine saves just the first four registers, plus the
+	 * stack context so we can do proper backtracing right away,
+	 * and defers to handle_interrupt to save the rest.
+	 * The backtracer needs pc, ex1, lr, sp, r52, and faultnum.
+	 */
+	addli   r0, r0, PTREGS_OFFSET_LR - (PTREGS_SIZE + KSTK_PTREGS_GAP)
+	wh64    r0    /* cache line 3 */
+	{
+	 sw     r0, lr
+	 addli  r0, r0, PTREGS_OFFSET_SP - PTREGS_OFFSET_LR
+	}
+	{
+	 sw     r0, sp
+	 addli  sp, r0, PTREGS_OFFSET_REG(52) - PTREGS_OFFSET_SP
+	}
+	{
+	 sw     sp, r52
+	 addli  sp, sp, PTREGS_OFFSET_REG(1) - PTREGS_OFFSET_REG(52)
+	}
+	wh64    sp    /* cache line 0 */
+	{
+	 sw     sp, r1
+	 addli  sp, sp, PTREGS_OFFSET_REG(2) - PTREGS_OFFSET_REG(1)
+	}
+	{
+	 sw     sp, r2
+	 addli  sp, sp, PTREGS_OFFSET_REG(3) - PTREGS_OFFSET_REG(2)
+	}
+	{
+	 sw     sp, r3
+	 addli  sp, sp, PTREGS_OFFSET_PC - PTREGS_OFFSET_REG(3)
+	}
+	mfspr   r0, SPR_EX_CONTEXT_K_0
+	.ifc \processing,handle_syscall
+	/*
+	 * Bump the saved PC by one bundle so that when we return, we won't
+	 * execute the same swint instruction again.  We need to do this while
+	 * we're in the critical section.
+	 */
+	addi    r0, r0, 8
+	.endif
+	{
+	 sw     sp, r0
+	 addli  sp, sp, PTREGS_OFFSET_EX1 - PTREGS_OFFSET_PC
+	}
+	mfspr   r0, SPR_EX_CONTEXT_K_1
+	{
+	 sw     sp, r0
+	 addi   sp, sp, PTREGS_OFFSET_FAULTNUM - PTREGS_OFFSET_EX1
+	/*
+	 * Use r0 for syscalls so it's a temporary; use r1 for interrupts
+	 * so that it gets passed through unchanged to the handler routine.
+	 * Note that the .if conditional confusingly spans bundles.
+	 */
+	 .ifc \processing,handle_syscall
+	 movei  r0, \vecnum
+	}
+	{
+	 sw     sp, r0
+	 .else
+	 movei  r1, \vecnum
+	}
+	{
+	 sw     sp, r1
+	 .endif
+	 addli  sp, sp, PTREGS_OFFSET_REG(0) - PTREGS_OFFSET_FAULTNUM
+	}
+	mfspr   r0, SPR_SYSTEM_SAVE_K_1    /* Original r0 */
+	{
+	 sw     sp, r0
+	 addi   sp, sp, -PTREGS_OFFSET_REG(0) - 4
+	}
+	{
+	 sw     sp, zero        /* write zero into "Next SP" frame pointer */
+	 addi   sp, sp, -4      /* leave SP pointing at bottom of frame */
+	}
+	.ifc \processing,handle_syscall
+	j       handle_syscall
+	.else
+	/*
+	 * Capture per-interrupt SPR context to registers.
+	 * We overload the meaning of r3 on this path such that if its bit 31
+	 * is set, we have to mask all interrupts including NMIs before
+	 * clearing the interrupt critical section bit.
+	 * See discussion below at "finish_interrupt_save".
+	 */
+	.ifc \c_routine, do_page_fault
+	mfspr   r2, SPR_SYSTEM_SAVE_K_3   /* address of page fault */
+	mfspr   r3, SPR_SYSTEM_SAVE_K_2   /* info about page fault */
+	.else
+	.ifc \vecnum, INT_DOUBLE_FAULT
+	{
+	 mfspr  r2, SPR_SYSTEM_SAVE_K_2   /* double fault info from HV */
+	 movei  r3, 0
+	}
+	.else
+	.ifc \c_routine, do_trap
+	{
+	 mfspr  r2, GPV_REASON
+	 movei  r3, 0
+	}
+	.else
+	.ifc \c_routine, handle_perf_interrupt
+	{
+	 mfspr  r2, PERF_COUNT_STS
+	 movei  r3, -1   /* not used, but set for consistency */
+	}
+	.else
+	.ifc \c_routine, handle_perf_interrupt
+	{
+	 mfspr  r2, AUX_PERF_COUNT_STS
+	 movei  r3, -1   /* not used, but set for consistency */
+	}
+	.else
+	movei   r3, 0
+	.endif
+	.endif
+	.endif
+	.endif
+	.endif
+	/* Put function pointer in r0 */
+	moveli  r0, lo16(\c_routine)
+	{
+	 auli   r0, r0, ha16(\c_routine)
+	 j       \processing
+	}
+	.endif
+	ENDPROC(intvec_\vecname)
+
+#ifdef __COLLECT_LINKER_FEEDBACK__
+	.pushsection .text.intvec_feedback,"ax"
+	.org    (\vecnum << 5)
+	FEEDBACK_ENTER_EXPLICIT(intvec_\vecname, .intrpt, 1 << 8)
+	jrp     lr
+	.popsection
+#endif
+
+	.endm
+
+
+	/*
+	 * Save the rest of the registers that we didn't save in the actual
+	 * vector itself.  We can't use r0-r10 inclusive here.
+	 */
+	.macro  finish_interrupt_save, function
+
+	/* If it's a syscall, save a proper orig_r0, otherwise just zero. */
+	PTREGS_PTR(r52, PTREGS_OFFSET_ORIG_R0)
+	{
+	 .ifc \function,handle_syscall
+	 sw     r52, r0
+	 .else
+	 sw     r52, zero
+	 .endif
+	 PTREGS_PTR(r52, PTREGS_OFFSET_TP)
+	}
+
+	/*
+	 * For ordinary syscalls, we save neither caller- nor callee-
+	 * save registers, since the syscall invoker doesn't expect the
+	 * caller-saves to be saved, and the called kernel functions will
+	 * take care of saving the callee-saves for us.
+	 *
+	 * For interrupts we save just the caller-save registers.  Saving
+	 * them is required (since the "caller" can't save them).  Again,
+	 * the called kernel functions will restore the callee-save
+	 * registers for us appropriately.
+	 *
+	 * On return, we normally restore nothing special for syscalls,
+	 * and just the caller-save registers for interrupts.
+	 *
+	 * However, there are some important caveats to all this:
+	 *
+	 * - We always save a few callee-save registers to give us
+	 *   some scratchpad registers to carry across function calls.
+	 *
+	 * - fork/vfork/etc require us to save all the callee-save
+	 *   registers, which we do in PTREGS_SYSCALL_ALL_REGS, below.
+	 *
+	 * - We always save r0..r5 and r10 for syscalls, since we need
+	 *   to reload them a bit later for the actual kernel call, and
+	 *   since we might need them for -ERESTARTNOINTR, etc.
+	 *
+	 * - Before invoking a signal handler, we save the unsaved
+	 *   callee-save registers so they are visible to the
+	 *   signal handler or any ptracer.
+	 *
+	 * - If the unsaved callee-save registers are modified, we set
+	 *   a bit in pt_regs so we know to reload them from pt_regs
+	 *   and not just rely on the kernel function unwinding.
+	 *   (Done for ptrace register writes and SA_SIGINFO handler.)
+	 */
+	{
+	 sw     r52, tp
+	 PTREGS_PTR(r52, PTREGS_OFFSET_REG(33))
+	}
+	wh64    r52    /* cache line 2 */
+	push_reg r33, r52
+	push_reg r32, r52
+	push_reg r31, r52
+	.ifc \function,handle_syscall
+	push_reg r30, r52, PTREGS_OFFSET_SYSCALL - PTREGS_OFFSET_REG(30)
+	push_reg TREG_SYSCALL_NR_NAME, r52, \
+	  PTREGS_OFFSET_REG(5) - PTREGS_OFFSET_SYSCALL
+	.else
+
+	push_reg r30, r52, PTREGS_OFFSET_REG(29) - PTREGS_OFFSET_REG(30)
+	wh64    r52    /* cache line 1 */
+	push_reg r29, r52
+	push_reg r28, r52
+	push_reg r27, r52
+	push_reg r26, r52
+	push_reg r25, r52
+	push_reg r24, r52
+	push_reg r23, r52
+	push_reg r22, r52
+	push_reg r21, r52
+	push_reg r20, r52
+	push_reg r19, r52
+	push_reg r18, r52
+	push_reg r17, r52
+	push_reg r16, r52
+	push_reg r15, r52
+	push_reg r14, r52
+	push_reg r13, r52
+	push_reg r12, r52
+	push_reg r11, r52
+	push_reg r10, r52
+	push_reg r9, r52
+	push_reg r8, r52
+	push_reg r7, r52
+	push_reg r6, r52
+
+	.endif
+
+	push_reg r5, r52
+	sw      r52, r4
+
+	/* Load tp with our per-cpu offset. */
+#ifdef CONFIG_SMP
+	{
+	 mfspr  r20, SPR_SYSTEM_SAVE_K_0
+	 moveli r21, lo16(__per_cpu_offset)
+	}
+	{
+	 auli   r21, r21, ha16(__per_cpu_offset)
+	 mm     r20, r20, zero, 0, LOG2_NR_CPU_IDS-1
+	}
+	s2a     r20, r20, r21
+	lw      tp, r20
+#else
+	move    tp, zero
+#endif
+
+	/*
+	 * If we will be returning to the kernel, we will need to
+	 * reset the interrupt masks to the state they had before.
+	 * Set DISABLE_IRQ in flags iff we came from PL1 with irqs disabled.
+	 * We load flags in r32 here so we can jump to .Lrestore_regs
+	 * directly after do_page_fault_ics() if necessary.
+	 */
+	mfspr   r32, SPR_EX_CONTEXT_K_1
+	{
+	 andi   r32, r32, SPR_EX_CONTEXT_1_1__PL_MASK  /* mask off ICS */
+	 PTREGS_PTR(r21, PTREGS_OFFSET_FLAGS)
+	}
+	bzt     r32, 1f       /* zero if from user space */
+	IRQS_DISABLED(r32)    /* zero if irqs enabled */
+#if PT_FLAGS_DISABLE_IRQ != 1
+# error Value of IRQS_DISABLED used to set PT_FLAGS_DISABLE_IRQ; fix
+#endif
+1:
+	.ifnc \function,handle_syscall
+	/* Record the fact that we saved the caller-save registers above. */
+	ori     r32, r32, PT_FLAGS_CALLER_SAVES
+	.endif
+	sw      r21, r32
+
+#ifdef __COLLECT_LINKER_FEEDBACK__
+	/*
+	 * Notify the feedback routines that we were in the
+	 * appropriate fixed interrupt vector area.  Note that we
+	 * still have ICS set at this point, so we can't invoke any
+	 * atomic operations or we will panic.  The feedback
+	 * routines internally preserve r0..r10 and r30 up.
+	 */
+	.ifnc \function,handle_syscall
+	shli    r20, r1, 5
+	.else
+	moveli  r20, INT_SWINT_1 << 5
+	.endif
+	addli   r20, r20, lo16(intvec_feedback)
+	auli    r20, r20, ha16(intvec_feedback)
+	jalr    r20
+
+	/* And now notify the feedback routines that we are here. */
+	FEEDBACK_ENTER(\function)
+#endif
+
+	/*
+	 * we've captured enough state to the stack (including in
+	 * particular our EX_CONTEXT state) that we can now release
+	 * the interrupt critical section and replace it with our
+	 * standard "interrupts disabled" mask value.  This allows
+	 * synchronous interrupts (and profile interrupts) to punch
+	 * through from this point onwards.
+	 *
+	 * If bit 31 of r3 is set during a non-NMI interrupt, we know we
+	 * are on the path where the hypervisor has punched through our
+	 * ICS with a page fault, so we call out to do_page_fault_ics()
+	 * to figure out what to do with it.  If the fault was in
+	 * an atomic op, we unlock the atomic lock, adjust the
+	 * saved register state a little, and return "zero" in r4,
+	 * falling through into the normal page-fault interrupt code.
+	 * If the fault was in a kernel-space atomic operation, then
+	 * do_page_fault_ics() resolves it itself, returns "one" in r4,
+	 * and as a result goes directly to restoring registers and iret,
+	 * without trying to adjust the interrupt masks at all.
+	 * The do_page_fault_ics() API involves passing and returning
+	 * a five-word struct (in registers) to avoid writing the
+	 * save and restore code here.
+	 */
+	.ifc \function,handle_nmi
+	IRQ_DISABLE_ALL(r20)
+	.else
+	.ifnc \function,handle_syscall
+	bgezt   r3, 1f
+	{
+	 PTREGS_PTR(r0, PTREGS_OFFSET_BASE)
+	 jal    do_page_fault_ics
+	}
+	FEEDBACK_REENTER(\function)
+	bzt     r4, 1f
+	j       .Lrestore_regs
+1:
+	.endif
+	IRQ_DISABLE(r20, r21)
+	.endif
+	mtspr   INTERRUPT_CRITICAL_SECTION, zero
+
+	/*
+	 * Prepare the first 256 stack bytes to be rapidly accessible
+	 * without having to fetch the background data.  We don't really
+	 * know how far to write-hint, but kernel stacks generally
+	 * aren't that big, and write-hinting here does take some time.
+	 */
+	addi    r52, sp, -64
+	{
+	 wh64   r52
+	 addi   r52, r52, -64
+	}
+	{
+	 wh64   r52
+	 addi   r52, r52, -64
+	}
+	{
+	 wh64   r52
+	 addi   r52, r52, -64
+	}
+	wh64    r52
+
+#ifdef CONFIG_TRACE_IRQFLAGS
+	.ifnc \function,handle_nmi
+	/*
+	 * We finally have enough state set up to notify the irq
+	 * tracing code that irqs were disabled on entry to the handler.
+	 * The TRACE_IRQS_OFF call clobbers registers r0-r29.
+	 * For syscalls, we already have the register state saved away
+	 * on the stack, so we don't bother to do any register saves here,
+	 * and later we pop the registers back off the kernel stack.
+	 * For interrupt handlers, save r0-r3 in callee-saved registers.
+	 */
+	.ifnc \function,handle_syscall
+	{ move r30, r0; move r31, r1 }
+	{ move r32, r2; move r33, r3 }
+	.endif
+	TRACE_IRQS_OFF
+	.ifnc \function,handle_syscall
+	{ move r0, r30; move r1, r31 }
+	{ move r2, r32; move r3, r33 }
+	.endif
+	.endif
+#endif
+
+	.endm
+
+	.macro  check_single_stepping, kind, not_single_stepping
+	/*
+	 * Check for single stepping in user-level priv
+	 *   kind can be "normal", "ill", or "syscall"
+	 * At end, if fall-thru
+	 *   r29: thread_info->step_state
+	 *   r28: &pt_regs->pc
+	 *   r27: pt_regs->pc
+	 *   r26: thread_info->step_state->buffer
+	 */
+
+	/* Check for single stepping */
+	GET_THREAD_INFO(r29)
+	{
+	 /* Get pointer to field holding step state */
+	 addi   r29, r29, THREAD_INFO_STEP_STATE_OFFSET
+
+	 /* Get pointer to EX1 in register state */
+	 PTREGS_PTR(r27, PTREGS_OFFSET_EX1)
+	}
+	{
+	 /* Get pointer to field holding PC */
+	 PTREGS_PTR(r28, PTREGS_OFFSET_PC)
+
+	 /* Load the pointer to the step state */
+	 lw     r29, r29
+	}
+	/* Load EX1 */
+	lw      r27, r27
+	{
+	 /* Points to flags */
+	 addi   r23, r29, SINGLESTEP_STATE_FLAGS_OFFSET
+
+	 /* No single stepping if there is no step state structure */
+	 bzt    r29, \not_single_stepping
+	}
+	{
+	 /* mask off ICS and any other high bits */
+	 andi   r27, r27, SPR_EX_CONTEXT_1_1__PL_MASK
+
+	 /* Load pointer to single step instruction buffer */
+	 lw     r26, r29
+	}
+	/* Check priv state */
+	bnz     r27, \not_single_stepping
+
+	/* Get flags */
+	lw      r22, r23
+	{
+	 /* Branch if single-step mode not enabled */
+	 bbnst  r22, \not_single_stepping
+
+	 /* Clear enabled flag */
+	 andi   r22, r22, ~SINGLESTEP_STATE_MASK_IS_ENABLED
+	}
+	.ifc \kind,normal
+	{
+	 /* Load PC */
+	 lw     r27, r28
+
+	 /* Point to the entry containing the original PC */
+	 addi   r24, r29, SINGLESTEP_STATE_ORIG_PC_OFFSET
+	}
+	{
+	 /* Disable single stepping flag */
+	 sw     r23, r22
+	}
+	{
+	 /* Get the original pc */
+	 lw     r24, r24
+
+	 /* See if the PC is at the start of the single step buffer */
+	 seq    r25, r26, r27
+	}
+	/*
+	 * NOTE: it is really expected that the PC be in the single step buffer
+	 *       at this point
+	 */
+	bzt     r25, \not_single_stepping
+
+	/* Restore the original PC */
+	sw      r28, r24
+	.else
+	.ifc \kind,syscall
+	{
+	 /* Load PC */
+	 lw     r27, r28
+
+	 /* Point to the entry containing the next PC */
+	 addi   r24, r29, SINGLESTEP_STATE_NEXT_PC_OFFSET
+	}
+	{
+	 /* Increment the stopped PC by the bundle size */
+	 addi   r26, r26, 8
+
+	 /* Disable single stepping flag */
+	 sw     r23, r22
+	}
+	{
+	 /* Get the next pc */
+	 lw     r24, r24
+
+	 /*
+	  * See if the PC is one bundle past the start of the
+	  * single step buffer
+	  */
+	 seq    r25, r26, r27
+	}
+	{
+	 /*
+	  * NOTE: it is really expected that the PC be in the
+	  * single step buffer at this point
+	  */
+	 bzt    r25, \not_single_stepping
+	}
+	/* Set to the next PC */
+	sw      r28, r24
+	.else
+	{
+	 /* Point to 3rd bundle in buffer */
+	 addi   r25, r26, 16
+
+	 /* Load PC */
+	 lw      r27, r28
+	}
+	{
+	 /* Disable single stepping flag */
+	 sw      r23, r22
+
+	 /* See if the PC is in the single step buffer */
+	 slte_u  r24, r26, r27
+	}
+	{
+	 slte_u r25, r27, r25
+
+	 /*
+	  * NOTE: it is really expected that the PC be in the
+	  * single step buffer at this point
+	  */
+	 bzt    r24, \not_single_stepping
+	}
+	bzt     r25, \not_single_stepping
+	.endif
+	.endif
+	.endm
+
+	/*
+	 * Redispatch a downcall.
+	 */
+	.macro  dc_dispatch vecnum, vecname
+	.org    (\vecnum << 8)
+intvec_\vecname:
+	j       _hv_downcall_dispatch
+	ENDPROC(intvec_\vecname)
+	.endm
+
+	/*
+	 * Common code for most interrupts.  The C function we're eventually
+	 * going to is in r0, and the faultnum is in r1; the original
+	 * values for those registers are on the stack.
+	 */
+	.pushsection .text.handle_interrupt,"ax"
+handle_interrupt:
+	finish_interrupt_save handle_interrupt
+
+	/*
+	 * Check for if we are single stepping in user level. If so, then
+	 * we need to restore the PC.
+	 */
+
+	check_single_stepping normal, .Ldispatch_interrupt
+.Ldispatch_interrupt:
+
+	/* Jump to the C routine; it should enable irqs as soon as possible. */
+	{
+	 jalr   r0
+	 PTREGS_PTR(r0, PTREGS_OFFSET_BASE)
+	}
+	FEEDBACK_REENTER(handle_interrupt)
+	{
+	 movei  r30, 0   /* not an NMI */
+	 j      interrupt_return
+	}
+	STD_ENDPROC(handle_interrupt)
+
+/*
+ * This routine takes a boolean in r30 indicating if this is an NMI.
+ * If so, we also expect a boolean in r31 indicating whether to
+ * re-enable the oprofile interrupts.
+ *
+ * Note that .Lresume_userspace is jumped to directly in several
+ * places, and we need to make sure r30 is set correctly in those
+ * callers as well.
+ */
+STD_ENTRY(interrupt_return)
+	/* If we're resuming to kernel space, don't check thread flags. */
+	{
+	 bnz    r30, .Lrestore_all  /* NMIs don't special-case user-space */
+	 PTREGS_PTR(r29, PTREGS_OFFSET_EX1)
+	}
+	lw      r29, r29
+	andi    r29, r29, SPR_EX_CONTEXT_1_1__PL_MASK  /* mask off ICS */
+	bzt     r29, .Lresume_userspace
+
+#ifdef CONFIG_PREEMPT
+	/* Returning to kernel space. Check if we need preemption. */
+	GET_THREAD_INFO(r29)
+	addli   r28, r29, THREAD_INFO_FLAGS_OFFSET
+	{
+	 lw     r28, r28
+	 addli  r29, r29, THREAD_INFO_PREEMPT_COUNT_OFFSET
+	}
+	{
+	 andi   r28, r28, _TIF_NEED_RESCHED
+	 lw     r29, r29
+	}
+	bzt     r28, 1f
+	bnz     r29, 1f
+	/* Disable interrupts explicitly for preemption. */
+	IRQ_DISABLE(r20,r21)
+	TRACE_IRQS_OFF
+	jal     preempt_schedule_irq
+	FEEDBACK_REENTER(interrupt_return)
+1:
+#endif
+
+	/* If we're resuming to _cpu_idle_nap, bump PC forward by 8. */
+	{
+	 PTREGS_PTR(r29, PTREGS_OFFSET_PC)
+	 moveli r27, lo16(_cpu_idle_nap)
+	}
+	{
+	 lw     r28, r29
+	 auli   r27, r27, ha16(_cpu_idle_nap)
+	}
+	{
+	 seq    r27, r27, r28
+	}
+	{
+	 bbns   r27, .Lrestore_all
+	 addi   r28, r28, 8
+	}
+	sw      r29, r28
+	j       .Lrestore_all
+
+.Lresume_userspace:
+	FEEDBACK_REENTER(interrupt_return)
+
+	/*
+	 * Use r33 to hold whether we have already loaded the callee-saves
+	 * into ptregs.  We don't want to do it twice in this loop, since
+	 * then we'd clobber whatever changes are made by ptrace, etc.
+	 * Get base of stack in r32.
+	 */
+	{
+	 GET_THREAD_INFO(r32)
+	 movei  r33, 0
+	}
+
+.Lretry_work_pending:
+	/*
+	 * Disable interrupts so as to make sure we don't
+	 * miss an interrupt that sets any of the thread flags (like
+	 * need_resched or sigpending) between sampling and the iret.
+	 * Routines like schedule() or do_signal() may re-enable
+	 * interrupts before returning.
+	 */
+	IRQ_DISABLE(r20, r21)
+	TRACE_IRQS_OFF  /* Note: clobbers registers r0-r29 */
+
+
+	/* Check to see if there is any work to do before returning to user. */
+	{
+	 addi   r29, r32, THREAD_INFO_FLAGS_OFFSET
+	 moveli r1, lo16(_TIF_ALLWORK_MASK)
+	}
+	{
+	 lw     r29, r29
+	 auli   r1, r1, ha16(_TIF_ALLWORK_MASK)
+	}
+	and     r1, r29, r1
+	bzt     r1, .Lrestore_all
+
+	/*
+	 * Make sure we have all the registers saved for signal
+	 * handling, notify-resume, or single-step.  Call out to C
+	 * code to figure out exactly what we need to do for each flag bit,
+	 * then if necessary, reload the flags and recheck.
+	 */
+	{
+	 PTREGS_PTR(r0, PTREGS_OFFSET_BASE)
+	 bnz    r33, 1f
+	}
+	push_extra_callee_saves r0
+	movei   r33, 1
+1:	jal     do_work_pending
+	bnz     r0, .Lretry_work_pending
+
+	/*
+	 * In the NMI case we
+	 * omit the call to single_process_check_nohz, which normally checks
+	 * to see if we should start or stop the scheduler tick, because
+	 * we can't call arbitrary Linux code from an NMI context.
+	 * We always call the homecache TLB deferral code to re-trigger
+	 * the deferral mechanism.
+	 *
+	 * The other chunk of responsibility this code has is to reset the
+	 * interrupt masks appropriately to reset irqs and NMIs.  We have
+	 * to call TRACE_IRQS_OFF and TRACE_IRQS_ON to support all the
+	 * lockdep-type stuff, but we can't set ICS until afterwards, since
+	 * ICS can only be used in very tight chunks of code to avoid
+	 * tripping over various assertions that it is off.
+	 *
+	 * (There is what looks like a window of vulnerability here since
+	 * we might take a profile interrupt between the two SPR writes
+	 * that set the mask, but since we write the low SPR word first,
+	 * and our interrupt entry code checks the low SPR word, any
+	 * profile interrupt will actually disable interrupts in both SPRs
+	 * before returning, which is OK.)
+	 */
+.Lrestore_all:
+	PTREGS_PTR(r0, PTREGS_OFFSET_EX1)
+	{
+	 lw     r0, r0
+	 PTREGS_PTR(r32, PTREGS_OFFSET_FLAGS)
+	}
+	{
+	 andi   r0, r0, SPR_EX_CONTEXT_1_1__PL_MASK
+	 lw     r32, r32
+	}
+	bnz    r0, 1f
+	j       2f
+#if PT_FLAGS_DISABLE_IRQ != 1
+# error Assuming PT_FLAGS_DISABLE_IRQ == 1 so we can use bbnst below
+#endif
+1:	bbnst   r32, 2f
+	IRQ_DISABLE(r20,r21)
+	TRACE_IRQS_OFF
+	movei   r0, 1
+	mtspr   INTERRUPT_CRITICAL_SECTION, r0
+	bzt     r30, .Lrestore_regs
+	j       3f
+2:	TRACE_IRQS_ON
+	movei   r0, 1
+	mtspr   INTERRUPT_CRITICAL_SECTION, r0
+	IRQ_ENABLE(r20, r21)
+	bzt     r30, .Lrestore_regs
+3:
+
+	/* We are relying on INT_PERF_COUNT at 33, and AUX_PERF_COUNT at 48 */
+	{
+	 moveli r0, lo16(1 << (INT_PERF_COUNT - 32))
+	 bz     r31, .Lrestore_regs
+	}
+	auli    r0, r0, ha16(1 << (INT_AUX_PERF_COUNT - 32))
+	mtspr   SPR_INTERRUPT_MASK_RESET_K_1, r0
+
+	/*
+	 * We now commit to returning from this interrupt, since we will be
+	 * doing things like setting EX_CONTEXT SPRs and unwinding the stack
+	 * frame.  No calls should be made to any other code after this point.
+	 * This code should only be entered with ICS set.
+	 * r32 must still be set to ptregs.flags.
+	 * We launch loads to each cache line separately first, so we can
+	 * get some parallelism out of the memory subsystem.
+	 * We start zeroing caller-saved registers throughout, since
+	 * that will save some cycles if this turns out to be a syscall.
+	 */
+.Lrestore_regs:
+	FEEDBACK_REENTER(interrupt_return)   /* called from elsewhere */
+
+	/*
+	 * Rotate so we have one high bit and one low bit to test.
+	 * - low bit says whether to restore all the callee-saved registers,
+	 *   or just r30-r33, and r52 up.
+	 * - high bit (i.e. sign bit) says whether to restore all the
+	 *   caller-saved registers, or just r0.
+	 */
+#if PT_FLAGS_CALLER_SAVES != 2 || PT_FLAGS_RESTORE_REGS != 4
+# error Rotate trick does not work :-)
+#endif
+	{
+	 rli    r20, r32, 30
+	 PTREGS_PTR(sp, PTREGS_OFFSET_REG(0))
+	}
+
+	/*
+	 * Load cache lines 0, 2, and 3 in that order, then use
+	 * the last loaded value, which makes it likely that the other
+	 * cache lines have also loaded, at which point we should be
+	 * able to safely read all the remaining words on those cache
+	 * lines without waiting for the memory subsystem.
+	 */
+	pop_reg_zero r0, r28, sp, PTREGS_OFFSET_REG(30) - PTREGS_OFFSET_REG(0)
+	pop_reg_zero r30, r2, sp, PTREGS_OFFSET_PC - PTREGS_OFFSET_REG(30)
+	pop_reg_zero r21, r3, sp, PTREGS_OFFSET_EX1 - PTREGS_OFFSET_PC
+	pop_reg_zero lr, r4, sp, PTREGS_OFFSET_REG(52) - PTREGS_OFFSET_EX1
+	{
+	 mtspr  SPR_EX_CONTEXT_K_0, r21
+	 move   r5, zero
+	}
+	{
+	 mtspr  SPR_EX_CONTEXT_K_1, lr
+	 andi   lr, lr, SPR_EX_CONTEXT_1_1__PL_MASK  /* mask off ICS */
+	}
+
+	/* Restore callee-saveds that we actually use. */
+	pop_reg_zero r52, r6, sp, PTREGS_OFFSET_REG(31) - PTREGS_OFFSET_REG(52)
+	pop_reg_zero r31, r7
+	pop_reg_zero r32, r8
+	pop_reg_zero r33, r9, sp, PTREGS_OFFSET_REG(29) - PTREGS_OFFSET_REG(33)
+
+	/*
+	 * If we modified other callee-saveds, restore them now.
+	 * This is rare, but could be via ptrace or signal handler.
+	 */
+	{
+	 move   r10, zero
+	 bbs    r20, .Lrestore_callees
+	}
+.Lcontinue_restore_regs:
+
+	/* Check if we're returning from a syscall. */
+	{
+	 move   r11, zero
+	 blzt   r20, 1f  /* no, so go restore callee-save registers */
+	}
+
+	/*
+	 * Check if we're returning to userspace.
+	 * Note that if we're not, we don't worry about zeroing everything.
+	 */
+	{
+	 addli  sp, sp, PTREGS_OFFSET_LR - PTREGS_OFFSET_REG(29)
+	 bnz    lr, .Lkernel_return
+	}
+
+	/*
+	 * On return from syscall, we've restored r0 from pt_regs, but we
+	 * clear the remainder of the caller-saved registers.  We could
+	 * restore the syscall arguments, but there's not much point,
+	 * and it ensures user programs aren't trying to use the
+	 * caller-saves if we clear them, as well as avoiding leaking
+	 * kernel pointers into userspace.
+	 */
+	pop_reg_zero lr, r12, sp, PTREGS_OFFSET_TP - PTREGS_OFFSET_LR
+	pop_reg_zero tp, r13, sp, PTREGS_OFFSET_SP - PTREGS_OFFSET_TP
+	{
+	 lw     sp, sp
+	 move   r14, zero
+	 move   r15, zero
+	}
+	{ move r16, zero; move r17, zero }
+	{ move r18, zero; move r19, zero }
+	{ move r20, zero; move r21, zero }
+	{ move r22, zero; move r23, zero }
+	{ move r24, zero; move r25, zero }
+	{ move r26, zero; move r27, zero }
+
+	/* Set r1 to errno if we are returning an error, otherwise zero. */
+	{
+	 moveli r29, 4096
+	 sub    r1, zero, r0
+	}
+	slt_u   r29, r1, r29
+	{
+	 mnz    r1, r29, r1
+	 move   r29, zero
+	}
+	iret
+
+	/*
+	 * Not a syscall, so restore caller-saved registers.
+	 * First kick off a load for cache line 1, which we're touching
+	 * for the first time here.
+	 */
+	.align 64
+1:	pop_reg r29, sp, PTREGS_OFFSET_REG(1) - PTREGS_OFFSET_REG(29)
+	pop_reg r1
+	pop_reg r2
+	pop_reg r3
+	pop_reg r4
+	pop_reg r5
+	pop_reg r6
+	pop_reg r7
+	pop_reg r8
+	pop_reg r9
+	pop_reg r10
+	pop_reg r11
+	pop_reg r12
+	pop_reg r13
+	pop_reg r14
+	pop_reg r15
+	pop_reg r16
+	pop_reg r17
+	pop_reg r18
+	pop_reg r19
+	pop_reg r20
+	pop_reg r21
+	pop_reg r22
+	pop_reg r23
+	pop_reg r24
+	pop_reg r25
+	pop_reg r26
+	pop_reg r27
+	pop_reg r28, sp, PTREGS_OFFSET_LR - PTREGS_OFFSET_REG(28)
+	/* r29 already restored above */
+	bnz     lr, .Lkernel_return
+	pop_reg lr, sp, PTREGS_OFFSET_TP - PTREGS_OFFSET_LR
+	pop_reg tp, sp, PTREGS_OFFSET_SP - PTREGS_OFFSET_TP
+	lw      sp, sp
+	iret
+
+	/*
+	 * We can't restore tp when in kernel mode, since a thread might
+	 * have migrated from another cpu and brought a stale tp value.
+	 */
+.Lkernel_return:
+	pop_reg lr, sp, PTREGS_OFFSET_SP - PTREGS_OFFSET_LR
+	lw      sp, sp
+	iret
+
+	/* Restore callee-saved registers from r34 to r51. */
+.Lrestore_callees:
+	addli  sp, sp, PTREGS_OFFSET_REG(34) - PTREGS_OFFSET_REG(29)
+	pop_reg r34
+	pop_reg r35
+	pop_reg r36
+	pop_reg r37
+	pop_reg r38
+	pop_reg r39
+	pop_reg r40
+	pop_reg r41
+	pop_reg r42
+	pop_reg r43
+	pop_reg r44
+	pop_reg r45
+	pop_reg r46
+	pop_reg r47
+	pop_reg r48
+	pop_reg r49
+	pop_reg r50
+	pop_reg r51, sp, PTREGS_OFFSET_REG(29) - PTREGS_OFFSET_REG(51)
+	j .Lcontinue_restore_regs
+	STD_ENDPROC(interrupt_return)
+
+	/*
+	 * Some interrupts don't check for single stepping
+	 */
+	.pushsection .text.handle_interrupt_no_single_step,"ax"
+handle_interrupt_no_single_step:
+	finish_interrupt_save handle_interrupt_no_single_step
+	{
+	 jalr   r0
+	 PTREGS_PTR(r0, PTREGS_OFFSET_BASE)
+	}
+	FEEDBACK_REENTER(handle_interrupt_no_single_step)
+	{
+	 movei  r30, 0   /* not an NMI */
+	 j      interrupt_return
+	}
+	STD_ENDPROC(handle_interrupt_no_single_step)
+
+	/*
+	 * "NMI" interrupts mask ALL interrupts before calling the
+	 * handler, and don't check thread flags, etc., on the way
+	 * back out.  In general, the only things we do here for NMIs
+	 * are the register save/restore, fixing the PC if we were
+	 * doing single step, and the dataplane kernel-TLB management.
+	 * We don't (for example) deal with start/stop of the sched tick.
+	 */
+	.pushsection .text.handle_nmi,"ax"
+handle_nmi:
+	finish_interrupt_save handle_nmi
+	check_single_stepping normal, .Ldispatch_nmi
+.Ldispatch_nmi:
+	{
+	 jalr   r0
+	 PTREGS_PTR(r0, PTREGS_OFFSET_BASE)
+	}
+	FEEDBACK_REENTER(handle_nmi)
+	{
+	 movei  r30, 1
+	 seq    r31, r0, zero
+	}
+	j       interrupt_return
+	STD_ENDPROC(handle_nmi)
+
+	/*
+	 * Parallel code for syscalls to handle_interrupt.
+	 */
+	.pushsection .text.handle_syscall,"ax"
+handle_syscall:
+	finish_interrupt_save handle_syscall
+
+	/*
+	 * Check for if we are single stepping in user level. If so, then
+	 * we need to restore the PC.
+	 */
+	check_single_stepping syscall, .Ldispatch_syscall
+.Ldispatch_syscall:
+
+	/* Enable irqs. */
+	TRACE_IRQS_ON
+	IRQ_ENABLE(r20, r21)
+
+	/* Bump the counter for syscalls made on this tile. */
+	moveli  r20, lo16(irq_stat + IRQ_CPUSTAT_SYSCALL_COUNT_OFFSET)
+	auli    r20, r20, ha16(irq_stat + IRQ_CPUSTAT_SYSCALL_COUNT_OFFSET)
+	add     r20, r20, tp
+	lw      r21, r20
+	addi    r21, r21, 1
+	{
+	 sw     r20, r21
+	 GET_THREAD_INFO(r31)
+	}
+
+	/* Trace syscalls, if requested. */
+	addi	r31, r31, THREAD_INFO_FLAGS_OFFSET
+	lw	r30, r31
+	andi    r30, r30, _TIF_SYSCALL_TRACE
+	bzt	r30, .Lrestore_syscall_regs
+	{
+	 PTREGS_PTR(r0, PTREGS_OFFSET_BASE)
+	 jal    do_syscall_trace_enter
+	}
+	FEEDBACK_REENTER(handle_syscall)
+
+	/*
+	 * We always reload our registers from the stack at this
+	 * point.  They might be valid, if we didn't build with
+	 * TRACE_IRQFLAGS, and this isn't a dataplane tile, and we're not
+	 * doing syscall tracing, but there are enough cases now that it
+	 * seems simplest just to do the reload unconditionally.
+	 */
+.Lrestore_syscall_regs:
+	PTREGS_PTR(r11, PTREGS_OFFSET_REG(0))
+	pop_reg r0, r11
+	pop_reg r1, r11
+	pop_reg r2, r11
+	pop_reg r3, r11
+	pop_reg r4, r11
+	pop_reg r5, r11, PTREGS_OFFSET_SYSCALL - PTREGS_OFFSET_REG(5)
+	pop_reg TREG_SYSCALL_NR_NAME, r11
+
+	/* Ensure that the syscall number is within the legal range. */
+	moveli  r21, __NR_syscalls
+	{
+	 slt_u  r21, TREG_SYSCALL_NR_NAME, r21
+	 moveli r20, lo16(sys_call_table)
+	}
+	{
+	 bbns   r21, .Linvalid_syscall
+	 auli   r20, r20, ha16(sys_call_table)
+	}
+	s2a     r20, TREG_SYSCALL_NR_NAME, r20
+	lw      r20, r20
+
+	/* Jump to syscall handler. */
+	jalr    r20
+.Lhandle_syscall_link: /* value of "lr" after "jalr r20" above */
+
+	/*
+	 * Write our r0 onto the stack so it gets restored instead
+	 * of whatever the user had there before.
+	 */
+	PTREGS_PTR(r29, PTREGS_OFFSET_REG(0))
+	sw      r29, r0
+
+.Lsyscall_sigreturn_skip:
+	FEEDBACK_REENTER(handle_syscall)
+
+	/* Do syscall trace again, if requested. */
+	lw	r30, r31
+	andi    r30, r30, _TIF_SYSCALL_TRACE
+	bzt     r30, 1f
+	{
+	 PTREGS_PTR(r0, PTREGS_OFFSET_BASE)
+	 jal    do_syscall_trace_exit
+	}
+	FEEDBACK_REENTER(handle_syscall)
+1:	{
+	 movei  r30, 0               /* not an NMI */
+	 j      .Lresume_userspace   /* jump into middle of interrupt_return */
+	}
+
+.Linvalid_syscall:
+	/* Report an invalid syscall back to the user program */
+	{
+	 PTREGS_PTR(r29, PTREGS_OFFSET_REG(0))
+	 movei  r28, -ENOSYS
+	}
+	sw      r29, r28
+	{
+	 movei  r30, 0               /* not an NMI */
+	 j      .Lresume_userspace   /* jump into middle of interrupt_return */
+	}
+	STD_ENDPROC(handle_syscall)
+
+	/* Return the address for oprofile to suppress in backtraces. */
+STD_ENTRY_SECTION(handle_syscall_link_address, .text.handle_syscall)
+	lnk     r0
+	{
+	 addli  r0, r0, .Lhandle_syscall_link - .
+	 jrp    lr
+	}
+	STD_ENDPROC(handle_syscall_link_address)
+
+STD_ENTRY(ret_from_fork)
+	jal     sim_notify_fork
+	jal     schedule_tail
+	FEEDBACK_REENTER(ret_from_fork)
+	{
+	 movei  r30, 0               /* not an NMI */
+	 j      .Lresume_userspace   /* jump into middle of interrupt_return */
+	}
+	STD_ENDPROC(ret_from_fork)
+
+STD_ENTRY(ret_from_kernel_thread)
+	jal     sim_notify_fork
+	jal     schedule_tail
+	FEEDBACK_REENTER(ret_from_fork)
+	{
+	 move   r0, r31
+	 jalr   r30
+	}
+	FEEDBACK_REENTER(ret_from_kernel_thread)
+	{
+	 movei  r30, 0               /* not an NMI */
+	 j      .Lresume_userspace   /* jump into middle of interrupt_return */
+	}
+	STD_ENDPROC(ret_from_kernel_thread)
+
+	/*
+	 * Code for ill interrupt.
+	 */
+	.pushsection .text.handle_ill,"ax"
+handle_ill:
+	finish_interrupt_save handle_ill
+
+	/*
+	 * Check for if we are single stepping in user level. If so, then
+	 * we need to restore the PC.
+	 */
+	check_single_stepping ill, .Ldispatch_normal_ill
+
+	{
+	 /* See if the PC is the 1st bundle in the buffer */
+	 seq    r25, r27, r26
+
+	 /* Point to the 2nd bundle in the buffer */
+	 addi   r26, r26, 8
+	}
+	{
+	 /* Point to the original pc */
+	 addi   r24, r29, SINGLESTEP_STATE_ORIG_PC_OFFSET
+
+	 /* Branch if the PC is the 1st bundle in the buffer */
+	 bnz    r25, 3f
+	}
+	{
+	 /* See if the PC is the 2nd bundle of the buffer */
+	 seq    r25, r27, r26
+
+	 /* Set PC to next instruction */
+	 addi   r24, r29, SINGLESTEP_STATE_NEXT_PC_OFFSET
+	}
+	{
+	 /* Point to flags */
+	 addi   r25, r29, SINGLESTEP_STATE_FLAGS_OFFSET
+
+	 /* Branch if PC is in the second bundle */
+	 bz     r25, 2f
+	}
+	/* Load flags */
+	lw      r25, r25
+	{
+	 /*
+	  * Get the offset for the register to restore
+	  * Note: the lower bound is 2, so we have implicit scaling by 4.
+	  *  No multiplication of the register number by the size of a register
+	  *  is needed.
+	  */
+	 mm     r27, r25, zero, SINGLESTEP_STATE_TARGET_LB, \
+		SINGLESTEP_STATE_TARGET_UB
+
+	 /* Mask Rewrite_LR */
+	 andi   r25, r25, SINGLESTEP_STATE_MASK_UPDATE
+	}
+	{
+	 addi   r29, r29, SINGLESTEP_STATE_UPDATE_VALUE_OFFSET
+
+	 /* Don't rewrite temp register */
+	 bz     r25, 3f
+	}
+	{
+	 /* Get the temp value */
+	 lw     r29, r29
+
+	 /* Point to where the register is stored */
+	 add    r27, r27, sp
+	}
+
+	/* Add in the C ABI save area size to the register offset */
+	addi    r27, r27, C_ABI_SAVE_AREA_SIZE
+
+	/* Restore the user's register with the temp value */
+	sw      r27, r29
+	j       3f
+
+2:
+	/* Must be in the third bundle */
+	addi    r24, r29, SINGLESTEP_STATE_BRANCH_NEXT_PC_OFFSET
+
+3:
+	/* set PC and continue */
+	lw      r26, r24
+	{
+	 sw     r28, r26
+	 GET_THREAD_INFO(r0)
+	}
+
+	/*
+	 * Clear TIF_SINGLESTEP to prevent recursion if we execute an ill.
+	 * The normal non-arch flow redundantly clears TIF_SINGLESTEP, but we
+	 * need to clear it here and can't really impose on all other arches.
+	 * So what's another write between friends?
+	 */
+
+	addi    r1, r0, THREAD_INFO_FLAGS_OFFSET
+	{
+	 lw     r2, r1
+	 addi   r0, r0, THREAD_INFO_TASK_OFFSET  /* currently a no-op */
+	}
+	andi    r2, r2, ~_TIF_SINGLESTEP
+	sw      r1, r2
+
+	/* Issue a sigtrap */
+	{
+	 lw     r0, r0          /* indirect thru thread_info to get task_info*/
+	 addi   r1, sp, C_ABI_SAVE_AREA_SIZE  /* put ptregs pointer into r1 */
+	}
+
+	jal     send_sigtrap    /* issue a SIGTRAP */
+	FEEDBACK_REENTER(handle_ill)
+	{
+	 movei  r30, 0               /* not an NMI */
+	 j      .Lresume_userspace   /* jump into middle of interrupt_return */
+	}
+
+.Ldispatch_normal_ill:
+	{
+	 jalr   r0
+	 PTREGS_PTR(r0, PTREGS_OFFSET_BASE)
+	}
+	FEEDBACK_REENTER(handle_ill)
+	{
+	 movei  r30, 0   /* not an NMI */
+	 j      interrupt_return
+	}
+	STD_ENDPROC(handle_ill)
+
+/* Various stub interrupt handlers and syscall handlers */
+
+STD_ENTRY_LOCAL(_kernel_double_fault)
+	mfspr   r1, SPR_EX_CONTEXT_K_0
+	move    r2, lr
+	move    r3, sp
+	move    r4, r52
+	addi    sp, sp, -C_ABI_SAVE_AREA_SIZE
+	j       kernel_double_fault
+	STD_ENDPROC(_kernel_double_fault)
+
+STD_ENTRY_LOCAL(bad_intr)
+	mfspr   r2, SPR_EX_CONTEXT_K_0
+	panic   "Unhandled interrupt %#x: PC %#lx"
+	STD_ENDPROC(bad_intr)
+
+/*
+ * Special-case sigreturn to not write r0 to the stack on return.
+ * This is technically more efficient, but it also avoids difficulties
+ * in the 64-bit OS when handling 32-bit compat code, since we must not
+ * sign-extend r0 for the sigreturn return-value case.
+ */
+#define PTREGS_SYSCALL_SIGRETURN(x, reg)                \
+	STD_ENTRY(_##x);                                \
+	addli   lr, lr, .Lsyscall_sigreturn_skip - .Lhandle_syscall_link; \
+	{                                               \
+	 PTREGS_PTR(reg, PTREGS_OFFSET_BASE);           \
+	 j      x                                       \
+	};                                              \
+	STD_ENDPROC(_##x)
+
+PTREGS_SYSCALL_SIGRETURN(sys_rt_sigreturn, r0)
+
+/* Save additional callee-saves to pt_regs and jump to standard function. */
+STD_ENTRY(_sys_clone)
+	push_extra_callee_saves r4
+	j       sys_clone
+	STD_ENDPROC(_sys_clone)
+
+/*
+ * This entrypoint is taken for the cmpxchg and atomic_update fast
+ * swints.  We may wish to generalize it to other fast swints at some
+ * point, but for now there are just two very similar ones, which
+ * makes it faster.
+ *
+ * The fast swint code is designed to have a small footprint.  It does
+ * not save or restore any GPRs, counting on the caller-save registers
+ * to be available to it on entry.  It does not modify any callee-save
+ * registers (including "lr").  It does not check what PL it is being
+ * called at, so you'd better not call it other than at PL0.
+ * The <atomic.h> wrapper assumes it only clobbers r20-r29, so if
+ * it ever is necessary to use more registers, be aware.
+ *
+ * It does not use the stack, but since it might be re-interrupted by
+ * a page fault which would assume the stack was valid, it does
+ * save/restore the stack pointer and zero it out to make sure it gets reset.
+ * Since we always keep interrupts disabled, the hypervisor won't
+ * clobber our EX_CONTEXT_K_x registers, so we don't save/restore them
+ * (other than to advance the PC on return).
+ *
+ * We have to manually validate the user vs kernel address range
+ * (since at PL1 we can read/write both), and for performance reasons
+ * we don't allow cmpxchg on the fc000000 memory region, since we only
+ * validate that the user address is below PAGE_OFFSET.
+ *
+ * We place it in the __HEAD section to ensure it is relatively
+ * near to the intvec_SWINT_1 code (reachable by a conditional branch).
+ *
+ * Our use of ATOMIC_LOCK_REG here must match do_page_fault_ics().
+ *
+ * As we do in lib/atomic_asm_32.S, we bypass a store if the value we
+ * would store is the same as the value we just loaded.
+ */
+	__HEAD
+	.align 64
+	/* Align much later jump on the start of a cache line. */
+	nop
+#if PAGE_SIZE >= 0x10000
+	nop
+#endif
+ENTRY(sys_cmpxchg)
+
+	/*
+	 * Save "sp" and set it zero for any possible page fault.
+	 *
+	 * HACK: We want to both zero sp and check r0's alignment,
+	 * so we do both at once. If "sp" becomes nonzero we
+	 * know r0 is unaligned and branch to the error handler that
+	 * restores sp, so this is OK.
+	 *
+	 * ICS is disabled right now so having a garbage but nonzero
+	 * sp is OK, since we won't execute any faulting instructions
+	 * when it is nonzero.
+	 */
+	{
+	 move   r27, sp
+	 andi	sp, r0, 3
+	}
+
+	/*
+	 * Get the lock address in ATOMIC_LOCK_REG, and also validate that the
+	 * address is less than PAGE_OFFSET, since that won't trap at PL1.
+	 * We only use bits less than PAGE_SHIFT to avoid having to worry
+	 * about aliasing among multiple mappings of the same physical page,
+	 * and we ignore the low 3 bits so we have one lock that covers
+	 * both a cmpxchg64() and a cmpxchg() on either its low or high word.
+	 * NOTE: this must match __atomic_hashed_lock() in lib/atomic_32.c.
+	 */
+
+#if (PAGE_OFFSET & 0xffff) != 0
+# error Code here assumes PAGE_OFFSET can be loaded with just hi16()
+#endif
+
+	{
+	 /* Check for unaligned input. */
+	 bnz    sp, .Lcmpxchg_badaddr
+	 auli   r23, zero, hi16(PAGE_OFFSET)  /* hugepage-aligned */
+	}
+	{
+	 /*
+	  * Slide bits into position for 'mm'. We want to ignore
+	  * the low 3 bits of r0, and consider only the next
+	  * ATOMIC_HASH_SHIFT bits.
+	  * Because of C pointer arithmetic, we want to compute this:
+	  *
+	  * ((char*)atomic_locks +
+	  *  (((r0 >> 3) & ((1 << ATOMIC_HASH_SHIFT) - 1)) << 2))
+	  *
+	  * Instead of two shifts we just ">> 1", and use 'mm'
+	  * to ignore the low and high bits we don't want.
+	  */
+	 shri	r25, r0, 1
+
+	 slt_u  r23, r0, r23
+
+	 /*
+	  * Ensure that the TLB is loaded before we take out the lock.
+	  * This will start fetching the value all the way into our L1
+	  * as well (and if it gets modified before we grab the lock,
+	  * it will be invalidated from our cache before we reload it).
+	  */
+	 lw	r26, r0
+	}
+	{
+	 auli	r21, zero, ha16(atomic_locks)
+
+	 bbns   r23, .Lcmpxchg_badaddr
+	}
+#if PAGE_SIZE < 0x10000
+	/* atomic_locks is page-aligned so for big pages we don't need this. */
+	addli   r21, r21, lo16(atomic_locks)
+#endif
+	{
+	 /*
+	  * Insert the hash bits into the page-aligned pointer.
+	  * ATOMIC_HASH_SHIFT is so big that we don't actually hash
+	  * the unmasked address bits, as that may cause unnecessary
+	  * collisions.
+	  */
+	 mm	ATOMIC_LOCK_REG_NAME, r25, r21, 2, (ATOMIC_HASH_SHIFT + 2) - 1
+
+	 seqi	r23, TREG_SYSCALL_NR_NAME, __NR_FAST_cmpxchg64
+	}
+	{
+	 /* Branch away at this point if we're doing a 64-bit cmpxchg. */
+	 bbs    r23, .Lcmpxchg64
+	 andi   r23, r0, 7       /* Precompute alignment for cmpxchg64. */
+	}
+	{
+	 /*
+	  * We very carefully align the code that actually runs with
+	  * the lock held (twelve bundles) so that we know it is all in
+	  * the icache when we start.  This instruction (the jump) is
+	  * at the start of the first cache line, address zero mod 64;
+	  * we jump to the very end of the second cache line to get that
+	  * line loaded in the icache, then fall through to issue the tns
+	  * in the third cache line, at which point it's all cached.
+	  * Note that is for performance, not correctness.
+	  */
+	 j      .Lcmpxchg32_tns
+	}
+
+/* Symbol for do_page_fault_ics() to use to compare against the PC. */
+.global __sys_cmpxchg_grab_lock
+__sys_cmpxchg_grab_lock:
+
+	/*
+	 * Perform the actual cmpxchg or atomic_update.
+	 */
+.Ldo_cmpxchg32:
+	{
+	 lw     r21, r0
+	 seqi	r23, TREG_SYSCALL_NR_NAME, __NR_FAST_atomic_update
+	 move	r24, r2
+	}
+	{
+	 seq    r22, r21, r1     /* See if cmpxchg matches. */
+	 and	r25, r21, r1     /* If atomic_update, compute (*mem & mask) */
+	}
+	{
+	 or	r22, r22, r23    /* Skip compare branch for atomic_update. */
+	 add	r25, r25, r2     /* Compute (*mem & mask) + addend. */
+	}
+	{
+	 mvnz	r24, r23, r25    /* Use atomic_update value if appropriate. */
+	 bbns   r22, .Lcmpxchg32_nostore
+	}
+	seq     r22, r24, r21    /* Are we storing the value we loaded? */
+	bbs     r22, .Lcmpxchg32_nostore
+	sw      r0, r24
+
+	/* The following instruction is the start of the second cache line. */
+	/* Do slow mtspr here so the following "mf" waits less. */
+	{
+	 move   sp, r27
+	 mtspr  SPR_EX_CONTEXT_K_0, r28
+	}
+	mf
+
+	{
+	 move   r0, r21
+	 sw     ATOMIC_LOCK_REG_NAME, zero
+	}
+	iret
+
+	/* Duplicated code here in the case where we don't overlap "mf" */
+.Lcmpxchg32_nostore:
+	{
+	 move   r0, r21
+	 sw     ATOMIC_LOCK_REG_NAME, zero
+	}
+	{
+	 move   sp, r27
+	 mtspr  SPR_EX_CONTEXT_K_0, r28
+	}
+	iret
+
+	/*
+	 * The locking code is the same for 32-bit cmpxchg/atomic_update,
+	 * and for 64-bit cmpxchg.  We provide it as a macro and put
+	 * it into both versions.  We can't share the code literally
+	 * since it depends on having the right branch-back address.
+	 */
+	.macro  cmpxchg_lock, bitwidth
+
+	/* Lock; if we succeed, jump back up to the read-modify-write. */
+#ifdef CONFIG_SMP
+	tns     r21, ATOMIC_LOCK_REG_NAME
+#else
+	/*
+	 * Non-SMP preserves all the lock infrastructure, to keep the
+	 * code simpler for the interesting (SMP) case.  However, we do
+	 * one small optimization here and in atomic_asm.S, which is
+	 * to fake out acquiring the actual lock in the atomic_lock table.
+	 */
+	movei	r21, 0
+#endif
+
+	/* Issue the slow SPR here while the tns result is in flight. */
+	mfspr   r28, SPR_EX_CONTEXT_K_0
+
+	{
+	 addi   r28, r28, 8    /* return to the instruction after the swint1 */
+	 bzt    r21, .Ldo_cmpxchg\bitwidth
+	}
+	/*
+	 * The preceding instruction is the last thing that must be
+	 * hot in the icache before we do the "tns" above.
+	 */
+
+#ifdef CONFIG_SMP
+	/*
+	 * We failed to acquire the tns lock on our first try.  Now use
+	 * bounded exponential backoff to retry, like __atomic_spinlock().
+	 */
+	{
+	 moveli r23, 2048       /* maximum backoff time in cycles */
+	 moveli r25, 32         /* starting backoff time in cycles */
+	}
+1:	mfspr   r26, CYCLE_LOW  /* get start point for this backoff */
+2:	mfspr   r22, CYCLE_LOW  /* test to see if we've backed off enough */
+	sub     r22, r22, r26
+	slt     r22, r22, r25
+	bbst    r22, 2b
+	{
+	 shli   r25, r25, 1     /* double the backoff; retry the tns */
+	 tns    r21, ATOMIC_LOCK_REG_NAME
+	}
+	slt     r26, r23, r25   /* is the proposed backoff too big? */
+	{
+	 mvnz   r25, r26, r23
+	 bzt    r21, .Ldo_cmpxchg\bitwidth
+	}
+	j       1b
+#endif /* CONFIG_SMP */
+	.endm
+
+.Lcmpxchg32_tns:
+	/*
+	 * This is the last instruction on the second cache line.
+	 * The nop here loads the second line, then we fall through
+	 * to the tns to load the third line before we take the lock.
+	 */
+	nop
+	cmpxchg_lock 32
+
+	/*
+	 * This code is invoked from sys_cmpxchg after most of the
+	 * preconditions have been checked.  We still need to check
+	 * that r0 is 8-byte aligned, since if it's not we won't
+	 * actually be atomic.  However, ATOMIC_LOCK_REG has the atomic
+	 * lock pointer and r27/r28 have the saved SP/PC.
+	 * r23 is holding "r0 & 7" so we can test for alignment.
+	 * The compare value is in r2/r3; the new value is in r4/r5.
+	 * On return, we must put the old value in r0/r1.
+	 */
+	.align 64
+.Lcmpxchg64:
+	{
+	 bzt     r23, .Lcmpxchg64_tns
+	}
+	j       .Lcmpxchg_badaddr
+
+.Ldo_cmpxchg64:
+	{
+	 lw     r21, r0
+	 addi   r25, r0, 4
+	}
+	{
+	 lw     r1, r25
+	}
+	seq     r26, r21, r2
+	{
+	 bz     r26, .Lcmpxchg64_mismatch
+	 seq    r26, r1, r3
+	}
+	{
+	 bz     r26, .Lcmpxchg64_mismatch
+	}
+	sw      r0, r4
+	sw      r25, r5
+
+	/*
+	 * The 32-bit path provides optimized "match" and "mismatch"
+	 * iret paths, but we don't have enough bundles in this cache line
+	 * to do that, so we just make even the "mismatch" path do an "mf".
+	 */
+.Lcmpxchg64_mismatch:
+	{
+	 move   sp, r27
+	 mtspr  SPR_EX_CONTEXT_K_0, r28
+	}
+	mf
+	{
+	 move   r0, r21
+	 sw     ATOMIC_LOCK_REG_NAME, zero
+	}
+	iret
+
+.Lcmpxchg64_tns:
+	cmpxchg_lock 64
+
+
+	/*
+	 * Reset sp and revector to sys_cmpxchg_badaddr(), which will
+	 * just raise the appropriate signal and exit.  Doing it this
+	 * way means we don't have to duplicate the code in intvec.S's
+	 * int_hand macro that locates the top of the stack.
+	 */
+.Lcmpxchg_badaddr:
+	{
+	 moveli TREG_SYSCALL_NR_NAME, __NR_cmpxchg_badaddr
+	 move   sp, r27
+	}
+	j       intvec_SWINT_1
+	ENDPROC(sys_cmpxchg)
+	ENTRY(__sys_cmpxchg_end)
+
+
+/* The single-step support may need to read all the registers. */
+int_unalign:
+	push_extra_callee_saves r0
+	j       do_trap
+
+/* Include .intrpt array of interrupt vectors */
+	.section ".intrpt", "ax"
+
+#ifndef CONFIG_USE_PMC
+#define handle_perf_interrupt bad_intr
+#endif
+
+#ifndef CONFIG_HARDWALL
+#define do_hardwall_trap bad_intr
+#endif
+
+	int_hand     INT_ITLB_MISS, ITLB_MISS, \
+		     do_page_fault, handle_interrupt_no_single_step
+	int_hand     INT_MEM_ERROR, MEM_ERROR, bad_intr
+	int_hand     INT_ILL, ILL, do_trap, handle_ill
+	int_hand     INT_GPV, GPV, do_trap
+	int_hand     INT_SN_ACCESS, SN_ACCESS, do_trap
+	int_hand     INT_IDN_ACCESS, IDN_ACCESS, do_trap
+	int_hand     INT_UDN_ACCESS, UDN_ACCESS, do_trap
+	int_hand     INT_IDN_REFILL, IDN_REFILL, bad_intr
+	int_hand     INT_UDN_REFILL, UDN_REFILL, bad_intr
+	int_hand     INT_IDN_COMPLETE, IDN_COMPLETE, bad_intr
+	int_hand     INT_UDN_COMPLETE, UDN_COMPLETE, bad_intr
+	int_hand     INT_SWINT_3, SWINT_3, do_trap
+	int_hand     INT_SWINT_2, SWINT_2, do_trap
+	int_hand     INT_SWINT_1, SWINT_1, SYSCALL, handle_syscall
+	int_hand     INT_SWINT_0, SWINT_0, do_trap
+	int_hand     INT_UNALIGN_DATA, UNALIGN_DATA, int_unalign
+	int_hand     INT_DTLB_MISS, DTLB_MISS, do_page_fault
+	int_hand     INT_DTLB_ACCESS, DTLB_ACCESS, do_page_fault
+	int_hand     INT_DMATLB_MISS, DMATLB_MISS, do_page_fault
+	int_hand     INT_DMATLB_ACCESS, DMATLB_ACCESS, do_page_fault
+	int_hand     INT_SNITLB_MISS, SNITLB_MISS, do_page_fault
+	int_hand     INT_SN_NOTIFY, SN_NOTIFY, bad_intr
+	int_hand     INT_SN_FIREWALL, SN_FIREWALL, do_hardwall_trap
+	int_hand     INT_IDN_FIREWALL, IDN_FIREWALL, bad_intr
+	int_hand     INT_UDN_FIREWALL, UDN_FIREWALL, do_hardwall_trap
+	int_hand     INT_TILE_TIMER, TILE_TIMER, do_timer_interrupt
+	int_hand     INT_IDN_TIMER, IDN_TIMER, bad_intr
+	int_hand     INT_UDN_TIMER, UDN_TIMER, bad_intr
+	int_hand     INT_DMA_NOTIFY, DMA_NOTIFY, bad_intr
+	int_hand     INT_IDN_CA, IDN_CA, bad_intr
+	int_hand     INT_UDN_CA, UDN_CA, bad_intr
+	int_hand     INT_IDN_AVAIL, IDN_AVAIL, bad_intr
+	int_hand     INT_UDN_AVAIL, UDN_AVAIL, bad_intr
+	int_hand     INT_PERF_COUNT, PERF_COUNT, \
+		     handle_perf_interrupt, handle_nmi
+	int_hand     INT_INTCTRL_3, INTCTRL_3, bad_intr
+#if CONFIG_KERNEL_PL == 2
+	dc_dispatch  INT_INTCTRL_2, INTCTRL_2
+	int_hand     INT_INTCTRL_1, INTCTRL_1, bad_intr
+#else
+	int_hand     INT_INTCTRL_2, INTCTRL_2, bad_intr
+	dc_dispatch  INT_INTCTRL_1, INTCTRL_1
+#endif
+	int_hand     INT_INTCTRL_0, INTCTRL_0, bad_intr
+	int_hand     INT_MESSAGE_RCV_DWNCL, MESSAGE_RCV_DWNCL, \
+		     hv_message_intr
+	int_hand     INT_DEV_INTR_DWNCL, DEV_INTR_DWNCL, \
+		     tile_dev_intr
+	int_hand     INT_I_ASID, I_ASID, bad_intr
+	int_hand     INT_D_ASID, D_ASID, bad_intr
+	int_hand     INT_DMATLB_MISS_DWNCL, DMATLB_MISS_DWNCL, \
+		     do_page_fault
+	int_hand     INT_SNITLB_MISS_DWNCL, SNITLB_MISS_DWNCL, \
+		     do_page_fault
+	int_hand     INT_DMATLB_ACCESS_DWNCL, DMATLB_ACCESS_DWNCL, \
+		     do_page_fault
+	int_hand     INT_SN_CPL, SN_CPL, bad_intr
+	int_hand     INT_DOUBLE_FAULT, DOUBLE_FAULT, do_trap
+	int_hand     INT_AUX_PERF_COUNT, AUX_PERF_COUNT, \
+		     handle_perf_interrupt, handle_nmi
+
+	/* Synthetic interrupt delivered only by the simulator */
+	int_hand     INT_BREAKPOINT, BREAKPOINT, do_breakpoint
diff --git a/kernel/arch/tile/kernel/intvec_64.S b/kernel/arch/tile/kernel/intvec_64.S
new file mode 100644
index 000000000..5b67efcec
--- /dev/null
+++ b/kernel/arch/tile/kernel/intvec_64.S
@@ -0,0 +1,1573 @@
+/*
+ * Copyright 2011 Tilera Corporation. All Rights Reserved.
+ *
+ *   This program is free software; you can redistribute it and/or
+ *   modify it under the terms of the GNU General Public License
+ *   as published by the Free Software Foundation, version 2.
+ *
+ *   This program is distributed in the hope that it will be useful, but
+ *   WITHOUT ANY WARRANTY; without even the implied warranty of
+ *   MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
+ *   NON INFRINGEMENT.  See the GNU General Public License for
+ *   more details.
+ *
+ * Linux interrupt vectors.
+ */
+
+#include <linux/linkage.h>
+#include <linux/errno.h>
+#include <linux/unistd.h>
+#include <linux/init.h>
+#include <asm/ptrace.h>
+#include <asm/thread_info.h>
+#include <asm/irqflags.h>
+#include <asm/asm-offsets.h>
+#include <asm/types.h>
+#include <asm/traps.h>
+#include <asm/signal.h>
+#include <hv/hypervisor.h>
+#include <arch/abi.h>
+#include <arch/interrupts.h>
+#include <arch/spr_def.h>
+
+#define PTREGS_PTR(reg, ptreg) addli reg, sp, C_ABI_SAVE_AREA_SIZE + (ptreg)
+
+#define PTREGS_OFFSET_SYSCALL PTREGS_OFFSET_REG(TREG_SYSCALL_NR)
+
+#if CONFIG_KERNEL_PL == 1 || CONFIG_KERNEL_PL == 2
+/*
+ * Set "result" non-zero if ex1 holds the PL of the kernel
+ * (with or without ICS being set).  Note this works only
+ * because we never find the PL at level 3.
+ */
+# define IS_KERNEL_EX1(result, ex1) andi result, ex1, CONFIG_KERNEL_PL
+#else
+# error Recode IS_KERNEL_EX1 for CONFIG_KERNEL_PL
+#endif
+
+	.macro  push_reg reg, ptr=sp, delta=-8
+	{
+	 st     \ptr, \reg
+	 addli  \ptr, \ptr, \delta
+	}
+	.endm
+
+	.macro  pop_reg reg, ptr=sp, delta=8
+	{
+	 ld     \reg, \ptr
+	 addli  \ptr, \ptr, \delta
+	}
+	.endm
+
+	.macro  pop_reg_zero reg, zreg, ptr=sp, delta=8
+	{
+	 move   \zreg, zero
+	 ld     \reg, \ptr
+	 addi   \ptr, \ptr, \delta
+	}
+	.endm
+
+	.macro  push_extra_callee_saves reg
+	PTREGS_PTR(\reg, PTREGS_OFFSET_REG(51))
+	push_reg r51, \reg
+	push_reg r50, \reg
+	push_reg r49, \reg
+	push_reg r48, \reg
+	push_reg r47, \reg
+	push_reg r46, \reg
+	push_reg r45, \reg
+	push_reg r44, \reg
+	push_reg r43, \reg
+	push_reg r42, \reg
+	push_reg r41, \reg
+	push_reg r40, \reg
+	push_reg r39, \reg
+	push_reg r38, \reg
+	push_reg r37, \reg
+	push_reg r36, \reg
+	push_reg r35, \reg
+	push_reg r34, \reg, PTREGS_OFFSET_BASE - PTREGS_OFFSET_REG(34)
+	.endm
+
+	.macro  panic str
+	.pushsection .rodata, "a"
+1:
+	.asciz  "\str"
+	.popsection
+	{
+	 moveli r0, hw2_last(1b)
+	}
+	{
+	 shl16insli r0, r0, hw1(1b)
+	}
+	{
+	 shl16insli r0, r0, hw0(1b)
+	 jal    panic
+	}
+	.endm
+
+	/*
+	 * Unalign data exception fast handling: In order to handle
+	 * unaligned data access, a fast JIT version is generated and stored
+	 * in a specific area in user space. We first need to do a quick poke
+	 * to see if the JIT is available. We use certain bits in the fault
+	 * PC (3 to 9 is used for 16KB page size) as index to address the JIT
+	 * code area. The first 64bit word is the fault PC, and the 2nd one is
+	 * the fault bundle itself. If these 2 words both match, then we
+	 * directly "iret" to JIT code. If not, a slow path is invoked to
+	 * generate new JIT code. Note: the current JIT code WILL be
+	 * overwritten if it existed. So, ideally we can handle 128 unalign
+	 * fixups via JIT. For lookup efficiency and to effectively support
+	 * tight loops with multiple unaligned reference, a simple
+	 * direct-mapped cache is used.
+	 *
+	 * SPR_EX_CONTEXT_K_0 is modified to return to JIT code.
+	 * SPR_EX_CONTEXT_K_1 has ICS set.
+	 * SPR_EX_CONTEXT_0_0 is setup to user program's next PC.
+	 * SPR_EX_CONTEXT_0_1 = 0.
+	 */
+	.macro int_hand_unalign_fast  vecnum, vecname
+	.org  (\vecnum << 8)
+intvec_\vecname:
+	/* Put r3 in SPR_SYSTEM_SAVE_K_1.  */
+	mtspr   SPR_SYSTEM_SAVE_K_1, r3
+
+	mfspr   r3, SPR_EX_CONTEXT_K_1
+	/*
+	 * Examine if exception comes from user without ICS set.
+	 * If not, just go directly to the slow path.
+	 */
+	bnez    r3, hand_unalign_slow_nonuser
+
+	mfspr   r3, SPR_SYSTEM_SAVE_K_0
+
+	/* Get &thread_info->unalign_jit_tmp[0] in r3. */
+	bfexts  r3, r3, 0, CPU_SHIFT-1
+	mm      r3, zero, LOG2_THREAD_SIZE, 63
+	addli   r3, r3, THREAD_INFO_UNALIGN_JIT_TMP_OFFSET
+
+	/*
+	 * Save r0, r1, r2 into thread_info array r3 points to
+	 * from low to high memory in order.
+	 */
+	st_add  r3, r0, 8
+	st_add  r3, r1, 8
+	{
+	 st_add r3, r2, 8
+	 andi   r2, sp, 7
+	}
+
+	/* Save stored r3 value so we can revert it on a page fault. */
+	mfspr   r1, SPR_SYSTEM_SAVE_K_1
+	st      r3, r1
+
+	{
+	 /* Generate a SIGBUS if sp is not 8-byte aligned. */
+	 bnez   r2, hand_unalign_slow_badsp
+	}
+
+	/*
+	 * Get the thread_info in r0; load r1 with pc. Set the low bit of sp
+	 * as an indicator to the page fault code in case we fault.
+	 */
+	{
+	 ori    sp, sp, 1
+	 mfspr  r1, SPR_EX_CONTEXT_K_0
+	}
+
+	/* Add the jit_info offset in thread_info; extract r1 [3:9] into r2. */
+	{
+	 addli  r0, r3, THREAD_INFO_UNALIGN_JIT_BASE_OFFSET - \
+	  (THREAD_INFO_UNALIGN_JIT_TMP_OFFSET + (3 * 8))
+	 bfextu r2, r1, 3, (2 + PAGE_SHIFT - UNALIGN_JIT_SHIFT)
+	}
+
+	/* Load the jit_info; multiply r2 by 128. */
+	{
+	 ld     r0, r0
+	 shli   r2, r2, UNALIGN_JIT_SHIFT
+	}
+
+	/*
+	 * If r0 is NULL, the JIT page is not mapped, so go to slow path;
+	 * add offset r2 to r0 at the same time.
+	 */
+	{
+	 beqz   r0, hand_unalign_slow
+	 add    r2, r0, r2
+	}
+
+        /*
+	 * We are loading from userspace (both the JIT info PC and
+	 * instruction word, and the instruction word we executed)
+	 * and since either could fault while holding the interrupt
+	 * critical section, we must tag this region and check it in
+	 * do_page_fault() to handle it properly.
+	 */
+ENTRY(__start_unalign_asm_code)
+
+	/* Load first word of JIT in r0 and increment r2 by 8. */
+	ld_add  r0, r2, 8
+
+	/*
+	 * Compare the PC with the 1st word in JIT; load the fault bundle
+	 * into r1.
+	 */
+	{
+	 cmpeq  r0, r0, r1
+	 ld     r1, r1
+	}
+
+	/* Go to slow path if PC doesn't match. */
+	beqz    r0, hand_unalign_slow
+
+	/*
+	 * Load the 2nd word of JIT, which is supposed to be the fault
+	 * bundle for a cache hit. Increment r2; after this bundle r2 will
+	 * point to the potential start of the JIT code we want to run.
+	 */
+	ld_add  r0, r2, 8
+
+	/* No further accesses to userspace are done after this point. */
+ENTRY(__end_unalign_asm_code)
+
+	/* Compare the real bundle with what is saved in the JIT area. */
+	{
+	 cmpeq  r0, r1, r0
+	 mtspr  SPR_EX_CONTEXT_0_1, zero
+	}
+
+	/* Go to slow path if the fault bundle does not match. */
+	beqz    r0, hand_unalign_slow
+
+	/*
+	 * A cache hit is found.
+	 * r2 points to start of JIT code (3rd word).
+	 * r0 is the fault pc.
+	 * r1 is the fault bundle.
+	 * Reset the low bit of sp.
+	 */
+	{
+	 mfspr  r0, SPR_EX_CONTEXT_K_0
+	 andi   sp, sp, ~1
+	}
+
+	/* Write r2 into EX_CONTEXT_K_0 and increment PC. */
+	{
+	 mtspr  SPR_EX_CONTEXT_K_0, r2
+	 addi   r0, r0, 8
+	}
+
+	/*
+	 * Set ICS on kernel EX_CONTEXT_K_1 in order to "iret" to
+	 * user with ICS set. This way, if the JIT fixup causes another
+	 * unalign exception (which shouldn't be possible) the user
+	 * process will be terminated with SIGBUS. Also, our fixup will
+	 * run without interleaving with external interrupts.
+	 * Each fixup is at most 14 bundles, so it won't hold ICS for long.
+	 */
+	{
+	 movei  r1, PL_ICS_EX1(USER_PL, 1)
+	 mtspr  SPR_EX_CONTEXT_0_0, r0
+	}
+
+	{
+	 mtspr  SPR_EX_CONTEXT_K_1, r1
+	 addi   r3, r3, -(3 * 8)
+	}
+
+	/* Restore r0..r3. */
+	ld_add  r0, r3, 8
+	ld_add  r1, r3, 8
+	ld_add  r2, r3, 8
+	ld      r3, r3
+
+	iret
+	ENDPROC(intvec_\vecname)
+	.endm
+
+#ifdef __COLLECT_LINKER_FEEDBACK__
+	.pushsection .text.intvec_feedback,"ax"
+intvec_feedback:
+	.popsection
+#endif
+
+	/*
+	 * Default interrupt handler.
+	 *
+	 * vecnum is where we'll put this code.
+	 * c_routine is the C routine we'll call.
+	 *
+	 * The C routine is passed two arguments:
+	 * - A pointer to the pt_regs state.
+	 * - The interrupt vector number.
+	 *
+	 * The "processing" argument specifies the code for processing
+	 * the interrupt. Defaults to "handle_interrupt".
+	 */
+	.macro __int_hand vecnum, vecname, c_routine,processing=handle_interrupt
+intvec_\vecname:
+	/* Temporarily save a register so we have somewhere to work. */
+
+	mtspr   SPR_SYSTEM_SAVE_K_1, r0
+	mfspr   r0, SPR_EX_CONTEXT_K_1
+
+	/*
+	 * The unalign data fastpath code sets the low bit in sp to
+	 * force us to reset it here on fault.
+	 */
+	{
+	 blbs   sp, 2f
+	 IS_KERNEL_EX1(r0, r0)
+	}
+
+	.ifc    \vecnum, INT_DOUBLE_FAULT
+	/*
+	 * For double-faults from user-space, fall through to the normal
+	 * register save and stack setup path.  Otherwise, it's the
+	 * hypervisor giving us one last chance to dump diagnostics, and we
+	 * branch to the kernel_double_fault routine to do so.
+	 */
+	beqz    r0, 1f
+	j       _kernel_double_fault
+1:
+	.else
+	/*
+	 * If we're coming from user-space, then set sp to the top of
+	 * the kernel stack.  Otherwise, assume sp is already valid.
+	 */
+	{
+	 bnez   r0, 0f
+	 move   r0, sp
+	}
+	.endif
+
+	.ifc    \c_routine, do_page_fault
+	/*
+	 * The page_fault handler may be downcalled directly by the
+	 * hypervisor even when Linux is running and has ICS set.
+	 *
+	 * In this case the contents of EX_CONTEXT_K_1 reflect the
+	 * previous fault and can't be relied on to choose whether or
+	 * not to reinitialize the stack pointer.  So we add a test
+	 * to see whether SYSTEM_SAVE_K_2 has the high bit set,
+	 * and if so we don't reinitialize sp, since we must be coming
+	 * from Linux.  (In fact the precise case is !(val & ~1),
+	 * but any Linux PC has to have the high bit set.)
+	 *
+	 * Note that the hypervisor *always* sets SYSTEM_SAVE_K_2 for
+	 * any path that turns into a downcall to one of our TLB handlers.
+	 *
+	 * FIXME: if we end up never using this path, perhaps we should
+	 * prevent the hypervisor from generating downcalls in this case.
+	 * The advantage of getting a downcall is we can panic in Linux.
+	 */
+	mfspr   r0, SPR_SYSTEM_SAVE_K_2
+	{
+	 bltz   r0, 0f    /* high bit in S_S_1_2 is for a PC to use */
+	 move   r0, sp
+	}
+	.endif
+
+2:
+	/*
+	 * SYSTEM_SAVE_K_0 holds the cpu number in the high bits, and
+	 * the current stack top in the lower bits.  So we recover
+	 * our starting stack value by sign-extending the low bits, then
+	 * point sp at the top aligned address on the actual stack page.
+	 */
+	mfspr   r0, SPR_SYSTEM_SAVE_K_0
+	bfexts  r0, r0, 0, CPU_SHIFT-1
+
+0:
+	/*
+	 * Align the stack mod 64 so we can properly predict what
+	 * cache lines we need to write-hint to reduce memory fetch
+	 * latency as we enter the kernel.  The layout of memory is
+	 * as follows, with cache line 0 at the lowest VA, and cache
+	 * line 8 just below the r0 value this "andi" computes.
+	 * Note that we never write to cache line 8, and we skip
+	 * cache lines 1-3 for syscalls.
+	 *
+	 *    cache line 8: ptregs padding (two words)
+	 *    cache line 7: sp, lr, pc, ex1, faultnum, orig_r0, flags, cmpexch
+	 *    cache line 6: r46...r53 (tp)
+	 *    cache line 5: r38...r45
+	 *    cache line 4: r30...r37
+	 *    cache line 3: r22...r29
+	 *    cache line 2: r14...r21
+	 *    cache line 1: r6...r13
+	 *    cache line 0: 2 x frame, r0..r5
+	 */
+#if STACK_TOP_DELTA != 64
+#error STACK_TOP_DELTA must be 64 for assumptions here and in task_pt_regs()
+#endif
+	andi    r0, r0, -64
+
+	/*
+	 * Push the first four registers on the stack, so that we can set
+	 * them to vector-unique values before we jump to the common code.
+	 *
+	 * Registers are pushed on the stack as a struct pt_regs,
+	 * with the sp initially just above the struct, and when we're
+	 * done, sp points to the base of the struct, minus
+	 * C_ABI_SAVE_AREA_SIZE, so we can directly jal to C code.
+	 *
+	 * This routine saves just the first four registers, plus the
+	 * stack context so we can do proper backtracing right away,
+	 * and defers to handle_interrupt to save the rest.
+	 * The backtracer needs pc, ex1, lr, sp, r52, and faultnum,
+	 * and needs sp set to its final location at the bottom of
+	 * the stack frame.
+	 */
+	addli   r0, r0, PTREGS_OFFSET_LR - (PTREGS_SIZE + KSTK_PTREGS_GAP)
+	wh64    r0   /* cache line 7 */
+	{
+	 st     r0, lr
+	 addli  r0, r0, PTREGS_OFFSET_SP - PTREGS_OFFSET_LR
+	}
+	{
+	 st     r0, sp
+	 addli  sp, r0, PTREGS_OFFSET_REG(52) - PTREGS_OFFSET_SP
+	}
+	wh64    sp   /* cache line 6 */
+	{
+	 st     sp, r52
+	 addli  sp, sp, PTREGS_OFFSET_REG(1) - PTREGS_OFFSET_REG(52)
+	}
+	wh64    sp   /* cache line 0 */
+	{
+	 st     sp, r1
+	 addli  sp, sp, PTREGS_OFFSET_REG(2) - PTREGS_OFFSET_REG(1)
+	}
+	{
+	 st     sp, r2
+	 addli  sp, sp, PTREGS_OFFSET_REG(3) - PTREGS_OFFSET_REG(2)
+	}
+	{
+	 st     sp, r3
+	 addli  sp, sp, PTREGS_OFFSET_PC - PTREGS_OFFSET_REG(3)
+	}
+	mfspr   r0, SPR_EX_CONTEXT_K_0
+	.ifc \processing,handle_syscall
+	/*
+	 * Bump the saved PC by one bundle so that when we return, we won't
+	 * execute the same swint instruction again.  We need to do this while
+	 * we're in the critical section.
+	 */
+	addi    r0, r0, 8
+	.endif
+	{
+	 st     sp, r0
+	 addli  sp, sp, PTREGS_OFFSET_EX1 - PTREGS_OFFSET_PC
+	}
+	mfspr   r0, SPR_EX_CONTEXT_K_1
+	{
+	 st     sp, r0
+	 addi   sp, sp, PTREGS_OFFSET_FAULTNUM - PTREGS_OFFSET_EX1
+	/*
+	 * Use r0 for syscalls so it's a temporary; use r1 for interrupts
+	 * so that it gets passed through unchanged to the handler routine.
+	 * Note that the .if conditional confusingly spans bundles.
+	 */
+	 .ifc \processing,handle_syscall
+	 movei  r0, \vecnum
+	}
+	{
+	 st     sp, r0
+	 .else
+	 movei  r1, \vecnum
+	}
+	{
+	 st     sp, r1
+	 .endif
+	 addli  sp, sp, PTREGS_OFFSET_REG(0) - PTREGS_OFFSET_FAULTNUM
+	}
+	mfspr   r0, SPR_SYSTEM_SAVE_K_1    /* Original r0 */
+	{
+	 st     sp, r0
+	 addi   sp, sp, -PTREGS_OFFSET_REG(0) - 8
+	}
+	{
+	 st     sp, zero        /* write zero into "Next SP" frame pointer */
+	 addi   sp, sp, -8      /* leave SP pointing at bottom of frame */
+	}
+	.ifc \processing,handle_syscall
+	j       handle_syscall
+	.else
+	/* Capture per-interrupt SPR context to registers. */
+	.ifc \c_routine, do_page_fault
+	mfspr   r2, SPR_SYSTEM_SAVE_K_3   /* address of page fault */
+	mfspr   r3, SPR_SYSTEM_SAVE_K_2   /* info about page fault */
+	.else
+	.ifc \vecnum, INT_ILL_TRANS
+	mfspr   r2, ILL_VA_PC
+	.else
+	.ifc \vecnum, INT_DOUBLE_FAULT
+	mfspr   r2, SPR_SYSTEM_SAVE_K_2   /* double fault info from HV */
+	.else
+	.ifc \c_routine, do_trap
+	mfspr   r2, GPV_REASON
+	.else
+	.ifc \c_routine, handle_perf_interrupt
+	mfspr   r2, PERF_COUNT_STS
+	.else
+	.ifc \c_routine, handle_perf_interrupt
+	mfspr   r2, AUX_PERF_COUNT_STS
+	.endif
+	.endif
+	.endif
+	.endif
+	.endif
+	.endif
+	/* Put function pointer in r0 */
+	moveli  r0, hw2_last(\c_routine)
+	shl16insli r0, r0, hw1(\c_routine)
+	{
+	 shl16insli r0, r0, hw0(\c_routine)
+	 j       \processing
+	}
+	.endif
+	ENDPROC(intvec_\vecname)
+
+#ifdef __COLLECT_LINKER_FEEDBACK__
+	.pushsection .text.intvec_feedback,"ax"
+	.org    (\vecnum << 5)
+	FEEDBACK_ENTER_EXPLICIT(intvec_\vecname, .intrpt, 1 << 8)
+	jrp     lr
+	.popsection
+#endif
+
+	.endm
+
+
+	/*
+	 * Save the rest of the registers that we didn't save in the actual
+	 * vector itself.  We can't use r0-r10 inclusive here.
+	 */
+	.macro  finish_interrupt_save, function
+
+	/* If it's a syscall, save a proper orig_r0, otherwise just zero. */
+	PTREGS_PTR(r52, PTREGS_OFFSET_ORIG_R0)
+	{
+	 .ifc \function,handle_syscall
+	 st     r52, r0
+	 .else
+	 st     r52, zero
+	 .endif
+	 PTREGS_PTR(r52, PTREGS_OFFSET_TP)
+	}
+	st      r52, tp
+	{
+	 mfspr  tp, CMPEXCH_VALUE
+	 PTREGS_PTR(r52, PTREGS_OFFSET_CMPEXCH)
+	}
+
+	/*
+	 * For ordinary syscalls, we save neither caller- nor callee-
+	 * save registers, since the syscall invoker doesn't expect the
+	 * caller-saves to be saved, and the called kernel functions will
+	 * take care of saving the callee-saves for us.
+	 *
+	 * For interrupts we save just the caller-save registers.  Saving
+	 * them is required (since the "caller" can't save them).  Again,
+	 * the called kernel functions will restore the callee-save
+	 * registers for us appropriately.
+	 *
+	 * On return, we normally restore nothing special for syscalls,
+	 * and just the caller-save registers for interrupts.
+	 *
+	 * However, there are some important caveats to all this:
+	 *
+	 * - We always save a few callee-save registers to give us
+	 *   some scratchpad registers to carry across function calls.
+	 *
+	 * - fork/vfork/etc require us to save all the callee-save
+	 *   registers, which we do in PTREGS_SYSCALL_ALL_REGS, below.
+	 *
+	 * - We always save r0..r5 and r10 for syscalls, since we need
+	 *   to reload them a bit later for the actual kernel call, and
+	 *   since we might need them for -ERESTARTNOINTR, etc.
+	 *
+	 * - Before invoking a signal handler, we save the unsaved
+	 *   callee-save registers so they are visible to the
+	 *   signal handler or any ptracer.
+	 *
+	 * - If the unsaved callee-save registers are modified, we set
+	 *   a bit in pt_regs so we know to reload them from pt_regs
+	 *   and not just rely on the kernel function unwinding.
+	 *   (Done for ptrace register writes and SA_SIGINFO handler.)
+	 */
+	{
+	 st     r52, tp
+	 PTREGS_PTR(r52, PTREGS_OFFSET_REG(33))
+	}
+	wh64    r52    /* cache line 4 */
+	push_reg r33, r52
+	push_reg r32, r52
+	push_reg r31, r52
+	.ifc \function,handle_syscall
+	push_reg r30, r52, PTREGS_OFFSET_SYSCALL - PTREGS_OFFSET_REG(30)
+	push_reg TREG_SYSCALL_NR_NAME, r52, \
+	  PTREGS_OFFSET_REG(5) - PTREGS_OFFSET_SYSCALL
+	.else
+
+	push_reg r30, r52, PTREGS_OFFSET_REG(29) - PTREGS_OFFSET_REG(30)
+	wh64    r52   /* cache line 3 */
+	push_reg r29, r52
+	push_reg r28, r52
+	push_reg r27, r52
+	push_reg r26, r52
+	push_reg r25, r52
+	push_reg r24, r52
+	push_reg r23, r52
+	push_reg r22, r52
+	wh64    r52   /* cache line 2 */
+	push_reg r21, r52
+	push_reg r20, r52
+	push_reg r19, r52
+	push_reg r18, r52
+	push_reg r17, r52
+	push_reg r16, r52
+	push_reg r15, r52
+	push_reg r14, r52
+	wh64    r52   /* cache line 1 */
+	push_reg r13, r52
+	push_reg r12, r52
+	push_reg r11, r52
+	push_reg r10, r52
+	push_reg r9, r52
+	push_reg r8, r52
+	push_reg r7, r52
+	push_reg r6, r52
+
+	.endif
+
+	push_reg r5, r52
+	st      r52, r4
+
+	/*
+	 * If we will be returning to the kernel, we will need to
+	 * reset the interrupt masks to the state they had before.
+	 * Set DISABLE_IRQ in flags iff we came from kernel pl with
+	 * irqs disabled.
+	 */
+	mfspr   r32, SPR_EX_CONTEXT_K_1
+	{
+	 IS_KERNEL_EX1(r22, r22)
+	 PTREGS_PTR(r21, PTREGS_OFFSET_FLAGS)
+	}
+	beqzt   r32, 1f       /* zero if from user space */
+	IRQS_DISABLED(r32)    /* zero if irqs enabled */
+#if PT_FLAGS_DISABLE_IRQ != 1
+# error Value of IRQS_DISABLED used to set PT_FLAGS_DISABLE_IRQ; fix
+#endif
+1:
+	.ifnc \function,handle_syscall
+	/* Record the fact that we saved the caller-save registers above. */
+	ori     r32, r32, PT_FLAGS_CALLER_SAVES
+	.endif
+	st      r21, r32
+
+	/*
+	 * we've captured enough state to the stack (including in
+	 * particular our EX_CONTEXT state) that we can now release
+	 * the interrupt critical section and replace it with our
+	 * standard "interrupts disabled" mask value.  This allows
+	 * synchronous interrupts (and profile interrupts) to punch
+	 * through from this point onwards.
+	 *
+	 * It's important that no code before this point touch memory
+	 * other than our own stack (to keep the invariant that this
+	 * is all that gets touched under ICS), and that no code after
+	 * this point reference any interrupt-specific SPR, in particular
+	 * the EX_CONTEXT_K_ values.
+	 */
+	.ifc \function,handle_nmi
+	IRQ_DISABLE_ALL(r20)
+	.else
+	IRQ_DISABLE(r20, r21)
+	.endif
+	mtspr   INTERRUPT_CRITICAL_SECTION, zero
+
+	/* Load tp with our per-cpu offset. */
+#ifdef CONFIG_SMP
+	{
+	 mfspr  r20, SPR_SYSTEM_SAVE_K_0
+	 moveli r21, hw2_last(__per_cpu_offset)
+	}
+	{
+	 shl16insli r21, r21, hw1(__per_cpu_offset)
+	 bfextu r20, r20, CPU_SHIFT, 63
+	}
+	shl16insli r21, r21, hw0(__per_cpu_offset)
+	shl3add r20, r20, r21
+	ld      tp, r20
+#else
+	move    tp, zero
+#endif
+
+#ifdef __COLLECT_LINKER_FEEDBACK__
+	/*
+	 * Notify the feedback routines that we were in the
+	 * appropriate fixed interrupt vector area.  Note that we
+	 * still have ICS set at this point, so we can't invoke any
+	 * atomic operations or we will panic.  The feedback
+	 * routines internally preserve r0..r10 and r30 up.
+	 */
+	.ifnc \function,handle_syscall
+	shli    r20, r1, 5
+	.else
+	moveli  r20, INT_SWINT_1 << 5
+	.endif
+	moveli  r21, hw2_last(intvec_feedback)
+	shl16insli r21, r21, hw1(intvec_feedback)
+	shl16insli r21, r21, hw0(intvec_feedback)
+	add     r20, r20, r21
+	jalr    r20
+
+	/* And now notify the feedback routines that we are here. */
+	FEEDBACK_ENTER(\function)
+#endif
+
+	/*
+	 * Prepare the first 256 stack bytes to be rapidly accessible
+	 * without having to fetch the background data.
+	 */
+	addi    r52, sp, -64
+	{
+	 wh64   r52
+	 addi   r52, r52, -64
+	}
+	{
+	 wh64   r52
+	 addi   r52, r52, -64
+	}
+	{
+	 wh64   r52
+	 addi   r52, r52, -64
+	}
+	wh64    r52
+
+#ifdef CONFIG_TRACE_IRQFLAGS
+	.ifnc \function,handle_nmi
+	/*
+	 * We finally have enough state set up to notify the irq
+	 * tracing code that irqs were disabled on entry to the handler.
+	 * The TRACE_IRQS_OFF call clobbers registers r0-r29.
+	 * For syscalls, we already have the register state saved away
+	 * on the stack, so we don't bother to do any register saves here,
+	 * and later we pop the registers back off the kernel stack.
+	 * For interrupt handlers, save r0-r3 in callee-saved registers.
+	 */
+	.ifnc \function,handle_syscall
+	{ move r30, r0; move r31, r1 }
+	{ move r32, r2; move r33, r3 }
+	.endif
+	TRACE_IRQS_OFF
+	.ifnc \function,handle_syscall
+	{ move r0, r30; move r1, r31 }
+	{ move r2, r32; move r3, r33 }
+	.endif
+	.endif
+#endif
+
+	.endm
+
+	/*
+	 * Redispatch a downcall.
+	 */
+	.macro  dc_dispatch vecnum, vecname
+	.org    (\vecnum << 8)
+intvec_\vecname:
+	j       _hv_downcall_dispatch
+	ENDPROC(intvec_\vecname)
+	.endm
+
+	/*
+	 * Common code for most interrupts.  The C function we're eventually
+	 * going to is in r0, and the faultnum is in r1; the original
+	 * values for those registers are on the stack.
+	 */
+	.pushsection .text.handle_interrupt,"ax"
+handle_interrupt:
+	finish_interrupt_save handle_interrupt
+
+	/* Jump to the C routine; it should enable irqs as soon as possible. */
+	{
+	 jalr   r0
+	 PTREGS_PTR(r0, PTREGS_OFFSET_BASE)
+	}
+	FEEDBACK_REENTER(handle_interrupt)
+	{
+	 movei  r30, 0   /* not an NMI */
+	 j      interrupt_return
+	}
+	STD_ENDPROC(handle_interrupt)
+
+/*
+ * This routine takes a boolean in r30 indicating if this is an NMI.
+ * If so, we also expect a boolean in r31 indicating whether to
+ * re-enable the oprofile interrupts.
+ *
+ * Note that .Lresume_userspace is jumped to directly in several
+ * places, and we need to make sure r30 is set correctly in those
+ * callers as well.
+ */
+STD_ENTRY(interrupt_return)
+	/* If we're resuming to kernel space, don't check thread flags. */
+	{
+	 bnez   r30, .Lrestore_all  /* NMIs don't special-case user-space */
+	 PTREGS_PTR(r29, PTREGS_OFFSET_EX1)
+	}
+	ld      r29, r29
+	IS_KERNEL_EX1(r29, r29)
+	{
+	 beqzt  r29, .Lresume_userspace
+	 move   r29, sp
+	}
+
+#ifdef CONFIG_PREEMPT
+	/* Returning to kernel space. Check if we need preemption. */
+	EXTRACT_THREAD_INFO(r29)
+	addli   r28, r29, THREAD_INFO_FLAGS_OFFSET
+	{
+	 ld     r28, r28
+	 addli  r29, r29, THREAD_INFO_PREEMPT_COUNT_OFFSET
+	}
+	{
+	 andi   r28, r28, _TIF_NEED_RESCHED
+	 ld4s   r29, r29
+	}
+	beqzt   r28, 1f
+	bnez    r29, 1f
+	/* Disable interrupts explicitly for preemption. */
+	IRQ_DISABLE(r20,r21)
+	TRACE_IRQS_OFF
+	jal     preempt_schedule_irq
+	FEEDBACK_REENTER(interrupt_return)
+1:
+#endif
+
+	/* If we're resuming to _cpu_idle_nap, bump PC forward by 8. */
+	{
+	 moveli r27, hw2_last(_cpu_idle_nap)
+	 PTREGS_PTR(r29, PTREGS_OFFSET_PC)
+	}
+	{
+	 ld     r28, r29
+	 shl16insli r27, r27, hw1(_cpu_idle_nap)
+	}
+	{
+	 shl16insli r27, r27, hw0(_cpu_idle_nap)
+	}
+	{
+	 cmpeq  r27, r27, r28
+	}
+	{
+	 blbc   r27, .Lrestore_all
+	 addi   r28, r28, 8
+	}
+	st      r29, r28
+	j       .Lrestore_all
+
+.Lresume_userspace:
+	FEEDBACK_REENTER(interrupt_return)
+
+	/*
+	 * Use r33 to hold whether we have already loaded the callee-saves
+	 * into ptregs.  We don't want to do it twice in this loop, since
+	 * then we'd clobber whatever changes are made by ptrace, etc.
+	 */
+	{
+	 movei  r33, 0
+	 move   r32, sp
+	}
+
+	/* Get base of stack in r32. */
+	EXTRACT_THREAD_INFO(r32)
+
+.Lretry_work_pending:
+	/*
+	 * Disable interrupts so as to make sure we don't
+	 * miss an interrupt that sets any of the thread flags (like
+	 * need_resched or sigpending) between sampling and the iret.
+	 * Routines like schedule() or do_signal() may re-enable
+	 * interrupts before returning.
+	 */
+	IRQ_DISABLE(r20, r21)
+	TRACE_IRQS_OFF  /* Note: clobbers registers r0-r29 */
+
+
+	/* Check to see if there is any work to do before returning to user. */
+	{
+	 addi   r29, r32, THREAD_INFO_FLAGS_OFFSET
+	 moveli r1, hw1_last(_TIF_ALLWORK_MASK)
+	}
+	{
+	 ld     r29, r29
+	 shl16insli r1, r1, hw0(_TIF_ALLWORK_MASK)
+	}
+	and     r1, r29, r1
+	beqzt   r1, .Lrestore_all
+
+	/*
+	 * Make sure we have all the registers saved for signal
+	 * handling or notify-resume.  Call out to C code to figure out
+	 * exactly what we need to do for each flag bit, then if
+	 * necessary, reload the flags and recheck.
+	 */
+	{
+	 PTREGS_PTR(r0, PTREGS_OFFSET_BASE)
+	 bnez   r33, 1f
+	}
+	push_extra_callee_saves r0
+	movei   r33, 1
+1:	jal     do_work_pending
+	bnez    r0, .Lretry_work_pending
+
+	/*
+	 * In the NMI case we
+	 * omit the call to single_process_check_nohz, which normally checks
+	 * to see if we should start or stop the scheduler tick, because
+	 * we can't call arbitrary Linux code from an NMI context.
+	 * We always call the homecache TLB deferral code to re-trigger
+	 * the deferral mechanism.
+	 *
+	 * The other chunk of responsibility this code has is to reset the
+	 * interrupt masks appropriately to reset irqs and NMIs.  We have
+	 * to call TRACE_IRQS_OFF and TRACE_IRQS_ON to support all the
+	 * lockdep-type stuff, but we can't set ICS until afterwards, since
+	 * ICS can only be used in very tight chunks of code to avoid
+	 * tripping over various assertions that it is off.
+	 */
+.Lrestore_all:
+	PTREGS_PTR(r0, PTREGS_OFFSET_EX1)
+	{
+	 ld      r0, r0
+	 PTREGS_PTR(r32, PTREGS_OFFSET_FLAGS)
+	}
+	{
+	 IS_KERNEL_EX1(r0, r0)
+	 ld     r32, r32
+	}
+	bnez    r0, 1f
+	j       2f
+#if PT_FLAGS_DISABLE_IRQ != 1
+# error Assuming PT_FLAGS_DISABLE_IRQ == 1 so we can use blbct below
+#endif
+1:	blbct   r32, 2f
+	IRQ_DISABLE(r20,r21)
+	TRACE_IRQS_OFF
+	movei   r0, 1
+	mtspr   INTERRUPT_CRITICAL_SECTION, r0
+	beqzt   r30, .Lrestore_regs
+	j       3f
+2:	TRACE_IRQS_ON
+	IRQ_ENABLE_LOAD(r20, r21)
+	movei   r0, 1
+	mtspr   INTERRUPT_CRITICAL_SECTION, r0
+	IRQ_ENABLE_APPLY(r20, r21)
+	beqzt   r30, .Lrestore_regs
+3:
+
+#if INT_PERF_COUNT + 1 != INT_AUX_PERF_COUNT
+# error Bad interrupt assumption
+#endif
+	{
+	 movei  r0, 3   /* two adjacent bits for the PERF_COUNT mask */
+	 beqz   r31, .Lrestore_regs
+	}
+	shli    r0, r0, INT_PERF_COUNT
+	mtspr   SPR_INTERRUPT_MASK_RESET_K, r0
+
+	/*
+	 * We now commit to returning from this interrupt, since we will be
+	 * doing things like setting EX_CONTEXT SPRs and unwinding the stack
+	 * frame.  No calls should be made to any other code after this point.
+	 * This code should only be entered with ICS set.
+	 * r32 must still be set to ptregs.flags.
+	 * We launch loads to each cache line separately first, so we can
+	 * get some parallelism out of the memory subsystem.
+	 * We start zeroing caller-saved registers throughout, since
+	 * that will save some cycles if this turns out to be a syscall.
+	 */
+.Lrestore_regs:
+
+	/*
+	 * Rotate so we have one high bit and one low bit to test.
+	 * - low bit says whether to restore all the callee-saved registers,
+	 *   or just r30-r33, and r52 up.
+	 * - high bit (i.e. sign bit) says whether to restore all the
+	 *   caller-saved registers, or just r0.
+	 */
+#if PT_FLAGS_CALLER_SAVES != 2 || PT_FLAGS_RESTORE_REGS != 4
+# error Rotate trick does not work :-)
+#endif
+	{
+	 rotli  r20, r32, 62
+	 PTREGS_PTR(sp, PTREGS_OFFSET_REG(0))
+	}
+
+	/*
+	 * Load cache lines 0, 4, 6 and 7, in that order, then use
+	 * the last loaded value, which makes it likely that the other
+	 * cache lines have also loaded, at which point we should be
+	 * able to safely read all the remaining words on those cache
+	 * lines without waiting for the memory subsystem.
+	 */
+	pop_reg r0, sp, PTREGS_OFFSET_REG(30) - PTREGS_OFFSET_REG(0)
+	pop_reg r30, sp, PTREGS_OFFSET_REG(52) - PTREGS_OFFSET_REG(30)
+	pop_reg_zero r52, r3, sp, PTREGS_OFFSET_CMPEXCH - PTREGS_OFFSET_REG(52)
+	pop_reg_zero r21, r27, sp, PTREGS_OFFSET_EX1 - PTREGS_OFFSET_CMPEXCH
+	pop_reg_zero lr, r2, sp, PTREGS_OFFSET_PC - PTREGS_OFFSET_EX1
+	{
+	 mtspr  CMPEXCH_VALUE, r21
+	 move   r4, zero
+	}
+	pop_reg r21, sp, PTREGS_OFFSET_REG(31) - PTREGS_OFFSET_PC
+	{
+	 mtspr  SPR_EX_CONTEXT_K_1, lr
+	 IS_KERNEL_EX1(lr, lr)
+	}
+	{
+	 mtspr  SPR_EX_CONTEXT_K_0, r21
+	 move   r5, zero
+	}
+
+	/* Restore callee-saveds that we actually use. */
+	pop_reg_zero r31, r6
+	pop_reg_zero r32, r7
+	pop_reg_zero r33, r8, sp, PTREGS_OFFSET_REG(29) - PTREGS_OFFSET_REG(33)
+
+	/*
+	 * If we modified other callee-saveds, restore them now.
+	 * This is rare, but could be via ptrace or signal handler.
+	 */
+	{
+	 move   r9, zero
+	 blbs   r20, .Lrestore_callees
+	}
+.Lcontinue_restore_regs:
+
+	/* Check if we're returning from a syscall. */
+	{
+	 move   r10, zero
+	 bltzt  r20, 1f  /* no, so go restore callee-save registers */
+	}
+
+	/*
+	 * Check if we're returning to userspace.
+	 * Note that if we're not, we don't worry about zeroing everything.
+	 */
+	{
+	 addli  sp, sp, PTREGS_OFFSET_LR - PTREGS_OFFSET_REG(29)
+	 bnez   lr, .Lkernel_return
+	}
+
+	/*
+	 * On return from syscall, we've restored r0 from pt_regs, but we
+	 * clear the remainder of the caller-saved registers.  We could
+	 * restore the syscall arguments, but there's not much point,
+	 * and it ensures user programs aren't trying to use the
+	 * caller-saves if we clear them, as well as avoiding leaking
+	 * kernel pointers into userspace.
+	 */
+	pop_reg_zero lr, r11, sp, PTREGS_OFFSET_TP - PTREGS_OFFSET_LR
+	pop_reg_zero tp, r12, sp, PTREGS_OFFSET_SP - PTREGS_OFFSET_TP
+	{
+	 ld     sp, sp
+	 move   r13, zero
+	 move   r14, zero
+	}
+	{ move r15, zero; move r16, zero }
+	{ move r17, zero; move r18, zero }
+	{ move r19, zero; move r20, zero }
+	{ move r21, zero; move r22, zero }
+	{ move r23, zero; move r24, zero }
+	{ move r25, zero; move r26, zero }
+
+	/* Set r1 to errno if we are returning an error, otherwise zero. */
+	{
+	 moveli r29, 4096
+	 sub    r1, zero, r0
+	}
+	{
+	 move   r28, zero
+	 cmpltu r29, r1, r29
+	}
+	{
+	 mnz    r1, r29, r1
+	 move   r29, zero
+	}
+	iret
+
+	/*
+	 * Not a syscall, so restore caller-saved registers.
+	 * First kick off loads for cache lines 1-3, which we're touching
+	 * for the first time here.
+	 */
+	.align 64
+1:	pop_reg r29, sp, PTREGS_OFFSET_REG(21) - PTREGS_OFFSET_REG(29)
+	pop_reg r21, sp, PTREGS_OFFSET_REG(13) - PTREGS_OFFSET_REG(21)
+	pop_reg r13, sp, PTREGS_OFFSET_REG(1) - PTREGS_OFFSET_REG(13)
+	pop_reg r1
+	pop_reg r2
+	pop_reg r3
+	pop_reg r4
+	pop_reg r5
+	pop_reg r6
+	pop_reg r7
+	pop_reg r8
+	pop_reg r9
+	pop_reg r10
+	pop_reg r11
+	pop_reg r12, sp, 16
+	/* r13 already restored above */
+	pop_reg r14
+	pop_reg r15
+	pop_reg r16
+	pop_reg r17
+	pop_reg r18
+	pop_reg r19
+	pop_reg r20, sp, 16
+	/* r21 already restored above */
+	pop_reg r22
+	pop_reg r23
+	pop_reg r24
+	pop_reg r25
+	pop_reg r26
+	pop_reg r27
+	pop_reg r28, sp, PTREGS_OFFSET_LR - PTREGS_OFFSET_REG(28)
+	/* r29 already restored above */
+	bnez    lr, .Lkernel_return
+	pop_reg lr, sp, PTREGS_OFFSET_TP - PTREGS_OFFSET_LR
+	pop_reg tp, sp, PTREGS_OFFSET_SP - PTREGS_OFFSET_TP
+	ld      sp, sp
+	iret
+
+	/*
+	 * We can't restore tp when in kernel mode, since a thread might
+	 * have migrated from another cpu and brought a stale tp value.
+	 */
+.Lkernel_return:
+	pop_reg lr, sp, PTREGS_OFFSET_SP - PTREGS_OFFSET_LR
+	ld      sp, sp
+	iret
+
+	/* Restore callee-saved registers from r34 to r51. */
+.Lrestore_callees:
+	addli  sp, sp, PTREGS_OFFSET_REG(34) - PTREGS_OFFSET_REG(29)
+	pop_reg r34
+	pop_reg r35
+	pop_reg r36
+	pop_reg r37
+	pop_reg r38
+	pop_reg r39
+	pop_reg r40
+	pop_reg r41
+	pop_reg r42
+	pop_reg r43
+	pop_reg r44
+	pop_reg r45
+	pop_reg r46
+	pop_reg r47
+	pop_reg r48
+	pop_reg r49
+	pop_reg r50
+	pop_reg r51, sp, PTREGS_OFFSET_REG(29) - PTREGS_OFFSET_REG(51)
+	j .Lcontinue_restore_regs
+	STD_ENDPROC(interrupt_return)
+
+	/*
+	 * "NMI" interrupts mask ALL interrupts before calling the
+	 * handler, and don't check thread flags, etc., on the way
+	 * back out.  In general, the only things we do here for NMIs
+	 * are register save/restore and dataplane kernel-TLB management.
+	 * We don't (for example) deal with start/stop of the sched tick.
+	 */
+	.pushsection .text.handle_nmi,"ax"
+handle_nmi:
+	finish_interrupt_save handle_nmi
+	{
+	 jalr   r0
+	 PTREGS_PTR(r0, PTREGS_OFFSET_BASE)
+	}
+	FEEDBACK_REENTER(handle_nmi)
+	{
+	 movei  r30, 1
+	 cmpeq  r31, r0, zero
+	}
+	j       interrupt_return
+	STD_ENDPROC(handle_nmi)
+
+	/*
+	 * Parallel code for syscalls to handle_interrupt.
+	 */
+	.pushsection .text.handle_syscall,"ax"
+handle_syscall:
+	finish_interrupt_save handle_syscall
+
+	/* Enable irqs. */
+	TRACE_IRQS_ON
+	IRQ_ENABLE(r20, r21)
+
+	/* Bump the counter for syscalls made on this tile. */
+	moveli r20, hw2_last(irq_stat + IRQ_CPUSTAT_SYSCALL_COUNT_OFFSET)
+	shl16insli r20, r20, hw1(irq_stat + IRQ_CPUSTAT_SYSCALL_COUNT_OFFSET)
+	shl16insli r20, r20, hw0(irq_stat + IRQ_CPUSTAT_SYSCALL_COUNT_OFFSET)
+	add     r20, r20, tp
+	ld4s    r21, r20
+	{
+	 addi   r21, r21, 1
+	 move   r31, sp
+	}
+	{
+	 st4    r20, r21
+	 EXTRACT_THREAD_INFO(r31)
+	}
+
+	/* Trace syscalls, if requested. */
+	addi	r31, r31, THREAD_INFO_FLAGS_OFFSET
+	{
+	 ld     r30, r31
+	 moveli r32, _TIF_SYSCALL_ENTRY_WORK
+	}
+	and     r30, r30, r32
+	{
+	 addi   r30, r31, THREAD_INFO_STATUS_OFFSET - THREAD_INFO_FLAGS_OFFSET
+	 beqzt	r30, .Lrestore_syscall_regs
+	}
+	{
+	 PTREGS_PTR(r0, PTREGS_OFFSET_BASE)
+	 jal    do_syscall_trace_enter
+	}
+	FEEDBACK_REENTER(handle_syscall)
+
+	/*
+	 * We always reload our registers from the stack at this
+	 * point.  They might be valid, if we didn't build with
+	 * TRACE_IRQFLAGS, and this isn't a dataplane tile, and we're not
+	 * doing syscall tracing, but there are enough cases now that it
+	 * seems simplest just to do the reload unconditionally.
+	 */
+.Lrestore_syscall_regs:
+	{
+	 ld     r30, r30
+	 PTREGS_PTR(r11, PTREGS_OFFSET_REG(0))
+	}
+	pop_reg r0,  r11
+	pop_reg r1,  r11
+	pop_reg r2,  r11
+	pop_reg r3,  r11
+	pop_reg r4,  r11
+	pop_reg r5,  r11, PTREGS_OFFSET_SYSCALL - PTREGS_OFFSET_REG(5)
+	{
+	 ld     TREG_SYSCALL_NR_NAME, r11
+	 moveli r21, __NR_syscalls
+	}
+
+	/* Ensure that the syscall number is within the legal range. */
+	{
+	 moveli r20, hw2(sys_call_table)
+#ifdef CONFIG_COMPAT
+	 blbs   r30, .Lcompat_syscall
+#endif
+	}
+	{
+	 cmpltu r21, TREG_SYSCALL_NR_NAME, r21
+	 shl16insli r20, r20, hw1(sys_call_table)
+	}
+	{
+	 blbc   r21, .Linvalid_syscall
+	 shl16insli r20, r20, hw0(sys_call_table)
+	}
+.Lload_syscall_pointer:
+	shl3add r20, TREG_SYSCALL_NR_NAME, r20
+	ld      r20, r20
+
+	/* Jump to syscall handler. */
+	jalr    r20
+.Lhandle_syscall_link: /* value of "lr" after "jalr r20" above */
+
+	/*
+	 * Write our r0 onto the stack so it gets restored instead
+	 * of whatever the user had there before.
+	 * In compat mode, sign-extend r0 before storing it.
+	 */
+	{
+	 PTREGS_PTR(r29, PTREGS_OFFSET_REG(0))
+	 blbct  r30, 1f
+	}
+	addxi   r0, r0, 0
+1:	st      r29, r0
+
+.Lsyscall_sigreturn_skip:
+	FEEDBACK_REENTER(handle_syscall)
+
+	/* Do syscall trace again, if requested. */
+	{
+	 ld      r30, r31
+	 moveli  r32, _TIF_SYSCALL_EXIT_WORK
+	}
+	and      r0, r30, r32
+	{
+	 andi    r0, r30, _TIF_SINGLESTEP
+	 beqzt   r0, 1f
+	}
+	{
+	 PTREGS_PTR(r0, PTREGS_OFFSET_BASE)
+	 jal    do_syscall_trace_exit
+	}
+	FEEDBACK_REENTER(handle_syscall)
+	andi    r0, r30, _TIF_SINGLESTEP
+
+1:	beqzt	r0, 2f
+
+	/* Single stepping -- notify ptrace. */
+	{
+	 movei   r0, SIGTRAP
+	 jal     ptrace_notify
+	}
+	FEEDBACK_REENTER(handle_syscall)
+
+2:	{
+	 movei  r30, 0               /* not an NMI */
+	 j      .Lresume_userspace   /* jump into middle of interrupt_return */
+	}
+
+#ifdef CONFIG_COMPAT
+.Lcompat_syscall:
+	/*
+	 * Load the base of the compat syscall table in r20, and
+	 * range-check the syscall number (duplicated from 64-bit path).
+	 * Sign-extend all the user's passed arguments to make them consistent.
+	 * Also save the original "r(n)" values away in "r(11+n)" in
+	 * case the syscall table entry wants to validate them.
+	 */
+	moveli  r20, hw2(compat_sys_call_table)
+	{
+	 cmpltu r21, TREG_SYSCALL_NR_NAME, r21
+	 shl16insli r20, r20, hw1(compat_sys_call_table)
+	}
+	{
+	 blbc   r21, .Linvalid_syscall
+	 shl16insli r20, r20, hw0(compat_sys_call_table)
+	}
+	{ move r11, r0; addxi r0, r0, 0 }
+	{ move r12, r1; addxi r1, r1, 0 }
+	{ move r13, r2; addxi r2, r2, 0 }
+	{ move r14, r3; addxi r3, r3, 0 }
+	{ move r15, r4; addxi r4, r4, 0 }
+	{ move r16, r5; addxi r5, r5, 0 }
+	j .Lload_syscall_pointer
+#endif
+
+.Linvalid_syscall:
+	/* Report an invalid syscall back to the user program */
+	{
+	 PTREGS_PTR(r29, PTREGS_OFFSET_REG(0))
+	 movei  r28, -ENOSYS
+	}
+	st      r29, r28
+	{
+	 movei  r30, 0               /* not an NMI */
+	 j      .Lresume_userspace   /* jump into middle of interrupt_return */
+	}
+	STD_ENDPROC(handle_syscall)
+
+	/* Return the address for oprofile to suppress in backtraces. */
+STD_ENTRY_SECTION(handle_syscall_link_address, .text.handle_syscall)
+	lnk     r0
+	{
+	 addli  r0, r0, .Lhandle_syscall_link - .
+	 jrp    lr
+	}
+	STD_ENDPROC(handle_syscall_link_address)
+
+STD_ENTRY(ret_from_fork)
+	jal     sim_notify_fork
+	jal     schedule_tail
+	FEEDBACK_REENTER(ret_from_fork)
+	{
+	 movei  r30, 0               /* not an NMI */
+	 j      .Lresume_userspace   /* jump into middle of interrupt_return */
+	}
+	STD_ENDPROC(ret_from_fork)
+
+STD_ENTRY(ret_from_kernel_thread)
+	jal     sim_notify_fork
+	jal     schedule_tail
+	FEEDBACK_REENTER(ret_from_fork)
+	{
+	 move   r0, r31
+	 jalr   r30
+	}
+	FEEDBACK_REENTER(ret_from_kernel_thread)
+	{
+	 movei  r30, 0               /* not an NMI */
+	 j      .Lresume_userspace   /* jump into middle of interrupt_return */
+	}
+	STD_ENDPROC(ret_from_kernel_thread)
+
+/* Various stub interrupt handlers and syscall handlers */
+
+STD_ENTRY_LOCAL(_kernel_double_fault)
+	mfspr   r1, SPR_EX_CONTEXT_K_0
+	move    r2, lr
+	move    r3, sp
+	move    r4, r52
+	addi    sp, sp, -C_ABI_SAVE_AREA_SIZE
+	j       kernel_double_fault
+	STD_ENDPROC(_kernel_double_fault)
+
+STD_ENTRY_LOCAL(bad_intr)
+	mfspr   r2, SPR_EX_CONTEXT_K_0
+	panic   "Unhandled interrupt %#x: PC %#lx"
+	STD_ENDPROC(bad_intr)
+
+/*
+ * Special-case sigreturn to not write r0 to the stack on return.
+ * This is technically more efficient, but it also avoids difficulties
+ * in the 64-bit OS when handling 32-bit compat code, since we must not
+ * sign-extend r0 for the sigreturn return-value case.
+ */
+#define PTREGS_SYSCALL_SIGRETURN(x, reg)                \
+	STD_ENTRY(_##x);                                \
+	addli   lr, lr, .Lsyscall_sigreturn_skip - .Lhandle_syscall_link; \
+	{                                               \
+	 PTREGS_PTR(reg, PTREGS_OFFSET_BASE);           \
+	 j      x                                       \
+	};                                              \
+	STD_ENDPROC(_##x)
+
+PTREGS_SYSCALL_SIGRETURN(sys_rt_sigreturn, r0)
+#ifdef CONFIG_COMPAT
+PTREGS_SYSCALL_SIGRETURN(compat_sys_rt_sigreturn, r0)
+#endif
+
+/* Save additional callee-saves to pt_regs and jump to standard function. */
+STD_ENTRY(_sys_clone)
+	push_extra_callee_saves r4
+	j       sys_clone
+	STD_ENDPROC(_sys_clone)
+
+	/*
+	 * Recover r3, r2, r1 and r0 here saved by unalign fast vector.
+	 * The vector area limit is 32 bundles, so we handle the reload here.
+	 * r0, r1, r2 are in thread_info from low to high memory in order.
+	 * r3 points to location the original r3 was saved.
+	 * We put this code in the __HEAD section so it can be reached
+	 * via a conditional branch from the fast path.
+	 */
+	__HEAD
+hand_unalign_slow:
+	andi    sp, sp, ~1
+hand_unalign_slow_badsp:
+	addi    r3, r3, -(3 * 8)
+	ld_add  r0, r3, 8
+	ld_add  r1, r3, 8
+	ld      r2, r3
+hand_unalign_slow_nonuser:
+	mfspr   r3, SPR_SYSTEM_SAVE_K_1
+	__int_hand     INT_UNALIGN_DATA, UNALIGN_DATA_SLOW, int_unalign
+
+/* The unaligned data support needs to read all the registers. */
+int_unalign:
+	push_extra_callee_saves r0
+	j       do_unaligned
+ENDPROC(hand_unalign_slow)
+
+/* Fill the return address stack with nonzero entries. */
+STD_ENTRY(fill_ra_stack)
+	{
+	 move	r0, lr
+	 jal	1f
+	}
+1:	jal	2f
+2:	jal	3f
+3:	jal	4f
+4:	jrp	r0
+	STD_ENDPROC(fill_ra_stack)
+
+	.macro int_hand  vecnum, vecname, c_routine, processing=handle_interrupt
+	.org   (\vecnum << 8)
+		__int_hand   \vecnum, \vecname, \c_routine, \processing
+	.endm
+
+/* Include .intrpt array of interrupt vectors */
+	.section ".intrpt", "ax"
+	.global intrpt_start
+intrpt_start:
+
+#ifndef CONFIG_USE_PMC
+#define handle_perf_interrupt bad_intr
+#endif
+
+#ifndef CONFIG_HARDWALL
+#define do_hardwall_trap bad_intr
+#endif
+
+	int_hand     INT_MEM_ERROR, MEM_ERROR, do_trap
+	int_hand     INT_SINGLE_STEP_3, SINGLE_STEP_3, bad_intr
+#if CONFIG_KERNEL_PL == 2
+	int_hand     INT_SINGLE_STEP_2, SINGLE_STEP_2, gx_singlestep_handle
+	int_hand     INT_SINGLE_STEP_1, SINGLE_STEP_1, bad_intr
+#else
+	int_hand     INT_SINGLE_STEP_2, SINGLE_STEP_2, bad_intr
+	int_hand     INT_SINGLE_STEP_1, SINGLE_STEP_1, gx_singlestep_handle
+#endif
+	int_hand     INT_SINGLE_STEP_0, SINGLE_STEP_0, bad_intr
+	int_hand     INT_IDN_COMPLETE, IDN_COMPLETE, bad_intr
+	int_hand     INT_UDN_COMPLETE, UDN_COMPLETE, bad_intr
+	int_hand     INT_ITLB_MISS, ITLB_MISS, do_page_fault
+	int_hand     INT_ILL, ILL, do_trap
+	int_hand     INT_GPV, GPV, do_trap
+	int_hand     INT_IDN_ACCESS, IDN_ACCESS, do_trap
+	int_hand     INT_UDN_ACCESS, UDN_ACCESS, do_trap
+	int_hand     INT_SWINT_3, SWINT_3, do_trap
+	int_hand     INT_SWINT_2, SWINT_2, do_trap
+	int_hand     INT_SWINT_1, SWINT_1, SYSCALL, handle_syscall
+	int_hand     INT_SWINT_0, SWINT_0, do_trap
+	int_hand     INT_ILL_TRANS, ILL_TRANS, do_trap
+	int_hand_unalign_fast INT_UNALIGN_DATA, UNALIGN_DATA
+	int_hand     INT_DTLB_MISS, DTLB_MISS, do_page_fault
+	int_hand     INT_DTLB_ACCESS, DTLB_ACCESS, do_page_fault
+	int_hand     INT_IDN_FIREWALL, IDN_FIREWALL, do_hardwall_trap
+	int_hand     INT_UDN_FIREWALL, UDN_FIREWALL, do_hardwall_trap
+	int_hand     INT_TILE_TIMER, TILE_TIMER, do_timer_interrupt
+	int_hand     INT_IDN_TIMER, IDN_TIMER, bad_intr
+	int_hand     INT_UDN_TIMER, UDN_TIMER, bad_intr
+	int_hand     INT_IDN_AVAIL, IDN_AVAIL, bad_intr
+	int_hand     INT_UDN_AVAIL, UDN_AVAIL, bad_intr
+	int_hand     INT_IPI_3, IPI_3, bad_intr
+#if CONFIG_KERNEL_PL == 2
+	int_hand     INT_IPI_2, IPI_2, tile_dev_intr
+	int_hand     INT_IPI_1, IPI_1, bad_intr
+#else
+	int_hand     INT_IPI_2, IPI_2, bad_intr
+	int_hand     INT_IPI_1, IPI_1, tile_dev_intr
+#endif
+	int_hand     INT_IPI_0, IPI_0, bad_intr
+	int_hand     INT_PERF_COUNT, PERF_COUNT, \
+		     handle_perf_interrupt, handle_nmi
+	int_hand     INT_AUX_PERF_COUNT, AUX_PERF_COUNT, \
+		     handle_perf_interrupt, handle_nmi
+	int_hand     INT_INTCTRL_3, INTCTRL_3, bad_intr
+#if CONFIG_KERNEL_PL == 2
+	dc_dispatch  INT_INTCTRL_2, INTCTRL_2
+	int_hand     INT_INTCTRL_1, INTCTRL_1, bad_intr
+#else
+	int_hand     INT_INTCTRL_2, INTCTRL_2, bad_intr
+	dc_dispatch  INT_INTCTRL_1, INTCTRL_1
+#endif
+	int_hand     INT_INTCTRL_0, INTCTRL_0, bad_intr
+	int_hand     INT_MESSAGE_RCV_DWNCL, MESSAGE_RCV_DWNCL, \
+		     hv_message_intr
+	int_hand     INT_DEV_INTR_DWNCL, DEV_INTR_DWNCL, bad_intr
+	int_hand     INT_I_ASID, I_ASID, bad_intr
+	int_hand     INT_D_ASID, D_ASID, bad_intr
+	int_hand     INT_DOUBLE_FAULT, DOUBLE_FAULT, do_trap
+
+	/* Synthetic interrupt delivered only by the simulator */
+	int_hand     INT_BREAKPOINT, BREAKPOINT, do_breakpoint
diff --git a/kernel/arch/tile/kernel/irq.c b/kernel/arch/tile/kernel/irq.c
new file mode 100644
index 000000000..22044fc69
--- /dev/null
+++ b/kernel/arch/tile/kernel/irq.c
@@ -0,0 +1,280 @@
+/*
+ * Copyright 2010 Tilera Corporation. All Rights Reserved.
+ *
+ *   This program is free software; you can redistribute it and/or
+ *   modify it under the terms of the GNU General Public License
+ *   as published by the Free Software Foundation, version 2.
+ *
+ *   This program is distributed in the hope that it will be useful, but
+ *   WITHOUT ANY WARRANTY; without even the implied warranty of
+ *   MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
+ *   NON INFRINGEMENT.  See the GNU General Public License for
+ *   more details.
+ */
+
+#include <linux/module.h>
+#include <linux/seq_file.h>
+#include <linux/interrupt.h>
+#include <linux/irq.h>
+#include <linux/kernel_stat.h>
+#include <linux/uaccess.h>
+#include <hv/drv_pcie_rc_intf.h>
+#include <arch/spr_def.h>
+#include <asm/traps.h>
+#include <linux/perf_event.h>
+
+/* Bit-flag stored in irq_desc->chip_data to indicate HW-cleared irqs. */
+#define IS_HW_CLEARED 1
+
+/*
+ * The set of interrupts we enable for arch_local_irq_enable().
+ * This is initialized to have just a single interrupt that the kernel
+ * doesn't actually use as a sentinel.  During kernel init,
+ * interrupts are added as the kernel gets prepared to support them.
+ * NOTE: we could probably initialize them all statically up front.
+ */
+DEFINE_PER_CPU(unsigned long long, interrupts_enabled_mask) =
+  INITIAL_INTERRUPTS_ENABLED;
+EXPORT_PER_CPU_SYMBOL(interrupts_enabled_mask);
+
+/* Define per-tile device interrupt statistics state. */
+DEFINE_PER_CPU(irq_cpustat_t, irq_stat) ____cacheline_internodealigned_in_smp;
+EXPORT_PER_CPU_SYMBOL(irq_stat);
+
+/*
+ * Define per-tile irq disable mask; the hardware/HV only has a single
+ * mask that we use to implement both masking and disabling.
+ */
+static DEFINE_PER_CPU(unsigned long, irq_disable_mask)
+	____cacheline_internodealigned_in_smp;
+
+/*
+ * Per-tile IRQ nesting depth.  Used to make sure we enable newly
+ * enabled IRQs before exiting the outermost interrupt.
+ */
+static DEFINE_PER_CPU(int, irq_depth);
+
+#if CHIP_HAS_IPI()
+/* Use SPRs to manipulate device interrupts. */
+#define mask_irqs(irq_mask) __insn_mtspr(SPR_IPI_MASK_SET_K, irq_mask)
+#define unmask_irqs(irq_mask) __insn_mtspr(SPR_IPI_MASK_RESET_K, irq_mask)
+#define clear_irqs(irq_mask) __insn_mtspr(SPR_IPI_EVENT_RESET_K, irq_mask)
+#else
+/* Use HV to manipulate device interrupts. */
+#define mask_irqs(irq_mask) hv_disable_intr(irq_mask)
+#define unmask_irqs(irq_mask) hv_enable_intr(irq_mask)
+#define clear_irqs(irq_mask) hv_clear_intr(irq_mask)
+#endif
+
+/*
+ * The interrupt handling path, implemented in terms of HV interrupt
+ * emulation on TILEPro, and IPI hardware on TILE-Gx.
+ * Entered with interrupts disabled.
+ */
+void tile_dev_intr(struct pt_regs *regs, int intnum)
+{
+	int depth = __this_cpu_inc_return(irq_depth);
+	unsigned long original_irqs;
+	unsigned long remaining_irqs;
+	struct pt_regs *old_regs;
+
+#if CHIP_HAS_IPI()
+	/*
+	 * Pending interrupts are listed in an SPR.  We might be
+	 * nested, so be sure to only handle irqs that weren't already
+	 * masked by a previous interrupt.  Then, mask out the ones
+	 * we're going to handle.
+	 */
+	unsigned long masked = __insn_mfspr(SPR_IPI_MASK_K);
+	original_irqs = __insn_mfspr(SPR_IPI_EVENT_K) & ~masked;
+	__insn_mtspr(SPR_IPI_MASK_SET_K, original_irqs);
+#else
+	/*
+	 * Hypervisor performs the equivalent of the Gx code above and
+	 * then puts the pending interrupt mask into a system save reg
+	 * for us to find.
+	 */
+	original_irqs = __insn_mfspr(SPR_SYSTEM_SAVE_K_3);
+#endif
+	remaining_irqs = original_irqs;
+
+	/* Track time spent here in an interrupt context. */
+	old_regs = set_irq_regs(regs);
+	irq_enter();
+
+#ifdef CONFIG_DEBUG_STACKOVERFLOW
+	/* Debugging check for stack overflow: less than 1/8th stack free? */
+	{
+		long sp = stack_pointer - (long) current_thread_info();
+		if (unlikely(sp < (sizeof(struct thread_info) + STACK_WARN))) {
+			pr_emerg("%s: stack overflow: %ld\n",
+				 __func__, sp - sizeof(struct thread_info));
+			dump_stack();
+		}
+	}
+#endif
+	while (remaining_irqs) {
+		unsigned long irq = __ffs(remaining_irqs);
+		remaining_irqs &= ~(1UL << irq);
+
+		/* Count device irqs; Linux IPIs are counted elsewhere. */
+		if (irq != IRQ_RESCHEDULE)
+			__this_cpu_inc(irq_stat.irq_dev_intr_count);
+
+		generic_handle_irq(irq);
+	}
+
+	/*
+	 * If we weren't nested, turn on all enabled interrupts,
+	 * including any that were reenabled during interrupt
+	 * handling.
+	 */
+	if (depth == 1)
+		unmask_irqs(~__this_cpu_read(irq_disable_mask));
+
+	__this_cpu_dec(irq_depth);
+
+	/*
+	 * Track time spent against the current process again and
+	 * process any softirqs if they are waiting.
+	 */
+	irq_exit();
+	set_irq_regs(old_regs);
+}
+
+
+/*
+ * Remove an irq from the disabled mask.  If we're in an interrupt
+ * context, defer enabling the HW interrupt until we leave.
+ */
+static void tile_irq_chip_enable(struct irq_data *d)
+{
+	get_cpu_var(irq_disable_mask) &= ~(1UL << d->irq);
+	if (__this_cpu_read(irq_depth) == 0)
+		unmask_irqs(1UL << d->irq);
+	put_cpu_var(irq_disable_mask);
+}
+
+/*
+ * Add an irq to the disabled mask.  We disable the HW interrupt
+ * immediately so that there's no possibility of it firing.  If we're
+ * in an interrupt context, the return path is careful to avoid
+ * unmasking a newly disabled interrupt.
+ */
+static void tile_irq_chip_disable(struct irq_data *d)
+{
+	get_cpu_var(irq_disable_mask) |= (1UL << d->irq);
+	mask_irqs(1UL << d->irq);
+	put_cpu_var(irq_disable_mask);
+}
+
+/* Mask an interrupt. */
+static void tile_irq_chip_mask(struct irq_data *d)
+{
+	mask_irqs(1UL << d->irq);
+}
+
+/* Unmask an interrupt. */
+static void tile_irq_chip_unmask(struct irq_data *d)
+{
+	unmask_irqs(1UL << d->irq);
+}
+
+/*
+ * Clear an interrupt before processing it so that any new assertions
+ * will trigger another irq.
+ */
+static void tile_irq_chip_ack(struct irq_data *d)
+{
+	if ((unsigned long)irq_data_get_irq_chip_data(d) != IS_HW_CLEARED)
+		clear_irqs(1UL << d->irq);
+}
+
+/*
+ * For per-cpu interrupts, we need to avoid unmasking any interrupts
+ * that we disabled via disable_percpu_irq().
+ */
+static void tile_irq_chip_eoi(struct irq_data *d)
+{
+	if (!(__this_cpu_read(irq_disable_mask) & (1UL << d->irq)))
+		unmask_irqs(1UL << d->irq);
+}
+
+static struct irq_chip tile_irq_chip = {
+	.name = "tile_irq_chip",
+	.irq_enable = tile_irq_chip_enable,
+	.irq_disable = tile_irq_chip_disable,
+	.irq_ack = tile_irq_chip_ack,
+	.irq_eoi = tile_irq_chip_eoi,
+	.irq_mask = tile_irq_chip_mask,
+	.irq_unmask = tile_irq_chip_unmask,
+};
+
+void __init init_IRQ(void)
+{
+	ipi_init();
+}
+
+void setup_irq_regs(void)
+{
+	/* Enable interrupt delivery. */
+	unmask_irqs(~0UL);
+#if CHIP_HAS_IPI()
+	arch_local_irq_unmask(INT_IPI_K);
+#endif
+}
+
+void tile_irq_activate(unsigned int irq, int tile_irq_type)
+{
+	/*
+	 * We use handle_level_irq() by default because the pending
+	 * interrupt vector (whether modeled by the HV on
+	 * TILEPro or implemented in hardware on TILE-Gx) has
+	 * level-style semantics for each bit.  An interrupt fires
+	 * whenever a bit is high, not just at edges.
+	 */
+	irq_flow_handler_t handle = handle_level_irq;
+	if (tile_irq_type == TILE_IRQ_PERCPU)
+		handle = handle_percpu_irq;
+	irq_set_chip_and_handler(irq, &tile_irq_chip, handle);
+
+	/*
+	 * Flag interrupts that are hardware-cleared so that ack()
+	 * won't clear them.
+	 */
+	if (tile_irq_type == TILE_IRQ_HW_CLEAR)
+		irq_set_chip_data(irq, (void *)IS_HW_CLEARED);
+}
+EXPORT_SYMBOL(tile_irq_activate);
+
+
+void ack_bad_irq(unsigned int irq)
+{
+	pr_err("unexpected IRQ trap at vector %02x\n", irq);
+}
+
+/*
+ * /proc/interrupts printing:
+ */
+int arch_show_interrupts(struct seq_file *p, int prec)
+{
+#ifdef CONFIG_PERF_EVENTS
+	int i;
+
+	seq_printf(p, "%*s: ", prec, "PMI");
+
+	for_each_online_cpu(i)
+		seq_printf(p, "%10llu ", per_cpu(perf_irqs, i));
+	seq_puts(p, "  perf_events\n");
+#endif
+	return 0;
+}
+
+#if CHIP_HAS_IPI()
+int arch_setup_hwirq(unsigned int irq, int node)
+{
+	return irq >= NR_IRQS ? -EINVAL : 0;
+}
+
+void arch_teardown_hwirq(unsigned int irq) { }
+#endif
diff --git a/kernel/arch/tile/kernel/kgdb.c b/kernel/arch/tile/kernel/kgdb.c
new file mode 100644
index 000000000..ff5335ae0
--- /dev/null
+++ b/kernel/arch/tile/kernel/kgdb.c
@@ -0,0 +1,495 @@
+/*
+ * Copyright 2013 Tilera Corporation. All Rights Reserved.
+ *
+ *   This program is free software; you can redistribute it and/or
+ *   modify it under the terms of the GNU General Public License
+ *   as published by the Free Software Foundation, version 2.
+ *
+ *   This program is distributed in the hope that it will be useful, but
+ *   WITHOUT ANY WARRANTY; without even the implied warranty of
+ *   MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
+ *   NON INFRINGEMENT.  See the GNU General Public License for
+ *   more details.
+ *
+ * TILE-Gx KGDB support.
+ */
+
+#include <linux/ptrace.h>
+#include <linux/kgdb.h>
+#include <linux/kdebug.h>
+#include <linux/uaccess.h>
+#include <linux/module.h>
+#include <asm/cacheflush.h>
+
+static tile_bundle_bits singlestep_insn = TILEGX_BPT_BUNDLE | DIE_SSTEPBP;
+static unsigned long stepped_addr;
+static tile_bundle_bits stepped_instr;
+
+struct dbg_reg_def_t dbg_reg_def[DBG_MAX_REG_NUM] = {
+	{ "r0", GDB_SIZEOF_REG, offsetof(struct pt_regs, regs[0])},
+	{ "r1", GDB_SIZEOF_REG, offsetof(struct pt_regs, regs[1])},
+	{ "r2", GDB_SIZEOF_REG, offsetof(struct pt_regs, regs[2])},
+	{ "r3", GDB_SIZEOF_REG, offsetof(struct pt_regs, regs[3])},
+	{ "r4", GDB_SIZEOF_REG, offsetof(struct pt_regs, regs[4])},
+	{ "r5", GDB_SIZEOF_REG, offsetof(struct pt_regs, regs[5])},
+	{ "r6", GDB_SIZEOF_REG, offsetof(struct pt_regs, regs[6])},
+	{ "r7", GDB_SIZEOF_REG, offsetof(struct pt_regs, regs[7])},
+	{ "r8", GDB_SIZEOF_REG, offsetof(struct pt_regs, regs[8])},
+	{ "r9", GDB_SIZEOF_REG, offsetof(struct pt_regs, regs[9])},
+	{ "r10", GDB_SIZEOF_REG, offsetof(struct pt_regs, regs[10])},
+	{ "r11", GDB_SIZEOF_REG, offsetof(struct pt_regs, regs[11])},
+	{ "r12", GDB_SIZEOF_REG, offsetof(struct pt_regs, regs[12])},
+	{ "r13", GDB_SIZEOF_REG, offsetof(struct pt_regs, regs[13])},
+	{ "r14", GDB_SIZEOF_REG, offsetof(struct pt_regs, regs[14])},
+	{ "r15", GDB_SIZEOF_REG, offsetof(struct pt_regs, regs[15])},
+	{ "r16", GDB_SIZEOF_REG, offsetof(struct pt_regs, regs[16])},
+	{ "r17", GDB_SIZEOF_REG, offsetof(struct pt_regs, regs[17])},
+	{ "r18", GDB_SIZEOF_REG, offsetof(struct pt_regs, regs[18])},
+	{ "r19", GDB_SIZEOF_REG, offsetof(struct pt_regs, regs[19])},
+	{ "r20", GDB_SIZEOF_REG, offsetof(struct pt_regs, regs[20])},
+	{ "r21", GDB_SIZEOF_REG, offsetof(struct pt_regs, regs[21])},
+	{ "r22", GDB_SIZEOF_REG, offsetof(struct pt_regs, regs[22])},
+	{ "r23", GDB_SIZEOF_REG, offsetof(struct pt_regs, regs[23])},
+	{ "r24", GDB_SIZEOF_REG, offsetof(struct pt_regs, regs[24])},
+	{ "r25", GDB_SIZEOF_REG, offsetof(struct pt_regs, regs[25])},
+	{ "r26", GDB_SIZEOF_REG, offsetof(struct pt_regs, regs[26])},
+	{ "r27", GDB_SIZEOF_REG, offsetof(struct pt_regs, regs[27])},
+	{ "r28", GDB_SIZEOF_REG, offsetof(struct pt_regs, regs[28])},
+	{ "r29", GDB_SIZEOF_REG, offsetof(struct pt_regs, regs[29])},
+	{ "r30", GDB_SIZEOF_REG, offsetof(struct pt_regs, regs[30])},
+	{ "r31", GDB_SIZEOF_REG, offsetof(struct pt_regs, regs[31])},
+	{ "r32", GDB_SIZEOF_REG, offsetof(struct pt_regs, regs[32])},
+	{ "r33", GDB_SIZEOF_REG, offsetof(struct pt_regs, regs[33])},
+	{ "r34", GDB_SIZEOF_REG, offsetof(struct pt_regs, regs[34])},
+	{ "r35", GDB_SIZEOF_REG, offsetof(struct pt_regs, regs[35])},
+	{ "r36", GDB_SIZEOF_REG, offsetof(struct pt_regs, regs[36])},
+	{ "r37", GDB_SIZEOF_REG, offsetof(struct pt_regs, regs[37])},
+	{ "r38", GDB_SIZEOF_REG, offsetof(struct pt_regs, regs[38])},
+	{ "r39", GDB_SIZEOF_REG, offsetof(struct pt_regs, regs[39])},
+	{ "r40", GDB_SIZEOF_REG, offsetof(struct pt_regs, regs[40])},
+	{ "r41", GDB_SIZEOF_REG, offsetof(struct pt_regs, regs[41])},
+	{ "r42", GDB_SIZEOF_REG, offsetof(struct pt_regs, regs[42])},
+	{ "r43", GDB_SIZEOF_REG, offsetof(struct pt_regs, regs[43])},
+	{ "r44", GDB_SIZEOF_REG, offsetof(struct pt_regs, regs[44])},
+	{ "r45", GDB_SIZEOF_REG, offsetof(struct pt_regs, regs[45])},
+	{ "r46", GDB_SIZEOF_REG, offsetof(struct pt_regs, regs[46])},
+	{ "r47", GDB_SIZEOF_REG, offsetof(struct pt_regs, regs[47])},
+	{ "r48", GDB_SIZEOF_REG, offsetof(struct pt_regs, regs[48])},
+	{ "r49", GDB_SIZEOF_REG, offsetof(struct pt_regs, regs[49])},
+	{ "r50", GDB_SIZEOF_REG, offsetof(struct pt_regs, regs[50])},
+	{ "r51", GDB_SIZEOF_REG, offsetof(struct pt_regs, regs[51])},
+	{ "r52", GDB_SIZEOF_REG, offsetof(struct pt_regs, regs[52])},
+	{ "tp", GDB_SIZEOF_REG, offsetof(struct pt_regs, tp)},
+	{ "sp", GDB_SIZEOF_REG, offsetof(struct pt_regs, sp)},
+	{ "lr", GDB_SIZEOF_REG, offsetof(struct pt_regs, lr)},
+	{ "sn", GDB_SIZEOF_REG, -1},
+	{ "idn0", GDB_SIZEOF_REG, -1},
+	{ "idn1", GDB_SIZEOF_REG, -1},
+	{ "udn0", GDB_SIZEOF_REG, -1},
+	{ "udn1", GDB_SIZEOF_REG, -1},
+	{ "udn2", GDB_SIZEOF_REG, -1},
+	{ "udn3", GDB_SIZEOF_REG, -1},
+	{ "zero", GDB_SIZEOF_REG, -1},
+	{ "pc", GDB_SIZEOF_REG, offsetof(struct pt_regs, pc)},
+	{ "faultnum", GDB_SIZEOF_REG, offsetof(struct pt_regs, faultnum)},
+};
+
+char *dbg_get_reg(int regno, void *mem, struct pt_regs *regs)
+{
+	if (regno >= DBG_MAX_REG_NUM || regno < 0)
+		return NULL;
+
+	if (dbg_reg_def[regno].offset != -1)
+		memcpy(mem, (void *)regs + dbg_reg_def[regno].offset,
+		       dbg_reg_def[regno].size);
+	else
+		memset(mem, 0, dbg_reg_def[regno].size);
+	return dbg_reg_def[regno].name;
+}
+
+int dbg_set_reg(int regno, void *mem, struct pt_regs *regs)
+{
+	if (regno >= DBG_MAX_REG_NUM || regno < 0)
+		return -EINVAL;
+
+	if (dbg_reg_def[regno].offset != -1)
+		memcpy((void *)regs + dbg_reg_def[regno].offset, mem,
+		       dbg_reg_def[regno].size);
+	return 0;
+}
+
+/*
+ * Similar to pt_regs_to_gdb_regs() except that process is sleeping and so
+ * we may not be able to get all the info.
+ */
+void
+sleeping_thread_to_gdb_regs(unsigned long *gdb_regs, struct task_struct *task)
+{
+	struct pt_regs *thread_regs;
+
+	if (task == NULL)
+		return;
+
+	/* Initialize to zero. */
+	memset(gdb_regs, 0, NUMREGBYTES);
+
+	thread_regs = task_pt_regs(task);
+	memcpy(gdb_regs, thread_regs, TREG_LAST_GPR * sizeof(unsigned long));
+	gdb_regs[TILEGX_PC_REGNUM] = thread_regs->pc;
+	gdb_regs[TILEGX_FAULTNUM_REGNUM] = thread_regs->faultnum;
+}
+
+void kgdb_arch_set_pc(struct pt_regs *regs, unsigned long pc)
+{
+	regs->pc = pc;
+}
+
+static void kgdb_call_nmi_hook(void *ignored)
+{
+	kgdb_nmicallback(raw_smp_processor_id(), NULL);
+}
+
+void kgdb_roundup_cpus(unsigned long flags)
+{
+	local_irq_enable();
+	smp_call_function(kgdb_call_nmi_hook, NULL, 0);
+	local_irq_disable();
+}
+
+/*
+ * Convert a kernel address to the writable kernel text mapping.
+ */
+static unsigned long writable_address(unsigned long addr)
+{
+	unsigned long ret = 0;
+
+	if (core_kernel_text(addr))
+		ret = addr - MEM_SV_START + PAGE_OFFSET;
+	else if (is_module_text_address(addr))
+		ret = addr;
+	else
+		pr_err("Unknown virtual address 0x%lx\n", addr);
+
+	return ret;
+}
+
+/*
+ * Calculate the new address for after a step.
+ */
+static unsigned long get_step_address(struct pt_regs *regs)
+{
+	int src_reg;
+	int jump_off;
+	int br_off;
+	unsigned long addr;
+	unsigned int opcode;
+	tile_bundle_bits bundle;
+
+	/* Move to the next instruction by default. */
+	addr = regs->pc + TILEGX_BUNDLE_SIZE_IN_BYTES;
+	bundle = *(unsigned long *)instruction_pointer(regs);
+
+	/* 0: X mode, Otherwise: Y mode. */
+	if (bundle & TILEGX_BUNDLE_MODE_MASK) {
+		if (get_Opcode_Y1(bundle) == RRR_1_OPCODE_Y1 &&
+		    get_RRROpcodeExtension_Y1(bundle) ==
+		    UNARY_RRR_1_OPCODE_Y1) {
+			opcode = get_UnaryOpcodeExtension_Y1(bundle);
+
+			switch (opcode) {
+			case JALR_UNARY_OPCODE_Y1:
+			case JALRP_UNARY_OPCODE_Y1:
+			case JR_UNARY_OPCODE_Y1:
+			case JRP_UNARY_OPCODE_Y1:
+				src_reg = get_SrcA_Y1(bundle);
+				dbg_get_reg(src_reg, &addr, regs);
+				break;
+			}
+		}
+	} else if (get_Opcode_X1(bundle) == RRR_0_OPCODE_X1) {
+		if (get_RRROpcodeExtension_X1(bundle) ==
+		    UNARY_RRR_0_OPCODE_X1) {
+			opcode = get_UnaryOpcodeExtension_X1(bundle);
+
+			switch (opcode) {
+			case JALR_UNARY_OPCODE_X1:
+			case JALRP_UNARY_OPCODE_X1:
+			case JR_UNARY_OPCODE_X1:
+			case JRP_UNARY_OPCODE_X1:
+				src_reg = get_SrcA_X1(bundle);
+				dbg_get_reg(src_reg, &addr, regs);
+				break;
+			}
+		}
+	} else if (get_Opcode_X1(bundle) == JUMP_OPCODE_X1) {
+		opcode = get_JumpOpcodeExtension_X1(bundle);
+
+		switch (opcode) {
+		case JAL_JUMP_OPCODE_X1:
+		case J_JUMP_OPCODE_X1:
+			jump_off = sign_extend(get_JumpOff_X1(bundle), 27);
+			addr = regs->pc +
+				(jump_off << TILEGX_LOG2_BUNDLE_SIZE_IN_BYTES);
+			break;
+		}
+	} else if (get_Opcode_X1(bundle) == BRANCH_OPCODE_X1) {
+		br_off = 0;
+		opcode = get_BrType_X1(bundle);
+
+		switch (opcode) {
+		case BEQZT_BRANCH_OPCODE_X1:
+		case BEQZ_BRANCH_OPCODE_X1:
+			if (get_SrcA_X1(bundle) == 0)
+				br_off = get_BrOff_X1(bundle);
+			break;
+		case BGEZT_BRANCH_OPCODE_X1:
+		case BGEZ_BRANCH_OPCODE_X1:
+			if (get_SrcA_X1(bundle) >= 0)
+				br_off = get_BrOff_X1(bundle);
+			break;
+		case BGTZT_BRANCH_OPCODE_X1:
+		case BGTZ_BRANCH_OPCODE_X1:
+			if (get_SrcA_X1(bundle) > 0)
+				br_off = get_BrOff_X1(bundle);
+			break;
+		case BLBCT_BRANCH_OPCODE_X1:
+		case BLBC_BRANCH_OPCODE_X1:
+			if (!(get_SrcA_X1(bundle) & 1))
+				br_off = get_BrOff_X1(bundle);
+			break;
+		case BLBST_BRANCH_OPCODE_X1:
+		case BLBS_BRANCH_OPCODE_X1:
+			if (get_SrcA_X1(bundle) & 1)
+				br_off = get_BrOff_X1(bundle);
+			break;
+		case BLEZT_BRANCH_OPCODE_X1:
+		case BLEZ_BRANCH_OPCODE_X1:
+			if (get_SrcA_X1(bundle) <= 0)
+				br_off = get_BrOff_X1(bundle);
+			break;
+		case BLTZT_BRANCH_OPCODE_X1:
+		case BLTZ_BRANCH_OPCODE_X1:
+			if (get_SrcA_X1(bundle) < 0)
+				br_off = get_BrOff_X1(bundle);
+			break;
+		case BNEZT_BRANCH_OPCODE_X1:
+		case BNEZ_BRANCH_OPCODE_X1:
+			if (get_SrcA_X1(bundle) != 0)
+				br_off = get_BrOff_X1(bundle);
+			break;
+		}
+
+		if (br_off != 0) {
+			br_off = sign_extend(br_off, 17);
+			addr = regs->pc +
+				(br_off << TILEGX_LOG2_BUNDLE_SIZE_IN_BYTES);
+		}
+	}
+
+	return addr;
+}
+
+/*
+ * Replace the next instruction after the current instruction with a
+ * breakpoint instruction.
+ */
+static void do_single_step(struct pt_regs *regs)
+{
+	unsigned long addr_wr;
+
+	/* Determine where the target instruction will send us to. */
+	stepped_addr = get_step_address(regs);
+	probe_kernel_read((char *)&stepped_instr, (char *)stepped_addr,
+			  BREAK_INSTR_SIZE);
+
+	addr_wr = writable_address(stepped_addr);
+	probe_kernel_write((char *)addr_wr, (char *)&singlestep_insn,
+			   BREAK_INSTR_SIZE);
+	smp_wmb();
+	flush_icache_range(stepped_addr, stepped_addr + BREAK_INSTR_SIZE);
+}
+
+static void undo_single_step(struct pt_regs *regs)
+{
+	unsigned long addr_wr;
+
+	if (stepped_instr == 0)
+		return;
+
+	addr_wr = writable_address(stepped_addr);
+	probe_kernel_write((char *)addr_wr, (char *)&stepped_instr,
+			   BREAK_INSTR_SIZE);
+	stepped_instr = 0;
+	smp_wmb();
+	flush_icache_range(stepped_addr, stepped_addr + BREAK_INSTR_SIZE);
+}
+
+/*
+ * Calls linux_debug_hook before the kernel dies. If KGDB is enabled,
+ * then try to fall into the debugger.
+ */
+static int
+kgdb_notify(struct notifier_block *self, unsigned long cmd, void *ptr)
+{
+	int ret;
+	unsigned long flags;
+	struct die_args *args = (struct die_args *)ptr;
+	struct pt_regs *regs = args->regs;
+
+#ifdef CONFIG_KPROBES
+	/*
+	 * Return immediately if the kprobes fault notifier has set
+	 * DIE_PAGE_FAULT.
+	 */
+	if (cmd == DIE_PAGE_FAULT)
+		return NOTIFY_DONE;
+#endif /* CONFIG_KPROBES */
+
+	switch (cmd) {
+	case DIE_BREAK:
+	case DIE_COMPILED_BPT:
+		break;
+	case DIE_SSTEPBP:
+		local_irq_save(flags);
+		kgdb_handle_exception(0, SIGTRAP, 0, regs);
+		local_irq_restore(flags);
+		return NOTIFY_STOP;
+	default:
+		/* Userspace events, ignore. */
+		if (user_mode(regs))
+			return NOTIFY_DONE;
+	}
+
+	local_irq_save(flags);
+	ret = kgdb_handle_exception(args->trapnr, args->signr, args->err, regs);
+	local_irq_restore(flags);
+	if (ret)
+		return NOTIFY_DONE;
+
+	return NOTIFY_STOP;
+}
+
+static struct notifier_block kgdb_notifier = {
+	.notifier_call = kgdb_notify,
+};
+
+/*
+ * kgdb_arch_handle_exception - Handle architecture specific GDB packets.
+ * @vector: The error vector of the exception that happened.
+ * @signo: The signal number of the exception that happened.
+ * @err_code: The error code of the exception that happened.
+ * @remcom_in_buffer: The buffer of the packet we have read.
+ * @remcom_out_buffer: The buffer of %BUFMAX bytes to write a packet into.
+ * @regs: The &struct pt_regs of the current process.
+ *
+ * This function MUST handle the 'c' and 's' command packets,
+ * as well packets to set / remove a hardware breakpoint, if used.
+ * If there are additional packets which the hardware needs to handle,
+ * they are handled here. The code should return -1 if it wants to
+ * process more packets, and a %0 or %1 if it wants to exit from the
+ * kgdb callback.
+ */
+int kgdb_arch_handle_exception(int vector, int signo, int err_code,
+			       char *remcom_in_buffer, char *remcom_out_buffer,
+			       struct pt_regs *regs)
+{
+	char *ptr;
+	unsigned long address;
+
+	/* Undo any stepping we may have done. */
+	undo_single_step(regs);
+
+	switch (remcom_in_buffer[0]) {
+	case 'c':
+	case 's':
+	case 'D':
+	case 'k':
+		/*
+		 * Try to read optional parameter, pc unchanged if no parm.
+		 * If this was a compiled-in breakpoint, we need to move
+		 * to the next instruction or we will just breakpoint
+		 * over and over again.
+		 */
+		ptr = &remcom_in_buffer[1];
+		if (kgdb_hex2long(&ptr, &address))
+			regs->pc = address;
+		else if (*(unsigned long *)regs->pc == compiled_bpt)
+			regs->pc += BREAK_INSTR_SIZE;
+
+		if (remcom_in_buffer[0] == 's') {
+			do_single_step(regs);
+			kgdb_single_step = 1;
+			atomic_set(&kgdb_cpu_doing_single_step,
+				   raw_smp_processor_id());
+		} else
+			atomic_set(&kgdb_cpu_doing_single_step, -1);
+
+		return 0;
+	}
+
+	return -1; /* this means that we do not want to exit from the handler */
+}
+
+struct kgdb_arch arch_kgdb_ops;
+
+/*
+ * kgdb_arch_init - Perform any architecture specific initalization.
+ *
+ * This function will handle the initalization of any architecture
+ * specific callbacks.
+ */
+int kgdb_arch_init(void)
+{
+	tile_bundle_bits bundle = TILEGX_BPT_BUNDLE;
+
+	memcpy(arch_kgdb_ops.gdb_bpt_instr, &bundle, BREAK_INSTR_SIZE);
+	return register_die_notifier(&kgdb_notifier);
+}
+
+/*
+ * kgdb_arch_exit - Perform any architecture specific uninitalization.
+ *
+ * This function will handle the uninitalization of any architecture
+ * specific callbacks, for dynamic registration and unregistration.
+ */
+void kgdb_arch_exit(void)
+{
+	unregister_die_notifier(&kgdb_notifier);
+}
+
+int kgdb_arch_set_breakpoint(struct kgdb_bkpt *bpt)
+{
+	int err;
+	unsigned long addr_wr = writable_address(bpt->bpt_addr);
+
+	if (addr_wr == 0)
+		return -1;
+
+	err = probe_kernel_read(bpt->saved_instr, (char *)bpt->bpt_addr,
+				BREAK_INSTR_SIZE);
+	if (err)
+		return err;
+
+	err = probe_kernel_write((char *)addr_wr, arch_kgdb_ops.gdb_bpt_instr,
+				 BREAK_INSTR_SIZE);
+	smp_wmb();
+	flush_icache_range((unsigned long)bpt->bpt_addr,
+			   (unsigned long)bpt->bpt_addr + BREAK_INSTR_SIZE);
+	return err;
+}
+
+int kgdb_arch_remove_breakpoint(struct kgdb_bkpt *bpt)
+{
+	int err;
+	unsigned long addr_wr = writable_address(bpt->bpt_addr);
+
+	if (addr_wr == 0)
+		return -1;
+
+	err = probe_kernel_write((char *)addr_wr, (char *)bpt->saved_instr,
+				 BREAK_INSTR_SIZE);
+	smp_wmb();
+	flush_icache_range((unsigned long)bpt->bpt_addr,
+			   (unsigned long)bpt->bpt_addr + BREAK_INSTR_SIZE);
+	return err;
+}
diff --git a/kernel/arch/tile/kernel/kprobes.c b/kernel/arch/tile/kernel/kprobes.c
new file mode 100644
index 000000000..f8a45c51e
--- /dev/null
+++ b/kernel/arch/tile/kernel/kprobes.c
@@ -0,0 +1,527 @@
+/*
+ * arch/tile/kernel/kprobes.c
+ * Kprobes on TILE-Gx
+ *
+ * Some portions copied from the MIPS version.
+ *
+ * Copyright (C) IBM Corporation, 2002, 2004
+ * Copyright 2006 Sony Corp.
+ * Copyright 2010 Cavium Networks
+ *
+ * Copyright 2012 Tilera Corporation. All Rights Reserved.
+ *
+ *   This program is free software; you can redistribute it and/or
+ *   modify it under the terms of the GNU General Public License
+ *   as published by the Free Software Foundation, version 2.
+ *
+ *   This program is distributed in the hope that it will be useful, but
+ *   WITHOUT ANY WARRANTY; without even the implied warranty of
+ *   MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
+ *   NON INFRINGEMENT.  See the GNU General Public License for
+ *   more details.
+ */
+
+#include <linux/kprobes.h>
+#include <linux/kdebug.h>
+#include <linux/module.h>
+#include <linux/slab.h>
+#include <linux/uaccess.h>
+#include <asm/cacheflush.h>
+
+#include <arch/opcode.h>
+
+DEFINE_PER_CPU(struct kprobe *, current_kprobe) = NULL;
+DEFINE_PER_CPU(struct kprobe_ctlblk, kprobe_ctlblk);
+
+tile_bundle_bits breakpoint_insn = TILEGX_BPT_BUNDLE;
+tile_bundle_bits breakpoint2_insn = TILEGX_BPT_BUNDLE | DIE_SSTEPBP;
+
+/*
+ * Check whether instruction is branch or jump, or if executing it
+ * has different results depending on where it is executed (e.g. lnk).
+ */
+static int __kprobes insn_has_control(kprobe_opcode_t insn)
+{
+	if (get_Mode(insn) != 0) {   /* Y-format bundle */
+		if (get_Opcode_Y1(insn) != RRR_1_OPCODE_Y1 ||
+		    get_RRROpcodeExtension_Y1(insn) != UNARY_RRR_1_OPCODE_Y1)
+			return 0;
+
+		switch (get_UnaryOpcodeExtension_Y1(insn)) {
+		case JALRP_UNARY_OPCODE_Y1:
+		case JALR_UNARY_OPCODE_Y1:
+		case JRP_UNARY_OPCODE_Y1:
+		case JR_UNARY_OPCODE_Y1:
+		case LNK_UNARY_OPCODE_Y1:
+			return 1;
+		default:
+			return 0;
+		}
+	}
+
+	switch (get_Opcode_X1(insn)) {
+	case BRANCH_OPCODE_X1:	/* branch instructions */
+	case JUMP_OPCODE_X1:	/* jump instructions: j and jal */
+		return 1;
+
+	case RRR_0_OPCODE_X1:   /* other jump instructions */
+		if (get_RRROpcodeExtension_X1(insn) != UNARY_RRR_0_OPCODE_X1)
+			return 0;
+		switch (get_UnaryOpcodeExtension_X1(insn)) {
+		case JALRP_UNARY_OPCODE_X1:
+		case JALR_UNARY_OPCODE_X1:
+		case JRP_UNARY_OPCODE_X1:
+		case JR_UNARY_OPCODE_X1:
+		case LNK_UNARY_OPCODE_X1:
+			return 1;
+		default:
+			return 0;
+		}
+	default:
+		return 0;
+	}
+}
+
+int __kprobes arch_prepare_kprobe(struct kprobe *p)
+{
+	unsigned long addr = (unsigned long)p->addr;
+
+	if (addr & (sizeof(kprobe_opcode_t) - 1))
+		return -EINVAL;
+
+	if (insn_has_control(*p->addr)) {
+		pr_notice("Kprobes for control instructions are not supported\n");
+		return -EINVAL;
+	}
+
+	/* insn: must be on special executable page on tile. */
+	p->ainsn.insn = get_insn_slot();
+	if (!p->ainsn.insn)
+		return -ENOMEM;
+
+	/*
+	 * In the kprobe->ainsn.insn[] array we store the original
+	 * instruction at index zero and a break trap instruction at
+	 * index one.
+	 */
+	memcpy(&p->ainsn.insn[0], p->addr, sizeof(kprobe_opcode_t));
+	p->ainsn.insn[1] = breakpoint2_insn;
+	p->opcode = *p->addr;
+
+	return 0;
+}
+
+void __kprobes arch_arm_kprobe(struct kprobe *p)
+{
+	unsigned long addr_wr;
+
+	/* Operate on writable kernel text mapping. */
+	addr_wr = (unsigned long)p->addr - MEM_SV_START + PAGE_OFFSET;
+
+	if (probe_kernel_write((void *)addr_wr, &breakpoint_insn,
+		sizeof(breakpoint_insn)))
+		pr_err("%s: failed to enable kprobe\n", __func__);
+
+	smp_wmb();
+	flush_insn_slot(p);
+}
+
+void __kprobes arch_disarm_kprobe(struct kprobe *kp)
+{
+	unsigned long addr_wr;
+
+	/* Operate on writable kernel text mapping. */
+	addr_wr = (unsigned long)kp->addr - MEM_SV_START + PAGE_OFFSET;
+
+	if (probe_kernel_write((void *)addr_wr, &kp->opcode,
+		sizeof(kp->opcode)))
+		pr_err("%s: failed to enable kprobe\n", __func__);
+
+	smp_wmb();
+	flush_insn_slot(kp);
+}
+
+void __kprobes arch_remove_kprobe(struct kprobe *p)
+{
+	if (p->ainsn.insn) {
+		free_insn_slot(p->ainsn.insn, 0);
+		p->ainsn.insn = NULL;
+	}
+}
+
+static void __kprobes save_previous_kprobe(struct kprobe_ctlblk *kcb)
+{
+	kcb->prev_kprobe.kp = kprobe_running();
+	kcb->prev_kprobe.status = kcb->kprobe_status;
+	kcb->prev_kprobe.saved_pc = kcb->kprobe_saved_pc;
+}
+
+static void __kprobes restore_previous_kprobe(struct kprobe_ctlblk *kcb)
+{
+	__this_cpu_write(current_kprobe, kcb->prev_kprobe.kp);
+	kcb->kprobe_status = kcb->prev_kprobe.status;
+	kcb->kprobe_saved_pc = kcb->prev_kprobe.saved_pc;
+}
+
+static void __kprobes set_current_kprobe(struct kprobe *p, struct pt_regs *regs,
+			struct kprobe_ctlblk *kcb)
+{
+	__this_cpu_write(current_kprobe, p);
+	kcb->kprobe_saved_pc = regs->pc;
+}
+
+static void __kprobes prepare_singlestep(struct kprobe *p, struct pt_regs *regs)
+{
+	/* Single step inline if the instruction is a break. */
+	if (p->opcode == breakpoint_insn ||
+	    p->opcode == breakpoint2_insn)
+		regs->pc = (unsigned long)p->addr;
+	else
+		regs->pc = (unsigned long)&p->ainsn.insn[0];
+}
+
+static int __kprobes kprobe_handler(struct pt_regs *regs)
+{
+	struct kprobe *p;
+	int ret = 0;
+	kprobe_opcode_t *addr;
+	struct kprobe_ctlblk *kcb;
+
+	addr = (kprobe_opcode_t *)regs->pc;
+
+	/*
+	 * We don't want to be preempted for the entire
+	 * duration of kprobe processing.
+	 */
+	preempt_disable();
+	kcb = get_kprobe_ctlblk();
+
+	/* Check we're not actually recursing. */
+	if (kprobe_running()) {
+		p = get_kprobe(addr);
+		if (p) {
+			if (kcb->kprobe_status == KPROBE_HIT_SS &&
+			    p->ainsn.insn[0] == breakpoint_insn) {
+				goto no_kprobe;
+			}
+			/*
+			 * We have reentered the kprobe_handler(), since
+			 * another probe was hit while within the handler.
+			 * We here save the original kprobes variables and
+			 * just single step on the instruction of the new probe
+			 * without calling any user handlers.
+			 */
+			save_previous_kprobe(kcb);
+			set_current_kprobe(p, regs, kcb);
+			kprobes_inc_nmissed_count(p);
+			prepare_singlestep(p, regs);
+			kcb->kprobe_status = KPROBE_REENTER;
+			return 1;
+		} else {
+			if (*addr != breakpoint_insn) {
+				/*
+				 * The breakpoint instruction was removed by
+				 * another cpu right after we hit, no further
+				 * handling of this interrupt is appropriate.
+				 */
+				ret = 1;
+				goto no_kprobe;
+			}
+			p = __this_cpu_read(current_kprobe);
+			if (p->break_handler && p->break_handler(p, regs))
+				goto ss_probe;
+		}
+		goto no_kprobe;
+	}
+
+	p = get_kprobe(addr);
+	if (!p) {
+		if (*addr != breakpoint_insn) {
+			/*
+			 * The breakpoint instruction was removed right
+			 * after we hit it.  Another cpu has removed
+			 * either a probepoint or a debugger breakpoint
+			 * at this address.  In either case, no further
+			 * handling of this interrupt is appropriate.
+			 */
+			ret = 1;
+		}
+		/* Not one of ours: let kernel handle it. */
+		goto no_kprobe;
+	}
+
+	set_current_kprobe(p, regs, kcb);
+	kcb->kprobe_status = KPROBE_HIT_ACTIVE;
+
+	if (p->pre_handler && p->pre_handler(p, regs)) {
+		/* Handler has already set things up, so skip ss setup. */
+		return 1;
+	}
+
+ss_probe:
+	prepare_singlestep(p, regs);
+	kcb->kprobe_status = KPROBE_HIT_SS;
+	return 1;
+
+no_kprobe:
+	preempt_enable_no_resched();
+	return ret;
+}
+
+/*
+ * Called after single-stepping.  p->addr is the address of the
+ * instruction that has been replaced by the breakpoint. To avoid the
+ * SMP problems that can occur when we temporarily put back the
+ * original opcode to single-step, we single-stepped a copy of the
+ * instruction.  The address of this copy is p->ainsn.insn.
+ *
+ * This function prepares to return from the post-single-step
+ * breakpoint trap.
+ */
+static void __kprobes resume_execution(struct kprobe *p,
+				       struct pt_regs *regs,
+				       struct kprobe_ctlblk *kcb)
+{
+	unsigned long orig_pc = kcb->kprobe_saved_pc;
+	regs->pc = orig_pc + 8;
+}
+
+static inline int post_kprobe_handler(struct pt_regs *regs)
+{
+	struct kprobe *cur = kprobe_running();
+	struct kprobe_ctlblk *kcb = get_kprobe_ctlblk();
+
+	if (!cur)
+		return 0;
+
+	if ((kcb->kprobe_status != KPROBE_REENTER) && cur->post_handler) {
+		kcb->kprobe_status = KPROBE_HIT_SSDONE;
+		cur->post_handler(cur, regs, 0);
+	}
+
+	resume_execution(cur, regs, kcb);
+
+	/* Restore back the original saved kprobes variables and continue. */
+	if (kcb->kprobe_status == KPROBE_REENTER) {
+		restore_previous_kprobe(kcb);
+		goto out;
+	}
+	reset_current_kprobe();
+out:
+	preempt_enable_no_resched();
+
+	return 1;
+}
+
+static inline int kprobe_fault_handler(struct pt_regs *regs, int trapnr)
+{
+	struct kprobe *cur = kprobe_running();
+	struct kprobe_ctlblk *kcb = get_kprobe_ctlblk();
+
+	if (cur->fault_handler && cur->fault_handler(cur, regs, trapnr))
+		return 1;
+
+	if (kcb->kprobe_status & KPROBE_HIT_SS) {
+		/*
+		 * We are here because the instruction being single
+		 * stepped caused a page fault. We reset the current
+		 * kprobe and the ip points back to the probe address
+		 * and allow the page fault handler to continue as a
+		 * normal page fault.
+		 */
+		resume_execution(cur, regs, kcb);
+		reset_current_kprobe();
+		preempt_enable_no_resched();
+	}
+	return 0;
+}
+
+/*
+ * Wrapper routine for handling exceptions.
+ */
+int __kprobes kprobe_exceptions_notify(struct notifier_block *self,
+				       unsigned long val, void *data)
+{
+	struct die_args *args = (struct die_args *)data;
+	int ret = NOTIFY_DONE;
+
+	switch (val) {
+	case DIE_BREAK:
+		if (kprobe_handler(args->regs))
+			ret = NOTIFY_STOP;
+		break;
+	case DIE_SSTEPBP:
+		if (post_kprobe_handler(args->regs))
+			ret = NOTIFY_STOP;
+		break;
+	case DIE_PAGE_FAULT:
+		/* kprobe_running() needs smp_processor_id(). */
+		preempt_disable();
+
+		if (kprobe_running()
+		    && kprobe_fault_handler(args->regs, args->trapnr))
+			ret = NOTIFY_STOP;
+		preempt_enable();
+		break;
+	default:
+		break;
+	}
+	return ret;
+}
+
+int __kprobes setjmp_pre_handler(struct kprobe *p, struct pt_regs *regs)
+{
+	struct jprobe *jp = container_of(p, struct jprobe, kp);
+	struct kprobe_ctlblk *kcb = get_kprobe_ctlblk();
+
+	kcb->jprobe_saved_regs = *regs;
+	kcb->jprobe_saved_sp = regs->sp;
+
+	memcpy(kcb->jprobes_stack, (void *)kcb->jprobe_saved_sp,
+	       MIN_JPROBES_STACK_SIZE(kcb->jprobe_saved_sp));
+
+	regs->pc = (unsigned long)(jp->entry);
+
+	return 1;
+}
+
+/* Defined in the inline asm below. */
+void jprobe_return_end(void);
+
+void __kprobes jprobe_return(void)
+{
+	asm volatile(
+		"bpt\n\t"
+		".globl jprobe_return_end\n"
+		"jprobe_return_end:\n");
+}
+
+int __kprobes longjmp_break_handler(struct kprobe *p, struct pt_regs *regs)
+{
+	struct kprobe_ctlblk *kcb = get_kprobe_ctlblk();
+
+	if (regs->pc >= (unsigned long)jprobe_return &&
+	    regs->pc <= (unsigned long)jprobe_return_end) {
+		*regs = kcb->jprobe_saved_regs;
+		memcpy((void *)kcb->jprobe_saved_sp, kcb->jprobes_stack,
+		       MIN_JPROBES_STACK_SIZE(kcb->jprobe_saved_sp));
+		preempt_enable_no_resched();
+
+		return 1;
+	}
+	return 0;
+}
+
+/*
+ * Function return probe trampoline:
+ * - init_kprobes() establishes a probepoint here
+ * - When the probed function returns, this probe causes the
+ *   handlers to fire
+ */
+static void __used kretprobe_trampoline_holder(void)
+{
+	asm volatile(
+		"nop\n\t"
+		".global kretprobe_trampoline\n"
+		"kretprobe_trampoline:\n\t"
+		"nop\n\t"
+		: : : "memory");
+}
+
+void kretprobe_trampoline(void);
+
+void __kprobes arch_prepare_kretprobe(struct kretprobe_instance *ri,
+				      struct pt_regs *regs)
+{
+	ri->ret_addr = (kprobe_opcode_t *) regs->lr;
+
+	/* Replace the return addr with trampoline addr */
+	regs->lr = (unsigned long)kretprobe_trampoline;
+}
+
+/*
+ * Called when the probe at kretprobe trampoline is hit.
+ */
+static int __kprobes trampoline_probe_handler(struct kprobe *p,
+						struct pt_regs *regs)
+{
+	struct kretprobe_instance *ri = NULL;
+	struct hlist_head *head, empty_rp;
+	struct hlist_node *tmp;
+	unsigned long flags, orig_ret_address = 0;
+	unsigned long trampoline_address = (unsigned long)kretprobe_trampoline;
+
+	INIT_HLIST_HEAD(&empty_rp);
+	kretprobe_hash_lock(current, &head, &flags);
+
+	/*
+	 * It is possible to have multiple instances associated with a given
+	 * task either because multiple functions in the call path have
+	 * a return probe installed on them, and/or more than one return
+	 * return probe was registered for a target function.
+	 *
+	 * We can handle this because:
+	 *     - instances are always inserted at the head of the list
+	 *     - when multiple return probes are registered for the same
+	 *       function, the first instance's ret_addr will point to the
+	 *       real return address, and all the rest will point to
+	 *       kretprobe_trampoline
+	 */
+	hlist_for_each_entry_safe(ri, tmp, head, hlist) {
+		if (ri->task != current)
+			/* another task is sharing our hash bucket */
+			continue;
+
+		if (ri->rp && ri->rp->handler)
+			ri->rp->handler(ri, regs);
+
+		orig_ret_address = (unsigned long)ri->ret_addr;
+		recycle_rp_inst(ri, &empty_rp);
+
+		if (orig_ret_address != trampoline_address) {
+			/*
+			 * This is the real return address. Any other
+			 * instances associated with this task are for
+			 * other calls deeper on the call stack
+			 */
+			break;
+		}
+	}
+
+	kretprobe_assert(ri, orig_ret_address, trampoline_address);
+	instruction_pointer(regs) = orig_ret_address;
+
+	reset_current_kprobe();
+	kretprobe_hash_unlock(current, &flags);
+	preempt_enable_no_resched();
+
+	hlist_for_each_entry_safe(ri, tmp, &empty_rp, hlist) {
+		hlist_del(&ri->hlist);
+		kfree(ri);
+	}
+	/*
+	 * By returning a non-zero value, we are telling
+	 * kprobe_handler() that we don't want the post_handler
+	 * to run (and have re-enabled preemption)
+	 */
+	return 1;
+}
+
+int __kprobes arch_trampoline_kprobe(struct kprobe *p)
+{
+	if (p->addr == (kprobe_opcode_t *)kretprobe_trampoline)
+		return 1;
+
+	return 0;
+}
+
+static struct kprobe trampoline_p = {
+	.addr = (kprobe_opcode_t *)kretprobe_trampoline,
+	.pre_handler = trampoline_probe_handler
+};
+
+int __init arch_init_kprobes(void)
+{
+	register_kprobe(&trampoline_p);
+	return 0;
+}
diff --git a/kernel/arch/tile/kernel/machine_kexec.c b/kernel/arch/tile/kernel/machine_kexec.c
new file mode 100644
index 000000000..008aa2fae
--- /dev/null
+++ b/kernel/arch/tile/kernel/machine_kexec.c
@@ -0,0 +1,298 @@
+/*
+ * Copyright 2010 Tilera Corporation. All Rights Reserved.
+ *
+ *   This program is free software; you can redistribute it and/or
+ *   modify it under the terms of the GNU General Public License
+ *   as published by the Free Software Foundation, version 2.
+ *
+ *   This program is distributed in the hope that it will be useful, but
+ *   WITHOUT ANY WARRANTY; without even the implied warranty of
+ *   MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
+ *   NON INFRINGEMENT.  See the GNU General Public License for
+ *   more details.
+ *
+ * based on machine_kexec.c from other architectures in linux-2.6.18
+ */
+
+#include <linux/mm.h>
+#include <linux/kexec.h>
+#include <linux/delay.h>
+#include <linux/reboot.h>
+#include <linux/errno.h>
+#include <linux/vmalloc.h>
+#include <linux/cpumask.h>
+#include <linux/kernel.h>
+#include <linux/elf.h>
+#include <linux/highmem.h>
+#include <linux/mmu_context.h>
+#include <linux/io.h>
+#include <linux/timex.h>
+#include <asm/pgtable.h>
+#include <asm/pgalloc.h>
+#include <asm/cacheflush.h>
+#include <asm/checksum.h>
+#include <asm/tlbflush.h>
+#include <asm/homecache.h>
+#include <hv/hypervisor.h>
+
+
+/*
+ * This stuff is not in elf.h and is not in any other kernel include.
+ * This stuff is needed below in the little boot notes parser to
+ * extract the command line so we can pass it to the hypervisor.
+ */
+struct Elf32_Bhdr {
+	Elf32_Word b_signature;
+	Elf32_Word b_size;
+	Elf32_Half b_checksum;
+	Elf32_Half b_records;
+};
+#define ELF_BOOT_MAGIC		0x0E1FB007
+#define EBN_COMMAND_LINE	0x00000004
+#define roundupsz(X) (((X) + 3) & ~3)
+
+/* - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - */
+
+
+void machine_shutdown(void)
+{
+	/*
+	 * Normally we would stop all the other processors here, but
+	 * the check in machine_kexec_prepare below ensures we'll only
+	 * get this far if we've been booted with "nosmp" on the
+	 * command line or without CONFIG_SMP so there's nothing to do
+	 * here (for now).
+	 */
+}
+
+void machine_crash_shutdown(struct pt_regs *regs)
+{
+	/*
+	 * Cannot happen.  This type of kexec is disabled on this
+	 * architecture (and enforced in machine_kexec_prepare below).
+	 */
+}
+
+
+int machine_kexec_prepare(struct kimage *image)
+{
+	if (num_online_cpus() > 1) {
+		pr_warn("%s: detected attempt to kexec with num_online_cpus() > 1\n",
+			__func__);
+		return -ENOSYS;
+	}
+	if (image->type != KEXEC_TYPE_DEFAULT) {
+		pr_warn("%s: detected attempt to kexec with unsupported type: %d\n",
+			__func__, image->type);
+		return -ENOSYS;
+	}
+	return 0;
+}
+
+void machine_kexec_cleanup(struct kimage *image)
+{
+	/*
+	 * We did nothing in machine_kexec_prepare,
+	 * so we have nothing to do here.
+	 */
+}
+
+/*
+ * If we can find elf boot notes on this page, return the command
+ * line.  Otherwise, silently return null.  Somewhat kludgy, but no
+ * good way to do this without significantly rearchitecting the
+ * architecture-independent kexec code.
+ */
+
+static unsigned char *kexec_bn2cl(void *pg)
+{
+	struct Elf32_Bhdr *bhdrp;
+	Elf32_Nhdr *nhdrp;
+	unsigned char *desc;
+	unsigned char *command_line;
+	__sum16 csum;
+
+	bhdrp = (struct Elf32_Bhdr *) pg;
+
+	/*
+	 * This routine is invoked for every source page, so make
+	 * sure to quietly ignore every impossible page.
+	 */
+	if (bhdrp->b_signature != ELF_BOOT_MAGIC ||
+	    bhdrp->b_size > PAGE_SIZE)
+		return 0;
+
+	/*
+	 * If we get a checksum mismatch, warn with the checksum
+	 * so we can diagnose better.
+	 */
+	csum = ip_compute_csum(pg, bhdrp->b_size);
+	if (csum != 0) {
+		pr_warn("%s: bad checksum %#x (size %d)\n",
+			__func__, csum, bhdrp->b_size);
+		return 0;
+	}
+
+	nhdrp = (Elf32_Nhdr *) (bhdrp + 1);
+
+	while (nhdrp->n_type != EBN_COMMAND_LINE) {
+
+		desc = (unsigned char *) (nhdrp + 1);
+		desc += roundupsz(nhdrp->n_descsz);
+
+		nhdrp = (Elf32_Nhdr *) desc;
+
+		/* still in bounds? */
+		if ((unsigned char *) (nhdrp + 1) >
+		    ((unsigned char *) pg) + bhdrp->b_size) {
+
+			pr_info("%s: out of bounds\n", __func__);
+			return 0;
+		}
+	}
+
+	command_line = (unsigned char *) (nhdrp + 1);
+	desc = command_line;
+
+	while (*desc != '\0') {
+		desc++;
+		if (((unsigned long)desc & PAGE_MASK) != (unsigned long)pg) {
+			pr_info("%s: ran off end of page\n", __func__);
+			return 0;
+		}
+	}
+
+	return command_line;
+}
+
+static void kexec_find_and_set_command_line(struct kimage *image)
+{
+	kimage_entry_t *ptr, entry;
+
+	unsigned char *command_line = 0;
+	unsigned char *r;
+	HV_Errno hverr;
+
+	for (ptr = &image->head;
+	     (entry = *ptr) && !(entry & IND_DONE);
+	     ptr = (entry & IND_INDIRECTION) ?
+		     phys_to_virt((entry & PAGE_MASK)) : ptr + 1) {
+
+		if ((entry & IND_SOURCE)) {
+			void *va =
+				kmap_atomic_pfn(entry >> PAGE_SHIFT);
+			r = kexec_bn2cl(va);
+			if (r) {
+				command_line = r;
+				break;
+			}
+			kunmap_atomic(va);
+		}
+	}
+
+	if (command_line != 0) {
+		pr_info("setting new command line to \"%s\"\n", command_line);
+
+		hverr = hv_set_command_line(
+			(HV_VirtAddr) command_line, strlen(command_line));
+		kunmap_atomic(command_line);
+	} else {
+		pr_info("%s: no command line found; making empty\n", __func__);
+		hverr = hv_set_command_line((HV_VirtAddr) command_line, 0);
+	}
+	if (hverr)
+		pr_warn("%s: hv_set_command_line returned error: %d\n",
+			__func__, hverr);
+}
+
+/*
+ * The kexec code range-checks all its PAs, so to avoid having it run
+ * amok and allocate memory and then sequester it from every other
+ * controller, we force it to come from controller zero.  We also
+ * disable the oom-killer since if we do end up running out of memory,
+ * that almost certainly won't help.
+ */
+struct page *kimage_alloc_pages_arch(gfp_t gfp_mask, unsigned int order)
+{
+	gfp_mask |= __GFP_THISNODE | __GFP_NORETRY;
+	return alloc_pages_node(0, gfp_mask, order);
+}
+
+/*
+ * Address range in which pa=va mapping is set in setup_quasi_va_is_pa().
+ * For tilepro, PAGE_OFFSET is used since this is the largest possbile value
+ * for tilepro, while for tilegx, we limit it to entire middle level page
+ * table which we assume has been allocated and is undoubtedly large enough.
+ */
+#ifndef __tilegx__
+#define	QUASI_VA_IS_PA_ADDR_RANGE PAGE_OFFSET
+#else
+#define	QUASI_VA_IS_PA_ADDR_RANGE PGDIR_SIZE
+#endif
+
+static void setup_quasi_va_is_pa(void)
+{
+	HV_PTE pte;
+	unsigned long i;
+
+	/*
+	 * Flush our TLB to prevent conflicts between the previous contents
+	 * and the new stuff we're about to add.
+	 */
+	local_flush_tlb_all();
+
+	/*
+	 * setup VA is PA, at least up to QUASI_VA_IS_PA_ADDR_RANGE.
+	 * Note here we assume that level-1 page table is defined by
+	 * HPAGE_SIZE.
+	 */
+	pte = hv_pte(_PAGE_KERNEL | _PAGE_HUGE_PAGE);
+	pte = hv_pte_set_mode(pte, HV_PTE_MODE_CACHE_NO_L3);
+	for (i = 0; i < (QUASI_VA_IS_PA_ADDR_RANGE >> HPAGE_SHIFT); i++) {
+		unsigned long vaddr = i << HPAGE_SHIFT;
+		pgd_t *pgd = pgd_offset(current->mm, vaddr);
+		pud_t *pud = pud_offset(pgd, vaddr);
+		pte_t *ptep = (pte_t *) pmd_offset(pud, vaddr);
+		unsigned long pfn = i << (HPAGE_SHIFT - PAGE_SHIFT);
+
+		if (pfn_valid(pfn))
+			__set_pte(ptep, pfn_pte(pfn, pte));
+	}
+}
+
+
+void machine_kexec(struct kimage *image)
+{
+	void *reboot_code_buffer;
+	pte_t *ptep;
+	void (*rnk)(unsigned long, void *, unsigned long)
+		__noreturn;
+
+	/* Mask all interrupts before starting to reboot. */
+	interrupt_mask_set_mask(~0ULL);
+
+	kexec_find_and_set_command_line(image);
+
+	/*
+	 * Adjust the home caching of the control page to be cached on
+	 * this cpu, and copy the assembly helper into the control
+	 * code page, which we map in the vmalloc area.
+	 */
+	homecache_change_page_home(image->control_code_page, 0,
+				   smp_processor_id());
+	reboot_code_buffer = page_address(image->control_code_page);
+	BUG_ON(reboot_code_buffer == NULL);
+	ptep = virt_to_pte(NULL, (unsigned long)reboot_code_buffer);
+	__set_pte(ptep, pte_mkexec(*ptep));
+	memcpy(reboot_code_buffer, relocate_new_kernel,
+	       relocate_new_kernel_size);
+	__flush_icache_range(
+		(unsigned long) reboot_code_buffer,
+		(unsigned long) reboot_code_buffer + relocate_new_kernel_size);
+
+	setup_quasi_va_is_pa();
+
+	/* now call it */
+	rnk = reboot_code_buffer;
+	(*rnk)(image->head, reboot_code_buffer, image->start);
+}
diff --git a/kernel/arch/tile/kernel/mcount_64.S b/kernel/arch/tile/kernel/mcount_64.S
new file mode 100644
index 000000000..6c6702451
--- /dev/null
+++ b/kernel/arch/tile/kernel/mcount_64.S
@@ -0,0 +1,211 @@
+/*
+ * Copyright 2012 Tilera Corporation. All Rights Reserved.
+ *
+ *   This program is free software; you can redistribute it and/or
+ *   modify it under the terms of the GNU General Public License
+ *   as published by the Free Software Foundation, version 2.
+ *
+ *   This program is distributed in the hope that it will be useful, but
+ *   WITHOUT ANY WARRANTY; without even the implied warranty of
+ *   MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
+ *   NON INFRINGEMENT.  See the GNU General Public License for
+ *   more details.
+ *
+ * TILE-Gx specific __mcount support
+ */
+
+#include <linux/linkage.h>
+#include <asm/ftrace.h>
+
+#define REGSIZE 8
+
+	.text
+	.global __mcount
+
+	.macro	MCOUNT_SAVE_REGS
+	addli	sp, sp, -REGSIZE
+	{
+	 st     sp, lr
+	 addli	r29, sp, - (12 * REGSIZE)
+	}
+	{
+	 addli	sp, sp, - (13 * REGSIZE)
+	 st     r29, sp
+	}
+	addli	r29, r29, REGSIZE
+	{ st	r29, r0; addli	r29, r29, REGSIZE }
+	{ st	r29, r1; addli	r29, r29, REGSIZE }
+	{ st	r29, r2; addli	r29, r29, REGSIZE }
+	{ st	r29, r3; addli	r29, r29, REGSIZE }
+	{ st	r29, r4; addli	r29, r29, REGSIZE }
+	{ st	r29, r5; addli	r29, r29, REGSIZE }
+	{ st	r29, r6; addli	r29, r29, REGSIZE }
+	{ st	r29, r7; addli	r29, r29, REGSIZE }
+	{ st	r29, r8; addli	r29, r29, REGSIZE }
+	{ st	r29, r9; addli	r29, r29, REGSIZE }
+	{ st	r29, r10; addli	r29, r29, REGSIZE }
+	.endm
+
+	.macro	MCOUNT_RESTORE_REGS
+	addli	r29, sp, (2 * REGSIZE)
+	{ ld	r0, r29; addli	r29, r29, REGSIZE }
+	{ ld	r1, r29; addli	r29, r29, REGSIZE }
+	{ ld	r2, r29; addli	r29, r29, REGSIZE }
+	{ ld	r3, r29; addli	r29, r29, REGSIZE }
+	{ ld	r4, r29; addli	r29, r29, REGSIZE }
+	{ ld	r5, r29; addli	r29, r29, REGSIZE }
+	{ ld	r6, r29; addli	r29, r29, REGSIZE }
+	{ ld	r7, r29; addli	r29, r29, REGSIZE }
+	{ ld	r8, r29; addli	r29, r29, REGSIZE }
+	{ ld	r9, r29; addli	r29, r29, REGSIZE }
+	{ ld	r10, r29; addli	lr, sp, (13 * REGSIZE) }
+	{ ld	lr, lr;  addli	sp, sp, (14 * REGSIZE) }
+	.endm
+
+	.macro  RETURN_BACK
+	{ move	r12, lr; move	lr, r10 }
+	jrp	r12
+	.endm
+
+#ifdef CONFIG_DYNAMIC_FTRACE
+
+	.align	64
+STD_ENTRY(__mcount)
+__mcount:
+	j	ftrace_stub
+STD_ENDPROC(__mcount)
+
+	.align	64
+STD_ENTRY(ftrace_caller)
+	MCOUNT_SAVE_REGS
+
+	/* arg1: self return address */
+	/* arg2: parent's return address */
+	/* arg3: ftrace_ops */
+	/* arg4: regs (but make it NULL) */
+	{ move	r0, lr;  moveli        r2, hw2_last(function_trace_op) }
+	{ move	r1, r10; shl16insli    r2, r2, hw1(function_trace_op) }
+	{ movei	r3, 0;   shl16insli    r2, r2, hw0(function_trace_op) }
+	ld	r2,r2
+
+	.global	ftrace_call
+ftrace_call:
+	/*
+	 * a placeholder for the call to a real tracing function, i.e.
+	 * ftrace_trace_function()
+	 */
+	nop
+
+#ifdef CONFIG_FUNCTION_GRAPH_TRACER
+	.global	ftrace_graph_call
+ftrace_graph_call:
+	/*
+	 * a placeholder for the call to a real tracing function, i.e.
+	 * ftrace_graph_caller()
+	 */
+	nop
+#endif
+	MCOUNT_RESTORE_REGS
+	.global	ftrace_stub
+ftrace_stub:
+	RETURN_BACK
+STD_ENDPROC(ftrace_caller)
+
+#else /* ! CONFIG_DYNAMIC_FTRACE */
+
+	.align	64
+STD_ENTRY(__mcount)
+	{
+	 moveli	r11, hw2_last(ftrace_trace_function)
+	 moveli	r13, hw2_last(ftrace_stub)
+	}
+	{
+	 shl16insli	r11, r11, hw1(ftrace_trace_function)
+	 shl16insli	r13, r13, hw1(ftrace_stub)
+	}
+	{
+	 shl16insli	r11, r11, hw0(ftrace_trace_function)
+	 shl16insli	r13, r13, hw0(ftrace_stub)
+	}
+
+	ld	r11, r11
+	sub	r14, r13, r11
+	bnez	r14, static_trace
+
+#ifdef	CONFIG_FUNCTION_GRAPH_TRACER
+	moveli	r15, hw2_last(ftrace_graph_return)
+	shl16insli	r15, r15, hw1(ftrace_graph_return)
+	shl16insli	r15, r15, hw0(ftrace_graph_return)
+	ld	r15, r15
+	sub	r15, r15, r13
+	bnez	r15, ftrace_graph_caller
+
+	{
+	 moveli	r16, hw2_last(ftrace_graph_entry)
+	 moveli	r17, hw2_last(ftrace_graph_entry_stub)
+	}
+	{
+	 shl16insli	r16, r16, hw1(ftrace_graph_entry)
+	 shl16insli	r17, r17, hw1(ftrace_graph_entry_stub)
+	}
+	{
+	 shl16insli	r16, r16, hw0(ftrace_graph_entry)
+	 shl16insli	r17, r17, hw0(ftrace_graph_entry_stub)
+	}
+	ld	r16, r16
+	sub	r17, r16, r17
+	bnez	r17, ftrace_graph_caller
+
+#endif
+	RETURN_BACK
+
+static_trace:
+	MCOUNT_SAVE_REGS
+
+	/* arg1: self return address */
+	/* arg2: parent's return address */
+	{ move	r0, lr; move	r1, r10 }
+
+	/* call ftrace_trace_function() */
+	jalr	r11
+
+	MCOUNT_RESTORE_REGS
+
+	.global ftrace_stub
+ftrace_stub:
+	RETURN_BACK
+STD_ENDPROC(__mcount)
+
+#endif	/* ! CONFIG_DYNAMIC_FTRACE */
+
+#ifdef CONFIG_FUNCTION_GRAPH_TRACER
+
+STD_ENTRY(ftrace_graph_caller)
+ftrace_graph_caller:
+#ifndef CONFIG_DYNAMIC_FTRACE
+	MCOUNT_SAVE_REGS
+#endif
+
+	/* arg1: Get the location of the parent's return address */
+	addi	r0, sp, 12 * REGSIZE
+	/* arg2: Get self return address */
+	move	r1, lr
+
+	jal prepare_ftrace_return
+
+	MCOUNT_RESTORE_REGS
+	RETURN_BACK
+STD_ENDPROC(ftrace_graph_caller)
+
+	.global return_to_handler
+return_to_handler:
+	MCOUNT_SAVE_REGS
+
+	jal	ftrace_return_to_handler
+	/* restore the real parent address */
+	move	r11, r0
+
+	MCOUNT_RESTORE_REGS
+	jr	r11
+
+#endif /* CONFIG_FUNCTION_GRAPH_TRACER */
diff --git a/kernel/arch/tile/kernel/messaging.c b/kernel/arch/tile/kernel/messaging.c
new file mode 100644
index 000000000..7475af3aa
--- /dev/null
+++ b/kernel/arch/tile/kernel/messaging.c
@@ -0,0 +1,115 @@
+/*
+ * Copyright 2010 Tilera Corporation. All Rights Reserved.
+ *
+ *   This program is free software; you can redistribute it and/or
+ *   modify it under the terms of the GNU General Public License
+ *   as published by the Free Software Foundation, version 2.
+ *
+ *   This program is distributed in the hope that it will be useful, but
+ *   WITHOUT ANY WARRANTY; without even the implied warranty of
+ *   MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
+ *   NON INFRINGEMENT.  See the GNU General Public License for
+ *   more details.
+ */
+
+#include <linux/percpu.h>
+#include <linux/smp.h>
+#include <linux/hardirq.h>
+#include <linux/ptrace.h>
+#include <asm/hv_driver.h>
+#include <asm/irq_regs.h>
+#include <asm/traps.h>
+#include <hv/hypervisor.h>
+#include <arch/interrupts.h>
+
+/* All messages are stored here */
+static DEFINE_PER_CPU(HV_MsgState, msg_state);
+
+void init_messaging(void)
+{
+	/* Allocate storage for messages in kernel space */
+	HV_MsgState *state = this_cpu_ptr(&msg_state);
+	int rc = hv_register_message_state(state);
+	if (rc != HV_OK)
+		panic("hv_register_message_state: error %d", rc);
+
+	/* Make sure downcall interrupts will be enabled. */
+	arch_local_irq_unmask(INT_INTCTRL_K);
+}
+
+void hv_message_intr(struct pt_regs *regs, int intnum)
+{
+	/*
+	 * We enter with interrupts disabled and leave them disabled,
+	 * to match expectations of called functions (e.g.
+	 * do_ccupdate_local() in mm/slab.c).  This is also consistent
+	 * with normal call entry for device interrupts.
+	 */
+
+	int message[HV_MAX_MESSAGE_SIZE/sizeof(int)];
+	HV_RcvMsgInfo rmi;
+	int nmsgs = 0;
+
+	/* Track time spent here in an interrupt context */
+	struct pt_regs *old_regs = set_irq_regs(regs);
+	irq_enter();
+
+#ifdef CONFIG_DEBUG_STACKOVERFLOW
+	/* Debugging check for stack overflow: less than 1/8th stack free? */
+	{
+		long sp = stack_pointer - (long) current_thread_info();
+		if (unlikely(sp < (sizeof(struct thread_info) + STACK_WARN))) {
+			pr_emerg("%s: stack overflow: %ld\n",
+				 __func__, sp - sizeof(struct thread_info));
+			dump_stack();
+		}
+	}
+#endif
+
+	while (1) {
+		HV_MsgState *state = this_cpu_ptr(&msg_state);
+		rmi = hv_receive_message(*state, (HV_VirtAddr) message,
+					 sizeof(message));
+		if (rmi.msglen == 0)
+			break;
+
+		if (rmi.msglen < 0)
+			panic("hv_receive_message failed: %d", rmi.msglen);
+
+		++nmsgs;
+
+		if (rmi.source == HV_MSG_TILE) {
+			int tag;
+
+			/* we just send tags for now */
+			BUG_ON(rmi.msglen != sizeof(int));
+
+			tag = message[0];
+#ifdef CONFIG_SMP
+			evaluate_message(message[0]);
+#else
+			panic("Received IPI message %d in UP mode", tag);
+#endif
+		} else if (rmi.source == HV_MSG_INTR) {
+			HV_IntrMsg *him = (HV_IntrMsg *)message;
+			struct hv_driver_cb *cb =
+				(struct hv_driver_cb *)him->intarg;
+			cb->callback(cb, him->intdata);
+			__this_cpu_inc(irq_stat.irq_hv_msg_count);
+		}
+	}
+
+	/*
+	 * We shouldn't have gotten a message downcall with no
+	 * messages available.
+	 */
+	if (nmsgs == 0)
+		panic("Message downcall invoked with no messages!");
+
+	/*
+	 * Track time spent against the current process again and
+	 * process any softirqs if they are waiting.
+	 */
+	irq_exit();
+	set_irq_regs(old_regs);
+}
diff --git a/kernel/arch/tile/kernel/module.c b/kernel/arch/tile/kernel/module.c
new file mode 100644
index 000000000..2305084c9
--- /dev/null
+++ b/kernel/arch/tile/kernel/module.c
@@ -0,0 +1,232 @@
+/*
+ * Copyright 2010 Tilera Corporation. All Rights Reserved.
+ *
+ *   This program is free software; you can redistribute it and/or
+ *   modify it under the terms of the GNU General Public License
+ *   as published by the Free Software Foundation, version 2.
+ *
+ *   This program is distributed in the hope that it will be useful, but
+ *   WITHOUT ANY WARRANTY; without even the implied warranty of
+ *   MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
+ *   NON INFRINGEMENT.  See the GNU General Public License for
+ *   more details.
+ *
+ * Based on i386 version, copyright (C) 2001 Rusty Russell.
+ */
+
+#include <linux/moduleloader.h>
+#include <linux/elf.h>
+#include <linux/vmalloc.h>
+#include <linux/fs.h>
+#include <linux/string.h>
+#include <linux/kernel.h>
+#include <asm/pgtable.h>
+#include <asm/homecache.h>
+#include <arch/opcode.h>
+
+#ifdef MODULE_DEBUG
+#define DEBUGP printk
+#else
+#define DEBUGP(fmt...)
+#endif
+
+/*
+ * Allocate some address space in the range MEM_MODULE_START to
+ * MEM_MODULE_END and populate it with memory.
+ */
+void *module_alloc(unsigned long size)
+{
+	struct page **pages;
+	pgprot_t prot_rwx = __pgprot(_PAGE_KERNEL | _PAGE_KERNEL_EXEC);
+	struct vm_struct *area;
+	int i = 0;
+	int npages;
+
+	npages = (size + PAGE_SIZE - 1) / PAGE_SIZE;
+	pages = kmalloc(npages * sizeof(struct page *), GFP_KERNEL);
+	if (pages == NULL)
+		return NULL;
+	for (; i < npages; ++i) {
+		pages[i] = alloc_page(GFP_KERNEL | __GFP_HIGHMEM);
+		if (!pages[i])
+			goto error;
+	}
+
+	area = __get_vm_area(size, VM_ALLOC, MEM_MODULE_START, MEM_MODULE_END);
+	if (!area)
+		goto error;
+	area->nr_pages = npages;
+	area->pages = pages;
+
+	if (map_vm_area(area, prot_rwx, pages)) {
+		vunmap(area->addr);
+		goto error;
+	}
+
+	return area->addr;
+
+error:
+	while (--i >= 0)
+		__free_page(pages[i]);
+	kfree(pages);
+	return NULL;
+}
+
+
+/* Free memory returned from module_alloc */
+void module_memfree(void *module_region)
+{
+	vfree(module_region);
+
+	/* Globally flush the L1 icache. */
+	flush_remote(0, HV_FLUSH_EVICT_L1I, cpu_online_mask,
+		     0, 0, 0, NULL, NULL, 0);
+
+	/*
+	 * FIXME: Add module_arch_freeing_init to trim exception
+	 * table entries.
+	 */
+}
+
+#ifdef __tilegx__
+/*
+ * Validate that the high 16 bits of "value" is just the sign-extension of
+ * the low 48 bits.
+ */
+static int validate_hw2_last(long value, struct module *me)
+{
+	if (((value << 16) >> 16) != value) {
+		pr_warn("module %s: Out of range HW2_LAST value %#lx\n",
+			me->name, value);
+		return 0;
+	}
+	return 1;
+}
+
+/*
+ * Validate that "value" isn't too big to hold in a JumpOff relocation.
+ */
+static int validate_jumpoff(long value)
+{
+	/* Determine size of jump offset. */
+	int shift = __builtin_clzl(get_JumpOff_X1(create_JumpOff_X1(-1)));
+
+	/* Check to see if it fits into the relocation slot. */
+	long f = get_JumpOff_X1(create_JumpOff_X1(value));
+	f = (f << shift) >> shift;
+
+	return f == value;
+}
+#endif
+
+int apply_relocate_add(Elf_Shdr *sechdrs,
+		       const char *strtab,
+		       unsigned int symindex,
+		       unsigned int relsec,
+		       struct module *me)
+{
+	unsigned int i;
+	Elf_Rela *rel = (void *)sechdrs[relsec].sh_addr;
+	Elf_Sym *sym;
+	u64 *location;
+	unsigned long value;
+
+	DEBUGP("Applying relocate section %u to %u\n", relsec,
+	       sechdrs[relsec].sh_info);
+	for (i = 0; i < sechdrs[relsec].sh_size / sizeof(*rel); i++) {
+		/* This is where to make the change */
+		location = (void *)sechdrs[sechdrs[relsec].sh_info].sh_addr
+			+ rel[i].r_offset;
+		/*
+		 * This is the symbol it is referring to.
+		 * Note that all undefined symbols have been resolved.
+		 */
+		sym = (Elf_Sym *)sechdrs[symindex].sh_addr
+			+ ELF_R_SYM(rel[i].r_info);
+		value = sym->st_value + rel[i].r_addend;
+
+		switch (ELF_R_TYPE(rel[i].r_info)) {
+
+#ifdef __LITTLE_ENDIAN
+# define MUNGE(func) \
+	(*location = ((*location & ~func(-1)) | func(value)))
+#else
+/*
+ * Instructions are always little-endian, so when we read them as data,
+ * we have to swap them around before and after modifying them.
+ */
+# define MUNGE(func) \
+	(*location = swab64((swab64(*location) & ~func(-1)) | func(value)))
+#endif
+
+#ifndef __tilegx__
+		case R_TILE_32:
+			*(uint32_t *)location = value;
+			break;
+		case R_TILE_IMM16_X0_HA:
+			value = (value + 0x8000) >> 16;
+			/*FALLTHROUGH*/
+		case R_TILE_IMM16_X0_LO:
+			MUNGE(create_Imm16_X0);
+			break;
+		case R_TILE_IMM16_X1_HA:
+			value = (value + 0x8000) >> 16;
+			/*FALLTHROUGH*/
+		case R_TILE_IMM16_X1_LO:
+			MUNGE(create_Imm16_X1);
+			break;
+		case R_TILE_JOFFLONG_X1:
+			value -= (unsigned long) location;  /* pc-relative */
+			value = (long) value >> 3;     /* count by instrs */
+			MUNGE(create_JOffLong_X1);
+			break;
+#else
+		case R_TILEGX_64:
+			*location = value;
+			break;
+		case R_TILEGX_IMM16_X0_HW2_LAST:
+			if (!validate_hw2_last(value, me))
+				return -ENOEXEC;
+			value >>= 16;
+			/*FALLTHROUGH*/
+		case R_TILEGX_IMM16_X0_HW1:
+			value >>= 16;
+			/*FALLTHROUGH*/
+		case R_TILEGX_IMM16_X0_HW0:
+			MUNGE(create_Imm16_X0);
+			break;
+		case R_TILEGX_IMM16_X1_HW2_LAST:
+			if (!validate_hw2_last(value, me))
+				return -ENOEXEC;
+			value >>= 16;
+			/*FALLTHROUGH*/
+		case R_TILEGX_IMM16_X1_HW1:
+			value >>= 16;
+			/*FALLTHROUGH*/
+		case R_TILEGX_IMM16_X1_HW0:
+			MUNGE(create_Imm16_X1);
+			break;
+		case R_TILEGX_JUMPOFF_X1:
+			value -= (unsigned long) location;  /* pc-relative */
+			value = (long) value >> 3;     /* count by instrs */
+			if (!validate_jumpoff(value)) {
+				pr_warn("module %s: Out of range jump to %#llx at %#llx (%p)\n",
+					me->name,
+					sym->st_value + rel[i].r_addend,
+					rel[i].r_offset, location);
+				return -ENOEXEC;
+			}
+			MUNGE(create_JumpOff_X1);
+			break;
+#endif
+
+#undef MUNGE
+
+		default:
+			pr_err("module %s: Unknown relocation: %d\n",
+			       me->name, (int) ELF_R_TYPE(rel[i].r_info));
+			return -ENOEXEC;
+		}
+	}
+	return 0;
+}
diff --git a/kernel/arch/tile/kernel/pci-dma.c b/kernel/arch/tile/kernel/pci-dma.c
new file mode 100644
index 000000000..09b58703a
--- /dev/null
+++ b/kernel/arch/tile/kernel/pci-dma.c
@@ -0,0 +1,630 @@
+/*
+ * Copyright 2010 Tilera Corporation. All Rights Reserved.
+ *
+ *   This program is free software; you can redistribute it and/or
+ *   modify it under the terms of the GNU General Public License
+ *   as published by the Free Software Foundation, version 2.
+ *
+ *   This program is distributed in the hope that it will be useful, but
+ *   WITHOUT ANY WARRANTY; without even the implied warranty of
+ *   MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
+ *   NON INFRINGEMENT.  See the GNU General Public License for
+ *   more details.
+ */
+
+#include <linux/mm.h>
+#include <linux/dma-mapping.h>
+#include <linux/swiotlb.h>
+#include <linux/vmalloc.h>
+#include <linux/export.h>
+#include <asm/tlbflush.h>
+#include <asm/homecache.h>
+
+/* Generic DMA mapping functions: */
+
+/*
+ * Allocate what Linux calls "coherent" memory.  On TILEPro this is
+ * uncached memory; on TILE-Gx it is hash-for-home memory.
+ */
+#ifdef __tilepro__
+#define PAGE_HOME_DMA PAGE_HOME_UNCACHED
+#else
+#define PAGE_HOME_DMA PAGE_HOME_HASH
+#endif
+
+static void *tile_dma_alloc_coherent(struct device *dev, size_t size,
+				     dma_addr_t *dma_handle, gfp_t gfp,
+				     struct dma_attrs *attrs)
+{
+	u64 dma_mask = (dev && dev->coherent_dma_mask) ?
+		dev->coherent_dma_mask : DMA_BIT_MASK(32);
+	int node = dev ? dev_to_node(dev) : 0;
+	int order = get_order(size);
+	struct page *pg;
+	dma_addr_t addr;
+
+	gfp |= __GFP_ZERO;
+
+	/*
+	 * If the mask specifies that the memory be in the first 4 GB, then
+	 * we force the allocation to come from the DMA zone.  We also
+	 * force the node to 0 since that's the only node where the DMA
+	 * zone isn't empty.  If the mask size is smaller than 32 bits, we
+	 * may still not be able to guarantee a suitable memory address, in
+	 * which case we will return NULL.  But such devices are uncommon.
+	 */
+	if (dma_mask <= DMA_BIT_MASK(32)) {
+		gfp |= GFP_DMA;
+		node = 0;
+	}
+
+	pg = homecache_alloc_pages_node(node, gfp, order, PAGE_HOME_DMA);
+	if (pg == NULL)
+		return NULL;
+
+	addr = page_to_phys(pg);
+	if (addr + size > dma_mask) {
+		__homecache_free_pages(pg, order);
+		return NULL;
+	}
+
+	*dma_handle = addr;
+
+	return page_address(pg);
+}
+
+/*
+ * Free memory that was allocated with tile_dma_alloc_coherent.
+ */
+static void tile_dma_free_coherent(struct device *dev, size_t size,
+				   void *vaddr, dma_addr_t dma_handle,
+				   struct dma_attrs *attrs)
+{
+	homecache_free_pages((unsigned long)vaddr, get_order(size));
+}
+
+/*
+ * The map routines "map" the specified address range for DMA
+ * accesses.  The memory belongs to the device after this call is
+ * issued, until it is unmapped with dma_unmap_single.
+ *
+ * We don't need to do any mapping, we just flush the address range
+ * out of the cache and return a DMA address.
+ *
+ * The unmap routines do whatever is necessary before the processor
+ * accesses the memory again, and must be called before the driver
+ * touches the memory.  We can get away with a cache invalidate if we
+ * can count on nothing having been touched.
+ */
+
+/* Set up a single page for DMA access. */
+static void __dma_prep_page(struct page *page, unsigned long offset,
+			    size_t size, enum dma_data_direction direction)
+{
+	/*
+	 * Flush the page from cache if necessary.
+	 * On tilegx, data is delivered to hash-for-home L3; on tilepro,
+	 * data is delivered direct to memory.
+	 *
+	 * NOTE: If we were just doing DMA_TO_DEVICE we could optimize
+	 * this to be a "flush" not a "finv" and keep some of the
+	 * state in cache across the DMA operation, but it doesn't seem
+	 * worth creating the necessary flush_buffer_xxx() infrastructure.
+	 */
+	int home = page_home(page);
+	switch (home) {
+	case PAGE_HOME_HASH:
+#ifdef __tilegx__
+		return;
+#endif
+		break;
+	case PAGE_HOME_UNCACHED:
+#ifdef __tilepro__
+		return;
+#endif
+		break;
+	case PAGE_HOME_IMMUTABLE:
+		/* Should be going to the device only. */
+		BUG_ON(direction == DMA_FROM_DEVICE ||
+		       direction == DMA_BIDIRECTIONAL);
+		return;
+	case PAGE_HOME_INCOHERENT:
+		/* Incoherent anyway, so no need to work hard here. */
+		return;
+	default:
+		BUG_ON(home < 0 || home >= NR_CPUS);
+		break;
+	}
+	homecache_finv_page(page);
+
+#ifdef DEBUG_ALIGNMENT
+	/* Warn if the region isn't cacheline aligned. */
+	if (offset & (L2_CACHE_BYTES - 1) || (size & (L2_CACHE_BYTES - 1)))
+		pr_warn("Unaligned DMA to non-hfh memory: PA %#llx/%#lx\n",
+			PFN_PHYS(page_to_pfn(page)) + offset, size);
+#endif
+}
+
+/* Make the page ready to be read by the core. */
+static void __dma_complete_page(struct page *page, unsigned long offset,
+				size_t size, enum dma_data_direction direction)
+{
+#ifdef __tilegx__
+	switch (page_home(page)) {
+	case PAGE_HOME_HASH:
+		/* I/O device delivered data the way the cpu wanted it. */
+		break;
+	case PAGE_HOME_INCOHERENT:
+		/* Incoherent anyway, so no need to work hard here. */
+		break;
+	case PAGE_HOME_IMMUTABLE:
+		/* Extra read-only copies are not a problem. */
+		break;
+	default:
+		/* Flush the bogus hash-for-home I/O entries to memory. */
+		homecache_finv_map_page(page, PAGE_HOME_HASH);
+		break;
+	}
+#endif
+}
+
+static void __dma_prep_pa_range(dma_addr_t dma_addr, size_t size,
+				enum dma_data_direction direction)
+{
+	struct page *page = pfn_to_page(PFN_DOWN(dma_addr));
+	unsigned long offset = dma_addr & (PAGE_SIZE - 1);
+	size_t bytes = min(size, (size_t)(PAGE_SIZE - offset));
+
+	while (size != 0) {
+		__dma_prep_page(page, offset, bytes, direction);
+		size -= bytes;
+		++page;
+		offset = 0;
+		bytes = min((size_t)PAGE_SIZE, size);
+	}
+}
+
+static void __dma_complete_pa_range(dma_addr_t dma_addr, size_t size,
+				    enum dma_data_direction direction)
+{
+	struct page *page = pfn_to_page(PFN_DOWN(dma_addr));
+	unsigned long offset = dma_addr & (PAGE_SIZE - 1);
+	size_t bytes = min(size, (size_t)(PAGE_SIZE - offset));
+
+	while (size != 0) {
+		__dma_complete_page(page, offset, bytes, direction);
+		size -= bytes;
+		++page;
+		offset = 0;
+		bytes = min((size_t)PAGE_SIZE, size);
+	}
+}
+
+static int tile_dma_map_sg(struct device *dev, struct scatterlist *sglist,
+			   int nents, enum dma_data_direction direction,
+			   struct dma_attrs *attrs)
+{
+	struct scatterlist *sg;
+	int i;
+
+	BUG_ON(!valid_dma_direction(direction));
+
+	WARN_ON(nents == 0 || sglist->length == 0);
+
+	for_each_sg(sglist, sg, nents, i) {
+		sg->dma_address = sg_phys(sg);
+		__dma_prep_pa_range(sg->dma_address, sg->length, direction);
+#ifdef CONFIG_NEED_SG_DMA_LENGTH
+		sg->dma_length = sg->length;
+#endif
+	}
+
+	return nents;
+}
+
+static void tile_dma_unmap_sg(struct device *dev, struct scatterlist *sglist,
+			      int nents, enum dma_data_direction direction,
+			      struct dma_attrs *attrs)
+{
+	struct scatterlist *sg;
+	int i;
+
+	BUG_ON(!valid_dma_direction(direction));
+	for_each_sg(sglist, sg, nents, i) {
+		sg->dma_address = sg_phys(sg);
+		__dma_complete_pa_range(sg->dma_address, sg->length,
+					direction);
+	}
+}
+
+static dma_addr_t tile_dma_map_page(struct device *dev, struct page *page,
+				    unsigned long offset, size_t size,
+				    enum dma_data_direction direction,
+				    struct dma_attrs *attrs)
+{
+	BUG_ON(!valid_dma_direction(direction));
+
+	BUG_ON(offset + size > PAGE_SIZE);
+	__dma_prep_page(page, offset, size, direction);
+
+	return page_to_pa(page) + offset;
+}
+
+static void tile_dma_unmap_page(struct device *dev, dma_addr_t dma_address,
+				size_t size, enum dma_data_direction direction,
+				struct dma_attrs *attrs)
+{
+	BUG_ON(!valid_dma_direction(direction));
+
+	__dma_complete_page(pfn_to_page(PFN_DOWN(dma_address)),
+			    dma_address & (PAGE_SIZE - 1), size, direction);
+}
+
+static void tile_dma_sync_single_for_cpu(struct device *dev,
+					 dma_addr_t dma_handle,
+					 size_t size,
+					 enum dma_data_direction direction)
+{
+	BUG_ON(!valid_dma_direction(direction));
+
+	__dma_complete_pa_range(dma_handle, size, direction);
+}
+
+static void tile_dma_sync_single_for_device(struct device *dev,
+					    dma_addr_t dma_handle, size_t size,
+					    enum dma_data_direction direction)
+{
+	__dma_prep_pa_range(dma_handle, size, direction);
+}
+
+static void tile_dma_sync_sg_for_cpu(struct device *dev,
+				     struct scatterlist *sglist, int nelems,
+				     enum dma_data_direction direction)
+{
+	struct scatterlist *sg;
+	int i;
+
+	BUG_ON(!valid_dma_direction(direction));
+	WARN_ON(nelems == 0 || sglist->length == 0);
+
+	for_each_sg(sglist, sg, nelems, i) {
+		dma_sync_single_for_cpu(dev, sg->dma_address,
+					sg_dma_len(sg), direction);
+	}
+}
+
+static void tile_dma_sync_sg_for_device(struct device *dev,
+					struct scatterlist *sglist, int nelems,
+					enum dma_data_direction direction)
+{
+	struct scatterlist *sg;
+	int i;
+
+	BUG_ON(!valid_dma_direction(direction));
+	WARN_ON(nelems == 0 || sglist->length == 0);
+
+	for_each_sg(sglist, sg, nelems, i) {
+		dma_sync_single_for_device(dev, sg->dma_address,
+					   sg_dma_len(sg), direction);
+	}
+}
+
+static inline int
+tile_dma_mapping_error(struct device *dev, dma_addr_t dma_addr)
+{
+	return 0;
+}
+
+static inline int
+tile_dma_supported(struct device *dev, u64 mask)
+{
+	return 1;
+}
+
+static struct dma_map_ops tile_default_dma_map_ops = {
+	.alloc = tile_dma_alloc_coherent,
+	.free = tile_dma_free_coherent,
+	.map_page = tile_dma_map_page,
+	.unmap_page = tile_dma_unmap_page,
+	.map_sg = tile_dma_map_sg,
+	.unmap_sg = tile_dma_unmap_sg,
+	.sync_single_for_cpu = tile_dma_sync_single_for_cpu,
+	.sync_single_for_device = tile_dma_sync_single_for_device,
+	.sync_sg_for_cpu = tile_dma_sync_sg_for_cpu,
+	.sync_sg_for_device = tile_dma_sync_sg_for_device,
+	.mapping_error = tile_dma_mapping_error,
+	.dma_supported = tile_dma_supported
+};
+
+struct dma_map_ops *tile_dma_map_ops = &tile_default_dma_map_ops;
+EXPORT_SYMBOL(tile_dma_map_ops);
+
+/* Generic PCI DMA mapping functions */
+
+static void *tile_pci_dma_alloc_coherent(struct device *dev, size_t size,
+					 dma_addr_t *dma_handle, gfp_t gfp,
+					 struct dma_attrs *attrs)
+{
+	int node = dev_to_node(dev);
+	int order = get_order(size);
+	struct page *pg;
+	dma_addr_t addr;
+
+	gfp |= __GFP_ZERO;
+
+	pg = homecache_alloc_pages_node(node, gfp, order, PAGE_HOME_DMA);
+	if (pg == NULL)
+		return NULL;
+
+	addr = page_to_phys(pg);
+
+	*dma_handle = addr + get_dma_offset(dev);
+
+	return page_address(pg);
+}
+
+/*
+ * Free memory that was allocated with tile_pci_dma_alloc_coherent.
+ */
+static void tile_pci_dma_free_coherent(struct device *dev, size_t size,
+				       void *vaddr, dma_addr_t dma_handle,
+				       struct dma_attrs *attrs)
+{
+	homecache_free_pages((unsigned long)vaddr, get_order(size));
+}
+
+static int tile_pci_dma_map_sg(struct device *dev, struct scatterlist *sglist,
+			       int nents, enum dma_data_direction direction,
+			       struct dma_attrs *attrs)
+{
+	struct scatterlist *sg;
+	int i;
+
+	BUG_ON(!valid_dma_direction(direction));
+
+	WARN_ON(nents == 0 || sglist->length == 0);
+
+	for_each_sg(sglist, sg, nents, i) {
+		sg->dma_address = sg_phys(sg);
+		__dma_prep_pa_range(sg->dma_address, sg->length, direction);
+
+		sg->dma_address = sg->dma_address + get_dma_offset(dev);
+#ifdef CONFIG_NEED_SG_DMA_LENGTH
+		sg->dma_length = sg->length;
+#endif
+	}
+
+	return nents;
+}
+
+static void tile_pci_dma_unmap_sg(struct device *dev,
+				  struct scatterlist *sglist, int nents,
+				  enum dma_data_direction direction,
+				  struct dma_attrs *attrs)
+{
+	struct scatterlist *sg;
+	int i;
+
+	BUG_ON(!valid_dma_direction(direction));
+	for_each_sg(sglist, sg, nents, i) {
+		sg->dma_address = sg_phys(sg);
+		__dma_complete_pa_range(sg->dma_address, sg->length,
+					direction);
+	}
+}
+
+static dma_addr_t tile_pci_dma_map_page(struct device *dev, struct page *page,
+					unsigned long offset, size_t size,
+					enum dma_data_direction direction,
+					struct dma_attrs *attrs)
+{
+	BUG_ON(!valid_dma_direction(direction));
+
+	BUG_ON(offset + size > PAGE_SIZE);
+	__dma_prep_page(page, offset, size, direction);
+
+	return page_to_pa(page) + offset + get_dma_offset(dev);
+}
+
+static void tile_pci_dma_unmap_page(struct device *dev, dma_addr_t dma_address,
+				    size_t size,
+				    enum dma_data_direction direction,
+				    struct dma_attrs *attrs)
+{
+	BUG_ON(!valid_dma_direction(direction));
+
+	dma_address -= get_dma_offset(dev);
+
+	__dma_complete_page(pfn_to_page(PFN_DOWN(dma_address)),
+			    dma_address & (PAGE_SIZE - 1), size, direction);
+}
+
+static void tile_pci_dma_sync_single_for_cpu(struct device *dev,
+					     dma_addr_t dma_handle,
+					     size_t size,
+					     enum dma_data_direction direction)
+{
+	BUG_ON(!valid_dma_direction(direction));
+
+	dma_handle -= get_dma_offset(dev);
+
+	__dma_complete_pa_range(dma_handle, size, direction);
+}
+
+static void tile_pci_dma_sync_single_for_device(struct device *dev,
+						dma_addr_t dma_handle,
+						size_t size,
+						enum dma_data_direction
+						direction)
+{
+	dma_handle -= get_dma_offset(dev);
+
+	__dma_prep_pa_range(dma_handle, size, direction);
+}
+
+static void tile_pci_dma_sync_sg_for_cpu(struct device *dev,
+					 struct scatterlist *sglist,
+					 int nelems,
+					 enum dma_data_direction direction)
+{
+	struct scatterlist *sg;
+	int i;
+
+	BUG_ON(!valid_dma_direction(direction));
+	WARN_ON(nelems == 0 || sglist->length == 0);
+
+	for_each_sg(sglist, sg, nelems, i) {
+		dma_sync_single_for_cpu(dev, sg->dma_address,
+					sg_dma_len(sg), direction);
+	}
+}
+
+static void tile_pci_dma_sync_sg_for_device(struct device *dev,
+					    struct scatterlist *sglist,
+					    int nelems,
+					    enum dma_data_direction direction)
+{
+	struct scatterlist *sg;
+	int i;
+
+	BUG_ON(!valid_dma_direction(direction));
+	WARN_ON(nelems == 0 || sglist->length == 0);
+
+	for_each_sg(sglist, sg, nelems, i) {
+		dma_sync_single_for_device(dev, sg->dma_address,
+					   sg_dma_len(sg), direction);
+	}
+}
+
+static inline int
+tile_pci_dma_mapping_error(struct device *dev, dma_addr_t dma_addr)
+{
+	return 0;
+}
+
+static inline int
+tile_pci_dma_supported(struct device *dev, u64 mask)
+{
+	return 1;
+}
+
+static struct dma_map_ops tile_pci_default_dma_map_ops = {
+	.alloc = tile_pci_dma_alloc_coherent,
+	.free = tile_pci_dma_free_coherent,
+	.map_page = tile_pci_dma_map_page,
+	.unmap_page = tile_pci_dma_unmap_page,
+	.map_sg = tile_pci_dma_map_sg,
+	.unmap_sg = tile_pci_dma_unmap_sg,
+	.sync_single_for_cpu = tile_pci_dma_sync_single_for_cpu,
+	.sync_single_for_device = tile_pci_dma_sync_single_for_device,
+	.sync_sg_for_cpu = tile_pci_dma_sync_sg_for_cpu,
+	.sync_sg_for_device = tile_pci_dma_sync_sg_for_device,
+	.mapping_error = tile_pci_dma_mapping_error,
+	.dma_supported = tile_pci_dma_supported
+};
+
+struct dma_map_ops *gx_pci_dma_map_ops = &tile_pci_default_dma_map_ops;
+EXPORT_SYMBOL(gx_pci_dma_map_ops);
+
+/* PCI DMA mapping functions for legacy PCI devices */
+
+#ifdef CONFIG_SWIOTLB
+static void *tile_swiotlb_alloc_coherent(struct device *dev, size_t size,
+					 dma_addr_t *dma_handle, gfp_t gfp,
+					 struct dma_attrs *attrs)
+{
+	gfp |= GFP_DMA;
+	return swiotlb_alloc_coherent(dev, size, dma_handle, gfp);
+}
+
+static void tile_swiotlb_free_coherent(struct device *dev, size_t size,
+				       void *vaddr, dma_addr_t dma_addr,
+				       struct dma_attrs *attrs)
+{
+	swiotlb_free_coherent(dev, size, vaddr, dma_addr);
+}
+
+static struct dma_map_ops pci_swiotlb_dma_ops = {
+	.alloc = tile_swiotlb_alloc_coherent,
+	.free = tile_swiotlb_free_coherent,
+	.map_page = swiotlb_map_page,
+	.unmap_page = swiotlb_unmap_page,
+	.map_sg = swiotlb_map_sg_attrs,
+	.unmap_sg = swiotlb_unmap_sg_attrs,
+	.sync_single_for_cpu = swiotlb_sync_single_for_cpu,
+	.sync_single_for_device = swiotlb_sync_single_for_device,
+	.sync_sg_for_cpu = swiotlb_sync_sg_for_cpu,
+	.sync_sg_for_device = swiotlb_sync_sg_for_device,
+	.dma_supported = swiotlb_dma_supported,
+	.mapping_error = swiotlb_dma_mapping_error,
+};
+
+static struct dma_map_ops pci_hybrid_dma_ops = {
+	.alloc = tile_swiotlb_alloc_coherent,
+	.free = tile_swiotlb_free_coherent,
+	.map_page = tile_pci_dma_map_page,
+	.unmap_page = tile_pci_dma_unmap_page,
+	.map_sg = tile_pci_dma_map_sg,
+	.unmap_sg = tile_pci_dma_unmap_sg,
+	.sync_single_for_cpu = tile_pci_dma_sync_single_for_cpu,
+	.sync_single_for_device = tile_pci_dma_sync_single_for_device,
+	.sync_sg_for_cpu = tile_pci_dma_sync_sg_for_cpu,
+	.sync_sg_for_device = tile_pci_dma_sync_sg_for_device,
+	.mapping_error = tile_pci_dma_mapping_error,
+	.dma_supported = tile_pci_dma_supported
+};
+
+struct dma_map_ops *gx_legacy_pci_dma_map_ops = &pci_swiotlb_dma_ops;
+struct dma_map_ops *gx_hybrid_pci_dma_map_ops = &pci_hybrid_dma_ops;
+#else
+struct dma_map_ops *gx_legacy_pci_dma_map_ops;
+struct dma_map_ops *gx_hybrid_pci_dma_map_ops;
+#endif
+EXPORT_SYMBOL(gx_legacy_pci_dma_map_ops);
+EXPORT_SYMBOL(gx_hybrid_pci_dma_map_ops);
+
+#ifdef CONFIG_ARCH_HAS_DMA_SET_COHERENT_MASK
+int dma_set_coherent_mask(struct device *dev, u64 mask)
+{
+	struct dma_map_ops *dma_ops = get_dma_ops(dev);
+
+	/*
+	 * For PCI devices with 64-bit DMA addressing capability, promote
+	 * the dma_ops to full capability for both streams and consistent
+	 * memory access. For 32-bit capable devices, limit the consistent 
+	 * memory DMA range to max_direct_dma_addr.
+	 */
+	if (dma_ops == gx_pci_dma_map_ops ||
+	    dma_ops == gx_hybrid_pci_dma_map_ops ||
+	    dma_ops == gx_legacy_pci_dma_map_ops) {
+		if (mask == DMA_BIT_MASK(64))
+			set_dma_ops(dev, gx_pci_dma_map_ops);
+		else if (mask > dev->archdata.max_direct_dma_addr)
+			mask = dev->archdata.max_direct_dma_addr;
+	}
+
+	if (!dma_supported(dev, mask))
+		return -EIO;
+	dev->coherent_dma_mask = mask;
+	return 0;
+}
+EXPORT_SYMBOL(dma_set_coherent_mask);
+#endif
+
+#ifdef ARCH_HAS_DMA_GET_REQUIRED_MASK
+/*
+ * The generic dma_get_required_mask() uses the highest physical address
+ * (max_pfn) to provide the hint to the PCI drivers regarding 32-bit or
+ * 64-bit DMA configuration. Since TILEGx has I/O TLB/MMU, allowing the
+ * DMAs to use the full 64-bit PCI address space and not limited by
+ * the physical memory space, we always let the PCI devices use
+ * 64-bit DMA if they have that capability, by returning the 64-bit
+ * DMA mask here. The device driver has the option to use 32-bit DMA if
+ * the device is not capable of 64-bit DMA.
+ */
+u64 dma_get_required_mask(struct device *dev)
+{
+	return DMA_BIT_MASK(64);
+}
+EXPORT_SYMBOL_GPL(dma_get_required_mask);
+#endif
diff --git a/kernel/arch/tile/kernel/pci.c b/kernel/arch/tile/kernel/pci.c
new file mode 100644
index 000000000..9475a74cd
--- /dev/null
+++ b/kernel/arch/tile/kernel/pci.c
@@ -0,0 +1,599 @@
+/*
+ * Copyright 2011 Tilera Corporation. All Rights Reserved.
+ *
+ *   This program is free software; you can redistribute it and/or
+ *   modify it under the terms of the GNU General Public License
+ *   as published by the Free Software Foundation, version 2.
+ *
+ *   This program is distributed in the hope that it will be useful, but
+ *   WITHOUT ANY WARRANTY; without even the implied warranty of
+ *   MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
+ *   NON INFRINGEMENT.  See the GNU General Public License for
+ *   more details.
+ */
+
+#include <linux/kernel.h>
+#include <linux/pci.h>
+#include <linux/delay.h>
+#include <linux/string.h>
+#include <linux/init.h>
+#include <linux/capability.h>
+#include <linux/sched.h>
+#include <linux/errno.h>
+#include <linux/irq.h>
+#include <linux/io.h>
+#include <linux/uaccess.h>
+#include <linux/export.h>
+
+#include <asm/processor.h>
+#include <asm/sections.h>
+#include <asm/byteorder.h>
+#include <asm/hv_driver.h>
+#include <hv/drv_pcie_rc_intf.h>
+
+
+/*
+ * Initialization flow and process
+ * -------------------------------
+ *
+ * This files contains the routines to search for PCI buses,
+ * enumerate the buses, and configure any attached devices.
+ *
+ * There are two entry points here:
+ * 1) tile_pci_init
+ *    This sets up the pci_controller structs, and opens the
+ *    FDs to the hypervisor.  This is called from setup_arch() early
+ *    in the boot process.
+ * 2) pcibios_init
+ *    This probes the PCI bus(es) for any attached hardware.  It's
+ *    called by subsys_initcall.  All of the real work is done by the
+ *    generic Linux PCI layer.
+ *
+ */
+
+static int pci_probe = 1;
+
+/*
+ * This flag tells if the platform is TILEmpower that needs
+ * special configuration for the PLX switch chip.
+ */
+int __write_once tile_plx_gen1;
+
+static struct pci_controller controllers[TILE_NUM_PCIE];
+static int num_controllers;
+static int pci_scan_flags[TILE_NUM_PCIE];
+
+static struct pci_ops tile_cfg_ops;
+
+
+/*
+ * We don't need to worry about the alignment of resources.
+ */
+resource_size_t pcibios_align_resource(void *data, const struct resource *res,
+			    resource_size_t size, resource_size_t align)
+{
+	return res->start;
+}
+EXPORT_SYMBOL(pcibios_align_resource);
+
+/*
+ * Open a FD to the hypervisor PCI device.
+ *
+ * controller_id is the controller number, config type is 0 or 1 for
+ * config0 or config1 operations.
+ */
+static int tile_pcie_open(int controller_id, int config_type)
+{
+	char filename[32];
+	int fd;
+
+	sprintf(filename, "pcie/%d/config%d", controller_id, config_type);
+
+	fd = hv_dev_open((HV_VirtAddr)filename, 0);
+
+	return fd;
+}
+
+
+/*
+ * Get the IRQ numbers from the HV and set up the handlers for them.
+ */
+static int tile_init_irqs(int controller_id, struct pci_controller *controller)
+{
+	char filename[32];
+	int fd;
+	int ret;
+	int x;
+	struct pcie_rc_config rc_config;
+
+	sprintf(filename, "pcie/%d/ctl", controller_id);
+	fd = hv_dev_open((HV_VirtAddr)filename, 0);
+	if (fd < 0) {
+		pr_err("PCI: hv_dev_open(%s) failed\n", filename);
+		return -1;
+	}
+	ret = hv_dev_pread(fd, 0, (HV_VirtAddr)(&rc_config),
+			   sizeof(rc_config), PCIE_RC_CONFIG_MASK_OFF);
+	hv_dev_close(fd);
+	if (ret != sizeof(rc_config)) {
+		pr_err("PCI: wanted %zd bytes, got %d\n",
+		       sizeof(rc_config), ret);
+		return -1;
+	}
+	/* Record irq_base so that we can map INTx to IRQ # later. */
+	controller->irq_base = rc_config.intr;
+
+	for (x = 0; x < 4; x++)
+		tile_irq_activate(rc_config.intr + x,
+				  TILE_IRQ_HW_CLEAR);
+
+	if (rc_config.plx_gen1)
+		controller->plx_gen1 = 1;
+
+	return 0;
+}
+
+/*
+ * First initialization entry point, called from setup_arch().
+ *
+ * Find valid controllers and fill in pci_controller structs for each
+ * of them.
+ *
+ * Returns the number of controllers discovered.
+ */
+int __init tile_pci_init(void)
+{
+	int i;
+
+	if (!pci_probe) {
+		pr_info("PCI: disabled by boot argument\n");
+		return 0;
+	}
+
+	pr_info("PCI: Searching for controllers...\n");
+
+	/* Re-init number of PCIe controllers to support hot-plug feature. */
+	num_controllers = 0;
+
+	/* Do any configuration we need before using the PCIe */
+
+	for (i = 0; i < TILE_NUM_PCIE; i++) {
+		/*
+		 * To see whether we need a real config op based on
+		 * the results of pcibios_init(), to support PCIe hot-plug.
+		 */
+		if (pci_scan_flags[i] == 0) {
+			int hv_cfg_fd0 = -1;
+			int hv_cfg_fd1 = -1;
+			int hv_mem_fd = -1;
+			char name[32];
+			struct pci_controller *controller;
+
+			/*
+			 * Open the fd to the HV.  If it fails then this
+			 * device doesn't exist.
+			 */
+			hv_cfg_fd0 = tile_pcie_open(i, 0);
+			if (hv_cfg_fd0 < 0)
+				continue;
+			hv_cfg_fd1 = tile_pcie_open(i, 1);
+			if (hv_cfg_fd1 < 0) {
+				pr_err("PCI: Couldn't open config fd to HV for controller %d\n",
+				       i);
+				goto err_cont;
+			}
+
+			sprintf(name, "pcie/%d/mem", i);
+			hv_mem_fd = hv_dev_open((HV_VirtAddr)name, 0);
+			if (hv_mem_fd < 0) {
+				pr_err("PCI: Could not open mem fd to HV!\n");
+				goto err_cont;
+			}
+
+			pr_info("PCI: Found PCI controller #%d\n", i);
+
+			controller = &controllers[i];
+
+			controller->index = i;
+			controller->hv_cfg_fd[0] = hv_cfg_fd0;
+			controller->hv_cfg_fd[1] = hv_cfg_fd1;
+			controller->hv_mem_fd = hv_mem_fd;
+			controller->last_busno = 0xff;
+			controller->ops = &tile_cfg_ops;
+
+			num_controllers++;
+			continue;
+
+err_cont:
+			if (hv_cfg_fd0 >= 0)
+				hv_dev_close(hv_cfg_fd0);
+			if (hv_cfg_fd1 >= 0)
+				hv_dev_close(hv_cfg_fd1);
+			if (hv_mem_fd >= 0)
+				hv_dev_close(hv_mem_fd);
+			continue;
+		}
+	}
+
+	/*
+	 * Before using the PCIe, see if we need to do any platform-specific
+	 * configuration, such as the PLX switch Gen 1 issue on TILEmpower.
+	 */
+	for (i = 0; i < num_controllers; i++) {
+		struct pci_controller *controller = &controllers[i];
+
+		if (controller->plx_gen1)
+			tile_plx_gen1 = 1;
+	}
+
+	return num_controllers;
+}
+
+/*
+ * (pin - 1) converts from the PCI standard's [1:4] convention to
+ * a normal [0:3] range.
+ */
+static int tile_map_irq(const struct pci_dev *dev, u8 slot, u8 pin)
+{
+	struct pci_controller *controller =
+		(struct pci_controller *)dev->sysdata;
+	return (pin - 1) + controller->irq_base;
+}
+
+
+static void fixup_read_and_payload_sizes(void)
+{
+	struct pci_dev *dev = NULL;
+	int smallest_max_payload = 0x1; /* Tile maxes out at 256 bytes. */
+	int max_read_size = PCI_EXP_DEVCTL_READRQ_512B;
+	u16 new_values;
+
+	/* Scan for the smallest maximum payload size. */
+	for_each_pci_dev(dev) {
+		if (!pci_is_pcie(dev))
+			continue;
+
+		if (dev->pcie_mpss < smallest_max_payload)
+			smallest_max_payload = dev->pcie_mpss;
+	}
+
+	/* Now, set the max_payload_size for all devices to that value. */
+	new_values = max_read_size | (smallest_max_payload << 5);
+	for_each_pci_dev(dev)
+		pcie_capability_clear_and_set_word(dev, PCI_EXP_DEVCTL,
+				PCI_EXP_DEVCTL_PAYLOAD | PCI_EXP_DEVCTL_READRQ,
+				new_values);
+}
+
+
+/*
+ * Second PCI initialization entry point, called by subsys_initcall.
+ *
+ * The controllers have been set up by the time we get here, by a call to
+ * tile_pci_init.
+ */
+int __init pcibios_init(void)
+{
+	int i;
+
+	pr_info("PCI: Probing PCI hardware\n");
+
+	/*
+	 * Delay a bit in case devices aren't ready.  Some devices are
+	 * known to require at least 20ms here, but we use a more
+	 * conservative value.
+	 */
+	msleep(250);
+
+	/* Scan all of the recorded PCI controllers.  */
+	for (i = 0; i < TILE_NUM_PCIE; i++) {
+		/*
+		 * Do real pcibios init ops if the controller is initialized
+		 * by tile_pci_init() successfully and not initialized by
+		 * pcibios_init() yet to support PCIe hot-plug.
+		 */
+		if (pci_scan_flags[i] == 0 && controllers[i].ops != NULL) {
+			struct pci_controller *controller = &controllers[i];
+			struct pci_bus *bus;
+			LIST_HEAD(resources);
+
+			if (tile_init_irqs(i, controller)) {
+				pr_err("PCI: Could not initialize IRQs\n");
+				continue;
+			}
+
+			pr_info("PCI: initializing controller #%d\n", i);
+
+			pci_add_resource(&resources, &ioport_resource);
+			pci_add_resource(&resources, &iomem_resource);
+			bus = pci_scan_root_bus(NULL, 0, controller->ops,
+						controller, &resources);
+			controller->root_bus = bus;
+			controller->last_busno = bus->busn_res.end;
+		}
+	}
+
+	/* Do machine dependent PCI interrupt routing */
+	pci_fixup_irqs(pci_common_swizzle, tile_map_irq);
+
+	/*
+	 * This comes from the generic Linux PCI driver.
+	 *
+	 * It allocates all of the resources (I/O memory, etc)
+	 * associated with the devices read in above.
+	 */
+	pci_assign_unassigned_resources();
+
+	/* Configure the max_read_size and max_payload_size values. */
+	fixup_read_and_payload_sizes();
+
+	/* Record the I/O resources in the PCI controller structure. */
+	for (i = 0; i < TILE_NUM_PCIE; i++) {
+		/*
+		 * Do real pcibios init ops if the controller is initialized
+		 * by tile_pci_init() successfully and not initialized by
+		 * pcibios_init() yet to support PCIe hot-plug.
+		 */
+		if (pci_scan_flags[i] == 0 && controllers[i].ops != NULL) {
+			struct pci_bus *root_bus = controllers[i].root_bus;
+			struct pci_bus *next_bus;
+			struct pci_dev *dev;
+
+			pci_bus_add_devices(root_bus);
+
+			list_for_each_entry(dev, &root_bus->devices, bus_list) {
+				/*
+				 * Find the PCI host controller, ie. the 1st
+				 * bridge.
+				 */
+				if ((dev->class >> 8) == PCI_CLASS_BRIDGE_PCI &&
+					(PCI_SLOT(dev->devfn) == 0)) {
+					next_bus = dev->subordinate;
+					controllers[i].mem_resources[0] =
+						*next_bus->resource[0];
+					controllers[i].mem_resources[1] =
+						 *next_bus->resource[1];
+					controllers[i].mem_resources[2] =
+						 *next_bus->resource[2];
+
+					/* Setup flags. */
+					pci_scan_flags[i] = 1;
+
+					break;
+				}
+			}
+		}
+	}
+
+	return 0;
+}
+subsys_initcall(pcibios_init);
+
+/*
+ * No bus fixups needed.
+ */
+void pcibios_fixup_bus(struct pci_bus *bus)
+{
+	/* Nothing needs to be done. */
+}
+
+void pcibios_set_master(struct pci_dev *dev)
+{
+	/* No special bus mastering setup handling. */
+}
+
+/* Process any "pci=" kernel boot arguments. */
+char *__init pcibios_setup(char *str)
+{
+	if (!strcmp(str, "off")) {
+		pci_probe = 0;
+		return NULL;
+	}
+	return str;
+}
+
+/*
+ * Enable memory and/or address decoding, as appropriate, for the
+ * device described by the 'dev' struct.
+ *
+ * This is called from the generic PCI layer, and can be called
+ * for bridges or endpoints.
+ */
+int pcibios_enable_device(struct pci_dev *dev, int mask)
+{
+	u16 cmd, old_cmd;
+	u8 header_type;
+	int i;
+	struct resource *r;
+
+	pci_read_config_byte(dev, PCI_HEADER_TYPE, &header_type);
+
+	pci_read_config_word(dev, PCI_COMMAND, &cmd);
+	old_cmd = cmd;
+	if ((header_type & 0x7F) == PCI_HEADER_TYPE_BRIDGE) {
+		/*
+		 * For bridges, we enable both memory and I/O decoding
+		 * in call cases.
+		 */
+		cmd |= PCI_COMMAND_IO;
+		cmd |= PCI_COMMAND_MEMORY;
+	} else {
+		/*
+		 * For endpoints, we enable memory and/or I/O decoding
+		 * only if they have a memory resource of that type.
+		 */
+		for (i = 0; i < 6; i++) {
+			r = &dev->resource[i];
+			if (r->flags & IORESOURCE_UNSET) {
+				pr_err("PCI: Device %s not available because of resource collisions\n",
+				       pci_name(dev));
+				return -EINVAL;
+			}
+			if (r->flags & IORESOURCE_IO)
+				cmd |= PCI_COMMAND_IO;
+			if (r->flags & IORESOURCE_MEM)
+				cmd |= PCI_COMMAND_MEMORY;
+		}
+	}
+
+	/*
+	 * We only write the command if it changed.
+	 */
+	if (cmd != old_cmd)
+		pci_write_config_word(dev, PCI_COMMAND, cmd);
+	return 0;
+}
+
+/****************************************************************
+ *
+ * Tile PCI config space read/write routines
+ *
+ ****************************************************************/
+
+/*
+ * These are the normal read and write ops
+ * These are expanded with macros from  pci_bus_read_config_byte() etc.
+ *
+ * devfn is the combined PCI slot & function.
+ *
+ * offset is in bytes, from the start of config space for the
+ * specified bus & slot.
+ */
+
+static int tile_cfg_read(struct pci_bus *bus, unsigned int devfn, int offset,
+			 int size, u32 *val)
+{
+	struct pci_controller *controller = bus->sysdata;
+	int busnum = bus->number & 0xff;
+	int slot = (devfn >> 3) & 0x1f;
+	int function = devfn & 0x7;
+	u32 addr;
+	int config_mode = 1;
+
+	/*
+	 * There is no bridge between the Tile and bus 0, so we
+	 * use config0 to talk to bus 0.
+	 *
+	 * If we're talking to a bus other than zero then we
+	 * must have found a bridge.
+	 */
+	if (busnum == 0) {
+		/*
+		 * We fake an empty slot for (busnum == 0) && (slot > 0),
+		 * since there is only one slot on bus 0.
+		 */
+		if (slot) {
+			*val = 0xFFFFFFFF;
+			return 0;
+		}
+		config_mode = 0;
+	}
+
+	addr = busnum << 20;		/* Bus in 27:20 */
+	addr |= slot << 15;		/* Slot (device) in 19:15 */
+	addr |= function << 12;		/* Function is in 14:12 */
+	addr |= (offset & 0xFFF);	/* byte address in 0:11 */
+
+	return hv_dev_pread(controller->hv_cfg_fd[config_mode], 0,
+			    (HV_VirtAddr)(val), size, addr);
+}
+
+
+/*
+ * See tile_cfg_read() for relevant comments.
+ * Note that "val" is the value to write, not a pointer to that value.
+ */
+static int tile_cfg_write(struct pci_bus *bus, unsigned int devfn, int offset,
+			  int size, u32 val)
+{
+	struct pci_controller *controller = bus->sysdata;
+	int busnum = bus->number & 0xff;
+	int slot = (devfn >> 3) & 0x1f;
+	int function = devfn & 0x7;
+	u32 addr;
+	int config_mode = 1;
+	HV_VirtAddr valp = (HV_VirtAddr)&val;
+
+	/*
+	 * For bus 0 slot 0 we use config 0 accesses.
+	 */
+	if (busnum == 0) {
+		/*
+		 * We fake an empty slot for (busnum == 0) && (slot > 0),
+		 * since there is only one slot on bus 0.
+		 */
+		if (slot)
+			return 0;
+		config_mode = 0;
+	}
+
+	addr = busnum << 20;		/* Bus in 27:20 */
+	addr |= slot << 15;		/* Slot (device) in 19:15 */
+	addr |= function << 12;		/* Function is in 14:12 */
+	addr |= (offset & 0xFFF);	/* byte address in 0:11 */
+
+#ifdef __BIG_ENDIAN
+	/* Point to the correct part of the 32-bit "val". */
+	valp += 4 - size;
+#endif
+
+	return hv_dev_pwrite(controller->hv_cfg_fd[config_mode], 0,
+			     valp, size, addr);
+}
+
+
+static struct pci_ops tile_cfg_ops = {
+	.read =         tile_cfg_read,
+	.write =        tile_cfg_write,
+};
+
+
+/*
+ * In the following, each PCI controller's mem_resources[1]
+ * represents its (non-prefetchable) PCI memory resource.
+ * mem_resources[0] and mem_resources[2] refer to its PCI I/O and
+ * prefetchable PCI memory resources, respectively.
+ * For more details, see pci_setup_bridge() in setup-bus.c.
+ * By comparing the target PCI memory address against the
+ * end address of controller 0, we can determine the controller
+ * that should accept the PCI memory access.
+ */
+#define TILE_READ(size, type)						\
+type _tile_read##size(unsigned long addr)				\
+{									\
+	type val;							\
+	int idx = 0;							\
+	if (addr > controllers[0].mem_resources[1].end &&		\
+	    addr > controllers[0].mem_resources[2].end)			\
+		idx = 1;                                                \
+	if (hv_dev_pread(controllers[idx].hv_mem_fd, 0,			\
+			 (HV_VirtAddr)(&val), sizeof(type), addr))	\
+		pr_err("PCI: read %zd bytes at 0x%lX failed\n",		\
+		       sizeof(type), addr);				\
+	return val;							\
+}									\
+EXPORT_SYMBOL(_tile_read##size)
+
+TILE_READ(b, u8);
+TILE_READ(w, u16);
+TILE_READ(l, u32);
+TILE_READ(q, u64);
+
+#define TILE_WRITE(size, type)						\
+void _tile_write##size(type val, unsigned long addr)			\
+{									\
+	int idx = 0;							\
+	if (addr > controllers[0].mem_resources[1].end &&		\
+	    addr > controllers[0].mem_resources[2].end)			\
+		idx = 1;                                                \
+	if (hv_dev_pwrite(controllers[idx].hv_mem_fd, 0,		\
+			  (HV_VirtAddr)(&val), sizeof(type), addr))	\
+		pr_err("PCI: write %zd bytes at 0x%lX failed\n",	\
+		       sizeof(type), addr);				\
+}									\
+EXPORT_SYMBOL(_tile_write##size)
+
+TILE_WRITE(b, u8);
+TILE_WRITE(w, u16);
+TILE_WRITE(l, u32);
+TILE_WRITE(q, u64);
diff --git a/kernel/arch/tile/kernel/pci_gx.c b/kernel/arch/tile/kernel/pci_gx.c
new file mode 100644
index 000000000..b1df847d0
--- /dev/null
+++ b/kernel/arch/tile/kernel/pci_gx.c
@@ -0,0 +1,1595 @@
+/*
+ * Copyright 2012 Tilera Corporation. All Rights Reserved.
+ *
+ *   This program is free software; you can redistribute it and/or
+ *   modify it under the terms of the GNU General Public License
+ *   as published by the Free Software Foundation, version 2.
+ *
+ *   This program is distributed in the hope that it will be useful, but
+ *   WITHOUT ANY WARRANTY; without even the implied warranty of
+ *   MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
+ *   NON INFRINGEMENT.  See the GNU General Public License for
+ *   more details.
+ */
+
+#include <linux/kernel.h>
+#include <linux/mmzone.h>
+#include <linux/pci.h>
+#include <linux/delay.h>
+#include <linux/string.h>
+#include <linux/init.h>
+#include <linux/capability.h>
+#include <linux/sched.h>
+#include <linux/errno.h>
+#include <linux/irq.h>
+#include <linux/msi.h>
+#include <linux/io.h>
+#include <linux/uaccess.h>
+#include <linux/ctype.h>
+
+#include <asm/processor.h>
+#include <asm/sections.h>
+#include <asm/byteorder.h>
+
+#include <gxio/iorpc_globals.h>
+#include <gxio/kiorpc.h>
+#include <gxio/trio.h>
+#include <gxio/iorpc_trio.h>
+#include <hv/drv_trio_intf.h>
+
+#include <arch/sim.h>
+
+/*
+ * This file containes the routines to search for PCI buses,
+ * enumerate the buses, and configure any attached devices.
+ */
+
+#define DEBUG_PCI_CFG	0
+
+#if DEBUG_PCI_CFG
+#define TRACE_CFG_WR(size, val, bus, dev, func, offset) \
+	pr_info("CFG WR %d-byte VAL %#x to bus %d dev %d func %d addr %u\n", \
+		size, val, bus, dev, func, offset & 0xFFF);
+#define TRACE_CFG_RD(size, val, bus, dev, func, offset) \
+	pr_info("CFG RD %d-byte VAL %#x from bus %d dev %d func %d addr %u\n", \
+		size, val, bus, dev, func, offset & 0xFFF);
+#else
+#define TRACE_CFG_WR(...)
+#define TRACE_CFG_RD(...)
+#endif
+
+static int pci_probe = 1;
+
+/* Information on the PCIe RC ports configuration. */
+static int pcie_rc[TILEGX_NUM_TRIO][TILEGX_TRIO_PCIES];
+
+/*
+ * On some platforms with one or more Gx endpoint ports, we need to
+ * delay the PCIe RC port probe for a few seconds to work around
+ * a HW PCIe link-training bug. The exact delay is specified with
+ * a kernel boot argument in the form of "pcie_rc_delay=T,P,S",
+ * where T is the TRIO instance number, P is the port number and S is
+ * the delay in seconds. If the argument is specified, but the delay is
+ * not provided, the value will be DEFAULT_RC_DELAY.
+ */
+static int rc_delay[TILEGX_NUM_TRIO][TILEGX_TRIO_PCIES];
+
+/* Default number of seconds that the PCIe RC port probe can be delayed. */
+#define DEFAULT_RC_DELAY	10
+
+/* The PCI I/O space size in each PCI domain. */
+#define IO_SPACE_SIZE		0x10000
+
+/* Provide shorter versions of some very long constant names. */
+#define AUTO_CONFIG_RC	\
+	TRIO_PCIE_INTFC_PORT_CONFIG__STRAP_STATE_VAL_AUTO_CONFIG_RC
+#define AUTO_CONFIG_RC_G1	\
+	TRIO_PCIE_INTFC_PORT_CONFIG__STRAP_STATE_VAL_AUTO_CONFIG_RC_G1
+#define AUTO_CONFIG_EP	\
+	TRIO_PCIE_INTFC_PORT_CONFIG__STRAP_STATE_VAL_AUTO_CONFIG_ENDPOINT
+#define AUTO_CONFIG_EP_G1	\
+	TRIO_PCIE_INTFC_PORT_CONFIG__STRAP_STATE_VAL_AUTO_CONFIG_ENDPOINT_G1
+
+/* Array of the PCIe ports configuration info obtained from the BIB. */
+struct pcie_trio_ports_property pcie_ports[TILEGX_NUM_TRIO];
+
+/* Number of configured TRIO instances. */
+int num_trio_shims;
+
+/* All drivers share the TRIO contexts defined here. */
+gxio_trio_context_t trio_contexts[TILEGX_NUM_TRIO];
+
+/* Pointer to an array of PCIe RC controllers. */
+struct pci_controller pci_controllers[TILEGX_NUM_TRIO * TILEGX_TRIO_PCIES];
+int num_rc_controllers;
+
+static struct pci_ops tile_cfg_ops;
+
+/* Mask of CPUs that should receive PCIe interrupts. */
+static struct cpumask intr_cpus_map;
+
+/* We don't need to worry about the alignment of resources. */
+resource_size_t pcibios_align_resource(void *data, const struct resource *res,
+				       resource_size_t size,
+				       resource_size_t align)
+{
+	return res->start;
+}
+EXPORT_SYMBOL(pcibios_align_resource);
+
+/*
+ * Pick a CPU to receive and handle the PCIe interrupts, based on the IRQ #.
+ * For now, we simply send interrupts to non-dataplane CPUs.
+ * We may implement methods to allow user to specify the target CPUs,
+ * e.g. via boot arguments.
+ */
+static int tile_irq_cpu(int irq)
+{
+	unsigned int count;
+	int i = 0;
+	int cpu;
+
+	count = cpumask_weight(&intr_cpus_map);
+	if (unlikely(count == 0)) {
+		pr_warn("intr_cpus_map empty, interrupts will be delievered to dataplane tiles\n");
+		return irq % (smp_height * smp_width);
+	}
+
+	count = irq % count;
+	for_each_cpu(cpu, &intr_cpus_map) {
+		if (i++ == count)
+			break;
+	}
+	return cpu;
+}
+
+/* Open a file descriptor to the TRIO shim. */
+static int tile_pcie_open(int trio_index)
+{
+	gxio_trio_context_t *context = &trio_contexts[trio_index];
+	int ret;
+	int mac;
+
+	/* This opens a file descriptor to the TRIO shim. */
+	ret = gxio_trio_init(context, trio_index);
+	if (ret < 0)
+		goto gxio_trio_init_failure;
+
+	/* Allocate an ASID for the kernel. */
+	ret = gxio_trio_alloc_asids(context, 1, 0, 0);
+	if (ret < 0) {
+		pr_err("PCI: ASID alloc failure on TRIO %d, give up\n",
+			trio_index);
+		goto asid_alloc_failure;
+	}
+
+	context->asid = ret;
+
+#ifdef USE_SHARED_PCIE_CONFIG_REGION
+	/*
+	 * Alloc a PIO region for config access, shared by all MACs per TRIO.
+	 * This shouldn't fail since the kernel is supposed to the first
+	 * client of the TRIO's PIO regions.
+	 */
+	ret = gxio_trio_alloc_pio_regions(context, 1, 0, 0);
+	if (ret < 0) {
+		pr_err("PCI: CFG PIO alloc failure on TRIO %d, give up\n",
+			trio_index);
+		goto pio_alloc_failure;
+	}
+
+	context->pio_cfg_index = ret;
+
+	/*
+	 * For PIO CFG, the bus_address_hi parameter is 0. The mac parameter
+	 * is also 0 because it is specified in PIO_REGION_SETUP_CFG_ADDR.
+	 */
+	ret = gxio_trio_init_pio_region_aux(context, context->pio_cfg_index,
+		0, 0, HV_TRIO_PIO_FLAG_CONFIG_SPACE);
+	if (ret < 0) {
+		pr_err("PCI: CFG PIO init failure on TRIO %d, give up\n",
+			trio_index);
+		goto pio_alloc_failure;
+	}
+#endif
+
+	/* Get the properties of the PCIe ports on this TRIO instance. */
+	ret = gxio_trio_get_port_property(context, &pcie_ports[trio_index]);
+	if (ret < 0) {
+		pr_err("PCI: PCIE_GET_PORT_PROPERTY failure, error %d, on TRIO %d\n",
+		       ret, trio_index);
+		goto get_port_property_failure;
+	}
+
+	context->mmio_base_mac =
+		iorpc_ioremap(context->fd, 0, HV_TRIO_CONFIG_IOREMAP_SIZE);
+	if (context->mmio_base_mac == NULL) {
+		pr_err("PCI: TRIO config space mapping failure, error %d, on TRIO %d\n",
+		       ret, trio_index);
+		ret = -ENOMEM;
+
+		goto trio_mmio_mapping_failure;
+	}
+
+	/* Check the port strap state which will override the BIB setting. */
+	for (mac = 0; mac < TILEGX_TRIO_PCIES; mac++) {
+		TRIO_PCIE_INTFC_PORT_CONFIG_t port_config;
+		unsigned int reg_offset;
+
+		/* Ignore ports that are not specified in the BIB. */
+		if (!pcie_ports[trio_index].ports[mac].allow_rc &&
+		    !pcie_ports[trio_index].ports[mac].allow_ep)
+			continue;
+
+		reg_offset =
+			(TRIO_PCIE_INTFC_PORT_CONFIG <<
+				TRIO_CFG_REGION_ADDR__REG_SHIFT) |
+			(TRIO_CFG_REGION_ADDR__INTFC_VAL_MAC_INTERFACE <<
+				TRIO_CFG_REGION_ADDR__INTFC_SHIFT) |
+			(mac << TRIO_CFG_REGION_ADDR__MAC_SEL_SHIFT);
+
+		port_config.word =
+			__gxio_mmio_read(context->mmio_base_mac + reg_offset);
+
+		if (port_config.strap_state != AUTO_CONFIG_RC &&
+		    port_config.strap_state != AUTO_CONFIG_RC_G1) {
+			/*
+			 * If this is really intended to be an EP port, record
+			 * it so that the endpoint driver will know about it.
+			 */
+			if (port_config.strap_state == AUTO_CONFIG_EP ||
+			    port_config.strap_state == AUTO_CONFIG_EP_G1)
+				pcie_ports[trio_index].ports[mac].allow_ep = 1;
+		}
+	}
+
+	return ret;
+
+trio_mmio_mapping_failure:
+get_port_property_failure:
+asid_alloc_failure:
+#ifdef USE_SHARED_PCIE_CONFIG_REGION
+pio_alloc_failure:
+#endif
+	hv_dev_close(context->fd);
+gxio_trio_init_failure:
+	context->fd = -1;
+
+	return ret;
+}
+
+static int __init tile_trio_init(void)
+{
+	int i;
+
+	/* We loop over all the TRIO shims. */
+	for (i = 0; i < TILEGX_NUM_TRIO; i++) {
+		if (tile_pcie_open(i) < 0)
+			continue;
+		num_trio_shims++;
+	}
+
+	return 0;
+}
+postcore_initcall(tile_trio_init);
+
+static void tilegx_legacy_irq_ack(struct irq_data *d)
+{
+	__insn_mtspr(SPR_IPI_EVENT_RESET_K, 1UL << d->irq);
+}
+
+static void tilegx_legacy_irq_mask(struct irq_data *d)
+{
+	__insn_mtspr(SPR_IPI_MASK_SET_K, 1UL << d->irq);
+}
+
+static void tilegx_legacy_irq_unmask(struct irq_data *d)
+{
+	__insn_mtspr(SPR_IPI_MASK_RESET_K, 1UL << d->irq);
+}
+
+static struct irq_chip tilegx_legacy_irq_chip = {
+	.name			= "tilegx_legacy_irq",
+	.irq_ack		= tilegx_legacy_irq_ack,
+	.irq_mask		= tilegx_legacy_irq_mask,
+	.irq_unmask		= tilegx_legacy_irq_unmask,
+
+	/* TBD: support set_affinity. */
+};
+
+/*
+ * This is a wrapper function of the kernel level-trigger interrupt
+ * handler handle_level_irq() for PCI legacy interrupts. The TRIO
+ * is configured such that only INTx Assert interrupts are proxied
+ * to Linux which just calls handle_level_irq() after clearing the
+ * MAC INTx Assert status bit associated with this interrupt.
+ */
+static void trio_handle_level_irq(unsigned int irq, struct irq_desc *desc)
+{
+	struct pci_controller *controller = irq_desc_get_handler_data(desc);
+	gxio_trio_context_t *trio_context = controller->trio;
+	uint64_t intx = (uint64_t)irq_desc_get_chip_data(desc);
+	int mac = controller->mac;
+	unsigned int reg_offset;
+	uint64_t level_mask;
+
+	handle_level_irq(irq, desc);
+
+	/*
+	 * Clear the INTx Level status, otherwise future interrupts are
+	 * not sent.
+	 */
+	reg_offset = (TRIO_PCIE_INTFC_MAC_INT_STS <<
+		TRIO_CFG_REGION_ADDR__REG_SHIFT) |
+		(TRIO_CFG_REGION_ADDR__INTFC_VAL_MAC_INTERFACE <<
+		TRIO_CFG_REGION_ADDR__INTFC_SHIFT ) |
+		(mac << TRIO_CFG_REGION_ADDR__MAC_SEL_SHIFT);
+
+	level_mask = TRIO_PCIE_INTFC_MAC_INT_STS__INT_LEVEL_MASK << intx;
+
+	__gxio_mmio_write(trio_context->mmio_base_mac + reg_offset, level_mask);
+}
+
+/*
+ * Create kernel irqs and set up the handlers for the legacy interrupts.
+ * Also some minimum initialization for the MSI support.
+ */
+static int tile_init_irqs(struct pci_controller *controller)
+{
+	int i;
+	int j;
+	int irq;
+	int result;
+
+	cpumask_copy(&intr_cpus_map, cpu_online_mask);
+
+
+	for (i = 0; i < 4; i++) {
+		gxio_trio_context_t *context = controller->trio;
+		int cpu;
+
+		/* Ask the kernel to allocate an IRQ. */
+		irq = irq_alloc_hwirq(-1);
+		if (!irq) {
+			pr_err("PCI: no free irq vectors, failed for %d\n", i);
+			goto free_irqs;
+		}
+		controller->irq_intx_table[i] = irq;
+
+		/* Distribute the 4 IRQs to different tiles. */
+		cpu = tile_irq_cpu(irq);
+
+		/* Configure the TRIO intr binding for this IRQ. */
+		result = gxio_trio_config_legacy_intr(context, cpu_x(cpu),
+						      cpu_y(cpu), KERNEL_PL,
+						      irq, controller->mac, i);
+		if (result < 0) {
+			pr_err("PCI: MAC intx config failed for %d\n", i);
+
+			goto free_irqs;
+		}
+
+		/* Register the IRQ handler with the kernel. */
+		irq_set_chip_and_handler(irq, &tilegx_legacy_irq_chip,
+					trio_handle_level_irq);
+		irq_set_chip_data(irq, (void *)(uint64_t)i);
+		irq_set_handler_data(irq, controller);
+	}
+
+	return 0;
+
+free_irqs:
+	for (j = 0; j < i; j++)
+		irq_free_hwirq(controller->irq_intx_table[j]);
+
+	return -1;
+}
+
+/*
+ * Return 1 if the port is strapped to operate in RC mode.
+ */
+static int
+strapped_for_rc(gxio_trio_context_t *trio_context, int mac)
+{
+	TRIO_PCIE_INTFC_PORT_CONFIG_t port_config;
+	unsigned int reg_offset;
+
+	/* Check the port configuration. */
+	reg_offset =
+		(TRIO_PCIE_INTFC_PORT_CONFIG <<
+			TRIO_CFG_REGION_ADDR__REG_SHIFT) |
+		(TRIO_CFG_REGION_ADDR__INTFC_VAL_MAC_INTERFACE <<
+			TRIO_CFG_REGION_ADDR__INTFC_SHIFT) |
+		(mac << TRIO_CFG_REGION_ADDR__MAC_SEL_SHIFT);
+	port_config.word =
+		__gxio_mmio_read(trio_context->mmio_base_mac + reg_offset);
+
+	if (port_config.strap_state == AUTO_CONFIG_RC ||
+	    port_config.strap_state == AUTO_CONFIG_RC_G1)
+		return 1;
+	else
+		return 0;
+}
+
+/*
+ * Find valid controllers and fill in pci_controller structs for each
+ * of them.
+ *
+ * Return the number of controllers discovered.
+ */
+int __init tile_pci_init(void)
+{
+	int ctl_index = 0;
+	int i, j;
+
+	if (!pci_probe) {
+		pr_info("PCI: disabled by boot argument\n");
+		return 0;
+	}
+
+	pr_info("PCI: Searching for controllers...\n");
+
+	if (num_trio_shims == 0 || sim_is_simulator())
+		return 0;
+
+	/*
+	 * Now determine which PCIe ports are configured to operate in RC
+	 * mode. There is a differece in the port configuration capability
+	 * between the Gx36 and Gx72 devices.
+	 *
+	 * The Gx36 has configuration capability for each of the 3 PCIe
+	 * interfaces (disable, auto endpoint, auto RC, etc.).
+	 * On the Gx72, you can only select one of the 3 PCIe interfaces per
+	 * TRIO to train automatically. Further, the allowable training modes
+	 * are reduced to four options (auto endpoint, auto RC, stream x1,
+	 * stream x4).
+	 *
+	 * For Gx36 ports, it must be allowed to be in RC mode by the
+	 * Board Information Block, and the hardware strapping pins must be
+	 * set to RC mode.
+	 *
+	 * For Gx72 ports, the port will operate in RC mode if either of the
+	 * following is true:
+	 * 1. It is allowed to be in RC mode by the Board Information Block,
+	 *    and the BIB doesn't allow the EP mode.
+	 * 2. It is allowed to be in either the RC or the EP mode by the BIB,
+	 *    and the hardware strapping pin is set to RC mode.
+	 */
+	for (i = 0; i < TILEGX_NUM_TRIO; i++) {
+		gxio_trio_context_t *context = &trio_contexts[i];
+
+		if (context->fd < 0)
+			continue;
+
+		for (j = 0; j < TILEGX_TRIO_PCIES; j++) {
+			int is_rc = 0;
+
+			if (pcie_ports[i].is_gx72 &&
+			    pcie_ports[i].ports[j].allow_rc) {
+				if (!pcie_ports[i].ports[j].allow_ep ||
+				    strapped_for_rc(context, j))
+					is_rc = 1;
+			} else if (pcie_ports[i].ports[j].allow_rc &&
+				   strapped_for_rc(context, j)) {
+				is_rc = 1;
+			}
+			if (is_rc) {
+				pcie_rc[i][j] = 1;
+				num_rc_controllers++;
+			}
+		}
+	}
+
+	/* Return if no PCIe ports are configured to operate in RC mode. */
+	if (num_rc_controllers == 0)
+		return 0;
+
+	/* Set the TRIO pointer and MAC index for each PCIe RC port. */
+	for (i = 0; i < TILEGX_NUM_TRIO; i++) {
+		for (j = 0; j < TILEGX_TRIO_PCIES; j++) {
+			if (pcie_rc[i][j]) {
+				pci_controllers[ctl_index].trio =
+					&trio_contexts[i];
+				pci_controllers[ctl_index].mac = j;
+				pci_controllers[ctl_index].trio_index = i;
+				ctl_index++;
+				if (ctl_index == num_rc_controllers)
+					goto out;
+			}
+		}
+	}
+
+out:
+	/* Configure each PCIe RC port. */
+	for (i = 0; i < num_rc_controllers; i++) {
+
+		/* Configure the PCIe MAC to run in RC mode. */
+		struct pci_controller *controller = &pci_controllers[i];
+
+		controller->index = i;
+		controller->ops = &tile_cfg_ops;
+
+		controller->io_space.start = PCIBIOS_MIN_IO +
+			(i * IO_SPACE_SIZE);
+		controller->io_space.end = controller->io_space.start +
+			IO_SPACE_SIZE - 1;
+		BUG_ON(controller->io_space.end > IO_SPACE_LIMIT);
+		controller->io_space.flags = IORESOURCE_IO;
+		snprintf(controller->io_space_name,
+			 sizeof(controller->io_space_name),
+			 "PCI I/O domain %d", i);
+		controller->io_space.name = controller->io_space_name;
+
+		/*
+		 * The PCI memory resource is located above the PA space.
+		 * For every host bridge, the BAR window or the MMIO aperture
+		 * is in range [3GB, 4GB - 1] of a 4GB space beyond the
+		 * PA space.
+		 */
+		controller->mem_offset = TILE_PCI_MEM_START +
+			(i * TILE_PCI_BAR_WINDOW_TOP);
+		controller->mem_space.start = controller->mem_offset +
+			TILE_PCI_BAR_WINDOW_TOP - TILE_PCI_BAR_WINDOW_SIZE;
+		controller->mem_space.end = controller->mem_offset +
+			TILE_PCI_BAR_WINDOW_TOP - 1;
+		controller->mem_space.flags = IORESOURCE_MEM;
+		snprintf(controller->mem_space_name,
+			 sizeof(controller->mem_space_name),
+			 "PCI mem domain %d", i);
+		controller->mem_space.name = controller->mem_space_name;
+	}
+
+	return num_rc_controllers;
+}
+
+/*
+ * (pin - 1) converts from the PCI standard's [1:4] convention to
+ * a normal [0:3] range.
+ */
+static int tile_map_irq(const struct pci_dev *dev, u8 device, u8 pin)
+{
+	struct pci_controller *controller =
+		(struct pci_controller *)dev->sysdata;
+	return controller->irq_intx_table[pin - 1];
+}
+
+static void fixup_read_and_payload_sizes(struct pci_controller *controller)
+{
+	gxio_trio_context_t *trio_context = controller->trio;
+	struct pci_bus *root_bus = controller->root_bus;
+	TRIO_PCIE_RC_DEVICE_CONTROL_t dev_control;
+	TRIO_PCIE_RC_DEVICE_CAP_t rc_dev_cap;
+	unsigned int reg_offset;
+	struct pci_bus *child;
+	int mac;
+	int err;
+
+	mac = controller->mac;
+
+	/* Set our max read request size to be 4KB. */
+	reg_offset =
+		(TRIO_PCIE_RC_DEVICE_CONTROL <<
+			TRIO_CFG_REGION_ADDR__REG_SHIFT) |
+		(TRIO_CFG_REGION_ADDR__INTFC_VAL_MAC_STANDARD <<
+			TRIO_CFG_REGION_ADDR__INTFC_SHIFT ) |
+		(mac << TRIO_CFG_REGION_ADDR__MAC_SEL_SHIFT);
+
+	dev_control.word = __gxio_mmio_read32(trio_context->mmio_base_mac +
+					      reg_offset);
+	dev_control.max_read_req_sz = 5;
+	__gxio_mmio_write32(trio_context->mmio_base_mac + reg_offset,
+			    dev_control.word);
+
+	/*
+	 * Set the max payload size supported by this Gx PCIe MAC.
+	 * Though Gx PCIe supports Max Payload Size of up to 1024 bytes,
+	 * experiments have shown that setting MPS to 256 yields the
+	 * best performance.
+	 */
+	reg_offset =
+		(TRIO_PCIE_RC_DEVICE_CAP <<
+			TRIO_CFG_REGION_ADDR__REG_SHIFT) |
+		(TRIO_CFG_REGION_ADDR__INTFC_VAL_MAC_STANDARD <<
+			TRIO_CFG_REGION_ADDR__INTFC_SHIFT ) |
+		(mac << TRIO_CFG_REGION_ADDR__MAC_SEL_SHIFT);
+
+	rc_dev_cap.word = __gxio_mmio_read32(trio_context->mmio_base_mac +
+					     reg_offset);
+	rc_dev_cap.mps_sup = 1;
+	__gxio_mmio_write32(trio_context->mmio_base_mac + reg_offset,
+			    rc_dev_cap.word);
+
+	/* Configure PCI Express MPS setting. */
+	list_for_each_entry(child, &root_bus->children, node)
+		pcie_bus_configure_settings(child);
+
+	/*
+	 * Set the mac_config register in trio based on the MPS/MRS of the link.
+	 */
+	reg_offset =
+		(TRIO_PCIE_RC_DEVICE_CONTROL <<
+			TRIO_CFG_REGION_ADDR__REG_SHIFT) |
+		(TRIO_CFG_REGION_ADDR__INTFC_VAL_MAC_STANDARD <<
+			TRIO_CFG_REGION_ADDR__INTFC_SHIFT ) |
+		(mac << TRIO_CFG_REGION_ADDR__MAC_SEL_SHIFT);
+
+	dev_control.word = __gxio_mmio_read32(trio_context->mmio_base_mac +
+						reg_offset);
+
+	err = gxio_trio_set_mps_mrs(trio_context,
+				    dev_control.max_payload_size,
+				    dev_control.max_read_req_sz,
+				    mac);
+	if (err < 0) {
+		pr_err("PCI: PCIE_CONFIGURE_MAC_MPS_MRS failure, MAC %d on TRIO %d\n",
+		       mac, controller->trio_index);
+	}
+}
+
+static int setup_pcie_rc_delay(char *str)
+{
+	unsigned long delay = 0;
+	unsigned long trio_index;
+	unsigned long mac;
+
+	if (str == NULL || !isdigit(*str))
+		return -EINVAL;
+	trio_index = simple_strtoul(str, (char **)&str, 10);
+	if (trio_index >= TILEGX_NUM_TRIO)
+		return -EINVAL;
+
+	if (*str != ',')
+		return -EINVAL;
+
+	str++;
+	if (!isdigit(*str))
+		return -EINVAL;
+	mac = simple_strtoul(str, (char **)&str, 10);
+	if (mac >= TILEGX_TRIO_PCIES)
+		return -EINVAL;
+
+	if (*str != '\0') {
+		if (*str != ',')
+			return -EINVAL;
+
+		str++;
+		if (!isdigit(*str))
+			return -EINVAL;
+		delay = simple_strtoul(str, (char **)&str, 10);
+	}
+
+	rc_delay[trio_index][mac] = delay ? : DEFAULT_RC_DELAY;
+	return 0;
+}
+early_param("pcie_rc_delay", setup_pcie_rc_delay);
+
+/* PCI initialization entry point, called by subsys_initcall. */
+int __init pcibios_init(void)
+{
+	resource_size_t offset;
+	LIST_HEAD(resources);
+	int next_busno;
+	int i;
+
+	tile_pci_init();
+
+	if (num_rc_controllers == 0)
+		return 0;
+
+	/*
+	 * Delay a bit in case devices aren't ready.  Some devices are
+	 * known to require at least 20ms here, but we use a more
+	 * conservative value.
+	 */
+	msleep(250);
+
+	/* Scan all of the recorded PCI controllers.  */
+	for (next_busno = 0, i = 0; i < num_rc_controllers; i++) {
+		struct pci_controller *controller = &pci_controllers[i];
+		gxio_trio_context_t *trio_context = controller->trio;
+		TRIO_PCIE_INTFC_PORT_STATUS_t port_status;
+		TRIO_PCIE_INTFC_TX_FIFO_CTL_t tx_fifo_ctl;
+		struct pci_bus *bus;
+		unsigned int reg_offset;
+		unsigned int class_code_revision;
+		int trio_index;
+		int mac;
+		int ret;
+
+		if (trio_context->fd < 0)
+			continue;
+
+		trio_index = controller->trio_index;
+		mac = controller->mac;
+
+		/*
+		 * Check for PCIe link-up status to decide if we need
+		 * to force the link to come up.
+		 */
+		reg_offset =
+			(TRIO_PCIE_INTFC_PORT_STATUS <<
+				TRIO_CFG_REGION_ADDR__REG_SHIFT) |
+			(TRIO_CFG_REGION_ADDR__INTFC_VAL_MAC_INTERFACE <<
+				TRIO_CFG_REGION_ADDR__INTFC_SHIFT) |
+			(mac << TRIO_CFG_REGION_ADDR__MAC_SEL_SHIFT);
+
+		port_status.word =
+			__gxio_mmio_read(trio_context->mmio_base_mac +
+					 reg_offset);
+		if (!port_status.dl_up) {
+			if (rc_delay[trio_index][mac]) {
+				pr_info("Delaying PCIe RC TRIO init %d sec on MAC %d on TRIO %d\n",
+					rc_delay[trio_index][mac], mac,
+					trio_index);
+				msleep(rc_delay[trio_index][mac] * 1000);
+			}
+			ret = gxio_trio_force_rc_link_up(trio_context, mac);
+			if (ret < 0)
+				pr_err("PCI: PCIE_FORCE_LINK_UP failure, MAC %d on TRIO %d\n",
+				       mac, trio_index);
+		}
+
+		pr_info("PCI: Found PCI controller #%d on TRIO %d MAC %d\n",
+			i, trio_index, controller->mac);
+
+		/* Delay the bus probe if needed. */
+		if (rc_delay[trio_index][mac]) {
+			pr_info("Delaying PCIe RC bus enumerating %d sec on MAC %d on TRIO %d\n",
+				rc_delay[trio_index][mac], mac, trio_index);
+			msleep(rc_delay[trio_index][mac] * 1000);
+		} else {
+			/*
+			 * Wait a bit here because some EP devices
+			 * take longer to come up.
+			 */
+			msleep(1000);
+		}
+
+		/* Check for PCIe link-up status again. */
+		port_status.word =
+			__gxio_mmio_read(trio_context->mmio_base_mac +
+					 reg_offset);
+		if (!port_status.dl_up) {
+			if (pcie_ports[trio_index].ports[mac].removable) {
+				pr_info("PCI: link is down, MAC %d on TRIO %d\n",
+					mac, trio_index);
+				pr_info("This is expected if no PCIe card is connected to this link\n");
+			} else
+				pr_err("PCI: link is down, MAC %d on TRIO %d\n",
+				       mac, trio_index);
+			continue;
+		}
+
+		/*
+		 * Ensure that the link can come out of L1 power down state.
+		 * Strictly speaking, this is needed only in the case of
+		 * heavy RC-initiated DMAs.
+		 */
+		reg_offset =
+			(TRIO_PCIE_INTFC_TX_FIFO_CTL <<
+				TRIO_CFG_REGION_ADDR__REG_SHIFT) |
+			(TRIO_CFG_REGION_ADDR__INTFC_VAL_MAC_INTERFACE <<
+				TRIO_CFG_REGION_ADDR__INTFC_SHIFT ) |
+			(mac << TRIO_CFG_REGION_ADDR__MAC_SEL_SHIFT);
+		tx_fifo_ctl.word =
+			__gxio_mmio_read(trio_context->mmio_base_mac +
+					 reg_offset);
+		tx_fifo_ctl.min_p_credits = 0;
+		__gxio_mmio_write(trio_context->mmio_base_mac + reg_offset,
+				  tx_fifo_ctl.word);
+
+		/*
+		 * Change the device ID so that Linux bus crawl doesn't confuse
+		 * the internal bridge with any Tilera endpoints.
+		 */
+		reg_offset =
+			(TRIO_PCIE_RC_DEVICE_ID_VEN_ID <<
+				TRIO_CFG_REGION_ADDR__REG_SHIFT) |
+			(TRIO_CFG_REGION_ADDR__INTFC_VAL_MAC_STANDARD <<
+				TRIO_CFG_REGION_ADDR__INTFC_SHIFT ) |
+			(mac << TRIO_CFG_REGION_ADDR__MAC_SEL_SHIFT);
+
+		__gxio_mmio_write32(trio_context->mmio_base_mac + reg_offset,
+				    (TILERA_GX36_RC_DEV_ID <<
+				    TRIO_PCIE_RC_DEVICE_ID_VEN_ID__DEV_ID_SHIFT) |
+				    TILERA_VENDOR_ID);
+
+		/* Set the internal P2P bridge class code. */
+		reg_offset =
+			(TRIO_PCIE_RC_REVISION_ID <<
+				TRIO_CFG_REGION_ADDR__REG_SHIFT) |
+			(TRIO_CFG_REGION_ADDR__INTFC_VAL_MAC_STANDARD <<
+				TRIO_CFG_REGION_ADDR__INTFC_SHIFT ) |
+			(mac << TRIO_CFG_REGION_ADDR__MAC_SEL_SHIFT);
+
+		class_code_revision =
+			__gxio_mmio_read32(trio_context->mmio_base_mac +
+					   reg_offset);
+		class_code_revision = (class_code_revision & 0xff) |
+			(PCI_CLASS_BRIDGE_PCI << 16);
+
+		__gxio_mmio_write32(trio_context->mmio_base_mac +
+				    reg_offset, class_code_revision);
+
+#ifdef USE_SHARED_PCIE_CONFIG_REGION
+
+		/* Map in the MMIO space for the PIO region. */
+		offset = HV_TRIO_PIO_OFFSET(trio_context->pio_cfg_index) |
+			(((unsigned long long)mac) <<
+			TRIO_TILE_PIO_REGION_SETUP_CFG_ADDR__MAC_SHIFT);
+
+#else
+
+		/* Alloc a PIO region for PCI config access per MAC. */
+		ret = gxio_trio_alloc_pio_regions(trio_context, 1, 0, 0);
+		if (ret < 0) {
+			pr_err("PCI: PCI CFG PIO alloc failure for mac %d on TRIO %d, give up\n",
+			       mac, trio_index);
+
+			continue;
+		}
+
+		trio_context->pio_cfg_index[mac] = ret;
+
+		/* For PIO CFG, the bus_address_hi parameter is 0. */
+		ret = gxio_trio_init_pio_region_aux(trio_context,
+			trio_context->pio_cfg_index[mac],
+			mac, 0, HV_TRIO_PIO_FLAG_CONFIG_SPACE);
+		if (ret < 0) {
+			pr_err("PCI: PCI CFG PIO init failure for mac %d on TRIO %d, give up\n",
+			       mac, trio_index);
+
+			continue;
+		}
+
+		offset = HV_TRIO_PIO_OFFSET(trio_context->pio_cfg_index[mac]) |
+			(((unsigned long long)mac) <<
+			TRIO_TILE_PIO_REGION_SETUP_CFG_ADDR__MAC_SHIFT);
+
+#endif
+
+		/*
+		 * To save VMALLOC space, we take advantage of the fact that
+		 * bit 29 in the PIO CFG address format is reserved 0. With
+		 * TRIO_TILE_PIO_REGION_SETUP_CFG_ADDR__MAC_SHIFT being 30,
+		 * this cuts VMALLOC space usage from 1GB to 512MB per mac.
+		 */
+		trio_context->mmio_base_pio_cfg[mac] =
+			iorpc_ioremap(trio_context->fd, offset, (1UL <<
+			(TRIO_TILE_PIO_REGION_SETUP_CFG_ADDR__MAC_SHIFT - 1)));
+		if (trio_context->mmio_base_pio_cfg[mac] == NULL) {
+			pr_err("PCI: PIO map failure for mac %d on TRIO %d\n",
+			       mac, trio_index);
+
+			continue;
+		}
+
+		/* Initialize the PCIe interrupts. */
+		if (tile_init_irqs(controller)) {
+			pr_err("PCI: IRQs init failure for mac %d on TRIO %d\n",
+				mac, trio_index);
+
+			continue;
+		}
+
+		/*
+		 * The PCI memory resource is located above the PA space.
+		 * The memory range for the PCI root bus should not overlap
+		 * with the physical RAM.
+		 */
+		pci_add_resource_offset(&resources, &controller->mem_space,
+					controller->mem_offset);
+		pci_add_resource(&resources, &controller->io_space);
+		controller->first_busno = next_busno;
+		bus = pci_scan_root_bus(NULL, next_busno, controller->ops,
+					controller, &resources);
+		controller->root_bus = bus;
+		next_busno = bus->busn_res.end + 1;
+	}
+
+	/* Do machine dependent PCI interrupt routing */
+	pci_fixup_irqs(pci_common_swizzle, tile_map_irq);
+
+	/*
+	 * This comes from the generic Linux PCI driver.
+	 *
+	 * It allocates all of the resources (I/O memory, etc)
+	 * associated with the devices read in above.
+	 */
+	pci_assign_unassigned_resources();
+
+	/* Record the I/O resources in the PCI controller structure. */
+	for (i = 0; i < num_rc_controllers; i++) {
+		struct pci_controller *controller = &pci_controllers[i];
+		gxio_trio_context_t *trio_context = controller->trio;
+		struct pci_bus *root_bus = pci_controllers[i].root_bus;
+		int ret;
+		int j;
+
+		/*
+		 * Skip controllers that are not properly initialized or
+		 * have down links.
+		 */
+		if (root_bus == NULL)
+			continue;
+
+		/* Configure the max_payload_size values for this domain. */
+		fixup_read_and_payload_sizes(controller);
+
+		/* Alloc a PIO region for PCI memory access for each RC port. */
+		ret = gxio_trio_alloc_pio_regions(trio_context, 1, 0, 0);
+		if (ret < 0) {
+			pr_err("PCI: MEM PIO alloc failure on TRIO %d mac %d, give up\n",
+			       controller->trio_index, controller->mac);
+
+			continue;
+		}
+
+		controller->pio_mem_index = ret;
+
+		/*
+		 * For PIO MEM, the bus_address_hi parameter is hard-coded 0
+		 * because we always assign 32-bit PCI bus BAR ranges.
+		 */
+		ret = gxio_trio_init_pio_region_aux(trio_context,
+						    controller->pio_mem_index,
+						    controller->mac,
+						    0,
+						    0);
+		if (ret < 0) {
+			pr_err("PCI: MEM PIO init failure on TRIO %d mac %d, give up\n",
+			       controller->trio_index, controller->mac);
+
+			continue;
+		}
+
+#ifdef CONFIG_TILE_PCI_IO
+		/*
+		 * Alloc a PIO region for PCI I/O space access for each RC port.
+		 */
+		ret = gxio_trio_alloc_pio_regions(trio_context, 1, 0, 0);
+		if (ret < 0) {
+			pr_err("PCI: I/O PIO alloc failure on TRIO %d mac %d, give up\n",
+			       controller->trio_index, controller->mac);
+
+			continue;
+		}
+
+		controller->pio_io_index = ret;
+
+		/*
+		 * For PIO IO, the bus_address_hi parameter is hard-coded 0
+		 * because PCI I/O address space is 32-bit.
+		 */
+		ret = gxio_trio_init_pio_region_aux(trio_context,
+						    controller->pio_io_index,
+						    controller->mac,
+						    0,
+						    HV_TRIO_PIO_FLAG_IO_SPACE);
+		if (ret < 0) {
+			pr_err("PCI: I/O PIO init failure on TRIO %d mac %d, give up\n",
+			       controller->trio_index, controller->mac);
+
+			continue;
+		}
+#endif
+
+		/*
+		 * Configure a Mem-Map region for each memory controller so
+		 * that Linux can map all of its PA space to the PCI bus.
+		 * Use the IOMMU to handle hash-for-home memory.
+		 */
+		for_each_online_node(j) {
+			unsigned long start_pfn = node_start_pfn[j];
+			unsigned long end_pfn = node_end_pfn[j];
+			unsigned long nr_pages = end_pfn - start_pfn;
+
+			ret = gxio_trio_alloc_memory_maps(trio_context, 1, 0,
+							  0);
+			if (ret < 0) {
+				pr_err("PCI: Mem-Map alloc failure on TRIO %d mac %d for MC %d, give up\n",
+				       controller->trio_index, controller->mac,
+				       j);
+
+				goto alloc_mem_map_failed;
+			}
+
+			controller->mem_maps[j] = ret;
+
+			/*
+			 * Initialize the Mem-Map and the I/O MMU so that all
+			 * the physical memory can be accessed by the endpoint
+			 * devices. The base bus address is set to the base CPA
+			 * of this memory controller plus an offset (see pci.h).
+			 * The region's base VA is set to the base CPA. The
+			 * I/O MMU table essentially translates the CPA to
+			 * the real PA. Implicitly, for node 0, we create
+			 * a separate Mem-Map region that serves as the inbound
+			 * window for legacy 32-bit devices. This is a direct
+			 * map of the low 4GB CPA space.
+			 */
+			ret = gxio_trio_init_memory_map_mmu_aux(trio_context,
+				controller->mem_maps[j],
+				start_pfn << PAGE_SHIFT,
+				nr_pages << PAGE_SHIFT,
+				trio_context->asid,
+				controller->mac,
+				(start_pfn << PAGE_SHIFT) +
+				TILE_PCI_MEM_MAP_BASE_OFFSET,
+				j,
+				GXIO_TRIO_ORDER_MODE_UNORDERED);
+			if (ret < 0) {
+				pr_err("PCI: Mem-Map init failure on TRIO %d mac %d for MC %d, give up\n",
+				       controller->trio_index, controller->mac,
+				       j);
+
+				goto alloc_mem_map_failed;
+			}
+			continue;
+
+alloc_mem_map_failed:
+			break;
+		}
+
+		pci_bus_add_devices(root_bus);
+	}
+
+	return 0;
+}
+subsys_initcall(pcibios_init);
+
+/* No bus fixups needed. */
+void pcibios_fixup_bus(struct pci_bus *bus)
+{
+}
+
+/* Process any "pci=" kernel boot arguments. */
+char *__init pcibios_setup(char *str)
+{
+	if (!strcmp(str, "off")) {
+		pci_probe = 0;
+		return NULL;
+	}
+	return str;
+}
+
+/*
+ * Called for each device after PCI setup is done.
+ * We initialize the PCI device capabilities conservatively, assuming that
+ * all devices can only address the 32-bit DMA space. The exception here is
+ * that the device dma_offset is set to the value that matches the 64-bit
+ * capable devices. This is OK because dma_offset is not used by legacy
+ * dma_ops, nor by the hybrid dma_ops's streaming DMAs, which are 64-bit ops.
+ * This implementation matches the kernel design of setting PCI devices'
+ * coherent_dma_mask to 0xffffffffull by default, allowing the device drivers
+ * to skip calling pci_set_consistent_dma_mask(DMA_BIT_MASK(32)).
+ */
+static void pcibios_fixup_final(struct pci_dev *pdev)
+{
+	set_dma_ops(&pdev->dev, gx_legacy_pci_dma_map_ops);
+	set_dma_offset(&pdev->dev, TILE_PCI_MEM_MAP_BASE_OFFSET);
+	pdev->dev.archdata.max_direct_dma_addr =
+		TILE_PCI_MAX_DIRECT_DMA_ADDRESS;
+	pdev->dev.coherent_dma_mask = TILE_PCI_MAX_DIRECT_DMA_ADDRESS;
+}
+DECLARE_PCI_FIXUP_FINAL(PCI_ANY_ID, PCI_ANY_ID, pcibios_fixup_final);
+
+/* Map a PCI MMIO bus address into VA space. */
+void __iomem *ioremap(resource_size_t phys_addr, unsigned long size)
+{
+	struct pci_controller *controller = NULL;
+	resource_size_t bar_start;
+	resource_size_t bar_end;
+	resource_size_t offset;
+	resource_size_t start;
+	resource_size_t end;
+	int trio_fd;
+	int i;
+
+	start = phys_addr;
+	end = phys_addr + size - 1;
+
+	/*
+	 * By searching phys_addr in each controller's mem_space, we can
+	 * determine the controller that should accept the PCI memory access.
+	 */
+	for (i = 0; i < num_rc_controllers; i++) {
+		/*
+		 * Skip controllers that are not properly initialized or
+		 * have down links.
+		 */
+		if (pci_controllers[i].root_bus == NULL)
+			continue;
+
+		bar_start = pci_controllers[i].mem_space.start;
+		bar_end = pci_controllers[i].mem_space.end;
+
+		if ((start >= bar_start) && (end <= bar_end)) {
+			controller = &pci_controllers[i];
+			break;
+		}
+	}
+
+	if (controller == NULL)
+		return NULL;
+
+	trio_fd = controller->trio->fd;
+
+	/* Convert the resource start to the bus address offset. */
+	start = phys_addr - controller->mem_offset;
+
+	offset = HV_TRIO_PIO_OFFSET(controller->pio_mem_index) + start;
+
+	/* We need to keep the PCI bus address's in-page offset in the VA. */
+	return iorpc_ioremap(trio_fd, offset, size) +
+		(start & (PAGE_SIZE - 1));
+}
+EXPORT_SYMBOL(ioremap);
+
+#ifdef CONFIG_TILE_PCI_IO
+/* Map a PCI I/O address into VA space. */
+void __iomem *ioport_map(unsigned long port, unsigned int size)
+{
+	struct pci_controller *controller = NULL;
+	resource_size_t bar_start;
+	resource_size_t bar_end;
+	resource_size_t offset;
+	resource_size_t start;
+	resource_size_t end;
+	int trio_fd;
+	int i;
+
+	start = port;
+	end = port + size - 1;
+
+	/*
+	 * By searching the port in each controller's io_space, we can
+	 * determine the controller that should accept the PCI I/O access.
+	 */
+	for (i = 0; i < num_rc_controllers; i++) {
+		/*
+		 * Skip controllers that are not properly initialized or
+		 * have down links.
+		 */
+		if (pci_controllers[i].root_bus == NULL)
+			continue;
+
+		bar_start = pci_controllers[i].io_space.start;
+		bar_end = pci_controllers[i].io_space.end;
+
+		if ((start >= bar_start) && (end <= bar_end)) {
+			controller = &pci_controllers[i];
+			break;
+		}
+	}
+
+	if (controller == NULL)
+		return NULL;
+
+	trio_fd = controller->trio->fd;
+
+	/* Convert the resource start to the bus address offset. */
+	port -= controller->io_space.start;
+
+	offset = HV_TRIO_PIO_OFFSET(controller->pio_io_index) + port;
+
+	/* We need to keep the PCI bus address's in-page offset in the VA. */
+	return iorpc_ioremap(trio_fd, offset, size) + (port & (PAGE_SIZE - 1));
+}
+EXPORT_SYMBOL(ioport_map);
+
+void ioport_unmap(void __iomem *addr)
+{
+	iounmap(addr);
+}
+EXPORT_SYMBOL(ioport_unmap);
+#endif
+
+void pci_iounmap(struct pci_dev *dev, void __iomem *addr)
+{
+	iounmap(addr);
+}
+EXPORT_SYMBOL(pci_iounmap);
+
+/****************************************************************
+ *
+ * Tile PCI config space read/write routines
+ *
+ ****************************************************************/
+
+/*
+ * These are the normal read and write ops
+ * These are expanded with macros from  pci_bus_read_config_byte() etc.
+ *
+ * devfn is the combined PCI device & function.
+ *
+ * offset is in bytes, from the start of config space for the
+ * specified bus & device.
+ */
+static int tile_cfg_read(struct pci_bus *bus, unsigned int devfn, int offset,
+			 int size, u32 *val)
+{
+	struct pci_controller *controller = bus->sysdata;
+	gxio_trio_context_t *trio_context = controller->trio;
+	int busnum = bus->number & 0xff;
+	int device = PCI_SLOT(devfn);
+	int function = PCI_FUNC(devfn);
+	int config_type = 1;
+	TRIO_TILE_PIO_REGION_SETUP_CFG_ADDR_t cfg_addr;
+	void *mmio_addr;
+
+	/*
+	 * Map all accesses to the local device on root bus into the
+	 * MMIO space of the MAC. Accesses to the downstream devices
+	 * go to the PIO space.
+	 */
+	if (pci_is_root_bus(bus)) {
+		if (device == 0) {
+			/*
+			 * This is the internal downstream P2P bridge,
+			 * access directly.
+			 */
+			unsigned int reg_offset;
+
+			reg_offset = ((offset & 0xFFF) <<
+				TRIO_CFG_REGION_ADDR__REG_SHIFT) |
+				(TRIO_CFG_REGION_ADDR__INTFC_VAL_MAC_PROTECTED
+				<< TRIO_CFG_REGION_ADDR__INTFC_SHIFT ) |
+				(controller->mac <<
+					TRIO_CFG_REGION_ADDR__MAC_SEL_SHIFT);
+
+			mmio_addr = trio_context->mmio_base_mac + reg_offset;
+
+			goto valid_device;
+
+		} else {
+			/*
+			 * We fake an empty device for (device > 0),
+			 * since there is only one device on bus 0.
+			 */
+			goto invalid_device;
+		}
+	}
+
+	/*
+	 * Accesses to the directly attached device have to be
+	 * sent as type-0 configs.
+	 */
+	if (busnum == (controller->first_busno + 1)) {
+		/*
+		 * There is only one device off of our built-in P2P bridge.
+		 */
+		if (device != 0)
+			goto invalid_device;
+
+		config_type = 0;
+	}
+
+	cfg_addr.word = 0;
+	cfg_addr.reg_addr = (offset & 0xFFF);
+	cfg_addr.fn = function;
+	cfg_addr.dev = device;
+	cfg_addr.bus = busnum;
+	cfg_addr.type = config_type;
+
+	/*
+	 * Note that we don't set the mac field in cfg_addr because the
+	 * mapping is per port.
+	 */
+	mmio_addr = trio_context->mmio_base_pio_cfg[controller->mac] +
+		cfg_addr.word;
+
+valid_device:
+
+	switch (size) {
+	case 4:
+		*val = __gxio_mmio_read32(mmio_addr);
+		break;
+
+	case 2:
+		*val = __gxio_mmio_read16(mmio_addr);
+		break;
+
+	case 1:
+		*val = __gxio_mmio_read8(mmio_addr);
+		break;
+
+	default:
+		return PCIBIOS_FUNC_NOT_SUPPORTED;
+	}
+
+	TRACE_CFG_RD(size, *val, busnum, device, function, offset);
+
+	return 0;
+
+invalid_device:
+
+	switch (size) {
+	case 4:
+		*val = 0xFFFFFFFF;
+		break;
+
+	case 2:
+		*val = 0xFFFF;
+		break;
+
+	case 1:
+		*val = 0xFF;
+		break;
+
+	default:
+		return PCIBIOS_FUNC_NOT_SUPPORTED;
+	}
+
+	return 0;
+}
+
+
+/*
+ * See tile_cfg_read() for relevent comments.
+ * Note that "val" is the value to write, not a pointer to that value.
+ */
+static int tile_cfg_write(struct pci_bus *bus, unsigned int devfn, int offset,
+			  int size, u32 val)
+{
+	struct pci_controller *controller = bus->sysdata;
+	gxio_trio_context_t *trio_context = controller->trio;
+	int busnum = bus->number & 0xff;
+	int device = PCI_SLOT(devfn);
+	int function = PCI_FUNC(devfn);
+	int config_type = 1;
+	TRIO_TILE_PIO_REGION_SETUP_CFG_ADDR_t cfg_addr;
+	void *mmio_addr;
+	u32 val_32 = (u32)val;
+	u16 val_16 = (u16)val;
+	u8 val_8 = (u8)val;
+
+	/*
+	 * Map all accesses to the local device on root bus into the
+	 * MMIO space of the MAC. Accesses to the downstream devices
+	 * go to the PIO space.
+	 */
+	if (pci_is_root_bus(bus)) {
+		if (device == 0) {
+			/*
+			 * This is the internal downstream P2P bridge,
+			 * access directly.
+			 */
+			unsigned int reg_offset;
+
+			reg_offset = ((offset & 0xFFF) <<
+				TRIO_CFG_REGION_ADDR__REG_SHIFT) |
+				(TRIO_CFG_REGION_ADDR__INTFC_VAL_MAC_PROTECTED
+				<< TRIO_CFG_REGION_ADDR__INTFC_SHIFT ) |
+				(controller->mac <<
+					TRIO_CFG_REGION_ADDR__MAC_SEL_SHIFT);
+
+			mmio_addr = trio_context->mmio_base_mac + reg_offset;
+
+			goto valid_device;
+
+		} else {
+			/*
+			 * We fake an empty device for (device > 0),
+			 * since there is only one device on bus 0.
+			 */
+			goto invalid_device;
+		}
+	}
+
+	/*
+	 * Accesses to the directly attached device have to be
+	 * sent as type-0 configs.
+	 */
+	if (busnum == (controller->first_busno + 1)) {
+		/*
+		 * There is only one device off of our built-in P2P bridge.
+		 */
+		if (device != 0)
+			goto invalid_device;
+
+		config_type = 0;
+	}
+
+	cfg_addr.word = 0;
+	cfg_addr.reg_addr = (offset & 0xFFF);
+	cfg_addr.fn = function;
+	cfg_addr.dev = device;
+	cfg_addr.bus = busnum;
+	cfg_addr.type = config_type;
+
+	/*
+	 * Note that we don't set the mac field in cfg_addr because the
+	 * mapping is per port.
+	 */
+	mmio_addr = trio_context->mmio_base_pio_cfg[controller->mac] +
+			cfg_addr.word;
+
+valid_device:
+
+	switch (size) {
+	case 4:
+		__gxio_mmio_write32(mmio_addr, val_32);
+		TRACE_CFG_WR(size, val_32, busnum, device, function, offset);
+		break;
+
+	case 2:
+		__gxio_mmio_write16(mmio_addr, val_16);
+		TRACE_CFG_WR(size, val_16, busnum, device, function, offset);
+		break;
+
+	case 1:
+		__gxio_mmio_write8(mmio_addr, val_8);
+		TRACE_CFG_WR(size, val_8, busnum, device, function, offset);
+		break;
+
+	default:
+		return PCIBIOS_FUNC_NOT_SUPPORTED;
+	}
+
+invalid_device:
+
+	return 0;
+}
+
+
+static struct pci_ops tile_cfg_ops = {
+	.read =         tile_cfg_read,
+	.write =        tile_cfg_write,
+};
+
+
+/* MSI support starts here. */
+static unsigned int tilegx_msi_startup(struct irq_data *d)
+{
+	if (d->msi_desc)
+		pci_msi_unmask_irq(d);
+
+	return 0;
+}
+
+static void tilegx_msi_ack(struct irq_data *d)
+{
+	__insn_mtspr(SPR_IPI_EVENT_RESET_K, 1UL << d->irq);
+}
+
+static void tilegx_msi_mask(struct irq_data *d)
+{
+	pci_msi_mask_irq(d);
+	__insn_mtspr(SPR_IPI_MASK_SET_K, 1UL << d->irq);
+}
+
+static void tilegx_msi_unmask(struct irq_data *d)
+{
+	__insn_mtspr(SPR_IPI_MASK_RESET_K, 1UL << d->irq);
+	pci_msi_unmask_irq(d);
+}
+
+static struct irq_chip tilegx_msi_chip = {
+	.name			= "tilegx_msi",
+	.irq_startup		= tilegx_msi_startup,
+	.irq_ack		= tilegx_msi_ack,
+	.irq_mask		= tilegx_msi_mask,
+	.irq_unmask		= tilegx_msi_unmask,
+
+	/* TBD: support set_affinity. */
+};
+
+int arch_setup_msi_irq(struct pci_dev *pdev, struct msi_desc *desc)
+{
+	struct pci_controller *controller;
+	gxio_trio_context_t *trio_context;
+	struct msi_msg msg;
+	int default_irq;
+	uint64_t mem_map_base;
+	uint64_t mem_map_limit;
+	u64 msi_addr;
+	int mem_map;
+	int cpu;
+	int irq;
+	int ret;
+
+	irq = irq_alloc_hwirq(-1);
+	if (!irq)
+		return -ENOSPC;
+
+	/*
+	 * Since we use a 64-bit Mem-Map to accept the MSI write, we fail
+	 * devices that are not capable of generating a 64-bit message address.
+	 * These devices will fall back to using the legacy interrupts.
+	 * Most PCIe endpoint devices do support 64-bit message addressing.
+	 */
+	if (desc->msi_attrib.is_64 == 0) {
+		dev_info(&pdev->dev, "64-bit MSI message address not supported, falling back to legacy interrupts\n");
+
+		ret = -ENOMEM;
+		goto is_64_failure;
+	}
+
+	default_irq = desc->msi_attrib.default_irq;
+	controller = irq_get_handler_data(default_irq);
+
+	BUG_ON(!controller);
+
+	trio_context = controller->trio;
+
+	/*
+	 * Allocate a scatter-queue that will accept the MSI write and
+	 * trigger the TILE-side interrupts. We use the scatter-queue regions
+	 * before the mem map regions, because the latter are needed by more
+	 * applications.
+	 */
+	mem_map = gxio_trio_alloc_scatter_queues(trio_context, 1, 0, 0);
+	if (mem_map >= 0) {
+		TRIO_MAP_SQ_DOORBELL_FMT_t doorbell_template = {{
+			.pop = 0,
+			.doorbell = 1,
+		}};
+
+		mem_map += TRIO_NUM_MAP_MEM_REGIONS;
+		mem_map_base = MEM_MAP_INTR_REGIONS_BASE +
+			mem_map * MEM_MAP_INTR_REGION_SIZE;
+		mem_map_limit = mem_map_base + MEM_MAP_INTR_REGION_SIZE - 1;
+
+		msi_addr = mem_map_base + MEM_MAP_INTR_REGION_SIZE - 8;
+		msg.data = (unsigned int)doorbell_template.word;
+	} else {
+		/* SQ regions are out, allocate from map mem regions. */
+		mem_map = gxio_trio_alloc_memory_maps(trio_context, 1, 0, 0);
+		if (mem_map < 0) {
+			dev_info(&pdev->dev, "%s Mem-Map alloc failure - failed to initialize MSI interrupts - falling back to legacy interrupts\n",
+				 desc->msi_attrib.is_msix ? "MSI-X" : "MSI");
+			ret = -ENOMEM;
+			goto msi_mem_map_alloc_failure;
+		}
+
+		mem_map_base = MEM_MAP_INTR_REGIONS_BASE +
+			mem_map * MEM_MAP_INTR_REGION_SIZE;
+		mem_map_limit = mem_map_base + MEM_MAP_INTR_REGION_SIZE - 1;
+
+		msi_addr = mem_map_base + TRIO_MAP_MEM_REG_INT3 -
+			TRIO_MAP_MEM_REG_INT0;
+
+		msg.data = mem_map;
+	}
+
+	/* We try to distribute different IRQs to different tiles. */
+	cpu = tile_irq_cpu(irq);
+
+	/*
+	 * Now call up to the HV to configure the MSI interrupt and
+	 * set up the IPI binding.
+	 */
+	ret = gxio_trio_config_msi_intr(trio_context, cpu_x(cpu), cpu_y(cpu),
+					KERNEL_PL, irq, controller->mac,
+					mem_map, mem_map_base, mem_map_limit,
+					trio_context->asid);
+	if (ret < 0) {
+		dev_info(&pdev->dev, "HV MSI config failed\n");
+
+		goto hv_msi_config_failure;
+	}
+
+	irq_set_msi_desc(irq, desc);
+
+	msg.address_hi = msi_addr >> 32;
+	msg.address_lo = msi_addr & 0xffffffff;
+
+	pci_write_msi_msg(irq, &msg);
+	irq_set_chip_and_handler(irq, &tilegx_msi_chip, handle_level_irq);
+	irq_set_handler_data(irq, controller);
+
+	return 0;
+
+hv_msi_config_failure:
+	/* Free mem-map */
+msi_mem_map_alloc_failure:
+is_64_failure:
+	irq_free_hwirq(irq);
+	return ret;
+}
+
+void arch_teardown_msi_irq(unsigned int irq)
+{
+	irq_free_hwirq(irq);
+}
diff --git a/kernel/arch/tile/kernel/perf_event.c b/kernel/arch/tile/kernel/perf_event.c
new file mode 100644
index 000000000..bb509cee3
--- /dev/null
+++ b/kernel/arch/tile/kernel/perf_event.c
@@ -0,0 +1,1005 @@
+/*
+ * Copyright 2014 Tilera Corporation. All Rights Reserved.
+ *
+ *   This program is free software; you can redistribute it and/or
+ *   modify it under the terms of the GNU General Public License
+ *   as published by the Free Software Foundation, version 2.
+ *
+ *   This program is distributed in the hope that it will be useful, but
+ *   WITHOUT ANY WARRANTY; without even the implied warranty of
+ *   MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
+ *   NON INFRINGEMENT.  See the GNU General Public License for
+ *   more details.
+ *
+ *
+ * Perf_events support for Tile processor.
+ *
+ * This code is based upon the x86 perf event
+ * code, which is:
+ *
+ *  Copyright (C) 2008 Thomas Gleixner <tglx@linutronix.de>
+ *  Copyright (C) 2008-2009 Red Hat, Inc., Ingo Molnar
+ *  Copyright (C) 2009 Jaswinder Singh Rajput
+ *  Copyright (C) 2009 Advanced Micro Devices, Inc., Robert Richter
+ *  Copyright (C) 2008-2009 Red Hat, Inc., Peter Zijlstra <pzijlstr@redhat.com>
+ *  Copyright (C) 2009 Intel Corporation, <markus.t.metzger@intel.com>
+ *  Copyright (C) 2009 Google, Inc., Stephane Eranian
+ */
+
+#include <linux/kprobes.h>
+#include <linux/kernel.h>
+#include <linux/kdebug.h>
+#include <linux/mutex.h>
+#include <linux/bitmap.h>
+#include <linux/irq.h>
+#include <linux/interrupt.h>
+#include <linux/perf_event.h>
+#include <linux/atomic.h>
+#include <asm/traps.h>
+#include <asm/stack.h>
+#include <asm/pmc.h>
+#include <hv/hypervisor.h>
+
+#define TILE_MAX_COUNTERS	4
+
+#define PERF_COUNT_0_IDX	0
+#define PERF_COUNT_1_IDX	1
+#define AUX_PERF_COUNT_0_IDX	2
+#define AUX_PERF_COUNT_1_IDX	3
+
+struct cpu_hw_events {
+	int			n_events;
+	struct perf_event	*events[TILE_MAX_COUNTERS]; /* counter order */
+	struct perf_event	*event_list[TILE_MAX_COUNTERS]; /* enabled
+								order */
+	int			assign[TILE_MAX_COUNTERS];
+	unsigned long		active_mask[BITS_TO_LONGS(TILE_MAX_COUNTERS)];
+	unsigned long		used_mask;
+};
+
+/* TILE arch specific performance monitor unit */
+struct tile_pmu {
+	const char	*name;
+	int		version;
+	const int	*hw_events;	/* generic hw events table */
+	/* generic hw cache events table */
+	const int	(*cache_events)[PERF_COUNT_HW_CACHE_MAX]
+				       [PERF_COUNT_HW_CACHE_OP_MAX]
+				       [PERF_COUNT_HW_CACHE_RESULT_MAX];
+	int		(*map_hw_event)(u64);	 /*method used to map
+						  hw events */
+	int		(*map_cache_event)(u64); /*method used to map
+						  cache events */
+
+	u64		max_period;		/* max sampling period */
+	u64		cntval_mask;		/* counter width mask */
+	int		cntval_bits;		/* counter width */
+	int		max_events;		/* max generic hw events
+						in map */
+	int		num_counters;		/* number base + aux counters */
+	int		num_base_counters;	/* number base counters */
+};
+
+DEFINE_PER_CPU(u64, perf_irqs);
+static DEFINE_PER_CPU(struct cpu_hw_events, cpu_hw_events);
+
+#define TILE_OP_UNSUPP		(-1)
+
+#ifndef __tilegx__
+/* TILEPro hardware events map */
+static const int tile_hw_event_map[] = {
+	[PERF_COUNT_HW_CPU_CYCLES]		= 0x01, /* ONE */
+	[PERF_COUNT_HW_INSTRUCTIONS]		= 0x06, /* MP_BUNDLE_RETIRED */
+	[PERF_COUNT_HW_CACHE_REFERENCES]	= TILE_OP_UNSUPP,
+	[PERF_COUNT_HW_CACHE_MISSES]		= TILE_OP_UNSUPP,
+	[PERF_COUNT_HW_BRANCH_INSTRUCTIONS]	= 0x16, /*
+					  MP_CONDITIONAL_BRANCH_ISSUED */
+	[PERF_COUNT_HW_BRANCH_MISSES]		= 0x14, /*
+					  MP_CONDITIONAL_BRANCH_MISSPREDICT */
+	[PERF_COUNT_HW_BUS_CYCLES]		= TILE_OP_UNSUPP,
+};
+#else
+/* TILEGx hardware events map */
+static const int tile_hw_event_map[] = {
+	[PERF_COUNT_HW_CPU_CYCLES]		= 0x181, /* ONE */
+	[PERF_COUNT_HW_INSTRUCTIONS]		= 0xdb, /* INSTRUCTION_BUNDLE */
+	[PERF_COUNT_HW_CACHE_REFERENCES]	= TILE_OP_UNSUPP,
+	[PERF_COUNT_HW_CACHE_MISSES]		= TILE_OP_UNSUPP,
+	[PERF_COUNT_HW_BRANCH_INSTRUCTIONS]	= 0xd9, /*
+						COND_BRANCH_PRED_CORRECT */
+	[PERF_COUNT_HW_BRANCH_MISSES]		= 0xda, /*
+						COND_BRANCH_PRED_INCORRECT */
+	[PERF_COUNT_HW_BUS_CYCLES]		= TILE_OP_UNSUPP,
+};
+#endif
+
+#define C(x) PERF_COUNT_HW_CACHE_##x
+
+/*
+ * Generalized hw caching related hw_event table, filled
+ * in on a per model basis. A value of -1 means
+ * 'not supported', any other value means the
+ * raw hw_event ID.
+ */
+#ifndef __tilegx__
+/* TILEPro hardware cache event map */
+static const int tile_cache_event_map[PERF_COUNT_HW_CACHE_MAX]
+				     [PERF_COUNT_HW_CACHE_OP_MAX]
+				     [PERF_COUNT_HW_CACHE_RESULT_MAX] = {
+[C(L1D)] = {
+	[C(OP_READ)] = {
+		[C(RESULT_ACCESS)] = TILE_OP_UNSUPP,
+		[C(RESULT_MISS)] = 0x21, /* RD_MISS */
+	},
+	[C(OP_WRITE)] = {
+		[C(RESULT_ACCESS)] = TILE_OP_UNSUPP,
+		[C(RESULT_MISS)] = 0x22, /* WR_MISS */
+	},
+	[C(OP_PREFETCH)] = {
+		[C(RESULT_ACCESS)] = TILE_OP_UNSUPP,
+		[C(RESULT_MISS)] = TILE_OP_UNSUPP,
+	},
+},
+[C(L1I)] = {
+	[C(OP_READ)] = {
+		[C(RESULT_ACCESS)] = 0x12, /* MP_ICACHE_HIT_ISSUED */
+		[C(RESULT_MISS)] = TILE_OP_UNSUPP,
+	},
+	[C(OP_WRITE)] = {
+		[C(RESULT_ACCESS)] = TILE_OP_UNSUPP,
+		[C(RESULT_MISS)] = TILE_OP_UNSUPP,
+	},
+	[C(OP_PREFETCH)] = {
+		[C(RESULT_ACCESS)] = TILE_OP_UNSUPP,
+		[C(RESULT_MISS)] = TILE_OP_UNSUPP,
+	},
+},
+[C(LL)] = {
+	[C(OP_READ)] = {
+		[C(RESULT_ACCESS)] = TILE_OP_UNSUPP,
+		[C(RESULT_MISS)] = TILE_OP_UNSUPP,
+	},
+	[C(OP_WRITE)] = {
+		[C(RESULT_ACCESS)] = TILE_OP_UNSUPP,
+		[C(RESULT_MISS)] = TILE_OP_UNSUPP,
+	},
+	[C(OP_PREFETCH)] = {
+		[C(RESULT_ACCESS)] = TILE_OP_UNSUPP,
+		[C(RESULT_MISS)] = TILE_OP_UNSUPP,
+	},
+},
+[C(DTLB)] = {
+	[C(OP_READ)] = {
+		[C(RESULT_ACCESS)] = 0x1d, /* TLB_CNT */
+		[C(RESULT_MISS)] = 0x20, /* TLB_EXCEPTION */
+	},
+	[C(OP_WRITE)] = {
+		[C(RESULT_ACCESS)] = TILE_OP_UNSUPP,
+		[C(RESULT_MISS)] = TILE_OP_UNSUPP,
+	},
+	[C(OP_PREFETCH)] = {
+		[C(RESULT_ACCESS)] = TILE_OP_UNSUPP,
+		[C(RESULT_MISS)] = TILE_OP_UNSUPP,
+	},
+},
+[C(ITLB)] = {
+	[C(OP_READ)] = {
+		[C(RESULT_ACCESS)] = 0x13, /* MP_ITLB_HIT_ISSUED */
+		[C(RESULT_MISS)] = TILE_OP_UNSUPP,
+	},
+	[C(OP_WRITE)] = {
+		[C(RESULT_ACCESS)] = TILE_OP_UNSUPP,
+		[C(RESULT_MISS)] = TILE_OP_UNSUPP,
+	},
+	[C(OP_PREFETCH)] = {
+		[C(RESULT_ACCESS)] = TILE_OP_UNSUPP,
+		[C(RESULT_MISS)] = TILE_OP_UNSUPP,
+	},
+},
+[C(BPU)] = {
+	[C(OP_READ)] = {
+		[C(RESULT_ACCESS)] = TILE_OP_UNSUPP,
+		[C(RESULT_MISS)] = TILE_OP_UNSUPP,
+	},
+	[C(OP_WRITE)] = {
+		[C(RESULT_ACCESS)] = TILE_OP_UNSUPP,
+		[C(RESULT_MISS)] = TILE_OP_UNSUPP,
+	},
+	[C(OP_PREFETCH)] = {
+		[C(RESULT_ACCESS)] = TILE_OP_UNSUPP,
+		[C(RESULT_MISS)] = TILE_OP_UNSUPP,
+	},
+},
+};
+#else
+/* TILEGx hardware events map */
+static const int tile_cache_event_map[PERF_COUNT_HW_CACHE_MAX]
+				     [PERF_COUNT_HW_CACHE_OP_MAX]
+				     [PERF_COUNT_HW_CACHE_RESULT_MAX] = {
+[C(L1D)] = {
+	/*
+	 * Like some other architectures (e.g. ARM), the performance
+	 * counters don't differentiate between read and write
+	 * accesses/misses, so this isn't strictly correct, but it's the
+	 * best we can do. Writes and reads get combined.
+	 */
+	[C(OP_READ)] = {
+		[C(RESULT_ACCESS)] = TILE_OP_UNSUPP,
+		[C(RESULT_MISS)] = 0x44, /* RD_MISS */
+	},
+	[C(OP_WRITE)] = {
+		[C(RESULT_ACCESS)] = TILE_OP_UNSUPP,
+		[C(RESULT_MISS)] = 0x45, /* WR_MISS */
+	},
+	[C(OP_PREFETCH)] = {
+		[C(RESULT_ACCESS)] = TILE_OP_UNSUPP,
+		[C(RESULT_MISS)] = TILE_OP_UNSUPP,
+	},
+},
+[C(L1I)] = {
+	[C(OP_READ)] = {
+		[C(RESULT_ACCESS)] = TILE_OP_UNSUPP,
+		[C(RESULT_MISS)] = TILE_OP_UNSUPP,
+	},
+	[C(OP_WRITE)] = {
+		[C(RESULT_ACCESS)] = TILE_OP_UNSUPP,
+		[C(RESULT_MISS)] = TILE_OP_UNSUPP,
+	},
+	[C(OP_PREFETCH)] = {
+		[C(RESULT_ACCESS)] = TILE_OP_UNSUPP,
+		[C(RESULT_MISS)] = TILE_OP_UNSUPP,
+	},
+},
+[C(LL)] = {
+	[C(OP_READ)] = {
+		[C(RESULT_ACCESS)] = TILE_OP_UNSUPP,
+		[C(RESULT_MISS)] = TILE_OP_UNSUPP,
+	},
+	[C(OP_WRITE)] = {
+		[C(RESULT_ACCESS)] = TILE_OP_UNSUPP,
+		[C(RESULT_MISS)] = TILE_OP_UNSUPP,
+	},
+	[C(OP_PREFETCH)] = {
+		[C(RESULT_ACCESS)] = TILE_OP_UNSUPP,
+		[C(RESULT_MISS)] = TILE_OP_UNSUPP,
+	},
+},
+[C(DTLB)] = {
+	[C(OP_READ)] = {
+		[C(RESULT_ACCESS)] = 0x40, /* TLB_CNT */
+		[C(RESULT_MISS)] = 0x43, /* TLB_EXCEPTION */
+	},
+	[C(OP_WRITE)] = {
+		[C(RESULT_ACCESS)] = 0x40, /* TLB_CNT */
+		[C(RESULT_MISS)] = 0x43, /* TLB_EXCEPTION */
+	},
+	[C(OP_PREFETCH)] = {
+		[C(RESULT_ACCESS)] = TILE_OP_UNSUPP,
+		[C(RESULT_MISS)] = TILE_OP_UNSUPP,
+	},
+},
+[C(ITLB)] = {
+	[C(OP_READ)] = {
+		[C(RESULT_ACCESS)] = TILE_OP_UNSUPP,
+		[C(RESULT_MISS)] = 0xd4, /* ITLB_MISS_INT */
+	},
+	[C(OP_WRITE)] = {
+		[C(RESULT_ACCESS)] = TILE_OP_UNSUPP,
+		[C(RESULT_MISS)] = 0xd4, /* ITLB_MISS_INT */
+	},
+	[C(OP_PREFETCH)] = {
+		[C(RESULT_ACCESS)] = TILE_OP_UNSUPP,
+		[C(RESULT_MISS)] = TILE_OP_UNSUPP,
+	},
+},
+[C(BPU)] = {
+	[C(OP_READ)] = {
+		[C(RESULT_ACCESS)] = TILE_OP_UNSUPP,
+		[C(RESULT_MISS)] = TILE_OP_UNSUPP,
+	},
+	[C(OP_WRITE)] = {
+		[C(RESULT_ACCESS)] = TILE_OP_UNSUPP,
+		[C(RESULT_MISS)] = TILE_OP_UNSUPP,
+	},
+	[C(OP_PREFETCH)] = {
+		[C(RESULT_ACCESS)] = TILE_OP_UNSUPP,
+		[C(RESULT_MISS)] = TILE_OP_UNSUPP,
+	},
+},
+};
+#endif
+
+static atomic_t tile_active_events;
+static DEFINE_MUTEX(perf_intr_reserve_mutex);
+
+static int tile_map_hw_event(u64 config);
+static int tile_map_cache_event(u64 config);
+
+static int tile_pmu_handle_irq(struct pt_regs *regs, int fault);
+
+/*
+ * To avoid new_raw_count getting larger then pre_raw_count
+ * in tile_perf_event_update(), we limit the value of max_period to 2^31 - 1.
+ */
+static const struct tile_pmu tilepmu = {
+#ifndef __tilegx__
+	.name = "tilepro",
+#else
+	.name = "tilegx",
+#endif
+	.max_events = ARRAY_SIZE(tile_hw_event_map),
+	.map_hw_event = tile_map_hw_event,
+	.hw_events = tile_hw_event_map,
+	.map_cache_event = tile_map_cache_event,
+	.cache_events = &tile_cache_event_map,
+	.cntval_bits = 32,
+	.cntval_mask = (1ULL << 32) - 1,
+	.max_period = (1ULL << 31) - 1,
+	.num_counters = TILE_MAX_COUNTERS,
+	.num_base_counters = TILE_BASE_COUNTERS,
+};
+
+static const struct tile_pmu *tile_pmu __read_mostly;
+
+/*
+ * Check whether perf event is enabled.
+ */
+int tile_perf_enabled(void)
+{
+	return atomic_read(&tile_active_events) != 0;
+}
+
+/*
+ * Read Performance Counters.
+ */
+static inline u64 read_counter(int idx)
+{
+	u64 val = 0;
+
+	/* __insn_mfspr() only takes an immediate argument */
+	switch (idx) {
+	case PERF_COUNT_0_IDX:
+		val = __insn_mfspr(SPR_PERF_COUNT_0);
+		break;
+	case PERF_COUNT_1_IDX:
+		val = __insn_mfspr(SPR_PERF_COUNT_1);
+		break;
+	case AUX_PERF_COUNT_0_IDX:
+		val = __insn_mfspr(SPR_AUX_PERF_COUNT_0);
+		break;
+	case AUX_PERF_COUNT_1_IDX:
+		val = __insn_mfspr(SPR_AUX_PERF_COUNT_1);
+		break;
+	default:
+		WARN_ON_ONCE(idx > AUX_PERF_COUNT_1_IDX ||
+				idx < PERF_COUNT_0_IDX);
+	}
+
+	return val;
+}
+
+/*
+ * Write Performance Counters.
+ */
+static inline void write_counter(int idx, u64 value)
+{
+	/* __insn_mtspr() only takes an immediate argument */
+	switch (idx) {
+	case PERF_COUNT_0_IDX:
+		__insn_mtspr(SPR_PERF_COUNT_0, value);
+		break;
+	case PERF_COUNT_1_IDX:
+		__insn_mtspr(SPR_PERF_COUNT_1, value);
+		break;
+	case AUX_PERF_COUNT_0_IDX:
+		__insn_mtspr(SPR_AUX_PERF_COUNT_0, value);
+		break;
+	case AUX_PERF_COUNT_1_IDX:
+		__insn_mtspr(SPR_AUX_PERF_COUNT_1, value);
+		break;
+	default:
+		WARN_ON_ONCE(idx > AUX_PERF_COUNT_1_IDX ||
+				idx < PERF_COUNT_0_IDX);
+	}
+}
+
+/*
+ * Enable performance event by setting
+ * Performance Counter Control registers.
+ */
+static inline void tile_pmu_enable_event(struct perf_event *event)
+{
+	struct hw_perf_event *hwc = &event->hw;
+	unsigned long cfg, mask;
+	int shift, idx = hwc->idx;
+
+	/*
+	 * prevent early activation from tile_pmu_start() in hw_perf_enable
+	 */
+
+	if (WARN_ON_ONCE(idx == -1))
+		return;
+
+	if (idx < tile_pmu->num_base_counters)
+		cfg = __insn_mfspr(SPR_PERF_COUNT_CTL);
+	else
+		cfg = __insn_mfspr(SPR_AUX_PERF_COUNT_CTL);
+
+	switch (idx) {
+	case PERF_COUNT_0_IDX:
+	case AUX_PERF_COUNT_0_IDX:
+		mask = TILE_EVENT_MASK;
+		shift = 0;
+		break;
+	case PERF_COUNT_1_IDX:
+	case AUX_PERF_COUNT_1_IDX:
+		mask = TILE_EVENT_MASK << 16;
+		shift = 16;
+		break;
+	default:
+		WARN_ON_ONCE(idx < PERF_COUNT_0_IDX ||
+			idx > AUX_PERF_COUNT_1_IDX);
+		return;
+	}
+
+	/* Clear mask bits to enable the event. */
+	cfg &= ~mask;
+	cfg |= hwc->config << shift;
+
+	if (idx < tile_pmu->num_base_counters)
+		__insn_mtspr(SPR_PERF_COUNT_CTL, cfg);
+	else
+		__insn_mtspr(SPR_AUX_PERF_COUNT_CTL, cfg);
+}
+
+/*
+ * Disable performance event by clearing
+ * Performance Counter Control registers.
+ */
+static inline void tile_pmu_disable_event(struct perf_event *event)
+{
+	struct hw_perf_event *hwc = &event->hw;
+	unsigned long cfg, mask;
+	int idx = hwc->idx;
+
+	if (idx == -1)
+		return;
+
+	if (idx < tile_pmu->num_base_counters)
+		cfg = __insn_mfspr(SPR_PERF_COUNT_CTL);
+	else
+		cfg = __insn_mfspr(SPR_AUX_PERF_COUNT_CTL);
+
+	switch (idx) {
+	case PERF_COUNT_0_IDX:
+	case AUX_PERF_COUNT_0_IDX:
+		mask = TILE_PLM_MASK;
+		break;
+	case PERF_COUNT_1_IDX:
+	case AUX_PERF_COUNT_1_IDX:
+		mask = TILE_PLM_MASK << 16;
+		break;
+	default:
+		WARN_ON_ONCE(idx < PERF_COUNT_0_IDX ||
+			idx > AUX_PERF_COUNT_1_IDX);
+		return;
+	}
+
+	/* Set mask bits to disable the event. */
+	cfg |= mask;
+
+	if (idx < tile_pmu->num_base_counters)
+		__insn_mtspr(SPR_PERF_COUNT_CTL, cfg);
+	else
+		__insn_mtspr(SPR_AUX_PERF_COUNT_CTL, cfg);
+}
+
+/*
+ * Propagate event elapsed time into the generic event.
+ * Can only be executed on the CPU where the event is active.
+ * Returns the delta events processed.
+ */
+static u64 tile_perf_event_update(struct perf_event *event)
+{
+	struct hw_perf_event *hwc = &event->hw;
+	int shift = 64 - tile_pmu->cntval_bits;
+	u64 prev_raw_count, new_raw_count;
+	u64 oldval;
+	int idx = hwc->idx;
+	u64 delta;
+
+	/*
+	 * Careful: an NMI might modify the previous event value.
+	 *
+	 * Our tactic to handle this is to first atomically read and
+	 * exchange a new raw count - then add that new-prev delta
+	 * count to the generic event atomically:
+	 */
+again:
+	prev_raw_count = local64_read(&hwc->prev_count);
+	new_raw_count = read_counter(idx);
+
+	oldval = local64_cmpxchg(&hwc->prev_count, prev_raw_count,
+				 new_raw_count);
+	if (oldval != prev_raw_count)
+		goto again;
+
+	/*
+	 * Now we have the new raw value and have updated the prev
+	 * timestamp already. We can now calculate the elapsed delta
+	 * (event-)time and add that to the generic event.
+	 *
+	 * Careful, not all hw sign-extends above the physical width
+	 * of the count.
+	 */
+	delta = (new_raw_count << shift) - (prev_raw_count << shift);
+	delta >>= shift;
+
+	local64_add(delta, &event->count);
+	local64_sub(delta, &hwc->period_left);
+
+	return new_raw_count;
+}
+
+/*
+ * Set the next IRQ period, based on the hwc->period_left value.
+ * To be called with the event disabled in hw:
+ */
+static int tile_event_set_period(struct perf_event *event)
+{
+	struct hw_perf_event *hwc = &event->hw;
+	int idx = hwc->idx;
+	s64 left = local64_read(&hwc->period_left);
+	s64 period = hwc->sample_period;
+	int ret = 0;
+
+	/*
+	 * If we are way outside a reasonable range then just skip forward:
+	 */
+	if (unlikely(left <= -period)) {
+		left = period;
+		local64_set(&hwc->period_left, left);
+		hwc->last_period = period;
+		ret = 1;
+	}
+
+	if (unlikely(left <= 0)) {
+		left += period;
+		local64_set(&hwc->period_left, left);
+		hwc->last_period = period;
+		ret = 1;
+	}
+	if (left > tile_pmu->max_period)
+		left = tile_pmu->max_period;
+
+	/*
+	 * The hw event starts counting from this event offset,
+	 * mark it to be able to extra future deltas:
+	 */
+	local64_set(&hwc->prev_count, (u64)-left);
+
+	write_counter(idx, (u64)(-left) & tile_pmu->cntval_mask);
+
+	perf_event_update_userpage(event);
+
+	return ret;
+}
+
+/*
+ * Stop the event but do not release the PMU counter
+ */
+static void tile_pmu_stop(struct perf_event *event, int flags)
+{
+	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
+	struct hw_perf_event *hwc = &event->hw;
+	int idx = hwc->idx;
+
+	if (__test_and_clear_bit(idx, cpuc->active_mask)) {
+		tile_pmu_disable_event(event);
+		cpuc->events[hwc->idx] = NULL;
+		WARN_ON_ONCE(hwc->state & PERF_HES_STOPPED);
+		hwc->state |= PERF_HES_STOPPED;
+	}
+
+	if ((flags & PERF_EF_UPDATE) && !(hwc->state & PERF_HES_UPTODATE)) {
+		/*
+		 * Drain the remaining delta count out of a event
+		 * that we are disabling:
+		 */
+		tile_perf_event_update(event);
+		hwc->state |= PERF_HES_UPTODATE;
+	}
+}
+
+/*
+ * Start an event (without re-assigning counter)
+ */
+static void tile_pmu_start(struct perf_event *event, int flags)
+{
+	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
+	int idx = event->hw.idx;
+
+	if (WARN_ON_ONCE(!(event->hw.state & PERF_HES_STOPPED)))
+		return;
+
+	if (WARN_ON_ONCE(idx == -1))
+		return;
+
+	if (flags & PERF_EF_RELOAD) {
+		WARN_ON_ONCE(!(event->hw.state & PERF_HES_UPTODATE));
+		tile_event_set_period(event);
+	}
+
+	event->hw.state = 0;
+
+	cpuc->events[idx] = event;
+	__set_bit(idx, cpuc->active_mask);
+
+	unmask_pmc_interrupts();
+
+	tile_pmu_enable_event(event);
+
+	perf_event_update_userpage(event);
+}
+
+/*
+ * Add a single event to the PMU.
+ *
+ * The event is added to the group of enabled events
+ * but only if it can be scehduled with existing events.
+ */
+static int tile_pmu_add(struct perf_event *event, int flags)
+{
+	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
+	struct hw_perf_event *hwc;
+	unsigned long mask;
+	int b, max_cnt;
+
+	hwc = &event->hw;
+
+	/*
+	 * We are full.
+	 */
+	if (cpuc->n_events == tile_pmu->num_counters)
+		return -ENOSPC;
+
+	cpuc->event_list[cpuc->n_events] = event;
+	cpuc->n_events++;
+
+	hwc->state = PERF_HES_UPTODATE | PERF_HES_STOPPED;
+	if (!(flags & PERF_EF_START))
+		hwc->state |= PERF_HES_ARCH;
+
+	/*
+	 * Find first empty counter.
+	 */
+	max_cnt = tile_pmu->num_counters;
+	mask = ~cpuc->used_mask;
+
+	/* Find next free counter. */
+	b = find_next_bit(&mask, max_cnt, 0);
+
+	/* Should not happen. */
+	if (WARN_ON_ONCE(b == max_cnt))
+		return -ENOSPC;
+
+	/*
+	 * Assign counter to event.
+	 */
+	event->hw.idx = b;
+	__set_bit(b, &cpuc->used_mask);
+
+	/*
+	 * Start if requested.
+	 */
+	if (flags & PERF_EF_START)
+		tile_pmu_start(event, PERF_EF_RELOAD);
+
+	return 0;
+}
+
+/*
+ * Delete a single event from the PMU.
+ *
+ * The event is deleted from the group of enabled events.
+ * If it is the last event, disable PMU interrupt.
+ */
+static void tile_pmu_del(struct perf_event *event, int flags)
+{
+	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
+	int i;
+
+	/*
+	 * Remove event from list, compact list if necessary.
+	 */
+	for (i = 0; i < cpuc->n_events; i++) {
+		if (cpuc->event_list[i] == event) {
+			while (++i < cpuc->n_events)
+				cpuc->event_list[i-1] = cpuc->event_list[i];
+			--cpuc->n_events;
+			cpuc->events[event->hw.idx] = NULL;
+			__clear_bit(event->hw.idx, &cpuc->used_mask);
+			tile_pmu_stop(event, PERF_EF_UPDATE);
+			break;
+		}
+	}
+	/*
+	 * If there are no events left, then mask PMU interrupt.
+	 */
+	if (cpuc->n_events == 0)
+		mask_pmc_interrupts();
+	perf_event_update_userpage(event);
+}
+
+/*
+ * Propagate event elapsed time into the event.
+ */
+static inline void tile_pmu_read(struct perf_event *event)
+{
+	tile_perf_event_update(event);
+}
+
+/*
+ * Map generic events to Tile PMU.
+ */
+static int tile_map_hw_event(u64 config)
+{
+	if (config >= tile_pmu->max_events)
+		return -EINVAL;
+	return tile_pmu->hw_events[config];
+}
+
+/*
+ * Map generic hardware cache events to Tile PMU.
+ */
+static int tile_map_cache_event(u64 config)
+{
+	unsigned int cache_type, cache_op, cache_result;
+	int code;
+
+	if (!tile_pmu->cache_events)
+		return -ENOENT;
+
+	cache_type = (config >>  0) & 0xff;
+	if (cache_type >= PERF_COUNT_HW_CACHE_MAX)
+		return -EINVAL;
+
+	cache_op = (config >>  8) & 0xff;
+	if (cache_op >= PERF_COUNT_HW_CACHE_OP_MAX)
+		return -EINVAL;
+
+	cache_result = (config >> 16) & 0xff;
+	if (cache_result >= PERF_COUNT_HW_CACHE_RESULT_MAX)
+		return -EINVAL;
+
+	code = (*tile_pmu->cache_events)[cache_type][cache_op][cache_result];
+	if (code == TILE_OP_UNSUPP)
+		return -EINVAL;
+
+	return code;
+}
+
+static void tile_event_destroy(struct perf_event *event)
+{
+	if (atomic_dec_return(&tile_active_events) == 0)
+		release_pmc_hardware();
+}
+
+static int __tile_event_init(struct perf_event *event)
+{
+	struct perf_event_attr *attr = &event->attr;
+	struct hw_perf_event *hwc = &event->hw;
+	int code;
+
+	switch (attr->type) {
+	case PERF_TYPE_HARDWARE:
+		code = tile_pmu->map_hw_event(attr->config);
+		break;
+	case PERF_TYPE_HW_CACHE:
+		code = tile_pmu->map_cache_event(attr->config);
+		break;
+	case PERF_TYPE_RAW:
+		code = attr->config & TILE_EVENT_MASK;
+		break;
+	default:
+		/* Should not happen. */
+		return -EOPNOTSUPP;
+	}
+
+	if (code < 0)
+		return code;
+
+	hwc->config = code;
+	hwc->idx = -1;
+
+	if (attr->exclude_user)
+		hwc->config |= TILE_CTL_EXCL_USER;
+
+	if (attr->exclude_kernel)
+		hwc->config |= TILE_CTL_EXCL_KERNEL;
+
+	if (attr->exclude_hv)
+		hwc->config |= TILE_CTL_EXCL_HV;
+
+	if (!hwc->sample_period) {
+		hwc->sample_period = tile_pmu->max_period;
+		hwc->last_period = hwc->sample_period;
+		local64_set(&hwc->period_left, hwc->sample_period);
+	}
+	event->destroy = tile_event_destroy;
+	return 0;
+}
+
+static int tile_event_init(struct perf_event *event)
+{
+	int err = 0;
+	perf_irq_t old_irq_handler = NULL;
+
+	if (atomic_inc_return(&tile_active_events) == 1)
+		old_irq_handler = reserve_pmc_hardware(tile_pmu_handle_irq);
+
+	if (old_irq_handler) {
+		pr_warn("PMC hardware busy (reserved by oprofile)\n");
+
+		atomic_dec(&tile_active_events);
+		return -EBUSY;
+	}
+
+	switch (event->attr.type) {
+	case PERF_TYPE_RAW:
+	case PERF_TYPE_HARDWARE:
+	case PERF_TYPE_HW_CACHE:
+		break;
+
+	default:
+		return -ENOENT;
+	}
+
+	err = __tile_event_init(event);
+	if (err) {
+		if (event->destroy)
+			event->destroy(event);
+	}
+	return err;
+}
+
+static struct pmu tilera_pmu = {
+	.event_init	= tile_event_init,
+	.add		= tile_pmu_add,
+	.del		= tile_pmu_del,
+
+	.start		= tile_pmu_start,
+	.stop		= tile_pmu_stop,
+
+	.read		= tile_pmu_read,
+};
+
+/*
+ * PMU's IRQ handler, PMU has 2 interrupts, they share the same handler.
+ */
+int tile_pmu_handle_irq(struct pt_regs *regs, int fault)
+{
+	struct perf_sample_data data;
+	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
+	struct perf_event *event;
+	struct hw_perf_event *hwc;
+	u64 val;
+	unsigned long status;
+	int bit;
+
+	__this_cpu_inc(perf_irqs);
+
+	if (!atomic_read(&tile_active_events))
+		return 0;
+
+	status = pmc_get_overflow();
+	pmc_ack_overflow(status);
+
+	for_each_set_bit(bit, &status, tile_pmu->num_counters) {
+
+		event = cpuc->events[bit];
+
+		if (!event)
+			continue;
+
+		if (!test_bit(bit, cpuc->active_mask))
+			continue;
+
+		hwc = &event->hw;
+
+		val = tile_perf_event_update(event);
+		if (val & (1ULL << (tile_pmu->cntval_bits - 1)))
+			continue;
+
+		perf_sample_data_init(&data, 0, event->hw.last_period);
+		if (!tile_event_set_period(event))
+			continue;
+
+		if (perf_event_overflow(event, &data, regs))
+			tile_pmu_stop(event, 0);
+	}
+
+	return 0;
+}
+
+static bool __init supported_pmu(void)
+{
+	tile_pmu = &tilepmu;
+	return true;
+}
+
+int __init init_hw_perf_events(void)
+{
+	supported_pmu();
+	perf_pmu_register(&tilera_pmu, "cpu", PERF_TYPE_RAW);
+	return 0;
+}
+arch_initcall(init_hw_perf_events);
+
+/* Callchain handling code. */
+
+/*
+ * Tile specific backtracing code for perf_events.
+ */
+static inline void perf_callchain(struct perf_callchain_entry *entry,
+		    struct pt_regs *regs)
+{
+	struct KBacktraceIterator kbt;
+	unsigned int i;
+
+	/*
+	 * Get the address just after the "jalr" instruction that
+	 * jumps to the handler for a syscall.  When we find this
+	 * address in a backtrace, we silently ignore it, which gives
+	 * us a one-step backtrace connection from the sys_xxx()
+	 * function in the kernel to the xxx() function in libc.
+	 * Otherwise, we lose the ability to properly attribute time
+	 * from the libc calls to the kernel implementations, since
+	 * oprofile only considers PCs from backtraces a pair at a time.
+	 */
+	unsigned long handle_syscall_pc = handle_syscall_link_address();
+
+	KBacktraceIterator_init(&kbt, NULL, regs);
+	kbt.profile = 1;
+
+	/*
+	 * The sample for the pc is already recorded.  Now we are adding the
+	 * address of the callsites on the stack.  Our iterator starts
+	 * with the frame of the (already sampled) call site.  If our
+	 * iterator contained a "return address" field, we could have just
+	 * used it and wouldn't have needed to skip the first
+	 * frame.  That's in effect what the arm and x86 versions do.
+	 * Instead we peel off the first iteration to get the equivalent
+	 * behavior.
+	 */
+
+	if (KBacktraceIterator_end(&kbt))
+		return;
+	KBacktraceIterator_next(&kbt);
+
+	/*
+	 * Set stack depth to 16 for user and kernel space respectively, that
+	 * is, total 32 stack frames.
+	 */
+	for (i = 0; i < 16; ++i) {
+		unsigned long pc;
+		if (KBacktraceIterator_end(&kbt))
+			break;
+		pc = kbt.it.pc;
+		if (pc != handle_syscall_pc)
+			perf_callchain_store(entry, pc);
+		KBacktraceIterator_next(&kbt);
+	}
+}
+
+void perf_callchain_user(struct perf_callchain_entry *entry,
+		    struct pt_regs *regs)
+{
+	perf_callchain(entry, regs);
+}
+
+void perf_callchain_kernel(struct perf_callchain_entry *entry,
+		      struct pt_regs *regs)
+{
+	perf_callchain(entry, regs);
+}
diff --git a/kernel/arch/tile/kernel/pmc.c b/kernel/arch/tile/kernel/pmc.c
new file mode 100644
index 000000000..db62cc34b
--- /dev/null
+++ b/kernel/arch/tile/kernel/pmc.c
@@ -0,0 +1,121 @@
+/*
+ * Copyright 2014 Tilera Corporation. All Rights Reserved.
+ *
+ *   This program is free software; you can redistribute it and/or
+ *   modify it under the terms of the GNU General Public License
+ *   as published by the Free Software Foundation, version 2.
+ *
+ *   This program is distributed in the hope that it will be useful, but
+ *   WITHOUT ANY WARRANTY; without even the implied warranty of
+ *   MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
+ *   NON INFRINGEMENT.  See the GNU General Public License for
+ *   more details.
+ */
+
+#include <linux/errno.h>
+#include <linux/spinlock.h>
+#include <linux/module.h>
+#include <linux/atomic.h>
+#include <linux/interrupt.h>
+
+#include <asm/processor.h>
+#include <asm/pmc.h>
+
+perf_irq_t perf_irq = NULL;
+int handle_perf_interrupt(struct pt_regs *regs, int fault)
+{
+	int retval;
+
+	if (!perf_irq)
+		panic("Unexpected PERF_COUNT interrupt %d\n", fault);
+
+	nmi_enter();
+	retval = perf_irq(regs, fault);
+	nmi_exit();
+	return retval;
+}
+
+/* Reserve PMC hardware if it is available. */
+perf_irq_t reserve_pmc_hardware(perf_irq_t new_perf_irq)
+{
+	return cmpxchg(&perf_irq, NULL, new_perf_irq);
+}
+EXPORT_SYMBOL(reserve_pmc_hardware);
+
+/* Release PMC hardware. */
+void release_pmc_hardware(void)
+{
+	perf_irq = NULL;
+}
+EXPORT_SYMBOL(release_pmc_hardware);
+
+
+/*
+ * Get current overflow status of each performance counter,
+ * and auxiliary performance counter.
+ */
+unsigned long
+pmc_get_overflow(void)
+{
+	unsigned long status;
+
+	/*
+	 * merge base+aux into a single vector
+	 */
+	status = __insn_mfspr(SPR_PERF_COUNT_STS);
+	status |= __insn_mfspr(SPR_AUX_PERF_COUNT_STS) << TILE_BASE_COUNTERS;
+	return status;
+}
+
+/*
+ * Clear the status bit for the corresponding counter, if written
+ * with a one.
+ */
+void
+pmc_ack_overflow(unsigned long status)
+{
+	/*
+	 * clear overflow status by writing ones
+	 */
+	__insn_mtspr(SPR_PERF_COUNT_STS, status);
+	__insn_mtspr(SPR_AUX_PERF_COUNT_STS, status >> TILE_BASE_COUNTERS);
+}
+
+/*
+ * The perf count interrupts are masked and unmasked explicitly,
+ * and only here.  The normal irq_enable() does not enable them,
+ * and irq_disable() does not disable them.  That lets these
+ * routines drive the perf count interrupts orthogonally.
+ *
+ * We also mask the perf count interrupts on entry to the perf count
+ * interrupt handler in assembly code, and by default unmask them
+ * again (with interrupt critical section protection) just before
+ * returning from the interrupt.  If the perf count handler returns
+ * a non-zero error code, then we don't re-enable them before returning.
+ *
+ * For Pro, we rely on both interrupts being in the same word to update
+ * them atomically so we never have one enabled and one disabled.
+ */
+
+#if CHIP_HAS_SPLIT_INTR_MASK()
+# if INT_PERF_COUNT < 32 || INT_AUX_PERF_COUNT < 32
+#  error Fix assumptions about which word PERF_COUNT interrupts are in
+# endif
+#endif
+
+static inline unsigned long long pmc_mask(void)
+{
+	unsigned long long mask = 1ULL << INT_PERF_COUNT;
+	mask |= 1ULL << INT_AUX_PERF_COUNT;
+	return mask;
+}
+
+void unmask_pmc_interrupts(void)
+{
+	interrupt_mask_reset_mask(pmc_mask());
+}
+
+void mask_pmc_interrupts(void)
+{
+	interrupt_mask_set_mask(pmc_mask());
+}
diff --git a/kernel/arch/tile/kernel/proc.c b/kernel/arch/tile/kernel/proc.c
new file mode 100644
index 000000000..7983e9868
--- /dev/null
+++ b/kernel/arch/tile/kernel/proc.c
@@ -0,0 +1,160 @@
+/*
+ * Copyright 2010 Tilera Corporation. All Rights Reserved.
+ *
+ *   This program is free software; you can redistribute it and/or
+ *   modify it under the terms of the GNU General Public License
+ *   as published by the Free Software Foundation, version 2.
+ *
+ *   This program is distributed in the hope that it will be useful, but
+ *   WITHOUT ANY WARRANTY; without even the implied warranty of
+ *   MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
+ *   NON INFRINGEMENT.  See the GNU General Public License for
+ *   more details.
+ */
+
+#include <linux/smp.h>
+#include <linux/seq_file.h>
+#include <linux/threads.h>
+#include <linux/cpumask.h>
+#include <linux/timex.h>
+#include <linux/delay.h>
+#include <linux/fs.h>
+#include <linux/proc_fs.h>
+#include <linux/sysctl.h>
+#include <linux/hardirq.h>
+#include <linux/hugetlb.h>
+#include <linux/mman.h>
+#include <asm/unaligned.h>
+#include <asm/pgtable.h>
+#include <asm/processor.h>
+#include <asm/sections.h>
+#include <asm/homecache.h>
+#include <asm/hardwall.h>
+#include <arch/chip.h>
+
+
+/*
+ * Support /proc/cpuinfo
+ */
+
+#define cpu_to_ptr(n) ((void *)((long)(n)+1))
+#define ptr_to_cpu(p) ((long)(p) - 1)
+
+static int show_cpuinfo(struct seq_file *m, void *v)
+{
+	int n = ptr_to_cpu(v);
+
+	if (n == 0) {
+		seq_printf(m, "cpu count\t: %d\n", num_online_cpus());
+		seq_printf(m, "cpu list\t: %*pbl\n",
+			   cpumask_pr_args(cpu_online_mask));
+		seq_printf(m, "model name\t: %s\n", chip_model);
+		seq_printf(m, "flags\t\t:\n");  /* nothing for now */
+		seq_printf(m, "cpu MHz\t\t: %llu.%06llu\n",
+			   get_clock_rate() / 1000000,
+			   (get_clock_rate() % 1000000));
+		seq_printf(m, "bogomips\t: %lu.%02lu\n\n",
+			   loops_per_jiffy/(500000/HZ),
+			   (loops_per_jiffy/(5000/HZ)) % 100);
+	}
+
+#ifdef CONFIG_SMP
+	if (!cpu_online(n))
+		return 0;
+#endif
+
+	seq_printf(m, "processor\t: %d\n", n);
+
+	/* Print only num_online_cpus() blank lines total. */
+	if (cpumask_next(n, cpu_online_mask) < nr_cpu_ids)
+		seq_printf(m, "\n");
+
+	return 0;
+}
+
+static void *c_start(struct seq_file *m, loff_t *pos)
+{
+	return *pos < nr_cpu_ids ? cpu_to_ptr(*pos) : NULL;
+}
+static void *c_next(struct seq_file *m, void *v, loff_t *pos)
+{
+	++*pos;
+	return c_start(m, pos);
+}
+static void c_stop(struct seq_file *m, void *v)
+{
+}
+const struct seq_operations cpuinfo_op = {
+	.start	= c_start,
+	.next	= c_next,
+	.stop	= c_stop,
+	.show	= show_cpuinfo,
+};
+
+/*
+ * Support /proc/tile directory
+ */
+
+static int __init proc_tile_init(void)
+{
+	struct proc_dir_entry *root = proc_mkdir("tile", NULL);
+	if (root == NULL)
+		return 0;
+
+	proc_tile_hardwall_init(root);
+
+	return 0;
+}
+
+arch_initcall(proc_tile_init);
+
+/*
+ * Support /proc/sys/tile directory
+ */
+
+static struct ctl_table unaligned_subtable[] = {
+	{
+		.procname	= "enabled",
+		.data		= &unaligned_fixup,
+		.maxlen		= sizeof(int),
+		.mode		= 0644,
+		.proc_handler	= &proc_dointvec
+	},
+	{
+		.procname	= "printk",
+		.data		= &unaligned_printk,
+		.maxlen		= sizeof(int),
+		.mode		= 0644,
+		.proc_handler	= &proc_dointvec
+	},
+	{
+		.procname	= "count",
+		.data		= &unaligned_fixup_count,
+		.maxlen		= sizeof(int),
+		.mode		= 0644,
+		.proc_handler	= &proc_dointvec
+	},
+	{}
+};
+
+static struct ctl_table unaligned_table[] = {
+	{
+		.procname	= "unaligned_fixup",
+		.mode		= 0555,
+		.child		= unaligned_subtable
+	},
+	{}
+};
+
+static struct ctl_path tile_path[] = {
+	{ .procname = "tile" },
+	{ }
+};
+
+static int __init proc_sys_tile_init(void)
+{
+	register_sysctl_paths(tile_path, unaligned_table);
+	return 0;
+}
+
+arch_initcall(proc_sys_tile_init);
diff --git a/kernel/arch/tile/kernel/process.c b/kernel/arch/tile/kernel/process.c
new file mode 100644
index 000000000..b403c2e3e
--- /dev/null
+++ b/kernel/arch/tile/kernel/process.c
@@ -0,0 +1,576 @@
+/*
+ * Copyright 2010 Tilera Corporation. All Rights Reserved.
+ *
+ *   This program is free software; you can redistribute it and/or
+ *   modify it under the terms of the GNU General Public License
+ *   as published by the Free Software Foundation, version 2.
+ *
+ *   This program is distributed in the hope that it will be useful, but
+ *   WITHOUT ANY WARRANTY; without even the implied warranty of
+ *   MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
+ *   NON INFRINGEMENT.  See the GNU General Public License for
+ *   more details.
+ */
+
+#include <linux/sched.h>
+#include <linux/preempt.h>
+#include <linux/module.h>
+#include <linux/fs.h>
+#include <linux/kprobes.h>
+#include <linux/elfcore.h>
+#include <linux/tick.h>
+#include <linux/init.h>
+#include <linux/mm.h>
+#include <linux/compat.h>
+#include <linux/hardirq.h>
+#include <linux/syscalls.h>
+#include <linux/kernel.h>
+#include <linux/tracehook.h>
+#include <linux/signal.h>
+#include <linux/context_tracking.h>
+#include <asm/stack.h>
+#include <asm/switch_to.h>
+#include <asm/homecache.h>
+#include <asm/syscalls.h>
+#include <asm/traps.h>
+#include <asm/setup.h>
+#include <asm/uaccess.h>
+#ifdef CONFIG_HARDWALL
+#include <asm/hardwall.h>
+#endif
+#include <arch/chip.h>
+#include <arch/abi.h>
+#include <arch/sim_def.h>
+
+/*
+ * Use the (x86) "idle=poll" option to prefer low latency when leaving the
+ * idle loop over low power while in the idle loop, e.g. if we have
+ * one thread per core and we want to get threads out of futex waits fast.
+ */
+static int __init idle_setup(char *str)
+{
+	if (!str)
+		return -EINVAL;
+
+	if (!strcmp(str, "poll")) {
+		pr_info("using polling idle threads\n");
+		cpu_idle_poll_ctrl(true);
+		return 0;
+	} else if (!strcmp(str, "halt")) {
+		return 0;
+	}
+	return -1;
+}
+early_param("idle", idle_setup);
+
+void arch_cpu_idle(void)
+{
+	__this_cpu_write(irq_stat.idle_timestamp, jiffies);
+	_cpu_idle();
+}
+
+/*
+ * Release a thread_info structure
+ */
+void arch_release_thread_info(struct thread_info *info)
+{
+	struct single_step_state *step_state = info->step_state;
+
+	if (step_state) {
+
+		/*
+		 * FIXME: we don't munmap step_state->buffer
+		 * because the mm_struct for this process (info->task->mm)
+		 * has already been zeroed in exit_mm().  Keeping a
+		 * reference to it here seems like a bad move, so this
+		 * means we can't munmap() the buffer, and therefore if we
+		 * ptrace multiple threads in a process, we will slowly
+		 * leak user memory.  (Note that as soon as the last
+		 * thread in a process dies, we will reclaim all user
+		 * memory including single-step buffers in the usual way.)
+		 * We should either assign a kernel VA to this buffer
+		 * somehow, or we should associate the buffer(s) with the
+		 * mm itself so we can clean them up that way.
+		 */
+		kfree(step_state);
+	}
+}
+
+static void save_arch_state(struct thread_struct *t);
+
+int copy_thread(unsigned long clone_flags, unsigned long sp,
+		unsigned long arg, struct task_struct *p)
+{
+	struct pt_regs *childregs = task_pt_regs(p);
+	unsigned long ksp;
+	unsigned long *callee_regs;
+
+	/*
+	 * Set up the stack and stack pointer appropriately for the
+	 * new child to find itself woken up in __switch_to().
+	 * The callee-saved registers must be on the stack to be read;
+	 * the new task will then jump to assembly support to handle
+	 * calling schedule_tail(), etc., and (for userspace tasks)
+	 * returning to the context set up in the pt_regs.
+	 */
+	ksp = (unsigned long) childregs;
+	ksp -= C_ABI_SAVE_AREA_SIZE;   /* interrupt-entry save area */
+	((long *)ksp)[0] = ((long *)ksp)[1] = 0;
+	ksp -= CALLEE_SAVED_REGS_COUNT * sizeof(unsigned long);
+	callee_regs = (unsigned long *)ksp;
+	ksp -= C_ABI_SAVE_AREA_SIZE;   /* __switch_to() save area */
+	((long *)ksp)[0] = ((long *)ksp)[1] = 0;
+	p->thread.ksp = ksp;
+
+	/* Record the pid of the task that created this one. */
+	p->thread.creator_pid = current->pid;
+
+	if (unlikely(p->flags & PF_KTHREAD)) {
+		/* kernel thread */
+		memset(childregs, 0, sizeof(struct pt_regs));
+		memset(&callee_regs[2], 0,
+		       (CALLEE_SAVED_REGS_COUNT - 2) * sizeof(unsigned long));
+		callee_regs[0] = sp;   /* r30 = function */
+		callee_regs[1] = arg;  /* r31 = arg */
+		childregs->ex1 = PL_ICS_EX1(KERNEL_PL, 0);
+		p->thread.pc = (unsigned long) ret_from_kernel_thread;
+		return 0;
+	}
+
+	/*
+	 * Start new thread in ret_from_fork so it schedules properly
+	 * and then return from interrupt like the parent.
+	 */
+	p->thread.pc = (unsigned long) ret_from_fork;
+
+	/*
+	 * Do not clone step state from the parent; each thread
+	 * must make its own lazily.
+	 */
+	task_thread_info(p)->step_state = NULL;
+
+#ifdef __tilegx__
+	/*
+	 * Do not clone unalign jit fixup from the parent; each thread
+	 * must allocate its own on demand.
+	 */
+	task_thread_info(p)->unalign_jit_base = NULL;
+#endif
+
+	/*
+	 * Copy the registers onto the kernel stack so the
+	 * return-from-interrupt code will reload it into registers.
+	 */
+	*childregs = *current_pt_regs();
+	childregs->regs[0] = 0;         /* return value is zero */
+	if (sp)
+		childregs->sp = sp;  /* override with new user stack pointer */
+	memcpy(callee_regs, &childregs->regs[CALLEE_SAVED_FIRST_REG],
+	       CALLEE_SAVED_REGS_COUNT * sizeof(unsigned long));
+
+	/* Save user stack top pointer so we can ID the stack vm area later. */
+	p->thread.usp0 = childregs->sp;
+
+	/*
+	 * If CLONE_SETTLS is set, set "tp" in the new task to "r4",
+	 * which is passed in as arg #5 to sys_clone().
+	 */
+	if (clone_flags & CLONE_SETTLS)
+		childregs->tp = childregs->regs[4];
+
+
+#if CHIP_HAS_TILE_DMA()
+	/*
+	 * No DMA in the new thread.  We model this on the fact that
+	 * fork() clears the pending signals, alarms, and aio for the child.
+	 */
+	memset(&p->thread.tile_dma_state, 0, sizeof(struct tile_dma_state));
+	memset(&p->thread.dma_async_tlb, 0, sizeof(struct async_tlb));
+#endif
+
+	/* New thread has its miscellaneous processor state bits clear. */
+	p->thread.proc_status = 0;
+
+#ifdef CONFIG_HARDWALL
+	/* New thread does not own any networks. */
+	memset(&p->thread.hardwall[0], 0,
+	       sizeof(struct hardwall_task) * HARDWALL_TYPES);
+#endif
+
+
+	/*
+	 * Start the new thread with the current architecture state
+	 * (user interrupt masks, etc.).
+	 */
+	save_arch_state(&p->thread);
+
+	return 0;
+}
+
+int set_unalign_ctl(struct task_struct *tsk, unsigned int val)
+{
+	task_thread_info(tsk)->align_ctl = val;
+	return 0;
+}
+
+int get_unalign_ctl(struct task_struct *tsk, unsigned long adr)
+{
+	return put_user(task_thread_info(tsk)->align_ctl,
+			(unsigned int __user *)adr);
+}
+
+static struct task_struct corrupt_current = { .comm = "<corrupt>" };
+
+/*
+ * Return "current" if it looks plausible, or else a pointer to a dummy.
+ * This can be helpful if we are just trying to emit a clean panic.
+ */
+struct task_struct *validate_current(void)
+{
+	struct task_struct *tsk = current;
+	if (unlikely((unsigned long)tsk < PAGE_OFFSET ||
+		     (high_memory && (void *)tsk > high_memory) ||
+		     ((unsigned long)tsk & (__alignof__(*tsk) - 1)) != 0)) {
+		pr_err("Corrupt 'current' %p (sp %#lx)\n", tsk, stack_pointer);
+		tsk = &corrupt_current;
+	}
+	return tsk;
+}
+
+/* Take and return the pointer to the previous task, for schedule_tail(). */
+struct task_struct *sim_notify_fork(struct task_struct *prev)
+{
+	struct task_struct *tsk = current;
+	__insn_mtspr(SPR_SIM_CONTROL, SIM_CONTROL_OS_FORK_PARENT |
+		     (tsk->thread.creator_pid << _SIM_CONTROL_OPERATOR_BITS));
+	__insn_mtspr(SPR_SIM_CONTROL, SIM_CONTROL_OS_FORK |
+		     (tsk->pid << _SIM_CONTROL_OPERATOR_BITS));
+	return prev;
+}
+
+int dump_task_regs(struct task_struct *tsk, elf_gregset_t *regs)
+{
+	struct pt_regs *ptregs = task_pt_regs(tsk);
+	elf_core_copy_regs(regs, ptregs);
+	return 1;
+}
+
+#if CHIP_HAS_TILE_DMA()
+
+/* Allow user processes to access the DMA SPRs */
+void grant_dma_mpls(void)
+{
+#if CONFIG_KERNEL_PL == 2
+	__insn_mtspr(SPR_MPL_DMA_CPL_SET_1, 1);
+	__insn_mtspr(SPR_MPL_DMA_NOTIFY_SET_1, 1);
+#else
+	__insn_mtspr(SPR_MPL_DMA_CPL_SET_0, 1);
+	__insn_mtspr(SPR_MPL_DMA_NOTIFY_SET_0, 1);
+#endif
+}
+
+/* Forbid user processes from accessing the DMA SPRs */
+void restrict_dma_mpls(void)
+{
+#if CONFIG_KERNEL_PL == 2
+	__insn_mtspr(SPR_MPL_DMA_CPL_SET_2, 1);
+	__insn_mtspr(SPR_MPL_DMA_NOTIFY_SET_2, 1);
+#else
+	__insn_mtspr(SPR_MPL_DMA_CPL_SET_1, 1);
+	__insn_mtspr(SPR_MPL_DMA_NOTIFY_SET_1, 1);
+#endif
+}
+
+/* Pause the DMA engine, then save off its state registers. */
+static void save_tile_dma_state(struct tile_dma_state *dma)
+{
+	unsigned long state = __insn_mfspr(SPR_DMA_USER_STATUS);
+	unsigned long post_suspend_state;
+
+	/* If we're running, suspend the engine. */
+	if ((state & DMA_STATUS_MASK) == SPR_DMA_STATUS__RUNNING_MASK)
+		__insn_mtspr(SPR_DMA_CTR, SPR_DMA_CTR__SUSPEND_MASK);
+
+	/*
+	 * Wait for the engine to idle, then save regs.  Note that we
+	 * want to record the "running" bit from before suspension,
+	 * and the "done" bit from after, so that we can properly
+	 * distinguish a case where the user suspended the engine from
+	 * the case where the kernel suspended as part of the context
+	 * swap.
+	 */
+	do {
+		post_suspend_state = __insn_mfspr(SPR_DMA_USER_STATUS);
+	} while (post_suspend_state & SPR_DMA_STATUS__BUSY_MASK);
+
+	dma->src = __insn_mfspr(SPR_DMA_SRC_ADDR);
+	dma->src_chunk = __insn_mfspr(SPR_DMA_SRC_CHUNK_ADDR);
+	dma->dest = __insn_mfspr(SPR_DMA_DST_ADDR);
+	dma->dest_chunk = __insn_mfspr(SPR_DMA_DST_CHUNK_ADDR);
+	dma->strides = __insn_mfspr(SPR_DMA_STRIDE);
+	dma->chunk_size = __insn_mfspr(SPR_DMA_CHUNK_SIZE);
+	dma->byte = __insn_mfspr(SPR_DMA_BYTE);
+	dma->status = (state & SPR_DMA_STATUS__RUNNING_MASK) |
+		(post_suspend_state & SPR_DMA_STATUS__DONE_MASK);
+}
+
+/* Restart a DMA that was running before we were context-switched out. */
+static void restore_tile_dma_state(struct thread_struct *t)
+{
+	const struct tile_dma_state *dma = &t->tile_dma_state;
+
+	/*
+	 * The only way to restore the done bit is to run a zero
+	 * length transaction.
+	 */
+	if ((dma->status & SPR_DMA_STATUS__DONE_MASK) &&
+	    !(__insn_mfspr(SPR_DMA_USER_STATUS) & SPR_DMA_STATUS__DONE_MASK)) {
+		__insn_mtspr(SPR_DMA_BYTE, 0);
+		__insn_mtspr(SPR_DMA_CTR, SPR_DMA_CTR__REQUEST_MASK);
+		while (__insn_mfspr(SPR_DMA_USER_STATUS) &
+		       SPR_DMA_STATUS__BUSY_MASK)
+			;
+	}
+
+	__insn_mtspr(SPR_DMA_SRC_ADDR, dma->src);
+	__insn_mtspr(SPR_DMA_SRC_CHUNK_ADDR, dma->src_chunk);
+	__insn_mtspr(SPR_DMA_DST_ADDR, dma->dest);
+	__insn_mtspr(SPR_DMA_DST_CHUNK_ADDR, dma->dest_chunk);
+	__insn_mtspr(SPR_DMA_STRIDE, dma->strides);
+	__insn_mtspr(SPR_DMA_CHUNK_SIZE, dma->chunk_size);
+	__insn_mtspr(SPR_DMA_BYTE, dma->byte);
+
+	/*
+	 * Restart the engine if we were running and not done.
+	 * Clear a pending async DMA fault that we were waiting on return
+	 * to user space to execute, since we expect the DMA engine
+	 * to regenerate those faults for us now.  Note that we don't
+	 * try to clear the TIF_ASYNC_TLB flag, since it's relatively
+	 * harmless if set, and it covers both DMA and the SN processor.
+	 */
+	if ((dma->status & DMA_STATUS_MASK) == SPR_DMA_STATUS__RUNNING_MASK) {
+		t->dma_async_tlb.fault_num = 0;
+		__insn_mtspr(SPR_DMA_CTR, SPR_DMA_CTR__REQUEST_MASK);
+	}
+}
+
+#endif
+
+static void save_arch_state(struct thread_struct *t)
+{
+#if CHIP_HAS_SPLIT_INTR_MASK()
+	t->interrupt_mask = __insn_mfspr(SPR_INTERRUPT_MASK_0_0) |
+		((u64)__insn_mfspr(SPR_INTERRUPT_MASK_0_1) << 32);
+#else
+	t->interrupt_mask = __insn_mfspr(SPR_INTERRUPT_MASK_0);
+#endif
+	t->ex_context[0] = __insn_mfspr(SPR_EX_CONTEXT_0_0);
+	t->ex_context[1] = __insn_mfspr(SPR_EX_CONTEXT_0_1);
+	t->system_save[0] = __insn_mfspr(SPR_SYSTEM_SAVE_0_0);
+	t->system_save[1] = __insn_mfspr(SPR_SYSTEM_SAVE_0_1);
+	t->system_save[2] = __insn_mfspr(SPR_SYSTEM_SAVE_0_2);
+	t->system_save[3] = __insn_mfspr(SPR_SYSTEM_SAVE_0_3);
+	t->intctrl_0 = __insn_mfspr(SPR_INTCTRL_0_STATUS);
+	t->proc_status = __insn_mfspr(SPR_PROC_STATUS);
+#if !CHIP_HAS_FIXED_INTVEC_BASE()
+	t->interrupt_vector_base = __insn_mfspr(SPR_INTERRUPT_VECTOR_BASE_0);
+#endif
+	t->tile_rtf_hwm = __insn_mfspr(SPR_TILE_RTF_HWM);
+#if CHIP_HAS_DSTREAM_PF()
+	t->dstream_pf = __insn_mfspr(SPR_DSTREAM_PF);
+#endif
+}
+
+static void restore_arch_state(const struct thread_struct *t)
+{
+#if CHIP_HAS_SPLIT_INTR_MASK()
+	__insn_mtspr(SPR_INTERRUPT_MASK_0_0, (u32) t->interrupt_mask);
+	__insn_mtspr(SPR_INTERRUPT_MASK_0_1, t->interrupt_mask >> 32);
+#else
+	__insn_mtspr(SPR_INTERRUPT_MASK_0, t->interrupt_mask);
+#endif
+	__insn_mtspr(SPR_EX_CONTEXT_0_0, t->ex_context[0]);
+	__insn_mtspr(SPR_EX_CONTEXT_0_1, t->ex_context[1]);
+	__insn_mtspr(SPR_SYSTEM_SAVE_0_0, t->system_save[0]);
+	__insn_mtspr(SPR_SYSTEM_SAVE_0_1, t->system_save[1]);
+	__insn_mtspr(SPR_SYSTEM_SAVE_0_2, t->system_save[2]);
+	__insn_mtspr(SPR_SYSTEM_SAVE_0_3, t->system_save[3]);
+	__insn_mtspr(SPR_INTCTRL_0_STATUS, t->intctrl_0);
+	__insn_mtspr(SPR_PROC_STATUS, t->proc_status);
+#if !CHIP_HAS_FIXED_INTVEC_BASE()
+	__insn_mtspr(SPR_INTERRUPT_VECTOR_BASE_0, t->interrupt_vector_base);
+#endif
+	__insn_mtspr(SPR_TILE_RTF_HWM, t->tile_rtf_hwm);
+#if CHIP_HAS_DSTREAM_PF()
+	__insn_mtspr(SPR_DSTREAM_PF, t->dstream_pf);
+#endif
+}
+
+
+void _prepare_arch_switch(struct task_struct *next)
+{
+#if CHIP_HAS_TILE_DMA()
+	struct tile_dma_state *dma = &current->thread.tile_dma_state;
+	if (dma->enabled)
+		save_tile_dma_state(dma);
+#endif
+}
+
+
+struct task_struct *__sched _switch_to(struct task_struct *prev,
+				       struct task_struct *next)
+{
+	/* DMA state is already saved; save off other arch state. */
+	save_arch_state(&prev->thread);
+
+#if CHIP_HAS_TILE_DMA()
+	/*
+	 * Restore DMA in new task if desired.
+	 * Note that it is only safe to restart here since interrupts
+	 * are disabled, so we can't take any DMATLB miss or access
+	 * interrupts before we have finished switching stacks.
+	 */
+	if (next->thread.tile_dma_state.enabled) {
+		restore_tile_dma_state(&next->thread);
+		grant_dma_mpls();
+	} else {
+		restrict_dma_mpls();
+	}
+#endif
+
+	/* Restore other arch state. */
+	restore_arch_state(&next->thread);
+
+#ifdef CONFIG_HARDWALL
+	/* Enable or disable access to the network registers appropriately. */
+	hardwall_switch_tasks(prev, next);
+#endif
+
+	/*
+	 * Switch kernel SP, PC, and callee-saved registers.
+	 * In the context of the new task, return the old task pointer
+	 * (i.e. the task that actually called __switch_to).
+	 * Pass the value to use for SYSTEM_SAVE_K_0 when we reset our sp.
+	 */
+	return __switch_to(prev, next, next_current_ksp0(next));
+}
+
+/*
+ * This routine is called on return from interrupt if any of the
+ * TIF_WORK_MASK flags are set in thread_info->flags.  It is
+ * entered with interrupts disabled so we don't miss an event
+ * that modified the thread_info flags.  If any flag is set, we
+ * handle it and return, and the calling assembly code will
+ * re-disable interrupts, reload the thread flags, and call back
+ * if more flags need to be handled.
+ *
+ * We return whether we need to check the thread_info flags again
+ * or not.  Note that we don't clear TIF_SINGLESTEP here, so it's
+ * important that it be tested last, and then claim that we don't
+ * need to recheck the flags.
+ */
+int do_work_pending(struct pt_regs *regs, u32 thread_info_flags)
+{
+	/* If we enter in kernel mode, do nothing and exit the caller loop. */
+	if (!user_mode(regs))
+		return 0;
+
+	user_exit();
+
+	/* Enable interrupts; they are disabled again on return to caller. */
+	local_irq_enable();
+
+	if (thread_info_flags & _TIF_NEED_RESCHED) {
+		schedule();
+		return 1;
+	}
+#if CHIP_HAS_TILE_DMA()
+	if (thread_info_flags & _TIF_ASYNC_TLB) {
+		do_async_page_fault(regs);
+		return 1;
+	}
+#endif
+	if (thread_info_flags & _TIF_SIGPENDING) {
+		do_signal(regs);
+		return 1;
+	}
+	if (thread_info_flags & _TIF_NOTIFY_RESUME) {
+		clear_thread_flag(TIF_NOTIFY_RESUME);
+		tracehook_notify_resume(regs);
+		return 1;
+	}
+	if (thread_info_flags & _TIF_SINGLESTEP)
+		single_step_once(regs);
+
+	user_enter();
+
+	return 0;
+}
+
+unsigned long get_wchan(struct task_struct *p)
+{
+	struct KBacktraceIterator kbt;
+
+	if (!p || p == current || p->state == TASK_RUNNING)
+		return 0;
+
+	for (KBacktraceIterator_init(&kbt, p, NULL);
+	     !KBacktraceIterator_end(&kbt);
+	     KBacktraceIterator_next(&kbt)) {
+		if (!in_sched_functions(kbt.it.pc))
+			return kbt.it.pc;
+	}
+
+	return 0;
+}
+
+/* Flush thread state. */
+void flush_thread(void)
+{
+	/* Nothing */
+}
+
+/*
+ * Free current thread data structures etc..
+ */
+void exit_thread(void)
+{
+#ifdef CONFIG_HARDWALL
+	/*
+	 * Remove the task from the list of tasks that are associated
+	 * with any live hardwalls.  (If the task that is exiting held
+	 * the last reference to a hardwall fd, it would already have
+	 * been released and deactivated at this point.)
+	 */
+	hardwall_deactivate_all(current);
+#endif
+}
+
+void show_regs(struct pt_regs *regs)
+{
+	struct task_struct *tsk = validate_current();
+	int i;
+
+	if (tsk != &corrupt_current)
+		show_regs_print_info(KERN_ERR);
+#ifdef __tilegx__
+	for (i = 0; i < 17; i++)
+		pr_err(" r%-2d: " REGFMT " r%-2d: " REGFMT " r%-2d: " REGFMT "\n",
+		       i, regs->regs[i], i+18, regs->regs[i+18],
+		       i+36, regs->regs[i+36]);
+	pr_err(" r17: " REGFMT " r35: " REGFMT " tp : " REGFMT "\n",
+	       regs->regs[17], regs->regs[35], regs->tp);
+	pr_err(" sp : " REGFMT " lr : " REGFMT "\n", regs->sp, regs->lr);
+#else
+	for (i = 0; i < 13; i++)
+		pr_err(" r%-2d: " REGFMT " r%-2d: " REGFMT " r%-2d: " REGFMT " r%-2d: " REGFMT "\n",
+		       i, regs->regs[i], i+14, regs->regs[i+14],
+		       i+27, regs->regs[i+27], i+40, regs->regs[i+40]);
+	pr_err(" r13: " REGFMT " tp : " REGFMT " sp : " REGFMT " lr : " REGFMT "\n",
+	       regs->regs[13], regs->tp, regs->sp, regs->lr);
+#endif
+	pr_err(" pc : " REGFMT " ex1: %ld     faultnum: %ld\n",
+	       regs->pc, regs->ex1, regs->faultnum);
+
+	dump_stack_regs(regs);
+}
diff --git a/kernel/arch/tile/kernel/ptrace.c b/kernel/arch/tile/kernel/ptrace.c
new file mode 100644
index 000000000..f84eed824
--- /dev/null
+++ b/kernel/arch/tile/kernel/ptrace.c
@@ -0,0 +1,325 @@
+/*
+ * Copyright 2010 Tilera Corporation. All Rights Reserved.
+ *
+ *   This program is free software; you can redistribute it and/or
+ *   modify it under the terms of the GNU General Public License
+ *   as published by the Free Software Foundation, version 2.
+ *
+ *   This program is distributed in the hope that it will be useful, but
+ *   WITHOUT ANY WARRANTY; without even the implied warranty of
+ *   MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
+ *   NON INFRINGEMENT.  See the GNU General Public License for
+ *   more details.
+ *
+ * Copied from i386: Ross Biro 1/23/92
+ */
+
+#include <linux/kernel.h>
+#include <linux/ptrace.h>
+#include <linux/kprobes.h>
+#include <linux/compat.h>
+#include <linux/uaccess.h>
+#include <linux/regset.h>
+#include <linux/elf.h>
+#include <linux/tracehook.h>
+#include <linux/context_tracking.h>
+#include <asm/traps.h>
+#include <arch/chip.h>
+
+#define CREATE_TRACE_POINTS
+#include <trace/events/syscalls.h>
+
+void user_enable_single_step(struct task_struct *child)
+{
+	set_tsk_thread_flag(child, TIF_SINGLESTEP);
+}
+
+void user_disable_single_step(struct task_struct *child)
+{
+	clear_tsk_thread_flag(child, TIF_SINGLESTEP);
+}
+
+/*
+ * Called by kernel/ptrace.c when detaching..
+ */
+void ptrace_disable(struct task_struct *child)
+{
+	clear_tsk_thread_flag(child, TIF_SINGLESTEP);
+
+	/*
+	 * These two are currently unused, but will be set by arch_ptrace()
+	 * and used in the syscall assembly when we do support them.
+	 */
+	clear_tsk_thread_flag(child, TIF_SYSCALL_TRACE);
+}
+
+/*
+ * Get registers from task and ready the result for userspace.
+ * Note that we localize the API issues to getregs() and putregs() at
+ * some cost in performance, e.g. we need a full pt_regs copy for
+ * PEEKUSR, and two copies for POKEUSR.  But in general we expect
+ * GETREGS/PUTREGS to be the API of choice anyway.
+ */
+static char *getregs(struct task_struct *child, struct pt_regs *uregs)
+{
+	*uregs = *task_pt_regs(child);
+
+	/* Set up flags ABI bits. */
+	uregs->flags = 0;
+#ifdef CONFIG_COMPAT
+	if (task_thread_info(child)->status & TS_COMPAT)
+		uregs->flags |= PT_FLAGS_COMPAT;
+#endif
+
+	return (char *)uregs;
+}
+
+/* Put registers back to task. */
+static void putregs(struct task_struct *child, struct pt_regs *uregs)
+{
+	struct pt_regs *regs = task_pt_regs(child);
+
+	/* Don't allow overwriting the kernel-internal flags word. */
+	uregs->flags = regs->flags;
+
+	/* Only allow setting the ICS bit in the ex1 word. */
+	uregs->ex1 = PL_ICS_EX1(USER_PL, EX1_ICS(uregs->ex1));
+
+	*regs = *uregs;
+}
+
+enum tile_regset {
+	REGSET_GPR,
+};
+
+static int tile_gpr_get(struct task_struct *target,
+			  const struct user_regset *regset,
+			  unsigned int pos, unsigned int count,
+			  void *kbuf, void __user *ubuf)
+{
+	struct pt_regs regs;
+
+	getregs(target, &regs);
+
+	return user_regset_copyout(&pos, &count, &kbuf, &ubuf, &regs, 0,
+				   sizeof(regs));
+}
+
+static int tile_gpr_set(struct task_struct *target,
+			  const struct user_regset *regset,
+			  unsigned int pos, unsigned int count,
+			  const void *kbuf, const void __user *ubuf)
+{
+	int ret;
+	struct pt_regs regs;
+
+	ret = user_regset_copyin(&pos, &count, &kbuf, &ubuf, &regs, 0,
+				 sizeof(regs));
+	if (ret)
+		return ret;
+
+	putregs(target, &regs);
+
+	return 0;
+}
+
+static const struct user_regset tile_user_regset[] = {
+	[REGSET_GPR] = {
+		.core_note_type = NT_PRSTATUS,
+		.n = ELF_NGREG,
+		.size = sizeof(elf_greg_t),
+		.align = sizeof(elf_greg_t),
+		.get = tile_gpr_get,
+		.set = tile_gpr_set,
+	},
+};
+
+static const struct user_regset_view tile_user_regset_view = {
+	.name = CHIP_ARCH_NAME,
+	.e_machine = ELF_ARCH,
+	.ei_osabi = ELF_OSABI,
+	.regsets = tile_user_regset,
+	.n = ARRAY_SIZE(tile_user_regset),
+};
+
+const struct user_regset_view *task_user_regset_view(struct task_struct *task)
+{
+	return &tile_user_regset_view;
+}
+
+long arch_ptrace(struct task_struct *child, long request,
+		 unsigned long addr, unsigned long data)
+{
+	unsigned long __user *datap = (long __user __force *)data;
+	unsigned long tmp;
+	long ret = -EIO;
+	char *childreg;
+	struct pt_regs copyregs;
+
+	switch (request) {
+
+	case PTRACE_PEEKUSR:  /* Read register from pt_regs. */
+		if (addr >= PTREGS_SIZE)
+			break;
+		childreg = getregs(child, &copyregs) + addr;
+#ifdef CONFIG_COMPAT
+		if (is_compat_task()) {
+			if (addr & (sizeof(compat_long_t)-1))
+				break;
+			ret = put_user(*(compat_long_t *)childreg,
+				       (compat_long_t __user *)datap);
+		} else
+#endif
+		{
+			if (addr & (sizeof(long)-1))
+				break;
+			ret = put_user(*(long *)childreg, datap);
+		}
+		break;
+
+	case PTRACE_POKEUSR:  /* Write register in pt_regs. */
+		if (addr >= PTREGS_SIZE)
+			break;
+		childreg = getregs(child, &copyregs) + addr;
+#ifdef CONFIG_COMPAT
+		if (is_compat_task()) {
+			if (addr & (sizeof(compat_long_t)-1))
+				break;
+			*(compat_long_t *)childreg = data;
+		} else
+#endif
+		{
+			if (addr & (sizeof(long)-1))
+				break;
+			*(long *)childreg = data;
+		}
+		putregs(child, &copyregs);
+		ret = 0;
+		break;
+
+	case PTRACE_GETREGS:  /* Get all registers from the child. */
+		ret = copy_regset_to_user(child, &tile_user_regset_view,
+					  REGSET_GPR, 0,
+					  sizeof(struct pt_regs), datap);
+		break;
+
+	case PTRACE_SETREGS:  /* Set all registers in the child. */
+		ret = copy_regset_from_user(child, &tile_user_regset_view,
+					    REGSET_GPR, 0,
+					    sizeof(struct pt_regs), datap);
+		break;
+
+	case PTRACE_GETFPREGS:  /* Get the child FPU state. */
+	case PTRACE_SETFPREGS:  /* Set the child FPU state. */
+		break;
+
+	case PTRACE_SETOPTIONS:
+		/* Support TILE-specific ptrace options. */
+		BUILD_BUG_ON(PTRACE_O_MASK_TILE & PTRACE_O_MASK);
+		tmp = data & PTRACE_O_MASK_TILE;
+		data &= ~PTRACE_O_MASK_TILE;
+		ret = ptrace_request(child, request, addr, data);
+		if (ret == 0) {
+			unsigned int flags = child->ptrace;
+			flags &= ~(PTRACE_O_MASK_TILE << PT_OPT_FLAG_SHIFT);
+			flags |= (tmp << PT_OPT_FLAG_SHIFT);
+			child->ptrace = flags;
+		}
+		break;
+
+	default:
+#ifdef CONFIG_COMPAT
+		if (task_thread_info(current)->status & TS_COMPAT) {
+			ret = compat_ptrace_request(child, request,
+						    addr, data);
+			break;
+		}
+#endif
+		ret = ptrace_request(child, request, addr, data);
+		break;
+	}
+
+	return ret;
+}
+
+#ifdef CONFIG_COMPAT
+/* Not used; we handle compat issues in arch_ptrace() directly. */
+long compat_arch_ptrace(struct task_struct *child, compat_long_t request,
+			       compat_ulong_t addr, compat_ulong_t data)
+{
+	BUG();
+}
+#endif
+
+int do_syscall_trace_enter(struct pt_regs *regs)
+{
+	u32 work = ACCESS_ONCE(current_thread_info()->flags);
+
+	/*
+	 * If TIF_NOHZ is set, we are required to call user_exit() before
+	 * doing anything that could touch RCU.
+	 */
+	if (work & _TIF_NOHZ)
+		user_exit();
+
+	if (work & _TIF_SYSCALL_TRACE) {
+		if (tracehook_report_syscall_entry(regs))
+			regs->regs[TREG_SYSCALL_NR] = -1;
+	}
+
+	if (work & _TIF_SYSCALL_TRACEPOINT)
+		trace_sys_enter(regs, regs->regs[TREG_SYSCALL_NR]);
+
+	return regs->regs[TREG_SYSCALL_NR];
+}
+
+void do_syscall_trace_exit(struct pt_regs *regs)
+{
+	long errno;
+
+	/*
+	 * We may come here right after calling schedule_user()
+	 * in which case we can be in RCU user mode.
+	 */
+	user_exit();
+
+	/*
+	 * The standard tile calling convention returns the value (or negative
+	 * errno) in r0, and zero (or positive errno) in r1.
+	 * It saves a couple of cycles on the hot path to do this work in
+	 * registers only as we return, rather than updating the in-memory
+	 * struct ptregs.
+	 */
+	errno = (long) regs->regs[0];
+	if (errno < 0 && errno > -4096)
+		regs->regs[1] = -errno;
+	else
+		regs->regs[1] = 0;
+
+	if (test_thread_flag(TIF_SYSCALL_TRACE))
+		tracehook_report_syscall_exit(regs, 0);
+
+	if (test_thread_flag(TIF_SYSCALL_TRACEPOINT))
+		trace_sys_exit(regs, regs->regs[0]);
+}
+
+void send_sigtrap(struct task_struct *tsk, struct pt_regs *regs)
+{
+	struct siginfo info;
+
+	memset(&info, 0, sizeof(info));
+	info.si_signo = SIGTRAP;
+	info.si_code  = TRAP_BRKPT;
+	info.si_addr  = (void __user *) regs->pc;
+
+	/* Send us the fakey SIGTRAP */
+	force_sig_info(SIGTRAP, &info, tsk);
+}
+
+/* Handle synthetic interrupt delivered only by the simulator. */
+void __kprobes do_breakpoint(struct pt_regs* regs, int fault_num)
+{
+	enum ctx_state prev_state = exception_enter();
+	send_sigtrap(current, regs);
+	exception_exit(prev_state);
+}
diff --git a/kernel/arch/tile/kernel/reboot.c b/kernel/arch/tile/kernel/reboot.c
new file mode 100644
index 000000000..6c5d2c070
--- /dev/null
+++ b/kernel/arch/tile/kernel/reboot.c
@@ -0,0 +1,51 @@
+/*
+ * Copyright 2010 Tilera Corporation. All Rights Reserved.
+ *
+ *   This program is free software; you can redistribute it and/or
+ *   modify it under the terms of the GNU General Public License
+ *   as published by the Free Software Foundation, version 2.
+ *
+ *   This program is distributed in the hope that it will be useful, but
+ *   WITHOUT ANY WARRANTY; without even the implied warranty of
+ *   MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
+ *   NON INFRINGEMENT.  See the GNU General Public License for
+ *   more details.
+ */
+
+#include <linux/stddef.h>
+#include <linux/reboot.h>
+#include <linux/smp.h>
+#include <linux/pm.h>
+#include <linux/export.h>
+#include <asm/page.h>
+#include <asm/setup.h>
+#include <hv/hypervisor.h>
+
+#ifndef CONFIG_SMP
+#define smp_send_stop()
+#endif
+
+void machine_halt(void)
+{
+	arch_local_irq_disable_all();
+	smp_send_stop();
+	hv_halt();
+}
+
+void machine_power_off(void)
+{
+	arch_local_irq_disable_all();
+	smp_send_stop();
+	hv_power_off();
+}
+
+void machine_restart(char *cmd)
+{
+	arch_local_irq_disable_all();
+	smp_send_stop();
+	hv_restart((HV_VirtAddr) "vmlinux", (HV_VirtAddr) cmd);
+}
+
+/* No interesting distinction to be made here. */
+void (*pm_power_off)(void) = NULL;
+EXPORT_SYMBOL(pm_power_off);
diff --git a/kernel/arch/tile/kernel/regs_32.S b/kernel/arch/tile/kernel/regs_32.S
new file mode 100644
index 000000000..542cae17a
--- /dev/null
+++ b/kernel/arch/tile/kernel/regs_32.S
@@ -0,0 +1,145 @@
+/*
+ * Copyright 2010 Tilera Corporation. All Rights Reserved.
+ *
+ *   This program is free software; you can redistribute it and/or
+ *   modify it under the terms of the GNU General Public License
+ *   as published by the Free Software Foundation, version 2.
+ *
+ *   This program is distributed in the hope that it will be useful, but
+ *   WITHOUT ANY WARRANTY; without even the implied warranty of
+ *   MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
+ *   NON INFRINGEMENT.  See the GNU General Public License for
+ *   more details.
+ */
+
+#include <linux/linkage.h>
+#include <asm/ptrace.h>
+#include <asm/asm-offsets.h>
+#include <arch/spr_def.h>
+#include <asm/processor.h>
+#include <asm/switch_to.h>
+
+/*
+ * See <asm/switch_to.h>; called with prev and next task_struct pointers.
+ * "prev" is returned in r0 for _switch_to and also for ret_from_fork.
+ *
+ * We want to save pc/sp in "prev", and get the new pc/sp from "next".
+ * We also need to save all the callee-saved registers on the stack.
+ *
+ * Intel enables/disables access to the hardware cycle counter in
+ * seccomp (secure computing) environments if necessary, based on
+ * has_secure_computing().  We might want to do this at some point,
+ * though it would require virtualizing the other SPRs under WORLD_ACCESS.
+ *
+ * Since we're saving to the stack, we omit sp from this list.
+ * And for parallels with other architectures, we save lr separately,
+ * in the thread_struct itself (as the "pc" field).
+ *
+ * This code also needs to be aligned with process.c copy_thread()
+ */
+
+#if CALLEE_SAVED_REGS_COUNT != 24
+# error Mismatch between <asm/switch_to.h> and kernel/entry.S
+#endif
+#define FRAME_SIZE ((2 + CALLEE_SAVED_REGS_COUNT) * 4)
+
+#define SAVE_REG(r) { sw r12, r; addi r12, r12, 4 }
+#define LOAD_REG(r) { lw r, r12; addi r12, r12, 4 }
+#define FOR_EACH_CALLEE_SAVED_REG(f)					\
+							f(r30); f(r31); \
+	f(r32); f(r33); f(r34); f(r35);	f(r36); f(r37); f(r38); f(r39); \
+	f(r40); f(r41); f(r42); f(r43); f(r44); f(r45); f(r46); f(r47); \
+	f(r48); f(r49); f(r50); f(r51); f(r52);
+
+STD_ENTRY_SECTION(__switch_to, .sched.text)
+	{
+	  move r10, sp
+	  sw sp, lr
+	  addi sp, sp, -FRAME_SIZE
+	}
+	{
+	  addi r11, sp, 4
+	  addi r12, sp, 8
+	}
+	{
+	  sw r11, r10
+	  addli r4, r1, TASK_STRUCT_THREAD_KSP_OFFSET
+	}
+	{
+	  lw r13, r4   /* Load new sp to a temp register early. */
+	  addli r3, r0, TASK_STRUCT_THREAD_KSP_OFFSET
+	}
+	FOR_EACH_CALLEE_SAVED_REG(SAVE_REG)
+	{
+	  sw r3, sp
+	  addli r3, r0, TASK_STRUCT_THREAD_PC_OFFSET
+	}
+	{
+	  sw r3, lr
+	  addli r4, r1, TASK_STRUCT_THREAD_PC_OFFSET
+	}
+	{
+	  lw lr, r4
+	  addi r12, r13, 8
+	}
+	{
+	  /* Update sp and ksp0 simultaneously to avoid backtracer warnings. */
+	  move sp, r13
+	  mtspr SPR_SYSTEM_SAVE_K_0, r2
+	}
+	FOR_EACH_CALLEE_SAVED_REG(LOAD_REG)
+.L__switch_to_pc:
+	{
+	  addi sp, sp, FRAME_SIZE
+	  jrp lr   /* r0 is still valid here, so return it */
+	}
+	STD_ENDPROC(__switch_to)
+
+/* Return a suitable address for the backtracer for suspended threads */
+STD_ENTRY_SECTION(get_switch_to_pc, .sched.text)
+	lnk r0
+	{
+	  addli r0, r0, .L__switch_to_pc - .
+	  jrp lr
+	}
+	STD_ENDPROC(get_switch_to_pc)
+
+STD_ENTRY(get_pt_regs)
+	.irp reg, r0, r1, r2, r3, r4, r5, r6, r7, \
+		 r8, r9, r10, r11, r12, r13, r14, r15, \
+		 r16, r17, r18, r19, r20, r21, r22, r23, \
+		 r24, r25, r26, r27, r28, r29, r30, r31, \
+		 r32, r33, r34, r35, r36, r37, r38, r39, \
+		 r40, r41, r42, r43, r44, r45, r46, r47, \
+		 r48, r49, r50, r51, r52, tp, sp
+	{
+	 sw r0, \reg
+	 addi r0, r0, 4
+	}
+	.endr
+	{
+	 sw r0, lr
+	 addi r0, r0, PTREGS_OFFSET_PC - PTREGS_OFFSET_LR
+	}
+	lnk r1
+	{
+	 sw r0, r1
+	 addi r0, r0, PTREGS_OFFSET_EX1 - PTREGS_OFFSET_PC
+	}
+	mfspr r1, INTERRUPT_CRITICAL_SECTION
+	shli r1, r1, SPR_EX_CONTEXT_1_1__ICS_SHIFT
+	ori r1, r1, KERNEL_PL
+	{
+	 sw r0, r1
+	 addi r0, r0, PTREGS_OFFSET_FAULTNUM - PTREGS_OFFSET_EX1
+	}
+	{
+	 sw r0, zero       /* clear faultnum */
+	 addi r0, r0, PTREGS_OFFSET_ORIG_R0 - PTREGS_OFFSET_FAULTNUM
+	}
+	{
+	 sw r0, zero       /* clear orig_r0 */
+	 addli r0, r0, -PTREGS_OFFSET_ORIG_R0    /* restore r0 to base */
+	}
+	jrp lr
+	STD_ENDPROC(get_pt_regs)
diff --git a/kernel/arch/tile/kernel/regs_64.S b/kernel/arch/tile/kernel/regs_64.S
new file mode 100644
index 000000000..bbffcc6f3
--- /dev/null
+++ b/kernel/arch/tile/kernel/regs_64.S
@@ -0,0 +1,145 @@
+/*
+ * Copyright 2011 Tilera Corporation. All Rights Reserved.
+ *
+ *   This program is free software; you can redistribute it and/or
+ *   modify it under the terms of the GNU General Public License
+ *   as published by the Free Software Foundation, version 2.
+ *
+ *   This program is distributed in the hope that it will be useful, but
+ *   WITHOUT ANY WARRANTY; without even the implied warranty of
+ *   MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
+ *   NON INFRINGEMENT.  See the GNU General Public License for
+ *   more details.
+ */
+
+#include <linux/linkage.h>
+#include <asm/ptrace.h>
+#include <asm/asm-offsets.h>
+#include <arch/spr_def.h>
+#include <asm/processor.h>
+#include <asm/switch_to.h>
+
+/*
+ * See <asm/switch_to.h>; called with prev and next task_struct pointers.
+ * "prev" is returned in r0 for _switch_to and also for ret_from_fork.
+ *
+ * We want to save pc/sp in "prev", and get the new pc/sp from "next".
+ * We also need to save all the callee-saved registers on the stack.
+ *
+ * Intel enables/disables access to the hardware cycle counter in
+ * seccomp (secure computing) environments if necessary, based on
+ * has_secure_computing().  We might want to do this at some point,
+ * though it would require virtualizing the other SPRs under WORLD_ACCESS.
+ *
+ * Since we're saving to the stack, we omit sp from this list.
+ * And for parallels with other architectures, we save lr separately,
+ * in the thread_struct itself (as the "pc" field).
+ *
+ * This code also needs to be aligned with process.c copy_thread()
+ */
+
+#if CALLEE_SAVED_REGS_COUNT != 24
+# error Mismatch between <asm/switch_to.h> and kernel/entry.S
+#endif
+#define FRAME_SIZE ((2 + CALLEE_SAVED_REGS_COUNT) * 8)
+
+#define SAVE_REG(r) { st r12, r; addi r12, r12, 8 }
+#define LOAD_REG(r) { ld r, r12; addi r12, r12, 8 }
+#define FOR_EACH_CALLEE_SAVED_REG(f)					\
+							f(r30); f(r31); \
+	f(r32); f(r33); f(r34); f(r35);	f(r36); f(r37); f(r38); f(r39); \
+	f(r40); f(r41); f(r42); f(r43); f(r44); f(r45); f(r46); f(r47); \
+	f(r48); f(r49); f(r50); f(r51); f(r52);
+
+STD_ENTRY_SECTION(__switch_to, .sched.text)
+	{
+	  move r10, sp
+	  st sp, lr
+	}
+	{
+	  addli r11, sp, -FRAME_SIZE + 8
+	  addli sp, sp, -FRAME_SIZE
+	}
+	{
+	  st r11, r10
+	  addli r4, r1, TASK_STRUCT_THREAD_KSP_OFFSET
+	}
+	{
+	  ld r13, r4   /* Load new sp to a temp register early. */
+	  addi r12, sp, 16
+	}
+	FOR_EACH_CALLEE_SAVED_REG(SAVE_REG)
+	addli r3, r0, TASK_STRUCT_THREAD_KSP_OFFSET
+	{
+	  st r3, sp
+	  addli r3, r0, TASK_STRUCT_THREAD_PC_OFFSET
+	}
+	{
+	  st r3, lr
+	  addli r4, r1, TASK_STRUCT_THREAD_PC_OFFSET
+	}
+	{
+	  ld lr, r4
+	  addi r12, r13, 16
+	}
+	{
+	  /* Update sp and ksp0 simultaneously to avoid backtracer warnings. */
+	  move sp, r13
+	  mtspr SPR_SYSTEM_SAVE_K_0, r2
+	}
+	FOR_EACH_CALLEE_SAVED_REG(LOAD_REG)
+.L__switch_to_pc:
+	{
+	  addli sp, sp, FRAME_SIZE
+	  jrp lr   /* r0 is still valid here, so return it */
+	}
+	STD_ENDPROC(__switch_to)
+
+/* Return a suitable address for the backtracer for suspended threads */
+STD_ENTRY_SECTION(get_switch_to_pc, .sched.text)
+	lnk r0
+	{
+	  addli r0, r0, .L__switch_to_pc - .
+	  jrp lr
+	}
+	STD_ENDPROC(get_switch_to_pc)
+
+STD_ENTRY(get_pt_regs)
+	.irp reg, r0, r1, r2, r3, r4, r5, r6, r7, \
+		 r8, r9, r10, r11, r12, r13, r14, r15, \
+		 r16, r17, r18, r19, r20, r21, r22, r23, \
+		 r24, r25, r26, r27, r28, r29, r30, r31, \
+		 r32, r33, r34, r35, r36, r37, r38, r39, \
+		 r40, r41, r42, r43, r44, r45, r46, r47, \
+		 r48, r49, r50, r51, r52, tp, sp
+	{
+	 st r0, \reg
+	 addi r0, r0, 8
+	}
+	.endr
+	{
+	 st r0, lr
+	 addi r0, r0, PTREGS_OFFSET_PC - PTREGS_OFFSET_LR
+	}
+	lnk r1
+	{
+	 st r0, r1
+	 addi r0, r0, PTREGS_OFFSET_EX1 - PTREGS_OFFSET_PC
+	}
+	mfspr r1, INTERRUPT_CRITICAL_SECTION
+	shli r1, r1, SPR_EX_CONTEXT_1_1__ICS_SHIFT
+	ori r1, r1, KERNEL_PL
+	{
+	 st r0, r1
+	 addi r0, r0, PTREGS_OFFSET_FAULTNUM - PTREGS_OFFSET_EX1
+	}
+	{
+	 st r0, zero       /* clear faultnum */
+	 addi r0, r0, PTREGS_OFFSET_ORIG_R0 - PTREGS_OFFSET_FAULTNUM
+	}
+	{
+	 st r0, zero       /* clear orig_r0 */
+	 addli r0, r0, -PTREGS_OFFSET_ORIG_R0    /* restore r0 to base */
+	}
+	jrp lr
+	STD_ENDPROC(get_pt_regs)
diff --git a/kernel/arch/tile/kernel/relocate_kernel_32.S b/kernel/arch/tile/kernel/relocate_kernel_32.S
new file mode 100644
index 000000000..e44fbcf8c
--- /dev/null
+++ b/kernel/arch/tile/kernel/relocate_kernel_32.S
@@ -0,0 +1,269 @@
+/*
+ * Copyright 2010 Tilera Corporation. All Rights Reserved.
+ *
+ *   This program is free software; you can redistribute it and/or
+ *   modify it under the terms of the GNU General Public License
+ *   as published by the Free Software Foundation, version 2.
+ *
+ *   This program is distributed in the hope that it will be useful, but
+ *   WITHOUT ANY WARRANTY; without even the implied warranty of
+ *   MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
+ *   NON INFRINGEMENT.  See the GNU General Public License for
+ *   more details.
+ *
+ * copy new kernel into place and then call hv_reexec
+ *
+ */
+
+#include <linux/linkage.h>
+#include <arch/chip.h>
+#include <asm/page.h>
+#include <hv/hypervisor.h>
+
+#undef RELOCATE_NEW_KERNEL_VERBOSE
+
+STD_ENTRY(relocate_new_kernel)
+
+	move	r30, r0		/* page list */
+	move	r31, r1		/* address of page we are on */
+	move	r32, r2		/* start address of new kernel */
+
+	shri	r1, r1, PAGE_SHIFT
+	addi	r1, r1, 1
+	shli	sp, r1, PAGE_SHIFT
+	addi	sp, sp, -8
+	/* we now have a stack (whether we need one or not) */
+
+	moveli	r40, lo16(hv_console_putc)
+	auli	r40, r40, ha16(hv_console_putc)
+
+#ifdef RELOCATE_NEW_KERNEL_VERBOSE
+	moveli	r0, 'r'
+	jalr	r40
+
+	moveli	r0, '_'
+	jalr	r40
+
+	moveli	r0, 'n'
+	jalr	r40
+
+	moveli	r0, '_'
+	jalr	r40
+
+	moveli	r0, 'k'
+	jalr	r40
+
+	moveli	r0, '\n'
+	jalr	r40
+#endif
+
+	/*
+	 * Throughout this code r30 is pointer to the element of page
+	 * list we are working on.
+	 *
+	 * Normally we get to the next element of the page list by
+	 * incrementing r30 by four.  The exception is if the element
+	 * on the page list is an IND_INDIRECTION in which case we use
+	 * the element with the low bits masked off as the new value
+	 * of r30.
+	 *
+	 * To get this started, we need the value passed to us (which
+	 * will always be an IND_INDIRECTION) in memory somewhere with
+	 * r30 pointing at it.  To do that, we push the value passed
+	 * to us on the stack and make r30 point to it.
+	 */
+
+	sw	sp, r30
+	move	r30, sp
+	addi	sp, sp, -8
+
+	/*
+	 * On TILEPro, we need to flush all tiles' caches, since we may
+	 * have been doing hash-for-home caching there.  Note that we
+	 * must do this _after_ we're completely done modifying any memory
+	 * other than our output buffer (which we know is locally cached).
+	 * We want the caches to be fully clean when we do the reexec,
+	 * because the hypervisor is going to do this flush again at that
+	 * point, and we don't want that second flush to overwrite any memory.
+	 */
+	{
+	 move	r0, zero	 /* cache_pa */
+	 move	r1, zero
+	}
+	{
+	 auli	r2, zero, ha16(HV_FLUSH_EVICT_L2) /* cache_control */
+	 movei	r3, -1		 /* cache_cpumask; -1 means all client tiles */
+	}
+	{
+	 move	r4, zero	 /* tlb_va */
+	 move	r5, zero	 /* tlb_length */
+	}
+	{
+	 move	r6, zero	 /* tlb_pgsize */
+	 move	r7, zero	 /* tlb_cpumask */
+	}
+	{
+	 move	r8, zero	 /* asids */
+	 moveli	r20, lo16(hv_flush_remote)
+	}
+	{
+	 move	r9, zero	 /* asidcount */
+	 auli	r20, r20, ha16(hv_flush_remote)
+	}
+
+	jalr	r20
+
+	/* r33 is destination pointer, default to zero */
+
+	moveli	r33, 0
+
+.Lloop:	lw	r10, r30
+
+	andi	r9, r10, 0xf	/* low 4 bits tell us what type it is */
+	xor	r10, r10, r9	/* r10 is now value with low 4 bits stripped */
+
+	seqi	r0, r9, 0x1	/* IND_DESTINATION */
+	bzt	r0, .Ltry2
+
+	move	r33, r10
+
+#ifdef RELOCATE_NEW_KERNEL_VERBOSE
+	moveli	r0, 'd'
+	jalr	r40
+#endif
+
+	addi	r30, r30, 4
+	j	.Lloop
+
+.Ltry2:
+	seqi	r0, r9, 0x2	/* IND_INDIRECTION */
+	bzt	r0, .Ltry4
+
+	move	r30, r10
+
+#ifdef RELOCATE_NEW_KERNEL_VERBOSE
+	moveli	r0, 'i'
+	jalr	r40
+#endif
+
+	j	.Lloop
+
+.Ltry4:
+	seqi	r0, r9, 0x4	/* IND_DONE */
+	bzt	r0, .Ltry8
+
+	mf
+
+#ifdef RELOCATE_NEW_KERNEL_VERBOSE
+	moveli	r0, 'D'
+	jalr	r40
+	moveli	r0, '\n'
+	jalr	r40
+#endif
+
+	move	r0, r32
+	moveli	r1, 0		/* arg to hv_reexec is 64 bits */
+
+	moveli	r41, lo16(hv_reexec)
+	auli	r41, r41, ha16(hv_reexec)
+
+	jalr	r41
+
+	/* we should not get here */
+
+	moveli	r0, '?'
+	jalr	r40
+	moveli	r0, '\n'
+	jalr	r40
+
+	j	.Lhalt
+
+.Ltry8:	seqi	r0, r9, 0x8	/* IND_SOURCE */
+	bz	r0, .Lerr	/* unknown type */
+
+	/* copy page at r10 to page at r33 */
+
+	move	r11, r33
+
+	moveli	r0, lo16(PAGE_SIZE)
+	auli	r0, r0, ha16(PAGE_SIZE)
+	add	r33, r33, r0
+
+	/* copy word at r10 to word at r11 until r11 equals r33 */
+
+	/* We know page size must be multiple of 16, so we can unroll
+	 * 16 times safely without any edge case checking.
+	 *
+	 * Issue a flush of the destination every 16 words to avoid
+	 * incoherence when starting the new kernel.  (Now this is
+	 * just good paranoia because the hv_reexec call will also
+	 * take care of this.)
+	 */
+
+1:
+	{ lw	r0, r10; addi	r10, r10, 4 }
+	{ sw	r11, r0; addi	r11, r11, 4 }
+	{ lw	r0, r10; addi	r10, r10, 4 }
+	{ sw	r11, r0; addi	r11, r11, 4 }
+	{ lw	r0, r10; addi	r10, r10, 4 }
+	{ sw	r11, r0; addi	r11, r11, 4 }
+	{ lw	r0, r10; addi	r10, r10, 4 }
+	{ sw	r11, r0; addi	r11, r11, 4 }
+	{ lw	r0, r10; addi	r10, r10, 4 }
+	{ sw	r11, r0; addi	r11, r11, 4 }
+	{ lw	r0, r10; addi	r10, r10, 4 }
+	{ sw	r11, r0; addi	r11, r11, 4 }
+	{ lw	r0, r10; addi	r10, r10, 4 }
+	{ sw	r11, r0; addi	r11, r11, 4 }
+	{ lw	r0, r10; addi	r10, r10, 4 }
+	{ sw	r11, r0; addi	r11, r11, 4 }
+	{ lw	r0, r10; addi	r10, r10, 4 }
+	{ sw	r11, r0; addi	r11, r11, 4 }
+	{ lw	r0, r10; addi	r10, r10, 4 }
+	{ sw	r11, r0; addi	r11, r11, 4 }
+	{ lw	r0, r10; addi	r10, r10, 4 }
+	{ sw	r11, r0; addi	r11, r11, 4 }
+	{ lw	r0, r10; addi	r10, r10, 4 }
+	{ sw	r11, r0; addi	r11, r11, 4 }
+	{ lw	r0, r10; addi	r10, r10, 4 }
+	{ sw	r11, r0; addi	r11, r11, 4 }
+	{ lw	r0, r10; addi	r10, r10, 4 }
+	{ sw	r11, r0; addi	r11, r11, 4 }
+	{ lw	r0, r10; addi	r10, r10, 4 }
+	{ sw	r11, r0; addi	r11, r11, 4 }
+	{ lw	r0, r10; addi	r10, r10, 4 }
+	{ sw	r11, r0 }
+	{ flush r11    ; addi	r11, r11, 4 }
+
+	seq	r0, r33, r11
+	bzt	r0, 1b
+
+#ifdef RELOCATE_NEW_KERNEL_VERBOSE
+	moveli	r0, 's'
+	jalr	r40
+#endif
+
+	addi	r30, r30, 4
+	j	.Lloop
+
+
+.Lerr:	moveli	r0, 'e'
+	jalr	r40
+	moveli	r0, 'r'
+	jalr	r40
+	moveli	r0, 'r'
+	jalr	r40
+	moveli	r0, '\n'
+	jalr	r40
+.Lhalt:
+	moveli	r41, lo16(hv_halt)
+	auli	r41, r41, ha16(hv_halt)
+
+	jalr	r41
+	STD_ENDPROC(relocate_new_kernel)
+
+	.section .rodata,"a"
+
+	.globl relocate_new_kernel_size
+relocate_new_kernel_size:
+	.long .Lend_relocate_new_kernel - relocate_new_kernel
diff --git a/kernel/arch/tile/kernel/relocate_kernel_64.S b/kernel/arch/tile/kernel/relocate_kernel_64.S
new file mode 100644
index 000000000..d9d8cf617
--- /dev/null
+++ b/kernel/arch/tile/kernel/relocate_kernel_64.S
@@ -0,0 +1,263 @@
+/*
+ * Copyright 2011 Tilera Corporation. All Rights Reserved.
+ *
+ *   This program is free software; you can redistribute it and/or
+ *   modify it under the terms of the GNU General Public License
+ *   as published by the Free Software Foundation, version 2.
+ *
+ *   This program is distributed in the hope that it will be useful, but
+ *   WITHOUT ANY WARRANTY; without even the implied warranty of
+ *   MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
+ *   NON INFRINGEMENT.  See the GNU General Public License for
+ *   more details.
+ *
+ * copy new kernel into place and then call hv_reexec
+ *
+ */
+
+#include <linux/linkage.h>
+#include <arch/chip.h>
+#include <asm/page.h>
+#include <hv/hypervisor.h>
+
+#undef RELOCATE_NEW_KERNEL_VERBOSE
+
+STD_ENTRY(relocate_new_kernel)
+
+	move	r30, r0		/* page list */
+	move	r31, r1		/* address of page we are on */
+	move	r32, r2		/* start address of new kernel */
+
+	shrui	r1, r1, PAGE_SHIFT
+	addi	r1, r1, 1
+	shli	sp, r1, PAGE_SHIFT
+	addi	sp, sp, -8
+	/* we now have a stack (whether we need one or not) */
+
+#ifdef RELOCATE_NEW_KERNEL_VERBOSE
+	moveli	r40, hw2_last(hv_console_putc)
+	shl16insli r40, r40, hw1(hv_console_putc)
+	shl16insli r40, r40, hw0(hv_console_putc)
+
+	moveli	r0, 'r'
+	jalr	r40
+
+	moveli	r0, '_'
+	jalr	r40
+
+	moveli	r0, 'n'
+	jalr	r40
+
+	moveli	r0, '_'
+	jalr	r40
+
+	moveli	r0, 'k'
+	jalr	r40
+
+	moveli	r0, '\n'
+	jalr	r40
+#endif
+
+	/*
+	 * Throughout this code r30 is pointer to the element of page
+	 * list we are working on.
+	 *
+	 * Normally we get to the next element of the page list by
+	 * incrementing r30 by eight.  The exception is if the element
+	 * on the page list is an IND_INDIRECTION in which case we use
+	 * the element with the low bits masked off as the new value
+	 * of r30.
+	 *
+	 * To get this started, we need the value passed to us (which
+	 * will always be an IND_INDIRECTION) in memory somewhere with
+	 * r30 pointing at it.  To do that, we push the value passed
+	 * to us on the stack and make r30 point to it.
+	 */
+
+	st	sp, r30
+	move	r30, sp
+	addi	sp, sp, -16
+
+	/*
+	 * On TILE-GX, we need to flush all tiles' caches, since we may
+	 * have been doing hash-for-home caching there.  Note that we
+	 * must do this _after_ we're completely done modifying any memory
+	 * other than our output buffer (which we know is locally cached).
+	 * We want the caches to be fully clean when we do the reexec,
+	 * because the hypervisor is going to do this flush again at that
+	 * point, and we don't want that second flush to overwrite any memory.
+	 */
+	{
+	 move	r0, zero	 /* cache_pa */
+	 moveli	r1, hw2_last(HV_FLUSH_EVICT_L2)
+	}
+	{
+	 shl16insli	r1, r1, hw1(HV_FLUSH_EVICT_L2)
+	 movei	r2, -1		 /* cache_cpumask; -1 means all client tiles */
+	}
+	{
+	 shl16insli	r1, r1, hw0(HV_FLUSH_EVICT_L2)  /* cache_control */
+	 move	r3, zero	 /* tlb_va */
+	}
+	{
+	 move	r4, zero	 /* tlb_length */
+	 move	r5, zero	 /* tlb_pgsize */
+	}
+	{
+	 move	r6, zero	 /* tlb_cpumask */
+	 move	r7, zero	 /* asids */
+	}
+	{
+	 moveli	r20, hw2_last(hv_flush_remote)
+	 move	r8, zero	 /* asidcount */
+	}
+	shl16insli	r20, r20, hw1(hv_flush_remote)
+	shl16insli	r20, r20, hw0(hv_flush_remote)
+
+	jalr	r20
+
+	/* r33 is destination pointer, default to zero */
+
+	moveli	r33, 0
+
+.Lloop:	ld	r10, r30
+
+	andi	r9, r10, 0xf	/* low 4 bits tell us what type it is */
+	xor	r10, r10, r9	/* r10 is now value with low 4 bits stripped */
+
+	cmpeqi	r0, r9, 0x1	/* IND_DESTINATION */
+	beqzt	r0, .Ltry2
+
+	move	r33, r10
+
+#ifdef RELOCATE_NEW_KERNEL_VERBOSE
+	moveli	r0, 'd'
+	jalr	r40
+#endif
+
+	addi	r30, r30, 8
+	j	.Lloop
+
+.Ltry2:
+	cmpeqi	r0, r9, 0x2	/* IND_INDIRECTION */
+	beqzt	r0, .Ltry4
+
+	move	r30, r10
+
+#ifdef RELOCATE_NEW_KERNEL_VERBOSE
+	moveli	r0, 'i'
+	jalr	r40
+#endif
+
+	j	.Lloop
+
+.Ltry4:
+	cmpeqi	r0, r9, 0x4	/* IND_DONE */
+	beqzt	r0, .Ltry8
+
+	mf
+
+#ifdef RELOCATE_NEW_KERNEL_VERBOSE
+	moveli	r0, 'D'
+	jalr	r40
+	moveli	r0, '\n'
+	jalr	r40
+#endif
+
+	move	r0, r32
+
+	moveli	r41, hw2_last(hv_reexec)
+	shl16insli	r41, r41, hw1(hv_reexec)
+	shl16insli	r41, r41, hw0(hv_reexec)
+
+	jalr	r41
+
+	/* we should not get here */
+
+#ifdef RELOCATE_NEW_KERNEL_VERBOSE
+	moveli	r0, '?'
+	jalr	r40
+	moveli	r0, '\n'
+	jalr	r40
+#endif
+
+	j	.Lhalt
+
+.Ltry8:	cmpeqi	r0, r9, 0x8	/* IND_SOURCE */
+	beqz	r0, .Lerr	/* unknown type */
+
+	/* copy page at r10 to page at r33 */
+
+	move	r11, r33
+
+	moveli	r0, hw2_last(PAGE_SIZE)
+	shl16insli	r0, r0, hw1(PAGE_SIZE)
+	shl16insli	r0, r0, hw0(PAGE_SIZE)
+	add	r33, r33, r0
+
+	/* copy word at r10 to word at r11 until r11 equals r33 */
+
+	/* We know page size must be multiple of 8, so we can unroll
+	 * 8 times safely without any edge case checking.
+	 *
+	 * Issue a flush of the destination every 8 words to avoid
+	 * incoherence when starting the new kernel.  (Now this is
+	 * just good paranoia because the hv_reexec call will also
+	 * take care of this.)
+	 */
+
+1:
+	{ ld	r0, r10; addi	r10, r10, 8 }
+	{ st	r11, r0; addi	r11, r11, 8 }
+	{ ld	r0, r10; addi	r10, r10, 8 }
+	{ st	r11, r0; addi	r11, r11, 8 }
+	{ ld	r0, r10; addi	r10, r10, 8 }
+	{ st	r11, r0; addi	r11, r11, 8 }
+	{ ld	r0, r10; addi	r10, r10, 8 }
+	{ st	r11, r0; addi	r11, r11, 8 }
+	{ ld	r0, r10; addi	r10, r10, 8 }
+	{ st	r11, r0; addi	r11, r11, 8 }
+	{ ld	r0, r10; addi	r10, r10, 8 }
+	{ st	r11, r0; addi	r11, r11, 8 }
+	{ ld	r0, r10; addi	r10, r10, 8 }
+	{ st	r11, r0; addi	r11, r11, 8 }
+	{ ld	r0, r10; addi	r10, r10, 8 }
+	{ st	r11, r0 }
+	{ flush r11    ; addi	r11, r11, 8 }
+
+	cmpeq	r0, r33, r11
+	beqzt	r0, 1b
+
+#ifdef RELOCATE_NEW_KERNEL_VERBOSE
+	moveli	r0, 's'
+	jalr	r40
+#endif
+
+	addi	r30, r30, 8
+	j	.Lloop
+
+
+.Lerr:
+#ifdef RELOCATE_NEW_KERNEL_VERBOSE
+	moveli	r0, 'e'
+	jalr	r40
+	moveli	r0, 'r'
+	jalr	r40
+	moveli	r0, 'r'
+	jalr	r40
+	moveli	r0, '\n'
+	jalr	r40
+#endif
+.Lhalt:
+	moveli r41, hw2_last(hv_halt)
+	shl16insli r41, r41, hw1(hv_halt)
+	shl16insli r41, r41, hw0(hv_halt)
+
+	jalr	r41
+	STD_ENDPROC(relocate_new_kernel)
+
+	.section .rodata,"a"
+
+	.globl relocate_new_kernel_size
+relocate_new_kernel_size:
+	.long .Lend_relocate_new_kernel - relocate_new_kernel
diff --git a/kernel/arch/tile/kernel/setup.c b/kernel/arch/tile/kernel/setup.c
new file mode 100644
index 000000000..d366675e4
--- /dev/null
+++ b/kernel/arch/tile/kernel/setup.c
@@ -0,0 +1,1740 @@
+/*
+ * Copyright 2010 Tilera Corporation. All Rights Reserved.
+ *
+ *   This program is free software; you can redistribute it and/or
+ *   modify it under the terms of the GNU General Public License
+ *   as published by the Free Software Foundation, version 2.
+ *
+ *   This program is distributed in the hope that it will be useful, but
+ *   WITHOUT ANY WARRANTY; without even the implied warranty of
+ *   MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
+ *   NON INFRINGEMENT.  See the GNU General Public License for
+ *   more details.
+ */
+
+#include <linux/sched.h>
+#include <linux/kernel.h>
+#include <linux/mmzone.h>
+#include <linux/bootmem.h>
+#include <linux/module.h>
+#include <linux/node.h>
+#include <linux/cpu.h>
+#include <linux/ioport.h>
+#include <linux/irq.h>
+#include <linux/kexec.h>
+#include <linux/pci.h>
+#include <linux/swiotlb.h>
+#include <linux/initrd.h>
+#include <linux/io.h>
+#include <linux/highmem.h>
+#include <linux/smp.h>
+#include <linux/timex.h>
+#include <linux/hugetlb.h>
+#include <linux/start_kernel.h>
+#include <linux/screen_info.h>
+#include <linux/tick.h>
+#include <asm/setup.h>
+#include <asm/sections.h>
+#include <asm/cacheflush.h>
+#include <asm/pgalloc.h>
+#include <asm/mmu_context.h>
+#include <hv/hypervisor.h>
+#include <arch/interrupts.h>
+
+/* <linux/smp.h> doesn't provide this definition. */
+#ifndef CONFIG_SMP
+#define setup_max_cpus 1
+#endif
+
+static inline int ABS(int x) { return x >= 0 ? x : -x; }
+
+/* Chip information */
+char chip_model[64] __write_once;
+
+#ifdef CONFIG_VT
+struct screen_info screen_info;
+#endif
+
+struct pglist_data node_data[MAX_NUMNODES] __read_mostly;
+EXPORT_SYMBOL(node_data);
+
+/* Information on the NUMA nodes that we compute early */
+unsigned long node_start_pfn[MAX_NUMNODES];
+unsigned long node_end_pfn[MAX_NUMNODES];
+unsigned long __initdata node_memmap_pfn[MAX_NUMNODES];
+unsigned long __initdata node_percpu_pfn[MAX_NUMNODES];
+unsigned long __initdata node_free_pfn[MAX_NUMNODES];
+
+static unsigned long __initdata node_percpu[MAX_NUMNODES];
+
+/*
+ * per-CPU stack and boot info.
+ */
+DEFINE_PER_CPU(unsigned long, boot_sp) =
+	(unsigned long)init_stack + THREAD_SIZE;
+
+#ifdef CONFIG_SMP
+DEFINE_PER_CPU(unsigned long, boot_pc) = (unsigned long)start_kernel;
+#else
+/*
+ * The variable must be __initdata since it references __init code.
+ * With CONFIG_SMP it is per-cpu data, which is exempt from validation.
+ */
+unsigned long __initdata boot_pc = (unsigned long)start_kernel;
+#endif
+
+#ifdef CONFIG_HIGHMEM
+/* Page frame index of end of lowmem on each controller. */
+unsigned long node_lowmem_end_pfn[MAX_NUMNODES];
+
+/* Number of pages that can be mapped into lowmem. */
+static unsigned long __initdata mappable_physpages;
+#endif
+
+/* Data on which physical memory controller corresponds to which NUMA node */
+int node_controller[MAX_NUMNODES] = { [0 ... MAX_NUMNODES-1] = -1 };
+
+#ifdef CONFIG_HIGHMEM
+/* Map information from VAs to PAs */
+unsigned long pbase_map[1 << (32 - HPAGE_SHIFT)]
+  __write_once __attribute__((aligned(L2_CACHE_BYTES)));
+EXPORT_SYMBOL(pbase_map);
+
+/* Map information from PAs to VAs */
+void *vbase_map[NR_PA_HIGHBIT_VALUES]
+  __write_once __attribute__((aligned(L2_CACHE_BYTES)));
+EXPORT_SYMBOL(vbase_map);
+#endif
+
+/* Node number as a function of the high PA bits */
+int highbits_to_node[NR_PA_HIGHBIT_VALUES] __write_once;
+EXPORT_SYMBOL(highbits_to_node);
+
+static unsigned int __initdata maxmem_pfn = -1U;
+static unsigned int __initdata maxnodemem_pfn[MAX_NUMNODES] = {
+	[0 ... MAX_NUMNODES-1] = -1U
+};
+static nodemask_t __initdata isolnodes;
+
+#if defined(CONFIG_PCI) && !defined(__tilegx__)
+enum { DEFAULT_PCI_RESERVE_MB = 64 };
+static unsigned int __initdata pci_reserve_mb = DEFAULT_PCI_RESERVE_MB;
+unsigned long __initdata pci_reserve_start_pfn = -1U;
+unsigned long __initdata pci_reserve_end_pfn = -1U;
+#endif
+
+static int __init setup_maxmem(char *str)
+{
+	unsigned long long maxmem;
+	if (str == NULL || (maxmem = memparse(str, NULL)) == 0)
+		return -EINVAL;
+
+	maxmem_pfn = (maxmem >> HPAGE_SHIFT) << (HPAGE_SHIFT - PAGE_SHIFT);
+	pr_info("Forcing RAM used to no more than %dMB\n",
+		maxmem_pfn >> (20 - PAGE_SHIFT));
+	return 0;
+}
+early_param("maxmem", setup_maxmem);
+
+static int __init setup_maxnodemem(char *str)
+{
+	char *endp;
+	unsigned long long maxnodemem;
+	long node;
+
+	node = str ? simple_strtoul(str, &endp, 0) : INT_MAX;
+	if (node >= MAX_NUMNODES || *endp != ':')
+		return -EINVAL;
+
+	maxnodemem = memparse(endp+1, NULL);
+	maxnodemem_pfn[node] = (maxnodemem >> HPAGE_SHIFT) <<
+		(HPAGE_SHIFT - PAGE_SHIFT);
+	pr_info("Forcing RAM used on node %ld to no more than %dMB\n",
+		node, maxnodemem_pfn[node] >> (20 - PAGE_SHIFT));
+	return 0;
+}
+early_param("maxnodemem", setup_maxnodemem);
+
+struct memmap_entry {
+	u64 addr;	/* start of memory segment */
+	u64 size;	/* size of memory segment */
+};
+static struct memmap_entry memmap_map[64];
+static int memmap_nr;
+
+static void add_memmap_region(u64 addr, u64 size)
+{
+	if (memmap_nr >= ARRAY_SIZE(memmap_map)) {
+		pr_err("Ooops! Too many entries in the memory map!\n");
+		return;
+	}
+	memmap_map[memmap_nr].addr = addr;
+	memmap_map[memmap_nr].size = size;
+	memmap_nr++;
+}
+
+static int __init setup_memmap(char *p)
+{
+	char *oldp;
+	u64 start_at, mem_size;
+
+	if (!p)
+		return -EINVAL;
+
+	if (!strncmp(p, "exactmap", 8)) {
+		pr_err("\"memmap=exactmap\" not valid on tile\n");
+		return 0;
+	}
+
+	oldp = p;
+	mem_size = memparse(p, &p);
+	if (p == oldp)
+		return -EINVAL;
+
+	if (*p == '@') {
+		pr_err("\"memmap=nn@ss\" (force RAM) invalid on tile\n");
+	} else if (*p == '#') {
+		pr_err("\"memmap=nn#ss\" (force ACPI data) invalid on tile\n");
+	} else if (*p == '$') {
+		start_at = memparse(p+1, &p);
+		add_memmap_region(start_at, mem_size);
+	} else {
+		if (mem_size == 0)
+			return -EINVAL;
+		maxmem_pfn = (mem_size >> HPAGE_SHIFT) <<
+			(HPAGE_SHIFT - PAGE_SHIFT);
+	}
+	return *p == '\0' ? 0 : -EINVAL;
+}
+early_param("memmap", setup_memmap);
+
+static int __init setup_mem(char *str)
+{
+	return setup_maxmem(str);
+}
+early_param("mem", setup_mem);  /* compatibility with x86 */
+
+static int __init setup_isolnodes(char *str)
+{
+	if (str == NULL || nodelist_parse(str, isolnodes) != 0)
+		return -EINVAL;
+
+	pr_info("Set isolnodes value to '%*pbl'\n",
+		nodemask_pr_args(&isolnodes));
+	return 0;
+}
+early_param("isolnodes", setup_isolnodes);
+
+#if defined(CONFIG_PCI) && !defined(__tilegx__)
+static int __init setup_pci_reserve(char* str)
+{
+	if (str == NULL || kstrtouint(str, 0, &pci_reserve_mb) != 0 ||
+	    pci_reserve_mb > 3 * 1024)
+		return -EINVAL;
+
+	pr_info("Reserving %dMB for PCIE root complex mappings\n",
+		pci_reserve_mb);
+	return 0;
+}
+early_param("pci_reserve", setup_pci_reserve);
+#endif
+
+#ifndef __tilegx__
+/*
+ * vmalloc=size forces the vmalloc area to be exactly 'size' bytes.
+ * This can be used to increase (or decrease) the vmalloc area.
+ */
+static int __init parse_vmalloc(char *arg)
+{
+	if (!arg)
+		return -EINVAL;
+
+	VMALLOC_RESERVE = (memparse(arg, &arg) + PGDIR_SIZE - 1) & PGDIR_MASK;
+
+	/* See validate_va() for more on this test. */
+	if ((long)_VMALLOC_START >= 0)
+		early_panic("\"vmalloc=%#lx\" value too large: maximum %#lx\n",
+			    VMALLOC_RESERVE, _VMALLOC_END - 0x80000000UL);
+
+	return 0;
+}
+early_param("vmalloc", parse_vmalloc);
+#endif
+
+#ifdef CONFIG_HIGHMEM
+/*
+ * Determine for each controller where its lowmem is mapped and how much of
+ * it is mapped there.  On controller zero, the first few megabytes are
+ * already mapped in as code at MEM_SV_START, so in principle we could
+ * start our data mappings higher up, but for now we don't bother, to avoid
+ * additional confusion.
+ *
+ * One question is whether, on systems with more than 768 Mb and
+ * controllers of different sizes, to map in a proportionate amount of
+ * each one, or to try to map the same amount from each controller.
+ * (E.g. if we have three controllers with 256MB, 1GB, and 256MB
+ * respectively, do we map 256MB from each, or do we map 128 MB, 512
+ * MB, and 128 MB respectively?)  For now we use a proportionate
+ * solution like the latter.
+ *
+ * The VA/PA mapping demands that we align our decisions at 16 MB
+ * boundaries so that we can rapidly convert VA to PA.
+ */
+static void *__init setup_pa_va_mapping(void)
+{
+	unsigned long curr_pages = 0;
+	unsigned long vaddr = PAGE_OFFSET;
+	nodemask_t highonlynodes = isolnodes;
+	int i, j;
+
+	memset(pbase_map, -1, sizeof(pbase_map));
+	memset(vbase_map, -1, sizeof(vbase_map));
+
+	/* Node zero cannot be isolated for LOWMEM purposes. */
+	node_clear(0, highonlynodes);
+
+	/* Count up the number of pages on non-highonlynodes controllers. */
+	mappable_physpages = 0;
+	for_each_online_node(i) {
+		if (!node_isset(i, highonlynodes))
+			mappable_physpages +=
+				node_end_pfn[i] - node_start_pfn[i];
+	}
+
+	for_each_online_node(i) {
+		unsigned long start = node_start_pfn[i];
+		unsigned long end = node_end_pfn[i];
+		unsigned long size = end - start;
+		unsigned long vaddr_end;
+
+		if (node_isset(i, highonlynodes)) {
+			/* Mark this controller as having no lowmem. */
+			node_lowmem_end_pfn[i] = start;
+			continue;
+		}
+
+		curr_pages += size;
+		if (mappable_physpages > MAXMEM_PFN) {
+			vaddr_end = PAGE_OFFSET +
+				(((u64)curr_pages * MAXMEM_PFN /
+				  mappable_physpages)
+				 << PAGE_SHIFT);
+		} else {
+			vaddr_end = PAGE_OFFSET + (curr_pages << PAGE_SHIFT);
+		}
+		for (j = 0; vaddr < vaddr_end; vaddr += HPAGE_SIZE, ++j) {
+			unsigned long this_pfn =
+				start + (j << HUGETLB_PAGE_ORDER);
+			pbase_map[vaddr >> HPAGE_SHIFT] = this_pfn;
+			if (vbase_map[__pfn_to_highbits(this_pfn)] ==
+			    (void *)-1)
+				vbase_map[__pfn_to_highbits(this_pfn)] =
+					(void *)(vaddr & HPAGE_MASK);
+		}
+		node_lowmem_end_pfn[i] = start + (j << HUGETLB_PAGE_ORDER);
+		BUG_ON(node_lowmem_end_pfn[i] > end);
+	}
+
+	/* Return highest address of any mapped memory. */
+	return (void *)vaddr;
+}
+#endif /* CONFIG_HIGHMEM */
+
+/*
+ * Register our most important memory mappings with the debug stub.
+ *
+ * This is up to 4 mappings for lowmem, one mapping per memory
+ * controller, plus one for our text segment.
+ */
+static void store_permanent_mappings(void)
+{
+	int i;
+
+	for_each_online_node(i) {
+		HV_PhysAddr pa = ((HV_PhysAddr)node_start_pfn[i]) << PAGE_SHIFT;
+#ifdef CONFIG_HIGHMEM
+		HV_PhysAddr high_mapped_pa = node_lowmem_end_pfn[i];
+#else
+		HV_PhysAddr high_mapped_pa = node_end_pfn[i];
+#endif
+
+		unsigned long pages = high_mapped_pa - node_start_pfn[i];
+		HV_VirtAddr addr = (HV_VirtAddr) __va(pa);
+		hv_store_mapping(addr, pages << PAGE_SHIFT, pa);
+	}
+
+	hv_store_mapping((HV_VirtAddr)_text,
+			 (uint32_t)(_einittext - _text), 0);
+}
+
+/*
+ * Use hv_inquire_physical() to populate node_{start,end}_pfn[]
+ * and node_online_map, doing suitable sanity-checking.
+ * Also set min_low_pfn, max_low_pfn, and max_pfn.
+ */
+static void __init setup_memory(void)
+{
+	int i, j;
+	int highbits_seen[NR_PA_HIGHBIT_VALUES] = { 0 };
+#ifdef CONFIG_HIGHMEM
+	long highmem_pages;
+#endif
+#ifndef __tilegx__
+	int cap;
+#endif
+#if defined(CONFIG_HIGHMEM) || defined(__tilegx__)
+	long lowmem_pages;
+#endif
+	unsigned long physpages = 0;
+
+	/* We are using a char to hold the cpu_2_node[] mapping */
+	BUILD_BUG_ON(MAX_NUMNODES > 127);
+
+	/* Discover the ranges of memory available to us */
+	for (i = 0; ; ++i) {
+		unsigned long start, size, end, highbits;
+		HV_PhysAddrRange range = hv_inquire_physical(i);
+		if (range.size == 0)
+			break;
+#ifdef CONFIG_FLATMEM
+		if (i > 0) {
+			pr_err("Can't use discontiguous PAs: %#llx..%#llx\n",
+			       range.size, range.start + range.size);
+			continue;
+		}
+#endif
+#ifndef __tilegx__
+		if ((unsigned long)range.start) {
+			pr_err("Range not at 4GB multiple: %#llx..%#llx\n",
+			       range.start, range.start + range.size);
+			continue;
+		}
+#endif
+		if ((range.start & (HPAGE_SIZE-1)) != 0 ||
+		    (range.size & (HPAGE_SIZE-1)) != 0) {
+			unsigned long long start_pa = range.start;
+			unsigned long long orig_size = range.size;
+			range.start = (start_pa + HPAGE_SIZE - 1) & HPAGE_MASK;
+			range.size -= (range.start - start_pa);
+			range.size &= HPAGE_MASK;
+			pr_err("Range not hugepage-aligned: %#llx..%#llx: now %#llx-%#llx\n",
+			       start_pa, start_pa + orig_size,
+			       range.start, range.start + range.size);
+		}
+		highbits = __pa_to_highbits(range.start);
+		if (highbits >= NR_PA_HIGHBIT_VALUES) {
+			pr_err("PA high bits too high: %#llx..%#llx\n",
+			       range.start, range.start + range.size);
+			continue;
+		}
+		if (highbits_seen[highbits]) {
+			pr_err("Range overlaps in high bits: %#llx..%#llx\n",
+			       range.start, range.start + range.size);
+			continue;
+		}
+		highbits_seen[highbits] = 1;
+		if (PFN_DOWN(range.size) > maxnodemem_pfn[i]) {
+			int max_size = maxnodemem_pfn[i];
+			if (max_size > 0) {
+				pr_err("Maxnodemem reduced node %d to %d pages\n",
+				       i, max_size);
+				range.size = PFN_PHYS(max_size);
+			} else {
+				pr_err("Maxnodemem disabled node %d\n", i);
+				continue;
+			}
+		}
+		if (physpages + PFN_DOWN(range.size) > maxmem_pfn) {
+			int max_size = maxmem_pfn - physpages;
+			if (max_size > 0) {
+				pr_err("Maxmem reduced node %d to %d pages\n",
+				       i, max_size);
+				range.size = PFN_PHYS(max_size);
+			} else {
+				pr_err("Maxmem disabled node %d\n", i);
+				continue;
+			}
+		}
+		if (i >= MAX_NUMNODES) {
+			pr_err("Too many PA nodes (#%d): %#llx...%#llx\n",
+			       i, range.size, range.size + range.start);
+			continue;
+		}
+
+		start = range.start >> PAGE_SHIFT;
+		size = range.size >> PAGE_SHIFT;
+		end = start + size;
+
+#ifndef __tilegx__
+		if (((HV_PhysAddr)end << PAGE_SHIFT) !=
+		    (range.start + range.size)) {
+			pr_err("PAs too high to represent: %#llx..%#llx\n",
+			       range.start, range.start + range.size);
+			continue;
+		}
+#endif
+#if defined(CONFIG_PCI) && !defined(__tilegx__)
+		/*
+		 * Blocks that overlap the pci reserved region must
+		 * have enough space to hold the maximum percpu data
+		 * region at the top of the range.  If there isn't
+		 * enough space above the reserved region, just
+		 * truncate the node.
+		 */
+		if (start <= pci_reserve_start_pfn &&
+		    end > pci_reserve_start_pfn) {
+			unsigned int per_cpu_size =
+				__per_cpu_end - __per_cpu_start;
+			unsigned int percpu_pages =
+				NR_CPUS * (PFN_UP(per_cpu_size) >> PAGE_SHIFT);
+			if (end < pci_reserve_end_pfn + percpu_pages) {
+				end = pci_reserve_start_pfn;
+				pr_err("PCI mapping region reduced node %d to %ld pages\n",
+				       i, end - start);
+			}
+		}
+#endif
+
+		for (j = __pfn_to_highbits(start);
+		     j <= __pfn_to_highbits(end - 1); j++)
+			highbits_to_node[j] = i;
+
+		node_start_pfn[i] = start;
+		node_end_pfn[i] = end;
+		node_controller[i] = range.controller;
+		physpages += size;
+		max_pfn = end;
+
+		/* Mark node as online */
+		node_set(i, node_online_map);
+		node_set(i, node_possible_map);
+	}
+
+#ifndef __tilegx__
+	/*
+	 * For 4KB pages, mem_map "struct page" data is 1% of the size
+	 * of the physical memory, so can be quite big (640 MB for
+	 * four 16G zones).  These structures must be mapped in
+	 * lowmem, and since we currently cap out at about 768 MB,
+	 * it's impractical to try to use this much address space.
+	 * For now, arbitrarily cap the amount of physical memory
+	 * we're willing to use at 8 million pages (32GB of 4KB pages).
+	 */
+	cap = 8 * 1024 * 1024;  /* 8 million pages */
+	if (physpages > cap) {
+		int num_nodes = num_online_nodes();
+		int cap_each = cap / num_nodes;
+		unsigned long dropped_pages = 0;
+		for (i = 0; i < num_nodes; ++i) {
+			int size = node_end_pfn[i] - node_start_pfn[i];
+			if (size > cap_each) {
+				dropped_pages += (size - cap_each);
+				node_end_pfn[i] = node_start_pfn[i] + cap_each;
+			}
+		}
+		physpages -= dropped_pages;
+		pr_warn("Only using %ldMB memory - ignoring %ldMB\n",
+			physpages >> (20 - PAGE_SHIFT),
+			dropped_pages >> (20 - PAGE_SHIFT));
+		pr_warn("Consider using a larger page size\n");
+	}
+#endif
+
+	/* Heap starts just above the last loaded address. */
+	min_low_pfn = PFN_UP((unsigned long)_end - PAGE_OFFSET);
+
+#ifdef CONFIG_HIGHMEM
+	/* Find where we map lowmem from each controller. */
+	high_memory = setup_pa_va_mapping();
+
+	/* Set max_low_pfn based on what node 0 can directly address. */
+	max_low_pfn = node_lowmem_end_pfn[0];
+
+	lowmem_pages = (mappable_physpages > MAXMEM_PFN) ?
+		MAXMEM_PFN : mappable_physpages;
+	highmem_pages = (long) (physpages - lowmem_pages);
+
+	pr_notice("%ldMB HIGHMEM available\n",
+		  pages_to_mb(highmem_pages > 0 ? highmem_pages : 0));
+	pr_notice("%ldMB LOWMEM available\n", pages_to_mb(lowmem_pages));
+#else
+	/* Set max_low_pfn based on what node 0 can directly address. */
+	max_low_pfn = node_end_pfn[0];
+
+#ifndef __tilegx__
+	if (node_end_pfn[0] > MAXMEM_PFN) {
+		pr_warn("Only using %ldMB LOWMEM\n", MAXMEM >> 20);
+		pr_warn("Use a HIGHMEM enabled kernel\n");
+		max_low_pfn = MAXMEM_PFN;
+		max_pfn = MAXMEM_PFN;
+		node_end_pfn[0] = MAXMEM_PFN;
+	} else {
+		pr_notice("%ldMB memory available\n",
+			  pages_to_mb(node_end_pfn[0]));
+	}
+	for (i = 1; i < MAX_NUMNODES; ++i) {
+		node_start_pfn[i] = 0;
+		node_end_pfn[i] = 0;
+	}
+	high_memory = __va(node_end_pfn[0]);
+#else
+	lowmem_pages = 0;
+	for (i = 0; i < MAX_NUMNODES; ++i) {
+		int pages = node_end_pfn[i] - node_start_pfn[i];
+		lowmem_pages += pages;
+		if (pages)
+			high_memory = pfn_to_kaddr(node_end_pfn[i]);
+	}
+	pr_notice("%ldMB memory available\n", pages_to_mb(lowmem_pages));
+#endif
+#endif
+}
+
+/*
+ * On 32-bit machines, we only put bootmem on the low controller,
+ * since PAs > 4GB can't be used in bootmem.  In principle one could
+ * imagine, e.g., multiple 1 GB controllers all of which could support
+ * bootmem, but in practice using controllers this small isn't a
+ * particularly interesting scenario, so we just keep it simple and
+ * use only the first controller for bootmem on 32-bit machines.
+ */
+static inline int node_has_bootmem(int nid)
+{
+#ifdef CONFIG_64BIT
+	return 1;
+#else
+	return nid == 0;
+#endif
+}
+
+static inline unsigned long alloc_bootmem_pfn(int nid,
+					      unsigned long size,
+					      unsigned long goal)
+{
+	void *kva = __alloc_bootmem_node(NODE_DATA(nid), size,
+					 PAGE_SIZE, goal);
+	unsigned long pfn = kaddr_to_pfn(kva);
+	BUG_ON(goal && PFN_PHYS(pfn) != goal);
+	return pfn;
+}
+
+static void __init setup_bootmem_allocator_node(int i)
+{
+	unsigned long start, end, mapsize, mapstart;
+
+	if (node_has_bootmem(i)) {
+		NODE_DATA(i)->bdata = &bootmem_node_data[i];
+	} else {
+		/* Share controller zero's bdata for now. */
+		NODE_DATA(i)->bdata = &bootmem_node_data[0];
+		return;
+	}
+
+	/* Skip up to after the bss in node 0. */
+	start = (i == 0) ? min_low_pfn : node_start_pfn[i];
+
+	/* Only lowmem, if we're a HIGHMEM build. */
+#ifdef CONFIG_HIGHMEM
+	end = node_lowmem_end_pfn[i];
+#else
+	end = node_end_pfn[i];
+#endif
+
+	/* No memory here. */
+	if (end == start)
+		return;
+
+	/* Figure out where the bootmem bitmap is located. */
+	mapsize = bootmem_bootmap_pages(end - start);
+	if (i == 0) {
+		/* Use some space right before the heap on node 0. */
+		mapstart = start;
+		start += mapsize;
+	} else {
+		/* Allocate bitmap on node 0 to avoid page table issues. */
+		mapstart = alloc_bootmem_pfn(0, PFN_PHYS(mapsize), 0);
+	}
+
+	/* Initialize a node. */
+	init_bootmem_node(NODE_DATA(i), mapstart, start, end);
+
+	/* Free all the space back into the allocator. */
+	free_bootmem(PFN_PHYS(start), PFN_PHYS(end - start));
+
+#if defined(CONFIG_PCI) && !defined(__tilegx__)
+	/*
+	 * Throw away any memory aliased by the PCI region.
+	 */
+	if (pci_reserve_start_pfn < end && pci_reserve_end_pfn > start) {
+		start = max(pci_reserve_start_pfn, start);
+		end = min(pci_reserve_end_pfn, end);
+		reserve_bootmem(PFN_PHYS(start), PFN_PHYS(end - start),
+				BOOTMEM_EXCLUSIVE);
+	}
+#endif
+}
+
+static void __init setup_bootmem_allocator(void)
+{
+	int i;
+	for (i = 0; i < MAX_NUMNODES; ++i)
+		setup_bootmem_allocator_node(i);
+
+	/* Reserve any memory excluded by "memmap" arguments. */
+	for (i = 0; i < memmap_nr; ++i) {
+		struct memmap_entry *m = &memmap_map[i];
+		reserve_bootmem(m->addr, m->size, BOOTMEM_DEFAULT);
+	}
+
+#ifdef CONFIG_BLK_DEV_INITRD
+	if (initrd_start) {
+		/* Make sure the initrd memory region is not modified. */
+		if (reserve_bootmem(initrd_start, initrd_end - initrd_start,
+				    BOOTMEM_EXCLUSIVE)) {
+			pr_crit("The initrd memory region has been polluted. Disabling it.\n");
+			initrd_start = 0;
+			initrd_end = 0;
+		} else {
+			/*
+			 * Translate initrd_start & initrd_end from PA to VA for
+			 * future access.
+			 */
+			initrd_start += PAGE_OFFSET;
+			initrd_end += PAGE_OFFSET;
+		}
+	}
+#endif
+
+#ifdef CONFIG_KEXEC
+	if (crashk_res.start != crashk_res.end)
+		reserve_bootmem(crashk_res.start, resource_size(&crashk_res),
+				BOOTMEM_DEFAULT);
+#endif
+}
+
+void *__init alloc_remap(int nid, unsigned long size)
+{
+	int pages = node_end_pfn[nid] - node_start_pfn[nid];
+	void *map = pfn_to_kaddr(node_memmap_pfn[nid]);
+	BUG_ON(size != pages * sizeof(struct page));
+	memset(map, 0, size);
+	return map;
+}
+
+static int __init percpu_size(void)
+{
+	int size = __per_cpu_end - __per_cpu_start;
+	size += PERCPU_MODULE_RESERVE;
+	size += PERCPU_DYNAMIC_EARLY_SIZE;
+	if (size < PCPU_MIN_UNIT_SIZE)
+		size = PCPU_MIN_UNIT_SIZE;
+	size = roundup(size, PAGE_SIZE);
+
+	/* In several places we assume the per-cpu data fits on a huge page. */
+	BUG_ON(kdata_huge && size > HPAGE_SIZE);
+	return size;
+}
+
+static void __init zone_sizes_init(void)
+{
+	unsigned long zones_size[MAX_NR_ZONES] = { 0 };
+	int size = percpu_size();
+	int num_cpus = smp_height * smp_width;
+	const unsigned long dma_end = (1UL << (32 - PAGE_SHIFT));
+
+	int i;
+
+	for (i = 0; i < num_cpus; ++i)
+		node_percpu[cpu_to_node(i)] += size;
+
+	for_each_online_node(i) {
+		unsigned long start = node_start_pfn[i];
+		unsigned long end = node_end_pfn[i];
+#ifdef CONFIG_HIGHMEM
+		unsigned long lowmem_end = node_lowmem_end_pfn[i];
+#else
+		unsigned long lowmem_end = end;
+#endif
+		int memmap_size = (end - start) * sizeof(struct page);
+		node_free_pfn[i] = start;
+
+		/*
+		 * Set aside pages for per-cpu data and the mem_map array.
+		 *
+		 * Since the per-cpu data requires special homecaching,
+		 * if we are in kdata_huge mode, we put it at the end of
+		 * the lowmem region.  If we're not in kdata_huge mode,
+		 * we take the per-cpu pages from the bottom of the
+		 * controller, since that avoids fragmenting a huge page
+		 * that users might want.  We always take the memmap
+		 * from the bottom of the controller, since with
+		 * kdata_huge that lets it be under a huge TLB entry.
+		 *
+		 * If the user has requested isolnodes for a controller,
+		 * though, there'll be no lowmem, so we just alloc_bootmem
+		 * the memmap.  There will be no percpu memory either.
+		 */
+		if (i != 0 && node_isset(i, isolnodes)) {
+			node_memmap_pfn[i] =
+				alloc_bootmem_pfn(0, memmap_size, 0);
+			BUG_ON(node_percpu[i] != 0);
+		} else if (node_has_bootmem(start)) {
+			unsigned long goal = 0;
+			node_memmap_pfn[i] =
+				alloc_bootmem_pfn(i, memmap_size, 0);
+			if (kdata_huge)
+				goal = PFN_PHYS(lowmem_end) - node_percpu[i];
+			if (node_percpu[i])
+				node_percpu_pfn[i] =
+					alloc_bootmem_pfn(i, node_percpu[i],
+							  goal);
+		} else {
+			/* In non-bootmem zones, just reserve some pages. */
+			node_memmap_pfn[i] = node_free_pfn[i];
+			node_free_pfn[i] += PFN_UP(memmap_size);
+			if (!kdata_huge) {
+				node_percpu_pfn[i] = node_free_pfn[i];
+				node_free_pfn[i] += PFN_UP(node_percpu[i]);
+			} else {
+				node_percpu_pfn[i] =
+					lowmem_end - PFN_UP(node_percpu[i]);
+			}
+		}
+
+#ifdef CONFIG_HIGHMEM
+		if (start > lowmem_end) {
+			zones_size[ZONE_NORMAL] = 0;
+			zones_size[ZONE_HIGHMEM] = end - start;
+		} else {
+			zones_size[ZONE_NORMAL] = lowmem_end - start;
+			zones_size[ZONE_HIGHMEM] = end - lowmem_end;
+		}
+#else
+		zones_size[ZONE_NORMAL] = end - start;
+#endif
+
+		if (start < dma_end) {
+			zones_size[ZONE_DMA] = min(zones_size[ZONE_NORMAL],
+						   dma_end - start);
+			zones_size[ZONE_NORMAL] -= zones_size[ZONE_DMA];
+		} else {
+			zones_size[ZONE_DMA] = 0;
+		}
+
+		/* Take zone metadata from controller 0 if we're isolnode. */
+		if (node_isset(i, isolnodes))
+			NODE_DATA(i)->bdata = &bootmem_node_data[0];
+
+		free_area_init_node(i, zones_size, start, NULL);
+		printk(KERN_DEBUG "  Normal zone: %ld per-cpu pages\n",
+		       PFN_UP(node_percpu[i]));
+
+		/* Track the type of memory on each node */
+		if (zones_size[ZONE_NORMAL] || zones_size[ZONE_DMA])
+			node_set_state(i, N_NORMAL_MEMORY);
+#ifdef CONFIG_HIGHMEM
+		if (end != start)
+			node_set_state(i, N_HIGH_MEMORY);
+#endif
+
+		node_set_online(i);
+	}
+}
+
+#ifdef CONFIG_NUMA
+
+/* which logical CPUs are on which nodes */
+struct cpumask node_2_cpu_mask[MAX_NUMNODES] __write_once;
+EXPORT_SYMBOL(node_2_cpu_mask);
+
+/* which node each logical CPU is on */
+char cpu_2_node[NR_CPUS] __write_once __attribute__((aligned(L2_CACHE_BYTES)));
+EXPORT_SYMBOL(cpu_2_node);
+
+/* Return cpu_to_node() except for cpus not yet assigned, which return -1 */
+static int __init cpu_to_bound_node(int cpu, struct cpumask* unbound_cpus)
+{
+	if (!cpu_possible(cpu) || cpumask_test_cpu(cpu, unbound_cpus))
+		return -1;
+	else
+		return cpu_to_node(cpu);
+}
+
+/* Return number of immediately-adjacent tiles sharing the same NUMA node. */
+static int __init node_neighbors(int node, int cpu,
+				 struct cpumask *unbound_cpus)
+{
+	int neighbors = 0;
+	int w = smp_width;
+	int h = smp_height;
+	int x = cpu % w;
+	int y = cpu / w;
+	if (x > 0 && cpu_to_bound_node(cpu-1, unbound_cpus) == node)
+		++neighbors;
+	if (x < w-1 && cpu_to_bound_node(cpu+1, unbound_cpus) == node)
+		++neighbors;
+	if (y > 0 && cpu_to_bound_node(cpu-w, unbound_cpus) == node)
+		++neighbors;
+	if (y < h-1 && cpu_to_bound_node(cpu+w, unbound_cpus) == node)
+		++neighbors;
+	return neighbors;
+}
+
+static void __init setup_numa_mapping(void)
+{
+	int distance[MAX_NUMNODES][NR_CPUS];
+	HV_Coord coord;
+	int cpu, node, cpus, i, x, y;
+	int num_nodes = num_online_nodes();
+	struct cpumask unbound_cpus;
+	nodemask_t default_nodes;
+
+	cpumask_clear(&unbound_cpus);
+
+	/* Get set of nodes we will use for defaults */
+	nodes_andnot(default_nodes, node_online_map, isolnodes);
+	if (nodes_empty(default_nodes)) {
+		BUG_ON(!node_isset(0, node_online_map));
+		pr_err("Forcing NUMA node zero available as a default node\n");
+		node_set(0, default_nodes);
+	}
+
+	/* Populate the distance[] array */
+	memset(distance, -1, sizeof(distance));
+	cpu = 0;
+	for (coord.y = 0; coord.y < smp_height; ++coord.y) {
+		for (coord.x = 0; coord.x < smp_width;
+		     ++coord.x, ++cpu) {
+			BUG_ON(cpu >= nr_cpu_ids);
+			if (!cpu_possible(cpu)) {
+				cpu_2_node[cpu] = -1;
+				continue;
+			}
+			for_each_node_mask(node, default_nodes) {
+				HV_MemoryControllerInfo info =
+					hv_inquire_memory_controller(
+						coord, node_controller[node]);
+				distance[node][cpu] =
+					ABS(info.coord.x) + ABS(info.coord.y);
+			}
+			cpumask_set_cpu(cpu, &unbound_cpus);
+		}
+	}
+	cpus = cpu;
+
+	/*
+	 * Round-robin through the NUMA nodes until all the cpus are
+	 * assigned.  We could be more clever here (e.g. create four
+	 * sorted linked lists on the same set of cpu nodes, and pull
+	 * off them in round-robin sequence, removing from all four
+	 * lists each time) but given the relatively small numbers
+	 * involved, O(n^2) seem OK for a one-time cost.
+	 */
+	node = first_node(default_nodes);
+	while (!cpumask_empty(&unbound_cpus)) {
+		int best_cpu = -1;
+		int best_distance = INT_MAX;
+		for (cpu = 0; cpu < cpus; ++cpu) {
+			if (cpumask_test_cpu(cpu, &unbound_cpus)) {
+				/*
+				 * Compute metric, which is how much
+				 * closer the cpu is to this memory
+				 * controller than the others, shifted
+				 * up, and then the number of
+				 * neighbors already in the node as an
+				 * epsilon adjustment to try to keep
+				 * the nodes compact.
+				 */
+				int d = distance[node][cpu] * num_nodes;
+				for_each_node_mask(i, default_nodes) {
+					if (i != node)
+						d -= distance[i][cpu];
+				}
+				d *= 8;  /* allow space for epsilon */
+				d -= node_neighbors(node, cpu, &unbound_cpus);
+				if (d < best_distance) {
+					best_cpu = cpu;
+					best_distance = d;
+				}
+			}
+		}
+		BUG_ON(best_cpu < 0);
+		cpumask_set_cpu(best_cpu, &node_2_cpu_mask[node]);
+		cpu_2_node[best_cpu] = node;
+		cpumask_clear_cpu(best_cpu, &unbound_cpus);
+		node = next_node(node, default_nodes);
+		if (node == MAX_NUMNODES)
+			node = first_node(default_nodes);
+	}
+
+	/* Print out node assignments and set defaults for disabled cpus */
+	cpu = 0;
+	for (y = 0; y < smp_height; ++y) {
+		printk(KERN_DEBUG "NUMA cpu-to-node row %d:", y);
+		for (x = 0; x < smp_width; ++x, ++cpu) {
+			if (cpu_to_node(cpu) < 0) {
+				pr_cont(" -");
+				cpu_2_node[cpu] = first_node(default_nodes);
+			} else {
+				pr_cont(" %d", cpu_to_node(cpu));
+			}
+		}
+		pr_cont("\n");
+	}
+}
+
+static struct cpu cpu_devices[NR_CPUS];
+
+static int __init topology_init(void)
+{
+	int i;
+
+	for_each_online_node(i)
+		register_one_node(i);
+
+	for (i = 0; i < smp_height * smp_width; ++i)
+		register_cpu(&cpu_devices[i], i);
+
+	return 0;
+}
+
+subsys_initcall(topology_init);
+
+#else /* !CONFIG_NUMA */
+
+#define setup_numa_mapping() do { } while (0)
+
+#endif /* CONFIG_NUMA */
+
+/*
+ * Initialize hugepage support on this cpu.  We do this on all cores
+ * early in boot: before argument parsing for the boot cpu, and after
+ * argument parsing but before the init functions run on the secondaries.
+ * So the values we set up here in the hypervisor may be overridden on
+ * the boot cpu as arguments are parsed.
+ */
+static void init_super_pages(void)
+{
+#ifdef CONFIG_HUGETLB_SUPER_PAGES
+	int i;
+	for (i = 0; i < HUGE_SHIFT_ENTRIES; ++i)
+		hv_set_pte_super_shift(i, huge_shift[i]);
+#endif
+}
+
+/**
+ * setup_cpu() - Do all necessary per-cpu, tile-specific initialization.
+ * @boot: Is this the boot cpu?
+ *
+ * Called from setup_arch() on the boot cpu, or online_secondary().
+ */
+void setup_cpu(int boot)
+{
+	/* The boot cpu sets up its permanent mappings much earlier. */
+	if (!boot)
+		store_permanent_mappings();
+
+	/* Allow asynchronous TLB interrupts. */
+#if CHIP_HAS_TILE_DMA()
+	arch_local_irq_unmask(INT_DMATLB_MISS);
+	arch_local_irq_unmask(INT_DMATLB_ACCESS);
+#endif
+#ifdef __tilegx__
+	arch_local_irq_unmask(INT_SINGLE_STEP_K);
+#endif
+
+	/*
+	 * Allow user access to many generic SPRs, like the cycle
+	 * counter, PASS/FAIL/DONE, INTERRUPT_CRITICAL_SECTION, etc.
+	 */
+	__insn_mtspr(SPR_MPL_WORLD_ACCESS_SET_0, 1);
+
+#if CHIP_HAS_SN()
+	/* Static network is not restricted. */
+	__insn_mtspr(SPR_MPL_SN_ACCESS_SET_0, 1);
+#endif
+
+	/*
+	 * Set the MPL for interrupt control 0 & 1 to the corresponding
+	 * values.  This includes access to the SYSTEM_SAVE and EX_CONTEXT
+	 * SPRs, as well as the interrupt mask.
+	 */
+	__insn_mtspr(SPR_MPL_INTCTRL_0_SET_0, 1);
+	__insn_mtspr(SPR_MPL_INTCTRL_1_SET_1, 1);
+
+	/* Initialize IRQ support for this cpu. */
+	setup_irq_regs();
+
+#ifdef CONFIG_HARDWALL
+	/* Reset the network state on this cpu. */
+	reset_network_state();
+#endif
+
+	init_super_pages();
+}
+
+#ifdef CONFIG_BLK_DEV_INITRD
+
+static int __initdata set_initramfs_file;
+static char __initdata initramfs_file[128] = "initramfs";
+
+static int __init setup_initramfs_file(char *str)
+{
+	if (str == NULL)
+		return -EINVAL;
+	strncpy(initramfs_file, str, sizeof(initramfs_file) - 1);
+	set_initramfs_file = 1;
+
+	return 0;
+}
+early_param("initramfs_file", setup_initramfs_file);
+
+/*
+ * We look for a file called "initramfs" in the hvfs.  If there is one, we
+ * allocate some memory for it and it will be unpacked to the initramfs.
+ * If it's compressed, the initd code will uncompress it first.
+ */
+static void __init load_hv_initrd(void)
+{
+	HV_FS_StatInfo stat;
+	int fd, rc;
+	void *initrd;
+
+	/* If initrd has already been set, skip initramfs file in hvfs. */
+	if (initrd_start)
+		return;
+
+	fd = hv_fs_findfile((HV_VirtAddr) initramfs_file);
+	if (fd == HV_ENOENT) {
+		if (set_initramfs_file) {
+			pr_warn("No such hvfs initramfs file '%s'\n",
+				initramfs_file);
+			return;
+		} else {
+			/* Try old backwards-compatible name. */
+			fd = hv_fs_findfile((HV_VirtAddr)"initramfs.cpio.gz");
+			if (fd == HV_ENOENT)
+				return;
+		}
+	}
+	BUG_ON(fd < 0);
+	stat = hv_fs_fstat(fd);
+	BUG_ON(stat.size < 0);
+	if (stat.flags & HV_FS_ISDIR) {
+		pr_warn("Ignoring hvfs file '%s': it's a directory\n",
+			initramfs_file);
+		return;
+	}
+	initrd = alloc_bootmem_pages(stat.size);
+	rc = hv_fs_pread(fd, (HV_VirtAddr) initrd, stat.size, 0);
+	if (rc != stat.size) {
+		pr_err("Error reading %d bytes from hvfs file '%s': %d\n",
+		       stat.size, initramfs_file, rc);
+		free_initrd_mem((unsigned long) initrd, stat.size);
+		return;
+	}
+	initrd_start = (unsigned long) initrd;
+	initrd_end = initrd_start + stat.size;
+}
+
+void __init free_initrd_mem(unsigned long begin, unsigned long end)
+{
+	free_bootmem(__pa(begin), end - begin);
+}
+
+static int __init setup_initrd(char *str)
+{
+	char *endp;
+	unsigned long initrd_size;
+
+	initrd_size = str ? simple_strtoul(str, &endp, 0) : 0;
+	if (initrd_size == 0 || *endp != '@')
+		return -EINVAL;
+
+	initrd_start = simple_strtoul(endp+1, &endp, 0);
+	if (initrd_start == 0)
+		return -EINVAL;
+
+	initrd_end = initrd_start + initrd_size;
+
+	return 0;
+}
+early_param("initrd", setup_initrd);
+
+#else
+static inline void load_hv_initrd(void) {}
+#endif /* CONFIG_BLK_DEV_INITRD */
+
+static void __init validate_hv(void)
+{
+	/*
+	 * It may already be too late, but let's check our built-in
+	 * configuration against what the hypervisor is providing.
+	 */
+	unsigned long glue_size = hv_sysconf(HV_SYSCONF_GLUE_SIZE);
+	int hv_page_size = hv_sysconf(HV_SYSCONF_PAGE_SIZE_SMALL);
+	int hv_hpage_size = hv_sysconf(HV_SYSCONF_PAGE_SIZE_LARGE);
+	HV_ASIDRange asid_range;
+
+#ifndef CONFIG_SMP
+	HV_Topology topology = hv_inquire_topology();
+	BUG_ON(topology.coord.x != 0 || topology.coord.y != 0);
+	if (topology.width != 1 || topology.height != 1) {
+		pr_warn("Warning: booting UP kernel on %dx%d grid; will ignore all but first tile\n",
+			topology.width, topology.height);
+	}
+#endif
+
+	if (PAGE_OFFSET + HV_GLUE_START_CPA + glue_size > (unsigned long)_text)
+		early_panic("Hypervisor glue size %ld is too big!\n",
+			    glue_size);
+	if (hv_page_size != PAGE_SIZE)
+		early_panic("Hypervisor page size %#x != our %#lx\n",
+			    hv_page_size, PAGE_SIZE);
+	if (hv_hpage_size != HPAGE_SIZE)
+		early_panic("Hypervisor huge page size %#x != our %#lx\n",
+			    hv_hpage_size, HPAGE_SIZE);
+
+#ifdef CONFIG_SMP
+	/*
+	 * Some hypervisor APIs take a pointer to a bitmap array
+	 * whose size is at least the number of cpus on the chip.
+	 * We use a struct cpumask for this, so it must be big enough.
+	 */
+	if ((smp_height * smp_width) > nr_cpu_ids)
+		early_panic("Hypervisor %d x %d grid too big for Linux NR_CPUS %d\n",
+			    smp_height, smp_width, nr_cpu_ids);
+#endif
+
+	/*
+	 * Check that we're using allowed ASIDs, and initialize the
+	 * various asid variables to their appropriate initial states.
+	 */
+	asid_range = hv_inquire_asid(0);
+	min_asid = asid_range.start;
+	__this_cpu_write(current_asid, min_asid);
+	max_asid = asid_range.start + asid_range.size - 1;
+
+	if (hv_confstr(HV_CONFSTR_CHIP_MODEL, (HV_VirtAddr)chip_model,
+		       sizeof(chip_model)) < 0) {
+		pr_err("Warning: HV_CONFSTR_CHIP_MODEL not available\n");
+		strlcpy(chip_model, "unknown", sizeof(chip_model));
+	}
+}
+
+static void __init validate_va(void)
+{
+#ifndef __tilegx__   /* FIXME: GX: probably some validation relevant here */
+	/*
+	 * Similarly, make sure we're only using allowed VAs.
+	 * We assume we can contiguously use MEM_USER_INTRPT .. MEM_HV_START,
+	 * and 0 .. KERNEL_HIGH_VADDR.
+	 * In addition, make sure we CAN'T use the end of memory, since
+	 * we use the last chunk of each pgd for the pgd_list.
+	 */
+	int i, user_kernel_ok = 0;
+	unsigned long max_va = 0;
+	unsigned long list_va =
+		((PGD_LIST_OFFSET / sizeof(pgd_t)) << PGDIR_SHIFT);
+
+	for (i = 0; ; ++i) {
+		HV_VirtAddrRange range = hv_inquire_virtual(i);
+		if (range.size == 0)
+			break;
+		if (range.start <= MEM_USER_INTRPT &&
+		    range.start + range.size >= MEM_HV_START)
+			user_kernel_ok = 1;
+		if (range.start == 0)
+			max_va = range.size;
+		BUG_ON(range.start + range.size > list_va);
+	}
+	if (!user_kernel_ok)
+		early_panic("Hypervisor not configured for user/kernel VAs\n");
+	if (max_va == 0)
+		early_panic("Hypervisor not configured for low VAs\n");
+	if (max_va < KERNEL_HIGH_VADDR)
+		early_panic("Hypervisor max VA %#lx smaller than %#lx\n",
+			    max_va, KERNEL_HIGH_VADDR);
+
+	/* Kernel PCs must have their high bit set; see intvec.S. */
+	if ((long)VMALLOC_START >= 0)
+		early_panic("Linux VMALLOC region below the 2GB line (%#lx)!\n"
+			    "Reconfigure the kernel with smaller VMALLOC_RESERVE\n",
+			    VMALLOC_START);
+#endif
+}
+
+/*
+ * cpu_lotar_map lists all the cpus that are valid for the supervisor
+ * to cache data on at a page level, i.e. what cpus can be placed in
+ * the LOTAR field of a PTE.  It is equivalent to the set of possible
+ * cpus plus any other cpus that are willing to share their cache.
+ * It is set by hv_inquire_tiles(HV_INQ_TILES_LOTAR).
+ */
+struct cpumask __write_once cpu_lotar_map;
+EXPORT_SYMBOL(cpu_lotar_map);
+
+/*
+ * hash_for_home_map lists all the tiles that hash-for-home data
+ * will be cached on.  Note that this may includes tiles that are not
+ * valid for this supervisor to use otherwise (e.g. if a hypervisor
+ * device is being shared between multiple supervisors).
+ * It is set by hv_inquire_tiles(HV_INQ_TILES_HFH_CACHE).
+ */
+struct cpumask hash_for_home_map;
+EXPORT_SYMBOL(hash_for_home_map);
+
+/*
+ * cpu_cacheable_map lists all the cpus whose caches the hypervisor can
+ * flush on our behalf.  It is set to cpu_possible_mask OR'ed with
+ * hash_for_home_map, and it is what should be passed to
+ * hv_flush_remote() to flush all caches.  Note that if there are
+ * dedicated hypervisor driver tiles that have authorized use of their
+ * cache, those tiles will only appear in cpu_lotar_map, NOT in
+ * cpu_cacheable_map, as they are a special case.
+ */
+struct cpumask __write_once cpu_cacheable_map;
+EXPORT_SYMBOL(cpu_cacheable_map);
+
+static __initdata struct cpumask disabled_map;
+
+static int __init disabled_cpus(char *str)
+{
+	int boot_cpu = smp_processor_id();
+
+	if (str == NULL || cpulist_parse_crop(str, &disabled_map) != 0)
+		return -EINVAL;
+	if (cpumask_test_cpu(boot_cpu, &disabled_map)) {
+		pr_err("disabled_cpus: can't disable boot cpu %d\n", boot_cpu);
+		cpumask_clear_cpu(boot_cpu, &disabled_map);
+	}
+	return 0;
+}
+
+early_param("disabled_cpus", disabled_cpus);
+
+void __init print_disabled_cpus(void)
+{
+	if (!cpumask_empty(&disabled_map))
+		pr_info("CPUs not available for Linux: %*pbl\n",
+			cpumask_pr_args(&disabled_map));
+}
+
+static void __init setup_cpu_maps(void)
+{
+	struct cpumask hv_disabled_map, cpu_possible_init;
+	int boot_cpu = smp_processor_id();
+	int cpus, i, rc;
+
+	/* Learn which cpus are allowed by the hypervisor. */
+	rc = hv_inquire_tiles(HV_INQ_TILES_AVAIL,
+			      (HV_VirtAddr) cpumask_bits(&cpu_possible_init),
+			      sizeof(cpu_cacheable_map));
+	if (rc < 0)
+		early_panic("hv_inquire_tiles(AVAIL) failed: rc %d\n", rc);
+	if (!cpumask_test_cpu(boot_cpu, &cpu_possible_init))
+		early_panic("Boot CPU %d disabled by hypervisor!\n", boot_cpu);
+
+	/* Compute the cpus disabled by the hvconfig file. */
+	cpumask_complement(&hv_disabled_map, &cpu_possible_init);
+
+	/* Include them with the cpus disabled by "disabled_cpus". */
+	cpumask_or(&disabled_map, &disabled_map, &hv_disabled_map);
+
+	/*
+	 * Disable every cpu after "setup_max_cpus".  But don't mark
+	 * as disabled the cpus that are outside of our initial rectangle,
+	 * since that turns out to be confusing.
+	 */
+	cpus = 1;                          /* this cpu */
+	cpumask_set_cpu(boot_cpu, &disabled_map);   /* ignore this cpu */
+	for (i = 0; cpus < setup_max_cpus; ++i)
+		if (!cpumask_test_cpu(i, &disabled_map))
+			++cpus;
+	for (; i < smp_height * smp_width; ++i)
+		cpumask_set_cpu(i, &disabled_map);
+	cpumask_clear_cpu(boot_cpu, &disabled_map); /* reset this cpu */
+	for (i = smp_height * smp_width; i < NR_CPUS; ++i)
+		cpumask_clear_cpu(i, &disabled_map);
+
+	/*
+	 * Setup cpu_possible map as every cpu allocated to us, minus
+	 * the results of any "disabled_cpus" settings.
+	 */
+	cpumask_andnot(&cpu_possible_init, &cpu_possible_init, &disabled_map);
+	init_cpu_possible(&cpu_possible_init);
+
+	/* Learn which cpus are valid for LOTAR caching. */
+	rc = hv_inquire_tiles(HV_INQ_TILES_LOTAR,
+			      (HV_VirtAddr) cpumask_bits(&cpu_lotar_map),
+			      sizeof(cpu_lotar_map));
+	if (rc < 0) {
+		pr_err("warning: no HV_INQ_TILES_LOTAR; using AVAIL\n");
+		cpu_lotar_map = *cpu_possible_mask;
+	}
+
+	/* Retrieve set of CPUs used for hash-for-home caching */
+	rc = hv_inquire_tiles(HV_INQ_TILES_HFH_CACHE,
+			      (HV_VirtAddr) hash_for_home_map.bits,
+			      sizeof(hash_for_home_map));
+	if (rc < 0)
+		early_panic("hv_inquire_tiles(HFH_CACHE) failed: rc %d\n", rc);
+	cpumask_or(&cpu_cacheable_map, cpu_possible_mask, &hash_for_home_map);
+}
+
+
+static int __init dataplane(char *str)
+{
+	pr_warn("WARNING: dataplane support disabled in this kernel\n");
+	return 0;
+}
+
+early_param("dataplane", dataplane);
+
+#ifdef CONFIG_NO_HZ_FULL
+/* Warn if hypervisor shared cpus are marked as nohz_full. */
+static int __init check_nohz_full_cpus(void)
+{
+	struct cpumask shared;
+	int cpu;
+
+	if (hv_inquire_tiles(HV_INQ_TILES_SHARED,
+			     (HV_VirtAddr) shared.bits, sizeof(shared)) < 0) {
+		pr_warn("WARNING: No support for inquiring hv shared tiles\n");
+		return 0;
+	}
+	for_each_cpu(cpu, &shared) {
+		if (tick_nohz_full_cpu(cpu))
+			pr_warn("WARNING: nohz_full cpu %d receives hypervisor interrupts!\n",
+			       cpu);
+	}
+	return 0;
+}
+arch_initcall(check_nohz_full_cpus);
+#endif
+
+#ifdef CONFIG_CMDLINE_BOOL
+static char __initdata builtin_cmdline[COMMAND_LINE_SIZE] = CONFIG_CMDLINE;
+#endif
+
+void __init setup_arch(char **cmdline_p)
+{
+	int len;
+
+#if defined(CONFIG_CMDLINE_BOOL) && defined(CONFIG_CMDLINE_OVERRIDE)
+	len = hv_get_command_line((HV_VirtAddr) boot_command_line,
+				  COMMAND_LINE_SIZE);
+	if (boot_command_line[0])
+		pr_warn("WARNING: ignoring dynamic command line \"%s\"\n",
+			boot_command_line);
+	strlcpy(boot_command_line, builtin_cmdline, COMMAND_LINE_SIZE);
+#else
+	char *hv_cmdline;
+#if defined(CONFIG_CMDLINE_BOOL)
+	if (builtin_cmdline[0]) {
+		int builtin_len = strlcpy(boot_command_line, builtin_cmdline,
+					  COMMAND_LINE_SIZE);
+		if (builtin_len < COMMAND_LINE_SIZE-1)
+			boot_command_line[builtin_len++] = ' ';
+		hv_cmdline = &boot_command_line[builtin_len];
+		len = COMMAND_LINE_SIZE - builtin_len;
+	} else
+#endif
+	{
+		hv_cmdline = boot_command_line;
+		len = COMMAND_LINE_SIZE;
+	}
+	len = hv_get_command_line((HV_VirtAddr) hv_cmdline, len);
+	if (len < 0 || len > COMMAND_LINE_SIZE)
+		early_panic("hv_get_command_line failed: %d\n", len);
+#endif
+
+	*cmdline_p = boot_command_line;
+
+	/* Set disabled_map and setup_max_cpus very early */
+	parse_early_param();
+
+	/* Make sure the kernel is compatible with the hypervisor. */
+	validate_hv();
+	validate_va();
+
+	setup_cpu_maps();
+
+
+#if defined(CONFIG_PCI) && !defined(__tilegx__)
+	/*
+	 * Initialize the PCI structures.  This is done before memory
+	 * setup so that we know whether or not a pci_reserve region
+	 * is necessary.
+	 */
+	if (tile_pci_init() == 0)
+		pci_reserve_mb = 0;
+
+	/* PCI systems reserve a region just below 4GB for mapping iomem. */
+	pci_reserve_end_pfn  = (1 << (32 - PAGE_SHIFT));
+	pci_reserve_start_pfn = pci_reserve_end_pfn -
+		(pci_reserve_mb << (20 - PAGE_SHIFT));
+#endif
+
+	init_mm.start_code = (unsigned long) _text;
+	init_mm.end_code = (unsigned long) _etext;
+	init_mm.end_data = (unsigned long) _edata;
+	init_mm.brk = (unsigned long) _end;
+
+	setup_memory();
+	store_permanent_mappings();
+	setup_bootmem_allocator();
+
+	/*
+	 * NOTE: before this point _nobody_ is allowed to allocate
+	 * any memory using the bootmem allocator.
+	 */
+
+#ifdef CONFIG_SWIOTLB
+	swiotlb_init(0);
+#endif
+
+	paging_init();
+	setup_numa_mapping();
+	zone_sizes_init();
+	set_page_homes();
+	setup_cpu(1);
+	setup_clock();
+	load_hv_initrd();
+}
+
+
+/*
+ * Set up per-cpu memory.
+ */
+
+unsigned long __per_cpu_offset[NR_CPUS] __write_once;
+EXPORT_SYMBOL(__per_cpu_offset);
+
+static size_t __initdata pfn_offset[MAX_NUMNODES] = { 0 };
+static unsigned long __initdata percpu_pfn[NR_CPUS] = { 0 };
+
+/*
+ * As the percpu code allocates pages, we return the pages from the
+ * end of the node for the specified cpu.
+ */
+static void *__init pcpu_fc_alloc(unsigned int cpu, size_t size, size_t align)
+{
+	int nid = cpu_to_node(cpu);
+	unsigned long pfn = node_percpu_pfn[nid] + pfn_offset[nid];
+
+	BUG_ON(size % PAGE_SIZE != 0);
+	pfn_offset[nid] += size / PAGE_SIZE;
+	BUG_ON(node_percpu[nid] < size);
+	node_percpu[nid] -= size;
+	if (percpu_pfn[cpu] == 0)
+		percpu_pfn[cpu] = pfn;
+	return pfn_to_kaddr(pfn);
+}
+
+/*
+ * Pages reserved for percpu memory are not freeable, and in any case we are
+ * on a short path to panic() in setup_per_cpu_area() at this point anyway.
+ */
+static void __init pcpu_fc_free(void *ptr, size_t size)
+{
+}
+
+/*
+ * Set up vmalloc page tables using bootmem for the percpu code.
+ */
+static void __init pcpu_fc_populate_pte(unsigned long addr)
+{
+	pgd_t *pgd;
+	pud_t *pud;
+	pmd_t *pmd;
+	pte_t *pte;
+
+	BUG_ON(pgd_addr_invalid(addr));
+	if (addr < VMALLOC_START || addr >= VMALLOC_END)
+		panic("PCPU addr %#lx outside vmalloc range %#lx..%#lx; try increasing CONFIG_VMALLOC_RESERVE\n",
+		      addr, VMALLOC_START, VMALLOC_END);
+
+	pgd = swapper_pg_dir + pgd_index(addr);
+	pud = pud_offset(pgd, addr);
+	BUG_ON(!pud_present(*pud));
+	pmd = pmd_offset(pud, addr);
+	if (pmd_present(*pmd)) {
+		BUG_ON(pmd_huge_page(*pmd));
+	} else {
+		pte = __alloc_bootmem(L2_KERNEL_PGTABLE_SIZE,
+				      HV_PAGE_TABLE_ALIGN, 0);
+		pmd_populate_kernel(&init_mm, pmd, pte);
+	}
+}
+
+void __init setup_per_cpu_areas(void)
+{
+	struct page *pg;
+	unsigned long delta, pfn, lowmem_va;
+	unsigned long size = percpu_size();
+	char *ptr;
+	int rc, cpu, i;
+
+	rc = pcpu_page_first_chunk(PERCPU_MODULE_RESERVE, pcpu_fc_alloc,
+				   pcpu_fc_free, pcpu_fc_populate_pte);
+	if (rc < 0)
+		panic("Cannot initialize percpu area (err=%d)", rc);
+
+	delta = (unsigned long)pcpu_base_addr - (unsigned long)__per_cpu_start;
+	for_each_possible_cpu(cpu) {
+		__per_cpu_offset[cpu] = delta + pcpu_unit_offsets[cpu];
+
+		/* finv the copy out of cache so we can change homecache */
+		ptr = pcpu_base_addr + pcpu_unit_offsets[cpu];
+		__finv_buffer(ptr, size);
+		pfn = percpu_pfn[cpu];
+
+		/* Rewrite the page tables to cache on that cpu */
+		pg = pfn_to_page(pfn);
+		for (i = 0; i < size; i += PAGE_SIZE, ++pfn, ++pg) {
+
+			/* Update the vmalloc mapping and page home. */
+			unsigned long addr = (unsigned long)ptr + i;
+			pte_t *ptep = virt_to_kpte(addr);
+			pte_t pte = *ptep;
+			BUG_ON(pfn != pte_pfn(pte));
+			pte = hv_pte_set_mode(pte, HV_PTE_MODE_CACHE_TILE_L3);
+			pte = set_remote_cache_cpu(pte, cpu);
+			set_pte_at(&init_mm, addr, ptep, pte);
+
+			/* Update the lowmem mapping for consistency. */
+			lowmem_va = (unsigned long)pfn_to_kaddr(pfn);
+			ptep = virt_to_kpte(lowmem_va);
+			if (pte_huge(*ptep)) {
+				printk(KERN_DEBUG "early shatter of huge page at %#lx\n",
+				       lowmem_va);
+				shatter_pmd((pmd_t *)ptep);
+				ptep = virt_to_kpte(lowmem_va);
+				BUG_ON(pte_huge(*ptep));
+			}
+			BUG_ON(pfn != pte_pfn(*ptep));
+			set_pte_at(&init_mm, lowmem_va, ptep, pte);
+		}
+	}
+
+	/* Set our thread pointer appropriately. */
+	set_my_cpu_offset(__per_cpu_offset[smp_processor_id()]);
+
+	/* Make sure the finv's have completed. */
+	mb_incoherent();
+
+	/* Flush the TLB so we reference it properly from here on out. */
+	local_flush_tlb_all();
+}
+
+static struct resource data_resource = {
+	.name	= "Kernel data",
+	.start	= 0,
+	.end	= 0,
+	.flags	= IORESOURCE_BUSY | IORESOURCE_MEM
+};
+
+static struct resource code_resource = {
+	.name	= "Kernel code",
+	.start	= 0,
+	.end	= 0,
+	.flags	= IORESOURCE_BUSY | IORESOURCE_MEM
+};
+
+/*
+ * On Pro, we reserve all resources above 4GB so that PCI won't try to put
+ * mappings above 4GB.
+ */
+#if defined(CONFIG_PCI) && !defined(__tilegx__)
+static struct resource* __init
+insert_non_bus_resource(void)
+{
+	struct resource *res =
+		kzalloc(sizeof(struct resource), GFP_ATOMIC);
+	if (!res)
+		return NULL;
+	res->name = "Non-Bus Physical Address Space";
+	res->start = (1ULL << 32);
+	res->end = -1LL;
+	res->flags = IORESOURCE_BUSY | IORESOURCE_MEM;
+	if (insert_resource(&iomem_resource, res)) {
+		kfree(res);
+		return NULL;
+	}
+	return res;
+}
+#endif
+
+static struct resource* __init
+insert_ram_resource(u64 start_pfn, u64 end_pfn, bool reserved)
+{
+	struct resource *res =
+		kzalloc(sizeof(struct resource), GFP_ATOMIC);
+	if (!res)
+		return NULL;
+	res->name = reserved ? "Reserved" : "System RAM";
+	res->start = start_pfn << PAGE_SHIFT;
+	res->end = (end_pfn << PAGE_SHIFT) - 1;
+	res->flags = IORESOURCE_BUSY | IORESOURCE_MEM;
+	if (insert_resource(&iomem_resource, res)) {
+		kfree(res);
+		return NULL;
+	}
+	return res;
+}
+
+/*
+ * Request address space for all standard resources
+ *
+ * If the system includes PCI root complex drivers, we need to create
+ * a window just below 4GB where PCI BARs can be mapped.
+ */
+static int __init request_standard_resources(void)
+{
+	int i;
+	enum { CODE_DELTA = MEM_SV_START - PAGE_OFFSET };
+
+#if defined(CONFIG_PCI) && !defined(__tilegx__)
+	insert_non_bus_resource();
+#endif
+
+	for_each_online_node(i) {
+		u64 start_pfn = node_start_pfn[i];
+		u64 end_pfn = node_end_pfn[i];
+
+#if defined(CONFIG_PCI) && !defined(__tilegx__)
+		if (start_pfn <= pci_reserve_start_pfn &&
+		    end_pfn > pci_reserve_start_pfn) {
+			if (end_pfn > pci_reserve_end_pfn)
+				insert_ram_resource(pci_reserve_end_pfn,
+						    end_pfn, 0);
+			end_pfn = pci_reserve_start_pfn;
+		}
+#endif
+		insert_ram_resource(start_pfn, end_pfn, 0);
+	}
+
+	code_resource.start = __pa(_text - CODE_DELTA);
+	code_resource.end = __pa(_etext - CODE_DELTA)-1;
+	data_resource.start = __pa(_sdata);
+	data_resource.end = __pa(_end)-1;
+
+	insert_resource(&iomem_resource, &code_resource);
+	insert_resource(&iomem_resource, &data_resource);
+
+	/* Mark any "memmap" regions busy for the resource manager. */
+	for (i = 0; i < memmap_nr; ++i) {
+		struct memmap_entry *m = &memmap_map[i];
+		insert_ram_resource(PFN_DOWN(m->addr),
+				    PFN_UP(m->addr + m->size - 1), 1);
+	}
+
+#ifdef CONFIG_KEXEC
+	insert_resource(&iomem_resource, &crashk_res);
+#endif
+
+	return 0;
+}
+
+subsys_initcall(request_standard_resources);
diff --git a/kernel/arch/tile/kernel/signal.c b/kernel/arch/tile/kernel/signal.c
new file mode 100644
index 000000000..87299a6cf
--- /dev/null
+++ b/kernel/arch/tile/kernel/signal.c
@@ -0,0 +1,409 @@
+/*
+ * Copyright (C) 1991, 1992  Linus Torvalds
+ * Copyright 2010 Tilera Corporation. All Rights Reserved.
+ *
+ *   This program is free software; you can redistribute it and/or
+ *   modify it under the terms of the GNU General Public License
+ *   as published by the Free Software Foundation, version 2.
+ *
+ *   This program is distributed in the hope that it will be useful, but
+ *   WITHOUT ANY WARRANTY; without even the implied warranty of
+ *   MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
+ *   NON INFRINGEMENT.  See the GNU General Public License for
+ *   more details.
+ */
+
+#include <linux/sched.h>
+#include <linux/mm.h>
+#include <linux/smp.h>
+#include <linux/kernel.h>
+#include <linux/signal.h>
+#include <linux/errno.h>
+#include <linux/wait.h>
+#include <linux/unistd.h>
+#include <linux/stddef.h>
+#include <linux/personality.h>
+#include <linux/suspend.h>
+#include <linux/ptrace.h>
+#include <linux/elf.h>
+#include <linux/compat.h>
+#include <linux/syscalls.h>
+#include <linux/uaccess.h>
+#include <asm/processor.h>
+#include <asm/ucontext.h>
+#include <asm/sigframe.h>
+#include <asm/syscalls.h>
+#include <asm/vdso.h>
+#include <arch/interrupts.h>
+
+#define DEBUG_SIG 0
+
+/*
+ * Do a signal return; undo the signal stack.
+ */
+
+int restore_sigcontext(struct pt_regs *regs,
+		       struct sigcontext __user *sc)
+{
+	int err;
+
+	/* Always make any pending restarted system calls return -EINTR */
+	current->restart_block.fn = do_no_restart_syscall;
+
+	/*
+	 * Enforce that sigcontext is like pt_regs, and doesn't mess
+	 * up our stack alignment rules.
+	 */
+	BUILD_BUG_ON(sizeof(struct sigcontext) != sizeof(struct pt_regs));
+	BUILD_BUG_ON(sizeof(struct sigcontext) % 8 != 0);
+	err = __copy_from_user(regs, sc, sizeof(*regs));
+
+	/* Ensure that the PL is always set to USER_PL. */
+	regs->ex1 = PL_ICS_EX1(USER_PL, EX1_ICS(regs->ex1));
+
+	regs->faultnum = INT_SWINT_1_SIGRETURN;
+
+	return err;
+}
+
+void signal_fault(const char *type, struct pt_regs *regs,
+		  void __user *frame, int sig)
+{
+	trace_unhandled_signal(type, regs, (unsigned long)frame, SIGSEGV);
+	force_sigsegv(sig, current);
+}
+
+/* The assembly shim for this function arranges to ignore the return value. */
+SYSCALL_DEFINE0(rt_sigreturn)
+{
+	struct pt_regs *regs = current_pt_regs();
+	struct rt_sigframe __user *frame =
+		(struct rt_sigframe __user *)(regs->sp);
+	sigset_t set;
+
+	if (!access_ok(VERIFY_READ, frame, sizeof(*frame)))
+		goto badframe;
+	if (__copy_from_user(&set, &frame->uc.uc_sigmask, sizeof(set)))
+		goto badframe;
+
+	set_current_blocked(&set);
+
+	if (restore_sigcontext(regs, &frame->uc.uc_mcontext))
+		goto badframe;
+
+	if (restore_altstack(&frame->uc.uc_stack))
+		goto badframe;
+
+	return 0;
+
+badframe:
+	signal_fault("bad sigreturn frame", regs, frame, 0);
+	return 0;
+}
+
+/*
+ * Set up a signal frame.
+ */
+
+int setup_sigcontext(struct sigcontext __user *sc, struct pt_regs *regs)
+{
+	return  __copy_to_user(sc, regs, sizeof(*regs));
+}
+
+/*
+ * Determine which stack to use..
+ */
+static inline void __user *get_sigframe(struct k_sigaction *ka,
+					struct pt_regs *regs,
+					size_t frame_size)
+{
+	unsigned long sp;
+
+	/* Default to using normal stack */
+	sp = regs->sp;
+
+	/*
+	 * If we are on the alternate signal stack and would overflow
+	 * it, don't.  Return an always-bogus address instead so we
+	 * will die with SIGSEGV.
+	 */
+	if (on_sig_stack(sp) && !likely(on_sig_stack(sp - frame_size)))
+		return (void __user __force *)-1UL;
+
+	/* This is the X/Open sanctioned signal stack switching.  */
+	if (ka->sa.sa_flags & SA_ONSTACK) {
+		if (sas_ss_flags(sp) == 0)
+			sp = current->sas_ss_sp + current->sas_ss_size;
+	}
+
+	sp -= frame_size;
+	/*
+	 * Align the stack pointer according to the TILE ABI,
+	 * i.e. so that on function entry (sp & 15) == 0.
+	 */
+	sp &= -16UL;
+	return (void __user *) sp;
+}
+
+static int setup_rt_frame(struct ksignal *ksig, sigset_t *set,
+			  struct pt_regs *regs)
+{
+	unsigned long restorer;
+	struct rt_sigframe __user *frame;
+	int err = 0, sig = ksig->sig;
+
+	frame = get_sigframe(&ksig->ka, regs, sizeof(*frame));
+
+	if (!access_ok(VERIFY_WRITE, frame, sizeof(*frame)))
+		goto err;
+
+	/* Always write at least the signal number for the stack backtracer. */
+	if (ksig->ka.sa.sa_flags & SA_SIGINFO) {
+		/* At sigreturn time, restore the callee-save registers too. */
+		err |= copy_siginfo_to_user(&frame->info, &ksig->info);
+		regs->flags |= PT_FLAGS_RESTORE_REGS;
+	} else {
+		err |= __put_user(ksig->info.si_signo, &frame->info.si_signo);
+	}
+
+	/* Create the ucontext.  */
+	err |= __clear_user(&frame->save_area, sizeof(frame->save_area));
+	err |= __put_user(0, &frame->uc.uc_flags);
+	err |= __put_user(NULL, &frame->uc.uc_link);
+	err |= __save_altstack(&frame->uc.uc_stack, regs->sp);
+	err |= setup_sigcontext(&frame->uc.uc_mcontext, regs);
+	err |= __copy_to_user(&frame->uc.uc_sigmask, set, sizeof(*set));
+	if (err)
+		goto err;
+
+	restorer = VDSO_SYM(&__vdso_rt_sigreturn);
+	if (ksig->ka.sa.sa_flags & SA_RESTORER)
+		restorer = (unsigned long) ksig->ka.sa.sa_restorer;
+
+	/*
+	 * Set up registers for signal handler.
+	 * Registers that we don't modify keep the value they had from
+	 * user-space at the time we took the signal.
+	 * We always pass siginfo and mcontext, regardless of SA_SIGINFO,
+	 * since some things rely on this (e.g. glibc's debug/segfault.c).
+	 */
+	regs->pc = (unsigned long) ksig->ka.sa.sa_handler;
+	regs->ex1 = PL_ICS_EX1(USER_PL, 1); /* set crit sec in handler */
+	regs->sp = (unsigned long) frame;
+	regs->lr = restorer;
+	regs->regs[0] = (unsigned long) sig;
+	regs->regs[1] = (unsigned long) &frame->info;
+	regs->regs[2] = (unsigned long) &frame->uc;
+	regs->flags |= PT_FLAGS_CALLER_SAVES;
+	return 0;
+
+err:
+	trace_unhandled_signal("bad sigreturn frame", regs,
+			      (unsigned long)frame, SIGSEGV);
+	return -EFAULT;
+}
+
+/*
+ * OK, we're invoking a handler
+ */
+
+static void handle_signal(struct ksignal *ksig, struct pt_regs *regs)
+{
+	sigset_t *oldset = sigmask_to_save();
+	int ret;
+
+	/* Are we from a system call? */
+	if (regs->faultnum == INT_SWINT_1) {
+		/* If so, check system call restarting.. */
+		switch (regs->regs[0]) {
+		case -ERESTART_RESTARTBLOCK:
+		case -ERESTARTNOHAND:
+			regs->regs[0] = -EINTR;
+			break;
+
+		case -ERESTARTSYS:
+			if (!(ksig->ka.sa.sa_flags & SA_RESTART)) {
+				regs->regs[0] = -EINTR;
+				break;
+			}
+			/* fallthrough */
+		case -ERESTARTNOINTR:
+			/* Reload caller-saves to restore r0..r5 and r10. */
+			regs->flags |= PT_FLAGS_CALLER_SAVES;
+			regs->regs[0] = regs->orig_r0;
+			regs->pc -= 8;
+		}
+	}
+
+	/* Set up the stack frame */
+#ifdef CONFIG_COMPAT
+	if (is_compat_task())
+		ret = compat_setup_rt_frame(ksig, oldset, regs);
+	else
+#endif
+		ret = setup_rt_frame(ksig, oldset, regs);
+
+	signal_setup_done(ret, ksig, test_thread_flag(TIF_SINGLESTEP));
+}
+
+/*
+ * Note that 'init' is a special process: it doesn't get signals it doesn't
+ * want to handle. Thus you cannot kill init even with a SIGKILL even by
+ * mistake.
+ */
+void do_signal(struct pt_regs *regs)
+{
+	struct ksignal ksig;
+
+	/*
+	 * i386 will check if we're coming from kernel mode and bail out
+	 * here.  In my experience this just turns weird crashes into
+	 * weird spin-hangs.  But if we find a case where this seems
+	 * helpful, we can reinstate the check on "!user_mode(regs)".
+	 */
+
+	if (get_signal(&ksig)) {
+		/* Whee! Actually deliver the signal.  */
+		handle_signal(&ksig, regs);
+		goto done;
+	}
+
+	/* Did we come from a system call? */
+	if (regs->faultnum == INT_SWINT_1) {
+		/* Restart the system call - no handlers present */
+		switch (regs->regs[0]) {
+		case -ERESTARTNOHAND:
+		case -ERESTARTSYS:
+		case -ERESTARTNOINTR:
+			regs->flags |= PT_FLAGS_CALLER_SAVES;
+			regs->regs[0] = regs->orig_r0;
+			regs->pc -= 8;
+			break;
+
+		case -ERESTART_RESTARTBLOCK:
+			regs->flags |= PT_FLAGS_CALLER_SAVES;
+			regs->regs[TREG_SYSCALL_NR] = __NR_restart_syscall;
+			regs->pc -= 8;
+			break;
+		}
+	}
+
+	/* If there's no signal to deliver, just put the saved sigmask back. */
+	restore_saved_sigmask();
+
+done:
+	/* Avoid double syscall restart if there are nested signals. */
+	regs->faultnum = INT_SWINT_1_SIGRETURN;
+}
+
+int show_unhandled_signals = 1;
+
+static int __init crashinfo(char *str)
+{
+	const char *word;
+
+	if (*str == '\0')
+		show_unhandled_signals = 2;
+	else if (*str != '=' || kstrtoint(++str, 0, &show_unhandled_signals) != 0)
+		return 0;
+
+	switch (show_unhandled_signals) {
+	case 0:
+		word = "No";
+		break;
+	case 1:
+		word = "One-line";
+		break;
+	default:
+		word = "Detailed";
+		break;
+	}
+	pr_info("%s crash reports will be generated on the console\n", word);
+	return 1;
+}
+__setup("crashinfo", crashinfo);
+
+static void dump_mem(void __user *address)
+{
+	void __user *addr;
+	enum { region_size = 256, bytes_per_line = 16 };
+	int i, j, k;
+	int found_readable_mem = 0;
+
+	if (!access_ok(VERIFY_READ, address, 1)) {
+		pr_err("Not dumping at address 0x%lx (kernel address)\n",
+		       (unsigned long)address);
+		return;
+	}
+
+	addr = (void __user *)
+		(((unsigned long)address & -bytes_per_line) - region_size/2);
+	if (addr > address)
+		addr = NULL;
+	for (i = 0; i < region_size;
+	     addr += bytes_per_line, i += bytes_per_line) {
+		unsigned char buf[bytes_per_line];
+		char line[100];
+		if (copy_from_user(buf, addr, bytes_per_line))
+			continue;
+		if (!found_readable_mem) {
+			pr_err("Dumping memory around address 0x%lx:\n",
+			       (unsigned long)address);
+			found_readable_mem = 1;
+		}
+		j = sprintf(line, REGFMT ":", (unsigned long)addr);
+		for (k = 0; k < bytes_per_line; ++k)
+			j += sprintf(&line[j], " %02x", buf[k]);
+		pr_err("%s\n", line);
+	}
+	if (!found_readable_mem)
+		pr_err("No readable memory around address 0x%lx\n",
+		       (unsigned long)address);
+}
+
+void trace_unhandled_signal(const char *type, struct pt_regs *regs,
+			    unsigned long address, int sig)
+{
+	struct task_struct *tsk = current;
+
+	if (show_unhandled_signals == 0)
+		return;
+
+	/* If the signal is handled, don't show it here. */
+	if (!is_global_init(tsk)) {
+		void __user *handler =
+			tsk->sighand->action[sig-1].sa.sa_handler;
+		if (handler != SIG_IGN && handler != SIG_DFL)
+			return;
+	}
+
+	/* Rate-limit the one-line output, not the detailed output. */
+	if (show_unhandled_signals <= 1 && !printk_ratelimit())
+		return;
+
+	printk("%s%s[%d]: %s at %lx pc "REGFMT" signal %d",
+	       task_pid_nr(tsk) > 1 ? KERN_INFO : KERN_EMERG,
+	       tsk->comm, task_pid_nr(tsk), type, address, regs->pc, sig);
+
+	print_vma_addr(KERN_CONT " in ", regs->pc);
+
+	printk(KERN_CONT "\n");
+
+	if (show_unhandled_signals > 1) {
+		switch (sig) {
+		case SIGILL:
+		case SIGFPE:
+		case SIGSEGV:
+		case SIGBUS:
+			pr_err("User crash: signal %d, trap %ld, address 0x%lx\n",
+			       sig, regs->faultnum, address);
+			show_regs(regs);
+			dump_mem((void __user *)address);
+			break;
+		default:
+			pr_err("User crash: signal %d, trap %ld\n",
+			       sig, regs->faultnum);
+			break;
+		}
+	}
+}
diff --git a/kernel/arch/tile/kernel/single_step.c b/kernel/arch/tile/kernel/single_step.c
new file mode 100644
index 000000000..53f7b9def
--- /dev/null
+++ b/kernel/arch/tile/kernel/single_step.c
@@ -0,0 +1,785 @@
+/*
+ * Copyright 2010 Tilera Corporation. All Rights Reserved.
+ *
+ *   This program is free software; you can redistribute it and/or
+ *   modify it under the terms of the GNU General Public License
+ *   as published by the Free Software Foundation, version 2.
+ *
+ *   This program is distributed in the hope that it will be useful, but
+ *   WITHOUT ANY WARRANTY; without even the implied warranty of
+ *   MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
+ *   NON INFRINGEMENT.  See the GNU General Public License for
+ *   more details.
+ *
+ * A code-rewriter that enables instruction single-stepping.
+ */
+
+#include <linux/smp.h>
+#include <linux/ptrace.h>
+#include <linux/slab.h>
+#include <linux/thread_info.h>
+#include <linux/uaccess.h>
+#include <linux/mman.h>
+#include <linux/types.h>
+#include <linux/err.h>
+#include <linux/prctl.h>
+#include <linux/context_tracking.h>
+#include <asm/cacheflush.h>
+#include <asm/traps.h>
+#include <asm/uaccess.h>
+#include <asm/unaligned.h>
+#include <arch/abi.h>
+#include <arch/spr_def.h>
+#include <arch/opcode.h>
+
+
+#ifndef __tilegx__   /* Hardware support for single step unavailable. */
+
+#define signExtend17(val) sign_extend((val), 17)
+#define TILE_X1_MASK (0xffffffffULL << 31)
+
+enum mem_op {
+	MEMOP_NONE,
+	MEMOP_LOAD,
+	MEMOP_STORE,
+	MEMOP_LOAD_POSTINCR,
+	MEMOP_STORE_POSTINCR
+};
+
+static inline tilepro_bundle_bits set_BrOff_X1(tilepro_bundle_bits n,
+	s32 offset)
+{
+	tilepro_bundle_bits result;
+
+	/* mask out the old offset */
+	tilepro_bundle_bits mask = create_BrOff_X1(-1);
+	result = n & (~mask);
+
+	/* or in the new offset */
+	result |= create_BrOff_X1(offset);
+
+	return result;
+}
+
+static inline tilepro_bundle_bits move_X1(tilepro_bundle_bits n, int dest,
+	int src)
+{
+	tilepro_bundle_bits result;
+	tilepro_bundle_bits op;
+
+	result = n & (~TILE_X1_MASK);
+
+	op = create_Opcode_X1(SPECIAL_0_OPCODE_X1) |
+		create_RRROpcodeExtension_X1(OR_SPECIAL_0_OPCODE_X1) |
+		create_Dest_X1(dest) |
+		create_SrcB_X1(TREG_ZERO) |
+		create_SrcA_X1(src) ;
+
+	result |= op;
+	return result;
+}
+
+static inline tilepro_bundle_bits nop_X1(tilepro_bundle_bits n)
+{
+	return move_X1(n, TREG_ZERO, TREG_ZERO);
+}
+
+static inline tilepro_bundle_bits addi_X1(
+	tilepro_bundle_bits n, int dest, int src, int imm)
+{
+	n &= ~TILE_X1_MASK;
+
+	n |=  (create_SrcA_X1(src) |
+	       create_Dest_X1(dest) |
+	       create_Imm8_X1(imm) |
+	       create_S_X1(0) |
+	       create_Opcode_X1(IMM_0_OPCODE_X1) |
+	       create_ImmOpcodeExtension_X1(ADDI_IMM_0_OPCODE_X1));
+
+	return n;
+}
+
+static tilepro_bundle_bits rewrite_load_store_unaligned(
+	struct single_step_state *state,
+	tilepro_bundle_bits bundle,
+	struct pt_regs *regs,
+	enum mem_op mem_op,
+	int size, int sign_ext)
+{
+	unsigned char __user *addr;
+	int val_reg, addr_reg, err, val;
+	int align_ctl;
+
+	align_ctl = unaligned_fixup;
+	switch (task_thread_info(current)->align_ctl) {
+	case PR_UNALIGN_NOPRINT:
+		align_ctl = 1;
+		break;
+	case PR_UNALIGN_SIGBUS:
+		align_ctl = 0;
+		break;
+	}
+
+	/* Get address and value registers */
+	if (bundle & TILEPRO_BUNDLE_Y_ENCODING_MASK) {
+		addr_reg = get_SrcA_Y2(bundle);
+		val_reg = get_SrcBDest_Y2(bundle);
+	} else if (mem_op == MEMOP_LOAD || mem_op == MEMOP_LOAD_POSTINCR) {
+		addr_reg = get_SrcA_X1(bundle);
+		val_reg  = get_Dest_X1(bundle);
+	} else {
+		addr_reg = get_SrcA_X1(bundle);
+		val_reg  = get_SrcB_X1(bundle);
+	}
+
+	/*
+	 * If registers are not GPRs, don't try to handle it.
+	 *
+	 * FIXME: we could handle non-GPR loads by getting the real value
+	 * from memory, writing it to the single step buffer, using a
+	 * temp_reg to hold a pointer to that memory, then executing that
+	 * instruction and resetting temp_reg.  For non-GPR stores, it's a
+	 * little trickier; we could use the single step buffer for that
+	 * too, but we'd have to add some more state bits so that we could
+	 * call back in here to copy that value to the real target.  For
+	 * now, we just handle the simple case.
+	 */
+	if ((val_reg >= PTREGS_NR_GPRS &&
+	     (val_reg != TREG_ZERO ||
+	      mem_op == MEMOP_LOAD ||
+	      mem_op == MEMOP_LOAD_POSTINCR)) ||
+	    addr_reg >= PTREGS_NR_GPRS)
+		return bundle;
+
+	/* If it's aligned, don't handle it specially */
+	addr = (void __user *)regs->regs[addr_reg];
+	if (((unsigned long)addr % size) == 0)
+		return bundle;
+
+	/*
+	 * Return SIGBUS with the unaligned address, if requested.
+	 * Note that we return SIGBUS even for completely invalid addresses
+	 * as long as they are in fact unaligned; this matches what the
+	 * tilepro hardware would be doing, if it could provide us with the
+	 * actual bad address in an SPR, which it doesn't.
+	 */
+	if (align_ctl == 0) {
+		siginfo_t info = {
+			.si_signo = SIGBUS,
+			.si_code = BUS_ADRALN,
+			.si_addr = addr
+		};
+		trace_unhandled_signal("unaligned trap", regs,
+				       (unsigned long)addr, SIGBUS);
+		force_sig_info(info.si_signo, &info, current);
+		return (tilepro_bundle_bits) 0;
+	}
+
+	/* Handle unaligned load/store */
+	if (mem_op == MEMOP_LOAD || mem_op == MEMOP_LOAD_POSTINCR) {
+		unsigned short val_16;
+		switch (size) {
+		case 2:
+			err = copy_from_user(&val_16, addr, sizeof(val_16));
+			val = sign_ext ? ((short)val_16) : val_16;
+			break;
+		case 4:
+			err = copy_from_user(&val, addr, sizeof(val));
+			break;
+		default:
+			BUG();
+		}
+		if (err == 0) {
+			state->update_reg = val_reg;
+			state->update_value = val;
+			state->update = 1;
+		}
+	} else {
+		unsigned short val_16;
+		val = (val_reg == TREG_ZERO) ? 0 : regs->regs[val_reg];
+		switch (size) {
+		case 2:
+			val_16 = val;
+			err = copy_to_user(addr, &val_16, sizeof(val_16));
+			break;
+		case 4:
+			err = copy_to_user(addr, &val, sizeof(val));
+			break;
+		default:
+			BUG();
+		}
+	}
+
+	if (err) {
+		siginfo_t info = {
+			.si_signo = SIGBUS,
+			.si_code = BUS_ADRALN,
+			.si_addr = addr
+		};
+		trace_unhandled_signal("bad address for unaligned fixup", regs,
+				       (unsigned long)addr, SIGBUS);
+		force_sig_info(info.si_signo, &info, current);
+		return (tilepro_bundle_bits) 0;
+	}
+
+	if (unaligned_printk || unaligned_fixup_count == 0) {
+		pr_info("Process %d/%s: PC %#lx: Fixup of unaligned %s at %#lx\n",
+			current->pid, current->comm, regs->pc,
+			mem_op == MEMOP_LOAD || mem_op == MEMOP_LOAD_POSTINCR ?
+			"load" : "store",
+			(unsigned long)addr);
+		if (!unaligned_printk) {
+#define P pr_info
+P("\n");
+P("Unaligned fixups in the kernel will slow your application considerably.\n");
+P("To find them, write a \"1\" to /proc/sys/tile/unaligned_fixup/printk,\n");
+P("which requests the kernel show all unaligned fixups, or write a \"0\"\n");
+P("to /proc/sys/tile/unaligned_fixup/enabled, in which case each unaligned\n");
+P("access will become a SIGBUS you can debug. No further warnings will be\n");
+P("shown so as to avoid additional slowdown, but you can track the number\n");
+P("of fixups performed via /proc/sys/tile/unaligned_fixup/count.\n");
+P("Use the tile-addr2line command (see \"info addr2line\") to decode PCs.\n");
+P("\n");
+#undef P
+		}
+	}
+	++unaligned_fixup_count;
+
+	if (bundle & TILEPRO_BUNDLE_Y_ENCODING_MASK) {
+		/* Convert the Y2 instruction to a prefetch. */
+		bundle &= ~(create_SrcBDest_Y2(-1) |
+			    create_Opcode_Y2(-1));
+		bundle |= (create_SrcBDest_Y2(TREG_ZERO) |
+			   create_Opcode_Y2(LW_OPCODE_Y2));
+	/* Replace the load postincr with an addi */
+	} else if (mem_op == MEMOP_LOAD_POSTINCR) {
+		bundle = addi_X1(bundle, addr_reg, addr_reg,
+				 get_Imm8_X1(bundle));
+	/* Replace the store postincr with an addi */
+	} else if (mem_op == MEMOP_STORE_POSTINCR) {
+		bundle = addi_X1(bundle, addr_reg, addr_reg,
+				 get_Dest_Imm8_X1(bundle));
+	} else {
+		/* Convert the X1 instruction to a nop. */
+		bundle &= ~(create_Opcode_X1(-1) |
+			    create_UnShOpcodeExtension_X1(-1) |
+			    create_UnOpcodeExtension_X1(-1));
+		bundle |= (create_Opcode_X1(SHUN_0_OPCODE_X1) |
+			   create_UnShOpcodeExtension_X1(
+				   UN_0_SHUN_0_OPCODE_X1) |
+			   create_UnOpcodeExtension_X1(
+				   NOP_UN_0_SHUN_0_OPCODE_X1));
+	}
+
+	return bundle;
+}
+
+/*
+ * Called after execve() has started the new image.  This allows us
+ * to reset the info state.  Note that the the mmap'ed memory, if there
+ * was any, has already been unmapped by the exec.
+ */
+void single_step_execve(void)
+{
+	struct thread_info *ti = current_thread_info();
+	kfree(ti->step_state);
+	ti->step_state = NULL;
+}
+
+/*
+ * single_step_once() - entry point when single stepping has been triggered.
+ * @regs: The machine register state
+ *
+ *  When we arrive at this routine via a trampoline, the single step
+ *  engine copies the executing bundle to the single step buffer.
+ *  If the instruction is a condition branch, then the target is
+ *  reset to one past the next instruction. If the instruction
+ *  sets the lr, then that is noted. If the instruction is a jump
+ *  or call, then the new target pc is preserved and the current
+ *  bundle instruction set to null.
+ *
+ *  The necessary post-single-step rewriting information is stored in
+ *  single_step_state->  We use data segment values because the
+ *  stack will be rewound when we run the rewritten single-stepped
+ *  instruction.
+ */
+void single_step_once(struct pt_regs *regs)
+{
+	extern tilepro_bundle_bits __single_step_ill_insn;
+	extern tilepro_bundle_bits __single_step_j_insn;
+	extern tilepro_bundle_bits __single_step_addli_insn;
+	extern tilepro_bundle_bits __single_step_auli_insn;
+	struct thread_info *info = (void *)current_thread_info();
+	struct single_step_state *state = info->step_state;
+	int is_single_step = test_ti_thread_flag(info, TIF_SINGLESTEP);
+	tilepro_bundle_bits __user *buffer, *pc;
+	tilepro_bundle_bits bundle;
+	int temp_reg;
+	int target_reg = TREG_LR;
+	int err;
+	enum mem_op mem_op = MEMOP_NONE;
+	int size = 0, sign_ext = 0;  /* happy compiler */
+	int align_ctl;
+
+	align_ctl = unaligned_fixup;
+	switch (task_thread_info(current)->align_ctl) {
+	case PR_UNALIGN_NOPRINT:
+		align_ctl = 1;
+		break;
+	case PR_UNALIGN_SIGBUS:
+		align_ctl = 0;
+		break;
+	}
+
+	asm(
+"    .pushsection .rodata.single_step\n"
+"    .align 8\n"
+"    .globl    __single_step_ill_insn\n"
+"__single_step_ill_insn:\n"
+"    ill\n"
+"    .globl    __single_step_addli_insn\n"
+"__single_step_addli_insn:\n"
+"    { nop; addli r0, zero, 0 }\n"
+"    .globl    __single_step_auli_insn\n"
+"__single_step_auli_insn:\n"
+"    { nop; auli r0, r0, 0 }\n"
+"    .globl    __single_step_j_insn\n"
+"__single_step_j_insn:\n"
+"    j .\n"
+"    .popsection\n"
+	);
+
+	/*
+	 * Enable interrupts here to allow touching userspace and the like.
+	 * The callers expect this: do_trap() already has interrupts
+	 * enabled, and do_work_pending() handles functions that enable
+	 * interrupts internally.
+	 */
+	local_irq_enable();
+
+	if (state == NULL) {
+		/* allocate a page of writable, executable memory */
+		state = kmalloc(sizeof(struct single_step_state), GFP_KERNEL);
+		if (state == NULL) {
+			pr_err("Out of kernel memory trying to single-step\n");
+			return;
+		}
+
+		/* allocate a cache line of writable, executable memory */
+		buffer = (void __user *) vm_mmap(NULL, 0, 64,
+					  PROT_EXEC | PROT_READ | PROT_WRITE,
+					  MAP_PRIVATE | MAP_ANONYMOUS,
+					  0);
+
+		if (IS_ERR((void __force *)buffer)) {
+			kfree(state);
+			pr_err("Out of kernel pages trying to single-step\n");
+			return;
+		}
+
+		state->buffer = buffer;
+		state->is_enabled = 0;
+
+		info->step_state = state;
+
+		/* Validate our stored instruction patterns */
+		BUG_ON(get_Opcode_X1(__single_step_addli_insn) !=
+		       ADDLI_OPCODE_X1);
+		BUG_ON(get_Opcode_X1(__single_step_auli_insn) !=
+		       AULI_OPCODE_X1);
+		BUG_ON(get_SrcA_X1(__single_step_addli_insn) != TREG_ZERO);
+		BUG_ON(get_Dest_X1(__single_step_addli_insn) != 0);
+		BUG_ON(get_JOffLong_X1(__single_step_j_insn) != 0);
+	}
+
+	/*
+	 * If we are returning from a syscall, we still haven't hit the
+	 * "ill" for the swint1 instruction.  So back the PC up to be
+	 * pointing at the swint1, but we'll actually return directly
+	 * back to the "ill" so we come back in via SIGILL as if we
+	 * had "executed" the swint1 without ever being in kernel space.
+	 */
+	if (regs->faultnum == INT_SWINT_1)
+		regs->pc -= 8;
+
+	pc = (tilepro_bundle_bits __user *)(regs->pc);
+	if (get_user(bundle, pc) != 0) {
+		pr_err("Couldn't read instruction at %p trying to step\n", pc);
+		return;
+	}
+
+	/* We'll follow the instruction with 2 ill op bundles */
+	state->orig_pc = (unsigned long)pc;
+	state->next_pc = (unsigned long)(pc + 1);
+	state->branch_next_pc = 0;
+	state->update = 0;
+
+	if (!(bundle & TILEPRO_BUNDLE_Y_ENCODING_MASK)) {
+		/* two wide, check for control flow */
+		int opcode = get_Opcode_X1(bundle);
+
+		switch (opcode) {
+		/* branches */
+		case BRANCH_OPCODE_X1:
+		{
+			s32 offset = signExtend17(get_BrOff_X1(bundle));
+
+			/*
+			 * For branches, we use a rewriting trick to let the
+			 * hardware evaluate whether the branch is taken or
+			 * untaken.  We record the target offset and then
+			 * rewrite the branch instruction to target 1 insn
+			 * ahead if the branch is taken.  We then follow the
+			 * rewritten branch with two bundles, each containing
+			 * an "ill" instruction. The supervisor examines the
+			 * pc after the single step code is executed, and if
+			 * the pc is the first ill instruction, then the
+			 * branch (if any) was not taken.  If the pc is the
+			 * second ill instruction, then the branch was
+			 * taken. The new pc is computed for these cases, and
+			 * inserted into the registers for the thread.  If
+			 * the pc is the start of the single step code, then
+			 * an exception or interrupt was taken before the
+			 * code started processing, and the same "original"
+			 * pc is restored.  This change, different from the
+			 * original implementation, has the advantage of
+			 * executing a single user instruction.
+			 */
+			state->branch_next_pc = (unsigned long)(pc + offset);
+
+			/* rewrite branch offset to go forward one bundle */
+			bundle = set_BrOff_X1(bundle, 2);
+		}
+		break;
+
+		/* jumps */
+		case JALB_OPCODE_X1:
+		case JALF_OPCODE_X1:
+			state->update = 1;
+			state->next_pc =
+				(unsigned long) (pc + get_JOffLong_X1(bundle));
+			break;
+
+		case JB_OPCODE_X1:
+		case JF_OPCODE_X1:
+			state->next_pc =
+				(unsigned long) (pc + get_JOffLong_X1(bundle));
+			bundle = nop_X1(bundle);
+			break;
+
+		case SPECIAL_0_OPCODE_X1:
+			switch (get_RRROpcodeExtension_X1(bundle)) {
+			/* jump-register */
+			case JALRP_SPECIAL_0_OPCODE_X1:
+			case JALR_SPECIAL_0_OPCODE_X1:
+				state->update = 1;
+				state->next_pc =
+					regs->regs[get_SrcA_X1(bundle)];
+				break;
+
+			case JRP_SPECIAL_0_OPCODE_X1:
+			case JR_SPECIAL_0_OPCODE_X1:
+				state->next_pc =
+					regs->regs[get_SrcA_X1(bundle)];
+				bundle = nop_X1(bundle);
+				break;
+
+			case LNK_SPECIAL_0_OPCODE_X1:
+				state->update = 1;
+				target_reg = get_Dest_X1(bundle);
+				break;
+
+			/* stores */
+			case SH_SPECIAL_0_OPCODE_X1:
+				mem_op = MEMOP_STORE;
+				size = 2;
+				break;
+
+			case SW_SPECIAL_0_OPCODE_X1:
+				mem_op = MEMOP_STORE;
+				size = 4;
+				break;
+			}
+			break;
+
+		/* loads and iret */
+		case SHUN_0_OPCODE_X1:
+			if (get_UnShOpcodeExtension_X1(bundle) ==
+			    UN_0_SHUN_0_OPCODE_X1) {
+				switch (get_UnOpcodeExtension_X1(bundle)) {
+				case LH_UN_0_SHUN_0_OPCODE_X1:
+					mem_op = MEMOP_LOAD;
+					size = 2;
+					sign_ext = 1;
+					break;
+
+				case LH_U_UN_0_SHUN_0_OPCODE_X1:
+					mem_op = MEMOP_LOAD;
+					size = 2;
+					sign_ext = 0;
+					break;
+
+				case LW_UN_0_SHUN_0_OPCODE_X1:
+					mem_op = MEMOP_LOAD;
+					size = 4;
+					break;
+
+				case IRET_UN_0_SHUN_0_OPCODE_X1:
+				{
+					unsigned long ex0_0 = __insn_mfspr(
+						SPR_EX_CONTEXT_0_0);
+					unsigned long ex0_1 = __insn_mfspr(
+						SPR_EX_CONTEXT_0_1);
+					/*
+					 * Special-case it if we're iret'ing
+					 * to PL0 again.  Otherwise just let
+					 * it run and it will generate SIGILL.
+					 */
+					if (EX1_PL(ex0_1) == USER_PL) {
+						state->next_pc = ex0_0;
+						regs->ex1 = ex0_1;
+						bundle = nop_X1(bundle);
+					}
+				}
+				}
+			}
+			break;
+
+		/* postincrement operations */
+		case IMM_0_OPCODE_X1:
+			switch (get_ImmOpcodeExtension_X1(bundle)) {
+			case LWADD_IMM_0_OPCODE_X1:
+				mem_op = MEMOP_LOAD_POSTINCR;
+				size = 4;
+				break;
+
+			case LHADD_IMM_0_OPCODE_X1:
+				mem_op = MEMOP_LOAD_POSTINCR;
+				size = 2;
+				sign_ext = 1;
+				break;
+
+			case LHADD_U_IMM_0_OPCODE_X1:
+				mem_op = MEMOP_LOAD_POSTINCR;
+				size = 2;
+				sign_ext = 0;
+				break;
+
+			case SWADD_IMM_0_OPCODE_X1:
+				mem_op = MEMOP_STORE_POSTINCR;
+				size = 4;
+				break;
+
+			case SHADD_IMM_0_OPCODE_X1:
+				mem_op = MEMOP_STORE_POSTINCR;
+				size = 2;
+				break;
+
+			default:
+				break;
+			}
+			break;
+		}
+
+		if (state->update) {
+			/*
+			 * Get an available register.  We start with a
+			 * bitmask with 1's for available registers.
+			 * We truncate to the low 32 registers since
+			 * we are guaranteed to have set bits in the
+			 * low 32 bits, then use ctz to pick the first.
+			 */
+			u32 mask = (u32) ~((1ULL << get_Dest_X0(bundle)) |
+					   (1ULL << get_SrcA_X0(bundle)) |
+					   (1ULL << get_SrcB_X0(bundle)) |
+					   (1ULL << target_reg));
+			temp_reg = __builtin_ctz(mask);
+			state->update_reg = temp_reg;
+			state->update_value = regs->regs[temp_reg];
+			regs->regs[temp_reg] = (unsigned long) (pc+1);
+			regs->flags |= PT_FLAGS_RESTORE_REGS;
+			bundle = move_X1(bundle, target_reg, temp_reg);
+		}
+	} else {
+		int opcode = get_Opcode_Y2(bundle);
+
+		switch (opcode) {
+		/* loads */
+		case LH_OPCODE_Y2:
+			mem_op = MEMOP_LOAD;
+			size = 2;
+			sign_ext = 1;
+			break;
+
+		case LH_U_OPCODE_Y2:
+			mem_op = MEMOP_LOAD;
+			size = 2;
+			sign_ext = 0;
+			break;
+
+		case LW_OPCODE_Y2:
+			mem_op = MEMOP_LOAD;
+			size = 4;
+			break;
+
+		/* stores */
+		case SH_OPCODE_Y2:
+			mem_op = MEMOP_STORE;
+			size = 2;
+			break;
+
+		case SW_OPCODE_Y2:
+			mem_op = MEMOP_STORE;
+			size = 4;
+			break;
+		}
+	}
+
+	/*
+	 * Check if we need to rewrite an unaligned load/store.
+	 * Returning zero is a special value meaning we generated a signal.
+	 */
+	if (mem_op != MEMOP_NONE && align_ctl >= 0) {
+		bundle = rewrite_load_store_unaligned(state, bundle, regs,
+						      mem_op, size, sign_ext);
+		if (bundle == 0)
+			return;
+	}
+
+	/* write the bundle to our execution area */
+	buffer = state->buffer;
+	err = __put_user(bundle, buffer++);
+
+	/*
+	 * If we're really single-stepping, we take an INT_ILL after.
+	 * If we're just handling an unaligned access, we can just
+	 * jump directly back to where we were in user code.
+	 */
+	if (is_single_step) {
+		err |= __put_user(__single_step_ill_insn, buffer++);
+		err |= __put_user(__single_step_ill_insn, buffer++);
+	} else {
+		long delta;
+
+		if (state->update) {
+			/* We have some state to update; do it inline */
+			int ha16;
+			bundle = __single_step_addli_insn;
+			bundle |= create_Dest_X1(state->update_reg);
+			bundle |= create_Imm16_X1(state->update_value);
+			err |= __put_user(bundle, buffer++);
+			bundle = __single_step_auli_insn;
+			bundle |= create_Dest_X1(state->update_reg);
+			bundle |= create_SrcA_X1(state->update_reg);
+			ha16 = (state->update_value + 0x8000) >> 16;
+			bundle |= create_Imm16_X1(ha16);
+			err |= __put_user(bundle, buffer++);
+			state->update = 0;
+		}
+
+		/* End with a jump back to the next instruction */
+		delta = ((regs->pc + TILEPRO_BUNDLE_SIZE_IN_BYTES) -
+			(unsigned long)buffer) >>
+			TILEPRO_LOG2_BUNDLE_ALIGNMENT_IN_BYTES;
+		bundle = __single_step_j_insn;
+		bundle |= create_JOffLong_X1(delta);
+		err |= __put_user(bundle, buffer++);
+	}
+
+	if (err) {
+		pr_err("Fault when writing to single-step buffer\n");
+		return;
+	}
+
+	/*
+	 * Flush the buffer.
+	 * We do a local flush only, since this is a thread-specific buffer.
+	 */
+	__flush_icache_range((unsigned long)state->buffer,
+			     (unsigned long)buffer);
+
+	/* Indicate enabled */
+	state->is_enabled = is_single_step;
+	regs->pc = (unsigned long)state->buffer;
+
+	/* Fault immediately if we are coming back from a syscall. */
+	if (regs->faultnum == INT_SWINT_1)
+		regs->pc += 8;
+}
+
+#else
+
+static DEFINE_PER_CPU(unsigned long, ss_saved_pc);
+
+
+/*
+ * Called directly on the occasion of an interrupt.
+ *
+ * If the process doesn't have single step set, then we use this as an
+ * opportunity to turn single step off.
+ *
+ * It has been mentioned that we could conditionally turn off single stepping
+ * on each entry into the kernel and rely on single_step_once to turn it
+ * on for the processes that matter (as we already do), but this
+ * implementation is somewhat more efficient in that we muck with registers
+ * once on a bum interrupt rather than on every entry into the kernel.
+ *
+ * If SINGLE_STEP_CONTROL_K has CANCELED set, then an interrupt occurred,
+ * so we have to run through this process again before we can say that an
+ * instruction has executed.
+ *
+ * swint will set CANCELED, but it's a legitimate instruction.  Fortunately
+ * it changes the PC.  If it hasn't changed, then we know that the interrupt
+ * wasn't generated by swint and we'll need to run this process again before
+ * we can say an instruction has executed.
+ *
+ * If either CANCELED == 0 or the PC's changed, we send out SIGTRAPs and get
+ * on with our lives.
+ */
+
+void gx_singlestep_handle(struct pt_regs *regs, int fault_num)
+{
+	enum ctx_state prev_state = exception_enter();
+	unsigned long *ss_pc = this_cpu_ptr(&ss_saved_pc);
+	struct thread_info *info = (void *)current_thread_info();
+	int is_single_step = test_ti_thread_flag(info, TIF_SINGLESTEP);
+	unsigned long control = __insn_mfspr(SPR_SINGLE_STEP_CONTROL_K);
+
+	if (is_single_step == 0) {
+		__insn_mtspr(SPR_SINGLE_STEP_EN_K_K, 0);
+
+	} else if ((*ss_pc != regs->pc) ||
+		   (!(control & SPR_SINGLE_STEP_CONTROL_1__CANCELED_MASK))) {
+
+		control |= SPR_SINGLE_STEP_CONTROL_1__CANCELED_MASK;
+		control |= SPR_SINGLE_STEP_CONTROL_1__INHIBIT_MASK;
+		__insn_mtspr(SPR_SINGLE_STEP_CONTROL_K, control);
+		send_sigtrap(current, regs);
+	}
+	exception_exit(prev_state);
+}
+
+
+/*
+ * Called from need_singlestep.  Set up the control registers and the enable
+ * register, then return back.
+ */
+
+void single_step_once(struct pt_regs *regs)
+{
+	unsigned long *ss_pc = this_cpu_ptr(&ss_saved_pc);
+	unsigned long control = __insn_mfspr(SPR_SINGLE_STEP_CONTROL_K);
+
+	*ss_pc = regs->pc;
+	control |= SPR_SINGLE_STEP_CONTROL_1__CANCELED_MASK;
+	control |= SPR_SINGLE_STEP_CONTROL_1__INHIBIT_MASK;
+	__insn_mtspr(SPR_SINGLE_STEP_CONTROL_K, control);
+	__insn_mtspr(SPR_SINGLE_STEP_EN_K_K, 1 << USER_PL);
+}
+
+void single_step_execve(void)
+{
+	/* Nothing */
+}
+
+#endif /* !__tilegx__ */
diff --git a/kernel/arch/tile/kernel/smp.c b/kernel/arch/tile/kernel/smp.c
new file mode 100644
index 000000000..07e3ff5cc
--- /dev/null
+++ b/kernel/arch/tile/kernel/smp.c
@@ -0,0 +1,287 @@
+/*
+ * Copyright 2010 Tilera Corporation. All Rights Reserved.
+ *
+ *   This program is free software; you can redistribute it and/or
+ *   modify it under the terms of the GNU General Public License
+ *   as published by the Free Software Foundation, version 2.
+ *
+ *   This program is distributed in the hope that it will be useful, but
+ *   WITHOUT ANY WARRANTY; without even the implied warranty of
+ *   MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
+ *   NON INFRINGEMENT.  See the GNU General Public License for
+ *   more details.
+ *
+ * TILE SMP support routines.
+ */
+
+#include <linux/smp.h>
+#include <linux/interrupt.h>
+#include <linux/io.h>
+#include <linux/irq.h>
+#include <linux/irq_work.h>
+#include <linux/module.h>
+#include <asm/cacheflush.h>
+#include <asm/homecache.h>
+
+/*
+ * We write to width and height with a single store in head_NN.S,
+ * so make the variable aligned to "long".
+ */
+HV_Topology smp_topology __write_once __aligned(sizeof(long));
+EXPORT_SYMBOL(smp_topology);
+
+#if CHIP_HAS_IPI()
+static unsigned long __iomem *ipi_mappings[NR_CPUS];
+#endif
+
+/* Does messaging work correctly to the local cpu? */
+bool self_interrupt_ok;
+
+/*
+ * Top-level send_IPI*() functions to send messages to other cpus.
+ */
+
+/* Set by smp_send_stop() to avoid recursive panics. */
+static int stopping_cpus;
+
+static void __send_IPI_many(HV_Recipient *recip, int nrecip, int tag)
+{
+	int sent = 0;
+	while (sent < nrecip) {
+		int rc = hv_send_message(recip, nrecip,
+					 (HV_VirtAddr)&tag, sizeof(tag));
+		if (rc < 0) {
+			if (!stopping_cpus)  /* avoid recursive panic */
+				panic("hv_send_message returned %d", rc);
+			break;
+		}
+		WARN_ONCE(rc == 0, "hv_send_message() returned zero\n");
+		sent += rc;
+	}
+}
+
+void send_IPI_single(int cpu, int tag)
+{
+	HV_Recipient recip = {
+		.y = cpu / smp_width,
+		.x = cpu % smp_width,
+		.state = HV_TO_BE_SENT
+	};
+	__send_IPI_many(&recip, 1, tag);
+}
+
+void send_IPI_many(const struct cpumask *mask, int tag)
+{
+	HV_Recipient recip[NR_CPUS];
+	int cpu;
+	int nrecip = 0;
+	int my_cpu = smp_processor_id();
+	for_each_cpu(cpu, mask) {
+		HV_Recipient *r;
+		BUG_ON(cpu == my_cpu);
+		r = &recip[nrecip++];
+		r->y = cpu / smp_width;
+		r->x = cpu % smp_width;
+		r->state = HV_TO_BE_SENT;
+	}
+	__send_IPI_many(recip, nrecip, tag);
+}
+
+void send_IPI_allbutself(int tag)
+{
+	struct cpumask mask;
+	cpumask_copy(&mask, cpu_online_mask);
+	cpumask_clear_cpu(smp_processor_id(), &mask);
+	send_IPI_many(&mask, tag);
+}
+
+/*
+ * Functions related to starting/stopping cpus.
+ */
+
+/* Handler to start the current cpu. */
+static void smp_start_cpu_interrupt(void)
+{
+	get_irq_regs()->pc = start_cpu_function_addr;
+}
+
+/* Handler to stop the current cpu. */
+static void smp_stop_cpu_interrupt(void)
+{
+	arch_local_irq_disable_all();
+	set_cpu_online(smp_processor_id(), 0);
+	for (;;)
+		asm("nap; nop");
+}
+
+/* This function calls the 'stop' function on all other CPUs in the system. */
+void smp_send_stop(void)
+{
+	stopping_cpus = 1;
+	send_IPI_allbutself(MSG_TAG_STOP_CPU);
+}
+
+/* On panic, just wait; we may get an smp_send_stop() later on. */
+void panic_smp_self_stop(void)
+{
+	while (1)
+		asm("nap; nop");
+}
+
+/*
+ * Dispatch code called from hv_message_intr() for HV_MSG_TILE hv messages.
+ */
+void evaluate_message(int tag)
+{
+	switch (tag) {
+	case MSG_TAG_START_CPU: /* Start up a cpu */
+		smp_start_cpu_interrupt();
+		break;
+
+	case MSG_TAG_STOP_CPU: /* Sent to shut down slave CPU's */
+		smp_stop_cpu_interrupt();
+		break;
+
+	case MSG_TAG_CALL_FUNCTION_MANY: /* Call function on cpumask */
+		generic_smp_call_function_interrupt();
+		break;
+
+	case MSG_TAG_CALL_FUNCTION_SINGLE: /* Call function on one other CPU */
+		generic_smp_call_function_single_interrupt();
+		break;
+
+	case MSG_TAG_IRQ_WORK: /* Invoke IRQ work */
+		irq_work_run();
+		break;
+
+	default:
+		panic("Unknown IPI message tag %d", tag);
+		break;
+	}
+}
+
+
+/*
+ * flush_icache_range() code uses smp_call_function().
+ */
+
+struct ipi_flush {
+	unsigned long start;
+	unsigned long end;
+};
+
+static void ipi_flush_icache_range(void *info)
+{
+	struct ipi_flush *flush = (struct ipi_flush *) info;
+	__flush_icache_range(flush->start, flush->end);
+}
+
+void flush_icache_range(unsigned long start, unsigned long end)
+{
+	struct ipi_flush flush = { start, end };
+
+	/* If invoked with irqs disabled, we can not issue IPIs. */
+	if (irqs_disabled())
+		flush_remote(0, HV_FLUSH_EVICT_L1I, NULL, 0, 0, 0,
+			NULL, NULL, 0);
+	else {
+		preempt_disable();
+		on_each_cpu(ipi_flush_icache_range, &flush, 1);
+		preempt_enable();
+	}
+}
+EXPORT_SYMBOL(flush_icache_range);
+
+
+#ifdef CONFIG_IRQ_WORK
+void arch_irq_work_raise(void)
+{
+	if (arch_irq_work_has_interrupt())
+		send_IPI_single(smp_processor_id(), MSG_TAG_IRQ_WORK);
+}
+#endif
+
+
+/* Called when smp_send_reschedule() triggers IRQ_RESCHEDULE. */
+static irqreturn_t handle_reschedule_ipi(int irq, void *token)
+{
+	__this_cpu_inc(irq_stat.irq_resched_count);
+	scheduler_ipi();
+
+	return IRQ_HANDLED;
+}
+
+static struct irqaction resched_action = {
+	.handler = handle_reschedule_ipi,
+	.name = "resched",
+	.dev_id = handle_reschedule_ipi /* unique token */,
+};
+
+void __init ipi_init(void)
+{
+	int cpu = smp_processor_id();
+	HV_Recipient recip = { .y = cpu_y(cpu), .x = cpu_x(cpu),
+			       .state = HV_TO_BE_SENT };
+	int tag = MSG_TAG_CALL_FUNCTION_SINGLE;
+
+	/*
+	 * Test if we can message ourselves for arch_irq_work_raise.
+	 * This functionality is only available in the Tilera hypervisor
+	 * in versions 4.3.4 and following.
+	 */
+	if (hv_send_message(&recip, 1, (HV_VirtAddr)&tag, sizeof(tag)) == 1)
+		self_interrupt_ok = true;
+	else
+		pr_warn("Older hypervisor: disabling fast irq_work_raise\n");
+
+#if CHIP_HAS_IPI()
+	/* Map IPI trigger MMIO addresses. */
+	for_each_possible_cpu(cpu) {
+		HV_Coord tile;
+		HV_PTE pte;
+		unsigned long offset;
+
+		tile.x = cpu_x(cpu);
+		tile.y = cpu_y(cpu);
+		if (hv_get_ipi_pte(tile, KERNEL_PL, &pte) != 0)
+			panic("Failed to initialize IPI for cpu %d\n", cpu);
+
+		offset = PFN_PHYS(pte_pfn(pte));
+		ipi_mappings[cpu] = ioremap_prot(offset, PAGE_SIZE, pte);
+	}
+#endif
+
+	/* Bind handle_reschedule_ipi() to IRQ_RESCHEDULE. */
+	tile_irq_activate(IRQ_RESCHEDULE, TILE_IRQ_PERCPU);
+	BUG_ON(setup_irq(IRQ_RESCHEDULE, &resched_action));
+}
+
+#if CHIP_HAS_IPI()
+
+void smp_send_reschedule(int cpu)
+{
+	WARN_ON(cpu_is_offline(cpu));
+
+	/*
+	 * We just want to do an MMIO store.  The traditional writeq()
+	 * functions aren't really correct here, since they're always
+	 * directed at the PCI shim.  For now, just do a raw store,
+	 * casting away the __iomem attribute.
+	 */
+	((unsigned long __force *)ipi_mappings[cpu])[IRQ_RESCHEDULE] = 0;
+}
+
+#else
+
+void smp_send_reschedule(int cpu)
+{
+	HV_Coord coord;
+
+	WARN_ON(cpu_is_offline(cpu));
+
+	coord.y = cpu_y(cpu);
+	coord.x = cpu_x(cpu);
+	hv_trigger_ipi(coord, IRQ_RESCHEDULE);
+}
+
+#endif /* CHIP_HAS_IPI() */
diff --git a/kernel/arch/tile/kernel/smpboot.c b/kernel/arch/tile/kernel/smpboot.c
new file mode 100644
index 000000000..20d52a98e
--- /dev/null
+++ b/kernel/arch/tile/kernel/smpboot.c
@@ -0,0 +1,268 @@
+/*
+ * Copyright 2010 Tilera Corporation. All Rights Reserved.
+ *
+ *   This program is free software; you can redistribute it and/or
+ *   modify it under the terms of the GNU General Public License
+ *   as published by the Free Software Foundation, version 2.
+ *
+ *   This program is distributed in the hope that it will be useful, but
+ *   WITHOUT ANY WARRANTY; without even the implied warranty of
+ *   MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
+ *   NON INFRINGEMENT.  See the GNU General Public License for
+ *   more details.
+ */
+
+#include <linux/module.h>
+#include <linux/init.h>
+#include <linux/kernel.h>
+#include <linux/mm.h>
+#include <linux/sched.h>
+#include <linux/kernel_stat.h>
+#include <linux/bootmem.h>
+#include <linux/notifier.h>
+#include <linux/cpu.h>
+#include <linux/percpu.h>
+#include <linux/delay.h>
+#include <linux/err.h>
+#include <linux/irq.h>
+#include <asm/mmu_context.h>
+#include <asm/tlbflush.h>
+#include <asm/sections.h>
+
+/* State of each CPU. */
+static DEFINE_PER_CPU(int, cpu_state) = { 0 };
+
+/* The messaging code jumps to this pointer during boot-up */
+unsigned long start_cpu_function_addr;
+
+/* Called very early during startup to mark boot cpu as online */
+void __init smp_prepare_boot_cpu(void)
+{
+	int cpu = smp_processor_id();
+	set_cpu_online(cpu, 1);
+	set_cpu_present(cpu, 1);
+	__this_cpu_write(cpu_state, CPU_ONLINE);
+
+	init_messaging();
+}
+
+static void start_secondary(void);
+
+/*
+ * Called at the top of init() to launch all the other CPUs.
+ * They run free to complete their initialization and then wait
+ * until they get an IPI from the boot cpu to come online.
+ */
+void __init smp_prepare_cpus(unsigned int max_cpus)
+{
+	long rc;
+	int cpu, cpu_count;
+	int boot_cpu = smp_processor_id();
+
+	current_thread_info()->cpu = boot_cpu;
+
+	/*
+	 * Pin this task to the boot CPU while we bring up the others,
+	 * just to make sure we don't uselessly migrate as they come up.
+	 */
+	rc = sched_setaffinity(current->pid, cpumask_of(boot_cpu));
+	if (rc != 0)
+		pr_err("Couldn't set init affinity to boot cpu (%ld)\n", rc);
+
+	/* Print information about disabled and dataplane cpus. */
+	print_disabled_cpus();
+
+	/*
+	 * Tell the messaging subsystem how to respond to the
+	 * startup message.  We use a level of indirection to avoid
+	 * confusing the linker with the fact that the messaging
+	 * subsystem is calling __init code.
+	 */
+	start_cpu_function_addr = (unsigned long) &online_secondary;
+
+	/* Set up thread context for all new processors. */
+	cpu_count = 1;
+	for (cpu = 0; cpu < NR_CPUS; ++cpu)	{
+		struct task_struct *idle;
+
+		if (cpu == boot_cpu)
+			continue;
+
+		if (!cpu_possible(cpu)) {
+			/*
+			 * Make this processor do nothing on boot.
+			 * Note that we don't give the boot_pc function
+			 * a stack, so it has to be assembly code.
+			 */
+			per_cpu(boot_sp, cpu) = 0;
+			per_cpu(boot_pc, cpu) = (unsigned long) smp_nap;
+			continue;
+		}
+
+		/* Create a new idle thread to run start_secondary() */
+		idle = fork_idle(cpu);
+		if (IS_ERR(idle))
+			panic("failed fork for CPU %d", cpu);
+		idle->thread.pc = (unsigned long) start_secondary;
+
+		/* Make this thread the boot thread for this processor */
+		per_cpu(boot_sp, cpu) = task_ksp0(idle);
+		per_cpu(boot_pc, cpu) = idle->thread.pc;
+
+		++cpu_count;
+	}
+	BUG_ON(cpu_count > (max_cpus ? max_cpus : 1));
+
+	/* Fire up the other tiles, if any */
+	init_cpu_present(cpu_possible_mask);
+	if (cpumask_weight(cpu_present_mask) > 1) {
+		mb();  /* make sure all data is visible to new processors */
+		hv_start_all_tiles();
+	}
+}
+
+static __initdata struct cpumask init_affinity;
+
+static __init int reset_init_affinity(void)
+{
+	long rc = sched_setaffinity(current->pid, &init_affinity);
+	if (rc != 0)
+		pr_warn("couldn't reset init affinity (%ld)\n", rc);
+	return 0;
+}
+late_initcall(reset_init_affinity);
+
+static struct cpumask cpu_started;
+
+/*
+ * Activate a secondary processor.  Very minimal; don't add anything
+ * to this path without knowing what you're doing, since SMP booting
+ * is pretty fragile.
+ */
+static void start_secondary(void)
+{
+	int cpuid;
+
+	preempt_disable();
+
+	cpuid = smp_processor_id();
+
+	/* Set our thread pointer appropriately. */
+	set_my_cpu_offset(__per_cpu_offset[cpuid]);
+
+	/*
+	 * In large machines even this will slow us down, since we
+	 * will be contending for for the printk spinlock.
+	 */
+	/* printk(KERN_DEBUG "Initializing CPU#%d\n", cpuid); */
+
+	/* Initialize the current asid for our first page table. */
+	__this_cpu_write(current_asid, min_asid);
+
+	/* Set up this thread as another owner of the init_mm */
+	atomic_inc(&init_mm.mm_count);
+	current->active_mm = &init_mm;
+	if (current->mm)
+		BUG();
+	enter_lazy_tlb(&init_mm, current);
+
+	/* Allow hypervisor messages to be received */
+	init_messaging();
+	local_irq_enable();
+
+	/* Indicate that we're ready to come up. */
+	/* Must not do this before we're ready to receive messages */
+	if (cpumask_test_and_set_cpu(cpuid, &cpu_started)) {
+		pr_warn("CPU#%d already started!\n", cpuid);
+		for (;;)
+			local_irq_enable();
+	}
+
+	smp_nap();
+}
+
+/*
+ * Bring a secondary processor online.
+ */
+void online_secondary(void)
+{
+	/*
+	 * low-memory mappings have been cleared, flush them from
+	 * the local TLBs too.
+	 */
+	local_flush_tlb();
+
+	BUG_ON(in_interrupt());
+
+	/* This must be done before setting cpu_online_mask */
+	wmb();
+
+	notify_cpu_starting(smp_processor_id());
+
+	set_cpu_online(smp_processor_id(), 1);
+	__this_cpu_write(cpu_state, CPU_ONLINE);
+
+	/* Set up tile-specific state for this cpu. */
+	setup_cpu(0);
+
+	/* Set up tile-timer clock-event device on this cpu */
+	setup_tile_timer();
+
+	cpu_startup_entry(CPUHP_ONLINE);
+}
+
+int __cpu_up(unsigned int cpu, struct task_struct *tidle)
+{
+	/* Wait 5s total for all CPUs for them to come online */
+	static int timeout;
+	for (; !cpumask_test_cpu(cpu, &cpu_started); timeout++) {
+		if (timeout >= 50000) {
+			pr_info("skipping unresponsive cpu%d\n", cpu);
+			local_irq_enable();
+			return -EIO;
+		}
+		udelay(100);
+	}
+
+	local_irq_enable();
+	per_cpu(cpu_state, cpu) = CPU_UP_PREPARE;
+
+	/* Unleash the CPU! */
+	send_IPI_single(cpu, MSG_TAG_START_CPU);
+	while (!cpumask_test_cpu(cpu, cpu_online_mask))
+		cpu_relax();
+	return 0;
+}
+
+static void panic_start_cpu(void)
+{
+	panic("Received a MSG_START_CPU IPI after boot finished.");
+}
+
+void __init smp_cpus_done(unsigned int max_cpus)
+{
+	int cpu, next, rc;
+
+	/* Reset the response to a (now illegal) MSG_START_CPU IPI. */
+	start_cpu_function_addr = (unsigned long) &panic_start_cpu;
+
+	cpumask_copy(&init_affinity, cpu_online_mask);
+
+	/*
+	 * Pin ourselves to a single cpu in the initial affinity set
+	 * so that kernel mappings for the rootfs are not in the dataplane,
+	 * if set, and to avoid unnecessary migrating during bringup.
+	 * Use the last cpu just in case the whole chip has been
+	 * isolated from the scheduler, to keep init away from likely
+	 * more useful user code.  This also ensures that work scheduled
+	 * via schedule_delayed_work() in the init routines will land
+	 * on this cpu.
+	 */
+	for (cpu = cpumask_first(&init_affinity);
+	     (next = cpumask_next(cpu, &init_affinity)) < nr_cpu_ids;
+	     cpu = next)
+		;
+	rc = sched_setaffinity(current->pid, cpumask_of(cpu));
+	if (rc != 0)
+		pr_err("Couldn't set init affinity to cpu %d (%d)\n", cpu, rc);
+}
diff --git a/kernel/arch/tile/kernel/stack.c b/kernel/arch/tile/kernel/stack.c
new file mode 100644
index 000000000..c42dce50a
--- /dev/null
+++ b/kernel/arch/tile/kernel/stack.c
@@ -0,0 +1,515 @@
+/*
+ * Copyright 2010 Tilera Corporation. All Rights Reserved.
+ *
+ *   This program is free software; you can redistribute it and/or
+ *   modify it under the terms of the GNU General Public License
+ *   as published by the Free Software Foundation, version 2.
+ *
+ *   This program is distributed in the hope that it will be useful, but
+ *   WITHOUT ANY WARRANTY; without even the implied warranty of
+ *   MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
+ *   NON INFRINGEMENT.  See the GNU General Public License for
+ *   more details.
+ */
+
+#include <linux/sched.h>
+#include <linux/kernel.h>
+#include <linux/kprobes.h>
+#include <linux/module.h>
+#include <linux/pfn.h>
+#include <linux/kallsyms.h>
+#include <linux/stacktrace.h>
+#include <linux/uaccess.h>
+#include <linux/mmzone.h>
+#include <linux/dcache.h>
+#include <linux/fs.h>
+#include <linux/string.h>
+#include <asm/backtrace.h>
+#include <asm/page.h>
+#include <asm/ucontext.h>
+#include <asm/switch_to.h>
+#include <asm/sigframe.h>
+#include <asm/stack.h>
+#include <asm/vdso.h>
+#include <arch/abi.h>
+#include <arch/interrupts.h>
+
+#define KBT_ONGOING	0  /* Backtrace still ongoing */
+#define KBT_DONE	1  /* Backtrace cleanly completed */
+#define KBT_RUNNING	2  /* Can't run backtrace on a running task */
+#define KBT_LOOP	3  /* Backtrace entered a loop */
+
+/* Is address on the specified kernel stack? */
+static int in_kernel_stack(struct KBacktraceIterator *kbt, unsigned long sp)
+{
+	ulong kstack_base = (ulong) kbt->task->stack;
+	if (kstack_base == 0)  /* corrupt task pointer; just follow stack... */
+		return sp >= PAGE_OFFSET && sp < (unsigned long)high_memory;
+	return sp >= kstack_base && sp < kstack_base + THREAD_SIZE;
+}
+
+/* Callback for backtracer; basically a glorified memcpy */
+static bool read_memory_func(void *result, unsigned long address,
+			     unsigned int size, void *vkbt)
+{
+	int retval;
+	struct KBacktraceIterator *kbt = (struct KBacktraceIterator *)vkbt;
+
+	if (address == 0)
+		return 0;
+	if (__kernel_text_address(address)) {
+		/* OK to read kernel code. */
+	} else if (address >= PAGE_OFFSET) {
+		/* We only tolerate kernel-space reads of this task's stack */
+		if (!in_kernel_stack(kbt, address))
+			return 0;
+	} else if (!kbt->is_current) {
+		return 0;	/* can't read from other user address spaces */
+	}
+	pagefault_disable();
+	retval = __copy_from_user_inatomic(result,
+					   (void __user __force *)address,
+					   size);
+	pagefault_enable();
+	return (retval == 0);
+}
+
+/* Return a pt_regs pointer for a valid fault handler frame */
+static struct pt_regs *valid_fault_handler(struct KBacktraceIterator* kbt)
+{
+	const char *fault = NULL;  /* happy compiler */
+	char fault_buf[64];
+	unsigned long sp = kbt->it.sp;
+	struct pt_regs *p;
+
+	if (sp % sizeof(long) != 0)
+		return NULL;
+	if (!in_kernel_stack(kbt, sp))
+		return NULL;
+	if (!in_kernel_stack(kbt, sp + C_ABI_SAVE_AREA_SIZE + PTREGS_SIZE-1))
+		return NULL;
+	p = (struct pt_regs *)(sp + C_ABI_SAVE_AREA_SIZE);
+	if (p->faultnum == INT_SWINT_1 || p->faultnum == INT_SWINT_1_SIGRETURN)
+		fault = "syscall";
+	else {
+		if (kbt->verbose) {     /* else we aren't going to use it */
+			snprintf(fault_buf, sizeof(fault_buf),
+				 "interrupt %ld", p->faultnum);
+			fault = fault_buf;
+		}
+	}
+	if (EX1_PL(p->ex1) == KERNEL_PL &&
+	    __kernel_text_address(p->pc) &&
+	    in_kernel_stack(kbt, p->sp) &&
+	    p->sp >= sp) {
+		if (kbt->verbose)
+			pr_err("  <%s while in kernel mode>\n", fault);
+	} else if (user_mode(p) &&
+		   p->sp < PAGE_OFFSET && p->sp != 0) {
+		if (kbt->verbose)
+			pr_err("  <%s while in user mode>\n", fault);
+	} else {
+		if (kbt->verbose)
+			pr_err("  (odd fault: pc %#lx, sp %#lx, ex1 %#lx?)\n",
+			       p->pc, p->sp, p->ex1);
+		return NULL;
+	}
+	if (kbt->profile && ((1ULL << p->faultnum) & QUEUED_INTERRUPTS) != 0)
+		return NULL;
+	return p;
+}
+
+/* Is the pc pointing to a sigreturn trampoline? */
+static int is_sigreturn(unsigned long pc)
+{
+	return current->mm && (pc == VDSO_SYM(&__vdso_rt_sigreturn));
+}
+
+/* Return a pt_regs pointer for a valid signal handler frame */
+static struct pt_regs *valid_sigframe(struct KBacktraceIterator* kbt,
+				      struct rt_sigframe* kframe)
+{
+	BacktraceIterator *b = &kbt->it;
+
+	if (is_sigreturn(b->pc) && b->sp < PAGE_OFFSET &&
+	    b->sp % sizeof(long) == 0) {
+		int retval;
+		pagefault_disable();
+		retval = __copy_from_user_inatomic(
+			kframe, (void __user __force *)b->sp,
+			sizeof(*kframe));
+		pagefault_enable();
+		if (retval != 0 ||
+		    (unsigned int)(kframe->info.si_signo) >= _NSIG)
+			return NULL;
+		if (kbt->verbose) {
+			pr_err("  <received signal %d>\n",
+			       kframe->info.si_signo);
+		}
+		return (struct pt_regs *)&kframe->uc.uc_mcontext;
+	}
+	return NULL;
+}
+
+static int KBacktraceIterator_is_sigreturn(struct KBacktraceIterator *kbt)
+{
+	return is_sigreturn(kbt->it.pc);
+}
+
+static int KBacktraceIterator_restart(struct KBacktraceIterator *kbt)
+{
+	struct pt_regs *p;
+	struct rt_sigframe kframe;
+
+	p = valid_fault_handler(kbt);
+	if (p == NULL)
+		p = valid_sigframe(kbt, &kframe);
+	if (p == NULL)
+		return 0;
+	backtrace_init(&kbt->it, read_memory_func, kbt,
+		       p->pc, p->lr, p->sp, p->regs[52]);
+	kbt->new_context = 1;
+	return 1;
+}
+
+/* Find a frame that isn't a sigreturn, if there is one. */
+static int KBacktraceIterator_next_item_inclusive(
+	struct KBacktraceIterator *kbt)
+{
+	for (;;) {
+		do {
+			if (!KBacktraceIterator_is_sigreturn(kbt))
+				return KBT_ONGOING;
+		} while (backtrace_next(&kbt->it));
+
+		if (!KBacktraceIterator_restart(kbt))
+			return KBT_DONE;
+	}
+}
+
+/*
+ * If the current sp is on a page different than what we recorded
+ * as the top-of-kernel-stack last time we context switched, we have
+ * probably blown the stack, and nothing is going to work out well.
+ * If we can at least get out a warning, that may help the debug,
+ * though we probably won't be able to backtrace into the code that
+ * actually did the recursive damage.
+ */
+static void validate_stack(struct pt_regs *regs)
+{
+	int cpu = raw_smp_processor_id();
+	unsigned long ksp0 = get_current_ksp0();
+	unsigned long ksp0_base = ksp0 & -THREAD_SIZE;
+	unsigned long sp = stack_pointer;
+
+	if (EX1_PL(regs->ex1) == KERNEL_PL && regs->sp >= ksp0) {
+		pr_err("WARNING: cpu %d: kernel stack %#lx..%#lx underrun!\n"
+		       "  sp %#lx (%#lx in caller), caller pc %#lx, lr %#lx\n",
+		       cpu, ksp0_base, ksp0, sp, regs->sp, regs->pc, regs->lr);
+	}
+
+	else if (sp < ksp0_base + sizeof(struct thread_info)) {
+		pr_err("WARNING: cpu %d: kernel stack %#lx..%#lx overrun!\n"
+		       "  sp %#lx (%#lx in caller), caller pc %#lx, lr %#lx\n",
+		       cpu, ksp0_base, ksp0, sp, regs->sp, regs->pc, regs->lr);
+	}
+}
+
+void KBacktraceIterator_init(struct KBacktraceIterator *kbt,
+			     struct task_struct *t, struct pt_regs *regs)
+{
+	unsigned long pc, lr, sp, r52;
+	int is_current;
+
+	/*
+	 * Set up callback information.  We grab the kernel stack base
+	 * so we will allow reads of that address range.
+	 */
+	is_current = (t == NULL || t == current);
+	kbt->is_current = is_current;
+	if (is_current)
+		t = validate_current();
+	kbt->task = t;
+	kbt->verbose = 0;   /* override in caller if desired */
+	kbt->profile = 0;   /* override in caller if desired */
+	kbt->end = KBT_ONGOING;
+	kbt->new_context = 1;
+	if (is_current)
+		validate_stack(regs);
+
+	if (regs == NULL) {
+		if (is_current || t->state == TASK_RUNNING) {
+			/* Can't do this; we need registers */
+			kbt->end = KBT_RUNNING;
+			return;
+		}
+		pc = get_switch_to_pc();
+		lr = t->thread.pc;
+		sp = t->thread.ksp;
+		r52 = 0;
+	} else {
+		pc = regs->pc;
+		lr = regs->lr;
+		sp = regs->sp;
+		r52 = regs->regs[52];
+	}
+
+	backtrace_init(&kbt->it, read_memory_func, kbt, pc, lr, sp, r52);
+	kbt->end = KBacktraceIterator_next_item_inclusive(kbt);
+}
+EXPORT_SYMBOL(KBacktraceIterator_init);
+
+int KBacktraceIterator_end(struct KBacktraceIterator *kbt)
+{
+	return kbt->end != KBT_ONGOING;
+}
+EXPORT_SYMBOL(KBacktraceIterator_end);
+
+void KBacktraceIterator_next(struct KBacktraceIterator *kbt)
+{
+	unsigned long old_pc = kbt->it.pc, old_sp = kbt->it.sp;
+	kbt->new_context = 0;
+	if (!backtrace_next(&kbt->it) && !KBacktraceIterator_restart(kbt)) {
+		kbt->end = KBT_DONE;
+		return;
+	}
+	kbt->end = KBacktraceIterator_next_item_inclusive(kbt);
+	if (old_pc == kbt->it.pc && old_sp == kbt->it.sp) {
+		/* Trapped in a loop; give up. */
+		kbt->end = KBT_LOOP;
+	}
+}
+EXPORT_SYMBOL(KBacktraceIterator_next);
+
+static void describe_addr(struct KBacktraceIterator *kbt,
+			  unsigned long address,
+			  int have_mmap_sem, char *buf, size_t bufsize)
+{
+	struct vm_area_struct *vma;
+	size_t namelen, remaining;
+	unsigned long size, offset, adjust;
+	char *p, *modname;
+	const char *name;
+	int rc;
+
+	/*
+	 * Look one byte back for every caller frame (i.e. those that
+	 * aren't a new context) so we look up symbol data for the
+	 * call itself, not the following instruction, which may be on
+	 * a different line (or in a different function).
+	 */
+	adjust = !kbt->new_context;
+	address -= adjust;
+
+	if (address >= PAGE_OFFSET) {
+		/* Handle kernel symbols. */
+		BUG_ON(bufsize < KSYM_NAME_LEN);
+		name = kallsyms_lookup(address, &size, &offset,
+				       &modname, buf);
+		if (name == NULL) {
+			buf[0] = '\0';
+			return;
+		}
+		namelen = strlen(buf);
+		remaining = (bufsize - 1) - namelen;
+		p = buf + namelen;
+		rc = snprintf(p, remaining, "+%#lx/%#lx ",
+			      offset + adjust, size);
+		if (modname && rc < remaining)
+			snprintf(p + rc, remaining - rc, "[%s] ", modname);
+		buf[bufsize-1] = '\0';
+		return;
+	}
+
+	/* If we don't have the mmap_sem, we can't show any more info. */
+	buf[0] = '\0';
+	if (!have_mmap_sem)
+		return;
+
+	/* Find vma info. */
+	vma = find_vma(kbt->task->mm, address);
+	if (vma == NULL || address < vma->vm_start) {
+		snprintf(buf, bufsize, "[unmapped address] ");
+		return;
+	}
+
+	if (vma->vm_file) {
+		p = d_path(&vma->vm_file->f_path, buf, bufsize);
+		if (IS_ERR(p))
+			p = "?";
+		name = kbasename(p);
+	} else {
+		name = "anon";
+	}
+
+	/* Generate a string description of the vma info. */
+	namelen = strlen(name);
+	remaining = (bufsize - 1) - namelen;
+	memmove(buf, name, namelen);
+	snprintf(buf + namelen, remaining, "[%lx+%lx] ",
+		 vma->vm_start, vma->vm_end - vma->vm_start);
+}
+
+/*
+ * Avoid possible crash recursion during backtrace.  If it happens, it
+ * makes it easy to lose the actual root cause of the failure, so we
+ * put a simple guard on all the backtrace loops.
+ */
+static bool start_backtrace(void)
+{
+	if (current->thread.in_backtrace) {
+		pr_err("Backtrace requested while in backtrace!\n");
+		return false;
+	}
+	current->thread.in_backtrace = true;
+	return true;
+}
+
+static void end_backtrace(void)
+{
+	current->thread.in_backtrace = false;
+}
+
+/*
+ * This method wraps the backtracer's more generic support.
+ * It is only invoked from the architecture-specific code; show_stack()
+ * and dump_stack() (in entry.S) are architecture-independent entry points.
+ */
+void tile_show_stack(struct KBacktraceIterator *kbt, int headers)
+{
+	int i;
+	int have_mmap_sem = 0;
+
+	if (!start_backtrace())
+		return;
+	if (headers) {
+		/*
+		 * Add a blank line since if we are called from panic(),
+		 * then bust_spinlocks() spit out a space in front of us
+		 * and it will mess up our KERN_ERR.
+		 */
+		pr_err("Starting stack dump of tid %d, pid %d (%s) on cpu %d at cycle %lld\n",
+		       kbt->task->pid, kbt->task->tgid, kbt->task->comm,
+		       raw_smp_processor_id(), get_cycles());
+	}
+	kbt->verbose = 1;
+	i = 0;
+	for (; !KBacktraceIterator_end(kbt); KBacktraceIterator_next(kbt)) {
+		char namebuf[KSYM_NAME_LEN+100];
+		unsigned long address = kbt->it.pc;
+
+		/* Try to acquire the mmap_sem as we pass into userspace. */
+		if (address < PAGE_OFFSET && !have_mmap_sem && kbt->task->mm)
+			have_mmap_sem =
+				down_read_trylock(&kbt->task->mm->mmap_sem);
+
+		describe_addr(kbt, address, have_mmap_sem,
+			      namebuf, sizeof(namebuf));
+
+		pr_err("  frame %d: 0x%lx %s(sp 0x%lx)\n",
+		       i++, address, namebuf, (unsigned long)(kbt->it.sp));
+
+		if (i >= 100) {
+			pr_err("Stack dump truncated (%d frames)\n", i);
+			break;
+		}
+	}
+	if (kbt->end == KBT_LOOP)
+		pr_err("Stack dump stopped; next frame identical to this one\n");
+	if (headers)
+		pr_err("Stack dump complete\n");
+	if (have_mmap_sem)
+		up_read(&kbt->task->mm->mmap_sem);
+	end_backtrace();
+}
+EXPORT_SYMBOL(tile_show_stack);
+
+
+/* This is called from show_regs() and _dump_stack() */
+void dump_stack_regs(struct pt_regs *regs)
+{
+	struct KBacktraceIterator kbt;
+	KBacktraceIterator_init(&kbt, NULL, regs);
+	tile_show_stack(&kbt, 1);
+}
+EXPORT_SYMBOL(dump_stack_regs);
+
+static struct pt_regs *regs_to_pt_regs(struct pt_regs *regs,
+				       ulong pc, ulong lr, ulong sp, ulong r52)
+{
+	memset(regs, 0, sizeof(struct pt_regs));
+	regs->pc = pc;
+	regs->lr = lr;
+	regs->sp = sp;
+	regs->regs[52] = r52;
+	return regs;
+}
+
+/* This is called from dump_stack() and just converts to pt_regs */
+void _dump_stack(int dummy, ulong pc, ulong lr, ulong sp, ulong r52)
+{
+	struct pt_regs regs;
+	dump_stack_regs(regs_to_pt_regs(&regs, pc, lr, sp, r52));
+}
+
+/* This is called from KBacktraceIterator_init_current() */
+void _KBacktraceIterator_init_current(struct KBacktraceIterator *kbt, ulong pc,
+				      ulong lr, ulong sp, ulong r52)
+{
+	struct pt_regs regs;
+	KBacktraceIterator_init(kbt, NULL,
+				regs_to_pt_regs(&regs, pc, lr, sp, r52));
+}
+
+/* This is called only from kernel/sched/core.c, with esp == NULL */
+void show_stack(struct task_struct *task, unsigned long *esp)
+{
+	struct KBacktraceIterator kbt;
+	if (task == NULL || task == current)
+		KBacktraceIterator_init_current(&kbt);
+	else
+		KBacktraceIterator_init(&kbt, task, NULL);
+	tile_show_stack(&kbt, 0);
+}
+
+#ifdef CONFIG_STACKTRACE
+
+/* Support generic Linux stack API too */
+
+void save_stack_trace_tsk(struct task_struct *task, struct stack_trace *trace)
+{
+	struct KBacktraceIterator kbt;
+	int skip = trace->skip;
+	int i = 0;
+
+	if (!start_backtrace())
+		goto done;
+	if (task == NULL || task == current)
+		KBacktraceIterator_init_current(&kbt);
+	else
+		KBacktraceIterator_init(&kbt, task, NULL);
+	for (; !KBacktraceIterator_end(&kbt); KBacktraceIterator_next(&kbt)) {
+		if (skip) {
+			--skip;
+			continue;
+		}
+		if (i >= trace->max_entries || kbt.it.pc < PAGE_OFFSET)
+			break;
+		trace->entries[i++] = kbt.it.pc;
+	}
+	end_backtrace();
+done:
+	trace->nr_entries = i;
+}
+EXPORT_SYMBOL(save_stack_trace_tsk);
+
+void save_stack_trace(struct stack_trace *trace)
+{
+	save_stack_trace_tsk(NULL, trace);
+}
+EXPORT_SYMBOL_GPL(save_stack_trace);
+
+#endif
+
+/* In entry.S */
+EXPORT_SYMBOL(KBacktraceIterator_init_current);
diff --git a/kernel/arch/tile/kernel/sys.c b/kernel/arch/tile/kernel/sys.c
new file mode 100644
index 000000000..38debe706
--- /dev/null
+++ b/kernel/arch/tile/kernel/sys.c
@@ -0,0 +1,123 @@
+/*
+ * Copyright 2010 Tilera Corporation. All Rights Reserved.
+ *
+ *   This program is free software; you can redistribute it and/or
+ *   modify it under the terms of the GNU General Public License
+ *   as published by the Free Software Foundation, version 2.
+ *
+ *   This program is distributed in the hope that it will be useful, but
+ *   WITHOUT ANY WARRANTY; without even the implied warranty of
+ *   MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
+ *   NON INFRINGEMENT.  See the GNU General Public License for
+ *   more details.
+ *
+ * This file contains various random system calls that
+ * have a non-standard calling sequence on the Linux/TILE
+ * platform.
+ */
+
+#include <linux/errno.h>
+#include <linux/sched.h>
+#include <linux/mm.h>
+#include <linux/smp.h>
+#include <linux/syscalls.h>
+#include <linux/mman.h>
+#include <linux/file.h>
+#include <linux/mempolicy.h>
+#include <linux/binfmts.h>
+#include <linux/fs.h>
+#include <linux/compat.h>
+#include <linux/uaccess.h>
+#include <linux/signal.h>
+#include <asm/syscalls.h>
+#include <asm/pgtable.h>
+#include <asm/homecache.h>
+#include <asm/cachectl.h>
+#include <arch/chip.h>
+
+SYSCALL_DEFINE3(cacheflush, unsigned long, addr, unsigned long, len,
+		unsigned long, flags)
+{
+	/* DCACHE is not particularly effective if not bound to one cpu. */
+	if (flags & DCACHE)
+		homecache_evict(cpumask_of(raw_smp_processor_id()));
+
+	if (flags & ICACHE)
+		flush_remote(0, HV_FLUSH_EVICT_L1I, mm_cpumask(current->mm),
+			     0, 0, 0, NULL, NULL, 0);
+	return 0;
+}
+
+/*
+ * Syscalls that pass 64-bit values on 32-bit systems normally
+ * pass them as (low,high) word packed into the immediately adjacent
+ * registers.  If the low word naturally falls on an even register,
+ * our ABI makes it work correctly; if not, we adjust it here.
+ * Handling it here means we don't have to fix uclibc AND glibc AND
+ * any other standard libcs we want to support.
+ */
+
+#if !defined(__tilegx__) || defined(CONFIG_COMPAT)
+
+ssize_t sys32_readahead(int fd, u32 offset_lo, u32 offset_hi, u32 count)
+{
+	return sys_readahead(fd, ((loff_t)offset_hi << 32) | offset_lo, count);
+}
+
+int sys32_fadvise64_64(int fd, u32 offset_lo, u32 offset_hi,
+		       u32 len_lo, u32 len_hi, int advice)
+{
+	return sys_fadvise64_64(fd, ((loff_t)offset_hi << 32) | offset_lo,
+				((loff_t)len_hi << 32) | len_lo, advice);
+}
+
+#endif /* 32-bit syscall wrappers */
+
+/* Note: used by the compat code even in 64-bit Linux. */
+SYSCALL_DEFINE6(mmap2, unsigned long, addr, unsigned long, len,
+		unsigned long, prot, unsigned long, flags,
+		unsigned long, fd, unsigned long, off_4k)
+{
+#define PAGE_ADJUST (PAGE_SHIFT - 12)
+	if (off_4k & ((1 << PAGE_ADJUST) - 1))
+		return -EINVAL;
+	return sys_mmap_pgoff(addr, len, prot, flags, fd,
+			      off_4k >> PAGE_ADJUST);
+}
+
+#ifdef __tilegx__
+SYSCALL_DEFINE6(mmap, unsigned long, addr, unsigned long, len,
+		unsigned long, prot, unsigned long, flags,
+		unsigned long, fd, off_t, offset)
+{
+	if (offset & ((1 << PAGE_SHIFT) - 1))
+		return -EINVAL;
+	return sys_mmap_pgoff(addr, len, prot, flags, fd,
+			      offset >> PAGE_SHIFT);
+}
+#endif
+
+
+/* Provide the actual syscall number to call mapping. */
+#undef __SYSCALL
+#define __SYSCALL(nr, call) [nr] = (call),
+
+#ifndef __tilegx__
+/* See comments at the top of the file. */
+#define sys_fadvise64_64 sys32_fadvise64_64
+#define sys_readahead sys32_readahead
+#endif
+
+/* Call the assembly trampolines where necessary. */
+#undef sys_rt_sigreturn
+#define sys_rt_sigreturn _sys_rt_sigreturn
+#define sys_clone _sys_clone
+
+/*
+ * Note that we can't include <linux/unistd.h> here since the header
+ * guard will defeat us; <asm/unistd.h> checks for __SYSCALL as well.
+ */
+void *sys_call_table[__NR_syscalls] = {
+	[0 ... __NR_syscalls-1] = sys_ni_syscall,
+#include <asm/unistd.h>
+};
diff --git a/kernel/arch/tile/kernel/sysfs.c b/kernel/arch/tile/kernel/sysfs.c
new file mode 100644
index 000000000..a3ed12f8f
--- /dev/null
+++ b/kernel/arch/tile/kernel/sysfs.c
@@ -0,0 +1,269 @@
+/*
+ * Copyright 2011 Tilera Corporation. All Rights Reserved.
+ *
+ *   This program is free software; you can redistribute it and/or
+ *   modify it under the terms of the GNU General Public License
+ *   as published by the Free Software Foundation, version 2.
+ *
+ *   This program is distributed in the hope that it will be useful, but
+ *   WITHOUT ANY WARRANTY; without even the implied warranty of
+ *   MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
+ *   NON INFRINGEMENT.  See the GNU General Public License for
+ *   more details.
+ *
+ * /sys entry support.
+ */
+
+#include <linux/device.h>
+#include <linux/cpu.h>
+#include <linux/slab.h>
+#include <linux/smp.h>
+#include <linux/stat.h>
+#include <hv/hypervisor.h>
+
+/* Return a string queried from the hypervisor, truncated to page size. */
+static ssize_t get_hv_confstr(char *page, int query)
+{
+	ssize_t n = hv_confstr(query, (unsigned long)page, PAGE_SIZE - 1);
+	n = n < 0 ? 0 : min(n, (ssize_t)PAGE_SIZE - 1) - 1;
+	if (n)
+		page[n++] = '\n';
+	page[n] = '\0';
+	return n;
+}
+
+static ssize_t chip_width_show(struct device *dev,
+			       struct device_attribute *attr,
+			       char *page)
+{
+	return sprintf(page, "%u\n", smp_width);
+}
+static DEVICE_ATTR(chip_width, 0444, chip_width_show, NULL);
+
+static ssize_t chip_height_show(struct device *dev,
+				struct device_attribute *attr,
+				char *page)
+{
+	return sprintf(page, "%u\n", smp_height);
+}
+static DEVICE_ATTR(chip_height, 0444, chip_height_show, NULL);
+
+static ssize_t chip_serial_show(struct device *dev,
+				struct device_attribute *attr,
+				char *page)
+{
+	return get_hv_confstr(page, HV_CONFSTR_CHIP_SERIAL_NUM);
+}
+static DEVICE_ATTR(chip_serial, 0444, chip_serial_show, NULL);
+
+static ssize_t chip_revision_show(struct device *dev,
+				  struct device_attribute *attr,
+				  char *page)
+{
+	return get_hv_confstr(page, HV_CONFSTR_CHIP_REV);
+}
+static DEVICE_ATTR(chip_revision, 0444, chip_revision_show, NULL);
+
+
+static ssize_t type_show(struct device *dev,
+			    struct device_attribute *attr,
+			    char *page)
+{
+	return sprintf(page, "tilera\n");
+}
+static DEVICE_ATTR(type, 0444, type_show, NULL);
+
+#define HV_CONF_ATTR(name, conf)					\
+	static ssize_t name ## _show(struct device *dev,		\
+				     struct device_attribute *attr, \
+				     char *page)			\
+	{								\
+		return get_hv_confstr(page, conf);			\
+	}								\
+	static DEVICE_ATTR(name, 0444, name ## _show, NULL);
+
+HV_CONF_ATTR(version,		HV_CONFSTR_HV_SW_VER)
+HV_CONF_ATTR(config_version,	HV_CONFSTR_HV_CONFIG_VER)
+
+HV_CONF_ATTR(board_part,	HV_CONFSTR_BOARD_PART_NUM)
+HV_CONF_ATTR(board_serial,	HV_CONFSTR_BOARD_SERIAL_NUM)
+HV_CONF_ATTR(board_revision,	HV_CONFSTR_BOARD_REV)
+HV_CONF_ATTR(board_description,	HV_CONFSTR_BOARD_DESC)
+HV_CONF_ATTR(mezz_part,		HV_CONFSTR_MEZZ_PART_NUM)
+HV_CONF_ATTR(mezz_serial,	HV_CONFSTR_MEZZ_SERIAL_NUM)
+HV_CONF_ATTR(mezz_revision,	HV_CONFSTR_MEZZ_REV)
+HV_CONF_ATTR(mezz_description,	HV_CONFSTR_MEZZ_DESC)
+HV_CONF_ATTR(cpumod_part,	HV_CONFSTR_CPUMOD_PART_NUM)
+HV_CONF_ATTR(cpumod_serial,	HV_CONFSTR_CPUMOD_SERIAL_NUM)
+HV_CONF_ATTR(cpumod_revision,	HV_CONFSTR_CPUMOD_REV)
+HV_CONF_ATTR(cpumod_description,HV_CONFSTR_CPUMOD_DESC)
+HV_CONF_ATTR(switch_control,	HV_CONFSTR_SWITCH_CONTROL)
+
+static struct attribute *board_attrs[] = {
+	&dev_attr_board_part.attr,
+	&dev_attr_board_serial.attr,
+	&dev_attr_board_revision.attr,
+	&dev_attr_board_description.attr,
+	&dev_attr_mezz_part.attr,
+	&dev_attr_mezz_serial.attr,
+	&dev_attr_mezz_revision.attr,
+	&dev_attr_mezz_description.attr,
+	&dev_attr_cpumod_part.attr,
+	&dev_attr_cpumod_serial.attr,
+	&dev_attr_cpumod_revision.attr,
+	&dev_attr_cpumod_description.attr,
+	&dev_attr_switch_control.attr,
+	NULL
+};
+
+static struct attribute_group board_attr_group = {
+	.name   = "board",
+	.attrs  = board_attrs,
+};
+
+
+static struct bin_attribute hvconfig_bin;
+
+static ssize_t
+hvconfig_bin_read(struct file *filp, struct kobject *kobj,
+		  struct bin_attribute *bin_attr,
+		  char *buf, loff_t off, size_t count)
+{
+	static size_t size;
+
+	/* Lazily learn the true size (minus the trailing NUL). */
+	if (size == 0)
+		size = hv_confstr(HV_CONFSTR_HV_CONFIG, 0, 0) - 1;
+
+	/* Check and adjust input parameters. */
+	if (off > size)
+		return -EINVAL;
+	if (count > size - off)
+		count = size - off;
+
+	if (count) {
+		/* Get a copy of the hvc and copy out the relevant portion. */
+		char *hvc;
+
+		size = off + count;
+		hvc = kmalloc(size, GFP_KERNEL);
+		if (hvc == NULL)
+			return -ENOMEM;
+		hv_confstr(HV_CONFSTR_HV_CONFIG, (unsigned long)hvc, size);
+		memcpy(buf, hvc + off, count);
+		kfree(hvc);
+	}
+
+	return count;
+}
+
+static ssize_t hv_stats_show(struct device *dev,
+			     struct device_attribute *attr,
+			     char *page)
+{
+	int cpu = dev->id;
+	long lotar = HV_XY_TO_LOTAR(cpu_x(cpu), cpu_y(cpu));
+
+	ssize_t n = hv_confstr(HV_CONFSTR_HV_STATS,
+			       (unsigned long)page, PAGE_SIZE - 1,
+			       lotar, 0);
+	n = n < 0 ? 0 : min(n, (ssize_t)PAGE_SIZE - 1);
+	page[n] = '\0';
+	return n;
+}
+
+static ssize_t hv_stats_store(struct device *dev,
+			      struct device_attribute *attr,
+			      const char *page,
+			      size_t count)
+{
+	int cpu = dev->id;
+	long lotar = HV_XY_TO_LOTAR(cpu_x(cpu), cpu_y(cpu));
+
+	ssize_t n = hv_confstr(HV_CONFSTR_HV_STATS, 0, 0, lotar, 1);
+	return n < 0 ? n : count;
+}
+
+static DEVICE_ATTR(hv_stats, 0644, hv_stats_show, hv_stats_store);
+
+static int hv_stats_device_add(struct device *dev, struct subsys_interface *sif)
+{
+	int err, cpu = dev->id;
+
+	if (!cpu_online(cpu))
+		return 0;
+
+	err = sysfs_create_file(&dev->kobj, &dev_attr_hv_stats.attr);
+
+	return err;
+}
+
+static int hv_stats_device_remove(struct device *dev,
+				  struct subsys_interface *sif)
+{
+	int cpu = dev->id;
+
+	if (!cpu_online(cpu))
+		return 0;
+
+	sysfs_remove_file(&dev->kobj, &dev_attr_hv_stats.attr);
+	return 0;
+}
+
+
+static struct subsys_interface hv_stats_interface = {
+	.name			= "hv_stats",
+	.subsys			= &cpu_subsys,
+	.add_dev		= hv_stats_device_add,
+	.remove_dev		= hv_stats_device_remove,
+};
+
+static int __init create_sysfs_entries(void)
+{
+	int err = 0;
+
+#define create_cpu_attr(name)						\
+	if (!err)							\
+		err = device_create_file(cpu_subsys.dev_root, &dev_attr_##name);
+	create_cpu_attr(chip_width);
+	create_cpu_attr(chip_height);
+	create_cpu_attr(chip_serial);
+	create_cpu_attr(chip_revision);
+
+#define create_hv_attr(name)						\
+	if (!err)							\
+		err = sysfs_create_file(hypervisor_kobj, &dev_attr_##name.attr);
+	create_hv_attr(type);
+	create_hv_attr(version);
+	create_hv_attr(config_version);
+
+	if (!err)
+		err = sysfs_create_group(hypervisor_kobj, &board_attr_group);
+
+	if (!err) {
+		sysfs_bin_attr_init(&hvconfig_bin);
+		hvconfig_bin.attr.name = "hvconfig";
+		hvconfig_bin.attr.mode = S_IRUGO;
+		hvconfig_bin.read = hvconfig_bin_read;
+		hvconfig_bin.size = PAGE_SIZE;
+		err = sysfs_create_bin_file(hypervisor_kobj, &hvconfig_bin);
+	}
+
+	if (!err) {
+		/*
+		 * Don't bother adding the hv_stats files on each CPU if
+		 * our hypervisor doesn't supply statistics.
+		 */
+		int cpu = raw_smp_processor_id();
+		long lotar = HV_XY_TO_LOTAR(cpu_x(cpu), cpu_y(cpu));
+		char dummy;
+		ssize_t n = hv_confstr(HV_CONFSTR_HV_STATS,
+				       (unsigned long) &dummy, 1,
+				       lotar, 0);
+		if (n >= 0)
+			err = subsys_interface_register(&hv_stats_interface);
+	}
+
+	return err;
+}
+subsys_initcall(create_sysfs_entries);
diff --git a/kernel/arch/tile/kernel/tile-desc_32.c b/kernel/arch/tile/kernel/tile-desc_32.c
new file mode 100644
index 000000000..dd7bd1d85
--- /dev/null
+++ b/kernel/arch/tile/kernel/tile-desc_32.c
@@ -0,0 +1,2605 @@
+/* TILEPro opcode information.
+ *
+ * Copyright 2011 Tilera Corporation. All Rights Reserved.
+ *
+ *   This program is free software; you can redistribute it and/or
+ *   modify it under the terms of the GNU General Public License
+ *   as published by the Free Software Foundation, version 2.
+ *
+ *   This program is distributed in the hope that it will be useful, but
+ *   WITHOUT ANY WARRANTY; without even the implied warranty of
+ *   MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
+ *   NON INFRINGEMENT.  See the GNU General Public License for
+ *   more details.
+ *
+ *
+ *
+ *
+ *
+ */
+
+/* This define is BFD_RELOC_##x for real bfd, or -1 for everyone else. */
+#define BFD_RELOC(x) -1
+
+/* Special registers. */
+#define TREG_LR 55
+#define TREG_SN 56
+#define TREG_ZERO 63
+
+#include <linux/stddef.h>
+#include <asm/tile-desc.h>
+
+const struct tilepro_opcode tilepro_opcodes[395] =
+{
+ { "bpt", TILEPRO_OPC_BPT, 0x2, 0, TREG_ZERO, 0,
+    { { 0, }, {  }, { 0, }, { 0, }, { 0, } },
+  },
+  { "info", TILEPRO_OPC_INFO, 0xf, 1, TREG_ZERO, 1,
+    { { 0 }, { 1 }, { 2 }, { 3 }, { 0, } },
+  },
+  { "infol", TILEPRO_OPC_INFOL, 0x3, 1, TREG_ZERO, 1,
+    { { 4 }, { 5 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "j", TILEPRO_OPC_J, 0x2, 1, TREG_ZERO, 1,
+    { { 0, }, { 6 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "jal", TILEPRO_OPC_JAL, 0x2, 1, TREG_LR, 1,
+    { { 0, }, { 6 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "move", TILEPRO_OPC_MOVE, 0xf, 2, TREG_ZERO, 1,
+    { { 7, 8 }, { 9, 10 }, { 11, 12 }, { 13, 14 }, { 0, } },
+  },
+  { "move.sn", TILEPRO_OPC_MOVE_SN, 0x3, 2, TREG_SN, 1,
+    { { 7, 8 }, { 9, 10 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "movei", TILEPRO_OPC_MOVEI, 0xf, 2, TREG_ZERO, 1,
+    { { 7, 0 }, { 9, 1 }, { 11, 2 }, { 13, 3 }, { 0, } },
+  },
+  { "movei.sn", TILEPRO_OPC_MOVEI_SN, 0x3, 2, TREG_SN, 1,
+    { { 7, 0 }, { 9, 1 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "moveli", TILEPRO_OPC_MOVELI, 0x3, 2, TREG_ZERO, 1,
+    { { 7, 4 }, { 9, 5 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "moveli.sn", TILEPRO_OPC_MOVELI_SN, 0x3, 2, TREG_SN, 1,
+    { { 7, 4 }, { 9, 5 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "movelis", TILEPRO_OPC_MOVELIS, 0x3, 2, TREG_SN, 1,
+    { { 7, 4 }, { 9, 5 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "prefetch", TILEPRO_OPC_PREFETCH, 0x12, 1, TREG_ZERO, 1,
+    { { 0, }, { 10 }, { 0, }, { 0, }, { 15 } },
+  },
+  { "raise", TILEPRO_OPC_RAISE, 0x2, 0, TREG_ZERO, 1,
+    { { 0, }, {  }, { 0, }, { 0, }, { 0, } },
+  },
+  { "add", TILEPRO_OPC_ADD, 0xf, 3, TREG_ZERO, 1,
+    { { 7, 8, 16 }, { 9, 10, 17 }, { 11, 12, 18 }, { 13, 14, 19 }, { 0, } },
+  },
+  { "add.sn", TILEPRO_OPC_ADD_SN, 0x3, 3, TREG_SN, 1,
+    { { 7, 8, 16 }, { 9, 10, 17 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "addb", TILEPRO_OPC_ADDB, 0x3, 3, TREG_ZERO, 1,
+    { { 7, 8, 16 }, { 9, 10, 17 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "addb.sn", TILEPRO_OPC_ADDB_SN, 0x3, 3, TREG_SN, 1,
+    { { 7, 8, 16 }, { 9, 10, 17 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "addbs_u", TILEPRO_OPC_ADDBS_U, 0x3, 3, TREG_ZERO, 1,
+    { { 7, 8, 16 }, { 9, 10, 17 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "addbs_u.sn", TILEPRO_OPC_ADDBS_U_SN, 0x3, 3, TREG_SN, 1,
+    { { 7, 8, 16 }, { 9, 10, 17 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "addh", TILEPRO_OPC_ADDH, 0x3, 3, TREG_ZERO, 1,
+    { { 7, 8, 16 }, { 9, 10, 17 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "addh.sn", TILEPRO_OPC_ADDH_SN, 0x3, 3, TREG_SN, 1,
+    { { 7, 8, 16 }, { 9, 10, 17 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "addhs", TILEPRO_OPC_ADDHS, 0x3, 3, TREG_ZERO, 1,
+    { { 7, 8, 16 }, { 9, 10, 17 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "addhs.sn", TILEPRO_OPC_ADDHS_SN, 0x3, 3, TREG_SN, 1,
+    { { 7, 8, 16 }, { 9, 10, 17 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "addi", TILEPRO_OPC_ADDI, 0xf, 3, TREG_ZERO, 1,
+    { { 7, 8, 0 }, { 9, 10, 1 }, { 11, 12, 2 }, { 13, 14, 3 }, { 0, } },
+  },
+  { "addi.sn", TILEPRO_OPC_ADDI_SN, 0x3, 3, TREG_SN, 1,
+    { { 7, 8, 0 }, { 9, 10, 1 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "addib", TILEPRO_OPC_ADDIB, 0x3, 3, TREG_ZERO, 1,
+    { { 7, 8, 0 }, { 9, 10, 1 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "addib.sn", TILEPRO_OPC_ADDIB_SN, 0x3, 3, TREG_SN, 1,
+    { { 7, 8, 0 }, { 9, 10, 1 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "addih", TILEPRO_OPC_ADDIH, 0x3, 3, TREG_ZERO, 1,
+    { { 7, 8, 0 }, { 9, 10, 1 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "addih.sn", TILEPRO_OPC_ADDIH_SN, 0x3, 3, TREG_SN, 1,
+    { { 7, 8, 0 }, { 9, 10, 1 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "addli", TILEPRO_OPC_ADDLI, 0x3, 3, TREG_ZERO, 1,
+    { { 7, 8, 4 }, { 9, 10, 5 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "addli.sn", TILEPRO_OPC_ADDLI_SN, 0x3, 3, TREG_SN, 1,
+    { { 7, 8, 4 }, { 9, 10, 5 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "addlis", TILEPRO_OPC_ADDLIS, 0x3, 3, TREG_SN, 1,
+    { { 7, 8, 4 }, { 9, 10, 5 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "adds", TILEPRO_OPC_ADDS, 0x3, 3, TREG_ZERO, 1,
+    { { 7, 8, 16 }, { 9, 10, 17 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "adds.sn", TILEPRO_OPC_ADDS_SN, 0x3, 3, TREG_SN, 1,
+    { { 7, 8, 16 }, { 9, 10, 17 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "adiffb_u", TILEPRO_OPC_ADIFFB_U, 0x1, 3, TREG_ZERO, 1,
+    { { 7, 8, 16 }, { 0, }, { 0, }, { 0, }, { 0, } },
+  },
+  { "adiffb_u.sn", TILEPRO_OPC_ADIFFB_U_SN, 0x1, 3, TREG_SN, 1,
+    { { 7, 8, 16 }, { 0, }, { 0, }, { 0, }, { 0, } },
+  },
+  { "adiffh", TILEPRO_OPC_ADIFFH, 0x1, 3, TREG_ZERO, 1,
+    { { 7, 8, 16 }, { 0, }, { 0, }, { 0, }, { 0, } },
+  },
+  { "adiffh.sn", TILEPRO_OPC_ADIFFH_SN, 0x1, 3, TREG_SN, 1,
+    { { 7, 8, 16 }, { 0, }, { 0, }, { 0, }, { 0, } },
+  },
+  { "and", TILEPRO_OPC_AND, 0xf, 3, TREG_ZERO, 1,
+    { { 7, 8, 16 }, { 9, 10, 17 }, { 11, 12, 18 }, { 13, 14, 19 }, { 0, } },
+  },
+  { "and.sn", TILEPRO_OPC_AND_SN, 0x3, 3, TREG_SN, 1,
+    { { 7, 8, 16 }, { 9, 10, 17 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "andi", TILEPRO_OPC_ANDI, 0xf, 3, TREG_ZERO, 1,
+    { { 7, 8, 0 }, { 9, 10, 1 }, { 11, 12, 2 }, { 13, 14, 3 }, { 0, } },
+  },
+  { "andi.sn", TILEPRO_OPC_ANDI_SN, 0x3, 3, TREG_SN, 1,
+    { { 7, 8, 0 }, { 9, 10, 1 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "auli", TILEPRO_OPC_AULI, 0x3, 3, TREG_ZERO, 1,
+    { { 7, 8, 4 }, { 9, 10, 5 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "avgb_u", TILEPRO_OPC_AVGB_U, 0x1, 3, TREG_ZERO, 1,
+    { { 7, 8, 16 }, { 0, }, { 0, }, { 0, }, { 0, } },
+  },
+  { "avgb_u.sn", TILEPRO_OPC_AVGB_U_SN, 0x1, 3, TREG_SN, 1,
+    { { 7, 8, 16 }, { 0, }, { 0, }, { 0, }, { 0, } },
+  },
+  { "avgh", TILEPRO_OPC_AVGH, 0x1, 3, TREG_ZERO, 1,
+    { { 7, 8, 16 }, { 0, }, { 0, }, { 0, }, { 0, } },
+  },
+  { "avgh.sn", TILEPRO_OPC_AVGH_SN, 0x1, 3, TREG_SN, 1,
+    { { 7, 8, 16 }, { 0, }, { 0, }, { 0, }, { 0, } },
+  },
+  { "bbns", TILEPRO_OPC_BBNS, 0x2, 2, TREG_ZERO, 1,
+    { { 0, }, { 10, 20 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "bbns.sn", TILEPRO_OPC_BBNS_SN, 0x2, 2, TREG_SN, 1,
+    { { 0, }, { 10, 20 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "bbnst", TILEPRO_OPC_BBNST, 0x2, 2, TREG_ZERO, 1,
+    { { 0, }, { 10, 20 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "bbnst.sn", TILEPRO_OPC_BBNST_SN, 0x2, 2, TREG_SN, 1,
+    { { 0, }, { 10, 20 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "bbs", TILEPRO_OPC_BBS, 0x2, 2, TREG_ZERO, 1,
+    { { 0, }, { 10, 20 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "bbs.sn", TILEPRO_OPC_BBS_SN, 0x2, 2, TREG_SN, 1,
+    { { 0, }, { 10, 20 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "bbst", TILEPRO_OPC_BBST, 0x2, 2, TREG_ZERO, 1,
+    { { 0, }, { 10, 20 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "bbst.sn", TILEPRO_OPC_BBST_SN, 0x2, 2, TREG_SN, 1,
+    { { 0, }, { 10, 20 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "bgez", TILEPRO_OPC_BGEZ, 0x2, 2, TREG_ZERO, 1,
+    { { 0, }, { 10, 20 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "bgez.sn", TILEPRO_OPC_BGEZ_SN, 0x2, 2, TREG_SN, 1,
+    { { 0, }, { 10, 20 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "bgezt", TILEPRO_OPC_BGEZT, 0x2, 2, TREG_ZERO, 1,
+    { { 0, }, { 10, 20 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "bgezt.sn", TILEPRO_OPC_BGEZT_SN, 0x2, 2, TREG_SN, 1,
+    { { 0, }, { 10, 20 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "bgz", TILEPRO_OPC_BGZ, 0x2, 2, TREG_ZERO, 1,
+    { { 0, }, { 10, 20 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "bgz.sn", TILEPRO_OPC_BGZ_SN, 0x2, 2, TREG_SN, 1,
+    { { 0, }, { 10, 20 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "bgzt", TILEPRO_OPC_BGZT, 0x2, 2, TREG_ZERO, 1,
+    { { 0, }, { 10, 20 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "bgzt.sn", TILEPRO_OPC_BGZT_SN, 0x2, 2, TREG_SN, 1,
+    { { 0, }, { 10, 20 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "bitx", TILEPRO_OPC_BITX, 0x5, 2, TREG_ZERO, 1,
+    { { 7, 8 }, { 0, }, { 11, 12 }, { 0, }, { 0, } },
+  },
+  { "bitx.sn", TILEPRO_OPC_BITX_SN, 0x1, 2, TREG_SN, 1,
+    { { 7, 8 }, { 0, }, { 0, }, { 0, }, { 0, } },
+  },
+  { "blez", TILEPRO_OPC_BLEZ, 0x2, 2, TREG_ZERO, 1,
+    { { 0, }, { 10, 20 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "blez.sn", TILEPRO_OPC_BLEZ_SN, 0x2, 2, TREG_SN, 1,
+    { { 0, }, { 10, 20 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "blezt", TILEPRO_OPC_BLEZT, 0x2, 2, TREG_ZERO, 1,
+    { { 0, }, { 10, 20 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "blezt.sn", TILEPRO_OPC_BLEZT_SN, 0x2, 2, TREG_SN, 1,
+    { { 0, }, { 10, 20 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "blz", TILEPRO_OPC_BLZ, 0x2, 2, TREG_ZERO, 1,
+    { { 0, }, { 10, 20 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "blz.sn", TILEPRO_OPC_BLZ_SN, 0x2, 2, TREG_SN, 1,
+    { { 0, }, { 10, 20 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "blzt", TILEPRO_OPC_BLZT, 0x2, 2, TREG_ZERO, 1,
+    { { 0, }, { 10, 20 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "blzt.sn", TILEPRO_OPC_BLZT_SN, 0x2, 2, TREG_SN, 1,
+    { { 0, }, { 10, 20 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "bnz", TILEPRO_OPC_BNZ, 0x2, 2, TREG_ZERO, 1,
+    { { 0, }, { 10, 20 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "bnz.sn", TILEPRO_OPC_BNZ_SN, 0x2, 2, TREG_SN, 1,
+    { { 0, }, { 10, 20 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "bnzt", TILEPRO_OPC_BNZT, 0x2, 2, TREG_ZERO, 1,
+    { { 0, }, { 10, 20 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "bnzt.sn", TILEPRO_OPC_BNZT_SN, 0x2, 2, TREG_SN, 1,
+    { { 0, }, { 10, 20 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "bytex", TILEPRO_OPC_BYTEX, 0x5, 2, TREG_ZERO, 1,
+    { { 7, 8 }, { 0, }, { 11, 12 }, { 0, }, { 0, } },
+  },
+  { "bytex.sn", TILEPRO_OPC_BYTEX_SN, 0x1, 2, TREG_SN, 1,
+    { { 7, 8 }, { 0, }, { 0, }, { 0, }, { 0, } },
+  },
+  { "bz", TILEPRO_OPC_BZ, 0x2, 2, TREG_ZERO, 1,
+    { { 0, }, { 10, 20 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "bz.sn", TILEPRO_OPC_BZ_SN, 0x2, 2, TREG_SN, 1,
+    { { 0, }, { 10, 20 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "bzt", TILEPRO_OPC_BZT, 0x2, 2, TREG_ZERO, 1,
+    { { 0, }, { 10, 20 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "bzt.sn", TILEPRO_OPC_BZT_SN, 0x2, 2, TREG_SN, 1,
+    { { 0, }, { 10, 20 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "clz", TILEPRO_OPC_CLZ, 0x5, 2, TREG_ZERO, 1,
+    { { 7, 8 }, { 0, }, { 11, 12 }, { 0, }, { 0, } },
+  },
+  { "clz.sn", TILEPRO_OPC_CLZ_SN, 0x1, 2, TREG_SN, 1,
+    { { 7, 8 }, { 0, }, { 0, }, { 0, }, { 0, } },
+  },
+  { "crc32_32", TILEPRO_OPC_CRC32_32, 0x1, 3, TREG_ZERO, 1,
+    { { 7, 8, 16 }, { 0, }, { 0, }, { 0, }, { 0, } },
+  },
+  { "crc32_32.sn", TILEPRO_OPC_CRC32_32_SN, 0x1, 3, TREG_SN, 1,
+    { { 7, 8, 16 }, { 0, }, { 0, }, { 0, }, { 0, } },
+  },
+  { "crc32_8", TILEPRO_OPC_CRC32_8, 0x1, 3, TREG_ZERO, 1,
+    { { 7, 8, 16 }, { 0, }, { 0, }, { 0, }, { 0, } },
+  },
+  { "crc32_8.sn", TILEPRO_OPC_CRC32_8_SN, 0x1, 3, TREG_SN, 1,
+    { { 7, 8, 16 }, { 0, }, { 0, }, { 0, }, { 0, } },
+  },
+  { "ctz", TILEPRO_OPC_CTZ, 0x5, 2, TREG_ZERO, 1,
+    { { 7, 8 }, { 0, }, { 11, 12 }, { 0, }, { 0, } },
+  },
+  { "ctz.sn", TILEPRO_OPC_CTZ_SN, 0x1, 2, TREG_SN, 1,
+    { { 7, 8 }, { 0, }, { 0, }, { 0, }, { 0, } },
+  },
+  { "drain", TILEPRO_OPC_DRAIN, 0x2, 0, TREG_ZERO, 0,
+    { { 0, }, {  }, { 0, }, { 0, }, { 0, } },
+  },
+  { "dtlbpr", TILEPRO_OPC_DTLBPR, 0x2, 1, TREG_ZERO, 1,
+    { { 0, }, { 10 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "dword_align", TILEPRO_OPC_DWORD_ALIGN, 0x1, 3, TREG_ZERO, 1,
+    { { 21, 8, 16 }, { 0, }, { 0, }, { 0, }, { 0, } },
+  },
+  { "dword_align.sn", TILEPRO_OPC_DWORD_ALIGN_SN, 0x1, 3, TREG_SN, 1,
+    { { 21, 8, 16 }, { 0, }, { 0, }, { 0, }, { 0, } },
+  },
+  { "finv", TILEPRO_OPC_FINV, 0x2, 1, TREG_ZERO, 1,
+    { { 0, }, { 10 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "flush", TILEPRO_OPC_FLUSH, 0x2, 1, TREG_ZERO, 1,
+    { { 0, }, { 10 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "fnop", TILEPRO_OPC_FNOP, 0xf, 0, TREG_ZERO, 1,
+    { {  }, {  }, {  }, {  }, { 0, } },
+  },
+  { "icoh", TILEPRO_OPC_ICOH, 0x2, 1, TREG_ZERO, 1,
+    { { 0, }, { 10 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "ill", TILEPRO_OPC_ILL, 0xa, 0, TREG_ZERO, 1,
+    { { 0, }, {  }, { 0, }, {  }, { 0, } },
+  },
+  { "inthb", TILEPRO_OPC_INTHB, 0x3, 3, TREG_ZERO, 1,
+    { { 7, 8, 16 }, { 9, 10, 17 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "inthb.sn", TILEPRO_OPC_INTHB_SN, 0x3, 3, TREG_SN, 1,
+    { { 7, 8, 16 }, { 9, 10, 17 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "inthh", TILEPRO_OPC_INTHH, 0x3, 3, TREG_ZERO, 1,
+    { { 7, 8, 16 }, { 9, 10, 17 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "inthh.sn", TILEPRO_OPC_INTHH_SN, 0x3, 3, TREG_SN, 1,
+    { { 7, 8, 16 }, { 9, 10, 17 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "intlb", TILEPRO_OPC_INTLB, 0x3, 3, TREG_ZERO, 1,
+    { { 7, 8, 16 }, { 9, 10, 17 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "intlb.sn", TILEPRO_OPC_INTLB_SN, 0x3, 3, TREG_SN, 1,
+    { { 7, 8, 16 }, { 9, 10, 17 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "intlh", TILEPRO_OPC_INTLH, 0x3, 3, TREG_ZERO, 1,
+    { { 7, 8, 16 }, { 9, 10, 17 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "intlh.sn", TILEPRO_OPC_INTLH_SN, 0x3, 3, TREG_SN, 1,
+    { { 7, 8, 16 }, { 9, 10, 17 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "inv", TILEPRO_OPC_INV, 0x2, 1, TREG_ZERO, 1,
+    { { 0, }, { 10 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "iret", TILEPRO_OPC_IRET, 0x2, 0, TREG_ZERO, 1,
+    { { 0, }, {  }, { 0, }, { 0, }, { 0, } },
+  },
+  { "jalb", TILEPRO_OPC_JALB, 0x2, 1, TREG_LR, 1,
+    { { 0, }, { 22 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "jalf", TILEPRO_OPC_JALF, 0x2, 1, TREG_LR, 1,
+    { { 0, }, { 22 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "jalr", TILEPRO_OPC_JALR, 0x2, 1, TREG_LR, 1,
+    { { 0, }, { 10 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "jalrp", TILEPRO_OPC_JALRP, 0x2, 1, TREG_LR, 1,
+    { { 0, }, { 10 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "jb", TILEPRO_OPC_JB, 0x2, 1, TREG_ZERO, 1,
+    { { 0, }, { 22 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "jf", TILEPRO_OPC_JF, 0x2, 1, TREG_ZERO, 1,
+    { { 0, }, { 22 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "jr", TILEPRO_OPC_JR, 0x2, 1, TREG_ZERO, 1,
+    { { 0, }, { 10 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "jrp", TILEPRO_OPC_JRP, 0x2, 1, TREG_ZERO, 1,
+    { { 0, }, { 10 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "lb", TILEPRO_OPC_LB, 0x12, 2, TREG_ZERO, 1,
+    { { 0, }, { 9, 10 }, { 0, }, { 0, }, { 23, 15 } },
+  },
+  { "lb.sn", TILEPRO_OPC_LB_SN, 0x2, 2, TREG_SN, 1,
+    { { 0, }, { 9, 10 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "lb_u", TILEPRO_OPC_LB_U, 0x12, 2, TREG_ZERO, 1,
+    { { 0, }, { 9, 10 }, { 0, }, { 0, }, { 23, 15 } },
+  },
+  { "lb_u.sn", TILEPRO_OPC_LB_U_SN, 0x2, 2, TREG_SN, 1,
+    { { 0, }, { 9, 10 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "lbadd", TILEPRO_OPC_LBADD, 0x2, 3, TREG_ZERO, 1,
+    { { 0, }, { 9, 24, 1 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "lbadd.sn", TILEPRO_OPC_LBADD_SN, 0x2, 3, TREG_SN, 1,
+    { { 0, }, { 9, 24, 1 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "lbadd_u", TILEPRO_OPC_LBADD_U, 0x2, 3, TREG_ZERO, 1,
+    { { 0, }, { 9, 24, 1 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "lbadd_u.sn", TILEPRO_OPC_LBADD_U_SN, 0x2, 3, TREG_SN, 1,
+    { { 0, }, { 9, 24, 1 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "lh", TILEPRO_OPC_LH, 0x12, 2, TREG_ZERO, 1,
+    { { 0, }, { 9, 10 }, { 0, }, { 0, }, { 23, 15 } },
+  },
+  { "lh.sn", TILEPRO_OPC_LH_SN, 0x2, 2, TREG_SN, 1,
+    { { 0, }, { 9, 10 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "lh_u", TILEPRO_OPC_LH_U, 0x12, 2, TREG_ZERO, 1,
+    { { 0, }, { 9, 10 }, { 0, }, { 0, }, { 23, 15 } },
+  },
+  { "lh_u.sn", TILEPRO_OPC_LH_U_SN, 0x2, 2, TREG_SN, 1,
+    { { 0, }, { 9, 10 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "lhadd", TILEPRO_OPC_LHADD, 0x2, 3, TREG_ZERO, 1,
+    { { 0, }, { 9, 24, 1 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "lhadd.sn", TILEPRO_OPC_LHADD_SN, 0x2, 3, TREG_SN, 1,
+    { { 0, }, { 9, 24, 1 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "lhadd_u", TILEPRO_OPC_LHADD_U, 0x2, 3, TREG_ZERO, 1,
+    { { 0, }, { 9, 24, 1 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "lhadd_u.sn", TILEPRO_OPC_LHADD_U_SN, 0x2, 3, TREG_SN, 1,
+    { { 0, }, { 9, 24, 1 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "lnk", TILEPRO_OPC_LNK, 0x2, 1, TREG_ZERO, 1,
+    { { 0, }, { 9 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "lnk.sn", TILEPRO_OPC_LNK_SN, 0x2, 1, TREG_SN, 1,
+    { { 0, }, { 9 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "lw", TILEPRO_OPC_LW, 0x12, 2, TREG_ZERO, 1,
+    { { 0, }, { 9, 10 }, { 0, }, { 0, }, { 23, 15 } },
+  },
+  { "lw.sn", TILEPRO_OPC_LW_SN, 0x2, 2, TREG_SN, 1,
+    { { 0, }, { 9, 10 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "lw_na", TILEPRO_OPC_LW_NA, 0x2, 2, TREG_ZERO, 1,
+    { { 0, }, { 9, 10 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "lw_na.sn", TILEPRO_OPC_LW_NA_SN, 0x2, 2, TREG_SN, 1,
+    { { 0, }, { 9, 10 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "lwadd", TILEPRO_OPC_LWADD, 0x2, 3, TREG_ZERO, 1,
+    { { 0, }, { 9, 24, 1 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "lwadd.sn", TILEPRO_OPC_LWADD_SN, 0x2, 3, TREG_SN, 1,
+    { { 0, }, { 9, 24, 1 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "lwadd_na", TILEPRO_OPC_LWADD_NA, 0x2, 3, TREG_ZERO, 1,
+    { { 0, }, { 9, 24, 1 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "lwadd_na.sn", TILEPRO_OPC_LWADD_NA_SN, 0x2, 3, TREG_SN, 1,
+    { { 0, }, { 9, 24, 1 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "maxb_u", TILEPRO_OPC_MAXB_U, 0x3, 3, TREG_ZERO, 1,
+    { { 7, 8, 16 }, { 9, 10, 17 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "maxb_u.sn", TILEPRO_OPC_MAXB_U_SN, 0x3, 3, TREG_SN, 1,
+    { { 7, 8, 16 }, { 9, 10, 17 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "maxh", TILEPRO_OPC_MAXH, 0x3, 3, TREG_ZERO, 1,
+    { { 7, 8, 16 }, { 9, 10, 17 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "maxh.sn", TILEPRO_OPC_MAXH_SN, 0x3, 3, TREG_SN, 1,
+    { { 7, 8, 16 }, { 9, 10, 17 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "maxib_u", TILEPRO_OPC_MAXIB_U, 0x3, 3, TREG_ZERO, 1,
+    { { 7, 8, 0 }, { 9, 10, 1 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "maxib_u.sn", TILEPRO_OPC_MAXIB_U_SN, 0x3, 3, TREG_SN, 1,
+    { { 7, 8, 0 }, { 9, 10, 1 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "maxih", TILEPRO_OPC_MAXIH, 0x3, 3, TREG_ZERO, 1,
+    { { 7, 8, 0 }, { 9, 10, 1 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "maxih.sn", TILEPRO_OPC_MAXIH_SN, 0x3, 3, TREG_SN, 1,
+    { { 7, 8, 0 }, { 9, 10, 1 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "mf", TILEPRO_OPC_MF, 0x2, 0, TREG_ZERO, 1,
+    { { 0, }, {  }, { 0, }, { 0, }, { 0, } },
+  },
+  { "mfspr", TILEPRO_OPC_MFSPR, 0x2, 2, TREG_ZERO, 1,
+    { { 0, }, { 9, 25 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "minb_u", TILEPRO_OPC_MINB_U, 0x3, 3, TREG_ZERO, 1,
+    { { 7, 8, 16 }, { 9, 10, 17 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "minb_u.sn", TILEPRO_OPC_MINB_U_SN, 0x3, 3, TREG_SN, 1,
+    { { 7, 8, 16 }, { 9, 10, 17 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "minh", TILEPRO_OPC_MINH, 0x3, 3, TREG_ZERO, 1,
+    { { 7, 8, 16 }, { 9, 10, 17 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "minh.sn", TILEPRO_OPC_MINH_SN, 0x3, 3, TREG_SN, 1,
+    { { 7, 8, 16 }, { 9, 10, 17 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "minib_u", TILEPRO_OPC_MINIB_U, 0x3, 3, TREG_ZERO, 1,
+    { { 7, 8, 0 }, { 9, 10, 1 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "minib_u.sn", TILEPRO_OPC_MINIB_U_SN, 0x3, 3, TREG_SN, 1,
+    { { 7, 8, 0 }, { 9, 10, 1 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "minih", TILEPRO_OPC_MINIH, 0x3, 3, TREG_ZERO, 1,
+    { { 7, 8, 0 }, { 9, 10, 1 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "minih.sn", TILEPRO_OPC_MINIH_SN, 0x3, 3, TREG_SN, 1,
+    { { 7, 8, 0 }, { 9, 10, 1 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "mm", TILEPRO_OPC_MM, 0x3, 5, TREG_ZERO, 1,
+    { { 7, 8, 16, 26, 27 }, { 9, 10, 17, 28, 29 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "mnz", TILEPRO_OPC_MNZ, 0xf, 3, TREG_ZERO, 1,
+    { { 7, 8, 16 }, { 9, 10, 17 }, { 11, 12, 18 }, { 13, 14, 19 }, { 0, } },
+  },
+  { "mnz.sn", TILEPRO_OPC_MNZ_SN, 0x3, 3, TREG_SN, 1,
+    { { 7, 8, 16 }, { 9, 10, 17 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "mnzb", TILEPRO_OPC_MNZB, 0x3, 3, TREG_ZERO, 1,
+    { { 7, 8, 16 }, { 9, 10, 17 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "mnzb.sn", TILEPRO_OPC_MNZB_SN, 0x3, 3, TREG_SN, 1,
+    { { 7, 8, 16 }, { 9, 10, 17 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "mnzh", TILEPRO_OPC_MNZH, 0x3, 3, TREG_ZERO, 1,
+    { { 7, 8, 16 }, { 9, 10, 17 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "mnzh.sn", TILEPRO_OPC_MNZH_SN, 0x3, 3, TREG_SN, 1,
+    { { 7, 8, 16 }, { 9, 10, 17 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "mtspr", TILEPRO_OPC_MTSPR, 0x2, 2, TREG_ZERO, 1,
+    { { 0, }, { 30, 10 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "mulhh_ss", TILEPRO_OPC_MULHH_SS, 0x5, 3, TREG_ZERO, 1,
+    { { 7, 8, 16 }, { 0, }, { 11, 12, 18 }, { 0, }, { 0, } },
+  },
+  { "mulhh_ss.sn", TILEPRO_OPC_MULHH_SS_SN, 0x1, 3, TREG_SN, 1,
+    { { 7, 8, 16 }, { 0, }, { 0, }, { 0, }, { 0, } },
+  },
+  { "mulhh_su", TILEPRO_OPC_MULHH_SU, 0x1, 3, TREG_ZERO, 1,
+    { { 7, 8, 16 }, { 0, }, { 0, }, { 0, }, { 0, } },
+  },
+  { "mulhh_su.sn", TILEPRO_OPC_MULHH_SU_SN, 0x1, 3, TREG_SN, 1,
+    { { 7, 8, 16 }, { 0, }, { 0, }, { 0, }, { 0, } },
+  },
+  { "mulhh_uu", TILEPRO_OPC_MULHH_UU, 0x5, 3, TREG_ZERO, 1,
+    { { 7, 8, 16 }, { 0, }, { 11, 12, 18 }, { 0, }, { 0, } },
+  },
+  { "mulhh_uu.sn", TILEPRO_OPC_MULHH_UU_SN, 0x1, 3, TREG_SN, 1,
+    { { 7, 8, 16 }, { 0, }, { 0, }, { 0, }, { 0, } },
+  },
+  { "mulhha_ss", TILEPRO_OPC_MULHHA_SS, 0x5, 3, TREG_ZERO, 1,
+    { { 21, 8, 16 }, { 0, }, { 31, 12, 18 }, { 0, }, { 0, } },
+  },
+  { "mulhha_ss.sn", TILEPRO_OPC_MULHHA_SS_SN, 0x1, 3, TREG_SN, 1,
+    { { 21, 8, 16 }, { 0, }, { 0, }, { 0, }, { 0, } },
+  },
+  { "mulhha_su", TILEPRO_OPC_MULHHA_SU, 0x1, 3, TREG_ZERO, 1,
+    { { 21, 8, 16 }, { 0, }, { 0, }, { 0, }, { 0, } },
+  },
+  { "mulhha_su.sn", TILEPRO_OPC_MULHHA_SU_SN, 0x1, 3, TREG_SN, 1,
+    { { 21, 8, 16 }, { 0, }, { 0, }, { 0, }, { 0, } },
+  },
+  { "mulhha_uu", TILEPRO_OPC_MULHHA_UU, 0x5, 3, TREG_ZERO, 1,
+    { { 21, 8, 16 }, { 0, }, { 31, 12, 18 }, { 0, }, { 0, } },
+  },
+  { "mulhha_uu.sn", TILEPRO_OPC_MULHHA_UU_SN, 0x1, 3, TREG_SN, 1,
+    { { 21, 8, 16 }, { 0, }, { 0, }, { 0, }, { 0, } },
+  },
+  { "mulhhsa_uu", TILEPRO_OPC_MULHHSA_UU, 0x1, 3, TREG_ZERO, 1,
+    { { 21, 8, 16 }, { 0, }, { 0, }, { 0, }, { 0, } },
+  },
+  { "mulhhsa_uu.sn", TILEPRO_OPC_MULHHSA_UU_SN, 0x1, 3, TREG_SN, 1,
+    { { 21, 8, 16 }, { 0, }, { 0, }, { 0, }, { 0, } },
+  },
+  { "mulhl_ss", TILEPRO_OPC_MULHL_SS, 0x1, 3, TREG_ZERO, 1,
+    { { 7, 8, 16 }, { 0, }, { 0, }, { 0, }, { 0, } },
+  },
+  { "mulhl_ss.sn", TILEPRO_OPC_MULHL_SS_SN, 0x1, 3, TREG_SN, 1,
+    { { 7, 8, 16 }, { 0, }, { 0, }, { 0, }, { 0, } },
+  },
+  { "mulhl_su", TILEPRO_OPC_MULHL_SU, 0x1, 3, TREG_ZERO, 1,
+    { { 7, 8, 16 }, { 0, }, { 0, }, { 0, }, { 0, } },
+  },
+  { "mulhl_su.sn", TILEPRO_OPC_MULHL_SU_SN, 0x1, 3, TREG_SN, 1,
+    { { 7, 8, 16 }, { 0, }, { 0, }, { 0, }, { 0, } },
+  },
+  { "mulhl_us", TILEPRO_OPC_MULHL_US, 0x1, 3, TREG_ZERO, 1,
+    { { 7, 8, 16 }, { 0, }, { 0, }, { 0, }, { 0, } },
+  },
+  { "mulhl_us.sn", TILEPRO_OPC_MULHL_US_SN, 0x1, 3, TREG_SN, 1,
+    { { 7, 8, 16 }, { 0, }, { 0, }, { 0, }, { 0, } },
+  },
+  { "mulhl_uu", TILEPRO_OPC_MULHL_UU, 0x1, 3, TREG_ZERO, 1,
+    { { 7, 8, 16 }, { 0, }, { 0, }, { 0, }, { 0, } },
+  },
+  { "mulhl_uu.sn", TILEPRO_OPC_MULHL_UU_SN, 0x1, 3, TREG_SN, 1,
+    { { 7, 8, 16 }, { 0, }, { 0, }, { 0, }, { 0, } },
+  },
+  { "mulhla_ss", TILEPRO_OPC_MULHLA_SS, 0x1, 3, TREG_ZERO, 1,
+    { { 21, 8, 16 }, { 0, }, { 0, }, { 0, }, { 0, } },
+  },
+  { "mulhla_ss.sn", TILEPRO_OPC_MULHLA_SS_SN, 0x1, 3, TREG_SN, 1,
+    { { 21, 8, 16 }, { 0, }, { 0, }, { 0, }, { 0, } },
+  },
+  { "mulhla_su", TILEPRO_OPC_MULHLA_SU, 0x1, 3, TREG_ZERO, 1,
+    { { 21, 8, 16 }, { 0, }, { 0, }, { 0, }, { 0, } },
+  },
+  { "mulhla_su.sn", TILEPRO_OPC_MULHLA_SU_SN, 0x1, 3, TREG_SN, 1,
+    { { 21, 8, 16 }, { 0, }, { 0, }, { 0, }, { 0, } },
+  },
+  { "mulhla_us", TILEPRO_OPC_MULHLA_US, 0x1, 3, TREG_ZERO, 1,
+    { { 21, 8, 16 }, { 0, }, { 0, }, { 0, }, { 0, } },
+  },
+  { "mulhla_us.sn", TILEPRO_OPC_MULHLA_US_SN, 0x1, 3, TREG_SN, 1,
+    { { 21, 8, 16 }, { 0, }, { 0, }, { 0, }, { 0, } },
+  },
+  { "mulhla_uu", TILEPRO_OPC_MULHLA_UU, 0x1, 3, TREG_ZERO, 1,
+    { { 21, 8, 16 }, { 0, }, { 0, }, { 0, }, { 0, } },
+  },
+  { "mulhla_uu.sn", TILEPRO_OPC_MULHLA_UU_SN, 0x1, 3, TREG_SN, 1,
+    { { 21, 8, 16 }, { 0, }, { 0, }, { 0, }, { 0, } },
+  },
+  { "mulhlsa_uu", TILEPRO_OPC_MULHLSA_UU, 0x5, 3, TREG_ZERO, 1,
+    { { 21, 8, 16 }, { 0, }, { 31, 12, 18 }, { 0, }, { 0, } },
+  },
+  { "mulhlsa_uu.sn", TILEPRO_OPC_MULHLSA_UU_SN, 0x1, 3, TREG_SN, 1,
+    { { 21, 8, 16 }, { 0, }, { 0, }, { 0, }, { 0, } },
+  },
+  { "mulll_ss", TILEPRO_OPC_MULLL_SS, 0x5, 3, TREG_ZERO, 1,
+    { { 7, 8, 16 }, { 0, }, { 11, 12, 18 }, { 0, }, { 0, } },
+  },
+  { "mulll_ss.sn", TILEPRO_OPC_MULLL_SS_SN, 0x1, 3, TREG_SN, 1,
+    { { 7, 8, 16 }, { 0, }, { 0, }, { 0, }, { 0, } },
+  },
+  { "mulll_su", TILEPRO_OPC_MULLL_SU, 0x1, 3, TREG_ZERO, 1,
+    { { 7, 8, 16 }, { 0, }, { 0, }, { 0, }, { 0, } },
+  },
+  { "mulll_su.sn", TILEPRO_OPC_MULLL_SU_SN, 0x1, 3, TREG_SN, 1,
+    { { 7, 8, 16 }, { 0, }, { 0, }, { 0, }, { 0, } },
+  },
+  { "mulll_uu", TILEPRO_OPC_MULLL_UU, 0x5, 3, TREG_ZERO, 1,
+    { { 7, 8, 16 }, { 0, }, { 11, 12, 18 }, { 0, }, { 0, } },
+  },
+  { "mulll_uu.sn", TILEPRO_OPC_MULLL_UU_SN, 0x1, 3, TREG_SN, 1,
+    { { 7, 8, 16 }, { 0, }, { 0, }, { 0, }, { 0, } },
+  },
+  { "mullla_ss", TILEPRO_OPC_MULLLA_SS, 0x5, 3, TREG_ZERO, 1,
+    { { 21, 8, 16 }, { 0, }, { 31, 12, 18 }, { 0, }, { 0, } },
+  },
+  { "mullla_ss.sn", TILEPRO_OPC_MULLLA_SS_SN, 0x1, 3, TREG_SN, 1,
+    { { 21, 8, 16 }, { 0, }, { 0, }, { 0, }, { 0, } },
+  },
+  { "mullla_su", TILEPRO_OPC_MULLLA_SU, 0x1, 3, TREG_ZERO, 1,
+    { { 21, 8, 16 }, { 0, }, { 0, }, { 0, }, { 0, } },
+  },
+  { "mullla_su.sn", TILEPRO_OPC_MULLLA_SU_SN, 0x1, 3, TREG_SN, 1,
+    { { 21, 8, 16 }, { 0, }, { 0, }, { 0, }, { 0, } },
+  },
+  { "mullla_uu", TILEPRO_OPC_MULLLA_UU, 0x5, 3, TREG_ZERO, 1,
+    { { 21, 8, 16 }, { 0, }, { 31, 12, 18 }, { 0, }, { 0, } },
+  },
+  { "mullla_uu.sn", TILEPRO_OPC_MULLLA_UU_SN, 0x1, 3, TREG_SN, 1,
+    { { 21, 8, 16 }, { 0, }, { 0, }, { 0, }, { 0, } },
+  },
+  { "mulllsa_uu", TILEPRO_OPC_MULLLSA_UU, 0x1, 3, TREG_ZERO, 1,
+    { { 21, 8, 16 }, { 0, }, { 0, }, { 0, }, { 0, } },
+  },
+  { "mulllsa_uu.sn", TILEPRO_OPC_MULLLSA_UU_SN, 0x1, 3, TREG_SN, 1,
+    { { 21, 8, 16 }, { 0, }, { 0, }, { 0, }, { 0, } },
+  },
+  { "mvnz", TILEPRO_OPC_MVNZ, 0x5, 3, TREG_ZERO, 1,
+    { { 21, 8, 16 }, { 0, }, { 31, 12, 18 }, { 0, }, { 0, } },
+  },
+  { "mvnz.sn", TILEPRO_OPC_MVNZ_SN, 0x1, 3, TREG_SN, 1,
+    { { 21, 8, 16 }, { 0, }, { 0, }, { 0, }, { 0, } },
+  },
+  { "mvz", TILEPRO_OPC_MVZ, 0x5, 3, TREG_ZERO, 1,
+    { { 21, 8, 16 }, { 0, }, { 31, 12, 18 }, { 0, }, { 0, } },
+  },
+  { "mvz.sn", TILEPRO_OPC_MVZ_SN, 0x1, 3, TREG_SN, 1,
+    { { 21, 8, 16 }, { 0, }, { 0, }, { 0, }, { 0, } },
+  },
+  { "mz", TILEPRO_OPC_MZ, 0xf, 3, TREG_ZERO, 1,
+    { { 7, 8, 16 }, { 9, 10, 17 }, { 11, 12, 18 }, { 13, 14, 19 }, { 0, } },
+  },
+  { "mz.sn", TILEPRO_OPC_MZ_SN, 0x3, 3, TREG_SN, 1,
+    { { 7, 8, 16 }, { 9, 10, 17 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "mzb", TILEPRO_OPC_MZB, 0x3, 3, TREG_ZERO, 1,
+    { { 7, 8, 16 }, { 9, 10, 17 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "mzb.sn", TILEPRO_OPC_MZB_SN, 0x3, 3, TREG_SN, 1,
+    { { 7, 8, 16 }, { 9, 10, 17 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "mzh", TILEPRO_OPC_MZH, 0x3, 3, TREG_ZERO, 1,
+    { { 7, 8, 16 }, { 9, 10, 17 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "mzh.sn", TILEPRO_OPC_MZH_SN, 0x3, 3, TREG_SN, 1,
+    { { 7, 8, 16 }, { 9, 10, 17 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "nap", TILEPRO_OPC_NAP, 0x2, 0, TREG_ZERO, 0,
+    { { 0, }, {  }, { 0, }, { 0, }, { 0, } },
+  },
+  { "nop", TILEPRO_OPC_NOP, 0xf, 0, TREG_ZERO, 1,
+    { {  }, {  }, {  }, {  }, { 0, } },
+  },
+  { "nor", TILEPRO_OPC_NOR, 0xf, 3, TREG_ZERO, 1,
+    { { 7, 8, 16 }, { 9, 10, 17 }, { 11, 12, 18 }, { 13, 14, 19 }, { 0, } },
+  },
+  { "nor.sn", TILEPRO_OPC_NOR_SN, 0x3, 3, TREG_SN, 1,
+    { { 7, 8, 16 }, { 9, 10, 17 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "or", TILEPRO_OPC_OR, 0xf, 3, TREG_ZERO, 1,
+    { { 7, 8, 16 }, { 9, 10, 17 }, { 11, 12, 18 }, { 13, 14, 19 }, { 0, } },
+  },
+  { "or.sn", TILEPRO_OPC_OR_SN, 0x3, 3, TREG_SN, 1,
+    { { 7, 8, 16 }, { 9, 10, 17 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "ori", TILEPRO_OPC_ORI, 0xf, 3, TREG_ZERO, 1,
+    { { 7, 8, 0 }, { 9, 10, 1 }, { 11, 12, 2 }, { 13, 14, 3 }, { 0, } },
+  },
+  { "ori.sn", TILEPRO_OPC_ORI_SN, 0x3, 3, TREG_SN, 1,
+    { { 7, 8, 0 }, { 9, 10, 1 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "packbs_u", TILEPRO_OPC_PACKBS_U, 0x3, 3, TREG_ZERO, 1,
+    { { 7, 8, 16 }, { 9, 10, 17 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "packbs_u.sn", TILEPRO_OPC_PACKBS_U_SN, 0x3, 3, TREG_SN, 1,
+    { { 7, 8, 16 }, { 9, 10, 17 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "packhb", TILEPRO_OPC_PACKHB, 0x3, 3, TREG_ZERO, 1,
+    { { 7, 8, 16 }, { 9, 10, 17 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "packhb.sn", TILEPRO_OPC_PACKHB_SN, 0x3, 3, TREG_SN, 1,
+    { { 7, 8, 16 }, { 9, 10, 17 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "packhs", TILEPRO_OPC_PACKHS, 0x3, 3, TREG_ZERO, 1,
+    { { 7, 8, 16 }, { 9, 10, 17 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "packhs.sn", TILEPRO_OPC_PACKHS_SN, 0x3, 3, TREG_SN, 1,
+    { { 7, 8, 16 }, { 9, 10, 17 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "packlb", TILEPRO_OPC_PACKLB, 0x3, 3, TREG_ZERO, 1,
+    { { 7, 8, 16 }, { 9, 10, 17 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "packlb.sn", TILEPRO_OPC_PACKLB_SN, 0x3, 3, TREG_SN, 1,
+    { { 7, 8, 16 }, { 9, 10, 17 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "pcnt", TILEPRO_OPC_PCNT, 0x5, 2, TREG_ZERO, 1,
+    { { 7, 8 }, { 0, }, { 11, 12 }, { 0, }, { 0, } },
+  },
+  { "pcnt.sn", TILEPRO_OPC_PCNT_SN, 0x1, 2, TREG_SN, 1,
+    { { 7, 8 }, { 0, }, { 0, }, { 0, }, { 0, } },
+  },
+  { "rl", TILEPRO_OPC_RL, 0xf, 3, TREG_ZERO, 1,
+    { { 7, 8, 16 }, { 9, 10, 17 }, { 11, 12, 18 }, { 13, 14, 19 }, { 0, } },
+  },
+  { "rl.sn", TILEPRO_OPC_RL_SN, 0x3, 3, TREG_SN, 1,
+    { { 7, 8, 16 }, { 9, 10, 17 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "rli", TILEPRO_OPC_RLI, 0xf, 3, TREG_ZERO, 1,
+    { { 7, 8, 32 }, { 9, 10, 33 }, { 11, 12, 34 }, { 13, 14, 35 }, { 0, } },
+  },
+  { "rli.sn", TILEPRO_OPC_RLI_SN, 0x3, 3, TREG_SN, 1,
+    { { 7, 8, 32 }, { 9, 10, 33 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "s1a", TILEPRO_OPC_S1A, 0xf, 3, TREG_ZERO, 1,
+    { { 7, 8, 16 }, { 9, 10, 17 }, { 11, 12, 18 }, { 13, 14, 19 }, { 0, } },
+  },
+  { "s1a.sn", TILEPRO_OPC_S1A_SN, 0x3, 3, TREG_SN, 1,
+    { { 7, 8, 16 }, { 9, 10, 17 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "s2a", TILEPRO_OPC_S2A, 0xf, 3, TREG_ZERO, 1,
+    { { 7, 8, 16 }, { 9, 10, 17 }, { 11, 12, 18 }, { 13, 14, 19 }, { 0, } },
+  },
+  { "s2a.sn", TILEPRO_OPC_S2A_SN, 0x3, 3, TREG_SN, 1,
+    { { 7, 8, 16 }, { 9, 10, 17 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "s3a", TILEPRO_OPC_S3A, 0xf, 3, TREG_ZERO, 1,
+    { { 7, 8, 16 }, { 9, 10, 17 }, { 11, 12, 18 }, { 13, 14, 19 }, { 0, } },
+  },
+  { "s3a.sn", TILEPRO_OPC_S3A_SN, 0x3, 3, TREG_SN, 1,
+    { { 7, 8, 16 }, { 9, 10, 17 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "sadab_u", TILEPRO_OPC_SADAB_U, 0x1, 3, TREG_ZERO, 1,
+    { { 21, 8, 16 }, { 0, }, { 0, }, { 0, }, { 0, } },
+  },
+  { "sadab_u.sn", TILEPRO_OPC_SADAB_U_SN, 0x1, 3, TREG_SN, 1,
+    { { 21, 8, 16 }, { 0, }, { 0, }, { 0, }, { 0, } },
+  },
+  { "sadah", TILEPRO_OPC_SADAH, 0x1, 3, TREG_ZERO, 1,
+    { { 21, 8, 16 }, { 0, }, { 0, }, { 0, }, { 0, } },
+  },
+  { "sadah.sn", TILEPRO_OPC_SADAH_SN, 0x1, 3, TREG_SN, 1,
+    { { 21, 8, 16 }, { 0, }, { 0, }, { 0, }, { 0, } },
+  },
+  { "sadah_u", TILEPRO_OPC_SADAH_U, 0x1, 3, TREG_ZERO, 1,
+    { { 21, 8, 16 }, { 0, }, { 0, }, { 0, }, { 0, } },
+  },
+  { "sadah_u.sn", TILEPRO_OPC_SADAH_U_SN, 0x1, 3, TREG_SN, 1,
+    { { 21, 8, 16 }, { 0, }, { 0, }, { 0, }, { 0, } },
+  },
+  { "sadb_u", TILEPRO_OPC_SADB_U, 0x1, 3, TREG_ZERO, 1,
+    { { 7, 8, 16 }, { 0, }, { 0, }, { 0, }, { 0, } },
+  },
+  { "sadb_u.sn", TILEPRO_OPC_SADB_U_SN, 0x1, 3, TREG_SN, 1,
+    { { 7, 8, 16 }, { 0, }, { 0, }, { 0, }, { 0, } },
+  },
+  { "sadh", TILEPRO_OPC_SADH, 0x1, 3, TREG_ZERO, 1,
+    { { 7, 8, 16 }, { 0, }, { 0, }, { 0, }, { 0, } },
+  },
+  { "sadh.sn", TILEPRO_OPC_SADH_SN, 0x1, 3, TREG_SN, 1,
+    { { 7, 8, 16 }, { 0, }, { 0, }, { 0, }, { 0, } },
+  },
+  { "sadh_u", TILEPRO_OPC_SADH_U, 0x1, 3, TREG_ZERO, 1,
+    { { 7, 8, 16 }, { 0, }, { 0, }, { 0, }, { 0, } },
+  },
+  { "sadh_u.sn", TILEPRO_OPC_SADH_U_SN, 0x1, 3, TREG_SN, 1,
+    { { 7, 8, 16 }, { 0, }, { 0, }, { 0, }, { 0, } },
+  },
+  { "sb", TILEPRO_OPC_SB, 0x12, 2, TREG_ZERO, 1,
+    { { 0, }, { 10, 17 }, { 0, }, { 0, }, { 15, 36 } },
+  },
+  { "sbadd", TILEPRO_OPC_SBADD, 0x2, 3, TREG_ZERO, 1,
+    { { 0, }, { 24, 17, 37 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "seq", TILEPRO_OPC_SEQ, 0xf, 3, TREG_ZERO, 1,
+    { { 7, 8, 16 }, { 9, 10, 17 }, { 11, 12, 18 }, { 13, 14, 19 }, { 0, } },
+  },
+  { "seq.sn", TILEPRO_OPC_SEQ_SN, 0x3, 3, TREG_SN, 1,
+    { { 7, 8, 16 }, { 9, 10, 17 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "seqb", TILEPRO_OPC_SEQB, 0x3, 3, TREG_ZERO, 1,
+    { { 7, 8, 16 }, { 9, 10, 17 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "seqb.sn", TILEPRO_OPC_SEQB_SN, 0x3, 3, TREG_SN, 1,
+    { { 7, 8, 16 }, { 9, 10, 17 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "seqh", TILEPRO_OPC_SEQH, 0x3, 3, TREG_ZERO, 1,
+    { { 7, 8, 16 }, { 9, 10, 17 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "seqh.sn", TILEPRO_OPC_SEQH_SN, 0x3, 3, TREG_SN, 1,
+    { { 7, 8, 16 }, { 9, 10, 17 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "seqi", TILEPRO_OPC_SEQI, 0xf, 3, TREG_ZERO, 1,
+    { { 7, 8, 0 }, { 9, 10, 1 }, { 11, 12, 2 }, { 13, 14, 3 }, { 0, } },
+  },
+  { "seqi.sn", TILEPRO_OPC_SEQI_SN, 0x3, 3, TREG_SN, 1,
+    { { 7, 8, 0 }, { 9, 10, 1 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "seqib", TILEPRO_OPC_SEQIB, 0x3, 3, TREG_ZERO, 1,
+    { { 7, 8, 0 }, { 9, 10, 1 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "seqib.sn", TILEPRO_OPC_SEQIB_SN, 0x3, 3, TREG_SN, 1,
+    { { 7, 8, 0 }, { 9, 10, 1 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "seqih", TILEPRO_OPC_SEQIH, 0x3, 3, TREG_ZERO, 1,
+    { { 7, 8, 0 }, { 9, 10, 1 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "seqih.sn", TILEPRO_OPC_SEQIH_SN, 0x3, 3, TREG_SN, 1,
+    { { 7, 8, 0 }, { 9, 10, 1 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "sh", TILEPRO_OPC_SH, 0x12, 2, TREG_ZERO, 1,
+    { { 0, }, { 10, 17 }, { 0, }, { 0, }, { 15, 36 } },
+  },
+  { "shadd", TILEPRO_OPC_SHADD, 0x2, 3, TREG_ZERO, 1,
+    { { 0, }, { 24, 17, 37 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "shl", TILEPRO_OPC_SHL, 0xf, 3, TREG_ZERO, 1,
+    { { 7, 8, 16 }, { 9, 10, 17 }, { 11, 12, 18 }, { 13, 14, 19 }, { 0, } },
+  },
+  { "shl.sn", TILEPRO_OPC_SHL_SN, 0x3, 3, TREG_SN, 1,
+    { { 7, 8, 16 }, { 9, 10, 17 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "shlb", TILEPRO_OPC_SHLB, 0x3, 3, TREG_ZERO, 1,
+    { { 7, 8, 16 }, { 9, 10, 17 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "shlb.sn", TILEPRO_OPC_SHLB_SN, 0x3, 3, TREG_SN, 1,
+    { { 7, 8, 16 }, { 9, 10, 17 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "shlh", TILEPRO_OPC_SHLH, 0x3, 3, TREG_ZERO, 1,
+    { { 7, 8, 16 }, { 9, 10, 17 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "shlh.sn", TILEPRO_OPC_SHLH_SN, 0x3, 3, TREG_SN, 1,
+    { { 7, 8, 16 }, { 9, 10, 17 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "shli", TILEPRO_OPC_SHLI, 0xf, 3, TREG_ZERO, 1,
+    { { 7, 8, 32 }, { 9, 10, 33 }, { 11, 12, 34 }, { 13, 14, 35 }, { 0, } },
+  },
+  { "shli.sn", TILEPRO_OPC_SHLI_SN, 0x3, 3, TREG_SN, 1,
+    { { 7, 8, 32 }, { 9, 10, 33 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "shlib", TILEPRO_OPC_SHLIB, 0x3, 3, TREG_ZERO, 1,
+    { { 7, 8, 32 }, { 9, 10, 33 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "shlib.sn", TILEPRO_OPC_SHLIB_SN, 0x3, 3, TREG_SN, 1,
+    { { 7, 8, 32 }, { 9, 10, 33 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "shlih", TILEPRO_OPC_SHLIH, 0x3, 3, TREG_ZERO, 1,
+    { { 7, 8, 32 }, { 9, 10, 33 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "shlih.sn", TILEPRO_OPC_SHLIH_SN, 0x3, 3, TREG_SN, 1,
+    { { 7, 8, 32 }, { 9, 10, 33 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "shr", TILEPRO_OPC_SHR, 0xf, 3, TREG_ZERO, 1,
+    { { 7, 8, 16 }, { 9, 10, 17 }, { 11, 12, 18 }, { 13, 14, 19 }, { 0, } },
+  },
+  { "shr.sn", TILEPRO_OPC_SHR_SN, 0x3, 3, TREG_SN, 1,
+    { { 7, 8, 16 }, { 9, 10, 17 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "shrb", TILEPRO_OPC_SHRB, 0x3, 3, TREG_ZERO, 1,
+    { { 7, 8, 16 }, { 9, 10, 17 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "shrb.sn", TILEPRO_OPC_SHRB_SN, 0x3, 3, TREG_SN, 1,
+    { { 7, 8, 16 }, { 9, 10, 17 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "shrh", TILEPRO_OPC_SHRH, 0x3, 3, TREG_ZERO, 1,
+    { { 7, 8, 16 }, { 9, 10, 17 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "shrh.sn", TILEPRO_OPC_SHRH_SN, 0x3, 3, TREG_SN, 1,
+    { { 7, 8, 16 }, { 9, 10, 17 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "shri", TILEPRO_OPC_SHRI, 0xf, 3, TREG_ZERO, 1,
+    { { 7, 8, 32 }, { 9, 10, 33 }, { 11, 12, 34 }, { 13, 14, 35 }, { 0, } },
+  },
+  { "shri.sn", TILEPRO_OPC_SHRI_SN, 0x3, 3, TREG_SN, 1,
+    { { 7, 8, 32 }, { 9, 10, 33 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "shrib", TILEPRO_OPC_SHRIB, 0x3, 3, TREG_ZERO, 1,
+    { { 7, 8, 32 }, { 9, 10, 33 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "shrib.sn", TILEPRO_OPC_SHRIB_SN, 0x3, 3, TREG_SN, 1,
+    { { 7, 8, 32 }, { 9, 10, 33 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "shrih", TILEPRO_OPC_SHRIH, 0x3, 3, TREG_ZERO, 1,
+    { { 7, 8, 32 }, { 9, 10, 33 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "shrih.sn", TILEPRO_OPC_SHRIH_SN, 0x3, 3, TREG_SN, 1,
+    { { 7, 8, 32 }, { 9, 10, 33 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "slt", TILEPRO_OPC_SLT, 0xf, 3, TREG_ZERO, 1,
+    { { 7, 8, 16 }, { 9, 10, 17 }, { 11, 12, 18 }, { 13, 14, 19 }, { 0, } },
+  },
+  { "slt.sn", TILEPRO_OPC_SLT_SN, 0x3, 3, TREG_SN, 1,
+    { { 7, 8, 16 }, { 9, 10, 17 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "slt_u", TILEPRO_OPC_SLT_U, 0xf, 3, TREG_ZERO, 1,
+    { { 7, 8, 16 }, { 9, 10, 17 }, { 11, 12, 18 }, { 13, 14, 19 }, { 0, } },
+  },
+  { "slt_u.sn", TILEPRO_OPC_SLT_U_SN, 0x3, 3, TREG_SN, 1,
+    { { 7, 8, 16 }, { 9, 10, 17 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "sltb", TILEPRO_OPC_SLTB, 0x3, 3, TREG_ZERO, 1,
+    { { 7, 8, 16 }, { 9, 10, 17 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "sltb.sn", TILEPRO_OPC_SLTB_SN, 0x3, 3, TREG_SN, 1,
+    { { 7, 8, 16 }, { 9, 10, 17 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "sltb_u", TILEPRO_OPC_SLTB_U, 0x3, 3, TREG_ZERO, 1,
+    { { 7, 8, 16 }, { 9, 10, 17 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "sltb_u.sn", TILEPRO_OPC_SLTB_U_SN, 0x3, 3, TREG_SN, 1,
+    { { 7, 8, 16 }, { 9, 10, 17 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "slte", TILEPRO_OPC_SLTE, 0xf, 3, TREG_ZERO, 1,
+    { { 7, 8, 16 }, { 9, 10, 17 }, { 11, 12, 18 }, { 13, 14, 19 }, { 0, } },
+  },
+  { "slte.sn", TILEPRO_OPC_SLTE_SN, 0x3, 3, TREG_SN, 1,
+    { { 7, 8, 16 }, { 9, 10, 17 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "slte_u", TILEPRO_OPC_SLTE_U, 0xf, 3, TREG_ZERO, 1,
+    { { 7, 8, 16 }, { 9, 10, 17 }, { 11, 12, 18 }, { 13, 14, 19 }, { 0, } },
+  },
+  { "slte_u.sn", TILEPRO_OPC_SLTE_U_SN, 0x3, 3, TREG_SN, 1,
+    { { 7, 8, 16 }, { 9, 10, 17 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "slteb", TILEPRO_OPC_SLTEB, 0x3, 3, TREG_ZERO, 1,
+    { { 7, 8, 16 }, { 9, 10, 17 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "slteb.sn", TILEPRO_OPC_SLTEB_SN, 0x3, 3, TREG_SN, 1,
+    { { 7, 8, 16 }, { 9, 10, 17 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "slteb_u", TILEPRO_OPC_SLTEB_U, 0x3, 3, TREG_ZERO, 1,
+    { { 7, 8, 16 }, { 9, 10, 17 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "slteb_u.sn", TILEPRO_OPC_SLTEB_U_SN, 0x3, 3, TREG_SN, 1,
+    { { 7, 8, 16 }, { 9, 10, 17 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "slteh", TILEPRO_OPC_SLTEH, 0x3, 3, TREG_ZERO, 1,
+    { { 7, 8, 16 }, { 9, 10, 17 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "slteh.sn", TILEPRO_OPC_SLTEH_SN, 0x3, 3, TREG_SN, 1,
+    { { 7, 8, 16 }, { 9, 10, 17 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "slteh_u", TILEPRO_OPC_SLTEH_U, 0x3, 3, TREG_ZERO, 1,
+    { { 7, 8, 16 }, { 9, 10, 17 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "slteh_u.sn", TILEPRO_OPC_SLTEH_U_SN, 0x3, 3, TREG_SN, 1,
+    { { 7, 8, 16 }, { 9, 10, 17 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "slth", TILEPRO_OPC_SLTH, 0x3, 3, TREG_ZERO, 1,
+    { { 7, 8, 16 }, { 9, 10, 17 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "slth.sn", TILEPRO_OPC_SLTH_SN, 0x3, 3, TREG_SN, 1,
+    { { 7, 8, 16 }, { 9, 10, 17 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "slth_u", TILEPRO_OPC_SLTH_U, 0x3, 3, TREG_ZERO, 1,
+    { { 7, 8, 16 }, { 9, 10, 17 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "slth_u.sn", TILEPRO_OPC_SLTH_U_SN, 0x3, 3, TREG_SN, 1,
+    { { 7, 8, 16 }, { 9, 10, 17 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "slti", TILEPRO_OPC_SLTI, 0xf, 3, TREG_ZERO, 1,
+    { { 7, 8, 0 }, { 9, 10, 1 }, { 11, 12, 2 }, { 13, 14, 3 }, { 0, } },
+  },
+  { "slti.sn", TILEPRO_OPC_SLTI_SN, 0x3, 3, TREG_SN, 1,
+    { { 7, 8, 0 }, { 9, 10, 1 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "slti_u", TILEPRO_OPC_SLTI_U, 0xf, 3, TREG_ZERO, 1,
+    { { 7, 8, 0 }, { 9, 10, 1 }, { 11, 12, 2 }, { 13, 14, 3 }, { 0, } },
+  },
+  { "slti_u.sn", TILEPRO_OPC_SLTI_U_SN, 0x3, 3, TREG_SN, 1,
+    { { 7, 8, 0 }, { 9, 10, 1 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "sltib", TILEPRO_OPC_SLTIB, 0x3, 3, TREG_ZERO, 1,
+    { { 7, 8, 0 }, { 9, 10, 1 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "sltib.sn", TILEPRO_OPC_SLTIB_SN, 0x3, 3, TREG_SN, 1,
+    { { 7, 8, 0 }, { 9, 10, 1 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "sltib_u", TILEPRO_OPC_SLTIB_U, 0x3, 3, TREG_ZERO, 1,
+    { { 7, 8, 0 }, { 9, 10, 1 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "sltib_u.sn", TILEPRO_OPC_SLTIB_U_SN, 0x3, 3, TREG_SN, 1,
+    { { 7, 8, 0 }, { 9, 10, 1 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "sltih", TILEPRO_OPC_SLTIH, 0x3, 3, TREG_ZERO, 1,
+    { { 7, 8, 0 }, { 9, 10, 1 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "sltih.sn", TILEPRO_OPC_SLTIH_SN, 0x3, 3, TREG_SN, 1,
+    { { 7, 8, 0 }, { 9, 10, 1 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "sltih_u", TILEPRO_OPC_SLTIH_U, 0x3, 3, TREG_ZERO, 1,
+    { { 7, 8, 0 }, { 9, 10, 1 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "sltih_u.sn", TILEPRO_OPC_SLTIH_U_SN, 0x3, 3, TREG_SN, 1,
+    { { 7, 8, 0 }, { 9, 10, 1 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "sne", TILEPRO_OPC_SNE, 0xf, 3, TREG_ZERO, 1,
+    { { 7, 8, 16 }, { 9, 10, 17 }, { 11, 12, 18 }, { 13, 14, 19 }, { 0, } },
+  },
+  { "sne.sn", TILEPRO_OPC_SNE_SN, 0x3, 3, TREG_SN, 1,
+    { { 7, 8, 16 }, { 9, 10, 17 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "sneb", TILEPRO_OPC_SNEB, 0x3, 3, TREG_ZERO, 1,
+    { { 7, 8, 16 }, { 9, 10, 17 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "sneb.sn", TILEPRO_OPC_SNEB_SN, 0x3, 3, TREG_SN, 1,
+    { { 7, 8, 16 }, { 9, 10, 17 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "sneh", TILEPRO_OPC_SNEH, 0x3, 3, TREG_ZERO, 1,
+    { { 7, 8, 16 }, { 9, 10, 17 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "sneh.sn", TILEPRO_OPC_SNEH_SN, 0x3, 3, TREG_SN, 1,
+    { { 7, 8, 16 }, { 9, 10, 17 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "sra", TILEPRO_OPC_SRA, 0xf, 3, TREG_ZERO, 1,
+    { { 7, 8, 16 }, { 9, 10, 17 }, { 11, 12, 18 }, { 13, 14, 19 }, { 0, } },
+  },
+  { "sra.sn", TILEPRO_OPC_SRA_SN, 0x3, 3, TREG_SN, 1,
+    { { 7, 8, 16 }, { 9, 10, 17 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "srab", TILEPRO_OPC_SRAB, 0x3, 3, TREG_ZERO, 1,
+    { { 7, 8, 16 }, { 9, 10, 17 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "srab.sn", TILEPRO_OPC_SRAB_SN, 0x3, 3, TREG_SN, 1,
+    { { 7, 8, 16 }, { 9, 10, 17 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "srah", TILEPRO_OPC_SRAH, 0x3, 3, TREG_ZERO, 1,
+    { { 7, 8, 16 }, { 9, 10, 17 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "srah.sn", TILEPRO_OPC_SRAH_SN, 0x3, 3, TREG_SN, 1,
+    { { 7, 8, 16 }, { 9, 10, 17 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "srai", TILEPRO_OPC_SRAI, 0xf, 3, TREG_ZERO, 1,
+    { { 7, 8, 32 }, { 9, 10, 33 }, { 11, 12, 34 }, { 13, 14, 35 }, { 0, } },
+  },
+  { "srai.sn", TILEPRO_OPC_SRAI_SN, 0x3, 3, TREG_SN, 1,
+    { { 7, 8, 32 }, { 9, 10, 33 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "sraib", TILEPRO_OPC_SRAIB, 0x3, 3, TREG_ZERO, 1,
+    { { 7, 8, 32 }, { 9, 10, 33 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "sraib.sn", TILEPRO_OPC_SRAIB_SN, 0x3, 3, TREG_SN, 1,
+    { { 7, 8, 32 }, { 9, 10, 33 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "sraih", TILEPRO_OPC_SRAIH, 0x3, 3, TREG_ZERO, 1,
+    { { 7, 8, 32 }, { 9, 10, 33 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "sraih.sn", TILEPRO_OPC_SRAIH_SN, 0x3, 3, TREG_SN, 1,
+    { { 7, 8, 32 }, { 9, 10, 33 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "sub", TILEPRO_OPC_SUB, 0xf, 3, TREG_ZERO, 1,
+    { { 7, 8, 16 }, { 9, 10, 17 }, { 11, 12, 18 }, { 13, 14, 19 }, { 0, } },
+  },
+  { "sub.sn", TILEPRO_OPC_SUB_SN, 0x3, 3, TREG_SN, 1,
+    { { 7, 8, 16 }, { 9, 10, 17 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "subb", TILEPRO_OPC_SUBB, 0x3, 3, TREG_ZERO, 1,
+    { { 7, 8, 16 }, { 9, 10, 17 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "subb.sn", TILEPRO_OPC_SUBB_SN, 0x3, 3, TREG_SN, 1,
+    { { 7, 8, 16 }, { 9, 10, 17 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "subbs_u", TILEPRO_OPC_SUBBS_U, 0x3, 3, TREG_ZERO, 1,
+    { { 7, 8, 16 }, { 9, 10, 17 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "subbs_u.sn", TILEPRO_OPC_SUBBS_U_SN, 0x3, 3, TREG_SN, 1,
+    { { 7, 8, 16 }, { 9, 10, 17 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "subh", TILEPRO_OPC_SUBH, 0x3, 3, TREG_ZERO, 1,
+    { { 7, 8, 16 }, { 9, 10, 17 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "subh.sn", TILEPRO_OPC_SUBH_SN, 0x3, 3, TREG_SN, 1,
+    { { 7, 8, 16 }, { 9, 10, 17 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "subhs", TILEPRO_OPC_SUBHS, 0x3, 3, TREG_ZERO, 1,
+    { { 7, 8, 16 }, { 9, 10, 17 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "subhs.sn", TILEPRO_OPC_SUBHS_SN, 0x3, 3, TREG_SN, 1,
+    { { 7, 8, 16 }, { 9, 10, 17 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "subs", TILEPRO_OPC_SUBS, 0x3, 3, TREG_ZERO, 1,
+    { { 7, 8, 16 }, { 9, 10, 17 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "subs.sn", TILEPRO_OPC_SUBS_SN, 0x3, 3, TREG_SN, 1,
+    { { 7, 8, 16 }, { 9, 10, 17 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "sw", TILEPRO_OPC_SW, 0x12, 2, TREG_ZERO, 1,
+    { { 0, }, { 10, 17 }, { 0, }, { 0, }, { 15, 36 } },
+  },
+  { "swadd", TILEPRO_OPC_SWADD, 0x2, 3, TREG_ZERO, 1,
+    { { 0, }, { 24, 17, 37 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "swint0", TILEPRO_OPC_SWINT0, 0x2, 0, TREG_ZERO, 0,
+    { { 0, }, {  }, { 0, }, { 0, }, { 0, } },
+  },
+  { "swint1", TILEPRO_OPC_SWINT1, 0x2, 0, TREG_ZERO, 0,
+    { { 0, }, {  }, { 0, }, { 0, }, { 0, } },
+  },
+  { "swint2", TILEPRO_OPC_SWINT2, 0x2, 0, TREG_ZERO, 0,
+    { { 0, }, {  }, { 0, }, { 0, }, { 0, } },
+  },
+  { "swint3", TILEPRO_OPC_SWINT3, 0x2, 0, TREG_ZERO, 0,
+    { { 0, }, {  }, { 0, }, { 0, }, { 0, } },
+  },
+  { "tblidxb0", TILEPRO_OPC_TBLIDXB0, 0x5, 2, TREG_ZERO, 1,
+    { { 21, 8 }, { 0, }, { 31, 12 }, { 0, }, { 0, } },
+  },
+  { "tblidxb0.sn", TILEPRO_OPC_TBLIDXB0_SN, 0x1, 2, TREG_SN, 1,
+    { { 21, 8 }, { 0, }, { 0, }, { 0, }, { 0, } },
+  },
+  { "tblidxb1", TILEPRO_OPC_TBLIDXB1, 0x5, 2, TREG_ZERO, 1,
+    { { 21, 8 }, { 0, }, { 31, 12 }, { 0, }, { 0, } },
+  },
+  { "tblidxb1.sn", TILEPRO_OPC_TBLIDXB1_SN, 0x1, 2, TREG_SN, 1,
+    { { 21, 8 }, { 0, }, { 0, }, { 0, }, { 0, } },
+  },
+  { "tblidxb2", TILEPRO_OPC_TBLIDXB2, 0x5, 2, TREG_ZERO, 1,
+    { { 21, 8 }, { 0, }, { 31, 12 }, { 0, }, { 0, } },
+  },
+  { "tblidxb2.sn", TILEPRO_OPC_TBLIDXB2_SN, 0x1, 2, TREG_SN, 1,
+    { { 21, 8 }, { 0, }, { 0, }, { 0, }, { 0, } },
+  },
+  { "tblidxb3", TILEPRO_OPC_TBLIDXB3, 0x5, 2, TREG_ZERO, 1,
+    { { 21, 8 }, { 0, }, { 31, 12 }, { 0, }, { 0, } },
+  },
+  { "tblidxb3.sn", TILEPRO_OPC_TBLIDXB3_SN, 0x1, 2, TREG_SN, 1,
+    { { 21, 8 }, { 0, }, { 0, }, { 0, }, { 0, } },
+  },
+  { "tns", TILEPRO_OPC_TNS, 0x2, 2, TREG_ZERO, 1,
+    { { 0, }, { 9, 10 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "tns.sn", TILEPRO_OPC_TNS_SN, 0x2, 2, TREG_SN, 1,
+    { { 0, }, { 9, 10 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "wh64", TILEPRO_OPC_WH64, 0x2, 1, TREG_ZERO, 1,
+    { { 0, }, { 10 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "xor", TILEPRO_OPC_XOR, 0xf, 3, TREG_ZERO, 1,
+    { { 7, 8, 16 }, { 9, 10, 17 }, { 11, 12, 18 }, { 13, 14, 19 }, { 0, } },
+  },
+  { "xor.sn", TILEPRO_OPC_XOR_SN, 0x3, 3, TREG_SN, 1,
+    { { 7, 8, 16 }, { 9, 10, 17 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "xori", TILEPRO_OPC_XORI, 0x3, 3, TREG_ZERO, 1,
+    { { 7, 8, 0 }, { 9, 10, 1 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "xori.sn", TILEPRO_OPC_XORI_SN, 0x3, 3, TREG_SN, 1,
+    { { 7, 8, 0 }, { 9, 10, 1 }, { 0, }, { 0, }, { 0, } },
+  },
+  { NULL, TILEPRO_OPC_NONE, 0, 0, TREG_ZERO, 0, { { 0, } },
+  }
+};
+#define BITFIELD(start, size) ((start) | (((1 << (size)) - 1) << 6))
+#define CHILD(array_index) (TILEPRO_OPC_NONE + (array_index))
+
+static const unsigned short decode_X0_fsm[1153] =
+{
+  BITFIELD(22, 9) /* index 0 */,
+  CHILD(513), CHILD(530), CHILD(547), CHILD(564), CHILD(596), CHILD(613),
+  CHILD(630), TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE,
+  TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE,
+  TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE,
+  TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE,
+  TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE,
+  TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE,
+  TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, CHILD(663), CHILD(680), CHILD(697),
+  CHILD(714), CHILD(746), CHILD(763), CHILD(780), TILEPRO_OPC_NONE,
+  TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE,
+  TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE,
+  TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE,
+  TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE,
+  TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE,
+  TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE,
+  CHILD(813), CHILD(813), CHILD(813), CHILD(813), CHILD(813), CHILD(813),
+  CHILD(813), CHILD(813), CHILD(813), CHILD(813), CHILD(813), CHILD(813),
+  CHILD(813), CHILD(813), CHILD(813), CHILD(813), CHILD(813), CHILD(813),
+  CHILD(813), CHILD(813), CHILD(813), CHILD(813), CHILD(813), CHILD(813),
+  CHILD(813), CHILD(813), CHILD(813), CHILD(813), CHILD(813), CHILD(813),
+  CHILD(813), CHILD(813), CHILD(813), CHILD(813), CHILD(813), CHILD(813),
+  CHILD(813), CHILD(813), CHILD(813), CHILD(813), CHILD(813), CHILD(813),
+  CHILD(813), CHILD(813), CHILD(813), CHILD(813), CHILD(813), CHILD(813),
+  CHILD(813), CHILD(813), CHILD(813), CHILD(813), CHILD(813), CHILD(813),
+  CHILD(813), CHILD(813), CHILD(813), CHILD(813), CHILD(813), CHILD(813),
+  CHILD(813), CHILD(813), CHILD(813), CHILD(813), CHILD(828), CHILD(828),
+  CHILD(828), CHILD(828), CHILD(828), CHILD(828), CHILD(828), CHILD(828),
+  CHILD(828), CHILD(828), CHILD(828), CHILD(828), CHILD(828), CHILD(828),
+  CHILD(828), CHILD(828), CHILD(828), CHILD(828), CHILD(828), CHILD(828),
+  CHILD(828), CHILD(828), CHILD(828), CHILD(828), CHILD(828), CHILD(828),
+  CHILD(828), CHILD(828), CHILD(828), CHILD(828), CHILD(828), CHILD(828),
+  CHILD(828), CHILD(828), CHILD(828), CHILD(828), CHILD(828), CHILD(828),
+  CHILD(828), CHILD(828), CHILD(828), CHILD(828), CHILD(828), CHILD(828),
+  CHILD(828), CHILD(828), CHILD(828), CHILD(828), CHILD(828), CHILD(828),
+  CHILD(828), CHILD(828), CHILD(828), CHILD(828), CHILD(828), CHILD(828),
+  CHILD(828), CHILD(828), CHILD(828), CHILD(828), CHILD(828), CHILD(828),
+  CHILD(828), CHILD(828), CHILD(843), CHILD(843), CHILD(843), CHILD(843),
+  CHILD(843), CHILD(843), CHILD(843), CHILD(843), CHILD(843), CHILD(843),
+  CHILD(843), CHILD(843), CHILD(843), CHILD(843), CHILD(843), CHILD(843),
+  CHILD(843), CHILD(843), CHILD(843), CHILD(843), CHILD(843), CHILD(843),
+  CHILD(843), CHILD(843), CHILD(843), CHILD(843), CHILD(843), CHILD(843),
+  CHILD(843), CHILD(843), CHILD(843), CHILD(843), CHILD(843), CHILD(843),
+  CHILD(843), CHILD(843), CHILD(843), CHILD(843), CHILD(843), CHILD(843),
+  CHILD(843), CHILD(843), CHILD(843), CHILD(843), CHILD(843), CHILD(843),
+  CHILD(843), CHILD(843), CHILD(843), CHILD(843), CHILD(843), CHILD(843),
+  CHILD(843), CHILD(843), CHILD(843), CHILD(843), CHILD(843), CHILD(843),
+  CHILD(843), CHILD(843), CHILD(843), CHILD(843), CHILD(843), CHILD(843),
+  CHILD(873), CHILD(878), CHILD(883), CHILD(903), CHILD(908),
+  TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE,
+  TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE,
+  TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE,
+  TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE,
+  TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE,
+  TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE,
+  TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, CHILD(913),
+  CHILD(918), CHILD(923), CHILD(943), CHILD(948), TILEPRO_OPC_NONE,
+  TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE,
+  TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE,
+  TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE,
+  TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE,
+  TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE,
+  TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE,
+  TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, CHILD(953), TILEPRO_OPC_NONE,
+  TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE,
+  TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE,
+  TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE,
+  TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE,
+  TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE,
+  TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE,
+  TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE,
+  TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, CHILD(988), TILEPRO_OPC_NONE,
+  TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE,
+  TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE,
+  TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE,
+  TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE,
+  TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE,
+  TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE,
+  TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE,
+  TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_MM, TILEPRO_OPC_MM,
+  TILEPRO_OPC_MM, TILEPRO_OPC_MM, TILEPRO_OPC_MM, TILEPRO_OPC_MM,
+  TILEPRO_OPC_MM, TILEPRO_OPC_MM, TILEPRO_OPC_MM, TILEPRO_OPC_MM,
+  TILEPRO_OPC_MM, TILEPRO_OPC_MM, TILEPRO_OPC_MM, TILEPRO_OPC_MM,
+  TILEPRO_OPC_MM, TILEPRO_OPC_MM, TILEPRO_OPC_MM, TILEPRO_OPC_MM,
+  TILEPRO_OPC_MM, TILEPRO_OPC_MM, TILEPRO_OPC_MM, TILEPRO_OPC_MM,
+  TILEPRO_OPC_MM, TILEPRO_OPC_MM, TILEPRO_OPC_MM, TILEPRO_OPC_MM,
+  TILEPRO_OPC_MM, TILEPRO_OPC_MM, TILEPRO_OPC_MM, TILEPRO_OPC_MM,
+  TILEPRO_OPC_MM, TILEPRO_OPC_MM, TILEPRO_OPC_MM, TILEPRO_OPC_MM,
+  TILEPRO_OPC_MM, TILEPRO_OPC_MM, TILEPRO_OPC_MM, TILEPRO_OPC_MM,
+  TILEPRO_OPC_MM, TILEPRO_OPC_MM, TILEPRO_OPC_MM, TILEPRO_OPC_MM,
+  TILEPRO_OPC_MM, TILEPRO_OPC_MM, TILEPRO_OPC_MM, TILEPRO_OPC_MM,
+  TILEPRO_OPC_MM, TILEPRO_OPC_MM, TILEPRO_OPC_MM, TILEPRO_OPC_MM,
+  TILEPRO_OPC_MM, TILEPRO_OPC_MM, TILEPRO_OPC_MM, TILEPRO_OPC_MM,
+  TILEPRO_OPC_MM, TILEPRO_OPC_MM, TILEPRO_OPC_MM, TILEPRO_OPC_MM,
+  TILEPRO_OPC_MM, TILEPRO_OPC_MM, TILEPRO_OPC_MM, TILEPRO_OPC_MM,
+  TILEPRO_OPC_MM, TILEPRO_OPC_MM, CHILD(993), TILEPRO_OPC_NONE,
+  TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE,
+  TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE,
+  TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE,
+  TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE,
+  TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE,
+  TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE,
+  TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE,
+  TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, CHILD(1076), TILEPRO_OPC_NONE,
+  TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE,
+  TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE,
+  TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE,
+  TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE,
+  TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE,
+  TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE,
+  TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE,
+  TILEPRO_OPC_NONE, TILEPRO_OPC_NONE,
+  BITFIELD(18, 4) /* index 513 */,
+  TILEPRO_OPC_NONE, TILEPRO_OPC_ADDB, TILEPRO_OPC_ADDH, TILEPRO_OPC_ADD,
+  TILEPRO_OPC_ADIFFB_U, TILEPRO_OPC_ADIFFH, TILEPRO_OPC_AND,
+  TILEPRO_OPC_AVGB_U, TILEPRO_OPC_AVGH, TILEPRO_OPC_CRC32_32,
+  TILEPRO_OPC_CRC32_8, TILEPRO_OPC_INTHB, TILEPRO_OPC_INTHH,
+  TILEPRO_OPC_INTLB, TILEPRO_OPC_INTLH, TILEPRO_OPC_MAXB_U,
+  BITFIELD(18, 4) /* index 530 */,
+  TILEPRO_OPC_MAXH, TILEPRO_OPC_MINB_U, TILEPRO_OPC_MINH, TILEPRO_OPC_MNZB,
+  TILEPRO_OPC_MNZH, TILEPRO_OPC_MNZ, TILEPRO_OPC_MULHHA_SS,
+  TILEPRO_OPC_MULHHA_SU, TILEPRO_OPC_MULHHA_UU, TILEPRO_OPC_MULHHSA_UU,
+  TILEPRO_OPC_MULHH_SS, TILEPRO_OPC_MULHH_SU, TILEPRO_OPC_MULHH_UU,
+  TILEPRO_OPC_MULHLA_SS, TILEPRO_OPC_MULHLA_SU, TILEPRO_OPC_MULHLA_US,
+  BITFIELD(18, 4) /* index 547 */,
+  TILEPRO_OPC_MULHLA_UU, TILEPRO_OPC_MULHLSA_UU, TILEPRO_OPC_MULHL_SS,
+  TILEPRO_OPC_MULHL_SU, TILEPRO_OPC_MULHL_US, TILEPRO_OPC_MULHL_UU,
+  TILEPRO_OPC_MULLLA_SS, TILEPRO_OPC_MULLLA_SU, TILEPRO_OPC_MULLLA_UU,
+  TILEPRO_OPC_MULLLSA_UU, TILEPRO_OPC_MULLL_SS, TILEPRO_OPC_MULLL_SU,
+  TILEPRO_OPC_MULLL_UU, TILEPRO_OPC_MVNZ, TILEPRO_OPC_MVZ, TILEPRO_OPC_MZB,
+  BITFIELD(18, 4) /* index 564 */,
+  TILEPRO_OPC_MZH, TILEPRO_OPC_MZ, TILEPRO_OPC_NOR, CHILD(581),
+  TILEPRO_OPC_PACKHB, TILEPRO_OPC_PACKLB, TILEPRO_OPC_RL, TILEPRO_OPC_S1A,
+  TILEPRO_OPC_S2A, TILEPRO_OPC_S3A, TILEPRO_OPC_SADAB_U, TILEPRO_OPC_SADAH,
+  TILEPRO_OPC_SADAH_U, TILEPRO_OPC_SADB_U, TILEPRO_OPC_SADH,
+  TILEPRO_OPC_SADH_U,
+  BITFIELD(12, 2) /* index 581 */,
+  TILEPRO_OPC_OR, TILEPRO_OPC_OR, TILEPRO_OPC_OR, CHILD(586),
+  BITFIELD(14, 2) /* index 586 */,
+  TILEPRO_OPC_OR, TILEPRO_OPC_OR, TILEPRO_OPC_OR, CHILD(591),
+  BITFIELD(16, 2) /* index 591 */,
+  TILEPRO_OPC_OR, TILEPRO_OPC_OR, TILEPRO_OPC_OR, TILEPRO_OPC_MOVE,
+  BITFIELD(18, 4) /* index 596 */,
+  TILEPRO_OPC_SEQB, TILEPRO_OPC_SEQH, TILEPRO_OPC_SEQ, TILEPRO_OPC_SHLB,
+  TILEPRO_OPC_SHLH, TILEPRO_OPC_SHL, TILEPRO_OPC_SHRB, TILEPRO_OPC_SHRH,
+  TILEPRO_OPC_SHR, TILEPRO_OPC_SLTB, TILEPRO_OPC_SLTB_U, TILEPRO_OPC_SLTEB,
+  TILEPRO_OPC_SLTEB_U, TILEPRO_OPC_SLTEH, TILEPRO_OPC_SLTEH_U,
+  TILEPRO_OPC_SLTE,
+  BITFIELD(18, 4) /* index 613 */,
+  TILEPRO_OPC_SLTE_U, TILEPRO_OPC_SLTH, TILEPRO_OPC_SLTH_U, TILEPRO_OPC_SLT,
+  TILEPRO_OPC_SLT_U, TILEPRO_OPC_SNEB, TILEPRO_OPC_SNEH, TILEPRO_OPC_SNE,
+  TILEPRO_OPC_SRAB, TILEPRO_OPC_SRAH, TILEPRO_OPC_SRA, TILEPRO_OPC_SUBB,
+  TILEPRO_OPC_SUBH, TILEPRO_OPC_SUB, TILEPRO_OPC_XOR, TILEPRO_OPC_DWORD_ALIGN,
+  BITFIELD(18, 3) /* index 630 */,
+  CHILD(639), CHILD(642), CHILD(645), CHILD(648), CHILD(651), CHILD(654),
+  CHILD(657), CHILD(660),
+  BITFIELD(21, 1) /* index 639 */,
+  TILEPRO_OPC_ADDS, TILEPRO_OPC_NONE,
+  BITFIELD(21, 1) /* index 642 */,
+  TILEPRO_OPC_SUBS, TILEPRO_OPC_NONE,
+  BITFIELD(21, 1) /* index 645 */,
+  TILEPRO_OPC_ADDBS_U, TILEPRO_OPC_NONE,
+  BITFIELD(21, 1) /* index 648 */,
+  TILEPRO_OPC_ADDHS, TILEPRO_OPC_NONE,
+  BITFIELD(21, 1) /* index 651 */,
+  TILEPRO_OPC_SUBBS_U, TILEPRO_OPC_NONE,
+  BITFIELD(21, 1) /* index 654 */,
+  TILEPRO_OPC_SUBHS, TILEPRO_OPC_NONE,
+  BITFIELD(21, 1) /* index 657 */,
+  TILEPRO_OPC_PACKHS, TILEPRO_OPC_NONE,
+  BITFIELD(21, 1) /* index 660 */,
+  TILEPRO_OPC_PACKBS_U, TILEPRO_OPC_NONE,
+  BITFIELD(18, 4) /* index 663 */,
+  TILEPRO_OPC_NONE, TILEPRO_OPC_ADDB_SN, TILEPRO_OPC_ADDH_SN,
+  TILEPRO_OPC_ADD_SN, TILEPRO_OPC_ADIFFB_U_SN, TILEPRO_OPC_ADIFFH_SN,
+  TILEPRO_OPC_AND_SN, TILEPRO_OPC_AVGB_U_SN, TILEPRO_OPC_AVGH_SN,
+  TILEPRO_OPC_CRC32_32_SN, TILEPRO_OPC_CRC32_8_SN, TILEPRO_OPC_INTHB_SN,
+  TILEPRO_OPC_INTHH_SN, TILEPRO_OPC_INTLB_SN, TILEPRO_OPC_INTLH_SN,
+  TILEPRO_OPC_MAXB_U_SN,
+  BITFIELD(18, 4) /* index 680 */,
+  TILEPRO_OPC_MAXH_SN, TILEPRO_OPC_MINB_U_SN, TILEPRO_OPC_MINH_SN,
+  TILEPRO_OPC_MNZB_SN, TILEPRO_OPC_MNZH_SN, TILEPRO_OPC_MNZ_SN,
+  TILEPRO_OPC_MULHHA_SS_SN, TILEPRO_OPC_MULHHA_SU_SN,
+  TILEPRO_OPC_MULHHA_UU_SN, TILEPRO_OPC_MULHHSA_UU_SN,
+  TILEPRO_OPC_MULHH_SS_SN, TILEPRO_OPC_MULHH_SU_SN, TILEPRO_OPC_MULHH_UU_SN,
+  TILEPRO_OPC_MULHLA_SS_SN, TILEPRO_OPC_MULHLA_SU_SN,
+  TILEPRO_OPC_MULHLA_US_SN,
+  BITFIELD(18, 4) /* index 697 */,
+  TILEPRO_OPC_MULHLA_UU_SN, TILEPRO_OPC_MULHLSA_UU_SN,
+  TILEPRO_OPC_MULHL_SS_SN, TILEPRO_OPC_MULHL_SU_SN, TILEPRO_OPC_MULHL_US_SN,
+  TILEPRO_OPC_MULHL_UU_SN, TILEPRO_OPC_MULLLA_SS_SN, TILEPRO_OPC_MULLLA_SU_SN,
+  TILEPRO_OPC_MULLLA_UU_SN, TILEPRO_OPC_MULLLSA_UU_SN,
+  TILEPRO_OPC_MULLL_SS_SN, TILEPRO_OPC_MULLL_SU_SN, TILEPRO_OPC_MULLL_UU_SN,
+  TILEPRO_OPC_MVNZ_SN, TILEPRO_OPC_MVZ_SN, TILEPRO_OPC_MZB_SN,
+  BITFIELD(18, 4) /* index 714 */,
+  TILEPRO_OPC_MZH_SN, TILEPRO_OPC_MZ_SN, TILEPRO_OPC_NOR_SN, CHILD(731),
+  TILEPRO_OPC_PACKHB_SN, TILEPRO_OPC_PACKLB_SN, TILEPRO_OPC_RL_SN,
+  TILEPRO_OPC_S1A_SN, TILEPRO_OPC_S2A_SN, TILEPRO_OPC_S3A_SN,
+  TILEPRO_OPC_SADAB_U_SN, TILEPRO_OPC_SADAH_SN, TILEPRO_OPC_SADAH_U_SN,
+  TILEPRO_OPC_SADB_U_SN, TILEPRO_OPC_SADH_SN, TILEPRO_OPC_SADH_U_SN,
+  BITFIELD(12, 2) /* index 731 */,
+  TILEPRO_OPC_OR_SN, TILEPRO_OPC_OR_SN, TILEPRO_OPC_OR_SN, CHILD(736),
+  BITFIELD(14, 2) /* index 736 */,
+  TILEPRO_OPC_OR_SN, TILEPRO_OPC_OR_SN, TILEPRO_OPC_OR_SN, CHILD(741),
+  BITFIELD(16, 2) /* index 741 */,
+  TILEPRO_OPC_OR_SN, TILEPRO_OPC_OR_SN, TILEPRO_OPC_OR_SN,
+  TILEPRO_OPC_MOVE_SN,
+  BITFIELD(18, 4) /* index 746 */,
+  TILEPRO_OPC_SEQB_SN, TILEPRO_OPC_SEQH_SN, TILEPRO_OPC_SEQ_SN,
+  TILEPRO_OPC_SHLB_SN, TILEPRO_OPC_SHLH_SN, TILEPRO_OPC_SHL_SN,
+  TILEPRO_OPC_SHRB_SN, TILEPRO_OPC_SHRH_SN, TILEPRO_OPC_SHR_SN,
+  TILEPRO_OPC_SLTB_SN, TILEPRO_OPC_SLTB_U_SN, TILEPRO_OPC_SLTEB_SN,
+  TILEPRO_OPC_SLTEB_U_SN, TILEPRO_OPC_SLTEH_SN, TILEPRO_OPC_SLTEH_U_SN,
+  TILEPRO_OPC_SLTE_SN,
+  BITFIELD(18, 4) /* index 763 */,
+  TILEPRO_OPC_SLTE_U_SN, TILEPRO_OPC_SLTH_SN, TILEPRO_OPC_SLTH_U_SN,
+  TILEPRO_OPC_SLT_SN, TILEPRO_OPC_SLT_U_SN, TILEPRO_OPC_SNEB_SN,
+  TILEPRO_OPC_SNEH_SN, TILEPRO_OPC_SNE_SN, TILEPRO_OPC_SRAB_SN,
+  TILEPRO_OPC_SRAH_SN, TILEPRO_OPC_SRA_SN, TILEPRO_OPC_SUBB_SN,
+  TILEPRO_OPC_SUBH_SN, TILEPRO_OPC_SUB_SN, TILEPRO_OPC_XOR_SN,
+  TILEPRO_OPC_DWORD_ALIGN_SN,
+  BITFIELD(18, 3) /* index 780 */,
+  CHILD(789), CHILD(792), CHILD(795), CHILD(798), CHILD(801), CHILD(804),
+  CHILD(807), CHILD(810),
+  BITFIELD(21, 1) /* index 789 */,
+  TILEPRO_OPC_ADDS_SN, TILEPRO_OPC_NONE,
+  BITFIELD(21, 1) /* index 792 */,
+  TILEPRO_OPC_SUBS_SN, TILEPRO_OPC_NONE,
+  BITFIELD(21, 1) /* index 795 */,
+  TILEPRO_OPC_ADDBS_U_SN, TILEPRO_OPC_NONE,
+  BITFIELD(21, 1) /* index 798 */,
+  TILEPRO_OPC_ADDHS_SN, TILEPRO_OPC_NONE,
+  BITFIELD(21, 1) /* index 801 */,
+  TILEPRO_OPC_SUBBS_U_SN, TILEPRO_OPC_NONE,
+  BITFIELD(21, 1) /* index 804 */,
+  TILEPRO_OPC_SUBHS_SN, TILEPRO_OPC_NONE,
+  BITFIELD(21, 1) /* index 807 */,
+  TILEPRO_OPC_PACKHS_SN, TILEPRO_OPC_NONE,
+  BITFIELD(21, 1) /* index 810 */,
+  TILEPRO_OPC_PACKBS_U_SN, TILEPRO_OPC_NONE,
+  BITFIELD(6, 2) /* index 813 */,
+  TILEPRO_OPC_ADDLI_SN, TILEPRO_OPC_ADDLI_SN, TILEPRO_OPC_ADDLI_SN,
+  CHILD(818),
+  BITFIELD(8, 2) /* index 818 */,
+  TILEPRO_OPC_ADDLI_SN, TILEPRO_OPC_ADDLI_SN, TILEPRO_OPC_ADDLI_SN,
+  CHILD(823),
+  BITFIELD(10, 2) /* index 823 */,
+  TILEPRO_OPC_ADDLI_SN, TILEPRO_OPC_ADDLI_SN, TILEPRO_OPC_ADDLI_SN,
+  TILEPRO_OPC_MOVELI_SN,
+  BITFIELD(6, 2) /* index 828 */,
+  TILEPRO_OPC_ADDLI, TILEPRO_OPC_ADDLI, TILEPRO_OPC_ADDLI, CHILD(833),
+  BITFIELD(8, 2) /* index 833 */,
+  TILEPRO_OPC_ADDLI, TILEPRO_OPC_ADDLI, TILEPRO_OPC_ADDLI, CHILD(838),
+  BITFIELD(10, 2) /* index 838 */,
+  TILEPRO_OPC_ADDLI, TILEPRO_OPC_ADDLI, TILEPRO_OPC_ADDLI, TILEPRO_OPC_MOVELI,
+  BITFIELD(0, 2) /* index 843 */,
+  TILEPRO_OPC_AULI, TILEPRO_OPC_AULI, TILEPRO_OPC_AULI, CHILD(848),
+  BITFIELD(2, 2) /* index 848 */,
+  TILEPRO_OPC_AULI, TILEPRO_OPC_AULI, TILEPRO_OPC_AULI, CHILD(853),
+  BITFIELD(4, 2) /* index 853 */,
+  TILEPRO_OPC_AULI, TILEPRO_OPC_AULI, TILEPRO_OPC_AULI, CHILD(858),
+  BITFIELD(6, 2) /* index 858 */,
+  TILEPRO_OPC_AULI, TILEPRO_OPC_AULI, TILEPRO_OPC_AULI, CHILD(863),
+  BITFIELD(8, 2) /* index 863 */,
+  TILEPRO_OPC_AULI, TILEPRO_OPC_AULI, TILEPRO_OPC_AULI, CHILD(868),
+  BITFIELD(10, 2) /* index 868 */,
+  TILEPRO_OPC_AULI, TILEPRO_OPC_AULI, TILEPRO_OPC_AULI, TILEPRO_OPC_INFOL,
+  BITFIELD(20, 2) /* index 873 */,
+  TILEPRO_OPC_NONE, TILEPRO_OPC_ADDIB, TILEPRO_OPC_ADDIH, TILEPRO_OPC_ADDI,
+  BITFIELD(20, 2) /* index 878 */,
+  TILEPRO_OPC_MAXIB_U, TILEPRO_OPC_MAXIH, TILEPRO_OPC_MINIB_U,
+  TILEPRO_OPC_MINIH,
+  BITFIELD(20, 2) /* index 883 */,
+  CHILD(888), TILEPRO_OPC_SEQIB, TILEPRO_OPC_SEQIH, TILEPRO_OPC_SEQI,
+  BITFIELD(6, 2) /* index 888 */,
+  TILEPRO_OPC_ORI, TILEPRO_OPC_ORI, TILEPRO_OPC_ORI, CHILD(893),
+  BITFIELD(8, 2) /* index 893 */,
+  TILEPRO_OPC_ORI, TILEPRO_OPC_ORI, TILEPRO_OPC_ORI, CHILD(898),
+  BITFIELD(10, 2) /* index 898 */,
+  TILEPRO_OPC_ORI, TILEPRO_OPC_ORI, TILEPRO_OPC_ORI, TILEPRO_OPC_MOVEI,
+  BITFIELD(20, 2) /* index 903 */,
+  TILEPRO_OPC_SLTIB, TILEPRO_OPC_SLTIB_U, TILEPRO_OPC_SLTIH,
+  TILEPRO_OPC_SLTIH_U,
+  BITFIELD(20, 2) /* index 908 */,
+  TILEPRO_OPC_SLTI, TILEPRO_OPC_SLTI_U, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE,
+  BITFIELD(20, 2) /* index 913 */,
+  TILEPRO_OPC_NONE, TILEPRO_OPC_ADDIB_SN, TILEPRO_OPC_ADDIH_SN,
+  TILEPRO_OPC_ADDI_SN,
+  BITFIELD(20, 2) /* index 918 */,
+  TILEPRO_OPC_MAXIB_U_SN, TILEPRO_OPC_MAXIH_SN, TILEPRO_OPC_MINIB_U_SN,
+  TILEPRO_OPC_MINIH_SN,
+  BITFIELD(20, 2) /* index 923 */,
+  CHILD(928), TILEPRO_OPC_SEQIB_SN, TILEPRO_OPC_SEQIH_SN, TILEPRO_OPC_SEQI_SN,
+  BITFIELD(6, 2) /* index 928 */,
+  TILEPRO_OPC_ORI_SN, TILEPRO_OPC_ORI_SN, TILEPRO_OPC_ORI_SN, CHILD(933),
+  BITFIELD(8, 2) /* index 933 */,
+  TILEPRO_OPC_ORI_SN, TILEPRO_OPC_ORI_SN, TILEPRO_OPC_ORI_SN, CHILD(938),
+  BITFIELD(10, 2) /* index 938 */,
+  TILEPRO_OPC_ORI_SN, TILEPRO_OPC_ORI_SN, TILEPRO_OPC_ORI_SN,
+  TILEPRO_OPC_MOVEI_SN,
+  BITFIELD(20, 2) /* index 943 */,
+  TILEPRO_OPC_SLTIB_SN, TILEPRO_OPC_SLTIB_U_SN, TILEPRO_OPC_SLTIH_SN,
+  TILEPRO_OPC_SLTIH_U_SN,
+  BITFIELD(20, 2) /* index 948 */,
+  TILEPRO_OPC_SLTI_SN, TILEPRO_OPC_SLTI_U_SN, TILEPRO_OPC_NONE,
+  TILEPRO_OPC_NONE,
+  BITFIELD(20, 2) /* index 953 */,
+  TILEPRO_OPC_NONE, CHILD(958), TILEPRO_OPC_XORI, TILEPRO_OPC_NONE,
+  BITFIELD(0, 2) /* index 958 */,
+  TILEPRO_OPC_ANDI, TILEPRO_OPC_ANDI, TILEPRO_OPC_ANDI, CHILD(963),
+  BITFIELD(2, 2) /* index 963 */,
+  TILEPRO_OPC_ANDI, TILEPRO_OPC_ANDI, TILEPRO_OPC_ANDI, CHILD(968),
+  BITFIELD(4, 2) /* index 968 */,
+  TILEPRO_OPC_ANDI, TILEPRO_OPC_ANDI, TILEPRO_OPC_ANDI, CHILD(973),
+  BITFIELD(6, 2) /* index 973 */,
+  TILEPRO_OPC_ANDI, TILEPRO_OPC_ANDI, TILEPRO_OPC_ANDI, CHILD(978),
+  BITFIELD(8, 2) /* index 978 */,
+  TILEPRO_OPC_ANDI, TILEPRO_OPC_ANDI, TILEPRO_OPC_ANDI, CHILD(983),
+  BITFIELD(10, 2) /* index 983 */,
+  TILEPRO_OPC_ANDI, TILEPRO_OPC_ANDI, TILEPRO_OPC_ANDI, TILEPRO_OPC_INFO,
+  BITFIELD(20, 2) /* index 988 */,
+  TILEPRO_OPC_NONE, TILEPRO_OPC_ANDI_SN, TILEPRO_OPC_XORI_SN,
+  TILEPRO_OPC_NONE,
+  BITFIELD(17, 5) /* index 993 */,
+  TILEPRO_OPC_NONE, TILEPRO_OPC_RLI, TILEPRO_OPC_SHLIB, TILEPRO_OPC_SHLIH,
+  TILEPRO_OPC_SHLI, TILEPRO_OPC_SHRIB, TILEPRO_OPC_SHRIH, TILEPRO_OPC_SHRI,
+  TILEPRO_OPC_SRAIB, TILEPRO_OPC_SRAIH, TILEPRO_OPC_SRAI, CHILD(1026),
+  TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE,
+  TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE,
+  TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE,
+  TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE,
+  TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE,
+  BITFIELD(12, 4) /* index 1026 */,
+  TILEPRO_OPC_NONE, CHILD(1043), CHILD(1046), CHILD(1049), CHILD(1052),
+  CHILD(1055), CHILD(1058), CHILD(1061), CHILD(1064), CHILD(1067),
+  CHILD(1070), CHILD(1073), TILEPRO_OPC_NONE, TILEPRO_OPC_NONE,
+  TILEPRO_OPC_NONE, TILEPRO_OPC_NONE,
+  BITFIELD(16, 1) /* index 1043 */,
+  TILEPRO_OPC_BITX, TILEPRO_OPC_NONE,
+  BITFIELD(16, 1) /* index 1046 */,
+  TILEPRO_OPC_BYTEX, TILEPRO_OPC_NONE,
+  BITFIELD(16, 1) /* index 1049 */,
+  TILEPRO_OPC_CLZ, TILEPRO_OPC_NONE,
+  BITFIELD(16, 1) /* index 1052 */,
+  TILEPRO_OPC_CTZ, TILEPRO_OPC_NONE,
+  BITFIELD(16, 1) /* index 1055 */,
+  TILEPRO_OPC_FNOP, TILEPRO_OPC_NONE,
+  BITFIELD(16, 1) /* index 1058 */,
+  TILEPRO_OPC_NOP, TILEPRO_OPC_NONE,
+  BITFIELD(16, 1) /* index 1061 */,
+  TILEPRO_OPC_PCNT, TILEPRO_OPC_NONE,
+  BITFIELD(16, 1) /* index 1064 */,
+  TILEPRO_OPC_TBLIDXB0, TILEPRO_OPC_NONE,
+  BITFIELD(16, 1) /* index 1067 */,
+  TILEPRO_OPC_TBLIDXB1, TILEPRO_OPC_NONE,
+  BITFIELD(16, 1) /* index 1070 */,
+  TILEPRO_OPC_TBLIDXB2, TILEPRO_OPC_NONE,
+  BITFIELD(16, 1) /* index 1073 */,
+  TILEPRO_OPC_TBLIDXB3, TILEPRO_OPC_NONE,
+  BITFIELD(17, 5) /* index 1076 */,
+  TILEPRO_OPC_NONE, TILEPRO_OPC_RLI_SN, TILEPRO_OPC_SHLIB_SN,
+  TILEPRO_OPC_SHLIH_SN, TILEPRO_OPC_SHLI_SN, TILEPRO_OPC_SHRIB_SN,
+  TILEPRO_OPC_SHRIH_SN, TILEPRO_OPC_SHRI_SN, TILEPRO_OPC_SRAIB_SN,
+  TILEPRO_OPC_SRAIH_SN, TILEPRO_OPC_SRAI_SN, CHILD(1109), TILEPRO_OPC_NONE,
+  TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE,
+  TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE,
+  TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE,
+  TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE,
+  TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE,
+  BITFIELD(12, 4) /* index 1109 */,
+  TILEPRO_OPC_NONE, CHILD(1126), CHILD(1129), CHILD(1132), CHILD(1135),
+  CHILD(1055), CHILD(1058), CHILD(1138), CHILD(1141), CHILD(1144),
+  CHILD(1147), CHILD(1150), TILEPRO_OPC_NONE, TILEPRO_OPC_NONE,
+  TILEPRO_OPC_NONE, TILEPRO_OPC_NONE,
+  BITFIELD(16, 1) /* index 1126 */,
+  TILEPRO_OPC_BITX_SN, TILEPRO_OPC_NONE,
+  BITFIELD(16, 1) /* index 1129 */,
+  TILEPRO_OPC_BYTEX_SN, TILEPRO_OPC_NONE,
+  BITFIELD(16, 1) /* index 1132 */,
+  TILEPRO_OPC_CLZ_SN, TILEPRO_OPC_NONE,
+  BITFIELD(16, 1) /* index 1135 */,
+  TILEPRO_OPC_CTZ_SN, TILEPRO_OPC_NONE,
+  BITFIELD(16, 1) /* index 1138 */,
+  TILEPRO_OPC_PCNT_SN, TILEPRO_OPC_NONE,
+  BITFIELD(16, 1) /* index 1141 */,
+  TILEPRO_OPC_TBLIDXB0_SN, TILEPRO_OPC_NONE,
+  BITFIELD(16, 1) /* index 1144 */,
+  TILEPRO_OPC_TBLIDXB1_SN, TILEPRO_OPC_NONE,
+  BITFIELD(16, 1) /* index 1147 */,
+  TILEPRO_OPC_TBLIDXB2_SN, TILEPRO_OPC_NONE,
+  BITFIELD(16, 1) /* index 1150 */,
+  TILEPRO_OPC_TBLIDXB3_SN, TILEPRO_OPC_NONE,
+};
+
+static const unsigned short decode_X1_fsm[1540] =
+{
+  BITFIELD(54, 9) /* index 0 */,
+  TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE,
+  TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE,
+  TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE,
+  TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE,
+  TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE,
+  TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE,
+  TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE,
+  TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE,
+  CHILD(513), CHILD(561), CHILD(594), TILEPRO_OPC_NONE, TILEPRO_OPC_NONE,
+  TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE,
+  TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE,
+  TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, CHILD(641),
+  CHILD(689), CHILD(722), TILEPRO_OPC_NONE, TILEPRO_OPC_NONE,
+  TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE,
+  TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE,
+  TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, CHILD(766),
+  CHILD(766), CHILD(766), CHILD(766), CHILD(766), CHILD(766), CHILD(766),
+  CHILD(766), CHILD(766), CHILD(766), CHILD(766), CHILD(766), CHILD(766),
+  CHILD(766), CHILD(766), CHILD(766), CHILD(766), CHILD(766), CHILD(766),
+  CHILD(766), CHILD(766), CHILD(766), CHILD(766), CHILD(766), CHILD(766),
+  CHILD(766), CHILD(766), CHILD(766), CHILD(766), CHILD(766), CHILD(766),
+  CHILD(766), CHILD(781), CHILD(781), CHILD(781), CHILD(781), CHILD(781),
+  CHILD(781), CHILD(781), CHILD(781), CHILD(781), CHILD(781), CHILD(781),
+  CHILD(781), CHILD(781), CHILD(781), CHILD(781), CHILD(781), CHILD(781),
+  CHILD(781), CHILD(781), CHILD(781), CHILD(781), CHILD(781), CHILD(781),
+  CHILD(781), CHILD(781), CHILD(781), CHILD(781), CHILD(781), CHILD(781),
+  CHILD(781), CHILD(781), CHILD(781), CHILD(796), CHILD(796), CHILD(796),
+  CHILD(796), CHILD(796), CHILD(796), CHILD(796), CHILD(796), CHILD(796),
+  CHILD(796), CHILD(796), CHILD(796), CHILD(796), CHILD(796), CHILD(796),
+  CHILD(796), CHILD(796), CHILD(796), CHILD(796), CHILD(796), CHILD(796),
+  CHILD(796), CHILD(796), CHILD(796), CHILD(796), CHILD(796), CHILD(796),
+  CHILD(796), CHILD(796), CHILD(796), CHILD(796), CHILD(796), CHILD(826),
+  CHILD(826), CHILD(826), CHILD(826), CHILD(826), CHILD(826), CHILD(826),
+  CHILD(826), CHILD(826), CHILD(826), CHILD(826), CHILD(826), CHILD(826),
+  CHILD(826), CHILD(826), CHILD(826), CHILD(843), CHILD(843), CHILD(843),
+  CHILD(843), CHILD(843), CHILD(843), CHILD(843), CHILD(843), CHILD(843),
+  CHILD(843), CHILD(843), CHILD(843), CHILD(843), CHILD(843), CHILD(843),
+  CHILD(843), CHILD(860), CHILD(899), CHILD(923), CHILD(932),
+  TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE,
+  TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE,
+  TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE,
+  CHILD(941), CHILD(950), CHILD(974), CHILD(983), TILEPRO_OPC_NONE,
+  TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE,
+  TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE,
+  TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_MM,
+  TILEPRO_OPC_MM, TILEPRO_OPC_MM, TILEPRO_OPC_MM, TILEPRO_OPC_MM,
+  TILEPRO_OPC_MM, TILEPRO_OPC_MM, TILEPRO_OPC_MM, TILEPRO_OPC_MM,
+  TILEPRO_OPC_MM, TILEPRO_OPC_MM, TILEPRO_OPC_MM, TILEPRO_OPC_MM,
+  TILEPRO_OPC_MM, TILEPRO_OPC_MM, TILEPRO_OPC_MM, TILEPRO_OPC_MM,
+  TILEPRO_OPC_MM, TILEPRO_OPC_MM, TILEPRO_OPC_MM, TILEPRO_OPC_MM,
+  TILEPRO_OPC_MM, TILEPRO_OPC_MM, TILEPRO_OPC_MM, TILEPRO_OPC_MM,
+  TILEPRO_OPC_MM, TILEPRO_OPC_MM, TILEPRO_OPC_MM, TILEPRO_OPC_MM,
+  TILEPRO_OPC_MM, TILEPRO_OPC_MM, TILEPRO_OPC_MM, CHILD(992),
+  TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE,
+  TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE,
+  TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE,
+  TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, CHILD(1334),
+  TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE,
+  TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE,
+  TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE,
+  TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE,
+  TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE,
+  TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE,
+  TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE,
+  TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE,
+  TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE,
+  TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE,
+  TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE,
+  TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_J,
+  TILEPRO_OPC_J, TILEPRO_OPC_J, TILEPRO_OPC_J, TILEPRO_OPC_J, TILEPRO_OPC_J,
+  TILEPRO_OPC_J, TILEPRO_OPC_J, TILEPRO_OPC_J, TILEPRO_OPC_J, TILEPRO_OPC_J,
+  TILEPRO_OPC_J, TILEPRO_OPC_J, TILEPRO_OPC_J, TILEPRO_OPC_J, TILEPRO_OPC_J,
+  TILEPRO_OPC_J, TILEPRO_OPC_J, TILEPRO_OPC_J, TILEPRO_OPC_J, TILEPRO_OPC_J,
+  TILEPRO_OPC_J, TILEPRO_OPC_J, TILEPRO_OPC_J, TILEPRO_OPC_J, TILEPRO_OPC_J,
+  TILEPRO_OPC_J, TILEPRO_OPC_J, TILEPRO_OPC_J, TILEPRO_OPC_J, TILEPRO_OPC_J,
+  TILEPRO_OPC_J, TILEPRO_OPC_J, TILEPRO_OPC_J, TILEPRO_OPC_J, TILEPRO_OPC_J,
+  TILEPRO_OPC_J, TILEPRO_OPC_J, TILEPRO_OPC_J, TILEPRO_OPC_J, TILEPRO_OPC_J,
+  TILEPRO_OPC_J, TILEPRO_OPC_J, TILEPRO_OPC_J, TILEPRO_OPC_J, TILEPRO_OPC_J,
+  TILEPRO_OPC_J, TILEPRO_OPC_J, TILEPRO_OPC_J, TILEPRO_OPC_J, TILEPRO_OPC_J,
+  TILEPRO_OPC_J, TILEPRO_OPC_J, TILEPRO_OPC_J, TILEPRO_OPC_J, TILEPRO_OPC_J,
+  TILEPRO_OPC_J, TILEPRO_OPC_J, TILEPRO_OPC_J, TILEPRO_OPC_J, TILEPRO_OPC_J,
+  TILEPRO_OPC_J, TILEPRO_OPC_J, TILEPRO_OPC_J, TILEPRO_OPC_JAL,
+  TILEPRO_OPC_JAL, TILEPRO_OPC_JAL, TILEPRO_OPC_JAL, TILEPRO_OPC_JAL,
+  TILEPRO_OPC_JAL, TILEPRO_OPC_JAL, TILEPRO_OPC_JAL, TILEPRO_OPC_JAL,
+  TILEPRO_OPC_JAL, TILEPRO_OPC_JAL, TILEPRO_OPC_JAL, TILEPRO_OPC_JAL,
+  TILEPRO_OPC_JAL, TILEPRO_OPC_JAL, TILEPRO_OPC_JAL, TILEPRO_OPC_JAL,
+  TILEPRO_OPC_JAL, TILEPRO_OPC_JAL, TILEPRO_OPC_JAL, TILEPRO_OPC_JAL,
+  TILEPRO_OPC_JAL, TILEPRO_OPC_JAL, TILEPRO_OPC_JAL, TILEPRO_OPC_JAL,
+  TILEPRO_OPC_JAL, TILEPRO_OPC_JAL, TILEPRO_OPC_JAL, TILEPRO_OPC_JAL,
+  TILEPRO_OPC_JAL, TILEPRO_OPC_JAL, TILEPRO_OPC_JAL, TILEPRO_OPC_JAL,
+  TILEPRO_OPC_JAL, TILEPRO_OPC_JAL, TILEPRO_OPC_JAL, TILEPRO_OPC_JAL,
+  TILEPRO_OPC_JAL, TILEPRO_OPC_JAL, TILEPRO_OPC_JAL, TILEPRO_OPC_JAL,
+  TILEPRO_OPC_JAL, TILEPRO_OPC_JAL, TILEPRO_OPC_JAL, TILEPRO_OPC_JAL,
+  TILEPRO_OPC_JAL, TILEPRO_OPC_JAL, TILEPRO_OPC_JAL, TILEPRO_OPC_JAL,
+  TILEPRO_OPC_JAL, TILEPRO_OPC_JAL, TILEPRO_OPC_JAL, TILEPRO_OPC_JAL,
+  TILEPRO_OPC_JAL, TILEPRO_OPC_JAL, TILEPRO_OPC_JAL, TILEPRO_OPC_JAL,
+  TILEPRO_OPC_JAL, TILEPRO_OPC_JAL, TILEPRO_OPC_JAL, TILEPRO_OPC_JAL,
+  TILEPRO_OPC_JAL, TILEPRO_OPC_JAL, TILEPRO_OPC_JAL, TILEPRO_OPC_NONE,
+  TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE,
+  TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE,
+  TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE,
+  TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE,
+  TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE,
+  TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE,
+  TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE,
+  TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE,
+  TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE,
+  TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE,
+  TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE,
+  TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE,
+  TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE,
+  TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE,
+  TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE,
+  TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE,
+  BITFIELD(49, 5) /* index 513 */,
+  TILEPRO_OPC_NONE, TILEPRO_OPC_ADDB, TILEPRO_OPC_ADDH, TILEPRO_OPC_ADD,
+  TILEPRO_OPC_AND, TILEPRO_OPC_INTHB, TILEPRO_OPC_INTHH, TILEPRO_OPC_INTLB,
+  TILEPRO_OPC_INTLH, TILEPRO_OPC_JALRP, TILEPRO_OPC_JALR, TILEPRO_OPC_JRP,
+  TILEPRO_OPC_JR, TILEPRO_OPC_LNK, TILEPRO_OPC_MAXB_U, TILEPRO_OPC_MAXH,
+  TILEPRO_OPC_MINB_U, TILEPRO_OPC_MINH, TILEPRO_OPC_MNZB, TILEPRO_OPC_MNZH,
+  TILEPRO_OPC_MNZ, TILEPRO_OPC_MZB, TILEPRO_OPC_MZH, TILEPRO_OPC_MZ,
+  TILEPRO_OPC_NOR, CHILD(546), TILEPRO_OPC_PACKHB, TILEPRO_OPC_PACKLB,
+  TILEPRO_OPC_RL, TILEPRO_OPC_S1A, TILEPRO_OPC_S2A, TILEPRO_OPC_S3A,
+  BITFIELD(43, 2) /* index 546 */,
+  TILEPRO_OPC_OR, TILEPRO_OPC_OR, TILEPRO_OPC_OR, CHILD(551),
+  BITFIELD(45, 2) /* index 551 */,
+  TILEPRO_OPC_OR, TILEPRO_OPC_OR, TILEPRO_OPC_OR, CHILD(556),
+  BITFIELD(47, 2) /* index 556 */,
+  TILEPRO_OPC_OR, TILEPRO_OPC_OR, TILEPRO_OPC_OR, TILEPRO_OPC_MOVE,
+  BITFIELD(49, 5) /* index 561 */,
+  TILEPRO_OPC_SB, TILEPRO_OPC_SEQB, TILEPRO_OPC_SEQH, TILEPRO_OPC_SEQ,
+  TILEPRO_OPC_SHLB, TILEPRO_OPC_SHLH, TILEPRO_OPC_SHL, TILEPRO_OPC_SHRB,
+  TILEPRO_OPC_SHRH, TILEPRO_OPC_SHR, TILEPRO_OPC_SH, TILEPRO_OPC_SLTB,
+  TILEPRO_OPC_SLTB_U, TILEPRO_OPC_SLTEB, TILEPRO_OPC_SLTEB_U,
+  TILEPRO_OPC_SLTEH, TILEPRO_OPC_SLTEH_U, TILEPRO_OPC_SLTE,
+  TILEPRO_OPC_SLTE_U, TILEPRO_OPC_SLTH, TILEPRO_OPC_SLTH_U, TILEPRO_OPC_SLT,
+  TILEPRO_OPC_SLT_U, TILEPRO_OPC_SNEB, TILEPRO_OPC_SNEH, TILEPRO_OPC_SNE,
+  TILEPRO_OPC_SRAB, TILEPRO_OPC_SRAH, TILEPRO_OPC_SRA, TILEPRO_OPC_SUBB,
+  TILEPRO_OPC_SUBH, TILEPRO_OPC_SUB,
+  BITFIELD(49, 4) /* index 594 */,
+  CHILD(611), CHILD(614), CHILD(617), CHILD(620), CHILD(623), CHILD(626),
+  CHILD(629), CHILD(632), CHILD(635), CHILD(638), TILEPRO_OPC_NONE,
+  TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE,
+  TILEPRO_OPC_NONE,
+  BITFIELD(53, 1) /* index 611 */,
+  TILEPRO_OPC_SW, TILEPRO_OPC_NONE,
+  BITFIELD(53, 1) /* index 614 */,
+  TILEPRO_OPC_XOR, TILEPRO_OPC_NONE,
+  BITFIELD(53, 1) /* index 617 */,
+  TILEPRO_OPC_ADDS, TILEPRO_OPC_NONE,
+  BITFIELD(53, 1) /* index 620 */,
+  TILEPRO_OPC_SUBS, TILEPRO_OPC_NONE,
+  BITFIELD(53, 1) /* index 623 */,
+  TILEPRO_OPC_ADDBS_U, TILEPRO_OPC_NONE,
+  BITFIELD(53, 1) /* index 626 */,
+  TILEPRO_OPC_ADDHS, TILEPRO_OPC_NONE,
+  BITFIELD(53, 1) /* index 629 */,
+  TILEPRO_OPC_SUBBS_U, TILEPRO_OPC_NONE,
+  BITFIELD(53, 1) /* index 632 */,
+  TILEPRO_OPC_SUBHS, TILEPRO_OPC_NONE,
+  BITFIELD(53, 1) /* index 635 */,
+  TILEPRO_OPC_PACKHS, TILEPRO_OPC_NONE,
+  BITFIELD(53, 1) /* index 638 */,
+  TILEPRO_OPC_PACKBS_U, TILEPRO_OPC_NONE,
+  BITFIELD(49, 5) /* index 641 */,
+  TILEPRO_OPC_NONE, TILEPRO_OPC_ADDB_SN, TILEPRO_OPC_ADDH_SN,
+  TILEPRO_OPC_ADD_SN, TILEPRO_OPC_AND_SN, TILEPRO_OPC_INTHB_SN,
+  TILEPRO_OPC_INTHH_SN, TILEPRO_OPC_INTLB_SN, TILEPRO_OPC_INTLH_SN,
+  TILEPRO_OPC_JALRP, TILEPRO_OPC_JALR, TILEPRO_OPC_JRP, TILEPRO_OPC_JR,
+  TILEPRO_OPC_LNK_SN, TILEPRO_OPC_MAXB_U_SN, TILEPRO_OPC_MAXH_SN,
+  TILEPRO_OPC_MINB_U_SN, TILEPRO_OPC_MINH_SN, TILEPRO_OPC_MNZB_SN,
+  TILEPRO_OPC_MNZH_SN, TILEPRO_OPC_MNZ_SN, TILEPRO_OPC_MZB_SN,
+  TILEPRO_OPC_MZH_SN, TILEPRO_OPC_MZ_SN, TILEPRO_OPC_NOR_SN, CHILD(674),
+  TILEPRO_OPC_PACKHB_SN, TILEPRO_OPC_PACKLB_SN, TILEPRO_OPC_RL_SN,
+  TILEPRO_OPC_S1A_SN, TILEPRO_OPC_S2A_SN, TILEPRO_OPC_S3A_SN,
+  BITFIELD(43, 2) /* index 674 */,
+  TILEPRO_OPC_OR_SN, TILEPRO_OPC_OR_SN, TILEPRO_OPC_OR_SN, CHILD(679),
+  BITFIELD(45, 2) /* index 679 */,
+  TILEPRO_OPC_OR_SN, TILEPRO_OPC_OR_SN, TILEPRO_OPC_OR_SN, CHILD(684),
+  BITFIELD(47, 2) /* index 684 */,
+  TILEPRO_OPC_OR_SN, TILEPRO_OPC_OR_SN, TILEPRO_OPC_OR_SN,
+  TILEPRO_OPC_MOVE_SN,
+  BITFIELD(49, 5) /* index 689 */,
+  TILEPRO_OPC_SB, TILEPRO_OPC_SEQB_SN, TILEPRO_OPC_SEQH_SN,
+  TILEPRO_OPC_SEQ_SN, TILEPRO_OPC_SHLB_SN, TILEPRO_OPC_SHLH_SN,
+  TILEPRO_OPC_SHL_SN, TILEPRO_OPC_SHRB_SN, TILEPRO_OPC_SHRH_SN,
+  TILEPRO_OPC_SHR_SN, TILEPRO_OPC_SH, TILEPRO_OPC_SLTB_SN,
+  TILEPRO_OPC_SLTB_U_SN, TILEPRO_OPC_SLTEB_SN, TILEPRO_OPC_SLTEB_U_SN,
+  TILEPRO_OPC_SLTEH_SN, TILEPRO_OPC_SLTEH_U_SN, TILEPRO_OPC_SLTE_SN,
+  TILEPRO_OPC_SLTE_U_SN, TILEPRO_OPC_SLTH_SN, TILEPRO_OPC_SLTH_U_SN,
+  TILEPRO_OPC_SLT_SN, TILEPRO_OPC_SLT_U_SN, TILEPRO_OPC_SNEB_SN,
+  TILEPRO_OPC_SNEH_SN, TILEPRO_OPC_SNE_SN, TILEPRO_OPC_SRAB_SN,
+  TILEPRO_OPC_SRAH_SN, TILEPRO_OPC_SRA_SN, TILEPRO_OPC_SUBB_SN,
+  TILEPRO_OPC_SUBH_SN, TILEPRO_OPC_SUB_SN,
+  BITFIELD(49, 4) /* index 722 */,
+  CHILD(611), CHILD(739), CHILD(742), CHILD(745), CHILD(748), CHILD(751),
+  CHILD(754), CHILD(757), CHILD(760), CHILD(763), TILEPRO_OPC_NONE,
+  TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE,
+  TILEPRO_OPC_NONE,
+  BITFIELD(53, 1) /* index 739 */,
+  TILEPRO_OPC_XOR_SN, TILEPRO_OPC_NONE,
+  BITFIELD(53, 1) /* index 742 */,
+  TILEPRO_OPC_ADDS_SN, TILEPRO_OPC_NONE,
+  BITFIELD(53, 1) /* index 745 */,
+  TILEPRO_OPC_SUBS_SN, TILEPRO_OPC_NONE,
+  BITFIELD(53, 1) /* index 748 */,
+  TILEPRO_OPC_ADDBS_U_SN, TILEPRO_OPC_NONE,
+  BITFIELD(53, 1) /* index 751 */,
+  TILEPRO_OPC_ADDHS_SN, TILEPRO_OPC_NONE,
+  BITFIELD(53, 1) /* index 754 */,
+  TILEPRO_OPC_SUBBS_U_SN, TILEPRO_OPC_NONE,
+  BITFIELD(53, 1) /* index 757 */,
+  TILEPRO_OPC_SUBHS_SN, TILEPRO_OPC_NONE,
+  BITFIELD(53, 1) /* index 760 */,
+  TILEPRO_OPC_PACKHS_SN, TILEPRO_OPC_NONE,
+  BITFIELD(53, 1) /* index 763 */,
+  TILEPRO_OPC_PACKBS_U_SN, TILEPRO_OPC_NONE,
+  BITFIELD(37, 2) /* index 766 */,
+  TILEPRO_OPC_ADDLI_SN, TILEPRO_OPC_ADDLI_SN, TILEPRO_OPC_ADDLI_SN,
+  CHILD(771),
+  BITFIELD(39, 2) /* index 771 */,
+  TILEPRO_OPC_ADDLI_SN, TILEPRO_OPC_ADDLI_SN, TILEPRO_OPC_ADDLI_SN,
+  CHILD(776),
+  BITFIELD(41, 2) /* index 776 */,
+  TILEPRO_OPC_ADDLI_SN, TILEPRO_OPC_ADDLI_SN, TILEPRO_OPC_ADDLI_SN,
+  TILEPRO_OPC_MOVELI_SN,
+  BITFIELD(37, 2) /* index 781 */,
+  TILEPRO_OPC_ADDLI, TILEPRO_OPC_ADDLI, TILEPRO_OPC_ADDLI, CHILD(786),
+  BITFIELD(39, 2) /* index 786 */,
+  TILEPRO_OPC_ADDLI, TILEPRO_OPC_ADDLI, TILEPRO_OPC_ADDLI, CHILD(791),
+  BITFIELD(41, 2) /* index 791 */,
+  TILEPRO_OPC_ADDLI, TILEPRO_OPC_ADDLI, TILEPRO_OPC_ADDLI, TILEPRO_OPC_MOVELI,
+  BITFIELD(31, 2) /* index 796 */,
+  TILEPRO_OPC_AULI, TILEPRO_OPC_AULI, TILEPRO_OPC_AULI, CHILD(801),
+  BITFIELD(33, 2) /* index 801 */,
+  TILEPRO_OPC_AULI, TILEPRO_OPC_AULI, TILEPRO_OPC_AULI, CHILD(806),
+  BITFIELD(35, 2) /* index 806 */,
+  TILEPRO_OPC_AULI, TILEPRO_OPC_AULI, TILEPRO_OPC_AULI, CHILD(811),
+  BITFIELD(37, 2) /* index 811 */,
+  TILEPRO_OPC_AULI, TILEPRO_OPC_AULI, TILEPRO_OPC_AULI, CHILD(816),
+  BITFIELD(39, 2) /* index 816 */,
+  TILEPRO_OPC_AULI, TILEPRO_OPC_AULI, TILEPRO_OPC_AULI, CHILD(821),
+  BITFIELD(41, 2) /* index 821 */,
+  TILEPRO_OPC_AULI, TILEPRO_OPC_AULI, TILEPRO_OPC_AULI, TILEPRO_OPC_INFOL,
+  BITFIELD(31, 4) /* index 826 */,
+  TILEPRO_OPC_BZ, TILEPRO_OPC_BZT, TILEPRO_OPC_BNZ, TILEPRO_OPC_BNZT,
+  TILEPRO_OPC_BGZ, TILEPRO_OPC_BGZT, TILEPRO_OPC_BGEZ, TILEPRO_OPC_BGEZT,
+  TILEPRO_OPC_BLZ, TILEPRO_OPC_BLZT, TILEPRO_OPC_BLEZ, TILEPRO_OPC_BLEZT,
+  TILEPRO_OPC_BBS, TILEPRO_OPC_BBST, TILEPRO_OPC_BBNS, TILEPRO_OPC_BBNST,
+  BITFIELD(31, 4) /* index 843 */,
+  TILEPRO_OPC_BZ_SN, TILEPRO_OPC_BZT_SN, TILEPRO_OPC_BNZ_SN,
+  TILEPRO_OPC_BNZT_SN, TILEPRO_OPC_BGZ_SN, TILEPRO_OPC_BGZT_SN,
+  TILEPRO_OPC_BGEZ_SN, TILEPRO_OPC_BGEZT_SN, TILEPRO_OPC_BLZ_SN,
+  TILEPRO_OPC_BLZT_SN, TILEPRO_OPC_BLEZ_SN, TILEPRO_OPC_BLEZT_SN,
+  TILEPRO_OPC_BBS_SN, TILEPRO_OPC_BBST_SN, TILEPRO_OPC_BBNS_SN,
+  TILEPRO_OPC_BBNST_SN,
+  BITFIELD(51, 3) /* index 860 */,
+  TILEPRO_OPC_NONE, TILEPRO_OPC_ADDIB, TILEPRO_OPC_ADDIH, TILEPRO_OPC_ADDI,
+  CHILD(869), TILEPRO_OPC_MAXIB_U, TILEPRO_OPC_MAXIH, TILEPRO_OPC_MFSPR,
+  BITFIELD(31, 2) /* index 869 */,
+  TILEPRO_OPC_ANDI, TILEPRO_OPC_ANDI, TILEPRO_OPC_ANDI, CHILD(874),
+  BITFIELD(33, 2) /* index 874 */,
+  TILEPRO_OPC_ANDI, TILEPRO_OPC_ANDI, TILEPRO_OPC_ANDI, CHILD(879),
+  BITFIELD(35, 2) /* index 879 */,
+  TILEPRO_OPC_ANDI, TILEPRO_OPC_ANDI, TILEPRO_OPC_ANDI, CHILD(884),
+  BITFIELD(37, 2) /* index 884 */,
+  TILEPRO_OPC_ANDI, TILEPRO_OPC_ANDI, TILEPRO_OPC_ANDI, CHILD(889),
+  BITFIELD(39, 2) /* index 889 */,
+  TILEPRO_OPC_ANDI, TILEPRO_OPC_ANDI, TILEPRO_OPC_ANDI, CHILD(894),
+  BITFIELD(41, 2) /* index 894 */,
+  TILEPRO_OPC_ANDI, TILEPRO_OPC_ANDI, TILEPRO_OPC_ANDI, TILEPRO_OPC_INFO,
+  BITFIELD(51, 3) /* index 899 */,
+  TILEPRO_OPC_MINIB_U, TILEPRO_OPC_MINIH, TILEPRO_OPC_MTSPR, CHILD(908),
+  TILEPRO_OPC_SEQIB, TILEPRO_OPC_SEQIH, TILEPRO_OPC_SEQI, TILEPRO_OPC_SLTIB,
+  BITFIELD(37, 2) /* index 908 */,
+  TILEPRO_OPC_ORI, TILEPRO_OPC_ORI, TILEPRO_OPC_ORI, CHILD(913),
+  BITFIELD(39, 2) /* index 913 */,
+  TILEPRO_OPC_ORI, TILEPRO_OPC_ORI, TILEPRO_OPC_ORI, CHILD(918),
+  BITFIELD(41, 2) /* index 918 */,
+  TILEPRO_OPC_ORI, TILEPRO_OPC_ORI, TILEPRO_OPC_ORI, TILEPRO_OPC_MOVEI,
+  BITFIELD(51, 3) /* index 923 */,
+  TILEPRO_OPC_SLTIB_U, TILEPRO_OPC_SLTIH, TILEPRO_OPC_SLTIH_U,
+  TILEPRO_OPC_SLTI, TILEPRO_OPC_SLTI_U, TILEPRO_OPC_XORI, TILEPRO_OPC_LBADD,
+  TILEPRO_OPC_LBADD_U,
+  BITFIELD(51, 3) /* index 932 */,
+  TILEPRO_OPC_LHADD, TILEPRO_OPC_LHADD_U, TILEPRO_OPC_LWADD,
+  TILEPRO_OPC_LWADD_NA, TILEPRO_OPC_SBADD, TILEPRO_OPC_SHADD,
+  TILEPRO_OPC_SWADD, TILEPRO_OPC_NONE,
+  BITFIELD(51, 3) /* index 941 */,
+  TILEPRO_OPC_NONE, TILEPRO_OPC_ADDIB_SN, TILEPRO_OPC_ADDIH_SN,
+  TILEPRO_OPC_ADDI_SN, TILEPRO_OPC_ANDI_SN, TILEPRO_OPC_MAXIB_U_SN,
+  TILEPRO_OPC_MAXIH_SN, TILEPRO_OPC_MFSPR,
+  BITFIELD(51, 3) /* index 950 */,
+  TILEPRO_OPC_MINIB_U_SN, TILEPRO_OPC_MINIH_SN, TILEPRO_OPC_MTSPR, CHILD(959),
+  TILEPRO_OPC_SEQIB_SN, TILEPRO_OPC_SEQIH_SN, TILEPRO_OPC_SEQI_SN,
+  TILEPRO_OPC_SLTIB_SN,
+  BITFIELD(37, 2) /* index 959 */,
+  TILEPRO_OPC_ORI_SN, TILEPRO_OPC_ORI_SN, TILEPRO_OPC_ORI_SN, CHILD(964),
+  BITFIELD(39, 2) /* index 964 */,
+  TILEPRO_OPC_ORI_SN, TILEPRO_OPC_ORI_SN, TILEPRO_OPC_ORI_SN, CHILD(969),
+  BITFIELD(41, 2) /* index 969 */,
+  TILEPRO_OPC_ORI_SN, TILEPRO_OPC_ORI_SN, TILEPRO_OPC_ORI_SN,
+  TILEPRO_OPC_MOVEI_SN,
+  BITFIELD(51, 3) /* index 974 */,
+  TILEPRO_OPC_SLTIB_U_SN, TILEPRO_OPC_SLTIH_SN, TILEPRO_OPC_SLTIH_U_SN,
+  TILEPRO_OPC_SLTI_SN, TILEPRO_OPC_SLTI_U_SN, TILEPRO_OPC_XORI_SN,
+  TILEPRO_OPC_LBADD_SN, TILEPRO_OPC_LBADD_U_SN,
+  BITFIELD(51, 3) /* index 983 */,
+  TILEPRO_OPC_LHADD_SN, TILEPRO_OPC_LHADD_U_SN, TILEPRO_OPC_LWADD_SN,
+  TILEPRO_OPC_LWADD_NA_SN, TILEPRO_OPC_SBADD, TILEPRO_OPC_SHADD,
+  TILEPRO_OPC_SWADD, TILEPRO_OPC_NONE,
+  BITFIELD(46, 7) /* index 992 */,
+  TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE,
+  CHILD(1121), CHILD(1121), CHILD(1121), CHILD(1121), CHILD(1124),
+  CHILD(1124), CHILD(1124), CHILD(1124), CHILD(1127), CHILD(1127),
+  CHILD(1127), CHILD(1127), CHILD(1130), CHILD(1130), CHILD(1130),
+  CHILD(1130), CHILD(1133), CHILD(1133), CHILD(1133), CHILD(1133),
+  CHILD(1136), CHILD(1136), CHILD(1136), CHILD(1136), CHILD(1139),
+  CHILD(1139), CHILD(1139), CHILD(1139), CHILD(1142), CHILD(1142),
+  CHILD(1142), CHILD(1142), CHILD(1145), CHILD(1145), CHILD(1145),
+  CHILD(1145), CHILD(1148), CHILD(1148), CHILD(1148), CHILD(1148),
+  CHILD(1151), CHILD(1242), CHILD(1290), CHILD(1323), TILEPRO_OPC_NONE,
+  TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE,
+  TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE,
+  TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE,
+  TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE,
+  TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE,
+  TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE,
+  TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE,
+  TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE,
+  TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE,
+  TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE,
+  TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE,
+  TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE,
+  TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE,
+  TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE,
+  TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE,
+  TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE,
+  TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE,
+  TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE,
+  TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE,
+  TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE,
+  BITFIELD(53, 1) /* index 1121 */,
+  TILEPRO_OPC_RLI, TILEPRO_OPC_NONE,
+  BITFIELD(53, 1) /* index 1124 */,
+  TILEPRO_OPC_SHLIB, TILEPRO_OPC_NONE,
+  BITFIELD(53, 1) /* index 1127 */,
+  TILEPRO_OPC_SHLIH, TILEPRO_OPC_NONE,
+  BITFIELD(53, 1) /* index 1130 */,
+  TILEPRO_OPC_SHLI, TILEPRO_OPC_NONE,
+  BITFIELD(53, 1) /* index 1133 */,
+  TILEPRO_OPC_SHRIB, TILEPRO_OPC_NONE,
+  BITFIELD(53, 1) /* index 1136 */,
+  TILEPRO_OPC_SHRIH, TILEPRO_OPC_NONE,
+  BITFIELD(53, 1) /* index 1139 */,
+  TILEPRO_OPC_SHRI, TILEPRO_OPC_NONE,
+  BITFIELD(53, 1) /* index 1142 */,
+  TILEPRO_OPC_SRAIB, TILEPRO_OPC_NONE,
+  BITFIELD(53, 1) /* index 1145 */,
+  TILEPRO_OPC_SRAIH, TILEPRO_OPC_NONE,
+  BITFIELD(53, 1) /* index 1148 */,
+  TILEPRO_OPC_SRAI, TILEPRO_OPC_NONE,
+  BITFIELD(43, 3) /* index 1151 */,
+  TILEPRO_OPC_NONE, CHILD(1160), CHILD(1163), CHILD(1166), CHILD(1169),
+  CHILD(1172), CHILD(1175), CHILD(1178),
+  BITFIELD(53, 1) /* index 1160 */,
+  TILEPRO_OPC_DRAIN, TILEPRO_OPC_NONE,
+  BITFIELD(53, 1) /* index 1163 */,
+  TILEPRO_OPC_DTLBPR, TILEPRO_OPC_NONE,
+  BITFIELD(53, 1) /* index 1166 */,
+  TILEPRO_OPC_FINV, TILEPRO_OPC_NONE,
+  BITFIELD(53, 1) /* index 1169 */,
+  TILEPRO_OPC_FLUSH, TILEPRO_OPC_NONE,
+  BITFIELD(53, 1) /* index 1172 */,
+  TILEPRO_OPC_FNOP, TILEPRO_OPC_NONE,
+  BITFIELD(53, 1) /* index 1175 */,
+  TILEPRO_OPC_ICOH, TILEPRO_OPC_NONE,
+  BITFIELD(31, 2) /* index 1178 */,
+  CHILD(1183), CHILD(1211), CHILD(1239), CHILD(1239),
+  BITFIELD(53, 1) /* index 1183 */,
+  CHILD(1186), TILEPRO_OPC_NONE,
+  BITFIELD(33, 2) /* index 1186 */,
+  TILEPRO_OPC_ILL, TILEPRO_OPC_ILL, TILEPRO_OPC_ILL, CHILD(1191),
+  BITFIELD(35, 2) /* index 1191 */,
+  TILEPRO_OPC_ILL, CHILD(1196), TILEPRO_OPC_ILL, TILEPRO_OPC_ILL,
+  BITFIELD(37, 2) /* index 1196 */,
+  TILEPRO_OPC_ILL, CHILD(1201), TILEPRO_OPC_ILL, TILEPRO_OPC_ILL,
+  BITFIELD(39, 2) /* index 1201 */,
+  TILEPRO_OPC_ILL, CHILD(1206), TILEPRO_OPC_ILL, TILEPRO_OPC_ILL,
+  BITFIELD(41, 2) /* index 1206 */,
+  TILEPRO_OPC_ILL, TILEPRO_OPC_ILL, TILEPRO_OPC_BPT, TILEPRO_OPC_ILL,
+  BITFIELD(53, 1) /* index 1211 */,
+  CHILD(1214), TILEPRO_OPC_NONE,
+  BITFIELD(33, 2) /* index 1214 */,
+  TILEPRO_OPC_ILL, TILEPRO_OPC_ILL, TILEPRO_OPC_ILL, CHILD(1219),
+  BITFIELD(35, 2) /* index 1219 */,
+  TILEPRO_OPC_ILL, CHILD(1224), TILEPRO_OPC_ILL, TILEPRO_OPC_ILL,
+  BITFIELD(37, 2) /* index 1224 */,
+  TILEPRO_OPC_ILL, CHILD(1229), TILEPRO_OPC_ILL, TILEPRO_OPC_ILL,
+  BITFIELD(39, 2) /* index 1229 */,
+  TILEPRO_OPC_ILL, CHILD(1234), TILEPRO_OPC_ILL, TILEPRO_OPC_ILL,
+  BITFIELD(41, 2) /* index 1234 */,
+  TILEPRO_OPC_ILL, TILEPRO_OPC_ILL, TILEPRO_OPC_RAISE, TILEPRO_OPC_ILL,
+  BITFIELD(53, 1) /* index 1239 */,
+  TILEPRO_OPC_ILL, TILEPRO_OPC_NONE,
+  BITFIELD(43, 3) /* index 1242 */,
+  CHILD(1251), CHILD(1254), CHILD(1257), CHILD(1275), CHILD(1278),
+  CHILD(1281), CHILD(1284), CHILD(1287),
+  BITFIELD(53, 1) /* index 1251 */,
+  TILEPRO_OPC_INV, TILEPRO_OPC_NONE,
+  BITFIELD(53, 1) /* index 1254 */,
+  TILEPRO_OPC_IRET, TILEPRO_OPC_NONE,
+  BITFIELD(53, 1) /* index 1257 */,
+  CHILD(1260), TILEPRO_OPC_NONE,
+  BITFIELD(31, 2) /* index 1260 */,
+  TILEPRO_OPC_LB, TILEPRO_OPC_LB, TILEPRO_OPC_LB, CHILD(1265),
+  BITFIELD(33, 2) /* index 1265 */,
+  TILEPRO_OPC_LB, TILEPRO_OPC_LB, TILEPRO_OPC_LB, CHILD(1270),
+  BITFIELD(35, 2) /* index 1270 */,
+  TILEPRO_OPC_LB, TILEPRO_OPC_LB, TILEPRO_OPC_LB, TILEPRO_OPC_PREFETCH,
+  BITFIELD(53, 1) /* index 1275 */,
+  TILEPRO_OPC_LB_U, TILEPRO_OPC_NONE,
+  BITFIELD(53, 1) /* index 1278 */,
+  TILEPRO_OPC_LH, TILEPRO_OPC_NONE,
+  BITFIELD(53, 1) /* index 1281 */,
+  TILEPRO_OPC_LH_U, TILEPRO_OPC_NONE,
+  BITFIELD(53, 1) /* index 1284 */,
+  TILEPRO_OPC_LW, TILEPRO_OPC_NONE,
+  BITFIELD(53, 1) /* index 1287 */,
+  TILEPRO_OPC_MF, TILEPRO_OPC_NONE,
+  BITFIELD(43, 3) /* index 1290 */,
+  CHILD(1299), CHILD(1302), CHILD(1305), CHILD(1308), CHILD(1311),
+  CHILD(1314), CHILD(1317), CHILD(1320),
+  BITFIELD(53, 1) /* index 1299 */,
+  TILEPRO_OPC_NAP, TILEPRO_OPC_NONE,
+  BITFIELD(53, 1) /* index 1302 */,
+  TILEPRO_OPC_NOP, TILEPRO_OPC_NONE,
+  BITFIELD(53, 1) /* index 1305 */,
+  TILEPRO_OPC_SWINT0, TILEPRO_OPC_NONE,
+  BITFIELD(53, 1) /* index 1308 */,
+  TILEPRO_OPC_SWINT1, TILEPRO_OPC_NONE,
+  BITFIELD(53, 1) /* index 1311 */,
+  TILEPRO_OPC_SWINT2, TILEPRO_OPC_NONE,
+  BITFIELD(53, 1) /* index 1314 */,
+  TILEPRO_OPC_SWINT3, TILEPRO_OPC_NONE,
+  BITFIELD(53, 1) /* index 1317 */,
+  TILEPRO_OPC_TNS, TILEPRO_OPC_NONE,
+  BITFIELD(53, 1) /* index 1320 */,
+  TILEPRO_OPC_WH64, TILEPRO_OPC_NONE,
+  BITFIELD(43, 2) /* index 1323 */,
+  CHILD(1328), TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE,
+  BITFIELD(45, 1) /* index 1328 */,
+  CHILD(1331), TILEPRO_OPC_NONE,
+  BITFIELD(53, 1) /* index 1331 */,
+  TILEPRO_OPC_LW_NA, TILEPRO_OPC_NONE,
+  BITFIELD(46, 7) /* index 1334 */,
+  TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE,
+  CHILD(1463), CHILD(1463), CHILD(1463), CHILD(1463), CHILD(1466),
+  CHILD(1466), CHILD(1466), CHILD(1466), CHILD(1469), CHILD(1469),
+  CHILD(1469), CHILD(1469), CHILD(1472), CHILD(1472), CHILD(1472),
+  CHILD(1472), CHILD(1475), CHILD(1475), CHILD(1475), CHILD(1475),
+  CHILD(1478), CHILD(1478), CHILD(1478), CHILD(1478), CHILD(1481),
+  CHILD(1481), CHILD(1481), CHILD(1481), CHILD(1484), CHILD(1484),
+  CHILD(1484), CHILD(1484), CHILD(1487), CHILD(1487), CHILD(1487),
+  CHILD(1487), CHILD(1490), CHILD(1490), CHILD(1490), CHILD(1490),
+  CHILD(1151), CHILD(1493), CHILD(1517), CHILD(1529), TILEPRO_OPC_NONE,
+  TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE,
+  TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE,
+  TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE,
+  TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE,
+  TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE,
+  TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE,
+  TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE,
+  TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE,
+  TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE,
+  TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE,
+  TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE,
+  TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE,
+  TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE,
+  TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE,
+  TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE,
+  TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE,
+  TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE,
+  TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE,
+  TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE,
+  TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE,
+  BITFIELD(53, 1) /* index 1463 */,
+  TILEPRO_OPC_RLI_SN, TILEPRO_OPC_NONE,
+  BITFIELD(53, 1) /* index 1466 */,
+  TILEPRO_OPC_SHLIB_SN, TILEPRO_OPC_NONE,
+  BITFIELD(53, 1) /* index 1469 */,
+  TILEPRO_OPC_SHLIH_SN, TILEPRO_OPC_NONE,
+  BITFIELD(53, 1) /* index 1472 */,
+  TILEPRO_OPC_SHLI_SN, TILEPRO_OPC_NONE,
+  BITFIELD(53, 1) /* index 1475 */,
+  TILEPRO_OPC_SHRIB_SN, TILEPRO_OPC_NONE,
+  BITFIELD(53, 1) /* index 1478 */,
+  TILEPRO_OPC_SHRIH_SN, TILEPRO_OPC_NONE,
+  BITFIELD(53, 1) /* index 1481 */,
+  TILEPRO_OPC_SHRI_SN, TILEPRO_OPC_NONE,
+  BITFIELD(53, 1) /* index 1484 */,
+  TILEPRO_OPC_SRAIB_SN, TILEPRO_OPC_NONE,
+  BITFIELD(53, 1) /* index 1487 */,
+  TILEPRO_OPC_SRAIH_SN, TILEPRO_OPC_NONE,
+  BITFIELD(53, 1) /* index 1490 */,
+  TILEPRO_OPC_SRAI_SN, TILEPRO_OPC_NONE,
+  BITFIELD(43, 3) /* index 1493 */,
+  CHILD(1251), CHILD(1254), CHILD(1502), CHILD(1505), CHILD(1508),
+  CHILD(1511), CHILD(1514), CHILD(1287),
+  BITFIELD(53, 1) /* index 1502 */,
+  TILEPRO_OPC_LB_SN, TILEPRO_OPC_NONE,
+  BITFIELD(53, 1) /* index 1505 */,
+  TILEPRO_OPC_LB_U_SN, TILEPRO_OPC_NONE,
+  BITFIELD(53, 1) /* index 1508 */,
+  TILEPRO_OPC_LH_SN, TILEPRO_OPC_NONE,
+  BITFIELD(53, 1) /* index 1511 */,
+  TILEPRO_OPC_LH_U_SN, TILEPRO_OPC_NONE,
+  BITFIELD(53, 1) /* index 1514 */,
+  TILEPRO_OPC_LW_SN, TILEPRO_OPC_NONE,
+  BITFIELD(43, 3) /* index 1517 */,
+  CHILD(1299), CHILD(1302), CHILD(1305), CHILD(1308), CHILD(1311),
+  CHILD(1314), CHILD(1526), CHILD(1320),
+  BITFIELD(53, 1) /* index 1526 */,
+  TILEPRO_OPC_TNS_SN, TILEPRO_OPC_NONE,
+  BITFIELD(43, 2) /* index 1529 */,
+  CHILD(1534), TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE,
+  BITFIELD(45, 1) /* index 1534 */,
+  CHILD(1537), TILEPRO_OPC_NONE,
+  BITFIELD(53, 1) /* index 1537 */,
+  TILEPRO_OPC_LW_NA_SN, TILEPRO_OPC_NONE,
+};
+
+static const unsigned short decode_Y0_fsm[168] =
+{
+  BITFIELD(27, 4) /* index 0 */,
+  TILEPRO_OPC_NONE, CHILD(17), CHILD(22), CHILD(27), CHILD(47), CHILD(52),
+  CHILD(57), CHILD(62), CHILD(67), TILEPRO_OPC_ADDI, CHILD(72), CHILD(102),
+  TILEPRO_OPC_SEQI, CHILD(117), TILEPRO_OPC_SLTI, TILEPRO_OPC_SLTI_U,
+  BITFIELD(18, 2) /* index 17 */,
+  TILEPRO_OPC_ADD, TILEPRO_OPC_S1A, TILEPRO_OPC_S2A, TILEPRO_OPC_SUB,
+  BITFIELD(18, 2) /* index 22 */,
+  TILEPRO_OPC_MNZ, TILEPRO_OPC_MVNZ, TILEPRO_OPC_MVZ, TILEPRO_OPC_MZ,
+  BITFIELD(18, 2) /* index 27 */,
+  TILEPRO_OPC_AND, TILEPRO_OPC_NOR, CHILD(32), TILEPRO_OPC_XOR,
+  BITFIELD(12, 2) /* index 32 */,
+  TILEPRO_OPC_OR, TILEPRO_OPC_OR, TILEPRO_OPC_OR, CHILD(37),
+  BITFIELD(14, 2) /* index 37 */,
+  TILEPRO_OPC_OR, TILEPRO_OPC_OR, TILEPRO_OPC_OR, CHILD(42),
+  BITFIELD(16, 2) /* index 42 */,
+  TILEPRO_OPC_OR, TILEPRO_OPC_OR, TILEPRO_OPC_OR, TILEPRO_OPC_MOVE,
+  BITFIELD(18, 2) /* index 47 */,
+  TILEPRO_OPC_RL, TILEPRO_OPC_SHL, TILEPRO_OPC_SHR, TILEPRO_OPC_SRA,
+  BITFIELD(18, 2) /* index 52 */,
+  TILEPRO_OPC_SLTE, TILEPRO_OPC_SLTE_U, TILEPRO_OPC_SLT, TILEPRO_OPC_SLT_U,
+  BITFIELD(18, 2) /* index 57 */,
+  TILEPRO_OPC_MULHLSA_UU, TILEPRO_OPC_S3A, TILEPRO_OPC_SEQ, TILEPRO_OPC_SNE,
+  BITFIELD(18, 2) /* index 62 */,
+  TILEPRO_OPC_MULHH_SS, TILEPRO_OPC_MULHH_UU, TILEPRO_OPC_MULLL_SS,
+  TILEPRO_OPC_MULLL_UU,
+  BITFIELD(18, 2) /* index 67 */,
+  TILEPRO_OPC_MULHHA_SS, TILEPRO_OPC_MULHHA_UU, TILEPRO_OPC_MULLLA_SS,
+  TILEPRO_OPC_MULLLA_UU,
+  BITFIELD(0, 2) /* index 72 */,
+  TILEPRO_OPC_ANDI, TILEPRO_OPC_ANDI, TILEPRO_OPC_ANDI, CHILD(77),
+  BITFIELD(2, 2) /* index 77 */,
+  TILEPRO_OPC_ANDI, TILEPRO_OPC_ANDI, TILEPRO_OPC_ANDI, CHILD(82),
+  BITFIELD(4, 2) /* index 82 */,
+  TILEPRO_OPC_ANDI, TILEPRO_OPC_ANDI, TILEPRO_OPC_ANDI, CHILD(87),
+  BITFIELD(6, 2) /* index 87 */,
+  TILEPRO_OPC_ANDI, TILEPRO_OPC_ANDI, TILEPRO_OPC_ANDI, CHILD(92),
+  BITFIELD(8, 2) /* index 92 */,
+  TILEPRO_OPC_ANDI, TILEPRO_OPC_ANDI, TILEPRO_OPC_ANDI, CHILD(97),
+  BITFIELD(10, 2) /* index 97 */,
+  TILEPRO_OPC_ANDI, TILEPRO_OPC_ANDI, TILEPRO_OPC_ANDI, TILEPRO_OPC_INFO,
+  BITFIELD(6, 2) /* index 102 */,
+  TILEPRO_OPC_ORI, TILEPRO_OPC_ORI, TILEPRO_OPC_ORI, CHILD(107),
+  BITFIELD(8, 2) /* index 107 */,
+  TILEPRO_OPC_ORI, TILEPRO_OPC_ORI, TILEPRO_OPC_ORI, CHILD(112),
+  BITFIELD(10, 2) /* index 112 */,
+  TILEPRO_OPC_ORI, TILEPRO_OPC_ORI, TILEPRO_OPC_ORI, TILEPRO_OPC_MOVEI,
+  BITFIELD(15, 5) /* index 117 */,
+  TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE,
+  TILEPRO_OPC_RLI, TILEPRO_OPC_RLI, TILEPRO_OPC_RLI, TILEPRO_OPC_RLI,
+  TILEPRO_OPC_SHLI, TILEPRO_OPC_SHLI, TILEPRO_OPC_SHLI, TILEPRO_OPC_SHLI,
+  TILEPRO_OPC_SHRI, TILEPRO_OPC_SHRI, TILEPRO_OPC_SHRI, TILEPRO_OPC_SHRI,
+  TILEPRO_OPC_SRAI, TILEPRO_OPC_SRAI, TILEPRO_OPC_SRAI, TILEPRO_OPC_SRAI,
+  CHILD(150), CHILD(159), TILEPRO_OPC_NONE, TILEPRO_OPC_NONE,
+  TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE,
+  TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE,
+  BITFIELD(12, 3) /* index 150 */,
+  TILEPRO_OPC_NONE, TILEPRO_OPC_BITX, TILEPRO_OPC_BYTEX, TILEPRO_OPC_CLZ,
+  TILEPRO_OPC_CTZ, TILEPRO_OPC_FNOP, TILEPRO_OPC_NOP, TILEPRO_OPC_PCNT,
+  BITFIELD(12, 3) /* index 159 */,
+  TILEPRO_OPC_TBLIDXB0, TILEPRO_OPC_TBLIDXB1, TILEPRO_OPC_TBLIDXB2,
+  TILEPRO_OPC_TBLIDXB3, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE,
+  TILEPRO_OPC_NONE,
+};
+
+static const unsigned short decode_Y1_fsm[140] =
+{
+  BITFIELD(59, 4) /* index 0 */,
+  TILEPRO_OPC_NONE, CHILD(17), CHILD(22), CHILD(27), CHILD(47), CHILD(52),
+  CHILD(57), TILEPRO_OPC_ADDI, CHILD(62), CHILD(92), TILEPRO_OPC_SEQI,
+  CHILD(107), TILEPRO_OPC_SLTI, TILEPRO_OPC_SLTI_U, TILEPRO_OPC_NONE,
+  TILEPRO_OPC_NONE,
+  BITFIELD(49, 2) /* index 17 */,
+  TILEPRO_OPC_ADD, TILEPRO_OPC_S1A, TILEPRO_OPC_S2A, TILEPRO_OPC_SUB,
+  BITFIELD(49, 2) /* index 22 */,
+  TILEPRO_OPC_NONE, TILEPRO_OPC_MNZ, TILEPRO_OPC_MZ, TILEPRO_OPC_NONE,
+  BITFIELD(49, 2) /* index 27 */,
+  TILEPRO_OPC_AND, TILEPRO_OPC_NOR, CHILD(32), TILEPRO_OPC_XOR,
+  BITFIELD(43, 2) /* index 32 */,
+  TILEPRO_OPC_OR, TILEPRO_OPC_OR, TILEPRO_OPC_OR, CHILD(37),
+  BITFIELD(45, 2) /* index 37 */,
+  TILEPRO_OPC_OR, TILEPRO_OPC_OR, TILEPRO_OPC_OR, CHILD(42),
+  BITFIELD(47, 2) /* index 42 */,
+  TILEPRO_OPC_OR, TILEPRO_OPC_OR, TILEPRO_OPC_OR, TILEPRO_OPC_MOVE,
+  BITFIELD(49, 2) /* index 47 */,
+  TILEPRO_OPC_RL, TILEPRO_OPC_SHL, TILEPRO_OPC_SHR, TILEPRO_OPC_SRA,
+  BITFIELD(49, 2) /* index 52 */,
+  TILEPRO_OPC_SLTE, TILEPRO_OPC_SLTE_U, TILEPRO_OPC_SLT, TILEPRO_OPC_SLT_U,
+  BITFIELD(49, 2) /* index 57 */,
+  TILEPRO_OPC_NONE, TILEPRO_OPC_S3A, TILEPRO_OPC_SEQ, TILEPRO_OPC_SNE,
+  BITFIELD(31, 2) /* index 62 */,
+  TILEPRO_OPC_ANDI, TILEPRO_OPC_ANDI, TILEPRO_OPC_ANDI, CHILD(67),
+  BITFIELD(33, 2) /* index 67 */,
+  TILEPRO_OPC_ANDI, TILEPRO_OPC_ANDI, TILEPRO_OPC_ANDI, CHILD(72),
+  BITFIELD(35, 2) /* index 72 */,
+  TILEPRO_OPC_ANDI, TILEPRO_OPC_ANDI, TILEPRO_OPC_ANDI, CHILD(77),
+  BITFIELD(37, 2) /* index 77 */,
+  TILEPRO_OPC_ANDI, TILEPRO_OPC_ANDI, TILEPRO_OPC_ANDI, CHILD(82),
+  BITFIELD(39, 2) /* index 82 */,
+  TILEPRO_OPC_ANDI, TILEPRO_OPC_ANDI, TILEPRO_OPC_ANDI, CHILD(87),
+  BITFIELD(41, 2) /* index 87 */,
+  TILEPRO_OPC_ANDI, TILEPRO_OPC_ANDI, TILEPRO_OPC_ANDI, TILEPRO_OPC_INFO,
+  BITFIELD(37, 2) /* index 92 */,
+  TILEPRO_OPC_ORI, TILEPRO_OPC_ORI, TILEPRO_OPC_ORI, CHILD(97),
+  BITFIELD(39, 2) /* index 97 */,
+  TILEPRO_OPC_ORI, TILEPRO_OPC_ORI, TILEPRO_OPC_ORI, CHILD(102),
+  BITFIELD(41, 2) /* index 102 */,
+  TILEPRO_OPC_ORI, TILEPRO_OPC_ORI, TILEPRO_OPC_ORI, TILEPRO_OPC_MOVEI,
+  BITFIELD(48, 3) /* index 107 */,
+  TILEPRO_OPC_NONE, TILEPRO_OPC_RLI, TILEPRO_OPC_SHLI, TILEPRO_OPC_SHRI,
+  TILEPRO_OPC_SRAI, CHILD(116), TILEPRO_OPC_NONE, TILEPRO_OPC_NONE,
+  BITFIELD(43, 3) /* index 116 */,
+  TILEPRO_OPC_NONE, CHILD(125), CHILD(130), CHILD(135), TILEPRO_OPC_NONE,
+  TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE,
+  BITFIELD(46, 2) /* index 125 */,
+  TILEPRO_OPC_FNOP, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE,
+  BITFIELD(46, 2) /* index 130 */,
+  TILEPRO_OPC_ILL, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE,
+  BITFIELD(46, 2) /* index 135 */,
+  TILEPRO_OPC_NOP, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE,
+};
+
+static const unsigned short decode_Y2_fsm[24] =
+{
+  BITFIELD(56, 3) /* index 0 */,
+  CHILD(9), TILEPRO_OPC_LB_U, TILEPRO_OPC_LH, TILEPRO_OPC_LH_U,
+  TILEPRO_OPC_LW, TILEPRO_OPC_SB, TILEPRO_OPC_SH, TILEPRO_OPC_SW,
+  BITFIELD(20, 2) /* index 9 */,
+  TILEPRO_OPC_LB, TILEPRO_OPC_LB, TILEPRO_OPC_LB, CHILD(14),
+  BITFIELD(22, 2) /* index 14 */,
+  TILEPRO_OPC_LB, TILEPRO_OPC_LB, TILEPRO_OPC_LB, CHILD(19),
+  BITFIELD(24, 2) /* index 19 */,
+  TILEPRO_OPC_LB, TILEPRO_OPC_LB, TILEPRO_OPC_LB, TILEPRO_OPC_PREFETCH,
+};
+
+#undef BITFIELD
+#undef CHILD
+const unsigned short * const
+tilepro_bundle_decoder_fsms[TILEPRO_NUM_PIPELINE_ENCODINGS] =
+{
+  decode_X0_fsm,
+  decode_X1_fsm,
+  decode_Y0_fsm,
+  decode_Y1_fsm,
+  decode_Y2_fsm
+};
+const struct tilepro_operand tilepro_operands[43] =
+{
+  {
+    TILEPRO_OP_TYPE_IMMEDIATE, BFD_RELOC(TILEPRO_IMM8_X0),
+    8, 1, 0, 0, 0, 0,
+    create_Imm8_X0, get_Imm8_X0
+  },
+  {
+    TILEPRO_OP_TYPE_IMMEDIATE, BFD_RELOC(TILEPRO_IMM8_X1),
+    8, 1, 0, 0, 0, 0,
+    create_Imm8_X1, get_Imm8_X1
+  },
+  {
+    TILEPRO_OP_TYPE_IMMEDIATE, BFD_RELOC(TILEPRO_IMM8_Y0),
+    8, 1, 0, 0, 0, 0,
+    create_Imm8_Y0, get_Imm8_Y0
+  },
+  {
+    TILEPRO_OP_TYPE_IMMEDIATE, BFD_RELOC(TILEPRO_IMM8_Y1),
+    8, 1, 0, 0, 0, 0,
+    create_Imm8_Y1, get_Imm8_Y1
+  },
+  {
+    TILEPRO_OP_TYPE_IMMEDIATE, BFD_RELOC(TILEPRO_IMM16_X0),
+    16, 1, 0, 0, 0, 0,
+    create_Imm16_X0, get_Imm16_X0
+  },
+  {
+    TILEPRO_OP_TYPE_IMMEDIATE, BFD_RELOC(TILEPRO_IMM16_X1),
+    16, 1, 0, 0, 0, 0,
+    create_Imm16_X1, get_Imm16_X1
+  },
+  {
+    TILEPRO_OP_TYPE_ADDRESS, BFD_RELOC(TILEPRO_JOFFLONG_X1),
+    29, 1, 0, 0, 1, TILEPRO_LOG2_BUNDLE_ALIGNMENT_IN_BYTES,
+    create_JOffLong_X1, get_JOffLong_X1
+  },
+  {
+    TILEPRO_OP_TYPE_REGISTER, BFD_RELOC(NONE),
+    6, 0, 0, 1, 0, 0,
+    create_Dest_X0, get_Dest_X0
+  },
+  {
+    TILEPRO_OP_TYPE_REGISTER, BFD_RELOC(NONE),
+    6, 0, 1, 0, 0, 0,
+    create_SrcA_X0, get_SrcA_X0
+  },
+  {
+    TILEPRO_OP_TYPE_REGISTER, BFD_RELOC(NONE),
+    6, 0, 0, 1, 0, 0,
+    create_Dest_X1, get_Dest_X1
+  },
+  {
+    TILEPRO_OP_TYPE_REGISTER, BFD_RELOC(NONE),
+    6, 0, 1, 0, 0, 0,
+    create_SrcA_X1, get_SrcA_X1
+  },
+  {
+    TILEPRO_OP_TYPE_REGISTER, BFD_RELOC(NONE),
+    6, 0, 0, 1, 0, 0,
+    create_Dest_Y0, get_Dest_Y0
+  },
+  {
+    TILEPRO_OP_TYPE_REGISTER, BFD_RELOC(NONE),
+    6, 0, 1, 0, 0, 0,
+    create_SrcA_Y0, get_SrcA_Y0
+  },
+  {
+    TILEPRO_OP_TYPE_REGISTER, BFD_RELOC(NONE),
+    6, 0, 0, 1, 0, 0,
+    create_Dest_Y1, get_Dest_Y1
+  },
+  {
+    TILEPRO_OP_TYPE_REGISTER, BFD_RELOC(NONE),
+    6, 0, 1, 0, 0, 0,
+    create_SrcA_Y1, get_SrcA_Y1
+  },
+  {
+    TILEPRO_OP_TYPE_REGISTER, BFD_RELOC(NONE),
+    6, 0, 1, 0, 0, 0,
+    create_SrcA_Y2, get_SrcA_Y2
+  },
+  {
+    TILEPRO_OP_TYPE_REGISTER, BFD_RELOC(NONE),
+    6, 0, 1, 0, 0, 0,
+    create_SrcB_X0, get_SrcB_X0
+  },
+  {
+    TILEPRO_OP_TYPE_REGISTER, BFD_RELOC(NONE),
+    6, 0, 1, 0, 0, 0,
+    create_SrcB_X1, get_SrcB_X1
+  },
+  {
+    TILEPRO_OP_TYPE_REGISTER, BFD_RELOC(NONE),
+    6, 0, 1, 0, 0, 0,
+    create_SrcB_Y0, get_SrcB_Y0
+  },
+  {
+    TILEPRO_OP_TYPE_REGISTER, BFD_RELOC(NONE),
+    6, 0, 1, 0, 0, 0,
+    create_SrcB_Y1, get_SrcB_Y1
+  },
+  {
+    TILEPRO_OP_TYPE_ADDRESS, BFD_RELOC(TILEPRO_BROFF_X1),
+    17, 1, 0, 0, 1, TILEPRO_LOG2_BUNDLE_ALIGNMENT_IN_BYTES,
+    create_BrOff_X1, get_BrOff_X1
+  },
+  {
+    TILEPRO_OP_TYPE_REGISTER, BFD_RELOC(NONE),
+    6, 0, 1, 1, 0, 0,
+    create_Dest_X0, get_Dest_X0
+  },
+  {
+    TILEPRO_OP_TYPE_ADDRESS, BFD_RELOC(NONE),
+    28, 1, 0, 0, 1, TILEPRO_LOG2_BUNDLE_ALIGNMENT_IN_BYTES,
+    create_JOff_X1, get_JOff_X1
+  },
+  {
+    TILEPRO_OP_TYPE_REGISTER, BFD_RELOC(NONE),
+    6, 0, 0, 1, 0, 0,
+    create_SrcBDest_Y2, get_SrcBDest_Y2
+  },
+  {
+    TILEPRO_OP_TYPE_REGISTER, BFD_RELOC(NONE),
+    6, 0, 1, 1, 0, 0,
+    create_SrcA_X1, get_SrcA_X1
+  },
+  {
+    TILEPRO_OP_TYPE_SPR, BFD_RELOC(TILEPRO_MF_IMM15_X1),
+    15, 0, 0, 0, 0, 0,
+    create_MF_Imm15_X1, get_MF_Imm15_X1
+  },
+  {
+    TILEPRO_OP_TYPE_IMMEDIATE, BFD_RELOC(TILEPRO_MMSTART_X0),
+    5, 0, 0, 0, 0, 0,
+    create_MMStart_X0, get_MMStart_X0
+  },
+  {
+    TILEPRO_OP_TYPE_IMMEDIATE, BFD_RELOC(TILEPRO_MMEND_X0),
+    5, 0, 0, 0, 0, 0,
+    create_MMEnd_X0, get_MMEnd_X0
+  },
+  {
+    TILEPRO_OP_TYPE_IMMEDIATE, BFD_RELOC(TILEPRO_MMSTART_X1),
+    5, 0, 0, 0, 0, 0,
+    create_MMStart_X1, get_MMStart_X1
+  },
+  {
+    TILEPRO_OP_TYPE_IMMEDIATE, BFD_RELOC(TILEPRO_MMEND_X1),
+    5, 0, 0, 0, 0, 0,
+    create_MMEnd_X1, get_MMEnd_X1
+  },
+  {
+    TILEPRO_OP_TYPE_SPR, BFD_RELOC(TILEPRO_MT_IMM15_X1),
+    15, 0, 0, 0, 0, 0,
+    create_MT_Imm15_X1, get_MT_Imm15_X1
+  },
+  {
+    TILEPRO_OP_TYPE_REGISTER, BFD_RELOC(NONE),
+    6, 0, 1, 1, 0, 0,
+    create_Dest_Y0, get_Dest_Y0
+  },
+  {
+    TILEPRO_OP_TYPE_IMMEDIATE, BFD_RELOC(TILEPRO_SHAMT_X0),
+    5, 0, 0, 0, 0, 0,
+    create_ShAmt_X0, get_ShAmt_X0
+  },
+  {
+    TILEPRO_OP_TYPE_IMMEDIATE, BFD_RELOC(TILEPRO_SHAMT_X1),
+    5, 0, 0, 0, 0, 0,
+    create_ShAmt_X1, get_ShAmt_X1
+  },
+  {
+    TILEPRO_OP_TYPE_IMMEDIATE, BFD_RELOC(TILEPRO_SHAMT_Y0),
+    5, 0, 0, 0, 0, 0,
+    create_ShAmt_Y0, get_ShAmt_Y0
+  },
+  {
+    TILEPRO_OP_TYPE_IMMEDIATE, BFD_RELOC(TILEPRO_SHAMT_Y1),
+    5, 0, 0, 0, 0, 0,
+    create_ShAmt_Y1, get_ShAmt_Y1
+  },
+  {
+    TILEPRO_OP_TYPE_REGISTER, BFD_RELOC(NONE),
+    6, 0, 1, 0, 0, 0,
+    create_SrcBDest_Y2, get_SrcBDest_Y2
+  },
+  {
+    TILEPRO_OP_TYPE_IMMEDIATE, BFD_RELOC(TILEPRO_DEST_IMM8_X1),
+    8, 1, 0, 0, 0, 0,
+    create_Dest_Imm8_X1, get_Dest_Imm8_X1
+  },
+  {
+    TILEPRO_OP_TYPE_ADDRESS, BFD_RELOC(NONE),
+    10, 1, 0, 0, 1, TILEPRO_LOG2_SN_INSTRUCTION_SIZE_IN_BYTES,
+    create_BrOff_SN, get_BrOff_SN
+  },
+  {
+    TILEPRO_OP_TYPE_IMMEDIATE, BFD_RELOC(NONE),
+    8, 0, 0, 0, 0, 0,
+    create_Imm8_SN, get_Imm8_SN
+  },
+  {
+    TILEPRO_OP_TYPE_IMMEDIATE, BFD_RELOC(NONE),
+    8, 1, 0, 0, 0, 0,
+    create_Imm8_SN, get_Imm8_SN
+  },
+  {
+    TILEPRO_OP_TYPE_REGISTER, BFD_RELOC(NONE),
+    2, 0, 0, 1, 0, 0,
+    create_Dest_SN, get_Dest_SN
+  },
+  {
+    TILEPRO_OP_TYPE_REGISTER, BFD_RELOC(NONE),
+    2, 0, 1, 0, 0, 0,
+    create_Src_SN, get_Src_SN
+  }
+};
+
+
+
+
+/* Given a set of bundle bits and a specific pipe, returns which
+ * instruction the bundle contains in that pipe.
+ */
+const struct tilepro_opcode *
+find_opcode(tilepro_bundle_bits bits, tilepro_pipeline pipe)
+{
+  const unsigned short *table = tilepro_bundle_decoder_fsms[pipe];
+  int index = 0;
+
+  while (1)
+  {
+    unsigned short bitspec = table[index];
+    unsigned int bitfield =
+      ((unsigned int)(bits >> (bitspec & 63))) & (bitspec >> 6);
+
+    unsigned short next = table[index + 1 + bitfield];
+    if (next <= TILEPRO_OPC_NONE)
+      return &tilepro_opcodes[next];
+
+    index = next - TILEPRO_OPC_NONE;
+  }
+}
+
+
+int
+parse_insn_tilepro(tilepro_bundle_bits bits,
+                   unsigned int pc,
+                   struct tilepro_decoded_instruction
+                   decoded[TILEPRO_MAX_INSTRUCTIONS_PER_BUNDLE])
+{
+  int num_instructions = 0;
+  int pipe;
+
+  int min_pipe, max_pipe;
+  if ((bits & TILEPRO_BUNDLE_Y_ENCODING_MASK) == 0)
+  {
+    min_pipe = TILEPRO_PIPELINE_X0;
+    max_pipe = TILEPRO_PIPELINE_X1;
+  }
+  else
+  {
+    min_pipe = TILEPRO_PIPELINE_Y0;
+    max_pipe = TILEPRO_PIPELINE_Y2;
+  }
+
+  /* For each pipe, find an instruction that fits. */
+  for (pipe = min_pipe; pipe <= max_pipe; pipe++)
+  {
+    const struct tilepro_opcode *opc;
+    struct tilepro_decoded_instruction *d;
+    int i;
+
+    d = &decoded[num_instructions++];
+    opc = find_opcode (bits, (tilepro_pipeline)pipe);
+    d->opcode = opc;
+
+    /* Decode each operand, sign extending, etc. as appropriate. */
+    for (i = 0; i < opc->num_operands; i++)
+    {
+      const struct tilepro_operand *op =
+        &tilepro_operands[opc->operands[pipe][i]];
+      int opval = op->extract (bits);
+      if (op->is_signed)
+      {
+        /* Sign-extend the operand. */
+        int shift = (int)((sizeof(int) * 8) - op->num_bits);
+        opval = (opval << shift) >> shift;
+      }
+
+      /* Adjust PC-relative scaled branch offsets. */
+      if (op->type == TILEPRO_OP_TYPE_ADDRESS)
+      {
+        opval *= TILEPRO_BUNDLE_SIZE_IN_BYTES;
+        opval += (int)pc;
+      }
+
+      /* Record the final value. */
+      d->operands[i] = op;
+      d->operand_values[i] = opval;
+    }
+  }
+
+  return num_instructions;
+}
diff --git a/kernel/arch/tile/kernel/tile-desc_64.c b/kernel/arch/tile/kernel/tile-desc_64.c
new file mode 100644
index 000000000..65b5f8aca
--- /dev/null
+++ b/kernel/arch/tile/kernel/tile-desc_64.c
@@ -0,0 +1,2218 @@
+/* TILE-Gx opcode information.
+ *
+ * Copyright 2011 Tilera Corporation. All Rights Reserved.
+ *
+ *   This program is free software; you can redistribute it and/or
+ *   modify it under the terms of the GNU General Public License
+ *   as published by the Free Software Foundation, version 2.
+ *
+ *   This program is distributed in the hope that it will be useful, but
+ *   WITHOUT ANY WARRANTY; without even the implied warranty of
+ *   MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
+ *   NON INFRINGEMENT.  See the GNU General Public License for
+ *   more details.
+ *
+ *
+ *
+ *
+ *
+ */
+
+/* This define is BFD_RELOC_##x for real bfd, or -1 for everyone else. */
+#define BFD_RELOC(x) -1
+
+/* Special registers. */
+#define TREG_LR 55
+#define TREG_SN 56
+#define TREG_ZERO 63
+
+#include <linux/stddef.h>
+#include <asm/tile-desc.h>
+
+const struct tilegx_opcode tilegx_opcodes[334] =
+{
+ { "bpt", TILEGX_OPC_BPT, 0x2, 0, TREG_ZERO, 0,
+    { { 0, }, {  }, { 0, }, { 0, }, { 0, } },
+  },
+  { "info", TILEGX_OPC_INFO, 0xf, 1, TREG_ZERO, 1,
+    { { 0 }, { 1 }, { 2 }, { 3 }, { 0, } },
+  },
+  { "infol", TILEGX_OPC_INFOL, 0x3, 1, TREG_ZERO, 1,
+    { { 4 }, { 5 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "move", TILEGX_OPC_MOVE, 0xf, 2, TREG_ZERO, 1,
+    { { 6, 7 }, { 8, 9 }, { 10, 11 }, { 12, 13 }, { 0, } },
+  },
+  { "movei", TILEGX_OPC_MOVEI, 0xf, 2, TREG_ZERO, 1,
+    { { 6, 0 }, { 8, 1 }, { 10, 2 }, { 12, 3 }, { 0, } },
+  },
+  { "moveli", TILEGX_OPC_MOVELI, 0x3, 2, TREG_ZERO, 1,
+    { { 6, 4 }, { 8, 5 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "prefetch", TILEGX_OPC_PREFETCH, 0x12, 1, TREG_ZERO, 1,
+    { { 0, }, { 9 }, { 0, }, { 0, }, { 14 } },
+  },
+  { "prefetch_add_l1", TILEGX_OPC_PREFETCH_ADD_L1, 0x2, 2, TREG_ZERO, 1,
+    { { 0, }, { 15, 1 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "prefetch_add_l1_fault", TILEGX_OPC_PREFETCH_ADD_L1_FAULT, 0x2, 2, TREG_ZERO, 1,
+    { { 0, }, { 15, 1 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "prefetch_add_l2", TILEGX_OPC_PREFETCH_ADD_L2, 0x2, 2, TREG_ZERO, 1,
+    { { 0, }, { 15, 1 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "prefetch_add_l2_fault", TILEGX_OPC_PREFETCH_ADD_L2_FAULT, 0x2, 2, TREG_ZERO, 1,
+    { { 0, }, { 15, 1 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "prefetch_add_l3", TILEGX_OPC_PREFETCH_ADD_L3, 0x2, 2, TREG_ZERO, 1,
+    { { 0, }, { 15, 1 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "prefetch_add_l3_fault", TILEGX_OPC_PREFETCH_ADD_L3_FAULT, 0x2, 2, TREG_ZERO, 1,
+    { { 0, }, { 15, 1 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "prefetch_l1", TILEGX_OPC_PREFETCH_L1, 0x12, 1, TREG_ZERO, 1,
+    { { 0, }, { 9 }, { 0, }, { 0, }, { 14 } },
+  },
+  { "prefetch_l1_fault", TILEGX_OPC_PREFETCH_L1_FAULT, 0x12, 1, TREG_ZERO, 1,
+    { { 0, }, { 9 }, { 0, }, { 0, }, { 14 } },
+  },
+  { "prefetch_l2", TILEGX_OPC_PREFETCH_L2, 0x12, 1, TREG_ZERO, 1,
+    { { 0, }, { 9 }, { 0, }, { 0, }, { 14 } },
+  },
+  { "prefetch_l2_fault", TILEGX_OPC_PREFETCH_L2_FAULT, 0x12, 1, TREG_ZERO, 1,
+    { { 0, }, { 9 }, { 0, }, { 0, }, { 14 } },
+  },
+  { "prefetch_l3", TILEGX_OPC_PREFETCH_L3, 0x12, 1, TREG_ZERO, 1,
+    { { 0, }, { 9 }, { 0, }, { 0, }, { 14 } },
+  },
+  { "prefetch_l3_fault", TILEGX_OPC_PREFETCH_L3_FAULT, 0x12, 1, TREG_ZERO, 1,
+    { { 0, }, { 9 }, { 0, }, { 0, }, { 14 } },
+  },
+  { "raise", TILEGX_OPC_RAISE, 0x2, 0, TREG_ZERO, 1,
+    { { 0, }, {  }, { 0, }, { 0, }, { 0, } },
+  },
+  { "add", TILEGX_OPC_ADD, 0xf, 3, TREG_ZERO, 1,
+    { { 6, 7, 16 }, { 8, 9, 17 }, { 10, 11, 18 }, { 12, 13, 19 }, { 0, } },
+  },
+  { "addi", TILEGX_OPC_ADDI, 0xf, 3, TREG_ZERO, 1,
+    { { 6, 7, 0 }, { 8, 9, 1 }, { 10, 11, 2 }, { 12, 13, 3 }, { 0, } },
+  },
+  { "addli", TILEGX_OPC_ADDLI, 0x3, 3, TREG_ZERO, 1,
+    { { 6, 7, 4 }, { 8, 9, 5 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "addx", TILEGX_OPC_ADDX, 0xf, 3, TREG_ZERO, 1,
+    { { 6, 7, 16 }, { 8, 9, 17 }, { 10, 11, 18 }, { 12, 13, 19 }, { 0, } },
+  },
+  { "addxi", TILEGX_OPC_ADDXI, 0xf, 3, TREG_ZERO, 1,
+    { { 6, 7, 0 }, { 8, 9, 1 }, { 10, 11, 2 }, { 12, 13, 3 }, { 0, } },
+  },
+  { "addxli", TILEGX_OPC_ADDXLI, 0x3, 3, TREG_ZERO, 1,
+    { { 6, 7, 4 }, { 8, 9, 5 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "addxsc", TILEGX_OPC_ADDXSC, 0x3, 3, TREG_ZERO, 1,
+    { { 6, 7, 16 }, { 8, 9, 17 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "and", TILEGX_OPC_AND, 0xf, 3, TREG_ZERO, 1,
+    { { 6, 7, 16 }, { 8, 9, 17 }, { 10, 11, 18 }, { 12, 13, 19 }, { 0, } },
+  },
+  { "andi", TILEGX_OPC_ANDI, 0xf, 3, TREG_ZERO, 1,
+    { { 6, 7, 0 }, { 8, 9, 1 }, { 10, 11, 2 }, { 12, 13, 3 }, { 0, } },
+  },
+  { "beqz", TILEGX_OPC_BEQZ, 0x2, 2, TREG_ZERO, 1,
+    { { 0, }, { 9, 20 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "beqzt", TILEGX_OPC_BEQZT, 0x2, 2, TREG_ZERO, 1,
+    { { 0, }, { 9, 20 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "bfexts", TILEGX_OPC_BFEXTS, 0x1, 4, TREG_ZERO, 1,
+    { { 6, 7, 21, 22 }, { 0, }, { 0, }, { 0, }, { 0, } },
+  },
+  { "bfextu", TILEGX_OPC_BFEXTU, 0x1, 4, TREG_ZERO, 1,
+    { { 6, 7, 21, 22 }, { 0, }, { 0, }, { 0, }, { 0, } },
+  },
+  { "bfins", TILEGX_OPC_BFINS, 0x1, 4, TREG_ZERO, 1,
+    { { 23, 7, 21, 22 }, { 0, }, { 0, }, { 0, }, { 0, } },
+  },
+  { "bgez", TILEGX_OPC_BGEZ, 0x2, 2, TREG_ZERO, 1,
+    { { 0, }, { 9, 20 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "bgezt", TILEGX_OPC_BGEZT, 0x2, 2, TREG_ZERO, 1,
+    { { 0, }, { 9, 20 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "bgtz", TILEGX_OPC_BGTZ, 0x2, 2, TREG_ZERO, 1,
+    { { 0, }, { 9, 20 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "bgtzt", TILEGX_OPC_BGTZT, 0x2, 2, TREG_ZERO, 1,
+    { { 0, }, { 9, 20 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "blbc", TILEGX_OPC_BLBC, 0x2, 2, TREG_ZERO, 1,
+    { { 0, }, { 9, 20 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "blbct", TILEGX_OPC_BLBCT, 0x2, 2, TREG_ZERO, 1,
+    { { 0, }, { 9, 20 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "blbs", TILEGX_OPC_BLBS, 0x2, 2, TREG_ZERO, 1,
+    { { 0, }, { 9, 20 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "blbst", TILEGX_OPC_BLBST, 0x2, 2, TREG_ZERO, 1,
+    { { 0, }, { 9, 20 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "blez", TILEGX_OPC_BLEZ, 0x2, 2, TREG_ZERO, 1,
+    { { 0, }, { 9, 20 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "blezt", TILEGX_OPC_BLEZT, 0x2, 2, TREG_ZERO, 1,
+    { { 0, }, { 9, 20 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "bltz", TILEGX_OPC_BLTZ, 0x2, 2, TREG_ZERO, 1,
+    { { 0, }, { 9, 20 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "bltzt", TILEGX_OPC_BLTZT, 0x2, 2, TREG_ZERO, 1,
+    { { 0, }, { 9, 20 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "bnez", TILEGX_OPC_BNEZ, 0x2, 2, TREG_ZERO, 1,
+    { { 0, }, { 9, 20 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "bnezt", TILEGX_OPC_BNEZT, 0x2, 2, TREG_ZERO, 1,
+    { { 0, }, { 9, 20 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "clz", TILEGX_OPC_CLZ, 0x5, 2, TREG_ZERO, 1,
+    { { 6, 7 }, { 0, }, { 10, 11 }, { 0, }, { 0, } },
+  },
+  { "cmoveqz", TILEGX_OPC_CMOVEQZ, 0x5, 3, TREG_ZERO, 1,
+    { { 23, 7, 16 }, { 0, }, { 24, 11, 18 }, { 0, }, { 0, } },
+  },
+  { "cmovnez", TILEGX_OPC_CMOVNEZ, 0x5, 3, TREG_ZERO, 1,
+    { { 23, 7, 16 }, { 0, }, { 24, 11, 18 }, { 0, }, { 0, } },
+  },
+  { "cmpeq", TILEGX_OPC_CMPEQ, 0xf, 3, TREG_ZERO, 1,
+    { { 6, 7, 16 }, { 8, 9, 17 }, { 10, 11, 18 }, { 12, 13, 19 }, { 0, } },
+  },
+  { "cmpeqi", TILEGX_OPC_CMPEQI, 0xf, 3, TREG_ZERO, 1,
+    { { 6, 7, 0 }, { 8, 9, 1 }, { 10, 11, 2 }, { 12, 13, 3 }, { 0, } },
+  },
+  { "cmpexch", TILEGX_OPC_CMPEXCH, 0x2, 3, TREG_ZERO, 1,
+    { { 0, }, { 8, 9, 17 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "cmpexch4", TILEGX_OPC_CMPEXCH4, 0x2, 3, TREG_ZERO, 1,
+    { { 0, }, { 8, 9, 17 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "cmples", TILEGX_OPC_CMPLES, 0xf, 3, TREG_ZERO, 1,
+    { { 6, 7, 16 }, { 8, 9, 17 }, { 10, 11, 18 }, { 12, 13, 19 }, { 0, } },
+  },
+  { "cmpleu", TILEGX_OPC_CMPLEU, 0xf, 3, TREG_ZERO, 1,
+    { { 6, 7, 16 }, { 8, 9, 17 }, { 10, 11, 18 }, { 12, 13, 19 }, { 0, } },
+  },
+  { "cmplts", TILEGX_OPC_CMPLTS, 0xf, 3, TREG_ZERO, 1,
+    { { 6, 7, 16 }, { 8, 9, 17 }, { 10, 11, 18 }, { 12, 13, 19 }, { 0, } },
+  },
+  { "cmpltsi", TILEGX_OPC_CMPLTSI, 0xf, 3, TREG_ZERO, 1,
+    { { 6, 7, 0 }, { 8, 9, 1 }, { 10, 11, 2 }, { 12, 13, 3 }, { 0, } },
+  },
+  { "cmpltu", TILEGX_OPC_CMPLTU, 0xf, 3, TREG_ZERO, 1,
+    { { 6, 7, 16 }, { 8, 9, 17 }, { 10, 11, 18 }, { 12, 13, 19 }, { 0, } },
+  },
+  { "cmpltui", TILEGX_OPC_CMPLTUI, 0x3, 3, TREG_ZERO, 1,
+    { { 6, 7, 0 }, { 8, 9, 1 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "cmpne", TILEGX_OPC_CMPNE, 0xf, 3, TREG_ZERO, 1,
+    { { 6, 7, 16 }, { 8, 9, 17 }, { 10, 11, 18 }, { 12, 13, 19 }, { 0, } },
+  },
+  { "cmul", TILEGX_OPC_CMUL, 0x1, 3, TREG_ZERO, 1,
+    { { 6, 7, 16 }, { 0, }, { 0, }, { 0, }, { 0, } },
+  },
+  { "cmula", TILEGX_OPC_CMULA, 0x1, 3, TREG_ZERO, 1,
+    { { 23, 7, 16 }, { 0, }, { 0, }, { 0, }, { 0, } },
+  },
+  { "cmulaf", TILEGX_OPC_CMULAF, 0x1, 3, TREG_ZERO, 1,
+    { { 23, 7, 16 }, { 0, }, { 0, }, { 0, }, { 0, } },
+  },
+  { "cmulf", TILEGX_OPC_CMULF, 0x1, 3, TREG_ZERO, 1,
+    { { 6, 7, 16 }, { 0, }, { 0, }, { 0, }, { 0, } },
+  },
+  { "cmulfr", TILEGX_OPC_CMULFR, 0x1, 3, TREG_ZERO, 1,
+    { { 6, 7, 16 }, { 0, }, { 0, }, { 0, }, { 0, } },
+  },
+  { "cmulh", TILEGX_OPC_CMULH, 0x1, 3, TREG_ZERO, 1,
+    { { 6, 7, 16 }, { 0, }, { 0, }, { 0, }, { 0, } },
+  },
+  { "cmulhr", TILEGX_OPC_CMULHR, 0x1, 3, TREG_ZERO, 1,
+    { { 6, 7, 16 }, { 0, }, { 0, }, { 0, }, { 0, } },
+  },
+  { "crc32_32", TILEGX_OPC_CRC32_32, 0x1, 3, TREG_ZERO, 1,
+    { { 6, 7, 16 }, { 0, }, { 0, }, { 0, }, { 0, } },
+  },
+  { "crc32_8", TILEGX_OPC_CRC32_8, 0x1, 3, TREG_ZERO, 1,
+    { { 6, 7, 16 }, { 0, }, { 0, }, { 0, }, { 0, } },
+  },
+  { "ctz", TILEGX_OPC_CTZ, 0x5, 2, TREG_ZERO, 1,
+    { { 6, 7 }, { 0, }, { 10, 11 }, { 0, }, { 0, } },
+  },
+  { "dblalign", TILEGX_OPC_DBLALIGN, 0x1, 3, TREG_ZERO, 1,
+    { { 23, 7, 16 }, { 0, }, { 0, }, { 0, }, { 0, } },
+  },
+  { "dblalign2", TILEGX_OPC_DBLALIGN2, 0x3, 3, TREG_ZERO, 1,
+    { { 6, 7, 16 }, { 8, 9, 17 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "dblalign4", TILEGX_OPC_DBLALIGN4, 0x3, 3, TREG_ZERO, 1,
+    { { 6, 7, 16 }, { 8, 9, 17 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "dblalign6", TILEGX_OPC_DBLALIGN6, 0x3, 3, TREG_ZERO, 1,
+    { { 6, 7, 16 }, { 8, 9, 17 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "drain", TILEGX_OPC_DRAIN, 0x2, 0, TREG_ZERO, 0,
+    { { 0, }, {  }, { 0, }, { 0, }, { 0, } },
+  },
+  { "dtlbpr", TILEGX_OPC_DTLBPR, 0x2, 1, TREG_ZERO, 1,
+    { { 0, }, { 9 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "exch", TILEGX_OPC_EXCH, 0x2, 3, TREG_ZERO, 1,
+    { { 0, }, { 8, 9, 17 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "exch4", TILEGX_OPC_EXCH4, 0x2, 3, TREG_ZERO, 1,
+    { { 0, }, { 8, 9, 17 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "fdouble_add_flags", TILEGX_OPC_FDOUBLE_ADD_FLAGS, 0x1, 3, TREG_ZERO, 1,
+    { { 6, 7, 16 }, { 0, }, { 0, }, { 0, }, { 0, } },
+  },
+  { "fdouble_addsub", TILEGX_OPC_FDOUBLE_ADDSUB, 0x1, 3, TREG_ZERO, 1,
+    { { 23, 7, 16 }, { 0, }, { 0, }, { 0, }, { 0, } },
+  },
+  { "fdouble_mul_flags", TILEGX_OPC_FDOUBLE_MUL_FLAGS, 0x1, 3, TREG_ZERO, 1,
+    { { 6, 7, 16 }, { 0, }, { 0, }, { 0, }, { 0, } },
+  },
+  { "fdouble_pack1", TILEGX_OPC_FDOUBLE_PACK1, 0x1, 3, TREG_ZERO, 1,
+    { { 6, 7, 16 }, { 0, }, { 0, }, { 0, }, { 0, } },
+  },
+  { "fdouble_pack2", TILEGX_OPC_FDOUBLE_PACK2, 0x1, 3, TREG_ZERO, 1,
+    { { 23, 7, 16 }, { 0, }, { 0, }, { 0, }, { 0, } },
+  },
+  { "fdouble_sub_flags", TILEGX_OPC_FDOUBLE_SUB_FLAGS, 0x1, 3, TREG_ZERO, 1,
+    { { 6, 7, 16 }, { 0, }, { 0, }, { 0, }, { 0, } },
+  },
+  { "fdouble_unpack_max", TILEGX_OPC_FDOUBLE_UNPACK_MAX, 0x1, 3, TREG_ZERO, 1,
+    { { 6, 7, 16 }, { 0, }, { 0, }, { 0, }, { 0, } },
+  },
+  { "fdouble_unpack_min", TILEGX_OPC_FDOUBLE_UNPACK_MIN, 0x1, 3, TREG_ZERO, 1,
+    { { 6, 7, 16 }, { 0, }, { 0, }, { 0, }, { 0, } },
+  },
+  { "fetchadd", TILEGX_OPC_FETCHADD, 0x2, 3, TREG_ZERO, 1,
+    { { 0, }, { 8, 9, 17 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "fetchadd4", TILEGX_OPC_FETCHADD4, 0x2, 3, TREG_ZERO, 1,
+    { { 0, }, { 8, 9, 17 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "fetchaddgez", TILEGX_OPC_FETCHADDGEZ, 0x2, 3, TREG_ZERO, 1,
+    { { 0, }, { 8, 9, 17 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "fetchaddgez4", TILEGX_OPC_FETCHADDGEZ4, 0x2, 3, TREG_ZERO, 1,
+    { { 0, }, { 8, 9, 17 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "fetchand", TILEGX_OPC_FETCHAND, 0x2, 3, TREG_ZERO, 1,
+    { { 0, }, { 8, 9, 17 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "fetchand4", TILEGX_OPC_FETCHAND4, 0x2, 3, TREG_ZERO, 1,
+    { { 0, }, { 8, 9, 17 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "fetchor", TILEGX_OPC_FETCHOR, 0x2, 3, TREG_ZERO, 1,
+    { { 0, }, { 8, 9, 17 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "fetchor4", TILEGX_OPC_FETCHOR4, 0x2, 3, TREG_ZERO, 1,
+    { { 0, }, { 8, 9, 17 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "finv", TILEGX_OPC_FINV, 0x2, 1, TREG_ZERO, 1,
+    { { 0, }, { 9 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "flush", TILEGX_OPC_FLUSH, 0x2, 1, TREG_ZERO, 1,
+    { { 0, }, { 9 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "flushwb", TILEGX_OPC_FLUSHWB, 0x2, 0, TREG_ZERO, 1,
+    { { 0, }, {  }, { 0, }, { 0, }, { 0, } },
+  },
+  { "fnop", TILEGX_OPC_FNOP, 0xf, 0, TREG_ZERO, 1,
+    { {  }, {  }, {  }, {  }, { 0, } },
+  },
+  { "fsingle_add1", TILEGX_OPC_FSINGLE_ADD1, 0x1, 3, TREG_ZERO, 1,
+    { { 6, 7, 16 }, { 0, }, { 0, }, { 0, }, { 0, } },
+  },
+  { "fsingle_addsub2", TILEGX_OPC_FSINGLE_ADDSUB2, 0x1, 3, TREG_ZERO, 1,
+    { { 23, 7, 16 }, { 0, }, { 0, }, { 0, }, { 0, } },
+  },
+  { "fsingle_mul1", TILEGX_OPC_FSINGLE_MUL1, 0x1, 3, TREG_ZERO, 1,
+    { { 6, 7, 16 }, { 0, }, { 0, }, { 0, }, { 0, } },
+  },
+  { "fsingle_mul2", TILEGX_OPC_FSINGLE_MUL2, 0x1, 3, TREG_ZERO, 1,
+    { { 6, 7, 16 }, { 0, }, { 0, }, { 0, }, { 0, } },
+  },
+  { "fsingle_pack1", TILEGX_OPC_FSINGLE_PACK1, 0x5, 2, TREG_ZERO, 1,
+    { { 6, 7 }, { 0, }, { 10, 11 }, { 0, }, { 0, } },
+  },
+  { "fsingle_pack2", TILEGX_OPC_FSINGLE_PACK2, 0x1, 3, TREG_ZERO, 1,
+    { { 6, 7, 16 }, { 0, }, { 0, }, { 0, }, { 0, } },
+  },
+  { "fsingle_sub1", TILEGX_OPC_FSINGLE_SUB1, 0x1, 3, TREG_ZERO, 1,
+    { { 6, 7, 16 }, { 0, }, { 0, }, { 0, }, { 0, } },
+  },
+  { "icoh", TILEGX_OPC_ICOH, 0x2, 1, TREG_ZERO, 1,
+    { { 0, }, { 9 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "ill", TILEGX_OPC_ILL, 0xa, 0, TREG_ZERO, 1,
+    { { 0, }, {  }, { 0, }, {  }, { 0, } },
+  },
+  { "inv", TILEGX_OPC_INV, 0x2, 1, TREG_ZERO, 1,
+    { { 0, }, { 9 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "iret", TILEGX_OPC_IRET, 0x2, 0, TREG_ZERO, 1,
+    { { 0, }, {  }, { 0, }, { 0, }, { 0, } },
+  },
+  { "j", TILEGX_OPC_J, 0x2, 1, TREG_ZERO, 1,
+    { { 0, }, { 25 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "jal", TILEGX_OPC_JAL, 0x2, 1, TREG_LR, 1,
+    { { 0, }, { 25 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "jalr", TILEGX_OPC_JALR, 0xa, 1, TREG_LR, 1,
+    { { 0, }, { 9 }, { 0, }, { 13 }, { 0, } },
+  },
+  { "jalrp", TILEGX_OPC_JALRP, 0xa, 1, TREG_LR, 1,
+    { { 0, }, { 9 }, { 0, }, { 13 }, { 0, } },
+  },
+  { "jr", TILEGX_OPC_JR, 0xa, 1, TREG_ZERO, 1,
+    { { 0, }, { 9 }, { 0, }, { 13 }, { 0, } },
+  },
+  { "jrp", TILEGX_OPC_JRP, 0xa, 1, TREG_ZERO, 1,
+    { { 0, }, { 9 }, { 0, }, { 13 }, { 0, } },
+  },
+  { "ld", TILEGX_OPC_LD, 0x12, 2, TREG_ZERO, 1,
+    { { 0, }, { 8, 9 }, { 0, }, { 0, }, { 26, 14 } },
+  },
+  { "ld1s", TILEGX_OPC_LD1S, 0x12, 2, TREG_ZERO, 1,
+    { { 0, }, { 8, 9 }, { 0, }, { 0, }, { 26, 14 } },
+  },
+  { "ld1s_add", TILEGX_OPC_LD1S_ADD, 0x2, 3, TREG_ZERO, 1,
+    { { 0, }, { 8, 15, 1 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "ld1u", TILEGX_OPC_LD1U, 0x12, 2, TREG_ZERO, 1,
+    { { 0, }, { 8, 9 }, { 0, }, { 0, }, { 26, 14 } },
+  },
+  { "ld1u_add", TILEGX_OPC_LD1U_ADD, 0x2, 3, TREG_ZERO, 1,
+    { { 0, }, { 8, 15, 1 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "ld2s", TILEGX_OPC_LD2S, 0x12, 2, TREG_ZERO, 1,
+    { { 0, }, { 8, 9 }, { 0, }, { 0, }, { 26, 14 } },
+  },
+  { "ld2s_add", TILEGX_OPC_LD2S_ADD, 0x2, 3, TREG_ZERO, 1,
+    { { 0, }, { 8, 15, 1 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "ld2u", TILEGX_OPC_LD2U, 0x12, 2, TREG_ZERO, 1,
+    { { 0, }, { 8, 9 }, { 0, }, { 0, }, { 26, 14 } },
+  },
+  { "ld2u_add", TILEGX_OPC_LD2U_ADD, 0x2, 3, TREG_ZERO, 1,
+    { { 0, }, { 8, 15, 1 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "ld4s", TILEGX_OPC_LD4S, 0x12, 2, TREG_ZERO, 1,
+    { { 0, }, { 8, 9 }, { 0, }, { 0, }, { 26, 14 } },
+  },
+  { "ld4s_add", TILEGX_OPC_LD4S_ADD, 0x2, 3, TREG_ZERO, 1,
+    { { 0, }, { 8, 15, 1 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "ld4u", TILEGX_OPC_LD4U, 0x12, 2, TREG_ZERO, 1,
+    { { 0, }, { 8, 9 }, { 0, }, { 0, }, { 26, 14 } },
+  },
+  { "ld4u_add", TILEGX_OPC_LD4U_ADD, 0x2, 3, TREG_ZERO, 1,
+    { { 0, }, { 8, 15, 1 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "ld_add", TILEGX_OPC_LD_ADD, 0x2, 3, TREG_ZERO, 1,
+    { { 0, }, { 8, 15, 1 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "ldna", TILEGX_OPC_LDNA, 0x2, 2, TREG_ZERO, 1,
+    { { 0, }, { 8, 9 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "ldna_add", TILEGX_OPC_LDNA_ADD, 0x2, 3, TREG_ZERO, 1,
+    { { 0, }, { 8, 15, 1 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "ldnt", TILEGX_OPC_LDNT, 0x2, 2, TREG_ZERO, 1,
+    { { 0, }, { 8, 9 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "ldnt1s", TILEGX_OPC_LDNT1S, 0x2, 2, TREG_ZERO, 1,
+    { { 0, }, { 8, 9 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "ldnt1s_add", TILEGX_OPC_LDNT1S_ADD, 0x2, 3, TREG_ZERO, 1,
+    { { 0, }, { 8, 15, 1 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "ldnt1u", TILEGX_OPC_LDNT1U, 0x2, 2, TREG_ZERO, 1,
+    { { 0, }, { 8, 9 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "ldnt1u_add", TILEGX_OPC_LDNT1U_ADD, 0x2, 3, TREG_ZERO, 1,
+    { { 0, }, { 8, 15, 1 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "ldnt2s", TILEGX_OPC_LDNT2S, 0x2, 2, TREG_ZERO, 1,
+    { { 0, }, { 8, 9 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "ldnt2s_add", TILEGX_OPC_LDNT2S_ADD, 0x2, 3, TREG_ZERO, 1,
+    { { 0, }, { 8, 15, 1 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "ldnt2u", TILEGX_OPC_LDNT2U, 0x2, 2, TREG_ZERO, 1,
+    { { 0, }, { 8, 9 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "ldnt2u_add", TILEGX_OPC_LDNT2U_ADD, 0x2, 3, TREG_ZERO, 1,
+    { { 0, }, { 8, 15, 1 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "ldnt4s", TILEGX_OPC_LDNT4S, 0x2, 2, TREG_ZERO, 1,
+    { { 0, }, { 8, 9 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "ldnt4s_add", TILEGX_OPC_LDNT4S_ADD, 0x2, 3, TREG_ZERO, 1,
+    { { 0, }, { 8, 15, 1 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "ldnt4u", TILEGX_OPC_LDNT4U, 0x2, 2, TREG_ZERO, 1,
+    { { 0, }, { 8, 9 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "ldnt4u_add", TILEGX_OPC_LDNT4U_ADD, 0x2, 3, TREG_ZERO, 1,
+    { { 0, }, { 8, 15, 1 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "ldnt_add", TILEGX_OPC_LDNT_ADD, 0x2, 3, TREG_ZERO, 1,
+    { { 0, }, { 8, 15, 1 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "lnk", TILEGX_OPC_LNK, 0xa, 1, TREG_ZERO, 1,
+    { { 0, }, { 8 }, { 0, }, { 12 }, { 0, } },
+  },
+  { "mf", TILEGX_OPC_MF, 0x2, 0, TREG_ZERO, 1,
+    { { 0, }, {  }, { 0, }, { 0, }, { 0, } },
+  },
+  { "mfspr", TILEGX_OPC_MFSPR, 0x2, 2, TREG_ZERO, 1,
+    { { 0, }, { 8, 27 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "mm", TILEGX_OPC_MM, 0x1, 4, TREG_ZERO, 1,
+    { { 23, 7, 21, 22 }, { 0, }, { 0, }, { 0, }, { 0, } },
+  },
+  { "mnz", TILEGX_OPC_MNZ, 0xf, 3, TREG_ZERO, 1,
+    { { 6, 7, 16 }, { 8, 9, 17 }, { 10, 11, 18 }, { 12, 13, 19 }, { 0, } },
+  },
+  { "mtspr", TILEGX_OPC_MTSPR, 0x2, 2, TREG_ZERO, 1,
+    { { 0, }, { 28, 9 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "mul_hs_hs", TILEGX_OPC_MUL_HS_HS, 0x5, 3, TREG_ZERO, 1,
+    { { 6, 7, 16 }, { 0, }, { 10, 11, 18 }, { 0, }, { 0, } },
+  },
+  { "mul_hs_hu", TILEGX_OPC_MUL_HS_HU, 0x1, 3, TREG_ZERO, 1,
+    { { 6, 7, 16 }, { 0, }, { 0, }, { 0, }, { 0, } },
+  },
+  { "mul_hs_ls", TILEGX_OPC_MUL_HS_LS, 0x1, 3, TREG_ZERO, 1,
+    { { 6, 7, 16 }, { 0, }, { 0, }, { 0, }, { 0, } },
+  },
+  { "mul_hs_lu", TILEGX_OPC_MUL_HS_LU, 0x1, 3, TREG_ZERO, 1,
+    { { 6, 7, 16 }, { 0, }, { 0, }, { 0, }, { 0, } },
+  },
+  { "mul_hu_hu", TILEGX_OPC_MUL_HU_HU, 0x5, 3, TREG_ZERO, 1,
+    { { 6, 7, 16 }, { 0, }, { 10, 11, 18 }, { 0, }, { 0, } },
+  },
+  { "mul_hu_ls", TILEGX_OPC_MUL_HU_LS, 0x1, 3, TREG_ZERO, 1,
+    { { 6, 7, 16 }, { 0, }, { 0, }, { 0, }, { 0, } },
+  },
+  { "mul_hu_lu", TILEGX_OPC_MUL_HU_LU, 0x1, 3, TREG_ZERO, 1,
+    { { 6, 7, 16 }, { 0, }, { 0, }, { 0, }, { 0, } },
+  },
+  { "mul_ls_ls", TILEGX_OPC_MUL_LS_LS, 0x5, 3, TREG_ZERO, 1,
+    { { 6, 7, 16 }, { 0, }, { 10, 11, 18 }, { 0, }, { 0, } },
+  },
+  { "mul_ls_lu", TILEGX_OPC_MUL_LS_LU, 0x1, 3, TREG_ZERO, 1,
+    { { 6, 7, 16 }, { 0, }, { 0, }, { 0, }, { 0, } },
+  },
+  { "mul_lu_lu", TILEGX_OPC_MUL_LU_LU, 0x5, 3, TREG_ZERO, 1,
+    { { 6, 7, 16 }, { 0, }, { 10, 11, 18 }, { 0, }, { 0, } },
+  },
+  { "mula_hs_hs", TILEGX_OPC_MULA_HS_HS, 0x5, 3, TREG_ZERO, 1,
+    { { 23, 7, 16 }, { 0, }, { 24, 11, 18 }, { 0, }, { 0, } },
+  },
+  { "mula_hs_hu", TILEGX_OPC_MULA_HS_HU, 0x1, 3, TREG_ZERO, 1,
+    { { 23, 7, 16 }, { 0, }, { 0, }, { 0, }, { 0, } },
+  },
+  { "mula_hs_ls", TILEGX_OPC_MULA_HS_LS, 0x1, 3, TREG_ZERO, 1,
+    { { 23, 7, 16 }, { 0, }, { 0, }, { 0, }, { 0, } },
+  },
+  { "mula_hs_lu", TILEGX_OPC_MULA_HS_LU, 0x1, 3, TREG_ZERO, 1,
+    { { 23, 7, 16 }, { 0, }, { 0, }, { 0, }, { 0, } },
+  },
+  { "mula_hu_hu", TILEGX_OPC_MULA_HU_HU, 0x5, 3, TREG_ZERO, 1,
+    { { 23, 7, 16 }, { 0, }, { 24, 11, 18 }, { 0, }, { 0, } },
+  },
+  { "mula_hu_ls", TILEGX_OPC_MULA_HU_LS, 0x1, 3, TREG_ZERO, 1,
+    { { 23, 7, 16 }, { 0, }, { 0, }, { 0, }, { 0, } },
+  },
+  { "mula_hu_lu", TILEGX_OPC_MULA_HU_LU, 0x1, 3, TREG_ZERO, 1,
+    { { 23, 7, 16 }, { 0, }, { 0, }, { 0, }, { 0, } },
+  },
+  { "mula_ls_ls", TILEGX_OPC_MULA_LS_LS, 0x5, 3, TREG_ZERO, 1,
+    { { 23, 7, 16 }, { 0, }, { 24, 11, 18 }, { 0, }, { 0, } },
+  },
+  { "mula_ls_lu", TILEGX_OPC_MULA_LS_LU, 0x1, 3, TREG_ZERO, 1,
+    { { 23, 7, 16 }, { 0, }, { 0, }, { 0, }, { 0, } },
+  },
+  { "mula_lu_lu", TILEGX_OPC_MULA_LU_LU, 0x5, 3, TREG_ZERO, 1,
+    { { 23, 7, 16 }, { 0, }, { 24, 11, 18 }, { 0, }, { 0, } },
+  },
+  { "mulax", TILEGX_OPC_MULAX, 0x5, 3, TREG_ZERO, 1,
+    { { 23, 7, 16 }, { 0, }, { 24, 11, 18 }, { 0, }, { 0, } },
+  },
+  { "mulx", TILEGX_OPC_MULX, 0x5, 3, TREG_ZERO, 1,
+    { { 6, 7, 16 }, { 0, }, { 10, 11, 18 }, { 0, }, { 0, } },
+  },
+  { "mz", TILEGX_OPC_MZ, 0xf, 3, TREG_ZERO, 1,
+    { { 6, 7, 16 }, { 8, 9, 17 }, { 10, 11, 18 }, { 12, 13, 19 }, { 0, } },
+  },
+  { "nap", TILEGX_OPC_NAP, 0x2, 0, TREG_ZERO, 0,
+    { { 0, }, {  }, { 0, }, { 0, }, { 0, } },
+  },
+  { "nop", TILEGX_OPC_NOP, 0xf, 0, TREG_ZERO, 1,
+    { {  }, {  }, {  }, {  }, { 0, } },
+  },
+  { "nor", TILEGX_OPC_NOR, 0xf, 3, TREG_ZERO, 1,
+    { { 6, 7, 16 }, { 8, 9, 17 }, { 10, 11, 18 }, { 12, 13, 19 }, { 0, } },
+  },
+  { "or", TILEGX_OPC_OR, 0xf, 3, TREG_ZERO, 1,
+    { { 6, 7, 16 }, { 8, 9, 17 }, { 10, 11, 18 }, { 12, 13, 19 }, { 0, } },
+  },
+  { "ori", TILEGX_OPC_ORI, 0x3, 3, TREG_ZERO, 1,
+    { { 6, 7, 0 }, { 8, 9, 1 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "pcnt", TILEGX_OPC_PCNT, 0x5, 2, TREG_ZERO, 1,
+    { { 6, 7 }, { 0, }, { 10, 11 }, { 0, }, { 0, } },
+  },
+  { "revbits", TILEGX_OPC_REVBITS, 0x5, 2, TREG_ZERO, 1,
+    { { 6, 7 }, { 0, }, { 10, 11 }, { 0, }, { 0, } },
+  },
+  { "revbytes", TILEGX_OPC_REVBYTES, 0x5, 2, TREG_ZERO, 1,
+    { { 6, 7 }, { 0, }, { 10, 11 }, { 0, }, { 0, } },
+  },
+  { "rotl", TILEGX_OPC_ROTL, 0xf, 3, TREG_ZERO, 1,
+    { { 6, 7, 16 }, { 8, 9, 17 }, { 10, 11, 18 }, { 12, 13, 19 }, { 0, } },
+  },
+  { "rotli", TILEGX_OPC_ROTLI, 0xf, 3, TREG_ZERO, 1,
+    { { 6, 7, 29 }, { 8, 9, 30 }, { 10, 11, 31 }, { 12, 13, 32 }, { 0, } },
+  },
+  { "shl", TILEGX_OPC_SHL, 0xf, 3, TREG_ZERO, 1,
+    { { 6, 7, 16 }, { 8, 9, 17 }, { 10, 11, 18 }, { 12, 13, 19 }, { 0, } },
+  },
+  { "shl16insli", TILEGX_OPC_SHL16INSLI, 0x3, 3, TREG_ZERO, 1,
+    { { 6, 7, 4 }, { 8, 9, 5 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "shl1add", TILEGX_OPC_SHL1ADD, 0xf, 3, TREG_ZERO, 1,
+    { { 6, 7, 16 }, { 8, 9, 17 }, { 10, 11, 18 }, { 12, 13, 19 }, { 0, } },
+  },
+  { "shl1addx", TILEGX_OPC_SHL1ADDX, 0xf, 3, TREG_ZERO, 1,
+    { { 6, 7, 16 }, { 8, 9, 17 }, { 10, 11, 18 }, { 12, 13, 19 }, { 0, } },
+  },
+  { "shl2add", TILEGX_OPC_SHL2ADD, 0xf, 3, TREG_ZERO, 1,
+    { { 6, 7, 16 }, { 8, 9, 17 }, { 10, 11, 18 }, { 12, 13, 19 }, { 0, } },
+  },
+  { "shl2addx", TILEGX_OPC_SHL2ADDX, 0xf, 3, TREG_ZERO, 1,
+    { { 6, 7, 16 }, { 8, 9, 17 }, { 10, 11, 18 }, { 12, 13, 19 }, { 0, } },
+  },
+  { "shl3add", TILEGX_OPC_SHL3ADD, 0xf, 3, TREG_ZERO, 1,
+    { { 6, 7, 16 }, { 8, 9, 17 }, { 10, 11, 18 }, { 12, 13, 19 }, { 0, } },
+  },
+  { "shl3addx", TILEGX_OPC_SHL3ADDX, 0xf, 3, TREG_ZERO, 1,
+    { { 6, 7, 16 }, { 8, 9, 17 }, { 10, 11, 18 }, { 12, 13, 19 }, { 0, } },
+  },
+  { "shli", TILEGX_OPC_SHLI, 0xf, 3, TREG_ZERO, 1,
+    { { 6, 7, 29 }, { 8, 9, 30 }, { 10, 11, 31 }, { 12, 13, 32 }, { 0, } },
+  },
+  { "shlx", TILEGX_OPC_SHLX, 0x3, 3, TREG_ZERO, 1,
+    { { 6, 7, 16 }, { 8, 9, 17 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "shlxi", TILEGX_OPC_SHLXI, 0x3, 3, TREG_ZERO, 1,
+    { { 6, 7, 29 }, { 8, 9, 30 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "shrs", TILEGX_OPC_SHRS, 0xf, 3, TREG_ZERO, 1,
+    { { 6, 7, 16 }, { 8, 9, 17 }, { 10, 11, 18 }, { 12, 13, 19 }, { 0, } },
+  },
+  { "shrsi", TILEGX_OPC_SHRSI, 0xf, 3, TREG_ZERO, 1,
+    { { 6, 7, 29 }, { 8, 9, 30 }, { 10, 11, 31 }, { 12, 13, 32 }, { 0, } },
+  },
+  { "shru", TILEGX_OPC_SHRU, 0xf, 3, TREG_ZERO, 1,
+    { { 6, 7, 16 }, { 8, 9, 17 }, { 10, 11, 18 }, { 12, 13, 19 }, { 0, } },
+  },
+  { "shrui", TILEGX_OPC_SHRUI, 0xf, 3, TREG_ZERO, 1,
+    { { 6, 7, 29 }, { 8, 9, 30 }, { 10, 11, 31 }, { 12, 13, 32 }, { 0, } },
+  },
+  { "shrux", TILEGX_OPC_SHRUX, 0x3, 3, TREG_ZERO, 1,
+    { { 6, 7, 16 }, { 8, 9, 17 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "shruxi", TILEGX_OPC_SHRUXI, 0x3, 3, TREG_ZERO, 1,
+    { { 6, 7, 29 }, { 8, 9, 30 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "shufflebytes", TILEGX_OPC_SHUFFLEBYTES, 0x1, 3, TREG_ZERO, 1,
+    { { 23, 7, 16 }, { 0, }, { 0, }, { 0, }, { 0, } },
+  },
+  { "st", TILEGX_OPC_ST, 0x12, 2, TREG_ZERO, 1,
+    { { 0, }, { 9, 17 }, { 0, }, { 0, }, { 14, 33 } },
+  },
+  { "st1", TILEGX_OPC_ST1, 0x12, 2, TREG_ZERO, 1,
+    { { 0, }, { 9, 17 }, { 0, }, { 0, }, { 14, 33 } },
+  },
+  { "st1_add", TILEGX_OPC_ST1_ADD, 0x2, 3, TREG_ZERO, 1,
+    { { 0, }, { 15, 17, 34 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "st2", TILEGX_OPC_ST2, 0x12, 2, TREG_ZERO, 1,
+    { { 0, }, { 9, 17 }, { 0, }, { 0, }, { 14, 33 } },
+  },
+  { "st2_add", TILEGX_OPC_ST2_ADD, 0x2, 3, TREG_ZERO, 1,
+    { { 0, }, { 15, 17, 34 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "st4", TILEGX_OPC_ST4, 0x12, 2, TREG_ZERO, 1,
+    { { 0, }, { 9, 17 }, { 0, }, { 0, }, { 14, 33 } },
+  },
+  { "st4_add", TILEGX_OPC_ST4_ADD, 0x2, 3, TREG_ZERO, 1,
+    { { 0, }, { 15, 17, 34 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "st_add", TILEGX_OPC_ST_ADD, 0x2, 3, TREG_ZERO, 1,
+    { { 0, }, { 15, 17, 34 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "stnt", TILEGX_OPC_STNT, 0x2, 2, TREG_ZERO, 1,
+    { { 0, }, { 9, 17 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "stnt1", TILEGX_OPC_STNT1, 0x2, 2, TREG_ZERO, 1,
+    { { 0, }, { 9, 17 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "stnt1_add", TILEGX_OPC_STNT1_ADD, 0x2, 3, TREG_ZERO, 1,
+    { { 0, }, { 15, 17, 34 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "stnt2", TILEGX_OPC_STNT2, 0x2, 2, TREG_ZERO, 1,
+    { { 0, }, { 9, 17 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "stnt2_add", TILEGX_OPC_STNT2_ADD, 0x2, 3, TREG_ZERO, 1,
+    { { 0, }, { 15, 17, 34 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "stnt4", TILEGX_OPC_STNT4, 0x2, 2, TREG_ZERO, 1,
+    { { 0, }, { 9, 17 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "stnt4_add", TILEGX_OPC_STNT4_ADD, 0x2, 3, TREG_ZERO, 1,
+    { { 0, }, { 15, 17, 34 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "stnt_add", TILEGX_OPC_STNT_ADD, 0x2, 3, TREG_ZERO, 1,
+    { { 0, }, { 15, 17, 34 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "sub", TILEGX_OPC_SUB, 0xf, 3, TREG_ZERO, 1,
+    { { 6, 7, 16 }, { 8, 9, 17 }, { 10, 11, 18 }, { 12, 13, 19 }, { 0, } },
+  },
+  { "subx", TILEGX_OPC_SUBX, 0xf, 3, TREG_ZERO, 1,
+    { { 6, 7, 16 }, { 8, 9, 17 }, { 10, 11, 18 }, { 12, 13, 19 }, { 0, } },
+  },
+  { "subxsc", TILEGX_OPC_SUBXSC, 0x3, 3, TREG_ZERO, 1,
+    { { 6, 7, 16 }, { 8, 9, 17 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "swint0", TILEGX_OPC_SWINT0, 0x2, 0, TREG_ZERO, 0,
+    { { 0, }, {  }, { 0, }, { 0, }, { 0, } },
+  },
+  { "swint1", TILEGX_OPC_SWINT1, 0x2, 0, TREG_ZERO, 0,
+    { { 0, }, {  }, { 0, }, { 0, }, { 0, } },
+  },
+  { "swint2", TILEGX_OPC_SWINT2, 0x2, 0, TREG_ZERO, 0,
+    { { 0, }, {  }, { 0, }, { 0, }, { 0, } },
+  },
+  { "swint3", TILEGX_OPC_SWINT3, 0x2, 0, TREG_ZERO, 0,
+    { { 0, }, {  }, { 0, }, { 0, }, { 0, } },
+  },
+  { "tblidxb0", TILEGX_OPC_TBLIDXB0, 0x5, 2, TREG_ZERO, 1,
+    { { 23, 7 }, { 0, }, { 24, 11 }, { 0, }, { 0, } },
+  },
+  { "tblidxb1", TILEGX_OPC_TBLIDXB1, 0x5, 2, TREG_ZERO, 1,
+    { { 23, 7 }, { 0, }, { 24, 11 }, { 0, }, { 0, } },
+  },
+  { "tblidxb2", TILEGX_OPC_TBLIDXB2, 0x5, 2, TREG_ZERO, 1,
+    { { 23, 7 }, { 0, }, { 24, 11 }, { 0, }, { 0, } },
+  },
+  { "tblidxb3", TILEGX_OPC_TBLIDXB3, 0x5, 2, TREG_ZERO, 1,
+    { { 23, 7 }, { 0, }, { 24, 11 }, { 0, }, { 0, } },
+  },
+  { "v1add", TILEGX_OPC_V1ADD, 0x3, 3, TREG_ZERO, 1,
+    { { 6, 7, 16 }, { 8, 9, 17 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "v1addi", TILEGX_OPC_V1ADDI, 0x3, 3, TREG_ZERO, 1,
+    { { 6, 7, 0 }, { 8, 9, 1 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "v1adduc", TILEGX_OPC_V1ADDUC, 0x3, 3, TREG_ZERO, 1,
+    { { 6, 7, 16 }, { 8, 9, 17 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "v1adiffu", TILEGX_OPC_V1ADIFFU, 0x1, 3, TREG_ZERO, 1,
+    { { 6, 7, 16 }, { 0, }, { 0, }, { 0, }, { 0, } },
+  },
+  { "v1avgu", TILEGX_OPC_V1AVGU, 0x1, 3, TREG_ZERO, 1,
+    { { 6, 7, 16 }, { 0, }, { 0, }, { 0, }, { 0, } },
+  },
+  { "v1cmpeq", TILEGX_OPC_V1CMPEQ, 0x3, 3, TREG_ZERO, 1,
+    { { 6, 7, 16 }, { 8, 9, 17 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "v1cmpeqi", TILEGX_OPC_V1CMPEQI, 0x3, 3, TREG_ZERO, 1,
+    { { 6, 7, 0 }, { 8, 9, 1 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "v1cmples", TILEGX_OPC_V1CMPLES, 0x3, 3, TREG_ZERO, 1,
+    { { 6, 7, 16 }, { 8, 9, 17 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "v1cmpleu", TILEGX_OPC_V1CMPLEU, 0x3, 3, TREG_ZERO, 1,
+    { { 6, 7, 16 }, { 8, 9, 17 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "v1cmplts", TILEGX_OPC_V1CMPLTS, 0x3, 3, TREG_ZERO, 1,
+    { { 6, 7, 16 }, { 8, 9, 17 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "v1cmpltsi", TILEGX_OPC_V1CMPLTSI, 0x3, 3, TREG_ZERO, 1,
+    { { 6, 7, 0 }, { 8, 9, 1 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "v1cmpltu", TILEGX_OPC_V1CMPLTU, 0x3, 3, TREG_ZERO, 1,
+    { { 6, 7, 16 }, { 8, 9, 17 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "v1cmpltui", TILEGX_OPC_V1CMPLTUI, 0x3, 3, TREG_ZERO, 1,
+    { { 6, 7, 0 }, { 8, 9, 1 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "v1cmpne", TILEGX_OPC_V1CMPNE, 0x3, 3, TREG_ZERO, 1,
+    { { 6, 7, 16 }, { 8, 9, 17 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "v1ddotpu", TILEGX_OPC_V1DDOTPU, 0x1, 3, TREG_ZERO, 1,
+    { { 6, 7, 16 }, { 0, }, { 0, }, { 0, }, { 0, } },
+  },
+  { "v1ddotpua", TILEGX_OPC_V1DDOTPUA, 0x1, 3, TREG_ZERO, 1,
+    { { 23, 7, 16 }, { 0, }, { 0, }, { 0, }, { 0, } },
+  },
+  { "v1ddotpus", TILEGX_OPC_V1DDOTPUS, 0x1, 3, TREG_ZERO, 1,
+    { { 6, 7, 16 }, { 0, }, { 0, }, { 0, }, { 0, } },
+  },
+  { "v1ddotpusa", TILEGX_OPC_V1DDOTPUSA, 0x1, 3, TREG_ZERO, 1,
+    { { 23, 7, 16 }, { 0, }, { 0, }, { 0, }, { 0, } },
+  },
+  { "v1dotp", TILEGX_OPC_V1DOTP, 0x1, 3, TREG_ZERO, 1,
+    { { 6, 7, 16 }, { 0, }, { 0, }, { 0, }, { 0, } },
+  },
+  { "v1dotpa", TILEGX_OPC_V1DOTPA, 0x1, 3, TREG_ZERO, 1,
+    { { 23, 7, 16 }, { 0, }, { 0, }, { 0, }, { 0, } },
+  },
+  { "v1dotpu", TILEGX_OPC_V1DOTPU, 0x1, 3, TREG_ZERO, 1,
+    { { 6, 7, 16 }, { 0, }, { 0, }, { 0, }, { 0, } },
+  },
+  { "v1dotpua", TILEGX_OPC_V1DOTPUA, 0x1, 3, TREG_ZERO, 1,
+    { { 23, 7, 16 }, { 0, }, { 0, }, { 0, }, { 0, } },
+  },
+  { "v1dotpus", TILEGX_OPC_V1DOTPUS, 0x1, 3, TREG_ZERO, 1,
+    { { 6, 7, 16 }, { 0, }, { 0, }, { 0, }, { 0, } },
+  },
+  { "v1dotpusa", TILEGX_OPC_V1DOTPUSA, 0x1, 3, TREG_ZERO, 1,
+    { { 23, 7, 16 }, { 0, }, { 0, }, { 0, }, { 0, } },
+  },
+  { "v1int_h", TILEGX_OPC_V1INT_H, 0x3, 3, TREG_ZERO, 1,
+    { { 6, 7, 16 }, { 8, 9, 17 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "v1int_l", TILEGX_OPC_V1INT_L, 0x3, 3, TREG_ZERO, 1,
+    { { 6, 7, 16 }, { 8, 9, 17 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "v1maxu", TILEGX_OPC_V1MAXU, 0x3, 3, TREG_ZERO, 1,
+    { { 6, 7, 16 }, { 8, 9, 17 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "v1maxui", TILEGX_OPC_V1MAXUI, 0x3, 3, TREG_ZERO, 1,
+    { { 6, 7, 0 }, { 8, 9, 1 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "v1minu", TILEGX_OPC_V1MINU, 0x3, 3, TREG_ZERO, 1,
+    { { 6, 7, 16 }, { 8, 9, 17 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "v1minui", TILEGX_OPC_V1MINUI, 0x3, 3, TREG_ZERO, 1,
+    { { 6, 7, 0 }, { 8, 9, 1 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "v1mnz", TILEGX_OPC_V1MNZ, 0x3, 3, TREG_ZERO, 1,
+    { { 6, 7, 16 }, { 8, 9, 17 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "v1multu", TILEGX_OPC_V1MULTU, 0x1, 3, TREG_ZERO, 1,
+    { { 6, 7, 16 }, { 0, }, { 0, }, { 0, }, { 0, } },
+  },
+  { "v1mulu", TILEGX_OPC_V1MULU, 0x1, 3, TREG_ZERO, 1,
+    { { 6, 7, 16 }, { 0, }, { 0, }, { 0, }, { 0, } },
+  },
+  { "v1mulus", TILEGX_OPC_V1MULUS, 0x1, 3, TREG_ZERO, 1,
+    { { 6, 7, 16 }, { 0, }, { 0, }, { 0, }, { 0, } },
+  },
+  { "v1mz", TILEGX_OPC_V1MZ, 0x3, 3, TREG_ZERO, 1,
+    { { 6, 7, 16 }, { 8, 9, 17 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "v1sadau", TILEGX_OPC_V1SADAU, 0x1, 3, TREG_ZERO, 1,
+    { { 23, 7, 16 }, { 0, }, { 0, }, { 0, }, { 0, } },
+  },
+  { "v1sadu", TILEGX_OPC_V1SADU, 0x1, 3, TREG_ZERO, 1,
+    { { 6, 7, 16 }, { 0, }, { 0, }, { 0, }, { 0, } },
+  },
+  { "v1shl", TILEGX_OPC_V1SHL, 0x3, 3, TREG_ZERO, 1,
+    { { 6, 7, 16 }, { 8, 9, 17 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "v1shli", TILEGX_OPC_V1SHLI, 0x3, 3, TREG_ZERO, 1,
+    { { 6, 7, 29 }, { 8, 9, 30 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "v1shrs", TILEGX_OPC_V1SHRS, 0x3, 3, TREG_ZERO, 1,
+    { { 6, 7, 16 }, { 8, 9, 17 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "v1shrsi", TILEGX_OPC_V1SHRSI, 0x3, 3, TREG_ZERO, 1,
+    { { 6, 7, 29 }, { 8, 9, 30 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "v1shru", TILEGX_OPC_V1SHRU, 0x3, 3, TREG_ZERO, 1,
+    { { 6, 7, 16 }, { 8, 9, 17 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "v1shrui", TILEGX_OPC_V1SHRUI, 0x3, 3, TREG_ZERO, 1,
+    { { 6, 7, 29 }, { 8, 9, 30 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "v1sub", TILEGX_OPC_V1SUB, 0x3, 3, TREG_ZERO, 1,
+    { { 6, 7, 16 }, { 8, 9, 17 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "v1subuc", TILEGX_OPC_V1SUBUC, 0x3, 3, TREG_ZERO, 1,
+    { { 6, 7, 16 }, { 8, 9, 17 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "v2add", TILEGX_OPC_V2ADD, 0x3, 3, TREG_ZERO, 1,
+    { { 6, 7, 16 }, { 8, 9, 17 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "v2addi", TILEGX_OPC_V2ADDI, 0x3, 3, TREG_ZERO, 1,
+    { { 6, 7, 0 }, { 8, 9, 1 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "v2addsc", TILEGX_OPC_V2ADDSC, 0x3, 3, TREG_ZERO, 1,
+    { { 6, 7, 16 }, { 8, 9, 17 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "v2adiffs", TILEGX_OPC_V2ADIFFS, 0x1, 3, TREG_ZERO, 1,
+    { { 6, 7, 16 }, { 0, }, { 0, }, { 0, }, { 0, } },
+  },
+  { "v2avgs", TILEGX_OPC_V2AVGS, 0x1, 3, TREG_ZERO, 1,
+    { { 6, 7, 16 }, { 0, }, { 0, }, { 0, }, { 0, } },
+  },
+  { "v2cmpeq", TILEGX_OPC_V2CMPEQ, 0x3, 3, TREG_ZERO, 1,
+    { { 6, 7, 16 }, { 8, 9, 17 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "v2cmpeqi", TILEGX_OPC_V2CMPEQI, 0x3, 3, TREG_ZERO, 1,
+    { { 6, 7, 0 }, { 8, 9, 1 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "v2cmples", TILEGX_OPC_V2CMPLES, 0x3, 3, TREG_ZERO, 1,
+    { { 6, 7, 16 }, { 8, 9, 17 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "v2cmpleu", TILEGX_OPC_V2CMPLEU, 0x3, 3, TREG_ZERO, 1,
+    { { 6, 7, 16 }, { 8, 9, 17 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "v2cmplts", TILEGX_OPC_V2CMPLTS, 0x3, 3, TREG_ZERO, 1,
+    { { 6, 7, 16 }, { 8, 9, 17 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "v2cmpltsi", TILEGX_OPC_V2CMPLTSI, 0x3, 3, TREG_ZERO, 1,
+    { { 6, 7, 0 }, { 8, 9, 1 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "v2cmpltu", TILEGX_OPC_V2CMPLTU, 0x3, 3, TREG_ZERO, 1,
+    { { 6, 7, 16 }, { 8, 9, 17 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "v2cmpltui", TILEGX_OPC_V2CMPLTUI, 0x3, 3, TREG_ZERO, 1,
+    { { 6, 7, 0 }, { 8, 9, 1 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "v2cmpne", TILEGX_OPC_V2CMPNE, 0x3, 3, TREG_ZERO, 1,
+    { { 6, 7, 16 }, { 8, 9, 17 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "v2dotp", TILEGX_OPC_V2DOTP, 0x1, 3, TREG_ZERO, 1,
+    { { 6, 7, 16 }, { 0, }, { 0, }, { 0, }, { 0, } },
+  },
+  { "v2dotpa", TILEGX_OPC_V2DOTPA, 0x1, 3, TREG_ZERO, 1,
+    { { 23, 7, 16 }, { 0, }, { 0, }, { 0, }, { 0, } },
+  },
+  { "v2int_h", TILEGX_OPC_V2INT_H, 0x3, 3, TREG_ZERO, 1,
+    { { 6, 7, 16 }, { 8, 9, 17 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "v2int_l", TILEGX_OPC_V2INT_L, 0x3, 3, TREG_ZERO, 1,
+    { { 6, 7, 16 }, { 8, 9, 17 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "v2maxs", TILEGX_OPC_V2MAXS, 0x3, 3, TREG_ZERO, 1,
+    { { 6, 7, 16 }, { 8, 9, 17 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "v2maxsi", TILEGX_OPC_V2MAXSI, 0x3, 3, TREG_ZERO, 1,
+    { { 6, 7, 0 }, { 8, 9, 1 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "v2mins", TILEGX_OPC_V2MINS, 0x3, 3, TREG_ZERO, 1,
+    { { 6, 7, 16 }, { 8, 9, 17 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "v2minsi", TILEGX_OPC_V2MINSI, 0x3, 3, TREG_ZERO, 1,
+    { { 6, 7, 0 }, { 8, 9, 1 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "v2mnz", TILEGX_OPC_V2MNZ, 0x3, 3, TREG_ZERO, 1,
+    { { 6, 7, 16 }, { 8, 9, 17 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "v2mulfsc", TILEGX_OPC_V2MULFSC, 0x1, 3, TREG_ZERO, 1,
+    { { 6, 7, 16 }, { 0, }, { 0, }, { 0, }, { 0, } },
+  },
+  { "v2muls", TILEGX_OPC_V2MULS, 0x1, 3, TREG_ZERO, 1,
+    { { 6, 7, 16 }, { 0, }, { 0, }, { 0, }, { 0, } },
+  },
+  { "v2mults", TILEGX_OPC_V2MULTS, 0x1, 3, TREG_ZERO, 1,
+    { { 6, 7, 16 }, { 0, }, { 0, }, { 0, }, { 0, } },
+  },
+  { "v2mz", TILEGX_OPC_V2MZ, 0x3, 3, TREG_ZERO, 1,
+    { { 6, 7, 16 }, { 8, 9, 17 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "v2packh", TILEGX_OPC_V2PACKH, 0x3, 3, TREG_ZERO, 1,
+    { { 6, 7, 16 }, { 8, 9, 17 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "v2packl", TILEGX_OPC_V2PACKL, 0x3, 3, TREG_ZERO, 1,
+    { { 6, 7, 16 }, { 8, 9, 17 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "v2packuc", TILEGX_OPC_V2PACKUC, 0x3, 3, TREG_ZERO, 1,
+    { { 6, 7, 16 }, { 8, 9, 17 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "v2sadas", TILEGX_OPC_V2SADAS, 0x1, 3, TREG_ZERO, 1,
+    { { 23, 7, 16 }, { 0, }, { 0, }, { 0, }, { 0, } },
+  },
+  { "v2sadau", TILEGX_OPC_V2SADAU, 0x1, 3, TREG_ZERO, 1,
+    { { 23, 7, 16 }, { 0, }, { 0, }, { 0, }, { 0, } },
+  },
+  { "v2sads", TILEGX_OPC_V2SADS, 0x1, 3, TREG_ZERO, 1,
+    { { 6, 7, 16 }, { 0, }, { 0, }, { 0, }, { 0, } },
+  },
+  { "v2sadu", TILEGX_OPC_V2SADU, 0x1, 3, TREG_ZERO, 1,
+    { { 6, 7, 16 }, { 0, }, { 0, }, { 0, }, { 0, } },
+  },
+  { "v2shl", TILEGX_OPC_V2SHL, 0x3, 3, TREG_ZERO, 1,
+    { { 6, 7, 16 }, { 8, 9, 17 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "v2shli", TILEGX_OPC_V2SHLI, 0x3, 3, TREG_ZERO, 1,
+    { { 6, 7, 29 }, { 8, 9, 30 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "v2shlsc", TILEGX_OPC_V2SHLSC, 0x3, 3, TREG_ZERO, 1,
+    { { 6, 7, 16 }, { 8, 9, 17 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "v2shrs", TILEGX_OPC_V2SHRS, 0x3, 3, TREG_ZERO, 1,
+    { { 6, 7, 16 }, { 8, 9, 17 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "v2shrsi", TILEGX_OPC_V2SHRSI, 0x3, 3, TREG_ZERO, 1,
+    { { 6, 7, 29 }, { 8, 9, 30 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "v2shru", TILEGX_OPC_V2SHRU, 0x3, 3, TREG_ZERO, 1,
+    { { 6, 7, 16 }, { 8, 9, 17 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "v2shrui", TILEGX_OPC_V2SHRUI, 0x3, 3, TREG_ZERO, 1,
+    { { 6, 7, 29 }, { 8, 9, 30 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "v2sub", TILEGX_OPC_V2SUB, 0x3, 3, TREG_ZERO, 1,
+    { { 6, 7, 16 }, { 8, 9, 17 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "v2subsc", TILEGX_OPC_V2SUBSC, 0x3, 3, TREG_ZERO, 1,
+    { { 6, 7, 16 }, { 8, 9, 17 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "v4add", TILEGX_OPC_V4ADD, 0x3, 3, TREG_ZERO, 1,
+    { { 6, 7, 16 }, { 8, 9, 17 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "v4addsc", TILEGX_OPC_V4ADDSC, 0x3, 3, TREG_ZERO, 1,
+    { { 6, 7, 16 }, { 8, 9, 17 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "v4int_h", TILEGX_OPC_V4INT_H, 0x3, 3, TREG_ZERO, 1,
+    { { 6, 7, 16 }, { 8, 9, 17 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "v4int_l", TILEGX_OPC_V4INT_L, 0x3, 3, TREG_ZERO, 1,
+    { { 6, 7, 16 }, { 8, 9, 17 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "v4packsc", TILEGX_OPC_V4PACKSC, 0x3, 3, TREG_ZERO, 1,
+    { { 6, 7, 16 }, { 8, 9, 17 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "v4shl", TILEGX_OPC_V4SHL, 0x3, 3, TREG_ZERO, 1,
+    { { 6, 7, 16 }, { 8, 9, 17 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "v4shlsc", TILEGX_OPC_V4SHLSC, 0x3, 3, TREG_ZERO, 1,
+    { { 6, 7, 16 }, { 8, 9, 17 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "v4shrs", TILEGX_OPC_V4SHRS, 0x3, 3, TREG_ZERO, 1,
+    { { 6, 7, 16 }, { 8, 9, 17 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "v4shru", TILEGX_OPC_V4SHRU, 0x3, 3, TREG_ZERO, 1,
+    { { 6, 7, 16 }, { 8, 9, 17 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "v4sub", TILEGX_OPC_V4SUB, 0x3, 3, TREG_ZERO, 1,
+    { { 6, 7, 16 }, { 8, 9, 17 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "v4subsc", TILEGX_OPC_V4SUBSC, 0x3, 3, TREG_ZERO, 1,
+    { { 6, 7, 16 }, { 8, 9, 17 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "wh64", TILEGX_OPC_WH64, 0x2, 1, TREG_ZERO, 1,
+    { { 0, }, { 9 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "xor", TILEGX_OPC_XOR, 0xf, 3, TREG_ZERO, 1,
+    { { 6, 7, 16 }, { 8, 9, 17 }, { 10, 11, 18 }, { 12, 13, 19 }, { 0, } },
+  },
+  { "xori", TILEGX_OPC_XORI, 0x3, 3, TREG_ZERO, 1,
+    { { 6, 7, 0 }, { 8, 9, 1 }, { 0, }, { 0, }, { 0, } },
+  },
+  { NULL, TILEGX_OPC_NONE, 0, 0, TREG_ZERO, 0, { { 0, } },
+  }
+};
+#define BITFIELD(start, size) ((start) | (((1 << (size)) - 1) << 6))
+#define CHILD(array_index) (TILEGX_OPC_NONE + (array_index))
+
+static const unsigned short decode_X0_fsm[936] =
+{
+  BITFIELD(22, 9) /* index 0 */,
+  TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE,
+  TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE,
+  TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE,
+  TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE,
+  TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE,
+  TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE,
+  TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE,
+  TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE,
+  TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE,
+  TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE,
+  TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE,
+  TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE,
+  TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE,
+  TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE,
+  TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE,
+  TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE,
+  CHILD(513), CHILD(513), CHILD(513), CHILD(513), CHILD(513), CHILD(513),
+  CHILD(513), CHILD(513), CHILD(513), CHILD(513), CHILD(513), CHILD(513),
+  CHILD(513), CHILD(513), CHILD(513), CHILD(513), CHILD(513), CHILD(513),
+  CHILD(513), CHILD(513), CHILD(513), CHILD(513), CHILD(513), CHILD(513),
+  CHILD(513), CHILD(513), CHILD(513), CHILD(513), CHILD(513), CHILD(513),
+  CHILD(513), CHILD(513), CHILD(513), CHILD(513), CHILD(513), CHILD(513),
+  CHILD(513), CHILD(513), CHILD(513), CHILD(513), CHILD(513), CHILD(513),
+  CHILD(513), CHILD(513), CHILD(513), CHILD(513), CHILD(513), CHILD(513),
+  CHILD(513), CHILD(513), CHILD(513), CHILD(513), CHILD(513), CHILD(513),
+  CHILD(513), CHILD(513), CHILD(513), CHILD(513), CHILD(513), CHILD(513),
+  CHILD(513), CHILD(513), CHILD(513), CHILD(513), TILEGX_OPC_ADDXLI,
+  TILEGX_OPC_ADDXLI, TILEGX_OPC_ADDXLI, TILEGX_OPC_ADDXLI, TILEGX_OPC_ADDXLI,
+  TILEGX_OPC_ADDXLI, TILEGX_OPC_ADDXLI, TILEGX_OPC_ADDXLI, TILEGX_OPC_ADDXLI,
+  TILEGX_OPC_ADDXLI, TILEGX_OPC_ADDXLI, TILEGX_OPC_ADDXLI, TILEGX_OPC_ADDXLI,
+  TILEGX_OPC_ADDXLI, TILEGX_OPC_ADDXLI, TILEGX_OPC_ADDXLI, TILEGX_OPC_ADDXLI,
+  TILEGX_OPC_ADDXLI, TILEGX_OPC_ADDXLI, TILEGX_OPC_ADDXLI, TILEGX_OPC_ADDXLI,
+  TILEGX_OPC_ADDXLI, TILEGX_OPC_ADDXLI, TILEGX_OPC_ADDXLI, TILEGX_OPC_ADDXLI,
+  TILEGX_OPC_ADDXLI, TILEGX_OPC_ADDXLI, TILEGX_OPC_ADDXLI, TILEGX_OPC_ADDXLI,
+  TILEGX_OPC_ADDXLI, TILEGX_OPC_ADDXLI, TILEGX_OPC_ADDXLI, TILEGX_OPC_ADDXLI,
+  TILEGX_OPC_ADDXLI, TILEGX_OPC_ADDXLI, TILEGX_OPC_ADDXLI, TILEGX_OPC_ADDXLI,
+  TILEGX_OPC_ADDXLI, TILEGX_OPC_ADDXLI, TILEGX_OPC_ADDXLI, TILEGX_OPC_ADDXLI,
+  TILEGX_OPC_ADDXLI, TILEGX_OPC_ADDXLI, TILEGX_OPC_ADDXLI, TILEGX_OPC_ADDXLI,
+  TILEGX_OPC_ADDXLI, TILEGX_OPC_ADDXLI, TILEGX_OPC_ADDXLI, TILEGX_OPC_ADDXLI,
+  TILEGX_OPC_ADDXLI, TILEGX_OPC_ADDXLI, TILEGX_OPC_ADDXLI, TILEGX_OPC_ADDXLI,
+  TILEGX_OPC_ADDXLI, TILEGX_OPC_ADDXLI, TILEGX_OPC_ADDXLI, TILEGX_OPC_ADDXLI,
+  TILEGX_OPC_ADDXLI, TILEGX_OPC_ADDXLI, TILEGX_OPC_ADDXLI, TILEGX_OPC_ADDXLI,
+  TILEGX_OPC_ADDXLI, TILEGX_OPC_ADDXLI, TILEGX_OPC_ADDXLI, TILEGX_OPC_NONE,
+  TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE,
+  TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE,
+  TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE,
+  TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_BFEXTS,
+  TILEGX_OPC_BFEXTS, TILEGX_OPC_BFEXTS, TILEGX_OPC_BFEXTS, TILEGX_OPC_BFEXTU,
+  TILEGX_OPC_BFEXTU, TILEGX_OPC_BFEXTU, TILEGX_OPC_BFEXTU, TILEGX_OPC_BFINS,
+  TILEGX_OPC_BFINS, TILEGX_OPC_BFINS, TILEGX_OPC_BFINS, TILEGX_OPC_MM,
+  TILEGX_OPC_MM, TILEGX_OPC_MM, TILEGX_OPC_MM, TILEGX_OPC_NONE,
+  TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE,
+  TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE,
+  TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE,
+  TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE,
+  TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE,
+  TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE,
+  TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE,
+  TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE, CHILD(528), CHILD(578),
+  CHILD(583), CHILD(588), CHILD(593), CHILD(598), TILEGX_OPC_NONE,
+  TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE,
+  TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE,
+  TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE,
+  TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE,
+  TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE,
+  TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE,
+  TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE,
+  TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE,
+  TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE,
+  TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE,
+  TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE,
+  TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE,
+  TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE,
+  TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE,
+  TILEGX_OPC_NONE, CHILD(603), CHILD(620), CHILD(637), CHILD(654), CHILD(671),
+  CHILD(703), CHILD(797), CHILD(814), CHILD(831), CHILD(848), CHILD(865),
+  TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE,
+  TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE,
+  TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE,
+  TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE,
+  TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE,
+  TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE,
+  TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE,
+  TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE,
+  TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE,
+  TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE,
+  TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE,
+  TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE,
+  TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE,
+  TILEGX_OPC_NONE, CHILD(889), TILEGX_OPC_NONE, TILEGX_OPC_NONE,
+  TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE,
+  TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE,
+  TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE,
+  TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE,
+  TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE,
+  TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE,
+  TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE,
+  TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE,
+  TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE,
+  TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE,
+  TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE,
+  TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE,
+  TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE,
+  TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE,
+  TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE,
+  TILEGX_OPC_NONE, CHILD(906), CHILD(906), CHILD(906), CHILD(906), CHILD(906),
+  CHILD(906), CHILD(906), CHILD(906), CHILD(906), CHILD(906), CHILD(906),
+  CHILD(906), CHILD(906), CHILD(906), CHILD(906), CHILD(906), CHILD(906),
+  CHILD(906), CHILD(906), CHILD(906), CHILD(906), CHILD(906), CHILD(906),
+  CHILD(906), CHILD(906), CHILD(906), CHILD(906), CHILD(906), CHILD(906),
+  CHILD(906), CHILD(906), CHILD(906), CHILD(906), CHILD(906), CHILD(906),
+  CHILD(906), CHILD(906), CHILD(906), CHILD(906), CHILD(906), CHILD(906),
+  CHILD(906), CHILD(906), CHILD(906), CHILD(906), CHILD(906), CHILD(906),
+  CHILD(906), CHILD(906), CHILD(906), CHILD(906), CHILD(906), CHILD(906),
+  CHILD(906), CHILD(906), CHILD(906), CHILD(906), CHILD(906), CHILD(906),
+  CHILD(906), CHILD(906), CHILD(906), CHILD(906), CHILD(906),
+  BITFIELD(6, 2) /* index 513 */,
+  TILEGX_OPC_ADDLI, TILEGX_OPC_ADDLI, TILEGX_OPC_ADDLI, CHILD(518),
+  BITFIELD(8, 2) /* index 518 */,
+  TILEGX_OPC_ADDLI, TILEGX_OPC_ADDLI, TILEGX_OPC_ADDLI, CHILD(523),
+  BITFIELD(10, 2) /* index 523 */,
+  TILEGX_OPC_ADDLI, TILEGX_OPC_ADDLI, TILEGX_OPC_ADDLI, TILEGX_OPC_MOVELI,
+  BITFIELD(20, 2) /* index 528 */,
+  TILEGX_OPC_NONE, CHILD(533), TILEGX_OPC_ADDXI, CHILD(548),
+  BITFIELD(6, 2) /* index 533 */,
+  TILEGX_OPC_ADDI, TILEGX_OPC_ADDI, TILEGX_OPC_ADDI, CHILD(538),
+  BITFIELD(8, 2) /* index 538 */,
+  TILEGX_OPC_ADDI, TILEGX_OPC_ADDI, TILEGX_OPC_ADDI, CHILD(543),
+  BITFIELD(10, 2) /* index 543 */,
+  TILEGX_OPC_ADDI, TILEGX_OPC_ADDI, TILEGX_OPC_ADDI, TILEGX_OPC_MOVEI,
+  BITFIELD(0, 2) /* index 548 */,
+  TILEGX_OPC_ANDI, TILEGX_OPC_ANDI, TILEGX_OPC_ANDI, CHILD(553),
+  BITFIELD(2, 2) /* index 553 */,
+  TILEGX_OPC_ANDI, TILEGX_OPC_ANDI, TILEGX_OPC_ANDI, CHILD(558),
+  BITFIELD(4, 2) /* index 558 */,
+  TILEGX_OPC_ANDI, TILEGX_OPC_ANDI, TILEGX_OPC_ANDI, CHILD(563),
+  BITFIELD(6, 2) /* index 563 */,
+  TILEGX_OPC_ANDI, TILEGX_OPC_ANDI, TILEGX_OPC_ANDI, CHILD(568),
+  BITFIELD(8, 2) /* index 568 */,
+  TILEGX_OPC_ANDI, TILEGX_OPC_ANDI, TILEGX_OPC_ANDI, CHILD(573),
+  BITFIELD(10, 2) /* index 573 */,
+  TILEGX_OPC_ANDI, TILEGX_OPC_ANDI, TILEGX_OPC_ANDI, TILEGX_OPC_INFO,
+  BITFIELD(20, 2) /* index 578 */,
+  TILEGX_OPC_CMPEQI, TILEGX_OPC_CMPLTSI, TILEGX_OPC_CMPLTUI, TILEGX_OPC_ORI,
+  BITFIELD(20, 2) /* index 583 */,
+  TILEGX_OPC_V1ADDI, TILEGX_OPC_V1CMPEQI, TILEGX_OPC_V1CMPLTSI,
+  TILEGX_OPC_V1CMPLTUI,
+  BITFIELD(20, 2) /* index 588 */,
+  TILEGX_OPC_V1MAXUI, TILEGX_OPC_V1MINUI, TILEGX_OPC_V2ADDI,
+  TILEGX_OPC_V2CMPEQI,
+  BITFIELD(20, 2) /* index 593 */,
+  TILEGX_OPC_V2CMPLTSI, TILEGX_OPC_V2CMPLTUI, TILEGX_OPC_V2MAXSI,
+  TILEGX_OPC_V2MINSI,
+  BITFIELD(20, 2) /* index 598 */,
+  TILEGX_OPC_XORI, TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE,
+  BITFIELD(18, 4) /* index 603 */,
+  TILEGX_OPC_NONE, TILEGX_OPC_ADDXSC, TILEGX_OPC_ADDX, TILEGX_OPC_ADD,
+  TILEGX_OPC_AND, TILEGX_OPC_CMOVEQZ, TILEGX_OPC_CMOVNEZ, TILEGX_OPC_CMPEQ,
+  TILEGX_OPC_CMPLES, TILEGX_OPC_CMPLEU, TILEGX_OPC_CMPLTS, TILEGX_OPC_CMPLTU,
+  TILEGX_OPC_CMPNE, TILEGX_OPC_CMULAF, TILEGX_OPC_CMULA, TILEGX_OPC_CMULFR,
+  BITFIELD(18, 4) /* index 620 */,
+  TILEGX_OPC_CMULF, TILEGX_OPC_CMULHR, TILEGX_OPC_CMULH, TILEGX_OPC_CMUL,
+  TILEGX_OPC_CRC32_32, TILEGX_OPC_CRC32_8, TILEGX_OPC_DBLALIGN2,
+  TILEGX_OPC_DBLALIGN4, TILEGX_OPC_DBLALIGN6, TILEGX_OPC_DBLALIGN,
+  TILEGX_OPC_FDOUBLE_ADDSUB, TILEGX_OPC_FDOUBLE_ADD_FLAGS,
+  TILEGX_OPC_FDOUBLE_MUL_FLAGS, TILEGX_OPC_FDOUBLE_PACK1,
+  TILEGX_OPC_FDOUBLE_PACK2, TILEGX_OPC_FDOUBLE_SUB_FLAGS,
+  BITFIELD(18, 4) /* index 637 */,
+  TILEGX_OPC_FDOUBLE_UNPACK_MAX, TILEGX_OPC_FDOUBLE_UNPACK_MIN,
+  TILEGX_OPC_FSINGLE_ADD1, TILEGX_OPC_FSINGLE_ADDSUB2,
+  TILEGX_OPC_FSINGLE_MUL1, TILEGX_OPC_FSINGLE_MUL2, TILEGX_OPC_FSINGLE_PACK2,
+  TILEGX_OPC_FSINGLE_SUB1, TILEGX_OPC_MNZ, TILEGX_OPC_MULAX,
+  TILEGX_OPC_MULA_HS_HS, TILEGX_OPC_MULA_HS_HU, TILEGX_OPC_MULA_HS_LS,
+  TILEGX_OPC_MULA_HS_LU, TILEGX_OPC_MULA_HU_HU, TILEGX_OPC_MULA_HU_LS,
+  BITFIELD(18, 4) /* index 654 */,
+  TILEGX_OPC_MULA_HU_LU, TILEGX_OPC_MULA_LS_LS, TILEGX_OPC_MULA_LS_LU,
+  TILEGX_OPC_MULA_LU_LU, TILEGX_OPC_MULX, TILEGX_OPC_MUL_HS_HS,
+  TILEGX_OPC_MUL_HS_HU, TILEGX_OPC_MUL_HS_LS, TILEGX_OPC_MUL_HS_LU,
+  TILEGX_OPC_MUL_HU_HU, TILEGX_OPC_MUL_HU_LS, TILEGX_OPC_MUL_HU_LU,
+  TILEGX_OPC_MUL_LS_LS, TILEGX_OPC_MUL_LS_LU, TILEGX_OPC_MUL_LU_LU,
+  TILEGX_OPC_MZ,
+  BITFIELD(18, 4) /* index 671 */,
+  TILEGX_OPC_NOR, CHILD(688), TILEGX_OPC_ROTL, TILEGX_OPC_SHL1ADDX,
+  TILEGX_OPC_SHL1ADD, TILEGX_OPC_SHL2ADDX, TILEGX_OPC_SHL2ADD,
+  TILEGX_OPC_SHL3ADDX, TILEGX_OPC_SHL3ADD, TILEGX_OPC_SHLX, TILEGX_OPC_SHL,
+  TILEGX_OPC_SHRS, TILEGX_OPC_SHRUX, TILEGX_OPC_SHRU, TILEGX_OPC_SHUFFLEBYTES,
+  TILEGX_OPC_SUBXSC,
+  BITFIELD(12, 2) /* index 688 */,
+  TILEGX_OPC_OR, TILEGX_OPC_OR, TILEGX_OPC_OR, CHILD(693),
+  BITFIELD(14, 2) /* index 693 */,
+  TILEGX_OPC_OR, TILEGX_OPC_OR, TILEGX_OPC_OR, CHILD(698),
+  BITFIELD(16, 2) /* index 698 */,
+  TILEGX_OPC_OR, TILEGX_OPC_OR, TILEGX_OPC_OR, TILEGX_OPC_MOVE,
+  BITFIELD(18, 4) /* index 703 */,
+  TILEGX_OPC_SUBX, TILEGX_OPC_SUB, CHILD(720), TILEGX_OPC_V1ADDUC,
+  TILEGX_OPC_V1ADD, TILEGX_OPC_V1ADIFFU, TILEGX_OPC_V1AVGU,
+  TILEGX_OPC_V1CMPEQ, TILEGX_OPC_V1CMPLES, TILEGX_OPC_V1CMPLEU,
+  TILEGX_OPC_V1CMPLTS, TILEGX_OPC_V1CMPLTU, TILEGX_OPC_V1CMPNE,
+  TILEGX_OPC_V1DDOTPUSA, TILEGX_OPC_V1DDOTPUS, TILEGX_OPC_V1DOTPA,
+  BITFIELD(12, 4) /* index 720 */,
+  TILEGX_OPC_NONE, CHILD(737), CHILD(742), CHILD(747), CHILD(752), CHILD(757),
+  CHILD(762), CHILD(767), CHILD(772), CHILD(777), CHILD(782), CHILD(787),
+  CHILD(792), TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE,
+  BITFIELD(16, 2) /* index 737 */,
+  TILEGX_OPC_CLZ, TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE,
+  BITFIELD(16, 2) /* index 742 */,
+  TILEGX_OPC_CTZ, TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE,
+  BITFIELD(16, 2) /* index 747 */,
+  TILEGX_OPC_FNOP, TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE,
+  BITFIELD(16, 2) /* index 752 */,
+  TILEGX_OPC_FSINGLE_PACK1, TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE,
+  BITFIELD(16, 2) /* index 757 */,
+  TILEGX_OPC_NOP, TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE,
+  BITFIELD(16, 2) /* index 762 */,
+  TILEGX_OPC_PCNT, TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE,
+  BITFIELD(16, 2) /* index 767 */,
+  TILEGX_OPC_REVBITS, TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE,
+  BITFIELD(16, 2) /* index 772 */,
+  TILEGX_OPC_REVBYTES, TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE,
+  BITFIELD(16, 2) /* index 777 */,
+  TILEGX_OPC_TBLIDXB0, TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE,
+  BITFIELD(16, 2) /* index 782 */,
+  TILEGX_OPC_TBLIDXB1, TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE,
+  BITFIELD(16, 2) /* index 787 */,
+  TILEGX_OPC_TBLIDXB2, TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE,
+  BITFIELD(16, 2) /* index 792 */,
+  TILEGX_OPC_TBLIDXB3, TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE,
+  BITFIELD(18, 4) /* index 797 */,
+  TILEGX_OPC_V1DOTPUSA, TILEGX_OPC_V1DOTPUS, TILEGX_OPC_V1DOTP,
+  TILEGX_OPC_V1INT_H, TILEGX_OPC_V1INT_L, TILEGX_OPC_V1MAXU,
+  TILEGX_OPC_V1MINU, TILEGX_OPC_V1MNZ, TILEGX_OPC_V1MULTU, TILEGX_OPC_V1MULUS,
+  TILEGX_OPC_V1MULU, TILEGX_OPC_V1MZ, TILEGX_OPC_V1SADAU, TILEGX_OPC_V1SADU,
+  TILEGX_OPC_V1SHL, TILEGX_OPC_V1SHRS,
+  BITFIELD(18, 4) /* index 814 */,
+  TILEGX_OPC_V1SHRU, TILEGX_OPC_V1SUBUC, TILEGX_OPC_V1SUB, TILEGX_OPC_V2ADDSC,
+  TILEGX_OPC_V2ADD, TILEGX_OPC_V2ADIFFS, TILEGX_OPC_V2AVGS,
+  TILEGX_OPC_V2CMPEQ, TILEGX_OPC_V2CMPLES, TILEGX_OPC_V2CMPLEU,
+  TILEGX_OPC_V2CMPLTS, TILEGX_OPC_V2CMPLTU, TILEGX_OPC_V2CMPNE,
+  TILEGX_OPC_V2DOTPA, TILEGX_OPC_V2DOTP, TILEGX_OPC_V2INT_H,
+  BITFIELD(18, 4) /* index 831 */,
+  TILEGX_OPC_V2INT_L, TILEGX_OPC_V2MAXS, TILEGX_OPC_V2MINS, TILEGX_OPC_V2MNZ,
+  TILEGX_OPC_V2MULFSC, TILEGX_OPC_V2MULS, TILEGX_OPC_V2MULTS, TILEGX_OPC_V2MZ,
+  TILEGX_OPC_V2PACKH, TILEGX_OPC_V2PACKL, TILEGX_OPC_V2PACKUC,
+  TILEGX_OPC_V2SADAS, TILEGX_OPC_V2SADAU, TILEGX_OPC_V2SADS,
+  TILEGX_OPC_V2SADU, TILEGX_OPC_V2SHLSC,
+  BITFIELD(18, 4) /* index 848 */,
+  TILEGX_OPC_V2SHL, TILEGX_OPC_V2SHRS, TILEGX_OPC_V2SHRU, TILEGX_OPC_V2SUBSC,
+  TILEGX_OPC_V2SUB, TILEGX_OPC_V4ADDSC, TILEGX_OPC_V4ADD, TILEGX_OPC_V4INT_H,
+  TILEGX_OPC_V4INT_L, TILEGX_OPC_V4PACKSC, TILEGX_OPC_V4SHLSC,
+  TILEGX_OPC_V4SHL, TILEGX_OPC_V4SHRS, TILEGX_OPC_V4SHRU, TILEGX_OPC_V4SUBSC,
+  TILEGX_OPC_V4SUB,
+  BITFIELD(18, 3) /* index 865 */,
+  CHILD(874), CHILD(877), CHILD(880), CHILD(883), CHILD(886), TILEGX_OPC_NONE,
+  TILEGX_OPC_NONE, TILEGX_OPC_NONE,
+  BITFIELD(21, 1) /* index 874 */,
+  TILEGX_OPC_XOR, TILEGX_OPC_NONE,
+  BITFIELD(21, 1) /* index 877 */,
+  TILEGX_OPC_V1DDOTPUA, TILEGX_OPC_NONE,
+  BITFIELD(21, 1) /* index 880 */,
+  TILEGX_OPC_V1DDOTPU, TILEGX_OPC_NONE,
+  BITFIELD(21, 1) /* index 883 */,
+  TILEGX_OPC_V1DOTPUA, TILEGX_OPC_NONE,
+  BITFIELD(21, 1) /* index 886 */,
+  TILEGX_OPC_V1DOTPU, TILEGX_OPC_NONE,
+  BITFIELD(18, 4) /* index 889 */,
+  TILEGX_OPC_NONE, TILEGX_OPC_ROTLI, TILEGX_OPC_SHLI, TILEGX_OPC_SHLXI,
+  TILEGX_OPC_SHRSI, TILEGX_OPC_SHRUI, TILEGX_OPC_SHRUXI, TILEGX_OPC_V1SHLI,
+  TILEGX_OPC_V1SHRSI, TILEGX_OPC_V1SHRUI, TILEGX_OPC_V2SHLI,
+  TILEGX_OPC_V2SHRSI, TILEGX_OPC_V2SHRUI, TILEGX_OPC_NONE, TILEGX_OPC_NONE,
+  TILEGX_OPC_NONE,
+  BITFIELD(0, 2) /* index 906 */,
+  TILEGX_OPC_SHL16INSLI, TILEGX_OPC_SHL16INSLI, TILEGX_OPC_SHL16INSLI,
+  CHILD(911),
+  BITFIELD(2, 2) /* index 911 */,
+  TILEGX_OPC_SHL16INSLI, TILEGX_OPC_SHL16INSLI, TILEGX_OPC_SHL16INSLI,
+  CHILD(916),
+  BITFIELD(4, 2) /* index 916 */,
+  TILEGX_OPC_SHL16INSLI, TILEGX_OPC_SHL16INSLI, TILEGX_OPC_SHL16INSLI,
+  CHILD(921),
+  BITFIELD(6, 2) /* index 921 */,
+  TILEGX_OPC_SHL16INSLI, TILEGX_OPC_SHL16INSLI, TILEGX_OPC_SHL16INSLI,
+  CHILD(926),
+  BITFIELD(8, 2) /* index 926 */,
+  TILEGX_OPC_SHL16INSLI, TILEGX_OPC_SHL16INSLI, TILEGX_OPC_SHL16INSLI,
+  CHILD(931),
+  BITFIELD(10, 2) /* index 931 */,
+  TILEGX_OPC_SHL16INSLI, TILEGX_OPC_SHL16INSLI, TILEGX_OPC_SHL16INSLI,
+  TILEGX_OPC_INFOL,
+};
+
+static const unsigned short decode_X1_fsm[1206] =
+{
+  BITFIELD(53, 9) /* index 0 */,
+  CHILD(513), CHILD(513), CHILD(513), CHILD(513), CHILD(513), CHILD(513),
+  CHILD(513), CHILD(513), CHILD(513), CHILD(513), CHILD(513), CHILD(513),
+  CHILD(513), CHILD(513), CHILD(513), CHILD(513), CHILD(513), CHILD(513),
+  CHILD(513), CHILD(513), CHILD(513), CHILD(513), CHILD(513), CHILD(513),
+  CHILD(513), CHILD(513), CHILD(513), CHILD(513), CHILD(513), CHILD(513),
+  CHILD(513), CHILD(513), CHILD(513), CHILD(513), CHILD(513), CHILD(513),
+  CHILD(513), CHILD(513), CHILD(513), CHILD(513), CHILD(513), CHILD(513),
+  CHILD(513), CHILD(513), CHILD(513), CHILD(513), CHILD(513), CHILD(513),
+  CHILD(513), CHILD(513), CHILD(513), CHILD(513), CHILD(513), CHILD(513),
+  CHILD(513), CHILD(513), CHILD(513), CHILD(513), CHILD(513), CHILD(513),
+  CHILD(513), CHILD(513), CHILD(513), CHILD(513), TILEGX_OPC_ADDXLI,
+  TILEGX_OPC_ADDXLI, TILEGX_OPC_ADDXLI, TILEGX_OPC_ADDXLI, TILEGX_OPC_ADDXLI,
+  TILEGX_OPC_ADDXLI, TILEGX_OPC_ADDXLI, TILEGX_OPC_ADDXLI, TILEGX_OPC_ADDXLI,
+  TILEGX_OPC_ADDXLI, TILEGX_OPC_ADDXLI, TILEGX_OPC_ADDXLI, TILEGX_OPC_ADDXLI,
+  TILEGX_OPC_ADDXLI, TILEGX_OPC_ADDXLI, TILEGX_OPC_ADDXLI, TILEGX_OPC_ADDXLI,
+  TILEGX_OPC_ADDXLI, TILEGX_OPC_ADDXLI, TILEGX_OPC_ADDXLI, TILEGX_OPC_ADDXLI,
+  TILEGX_OPC_ADDXLI, TILEGX_OPC_ADDXLI, TILEGX_OPC_ADDXLI, TILEGX_OPC_ADDXLI,
+  TILEGX_OPC_ADDXLI, TILEGX_OPC_ADDXLI, TILEGX_OPC_ADDXLI, TILEGX_OPC_ADDXLI,
+  TILEGX_OPC_ADDXLI, TILEGX_OPC_ADDXLI, TILEGX_OPC_ADDXLI, TILEGX_OPC_ADDXLI,
+  TILEGX_OPC_ADDXLI, TILEGX_OPC_ADDXLI, TILEGX_OPC_ADDXLI, TILEGX_OPC_ADDXLI,
+  TILEGX_OPC_ADDXLI, TILEGX_OPC_ADDXLI, TILEGX_OPC_ADDXLI, TILEGX_OPC_ADDXLI,
+  TILEGX_OPC_ADDXLI, TILEGX_OPC_ADDXLI, TILEGX_OPC_ADDXLI, TILEGX_OPC_ADDXLI,
+  TILEGX_OPC_ADDXLI, TILEGX_OPC_ADDXLI, TILEGX_OPC_ADDXLI, TILEGX_OPC_ADDXLI,
+  TILEGX_OPC_ADDXLI, TILEGX_OPC_ADDXLI, TILEGX_OPC_ADDXLI, TILEGX_OPC_ADDXLI,
+  TILEGX_OPC_ADDXLI, TILEGX_OPC_ADDXLI, TILEGX_OPC_ADDXLI, TILEGX_OPC_ADDXLI,
+  TILEGX_OPC_ADDXLI, TILEGX_OPC_ADDXLI, TILEGX_OPC_ADDXLI, TILEGX_OPC_ADDXLI,
+  TILEGX_OPC_ADDXLI, TILEGX_OPC_ADDXLI, TILEGX_OPC_ADDXLI, TILEGX_OPC_NONE,
+  TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE,
+  TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE,
+  TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE,
+  TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE,
+  TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE,
+  TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE,
+  TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE,
+  TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_BEQZT,
+  TILEGX_OPC_BEQZT, TILEGX_OPC_BEQZ, TILEGX_OPC_BEQZ, TILEGX_OPC_BGEZT,
+  TILEGX_OPC_BGEZT, TILEGX_OPC_BGEZ, TILEGX_OPC_BGEZ, TILEGX_OPC_BGTZT,
+  TILEGX_OPC_BGTZT, TILEGX_OPC_BGTZ, TILEGX_OPC_BGTZ, TILEGX_OPC_BLBCT,
+  TILEGX_OPC_BLBCT, TILEGX_OPC_BLBC, TILEGX_OPC_BLBC, TILEGX_OPC_BLBST,
+  TILEGX_OPC_BLBST, TILEGX_OPC_BLBS, TILEGX_OPC_BLBS, TILEGX_OPC_BLEZT,
+  TILEGX_OPC_BLEZT, TILEGX_OPC_BLEZ, TILEGX_OPC_BLEZ, TILEGX_OPC_BLTZT,
+  TILEGX_OPC_BLTZT, TILEGX_OPC_BLTZ, TILEGX_OPC_BLTZ, TILEGX_OPC_BNEZT,
+  TILEGX_OPC_BNEZT, TILEGX_OPC_BNEZ, TILEGX_OPC_BNEZ, CHILD(528), CHILD(578),
+  CHILD(598), CHILD(663), CHILD(683), CHILD(688), CHILD(693), CHILD(698),
+  CHILD(703), CHILD(708), CHILD(713), CHILD(718), TILEGX_OPC_NONE,
+  TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE,
+  TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE,
+  TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE,
+  TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE,
+  TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE,
+  TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE,
+  TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE,
+  TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE,
+  TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE,
+  TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE,
+  TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE,
+  TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE,
+  TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_JAL,
+  TILEGX_OPC_JAL, TILEGX_OPC_JAL, TILEGX_OPC_JAL, TILEGX_OPC_JAL,
+  TILEGX_OPC_JAL, TILEGX_OPC_JAL, TILEGX_OPC_JAL, TILEGX_OPC_JAL,
+  TILEGX_OPC_JAL, TILEGX_OPC_JAL, TILEGX_OPC_JAL, TILEGX_OPC_JAL,
+  TILEGX_OPC_JAL, TILEGX_OPC_JAL, TILEGX_OPC_JAL, TILEGX_OPC_JAL,
+  TILEGX_OPC_JAL, TILEGX_OPC_JAL, TILEGX_OPC_JAL, TILEGX_OPC_JAL,
+  TILEGX_OPC_JAL, TILEGX_OPC_JAL, TILEGX_OPC_JAL, TILEGX_OPC_JAL,
+  TILEGX_OPC_JAL, TILEGX_OPC_JAL, TILEGX_OPC_JAL, TILEGX_OPC_JAL,
+  TILEGX_OPC_JAL, TILEGX_OPC_JAL, TILEGX_OPC_JAL, TILEGX_OPC_J, TILEGX_OPC_J,
+  TILEGX_OPC_J, TILEGX_OPC_J, TILEGX_OPC_J, TILEGX_OPC_J, TILEGX_OPC_J,
+  TILEGX_OPC_J, TILEGX_OPC_J, TILEGX_OPC_J, TILEGX_OPC_J, TILEGX_OPC_J,
+  TILEGX_OPC_J, TILEGX_OPC_J, TILEGX_OPC_J, TILEGX_OPC_J, TILEGX_OPC_J,
+  TILEGX_OPC_J, TILEGX_OPC_J, TILEGX_OPC_J, TILEGX_OPC_J, TILEGX_OPC_J,
+  TILEGX_OPC_J, TILEGX_OPC_J, TILEGX_OPC_J, TILEGX_OPC_J, TILEGX_OPC_J,
+  TILEGX_OPC_J, TILEGX_OPC_J, TILEGX_OPC_J, TILEGX_OPC_J, TILEGX_OPC_J,
+  CHILD(723), CHILD(740), CHILD(772), CHILD(789), CHILD(1108), CHILD(1125),
+  CHILD(1142), TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE,
+  TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE,
+  TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE,
+  TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE,
+  TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE,
+  TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE,
+  TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE,
+  TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE,
+  TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE,
+  TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE,
+  TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE,
+  TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE,
+  TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE,
+  TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE,
+  TILEGX_OPC_NONE, TILEGX_OPC_NONE, CHILD(1159), TILEGX_OPC_NONE,
+  TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE,
+  TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE,
+  TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE,
+  TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE,
+  TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE,
+  TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE,
+  TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE,
+  TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE,
+  TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE,
+  TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE,
+  TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE,
+  TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE,
+  TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE,
+  TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE,
+  TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE,
+  TILEGX_OPC_NONE, TILEGX_OPC_NONE, CHILD(1176), CHILD(1176), CHILD(1176),
+  CHILD(1176), CHILD(1176), CHILD(1176), CHILD(1176), CHILD(1176),
+  CHILD(1176), CHILD(1176), CHILD(1176), CHILD(1176), CHILD(1176),
+  CHILD(1176), CHILD(1176), CHILD(1176), CHILD(1176), CHILD(1176),
+  CHILD(1176), CHILD(1176), CHILD(1176), CHILD(1176), CHILD(1176),
+  CHILD(1176), CHILD(1176), CHILD(1176), CHILD(1176), CHILD(1176),
+  CHILD(1176), CHILD(1176), CHILD(1176), CHILD(1176), CHILD(1176),
+  CHILD(1176), CHILD(1176), CHILD(1176), CHILD(1176), CHILD(1176),
+  CHILD(1176), CHILD(1176), CHILD(1176), CHILD(1176), CHILD(1176),
+  CHILD(1176), CHILD(1176), CHILD(1176), CHILD(1176), CHILD(1176),
+  CHILD(1176), CHILD(1176), CHILD(1176), CHILD(1176), CHILD(1176),
+  CHILD(1176), CHILD(1176), CHILD(1176), CHILD(1176), CHILD(1176),
+  CHILD(1176), CHILD(1176), CHILD(1176), CHILD(1176), CHILD(1176),
+  CHILD(1176),
+  BITFIELD(37, 2) /* index 513 */,
+  TILEGX_OPC_ADDLI, TILEGX_OPC_ADDLI, TILEGX_OPC_ADDLI, CHILD(518),
+  BITFIELD(39, 2) /* index 518 */,
+  TILEGX_OPC_ADDLI, TILEGX_OPC_ADDLI, TILEGX_OPC_ADDLI, CHILD(523),
+  BITFIELD(41, 2) /* index 523 */,
+  TILEGX_OPC_ADDLI, TILEGX_OPC_ADDLI, TILEGX_OPC_ADDLI, TILEGX_OPC_MOVELI,
+  BITFIELD(51, 2) /* index 528 */,
+  TILEGX_OPC_NONE, CHILD(533), TILEGX_OPC_ADDXI, CHILD(548),
+  BITFIELD(37, 2) /* index 533 */,
+  TILEGX_OPC_ADDI, TILEGX_OPC_ADDI, TILEGX_OPC_ADDI, CHILD(538),
+  BITFIELD(39, 2) /* index 538 */,
+  TILEGX_OPC_ADDI, TILEGX_OPC_ADDI, TILEGX_OPC_ADDI, CHILD(543),
+  BITFIELD(41, 2) /* index 543 */,
+  TILEGX_OPC_ADDI, TILEGX_OPC_ADDI, TILEGX_OPC_ADDI, TILEGX_OPC_MOVEI,
+  BITFIELD(31, 2) /* index 548 */,
+  TILEGX_OPC_ANDI, TILEGX_OPC_ANDI, TILEGX_OPC_ANDI, CHILD(553),
+  BITFIELD(33, 2) /* index 553 */,
+  TILEGX_OPC_ANDI, TILEGX_OPC_ANDI, TILEGX_OPC_ANDI, CHILD(558),
+  BITFIELD(35, 2) /* index 558 */,
+  TILEGX_OPC_ANDI, TILEGX_OPC_ANDI, TILEGX_OPC_ANDI, CHILD(563),
+  BITFIELD(37, 2) /* index 563 */,
+  TILEGX_OPC_ANDI, TILEGX_OPC_ANDI, TILEGX_OPC_ANDI, CHILD(568),
+  BITFIELD(39, 2) /* index 568 */,
+  TILEGX_OPC_ANDI, TILEGX_OPC_ANDI, TILEGX_OPC_ANDI, CHILD(573),
+  BITFIELD(41, 2) /* index 573 */,
+  TILEGX_OPC_ANDI, TILEGX_OPC_ANDI, TILEGX_OPC_ANDI, TILEGX_OPC_INFO,
+  BITFIELD(51, 2) /* index 578 */,
+  TILEGX_OPC_CMPEQI, TILEGX_OPC_CMPLTSI, TILEGX_OPC_CMPLTUI, CHILD(583),
+  BITFIELD(31, 2) /* index 583 */,
+  TILEGX_OPC_LD1S_ADD, TILEGX_OPC_LD1S_ADD, TILEGX_OPC_LD1S_ADD, CHILD(588),
+  BITFIELD(33, 2) /* index 588 */,
+  TILEGX_OPC_LD1S_ADD, TILEGX_OPC_LD1S_ADD, TILEGX_OPC_LD1S_ADD, CHILD(593),
+  BITFIELD(35, 2) /* index 593 */,
+  TILEGX_OPC_LD1S_ADD, TILEGX_OPC_LD1S_ADD, TILEGX_OPC_LD1S_ADD,
+  TILEGX_OPC_PREFETCH_ADD_L1_FAULT,
+  BITFIELD(51, 2) /* index 598 */,
+  CHILD(603), CHILD(618), CHILD(633), CHILD(648),
+  BITFIELD(31, 2) /* index 603 */,
+  TILEGX_OPC_LD1U_ADD, TILEGX_OPC_LD1U_ADD, TILEGX_OPC_LD1U_ADD, CHILD(608),
+  BITFIELD(33, 2) /* index 608 */,
+  TILEGX_OPC_LD1U_ADD, TILEGX_OPC_LD1U_ADD, TILEGX_OPC_LD1U_ADD, CHILD(613),
+  BITFIELD(35, 2) /* index 613 */,
+  TILEGX_OPC_LD1U_ADD, TILEGX_OPC_LD1U_ADD, TILEGX_OPC_LD1U_ADD,
+  TILEGX_OPC_PREFETCH_ADD_L1,
+  BITFIELD(31, 2) /* index 618 */,
+  TILEGX_OPC_LD2S_ADD, TILEGX_OPC_LD2S_ADD, TILEGX_OPC_LD2S_ADD, CHILD(623),
+  BITFIELD(33, 2) /* index 623 */,
+  TILEGX_OPC_LD2S_ADD, TILEGX_OPC_LD2S_ADD, TILEGX_OPC_LD2S_ADD, CHILD(628),
+  BITFIELD(35, 2) /* index 628 */,
+  TILEGX_OPC_LD2S_ADD, TILEGX_OPC_LD2S_ADD, TILEGX_OPC_LD2S_ADD,
+  TILEGX_OPC_PREFETCH_ADD_L2_FAULT,
+  BITFIELD(31, 2) /* index 633 */,
+  TILEGX_OPC_LD2U_ADD, TILEGX_OPC_LD2U_ADD, TILEGX_OPC_LD2U_ADD, CHILD(638),
+  BITFIELD(33, 2) /* index 638 */,
+  TILEGX_OPC_LD2U_ADD, TILEGX_OPC_LD2U_ADD, TILEGX_OPC_LD2U_ADD, CHILD(643),
+  BITFIELD(35, 2) /* index 643 */,
+  TILEGX_OPC_LD2U_ADD, TILEGX_OPC_LD2U_ADD, TILEGX_OPC_LD2U_ADD,
+  TILEGX_OPC_PREFETCH_ADD_L2,
+  BITFIELD(31, 2) /* index 648 */,
+  TILEGX_OPC_LD4S_ADD, TILEGX_OPC_LD4S_ADD, TILEGX_OPC_LD4S_ADD, CHILD(653),
+  BITFIELD(33, 2) /* index 653 */,
+  TILEGX_OPC_LD4S_ADD, TILEGX_OPC_LD4S_ADD, TILEGX_OPC_LD4S_ADD, CHILD(658),
+  BITFIELD(35, 2) /* index 658 */,
+  TILEGX_OPC_LD4S_ADD, TILEGX_OPC_LD4S_ADD, TILEGX_OPC_LD4S_ADD,
+  TILEGX_OPC_PREFETCH_ADD_L3_FAULT,
+  BITFIELD(51, 2) /* index 663 */,
+  CHILD(668), TILEGX_OPC_LDNT1S_ADD, TILEGX_OPC_LDNT1U_ADD,
+  TILEGX_OPC_LDNT2S_ADD,
+  BITFIELD(31, 2) /* index 668 */,
+  TILEGX_OPC_LD4U_ADD, TILEGX_OPC_LD4U_ADD, TILEGX_OPC_LD4U_ADD, CHILD(673),
+  BITFIELD(33, 2) /* index 673 */,
+  TILEGX_OPC_LD4U_ADD, TILEGX_OPC_LD4U_ADD, TILEGX_OPC_LD4U_ADD, CHILD(678),
+  BITFIELD(35, 2) /* index 678 */,
+  TILEGX_OPC_LD4U_ADD, TILEGX_OPC_LD4U_ADD, TILEGX_OPC_LD4U_ADD,
+  TILEGX_OPC_PREFETCH_ADD_L3,
+  BITFIELD(51, 2) /* index 683 */,
+  TILEGX_OPC_LDNT2U_ADD, TILEGX_OPC_LDNT4S_ADD, TILEGX_OPC_LDNT4U_ADD,
+  TILEGX_OPC_LDNT_ADD,
+  BITFIELD(51, 2) /* index 688 */,
+  TILEGX_OPC_LD_ADD, TILEGX_OPC_LDNA_ADD, TILEGX_OPC_MFSPR, TILEGX_OPC_MTSPR,
+  BITFIELD(51, 2) /* index 693 */,
+  TILEGX_OPC_ORI, TILEGX_OPC_ST1_ADD, TILEGX_OPC_ST2_ADD, TILEGX_OPC_ST4_ADD,
+  BITFIELD(51, 2) /* index 698 */,
+  TILEGX_OPC_STNT1_ADD, TILEGX_OPC_STNT2_ADD, TILEGX_OPC_STNT4_ADD,
+  TILEGX_OPC_STNT_ADD,
+  BITFIELD(51, 2) /* index 703 */,
+  TILEGX_OPC_ST_ADD, TILEGX_OPC_V1ADDI, TILEGX_OPC_V1CMPEQI,
+  TILEGX_OPC_V1CMPLTSI,
+  BITFIELD(51, 2) /* index 708 */,
+  TILEGX_OPC_V1CMPLTUI, TILEGX_OPC_V1MAXUI, TILEGX_OPC_V1MINUI,
+  TILEGX_OPC_V2ADDI,
+  BITFIELD(51, 2) /* index 713 */,
+  TILEGX_OPC_V2CMPEQI, TILEGX_OPC_V2CMPLTSI, TILEGX_OPC_V2CMPLTUI,
+  TILEGX_OPC_V2MAXSI,
+  BITFIELD(51, 2) /* index 718 */,
+  TILEGX_OPC_V2MINSI, TILEGX_OPC_XORI, TILEGX_OPC_NONE, TILEGX_OPC_NONE,
+  BITFIELD(49, 4) /* index 723 */,
+  TILEGX_OPC_NONE, TILEGX_OPC_ADDXSC, TILEGX_OPC_ADDX, TILEGX_OPC_ADD,
+  TILEGX_OPC_AND, TILEGX_OPC_CMPEQ, TILEGX_OPC_CMPEXCH4, TILEGX_OPC_CMPEXCH,
+  TILEGX_OPC_CMPLES, TILEGX_OPC_CMPLEU, TILEGX_OPC_CMPLTS, TILEGX_OPC_CMPLTU,
+  TILEGX_OPC_CMPNE, TILEGX_OPC_DBLALIGN2, TILEGX_OPC_DBLALIGN4,
+  TILEGX_OPC_DBLALIGN6,
+  BITFIELD(49, 4) /* index 740 */,
+  TILEGX_OPC_EXCH4, TILEGX_OPC_EXCH, TILEGX_OPC_FETCHADD4,
+  TILEGX_OPC_FETCHADDGEZ4, TILEGX_OPC_FETCHADDGEZ, TILEGX_OPC_FETCHADD,
+  TILEGX_OPC_FETCHAND4, TILEGX_OPC_FETCHAND, TILEGX_OPC_FETCHOR4,
+  TILEGX_OPC_FETCHOR, TILEGX_OPC_MNZ, TILEGX_OPC_MZ, TILEGX_OPC_NOR,
+  CHILD(757), TILEGX_OPC_ROTL, TILEGX_OPC_SHL1ADDX,
+  BITFIELD(43, 2) /* index 757 */,
+  TILEGX_OPC_OR, TILEGX_OPC_OR, TILEGX_OPC_OR, CHILD(762),
+  BITFIELD(45, 2) /* index 762 */,
+  TILEGX_OPC_OR, TILEGX_OPC_OR, TILEGX_OPC_OR, CHILD(767),
+  BITFIELD(47, 2) /* index 767 */,
+  TILEGX_OPC_OR, TILEGX_OPC_OR, TILEGX_OPC_OR, TILEGX_OPC_MOVE,
+  BITFIELD(49, 4) /* index 772 */,
+  TILEGX_OPC_SHL1ADD, TILEGX_OPC_SHL2ADDX, TILEGX_OPC_SHL2ADD,
+  TILEGX_OPC_SHL3ADDX, TILEGX_OPC_SHL3ADD, TILEGX_OPC_SHLX, TILEGX_OPC_SHL,
+  TILEGX_OPC_SHRS, TILEGX_OPC_SHRUX, TILEGX_OPC_SHRU, TILEGX_OPC_ST1,
+  TILEGX_OPC_ST2, TILEGX_OPC_ST4, TILEGX_OPC_STNT1, TILEGX_OPC_STNT2,
+  TILEGX_OPC_STNT4,
+  BITFIELD(46, 7) /* index 789 */,
+  TILEGX_OPC_STNT, TILEGX_OPC_STNT, TILEGX_OPC_STNT, TILEGX_OPC_STNT,
+  TILEGX_OPC_STNT, TILEGX_OPC_STNT, TILEGX_OPC_STNT, TILEGX_OPC_STNT,
+  TILEGX_OPC_ST, TILEGX_OPC_ST, TILEGX_OPC_ST, TILEGX_OPC_ST, TILEGX_OPC_ST,
+  TILEGX_OPC_ST, TILEGX_OPC_ST, TILEGX_OPC_ST, TILEGX_OPC_SUBXSC,
+  TILEGX_OPC_SUBXSC, TILEGX_OPC_SUBXSC, TILEGX_OPC_SUBXSC, TILEGX_OPC_SUBXSC,
+  TILEGX_OPC_SUBXSC, TILEGX_OPC_SUBXSC, TILEGX_OPC_SUBXSC, TILEGX_OPC_SUBX,
+  TILEGX_OPC_SUBX, TILEGX_OPC_SUBX, TILEGX_OPC_SUBX, TILEGX_OPC_SUBX,
+  TILEGX_OPC_SUBX, TILEGX_OPC_SUBX, TILEGX_OPC_SUBX, TILEGX_OPC_SUB,
+  TILEGX_OPC_SUB, TILEGX_OPC_SUB, TILEGX_OPC_SUB, TILEGX_OPC_SUB,
+  TILEGX_OPC_SUB, TILEGX_OPC_SUB, TILEGX_OPC_SUB, CHILD(918), CHILD(927),
+  CHILD(1006), CHILD(1090), CHILD(1099), TILEGX_OPC_NONE, TILEGX_OPC_NONE,
+  TILEGX_OPC_NONE, TILEGX_OPC_V1ADDUC, TILEGX_OPC_V1ADDUC, TILEGX_OPC_V1ADDUC,
+  TILEGX_OPC_V1ADDUC, TILEGX_OPC_V1ADDUC, TILEGX_OPC_V1ADDUC,
+  TILEGX_OPC_V1ADDUC, TILEGX_OPC_V1ADDUC, TILEGX_OPC_V1ADD, TILEGX_OPC_V1ADD,
+  TILEGX_OPC_V1ADD, TILEGX_OPC_V1ADD, TILEGX_OPC_V1ADD, TILEGX_OPC_V1ADD,
+  TILEGX_OPC_V1ADD, TILEGX_OPC_V1ADD, TILEGX_OPC_V1CMPEQ, TILEGX_OPC_V1CMPEQ,
+  TILEGX_OPC_V1CMPEQ, TILEGX_OPC_V1CMPEQ, TILEGX_OPC_V1CMPEQ,
+  TILEGX_OPC_V1CMPEQ, TILEGX_OPC_V1CMPEQ, TILEGX_OPC_V1CMPEQ,
+  TILEGX_OPC_V1CMPLES, TILEGX_OPC_V1CMPLES, TILEGX_OPC_V1CMPLES,
+  TILEGX_OPC_V1CMPLES, TILEGX_OPC_V1CMPLES, TILEGX_OPC_V1CMPLES,
+  TILEGX_OPC_V1CMPLES, TILEGX_OPC_V1CMPLES, TILEGX_OPC_V1CMPLEU,
+  TILEGX_OPC_V1CMPLEU, TILEGX_OPC_V1CMPLEU, TILEGX_OPC_V1CMPLEU,
+  TILEGX_OPC_V1CMPLEU, TILEGX_OPC_V1CMPLEU, TILEGX_OPC_V1CMPLEU,
+  TILEGX_OPC_V1CMPLEU, TILEGX_OPC_V1CMPLTS, TILEGX_OPC_V1CMPLTS,
+  TILEGX_OPC_V1CMPLTS, TILEGX_OPC_V1CMPLTS, TILEGX_OPC_V1CMPLTS,
+  TILEGX_OPC_V1CMPLTS, TILEGX_OPC_V1CMPLTS, TILEGX_OPC_V1CMPLTS,
+  TILEGX_OPC_V1CMPLTU, TILEGX_OPC_V1CMPLTU, TILEGX_OPC_V1CMPLTU,
+  TILEGX_OPC_V1CMPLTU, TILEGX_OPC_V1CMPLTU, TILEGX_OPC_V1CMPLTU,
+  TILEGX_OPC_V1CMPLTU, TILEGX_OPC_V1CMPLTU, TILEGX_OPC_V1CMPNE,
+  TILEGX_OPC_V1CMPNE, TILEGX_OPC_V1CMPNE, TILEGX_OPC_V1CMPNE,
+  TILEGX_OPC_V1CMPNE, TILEGX_OPC_V1CMPNE, TILEGX_OPC_V1CMPNE,
+  TILEGX_OPC_V1CMPNE, TILEGX_OPC_V1INT_H, TILEGX_OPC_V1INT_H,
+  TILEGX_OPC_V1INT_H, TILEGX_OPC_V1INT_H, TILEGX_OPC_V1INT_H,
+  TILEGX_OPC_V1INT_H, TILEGX_OPC_V1INT_H, TILEGX_OPC_V1INT_H,
+  TILEGX_OPC_V1INT_L, TILEGX_OPC_V1INT_L, TILEGX_OPC_V1INT_L,
+  TILEGX_OPC_V1INT_L, TILEGX_OPC_V1INT_L, TILEGX_OPC_V1INT_L,
+  TILEGX_OPC_V1INT_L, TILEGX_OPC_V1INT_L,
+  BITFIELD(43, 3) /* index 918 */,
+  TILEGX_OPC_NONE, TILEGX_OPC_DRAIN, TILEGX_OPC_DTLBPR, TILEGX_OPC_FINV,
+  TILEGX_OPC_FLUSHWB, TILEGX_OPC_FLUSH, TILEGX_OPC_FNOP, TILEGX_OPC_ICOH,
+  BITFIELD(43, 3) /* index 927 */,
+  CHILD(936), TILEGX_OPC_INV, TILEGX_OPC_IRET, TILEGX_OPC_JALRP,
+  TILEGX_OPC_JALR, TILEGX_OPC_JRP, TILEGX_OPC_JR, CHILD(991),
+  BITFIELD(31, 2) /* index 936 */,
+  CHILD(941), CHILD(966), TILEGX_OPC_ILL, TILEGX_OPC_ILL,
+  BITFIELD(33, 2) /* index 941 */,
+  TILEGX_OPC_ILL, TILEGX_OPC_ILL, TILEGX_OPC_ILL, CHILD(946),
+  BITFIELD(35, 2) /* index 946 */,
+  TILEGX_OPC_ILL, CHILD(951), TILEGX_OPC_ILL, TILEGX_OPC_ILL,
+  BITFIELD(37, 2) /* index 951 */,
+  TILEGX_OPC_ILL, CHILD(956), TILEGX_OPC_ILL, TILEGX_OPC_ILL,
+  BITFIELD(39, 2) /* index 956 */,
+  TILEGX_OPC_ILL, CHILD(961), TILEGX_OPC_ILL, TILEGX_OPC_ILL,
+  BITFIELD(41, 2) /* index 961 */,
+  TILEGX_OPC_ILL, TILEGX_OPC_ILL, TILEGX_OPC_BPT, TILEGX_OPC_ILL,
+  BITFIELD(33, 2) /* index 966 */,
+  TILEGX_OPC_ILL, TILEGX_OPC_ILL, TILEGX_OPC_ILL, CHILD(971),
+  BITFIELD(35, 2) /* index 971 */,
+  TILEGX_OPC_ILL, CHILD(976), TILEGX_OPC_ILL, TILEGX_OPC_ILL,
+  BITFIELD(37, 2) /* index 976 */,
+  TILEGX_OPC_ILL, CHILD(981), TILEGX_OPC_ILL, TILEGX_OPC_ILL,
+  BITFIELD(39, 2) /* index 981 */,
+  TILEGX_OPC_ILL, CHILD(986), TILEGX_OPC_ILL, TILEGX_OPC_ILL,
+  BITFIELD(41, 2) /* index 986 */,
+  TILEGX_OPC_ILL, TILEGX_OPC_ILL, TILEGX_OPC_RAISE, TILEGX_OPC_ILL,
+  BITFIELD(31, 2) /* index 991 */,
+  TILEGX_OPC_LD1S, TILEGX_OPC_LD1S, TILEGX_OPC_LD1S, CHILD(996),
+  BITFIELD(33, 2) /* index 996 */,
+  TILEGX_OPC_LD1S, TILEGX_OPC_LD1S, TILEGX_OPC_LD1S, CHILD(1001),
+  BITFIELD(35, 2) /* index 1001 */,
+  TILEGX_OPC_LD1S, TILEGX_OPC_LD1S, TILEGX_OPC_LD1S,
+  TILEGX_OPC_PREFETCH_L1_FAULT,
+  BITFIELD(43, 3) /* index 1006 */,
+  CHILD(1015), CHILD(1030), CHILD(1045), CHILD(1060), CHILD(1075),
+  TILEGX_OPC_LDNA, TILEGX_OPC_LDNT1S, TILEGX_OPC_LDNT1U,
+  BITFIELD(31, 2) /* index 1015 */,
+  TILEGX_OPC_LD1U, TILEGX_OPC_LD1U, TILEGX_OPC_LD1U, CHILD(1020),
+  BITFIELD(33, 2) /* index 1020 */,
+  TILEGX_OPC_LD1U, TILEGX_OPC_LD1U, TILEGX_OPC_LD1U, CHILD(1025),
+  BITFIELD(35, 2) /* index 1025 */,
+  TILEGX_OPC_LD1U, TILEGX_OPC_LD1U, TILEGX_OPC_LD1U, TILEGX_OPC_PREFETCH,
+  BITFIELD(31, 2) /* index 1030 */,
+  TILEGX_OPC_LD2S, TILEGX_OPC_LD2S, TILEGX_OPC_LD2S, CHILD(1035),
+  BITFIELD(33, 2) /* index 1035 */,
+  TILEGX_OPC_LD2S, TILEGX_OPC_LD2S, TILEGX_OPC_LD2S, CHILD(1040),
+  BITFIELD(35, 2) /* index 1040 */,
+  TILEGX_OPC_LD2S, TILEGX_OPC_LD2S, TILEGX_OPC_LD2S,
+  TILEGX_OPC_PREFETCH_L2_FAULT,
+  BITFIELD(31, 2) /* index 1045 */,
+  TILEGX_OPC_LD2U, TILEGX_OPC_LD2U, TILEGX_OPC_LD2U, CHILD(1050),
+  BITFIELD(33, 2) /* index 1050 */,
+  TILEGX_OPC_LD2U, TILEGX_OPC_LD2U, TILEGX_OPC_LD2U, CHILD(1055),
+  BITFIELD(35, 2) /* index 1055 */,
+  TILEGX_OPC_LD2U, TILEGX_OPC_LD2U, TILEGX_OPC_LD2U, TILEGX_OPC_PREFETCH_L2,
+  BITFIELD(31, 2) /* index 1060 */,
+  TILEGX_OPC_LD4S, TILEGX_OPC_LD4S, TILEGX_OPC_LD4S, CHILD(1065),
+  BITFIELD(33, 2) /* index 1065 */,
+  TILEGX_OPC_LD4S, TILEGX_OPC_LD4S, TILEGX_OPC_LD4S, CHILD(1070),
+  BITFIELD(35, 2) /* index 1070 */,
+  TILEGX_OPC_LD4S, TILEGX_OPC_LD4S, TILEGX_OPC_LD4S,
+  TILEGX_OPC_PREFETCH_L3_FAULT,
+  BITFIELD(31, 2) /* index 1075 */,
+  TILEGX_OPC_LD4U, TILEGX_OPC_LD4U, TILEGX_OPC_LD4U, CHILD(1080),
+  BITFIELD(33, 2) /* index 1080 */,
+  TILEGX_OPC_LD4U, TILEGX_OPC_LD4U, TILEGX_OPC_LD4U, CHILD(1085),
+  BITFIELD(35, 2) /* index 1085 */,
+  TILEGX_OPC_LD4U, TILEGX_OPC_LD4U, TILEGX_OPC_LD4U, TILEGX_OPC_PREFETCH_L3,
+  BITFIELD(43, 3) /* index 1090 */,
+  TILEGX_OPC_LDNT2S, TILEGX_OPC_LDNT2U, TILEGX_OPC_LDNT4S, TILEGX_OPC_LDNT4U,
+  TILEGX_OPC_LDNT, TILEGX_OPC_LD, TILEGX_OPC_LNK, TILEGX_OPC_MF,
+  BITFIELD(43, 3) /* index 1099 */,
+  TILEGX_OPC_NAP, TILEGX_OPC_NOP, TILEGX_OPC_SWINT0, TILEGX_OPC_SWINT1,
+  TILEGX_OPC_SWINT2, TILEGX_OPC_SWINT3, TILEGX_OPC_WH64, TILEGX_OPC_NONE,
+  BITFIELD(49, 4) /* index 1108 */,
+  TILEGX_OPC_V1MAXU, TILEGX_OPC_V1MINU, TILEGX_OPC_V1MNZ, TILEGX_OPC_V1MZ,
+  TILEGX_OPC_V1SHL, TILEGX_OPC_V1SHRS, TILEGX_OPC_V1SHRU, TILEGX_OPC_V1SUBUC,
+  TILEGX_OPC_V1SUB, TILEGX_OPC_V2ADDSC, TILEGX_OPC_V2ADD, TILEGX_OPC_V2CMPEQ,
+  TILEGX_OPC_V2CMPLES, TILEGX_OPC_V2CMPLEU, TILEGX_OPC_V2CMPLTS,
+  TILEGX_OPC_V2CMPLTU,
+  BITFIELD(49, 4) /* index 1125 */,
+  TILEGX_OPC_V2CMPNE, TILEGX_OPC_V2INT_H, TILEGX_OPC_V2INT_L,
+  TILEGX_OPC_V2MAXS, TILEGX_OPC_V2MINS, TILEGX_OPC_V2MNZ, TILEGX_OPC_V2MZ,
+  TILEGX_OPC_V2PACKH, TILEGX_OPC_V2PACKL, TILEGX_OPC_V2PACKUC,
+  TILEGX_OPC_V2SHLSC, TILEGX_OPC_V2SHL, TILEGX_OPC_V2SHRS, TILEGX_OPC_V2SHRU,
+  TILEGX_OPC_V2SUBSC, TILEGX_OPC_V2SUB,
+  BITFIELD(49, 4) /* index 1142 */,
+  TILEGX_OPC_V4ADDSC, TILEGX_OPC_V4ADD, TILEGX_OPC_V4INT_H,
+  TILEGX_OPC_V4INT_L, TILEGX_OPC_V4PACKSC, TILEGX_OPC_V4SHLSC,
+  TILEGX_OPC_V4SHL, TILEGX_OPC_V4SHRS, TILEGX_OPC_V4SHRU, TILEGX_OPC_V4SUBSC,
+  TILEGX_OPC_V4SUB, TILEGX_OPC_XOR, TILEGX_OPC_NONE, TILEGX_OPC_NONE,
+  TILEGX_OPC_NONE, TILEGX_OPC_NONE,
+  BITFIELD(49, 4) /* index 1159 */,
+  TILEGX_OPC_NONE, TILEGX_OPC_ROTLI, TILEGX_OPC_SHLI, TILEGX_OPC_SHLXI,
+  TILEGX_OPC_SHRSI, TILEGX_OPC_SHRUI, TILEGX_OPC_SHRUXI, TILEGX_OPC_V1SHLI,
+  TILEGX_OPC_V1SHRSI, TILEGX_OPC_V1SHRUI, TILEGX_OPC_V2SHLI,
+  TILEGX_OPC_V2SHRSI, TILEGX_OPC_V2SHRUI, TILEGX_OPC_NONE, TILEGX_OPC_NONE,
+  TILEGX_OPC_NONE,
+  BITFIELD(31, 2) /* index 1176 */,
+  TILEGX_OPC_SHL16INSLI, TILEGX_OPC_SHL16INSLI, TILEGX_OPC_SHL16INSLI,
+  CHILD(1181),
+  BITFIELD(33, 2) /* index 1181 */,
+  TILEGX_OPC_SHL16INSLI, TILEGX_OPC_SHL16INSLI, TILEGX_OPC_SHL16INSLI,
+  CHILD(1186),
+  BITFIELD(35, 2) /* index 1186 */,
+  TILEGX_OPC_SHL16INSLI, TILEGX_OPC_SHL16INSLI, TILEGX_OPC_SHL16INSLI,
+  CHILD(1191),
+  BITFIELD(37, 2) /* index 1191 */,
+  TILEGX_OPC_SHL16INSLI, TILEGX_OPC_SHL16INSLI, TILEGX_OPC_SHL16INSLI,
+  CHILD(1196),
+  BITFIELD(39, 2) /* index 1196 */,
+  TILEGX_OPC_SHL16INSLI, TILEGX_OPC_SHL16INSLI, TILEGX_OPC_SHL16INSLI,
+  CHILD(1201),
+  BITFIELD(41, 2) /* index 1201 */,
+  TILEGX_OPC_SHL16INSLI, TILEGX_OPC_SHL16INSLI, TILEGX_OPC_SHL16INSLI,
+  TILEGX_OPC_INFOL,
+};
+
+static const unsigned short decode_Y0_fsm[178] =
+{
+  BITFIELD(27, 4) /* index 0 */,
+  CHILD(17), TILEGX_OPC_ADDXI, CHILD(32), TILEGX_OPC_CMPEQI,
+  TILEGX_OPC_CMPLTSI, CHILD(62), CHILD(67), CHILD(118), CHILD(123),
+  CHILD(128), CHILD(133), CHILD(153), CHILD(158), CHILD(163), CHILD(168),
+  CHILD(173),
+  BITFIELD(6, 2) /* index 17 */,
+  TILEGX_OPC_ADDI, TILEGX_OPC_ADDI, TILEGX_OPC_ADDI, CHILD(22),
+  BITFIELD(8, 2) /* index 22 */,
+  TILEGX_OPC_ADDI, TILEGX_OPC_ADDI, TILEGX_OPC_ADDI, CHILD(27),
+  BITFIELD(10, 2) /* index 27 */,
+  TILEGX_OPC_ADDI, TILEGX_OPC_ADDI, TILEGX_OPC_ADDI, TILEGX_OPC_MOVEI,
+  BITFIELD(0, 2) /* index 32 */,
+  TILEGX_OPC_ANDI, TILEGX_OPC_ANDI, TILEGX_OPC_ANDI, CHILD(37),
+  BITFIELD(2, 2) /* index 37 */,
+  TILEGX_OPC_ANDI, TILEGX_OPC_ANDI, TILEGX_OPC_ANDI, CHILD(42),
+  BITFIELD(4, 2) /* index 42 */,
+  TILEGX_OPC_ANDI, TILEGX_OPC_ANDI, TILEGX_OPC_ANDI, CHILD(47),
+  BITFIELD(6, 2) /* index 47 */,
+  TILEGX_OPC_ANDI, TILEGX_OPC_ANDI, TILEGX_OPC_ANDI, CHILD(52),
+  BITFIELD(8, 2) /* index 52 */,
+  TILEGX_OPC_ANDI, TILEGX_OPC_ANDI, TILEGX_OPC_ANDI, CHILD(57),
+  BITFIELD(10, 2) /* index 57 */,
+  TILEGX_OPC_ANDI, TILEGX_OPC_ANDI, TILEGX_OPC_ANDI, TILEGX_OPC_INFO,
+  BITFIELD(18, 2) /* index 62 */,
+  TILEGX_OPC_ADDX, TILEGX_OPC_ADD, TILEGX_OPC_SUBX, TILEGX_OPC_SUB,
+  BITFIELD(15, 5) /* index 67 */,
+  TILEGX_OPC_SHL1ADD, TILEGX_OPC_SHL1ADD, TILEGX_OPC_SHL1ADD,
+  TILEGX_OPC_SHL1ADD, TILEGX_OPC_SHL1ADD, TILEGX_OPC_SHL1ADD,
+  TILEGX_OPC_SHL1ADD, TILEGX_OPC_SHL1ADD, TILEGX_OPC_SHL2ADD,
+  TILEGX_OPC_SHL2ADD, TILEGX_OPC_SHL2ADD, TILEGX_OPC_SHL2ADD,
+  TILEGX_OPC_SHL2ADD, TILEGX_OPC_SHL2ADD, TILEGX_OPC_SHL2ADD,
+  TILEGX_OPC_SHL2ADD, TILEGX_OPC_SHL3ADD, TILEGX_OPC_SHL3ADD,
+  TILEGX_OPC_SHL3ADD, TILEGX_OPC_SHL3ADD, TILEGX_OPC_SHL3ADD,
+  TILEGX_OPC_SHL3ADD, TILEGX_OPC_SHL3ADD, TILEGX_OPC_SHL3ADD, CHILD(100),
+  CHILD(109), TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE,
+  TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE,
+  BITFIELD(12, 3) /* index 100 */,
+  TILEGX_OPC_NONE, TILEGX_OPC_CLZ, TILEGX_OPC_CTZ, TILEGX_OPC_FNOP,
+  TILEGX_OPC_FSINGLE_PACK1, TILEGX_OPC_NOP, TILEGX_OPC_PCNT,
+  TILEGX_OPC_REVBITS,
+  BITFIELD(12, 3) /* index 109 */,
+  TILEGX_OPC_REVBYTES, TILEGX_OPC_TBLIDXB0, TILEGX_OPC_TBLIDXB1,
+  TILEGX_OPC_TBLIDXB2, TILEGX_OPC_TBLIDXB3, TILEGX_OPC_NONE, TILEGX_OPC_NONE,
+  TILEGX_OPC_NONE,
+  BITFIELD(18, 2) /* index 118 */,
+  TILEGX_OPC_CMPLES, TILEGX_OPC_CMPLEU, TILEGX_OPC_CMPLTS, TILEGX_OPC_CMPLTU,
+  BITFIELD(18, 2) /* index 123 */,
+  TILEGX_OPC_CMPEQ, TILEGX_OPC_CMPNE, TILEGX_OPC_MULAX, TILEGX_OPC_MULX,
+  BITFIELD(18, 2) /* index 128 */,
+  TILEGX_OPC_CMOVEQZ, TILEGX_OPC_CMOVNEZ, TILEGX_OPC_MNZ, TILEGX_OPC_MZ,
+  BITFIELD(18, 2) /* index 133 */,
+  TILEGX_OPC_AND, TILEGX_OPC_NOR, CHILD(138), TILEGX_OPC_XOR,
+  BITFIELD(12, 2) /* index 138 */,
+  TILEGX_OPC_OR, TILEGX_OPC_OR, TILEGX_OPC_OR, CHILD(143),
+  BITFIELD(14, 2) /* index 143 */,
+  TILEGX_OPC_OR, TILEGX_OPC_OR, TILEGX_OPC_OR, CHILD(148),
+  BITFIELD(16, 2) /* index 148 */,
+  TILEGX_OPC_OR, TILEGX_OPC_OR, TILEGX_OPC_OR, TILEGX_OPC_MOVE,
+  BITFIELD(18, 2) /* index 153 */,
+  TILEGX_OPC_ROTL, TILEGX_OPC_SHL, TILEGX_OPC_SHRS, TILEGX_OPC_SHRU,
+  BITFIELD(18, 2) /* index 158 */,
+  TILEGX_OPC_NONE, TILEGX_OPC_SHL1ADDX, TILEGX_OPC_SHL2ADDX,
+  TILEGX_OPC_SHL3ADDX,
+  BITFIELD(18, 2) /* index 163 */,
+  TILEGX_OPC_MUL_HS_HS, TILEGX_OPC_MUL_HU_HU, TILEGX_OPC_MUL_LS_LS,
+  TILEGX_OPC_MUL_LU_LU,
+  BITFIELD(18, 2) /* index 168 */,
+  TILEGX_OPC_MULA_HS_HS, TILEGX_OPC_MULA_HU_HU, TILEGX_OPC_MULA_LS_LS,
+  TILEGX_OPC_MULA_LU_LU,
+  BITFIELD(18, 2) /* index 173 */,
+  TILEGX_OPC_ROTLI, TILEGX_OPC_SHLI, TILEGX_OPC_SHRSI, TILEGX_OPC_SHRUI,
+};
+
+static const unsigned short decode_Y1_fsm[167] =
+{
+  BITFIELD(58, 4) /* index 0 */,
+  TILEGX_OPC_NONE, CHILD(17), TILEGX_OPC_ADDXI, CHILD(32), TILEGX_OPC_CMPEQI,
+  TILEGX_OPC_CMPLTSI, CHILD(62), CHILD(67), CHILD(117), CHILD(122),
+  CHILD(127), CHILD(132), CHILD(152), CHILD(157), CHILD(162), TILEGX_OPC_NONE,
+  BITFIELD(37, 2) /* index 17 */,
+  TILEGX_OPC_ADDI, TILEGX_OPC_ADDI, TILEGX_OPC_ADDI, CHILD(22),
+  BITFIELD(39, 2) /* index 22 */,
+  TILEGX_OPC_ADDI, TILEGX_OPC_ADDI, TILEGX_OPC_ADDI, CHILD(27),
+  BITFIELD(41, 2) /* index 27 */,
+  TILEGX_OPC_ADDI, TILEGX_OPC_ADDI, TILEGX_OPC_ADDI, TILEGX_OPC_MOVEI,
+  BITFIELD(31, 2) /* index 32 */,
+  TILEGX_OPC_ANDI, TILEGX_OPC_ANDI, TILEGX_OPC_ANDI, CHILD(37),
+  BITFIELD(33, 2) /* index 37 */,
+  TILEGX_OPC_ANDI, TILEGX_OPC_ANDI, TILEGX_OPC_ANDI, CHILD(42),
+  BITFIELD(35, 2) /* index 42 */,
+  TILEGX_OPC_ANDI, TILEGX_OPC_ANDI, TILEGX_OPC_ANDI, CHILD(47),
+  BITFIELD(37, 2) /* index 47 */,
+  TILEGX_OPC_ANDI, TILEGX_OPC_ANDI, TILEGX_OPC_ANDI, CHILD(52),
+  BITFIELD(39, 2) /* index 52 */,
+  TILEGX_OPC_ANDI, TILEGX_OPC_ANDI, TILEGX_OPC_ANDI, CHILD(57),
+  BITFIELD(41, 2) /* index 57 */,
+  TILEGX_OPC_ANDI, TILEGX_OPC_ANDI, TILEGX_OPC_ANDI, TILEGX_OPC_INFO,
+  BITFIELD(49, 2) /* index 62 */,
+  TILEGX_OPC_ADDX, TILEGX_OPC_ADD, TILEGX_OPC_SUBX, TILEGX_OPC_SUB,
+  BITFIELD(47, 4) /* index 67 */,
+  TILEGX_OPC_SHL1ADD, TILEGX_OPC_SHL1ADD, TILEGX_OPC_SHL1ADD,
+  TILEGX_OPC_SHL1ADD, TILEGX_OPC_SHL2ADD, TILEGX_OPC_SHL2ADD,
+  TILEGX_OPC_SHL2ADD, TILEGX_OPC_SHL2ADD, TILEGX_OPC_SHL3ADD,
+  TILEGX_OPC_SHL3ADD, TILEGX_OPC_SHL3ADD, TILEGX_OPC_SHL3ADD, CHILD(84),
+  TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE,
+  BITFIELD(43, 3) /* index 84 */,
+  CHILD(93), CHILD(96), CHILD(99), CHILD(102), CHILD(105), CHILD(108),
+  CHILD(111), CHILD(114),
+  BITFIELD(46, 1) /* index 93 */,
+  TILEGX_OPC_NONE, TILEGX_OPC_FNOP,
+  BITFIELD(46, 1) /* index 96 */,
+  TILEGX_OPC_NONE, TILEGX_OPC_ILL,
+  BITFIELD(46, 1) /* index 99 */,
+  TILEGX_OPC_NONE, TILEGX_OPC_JALRP,
+  BITFIELD(46, 1) /* index 102 */,
+  TILEGX_OPC_NONE, TILEGX_OPC_JALR,
+  BITFIELD(46, 1) /* index 105 */,
+  TILEGX_OPC_NONE, TILEGX_OPC_JRP,
+  BITFIELD(46, 1) /* index 108 */,
+  TILEGX_OPC_NONE, TILEGX_OPC_JR,
+  BITFIELD(46, 1) /* index 111 */,
+  TILEGX_OPC_NONE, TILEGX_OPC_LNK,
+  BITFIELD(46, 1) /* index 114 */,
+  TILEGX_OPC_NONE, TILEGX_OPC_NOP,
+  BITFIELD(49, 2) /* index 117 */,
+  TILEGX_OPC_CMPLES, TILEGX_OPC_CMPLEU, TILEGX_OPC_CMPLTS, TILEGX_OPC_CMPLTU,
+  BITFIELD(49, 2) /* index 122 */,
+  TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_CMPEQ, TILEGX_OPC_CMPNE,
+  BITFIELD(49, 2) /* index 127 */,
+  TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_MNZ, TILEGX_OPC_MZ,
+  BITFIELD(49, 2) /* index 132 */,
+  TILEGX_OPC_AND, TILEGX_OPC_NOR, CHILD(137), TILEGX_OPC_XOR,
+  BITFIELD(43, 2) /* index 137 */,
+  TILEGX_OPC_OR, TILEGX_OPC_OR, TILEGX_OPC_OR, CHILD(142),
+  BITFIELD(45, 2) /* index 142 */,
+  TILEGX_OPC_OR, TILEGX_OPC_OR, TILEGX_OPC_OR, CHILD(147),
+  BITFIELD(47, 2) /* index 147 */,
+  TILEGX_OPC_OR, TILEGX_OPC_OR, TILEGX_OPC_OR, TILEGX_OPC_MOVE,
+  BITFIELD(49, 2) /* index 152 */,
+  TILEGX_OPC_ROTL, TILEGX_OPC_SHL, TILEGX_OPC_SHRS, TILEGX_OPC_SHRU,
+  BITFIELD(49, 2) /* index 157 */,
+  TILEGX_OPC_NONE, TILEGX_OPC_SHL1ADDX, TILEGX_OPC_SHL2ADDX,
+  TILEGX_OPC_SHL3ADDX,
+  BITFIELD(49, 2) /* index 162 */,
+  TILEGX_OPC_ROTLI, TILEGX_OPC_SHLI, TILEGX_OPC_SHRSI, TILEGX_OPC_SHRUI,
+};
+
+static const unsigned short decode_Y2_fsm[118] =
+{
+  BITFIELD(62, 2) /* index 0 */,
+  TILEGX_OPC_NONE, CHILD(5), CHILD(66), CHILD(109),
+  BITFIELD(55, 3) /* index 5 */,
+  CHILD(14), CHILD(14), CHILD(14), CHILD(17), CHILD(40), CHILD(40), CHILD(40),
+  CHILD(43),
+  BITFIELD(26, 1) /* index 14 */,
+  TILEGX_OPC_LD1S, TILEGX_OPC_LD1U,
+  BITFIELD(26, 1) /* index 17 */,
+  CHILD(20), CHILD(30),
+  BITFIELD(51, 2) /* index 20 */,
+  TILEGX_OPC_LD1S, TILEGX_OPC_LD1S, TILEGX_OPC_LD1S, CHILD(25),
+  BITFIELD(53, 2) /* index 25 */,
+  TILEGX_OPC_LD1S, TILEGX_OPC_LD1S, TILEGX_OPC_LD1S,
+  TILEGX_OPC_PREFETCH_L1_FAULT,
+  BITFIELD(51, 2) /* index 30 */,
+  TILEGX_OPC_LD1U, TILEGX_OPC_LD1U, TILEGX_OPC_LD1U, CHILD(35),
+  BITFIELD(53, 2) /* index 35 */,
+  TILEGX_OPC_LD1U, TILEGX_OPC_LD1U, TILEGX_OPC_LD1U, TILEGX_OPC_PREFETCH,
+  BITFIELD(26, 1) /* index 40 */,
+  TILEGX_OPC_LD2S, TILEGX_OPC_LD2U,
+  BITFIELD(26, 1) /* index 43 */,
+  CHILD(46), CHILD(56),
+  BITFIELD(51, 2) /* index 46 */,
+  TILEGX_OPC_LD2S, TILEGX_OPC_LD2S, TILEGX_OPC_LD2S, CHILD(51),
+  BITFIELD(53, 2) /* index 51 */,
+  TILEGX_OPC_LD2S, TILEGX_OPC_LD2S, TILEGX_OPC_LD2S,
+  TILEGX_OPC_PREFETCH_L2_FAULT,
+  BITFIELD(51, 2) /* index 56 */,
+  TILEGX_OPC_LD2U, TILEGX_OPC_LD2U, TILEGX_OPC_LD2U, CHILD(61),
+  BITFIELD(53, 2) /* index 61 */,
+  TILEGX_OPC_LD2U, TILEGX_OPC_LD2U, TILEGX_OPC_LD2U, TILEGX_OPC_PREFETCH_L2,
+  BITFIELD(56, 2) /* index 66 */,
+  CHILD(71), CHILD(74), CHILD(90), CHILD(93),
+  BITFIELD(26, 1) /* index 71 */,
+  TILEGX_OPC_NONE, TILEGX_OPC_LD4S,
+  BITFIELD(26, 1) /* index 74 */,
+  TILEGX_OPC_NONE, CHILD(77),
+  BITFIELD(51, 2) /* index 77 */,
+  TILEGX_OPC_LD4S, TILEGX_OPC_LD4S, TILEGX_OPC_LD4S, CHILD(82),
+  BITFIELD(53, 2) /* index 82 */,
+  TILEGX_OPC_LD4S, TILEGX_OPC_LD4S, TILEGX_OPC_LD4S, CHILD(87),
+  BITFIELD(55, 1) /* index 87 */,
+  TILEGX_OPC_LD4S, TILEGX_OPC_PREFETCH_L3_FAULT,
+  BITFIELD(26, 1) /* index 90 */,
+  TILEGX_OPC_LD4U, TILEGX_OPC_LD,
+  BITFIELD(26, 1) /* index 93 */,
+  CHILD(96), TILEGX_OPC_LD,
+  BITFIELD(51, 2) /* index 96 */,
+  TILEGX_OPC_LD4U, TILEGX_OPC_LD4U, TILEGX_OPC_LD4U, CHILD(101),
+  BITFIELD(53, 2) /* index 101 */,
+  TILEGX_OPC_LD4U, TILEGX_OPC_LD4U, TILEGX_OPC_LD4U, CHILD(106),
+  BITFIELD(55, 1) /* index 106 */,
+  TILEGX_OPC_LD4U, TILEGX_OPC_PREFETCH_L3,
+  BITFIELD(26, 1) /* index 109 */,
+  CHILD(112), CHILD(115),
+  BITFIELD(57, 1) /* index 112 */,
+  TILEGX_OPC_ST1, TILEGX_OPC_ST4,
+  BITFIELD(57, 1) /* index 115 */,
+  TILEGX_OPC_ST2, TILEGX_OPC_ST,
+};
+
+#undef BITFIELD
+#undef CHILD
+const unsigned short * const
+tilegx_bundle_decoder_fsms[TILEGX_NUM_PIPELINE_ENCODINGS] =
+{
+  decode_X0_fsm,
+  decode_X1_fsm,
+  decode_Y0_fsm,
+  decode_Y1_fsm,
+  decode_Y2_fsm
+};
+const struct tilegx_operand tilegx_operands[35] =
+{
+  {
+    TILEGX_OP_TYPE_IMMEDIATE, BFD_RELOC(TILEGX_IMM8_X0),
+    8, 1, 0, 0, 0, 0,
+    create_Imm8_X0, get_Imm8_X0
+  },
+  {
+    TILEGX_OP_TYPE_IMMEDIATE, BFD_RELOC(TILEGX_IMM8_X1),
+    8, 1, 0, 0, 0, 0,
+    create_Imm8_X1, get_Imm8_X1
+  },
+  {
+    TILEGX_OP_TYPE_IMMEDIATE, BFD_RELOC(TILEGX_IMM8_Y0),
+    8, 1, 0, 0, 0, 0,
+    create_Imm8_Y0, get_Imm8_Y0
+  },
+  {
+    TILEGX_OP_TYPE_IMMEDIATE, BFD_RELOC(TILEGX_IMM8_Y1),
+    8, 1, 0, 0, 0, 0,
+    create_Imm8_Y1, get_Imm8_Y1
+  },
+  {
+    TILEGX_OP_TYPE_IMMEDIATE, BFD_RELOC(TILEGX_IMM16_X0_HW0_LAST),
+    16, 1, 0, 0, 0, 0,
+    create_Imm16_X0, get_Imm16_X0
+  },
+  {
+    TILEGX_OP_TYPE_IMMEDIATE, BFD_RELOC(TILEGX_IMM16_X1_HW0_LAST),
+    16, 1, 0, 0, 0, 0,
+    create_Imm16_X1, get_Imm16_X1
+  },
+  {
+    TILEGX_OP_TYPE_REGISTER, BFD_RELOC(NONE),
+    6, 0, 0, 1, 0, 0,
+    create_Dest_X0, get_Dest_X0
+  },
+  {
+    TILEGX_OP_TYPE_REGISTER, BFD_RELOC(NONE),
+    6, 0, 1, 0, 0, 0,
+    create_SrcA_X0, get_SrcA_X0
+  },
+  {
+    TILEGX_OP_TYPE_REGISTER, BFD_RELOC(NONE),
+    6, 0, 0, 1, 0, 0,
+    create_Dest_X1, get_Dest_X1
+  },
+  {
+    TILEGX_OP_TYPE_REGISTER, BFD_RELOC(NONE),
+    6, 0, 1, 0, 0, 0,
+    create_SrcA_X1, get_SrcA_X1
+  },
+  {
+    TILEGX_OP_TYPE_REGISTER, BFD_RELOC(NONE),
+    6, 0, 0, 1, 0, 0,
+    create_Dest_Y0, get_Dest_Y0
+  },
+  {
+    TILEGX_OP_TYPE_REGISTER, BFD_RELOC(NONE),
+    6, 0, 1, 0, 0, 0,
+    create_SrcA_Y0, get_SrcA_Y0
+  },
+  {
+    TILEGX_OP_TYPE_REGISTER, BFD_RELOC(NONE),
+    6, 0, 0, 1, 0, 0,
+    create_Dest_Y1, get_Dest_Y1
+  },
+  {
+    TILEGX_OP_TYPE_REGISTER, BFD_RELOC(NONE),
+    6, 0, 1, 0, 0, 0,
+    create_SrcA_Y1, get_SrcA_Y1
+  },
+  {
+    TILEGX_OP_TYPE_REGISTER, BFD_RELOC(NONE),
+    6, 0, 1, 0, 0, 0,
+    create_SrcA_Y2, get_SrcA_Y2
+  },
+  {
+    TILEGX_OP_TYPE_REGISTER, BFD_RELOC(NONE),
+    6, 0, 1, 1, 0, 0,
+    create_SrcA_X1, get_SrcA_X1
+  },
+  {
+    TILEGX_OP_TYPE_REGISTER, BFD_RELOC(NONE),
+    6, 0, 1, 0, 0, 0,
+    create_SrcB_X0, get_SrcB_X0
+  },
+  {
+    TILEGX_OP_TYPE_REGISTER, BFD_RELOC(NONE),
+    6, 0, 1, 0, 0, 0,
+    create_SrcB_X1, get_SrcB_X1
+  },
+  {
+    TILEGX_OP_TYPE_REGISTER, BFD_RELOC(NONE),
+    6, 0, 1, 0, 0, 0,
+    create_SrcB_Y0, get_SrcB_Y0
+  },
+  {
+    TILEGX_OP_TYPE_REGISTER, BFD_RELOC(NONE),
+    6, 0, 1, 0, 0, 0,
+    create_SrcB_Y1, get_SrcB_Y1
+  },
+  {
+    TILEGX_OP_TYPE_ADDRESS, BFD_RELOC(TILEGX_BROFF_X1),
+    17, 1, 0, 0, 1, TILEGX_LOG2_BUNDLE_ALIGNMENT_IN_BYTES,
+    create_BrOff_X1, get_BrOff_X1
+  },
+  {
+    TILEGX_OP_TYPE_IMMEDIATE, BFD_RELOC(TILEGX_MMSTART_X0),
+    6, 0, 0, 0, 0, 0,
+    create_BFStart_X0, get_BFStart_X0
+  },
+  {
+    TILEGX_OP_TYPE_IMMEDIATE, BFD_RELOC(TILEGX_MMEND_X0),
+    6, 0, 0, 0, 0, 0,
+    create_BFEnd_X0, get_BFEnd_X0
+  },
+  {
+    TILEGX_OP_TYPE_REGISTER, BFD_RELOC(NONE),
+    6, 0, 1, 1, 0, 0,
+    create_Dest_X0, get_Dest_X0
+  },
+  {
+    TILEGX_OP_TYPE_REGISTER, BFD_RELOC(NONE),
+    6, 0, 1, 1, 0, 0,
+    create_Dest_Y0, get_Dest_Y0
+  },
+  {
+    TILEGX_OP_TYPE_ADDRESS, BFD_RELOC(TILEGX_JUMPOFF_X1),
+    27, 1, 0, 0, 1, TILEGX_LOG2_BUNDLE_ALIGNMENT_IN_BYTES,
+    create_JumpOff_X1, get_JumpOff_X1
+  },
+  {
+    TILEGX_OP_TYPE_REGISTER, BFD_RELOC(NONE),
+    6, 0, 0, 1, 0, 0,
+    create_SrcBDest_Y2, get_SrcBDest_Y2
+  },
+  {
+    TILEGX_OP_TYPE_SPR, BFD_RELOC(TILEGX_MF_IMM14_X1),
+    14, 0, 0, 0, 0, 0,
+    create_MF_Imm14_X1, get_MF_Imm14_X1
+  },
+  {
+    TILEGX_OP_TYPE_SPR, BFD_RELOC(TILEGX_MT_IMM14_X1),
+    14, 0, 0, 0, 0, 0,
+    create_MT_Imm14_X1, get_MT_Imm14_X1
+  },
+  {
+    TILEGX_OP_TYPE_IMMEDIATE, BFD_RELOC(TILEGX_SHAMT_X0),
+    6, 0, 0, 0, 0, 0,
+    create_ShAmt_X0, get_ShAmt_X0
+  },
+  {
+    TILEGX_OP_TYPE_IMMEDIATE, BFD_RELOC(TILEGX_SHAMT_X1),
+    6, 0, 0, 0, 0, 0,
+    create_ShAmt_X1, get_ShAmt_X1
+  },
+  {
+    TILEGX_OP_TYPE_IMMEDIATE, BFD_RELOC(TILEGX_SHAMT_Y0),
+    6, 0, 0, 0, 0, 0,
+    create_ShAmt_Y0, get_ShAmt_Y0
+  },
+  {
+    TILEGX_OP_TYPE_IMMEDIATE, BFD_RELOC(TILEGX_SHAMT_Y1),
+    6, 0, 0, 0, 0, 0,
+    create_ShAmt_Y1, get_ShAmt_Y1
+  },
+  {
+    TILEGX_OP_TYPE_REGISTER, BFD_RELOC(NONE),
+    6, 0, 1, 0, 0, 0,
+    create_SrcBDest_Y2, get_SrcBDest_Y2
+  },
+  {
+    TILEGX_OP_TYPE_IMMEDIATE, BFD_RELOC(TILEGX_DEST_IMM8_X1),
+    8, 1, 0, 0, 0, 0,
+    create_Dest_Imm8_X1, get_Dest_Imm8_X1
+  }
+};
+
+
+
+
+/* Given a set of bundle bits and the lookup FSM for a specific pipe,
+ * returns which instruction the bundle contains in that pipe.
+ */
+static const struct tilegx_opcode *
+find_opcode(tilegx_bundle_bits bits, const unsigned short *table)
+{
+  int index = 0;
+
+  while (1)
+  {
+    unsigned short bitspec = table[index];
+    unsigned int bitfield =
+      ((unsigned int)(bits >> (bitspec & 63))) & (bitspec >> 6);
+
+    unsigned short next = table[index + 1 + bitfield];
+    if (next <= TILEGX_OPC_NONE)
+      return &tilegx_opcodes[next];
+
+    index = next - TILEGX_OPC_NONE;
+  }
+}
+
+
+int
+parse_insn_tilegx(tilegx_bundle_bits bits,
+                  unsigned long long pc,
+                  struct tilegx_decoded_instruction
+                  decoded[TILEGX_MAX_INSTRUCTIONS_PER_BUNDLE])
+{
+  int num_instructions = 0;
+  int pipe;
+
+  int min_pipe, max_pipe;
+  if ((bits & TILEGX_BUNDLE_MODE_MASK) == 0)
+  {
+    min_pipe = TILEGX_PIPELINE_X0;
+    max_pipe = TILEGX_PIPELINE_X1;
+  }
+  else
+  {
+    min_pipe = TILEGX_PIPELINE_Y0;
+    max_pipe = TILEGX_PIPELINE_Y2;
+  }
+
+  /* For each pipe, find an instruction that fits. */
+  for (pipe = min_pipe; pipe <= max_pipe; pipe++)
+  {
+    const struct tilegx_opcode *opc;
+    struct tilegx_decoded_instruction *d;
+    int i;
+
+    d = &decoded[num_instructions++];
+    opc = find_opcode (bits, tilegx_bundle_decoder_fsms[pipe]);
+    d->opcode = opc;
+
+    /* Decode each operand, sign extending, etc. as appropriate. */
+    for (i = 0; i < opc->num_operands; i++)
+    {
+      const struct tilegx_operand *op =
+        &tilegx_operands[opc->operands[pipe][i]];
+      int raw_opval = op->extract (bits);
+      long long opval;
+
+      if (op->is_signed)
+      {
+        /* Sign-extend the operand. */
+        int shift = (int)((sizeof(int) * 8) - op->num_bits);
+        raw_opval = (raw_opval << shift) >> shift;
+      }
+
+      /* Adjust PC-relative scaled branch offsets. */
+      if (op->type == TILEGX_OP_TYPE_ADDRESS)
+        opval = (raw_opval * TILEGX_BUNDLE_SIZE_IN_BYTES) + pc;
+      else
+        opval = raw_opval;
+
+      /* Record the final value. */
+      d->operands[i] = op;
+      d->operand_values[i] = opval;
+    }
+  }
+
+  return num_instructions;
+}
diff --git a/kernel/arch/tile/kernel/time.c b/kernel/arch/tile/kernel/time.c
new file mode 100644
index 000000000..00178ecf9
--- /dev/null
+++ b/kernel/arch/tile/kernel/time.c
@@ -0,0 +1,300 @@
+/*
+ * Copyright 2010 Tilera Corporation. All Rights Reserved.
+ *
+ *   This program is free software; you can redistribute it and/or
+ *   modify it under the terms of the GNU General Public License
+ *   as published by the Free Software Foundation, version 2.
+ *
+ *   This program is distributed in the hope that it will be useful, but
+ *   WITHOUT ANY WARRANTY; without even the implied warranty of
+ *   MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
+ *   NON INFRINGEMENT.  See the GNU General Public License for
+ *   more details.
+ *
+ * Support the cycle counter clocksource and tile timer clock event device.
+ */
+
+#include <linux/time.h>
+#include <linux/timex.h>
+#include <linux/clocksource.h>
+#include <linux/clockchips.h>
+#include <linux/hardirq.h>
+#include <linux/sched.h>
+#include <linux/smp.h>
+#include <linux/delay.h>
+#include <linux/module.h>
+#include <linux/timekeeper_internal.h>
+#include <asm/irq_regs.h>
+#include <asm/traps.h>
+#include <asm/vdso.h>
+#include <hv/hypervisor.h>
+#include <arch/interrupts.h>
+#include <arch/spr_def.h>
+
+
+/*
+ * Define the cycle counter clock source.
+ */
+
+/* How many cycles per second we are running at. */
+static cycles_t cycles_per_sec __write_once;
+
+cycles_t get_clock_rate(void)
+{
+	return cycles_per_sec;
+}
+
+#if CHIP_HAS_SPLIT_CYCLE()
+cycles_t get_cycles(void)
+{
+	unsigned int high = __insn_mfspr(SPR_CYCLE_HIGH);
+	unsigned int low = __insn_mfspr(SPR_CYCLE_LOW);
+	unsigned int high2 = __insn_mfspr(SPR_CYCLE_HIGH);
+
+	while (unlikely(high != high2)) {
+		low = __insn_mfspr(SPR_CYCLE_LOW);
+		high = high2;
+		high2 = __insn_mfspr(SPR_CYCLE_HIGH);
+	}
+
+	return (((cycles_t)high) << 32) | low;
+}
+EXPORT_SYMBOL(get_cycles);
+#endif
+
+/*
+ * We use a relatively small shift value so that sched_clock()
+ * won't wrap around very often.
+ */
+#define SCHED_CLOCK_SHIFT 10
+
+static unsigned long sched_clock_mult __write_once;
+
+static cycles_t clocksource_get_cycles(struct clocksource *cs)
+{
+	return get_cycles();
+}
+
+static struct clocksource cycle_counter_cs = {
+	.name = "cycle counter",
+	.rating = 300,
+	.read = clocksource_get_cycles,
+	.mask = CLOCKSOURCE_MASK(64),
+	.flags = CLOCK_SOURCE_IS_CONTINUOUS,
+};
+
+/*
+ * Called very early from setup_arch() to set cycles_per_sec.
+ * We initialize it early so we can use it to set up loops_per_jiffy.
+ */
+void __init setup_clock(void)
+{
+	cycles_per_sec = hv_sysconf(HV_SYSCONF_CPU_SPEED);
+	sched_clock_mult =
+		clocksource_hz2mult(cycles_per_sec, SCHED_CLOCK_SHIFT);
+}
+
+void __init calibrate_delay(void)
+{
+	loops_per_jiffy = get_clock_rate() / HZ;
+	pr_info("Clock rate yields %lu.%02lu BogoMIPS (lpj=%lu)\n",
+		loops_per_jiffy / (500000 / HZ),
+		(loops_per_jiffy / (5000 / HZ)) % 100, loops_per_jiffy);
+}
+
+/* Called fairly late in init/main.c, but before we go smp. */
+void __init time_init(void)
+{
+	/* Initialize and register the clock source. */
+	clocksource_register_hz(&cycle_counter_cs, cycles_per_sec);
+
+	/* Start up the tile-timer interrupt source on the boot cpu. */
+	setup_tile_timer();
+}
+
+/*
+ * Define the tile timer clock event device.  The timer is driven by
+ * the TILE_TIMER_CONTROL register, which consists of a 31-bit down
+ * counter, plus bit 31, which signifies that the counter has wrapped
+ * from zero to (2**31) - 1.  The INT_TILE_TIMER interrupt will be
+ * raised as long as bit 31 is set.
+ *
+ * The TILE_MINSEC value represents the largest range of real-time
+ * we can possibly cover with the timer, based on MAX_TICK combined
+ * with the slowest reasonable clock rate we might run at.
+ */
+
+#define MAX_TICK 0x7fffffff   /* we have 31 bits of countdown timer */
+#define TILE_MINSEC 5         /* timer covers no more than 5 seconds */
+
+static int tile_timer_set_next_event(unsigned long ticks,
+				     struct clock_event_device *evt)
+{
+	BUG_ON(ticks > MAX_TICK);
+	__insn_mtspr(SPR_TILE_TIMER_CONTROL, ticks);
+	arch_local_irq_unmask_now(INT_TILE_TIMER);
+	return 0;
+}
+
+/*
+ * Whenever anyone tries to change modes, we just mask interrupts
+ * and wait for the next event to get set.
+ */
+static void tile_timer_set_mode(enum clock_event_mode mode,
+				struct clock_event_device *evt)
+{
+	arch_local_irq_mask_now(INT_TILE_TIMER);
+}
+
+/*
+ * Set min_delta_ns to 1 microsecond, since it takes about
+ * that long to fire the interrupt.
+ */
+static DEFINE_PER_CPU(struct clock_event_device, tile_timer) = {
+	.name = "tile timer",
+	.features = CLOCK_EVT_FEAT_ONESHOT,
+	.min_delta_ns = 1000,
+	.rating = 100,
+	.irq = -1,
+	.set_next_event = tile_timer_set_next_event,
+	.set_mode = tile_timer_set_mode,
+};
+
+void setup_tile_timer(void)
+{
+	struct clock_event_device *evt = this_cpu_ptr(&tile_timer);
+
+	/* Fill in fields that are speed-specific. */
+	clockevents_calc_mult_shift(evt, cycles_per_sec, TILE_MINSEC);
+	evt->max_delta_ns = clockevent_delta2ns(MAX_TICK, evt);
+
+	/* Mark as being for this cpu only. */
+	evt->cpumask = cpumask_of(smp_processor_id());
+
+	/* Start out with timer not firing. */
+	arch_local_irq_mask_now(INT_TILE_TIMER);
+
+	/* Register tile timer. */
+	clockevents_register_device(evt);
+}
+
+/* Called from the interrupt vector. */
+void do_timer_interrupt(struct pt_regs *regs, int fault_num)
+{
+	struct pt_regs *old_regs = set_irq_regs(regs);
+	struct clock_event_device *evt = this_cpu_ptr(&tile_timer);
+
+	/*
+	 * Mask the timer interrupt here, since we are a oneshot timer
+	 * and there are now by definition no events pending.
+	 */
+	arch_local_irq_mask(INT_TILE_TIMER);
+
+	/* Track time spent here in an interrupt context */
+	irq_enter();
+
+	/* Track interrupt count. */
+	__this_cpu_inc(irq_stat.irq_timer_count);
+
+	/* Call the generic timer handler */
+	evt->event_handler(evt);
+
+	/*
+	 * Track time spent against the current process again and
+	 * process any softirqs if they are waiting.
+	 */
+	irq_exit();
+
+	set_irq_regs(old_regs);
+}
+
+/*
+ * Scheduler clock - returns current time in nanosec units.
+ * Note that with LOCKDEP, this is called during lockdep_init(), and
+ * we will claim that sched_clock() is zero for a little while, until
+ * we run setup_clock(), above.
+ */
+unsigned long long sched_clock(void)
+{
+	return clocksource_cyc2ns(get_cycles(),
+				  sched_clock_mult, SCHED_CLOCK_SHIFT);
+}
+
+int setup_profiling_timer(unsigned int multiplier)
+{
+	return -EINVAL;
+}
+
+/*
+ * Use the tile timer to convert nsecs to core clock cycles, relying
+ * on it having the same frequency as SPR_CYCLE.
+ */
+cycles_t ns2cycles(unsigned long nsecs)
+{
+	/*
+	 * We do not have to disable preemption here as each core has the same
+	 * clock frequency.
+	 */
+	struct clock_event_device *dev = raw_cpu_ptr(&tile_timer);
+
+	/*
+	 * as in clocksource.h and x86's timer.h, we split the calculation
+	 * into 2 parts to avoid unecessary overflow of the intermediate
+	 * value. This will not lead to any loss of precision.
+	 */
+	u64 quot = (u64)nsecs >> dev->shift;
+	u64 rem  = (u64)nsecs & ((1ULL << dev->shift) - 1);
+	return quot * dev->mult + ((rem * dev->mult) >> dev->shift);
+}
+
+void update_vsyscall_tz(void)
+{
+	write_seqcount_begin(&vdso_data->tz_seq);
+	vdso_data->tz_minuteswest = sys_tz.tz_minuteswest;
+	vdso_data->tz_dsttime = sys_tz.tz_dsttime;
+	write_seqcount_end(&vdso_data->tz_seq);
+}
+
+void update_vsyscall(struct timekeeper *tk)
+{
+	if (tk->tkr_mono.clock != &cycle_counter_cs)
+		return;
+
+	write_seqcount_begin(&vdso_data->tb_seq);
+
+	vdso_data->cycle_last		= tk->tkr_mono.cycle_last;
+	vdso_data->mask			= tk->tkr_mono.mask;
+	vdso_data->mult			= tk->tkr_mono.mult;
+	vdso_data->shift		= tk->tkr_mono.shift;
+
+	vdso_data->wall_time_sec	= tk->xtime_sec;
+	vdso_data->wall_time_snsec	= tk->tkr_mono.xtime_nsec;
+
+	vdso_data->monotonic_time_sec	= tk->xtime_sec
+					+ tk->wall_to_monotonic.tv_sec;
+	vdso_data->monotonic_time_snsec	= tk->tkr_mono.xtime_nsec
+					+ ((u64)tk->wall_to_monotonic.tv_nsec
+						<< tk->tkr_mono.shift);
+	while (vdso_data->monotonic_time_snsec >=
+					(((u64)NSEC_PER_SEC) << tk->tkr_mono.shift)) {
+		vdso_data->monotonic_time_snsec -=
+					((u64)NSEC_PER_SEC) << tk->tkr_mono.shift;
+		vdso_data->monotonic_time_sec++;
+	}
+
+	vdso_data->wall_time_coarse_sec	= tk->xtime_sec;
+	vdso_data->wall_time_coarse_nsec = (long)(tk->tkr_mono.xtime_nsec >>
+						 tk->tkr_mono.shift);
+
+	vdso_data->monotonic_time_coarse_sec =
+		vdso_data->wall_time_coarse_sec + tk->wall_to_monotonic.tv_sec;
+	vdso_data->monotonic_time_coarse_nsec =
+		vdso_data->wall_time_coarse_nsec + tk->wall_to_monotonic.tv_nsec;
+
+	while (vdso_data->monotonic_time_coarse_nsec >= NSEC_PER_SEC) {
+		vdso_data->monotonic_time_coarse_nsec -= NSEC_PER_SEC;
+		vdso_data->monotonic_time_coarse_sec++;
+	}
+
+	write_seqcount_end(&vdso_data->tb_seq);
+}
diff --git a/kernel/arch/tile/kernel/tlb.c b/kernel/arch/tile/kernel/tlb.c
new file mode 100644
index 000000000..f23b53515
--- /dev/null
+++ b/kernel/arch/tile/kernel/tlb.c
@@ -0,0 +1,104 @@
+/*
+ * Copyright 2010 Tilera Corporation. All Rights Reserved.
+ *
+ *   This program is free software; you can redistribute it and/or
+ *   modify it under the terms of the GNU General Public License
+ *   as published by the Free Software Foundation, version 2.
+ *
+ *   This program is distributed in the hope that it will be useful, but
+ *   WITHOUT ANY WARRANTY; without even the implied warranty of
+ *   MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
+ *   NON INFRINGEMENT.  See the GNU General Public License for
+ *   more details.
+ *
+ */
+
+#include <linux/cpumask.h>
+#include <linux/module.h>
+#include <linux/hugetlb.h>
+#include <asm/tlbflush.h>
+#include <asm/homecache.h>
+#include <hv/hypervisor.h>
+
+/* From tlbflush.h */
+DEFINE_PER_CPU(int, current_asid);
+int min_asid, max_asid;
+
+/*
+ * Note that we flush the L1I (for VM_EXEC pages) as well as the TLB
+ * so that when we are unmapping an executable page, we also flush it.
+ * Combined with flushing the L1I at context switch time, this means
+ * we don't have to do any other icache flushes.
+ */
+
+void flush_tlb_mm(struct mm_struct *mm)
+{
+	HV_Remote_ASID asids[NR_CPUS];
+	int i = 0, cpu;
+	for_each_cpu(cpu, mm_cpumask(mm)) {
+		HV_Remote_ASID *asid = &asids[i++];
+		asid->y = cpu / smp_topology.width;
+		asid->x = cpu % smp_topology.width;
+		asid->asid = per_cpu(current_asid, cpu);
+	}
+	flush_remote(0, HV_FLUSH_EVICT_L1I, mm_cpumask(mm),
+		     0, 0, 0, NULL, asids, i);
+}
+
+void flush_tlb_current_task(void)
+{
+	flush_tlb_mm(current->mm);
+}
+
+void flush_tlb_page_mm(struct vm_area_struct *vma, struct mm_struct *mm,
+		       unsigned long va)
+{
+	unsigned long size = vma_kernel_pagesize(vma);
+	int cache = (vma->vm_flags & VM_EXEC) ? HV_FLUSH_EVICT_L1I : 0;
+	flush_remote(0, cache, mm_cpumask(mm),
+		     va, size, size, mm_cpumask(mm), NULL, 0);
+}
+
+void flush_tlb_page(struct vm_area_struct *vma, unsigned long va)
+{
+	flush_tlb_page_mm(vma, vma->vm_mm, va);
+}
+EXPORT_SYMBOL(flush_tlb_page);
+
+void flush_tlb_range(struct vm_area_struct *vma,
+		     unsigned long start, unsigned long end)
+{
+	unsigned long size = vma_kernel_pagesize(vma);
+	struct mm_struct *mm = vma->vm_mm;
+	int cache = (vma->vm_flags & VM_EXEC) ? HV_FLUSH_EVICT_L1I : 0;
+	flush_remote(0, cache, mm_cpumask(mm), start, end - start, size,
+		     mm_cpumask(mm), NULL, 0);
+}
+
+void flush_tlb_all(void)
+{
+	int i;
+	for (i = 0; ; ++i) {
+		HV_VirtAddrRange r = hv_inquire_virtual(i);
+		if (r.size == 0)
+			break;
+		flush_remote(0, HV_FLUSH_EVICT_L1I, cpu_online_mask,
+			     r.start, r.size, PAGE_SIZE, cpu_online_mask,
+			     NULL, 0);
+		flush_remote(0, 0, NULL,
+			     r.start, r.size, HPAGE_SIZE, cpu_online_mask,
+			     NULL, 0);
+	}
+}
+
+/*
+ * Callers need to flush the L1I themselves if necessary, e.g. for
+ * kernel module unload.  Otherwise we assume callers are not using
+ * executable pgprot_t's.  Using EVICT_L1I means that dataplane cpus
+ * will get an unnecessary interrupt otherwise.
+ */
+void flush_tlb_kernel_range(unsigned long start, unsigned long end)
+{
+	flush_remote(0, 0, NULL,
+		     start, end - start, PAGE_SIZE, cpu_online_mask, NULL, 0);
+}
diff --git a/kernel/arch/tile/kernel/traps.c b/kernel/arch/tile/kernel/traps.c
new file mode 100644
index 000000000..312fc134c
--- /dev/null
+++ b/kernel/arch/tile/kernel/traps.c
@@ -0,0 +1,403 @@
+/*
+ * Copyright 2010 Tilera Corporation. All Rights Reserved.
+ *
+ *   This program is free software; you can redistribute it and/or
+ *   modify it under the terms of the GNU General Public License
+ *   as published by the Free Software Foundation, version 2.
+ *
+ *   This program is distributed in the hope that it will be useful, but
+ *   WITHOUT ANY WARRANTY; without even the implied warranty of
+ *   MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
+ *   NON INFRINGEMENT.  See the GNU General Public License for
+ *   more details.
+ */
+
+#include <linux/sched.h>
+#include <linux/kernel.h>
+#include <linux/kprobes.h>
+#include <linux/kdebug.h>
+#include <linux/module.h>
+#include <linux/reboot.h>
+#include <linux/uaccess.h>
+#include <linux/ptrace.h>
+#include <linux/context_tracking.h>
+#include <asm/stack.h>
+#include <asm/traps.h>
+#include <asm/setup.h>
+
+#include <arch/interrupts.h>
+#include <arch/spr_def.h>
+#include <arch/opcode.h>
+
+void __init trap_init(void)
+{
+	/* Nothing needed here since we link code at .intrpt */
+}
+
+int unaligned_fixup = 1;
+
+static int __init setup_unaligned_fixup(char *str)
+{
+	/*
+	 * Say "=-1" to completely disable it.  If you just do "=0", we
+	 * will still parse the instruction, then fire a SIGBUS with
+	 * the correct address from inside the single_step code.
+	 */
+	if (kstrtoint(str, 0, &unaligned_fixup) != 0)
+		return 0;
+
+	pr_info("Fixups for unaligned data accesses are %s\n",
+		unaligned_fixup >= 0 ?
+		(unaligned_fixup ? "enabled" : "disabled") :
+		"completely disabled");
+	return 1;
+}
+__setup("unaligned_fixup=", setup_unaligned_fixup);
+
+#if CHIP_HAS_TILE_DMA()
+
+static int dma_disabled;
+
+static int __init nodma(char *str)
+{
+	pr_info("User-space DMA is disabled\n");
+	dma_disabled = 1;
+	return 1;
+}
+__setup("nodma", nodma);
+
+/* How to decode SPR_GPV_REASON */
+#define IRET_ERROR (1U << 31)
+#define MT_ERROR   (1U << 30)
+#define MF_ERROR   (1U << 29)
+#define SPR_INDEX  ((1U << 15) - 1)
+#define SPR_MPL_SHIFT  9  /* starting bit position for MPL encoded in SPR */
+
+/*
+ * See if this GPV is just to notify the kernel of SPR use and we can
+ * retry the user instruction after adjusting some MPLs suitably.
+ */
+static int retry_gpv(unsigned int gpv_reason)
+{
+	int mpl;
+
+	if (gpv_reason & IRET_ERROR)
+		return 0;
+
+	BUG_ON((gpv_reason & (MT_ERROR|MF_ERROR)) == 0);
+	mpl = (gpv_reason & SPR_INDEX) >> SPR_MPL_SHIFT;
+	if (mpl == INT_DMA_NOTIFY && !dma_disabled) {
+		/* User is turning on DMA. Allow it and retry. */
+		printk(KERN_DEBUG "Process %d/%s is now enabled for DMA\n",
+		       current->pid, current->comm);
+		BUG_ON(current->thread.tile_dma_state.enabled);
+		current->thread.tile_dma_state.enabled = 1;
+		grant_dma_mpls();
+		return 1;
+	}
+
+	return 0;
+}
+
+#endif /* CHIP_HAS_TILE_DMA() */
+
+extern tile_bundle_bits bpt_code;
+
+asm(".pushsection .rodata.bpt_code,\"a\";"
+    ".align 8;"
+    "bpt_code: bpt;"
+    ".size bpt_code,.-bpt_code;"
+    ".popsection");
+
+static int special_ill(tile_bundle_bits bundle, int *sigp, int *codep)
+{
+	int sig, code, maxcode;
+
+	if (bundle == bpt_code) {
+		*sigp = SIGTRAP;
+		*codep = TRAP_BRKPT;
+		return 1;
+	}
+
+	/* If it's a "raise" bundle, then "ill" must be in pipe X1. */
+#ifdef __tilegx__
+	if ((bundle & TILEGX_BUNDLE_MODE_MASK) != 0)
+		return 0;
+	if (get_Opcode_X1(bundle) != RRR_0_OPCODE_X1)
+		return 0;
+	if (get_RRROpcodeExtension_X1(bundle) != UNARY_RRR_0_OPCODE_X1)
+		return 0;
+	if (get_UnaryOpcodeExtension_X1(bundle) != ILL_UNARY_OPCODE_X1)
+		return 0;
+#else
+	if (bundle & TILEPRO_BUNDLE_Y_ENCODING_MASK)
+		return 0;
+	if (get_Opcode_X1(bundle) != SHUN_0_OPCODE_X1)
+		return 0;
+	if (get_UnShOpcodeExtension_X1(bundle) != UN_0_SHUN_0_OPCODE_X1)
+		return 0;
+	if (get_UnOpcodeExtension_X1(bundle) != ILL_UN_0_SHUN_0_OPCODE_X1)
+		return 0;
+#endif
+
+	/* Check that the magic distinguishers are set to mean "raise". */
+	if (get_Dest_X1(bundle) != 29 || get_SrcA_X1(bundle) != 37)
+		return 0;
+
+	/* There must be an "addli zero, zero, VAL" in X0. */
+	if (get_Opcode_X0(bundle) != ADDLI_OPCODE_X0)
+		return 0;
+	if (get_Dest_X0(bundle) != TREG_ZERO)
+		return 0;
+	if (get_SrcA_X0(bundle) != TREG_ZERO)
+		return 0;
+
+	/*
+	 * Validate the proposed signal number and si_code value.
+	 * Note that we embed these in the static instruction itself
+	 * so that we perturb the register state as little as possible
+	 * at the time of the actual fault; it's unlikely you'd ever
+	 * need to dynamically choose which kind of fault to raise
+	 * from user space.
+	 */
+	sig = get_Imm16_X0(bundle) & 0x3f;
+	switch (sig) {
+	case SIGILL:
+		maxcode = NSIGILL;
+		break;
+	case SIGFPE:
+		maxcode = NSIGFPE;
+		break;
+	case SIGSEGV:
+		maxcode = NSIGSEGV;
+		break;
+	case SIGBUS:
+		maxcode = NSIGBUS;
+		break;
+	case SIGTRAP:
+		maxcode = NSIGTRAP;
+		break;
+	default:
+		return 0;
+	}
+	code = (get_Imm16_X0(bundle) >> 6) & 0xf;
+	if (code <= 0 || code > maxcode)
+		return 0;
+
+	/* Make it the requested signal. */
+	*sigp = sig;
+	*codep = code | __SI_FAULT;
+	return 1;
+}
+
+static const char *const int_name[] = {
+	[INT_MEM_ERROR] = "Memory error",
+	[INT_ILL] = "Illegal instruction",
+	[INT_GPV] = "General protection violation",
+	[INT_UDN_ACCESS] = "UDN access",
+	[INT_IDN_ACCESS] = "IDN access",
+#if CHIP_HAS_SN()
+	[INT_SN_ACCESS] = "SN access",
+#endif
+	[INT_SWINT_3] = "Software interrupt 3",
+	[INT_SWINT_2] = "Software interrupt 2",
+	[INT_SWINT_0] = "Software interrupt 0",
+	[INT_UNALIGN_DATA] = "Unaligned data",
+	[INT_DOUBLE_FAULT] = "Double fault",
+#ifdef __tilegx__
+	[INT_ILL_TRANS] = "Illegal virtual address",
+#endif
+};
+
+static int do_bpt(struct pt_regs *regs)
+{
+	unsigned long bundle, bcode, bpt;
+
+	bundle = *(unsigned long *)instruction_pointer(regs);
+
+	/*
+	 * bpt shoule be { bpt; nop }, which is 0x286a44ae51485000ULL.
+	 * we encode the unused least significant bits for other purpose.
+	 */
+	bpt = bundle & ~((1ULL << 12) - 1);
+	if (bpt != TILE_BPT_BUNDLE)
+		return 0;
+
+	bcode = bundle & ((1ULL << 12) - 1);
+	/*
+	 * notify the kprobe handlers, if instruction is likely to
+	 * pertain to them.
+	 */
+	switch (bcode) {
+	/* breakpoint_insn */
+	case 0:
+		notify_die(DIE_BREAK, "debug", regs, bundle,
+			INT_ILL, SIGTRAP);
+		break;
+	/* compiled_bpt */
+	case DIE_COMPILED_BPT:
+		notify_die(DIE_COMPILED_BPT, "debug", regs, bundle,
+			INT_ILL, SIGTRAP);
+		break;
+	/* breakpoint2_insn */
+	case DIE_SSTEPBP:
+		notify_die(DIE_SSTEPBP, "single_step", regs, bundle,
+			INT_ILL, SIGTRAP);
+		break;
+	default:
+		return 0;
+	}
+
+	return 1;
+}
+
+void __kprobes do_trap(struct pt_regs *regs, int fault_num,
+		       unsigned long reason)
+{
+	enum ctx_state prev_state = exception_enter();
+	siginfo_t info = { 0 };
+	int signo, code;
+	unsigned long address = 0;
+	tile_bundle_bits instr;
+	int is_kernel = !user_mode(regs);
+
+	/* Handle breakpoints, etc. */
+	if (is_kernel && fault_num == INT_ILL && do_bpt(regs))
+		goto done;
+
+	/* Re-enable interrupts, if they were previously enabled. */
+	if (!(regs->flags & PT_FLAGS_DISABLE_IRQ))
+		local_irq_enable();
+
+	/*
+	 * If it hits in kernel mode and we can't fix it up, just exit the
+	 * current process and hope for the best.
+	 */
+	if (is_kernel) {
+		const char *name;
+		char buf[100];
+		if (fixup_exception(regs))  /* ILL_TRANS or UNALIGN_DATA */
+			goto done;
+		if (fault_num >= 0 &&
+		    fault_num < ARRAY_SIZE(int_name) &&
+		    int_name[fault_num] != NULL)
+			name = int_name[fault_num];
+		else
+			name = "Unknown interrupt";
+		if (fault_num == INT_GPV)
+			snprintf(buf, sizeof(buf), "; GPV_REASON %#lx", reason);
+#ifdef __tilegx__
+		else if (fault_num == INT_ILL_TRANS)
+			snprintf(buf, sizeof(buf), "; address %#lx", reason);
+#endif
+		else
+			buf[0] = '\0';
+		pr_alert("Kernel took bad trap %d (%s) at PC %#lx%s\n",
+			 fault_num, name, regs->pc, buf);
+		show_regs(regs);
+		do_exit(SIGKILL);  /* FIXME: implement i386 die() */
+	}
+
+	switch (fault_num) {
+	case INT_MEM_ERROR:
+		signo = SIGBUS;
+		code = BUS_OBJERR;
+		break;
+	case INT_ILL:
+		if (copy_from_user(&instr, (void __user *)regs->pc,
+				   sizeof(instr))) {
+			pr_err("Unreadable instruction for INT_ILL: %#lx\n",
+			       regs->pc);
+			do_exit(SIGKILL);
+		}
+		if (!special_ill(instr, &signo, &code)) {
+			signo = SIGILL;
+			code = ILL_ILLOPC;
+		}
+		address = regs->pc;
+		break;
+	case INT_GPV:
+#if CHIP_HAS_TILE_DMA()
+		if (retry_gpv(reason))
+			goto done;
+#endif
+		/*FALLTHROUGH*/
+	case INT_UDN_ACCESS:
+	case INT_IDN_ACCESS:
+#if CHIP_HAS_SN()
+	case INT_SN_ACCESS:
+#endif
+		signo = SIGILL;
+		code = ILL_PRVREG;
+		address = regs->pc;
+		break;
+	case INT_SWINT_3:
+	case INT_SWINT_2:
+	case INT_SWINT_0:
+		signo = SIGILL;
+		code = ILL_ILLTRP;
+		address = regs->pc;
+		break;
+	case INT_UNALIGN_DATA:
+#ifndef __tilegx__  /* Emulated support for single step debugging */
+		if (unaligned_fixup >= 0) {
+			struct single_step_state *state =
+				current_thread_info()->step_state;
+			if (!state ||
+			    (void __user *)(regs->pc) != state->buffer) {
+				single_step_once(regs);
+				goto done;
+			}
+		}
+#endif
+		signo = SIGBUS;
+		code = BUS_ADRALN;
+		address = 0;
+		break;
+	case INT_DOUBLE_FAULT:
+		/*
+		 * For double fault, "reason" is actually passed as
+		 * SYSTEM_SAVE_K_2, the hypervisor's double-fault info, so
+		 * we can provide the original fault number rather than
+		 * the uninteresting "INT_DOUBLE_FAULT" so the user can
+		 * learn what actually struck while PL0 ICS was set.
+		 */
+		fault_num = reason;
+		signo = SIGILL;
+		code = ILL_DBLFLT;
+		address = regs->pc;
+		break;
+#ifdef __tilegx__
+	case INT_ILL_TRANS: {
+		/* Avoid a hardware erratum with the return address stack. */
+		fill_ra_stack();
+
+		signo = SIGSEGV;
+		address = reason;
+		code = SEGV_MAPERR;
+		break;
+	}
+#endif
+	default:
+		panic("Unexpected do_trap interrupt number %d", fault_num);
+	}
+
+	info.si_signo = signo;
+	info.si_code = code;
+	info.si_addr = (void __user *)address;
+	if (signo == SIGILL)
+		info.si_trapno = fault_num;
+	if (signo != SIGTRAP)
+		trace_unhandled_signal("trap", regs, address, signo);
+	force_sig_info(signo, &info, current);
+
+done:
+	exception_exit(prev_state);
+}
+
+void kernel_double_fault(int dummy, ulong pc, ulong lr, ulong sp, ulong r52)
+{
+	_dump_stack(dummy, pc, lr, sp, r52);
+	pr_emerg("Double fault: exiting\n");
+	machine_halt();
+}
diff --git a/kernel/arch/tile/kernel/unaligned.c b/kernel/arch/tile/kernel/unaligned.c
new file mode 100644
index 000000000..d075f92cc
--- /dev/null
+++ b/kernel/arch/tile/kernel/unaligned.c
@@ -0,0 +1,1600 @@
+/*
+ * Copyright 2013 Tilera Corporation. All Rights Reserved.
+ *
+ *   This program is free software; you can redistribute it and/or
+ *   modify it under the terms of the GNU General Public License
+ *   as published by the Free Software Foundation, version 2.
+ *
+ *   This program is distributed in the hope that it will be useful, but
+ *   WITHOUT ANY WARRANTY; without even the implied warranty of
+ *   MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
+ *   NON INFRINGEMENT.  See the GNU General Public License for
+ *   more details.
+ *
+ * A code-rewriter that handles unaligned exception.
+ */
+
+#include <linux/smp.h>
+#include <linux/ptrace.h>
+#include <linux/slab.h>
+#include <linux/thread_info.h>
+#include <linux/uaccess.h>
+#include <linux/mman.h>
+#include <linux/types.h>
+#include <linux/err.h>
+#include <linux/module.h>
+#include <linux/compat.h>
+#include <linux/prctl.h>
+#include <linux/context_tracking.h>
+#include <asm/cacheflush.h>
+#include <asm/traps.h>
+#include <asm/uaccess.h>
+#include <asm/unaligned.h>
+#include <arch/abi.h>
+#include <arch/spr_def.h>
+#include <arch/opcode.h>
+
+
+/*
+ * This file handles unaligned exception for tile-Gx. The tilepro's unaligned
+ * exception is supported out of single_step.c
+ */
+
+int unaligned_printk;
+
+static int __init setup_unaligned_printk(char *str)
+{
+	long val;
+	if (kstrtol(str, 0, &val) != 0)
+		return 0;
+	unaligned_printk = val;
+	pr_info("Printk for each unaligned data accesses is %s\n",
+		unaligned_printk ? "enabled" : "disabled");
+	return 1;
+}
+__setup("unaligned_printk=", setup_unaligned_printk);
+
+unsigned int unaligned_fixup_count;
+
+#ifdef __tilegx__
+
+/*
+ * Unalign data jit fixup code fragement. Reserved space is 128 bytes.
+ * The 1st 64-bit word saves fault PC address, 2nd word is the fault
+ * instruction bundle followed by 14 JIT bundles.
+ */
+
+struct unaligned_jit_fragment {
+	unsigned long       pc;
+	tilegx_bundle_bits  bundle;
+	tilegx_bundle_bits  insn[14];
+};
+
+/*
+ * Check if a nop or fnop at bundle's pipeline X0.
+ */
+
+static bool is_bundle_x0_nop(tilegx_bundle_bits bundle)
+{
+	return (((get_UnaryOpcodeExtension_X0(bundle) ==
+		  NOP_UNARY_OPCODE_X0) &&
+		 (get_RRROpcodeExtension_X0(bundle) ==
+		  UNARY_RRR_0_OPCODE_X0) &&
+		 (get_Opcode_X0(bundle) ==
+		  RRR_0_OPCODE_X0)) ||
+		((get_UnaryOpcodeExtension_X0(bundle) ==
+		  FNOP_UNARY_OPCODE_X0) &&
+		 (get_RRROpcodeExtension_X0(bundle) ==
+		  UNARY_RRR_0_OPCODE_X0) &&
+		 (get_Opcode_X0(bundle) ==
+		  RRR_0_OPCODE_X0)));
+}
+
+/*
+ * Check if nop or fnop at bundle's pipeline X1.
+ */
+
+static bool is_bundle_x1_nop(tilegx_bundle_bits bundle)
+{
+	return (((get_UnaryOpcodeExtension_X1(bundle) ==
+		  NOP_UNARY_OPCODE_X1) &&
+		 (get_RRROpcodeExtension_X1(bundle) ==
+		  UNARY_RRR_0_OPCODE_X1) &&
+		 (get_Opcode_X1(bundle) ==
+		  RRR_0_OPCODE_X1)) ||
+		((get_UnaryOpcodeExtension_X1(bundle) ==
+		  FNOP_UNARY_OPCODE_X1) &&
+		 (get_RRROpcodeExtension_X1(bundle) ==
+		  UNARY_RRR_0_OPCODE_X1) &&
+		 (get_Opcode_X1(bundle) ==
+		  RRR_0_OPCODE_X1)));
+}
+
+/*
+ * Check if nop or fnop at bundle's Y0 pipeline.
+ */
+
+static bool is_bundle_y0_nop(tilegx_bundle_bits bundle)
+{
+	return (((get_UnaryOpcodeExtension_Y0(bundle) ==
+		  NOP_UNARY_OPCODE_Y0) &&
+		 (get_RRROpcodeExtension_Y0(bundle) ==
+		  UNARY_RRR_1_OPCODE_Y0) &&
+		 (get_Opcode_Y0(bundle) ==
+		  RRR_1_OPCODE_Y0)) ||
+		((get_UnaryOpcodeExtension_Y0(bundle) ==
+		  FNOP_UNARY_OPCODE_Y0) &&
+		 (get_RRROpcodeExtension_Y0(bundle) ==
+		  UNARY_RRR_1_OPCODE_Y0) &&
+		 (get_Opcode_Y0(bundle) ==
+		  RRR_1_OPCODE_Y0)));
+}
+
+/*
+ * Check if nop or fnop at bundle's pipeline Y1.
+ */
+
+static bool is_bundle_y1_nop(tilegx_bundle_bits bundle)
+{
+	return (((get_UnaryOpcodeExtension_Y1(bundle) ==
+		  NOP_UNARY_OPCODE_Y1) &&
+		 (get_RRROpcodeExtension_Y1(bundle) ==
+		  UNARY_RRR_1_OPCODE_Y1) &&
+		 (get_Opcode_Y1(bundle) ==
+		  RRR_1_OPCODE_Y1)) ||
+		((get_UnaryOpcodeExtension_Y1(bundle) ==
+		  FNOP_UNARY_OPCODE_Y1) &&
+		 (get_RRROpcodeExtension_Y1(bundle) ==
+		  UNARY_RRR_1_OPCODE_Y1) &&
+		 (get_Opcode_Y1(bundle) ==
+		  RRR_1_OPCODE_Y1)));
+}
+
+/*
+ * Test if a bundle's y0 and y1 pipelines are both nop or fnop.
+ */
+
+static bool is_y0_y1_nop(tilegx_bundle_bits bundle)
+{
+	return is_bundle_y0_nop(bundle) && is_bundle_y1_nop(bundle);
+}
+
+/*
+ * Test if a bundle's x0 and x1 pipelines are both nop or fnop.
+ */
+
+static bool is_x0_x1_nop(tilegx_bundle_bits bundle)
+{
+	return is_bundle_x0_nop(bundle) && is_bundle_x1_nop(bundle);
+}
+
+/*
+ * Find the destination, source registers of fault unalign access instruction
+ * at X1 or Y2. Also, allocate up to 3 scratch registers clob1, clob2 and
+ * clob3, which are guaranteed different from any register used in the fault
+ * bundle. r_alias is used to return if the other instructions other than the
+ * unalign load/store shares same register with ra, rb and rd.
+ */
+
+static void find_regs(tilegx_bundle_bits bundle, uint64_t *rd, uint64_t *ra,
+		      uint64_t *rb, uint64_t *clob1, uint64_t *clob2,
+		      uint64_t *clob3, bool *r_alias)
+{
+	int i;
+	uint64_t reg;
+	uint64_t reg_map = 0, alias_reg_map = 0, map;
+	bool alias = false;
+
+	/*
+	 * Parse fault bundle, find potential used registers and mark
+	 * corresponding bits in reg_map and alias_map. These 2 bit maps
+	 * are used to find the scratch registers and determine if there
+	 * is register alais.
+	 */
+	if (bundle & TILEGX_BUNDLE_MODE_MASK) {  /* Y Mode Bundle. */
+
+		reg = get_SrcA_Y2(bundle);
+		reg_map |= 1ULL << reg;
+		*ra = reg;
+		reg = get_SrcBDest_Y2(bundle);
+		reg_map |= 1ULL << reg;
+
+		if (rd) {
+			/* Load. */
+			*rd = reg;
+			alias_reg_map = (1ULL << *rd) | (1ULL << *ra);
+		} else {
+			/* Store. */
+			*rb = reg;
+			alias_reg_map = (1ULL << *ra) | (1ULL << *rb);
+		}
+
+		if (!is_bundle_y1_nop(bundle)) {
+			reg = get_SrcA_Y1(bundle);
+			reg_map |= (1ULL << reg);
+			map = (1ULL << reg);
+
+			reg = get_SrcB_Y1(bundle);
+			reg_map |= (1ULL << reg);
+			map |= (1ULL << reg);
+
+			reg = get_Dest_Y1(bundle);
+			reg_map |= (1ULL << reg);
+			map |= (1ULL << reg);
+
+			if (map & alias_reg_map)
+				alias = true;
+		}
+
+		if (!is_bundle_y0_nop(bundle)) {
+			reg = get_SrcA_Y0(bundle);
+			reg_map |= (1ULL << reg);
+			map = (1ULL << reg);
+
+			reg = get_SrcB_Y0(bundle);
+			reg_map |= (1ULL << reg);
+			map |= (1ULL << reg);
+
+			reg = get_Dest_Y0(bundle);
+			reg_map |= (1ULL << reg);
+			map |= (1ULL << reg);
+
+			if (map & alias_reg_map)
+				alias = true;
+		}
+	} else	{ /* X Mode Bundle. */
+
+		reg = get_SrcA_X1(bundle);
+		reg_map |= (1ULL << reg);
+		*ra = reg;
+		if (rd)	{
+			/* Load. */
+			reg = get_Dest_X1(bundle);
+			reg_map |= (1ULL << reg);
+			*rd = reg;
+			alias_reg_map = (1ULL << *rd) | (1ULL << *ra);
+		} else {
+			/* Store. */
+			reg = get_SrcB_X1(bundle);
+			reg_map |= (1ULL << reg);
+			*rb = reg;
+			alias_reg_map = (1ULL << *ra) | (1ULL << *rb);
+		}
+
+		if (!is_bundle_x0_nop(bundle)) {
+			reg = get_SrcA_X0(bundle);
+			reg_map |= (1ULL << reg);
+			map = (1ULL << reg);
+
+			reg = get_SrcB_X0(bundle);
+			reg_map |= (1ULL << reg);
+			map |= (1ULL << reg);
+
+			reg = get_Dest_X0(bundle);
+			reg_map |= (1ULL << reg);
+			map |= (1ULL << reg);
+
+			if (map & alias_reg_map)
+				alias = true;
+		}
+	}
+
+	/*
+	 * "alias" indicates if the unalign access registers have collision
+	 * with others in the same bundle. We jsut simply test all register
+	 * operands case (RRR), ignored the case with immidate. If a bundle
+	 * has no register alias, we may do fixup in a simple or fast manner.
+	 * So if an immidata field happens to hit with a register, we may end
+	 * up fall back to the generic handling.
+	 */
+
+	*r_alias = alias;
+
+	/* Flip bits on reg_map. */
+	reg_map ^= -1ULL;
+
+	/* Scan reg_map lower 54(TREG_SP) bits to find 3 set bits. */
+	for (i = 0; i < TREG_SP; i++) {
+		if (reg_map & (0x1ULL << i)) {
+			if (*clob1 == -1) {
+				*clob1 = i;
+			} else if (*clob2 == -1) {
+				*clob2 = i;
+			} else if (*clob3 == -1) {
+				*clob3 = i;
+				return;
+			}
+		}
+	}
+}
+
+/*
+ * Sanity check for register ra, rb, rd, clob1/2/3. Return true if any of them
+ * is unexpected.
+ */
+
+static bool check_regs(uint64_t rd, uint64_t ra, uint64_t rb,
+		       uint64_t clob1, uint64_t clob2,  uint64_t clob3)
+{
+	bool unexpected = false;
+	if ((ra >= 56) && (ra != TREG_ZERO))
+		unexpected = true;
+
+	if ((clob1 >= 56) || (clob2 >= 56) || (clob3 >= 56))
+		unexpected = true;
+
+	if (rd != -1) {
+		if ((rd >= 56) && (rd != TREG_ZERO))
+			unexpected = true;
+	} else {
+		if ((rb >= 56) && (rb != TREG_ZERO))
+			unexpected = true;
+	}
+	return unexpected;
+}
+
+
+#define  GX_INSN_X0_MASK   ((1ULL << 31) - 1)
+#define  GX_INSN_X1_MASK   (((1ULL << 31) - 1) << 31)
+#define  GX_INSN_Y0_MASK   ((0xFULL << 27) | (0xFFFFFULL))
+#define  GX_INSN_Y1_MASK   (GX_INSN_Y0_MASK << 31)
+#define  GX_INSN_Y2_MASK   ((0x7FULL << 51) | (0x7FULL << 20))
+
+#ifdef __LITTLE_ENDIAN
+#define  GX_INSN_BSWAP(_bundle_)    (_bundle_)
+#else
+#define  GX_INSN_BSWAP(_bundle_)    swab64(_bundle_)
+#endif /* __LITTLE_ENDIAN */
+
+/*
+ * __JIT_CODE(.) creates template bundles in .rodata.unalign_data section.
+ * The corresponding static function jix_x#_###(.) generates partial or
+ * whole bundle based on the template and given arguments.
+ */
+
+#define __JIT_CODE(_X_)						\
+	asm (".pushsection .rodata.unalign_data, \"a\"\n"	\
+	     _X_"\n"						\
+	     ".popsection\n")
+
+__JIT_CODE("__unalign_jit_x1_mtspr:   {mtspr 0,  r0}");
+static tilegx_bundle_bits jit_x1_mtspr(int spr, int reg)
+{
+	extern  tilegx_bundle_bits __unalign_jit_x1_mtspr;
+	return (GX_INSN_BSWAP(__unalign_jit_x1_mtspr) & GX_INSN_X1_MASK) |
+		create_MT_Imm14_X1(spr) | create_SrcA_X1(reg);
+}
+
+__JIT_CODE("__unalign_jit_x1_mfspr:   {mfspr r0, 0}");
+static tilegx_bundle_bits  jit_x1_mfspr(int reg, int spr)
+{
+	extern  tilegx_bundle_bits __unalign_jit_x1_mfspr;
+	return (GX_INSN_BSWAP(__unalign_jit_x1_mfspr) & GX_INSN_X1_MASK) |
+		create_MF_Imm14_X1(spr) | create_Dest_X1(reg);
+}
+
+__JIT_CODE("__unalign_jit_x0_addi:   {addi  r0, r0, 0; iret}");
+static tilegx_bundle_bits  jit_x0_addi(int rd, int ra, int imm8)
+{
+	extern  tilegx_bundle_bits __unalign_jit_x0_addi;
+	return (GX_INSN_BSWAP(__unalign_jit_x0_addi) & GX_INSN_X0_MASK) |
+		create_Dest_X0(rd) | create_SrcA_X0(ra) |
+		create_Imm8_X0(imm8);
+}
+
+__JIT_CODE("__unalign_jit_x1_ldna:   {ldna  r0, r0}");
+static tilegx_bundle_bits  jit_x1_ldna(int rd, int ra)
+{
+	extern  tilegx_bundle_bits __unalign_jit_x1_ldna;
+	return (GX_INSN_BSWAP(__unalign_jit_x1_ldna) &  GX_INSN_X1_MASK) |
+		create_Dest_X1(rd) | create_SrcA_X1(ra);
+}
+
+__JIT_CODE("__unalign_jit_x0_dblalign:   {dblalign r0, r0 ,r0}");
+static tilegx_bundle_bits  jit_x0_dblalign(int rd, int ra, int rb)
+{
+	extern  tilegx_bundle_bits __unalign_jit_x0_dblalign;
+	return (GX_INSN_BSWAP(__unalign_jit_x0_dblalign) & GX_INSN_X0_MASK) |
+		create_Dest_X0(rd) | create_SrcA_X0(ra) |
+		create_SrcB_X0(rb);
+}
+
+__JIT_CODE("__unalign_jit_x1_iret:   {iret}");
+static tilegx_bundle_bits  jit_x1_iret(void)
+{
+	extern  tilegx_bundle_bits __unalign_jit_x1_iret;
+	return GX_INSN_BSWAP(__unalign_jit_x1_iret) & GX_INSN_X1_MASK;
+}
+
+__JIT_CODE("__unalign_jit_x01_fnop:   {fnop;fnop}");
+static tilegx_bundle_bits  jit_x0_fnop(void)
+{
+	extern  tilegx_bundle_bits __unalign_jit_x01_fnop;
+	return GX_INSN_BSWAP(__unalign_jit_x01_fnop) & GX_INSN_X0_MASK;
+}
+
+static tilegx_bundle_bits  jit_x1_fnop(void)
+{
+	extern  tilegx_bundle_bits __unalign_jit_x01_fnop;
+	return GX_INSN_BSWAP(__unalign_jit_x01_fnop) & GX_INSN_X1_MASK;
+}
+
+__JIT_CODE("__unalign_jit_y2_dummy:   {fnop; fnop; ld zero, sp}");
+static tilegx_bundle_bits  jit_y2_dummy(void)
+{
+	extern  tilegx_bundle_bits __unalign_jit_y2_dummy;
+	return GX_INSN_BSWAP(__unalign_jit_y2_dummy) & GX_INSN_Y2_MASK;
+}
+
+static tilegx_bundle_bits  jit_y1_fnop(void)
+{
+	extern  tilegx_bundle_bits __unalign_jit_y2_dummy;
+	return GX_INSN_BSWAP(__unalign_jit_y2_dummy) & GX_INSN_Y1_MASK;
+}
+
+__JIT_CODE("__unalign_jit_x1_st1_add:  {st1_add r1, r0, 0}");
+static tilegx_bundle_bits  jit_x1_st1_add(int ra, int rb, int imm8)
+{
+	extern  tilegx_bundle_bits __unalign_jit_x1_st1_add;
+	return (GX_INSN_BSWAP(__unalign_jit_x1_st1_add) &
+		(~create_SrcA_X1(-1)) &
+		GX_INSN_X1_MASK) | create_SrcA_X1(ra) |
+		create_SrcB_X1(rb) | create_Dest_Imm8_X1(imm8);
+}
+
+__JIT_CODE("__unalign_jit_x1_st:  {crc32_8 r1, r0, r0; st  r0, r0}");
+static tilegx_bundle_bits  jit_x1_st(int ra, int rb)
+{
+	extern  tilegx_bundle_bits __unalign_jit_x1_st;
+	return (GX_INSN_BSWAP(__unalign_jit_x1_st) & GX_INSN_X1_MASK) |
+		create_SrcA_X1(ra) | create_SrcB_X1(rb);
+}
+
+__JIT_CODE("__unalign_jit_x1_st_add:  {st_add  r1, r0, 0}");
+static tilegx_bundle_bits  jit_x1_st_add(int ra, int rb, int imm8)
+{
+	extern  tilegx_bundle_bits __unalign_jit_x1_st_add;
+	return (GX_INSN_BSWAP(__unalign_jit_x1_st_add) &
+		(~create_SrcA_X1(-1)) &
+		GX_INSN_X1_MASK) | create_SrcA_X1(ra) |
+		create_SrcB_X1(rb) | create_Dest_Imm8_X1(imm8);
+}
+
+__JIT_CODE("__unalign_jit_x1_ld:  {crc32_8 r1, r0, r0; ld  r0, r0}");
+static tilegx_bundle_bits  jit_x1_ld(int rd, int ra)
+{
+	extern  tilegx_bundle_bits __unalign_jit_x1_ld;
+	return (GX_INSN_BSWAP(__unalign_jit_x1_ld) & GX_INSN_X1_MASK) |
+		create_Dest_X1(rd) | create_SrcA_X1(ra);
+}
+
+__JIT_CODE("__unalign_jit_x1_ld_add:  {ld_add  r1, r0, 0}");
+static tilegx_bundle_bits  jit_x1_ld_add(int rd, int ra, int imm8)
+{
+	extern  tilegx_bundle_bits __unalign_jit_x1_ld_add;
+	return (GX_INSN_BSWAP(__unalign_jit_x1_ld_add) &
+		(~create_Dest_X1(-1)) &
+		GX_INSN_X1_MASK) | create_Dest_X1(rd) |
+		create_SrcA_X1(ra) | create_Imm8_X1(imm8);
+}
+
+__JIT_CODE("__unalign_jit_x0_bfexts:  {bfexts r0, r0, 0, 0}");
+static tilegx_bundle_bits  jit_x0_bfexts(int rd, int ra, int bfs, int bfe)
+{
+	extern  tilegx_bundle_bits __unalign_jit_x0_bfexts;
+	return (GX_INSN_BSWAP(__unalign_jit_x0_bfexts) &
+		GX_INSN_X0_MASK) |
+		create_Dest_X0(rd) | create_SrcA_X0(ra) |
+		create_BFStart_X0(bfs) | create_BFEnd_X0(bfe);
+}
+
+__JIT_CODE("__unalign_jit_x0_bfextu:  {bfextu r0, r0, 0, 0}");
+static tilegx_bundle_bits  jit_x0_bfextu(int rd, int ra, int bfs, int bfe)
+{
+	extern  tilegx_bundle_bits __unalign_jit_x0_bfextu;
+	return (GX_INSN_BSWAP(__unalign_jit_x0_bfextu) &
+		GX_INSN_X0_MASK) |
+		create_Dest_X0(rd) | create_SrcA_X0(ra) |
+		create_BFStart_X0(bfs) | create_BFEnd_X0(bfe);
+}
+
+__JIT_CODE("__unalign_jit_x1_addi:  {bfextu r1, r1, 0, 0; addi r0, r0, 0}");
+static tilegx_bundle_bits  jit_x1_addi(int rd, int ra, int imm8)
+{
+	extern  tilegx_bundle_bits __unalign_jit_x1_addi;
+	return (GX_INSN_BSWAP(__unalign_jit_x1_addi) & GX_INSN_X1_MASK) |
+		create_Dest_X1(rd) | create_SrcA_X1(ra) |
+		create_Imm8_X1(imm8);
+}
+
+__JIT_CODE("__unalign_jit_x0_shrui:  {shrui r0, r0, 0; iret}");
+static tilegx_bundle_bits  jit_x0_shrui(int rd, int ra, int imm6)
+{
+	extern  tilegx_bundle_bits __unalign_jit_x0_shrui;
+	return (GX_INSN_BSWAP(__unalign_jit_x0_shrui) &
+		GX_INSN_X0_MASK) |
+		create_Dest_X0(rd) | create_SrcA_X0(ra) |
+		create_ShAmt_X0(imm6);
+}
+
+__JIT_CODE("__unalign_jit_x0_rotli:  {rotli r0, r0, 0; iret}");
+static tilegx_bundle_bits  jit_x0_rotli(int rd, int ra, int imm6)
+{
+	extern  tilegx_bundle_bits __unalign_jit_x0_rotli;
+	return (GX_INSN_BSWAP(__unalign_jit_x0_rotli) &
+		GX_INSN_X0_MASK) |
+		create_Dest_X0(rd) | create_SrcA_X0(ra) |
+		create_ShAmt_X0(imm6);
+}
+
+__JIT_CODE("__unalign_jit_x1_bnezt:  {bnezt r0, __unalign_jit_x1_bnezt}");
+static tilegx_bundle_bits  jit_x1_bnezt(int ra, int broff)
+{
+	extern  tilegx_bundle_bits __unalign_jit_x1_bnezt;
+	return (GX_INSN_BSWAP(__unalign_jit_x1_bnezt) &
+		GX_INSN_X1_MASK) |
+		create_SrcA_X1(ra) | create_BrOff_X1(broff);
+}
+
+#undef __JIT_CODE
+
+/*
+ * This function generates unalign fixup JIT.
+ *
+ * We first find unalign load/store instruction's destination, source
+ * registers: ra, rb and rd. and 3 scratch registers by calling
+ * find_regs(...). 3 scratch clobbers should not alias with any register
+ * used in the fault bundle. Then analyze the fault bundle to determine
+ * if it's a load or store, operand width, branch or address increment etc.
+ * At last generated JIT is copied into JIT code area in user space.
+ */
+
+static
+void jit_bundle_gen(struct pt_regs *regs, tilegx_bundle_bits bundle,
+		    int align_ctl)
+{
+	struct thread_info *info = current_thread_info();
+	struct unaligned_jit_fragment frag;
+	struct unaligned_jit_fragment *jit_code_area;
+	tilegx_bundle_bits bundle_2 = 0;
+	/* If bundle_2_enable = false, bundle_2 is fnop/nop operation. */
+	bool     bundle_2_enable = true;
+	uint64_t ra = -1, rb = -1, rd = -1, clob1 = -1, clob2 = -1, clob3 = -1;
+	/*
+	 * Indicate if the unalign access
+	 * instruction's registers hit with
+	 * others in the same bundle.
+	 */
+	bool     alias = false;
+	bool     load_n_store = true;
+	bool     load_store_signed = false;
+	unsigned int  load_store_size = 8;
+	bool     y1_br = false;  /* True, for a branch in same bundle at Y1.*/
+	int      y1_br_reg = 0;
+	/* True for link operation. i.e. jalr or lnk at Y1 */
+	bool     y1_lr = false;
+	int      y1_lr_reg = 0;
+	bool     x1_add = false;/* True, for load/store ADD instruction at X1*/
+	int      x1_add_imm8 = 0;
+	bool     unexpected = false;
+	int      n = 0, k;
+
+	jit_code_area =
+		(struct unaligned_jit_fragment *)(info->unalign_jit_base);
+
+	memset((void *)&frag, 0, sizeof(frag));
+
+	/* 0: X mode, Otherwise: Y mode. */
+	if (bundle & TILEGX_BUNDLE_MODE_MASK) {
+		unsigned int mod, opcode;
+
+		if (get_Opcode_Y1(bundle) == RRR_1_OPCODE_Y1 &&
+		    get_RRROpcodeExtension_Y1(bundle) ==
+		    UNARY_RRR_1_OPCODE_Y1) {
+
+			opcode = get_UnaryOpcodeExtension_Y1(bundle);
+
+			/*
+			 * Test "jalr", "jalrp", "jr", "jrp" instruction at Y1
+			 * pipeline.
+			 */
+			switch (opcode) {
+			case JALR_UNARY_OPCODE_Y1:
+			case JALRP_UNARY_OPCODE_Y1:
+				y1_lr = true;
+				y1_lr_reg = 55; /* Link register. */
+				/* FALLTHROUGH */
+			case JR_UNARY_OPCODE_Y1:
+			case JRP_UNARY_OPCODE_Y1:
+				y1_br = true;
+				y1_br_reg = get_SrcA_Y1(bundle);
+				break;
+			case LNK_UNARY_OPCODE_Y1:
+				/* "lnk" at Y1 pipeline. */
+				y1_lr = true;
+				y1_lr_reg = get_Dest_Y1(bundle);
+				break;
+			}
+		}
+
+		opcode = get_Opcode_Y2(bundle);
+		mod = get_Mode(bundle);
+
+		/*
+		 *  bundle_2 is bundle after making Y2 as a dummy operation
+		 *  - ld zero, sp
+		 */
+		bundle_2 = (bundle & (~GX_INSN_Y2_MASK)) | jit_y2_dummy();
+
+		/* Make Y1 as fnop if Y1 is a branch or lnk operation. */
+		if (y1_br || y1_lr) {
+			bundle_2 &= ~(GX_INSN_Y1_MASK);
+			bundle_2 |= jit_y1_fnop();
+		}
+
+		if (is_y0_y1_nop(bundle_2))
+			bundle_2_enable = false;
+
+		if (mod == MODE_OPCODE_YC2) {
+			/* Store. */
+			load_n_store = false;
+			load_store_size = 1 << opcode;
+			load_store_signed = false;
+			find_regs(bundle, 0, &ra, &rb, &clob1, &clob2,
+				  &clob3, &alias);
+			if (load_store_size > 8)
+				unexpected = true;
+		} else {
+			/* Load. */
+			load_n_store = true;
+			if (mod == MODE_OPCODE_YB2) {
+				switch (opcode) {
+				case LD_OPCODE_Y2:
+					load_store_signed = false;
+					load_store_size = 8;
+					break;
+				case LD4S_OPCODE_Y2:
+					load_store_signed = true;
+					load_store_size = 4;
+					break;
+				case LD4U_OPCODE_Y2:
+					load_store_signed = false;
+					load_store_size = 4;
+					break;
+				default:
+					unexpected = true;
+				}
+			} else if (mod == MODE_OPCODE_YA2) {
+				if (opcode == LD2S_OPCODE_Y2) {
+					load_store_signed = true;
+					load_store_size = 2;
+				} else if (opcode == LD2U_OPCODE_Y2) {
+					load_store_signed = false;
+					load_store_size = 2;
+				} else
+					unexpected = true;
+			} else
+				unexpected = true;
+			find_regs(bundle, &rd, &ra, &rb, &clob1, &clob2,
+				  &clob3, &alias);
+		}
+	} else {
+		unsigned int opcode;
+
+		/* bundle_2 is bundle after making X1 as "fnop". */
+		bundle_2 = (bundle & (~GX_INSN_X1_MASK)) | jit_x1_fnop();
+
+		if (is_x0_x1_nop(bundle_2))
+			bundle_2_enable = false;
+
+		if (get_Opcode_X1(bundle) == RRR_0_OPCODE_X1) {
+			opcode = get_UnaryOpcodeExtension_X1(bundle);
+
+			if (get_RRROpcodeExtension_X1(bundle) ==
+			    UNARY_RRR_0_OPCODE_X1) {
+				load_n_store = true;
+				find_regs(bundle, &rd, &ra, &rb, &clob1,
+					  &clob2, &clob3, &alias);
+
+				switch (opcode) {
+				case LD_UNARY_OPCODE_X1:
+					load_store_signed = false;
+					load_store_size = 8;
+					break;
+				case LD4S_UNARY_OPCODE_X1:
+					load_store_signed = true;
+					/* FALLTHROUGH */
+				case LD4U_UNARY_OPCODE_X1:
+					load_store_size = 4;
+					break;
+
+				case LD2S_UNARY_OPCODE_X1:
+					load_store_signed = true;
+					/* FALLTHROUGH */
+				case LD2U_UNARY_OPCODE_X1:
+					load_store_size = 2;
+					break;
+				default:
+					unexpected = true;
+				}
+			} else {
+				load_n_store = false;
+				load_store_signed = false;
+				find_regs(bundle, 0, &ra, &rb,
+					  &clob1, &clob2, &clob3,
+					  &alias);
+
+				opcode = get_RRROpcodeExtension_X1(bundle);
+				switch (opcode)	{
+				case ST_RRR_0_OPCODE_X1:
+					load_store_size = 8;
+					break;
+				case ST4_RRR_0_OPCODE_X1:
+					load_store_size = 4;
+					break;
+				case ST2_RRR_0_OPCODE_X1:
+					load_store_size = 2;
+					break;
+				default:
+					unexpected = true;
+				}
+			}
+		} else if (get_Opcode_X1(bundle) == IMM8_OPCODE_X1) {
+			load_n_store = true;
+			opcode = get_Imm8OpcodeExtension_X1(bundle);
+			switch (opcode)	{
+			case LD_ADD_IMM8_OPCODE_X1:
+				load_store_size = 8;
+				break;
+
+			case LD4S_ADD_IMM8_OPCODE_X1:
+				load_store_signed = true;
+				/* FALLTHROUGH */
+			case LD4U_ADD_IMM8_OPCODE_X1:
+				load_store_size = 4;
+				break;
+
+			case LD2S_ADD_IMM8_OPCODE_X1:
+				load_store_signed = true;
+				/* FALLTHROUGH */
+			case LD2U_ADD_IMM8_OPCODE_X1:
+				load_store_size = 2;
+				break;
+
+			case ST_ADD_IMM8_OPCODE_X1:
+				load_n_store = false;
+				load_store_size = 8;
+				break;
+			case ST4_ADD_IMM8_OPCODE_X1:
+				load_n_store = false;
+				load_store_size = 4;
+				break;
+			case ST2_ADD_IMM8_OPCODE_X1:
+				load_n_store = false;
+				load_store_size = 2;
+				break;
+			default:
+				unexpected = true;
+			}
+
+			if (!unexpected) {
+				x1_add = true;
+				if (load_n_store)
+					x1_add_imm8 = get_Imm8_X1(bundle);
+				else
+					x1_add_imm8 = get_Dest_Imm8_X1(bundle);
+			}
+
+			find_regs(bundle, load_n_store ? (&rd) : NULL,
+				  &ra, &rb, &clob1, &clob2, &clob3, &alias);
+		} else
+			unexpected = true;
+	}
+
+	/*
+	 * Some sanity check for register numbers extracted from fault bundle.
+	 */
+	if (check_regs(rd, ra, rb, clob1, clob2, clob3) == true)
+		unexpected = true;
+
+	/* Give warning if register ra has an aligned address. */
+	if (!unexpected)
+		WARN_ON(!((load_store_size - 1) & (regs->regs[ra])));
+
+
+	/*
+	 * Fault came from kernel space, here we only need take care of
+	 * unaligned "get_user/put_user" macros defined in "uaccess.h".
+	 * Basically, we will handle bundle like this:
+	 * {ld/2u/4s rd, ra; movei rx, 0} or {st/2/4 ra, rb; movei rx, 0}
+	 * (Refer to file "arch/tile/include/asm/uaccess.h" for details).
+	 * For either load or store, byte-wise operation is performed by calling
+	 * get_user() or put_user(). If the macro returns non-zero value,
+	 * set the value to rx, otherwise set zero to rx. Finally make pc point
+	 * to next bundle and return.
+	 */
+
+	if (EX1_PL(regs->ex1) != USER_PL) {
+
+		unsigned long rx = 0;
+		unsigned long x = 0, ret = 0;
+
+		if (y1_br || y1_lr || x1_add ||
+		    (load_store_signed !=
+		     (load_n_store && load_store_size == 4))) {
+			/* No branch, link, wrong sign-ext or load/store add. */
+			unexpected = true;
+		} else if (!unexpected) {
+			if (bundle & TILEGX_BUNDLE_MODE_MASK) {
+				/*
+				 * Fault bundle is Y mode.
+				 * Check if the Y1 and Y0 is the form of
+				 * { movei rx, 0; nop/fnop }, if yes,
+				 * find the rx.
+				 */
+
+				if ((get_Opcode_Y1(bundle) == ADDI_OPCODE_Y1)
+				    && (get_SrcA_Y1(bundle) == TREG_ZERO) &&
+				    (get_Imm8_Y1(bundle) == 0) &&
+				    is_bundle_y0_nop(bundle)) {
+					rx = get_Dest_Y1(bundle);
+				} else if ((get_Opcode_Y0(bundle) ==
+					    ADDI_OPCODE_Y0) &&
+					   (get_SrcA_Y0(bundle) == TREG_ZERO) &&
+					   (get_Imm8_Y0(bundle) == 0) &&
+					   is_bundle_y1_nop(bundle)) {
+					rx = get_Dest_Y0(bundle);
+				} else {
+					unexpected = true;
+				}
+			} else {
+				/*
+				 * Fault bundle is X mode.
+				 * Check if the X0 is 'movei rx, 0',
+				 * if yes, find the rx.
+				 */
+
+				if ((get_Opcode_X0(bundle) == IMM8_OPCODE_X0)
+				    && (get_Imm8OpcodeExtension_X0(bundle) ==
+					ADDI_IMM8_OPCODE_X0) &&
+				    (get_SrcA_X0(bundle) == TREG_ZERO) &&
+				    (get_Imm8_X0(bundle) == 0)) {
+					rx = get_Dest_X0(bundle);
+				} else {
+					unexpected = true;
+				}
+			}
+
+			/* rx should be less than 56. */
+			if (!unexpected && (rx >= 56))
+				unexpected = true;
+		}
+
+		if (!search_exception_tables(regs->pc)) {
+			/* No fixup in the exception tables for the pc. */
+			unexpected = true;
+		}
+
+		if (unexpected) {
+			/* Unexpected unalign kernel fault. */
+			struct task_struct *tsk = validate_current();
+
+			bust_spinlocks(1);
+
+			show_regs(regs);
+
+			if (unlikely(tsk->pid < 2)) {
+				panic("Kernel unalign fault running %s!",
+				      tsk->pid ? "init" : "the idle task");
+			}
+#ifdef SUPPORT_DIE
+			die("Oops", regs);
+#endif
+			bust_spinlocks(1);
+
+			do_group_exit(SIGKILL);
+
+		} else {
+			unsigned long i, b = 0;
+			unsigned char *ptr =
+				(unsigned char *)regs->regs[ra];
+			if (load_n_store) {
+				/* handle get_user(x, ptr) */
+				for (i = 0; i < load_store_size; i++) {
+					ret = get_user(b, ptr++);
+					if (!ret) {
+						/* Success! update x. */
+#ifdef __LITTLE_ENDIAN
+						x |= (b << (8 * i));
+#else
+						x <<= 8;
+						x |= b;
+#endif /* __LITTLE_ENDIAN */
+					} else {
+						x = 0;
+						break;
+					}
+				}
+
+				/* Sign-extend 4-byte loads. */
+				if (load_store_size == 4)
+					x = (long)(int)x;
+
+				/* Set register rd. */
+				regs->regs[rd] = x;
+
+				/* Set register rx. */
+				regs->regs[rx] = ret;
+
+				/* Bump pc. */
+				regs->pc += 8;
+
+			} else {
+				/* Handle put_user(x, ptr) */
+				x = regs->regs[rb];
+#ifdef __LITTLE_ENDIAN
+				b = x;
+#else
+				/*
+				 * Swap x in order to store x from low
+				 * to high memory same as the
+				 * little-endian case.
+				 */
+				switch (load_store_size) {
+				case 8:
+					b = swab64(x);
+					break;
+				case 4:
+					b = swab32(x);
+					break;
+				case 2:
+					b = swab16(x);
+					break;
+				}
+#endif /* __LITTLE_ENDIAN */
+				for (i = 0; i < load_store_size; i++) {
+					ret = put_user(b, ptr++);
+					if (ret)
+						break;
+					/* Success! shift 1 byte. */
+					b >>= 8;
+				}
+				/* Set register rx. */
+				regs->regs[rx] = ret;
+
+				/* Bump pc. */
+				regs->pc += 8;
+			}
+		}
+
+		unaligned_fixup_count++;
+
+		if (unaligned_printk) {
+			pr_info("%s/%d - Unalign fixup for kernel access to userspace %lx\n",
+				current->comm, current->pid, regs->regs[ra]);
+		}
+
+		/* Done! Return to the exception handler. */
+		return;
+	}
+
+	if ((align_ctl == 0) || unexpected) {
+		siginfo_t info = {
+			.si_signo = SIGBUS,
+			.si_code = BUS_ADRALN,
+			.si_addr = (unsigned char __user *)0
+		};
+		if (unaligned_printk)
+			pr_info("Unalign bundle: unexp @%llx, %llx\n",
+				(unsigned long long)regs->pc,
+				(unsigned long long)bundle);
+
+		if (ra < 56) {
+			unsigned long uaa = (unsigned long)regs->regs[ra];
+			/* Set bus Address. */
+			info.si_addr = (unsigned char __user *)uaa;
+		}
+
+		unaligned_fixup_count++;
+
+		trace_unhandled_signal("unaligned fixup trap", regs,
+				       (unsigned long)info.si_addr, SIGBUS);
+		force_sig_info(info.si_signo, &info, current);
+		return;
+	}
+
+#ifdef __LITTLE_ENDIAN
+#define UA_FIXUP_ADDR_DELTA          1
+#define UA_FIXUP_BFEXT_START(_B_)    0
+#define UA_FIXUP_BFEXT_END(_B_)     (8 * (_B_) - 1)
+#else /* __BIG_ENDIAN */
+#define UA_FIXUP_ADDR_DELTA          -1
+#define UA_FIXUP_BFEXT_START(_B_)   (64 - 8 * (_B_))
+#define UA_FIXUP_BFEXT_END(_B_)      63
+#endif /* __LITTLE_ENDIAN */
+
+
+
+	if ((ra != rb) && (rd != TREG_SP) && !alias &&
+	    !y1_br && !y1_lr && !x1_add) {
+		/*
+		 * Simple case: ra != rb and no register alias found,
+		 * and no branch or link. This will be the majority.
+		 * We can do a little better for simplae case than the
+		 * generic scheme below.
+		 */
+		if (!load_n_store) {
+			/*
+			 * Simple store: ra != rb, no need for scratch register.
+			 * Just store and rotate to right bytewise.
+			 */
+#ifdef __BIG_ENDIAN
+			frag.insn[n++] =
+				jit_x0_addi(ra, ra, load_store_size - 1) |
+				jit_x1_fnop();
+#endif /* __BIG_ENDIAN */
+			for (k = 0; k < load_store_size; k++) {
+				/* Store a byte. */
+				frag.insn[n++] =
+					jit_x0_rotli(rb, rb, 56) |
+					jit_x1_st1_add(ra, rb,
+						       UA_FIXUP_ADDR_DELTA);
+			}
+#ifdef __BIG_ENDIAN
+			frag.insn[n] = jit_x1_addi(ra, ra, 1);
+#else
+			frag.insn[n] = jit_x1_addi(ra, ra,
+						   -1 * load_store_size);
+#endif /* __LITTLE_ENDIAN */
+
+			if (load_store_size == 8) {
+				frag.insn[n] |= jit_x0_fnop();
+			} else if (load_store_size == 4) {
+				frag.insn[n] |= jit_x0_rotli(rb, rb, 32);
+			} else { /* = 2 */
+				frag.insn[n] |= jit_x0_rotli(rb, rb, 16);
+			}
+			n++;
+			if (bundle_2_enable)
+				frag.insn[n++] = bundle_2;
+			frag.insn[n++] = jit_x0_fnop() | jit_x1_iret();
+		} else {
+			if (rd == ra) {
+				/* Use two clobber registers: clob1/2. */
+				frag.insn[n++] =
+					jit_x0_addi(TREG_SP, TREG_SP, -16) |
+					jit_x1_fnop();
+				frag.insn[n++] =
+					jit_x0_addi(clob1, ra, 7) |
+					jit_x1_st_add(TREG_SP, clob1, -8);
+				frag.insn[n++] =
+					jit_x0_addi(clob2, ra, 0) |
+					jit_x1_st(TREG_SP, clob2);
+				frag.insn[n++] =
+					jit_x0_fnop() |
+					jit_x1_ldna(rd, ra);
+				frag.insn[n++] =
+					jit_x0_fnop() |
+					jit_x1_ldna(clob1, clob1);
+				/*
+				 * Note: we must make sure that rd must not
+				 * be sp. Recover clob1/2 from stack.
+				 */
+				frag.insn[n++] =
+					jit_x0_dblalign(rd, clob1, clob2) |
+					jit_x1_ld_add(clob2, TREG_SP, 8);
+				frag.insn[n++] =
+					jit_x0_fnop() |
+					jit_x1_ld_add(clob1, TREG_SP, 16);
+			} else {
+				/* Use one clobber register: clob1 only. */
+				frag.insn[n++] =
+					jit_x0_addi(TREG_SP, TREG_SP, -16) |
+					jit_x1_fnop();
+				frag.insn[n++] =
+					jit_x0_addi(clob1, ra, 7) |
+					jit_x1_st(TREG_SP, clob1);
+				frag.insn[n++] =
+					jit_x0_fnop() |
+					jit_x1_ldna(rd, ra);
+				frag.insn[n++] =
+					jit_x0_fnop() |
+					jit_x1_ldna(clob1, clob1);
+				/*
+				 * Note: we must make sure that rd must not
+				 * be sp. Recover clob1 from stack.
+				 */
+				frag.insn[n++] =
+					jit_x0_dblalign(rd, clob1, ra) |
+					jit_x1_ld_add(clob1, TREG_SP, 16);
+			}
+
+			if (bundle_2_enable)
+				frag.insn[n++] = bundle_2;
+			/*
+			 * For non 8-byte load, extract corresponding bytes and
+			 * signed extension.
+			 */
+			if (load_store_size == 4) {
+				if (load_store_signed)
+					frag.insn[n++] =
+						jit_x0_bfexts(
+							rd, rd,
+							UA_FIXUP_BFEXT_START(4),
+							UA_FIXUP_BFEXT_END(4)) |
+						jit_x1_fnop();
+				else
+					frag.insn[n++] =
+						jit_x0_bfextu(
+							rd, rd,
+							UA_FIXUP_BFEXT_START(4),
+							UA_FIXUP_BFEXT_END(4)) |
+						jit_x1_fnop();
+			} else if (load_store_size == 2) {
+				if (load_store_signed)
+					frag.insn[n++] =
+						jit_x0_bfexts(
+							rd, rd,
+							UA_FIXUP_BFEXT_START(2),
+							UA_FIXUP_BFEXT_END(2)) |
+						jit_x1_fnop();
+				else
+					frag.insn[n++] =
+						jit_x0_bfextu(
+							rd, rd,
+							UA_FIXUP_BFEXT_START(2),
+							UA_FIXUP_BFEXT_END(2)) |
+						jit_x1_fnop();
+			}
+
+			frag.insn[n++] =
+				jit_x0_fnop()  |
+				jit_x1_iret();
+		}
+	} else if (!load_n_store) {
+
+		/*
+		 * Generic memory store cases: use 3 clobber registers.
+		 *
+		 * Alloc space for saveing clob2,1,3 on user's stack.
+		 * register clob3 points to where clob2 saved, followed by
+		 * clob1 and 3 from high to low memory.
+		 */
+		frag.insn[n++] =
+			jit_x0_addi(TREG_SP, TREG_SP, -32)    |
+			jit_x1_fnop();
+		frag.insn[n++] =
+			jit_x0_addi(clob3, TREG_SP, 16)  |
+			jit_x1_st_add(TREG_SP, clob3, 8);
+#ifdef __LITTLE_ENDIAN
+		frag.insn[n++] =
+			jit_x0_addi(clob1, ra, 0)   |
+			jit_x1_st_add(TREG_SP, clob1, 8);
+#else
+		frag.insn[n++] =
+			jit_x0_addi(clob1, ra, load_store_size - 1)   |
+			jit_x1_st_add(TREG_SP, clob1, 8);
+#endif
+		if (load_store_size == 8) {
+			/*
+			 * We save one byte a time, not for fast, but compact
+			 * code. After each store, data source register shift
+			 * right one byte. unchanged after 8 stores.
+			 */
+			frag.insn[n++] =
+				jit_x0_addi(clob2, TREG_ZERO, 7)     |
+				jit_x1_st_add(TREG_SP, clob2, 16);
+			frag.insn[n++] =
+				jit_x0_rotli(rb, rb, 56)      |
+				jit_x1_st1_add(clob1, rb, UA_FIXUP_ADDR_DELTA);
+			frag.insn[n++] =
+				jit_x0_addi(clob2, clob2, -1) |
+				jit_x1_bnezt(clob2, -1);
+			frag.insn[n++] =
+				jit_x0_fnop()                 |
+				jit_x1_addi(clob2, y1_br_reg, 0);
+		} else if (load_store_size == 4) {
+			frag.insn[n++] =
+				jit_x0_addi(clob2, TREG_ZERO, 3)     |
+				jit_x1_st_add(TREG_SP, clob2, 16);
+			frag.insn[n++] =
+				jit_x0_rotli(rb, rb, 56)      |
+				jit_x1_st1_add(clob1, rb, UA_FIXUP_ADDR_DELTA);
+			frag.insn[n++] =
+				jit_x0_addi(clob2, clob2, -1) |
+				jit_x1_bnezt(clob2, -1);
+			/*
+			 * same as 8-byte case, but need shift another 4
+			 * byte to recover rb for 4-byte store.
+			 */
+			frag.insn[n++] = jit_x0_rotli(rb, rb, 32)      |
+				jit_x1_addi(clob2, y1_br_reg, 0);
+		} else { /* =2 */
+			frag.insn[n++] =
+				jit_x0_addi(clob2, rb, 0)     |
+				jit_x1_st_add(TREG_SP, clob2, 16);
+			for (k = 0; k < 2; k++) {
+				frag.insn[n++] =
+					jit_x0_shrui(rb, rb, 8)  |
+					jit_x1_st1_add(clob1, rb,
+						       UA_FIXUP_ADDR_DELTA);
+			}
+			frag.insn[n++] =
+				jit_x0_addi(rb, clob2, 0)       |
+				jit_x1_addi(clob2, y1_br_reg, 0);
+		}
+
+		if (bundle_2_enable)
+			frag.insn[n++] = bundle_2;
+
+		if (y1_lr) {
+			frag.insn[n++] =
+				jit_x0_fnop()                    |
+				jit_x1_mfspr(y1_lr_reg,
+					     SPR_EX_CONTEXT_0_0);
+		}
+		if (y1_br) {
+			frag.insn[n++] =
+				jit_x0_fnop()                    |
+				jit_x1_mtspr(SPR_EX_CONTEXT_0_0,
+					     clob2);
+		}
+		if (x1_add) {
+			frag.insn[n++] =
+				jit_x0_addi(ra, ra, x1_add_imm8) |
+				jit_x1_ld_add(clob2, clob3, -8);
+		} else {
+			frag.insn[n++] =
+				jit_x0_fnop()                    |
+				jit_x1_ld_add(clob2, clob3, -8);
+		}
+		frag.insn[n++] =
+			jit_x0_fnop()   |
+			jit_x1_ld_add(clob1, clob3, -8);
+		frag.insn[n++] = jit_x0_fnop()   | jit_x1_ld(clob3, clob3);
+		frag.insn[n++] = jit_x0_fnop()   | jit_x1_iret();
+
+	} else {
+		/*
+		 * Generic memory load cases.
+		 *
+		 * Alloc space for saveing clob1,2,3 on user's stack.
+		 * register clob3 points to where clob1 saved, followed
+		 * by clob2 and 3 from high to low memory.
+		 */
+
+		frag.insn[n++] =
+			jit_x0_addi(TREG_SP, TREG_SP, -32) |
+			jit_x1_fnop();
+		frag.insn[n++] =
+			jit_x0_addi(clob3, TREG_SP, 16) |
+			jit_x1_st_add(TREG_SP, clob3, 8);
+		frag.insn[n++] =
+			jit_x0_addi(clob2, ra, 0) |
+			jit_x1_st_add(TREG_SP, clob2, 8);
+
+		if (y1_br) {
+			frag.insn[n++] =
+				jit_x0_addi(clob1, y1_br_reg, 0) |
+				jit_x1_st_add(TREG_SP, clob1, 16);
+		} else {
+			frag.insn[n++] =
+				jit_x0_fnop() |
+				jit_x1_st_add(TREG_SP, clob1, 16);
+		}
+
+		if (bundle_2_enable)
+			frag.insn[n++] = bundle_2;
+
+		if (y1_lr) {
+			frag.insn[n++] =
+				jit_x0_fnop()  |
+				jit_x1_mfspr(y1_lr_reg,
+					     SPR_EX_CONTEXT_0_0);
+		}
+
+		if (y1_br) {
+			frag.insn[n++] =
+				jit_x0_fnop() |
+				jit_x1_mtspr(SPR_EX_CONTEXT_0_0,
+					     clob1);
+		}
+
+		frag.insn[n++] =
+			jit_x0_addi(clob1, clob2, 7)      |
+			jit_x1_ldna(rd, clob2);
+		frag.insn[n++] =
+			jit_x0_fnop()                     |
+			jit_x1_ldna(clob1, clob1);
+		frag.insn[n++] =
+			jit_x0_dblalign(rd, clob1, clob2) |
+			jit_x1_ld_add(clob1, clob3, -8);
+		if (x1_add) {
+			frag.insn[n++] =
+				jit_x0_addi(ra, ra, x1_add_imm8) |
+				jit_x1_ld_add(clob2, clob3, -8);
+		} else {
+			frag.insn[n++] =
+				jit_x0_fnop()  |
+				jit_x1_ld_add(clob2, clob3, -8);
+		}
+
+		frag.insn[n++] =
+			jit_x0_fnop() |
+			jit_x1_ld(clob3, clob3);
+
+		if (load_store_size == 4) {
+			if (load_store_signed)
+				frag.insn[n++] =
+					jit_x0_bfexts(
+						rd, rd,
+						UA_FIXUP_BFEXT_START(4),
+						UA_FIXUP_BFEXT_END(4)) |
+					jit_x1_fnop();
+			else
+				frag.insn[n++] =
+					jit_x0_bfextu(
+						rd, rd,
+						UA_FIXUP_BFEXT_START(4),
+						UA_FIXUP_BFEXT_END(4)) |
+					jit_x1_fnop();
+		} else if (load_store_size == 2) {
+			if (load_store_signed)
+				frag.insn[n++] =
+					jit_x0_bfexts(
+						rd, rd,
+						UA_FIXUP_BFEXT_START(2),
+						UA_FIXUP_BFEXT_END(2)) |
+					jit_x1_fnop();
+			else
+				frag.insn[n++] =
+					jit_x0_bfextu(
+						rd, rd,
+						UA_FIXUP_BFEXT_START(2),
+						UA_FIXUP_BFEXT_END(2)) |
+					jit_x1_fnop();
+		}
+
+		frag.insn[n++] = jit_x0_fnop() | jit_x1_iret();
+	}
+
+	/* Max JIT bundle count is 14. */
+	WARN_ON(n > 14);
+
+	if (!unexpected) {
+		int status = 0;
+		int idx = (regs->pc >> 3) &
+			((1ULL << (PAGE_SHIFT - UNALIGN_JIT_SHIFT)) - 1);
+
+		frag.pc = regs->pc;
+		frag.bundle = bundle;
+
+		if (unaligned_printk) {
+			pr_info("%s/%d, Unalign fixup: pc=%lx bundle=%lx %d %d %d %d %d %d %d %d\n",
+				current->comm, current->pid,
+				(unsigned long)frag.pc,
+				(unsigned long)frag.bundle,
+				(int)alias, (int)rd, (int)ra,
+				(int)rb, (int)bundle_2_enable,
+				(int)y1_lr, (int)y1_br, (int)x1_add);
+
+			for (k = 0; k < n; k += 2)
+				pr_info("[%d] %016llx %016llx\n",
+					k, (unsigned long long)frag.insn[k],
+					(unsigned long long)frag.insn[k+1]);
+		}
+
+		/* Swap bundle byte order for big endian sys. */
+#ifdef __BIG_ENDIAN
+		frag.bundle = GX_INSN_BSWAP(frag.bundle);
+		for (k = 0; k < n; k++)
+			frag.insn[k] = GX_INSN_BSWAP(frag.insn[k]);
+#endif /* __BIG_ENDIAN */
+
+		status = copy_to_user((void __user *)&jit_code_area[idx],
+				      &frag, sizeof(frag));
+		if (status) {
+			/* Fail to copy JIT into user land. send SIGSEGV. */
+			siginfo_t info = {
+				.si_signo = SIGSEGV,
+				.si_code = SEGV_MAPERR,
+				.si_addr = (void __user *)&jit_code_area[idx]
+			};
+
+			pr_warn("Unalign fixup: pid=%d %s jit_code_area=%llx\n",
+				current->pid, current->comm,
+				(unsigned long long)&jit_code_area[idx]);
+
+			trace_unhandled_signal("segfault in unalign fixup",
+					       regs,
+					       (unsigned long)info.si_addr,
+					       SIGSEGV);
+			force_sig_info(info.si_signo, &info, current);
+			return;
+		}
+
+
+		/* Do a cheaper increment, not accurate. */
+		unaligned_fixup_count++;
+		__flush_icache_range((unsigned long)&jit_code_area[idx],
+				     (unsigned long)&jit_code_area[idx] +
+				     sizeof(frag));
+
+		/* Setup SPR_EX_CONTEXT_0_0/1 for returning to user program.*/
+		__insn_mtspr(SPR_EX_CONTEXT_0_0, regs->pc + 8);
+		__insn_mtspr(SPR_EX_CONTEXT_0_1, PL_ICS_EX1(USER_PL, 0));
+
+		/* Modify pc at the start of new JIT. */
+		regs->pc = (unsigned long)&jit_code_area[idx].insn[0];
+		/* Set ICS in SPR_EX_CONTEXT_K_1. */
+		regs->ex1 = PL_ICS_EX1(USER_PL, 1);
+	}
+}
+
+
+/*
+ * C function to generate unalign data JIT. Called from unalign data
+ * interrupt handler.
+ *
+ * First check if unalign fix is disabled or exception did not not come from
+ * user space or sp register points to unalign address, if true, generate a
+ * SIGBUS. Then map a page into user space as JIT area if it is not mapped
+ * yet. Genenerate JIT code by calling jit_bundle_gen(). After that return
+ * back to exception handler.
+ *
+ * The exception handler will "iret" to new generated JIT code after
+ * restoring caller saved registers. In theory, the JIT code will perform
+ * another "iret" to resume user's program.
+ */
+
+void do_unaligned(struct pt_regs *regs, int vecnum)
+{
+	enum ctx_state prev_state = exception_enter();
+	tilegx_bundle_bits __user  *pc;
+	tilegx_bundle_bits bundle;
+	struct thread_info *info = current_thread_info();
+	int align_ctl;
+
+	/* Checks the per-process unaligned JIT flags */
+	align_ctl = unaligned_fixup;
+	switch (task_thread_info(current)->align_ctl) {
+	case PR_UNALIGN_NOPRINT:
+		align_ctl = 1;
+		break;
+	case PR_UNALIGN_SIGBUS:
+		align_ctl = 0;
+		break;
+	}
+
+	/* Enable iterrupt in order to access user land. */
+	local_irq_enable();
+
+	/*
+	 * The fault came from kernel space. Two choices:
+	 * (a) unaligned_fixup < 1, we will first call get/put_user fixup
+	 *     to return -EFAULT. If no fixup, simply panic the kernel.
+	 * (b) unaligned_fixup >=1, we will try to fix the unaligned access
+	 *     if it was triggered by get_user/put_user() macros. Panic the
+	 *     kernel if it is not fixable.
+	 */
+
+	if (EX1_PL(regs->ex1) != USER_PL) {
+
+		if (align_ctl < 1) {
+			unaligned_fixup_count++;
+			/* If exception came from kernel, try fix it up. */
+			if (fixup_exception(regs)) {
+				if (unaligned_printk)
+					pr_info("Unalign fixup: %d %llx @%llx\n",
+						(int)unaligned_fixup,
+						(unsigned long long)regs->ex1,
+						(unsigned long long)regs->pc);
+			} else {
+				/* Not fixable. Go panic. */
+				panic("Unalign exception in Kernel. pc=%lx",
+				      regs->pc);
+			}
+		} else {
+			/*
+			 * Try to fix the exception. If we can't, panic the
+			 * kernel.
+			 */
+			bundle = GX_INSN_BSWAP(
+				*((tilegx_bundle_bits *)(regs->pc)));
+			jit_bundle_gen(regs, bundle, align_ctl);
+		}
+		goto done;
+	}
+
+	/*
+	 * Fault came from user with ICS or stack is not aligned.
+	 * If so, we will trigger SIGBUS.
+	 */
+	if ((regs->sp & 0x7) || (regs->ex1) || (align_ctl < 0)) {
+		siginfo_t info = {
+			.si_signo = SIGBUS,
+			.si_code = BUS_ADRALN,
+			.si_addr = (unsigned char __user *)0
+		};
+
+		if (unaligned_printk)
+			pr_info("Unalign fixup: %d %llx @%llx\n",
+				(int)unaligned_fixup,
+				(unsigned long long)regs->ex1,
+				(unsigned long long)regs->pc);
+
+		unaligned_fixup_count++;
+
+		trace_unhandled_signal("unaligned fixup trap", regs, 0, SIGBUS);
+		force_sig_info(info.si_signo, &info, current);
+		goto done;
+	}
+
+
+	/* Read the bundle casued the exception! */
+	pc = (tilegx_bundle_bits __user *)(regs->pc);
+	if (get_user(bundle, pc) != 0) {
+		/* Probably never be here since pc is valid user address.*/
+		siginfo_t info = {
+			.si_signo = SIGSEGV,
+			.si_code = SEGV_MAPERR,
+			.si_addr = (void __user *)pc
+		};
+		pr_err("Couldn't read instruction at %p trying to step\n", pc);
+		trace_unhandled_signal("segfault in unalign fixup", regs,
+				       (unsigned long)info.si_addr, SIGSEGV);
+		force_sig_info(info.si_signo, &info, current);
+		goto done;
+	}
+
+	if (!info->unalign_jit_base) {
+		void __user *user_page;
+
+		/*
+		 * Allocate a page in userland.
+		 * For 64-bit processes we try to place the mapping far
+		 * from anything else that might be going on (specifically
+		 * 64 GB below the top of the user address space).  If it
+		 * happens not to be possible to put it there, it's OK;
+		 * the kernel will choose another location and we'll
+		 * remember it for later.
+		 */
+		if (is_compat_task())
+			user_page = NULL;
+		else
+			user_page = (void __user *)(TASK_SIZE - (1UL << 36)) +
+				(current->pid << PAGE_SHIFT);
+
+		user_page = (void __user *) vm_mmap(NULL,
+						    (unsigned long)user_page,
+						    PAGE_SIZE,
+						    PROT_EXEC | PROT_READ |
+						    PROT_WRITE,
+#ifdef CONFIG_HOMECACHE
+						    MAP_CACHE_HOME_TASK |
+#endif
+						    MAP_PRIVATE |
+						    MAP_ANONYMOUS,
+						    0);
+
+		if (IS_ERR((void __force *)user_page)) {
+			pr_err("Out of kernel pages trying do_mmap\n");
+			goto done;
+		}
+
+		/* Save the address in the thread_info struct */
+		info->unalign_jit_base = user_page;
+		if (unaligned_printk)
+			pr_info("Unalign bundle: %d:%d, allocate page @%llx\n",
+				raw_smp_processor_id(), current->pid,
+				(unsigned long long)user_page);
+	}
+
+	/* Generate unalign JIT */
+	jit_bundle_gen(regs, GX_INSN_BSWAP(bundle), align_ctl);
+
+done:
+	exception_exit(prev_state);
+}
+
+#endif /* __tilegx__ */
diff --git a/kernel/arch/tile/kernel/usb.c b/kernel/arch/tile/kernel/usb.c
new file mode 100644
index 000000000..5af8debc6
--- /dev/null
+++ b/kernel/arch/tile/kernel/usb.c
@@ -0,0 +1,69 @@
+/*
+ * Copyright 2012 Tilera Corporation. All Rights Reserved.
+ *
+ *   This program is free software; you can redistribute it and/or
+ *   modify it under the terms of the GNU General Public License
+ *   as published by the Free Software Foundation, version 2.
+ *
+ *   This program is distributed in the hope that it will be useful, but
+ *   WITHOUT ANY WARRANTY; without even the implied warranty of
+ *   MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
+ *   NON INFRINGEMENT.  See the GNU General Public License for
+ *   more details.
+ *
+ * Register the Tile-Gx USB interfaces as platform devices.
+ *
+ * The actual USB driver is just some glue (in
+ * drivers/usb/host/[eo]hci-tilegx.c) which makes the registers available
+ * to the standard kernel EHCI and OHCI drivers.
+ */
+
+#include <linux/dma-mapping.h>
+#include <linux/platform_device.h>
+#include <linux/usb/tilegx.h>
+#include <linux/types.h>
+
+static u64 ehci_dmamask = DMA_BIT_MASK(32);
+
+#define USB_HOST_DEF(unit, type, dmamask) \
+	static struct \
+	    tilegx_usb_platform_data tilegx_usb_platform_data_ ## type ## \
+		hci ## unit = { \
+		.dev_index = unit, \
+	}; \
+	\
+	static struct platform_device tilegx_usb_ ## type ## hci ## unit = { \
+		.name		= "tilegx-" #type "hci", \
+		.id		= unit, \
+		.dev = { \
+			.dma_mask		= dmamask, \
+			.coherent_dma_mask	= DMA_BIT_MASK(32), \
+			.platform_data = \
+				&tilegx_usb_platform_data_ ## type ## hci ## \
+				unit, \
+		}, \
+	};
+
+USB_HOST_DEF(0, e, &ehci_dmamask)
+USB_HOST_DEF(0, o, NULL)
+USB_HOST_DEF(1, e, &ehci_dmamask)
+USB_HOST_DEF(1, o, NULL)
+
+#undef USB_HOST_DEF
+
+static struct platform_device *tilegx_usb_devices[] __initdata = {
+	&tilegx_usb_ehci0,
+	&tilegx_usb_ehci1,
+	&tilegx_usb_ohci0,
+	&tilegx_usb_ohci1,
+};
+
+/** Add our set of possible USB devices. */
+static int __init tilegx_usb_init(void)
+{
+	platform_add_devices(tilegx_usb_devices,
+			     ARRAY_SIZE(tilegx_usb_devices));
+
+	return 0;
+}
+arch_initcall(tilegx_usb_init);
diff --git a/kernel/arch/tile/kernel/vdso.c b/kernel/arch/tile/kernel/vdso.c
new file mode 100644
index 000000000..5bc51d7df
--- /dev/null
+++ b/kernel/arch/tile/kernel/vdso.c
@@ -0,0 +1,197 @@
+/*
+ * Copyright 2012 Tilera Corporation. All Rights Reserved.
+ *
+ *   This program is free software; you can redistribute it and/or
+ *   modify it under the terms of the GNU General Public License
+ *   as published by the Free Software Foundation, version 2.
+ *
+ *   This program is distributed in the hope that it will be useful, but
+ *   WITHOUT ANY WARRANTY; without even the implied warranty of
+ *   MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
+ *   NON INFRINGEMENT.  See the GNU General Public License for
+ *   more details.
+ */
+
+#include <linux/binfmts.h>
+#include <linux/compat.h>
+#include <linux/elf.h>
+#include <linux/mm.h>
+#include <linux/pagemap.h>
+
+#include <asm/vdso.h>
+#include <asm/mman.h>
+#include <asm/sections.h>
+
+#include <arch/sim.h>
+
+/* The alignment of the vDSO. */
+#define VDSO_ALIGNMENT  PAGE_SIZE
+
+
+static unsigned int vdso_pages;
+static struct page **vdso_pagelist;
+
+#ifdef CONFIG_COMPAT
+static unsigned int vdso32_pages;
+static struct page **vdso32_pagelist;
+#endif
+static int vdso_ready;
+
+/*
+ * The vdso data page.
+ */
+static union {
+	struct vdso_data	data;
+	u8			page[PAGE_SIZE];
+} vdso_data_store __page_aligned_data;
+
+struct vdso_data *vdso_data = &vdso_data_store.data;
+
+static unsigned int __read_mostly vdso_enabled = 1;
+
+static struct page **vdso_setup(void *vdso_kbase, unsigned int pages)
+{
+	int i;
+	struct page **pagelist;
+
+	pagelist = kzalloc(sizeof(struct page *) * (pages + 1), GFP_KERNEL);
+	BUG_ON(pagelist == NULL);
+	for (i = 0; i < pages - 1; i++) {
+		struct page *pg = virt_to_page(vdso_kbase + i*PAGE_SIZE);
+		ClearPageReserved(pg);
+		pagelist[i] = pg;
+	}
+	pagelist[pages - 1] = virt_to_page(vdso_data);
+	pagelist[pages] = NULL;
+
+	return pagelist;
+}
+
+static int __init vdso_init(void)
+{
+	int data_pages = sizeof(vdso_data_store) >> PAGE_SHIFT;
+
+	/*
+	 * We can disable vDSO support generally, but we need to retain
+	 * one page to support the two-bundle (16-byte) rt_sigreturn path.
+	 */
+	if (!vdso_enabled) {
+		size_t offset = (unsigned long)&__vdso_rt_sigreturn;
+		static struct page *sigret_page;
+		sigret_page = alloc_page(GFP_KERNEL | __GFP_ZERO);
+		BUG_ON(sigret_page == NULL);
+		vdso_pagelist = &sigret_page;
+		vdso_pages = 1;
+		BUG_ON(offset >= PAGE_SIZE);
+		memcpy(page_address(sigret_page) + offset,
+		       vdso_start + offset, 16);
+#ifdef CONFIG_COMPAT
+		vdso32_pages = vdso_pages;
+		vdso32_pagelist = vdso_pagelist;
+#endif
+		vdso_ready = 1;
+		return 0;
+	}
+
+	vdso_pages = (vdso_end - vdso_start) >> PAGE_SHIFT;
+	vdso_pages += data_pages;
+	vdso_pagelist = vdso_setup(vdso_start, vdso_pages);
+
+#ifdef CONFIG_COMPAT
+	vdso32_pages = (vdso32_end - vdso32_start) >> PAGE_SHIFT;
+	vdso32_pages += data_pages;
+	vdso32_pagelist = vdso_setup(vdso32_start, vdso32_pages);
+#endif
+
+	smp_wmb();
+	vdso_ready = 1;
+
+	return 0;
+}
+arch_initcall(vdso_init);
+
+const char *arch_vma_name(struct vm_area_struct *vma)
+{
+	if (vma->vm_mm && vma->vm_start == VDSO_BASE)
+		return "[vdso]";
+#ifndef __tilegx__
+	if (vma->vm_start == MEM_USER_INTRPT)
+		return "[intrpt]";
+#endif
+	return NULL;
+}
+
+int setup_vdso_pages(void)
+{
+	struct page **pagelist;
+	unsigned long pages;
+	struct mm_struct *mm = current->mm;
+	unsigned long vdso_base = 0;
+	int retval = 0;
+
+	if (!vdso_ready)
+		return 0;
+
+	mm->context.vdso_base = 0;
+
+	pagelist = vdso_pagelist;
+	pages = vdso_pages;
+#ifdef CONFIG_COMPAT
+	if (is_compat_task()) {
+		pagelist = vdso32_pagelist;
+		pages = vdso32_pages;
+	}
+#endif
+
+	/*
+	 * vDSO has a problem and was disabled, just don't "enable" it for the
+	 * process.
+	 */
+	if (pages == 0)
+		return 0;
+
+	vdso_base = get_unmapped_area(NULL, vdso_base,
+				      (pages << PAGE_SHIFT) +
+				      ((VDSO_ALIGNMENT - 1) & PAGE_MASK),
+				      0, 0);
+	if (IS_ERR_VALUE(vdso_base)) {
+		retval = vdso_base;
+		return retval;
+	}
+
+	/* Add required alignment. */
+	vdso_base = ALIGN(vdso_base, VDSO_ALIGNMENT);
+
+	/*
+	 * Put vDSO base into mm struct. We need to do this before calling
+	 * install_special_mapping or the perf counter mmap tracking code
+	 * will fail to recognise it as a vDSO (since arch_vma_name fails).
+	 */
+	mm->context.vdso_base = vdso_base;
+
+	/*
+	 * our vma flags don't have VM_WRITE so by default, the process isn't
+	 * allowed to write those pages.
+	 * gdb can break that with ptrace interface, and thus trigger COW on
+	 * those pages but it's then your responsibility to never do that on
+	 * the "data" page of the vDSO or you'll stop getting kernel updates
+	 * and your nice userland gettimeofday will be totally dead.
+	 * It's fine to use that for setting breakpoints in the vDSO code
+	 * pages though
+	 */
+	retval = install_special_mapping(mm, vdso_base,
+					 pages << PAGE_SHIFT,
+					 VM_READ|VM_EXEC |
+					 VM_MAYREAD | VM_MAYWRITE | VM_MAYEXEC,
+					 pagelist);
+	if (retval)
+		mm->context.vdso_base = 0;
+
+	return retval;
+}
+
+static __init int vdso_func(char *s)
+{
+	return kstrtouint(s, 0, &vdso_enabled);
+}
+__setup("vdso=", vdso_func);
diff --git a/kernel/arch/tile/kernel/vdso/Makefile b/kernel/arch/tile/kernel/vdso/Makefile
new file mode 100644
index 000000000..a025f63d5
--- /dev/null
+++ b/kernel/arch/tile/kernel/vdso/Makefile
@@ -0,0 +1,118 @@
+# Symbols present in the vdso
+vdso-syms = rt_sigreturn gettimeofday
+
+# Files to link into the vdso
+obj-vdso = $(patsubst %, v%.o, $(vdso-syms))
+
+# Build rules
+targets := $(obj-vdso) vdso.so vdso.so.dbg vdso.lds
+obj-vdso := $(addprefix $(obj)/, $(obj-vdso))
+
+# vdso32 is only for tilegx -m32 compat task.
+VDSO32-$(CONFIG_COMPAT) := y
+
+obj-y += vdso.o
+obj-$(VDSO32-y) += vdso32.o
+extra-y += vdso.lds
+CPPFLAGS_vdso.lds += -P -C -U$(ARCH)
+
+# vDSO code runs in userspace and -pg doesn't help with profiling anyway.
+CFLAGS_REMOVE_vdso.o = -pg
+CFLAGS_REMOVE_vdso32.o = -pg
+CFLAGS_REMOVE_vrt_sigreturn.o = -pg
+CFLAGS_REMOVE_vrt_sigreturn32.o = -pg
+CFLAGS_REMOVE_vgettimeofday.o = -pg
+CFLAGS_REMOVE_vgettimeofday32.o = -pg
+
+ifdef CONFIG_FEEDBACK_COLLECT
+# vDSO code runs in userspace, not collecting feedback data.
+CFLAGS_REMOVE_vdso.o = -ffeedback-generate
+CFLAGS_REMOVE_vdso32.o = -ffeedback-generate
+CFLAGS_REMOVE_vrt_sigreturn.o = -ffeedback-generate
+CFLAGS_REMOVE_vrt_sigreturn32.o = -ffeedback-generate
+CFLAGS_REMOVE_vgettimeofday.o = -ffeedback-generate
+CFLAGS_REMOVE_vgettimeofday32.o = -ffeedback-generate
+endif
+
+# Disable gcov profiling for VDSO code
+GCOV_PROFILE := n
+
+# Force dependency
+$(obj)/vdso.o: $(obj)/vdso.so
+
+# link rule for the .so file, .lds has to be first
+SYSCFLAGS_vdso.so.dbg = $(c_flags)
+$(obj)/vdso.so.dbg: $(src)/vdso.lds $(obj-vdso)
+	$(call if_changed,vdsold)
+
+
+# We also create a special relocatable object that should mirror the symbol
+# table and layout of the linked DSO.  With ld -R we can then refer to
+# these symbols in the kernel code rather than hand-coded addresses.
+extra-y += vdso-syms.o
+$(obj)/built-in.o: $(obj)/vdso-syms.o
+$(obj)/built-in.o: ld_flags += -R $(obj)/vdso-syms.o
+
+SYSCFLAGS_vdso.so.dbg = -shared -s -Wl,-soname=linux-vdso.so.1 \
+                            $(call cc-ldoption, -Wl$(comma)--hash-style=sysv)
+SYSCFLAGS_vdso_syms.o = -r
+$(obj)/vdso-syms.o: $(src)/vdso.lds $(obj)/vrt_sigreturn.o FORCE
+	$(call if_changed,vdsold)
+
+
+# strip rule for the .so file
+$(obj)/%.so: OBJCOPYFLAGS := -S
+$(obj)/%.so: $(obj)/%.so.dbg FORCE
+	$(call if_changed,objcopy)
+
+# actual build commands
+# The DSO images are built using a special linker script
+# Add -lgcc so tilepro gets static muldi3 and lshrdi3 definitions.
+# Make sure only to export the intended __vdso_xxx symbol offsets.
+quiet_cmd_vdsold = VDSOLD  $@
+      cmd_vdsold = $(CC) $(KCFLAGS) -nostdlib $(SYSCFLAGS_$(@F)) \
+                           -Wl,-T,$(filter-out FORCE,$^) -o $@.tmp -lgcc && \
+                   $(CROSS_COMPILE)objcopy \
+                           $(patsubst %, -G __vdso_%, $(vdso-syms)) $@.tmp $@
+
+# install commands for the unstripped file
+quiet_cmd_vdso_install = INSTALL $@
+      cmd_vdso_install = cp $(obj)/$@.dbg $(MODLIB)/vdso/$@
+
+vdso.so: $(obj)/vdso.so.dbg
+	@mkdir -p $(MODLIB)/vdso
+	$(call cmd,vdso_install)
+
+vdso32.so: $(obj)/vdso32.so.dbg
+	$(call cmd,vdso_install)
+
+vdso_install: vdso.so
+vdso32_install: vdso32.so
+
+
+KBUILD_AFLAGS_32 := $(filter-out -m64,$(KBUILD_AFLAGS))
+KBUILD_AFLAGS_32 += -m32 -s
+KBUILD_CFLAGS_32 := $(filter-out -m64,$(KBUILD_CFLAGS))
+KBUILD_CFLAGS_32 += -m32 -fPIC -shared
+
+obj-vdso32 = $(patsubst %, v%32.o, $(vdso-syms))
+obj-vdso32 := $(addprefix $(obj)/, $(obj-vdso32))
+
+targets += $(obj-vdso32) vdso32.so vdso32.so.dbg
+
+$(obj-vdso32:%=%): KBUILD_AFLAGS = $(KBUILD_AFLAGS_32)
+$(obj-vdso32:%=%): KBUILD_CFLAGS = $(KBUILD_CFLAGS_32)
+
+$(obj)/vgettimeofday32.o: $(obj)/vgettimeofday.c
+	$(call if_changed_rule,cc_o_c)
+
+$(obj)/vrt_sigreturn32.o: $(obj)/vrt_sigreturn.S
+	$(call if_changed,as_o_S)
+
+# Force dependency
+$(obj)/vdso32.o: $(obj)/vdso32.so
+
+SYSCFLAGS_vdso32.so.dbg = -m32 -shared -s -Wl,-soname=linux-vdso32.so.1 \
+			    $(call cc-ldoption, -Wl$(comma)--hash-style=sysv)
+$(obj)/vdso32.so.dbg: $(src)/vdso.lds $(obj-vdso32)
+	$(call if_changed,vdsold)
diff --git a/kernel/arch/tile/kernel/vdso/vdso.S b/kernel/arch/tile/kernel/vdso/vdso.S
new file mode 100644
index 000000000..3467adb41
--- /dev/null
+++ b/kernel/arch/tile/kernel/vdso/vdso.S
@@ -0,0 +1,28 @@
+/*
+ * Copyright 2012 Tilera Corporation. All Rights Reserved.
+ *
+ *   This program is free software; you can redistribute it and/or
+ *   modify it under the terms of the GNU General Public License
+ *   as published by the Free Software Foundation, version 2.
+ *
+ *   This program is distributed in the hope that it will be useful, but
+ *   WITHOUT ANY WARRANTY; without even the implied warranty of
+ *   MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
+ *   NON INFRINGEMENT.  See the GNU General Public License for
+ *   more details.
+ */
+
+#include <linux/init.h>
+#include <linux/linkage.h>
+#include <asm/page.h>
+
+	__PAGE_ALIGNED_DATA
+
+	.global vdso_start, vdso_end
+	.align PAGE_SIZE
+vdso_start:
+	.incbin "arch/tile/kernel/vdso/vdso.so"
+	.align PAGE_SIZE
+vdso_end:
+
+	.previous
diff --git a/kernel/arch/tile/kernel/vdso/vdso.lds.S b/kernel/arch/tile/kernel/vdso/vdso.lds.S
new file mode 100644
index 000000000..731529f3f
--- /dev/null
+++ b/kernel/arch/tile/kernel/vdso/vdso.lds.S
@@ -0,0 +1,89 @@
+/*
+ * Copyright 2012 Tilera Corporation. All Rights Reserved.
+ *
+ *   This program is free software; you can redistribute it and/or
+ *   modify it under the terms of the GNU General Public License
+ *   as published by the Free Software Foundation, version 2.
+ *
+ *   This program is distributed in the hope that it will be useful, but
+ *   WITHOUT ANY WARRANTY; without even the implied warranty of
+ *   MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
+ *   NON INFRINGEMENT.  See the GNU General Public License for
+ *   more details.
+ */
+
+#define VDSO_VERSION_STRING	LINUX_2.6
+
+
+OUTPUT_ARCH(tile)
+
+/* The ELF entry point can be used to set the AT_SYSINFO value. */
+ENTRY(__vdso_rt_sigreturn);
+
+
+SECTIONS
+{
+	. = SIZEOF_HEADERS;
+
+	.hash		: { *(.hash) }			:text
+	.gnu.hash	: { *(.gnu.hash) }
+	.dynsym		: { *(.dynsym) }
+	.dynstr		: { *(.dynstr) }
+	.gnu.version	: { *(.gnu.version) }
+	.gnu.version_d	: { *(.gnu.version_d) }
+	.gnu.version_r	: { *(.gnu.version_r) }
+
+	.note		: { *(.note.*) }		:text	:note
+	.dynamic	: { *(.dynamic) }		:text	:dynamic
+
+	.eh_frame_hdr	: { *(.eh_frame_hdr) }		:text	:eh_frame_hdr
+	.eh_frame	: { KEEP (*(.eh_frame)) }	:text
+
+	.rodata	 : { *(.rodata .rodata.* .gnu.linkonce.r.*) }
+
+	/*
+	 * This linker script is used both with -r and with -shared.
+	 * For the layouts to match, we need to skip more than enough
+	 * space for the dynamic symbol table et al. If this amount
+	 * is insufficient, ld -shared will barf. Just increase it here.
+	 */
+	. = 0x1000;
+	.text		: { *(.text .text.*) }		:text
+
+	.data		: {
+		*(.got.plt) *(.got)
+		*(.data .data.* .gnu.linkonce.d.*)
+		*(.dynbss)
+		*(.bss .bss.* .gnu.linkonce.b.*)
+	}
+}
+
+
+/*
+ * We must supply the ELF program headers explicitly to get just one
+ * PT_LOAD segment, and set the flags explicitly to make segments read-only.
+ */
+PHDRS
+{
+	text		PT_LOAD		FLAGS(5) FILEHDR PHDRS;	/* PF_R|PF_X */
+	dynamic		PT_DYNAMIC	FLAGS(4);		/* PF_R */
+	note		PT_NOTE		FLAGS(4);		/* PF_R */
+	eh_frame_hdr	PT_GNU_EH_FRAME;
+}
+
+
+/*
+ * This controls what userland symbols we export from the vDSO.
+ */
+VERSION
+{
+	VDSO_VERSION_STRING {
+	global:
+		__vdso_rt_sigreturn;
+		__vdso_gettimeofday;
+		gettimeofday;
+		__vdso_clock_gettime;
+		clock_gettime;
+	local:*;
+	};
+}
diff --git a/kernel/arch/tile/kernel/vdso/vdso32.S b/kernel/arch/tile/kernel/vdso/vdso32.S
new file mode 100644
index 000000000..1d1ac3257
--- /dev/null
+++ b/kernel/arch/tile/kernel/vdso/vdso32.S
@@ -0,0 +1,28 @@
+/*
+ * Copyright 2013 Tilera Corporation. All Rights Reserved.
+ *
+ *   This program is free software; you can redistribute it and/or
+ *   modify it under the terms of the GNU General Public License
+ *   as published by the Free Software Foundation, version 2.
+ *
+ *   This program is distributed in the hope that it will be useful, but
+ *   WITHOUT ANY WARRANTY; without even the implied warranty of
+ *   MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
+ *   NON INFRINGEMENT.  See the GNU General Public License for
+ *   more details.
+ */
+
+#include <linux/init.h>
+#include <linux/linkage.h>
+#include <asm/page.h>
+
+	__PAGE_ALIGNED_DATA
+
+	.global vdso32_start, vdso32_end
+	.align PAGE_SIZE
+vdso32_start:
+	.incbin "arch/tile/kernel/vdso/vdso32.so"
+	.align PAGE_SIZE
+vdso32_end:
+
+	.previous
diff --git a/kernel/arch/tile/kernel/vdso/vgettimeofday.c b/kernel/arch/tile/kernel/vdso/vgettimeofday.c
new file mode 100644
index 000000000..8bb21eda0
--- /dev/null
+++ b/kernel/arch/tile/kernel/vdso/vgettimeofday.c
@@ -0,0 +1,198 @@
+/*
+ * Copyright 2012 Tilera Corporation. All Rights Reserved.
+ *
+ *   This program is free software; you can redistribute it and/or
+ *   modify it under the terms of the GNU General Public License
+ *   as published by the Free Software Foundation, version 2.
+ *
+ *   This program is distributed in the hope that it will be useful, but
+ *   WITHOUT ANY WARRANTY; without even the implied warranty of
+ *   MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
+ *   NON INFRINGEMENT.  See the GNU General Public License for
+ *   more details.
+ */
+
+#define VDSO_BUILD  /* avoid some shift warnings for -m32 in <asm/page.h> */
+#include <linux/time.h>
+#include <asm/timex.h>
+#include <asm/unistd.h>
+#include <asm/vdso.h>
+
+#if CHIP_HAS_SPLIT_CYCLE()
+static inline cycles_t get_cycles_inline(void)
+{
+	unsigned int high = __insn_mfspr(SPR_CYCLE_HIGH);
+	unsigned int low = __insn_mfspr(SPR_CYCLE_LOW);
+	unsigned int high2 = __insn_mfspr(SPR_CYCLE_HIGH);
+
+	while (unlikely(high != high2)) {
+		low = __insn_mfspr(SPR_CYCLE_LOW);
+		high = high2;
+		high2 = __insn_mfspr(SPR_CYCLE_HIGH);
+	}
+
+	return (((cycles_t)high) << 32) | low;
+}
+#define get_cycles get_cycles_inline
+#endif
+
+struct syscall_return_value {
+	long value;
+	long error;
+};
+
+/*
+ * Find out the vDSO data page address in the process address space.
+ */
+inline unsigned long get_datapage(void)
+{
+	unsigned long ret;
+
+	/* vdso data page located in the 2nd vDSO page. */
+	asm volatile ("lnk %0" : "=r"(ret));
+	ret &= ~(PAGE_SIZE - 1);
+	ret += PAGE_SIZE;
+
+	return ret;
+}
+
+static inline u64 vgetsns(struct vdso_data *vdso)
+{
+	return ((get_cycles() - vdso->cycle_last) & vdso->mask) * vdso->mult;
+}
+
+static inline int do_realtime(struct vdso_data *vdso, struct timespec *ts)
+{
+	unsigned count;
+	u64 ns;
+
+	do {
+		count = read_seqcount_begin(&vdso->tb_seq);
+		ts->tv_sec = vdso->wall_time_sec;
+		ns = vdso->wall_time_snsec;
+		ns += vgetsns(vdso);
+		ns >>= vdso->shift;
+	} while (unlikely(read_seqcount_retry(&vdso->tb_seq, count)));
+
+	ts->tv_sec += __iter_div_u64_rem(ns, NSEC_PER_SEC, &ns);
+	ts->tv_nsec = ns;
+
+	return 0;
+}
+
+static inline int do_monotonic(struct vdso_data *vdso, struct timespec *ts)
+{
+	unsigned count;
+	u64 ns;
+
+	do {
+		count = read_seqcount_begin(&vdso->tb_seq);
+		ts->tv_sec = vdso->monotonic_time_sec;
+		ns = vdso->monotonic_time_snsec;
+		ns += vgetsns(vdso);
+		ns >>= vdso->shift;
+	} while (unlikely(read_seqcount_retry(&vdso->tb_seq, count)));
+
+	ts->tv_sec += __iter_div_u64_rem(ns, NSEC_PER_SEC, &ns);
+	ts->tv_nsec = ns;
+
+	return 0;
+}
+
+static inline int do_realtime_coarse(struct vdso_data *vdso,
+				     struct timespec *ts)
+{
+	unsigned count;
+
+	do {
+		count = read_seqcount_begin(&vdso->tb_seq);
+		ts->tv_sec = vdso->wall_time_coarse_sec;
+		ts->tv_nsec = vdso->wall_time_coarse_nsec;
+	} while (unlikely(read_seqcount_retry(&vdso->tb_seq, count)));
+
+	return 0;
+}
+
+static inline int do_monotonic_coarse(struct vdso_data *vdso,
+				      struct timespec *ts)
+{
+	unsigned count;
+
+	do {
+		count = read_seqcount_begin(&vdso->tb_seq);
+		ts->tv_sec = vdso->monotonic_time_coarse_sec;
+		ts->tv_nsec = vdso->monotonic_time_coarse_nsec;
+	} while (unlikely(read_seqcount_retry(&vdso->tb_seq, count)));
+
+	return 0;
+}
+
+struct syscall_return_value __vdso_gettimeofday(struct timeval *tv,
+						struct timezone *tz)
+{
+	struct syscall_return_value ret = { 0, 0 };
+	unsigned count;
+	struct vdso_data *vdso = (struct vdso_data *)get_datapage();
+
+	/* The use of the timezone is obsolete, normally tz is NULL. */
+	if (unlikely(tz != NULL)) {
+		do {
+			count = read_seqcount_begin(&vdso->tz_seq);
+			tz->tz_minuteswest = vdso->tz_minuteswest;
+			tz->tz_dsttime = vdso->tz_dsttime;
+		} while (unlikely(read_seqcount_retry(&vdso->tz_seq, count)));
+	}
+
+	if (unlikely(tv == NULL))
+		return ret;
+
+	do_realtime(vdso, (struct timespec *)tv);
+	tv->tv_usec /= 1000;
+
+	return ret;
+}
+
+int gettimeofday(struct timeval *tv, struct timezone *tz)
+	__attribute__((weak, alias("__vdso_gettimeofday")));
+
+static struct syscall_return_value vdso_fallback_gettime(long clock,
+							 struct timespec *ts)
+{
+	struct syscall_return_value ret;
+	__asm__ __volatile__ (
+		"swint1"
+		: "=R00" (ret.value), "=R01" (ret.error)
+		: "R10" (__NR_clock_gettime), "R00" (clock), "R01" (ts)
+		: "r2", "r3", "r4", "r5", "r6", "r7",
+		"r8",  "r9", "r11", "r12", "r13", "r14", "r15",
+		"r16", "r17", "r18", "r19", "r20", "r21", "r22", "r23",
+		"r24", "r25", "r26", "r27", "r28", "r29", "memory");
+	return ret;
+}
+
+struct syscall_return_value __vdso_clock_gettime(clockid_t clock,
+						 struct timespec *ts)
+{
+	struct vdso_data *vdso = (struct vdso_data *)get_datapage();
+	struct syscall_return_value ret = { 0, 0 };
+
+	switch (clock) {
+	case CLOCK_REALTIME:
+		do_realtime(vdso, ts);
+		return ret;
+	case CLOCK_MONOTONIC:
+		do_monotonic(vdso, ts);
+		return ret;
+	case CLOCK_REALTIME_COARSE:
+		do_realtime_coarse(vdso, ts);
+		return ret;
+	case CLOCK_MONOTONIC_COARSE:
+		do_monotonic_coarse(vdso, ts);
+		return ret;
+	default:
+		return vdso_fallback_gettime(clock, ts);
+	}
+}
+
+int clock_gettime(clockid_t clock, struct timespec *ts)
+	__attribute__((weak, alias("__vdso_clock_gettime")));
diff --git a/kernel/arch/tile/kernel/vdso/vrt_sigreturn.S b/kernel/arch/tile/kernel/vdso/vrt_sigreturn.S
new file mode 100644
index 000000000..6326caf4a
--- /dev/null
+++ b/kernel/arch/tile/kernel/vdso/vrt_sigreturn.S
@@ -0,0 +1,30 @@
+/*
+ * Copyright 2012 Tilera Corporation. All Rights Reserved.
+ *
+ *   This program is free software; you can redistribute it and/or
+ *   modify it under the terms of the GNU General Public License
+ *   as published by the Free Software Foundation, version 2.
+ *
+ *   This program is distributed in the hope that it will be useful, but
+ *   WITHOUT ANY WARRANTY; without even the implied warranty of
+ *   MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
+ *   NON INFRINGEMENT.  See the GNU General Public License for
+ *   more details.
+ */
+
+#include <linux/linkage.h>
+#include <arch/abi.h>
+#include <asm/unistd.h>
+
+/*
+ * Note that libc has a copy of this function that it uses to compare
+ * against the PC when a stack backtrace ends, so if this code is
+ * changed, the libc implementation(s) should also be updated.
+ */
+ENTRY(__vdso_rt_sigreturn)
+	moveli TREG_SYSCALL_NR_NAME, __NR_rt_sigreturn
+	swint1
+	/* We don't use ENDPROC to avoid tagging this symbol as FUNC,
+	 * which confuses the perf tool.
+	 */
+	END(__vdso_rt_sigreturn)
diff --git a/kernel/arch/tile/kernel/vmlinux.lds.S b/kernel/arch/tile/kernel/vmlinux.lds.S
new file mode 100644
index 000000000..0e059a010
--- /dev/null
+++ b/kernel/arch/tile/kernel/vmlinux.lds.S
@@ -0,0 +1,90 @@
+#include <asm-generic/vmlinux.lds.h>
+#include <asm/page.h>
+#include <asm/cache.h>
+#include <asm/thread_info.h>
+#include <hv/hypervisor.h>
+
+/* Text loads starting from the supervisor interrupt vector address. */
+#define TEXT_OFFSET MEM_SV_START
+
+OUTPUT_ARCH(tile)
+ENTRY(_start)
+jiffies = jiffies_64;
+
+PHDRS
+{
+  intrpt PT_LOAD ;
+  text PT_LOAD ;
+  data PT_LOAD ;
+}
+SECTIONS
+{
+  /* Text is loaded with a different VA than data; start with text. */
+  #undef LOAD_OFFSET
+  #define LOAD_OFFSET TEXT_OFFSET
+
+  /* Interrupt vectors */
+  .intrpt (LOAD_OFFSET) : AT ( 0 )   /* put at the start of physical memory */
+  {
+    _text = .;
+    *(.intrpt)
+  } :intrpt =0
+
+  /* Hypervisor call vectors */
+  . = ALIGN(0x10000);
+  .hvglue : AT (ADDR(.hvglue) - LOAD_OFFSET) {
+    *(.hvglue)
+  } :NONE
+
+  /* Now the real code */
+  . = ALIGN(0x20000);
+  _stext = .;
+  .text : AT (ADDR(.text) - LOAD_OFFSET) {
+    HEAD_TEXT
+    SCHED_TEXT
+    LOCK_TEXT
+    KPROBES_TEXT
+    IRQENTRY_TEXT
+    __fix_text_end = .;   /* tile-cpack won't rearrange before this */
+    ALIGN_FUNCTION();
+    *(.hottext*)
+    TEXT_TEXT
+    *(.text.*)
+    *(.coldtext*)
+    *(.fixup)
+    *(.gnu.warning)
+  } :text =0
+  _etext = .;
+
+  /* "Init" is divided into two areas with very different virtual addresses. */
+  INIT_TEXT_SECTION(PAGE_SIZE)
+
+  /* Now we skip back to PAGE_OFFSET for the data. */
+  . = (. - TEXT_OFFSET + PAGE_OFFSET);
+  #undef LOAD_OFFSET
+  #define LOAD_OFFSET PAGE_OFFSET
+
+  . = ALIGN(PAGE_SIZE);
+  __init_begin = .;
+  INIT_DATA_SECTION(16) :data =0
+  PERCPU_SECTION(L2_CACHE_BYTES)
+  . = ALIGN(PAGE_SIZE);
+  __init_end = .;
+
+  _sdata = .;                   /* Start of data section */
+  RO_DATA_SECTION(PAGE_SIZE)
+  RW_DATA_SECTION(L2_CACHE_BYTES, PAGE_SIZE, THREAD_SIZE)
+  _edata = .;
+
+  EXCEPTION_TABLE(L2_CACHE_BYTES)
+  NOTES
+
+
+  BSS_SECTION(8, PAGE_SIZE, 1)
+  _end = . ;
+
+  STABS_DEBUG
+  DWARF_DEBUG
+
+  DISCARDS
+}
diff --git a/kernel/arch/tile/kvm/Kconfig b/kernel/arch/tile/kvm/Kconfig
new file mode 100644
index 000000000..1e968f755
--- /dev/null
+++ b/kernel/arch/tile/kvm/Kconfig
@@ -0,0 +1,38 @@
+#
+# KVM configuration
+#
+
+source "virt/kvm/Kconfig"
+
+menuconfig VIRTUALIZATION
+	bool "Virtualization"
+	---help---
+	  Say Y here to get to see options for using your Linux host to run
+	  other operating systems inside virtual machines (guests).
+	  This option alone does not add any kernel code.
+
+	  If you say N, all options in this submenu will be skipped and
+	  disabled.
+
+if VIRTUALIZATION
+
+config KVM
+	tristate "Kernel-based Virtual Machine (KVM) support"
+	depends on HAVE_KVM && MODULES
+	select PREEMPT_NOTIFIERS
+	select ANON_INODES
+	select SRCU
+	---help---
+	  Support hosting paravirtualized guest machines.
+
+	  This module provides access to the hardware capabilities through
+	  a character device node named /dev/kvm.
+
+	  To compile this as a module, choose M here: the module
+	  will be called kvm.
+
+	  If unsure, say N.
+
+source drivers/vhost/Kconfig
+
+endif # VIRTUALIZATION
diff --git a/kernel/arch/tile/lib/Makefile b/kernel/arch/tile/lib/Makefile
new file mode 100644
index 000000000..c4211cbb2
--- /dev/null
+++ b/kernel/arch/tile/lib/Makefile
@@ -0,0 +1,18 @@
+#
+# Makefile for TILE-specific library files..
+#
+
+lib-y = cacheflush.o checksum.o cpumask.o delay.o uaccess.o \
+	memmove.o memcpy_$(BITS).o memchr_$(BITS).o memset_$(BITS).o \
+	strchr_$(BITS).o strlen_$(BITS).o strnlen_$(BITS).o
+
+lib-$(CONFIG_TILEGX) += memcpy_user_64.o
+lib-$(CONFIG_TILEPRO) += atomic_32.o atomic_asm_32.o
+lib-$(CONFIG_SMP) += spinlock_$(BITS).o usercopy_$(BITS).o
+
+obj-$(CONFIG_MODULES) += exports.o
+
+# The finv_buffer_remote() and copy_{to,from}_user() routines can't
+# have -pg added, since they both rely on being leaf functions.
+CFLAGS_REMOVE_cacheflush.o = -pg
+CFLAGS_REMOVE_memcpy_user_64.o = -pg
diff --git a/kernel/arch/tile/lib/atomic_32.c b/kernel/arch/tile/lib/atomic_32.c
new file mode 100644
index 000000000..c89b211fd
--- /dev/null
+++ b/kernel/arch/tile/lib/atomic_32.c
@@ -0,0 +1,183 @@
+/*
+ * Copyright 2010 Tilera Corporation. All Rights Reserved.
+ *
+ *   This program is free software; you can redistribute it and/or
+ *   modify it under the terms of the GNU General Public License
+ *   as published by the Free Software Foundation, version 2.
+ *
+ *   This program is distributed in the hope that it will be useful, but
+ *   WITHOUT ANY WARRANTY; without even the implied warranty of
+ *   MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
+ *   NON INFRINGEMENT.  See the GNU General Public License for
+ *   more details.
+ */
+
+#include <linux/cache.h>
+#include <linux/delay.h>
+#include <linux/uaccess.h>
+#include <linux/module.h>
+#include <linux/mm.h>
+#include <linux/atomic.h>
+#include <arch/chip.h>
+
+/* This page is remapped on startup to be hash-for-home. */
+int atomic_locks[PAGE_SIZE / sizeof(int)] __page_aligned_bss;
+
+int *__atomic_hashed_lock(volatile void *v)
+{
+	/* NOTE: this code must match "sys_cmpxchg" in kernel/intvec_32.S */
+	/*
+	 * Use bits [3, 3 + ATOMIC_HASH_SHIFT) as the lock index.
+	 * Using mm works here because atomic_locks is page aligned.
+	 */
+	unsigned long ptr = __insn_mm((unsigned long)v >> 1,
+				      (unsigned long)atomic_locks,
+				      2, (ATOMIC_HASH_SHIFT + 2) - 1);
+	return (int *)ptr;
+}
+
+#ifdef CONFIG_SMP
+/* Return whether the passed pointer is a valid atomic lock pointer. */
+static int is_atomic_lock(int *p)
+{
+	return p >= &atomic_locks[0] && p < &atomic_locks[ATOMIC_HASH_SIZE];
+}
+
+void __atomic_fault_unlock(int *irqlock_word)
+{
+	BUG_ON(!is_atomic_lock(irqlock_word));
+	BUG_ON(*irqlock_word != 1);
+	*irqlock_word = 0;
+}
+
+#endif /* CONFIG_SMP */
+
+static inline int *__atomic_setup(volatile void *v)
+{
+	/* Issue a load to the target to bring it into cache. */
+	*(volatile int *)v;
+	return __atomic_hashed_lock(v);
+}
+
+int _atomic_xchg(int *v, int n)
+{
+	return __atomic_xchg(v, __atomic_setup(v), n).val;
+}
+EXPORT_SYMBOL(_atomic_xchg);
+
+int _atomic_xchg_add(int *v, int i)
+{
+	return __atomic_xchg_add(v, __atomic_setup(v), i).val;
+}
+EXPORT_SYMBOL(_atomic_xchg_add);
+
+int _atomic_xchg_add_unless(int *v, int a, int u)
+{
+	/*
+	 * Note: argument order is switched here since it is easier
+	 * to use the first argument consistently as the "old value"
+	 * in the assembly, as is done for _atomic_cmpxchg().
+	 */
+	return __atomic_xchg_add_unless(v, __atomic_setup(v), u, a).val;
+}
+EXPORT_SYMBOL(_atomic_xchg_add_unless);
+
+int _atomic_cmpxchg(int *v, int o, int n)
+{
+	return __atomic_cmpxchg(v, __atomic_setup(v), o, n).val;
+}
+EXPORT_SYMBOL(_atomic_cmpxchg);
+
+unsigned long _atomic_or(volatile unsigned long *p, unsigned long mask)
+{
+	return __atomic_or((int *)p, __atomic_setup(p), mask).val;
+}
+EXPORT_SYMBOL(_atomic_or);
+
+unsigned long _atomic_andn(volatile unsigned long *p, unsigned long mask)
+{
+	return __atomic_andn((int *)p, __atomic_setup(p), mask).val;
+}
+EXPORT_SYMBOL(_atomic_andn);
+
+unsigned long _atomic_xor(volatile unsigned long *p, unsigned long mask)
+{
+	return __atomic_xor((int *)p, __atomic_setup(p), mask).val;
+}
+EXPORT_SYMBOL(_atomic_xor);
+
+
+long long _atomic64_xchg(long long *v, long long n)
+{
+	return __atomic64_xchg(v, __atomic_setup(v), n);
+}
+EXPORT_SYMBOL(_atomic64_xchg);
+
+long long _atomic64_xchg_add(long long *v, long long i)
+{
+	return __atomic64_xchg_add(v, __atomic_setup(v), i);
+}
+EXPORT_SYMBOL(_atomic64_xchg_add);
+
+long long _atomic64_xchg_add_unless(long long *v, long long a, long long u)
+{
+	/*
+	 * Note: argument order is switched here since it is easier
+	 * to use the first argument consistently as the "old value"
+	 * in the assembly, as is done for _atomic_cmpxchg().
+	 */
+	return __atomic64_xchg_add_unless(v, __atomic_setup(v), u, a);
+}
+EXPORT_SYMBOL(_atomic64_xchg_add_unless);
+
+long long _atomic64_cmpxchg(long long *v, long long o, long long n)
+{
+	return __atomic64_cmpxchg(v, __atomic_setup(v), o, n);
+}
+EXPORT_SYMBOL(_atomic64_cmpxchg);
+
+
+/*
+ * If any of the atomic or futex routines hit a bad address (not in
+ * the page tables at kernel PL) this routine is called.  The futex
+ * routines are never used on kernel space, and the normal atomics and
+ * bitops are never used on user space.  So a fault on kernel space
+ * must be fatal, but a fault on userspace is a futex fault and we
+ * need to return -EFAULT.  Note that the context this routine is
+ * invoked in is the context of the "_atomic_xxx()" routines called
+ * by the functions in this file.
+ */
+struct __get_user __atomic_bad_address(int __user *addr)
+{
+	if (unlikely(!access_ok(VERIFY_WRITE, addr, sizeof(int))))
+		panic("Bad address used for kernel atomic op: %p\n", addr);
+	return (struct __get_user) { .err = -EFAULT };
+}
+
+
+void __init __init_atomic_per_cpu(void)
+{
+	/* Validate power-of-two and "bigger than cpus" assumption */
+	BUILD_BUG_ON(ATOMIC_HASH_SIZE & (ATOMIC_HASH_SIZE-1));
+	BUG_ON(ATOMIC_HASH_SIZE < nr_cpu_ids);
+
+	/*
+	 * On TILEPro we prefer to use a single hash-for-home
+	 * page, since this means atomic operations are less
+	 * likely to encounter a TLB fault and thus should
+	 * in general perform faster.  You may wish to disable
+	 * this in situations where few hash-for-home tiles
+	 * are configured.
+	 */
+	BUG_ON((unsigned long)atomic_locks % PAGE_SIZE != 0);
+
+	/* The locks must all fit on one page. */
+	BUILD_BUG_ON(ATOMIC_HASH_SIZE * sizeof(int) > PAGE_SIZE);
+
+	/*
+	 * We use the page offset of the atomic value's address as
+	 * an index into atomic_locks, excluding the low 3 bits.
+	 * That should not produce more indices than ATOMIC_HASH_SIZE.
+	 */
+	BUILD_BUG_ON((PAGE_SIZE >> 3) > ATOMIC_HASH_SIZE);
+}
diff --git a/kernel/arch/tile/lib/atomic_asm_32.S b/kernel/arch/tile/lib/atomic_asm_32.S
new file mode 100644
index 000000000..6bda3132c
--- /dev/null
+++ b/kernel/arch/tile/lib/atomic_asm_32.S
@@ -0,0 +1,197 @@
+/*
+ * Copyright 2010 Tilera Corporation. All Rights Reserved.
+ *
+ *   This program is free software; you can redistribute it and/or
+ *   modify it under the terms of the GNU General Public License
+ *   as published by the Free Software Foundation, version 2.
+ *
+ *   This program is distributed in the hope that it will be useful, but
+ *   WITHOUT ANY WARRANTY; without even the implied warranty of
+ *   MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
+ *   NON INFRINGEMENT.  See the GNU General Public License for
+ *   more details.
+ *
+ * Support routines for atomic operations.  Each function takes:
+ *
+ * r0: address to manipulate
+ * r1: pointer to atomic lock guarding this operation (for ATOMIC_LOCK_REG)
+ * r2: new value to write, or for cmpxchg/add_unless, value to compare against
+ * r3: (cmpxchg/xchg_add_unless) new value to write or add;
+ *     (atomic64 ops) high word of value to write
+ * r4/r5: (cmpxchg64/add_unless64) new value to write or add
+ *
+ * The 32-bit routines return a "struct __get_user" so that the futex code
+ * has an opportunity to return -EFAULT to the user if needed.
+ * The 64-bit routines just return a "long long" with the value,
+ * since they are only used from kernel space and don't expect to fault.
+ * Support for 16-bit ops is included in the framework but we don't provide
+ * any (x86_64 has an atomic_inc_short(), so we might want to some day).
+ *
+ * Note that the caller is advised to issue a suitable L1 or L2
+ * prefetch on the address being manipulated to avoid extra stalls.
+ * In addition, the hot path is on two icache lines, and we start with
+ * a jump to the second line to make sure they are both in cache so
+ * that we never stall waiting on icache fill while holding the lock.
+ * (This doesn't work out with most 64-bit ops, since they consume
+ * too many bundles, so may take an extra i-cache stall.)
+ *
+ * These routines set the INTERRUPT_CRITICAL_SECTION bit, just
+ * like sys_cmpxchg(), so that NMIs like PERF_COUNT will not interrupt
+ * the code, just page faults.
+ *
+ * If the load or store faults in a way that can be directly fixed in
+ * the do_page_fault_ics() handler (e.g. a vmalloc reference) we fix it
+ * directly, return to the instruction that faulted, and retry it.
+ *
+ * If the load or store faults in a way that potentially requires us
+ * to release the atomic lock, then retry (e.g. a migrating PTE), we
+ * reset the PC in do_page_fault_ics() to the "tns" instruction so
+ * that on return we will reacquire the lock and restart the op.  We
+ * are somewhat overloading the exception_table_entry notion by doing
+ * this, since those entries are not normally used for migrating PTEs.
+ *
+ * If the main page fault handler discovers a bad address, it will see
+ * the PC pointing to the "tns" instruction (due to the earlier
+ * exception_table_entry processing in do_page_fault_ics), and
+ * re-reset the PC to the fault handler, atomic_bad_address(), which
+ * effectively takes over from the atomic op and can either return a
+ * bad "struct __get_user" (for user addresses) or can just panic (for
+ * bad kernel addresses).
+ *
+ * Note that if the value we would store is the same as what we
+ * loaded, we bypass the store.  Other platforms with true atomics can
+ * make the guarantee that a non-atomic __clear_bit(), for example,
+ * can safely race with an atomic test_and_set_bit(); this example is
+ * from bit_spinlock.h in slub_lock() / slub_unlock().  We can't do
+ * that on Tile since the "atomic" op is really just a
+ * read/modify/write, and can race with the non-atomic
+ * read/modify/write.  However, if we can short-circuit the write when
+ * it is not needed, in the atomic case, we avoid the race.
+ */
+
+#include <linux/linkage.h>
+#include <asm/atomic_32.h>
+#include <asm/page.h>
+#include <asm/processor.h>
+
+	.section .text.atomic,"ax"
+ENTRY(__start_atomic_asm_code)
+
+	.macro  atomic_op, name, bitwidth, body
+	.align  64
+STD_ENTRY_SECTION(__atomic\name, .text.atomic)
+	{
+	 movei  r24, 1
+	 j      4f		/* branch to second cache line */
+	}
+1:	{
+	 .ifc \bitwidth,16
+	 lh     r22, r0
+	 .else
+	 lw     r22, r0
+	 addi   r28, r0, 4
+	 .endif
+	}
+	.ifc \bitwidth,64
+	lw      r23, r28
+	.endif
+	\body /* set r24, and r25 if 64-bit */
+	{
+	 seq    r26, r22, r24
+	 seq    r27, r23, r25
+	}
+	.ifc \bitwidth,64
+	bbnst   r27, 2f
+	.endif
+	bbs     r26, 3f		/* skip write-back if it's the same value */
+2:	{
+	 .ifc \bitwidth,16
+	 sh     r0, r24
+	 .else
+	 sw     r0, r24
+	 .endif
+	}
+	.ifc \bitwidth,64
+	sw      r28, r25
+	.endif
+	mf
+3:	{
+	 move   r0, r22
+	 .ifc \bitwidth,64
+	 move   r1, r23
+	 .else
+	 move   r1, zero
+	 .endif
+	 sw     ATOMIC_LOCK_REG_NAME, zero
+	}
+	mtspr   INTERRUPT_CRITICAL_SECTION, zero
+	jrp     lr
+4:	{
+	 move   ATOMIC_LOCK_REG_NAME, r1
+	 mtspr  INTERRUPT_CRITICAL_SECTION, r24
+	}
+#ifndef CONFIG_SMP
+	j       1b		/* no atomic locks */
+#else
+	{
+	 tns    r21, ATOMIC_LOCK_REG_NAME
+	 moveli r23, 2048       /* maximum backoff time in cycles */
+	}
+	{
+	 bzt    r21, 1b		/* branch if lock acquired */
+	 moveli r25, 32         /* starting backoff time in cycles */
+	}
+5:	mtspr   INTERRUPT_CRITICAL_SECTION, zero
+	mfspr   r26, CYCLE_LOW  /* get start point for this backoff */
+6:	mfspr   r22, CYCLE_LOW  /* test to see if we've backed off enough */
+	sub     r22, r22, r26
+	slt     r22, r22, r25
+	bbst    r22, 6b
+	{
+	 mtspr  INTERRUPT_CRITICAL_SECTION, r24
+	 shli   r25, r25, 1     /* double the backoff; retry the tns */
+	}
+	{
+	 tns    r21, ATOMIC_LOCK_REG_NAME
+	 slt    r26, r23, r25   /* is the proposed backoff too big? */
+	}
+	{
+	 bzt    r21, 1b		/* branch if lock acquired */
+	 mvnz   r25, r26, r23
+	}
+	j       5b
+#endif
+	STD_ENDPROC(__atomic\name)
+	.ifc \bitwidth,32
+	.pushsection __ex_table,"a"
+	.align  4
+	.word   1b, __atomic\name
+	.word   2b, __atomic\name
+	.word   __atomic\name, __atomic_bad_address
+	.popsection
+	.endif
+	.endm
+
+atomic_op _cmpxchg, 32, "seq r26, r22, r2; { bbns r26, 3f; move r24, r3 }"
+atomic_op _xchg, 32, "move r24, r2"
+atomic_op _xchg_add, 32, "add r24, r22, r2"
+atomic_op _xchg_add_unless, 32, \
+	"sne r26, r22, r2; { bbns r26, 3f; add r24, r22, r3 }"
+atomic_op _or, 32, "or r24, r22, r2"
+atomic_op _andn, 32, "nor r2, r2, zero; and r24, r22, r2"
+atomic_op _xor, 32, "xor r24, r22, r2"
+
+atomic_op 64_cmpxchg, 64, "{ seq r26, r22, r2; seq r27, r23, r3 }; \
+	{ bbns r26, 3f; move r24, r4 }; { bbns r27, 3f; move r25, r5 }"
+atomic_op 64_xchg, 64, "{ move r24, r2; move r25, r3 }"
+atomic_op 64_xchg_add, 64, "{ add r24, r22, r2; add r25, r23, r3 }; \
+	slt_u r26, r24, r22; add r25, r25, r26"
+atomic_op 64_xchg_add_unless, 64, \
+	"{ sne r26, r22, r2; sne r27, r23, r3 }; \
+	{ bbns r26, 3f; add r24, r22, r4 }; \
+	{ bbns r27, 3f; add r25, r23, r5 }; \
+	slt_u r26, r24, r22; add r25, r25, r26"
+
+	jrp     lr              /* happy backtracer */
+
+ENTRY(__end_atomic_asm_code)
diff --git a/kernel/arch/tile/lib/cacheflush.c b/kernel/arch/tile/lib/cacheflush.c
new file mode 100644
index 000000000..9c0ec2200
--- /dev/null
+++ b/kernel/arch/tile/lib/cacheflush.c
@@ -0,0 +1,173 @@
+/*
+ * Copyright 2010 Tilera Corporation. All Rights Reserved.
+ *
+ *   This program is free software; you can redistribute it and/or
+ *   modify it under the terms of the GNU General Public License
+ *   as published by the Free Software Foundation, version 2.
+ *
+ *   This program is distributed in the hope that it will be useful, but
+ *   WITHOUT ANY WARRANTY; without even the implied warranty of
+ *   MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
+ *   NON INFRINGEMENT.  See the GNU General Public License for
+ *   more details.
+ */
+
+#include <linux/export.h>
+#include <asm/page.h>
+#include <asm/cacheflush.h>
+#include <arch/icache.h>
+#include <arch/spr_def.h>
+
+
+void __flush_icache_range(unsigned long start, unsigned long end)
+{
+	invalidate_icache((const void *)start, end - start, PAGE_SIZE);
+}
+
+
+/* Force a load instruction to issue. */
+static inline void force_load(char *p)
+{
+	*(volatile char *)p;
+}
+
+/*
+ * Flush and invalidate a VA range that is homed remotely on a single
+ * core (if "!hfh") or homed via hash-for-home (if "hfh"), waiting
+ * until the memory controller holds the flushed values.
+ */
+void __attribute__((optimize("omit-frame-pointer")))
+finv_buffer_remote(void *buffer, size_t size, int hfh)
+{
+	char *p, *base;
+	size_t step_size, load_count;
+
+	/*
+	 * On TILEPro the striping granularity is a fixed 8KB; on
+	 * TILE-Gx it is configurable, and we rely on the fact that
+	 * the hypervisor always configures maximum striping, so that
+	 * bits 9 and 10 of the PA are part of the stripe function, so
+	 * every 512 bytes we hit a striping boundary.
+	 *
+	 */
+#ifdef __tilegx__
+	const unsigned long STRIPE_WIDTH = 512;
+#else
+	const unsigned long STRIPE_WIDTH = 8192;
+#endif
+
+#ifdef __tilegx__
+	/*
+	 * On TILE-Gx, we must disable the dstream prefetcher before doing
+	 * a cache flush; otherwise, we could end up with data in the cache
+	 * that we don't want there.  Note that normally we'd do an mf
+	 * after the SPR write to disabling the prefetcher, but we do one
+	 * below, before any further loads, so there's no need to do it
+	 * here.
+	 */
+	uint_reg_t old_dstream_pf = __insn_mfspr(SPR_DSTREAM_PF);
+	__insn_mtspr(SPR_DSTREAM_PF, 0);
+#endif
+
+	/*
+	 * Flush and invalidate the buffer out of the local L1/L2
+	 * and request the home cache to flush and invalidate as well.
+	 */
+	__finv_buffer(buffer, size);
+
+	/*
+	 * Wait for the home cache to acknowledge that it has processed
+	 * all the flush-and-invalidate requests.  This does not mean
+	 * that the flushed data has reached the memory controller yet,
+	 * but it does mean the home cache is processing the flushes.
+	 */
+	__insn_mf();
+
+	/*
+	 * Issue a load to the last cache line, which can't complete
+	 * until all the previously-issued flushes to the same memory
+	 * controller have also completed.  If we weren't striping
+	 * memory, that one load would be sufficient, but since we may
+	 * be, we also need to back up to the last load issued to
+	 * another memory controller, which would be the point where
+	 * we crossed a "striping" boundary (the granularity of striping
+	 * across memory controllers).  Keep backing up and doing this
+	 * until we are before the beginning of the buffer, or have
+	 * hit all the controllers.
+	 *
+	 * If we are flushing a hash-for-home buffer, it's even worse.
+	 * Each line may be homed on a different tile, and each tile
+	 * may have up to four lines that are on different
+	 * controllers.  So as we walk backwards, we have to touch
+	 * enough cache lines to satisfy these constraints.  In
+	 * practice this ends up being close enough to "load from
+	 * every cache line on a full memory stripe on each
+	 * controller" that we simply do that, to simplify the logic.
+	 *
+	 * On TILE-Gx the hash-for-home function is much more complex,
+	 * with the upshot being we can't readily guarantee we have
+	 * hit both entries in the 128-entry AMT that were hit by any
+	 * load in the entire range, so we just re-load them all.
+	 * With larger buffers, we may want to consider using a hypervisor
+	 * trap to issue loads directly to each hash-for-home tile for
+	 * each controller (doing it from Linux would trash the TLB).
+	 */
+	if (hfh) {
+		step_size = L2_CACHE_BYTES;
+#ifdef __tilegx__
+		load_count = (size + L2_CACHE_BYTES - 1) / L2_CACHE_BYTES;
+#else
+		load_count = (STRIPE_WIDTH / L2_CACHE_BYTES) *
+			      (1 << CHIP_LOG_NUM_MSHIMS());
+#endif
+	} else {
+		step_size = STRIPE_WIDTH;
+		load_count = (1 << CHIP_LOG_NUM_MSHIMS());
+	}
+
+	/* Load the last byte of the buffer. */
+	p = (char *)buffer + size - 1;
+	force_load(p);
+
+	/* Bump down to the end of the previous stripe or cache line. */
+	p -= step_size;
+	p = (char *)((unsigned long)p | (step_size - 1));
+
+	/* Figure out how far back we need to go. */
+	base = p - (step_size * (load_count - 2));
+	if ((unsigned long)base < (unsigned long)buffer)
+		base = buffer;
+
+	/*
+	 * Fire all the loads we need.  The MAF only has eight entries
+	 * so we can have at most eight outstanding loads, so we
+	 * unroll by that amount.
+	 */
+#pragma unroll 8
+	for (; p >= base; p -= step_size)
+		force_load(p);
+
+	/*
+	 * Repeat, but with finv's instead of loads, to get rid of the
+	 * data we just loaded into our own cache and the old home L3.
+	 * No need to unroll since finv's don't target a register.
+	 * The finv's are guaranteed not to actually flush the data in
+	 * the buffer back to their home, since we just read it, so the
+	 * lines are clean in cache; we will only invalidate those lines.
+	 */
+	p = (char *)buffer + size - 1;
+	__insn_finv(p);
+	p -= step_size;
+	p = (char *)((unsigned long)p | (step_size - 1));
+	for (; p >= base; p -= step_size)
+		__insn_finv(p);
+
+	/* Wait for these finv's (and thus the first finvs) to be done. */
+	__insn_mf();
+
+#ifdef __tilegx__
+	/* Reenable the prefetcher. */
+	__insn_mtspr(SPR_DSTREAM_PF, old_dstream_pf);
+#endif
+}
+EXPORT_SYMBOL_GPL(finv_buffer_remote);
diff --git a/kernel/arch/tile/lib/checksum.c b/kernel/arch/tile/lib/checksum.c
new file mode 100644
index 000000000..c3ca3e64d
--- /dev/null
+++ b/kernel/arch/tile/lib/checksum.c
@@ -0,0 +1,89 @@
+/*
+ * Copyright 2010 Tilera Corporation. All Rights Reserved.
+ *
+ *   This program is free software; you can redistribute it and/or
+ *   modify it under the terms of the GNU General Public License
+ *   as published by the Free Software Foundation, version 2.
+ *
+ *   This program is distributed in the hope that it will be useful, but
+ *   WITHOUT ANY WARRANTY; without even the implied warranty of
+ *   MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
+ *   NON INFRINGEMENT.  See the GNU General Public License for
+ *   more details.
+ * Support code for the main lib/checksum.c.
+ */
+
+#include <net/checksum.h>
+#include <linux/module.h>
+
+__wsum do_csum(const unsigned char *buff, int len)
+{
+	int odd, count;
+	unsigned long result = 0;
+
+	if (len <= 0)
+		goto out;
+	odd = 1 & (unsigned long) buff;
+	if (odd) {
+		result = (*buff << 8);
+		len--;
+		buff++;
+	}
+	count = len >> 1;		/* nr of 16-bit words.. */
+	if (count) {
+		if (2 & (unsigned long) buff) {
+			result += *(const unsigned short *)buff;
+			count--;
+			len -= 2;
+			buff += 2;
+		}
+		count >>= 1;		/* nr of 32-bit words.. */
+		if (count) {
+#ifdef __tilegx__
+			if (4 & (unsigned long) buff) {
+				unsigned int w = *(const unsigned int *)buff;
+				result = __insn_v2sadau(result, w, 0);
+				count--;
+				len -= 4;
+				buff += 4;
+			}
+			count >>= 1;		/* nr of 64-bit words.. */
+#endif
+
+			/*
+			 * This algorithm could wrap around for very
+			 * large buffers, but those should be impossible.
+			 */
+			BUG_ON(count >= 65530);
+
+			while (count) {
+				unsigned long w = *(const unsigned long *)buff;
+				count--;
+				buff += sizeof(w);
+#ifdef __tilegx__
+				result = __insn_v2sadau(result, w, 0);
+#else
+				result = __insn_sadah_u(result, w, 0);
+#endif
+			}
+#ifdef __tilegx__
+			if (len & 4) {
+				unsigned int w = *(const unsigned int *)buff;
+				result = __insn_v2sadau(result, w, 0);
+				buff += 4;
+			}
+#endif
+		}
+		if (len & 2) {
+			result += *(const unsigned short *) buff;
+			buff += 2;
+		}
+	}
+	if (len & 1)
+		result += *buff;
+	result = csum_long(result);
+	if (odd)
+		result = swab16(result);
+out:
+	return result;
+}
diff --git a/kernel/arch/tile/lib/cpumask.c b/kernel/arch/tile/lib/cpumask.c
new file mode 100644
index 000000000..75947edcc
--- /dev/null
+++ b/kernel/arch/tile/lib/cpumask.c
@@ -0,0 +1,54 @@
+/*
+ * Copyright 2010 Tilera Corporation. All Rights Reserved.
+ *
+ *   This program is free software; you can redistribute it and/or
+ *   modify it under the terms of the GNU General Public License
+ *   as published by the Free Software Foundation, version 2.
+ *
+ *   This program is distributed in the hope that it will be useful, but
+ *   WITHOUT ANY WARRANTY; without even the implied warranty of
+ *   MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
+ *   NON INFRINGEMENT.  See the GNU General Public License for
+ *   more details.
+ */
+
+#include <linux/cpumask.h>
+#include <linux/ctype.h>
+#include <linux/errno.h>
+#include <linux/smp.h>
+#include <linux/export.h>
+
+/*
+ * Allow cropping out bits beyond the end of the array.
+ * Move to "lib" directory if more clients want to use this routine.
+ */
+int bitmap_parselist_crop(const char *bp, unsigned long *maskp, int nmaskbits)
+{
+	unsigned a, b;
+
+	bitmap_zero(maskp, nmaskbits);
+	do {
+		if (!isdigit(*bp))
+			return -EINVAL;
+		a = simple_strtoul(bp, (char **)&bp, 10);
+		b = a;
+		if (*bp == '-') {
+			bp++;
+			if (!isdigit(*bp))
+				return -EINVAL;
+			b = simple_strtoul(bp, (char **)&bp, 10);
+		}
+		if (!(a <= b))
+			return -EINVAL;
+		if (b >= nmaskbits)
+			b = nmaskbits-1;
+		while (a <= b) {
+			set_bit(a, maskp);
+			a++;
+		}
+		if (*bp == ',')
+			bp++;
+	} while (*bp != '\0' && *bp != '\n');
+	return 0;
+}
+EXPORT_SYMBOL(bitmap_parselist_crop);
diff --git a/kernel/arch/tile/lib/delay.c b/kernel/arch/tile/lib/delay.c
new file mode 100644
index 000000000..cdacdd11d
--- /dev/null
+++ b/kernel/arch/tile/lib/delay.c
@@ -0,0 +1,45 @@
+/*
+ * Copyright 2010 Tilera Corporation. All Rights Reserved.
+ *
+ *   This program is free software; you can redistribute it and/or
+ *   modify it under the terms of the GNU General Public License
+ *   as published by the Free Software Foundation, version 2.
+ *
+ *   This program is distributed in the hope that it will be useful, but
+ *   WITHOUT ANY WARRANTY; without even the implied warranty of
+ *   MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
+ *   NON INFRINGEMENT.  See the GNU General Public License for
+ *   more details.
+ */
+
+#include <linux/module.h>
+#include <linux/delay.h>
+#include <linux/thread_info.h>
+#include <asm/timex.h>
+
+void __udelay(unsigned long usecs)
+{
+	if (usecs > ULONG_MAX / 1000) {
+		WARN_ON_ONCE(usecs > ULONG_MAX / 1000);
+		usecs = ULONG_MAX / 1000;
+	}
+	__ndelay(usecs * 1000);
+}
+EXPORT_SYMBOL(__udelay);
+
+void __ndelay(unsigned long nsecs)
+{
+	cycles_t target = get_cycles();
+	target += ns2cycles(nsecs);
+	while (get_cycles() < target)
+		cpu_relax();
+}
+EXPORT_SYMBOL(__ndelay);
+
+void __delay(unsigned long cycles)
+{
+	cycles_t target = get_cycles() + cycles;
+	while (get_cycles() < target)
+		cpu_relax();
+}
+EXPORT_SYMBOL(__delay);
diff --git a/kernel/arch/tile/lib/exports.c b/kernel/arch/tile/lib/exports.c
new file mode 100644
index 000000000..82733c87d
--- /dev/null
+++ b/kernel/arch/tile/lib/exports.c
@@ -0,0 +1,94 @@
+/*
+ * Copyright 2010 Tilera Corporation. All Rights Reserved.
+ *
+ *   This program is free software; you can redistribute it and/or
+ *   modify it under the terms of the GNU General Public License
+ *   as published by the Free Software Foundation, version 2.
+ *
+ *   This program is distributed in the hope that it will be useful, but
+ *   WITHOUT ANY WARRANTY; without even the implied warranty of
+ *   MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
+ *   NON INFRINGEMENT.  See the GNU General Public License for
+ *   more details.
+ *
+ * Exports from assembler code and from libtile-cc.
+ */
+
+#include <linux/module.h>
+
+/* arch/tile/lib/usercopy.S */
+#include <linux/uaccess.h>
+EXPORT_SYMBOL(strnlen_user_asm);
+EXPORT_SYMBOL(strncpy_from_user_asm);
+EXPORT_SYMBOL(clear_user_asm);
+EXPORT_SYMBOL(flush_user_asm);
+EXPORT_SYMBOL(finv_user_asm);
+
+/* arch/tile/kernel/entry.S */
+#include <linux/kernel.h>
+#include <asm/processor.h>
+EXPORT_SYMBOL(current_text_addr);
+EXPORT_SYMBOL(dump_stack);
+
+/* arch/tile/kernel/head.S */
+EXPORT_SYMBOL(empty_zero_page);
+
+#ifdef CONFIG_FUNCTION_TRACER
+/* arch/tile/kernel/mcount_64.S */
+#include <asm/ftrace.h>
+EXPORT_SYMBOL(__mcount);
+#endif /* CONFIG_FUNCTION_TRACER */
+
+/* arch/tile/lib/, various memcpy files */
+EXPORT_SYMBOL(memcpy);
+EXPORT_SYMBOL(__copy_to_user_inatomic);
+EXPORT_SYMBOL(__copy_from_user_inatomic);
+EXPORT_SYMBOL(__copy_from_user_zeroing);
+#ifdef __tilegx__
+EXPORT_SYMBOL(__copy_in_user_inatomic);
+#endif
+
+/* hypervisor glue */
+#include <hv/hypervisor.h>
+EXPORT_SYMBOL(hv_dev_open);
+EXPORT_SYMBOL(hv_dev_pread);
+EXPORT_SYMBOL(hv_dev_pwrite);
+EXPORT_SYMBOL(hv_dev_preada);
+EXPORT_SYMBOL(hv_dev_pwritea);
+EXPORT_SYMBOL(hv_dev_poll);
+EXPORT_SYMBOL(hv_dev_poll_cancel);
+EXPORT_SYMBOL(hv_dev_close);
+EXPORT_SYMBOL(hv_sysconf);
+EXPORT_SYMBOL(hv_confstr);
+EXPORT_SYMBOL(hv_get_rtc);
+EXPORT_SYMBOL(hv_set_rtc);
+
+/* libgcc.a */
+uint32_t __udivsi3(uint32_t dividend, uint32_t divisor);
+EXPORT_SYMBOL(__udivsi3);
+int32_t __divsi3(int32_t dividend, int32_t divisor);
+EXPORT_SYMBOL(__divsi3);
+uint64_t __udivdi3(uint64_t dividend, uint64_t divisor);
+EXPORT_SYMBOL(__udivdi3);
+int64_t __divdi3(int64_t dividend, int64_t divisor);
+EXPORT_SYMBOL(__divdi3);
+uint32_t __umodsi3(uint32_t dividend, uint32_t divisor);
+EXPORT_SYMBOL(__umodsi3);
+int32_t __modsi3(int32_t dividend, int32_t divisor);
+EXPORT_SYMBOL(__modsi3);
+uint64_t __umoddi3(uint64_t dividend, uint64_t divisor);
+EXPORT_SYMBOL(__umoddi3);
+int64_t __moddi3(int64_t dividend, int64_t divisor);
+EXPORT_SYMBOL(__moddi3);
+#ifndef __tilegx__
+int64_t __muldi3(int64_t, int64_t);
+EXPORT_SYMBOL(__muldi3);
+uint64_t __lshrdi3(uint64_t, unsigned int);
+EXPORT_SYMBOL(__lshrdi3);
+uint64_t __ashrdi3(uint64_t, unsigned int);
+EXPORT_SYMBOL(__ashrdi3);
+uint64_t __ashldi3(uint64_t, unsigned int);
+EXPORT_SYMBOL(__ashldi3);
+int __ffsdi2(uint64_t);
+EXPORT_SYMBOL(__ffsdi2);
+#endif
diff --git a/kernel/arch/tile/lib/memchr_32.c b/kernel/arch/tile/lib/memchr_32.c
new file mode 100644
index 000000000..cc3d9badf
--- /dev/null
+++ b/kernel/arch/tile/lib/memchr_32.c
@@ -0,0 +1,71 @@
+/*
+ * Copyright 2010 Tilera Corporation. All Rights Reserved.
+ *
+ *   This program is free software; you can redistribute it and/or
+ *   modify it under the terms of the GNU General Public License
+ *   as published by the Free Software Foundation, version 2.
+ *
+ *   This program is distributed in the hope that it will be useful, but
+ *   WITHOUT ANY WARRANTY; without even the implied warranty of
+ *   MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
+ *   NON INFRINGEMENT.  See the GNU General Public License for
+ *   more details.
+ */
+
+#include <linux/types.h>
+#include <linux/string.h>
+#include <linux/module.h>
+
+void *memchr(const void *s, int c, size_t n)
+{
+	const uint32_t *last_word_ptr;
+	const uint32_t *p;
+	const char *last_byte_ptr;
+	uintptr_t s_int;
+	uint32_t goal, before_mask, v, bits;
+	char *ret;
+
+	if (__builtin_expect(n == 0, 0)) {
+		/* Don't dereference any memory if the array is empty. */
+		return NULL;
+	}
+
+	/* Get an aligned pointer. */
+	s_int = (uintptr_t) s;
+	p = (const uint32_t *)(s_int & -4);
+
+	/* Create four copies of the byte for which we are looking. */
+	goal = 0x01010101 * (uint8_t) c;
+
+	/* Read the first word, but munge it so that bytes before the array
+	 * will not match goal.
+	 *
+	 * Note that this shift count expression works because we know
+	 * shift counts are taken mod 32.
+	 */
+	before_mask = (1 << (s_int << 3)) - 1;
+	v = (*p | before_mask) ^ (goal & before_mask);
+
+	/* Compute the address of the last byte. */
+	last_byte_ptr = (const char *)s + n - 1;
+
+	/* Compute the address of the word containing the last byte. */
+	last_word_ptr = (const uint32_t *)((uintptr_t) last_byte_ptr & -4);
+
+	while ((bits = __insn_seqb(v, goal)) == 0) {
+		if (__builtin_expect(p == last_word_ptr, 0)) {
+			/* We already read the last word in the array,
+			 * so give up.
+			 */
+			return NULL;
+		}
+		v = *++p;
+	}
+
+	/* We found a match, but it might be in a byte past the end
+	 * of the array.
+	 */
+	ret = ((char *)p) + (__insn_ctz(bits) >> 3);
+	return (ret <= last_byte_ptr) ? ret : NULL;
+}
+EXPORT_SYMBOL(memchr);
diff --git a/kernel/arch/tile/lib/memchr_64.c b/kernel/arch/tile/lib/memchr_64.c
new file mode 100644
index 000000000..f8196b3a9
--- /dev/null
+++ b/kernel/arch/tile/lib/memchr_64.c
@@ -0,0 +1,69 @@
+/*
+ * Copyright 2011 Tilera Corporation. All Rights Reserved.
+ *
+ *   This program is free software; you can redistribute it and/or
+ *   modify it under the terms of the GNU General Public License
+ *   as published by the Free Software Foundation, version 2.
+ *
+ *   This program is distributed in the hope that it will be useful, but
+ *   WITHOUT ANY WARRANTY; without even the implied warranty of
+ *   MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
+ *   NON INFRINGEMENT.  See the GNU General Public License for
+ *   more details.
+ */
+
+#include <linux/types.h>
+#include <linux/string.h>
+#include <linux/module.h>
+#include "string-endian.h"
+
+void *memchr(const void *s, int c, size_t n)
+{
+	const uint64_t *last_word_ptr;
+	const uint64_t *p;
+	const char *last_byte_ptr;
+	uintptr_t s_int;
+	uint64_t goal, before_mask, v, bits;
+	char *ret;
+
+	if (__builtin_expect(n == 0, 0)) {
+		/* Don't dereference any memory if the array is empty. */
+		return NULL;
+	}
+
+	/* Get an aligned pointer. */
+	s_int = (uintptr_t) s;
+	p = (const uint64_t *)(s_int & -8);
+
+	/* Create eight copies of the byte for which we are looking. */
+	goal = copy_byte(c);
+
+	/* Read the first word, but munge it so that bytes before the array
+	 * will not match goal.
+	 */
+	before_mask = MASK(s_int);
+	v = (*p | before_mask) ^ (goal & before_mask);
+
+	/* Compute the address of the last byte. */
+	last_byte_ptr = (const char *)s + n - 1;
+
+	/* Compute the address of the word containing the last byte. */
+	last_word_ptr = (const uint64_t *)((uintptr_t) last_byte_ptr & -8);
+
+	while ((bits = __insn_v1cmpeq(v, goal)) == 0) {
+		if (__builtin_expect(p == last_word_ptr, 0)) {
+			/* We already read the last word in the array,
+			 * so give up.
+			 */
+			return NULL;
+		}
+		v = *++p;
+	}
+
+	/* We found a match, but it might be in a byte past the end
+	 * of the array.
+	 */
+	ret = ((char *)p) + (CFZ(bits) >> 3);
+	return (ret <= last_byte_ptr) ? ret : NULL;
+}
+EXPORT_SYMBOL(memchr);
diff --git a/kernel/arch/tile/lib/memcpy_32.S b/kernel/arch/tile/lib/memcpy_32.S
new file mode 100644
index 000000000..a2771ae5d
--- /dev/null
+++ b/kernel/arch/tile/lib/memcpy_32.S
@@ -0,0 +1,559 @@
+/*
+ * Copyright 2010 Tilera Corporation. All Rights Reserved.
+ *
+ *   This program is free software; you can redistribute it and/or
+ *   modify it under the terms of the GNU General Public License
+ *   as published by the Free Software Foundation, version 2.
+ *
+ *   This program is distributed in the hope that it will be useful, but
+ *   WITHOUT ANY WARRANTY; without even the implied warranty of
+ *   MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
+ *   NON INFRINGEMENT.  See the GNU General Public License for
+ *   more details.
+ */
+
+#include <arch/chip.h>
+
+
+/*
+ * This file shares the implementation of the userspace memcpy and
+ * the kernel's memcpy, copy_to_user and copy_from_user.
+ */
+
+#include <linux/linkage.h>
+
+#define IS_MEMCPY	  0
+#define IS_COPY_FROM_USER  1
+#define IS_COPY_FROM_USER_ZEROING  2
+#define IS_COPY_TO_USER   -1
+
+	.section .text.memcpy_common, "ax"
+	.align 64
+
+/* Use this to preface each bundle that can cause an exception so
+ * the kernel can clean up properly. The special cleanup code should
+ * not use these, since it knows what it is doing.
+ */
+#define EX \
+	.pushsection __ex_table, "a"; \
+	.align 4; \
+	.word 9f, memcpy_common_fixup; \
+	.popsection; \
+	9
+
+
+/* __copy_from_user_inatomic takes the kernel target address in r0,
+ * the user source in r1, and the bytes to copy in r2.
+ * It returns the number of uncopiable bytes (hopefully zero) in r0.
+ */
+ENTRY(__copy_from_user_inatomic)
+.type __copy_from_user_inatomic, @function
+	FEEDBACK_ENTER_EXPLICIT(__copy_from_user_inatomic, \
+	  .text.memcpy_common, \
+	  .Lend_memcpy_common - __copy_from_user_inatomic)
+	{ movei r29, IS_COPY_FROM_USER; j memcpy_common }
+	.size __copy_from_user_inatomic, . - __copy_from_user_inatomic
+
+/* __copy_from_user_zeroing is like __copy_from_user_inatomic, but
+ * any uncopiable bytes are zeroed in the target.
+ */
+ENTRY(__copy_from_user_zeroing)
+.type __copy_from_user_zeroing, @function
+	FEEDBACK_REENTER(__copy_from_user_inatomic)
+	{ movei r29, IS_COPY_FROM_USER_ZEROING; j memcpy_common }
+	.size __copy_from_user_zeroing, . - __copy_from_user_zeroing
+
+/* __copy_to_user_inatomic takes the user target address in r0,
+ * the kernel source in r1, and the bytes to copy in r2.
+ * It returns the number of uncopiable bytes (hopefully zero) in r0.
+ */
+ENTRY(__copy_to_user_inatomic)
+.type __copy_to_user_inatomic, @function
+	FEEDBACK_REENTER(__copy_from_user_inatomic)
+	{ movei r29, IS_COPY_TO_USER; j memcpy_common }
+	.size __copy_to_user_inatomic, . - __copy_to_user_inatomic
+
+ENTRY(memcpy)
+.type memcpy, @function
+	FEEDBACK_REENTER(__copy_from_user_inatomic)
+	{ movei r29, IS_MEMCPY }
+	.size memcpy, . - memcpy
+	/* Fall through */
+
+	.type memcpy_common, @function
+memcpy_common:
+	/* On entry, r29 holds one of the IS_* macro values from above. */
+
+
+	/* r0 is the dest, r1 is the source, r2 is the size. */
+
+	/* Save aside original dest so we can return it at the end. */
+	{ sw sp, lr; move r23, r0; or r4, r0, r1 }
+
+	/* Check for an empty size. */
+	{ bz r2, .Ldone; andi r4, r4, 3 }
+
+	/* Save aside original values in case of a fault. */
+	{ move r24, r1; move r25, r2 }
+	move r27, lr
+
+	/* Check for an unaligned source or dest. */
+	{ bnz r4, .Lcopy_unaligned_maybe_many; addli r4, r2, -256 }
+
+.Lcheck_aligned_copy_size:
+	/* If we are copying < 256 bytes, branch to simple case. */
+	{ blzt r4, .Lcopy_8_check; slti_u r8, r2, 8 }
+
+	/* Copying >= 256 bytes, so jump to complex prefetching loop. */
+	{ andi r6, r1, 63; j .Lcopy_many }
+
+/*
+ *
+ * Aligned 4 byte at a time copy loop
+ *
+ */
+
+.Lcopy_8_loop:
+	/* Copy two words at a time to hide load latency. */
+EX:	{ lw r3, r1; addi r1, r1, 4; slti_u r8, r2, 16 }
+EX:	{ lw r4, r1; addi r1, r1, 4 }
+EX:	{ sw r0, r3; addi r0, r0, 4; addi r2, r2, -4 }
+EX:	{ sw r0, r4; addi r0, r0, 4; addi r2, r2, -4 }
+.Lcopy_8_check:
+	{ bzt r8, .Lcopy_8_loop; slti_u r4, r2, 4 }
+
+	/* Copy odd leftover word, if any. */
+	{ bnzt r4, .Lcheck_odd_stragglers }
+EX:	{ lw r3, r1; addi r1, r1, 4 }
+EX:	{ sw r0, r3; addi r0, r0, 4; addi r2, r2, -4 }
+
+.Lcheck_odd_stragglers:
+	{ bnz r2, .Lcopy_unaligned_few }
+
+.Ldone:
+	/* For memcpy return original dest address, else zero. */
+	{ mz r0, r29, r23; jrp lr }
+
+
+/*
+ *
+ * Prefetching multiple cache line copy handler (for large transfers).
+ *
+ */
+
+	/* Copy words until r1 is cache-line-aligned. */
+.Lalign_loop:
+EX:	{ lw r3, r1; addi r1, r1, 4 }
+	{ andi r6, r1, 63 }
+EX:	{ sw r0, r3; addi r0, r0, 4; addi r2, r2, -4 }
+.Lcopy_many:
+	{ bnzt r6, .Lalign_loop; addi r9, r0, 63 }
+
+	{ addi r3, r1, 60; andi r9, r9, -64 }
+
+	/* No need to prefetch dst, we'll just do the wh64
+	 * right before we copy a line.
+	 */
+EX:	{ lw r5, r3; addi r3, r3, 64; movei r4, 1 }
+	/* Intentionally stall for a few cycles to leave L2 cache alone. */
+	{ bnzt zero, .; move r27, lr }
+EX:	{ lw r6, r3; addi r3, r3, 64 }
+	/* Intentionally stall for a few cycles to leave L2 cache alone. */
+	{ bnzt zero, . }
+EX:	{ lw r7, r3; addi r3, r3, 64 }
+	/* Intentionally stall for a few cycles to leave L2 cache alone. */
+	{ bz zero, .Lbig_loop2 }
+
+	/* On entry to this loop:
+	 * - r0 points to the start of dst line 0
+	 * - r1 points to start of src line 0
+	 * - r2 >= (256 - 60), only the first time the loop trips.
+	 * - r3 contains r1 + 128 + 60    [pointer to end of source line 2]
+	 *   This is our prefetch address. When we get near the end
+	 *   rather than prefetching off the end this is changed to point
+	 *   to some "safe" recently loaded address.
+	 * - r5 contains *(r1 + 60)       [i.e. last word of source line 0]
+	 * - r6 contains *(r1 + 64 + 60)  [i.e. last word of source line 1]
+	 * - r9 contains ((r0 + 63) & -64)
+	 *     [start of next dst cache line.]
+	 */
+
+.Lbig_loop:
+	{ jal .Lcopy_line2; add r15, r1, r2 }
+
+.Lbig_loop2:
+	/* Copy line 0, first stalling until r5 is ready. */
+EX:	{ move r12, r5; lw r16, r1 }
+	{ bz r4, .Lcopy_8_check; slti_u r8, r2, 8 }
+	/* Prefetch several lines ahead. */
+EX:	{ lw r5, r3; addi r3, r3, 64 }
+	{ jal .Lcopy_line }
+
+	/* Copy line 1, first stalling until r6 is ready. */
+EX:	{ move r12, r6; lw r16, r1 }
+	{ bz r4, .Lcopy_8_check; slti_u r8, r2, 8 }
+	/* Prefetch several lines ahead. */
+EX:	{ lw r6, r3; addi r3, r3, 64 }
+	{ jal .Lcopy_line }
+
+	/* Copy line 2, first stalling until r7 is ready. */
+EX:	{ move r12, r7; lw r16, r1 }
+	{ bz r4, .Lcopy_8_check; slti_u r8, r2, 8 }
+	/* Prefetch several lines ahead. */
+EX:	{ lw r7, r3; addi r3, r3, 64 }
+	/* Use up a caches-busy cycle by jumping back to the top of the
+	 * loop. Might as well get it out of the way now.
+	 */
+	{ j .Lbig_loop }
+
+
+	/* On entry:
+	 * - r0 points to the destination line.
+	 * - r1 points to the source line.
+	 * - r3 is the next prefetch address.
+	 * - r9 holds the last address used for wh64.
+	 * - r12 = WORD_15
+	 * - r16 = WORD_0.
+	 * - r17 == r1 + 16.
+	 * - r27 holds saved lr to restore.
+	 *
+	 * On exit:
+	 * - r0 is incremented by 64.
+	 * - r1 is incremented by 64, unless that would point to a word
+	 *   beyond the end of the source array, in which case it is redirected
+	 *   to point to an arbitrary word already in the cache.
+	 * - r2 is decremented by 64.
+	 * - r3 is unchanged, unless it points to a word beyond the
+	 *   end of the source array, in which case it is redirected
+	 *   to point to an arbitrary word already in the cache.
+	 *   Redirecting is OK since if we are that close to the end
+	 *   of the array we will not come back to this subroutine
+	 *   and use the contents of the prefetched address.
+	 * - r4 is nonzero iff r2 >= 64.
+	 * - r9 is incremented by 64, unless it points beyond the
+	 *   end of the last full destination cache line, in which
+	 *   case it is redirected to a "safe address" that can be
+	 *   clobbered (sp - 64)
+	 * - lr contains the value in r27.
+	 */
+
+/* r26 unused */
+
+.Lcopy_line:
+	/* TODO: when r3 goes past the end, we would like to redirect it
+	 * to prefetch the last partial cache line (if any) just once, for the
+	 * benefit of the final cleanup loop. But we don't want to
+	 * prefetch that line more than once, or subsequent prefetches
+	 * will go into the RTF. But then .Lbig_loop should unconditionally
+	 * branch to top of loop to execute final prefetch, and its
+	 * nop should become a conditional branch.
+	 */
+
+	/* We need two non-memory cycles here to cover the resources
+	 * used by the loads initiated by the caller.
+	 */
+	{ add r15, r1, r2 }
+.Lcopy_line2:
+	{ slt_u r13, r3, r15; addi r17, r1, 16 }
+
+	/* NOTE: this will stall for one cycle as L1 is busy. */
+
+	/* Fill second L1D line. */
+EX:	{ lw r17, r17; addi r1, r1, 48; mvz r3, r13, r1 } /* r17 = WORD_4 */
+
+	/* Prepare destination line for writing. */
+EX:	{ wh64 r9; addi r9, r9, 64 }
+	/* Load seven words that are L1D hits to cover wh64 L2 usage. */
+
+	/* Load the three remaining words from the last L1D line, which
+	 * we know has already filled the L1D.
+	 */
+EX:	{ lw r4, r1;  addi r1, r1, 4;   addi r20, r1, 16 }   /* r4 = WORD_12 */
+EX:	{ lw r8, r1;  addi r1, r1, 4;   slt_u r13, r20, r15 }/* r8 = WORD_13 */
+EX:	{ lw r11, r1; addi r1, r1, -52; mvz r20, r13, r1 }  /* r11 = WORD_14 */
+
+	/* Load the three remaining words from the first L1D line, first
+	 * stalling until it has filled by "looking at" r16.
+	 */
+EX:	{ lw r13, r1; addi r1, r1, 4; move zero, r16 }   /* r13 = WORD_1 */
+EX:	{ lw r14, r1; addi r1, r1, 4 }                   /* r14 = WORD_2 */
+EX:	{ lw r15, r1; addi r1, r1, 8; addi r10, r0, 60 } /* r15 = WORD_3 */
+
+	/* Load second word from the second L1D line, first
+	 * stalling until it has filled by "looking at" r17.
+	 */
+EX:	{ lw r19, r1; addi r1, r1, 4; move zero, r17 }  /* r19 = WORD_5 */
+
+	/* Store last word to the destination line, potentially dirtying it
+	 * for the first time, which keeps the L2 busy for two cycles.
+	 */
+EX:	{ sw r10, r12 }                                 /* store(WORD_15) */
+
+	/* Use two L1D hits to cover the sw L2 access above. */
+EX:	{ lw r10, r1; addi r1, r1, 4 }                  /* r10 = WORD_6 */
+EX:	{ lw r12, r1; addi r1, r1, 4 }                  /* r12 = WORD_7 */
+
+	/* Fill third L1D line. */
+EX:	{ lw r18, r1; addi r1, r1, 4 }                  /* r18 = WORD_8 */
+
+	/* Store first L1D line. */
+EX:	{ sw r0, r16; addi r0, r0, 4; add r16, r0, r2 } /* store(WORD_0) */
+EX:	{ sw r0, r13; addi r0, r0, 4; andi r16, r16, -64 } /* store(WORD_1) */
+EX:	{ sw r0, r14; addi r0, r0, 4; slt_u r16, r9, r16 } /* store(WORD_2) */
+EX:	{ sw r0, r15; addi r0, r0, 4; addi r13, sp, -64 } /* store(WORD_3) */
+	/* Store second L1D line. */
+EX:	{ sw r0, r17; addi r0, r0, 4; mvz r9, r16, r13 }/* store(WORD_4) */
+EX:	{ sw r0, r19; addi r0, r0, 4 }                  /* store(WORD_5) */
+EX:	{ sw r0, r10; addi r0, r0, 4 }                  /* store(WORD_6) */
+EX:	{ sw r0, r12; addi r0, r0, 4 }                  /* store(WORD_7) */
+
+EX:	{ lw r13, r1; addi r1, r1, 4; move zero, r18 }  /* r13 = WORD_9 */
+EX:	{ lw r14, r1; addi r1, r1, 4 }                  /* r14 = WORD_10 */
+EX:	{ lw r15, r1; move r1, r20   }                  /* r15 = WORD_11 */
+
+	/* Store third L1D line. */
+EX:	{ sw r0, r18; addi r0, r0, 4 }                  /* store(WORD_8) */
+EX:	{ sw r0, r13; addi r0, r0, 4 }                  /* store(WORD_9) */
+EX:	{ sw r0, r14; addi r0, r0, 4 }                  /* store(WORD_10) */
+EX:	{ sw r0, r15; addi r0, r0, 4 }                  /* store(WORD_11) */
+
+	/* Store rest of fourth L1D line. */
+EX:	{ sw r0, r4;  addi r0, r0, 4 }                  /* store(WORD_12) */
+	{
+EX:	sw r0, r8                                       /* store(WORD_13) */
+	addi r0, r0, 4
+	/* Will r2 be > 64 after we subtract 64 below? */
+	shri r4, r2, 7
+	}
+	{
+EX:	sw r0, r11                                      /* store(WORD_14) */
+	addi r0, r0, 8
+	/* Record 64 bytes successfully copied. */
+	addi r2, r2, -64
+	}
+
+	{ jrp lr; move lr, r27 }
+
+	/* Convey to the backtrace library that the stack frame is size
+	 * zero, and the real return address is on the stack rather than
+	 * in 'lr'.
+	 */
+	{ info 8 }
+
+	.align 64
+.Lcopy_unaligned_maybe_many:
+	/* Skip the setup overhead if we aren't copying many bytes. */
+	{ slti_u r8, r2, 20; sub r4, zero, r0 }
+	{ bnzt r8, .Lcopy_unaligned_few; andi r4, r4, 3 }
+	{ bz r4, .Ldest_is_word_aligned; add r18, r1, r2 }
+
+/*
+ *
+ * unaligned 4 byte at a time copy handler.
+ *
+ */
+
+	/* Copy single bytes until r0 == 0 mod 4, so we can store words. */
+.Lalign_dest_loop:
+EX:	{ lb_u r3, r1; addi r1, r1, 1; addi r4, r4, -1 }
+EX:	{ sb r0, r3;   addi r0, r0, 1; addi r2, r2, -1 }
+	{ bnzt r4, .Lalign_dest_loop; andi r3, r1, 3 }
+
+	/* If source and dest are now *both* aligned, do an aligned copy. */
+	{ bz r3, .Lcheck_aligned_copy_size; addli r4, r2, -256 }
+
+.Ldest_is_word_aligned:
+
+EX:	{ andi r8, r0, 63; lwadd_na r6, r1, 4}
+	{ slti_u r9, r2, 64; bz r8, .Ldest_is_L2_line_aligned }
+
+	/* This copies unaligned words until either there are fewer
+	 * than 4 bytes left to copy, or until the destination pointer
+	 * is cache-aligned, whichever comes first.
+	 *
+	 * On entry:
+	 * - r0 is the next store address.
+	 * - r1 points 4 bytes past the load address corresponding to r0.
+	 * - r2 >= 4
+	 * - r6 is the next aligned word loaded.
+	 */
+.Lcopy_unaligned_src_words:
+EX:	{ lwadd_na r7, r1, 4; slti_u r8, r2, 4 + 4 }
+	/* stall */
+	{ dword_align r6, r7, r1; slti_u r9, r2, 64 + 4 }
+EX:	{ swadd r0, r6, 4; addi r2, r2, -4 }
+	{ bnz r8, .Lcleanup_unaligned_words; andi r8, r0, 63 }
+	{ bnzt r8, .Lcopy_unaligned_src_words; move r6, r7 }
+
+	/* On entry:
+	 * - r0 is the next store address.
+	 * - r1 points 4 bytes past the load address corresponding to r0.
+	 * - r2 >= 4 (# of bytes left to store).
+	 * - r6 is the next aligned src word value.
+	 * - r9 = (r2 < 64U).
+	 * - r18 points one byte past the end of source memory.
+	 */
+.Ldest_is_L2_line_aligned:
+
+	{
+	/* Not a full cache line remains. */
+	bnz r9, .Lcleanup_unaligned_words
+	move r7, r6
+	}
+
+	/* r2 >= 64 */
+
+	/* Kick off two prefetches, but don't go past the end. */
+	{ addi r3, r1, 63 - 4; addi r8, r1, 64 + 63 - 4 }
+	{ prefetch r3; move r3, r8; slt_u r8, r8, r18 }
+	{ mvz r3, r8, r1; addi r8, r3, 64 }
+	{ prefetch r3; move r3, r8; slt_u r8, r8, r18 }
+	{ mvz r3, r8, r1; movei r17, 0 }
+
+.Lcopy_unaligned_line:
+	/* Prefetch another line. */
+	{ prefetch r3; addi r15, r1, 60; addi r3, r3, 64 }
+	/* Fire off a load of the last word we are about to copy. */
+EX:	{ lw_na r15, r15; slt_u r8, r3, r18 }
+
+EX:	{ mvz r3, r8, r1; wh64 r0 }
+
+	/* This loop runs twice.
+	 *
+	 * On entry:
+	 * - r17 is even before the first iteration, and odd before
+	 *   the second.  It is incremented inside the loop.  Encountering
+	 *   an even value at the end of the loop makes it stop.
+	 */
+.Lcopy_half_an_unaligned_line:
+EX:	{
+	/* Stall until the last byte is ready. In the steady state this
+	 * guarantees all words to load below will be in the L2 cache, which
+	 * avoids shunting the loads to the RTF.
+	 */
+	move zero, r15
+	lwadd_na r7, r1, 16
+	}
+EX:	{ lwadd_na r11, r1, 12 }
+EX:	{ lwadd_na r14, r1, -24 }
+EX:	{ lwadd_na r8, r1, 4 }
+EX:	{ lwadd_na r9, r1, 4 }
+EX:	{
+	lwadd_na r10, r1, 8
+	/* r16 = (r2 < 64), after we subtract 32 from r2 below. */
+	slti_u r16, r2, 64 + 32
+	}
+EX:	{ lwadd_na r12, r1, 4; addi r17, r17, 1 }
+EX:	{ lwadd_na r13, r1, 8; dword_align r6, r7, r1 }
+EX:	{ swadd r0, r6,  4; dword_align r7,  r8,  r1 }
+EX:	{ swadd r0, r7,  4; dword_align r8,  r9,  r1 }
+EX:	{ swadd r0, r8,  4; dword_align r9,  r10, r1 }
+EX:	{ swadd r0, r9,  4; dword_align r10, r11, r1 }
+EX:	{ swadd r0, r10, 4; dword_align r11, r12, r1 }
+EX:	{ swadd r0, r11, 4; dword_align r12, r13, r1 }
+EX:	{ swadd r0, r12, 4; dword_align r13, r14, r1 }
+EX:	{ swadd r0, r13, 4; addi r2, r2, -32 }
+	{ move r6, r14; bbst r17, .Lcopy_half_an_unaligned_line }
+
+	{ bzt r16, .Lcopy_unaligned_line; move r7, r6 }
+
+	/* On entry:
+	 * - r0 is the next store address.
+	 * - r1 points 4 bytes past the load address corresponding to r0.
+	 * - r2 >= 0 (# of bytes left to store).
+	 * - r7 is the next aligned src word value.
+	 */
+.Lcleanup_unaligned_words:
+	/* Handle any trailing bytes. */
+	{ bz r2, .Lcopy_unaligned_done; slti_u r8, r2, 4 }
+	{ bzt r8, .Lcopy_unaligned_src_words; move r6, r7 }
+
+	/* Move r1 back to the point where it corresponds to r0. */
+	{ addi r1, r1, -4 }
+
+	/* Fall through */
+
+/*
+ *
+ * 1 byte at a time copy handler.
+ *
+ */
+
+.Lcopy_unaligned_few:
+EX:	{ lb_u r3, r1; addi r1, r1, 1 }
+EX:	{ sb r0, r3;   addi r0, r0, 1; addi r2, r2, -1 }
+	{ bnzt r2, .Lcopy_unaligned_few }
+
+.Lcopy_unaligned_done:
+
+	/* For memcpy return original dest address, else zero. */
+	{ mz r0, r29, r23; jrp lr }
+
+.Lend_memcpy_common:
+	.size memcpy_common, .Lend_memcpy_common - memcpy_common
+
+	.section .fixup,"ax"
+memcpy_common_fixup:
+	.type memcpy_common_fixup, @function
+
+	/* Skip any bytes we already successfully copied.
+	 * r2 (num remaining) is correct, but r0 (dst) and r1 (src)
+	 * may not be quite right because of unrolling and prefetching.
+	 * So we need to recompute their values as the address just
+	 * after the last byte we are sure was successfully loaded and
+	 * then stored.
+	 */
+
+	/* Determine how many bytes we successfully copied. */
+	{ sub r3, r25, r2 }
+
+	/* Add this to the original r0 and r1 to get their new values. */
+	{ add r0, r23, r3; add r1, r24, r3 }
+
+	{ bzt r29, memcpy_fixup_loop }
+	{ blzt r29, copy_to_user_fixup_loop }
+
+copy_from_user_fixup_loop:
+	/* Try copying the rest one byte at a time, expecting a load fault. */
+.Lcfu:	{ lb_u r3, r1; addi r1, r1, 1 }
+	{ sb r0, r3; addi r0, r0, 1; addi r2, r2, -1 }
+	{ bnzt r2, copy_from_user_fixup_loop }
+
+.Lcopy_from_user_fixup_zero_remainder:
+	{ bbs r29, 2f }  /* low bit set means IS_COPY_FROM_USER */
+	/* byte-at-a-time loop faulted, so zero the rest. */
+	{ move r3, r2; bz r2, 2f /* should be impossible, but handle it. */ }
+1:      { sb r0, zero; addi r0, r0, 1; addi r3, r3, -1 }
+	{ bnzt r3, 1b }
+2:	move lr, r27
+	{ move r0, r2; jrp lr }
+
+copy_to_user_fixup_loop:
+	/* Try copying the rest one byte at a time, expecting a store fault. */
+	{ lb_u r3, r1; addi r1, r1, 1 }
+.Lctu:	{ sb r0, r3; addi r0, r0, 1; addi r2, r2, -1 }
+	{ bnzt r2, copy_to_user_fixup_loop }
+.Lcopy_to_user_fixup_done:
+	move lr, r27
+	{ move r0, r2; jrp lr }
+
+memcpy_fixup_loop:
+	/* Try copying the rest one byte at a time. We expect a disastrous
+	 * fault to happen since we are in fixup code, but let it happen.
+	 */
+	{ lb_u r3, r1; addi r1, r1, 1 }
+	{ sb r0, r3; addi r0, r0, 1; addi r2, r2, -1 }
+	{ bnzt r2, memcpy_fixup_loop }
+	/* This should be unreachable, we should have faulted again.
+	 * But be paranoid and handle it in case some interrupt changed
+	 * the TLB or something.
+	 */
+	move lr, r27
+	{ move r0, r23; jrp lr }
+
+	.size memcpy_common_fixup, . - memcpy_common_fixup
+
+	.section __ex_table,"a"
+	.align 4
+	.word .Lcfu, .Lcopy_from_user_fixup_zero_remainder
+	.word .Lctu, .Lcopy_to_user_fixup_done
diff --git a/kernel/arch/tile/lib/memcpy_64.c b/kernel/arch/tile/lib/memcpy_64.c
new file mode 100644
index 000000000..4815354b8
--- /dev/null
+++ b/kernel/arch/tile/lib/memcpy_64.c
@@ -0,0 +1,367 @@
+/*
+ * Copyright 2011 Tilera Corporation. All Rights Reserved.
+ *
+ *   This program is free software; you can redistribute it and/or
+ *   modify it under the terms of the GNU General Public License
+ *   as published by the Free Software Foundation, version 2.
+ *
+ *   This program is distributed in the hope that it will be useful, but
+ *   WITHOUT ANY WARRANTY; without even the implied warranty of
+ *   MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
+ *   NON INFRINGEMENT.  See the GNU General Public License for
+ *   more details.
+ */
+
+#include <linux/types.h>
+#include <linux/string.h>
+#include <linux/module.h>
+/* EXPORT_SYMBOL() is in arch/tile/lib/exports.c since this should be asm. */
+
+/* Must be 8 bytes in size. */
+#define op_t uint64_t
+
+/* Threshold value for when to enter the unrolled loops. */
+#define	OP_T_THRES	16
+
+#if CHIP_L2_LINE_SIZE() != 64
+#error "Assumes 64 byte line size"
+#endif
+
+/* How many cache lines ahead should we prefetch? */
+#define PREFETCH_LINES_AHEAD 4
+
+/*
+ * Provide "base versions" of load and store for the normal code path.
+ * The kernel provides other versions for userspace copies.
+ */
+#define ST(p, v) (*(p) = (v))
+#define LD(p) (*(p))
+
+#ifndef USERCOPY_FUNC
+#define ST1 ST
+#define ST2 ST
+#define ST4 ST
+#define ST8 ST
+#define LD1 LD
+#define LD2 LD
+#define LD4 LD
+#define LD8 LD
+#define RETVAL dstv
+void *memcpy(void *__restrict dstv, const void *__restrict srcv, size_t n)
+#else
+/*
+ * Special kernel version will provide implementation of the LDn/STn
+ * macros to return a count of uncopied bytes due to mm fault.
+ */
+#define RETVAL 0
+int __attribute__((optimize("omit-frame-pointer")))
+USERCOPY_FUNC(void *__restrict dstv, const void *__restrict srcv, size_t n)
+#endif
+{
+	char *__restrict dst1 = (char *)dstv;
+	const char *__restrict src1 = (const char *)srcv;
+	const char *__restrict src1_end;
+	const char *__restrict prefetch;
+	op_t *__restrict dst8;    /* 8-byte pointer to destination memory. */
+	op_t final; /* Final bytes to write to trailing word, if any */
+	long i;
+
+	if (n < 16) {
+		for (; n; n--)
+			ST1(dst1++, LD1(src1++));
+		return RETVAL;
+	}
+
+	/*
+	 * Locate the end of source memory we will copy.  Don't
+	 * prefetch past this.
+	 */
+	src1_end = src1 + n - 1;
+
+	/* Prefetch ahead a few cache lines, but not past the end. */
+	prefetch = src1;
+	for (i = 0; i < PREFETCH_LINES_AHEAD; i++) {
+		__insn_prefetch(prefetch);
+		prefetch += CHIP_L2_LINE_SIZE();
+		prefetch = (prefetch < src1_end) ? prefetch : src1;
+	}
+
+	/* Copy bytes until dst is word-aligned. */
+	for (; (uintptr_t)dst1 & (sizeof(op_t) - 1); n--)
+		ST1(dst1++, LD1(src1++));
+
+	/* 8-byte pointer to destination memory. */
+	dst8 = (op_t *)dst1;
+
+	if (__builtin_expect((uintptr_t)src1 & (sizeof(op_t) - 1), 0)) {
+		/* Unaligned copy. */
+
+		op_t  tmp0 = 0, tmp1 = 0, tmp2, tmp3;
+		const op_t *src8 = (const op_t *) ((uintptr_t)src1 &
+						   -sizeof(op_t));
+		const void *srci = (void *)src1;
+		int m;
+
+		m = (CHIP_L2_LINE_SIZE() << 2) -
+			(((uintptr_t)dst8) & ((CHIP_L2_LINE_SIZE() << 2) - 1));
+		m = (n < m) ? n : m;
+		m /= sizeof(op_t);
+
+		/* Copy until 'dst' is cache-line-aligned. */
+		n -= (sizeof(op_t) * m);
+
+		switch (m % 4) {
+		case 0:
+			if (__builtin_expect(!m, 0))
+				goto _M0;
+			tmp1 = LD8(src8++);
+			tmp2 = LD8(src8++);
+			goto _8B3;
+		case 2:
+			m += 2;
+			tmp3 = LD8(src8++);
+			tmp0 = LD8(src8++);
+			goto _8B1;
+		case 3:
+			m += 1;
+			tmp2 = LD8(src8++);
+			tmp3 = LD8(src8++);
+			goto _8B2;
+		case 1:
+			m--;
+			tmp0 = LD8(src8++);
+			tmp1 = LD8(src8++);
+			if (__builtin_expect(!m, 0))
+				goto _8B0;
+		}
+
+		do {
+			tmp2 = LD8(src8++);
+			tmp0 =  __insn_dblalign(tmp0, tmp1, srci);
+			ST8(dst8++, tmp0);
+_8B3:
+			tmp3 = LD8(src8++);
+			tmp1 = __insn_dblalign(tmp1, tmp2, srci);
+			ST8(dst8++, tmp1);
+_8B2:
+			tmp0 = LD8(src8++);
+			tmp2 = __insn_dblalign(tmp2, tmp3, srci);
+			ST8(dst8++, tmp2);
+_8B1:
+			tmp1 = LD8(src8++);
+			tmp3 = __insn_dblalign(tmp3, tmp0, srci);
+			ST8(dst8++, tmp3);
+			m -= 4;
+		} while (m);
+
+_8B0:
+		tmp0 = __insn_dblalign(tmp0, tmp1, srci);
+		ST8(dst8++, tmp0);
+		src8--;
+
+_M0:
+		if (__builtin_expect(n >= CHIP_L2_LINE_SIZE(), 0)) {
+			op_t tmp4, tmp5, tmp6, tmp7, tmp8;
+
+			prefetch = ((const char *)src8) +
+				CHIP_L2_LINE_SIZE() * PREFETCH_LINES_AHEAD;
+
+			for (tmp0 = LD8(src8++); n >= CHIP_L2_LINE_SIZE();
+			     n -= CHIP_L2_LINE_SIZE()) {
+				/* Prefetch and advance to next line to
+				   prefetch, but don't go past the end.  */
+				__insn_prefetch(prefetch);
+
+				/* Make sure prefetch got scheduled
+				   earlier.  */
+				__asm__ ("" : : : "memory");
+
+				prefetch += CHIP_L2_LINE_SIZE();
+				prefetch = (prefetch < src1_end) ? prefetch :
+					(const char *) src8;
+
+				tmp1 = LD8(src8++);
+				tmp2 = LD8(src8++);
+				tmp3 = LD8(src8++);
+				tmp4 = LD8(src8++);
+				tmp5 = LD8(src8++);
+				tmp6 = LD8(src8++);
+				tmp7 = LD8(src8++);
+				tmp8 = LD8(src8++);
+
+				tmp0 = __insn_dblalign(tmp0, tmp1, srci);
+				tmp1 = __insn_dblalign(tmp1, tmp2, srci);
+				tmp2 = __insn_dblalign(tmp2, tmp3, srci);
+				tmp3 = __insn_dblalign(tmp3, tmp4, srci);
+				tmp4 = __insn_dblalign(tmp4, tmp5, srci);
+				tmp5 = __insn_dblalign(tmp5, tmp6, srci);
+				tmp6 = __insn_dblalign(tmp6, tmp7, srci);
+				tmp7 = __insn_dblalign(tmp7, tmp8, srci);
+
+				__insn_wh64(dst8);
+
+				ST8(dst8++, tmp0);
+				ST8(dst8++, tmp1);
+				ST8(dst8++, tmp2);
+				ST8(dst8++, tmp3);
+				ST8(dst8++, tmp4);
+				ST8(dst8++, tmp5);
+				ST8(dst8++, tmp6);
+				ST8(dst8++, tmp7);
+
+				tmp0 = tmp8;
+			}
+			src8--;
+		}
+
+		/* Copy the rest 8-byte chunks. */
+		if (n >= sizeof(op_t)) {
+			tmp0 = LD8(src8++);
+			for (; n >= sizeof(op_t); n -= sizeof(op_t)) {
+				tmp1 = LD8(src8++);
+				tmp0 = __insn_dblalign(tmp0, tmp1, srci);
+				ST8(dst8++, tmp0);
+				tmp0 = tmp1;
+			}
+			src8--;
+		}
+
+		if (n == 0)
+			return RETVAL;
+
+		tmp0 = LD8(src8++);
+		tmp1 = ((const char *)src8 <= src1_end)
+			? LD8((op_t *)src8) : 0;
+		final = __insn_dblalign(tmp0, tmp1, srci);
+
+	} else {
+		/* Aligned copy. */
+
+		const op_t *__restrict src8 = (const op_t *)src1;
+
+		/* src8 and dst8 are both word-aligned. */
+		if (n >= CHIP_L2_LINE_SIZE()) {
+			/* Copy until 'dst' is cache-line-aligned. */
+			for (; (uintptr_t)dst8 & (CHIP_L2_LINE_SIZE() - 1);
+			     n -= sizeof(op_t))
+				ST8(dst8++, LD8(src8++));
+
+			for (; n >= CHIP_L2_LINE_SIZE(); ) {
+				op_t tmp0, tmp1, tmp2, tmp3;
+				op_t tmp4, tmp5, tmp6, tmp7;
+
+				/*
+				 * Prefetch and advance to next line
+				 * to prefetch, but don't go past the
+				 * end.
+				 */
+				__insn_prefetch(prefetch);
+
+				/* Make sure prefetch got scheduled
+				   earlier.  */
+				__asm__ ("" : : : "memory");
+
+				prefetch += CHIP_L2_LINE_SIZE();
+				prefetch = (prefetch < src1_end) ? prefetch :
+					(const char *)src8;
+
+				/*
+				 * Do all the loads before wh64.  This
+				 * is necessary if [src8, src8+7] and
+				 * [dst8, dst8+7] share the same cache
+				 * line and dst8 <= src8, as can be
+				 * the case when called from memmove,
+				 * or with code tested on x86 whose
+				 * memcpy always works with forward
+				 * copies.
+				 */
+				tmp0 = LD8(src8++);
+				tmp1 = LD8(src8++);
+				tmp2 = LD8(src8++);
+				tmp3 = LD8(src8++);
+				tmp4 = LD8(src8++);
+				tmp5 = LD8(src8++);
+				tmp6 = LD8(src8++);
+				tmp7 = LD8(src8++);
+
+				/* wh64 and wait for tmp7 load completion. */
+				__asm__ ("move %0, %0; wh64 %1\n"
+					 : : "r"(tmp7), "r"(dst8));
+
+				ST8(dst8++, tmp0);
+				ST8(dst8++, tmp1);
+				ST8(dst8++, tmp2);
+				ST8(dst8++, tmp3);
+				ST8(dst8++, tmp4);
+				ST8(dst8++, tmp5);
+				ST8(dst8++, tmp6);
+				ST8(dst8++, tmp7);
+
+				n -= CHIP_L2_LINE_SIZE();
+			}
+#if CHIP_L2_LINE_SIZE() != 64
+# error "Fix code that assumes particular L2 cache line size."
+#endif
+		}
+
+		for (; n >= sizeof(op_t); n -= sizeof(op_t))
+			ST8(dst8++, LD8(src8++));
+
+		if (__builtin_expect(n == 0, 1))
+			return RETVAL;
+
+		final = LD8(src8);
+	}
+
+	/* n != 0 if we get here.  Write out any trailing bytes. */
+	dst1 = (char *)dst8;
+#ifndef __BIG_ENDIAN__
+	if (n & 4) {
+		ST4((uint32_t *)dst1, final);
+		dst1 += 4;
+		final >>= 32;
+		n &= 3;
+	}
+	if (n & 2) {
+		ST2((uint16_t *)dst1, final);
+		dst1 += 2;
+		final >>= 16;
+		n &= 1;
+	}
+	if (n)
+		ST1((uint8_t *)dst1, final);
+#else
+	if (n & 4) {
+		ST4((uint32_t *)dst1, final >> 32);
+		dst1 += 4;
+        }
+        else
+        {
+		final >>= 32;
+        }
+	if (n & 2) {
+		ST2((uint16_t *)dst1, final >> 16);
+		dst1 += 2;
+        }
+        else
+        {
+		final >>= 16;
+        }
+	if (n & 1)
+		ST1((uint8_t *)dst1, final >> 8);
+#endif
+
+	return RETVAL;
+}
+
+#ifdef USERCOPY_FUNC
+#undef ST1
+#undef ST2
+#undef ST4
+#undef ST8
+#undef LD1
+#undef LD2
+#undef LD4
+#undef LD8
+#undef USERCOPY_FUNC
+#endif
diff --git a/kernel/arch/tile/lib/memcpy_user_64.c b/kernel/arch/tile/lib/memcpy_user_64.c
new file mode 100644
index 000000000..88c701649
--- /dev/null
+++ b/kernel/arch/tile/lib/memcpy_user_64.c
@@ -0,0 +1,94 @@
+/*
+ * Copyright 2011 Tilera Corporation. All Rights Reserved.
+ *
+ *   This program is free software; you can redistribute it and/or
+ *   modify it under the terms of the GNU General Public License
+ *   as published by the Free Software Foundation, version 2.
+ *
+ *   This program is distributed in the hope that it will be useful, but
+ *   WITHOUT ANY WARRANTY; without even the implied warranty of
+ *   MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
+ *   NON INFRINGEMENT.  See the GNU General Public License for
+ *   more details.
+ *
+ * Do memcpy(), but trap and return "n" when a load or store faults.
+ *
+ * Note: this idiom only works when memcpy() compiles to a leaf function.
+ * Here leaf function not only means it does not have calls, but also
+ * requires no stack operations (sp, stack frame pointer) and no
+ * use of callee-saved registers, else "jrp lr" will be incorrect since
+ * unwinding stack frame is bypassed. Since memcpy() is not complex so
+ * these conditions are satisfied here, but we need to be careful when
+ * modifying this file. This is not a clean solution but is the best
+ * one so far.
+ *
+ * Also note that we are capturing "n" from the containing scope here.
+ */
+
+#define _ST(p, inst, v)						\
+	({							\
+		asm("1: " #inst " %0, %1;"			\
+		    ".pushsection .coldtext.memcpy,\"ax\";"	\
+		    "2: { move r0, %2; jrp lr };"		\
+		    ".section __ex_table,\"a\";"		\
+		    ".align 8;"					\
+		    ".quad 1b, 2b;"				\
+		    ".popsection"				\
+		    : "=m" (*(p)) : "r" (v), "r" (n));		\
+	})
+
+#define _LD(p, inst)						\
+	({							\
+		unsigned long __v;				\
+		asm("1: " #inst " %0, %1;"			\
+		    ".pushsection .coldtext.memcpy,\"ax\";"	\
+		    "2: { move r0, %2; jrp lr };"		\
+		    ".section __ex_table,\"a\";"		\
+		    ".align 8;"					\
+		    ".quad 1b, 2b;"				\
+		    ".popsection"				\
+		    : "=r" (__v) : "m" (*(p)), "r" (n));	\
+		__v;						\
+	})
+
+#define USERCOPY_FUNC __copy_to_user_inatomic
+#define ST1(p, v) _ST((p), st1, (v))
+#define ST2(p, v) _ST((p), st2, (v))
+#define ST4(p, v) _ST((p), st4, (v))
+#define ST8(p, v) _ST((p), st, (v))
+#define LD1 LD
+#define LD2 LD
+#define LD4 LD
+#define LD8 LD
+#include "memcpy_64.c"
+
+#define USERCOPY_FUNC __copy_from_user_inatomic
+#define ST1 ST
+#define ST2 ST
+#define ST4 ST
+#define ST8 ST
+#define LD1(p) _LD((p), ld1u)
+#define LD2(p) _LD((p), ld2u)
+#define LD4(p) _LD((p), ld4u)
+#define LD8(p) _LD((p), ld)
+#include "memcpy_64.c"
+
+#define USERCOPY_FUNC __copy_in_user_inatomic
+#define ST1(p, v) _ST((p), st1, (v))
+#define ST2(p, v) _ST((p), st2, (v))
+#define ST4(p, v) _ST((p), st4, (v))
+#define ST8(p, v) _ST((p), st, (v))
+#define LD1(p) _LD((p), ld1u)
+#define LD2(p) _LD((p), ld2u)
+#define LD4(p) _LD((p), ld4u)
+#define LD8(p) _LD((p), ld)
+#include "memcpy_64.c"
+
+unsigned long __copy_from_user_zeroing(void *to, const void __user *from,
+				       unsigned long n)
+{
+	unsigned long rc = __copy_from_user_inatomic(to, from, n);
+	if (unlikely(rc))
+		memset(to + n - rc, 0, rc);
+	return rc;
+}
diff --git a/kernel/arch/tile/lib/memmove.c b/kernel/arch/tile/lib/memmove.c
new file mode 100644
index 000000000..fd615ae6a
--- /dev/null
+++ b/kernel/arch/tile/lib/memmove.c
@@ -0,0 +1,63 @@
+/*
+ * Copyright 2010 Tilera Corporation. All Rights Reserved.
+ *
+ *   This program is free software; you can redistribute it and/or
+ *   modify it under the terms of the GNU General Public License
+ *   as published by the Free Software Foundation, version 2.
+ *
+ *   This program is distributed in the hope that it will be useful, but
+ *   WITHOUT ANY WARRANTY; without even the implied warranty of
+ *   MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
+ *   NON INFRINGEMENT.  See the GNU General Public License for
+ *   more details.
+ */
+
+#include <linux/types.h>
+#include <linux/string.h>
+#include <linux/module.h>
+
+void *memmove(void *dest, const void *src, size_t n)
+{
+	if ((const char *)src >= (char *)dest + n
+	    || (char *)dest >= (const char *)src + n) {
+		/* We found no overlap, so let memcpy do all the heavy
+		 * lifting (prefetching, etc.)
+		 */
+		return memcpy(dest, src, n);
+	}
+
+	if (n != 0) {
+		const uint8_t *in;
+		uint8_t x;
+		uint8_t *out;
+		int stride;
+
+		if (src < dest) {
+			/* copy backwards */
+			in = (const uint8_t *)src + n - 1;
+			out = (uint8_t *)dest + n - 1;
+			stride = -1;
+		} else {
+			/* copy forwards */
+			in = (const uint8_t *)src;
+			out = (uint8_t *)dest;
+			stride = 1;
+		}
+
+		/* Manually software-pipeline this loop. */
+		x = *in;
+		in += stride;
+
+		while (--n != 0) {
+			*out = x;
+			out += stride;
+			x = *in;
+			in += stride;
+		}
+
+		*out = x;
+	}
+
+	return dest;
+}
+EXPORT_SYMBOL(memmove);
diff --git a/kernel/arch/tile/lib/memset_32.c b/kernel/arch/tile/lib/memset_32.c
new file mode 100644
index 000000000..2042bfe65
--- /dev/null
+++ b/kernel/arch/tile/lib/memset_32.c
@@ -0,0 +1,143 @@
+/*
+ * Copyright 2010 Tilera Corporation. All Rights Reserved.
+ *
+ *   This program is free software; you can redistribute it and/or
+ *   modify it under the terms of the GNU General Public License
+ *   as published by the Free Software Foundation, version 2.
+ *
+ *   This program is distributed in the hope that it will be useful, but
+ *   WITHOUT ANY WARRANTY; without even the implied warranty of
+ *   MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
+ *   NON INFRINGEMENT.  See the GNU General Public License for
+ *   more details.
+ */
+
+#include <linux/types.h>
+#include <linux/string.h>
+#include <linux/module.h>
+#include <arch/chip.h>
+
+void *memset(void *s, int c, size_t n)
+{
+	uint32_t *out32;
+	int n32;
+	uint32_t v16, v32;
+	uint8_t *out8 = s;
+	int to_align32;
+
+	/* Experimentation shows that a trivial tight loop is a win up until
+	 * around a size of 20, where writing a word at a time starts to win.
+	 */
+#define BYTE_CUTOFF 20
+
+#if BYTE_CUTOFF < 3
+	/* This must be at least at least this big, or some code later
+	 * on doesn't work.
+	 */
+#error "BYTE_CUTOFF is too small"
+#endif
+
+	if (n < BYTE_CUTOFF) {
+		/* Strangely, this turns out to be the tightest way to
+		 * write this loop.
+		 */
+		if (n != 0) {
+			do {
+				/* Strangely, combining these into one line
+				 * performs worse.
+				 */
+				*out8 = c;
+				out8++;
+			} while (--n != 0);
+		}
+
+		return s;
+	}
+
+	/* Align 'out8'. We know n >= 3 so this won't write past the end. */
+	while (((uintptr_t) out8 & 3) != 0) {
+		*out8++ = c;
+		--n;
+	}
+
+	/* Align 'n'. */
+	while (n & 3)
+		out8[--n] = c;
+
+	out32 = (uint32_t *) out8;
+	n32 = n >> 2;
+
+	/* Tile input byte out to 32 bits. */
+	v16 = __insn_intlb(c, c);
+	v32 = __insn_intlh(v16, v16);
+
+	/* This must be at least 8 or the following loop doesn't work. */
+#define CACHE_LINE_SIZE_IN_WORDS (CHIP_L2_LINE_SIZE() / 4)
+
+	/* Determine how many words we need to emit before the 'out32'
+	 * pointer becomes aligned modulo the cache line size.
+	 */
+	to_align32 =
+		(-((uintptr_t)out32 >> 2)) & (CACHE_LINE_SIZE_IN_WORDS - 1);
+
+	/* Only bother aligning and using wh64 if there is at least
+	 * one full cache line to process.  This check also prevents
+	 * overrunning the end of the buffer with alignment words.
+	 */
+	if (to_align32 <= n32 - CACHE_LINE_SIZE_IN_WORDS) {
+		int lines_left;
+
+		/* Align out32 mod the cache line size so we can use wh64. */
+		n32 -= to_align32;
+		for (; to_align32 != 0; to_align32--) {
+			*out32 = v32;
+			out32++;
+		}
+
+		/* Use unsigned divide to turn this into a right shift. */
+		lines_left = (unsigned)n32 / CACHE_LINE_SIZE_IN_WORDS;
+
+		do {
+			/* Only wh64 a few lines at a time, so we don't
+			 * exceed the maximum number of victim lines.
+			 */
+			int x = ((lines_left < CHIP_MAX_OUTSTANDING_VICTIMS())
+				  ? lines_left
+				  : CHIP_MAX_OUTSTANDING_VICTIMS());
+			uint32_t *wh = out32;
+			int i = x;
+			int j;
+
+			lines_left -= x;
+
+			do {
+				__insn_wh64(wh);
+				wh += CACHE_LINE_SIZE_IN_WORDS;
+			} while (--i);
+
+			for (j = x * (CACHE_LINE_SIZE_IN_WORDS / 4);
+			     j != 0; j--) {
+				*out32++ = v32;
+				*out32++ = v32;
+				*out32++ = v32;
+				*out32++ = v32;
+			}
+		} while (lines_left != 0);
+
+		/* We processed all full lines above, so only this many
+		 * words remain to be processed.
+		 */
+		n32 &= CACHE_LINE_SIZE_IN_WORDS - 1;
+	}
+
+	/* Now handle any leftover values. */
+	if (n32 != 0) {
+		do {
+			*out32 = v32;
+			out32++;
+		} while (--n32 != 0);
+	}
+
+	return s;
+}
+EXPORT_SYMBOL(memset);
diff --git a/kernel/arch/tile/lib/memset_64.c b/kernel/arch/tile/lib/memset_64.c
new file mode 100644
index 000000000..03ef69cd7
--- /dev/null
+++ b/kernel/arch/tile/lib/memset_64.c
@@ -0,0 +1,142 @@
+/*
+ * Copyright 2011 Tilera Corporation. All Rights Reserved.
+ *
+ *   This program is free software; you can redistribute it and/or
+ *   modify it under the terms of the GNU General Public License
+ *   as published by the Free Software Foundation, version 2.
+ *
+ *   This program is distributed in the hope that it will be useful, but
+ *   WITHOUT ANY WARRANTY; without even the implied warranty of
+ *   MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
+ *   NON INFRINGEMENT.  See the GNU General Public License for
+ *   more details.
+ */
+
+#include <linux/types.h>
+#include <linux/string.h>
+#include <linux/module.h>
+#include <arch/chip.h>
+#include "string-endian.h"
+
+void *memset(void *s, int c, size_t n)
+{
+	uint64_t *out64;
+	int n64, to_align64;
+	uint64_t v64;
+	uint8_t *out8 = s;
+
+	/* Experimentation shows that a trivial tight loop is a win up until
+	 * around a size of 20, where writing a word at a time starts to win.
+	 */
+#define BYTE_CUTOFF 20
+
+#if BYTE_CUTOFF < 7
+	/* This must be at least at least this big, or some code later
+	 * on doesn't work.
+	 */
+#error "BYTE_CUTOFF is too small"
+#endif
+
+	if (n < BYTE_CUTOFF) {
+		/* Strangely, this turns out to be the tightest way to
+		 * write this loop.
+		 */
+		if (n != 0) {
+			do {
+				/* Strangely, combining these into one line
+				 * performs worse.
+				 */
+				*out8 = c;
+				out8++;
+			} while (--n != 0);
+		}
+
+		return s;
+	}
+
+	/* Align 'out8'. We know n >= 7 so this won't write past the end. */
+	while (((uintptr_t) out8 & 7) != 0) {
+		*out8++ = c;
+		--n;
+	}
+
+	/* Align 'n'. */
+	while (n & 7)
+		out8[--n] = c;
+
+	out64 = (uint64_t *) out8;
+	n64 = n >> 3;
+
+	/* Tile input byte out to 64 bits. */
+	v64 = copy_byte(c);
+
+	/* This must be at least 8 or the following loop doesn't work. */
+#define CACHE_LINE_SIZE_IN_DOUBLEWORDS (CHIP_L2_LINE_SIZE() / 8)
+
+	/* Determine how many words we need to emit before the 'out32'
+	 * pointer becomes aligned modulo the cache line size.
+	 */
+	to_align64 = (-((uintptr_t)out64 >> 3)) &
+		(CACHE_LINE_SIZE_IN_DOUBLEWORDS - 1);
+
+	/* Only bother aligning and using wh64 if there is at least
+	 * one full cache line to process.  This check also prevents
+	 * overrunning the end of the buffer with alignment words.
+	 */
+	if (to_align64 <= n64 - CACHE_LINE_SIZE_IN_DOUBLEWORDS) {
+		int lines_left;
+
+		/* Align out64 mod the cache line size so we can use wh64. */
+		n64 -= to_align64;
+		for (; to_align64 != 0; to_align64--) {
+			*out64 = v64;
+			out64++;
+		}
+
+		/* Use unsigned divide to turn this into a right shift. */
+		lines_left = (unsigned)n64 / CACHE_LINE_SIZE_IN_DOUBLEWORDS;
+
+		do {
+			/* Only wh64 a few lines at a time, so we don't
+			 * exceed the maximum number of victim lines.
+			 */
+			int x = ((lines_left < CHIP_MAX_OUTSTANDING_VICTIMS())
+				  ? lines_left
+				  : CHIP_MAX_OUTSTANDING_VICTIMS());
+			uint64_t *wh = out64;
+			int i = x;
+			int j;
+
+			lines_left -= x;
+
+			do {
+				__insn_wh64(wh);
+				wh += CACHE_LINE_SIZE_IN_DOUBLEWORDS;
+			} while (--i);
+
+			for (j = x * (CACHE_LINE_SIZE_IN_DOUBLEWORDS / 4);
+			     j != 0; j--) {
+				*out64++ = v64;
+				*out64++ = v64;
+				*out64++ = v64;
+				*out64++ = v64;
+			}
+		} while (lines_left != 0);
+
+		/* We processed all full lines above, so only this many
+		 * words remain to be processed.
+		 */
+		n64 &= CACHE_LINE_SIZE_IN_DOUBLEWORDS - 1;
+	}
+
+	/* Now handle any leftover values. */
+	if (n64 != 0) {
+		do {
+			*out64 = v64;
+			out64++;
+		} while (--n64 != 0);
+	}
+
+	return s;
+}
+EXPORT_SYMBOL(memset);
diff --git a/kernel/arch/tile/lib/spinlock_32.c b/kernel/arch/tile/lib/spinlock_32.c
new file mode 100644
index 000000000..b34f79aad
--- /dev/null
+++ b/kernel/arch/tile/lib/spinlock_32.c
@@ -0,0 +1,259 @@
+/*
+ * Copyright 2010 Tilera Corporation. All Rights Reserved.
+ *
+ *   This program is free software; you can redistribute it and/or
+ *   modify it under the terms of the GNU General Public License
+ *   as published by the Free Software Foundation, version 2.
+ *
+ *   This program is distributed in the hope that it will be useful, but
+ *   WITHOUT ANY WARRANTY; without even the implied warranty of
+ *   MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
+ *   NON INFRINGEMENT.  See the GNU General Public License for
+ *   more details.
+ */
+
+#include <linux/spinlock.h>
+#include <linux/module.h>
+#include <asm/processor.h>
+#include <arch/spr_def.h>
+
+#include "spinlock_common.h"
+
+void arch_spin_lock(arch_spinlock_t *lock)
+{
+	int my_ticket;
+	int iterations = 0;
+	int delta;
+
+	while ((my_ticket = __insn_tns((void *)&lock->next_ticket)) & 1)
+		delay_backoff(iterations++);
+
+	/* Increment the next ticket number, implicitly releasing tns lock. */
+	lock->next_ticket = my_ticket + TICKET_QUANTUM;
+
+	/* Wait until it's our turn. */
+	while ((delta = my_ticket - lock->current_ticket) != 0)
+		relax((128 / CYCLES_PER_RELAX_LOOP) * delta);
+}
+EXPORT_SYMBOL(arch_spin_lock);
+
+int arch_spin_trylock(arch_spinlock_t *lock)
+{
+	/*
+	 * Grab a ticket; no need to retry if it's busy, we'll just
+	 * treat that the same as "locked", since someone else
+	 * will lock it momentarily anyway.
+	 */
+	int my_ticket = __insn_tns((void *)&lock->next_ticket);
+
+	if (my_ticket == lock->current_ticket) {
+		/* Not currently locked, so lock it by keeping this ticket. */
+		lock->next_ticket = my_ticket + TICKET_QUANTUM;
+		/* Success! */
+		return 1;
+	}
+
+	if (!(my_ticket & 1)) {
+		/* Release next_ticket. */
+		lock->next_ticket = my_ticket;
+	}
+
+	return 0;
+}
+EXPORT_SYMBOL(arch_spin_trylock);
+
+void arch_spin_unlock_wait(arch_spinlock_t *lock)
+{
+	u32 iterations = 0;
+	while (arch_spin_is_locked(lock))
+		delay_backoff(iterations++);
+}
+EXPORT_SYMBOL(arch_spin_unlock_wait);
+
+/*
+ * The low byte is always reserved to be the marker for a "tns" operation
+ * since the low bit is set to "1" by a tns.  The next seven bits are
+ * zeroes.  The next byte holds the "next" writer value, i.e. the ticket
+ * available for the next task that wants to write.  The third byte holds
+ * the current writer value, i.e. the writer who holds the current ticket.
+ * If current == next == 0, there are no interested writers.
+ */
+#define WR_NEXT_SHIFT   _WR_NEXT_SHIFT
+#define WR_CURR_SHIFT   _WR_CURR_SHIFT
+#define WR_WIDTH        _WR_WIDTH
+#define WR_MASK         ((1 << WR_WIDTH) - 1)
+
+/*
+ * The last eight bits hold the active reader count.  This has to be
+ * zero before a writer can start to write.
+ */
+#define RD_COUNT_SHIFT  _RD_COUNT_SHIFT
+#define RD_COUNT_WIDTH  _RD_COUNT_WIDTH
+#define RD_COUNT_MASK   ((1 << RD_COUNT_WIDTH) - 1)
+
+
+/*
+ * We can get the read lock if everything but the reader bits (which
+ * are in the high part of the word) is zero, i.e. no active or
+ * waiting writers, no tns.
+ *
+ * We guard the tns/store-back with an interrupt critical section to
+ * preserve the semantic that the same read lock can be acquired in an
+ * interrupt context.
+ */
+int arch_read_trylock(arch_rwlock_t *rwlock)
+{
+	u32 val;
+	__insn_mtspr(SPR_INTERRUPT_CRITICAL_SECTION, 1);
+	val = __insn_tns((int *)&rwlock->lock);
+	if (likely((val << _RD_COUNT_WIDTH) == 0)) {
+		val += 1 << RD_COUNT_SHIFT;
+		rwlock->lock = val;
+		__insn_mtspr(SPR_INTERRUPT_CRITICAL_SECTION, 0);
+		BUG_ON(val == 0);  /* we don't expect wraparound */
+		return 1;
+	}
+	if ((val & 1) == 0)
+		rwlock->lock = val;
+	__insn_mtspr(SPR_INTERRUPT_CRITICAL_SECTION, 0);
+	return 0;
+}
+EXPORT_SYMBOL(arch_read_trylock);
+
+/*
+ * Spin doing arch_read_trylock() until we acquire the lock.
+ * ISSUE: This approach can permanently starve readers.  A reader who sees
+ * a writer could instead take a ticket lock (just like a writer would),
+ * and atomically enter read mode (with 1 reader) when it gets the ticket.
+ * This way both readers and writers would always make forward progress
+ * in a finite time.
+ */
+void arch_read_lock(arch_rwlock_t *rwlock)
+{
+	u32 iterations = 0;
+	while (unlikely(!arch_read_trylock(rwlock)))
+		delay_backoff(iterations++);
+}
+EXPORT_SYMBOL(arch_read_lock);
+
+void arch_read_unlock(arch_rwlock_t *rwlock)
+{
+	u32 val, iterations = 0;
+
+	mb();  /* guarantee anything modified under the lock is visible */
+	for (;;) {
+		__insn_mtspr(SPR_INTERRUPT_CRITICAL_SECTION, 1);
+		val = __insn_tns((int *)&rwlock->lock);
+		if (likely((val & 1) == 0)) {
+			rwlock->lock = val - (1 << _RD_COUNT_SHIFT);
+			__insn_mtspr(SPR_INTERRUPT_CRITICAL_SECTION, 0);
+			break;
+		}
+		__insn_mtspr(SPR_INTERRUPT_CRITICAL_SECTION, 0);
+		delay_backoff(iterations++);
+	}
+}
+EXPORT_SYMBOL(arch_read_unlock);
+
+/*
+ * We don't need an interrupt critical section here (unlike for
+ * arch_read_lock) since we should never use a bare write lock where
+ * it could be interrupted by code that could try to re-acquire it.
+ */
+void arch_write_lock(arch_rwlock_t *rwlock)
+{
+	/*
+	 * The trailing underscore on this variable (and curr_ below)
+	 * reminds us that the high bits are garbage; we mask them out
+	 * when we compare them.
+	 */
+	u32 my_ticket_;
+	u32 iterations = 0;
+	u32 val = __insn_tns((int *)&rwlock->lock);
+
+	if (likely(val == 0)) {
+		rwlock->lock = 1 << _WR_NEXT_SHIFT;
+		return;
+	}
+
+	/*
+	 * Wait until there are no readers, then bump up the next
+	 * field and capture the ticket value.
+	 */
+	for (;;) {
+		if (!(val & 1)) {
+			if ((val >> RD_COUNT_SHIFT) == 0)
+				break;
+			rwlock->lock = val;
+		}
+		delay_backoff(iterations++);
+		val = __insn_tns((int *)&rwlock->lock);
+	}
+
+	/* Take out the next ticket and extract my ticket value. */
+	rwlock->lock = __insn_addb(val, 1 << WR_NEXT_SHIFT);
+	my_ticket_ = val >> WR_NEXT_SHIFT;
+
+	/* Wait until the "current" field matches our ticket. */
+	for (;;) {
+		u32 curr_ = val >> WR_CURR_SHIFT;
+		u32 delta = ((my_ticket_ - curr_) & WR_MASK);
+		if (likely(delta == 0))
+			break;
+
+		/* Delay based on how many lock-holders are still out there. */
+		relax((256 / CYCLES_PER_RELAX_LOOP) * delta);
+
+		/*
+		 * Get a non-tns value to check; we don't need to tns
+		 * it ourselves.  Since we're not tns'ing, we retry
+		 * more rapidly to get a valid value.
+		 */
+		while ((val = rwlock->lock) & 1)
+			relax(4);
+	}
+}
+EXPORT_SYMBOL(arch_write_lock);
+
+int arch_write_trylock(arch_rwlock_t *rwlock)
+{
+	u32 val = __insn_tns((int *)&rwlock->lock);
+
+	/*
+	 * If a tns is in progress, or there's a waiting or active locker,
+	 * or active readers, we can't take the lock, so give up.
+	 */
+	if (unlikely(val != 0)) {
+		if (!(val & 1))
+			rwlock->lock = val;
+		return 0;
+	}
+
+	/* Set the "next" field to mark it locked. */
+	rwlock->lock = 1 << _WR_NEXT_SHIFT;
+	return 1;
+}
+EXPORT_SYMBOL(arch_write_trylock);
+
+void arch_write_unlock(arch_rwlock_t *rwlock)
+{
+	u32 val, eq, mask;
+
+	mb();  /* guarantee anything modified under the lock is visible */
+	val = __insn_tns((int *)&rwlock->lock);
+	if (likely(val == (1 << _WR_NEXT_SHIFT))) {
+		rwlock->lock = 0;
+		return;
+	}
+	while (unlikely(val & 1)) {
+		/* Limited backoff since we are the highest-priority task. */
+		relax(4);
+		val = __insn_tns((int *)&rwlock->lock);
+	}
+	mask = 1 << WR_CURR_SHIFT;
+	val = __insn_addb(val, mask);
+	eq = __insn_seqb(val, val << (WR_CURR_SHIFT - WR_NEXT_SHIFT));
+	val = __insn_mz(eq & mask, val);
+	rwlock->lock = val;
+}
+EXPORT_SYMBOL(arch_write_unlock);
diff --git a/kernel/arch/tile/lib/spinlock_64.c b/kernel/arch/tile/lib/spinlock_64.c
new file mode 100644
index 000000000..d6fb9581e
--- /dev/null
+++ b/kernel/arch/tile/lib/spinlock_64.c
@@ -0,0 +1,104 @@
+/*
+ * Copyright 2011 Tilera Corporation. All Rights Reserved.
+ *
+ *   This program is free software; you can redistribute it and/or
+ *   modify it under the terms of the GNU General Public License
+ *   as published by the Free Software Foundation, version 2.
+ *
+ *   This program is distributed in the hope that it will be useful, but
+ *   WITHOUT ANY WARRANTY; without even the implied warranty of
+ *   MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
+ *   NON INFRINGEMENT.  See the GNU General Public License for
+ *   more details.
+ */
+
+#include <linux/spinlock.h>
+#include <linux/module.h>
+#include <asm/processor.h>
+
+#include "spinlock_common.h"
+
+/*
+ * Read the spinlock value without allocating in our cache and without
+ * causing an invalidation to another cpu with a copy of the cacheline.
+ * This is important when we are spinning waiting for the lock.
+ */
+static inline u32 arch_spin_read_noalloc(void *lock)
+{
+	return atomic_cmpxchg((atomic_t *)lock, -1, -1);
+}
+
+/*
+ * Wait until the high bits (current) match my ticket.
+ * If we notice the overflow bit set on entry, we clear it.
+ */
+void arch_spin_lock_slow(arch_spinlock_t *lock, u32 my_ticket)
+{
+	if (unlikely(my_ticket & __ARCH_SPIN_NEXT_OVERFLOW)) {
+		__insn_fetchand4(&lock->lock, ~__ARCH_SPIN_NEXT_OVERFLOW);
+		my_ticket &= ~__ARCH_SPIN_NEXT_OVERFLOW;
+	}
+
+	for (;;) {
+		u32 val = arch_spin_read_noalloc(lock);
+		u32 delta = my_ticket - arch_spin_current(val);
+		if (delta == 0)
+			return;
+		relax((128 / CYCLES_PER_RELAX_LOOP) * delta);
+	}
+}
+EXPORT_SYMBOL(arch_spin_lock_slow);
+
+/*
+ * Check the lock to see if it is plausible, and try to get it with cmpxchg().
+ */
+int arch_spin_trylock(arch_spinlock_t *lock)
+{
+	u32 val = arch_spin_read_noalloc(lock);
+	if (unlikely(arch_spin_current(val) != arch_spin_next(val)))
+		return 0;
+	return cmpxchg(&lock->lock, val, (val + 1) & ~__ARCH_SPIN_NEXT_OVERFLOW)
+		== val;
+}
+EXPORT_SYMBOL(arch_spin_trylock);
+
+void arch_spin_unlock_wait(arch_spinlock_t *lock)
+{
+	u32 iterations = 0;
+	while (arch_spin_is_locked(lock))
+		delay_backoff(iterations++);
+}
+EXPORT_SYMBOL(arch_spin_unlock_wait);
+
+/*
+ * If the read lock fails due to a writer, we retry periodically
+ * until the value is positive and we write our incremented reader count.
+ */
+void __read_lock_failed(arch_rwlock_t *rw)
+{
+	u32 val;
+	int iterations = 0;
+	do {
+		delay_backoff(iterations++);
+		val = __insn_fetchaddgez4(&rw->lock, 1);
+	} while (unlikely(arch_write_val_locked(val)));
+}
+EXPORT_SYMBOL(__read_lock_failed);
+
+/*
+ * If we failed because there were readers, clear the "writer" bit
+ * so we don't block additional readers.  Otherwise, there was another
+ * writer anyway, so our "fetchor" made no difference.  Then wait,
+ * issuing periodic fetchor instructions, till we get the lock.
+ */
+void __write_lock_failed(arch_rwlock_t *rw, u32 val)
+{
+	int iterations = 0;
+	do {
+		if (!arch_write_val_locked(val))
+			val = __insn_fetchand4(&rw->lock, ~__WRITE_LOCK_BIT);
+		delay_backoff(iterations++);
+		val = __insn_fetchor4(&rw->lock, __WRITE_LOCK_BIT);
+	} while (val != 0);
+}
+EXPORT_SYMBOL(__write_lock_failed);
diff --git a/kernel/arch/tile/lib/spinlock_common.h b/kernel/arch/tile/lib/spinlock_common.h
new file mode 100644
index 000000000..6ac37509f
--- /dev/null
+++ b/kernel/arch/tile/lib/spinlock_common.h
@@ -0,0 +1,64 @@
+/*
+ * Copyright 2010 Tilera Corporation. All Rights Reserved.
+ *
+ *   This program is free software; you can redistribute it and/or
+ *   modify it under the terms of the GNU General Public License
+ *   as published by the Free Software Foundation, version 2.
+ *
+ *   This program is distributed in the hope that it will be useful, but
+ *   WITHOUT ANY WARRANTY; without even the implied warranty of
+ *   MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
+ *   NON INFRINGEMENT.  See the GNU General Public License for
+ *   more details.
+ * This file is included into spinlock_32.c or _64.c.
+ */
+
+/*
+ * The mfspr in __spinlock_relax() is 5 or 6 cycles plus 2 for loop
+ * overhead.
+ */
+#ifdef __tilegx__
+#define CYCLES_PER_RELAX_LOOP 7
+#else
+#define CYCLES_PER_RELAX_LOOP 8
+#endif
+
+/*
+ * Idle the core for CYCLES_PER_RELAX_LOOP * iterations cycles.
+ */
+static inline void
+relax(int iterations)
+{
+	for (/*above*/; iterations > 0; iterations--)
+		__insn_mfspr(SPR_PASS);
+	barrier();
+}
+
+/* Perform bounded exponential backoff.*/
+static void delay_backoff(int iterations)
+{
+	u32 exponent, loops;
+
+	/*
+	 * 2^exponent is how many times we go around the loop,
+	 * which takes 8 cycles.  We want to start with a 16- to 31-cycle
+	 * loop, so we need to go around minimum 2 = 2^1 times, so we
+	 * bias the original value up by 1.
+	 */
+	exponent = iterations + 1;
+
+	/*
+	 * Don't allow exponent to exceed 7, so we have 128 loops,
+	 * or 1,024 (to 2,047) cycles, as our maximum.
+	 */
+	if (exponent > 8)
+		exponent = 8;
+
+	loops = 1 << exponent;
+
+	/* Add a randomness factor so two cpus never get in lock step. */
+	loops += __insn_crc32_32(stack_pointer, get_cycles_low()) &
+		(loops - 1);
+
+	relax(loops);
+}
diff --git a/kernel/arch/tile/lib/strchr_32.c b/kernel/arch/tile/lib/strchr_32.c
new file mode 100644
index 000000000..841fe6963
--- /dev/null
+++ b/kernel/arch/tile/lib/strchr_32.c
@@ -0,0 +1,64 @@
+/*
+ * Copyright 2010 Tilera Corporation. All Rights Reserved.
+ *
+ *   This program is free software; you can redistribute it and/or
+ *   modify it under the terms of the GNU General Public License
+ *   as published by the Free Software Foundation, version 2.
+ *
+ *   This program is distributed in the hope that it will be useful, but
+ *   WITHOUT ANY WARRANTY; without even the implied warranty of
+ *   MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
+ *   NON INFRINGEMENT.  See the GNU General Public License for
+ *   more details.
+ */
+
+#include <linux/types.h>
+#include <linux/string.h>
+#include <linux/module.h>
+
+char *strchr(const char *s, int c)
+{
+	int z, g;
+
+	/* Get an aligned pointer. */
+	const uintptr_t s_int = (uintptr_t) s;
+	const uint32_t *p = (const uint32_t *)(s_int & -4);
+
+	/* Create four copies of the byte for which we are looking. */
+	const uint32_t goal = 0x01010101 * (uint8_t) c;
+
+	/* Read the first aligned word, but force bytes before the string to
+	 * match neither zero nor goal (we make sure the high bit of each
+	 * byte is 1, and the low 7 bits are all the opposite of the goal
+	 * byte).
+	 *
+	 * Note that this shift count expression works because we know shift
+	 * counts are taken mod 32.
+	 */
+	const uint32_t before_mask = (1 << (s_int << 3)) - 1;
+	uint32_t v = (*p | before_mask) ^ (goal & __insn_shrib(before_mask, 1));
+
+	uint32_t zero_matches, goal_matches;
+	while (1) {
+		/* Look for a terminating '\0'. */
+		zero_matches = __insn_seqb(v, 0);
+
+		/* Look for the goal byte. */
+		goal_matches = __insn_seqb(v, goal);
+
+		if (__builtin_expect(zero_matches | goal_matches, 0))
+			break;
+
+		v = *++p;
+	}
+
+	z = __insn_ctz(zero_matches);
+	g = __insn_ctz(goal_matches);
+
+	/* If we found c before '\0' we got a match. Note that if c == '\0'
+	 * then g == z, and we correctly return the address of the '\0'
+	 * rather than NULL.
+	 */
+	return (g <= z) ? ((char *)p) + (g >> 3) : NULL;
+}
+EXPORT_SYMBOL(strchr);
diff --git a/kernel/arch/tile/lib/strchr_64.c b/kernel/arch/tile/lib/strchr_64.c
new file mode 100644
index 000000000..fe6e31c06
--- /dev/null
+++ b/kernel/arch/tile/lib/strchr_64.c
@@ -0,0 +1,62 @@
+/*
+ * Copyright 2011 Tilera Corporation. All Rights Reserved.
+ *
+ *   This program is free software; you can redistribute it and/or
+ *   modify it under the terms of the GNU General Public License
+ *   as published by the Free Software Foundation, version 2.
+ *
+ *   This program is distributed in the hope that it will be useful, but
+ *   WITHOUT ANY WARRANTY; without even the implied warranty of
+ *   MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
+ *   NON INFRINGEMENT.  See the GNU General Public License for
+ *   more details.
+ */
+
+#include <linux/types.h>
+#include <linux/string.h>
+#include <linux/module.h>
+#include "string-endian.h"
+
+char *strchr(const char *s, int c)
+{
+	int z, g;
+
+	/* Get an aligned pointer. */
+	const uintptr_t s_int = (uintptr_t) s;
+	const uint64_t *p = (const uint64_t *)(s_int & -8);
+
+	/* Create eight copies of the byte for which we are looking. */
+	const uint64_t goal = copy_byte(c);
+
+	/* Read the first aligned word, but force bytes before the string to
+	 * match neither zero nor goal (we make sure the high bit of each
+	 * byte is 1, and the low 7 bits are all the opposite of the goal
+	 * byte).
+	 */
+	const uint64_t before_mask = MASK(s_int);
+	uint64_t v = (*p | before_mask) ^ (goal & __insn_v1shrui(before_mask, 1));
+
+	uint64_t zero_matches, goal_matches;
+	while (1) {
+		/* Look for a terminating '\0'. */
+		zero_matches = __insn_v1cmpeqi(v, 0);
+
+		/* Look for the goal byte. */
+		goal_matches = __insn_v1cmpeq(v, goal);
+
+		if (__builtin_expect((zero_matches | goal_matches) != 0, 0))
+			break;
+
+		v = *++p;
+	}
+
+	z = CFZ(zero_matches);
+	g = CFZ(goal_matches);
+
+	/* If we found c before '\0' we got a match. Note that if c == '\0'
+	 * then g == z, and we correctly return the address of the '\0'
+	 * rather than NULL.
+	 */
+	return (g <= z) ? ((char *)p) + (g >> 3) : NULL;
+}
+EXPORT_SYMBOL(strchr);
diff --git a/kernel/arch/tile/lib/string-endian.h b/kernel/arch/tile/lib/string-endian.h
new file mode 100644
index 000000000..2e49cbfe9
--- /dev/null
+++ b/kernel/arch/tile/lib/string-endian.h
@@ -0,0 +1,44 @@
+/*
+ * Copyright 2013 Tilera Corporation. All Rights Reserved.
+ *
+ *   This program is free software; you can redistribute it and/or
+ *   modify it under the terms of the GNU General Public License
+ *   as published by the Free Software Foundation, version 2.
+ *
+ *   This program is distributed in the hope that it will be useful, but
+ *   WITHOUT ANY WARRANTY; without even the implied warranty of
+ *   MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
+ *   NON INFRINGEMENT.  See the GNU General Public License for
+ *   more details.
+ *
+ * Provide a mask based on the pointer alignment that
+ * sets up non-zero bytes before the beginning of the string.
+ * The MASK expression works because shift counts are taken mod 64.
+ * Also, specify how to count "first" and "last" bits
+ * when the bits have been read as a word.
+ */
+
+#include <asm/byteorder.h>
+
+#ifdef __LITTLE_ENDIAN
+#define MASK(x) (__insn_shl(1ULL, (x << 3)) - 1)
+#define NULMASK(x) ((2ULL << x) - 1)
+#define CFZ(x) __insn_ctz(x)
+#define REVCZ(x) __insn_clz(x)
+#else
+#define MASK(x) (__insn_shl(-2LL, ((-x << 3) - 1)))
+#define NULMASK(x) (-2LL << (63 - x))
+#define CFZ(x) __insn_clz(x)
+#define REVCZ(x) __insn_ctz(x)
+#endif
+
+/*
+ * Create eight copies of the byte in a uint64_t.  Byte Shuffle uses
+ * the bytes of srcB as the index into the dest vector to select a
+ * byte.  With all indices of zero, the first byte is copied into all
+ * the other bytes.
+ */
+static inline uint64_t copy_byte(uint8_t byte)
+{
+	return __insn_shufflebytes(byte, 0, 0);
+}
diff --git a/kernel/arch/tile/lib/strlen_32.c b/kernel/arch/tile/lib/strlen_32.c
new file mode 100644
index 000000000..f26f88e11
--- /dev/null
+++ b/kernel/arch/tile/lib/strlen_32.c
@@ -0,0 +1,36 @@
+/*
+ * Copyright 2010 Tilera Corporation. All Rights Reserved.
+ *
+ *   This program is free software; you can redistribute it and/or
+ *   modify it under the terms of the GNU General Public License
+ *   as published by the Free Software Foundation, version 2.
+ *
+ *   This program is distributed in the hope that it will be useful, but
+ *   WITHOUT ANY WARRANTY; without even the implied warranty of
+ *   MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
+ *   NON INFRINGEMENT.  See the GNU General Public License for
+ *   more details.
+ */
+
+#include <linux/types.h>
+#include <linux/string.h>
+#include <linux/module.h>
+
+size_t strlen(const char *s)
+{
+	/* Get an aligned pointer. */
+	const uintptr_t s_int = (uintptr_t) s;
+	const uint32_t *p = (const uint32_t *)(s_int & -4);
+
+	/* Read the first word, but force bytes before the string to be nonzero.
+	 * This expression works because we know shift counts are taken mod 32.
+	 */
+	uint32_t v = *p | ((1 << (s_int << 3)) - 1);
+
+	uint32_t bits;
+	while ((bits = __insn_seqb(v, 0)) == 0)
+		v = *++p;
+
+	return ((const char *)p) + (__insn_ctz(bits) >> 3) - s;
+}
+EXPORT_SYMBOL(strlen);
diff --git a/kernel/arch/tile/lib/strlen_64.c b/kernel/arch/tile/lib/strlen_64.c
new file mode 100644
index 000000000..9583fc336
--- /dev/null
+++ b/kernel/arch/tile/lib/strlen_64.c
@@ -0,0 +1,35 @@
+/*
+ * Copyright 2011 Tilera Corporation. All Rights Reserved.
+ *
+ *   This program is free software; you can redistribute it and/or
+ *   modify it under the terms of the GNU General Public License
+ *   as published by the Free Software Foundation, version 2.
+ *
+ *   This program is distributed in the hope that it will be useful, but
+ *   WITHOUT ANY WARRANTY; without even the implied warranty of
+ *   MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
+ *   NON INFRINGEMENT.  See the GNU General Public License for
+ *   more details.
+ */
+
+#include <linux/types.h>
+#include <linux/string.h>
+#include <linux/module.h>
+#include "string-endian.h"
+
+size_t strlen(const char *s)
+{
+	/* Get an aligned pointer. */
+	const uintptr_t s_int = (uintptr_t) s;
+	const uint64_t *p = (const uint64_t *)(s_int & -8);
+
+	/* Read and MASK the first word. */
+	uint64_t v = *p | MASK(s_int);
+
+	uint64_t bits;
+	while ((bits = __insn_v1cmpeqi(v, 0)) == 0)
+		v = *++p;
+
+	return ((const char *)p) + (CFZ(bits) >> 3) - s;
+}
+EXPORT_SYMBOL(strlen);
diff --git a/kernel/arch/tile/lib/strnlen_32.c b/kernel/arch/tile/lib/strnlen_32.c
new file mode 100644
index 000000000..1434141d9
--- /dev/null
+++ b/kernel/arch/tile/lib/strnlen_32.c
@@ -0,0 +1,47 @@
+/*
+ * Copyright 2013 Tilera Corporation. All Rights Reserved.
+ *
+ *   This program is free software; you can redistribute it and/or
+ *   modify it under the terms of the GNU General Public License
+ *   as published by the Free Software Foundation, version 2.
+ *
+ *   This program is distributed in the hope that it will be useful, but
+ *   WITHOUT ANY WARRANTY; without even the implied warranty of
+ *   MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
+ *   NON INFRINGEMENT.  See the GNU General Public License for
+ *   more details.
+ */
+
+#include <linux/types.h>
+#include <linux/string.h>
+#include <linux/module.h>
+
+size_t strnlen(const char *s, size_t count)
+{
+	/* Get an aligned pointer. */
+	const uintptr_t s_int = (uintptr_t) s;
+	const uint32_t *p = (const uint32_t *)(s_int & -4);
+	size_t bytes_read = sizeof(*p) - (s_int & (sizeof(*p) - 1));
+	size_t len;
+	uint32_t v, bits;
+
+	/* Avoid page fault risk by not reading any bytes when count is 0. */
+	if (count == 0)
+		return 0;
+
+	/* Read first word, but force bytes before the string to be nonzero. */
+	v = *p | ((1 << ((s_int << 3) & 31)) - 1);
+
+	while ((bits = __insn_seqb(v, 0)) == 0) {
+		if (bytes_read >= count) {
+			/* Read COUNT bytes and didn't find the terminator. */
+			return count;
+		}
+		v = *++p;
+		bytes_read += sizeof(v);
+	}
+
+	len = ((const char *) p) + (__insn_ctz(bits) >> 3) - s;
+	return (len < count ? len : count);
+}
+EXPORT_SYMBOL(strnlen);
diff --git a/kernel/arch/tile/lib/strnlen_64.c b/kernel/arch/tile/lib/strnlen_64.c
new file mode 100644
index 000000000..2e8de6a51
--- /dev/null
+++ b/kernel/arch/tile/lib/strnlen_64.c
@@ -0,0 +1,48 @@
+/*
+ * Copyright 2013 Tilera Corporation. All Rights Reserved.
+ *
+ *   This program is free software; you can redistribute it and/or
+ *   modify it under the terms of the GNU General Public License
+ *   as published by the Free Software Foundation, version 2.
+ *
+ *   This program is distributed in the hope that it will be useful, but
+ *   WITHOUT ANY WARRANTY; without even the implied warranty of
+ *   MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
+ *   NON INFRINGEMENT.  See the GNU General Public License for
+ *   more details.
+ */
+
+#include <linux/types.h>
+#include <linux/string.h>
+#include <linux/module.h>
+#include "string-endian.h"
+
+size_t strnlen(const char *s, size_t count)
+{
+	/* Get an aligned pointer. */
+	const uintptr_t s_int = (uintptr_t) s;
+	const uint64_t *p = (const uint64_t *)(s_int & -8);
+	size_t bytes_read = sizeof(*p) - (s_int & (sizeof(*p) - 1));
+	size_t len;
+	uint64_t v, bits;
+
+	/* Avoid page fault risk by not reading any bytes when count is 0. */
+	if (count == 0)
+		return 0;
+
+	/* Read and MASK the first word. */
+	v = *p | MASK(s_int);
+
+	while ((bits = __insn_v1cmpeqi(v, 0)) == 0) {
+		if (bytes_read >= count) {
+			/* Read COUNT bytes and didn't find the terminator. */
+			return count;
+		}
+		v = *++p;
+		bytes_read += sizeof(v);
+	}
+
+	len = ((const char *) p) + (CFZ(bits) >> 3) - s;
+	return (len < count ? len : count);
+}
+EXPORT_SYMBOL(strnlen);
diff --git a/kernel/arch/tile/lib/uaccess.c b/kernel/arch/tile/lib/uaccess.c
new file mode 100644
index 000000000..030abe3ee
--- /dev/null
+++ b/kernel/arch/tile/lib/uaccess.c
@@ -0,0 +1,24 @@
+/*
+ * Copyright 2010 Tilera Corporation. All Rights Reserved.
+ *
+ *   This program is free software; you can redistribute it and/or
+ *   modify it under the terms of the GNU General Public License
+ *   as published by the Free Software Foundation, version 2.
+ *
+ *   This program is distributed in the hope that it will be useful, but
+ *   WITHOUT ANY WARRANTY; without even the implied warranty of
+ *   MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
+ *   NON INFRINGEMENT.  See the GNU General Public License for
+ *   more details.
+ */
+
+#include <linux/uaccess.h>
+#include <linux/module.h>
+
+int __range_ok(unsigned long addr, unsigned long size)
+{
+	unsigned long limit = current_thread_info()->addr_limit.seg;
+	return !((addr < limit && size <= limit - addr) ||
+		 is_arch_mappable_range(addr, size));
+}
+EXPORT_SYMBOL(__range_ok);
diff --git a/kernel/arch/tile/lib/usercopy_32.S b/kernel/arch/tile/lib/usercopy_32.S
new file mode 100644
index 000000000..1bc162224
--- /dev/null
+++ b/kernel/arch/tile/lib/usercopy_32.S
@@ -0,0 +1,135 @@
+/*
+ * Copyright 2010 Tilera Corporation. All Rights Reserved.
+ *
+ *   This program is free software; you can redistribute it and/or
+ *   modify it under the terms of the GNU General Public License
+ *   as published by the Free Software Foundation, version 2.
+ *
+ *   This program is distributed in the hope that it will be useful, but
+ *   WITHOUT ANY WARRANTY; without even the implied warranty of
+ *   MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
+ *   NON INFRINGEMENT.  See the GNU General Public License for
+ *   more details.
+ */
+
+#include <linux/linkage.h>
+#include <asm/errno.h>
+#include <asm/cache.h>
+#include <arch/chip.h>
+
+/* Access user memory, but use MMU to avoid propagating kernel exceptions. */
+
+/*
+ * strnlen_user_asm takes the pointer in r0, and the length bound in r1.
+ * It returns the length, including the terminating NUL, or zero on exception.
+ * If length is greater than the bound, returns one plus the bound.
+ */
+STD_ENTRY(strnlen_user_asm)
+	{ bz r1, 2f; addi r3, r0, -1 }  /* bias down to include NUL */
+1:      { lb_u r4, r0; addi r1, r1, -1 }
+	bz r4, 2f
+	{ bnzt r1, 1b; addi r0, r0, 1 }
+2:      { sub r0, r0, r3; jrp lr }
+	STD_ENDPROC(strnlen_user_asm)
+	.pushsection .fixup,"ax"
+strnlen_user_fault:
+	{ move r0, zero; jrp lr }
+	ENDPROC(strnlen_user_fault)
+	.section __ex_table,"a"
+	.align 4
+	.word 1b, strnlen_user_fault
+	.popsection
+
+/*
+ * strncpy_from_user_asm takes the kernel target pointer in r0,
+ * the userspace source pointer in r1, and the length bound (including
+ * the trailing NUL) in r2.  On success, it returns the string length
+ * (not including the trailing NUL), or -EFAULT on failure.
+ */
+STD_ENTRY(strncpy_from_user_asm)
+	{ bz r2, 2f; move r3, r0 }
+1:	{ lb_u r4, r1; addi r1, r1, 1; addi r2, r2, -1 }
+	{ sb r0, r4; addi r0, r0, 1 }
+	bz r4, 2f
+	bnzt r2, 1b
+	{ sub r0, r0, r3; jrp lr }
+2:	addi r0, r0, -1   /* don't count the trailing NUL */
+	{ sub r0, r0, r3; jrp lr }
+	STD_ENDPROC(strncpy_from_user_asm)
+	.pushsection .fixup,"ax"
+strncpy_from_user_fault:
+	{ movei r0, -EFAULT; jrp lr }
+	ENDPROC(strncpy_from_user_fault)
+	.section __ex_table,"a"
+	.align 4
+	.word 1b, strncpy_from_user_fault
+	.popsection
+
+/*
+ * clear_user_asm takes the user target address in r0 and the
+ * number of bytes to zero in r1.
+ * It returns the number of uncopiable bytes (hopefully zero) in r0.
+ * Note that we don't use a separate .fixup section here since we fall
+ * through into the "fixup" code as the last straight-line bundle anyway.
+ */
+STD_ENTRY(clear_user_asm)
+	{ bz r1, 2f; or r2, r0, r1 }
+	andi r2, r2, 3
+	bzt r2, .Lclear_aligned_user_asm
+1:      { sb r0, zero; addi r0, r0, 1; addi r1, r1, -1 }
+	bnzt r1, 1b
+2:      { move r0, r1; jrp lr }
+	.pushsection __ex_table,"a"
+	.align 4
+	.word 1b, 2b
+	.popsection
+
+.Lclear_aligned_user_asm:
+1:      { sw r0, zero; addi r0, r0, 4; addi r1, r1, -4 }
+	bnzt r1, 1b
+2:      { move r0, r1; jrp lr }
+	STD_ENDPROC(clear_user_asm)
+	.pushsection __ex_table,"a"
+	.align 4
+	.word 1b, 2b
+	.popsection
+
+/*
+ * flush_user_asm takes the user target address in r0 and the
+ * number of bytes to flush in r1.
+ * It returns the number of unflushable bytes (hopefully zero) in r0.
+ */
+STD_ENTRY(flush_user_asm)
+	bz r1, 2f
+	{ movei r2, L2_CACHE_BYTES; add r1, r0, r1 }
+	{ sub r2, zero, r2; addi r1, r1, L2_CACHE_BYTES-1 }
+	{ and r0, r0, r2; and r1, r1, r2 }
+	{ sub r1, r1, r0 }
+1:      { flush r0; addi r1, r1, -CHIP_FLUSH_STRIDE() }
+	{ addi r0, r0, CHIP_FLUSH_STRIDE(); bnzt r1, 1b }
+2:      { move r0, r1; jrp lr }
+	STD_ENDPROC(flush_user_asm)
+	.pushsection __ex_table,"a"
+	.align 4
+	.word 1b, 2b
+	.popsection
+
+/*
+ * finv_user_asm takes the user target address in r0 and the
+ * number of bytes to flush-invalidate in r1.
+ * It returns the number of not finv'able bytes (hopefully zero) in r0.
+ */
+STD_ENTRY(finv_user_asm)
+	bz r1, 2f
+	{ movei r2, L2_CACHE_BYTES; add r1, r0, r1 }
+	{ sub r2, zero, r2; addi r1, r1, L2_CACHE_BYTES-1 }
+	{ and r0, r0, r2; and r1, r1, r2 }
+	{ sub r1, r1, r0 }
+1:      { finv r0; addi r1, r1, -CHIP_FINV_STRIDE() }
+	{ addi r0, r0, CHIP_FINV_STRIDE(); bnzt r1, 1b }
+2:      { move r0, r1; jrp lr }
+	STD_ENDPROC(finv_user_asm)
+	.pushsection __ex_table,"a"
+	.align 4
+	.word 1b, 2b
+	.popsection
diff --git a/kernel/arch/tile/lib/usercopy_64.S b/kernel/arch/tile/lib/usercopy_64.S
new file mode 100644
index 000000000..b3b31a330
--- /dev/null
+++ b/kernel/arch/tile/lib/usercopy_64.S
@@ -0,0 +1,135 @@
+/*
+ * Copyright 2011 Tilera Corporation. All Rights Reserved.
+ *
+ *   This program is free software; you can redistribute it and/or
+ *   modify it under the terms of the GNU General Public License
+ *   as published by the Free Software Foundation, version 2.
+ *
+ *   This program is distributed in the hope that it will be useful, but
+ *   WITHOUT ANY WARRANTY; without even the implied warranty of
+ *   MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
+ *   NON INFRINGEMENT.  See the GNU General Public License for
+ *   more details.
+ */
+
+#include <linux/linkage.h>
+#include <asm/errno.h>
+#include <asm/cache.h>
+#include <arch/chip.h>
+
+/* Access user memory, but use MMU to avoid propagating kernel exceptions. */
+
+/*
+ * strnlen_user_asm takes the pointer in r0, and the length bound in r1.
+ * It returns the length, including the terminating NUL, or zero on exception.
+ * If length is greater than the bound, returns one plus the bound.
+ */
+STD_ENTRY(strnlen_user_asm)
+	{ beqz r1, 2f; addi r3, r0, -1 }  /* bias down to include NUL */
+1:      { ld1u r4, r0; addi r1, r1, -1 }
+	beqz r4, 2f
+	{ bnezt r1, 1b; addi r0, r0, 1 }
+2:      { sub r0, r0, r3; jrp lr }
+	STD_ENDPROC(strnlen_user_asm)
+	.pushsection .fixup,"ax"
+strnlen_user_fault:
+	{ move r0, zero; jrp lr }
+	ENDPROC(strnlen_user_fault)
+	.section __ex_table,"a"
+	.align 8
+	.quad 1b, strnlen_user_fault
+	.popsection
+
+/*
+ * strncpy_from_user_asm takes the kernel target pointer in r0,
+ * the userspace source pointer in r1, and the length bound (including
+ * the trailing NUL) in r2.  On success, it returns the string length
+ * (not including the trailing NUL), or -EFAULT on failure.
+ */
+STD_ENTRY(strncpy_from_user_asm)
+	{ beqz r2, 2f; move r3, r0 }
+1:	{ ld1u r4, r1; addi r1, r1, 1; addi r2, r2, -1 }
+	{ st1 r0, r4; addi r0, r0, 1 }
+	beqz r4, 2f
+	bnezt r2, 1b
+	{ sub r0, r0, r3; jrp lr }
+2:	addi r0, r0, -1   /* don't count the trailing NUL */
+	{ sub r0, r0, r3; jrp lr }
+	STD_ENDPROC(strncpy_from_user_asm)
+	.pushsection .fixup,"ax"
+strncpy_from_user_fault:
+	{ movei r0, -EFAULT; jrp lr }
+	ENDPROC(strncpy_from_user_fault)
+	.section __ex_table,"a"
+	.align 8
+	.quad 1b, strncpy_from_user_fault
+	.popsection
+
+/*
+ * clear_user_asm takes the user target address in r0 and the
+ * number of bytes to zero in r1.
+ * It returns the number of uncopiable bytes (hopefully zero) in r0.
+ * Note that we don't use a separate .fixup section here since we fall
+ * through into the "fixup" code as the last straight-line bundle anyway.
+ */
+STD_ENTRY(clear_user_asm)
+	{ beqz r1, 2f; or r2, r0, r1 }
+	andi r2, r2, 7
+	beqzt r2, .Lclear_aligned_user_asm
+1:      { st1 r0, zero; addi r0, r0, 1; addi r1, r1, -1 }
+	bnezt r1, 1b
+2:      { move r0, r1; jrp lr }
+	.pushsection __ex_table,"a"
+	.align 8
+	.quad 1b, 2b
+	.popsection
+
+.Lclear_aligned_user_asm:
+1:      { st r0, zero; addi r0, r0, 8; addi r1, r1, -8 }
+	bnezt r1, 1b
+2:      { move r0, r1; jrp lr }
+	STD_ENDPROC(clear_user_asm)
+	.pushsection __ex_table,"a"
+	.align 8
+	.quad 1b, 2b
+	.popsection
+
+/*
+ * flush_user_asm takes the user target address in r0 and the
+ * number of bytes to flush in r1.
+ * It returns the number of unflushable bytes (hopefully zero) in r0.
+ */
+STD_ENTRY(flush_user_asm)
+	beqz r1, 2f
+	{ movei r2, L2_CACHE_BYTES; add r1, r0, r1 }
+	{ sub r2, zero, r2; addi r1, r1, L2_CACHE_BYTES-1 }
+	{ and r0, r0, r2; and r1, r1, r2 }
+	{ sub r1, r1, r0 }
+1:      { flush r0; addi r1, r1, -CHIP_FLUSH_STRIDE() }
+	{ addi r0, r0, CHIP_FLUSH_STRIDE(); bnezt r1, 1b }
+2:      { move r0, r1; jrp lr }
+	STD_ENDPROC(flush_user_asm)
+	.pushsection __ex_table,"a"
+	.align 8
+	.quad 1b, 2b
+	.popsection
+
+/*
+ * finv_user_asm takes the user target address in r0 and the
+ * number of bytes to flush-invalidate in r1.
+ * It returns the number of not finv'able bytes (hopefully zero) in r0.
+ */
+STD_ENTRY(finv_user_asm)
+	beqz r1, 2f
+	{ movei r2, L2_CACHE_BYTES; add r1, r0, r1 }
+	{ sub r2, zero, r2; addi r1, r1, L2_CACHE_BYTES-1 }
+	{ and r0, r0, r2; and r1, r1, r2 }
+	{ sub r1, r1, r0 }
+1:      { finv r0; addi r1, r1, -CHIP_FINV_STRIDE() }
+	{ addi r0, r0, CHIP_FINV_STRIDE(); bnezt r1, 1b }
+2:      { move r0, r1; jrp lr }
+	STD_ENDPROC(finv_user_asm)
+	.pushsection __ex_table,"a"
+	.align 8
+	.quad 1b, 2b
+	.popsection
diff --git a/kernel/arch/tile/mm/Makefile b/kernel/arch/tile/mm/Makefile
new file mode 100644
index 000000000..e252aeddc
--- /dev/null
+++ b/kernel/arch/tile/mm/Makefile
@@ -0,0 +1,9 @@
+#
+# Makefile for the linux tile-specific parts of the memory manager.
+#
+
+obj-y	:= init.o pgtable.o fault.o extable.o elf.o \
+	   mmap.o homecache.o migrate_$(BITS).o
+
+obj-$(CONFIG_HUGETLB_PAGE) += hugetlbpage.o
+obj-$(CONFIG_HIGHMEM) += highmem.o
diff --git a/kernel/arch/tile/mm/elf.c b/kernel/arch/tile/mm/elf.c
new file mode 100644
index 000000000..f7ddae372
--- /dev/null
+++ b/kernel/arch/tile/mm/elf.c
@@ -0,0 +1,165 @@
+/*
+ * Copyright 2010 Tilera Corporation. All Rights Reserved.
+ *
+ *   This program is free software; you can redistribute it and/or
+ *   modify it under the terms of the GNU General Public License
+ *   as published by the Free Software Foundation, version 2.
+ *
+ *   This program is distributed in the hope that it will be useful, but
+ *   WITHOUT ANY WARRANTY; without even the implied warranty of
+ *   MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
+ *   NON INFRINGEMENT.  See the GNU General Public License for
+ *   more details.
+ */
+
+#include <linux/mm.h>
+#include <linux/pagemap.h>
+#include <linux/binfmts.h>
+#include <linux/compat.h>
+#include <linux/mman.h>
+#include <linux/file.h>
+#include <linux/elf.h>
+#include <asm/pgtable.h>
+#include <asm/pgalloc.h>
+#include <asm/sections.h>
+#include <asm/vdso.h>
+#include <arch/sim.h>
+
+/* Notify a running simulator, if any, that an exec just occurred. */
+static void sim_notify_exec(const char *binary_name)
+{
+	unsigned char c;
+	do {
+		c = *binary_name++;
+		__insn_mtspr(SPR_SIM_CONTROL,
+			     (SIM_CONTROL_OS_EXEC
+			      | (c << _SIM_CONTROL_OPERATOR_BITS)));
+
+	} while (c);
+}
+
+static int notify_exec(struct mm_struct *mm)
+{
+	int ret = 0;
+	char *buf, *path;
+	struct vm_area_struct *vma;
+	struct file *exe_file;
+
+	if (!sim_is_simulator())
+		return 1;
+
+	buf = (char *) __get_free_page(GFP_KERNEL);
+	if (buf == NULL)
+		return 0;
+
+	exe_file = get_mm_exe_file(mm);
+	if (exe_file == NULL)
+		goto done_free;
+
+	path = d_path(&exe_file->f_path, buf, PAGE_SIZE);
+	if (IS_ERR(path))
+		goto done_put;
+
+	down_read(&mm->mmap_sem);
+	for (vma = current->mm->mmap; ; vma = vma->vm_next) {
+		if (vma == NULL) {
+			up_read(&mm->mmap_sem);
+			goto done_put;
+		}
+		if (vma->vm_file == exe_file)
+			break;
+	}
+
+	/*
+	 * Notify simulator of an ET_DYN object so we know the load address.
+	 * The somewhat cryptic overuse of SIM_CONTROL_DLOPEN allows us
+	 * to be backward-compatible with older simulator releases.
+	 */
+	if (vma->vm_start == (ELF_ET_DYN_BASE & PAGE_MASK)) {
+		char buf[64];
+		int i;
+
+		snprintf(buf, sizeof(buf), "0x%lx:@", vma->vm_start);
+		for (i = 0; ; ++i) {
+			char c = buf[i];
+			__insn_mtspr(SPR_SIM_CONTROL,
+				     (SIM_CONTROL_DLOPEN
+				      | (c << _SIM_CONTROL_OPERATOR_BITS)));
+			if (c == '\0') {
+				ret = 1; /* success */
+				break;
+			}
+		}
+	}
+	up_read(&mm->mmap_sem);
+
+	sim_notify_exec(path);
+done_put:
+	fput(exe_file);
+done_free:
+	free_page((unsigned long)buf);
+	return ret;
+}
+
+/* Notify a running simulator, if any, that we loaded an interpreter. */
+static void sim_notify_interp(unsigned long load_addr)
+{
+	size_t i;
+	for (i = 0; i < sizeof(load_addr); i++) {
+		unsigned char c = load_addr >> (i * 8);
+		__insn_mtspr(SPR_SIM_CONTROL,
+			     (SIM_CONTROL_OS_INTERP
+			      | (c << _SIM_CONTROL_OPERATOR_BITS)));
+	}
+}
+
+
+int arch_setup_additional_pages(struct linux_binprm *bprm,
+				int executable_stack)
+{
+	struct mm_struct *mm = current->mm;
+	int retval = 0;
+
+	/*
+	 * Notify the simulator that an exec just occurred.
+	 * If we can't find the filename of the mapping, just use
+	 * whatever was passed as the linux_binprm filename.
+	 */
+	if (!notify_exec(mm))
+		sim_notify_exec(bprm->filename);
+
+	down_write(&mm->mmap_sem);
+
+	retval = setup_vdso_pages();
+
+#ifndef __tilegx__
+	/*
+	 * Set up a user-interrupt mapping here; the user can't
+	 * create one themselves since it is above TASK_SIZE.
+	 * We make it unwritable by default, so the model for adding
+	 * interrupt vectors always involves an mprotect.
+	 */
+	if (!retval) {
+		unsigned long addr = MEM_USER_INTRPT;
+		addr = mmap_region(NULL, addr, INTRPT_SIZE,
+				   VM_READ|VM_EXEC|
+				   VM_MAYREAD|VM_MAYWRITE|VM_MAYEXEC, 0);
+		if (addr > (unsigned long) -PAGE_SIZE)
+			retval = (int) addr;
+	}
+#endif
+
+	up_write(&mm->mmap_sem);
+
+	return retval;
+}
+
+
+void elf_plat_init(struct pt_regs *regs, unsigned long load_addr)
+{
+	/* Zero all registers. */
+	memset(regs, 0, sizeof(*regs));
+
+	/* Report the interpreter's load address. */
+	sim_notify_interp(load_addr);
+}
diff --git a/kernel/arch/tile/mm/extable.c b/kernel/arch/tile/mm/extable.c
new file mode 100644
index 000000000..4fb0acb9d
--- /dev/null
+++ b/kernel/arch/tile/mm/extable.c
@@ -0,0 +1,30 @@
+/*
+ * Copyright 2010 Tilera Corporation. All Rights Reserved.
+ *
+ *   This program is free software; you can redistribute it and/or
+ *   modify it under the terms of the GNU General Public License
+ *   as published by the Free Software Foundation, version 2.
+ *
+ *   This program is distributed in the hope that it will be useful, but
+ *   WITHOUT ANY WARRANTY; without even the implied warranty of
+ *   MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
+ *   NON INFRINGEMENT.  See the GNU General Public License for
+ *   more details.
+ */
+
+#include <linux/module.h>
+#include <linux/spinlock.h>
+#include <linux/uaccess.h>
+
+int fixup_exception(struct pt_regs *regs)
+{
+	const struct exception_table_entry *fixup;
+
+	fixup = search_exception_tables(regs->pc);
+	if (fixup) {
+		regs->pc = fixup->fixup;
+		return 1;
+	}
+
+	return 0;
+}
diff --git a/kernel/arch/tile/mm/fault.c b/kernel/arch/tile/mm/fault.c
new file mode 100644
index 000000000..3f4f58d34
--- /dev/null
+++ b/kernel/arch/tile/mm/fault.c
@@ -0,0 +1,921 @@
+/*
+ * Copyright 2010 Tilera Corporation. All Rights Reserved.
+ *
+ *   This program is free software; you can redistribute it and/or
+ *   modify it under the terms of the GNU General Public License
+ *   as published by the Free Software Foundation, version 2.
+ *
+ *   This program is distributed in the hope that it will be useful, but
+ *   WITHOUT ANY WARRANTY; without even the implied warranty of
+ *   MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
+ *   NON INFRINGEMENT.  See the GNU General Public License for
+ *   more details.
+ *
+ * From i386 code copyright (C) 1995  Linus Torvalds
+ */
+
+#include <linux/signal.h>
+#include <linux/sched.h>
+#include <linux/kernel.h>
+#include <linux/errno.h>
+#include <linux/string.h>
+#include <linux/types.h>
+#include <linux/ptrace.h>
+#include <linux/mman.h>
+#include <linux/mm.h>
+#include <linux/smp.h>
+#include <linux/interrupt.h>
+#include <linux/init.h>
+#include <linux/tty.h>
+#include <linux/vt_kern.h>		/* For unblank_screen() */
+#include <linux/highmem.h>
+#include <linux/module.h>
+#include <linux/kprobes.h>
+#include <linux/hugetlb.h>
+#include <linux/syscalls.h>
+#include <linux/uaccess.h>
+#include <linux/kdebug.h>
+#include <linux/context_tracking.h>
+
+#include <asm/pgalloc.h>
+#include <asm/sections.h>
+#include <asm/traps.h>
+#include <asm/syscalls.h>
+
+#include <arch/interrupts.h>
+
+static noinline void force_sig_info_fault(const char *type, int si_signo,
+					  int si_code, unsigned long address,
+					  int fault_num,
+					  struct task_struct *tsk,
+					  struct pt_regs *regs)
+{
+	siginfo_t info;
+
+	if (unlikely(tsk->pid < 2)) {
+		panic("Signal %d (code %d) at %#lx sent to %s!",
+		      si_signo, si_code & 0xffff, address,
+		      is_idle_task(tsk) ? "the idle task" : "init");
+	}
+
+	info.si_signo = si_signo;
+	info.si_errno = 0;
+	info.si_code = si_code;
+	info.si_addr = (void __user *)address;
+	info.si_trapno = fault_num;
+	trace_unhandled_signal(type, regs, address, si_signo);
+	force_sig_info(si_signo, &info, tsk);
+}
+
+#ifndef __tilegx__
+/*
+ * Synthesize the fault a PL0 process would get by doing a word-load of
+ * an unaligned address or a high kernel address.
+ */
+SYSCALL_DEFINE1(cmpxchg_badaddr, unsigned long, address)
+{
+	struct pt_regs *regs = current_pt_regs();
+
+	if (address >= PAGE_OFFSET)
+		force_sig_info_fault("atomic segfault", SIGSEGV, SEGV_MAPERR,
+				     address, INT_DTLB_MISS, current, regs);
+	else
+		force_sig_info_fault("atomic alignment fault", SIGBUS,
+				     BUS_ADRALN, address,
+				     INT_UNALIGN_DATA, current, regs);
+
+	/*
+	 * Adjust pc to point at the actual instruction, which is unusual
+	 * for syscalls normally, but is appropriate when we are claiming
+	 * that a syscall swint1 caused a page fault or bus error.
+	 */
+	regs->pc -= 8;
+
+	/*
+	 * Mark this as a caller-save interrupt, like a normal page fault,
+	 * so that when we go through the signal handler path we will
+	 * properly restore r0, r1, and r2 for the signal handler arguments.
+	 */
+	regs->flags |= PT_FLAGS_CALLER_SAVES;
+
+	return 0;
+}
+#endif
+
+static inline pmd_t *vmalloc_sync_one(pgd_t *pgd, unsigned long address)
+{
+	unsigned index = pgd_index(address);
+	pgd_t *pgd_k;
+	pud_t *pud, *pud_k;
+	pmd_t *pmd, *pmd_k;
+
+	pgd += index;
+	pgd_k = init_mm.pgd + index;
+
+	if (!pgd_present(*pgd_k))
+		return NULL;
+
+	pud = pud_offset(pgd, address);
+	pud_k = pud_offset(pgd_k, address);
+	if (!pud_present(*pud_k))
+		return NULL;
+
+	pmd = pmd_offset(pud, address);
+	pmd_k = pmd_offset(pud_k, address);
+	if (!pmd_present(*pmd_k))
+		return NULL;
+	if (!pmd_present(*pmd))
+		set_pmd(pmd, *pmd_k);
+	else
+		BUG_ON(pmd_ptfn(*pmd) != pmd_ptfn(*pmd_k));
+	return pmd_k;
+}
+
+/*
+ * Handle a fault on the vmalloc area.
+ */
+static inline int vmalloc_fault(pgd_t *pgd, unsigned long address)
+{
+	pmd_t *pmd_k;
+	pte_t *pte_k;
+
+	/* Make sure we are in vmalloc area */
+	if (!(address >= VMALLOC_START && address < VMALLOC_END))
+		return -1;
+
+	/*
+	 * Synchronize this task's top level page-table
+	 * with the 'reference' page table.
+	 */
+	pmd_k = vmalloc_sync_one(pgd, address);
+	if (!pmd_k)
+		return -1;
+	pte_k = pte_offset_kernel(pmd_k, address);
+	if (!pte_present(*pte_k))
+		return -1;
+	return 0;
+}
+
+/* Wait until this PTE has completed migration. */
+static void wait_for_migration(pte_t *pte)
+{
+	if (pte_migrating(*pte)) {
+		/*
+		 * Wait until the migrater fixes up this pte.
+		 * We scale the loop count by the clock rate so we'll wait for
+		 * a few seconds here.
+		 */
+		int retries = 0;
+		int bound = get_clock_rate();
+		while (pte_migrating(*pte)) {
+			barrier();
+			if (++retries > bound)
+				panic("Hit migrating PTE (%#llx) and page PFN %#lx still migrating",
+				      pte->val, pte_pfn(*pte));
+		}
+	}
+}
+
+/*
+ * It's not generally safe to use "current" to get the page table pointer,
+ * since we might be running an oprofile interrupt in the middle of a
+ * task switch.
+ */
+static pgd_t *get_current_pgd(void)
+{
+	HV_Context ctx = hv_inquire_context();
+	unsigned long pgd_pfn = ctx.page_table >> PAGE_SHIFT;
+	struct page *pgd_page = pfn_to_page(pgd_pfn);
+	BUG_ON(PageHighMem(pgd_page));
+	return (pgd_t *) __va(ctx.page_table);
+}
+
+/*
+ * We can receive a page fault from a migrating PTE at any time.
+ * Handle it by just waiting until the fault resolves.
+ *
+ * It's also possible to get a migrating kernel PTE that resolves
+ * itself during the downcall from hypervisor to Linux.  We just check
+ * here to see if the PTE seems valid, and if so we retry it.
+ *
+ * NOTE! We MUST NOT take any locks for this case.  We may be in an
+ * interrupt or a critical region, and must do as little as possible.
+ * Similarly, we can't use atomic ops here, since we may be handling a
+ * fault caused by an atomic op access.
+ *
+ * If we find a migrating PTE while we're in an NMI context, and we're
+ * at a PC that has a registered exception handler, we don't wait,
+ * since this thread may (e.g.) have been interrupted while migrating
+ * its own stack, which would then cause us to self-deadlock.
+ */
+static int handle_migrating_pte(pgd_t *pgd, int fault_num,
+				unsigned long address, unsigned long pc,
+				int is_kernel_mode, int write)
+{
+	pud_t *pud;
+	pmd_t *pmd;
+	pte_t *pte;
+	pte_t pteval;
+
+	if (pgd_addr_invalid(address))
+		return 0;
+
+	pgd += pgd_index(address);
+	pud = pud_offset(pgd, address);
+	if (!pud || !pud_present(*pud))
+		return 0;
+	pmd = pmd_offset(pud, address);
+	if (!pmd || !pmd_present(*pmd))
+		return 0;
+	pte = pmd_huge_page(*pmd) ? ((pte_t *)pmd) :
+		pte_offset_kernel(pmd, address);
+	pteval = *pte;
+	if (pte_migrating(pteval)) {
+		if (in_nmi() && search_exception_tables(pc))
+			return 0;
+		wait_for_migration(pte);
+		return 1;
+	}
+
+	if (!is_kernel_mode || !pte_present(pteval))
+		return 0;
+	if (fault_num == INT_ITLB_MISS) {
+		if (pte_exec(pteval))
+			return 1;
+	} else if (write) {
+		if (pte_write(pteval))
+			return 1;
+	} else {
+		if (pte_read(pteval))
+			return 1;
+	}
+
+	return 0;
+}
+
+/*
+ * This routine is responsible for faulting in user pages.
+ * It passes the work off to one of the appropriate routines.
+ * It returns true if the fault was successfully handled.
+ */
+static int handle_page_fault(struct pt_regs *regs,
+			     int fault_num,
+			     int is_page_fault,
+			     unsigned long address,
+			     int write)
+{
+	struct task_struct *tsk;
+	struct mm_struct *mm;
+	struct vm_area_struct *vma;
+	unsigned long stack_offset;
+	int fault;
+	int si_code;
+	int is_kernel_mode;
+	pgd_t *pgd;
+	unsigned int flags;
+
+	/* on TILE, protection faults are always writes */
+	if (!is_page_fault)
+		write = 1;
+
+	flags = FAULT_FLAG_ALLOW_RETRY | FAULT_FLAG_KILLABLE;
+
+	is_kernel_mode = !user_mode(regs);
+
+	tsk = validate_current();
+
+	/*
+	 * Check to see if we might be overwriting the stack, and bail
+	 * out if so.  The page fault code is a relatively likely
+	 * place to get trapped in an infinite regress, and once we
+	 * overwrite the whole stack, it becomes very hard to recover.
+	 */
+	stack_offset = stack_pointer & (THREAD_SIZE-1);
+	if (stack_offset < THREAD_SIZE / 8) {
+		pr_alert("Potential stack overrun: sp %#lx\n", stack_pointer);
+		show_regs(regs);
+		pr_alert("Killing current process %d/%s\n",
+			 tsk->pid, tsk->comm);
+		do_group_exit(SIGKILL);
+	}
+
+	/*
+	 * Early on, we need to check for migrating PTE entries;
+	 * see homecache.c.  If we find a migrating PTE, we wait until
+	 * the backing page claims to be done migrating, then we proceed.
+	 * For kernel PTEs, we rewrite the PTE and return and retry.
+	 * Otherwise, we treat the fault like a normal "no PTE" fault,
+	 * rather than trying to patch up the existing PTE.
+	 */
+	pgd = get_current_pgd();
+	if (handle_migrating_pte(pgd, fault_num, address, regs->pc,
+				 is_kernel_mode, write))
+		return 1;
+
+	si_code = SEGV_MAPERR;
+
+	/*
+	 * We fault-in kernel-space virtual memory on-demand. The
+	 * 'reference' page table is init_mm.pgd.
+	 *
+	 * NOTE! We MUST NOT take any locks for this case. We may
+	 * be in an interrupt or a critical region, and should
+	 * only copy the information from the master page table,
+	 * nothing more.
+	 *
+	 * This verifies that the fault happens in kernel space
+	 * and that the fault was not a protection fault.
+	 */
+	if (unlikely(address >= TASK_SIZE &&
+		     !is_arch_mappable_range(address, 0))) {
+		if (is_kernel_mode && is_page_fault &&
+		    vmalloc_fault(pgd, address) >= 0)
+			return 1;
+		/*
+		 * Don't take the mm semaphore here. If we fixup a prefetch
+		 * fault we could otherwise deadlock.
+		 */
+		mm = NULL;  /* happy compiler */
+		vma = NULL;
+		goto bad_area_nosemaphore;
+	}
+
+	/*
+	 * If we're trying to touch user-space addresses, we must
+	 * be either at PL0, or else with interrupts enabled in the
+	 * kernel, so either way we can re-enable interrupts here
+	 * unless we are doing atomic access to user space with
+	 * interrupts disabled.
+	 */
+	if (!(regs->flags & PT_FLAGS_DISABLE_IRQ))
+		local_irq_enable();
+
+	mm = tsk->mm;
+
+	/*
+	 * If we're in an interrupt, have no user context or are running in an
+	 * region with pagefaults disabled then we must not take the fault.
+	 */
+	if (pagefault_disabled() || !mm) {
+		vma = NULL;  /* happy compiler */
+		goto bad_area_nosemaphore;
+	}
+
+	if (!is_kernel_mode)
+		flags |= FAULT_FLAG_USER;
+
+	/*
+	 * When running in the kernel we expect faults to occur only to
+	 * addresses in user space.  All other faults represent errors in the
+	 * kernel and should generate an OOPS.  Unfortunately, in the case of an
+	 * erroneous fault occurring in a code path which already holds mmap_sem
+	 * we will deadlock attempting to validate the fault against the
+	 * address space.  Luckily the kernel only validly references user
+	 * space from well defined areas of code, which are listed in the
+	 * exceptions table.
+	 *
+	 * As the vast majority of faults will be valid we will only perform
+	 * the source reference check when there is a possibility of a deadlock.
+	 * Attempt to lock the address space, if we cannot we then validate the
+	 * source.  If this is invalid we can skip the address space check,
+	 * thus avoiding the deadlock.
+	 */
+	if (!down_read_trylock(&mm->mmap_sem)) {
+		if (is_kernel_mode &&
+		    !search_exception_tables(regs->pc)) {
+			vma = NULL;  /* happy compiler */
+			goto bad_area_nosemaphore;
+		}
+
+retry:
+		down_read(&mm->mmap_sem);
+	}
+
+	vma = find_vma(mm, address);
+	if (!vma)
+		goto bad_area;
+	if (vma->vm_start <= address)
+		goto good_area;
+	if (!(vma->vm_flags & VM_GROWSDOWN))
+		goto bad_area;
+	if (regs->sp < PAGE_OFFSET) {
+		/*
+		 * accessing the stack below sp is always a bug.
+		 */
+		if (address < regs->sp)
+			goto bad_area;
+	}
+	if (expand_stack(vma, address))
+		goto bad_area;
+
+/*
+ * Ok, we have a good vm_area for this memory access, so
+ * we can handle it..
+ */
+good_area:
+	si_code = SEGV_ACCERR;
+	if (fault_num == INT_ITLB_MISS) {
+		if (!(vma->vm_flags & VM_EXEC))
+			goto bad_area;
+	} else if (write) {
+#ifdef TEST_VERIFY_AREA
+		if (!is_page_fault && regs->cs == KERNEL_CS)
+			pr_err("WP fault at " REGFMT "\n", regs->eip);
+#endif
+		if (!(vma->vm_flags & VM_WRITE))
+			goto bad_area;
+		flags |= FAULT_FLAG_WRITE;
+	} else {
+		if (!is_page_fault || !(vma->vm_flags & VM_READ))
+			goto bad_area;
+	}
+
+	/*
+	 * If for any reason at all we couldn't handle the fault,
+	 * make sure we exit gracefully rather than endlessly redo
+	 * the fault.
+	 */
+	fault = handle_mm_fault(mm, vma, address, flags);
+
+	if ((fault & VM_FAULT_RETRY) && fatal_signal_pending(current))
+		return 0;
+
+	if (unlikely(fault & VM_FAULT_ERROR)) {
+		if (fault & VM_FAULT_OOM)
+			goto out_of_memory;
+		else if (fault & VM_FAULT_SIGSEGV)
+			goto bad_area;
+		else if (fault & VM_FAULT_SIGBUS)
+			goto do_sigbus;
+		BUG();
+	}
+	if (flags & FAULT_FLAG_ALLOW_RETRY) {
+		if (fault & VM_FAULT_MAJOR)
+			tsk->maj_flt++;
+		else
+			tsk->min_flt++;
+		if (fault & VM_FAULT_RETRY) {
+			flags &= ~FAULT_FLAG_ALLOW_RETRY;
+			flags |= FAULT_FLAG_TRIED;
+
+			 /*
+			  * No need to up_read(&mm->mmap_sem) as we would
+			  * have already released it in __lock_page_or_retry
+			  * in mm/filemap.c.
+			  */
+			goto retry;
+		}
+	}
+
+#if CHIP_HAS_TILE_DMA()
+	/* If this was a DMA TLB fault, restart the DMA engine. */
+	switch (fault_num) {
+	case INT_DMATLB_MISS:
+	case INT_DMATLB_MISS_DWNCL:
+	case INT_DMATLB_ACCESS:
+	case INT_DMATLB_ACCESS_DWNCL:
+		__insn_mtspr(SPR_DMA_CTR, SPR_DMA_CTR__REQUEST_MASK);
+		break;
+	}
+#endif
+
+	up_read(&mm->mmap_sem);
+	return 1;
+
+/*
+ * Something tried to access memory that isn't in our memory map..
+ * Fix it, but check if it's kernel or user first..
+ */
+bad_area:
+	up_read(&mm->mmap_sem);
+
+bad_area_nosemaphore:
+	/* User mode accesses just cause a SIGSEGV */
+	if (!is_kernel_mode) {
+		/*
+		 * It's possible to have interrupts off here.
+		 */
+		local_irq_enable();
+
+		force_sig_info_fault("segfault", SIGSEGV, si_code, address,
+				     fault_num, tsk, regs);
+		return 0;
+	}
+
+no_context:
+	/* Are we prepared to handle this kernel fault?  */
+	if (fixup_exception(regs))
+		return 0;
+
+/*
+ * Oops. The kernel tried to access some bad page. We'll have to
+ * terminate things with extreme prejudice.
+ */
+
+	bust_spinlocks(1);
+
+	/* FIXME: no lookup_address() yet */
+#ifdef SUPPORT_LOOKUP_ADDRESS
+	if (fault_num == INT_ITLB_MISS) {
+		pte_t *pte = lookup_address(address);
+
+		if (pte && pte_present(*pte) && !pte_exec_kernel(*pte))
+			pr_crit("kernel tried to execute non-executable page - exploit attempt? (uid: %d)\n",
+				current->uid);
+	}
+#endif
+	if (address < PAGE_SIZE)
+		pr_alert("Unable to handle kernel NULL pointer dereference\n");
+	else
+		pr_alert("Unable to handle kernel paging request\n");
+	pr_alert(" at virtual address " REGFMT ", pc " REGFMT "\n",
+		 address, regs->pc);
+
+	show_regs(regs);
+
+	if (unlikely(tsk->pid < 2)) {
+		panic("Kernel page fault running %s!",
+		      is_idle_task(tsk) ? "the idle task" : "init");
+	}
+
+	/*
+	 * More FIXME: we should probably copy the i386 here and
+	 * implement a generic die() routine.  Not today.
+	 */
+#ifdef SUPPORT_DIE
+	die("Oops", regs);
+#endif
+	bust_spinlocks(1);
+
+	do_group_exit(SIGKILL);
+
+/*
+ * We ran out of memory, or some other thing happened to us that made
+ * us unable to handle the page fault gracefully.
+ */
+out_of_memory:
+	up_read(&mm->mmap_sem);
+	if (is_kernel_mode)
+		goto no_context;
+	pagefault_out_of_memory();
+	return 0;
+
+do_sigbus:
+	up_read(&mm->mmap_sem);
+
+	/* Kernel mode? Handle exceptions or die */
+	if (is_kernel_mode)
+		goto no_context;
+
+	force_sig_info_fault("bus error", SIGBUS, BUS_ADRERR, address,
+			     fault_num, tsk, regs);
+	return 0;
+}
+
+#ifndef __tilegx__
+
+/* We must release ICS before panicking or we won't get anywhere. */
+#define ics_panic(fmt, ...)					\
+do {								\
+	__insn_mtspr(SPR_INTERRUPT_CRITICAL_SECTION, 0);	\
+	panic(fmt, ##__VA_ARGS__);				\
+} while (0)
+
+/*
+ * When we take an ITLB or DTLB fault or access violation in the
+ * supervisor while the critical section bit is set, the hypervisor is
+ * reluctant to write new values into the EX_CONTEXT_K_x registers,
+ * since that might indicate we have not yet squirreled the SPR
+ * contents away and can thus safely take a recursive interrupt.
+ * Accordingly, the hypervisor passes us the PC via SYSTEM_SAVE_K_2.
+ *
+ * Note that this routine is called before homecache_tlb_defer_enter(),
+ * which means that we can properly unlock any atomics that might
+ * be used there (good), but also means we must be very sensitive
+ * to not touch any data structures that might be located in memory
+ * that could migrate, as we could be entering the kernel on a dataplane
+ * cpu that has been deferring kernel TLB updates.  This means, for
+ * example, that we can't migrate init_mm or its pgd.
+ */
+struct intvec_state do_page_fault_ics(struct pt_regs *regs, int fault_num,
+				      unsigned long address,
+				      unsigned long info)
+{
+	unsigned long pc = info & ~1;
+	int write = info & 1;
+	pgd_t *pgd = get_current_pgd();
+
+	/* Retval is 1 at first since we will handle the fault fully. */
+	struct intvec_state state = {
+		do_page_fault, fault_num, address, write, 1
+	};
+
+	/* Validate that we are plausibly in the right routine. */
+	if ((pc & 0x7) != 0 || pc < PAGE_OFFSET ||
+	    (fault_num != INT_DTLB_MISS &&
+	     fault_num != INT_DTLB_ACCESS)) {
+		unsigned long old_pc = regs->pc;
+		regs->pc = pc;
+		ics_panic("Bad ICS page fault args: old PC %#lx, fault %d/%d at %#lx",
+			  old_pc, fault_num, write, address);
+	}
+
+	/* We might be faulting on a vmalloc page, so check that first. */
+	if (fault_num != INT_DTLB_ACCESS && vmalloc_fault(pgd, address) >= 0)
+		return state;
+
+	/*
+	 * If we faulted with ICS set in sys_cmpxchg, we are providing
+	 * a user syscall service that should generate a signal on
+	 * fault.  We didn't set up a kernel stack on initial entry to
+	 * sys_cmpxchg, but instead had one set up by the fault, which
+	 * (because sys_cmpxchg never releases ICS) came to us via the
+	 * SYSTEM_SAVE_K_2 mechanism, and thus EX_CONTEXT_K_[01] are
+	 * still referencing the original user code.  We release the
+	 * atomic lock and rewrite pt_regs so that it appears that we
+	 * came from user-space directly, and after we finish the
+	 * fault we'll go back to user space and re-issue the swint.
+	 * This way the backtrace information is correct if we need to
+	 * emit a stack dump at any point while handling this.
+	 *
+	 * Must match register use in sys_cmpxchg().
+	 */
+	if (pc >= (unsigned long) sys_cmpxchg &&
+	    pc < (unsigned long) __sys_cmpxchg_end) {
+#ifdef CONFIG_SMP
+		/* Don't unlock before we could have locked. */
+		if (pc >= (unsigned long)__sys_cmpxchg_grab_lock) {
+			int *lock_ptr = (int *)(regs->regs[ATOMIC_LOCK_REG]);
+			__atomic_fault_unlock(lock_ptr);
+		}
+#endif
+		regs->sp = regs->regs[27];
+	}
+
+	/*
+	 * We can also fault in the atomic assembly, in which
+	 * case we use the exception table to do the first-level fixup.
+	 * We may re-fixup again in the real fault handler if it
+	 * turns out the faulting address is just bad, and not,
+	 * for example, migrating.
+	 */
+	else if (pc >= (unsigned long) __start_atomic_asm_code &&
+		   pc < (unsigned long) __end_atomic_asm_code) {
+		const struct exception_table_entry *fixup;
+#ifdef CONFIG_SMP
+		/* Unlock the atomic lock. */
+		int *lock_ptr = (int *)(regs->regs[ATOMIC_LOCK_REG]);
+		__atomic_fault_unlock(lock_ptr);
+#endif
+		fixup = search_exception_tables(pc);
+		if (!fixup)
+			ics_panic("ICS atomic fault not in table: PC %#lx, fault %d",
+				  pc, fault_num);
+		regs->pc = fixup->fixup;
+		regs->ex1 = PL_ICS_EX1(KERNEL_PL, 0);
+	}
+
+	/*
+	 * Now that we have released the atomic lock (if necessary),
+	 * it's safe to spin if the PTE that caused the fault was migrating.
+	 */
+	if (fault_num == INT_DTLB_ACCESS)
+		write = 1;
+	if (handle_migrating_pte(pgd, fault_num, address, pc, 1, write))
+		return state;
+
+	/* Return zero so that we continue on with normal fault handling. */
+	state.retval = 0;
+	return state;
+}
+
+#endif /* !__tilegx__ */
+
+/*
+ * This routine handles page faults.  It determines the address, and the
+ * problem, and then passes it handle_page_fault() for normal DTLB and
+ * ITLB issues, and for DMA or SN processor faults when we are in user
+ * space.  For the latter, if we're in kernel mode, we just save the
+ * interrupt away appropriately and return immediately.  We can't do
+ * page faults for user code while in kernel mode.
+ */
+void do_page_fault(struct pt_regs *regs, int fault_num,
+		   unsigned long address, unsigned long write)
+{
+	int is_page_fault;
+	enum ctx_state prev_state = exception_enter();
+
+#ifdef CONFIG_KPROBES
+	/*
+	 * This is to notify the fault handler of the kprobes.  The
+	 * exception code is redundant as it is also carried in REGS,
+	 * but we pass it anyhow.
+	 */
+	if (notify_die(DIE_PAGE_FAULT, "page fault", regs, -1,
+		       regs->faultnum, SIGSEGV) == NOTIFY_STOP)
+		goto done;
+#endif
+
+#ifdef __tilegx__
+	/*
+	 * We don't need early do_page_fault_ics() support, since unlike
+	 * Pro we don't need to worry about unlocking the atomic locks.
+	 * There is only one current case in GX where we touch any memory
+	 * under ICS other than our own kernel stack, and we handle that
+	 * here.  (If we crash due to trying to touch our own stack,
+	 * we're in too much trouble for C code to help out anyway.)
+	 */
+	if (write & ~1) {
+		unsigned long pc = write & ~1;
+		if (pc >= (unsigned long) __start_unalign_asm_code &&
+		    pc < (unsigned long) __end_unalign_asm_code) {
+			struct thread_info *ti = current_thread_info();
+			/*
+			 * Our EX_CONTEXT is still what it was from the
+			 * initial unalign exception, but now we've faulted
+			 * on the JIT page.  We would like to complete the
+			 * page fault however is appropriate, and then retry
+			 * the instruction that caused the unalign exception.
+			 * Our state has been "corrupted" by setting the low
+			 * bit in "sp", and stashing r0..r3 in the
+			 * thread_info area, so we revert all of that, then
+			 * continue as if this were a normal page fault.
+			 */
+			regs->sp &= ~1UL;
+			regs->regs[0] = ti->unalign_jit_tmp[0];
+			regs->regs[1] = ti->unalign_jit_tmp[1];
+			regs->regs[2] = ti->unalign_jit_tmp[2];
+			regs->regs[3] = ti->unalign_jit_tmp[3];
+			write &= 1;
+		} else {
+			pr_alert("%s/%d: ICS set at page fault at %#lx: %#lx\n",
+				 current->comm, current->pid, pc, address);
+			show_regs(regs);
+			do_group_exit(SIGKILL);
+		}
+	}
+#else
+	/* This case should have been handled by do_page_fault_ics(). */
+	BUG_ON(write & ~1);
+#endif
+
+#if CHIP_HAS_TILE_DMA()
+	/*
+	 * If it's a DMA fault, suspend the transfer while we're
+	 * handling the miss; we'll restart after it's handled.  If we
+	 * don't suspend, it's possible that this process could swap
+	 * out and back in, and restart the engine since the DMA is
+	 * still 'running'.
+	 */
+	if (fault_num == INT_DMATLB_MISS ||
+	    fault_num == INT_DMATLB_ACCESS ||
+	    fault_num == INT_DMATLB_MISS_DWNCL ||
+	    fault_num == INT_DMATLB_ACCESS_DWNCL) {
+		__insn_mtspr(SPR_DMA_CTR, SPR_DMA_CTR__SUSPEND_MASK);
+		while (__insn_mfspr(SPR_DMA_USER_STATUS) &
+		       SPR_DMA_STATUS__BUSY_MASK)
+			;
+	}
+#endif
+
+	/* Validate fault num and decide if this is a first-time page fault. */
+	switch (fault_num) {
+	case INT_ITLB_MISS:
+	case INT_DTLB_MISS:
+#if CHIP_HAS_TILE_DMA()
+	case INT_DMATLB_MISS:
+	case INT_DMATLB_MISS_DWNCL:
+#endif
+		is_page_fault = 1;
+		break;
+
+	case INT_DTLB_ACCESS:
+#if CHIP_HAS_TILE_DMA()
+	case INT_DMATLB_ACCESS:
+	case INT_DMATLB_ACCESS_DWNCL:
+#endif
+		is_page_fault = 0;
+		break;
+
+	default:
+		panic("Bad fault number %d in do_page_fault", fault_num);
+	}
+
+#if CHIP_HAS_TILE_DMA()
+	if (!user_mode(regs)) {
+		struct async_tlb *async;
+		switch (fault_num) {
+#if CHIP_HAS_TILE_DMA()
+		case INT_DMATLB_MISS:
+		case INT_DMATLB_ACCESS:
+		case INT_DMATLB_MISS_DWNCL:
+		case INT_DMATLB_ACCESS_DWNCL:
+			async = &current->thread.dma_async_tlb;
+			break;
+#endif
+		default:
+			async = NULL;
+		}
+		if (async) {
+
+			/*
+			 * No vmalloc check required, so we can allow
+			 * interrupts immediately at this point.
+			 */
+			local_irq_enable();
+
+			set_thread_flag(TIF_ASYNC_TLB);
+			if (async->fault_num != 0) {
+				panic("Second async fault %d; old fault was %d (%#lx/%ld)",
+				      fault_num, async->fault_num,
+				      address, write);
+			}
+			BUG_ON(fault_num == 0);
+			async->fault_num = fault_num;
+			async->is_fault = is_page_fault;
+			async->is_write = write;
+			async->address = address;
+			goto done;
+		}
+	}
+#endif
+
+	handle_page_fault(regs, fault_num, is_page_fault, address, write);
+
+done:
+	exception_exit(prev_state);
+}
+
+
+#if CHIP_HAS_TILE_DMA()
+/*
+ * This routine effectively re-issues asynchronous page faults
+ * when we are returning to user space.
+ */
+void do_async_page_fault(struct pt_regs *regs)
+{
+	struct async_tlb *async = &current->thread.dma_async_tlb;
+
+	/*
+	 * Clear thread flag early.  If we re-interrupt while processing
+	 * code here, we will reset it and recall this routine before
+	 * returning to user space.
+	 */
+	clear_thread_flag(TIF_ASYNC_TLB);
+
+	if (async->fault_num) {
+		/*
+		 * Clear async->fault_num before calling the page-fault
+		 * handler so that if we re-interrupt before returning
+		 * from the function we have somewhere to put the
+		 * information from the new interrupt.
+		 */
+		int fault_num = async->fault_num;
+		async->fault_num = 0;
+		handle_page_fault(regs, fault_num, async->is_fault,
+				  async->address, async->is_write);
+	}
+}
+#endif /* CHIP_HAS_TILE_DMA() */
+
+
+void vmalloc_sync_all(void)
+{
+#ifdef __tilegx__
+	/* Currently all L1 kernel pmd's are static and shared. */
+	BUILD_BUG_ON(pgd_index(VMALLOC_END - PAGE_SIZE) !=
+		     pgd_index(VMALLOC_START));
+#else
+	/*
+	 * Note that races in the updates of insync and start aren't
+	 * problematic: insync can only get set bits added, and updates to
+	 * start are only improving performance (without affecting correctness
+	 * if undone).
+	 */
+	static DECLARE_BITMAP(insync, PTRS_PER_PGD);
+	static unsigned long start = PAGE_OFFSET;
+	unsigned long address;
+
+	BUILD_BUG_ON(PAGE_OFFSET & ~PGDIR_MASK);
+	for (address = start; address >= PAGE_OFFSET; address += PGDIR_SIZE) {
+		if (!test_bit(pgd_index(address), insync)) {
+			unsigned long flags;
+			struct list_head *pos;
+
+			spin_lock_irqsave(&pgd_lock, flags);
+			list_for_each(pos, &pgd_list)
+				if (!vmalloc_sync_one(list_to_pgd(pos),
+								address)) {
+					/* Must be at first entry in list. */
+					BUG_ON(pos != pgd_list.next);
+					break;
+				}
+			spin_unlock_irqrestore(&pgd_lock, flags);
+			if (pos != pgd_list.next)
+				set_bit(pgd_index(address), insync);
+		}
+		if (address == start && test_bit(pgd_index(address), insync))
+			start = address + PGDIR_SIZE;
+	}
+#endif
+}
diff --git a/kernel/arch/tile/mm/highmem.c b/kernel/arch/tile/mm/highmem.c
new file mode 100644
index 000000000..fcd545014
--- /dev/null
+++ b/kernel/arch/tile/mm/highmem.c
@@ -0,0 +1,289 @@
+/*
+ * Copyright 2010 Tilera Corporation. All Rights Reserved.
+ *
+ *   This program is free software; you can redistribute it and/or
+ *   modify it under the terms of the GNU General Public License
+ *   as published by the Free Software Foundation, version 2.
+ *
+ *   This program is distributed in the hope that it will be useful, but
+ *   WITHOUT ANY WARRANTY; without even the implied warranty of
+ *   MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
+ *   NON INFRINGEMENT.  See the GNU General Public License for
+ *   more details.
+ */
+
+#include <linux/highmem.h>
+#include <linux/module.h>
+#include <linux/pagemap.h>
+#include <asm/homecache.h>
+
+#define kmap_get_pte(vaddr) \
+	pte_offset_kernel(pmd_offset(pud_offset(pgd_offset_k(vaddr), (vaddr)),\
+		(vaddr)), (vaddr))
+
+
+void *kmap(struct page *page)
+{
+	void *kva;
+	unsigned long flags;
+	pte_t *ptep;
+
+	might_sleep();
+	if (!PageHighMem(page))
+		return page_address(page);
+	kva = kmap_high(page);
+
+	/*
+	 * Rewrite the PTE under the lock.  This ensures that the page
+	 * is not currently migrating.
+	 */
+	ptep = kmap_get_pte((unsigned long)kva);
+	flags = homecache_kpte_lock();
+	set_pte_at(&init_mm, kva, ptep, mk_pte(page, page_to_kpgprot(page)));
+	homecache_kpte_unlock(flags);
+
+	return kva;
+}
+EXPORT_SYMBOL(kmap);
+
+void kunmap(struct page *page)
+{
+	if (in_interrupt())
+		BUG();
+	if (!PageHighMem(page))
+		return;
+	kunmap_high(page);
+}
+EXPORT_SYMBOL(kunmap);
+
+/*
+ * Describe a single atomic mapping of a page on a given cpu at a
+ * given address, and allow it to be linked into a list.
+ */
+struct atomic_mapped_page {
+	struct list_head list;
+	struct page *page;
+	int cpu;
+	unsigned long va;
+};
+
+static spinlock_t amp_lock = __SPIN_LOCK_UNLOCKED(&amp_lock);
+static struct list_head amp_list = LIST_HEAD_INIT(amp_list);
+
+/*
+ * Combining this structure with a per-cpu declaration lets us give
+ * each cpu an atomic_mapped_page structure per type.
+ */
+struct kmap_amps {
+	struct atomic_mapped_page per_type[KM_TYPE_NR];
+};
+static DEFINE_PER_CPU(struct kmap_amps, amps);
+
+/*
+ * Add a page and va, on this cpu, to the list of kmap_atomic pages,
+ * and write the new pte to memory.  Writing the new PTE under the
+ * lock guarantees that it is either on the list before migration starts
+ * (if we won the race), or set_pte() sets the migrating bit in the PTE
+ * (if we lost the race).  And doing it under the lock guarantees
+ * that when kmap_atomic_fix_one_pte() comes along, it finds a valid
+ * PTE in memory, iff the mapping is still on the amp_list.
+ *
+ * Finally, doing it under the lock lets us safely examine the page
+ * to see if it is immutable or not, for the generic kmap_atomic() case.
+ * If we examine it earlier we are exposed to a race where it looks
+ * writable earlier, but becomes immutable before we write the PTE.
+ */
+static void kmap_atomic_register(struct page *page, int type,
+				 unsigned long va, pte_t *ptep, pte_t pteval)
+{
+	unsigned long flags;
+	struct atomic_mapped_page *amp;
+
+	flags = homecache_kpte_lock();
+	spin_lock(&amp_lock);
+
+	/* With interrupts disabled, now fill in the per-cpu info. */
+	amp = this_cpu_ptr(&amps.per_type[type]);
+	amp->page = page;
+	amp->cpu = smp_processor_id();
+	amp->va = va;
+
+	/* For generic kmap_atomic(), choose the PTE writability now. */
+	if (!pte_read(pteval))
+		pteval = mk_pte(page, page_to_kpgprot(page));
+
+	list_add(&amp->list, &amp_list);
+	set_pte(ptep, pteval);
+
+	spin_unlock(&amp_lock);
+	homecache_kpte_unlock(flags);
+}
+
+/*
+ * Remove a page and va, on this cpu, from the list of kmap_atomic pages.
+ * Linear-time search, but we count on the lists being short.
+ * We don't need to adjust the PTE under the lock (as opposed to the
+ * kmap_atomic_register() case), since we're just unconditionally
+ * zeroing the PTE after it's off the list.
+ */
+static void kmap_atomic_unregister(struct page *page, unsigned long va)
+{
+	unsigned long flags;
+	struct atomic_mapped_page *amp;
+	int cpu = smp_processor_id();
+	spin_lock_irqsave(&amp_lock, flags);
+	list_for_each_entry(amp, &amp_list, list) {
+		if (amp->page == page && amp->cpu == cpu && amp->va == va)
+			break;
+	}
+	BUG_ON(&amp->list == &amp_list);
+	list_del(&amp->list);
+	spin_unlock_irqrestore(&amp_lock, flags);
+}
+
+/* Helper routine for kmap_atomic_fix_kpte(), below. */
+static void kmap_atomic_fix_one_kpte(struct atomic_mapped_page *amp,
+				     int finished)
+{
+	pte_t *ptep = kmap_get_pte(amp->va);
+	if (!finished) {
+		set_pte(ptep, pte_mkmigrate(*ptep));
+		flush_remote(0, 0, NULL, amp->va, PAGE_SIZE, PAGE_SIZE,
+			     cpumask_of(amp->cpu), NULL, 0);
+	} else {
+		/*
+		 * Rewrite a default kernel PTE for this page.
+		 * We rely on the fact that set_pte() writes the
+		 * present+migrating bits last.
+		 */
+		pte_t pte = mk_pte(amp->page, page_to_kpgprot(amp->page));
+		set_pte(ptep, pte);
+	}
+}
+
+/*
+ * This routine is a helper function for homecache_fix_kpte(); see
+ * its comments for more information on the "finished" argument here.
+ *
+ * Note that we hold the lock while doing the remote flushes, which
+ * will stall any unrelated cpus trying to do kmap_atomic operations.
+ * We could just update the PTEs under the lock, and save away copies
+ * of the structs (or just the va+cpu), then flush them after we
+ * release the lock, but it seems easier just to do it all under the lock.
+ */
+void kmap_atomic_fix_kpte(struct page *page, int finished)
+{
+	struct atomic_mapped_page *amp;
+	unsigned long flags;
+	spin_lock_irqsave(&amp_lock, flags);
+	list_for_each_entry(amp, &amp_list, list) {
+		if (amp->page == page)
+			kmap_atomic_fix_one_kpte(amp, finished);
+	}
+	spin_unlock_irqrestore(&amp_lock, flags);
+}
+
+/*
+ * kmap_atomic/kunmap_atomic is significantly faster than kmap/kunmap
+ * because the kmap code must perform a global TLB invalidation when
+ * the kmap pool wraps.
+ *
+ * Note that they may be slower than on x86 (etc.) because unlike on
+ * those platforms, we do have to take a global lock to map and unmap
+ * pages on Tile (see above).
+ *
+ * When holding an atomic kmap is is not legal to sleep, so atomic
+ * kmaps are appropriate for short, tight code paths only.
+ */
+void *kmap_atomic_prot(struct page *page, pgprot_t prot)
+{
+	unsigned long vaddr;
+	int idx, type;
+	pte_t *pte;
+
+	preempt_disable();
+	pagefault_disable();
+
+	/* Avoid icache flushes by disallowing atomic executable mappings. */
+	BUG_ON(pte_exec(prot));
+
+	if (!PageHighMem(page))
+		return page_address(page);
+
+	type = kmap_atomic_idx_push();
+	idx = type + KM_TYPE_NR*smp_processor_id();
+	vaddr = __fix_to_virt(FIX_KMAP_BEGIN + idx);
+	pte = kmap_get_pte(vaddr);
+	BUG_ON(!pte_none(*pte));
+
+	/* Register that this page is mapped atomically on this cpu. */
+	kmap_atomic_register(page, type, vaddr, pte, mk_pte(page, prot));
+
+	return (void *)vaddr;
+}
+EXPORT_SYMBOL(kmap_atomic_prot);
+
+void *kmap_atomic(struct page *page)
+{
+	/* PAGE_NONE is a magic value that tells us to check immutability. */
+	return kmap_atomic_prot(page, PAGE_NONE);
+}
+EXPORT_SYMBOL(kmap_atomic);
+
+void __kunmap_atomic(void *kvaddr)
+{
+	unsigned long vaddr = (unsigned long) kvaddr & PAGE_MASK;
+
+	if (vaddr >= __fix_to_virt(FIX_KMAP_END) &&
+	    vaddr <= __fix_to_virt(FIX_KMAP_BEGIN)) {
+		pte_t *pte = kmap_get_pte(vaddr);
+		pte_t pteval = *pte;
+		int idx, type;
+
+		type = kmap_atomic_idx();
+		idx = type + KM_TYPE_NR*smp_processor_id();
+
+		/*
+		 * Force other mappings to Oops if they try to access this pte
+		 * without first remapping it.  Keeping stale mappings around
+		 * is a bad idea.
+		 */
+		BUG_ON(!pte_present(pteval) && !pte_migrating(pteval));
+		kmap_atomic_unregister(pte_page(pteval), vaddr);
+		kpte_clear_flush(pte, vaddr);
+		kmap_atomic_idx_pop();
+	} else {
+		/* Must be a lowmem page */
+		BUG_ON(vaddr < PAGE_OFFSET);
+		BUG_ON(vaddr >= (unsigned long)high_memory);
+	}
+
+	pagefault_enable();
+	preempt_enable();
+}
+EXPORT_SYMBOL(__kunmap_atomic);
+
+/*
+ * This API is supposed to allow us to map memory without a "struct page".
+ * Currently we don't support this, though this may change in the future.
+ */
+void *kmap_atomic_pfn(unsigned long pfn)
+{
+	return kmap_atomic(pfn_to_page(pfn));
+}
+void *kmap_atomic_prot_pfn(unsigned long pfn, pgprot_t prot)
+{
+	return kmap_atomic_prot(pfn_to_page(pfn), prot);
+}
+
+struct page *kmap_atomic_to_page(void *ptr)
+{
+	pte_t *pte;
+	unsigned long vaddr = (unsigned long)ptr;
+
+	if (vaddr < FIXADDR_START)
+		return virt_to_page(ptr);
+
+	pte = kmap_get_pte(vaddr);
+	return pte_page(*pte);
+}
diff --git a/kernel/arch/tile/mm/homecache.c b/kernel/arch/tile/mm/homecache.c
new file mode 100644
index 000000000..40ca30a9f
--- /dev/null
+++ b/kernel/arch/tile/mm/homecache.c
@@ -0,0 +1,428 @@
+/*
+ * Copyright 2010 Tilera Corporation. All Rights Reserved.
+ *
+ *   This program is free software; you can redistribute it and/or
+ *   modify it under the terms of the GNU General Public License
+ *   as published by the Free Software Foundation, version 2.
+ *
+ *   This program is distributed in the hope that it will be useful, but
+ *   WITHOUT ANY WARRANTY; without even the implied warranty of
+ *   MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
+ *   NON INFRINGEMENT.  See the GNU General Public License for
+ *   more details.
+ *
+ * This code maintains the "home" for each page in the system.
+ */
+
+#include <linux/kernel.h>
+#include <linux/mm.h>
+#include <linux/spinlock.h>
+#include <linux/list.h>
+#include <linux/bootmem.h>
+#include <linux/rmap.h>
+#include <linux/pagemap.h>
+#include <linux/mutex.h>
+#include <linux/interrupt.h>
+#include <linux/sysctl.h>
+#include <linux/pagevec.h>
+#include <linux/ptrace.h>
+#include <linux/timex.h>
+#include <linux/cache.h>
+#include <linux/smp.h>
+#include <linux/module.h>
+#include <linux/hugetlb.h>
+
+#include <asm/page.h>
+#include <asm/sections.h>
+#include <asm/tlbflush.h>
+#include <asm/pgalloc.h>
+#include <asm/homecache.h>
+
+#include <arch/sim.h>
+
+#include "migrate.h"
+
+
+/*
+ * The noallocl2 option suppresses all use of the L2 cache to cache
+ * locally from a remote home.
+ */
+static int __write_once noallocl2;
+static int __init set_noallocl2(char *str)
+{
+	noallocl2 = 1;
+	return 0;
+}
+early_param("noallocl2", set_noallocl2);
+
+
+/*
+ * Update the irq_stat for cpus that we are going to interrupt
+ * with TLB or cache flushes.  Also handle removing dataplane cpus
+ * from the TLB flush set, and setting dataplane_tlb_state instead.
+ */
+static void hv_flush_update(const struct cpumask *cache_cpumask,
+			    struct cpumask *tlb_cpumask,
+			    unsigned long tlb_va, unsigned long tlb_length,
+			    HV_Remote_ASID *asids, int asidcount)
+{
+	struct cpumask mask;
+	int i, cpu;
+
+	cpumask_clear(&mask);
+	if (cache_cpumask)
+		cpumask_or(&mask, &mask, cache_cpumask);
+	if (tlb_cpumask && tlb_length) {
+		cpumask_or(&mask, &mask, tlb_cpumask);
+	}
+
+	for (i = 0; i < asidcount; ++i)
+		cpumask_set_cpu(asids[i].y * smp_width + asids[i].x, &mask);
+
+	/*
+	 * Don't bother to update atomically; losing a count
+	 * here is not that critical.
+	 */
+	for_each_cpu(cpu, &mask)
+		++per_cpu(irq_stat, cpu).irq_hv_flush_count;
+}
+
+/*
+ * This wrapper function around hv_flush_remote() does several things:
+ *
+ *  - Provides a return value error-checking panic path, since
+ *    there's never any good reason for hv_flush_remote() to fail.
+ *  - Accepts a 32-bit PFN rather than a 64-bit PA, which generally
+ *    is the type that Linux wants to pass around anyway.
+ *  - Canonicalizes that lengths of zero make cpumasks NULL.
+ *  - Handles deferring TLB flushes for dataplane tiles.
+ *  - Tracks remote interrupts in the per-cpu irq_cpustat_t.
+ *
+ * Note that we have to wait until the cache flush completes before
+ * updating the per-cpu last_cache_flush word, since otherwise another
+ * concurrent flush can race, conclude the flush has already
+ * completed, and start to use the page while it's still dirty
+ * remotely (running concurrently with the actual evict, presumably).
+ */
+void flush_remote(unsigned long cache_pfn, unsigned long cache_control,
+		  const struct cpumask *cache_cpumask_orig,
+		  HV_VirtAddr tlb_va, unsigned long tlb_length,
+		  unsigned long tlb_pgsize,
+		  const struct cpumask *tlb_cpumask_orig,
+		  HV_Remote_ASID *asids, int asidcount)
+{
+	int rc;
+	struct cpumask cache_cpumask_copy, tlb_cpumask_copy;
+	struct cpumask *cache_cpumask, *tlb_cpumask;
+	HV_PhysAddr cache_pa;
+
+	mb();   /* provided just to simplify "magic hypervisor" mode */
+
+	/*
+	 * Canonicalize and copy the cpumasks.
+	 */
+	if (cache_cpumask_orig && cache_control) {
+		cpumask_copy(&cache_cpumask_copy, cache_cpumask_orig);
+		cache_cpumask = &cache_cpumask_copy;
+	} else {
+		cpumask_clear(&cache_cpumask_copy);
+		cache_cpumask = NULL;
+	}
+	if (cache_cpumask == NULL)
+		cache_control = 0;
+	if (tlb_cpumask_orig && tlb_length) {
+		cpumask_copy(&tlb_cpumask_copy, tlb_cpumask_orig);
+		tlb_cpumask = &tlb_cpumask_copy;
+	} else {
+		cpumask_clear(&tlb_cpumask_copy);
+		tlb_cpumask = NULL;
+	}
+
+	hv_flush_update(cache_cpumask, tlb_cpumask, tlb_va, tlb_length,
+			asids, asidcount);
+	cache_pa = (HV_PhysAddr)cache_pfn << PAGE_SHIFT;
+	rc = hv_flush_remote(cache_pa, cache_control,
+			     cpumask_bits(cache_cpumask),
+			     tlb_va, tlb_length, tlb_pgsize,
+			     cpumask_bits(tlb_cpumask),
+			     asids, asidcount);
+	if (rc == 0)
+		return;
+
+	pr_err("hv_flush_remote(%#llx, %#lx, %p [%*pb], %#lx, %#lx, %#lx, %p [%*pb], %p, %d) = %d\n",
+	       cache_pa, cache_control, cache_cpumask,
+	       cpumask_pr_args(&cache_cpumask_copy),
+	       (unsigned long)tlb_va, tlb_length, tlb_pgsize, tlb_cpumask,
+	       cpumask_pr_args(&tlb_cpumask_copy), asids, asidcount, rc);
+	panic("Unsafe to continue.");
+}
+
+static void homecache_finv_page_va(void* va, int home)
+{
+	int cpu = get_cpu();
+	if (home == cpu) {
+		finv_buffer_local(va, PAGE_SIZE);
+	} else if (home == PAGE_HOME_HASH) {
+		finv_buffer_remote(va, PAGE_SIZE, 1);
+	} else {
+		BUG_ON(home < 0 || home >= NR_CPUS);
+		finv_buffer_remote(va, PAGE_SIZE, 0);
+	}
+	put_cpu();
+}
+
+void homecache_finv_map_page(struct page *page, int home)
+{
+	unsigned long flags;
+	unsigned long va;
+	pte_t *ptep;
+	pte_t pte;
+
+	if (home == PAGE_HOME_UNCACHED)
+		return;
+	local_irq_save(flags);
+#ifdef CONFIG_HIGHMEM
+	va = __fix_to_virt(FIX_KMAP_BEGIN + kmap_atomic_idx_push() +
+			   (KM_TYPE_NR * smp_processor_id()));
+#else
+	va = __fix_to_virt(FIX_HOMECACHE_BEGIN + smp_processor_id());
+#endif
+	ptep = virt_to_kpte(va);
+	pte = pfn_pte(page_to_pfn(page), PAGE_KERNEL);
+	__set_pte(ptep, pte_set_home(pte, home));
+	homecache_finv_page_va((void *)va, home);
+	__pte_clear(ptep);
+	hv_flush_page(va, PAGE_SIZE);
+#ifdef CONFIG_HIGHMEM
+	kmap_atomic_idx_pop();
+#endif
+	local_irq_restore(flags);
+}
+
+static void homecache_finv_page_home(struct page *page, int home)
+{
+	if (!PageHighMem(page) && home == page_home(page))
+		homecache_finv_page_va(page_address(page), home);
+	else
+		homecache_finv_map_page(page, home);
+}
+
+static inline bool incoherent_home(int home)
+{
+	return home == PAGE_HOME_IMMUTABLE || home == PAGE_HOME_INCOHERENT;
+}
+
+static void homecache_finv_page_internal(struct page *page, int force_map)
+{
+	int home = page_home(page);
+	if (home == PAGE_HOME_UNCACHED)
+		return;
+	if (incoherent_home(home)) {
+		int cpu;
+		for_each_cpu(cpu, &cpu_cacheable_map)
+			homecache_finv_map_page(page, cpu);
+	} else if (force_map) {
+		/* Force if, e.g., the normal mapping is migrating. */
+		homecache_finv_map_page(page, home);
+	} else {
+		homecache_finv_page_home(page, home);
+	}
+	sim_validate_lines_evicted(PFN_PHYS(page_to_pfn(page)), PAGE_SIZE);
+}
+
+void homecache_finv_page(struct page *page)
+{
+	homecache_finv_page_internal(page, 0);
+}
+
+void homecache_evict(const struct cpumask *mask)
+{
+	flush_remote(0, HV_FLUSH_EVICT_L2, mask, 0, 0, 0, NULL, NULL, 0);
+}
+
+/* Report the home corresponding to a given PTE. */
+static int pte_to_home(pte_t pte)
+{
+	if (hv_pte_get_nc(pte))
+		return PAGE_HOME_IMMUTABLE;
+	switch (hv_pte_get_mode(pte)) {
+	case HV_PTE_MODE_CACHE_TILE_L3:
+		return get_remote_cache_cpu(pte);
+	case HV_PTE_MODE_CACHE_NO_L3:
+		return PAGE_HOME_INCOHERENT;
+	case HV_PTE_MODE_UNCACHED:
+		return PAGE_HOME_UNCACHED;
+	case HV_PTE_MODE_CACHE_HASH_L3:
+		return PAGE_HOME_HASH;
+	}
+	panic("Bad PTE %#llx\n", pte.val);
+}
+
+/* Update the home of a PTE if necessary (can also be used for a pgprot_t). */
+pte_t pte_set_home(pte_t pte, int home)
+{
+#if CHIP_HAS_MMIO()
+	/* Check for MMIO mappings and pass them through. */
+	if (hv_pte_get_mode(pte) == HV_PTE_MODE_MMIO)
+		return pte;
+#endif
+
+
+	/*
+	 * Only immutable pages get NC mappings.  If we have a
+	 * non-coherent PTE, but the underlying page is not
+	 * immutable, it's likely the result of a forced
+	 * caching setting running up against ptrace setting
+	 * the page to be writable underneath.  In this case,
+	 * just keep the PTE coherent.
+	 */
+	if (hv_pte_get_nc(pte) && home != PAGE_HOME_IMMUTABLE) {
+		pte = hv_pte_clear_nc(pte);
+		pr_err("non-immutable page incoherently referenced: %#llx\n",
+		       pte.val);
+	}
+
+	switch (home) {
+
+	case PAGE_HOME_UNCACHED:
+		pte = hv_pte_set_mode(pte, HV_PTE_MODE_UNCACHED);
+		break;
+
+	case PAGE_HOME_INCOHERENT:
+		pte = hv_pte_set_mode(pte, HV_PTE_MODE_CACHE_NO_L3);
+		break;
+
+	case PAGE_HOME_IMMUTABLE:
+		/*
+		 * We could home this page anywhere, since it's immutable,
+		 * but by default just home it to follow "hash_default".
+		 */
+		BUG_ON(hv_pte_get_writable(pte));
+		if (pte_get_forcecache(pte)) {
+			/* Upgrade "force any cpu" to "No L3" for immutable. */
+			if (hv_pte_get_mode(pte) == HV_PTE_MODE_CACHE_TILE_L3
+			    && pte_get_anyhome(pte)) {
+				pte = hv_pte_set_mode(pte,
+						      HV_PTE_MODE_CACHE_NO_L3);
+			}
+		} else
+		if (hash_default)
+			pte = hv_pte_set_mode(pte, HV_PTE_MODE_CACHE_HASH_L3);
+		else
+			pte = hv_pte_set_mode(pte, HV_PTE_MODE_CACHE_NO_L3);
+		pte = hv_pte_set_nc(pte);
+		break;
+
+	case PAGE_HOME_HASH:
+		pte = hv_pte_set_mode(pte, HV_PTE_MODE_CACHE_HASH_L3);
+		break;
+
+	default:
+		BUG_ON(home < 0 || home >= NR_CPUS ||
+		       !cpu_is_valid_lotar(home));
+		pte = hv_pte_set_mode(pte, HV_PTE_MODE_CACHE_TILE_L3);
+		pte = set_remote_cache_cpu(pte, home);
+		break;
+	}
+
+	if (noallocl2)
+		pte = hv_pte_set_no_alloc_l2(pte);
+
+	/* Simplify "no local and no l3" to "uncached" */
+	if (hv_pte_get_no_alloc_l2(pte) && hv_pte_get_no_alloc_l1(pte) &&
+	    hv_pte_get_mode(pte) == HV_PTE_MODE_CACHE_NO_L3) {
+		pte = hv_pte_set_mode(pte, HV_PTE_MODE_UNCACHED);
+	}
+
+	/* Checking this case here gives a better panic than from the hv. */
+	BUG_ON(hv_pte_get_mode(pte) == 0);
+
+	return pte;
+}
+EXPORT_SYMBOL(pte_set_home);
+
+/*
+ * The routines in this section are the "static" versions of the normal
+ * dynamic homecaching routines; they just set the home cache
+ * of a kernel page once, and require a full-chip cache/TLB flush,
+ * so they're not suitable for anything but infrequent use.
+ */
+
+int page_home(struct page *page)
+{
+	if (PageHighMem(page)) {
+		return PAGE_HOME_HASH;
+	} else {
+		unsigned long kva = (unsigned long)page_address(page);
+		return pte_to_home(*virt_to_kpte(kva));
+	}
+}
+EXPORT_SYMBOL(page_home);
+
+void homecache_change_page_home(struct page *page, int order, int home)
+{
+	int i, pages = (1 << order);
+	unsigned long kva;
+
+	BUG_ON(PageHighMem(page));
+	BUG_ON(page_count(page) > 1);
+	BUG_ON(page_mapcount(page) != 0);
+	kva = (unsigned long) page_address(page);
+	flush_remote(0, HV_FLUSH_EVICT_L2, &cpu_cacheable_map,
+		     kva, pages * PAGE_SIZE, PAGE_SIZE, cpu_online_mask,
+		     NULL, 0);
+
+	for (i = 0; i < pages; ++i, kva += PAGE_SIZE) {
+		pte_t *ptep = virt_to_kpte(kva);
+		pte_t pteval = *ptep;
+		BUG_ON(!pte_present(pteval) || pte_huge(pteval));
+		__set_pte(ptep, pte_set_home(pteval, home));
+	}
+}
+EXPORT_SYMBOL(homecache_change_page_home);
+
+struct page *homecache_alloc_pages(gfp_t gfp_mask,
+				   unsigned int order, int home)
+{
+	struct page *page;
+	BUG_ON(gfp_mask & __GFP_HIGHMEM);   /* must be lowmem */
+	page = alloc_pages(gfp_mask, order);
+	if (page)
+		homecache_change_page_home(page, order, home);
+	return page;
+}
+EXPORT_SYMBOL(homecache_alloc_pages);
+
+struct page *homecache_alloc_pages_node(int nid, gfp_t gfp_mask,
+					unsigned int order, int home)
+{
+	struct page *page;
+	BUG_ON(gfp_mask & __GFP_HIGHMEM);   /* must be lowmem */
+	page = alloc_pages_node(nid, gfp_mask, order);
+	if (page)
+		homecache_change_page_home(page, order, home);
+	return page;
+}
+
+void __homecache_free_pages(struct page *page, unsigned int order)
+{
+	if (put_page_testzero(page)) {
+		homecache_change_page_home(page, order, PAGE_HOME_HASH);
+		if (order == 0) {
+			free_hot_cold_page(page, false);
+		} else {
+			init_page_count(page);
+			__free_pages(page, order);
+		}
+	}
+}
+EXPORT_SYMBOL(__homecache_free_pages);
+
+void homecache_free_pages(unsigned long addr, unsigned int order)
+{
+	if (addr != 0) {
+		VM_BUG_ON(!virt_addr_valid((void *)addr));
+		__homecache_free_pages(virt_to_page((void *)addr), order);
+	}
+}
+EXPORT_SYMBOL(homecache_free_pages);
diff --git a/kernel/arch/tile/mm/hugetlbpage.c b/kernel/arch/tile/mm/hugetlbpage.c
new file mode 100644
index 000000000..8416240c3
--- /dev/null
+++ b/kernel/arch/tile/mm/hugetlbpage.c
@@ -0,0 +1,346 @@
+/*
+ * Copyright 2010 Tilera Corporation. All Rights Reserved.
+ *
+ *   This program is free software; you can redistribute it and/or
+ *   modify it under the terms of the GNU General Public License
+ *   as published by the Free Software Foundation, version 2.
+ *
+ *   This program is distributed in the hope that it will be useful, but
+ *   WITHOUT ANY WARRANTY; without even the implied warranty of
+ *   MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
+ *   NON INFRINGEMENT.  See the GNU General Public License for
+ *   more details.
+ *
+ * TILE Huge TLB Page Support for Kernel.
+ * Taken from i386 hugetlb implementation:
+ * Copyright (C) 2002, Rohit Seth <rohit.seth@intel.com>
+ */
+
+#include <linux/init.h>
+#include <linux/fs.h>
+#include <linux/mm.h>
+#include <linux/hugetlb.h>
+#include <linux/pagemap.h>
+#include <linux/slab.h>
+#include <linux/err.h>
+#include <linux/sysctl.h>
+#include <linux/mman.h>
+#include <asm/tlb.h>
+#include <asm/tlbflush.h>
+#include <asm/setup.h>
+
+#ifdef CONFIG_HUGETLB_SUPER_PAGES
+
+/*
+ * Provide an additional huge page size (in addition to the regular default
+ * huge page size) if no "hugepagesz" arguments are specified.
+ * Note that it must be smaller than the default huge page size so
+ * that it's possible to allocate them on demand from the buddy allocator.
+ * You can change this to 64K (on a 16K build), 256K, 1M, or 4M,
+ * or not define it at all.
+ */
+#define ADDITIONAL_HUGE_SIZE (1024 * 1024UL)
+
+/* "Extra" page-size multipliers, one per level of the page table. */
+int huge_shift[HUGE_SHIFT_ENTRIES] = {
+#ifdef ADDITIONAL_HUGE_SIZE
+#define ADDITIONAL_HUGE_SHIFT __builtin_ctzl(ADDITIONAL_HUGE_SIZE / PAGE_SIZE)
+	[HUGE_SHIFT_PAGE] = ADDITIONAL_HUGE_SHIFT
+#endif
+};
+
+#endif
+
+pte_t *huge_pte_alloc(struct mm_struct *mm,
+		      unsigned long addr, unsigned long sz)
+{
+	pgd_t *pgd;
+	pud_t *pud;
+
+	addr &= -sz;   /* Mask off any low bits in the address. */
+
+	pgd = pgd_offset(mm, addr);
+	pud = pud_alloc(mm, pgd, addr);
+
+#ifdef CONFIG_HUGETLB_SUPER_PAGES
+	if (sz >= PGDIR_SIZE) {
+		BUG_ON(sz != PGDIR_SIZE &&
+		       sz != PGDIR_SIZE << huge_shift[HUGE_SHIFT_PGDIR]);
+		return (pte_t *)pud;
+	} else {
+		pmd_t *pmd = pmd_alloc(mm, pud, addr);
+		if (sz >= PMD_SIZE) {
+			BUG_ON(sz != PMD_SIZE &&
+			       sz != (PMD_SIZE << huge_shift[HUGE_SHIFT_PMD]));
+			return (pte_t *)pmd;
+		}
+		else {
+			if (sz != PAGE_SIZE << huge_shift[HUGE_SHIFT_PAGE])
+				panic("Unexpected page size %#lx\n", sz);
+			return pte_alloc_map(mm, NULL, pmd, addr);
+		}
+	}
+#else
+	BUG_ON(sz != PMD_SIZE);
+	return (pte_t *) pmd_alloc(mm, pud, addr);
+#endif
+}
+
+static pte_t *get_pte(pte_t *base, int index, int level)
+{
+	pte_t *ptep = base + index;
+#ifdef CONFIG_HUGETLB_SUPER_PAGES
+	if (!pte_present(*ptep) && huge_shift[level] != 0) {
+		unsigned long mask = -1UL << huge_shift[level];
+		pte_t *super_ptep = base + (index & mask);
+		pte_t pte = *super_ptep;
+		if (pte_present(pte) && pte_super(pte))
+			ptep = super_ptep;
+	}
+#endif
+	return ptep;
+}
+
+pte_t *huge_pte_offset(struct mm_struct *mm, unsigned long addr)
+{
+	pgd_t *pgd;
+	pud_t *pud;
+	pmd_t *pmd;
+#ifdef CONFIG_HUGETLB_SUPER_PAGES
+	pte_t *pte;
+#endif
+
+	/* Get the top-level page table entry. */
+	pgd = (pgd_t *)get_pte((pte_t *)mm->pgd, pgd_index(addr), 0);
+
+	/* We don't have four levels. */
+	pud = pud_offset(pgd, addr);
+#ifndef __PAGETABLE_PUD_FOLDED
+# error support fourth page table level
+#endif
+	if (!pud_present(*pud))
+		return NULL;
+
+	/* Check for an L0 huge PTE, if we have three levels. */
+#ifndef __PAGETABLE_PMD_FOLDED
+	if (pud_huge(*pud))
+		return (pte_t *)pud;
+
+	pmd = (pmd_t *)get_pte((pte_t *)pud_page_vaddr(*pud),
+			       pmd_index(addr), 1);
+	if (!pmd_present(*pmd))
+		return NULL;
+#else
+	pmd = pmd_offset(pud, addr);
+#endif
+
+	/* Check for an L1 huge PTE. */
+	if (pmd_huge(*pmd))
+		return (pte_t *)pmd;
+
+#ifdef CONFIG_HUGETLB_SUPER_PAGES
+	/* Check for an L2 huge PTE. */
+	pte = get_pte((pte_t *)pmd_page_vaddr(*pmd), pte_index(addr), 2);
+	if (!pte_present(*pte))
+		return NULL;
+	if (pte_super(*pte))
+		return pte;
+#endif
+
+	return NULL;
+}
+
+int pmd_huge(pmd_t pmd)
+{
+	return !!(pmd_val(pmd) & _PAGE_HUGE_PAGE);
+}
+
+int pud_huge(pud_t pud)
+{
+	return !!(pud_val(pud) & _PAGE_HUGE_PAGE);
+}
+
+int huge_pmd_unshare(struct mm_struct *mm, unsigned long *addr, pte_t *ptep)
+{
+	return 0;
+}
+
+#ifdef HAVE_ARCH_HUGETLB_UNMAPPED_AREA
+static unsigned long hugetlb_get_unmapped_area_bottomup(struct file *file,
+		unsigned long addr, unsigned long len,
+		unsigned long pgoff, unsigned long flags)
+{
+	struct hstate *h = hstate_file(file);
+	struct vm_unmapped_area_info info;
+
+	info.flags = 0;
+	info.length = len;
+	info.low_limit = TASK_UNMAPPED_BASE;
+	info.high_limit = TASK_SIZE;
+	info.align_mask = PAGE_MASK & ~huge_page_mask(h);
+	info.align_offset = 0;
+	return vm_unmapped_area(&info);
+}
+
+static unsigned long hugetlb_get_unmapped_area_topdown(struct file *file,
+		unsigned long addr0, unsigned long len,
+		unsigned long pgoff, unsigned long flags)
+{
+	struct hstate *h = hstate_file(file);
+	struct vm_unmapped_area_info info;
+	unsigned long addr;
+
+	info.flags = VM_UNMAPPED_AREA_TOPDOWN;
+	info.length = len;
+	info.low_limit = PAGE_SIZE;
+	info.high_limit = current->mm->mmap_base;
+	info.align_mask = PAGE_MASK & ~huge_page_mask(h);
+	info.align_offset = 0;
+	addr = vm_unmapped_area(&info);
+
+	/*
+	 * A failed mmap() very likely causes application failure,
+	 * so fall back to the bottom-up function here. This scenario
+	 * can happen with large stack limits and large mmap()
+	 * allocations.
+	 */
+	if (addr & ~PAGE_MASK) {
+		VM_BUG_ON(addr != -ENOMEM);
+		info.flags = 0;
+		info.low_limit = TASK_UNMAPPED_BASE;
+		info.high_limit = TASK_SIZE;
+		addr = vm_unmapped_area(&info);
+	}
+
+	return addr;
+}
+
+unsigned long hugetlb_get_unmapped_area(struct file *file, unsigned long addr,
+		unsigned long len, unsigned long pgoff, unsigned long flags)
+{
+	struct hstate *h = hstate_file(file);
+	struct mm_struct *mm = current->mm;
+	struct vm_area_struct *vma;
+
+	if (len & ~huge_page_mask(h))
+		return -EINVAL;
+	if (len > TASK_SIZE)
+		return -ENOMEM;
+
+	if (flags & MAP_FIXED) {
+		if (prepare_hugepage_range(file, addr, len))
+			return -EINVAL;
+		return addr;
+	}
+
+	if (addr) {
+		addr = ALIGN(addr, huge_page_size(h));
+		vma = find_vma(mm, addr);
+		if (TASK_SIZE - len >= addr &&
+		    (!vma || addr + len <= vma->vm_start))
+			return addr;
+	}
+	if (current->mm->get_unmapped_area == arch_get_unmapped_area)
+		return hugetlb_get_unmapped_area_bottomup(file, addr, len,
+				pgoff, flags);
+	else
+		return hugetlb_get_unmapped_area_topdown(file, addr, len,
+				pgoff, flags);
+}
+#endif /* HAVE_ARCH_HUGETLB_UNMAPPED_AREA */
+
+#ifdef CONFIG_HUGETLB_SUPER_PAGES
+static __init int __setup_hugepagesz(unsigned long ps)
+{
+	int log_ps = __builtin_ctzl(ps);
+	int level, base_shift;
+
+	if ((1UL << log_ps) != ps || (log_ps & 1) != 0) {
+		pr_warn("Not enabling %ld byte huge pages; must be a power of four\n",
+			ps);
+		return -EINVAL;
+	}
+
+	if (ps > 64*1024*1024*1024UL) {
+		pr_warn("Not enabling %ld MB huge pages; largest legal value is 64 GB\n",
+			ps >> 20);
+		return -EINVAL;
+	} else if (ps >= PUD_SIZE) {
+		static long hv_jpage_size;
+		if (hv_jpage_size == 0)
+			hv_jpage_size = hv_sysconf(HV_SYSCONF_PAGE_SIZE_JUMBO);
+		if (hv_jpage_size != PUD_SIZE) {
+			pr_warn("Not enabling >= %ld MB huge pages: hypervisor reports size %ld\n",
+				PUD_SIZE >> 20, hv_jpage_size);
+			return -EINVAL;
+		}
+		level = 0;
+		base_shift = PUD_SHIFT;
+	} else if (ps >= PMD_SIZE) {
+		level = 1;
+		base_shift = PMD_SHIFT;
+	} else if (ps > PAGE_SIZE) {
+		level = 2;
+		base_shift = PAGE_SHIFT;
+	} else {
+		pr_err("hugepagesz: huge page size %ld too small\n", ps);
+		return -EINVAL;
+	}
+
+	if (log_ps != base_shift) {
+		int shift_val = log_ps - base_shift;
+		if (huge_shift[level] != 0) {
+			int old_shift = base_shift + huge_shift[level];
+			pr_warn("Not enabling %ld MB huge pages; already have size %ld MB\n",
+				ps >> 20, (1UL << old_shift) >> 20);
+			return -EINVAL;
+		}
+		if (hv_set_pte_super_shift(level, shift_val) != 0) {
+			pr_warn("Not enabling %ld MB huge pages; no hypervisor support\n",
+				ps >> 20);
+			return -EINVAL;
+		}
+		printk(KERN_DEBUG "Enabled %ld MB huge pages\n", ps >> 20);
+		huge_shift[level] = shift_val;
+	}
+
+	hugetlb_add_hstate(log_ps - PAGE_SHIFT);
+
+	return 0;
+}
+
+static bool saw_hugepagesz;
+
+static __init int setup_hugepagesz(char *opt)
+{
+	if (!saw_hugepagesz) {
+		saw_hugepagesz = true;
+		memset(huge_shift, 0, sizeof(huge_shift));
+	}
+	return __setup_hugepagesz(memparse(opt, NULL));
+}
+__setup("hugepagesz=", setup_hugepagesz);
+
+#ifdef ADDITIONAL_HUGE_SIZE
+/*
+ * Provide an additional huge page size if no "hugepagesz" args are given.
+ * In that case, all the cores have properly set up their hv super_shift
+ * already, but we need to notify the hugetlb code to enable the
+ * new huge page size from the Linux point of view.
+ */
+static __init int add_default_hugepagesz(void)
+{
+	if (!saw_hugepagesz) {
+		BUILD_BUG_ON(ADDITIONAL_HUGE_SIZE >= PMD_SIZE ||
+			     ADDITIONAL_HUGE_SIZE <= PAGE_SIZE);
+		BUILD_BUG_ON((PAGE_SIZE << ADDITIONAL_HUGE_SHIFT) !=
+			     ADDITIONAL_HUGE_SIZE);
+		BUILD_BUG_ON(ADDITIONAL_HUGE_SHIFT & 1);
+		hugetlb_add_hstate(ADDITIONAL_HUGE_SHIFT);
+	}
+	return 0;
+}
+arch_initcall(add_default_hugepagesz);
+#endif
+
+#endif /* CONFIG_HUGETLB_SUPER_PAGES */
diff --git a/kernel/arch/tile/mm/init.c b/kernel/arch/tile/mm/init.c
new file mode 100644
index 000000000..5bd252e3f
--- /dev/null
+++ b/kernel/arch/tile/mm/init.c
@@ -0,0 +1,983 @@
+/*
+ * Copyright (C) 1995  Linus Torvalds
+ * Copyright 2010 Tilera Corporation. All Rights Reserved.
+ *
+ *   This program is free software; you can redistribute it and/or
+ *   modify it under the terms of the GNU General Public License
+ *   as published by the Free Software Foundation, version 2.
+ *
+ *   This program is distributed in the hope that it will be useful, but
+ *   WITHOUT ANY WARRANTY; without even the implied warranty of
+ *   MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
+ *   NON INFRINGEMENT.  See the GNU General Public License for
+ *   more details.
+ */
+
+#include <linux/module.h>
+#include <linux/signal.h>
+#include <linux/sched.h>
+#include <linux/kernel.h>
+#include <linux/errno.h>
+#include <linux/string.h>
+#include <linux/types.h>
+#include <linux/ptrace.h>
+#include <linux/mman.h>
+#include <linux/mm.h>
+#include <linux/hugetlb.h>
+#include <linux/swap.h>
+#include <linux/smp.h>
+#include <linux/init.h>
+#include <linux/highmem.h>
+#include <linux/pagemap.h>
+#include <linux/poison.h>
+#include <linux/bootmem.h>
+#include <linux/slab.h>
+#include <linux/proc_fs.h>
+#include <linux/efi.h>
+#include <linux/memory_hotplug.h>
+#include <linux/uaccess.h>
+#include <asm/mmu_context.h>
+#include <asm/processor.h>
+#include <asm/pgtable.h>
+#include <asm/pgalloc.h>
+#include <asm/dma.h>
+#include <asm/fixmap.h>
+#include <asm/tlb.h>
+#include <asm/tlbflush.h>
+#include <asm/sections.h>
+#include <asm/setup.h>
+#include <asm/homecache.h>
+#include <hv/hypervisor.h>
+#include <arch/chip.h>
+
+#include "migrate.h"
+
+#define clear_pgd(pmdptr) (*(pmdptr) = hv_pte(0))
+
+#ifndef __tilegx__
+unsigned long VMALLOC_RESERVE = CONFIG_VMALLOC_RESERVE;
+EXPORT_SYMBOL(VMALLOC_RESERVE);
+#endif
+
+/* Create an L2 page table */
+static pte_t * __init alloc_pte(void)
+{
+	return __alloc_bootmem(L2_KERNEL_PGTABLE_SIZE, HV_PAGE_TABLE_ALIGN, 0);
+}
+
+/*
+ * L2 page tables per controller.  We allocate these all at once from
+ * the bootmem allocator and store them here.  This saves on kernel L2
+ * page table memory, compared to allocating a full 64K page per L2
+ * page table, and also means that in cases where we use huge pages,
+ * we are guaranteed to later be able to shatter those huge pages and
+ * switch to using these page tables instead, without requiring
+ * further allocation.  Each l2_ptes[] entry points to the first page
+ * table for the first hugepage-size piece of memory on the
+ * controller; other page tables are just indexed directly, i.e. the
+ * L2 page tables are contiguous in memory for each controller.
+ */
+static pte_t *l2_ptes[MAX_NUMNODES];
+static int num_l2_ptes[MAX_NUMNODES];
+
+static void init_prealloc_ptes(int node, int pages)
+{
+	BUG_ON(pages & (PTRS_PER_PTE - 1));
+	if (pages) {
+		num_l2_ptes[node] = pages;
+		l2_ptes[node] = __alloc_bootmem(pages * sizeof(pte_t),
+						HV_PAGE_TABLE_ALIGN, 0);
+	}
+}
+
+pte_t *get_prealloc_pte(unsigned long pfn)
+{
+	int node = pfn_to_nid(pfn);
+	pfn &= ~(-1UL << (NR_PA_HIGHBIT_SHIFT - PAGE_SHIFT));
+	BUG_ON(node >= MAX_NUMNODES);
+	BUG_ON(pfn >= num_l2_ptes[node]);
+	return &l2_ptes[node][pfn];
+}
+
+/*
+ * What caching do we expect pages from the heap to have when
+ * they are allocated during bootup?  (Once we've installed the
+ * "real" swapper_pg_dir.)
+ */
+static int initial_heap_home(void)
+{
+	if (hash_default)
+		return PAGE_HOME_HASH;
+	return smp_processor_id();
+}
+
+/*
+ * Place a pointer to an L2 page table in a middle page
+ * directory entry.
+ */
+static void __init assign_pte(pmd_t *pmd, pte_t *page_table)
+{
+	phys_addr_t pa = __pa(page_table);
+	unsigned long l2_ptfn = pa >> HV_LOG2_PAGE_TABLE_ALIGN;
+	pte_t pteval = hv_pte_set_ptfn(__pgprot(_PAGE_TABLE), l2_ptfn);
+	BUG_ON((pa & (HV_PAGE_TABLE_ALIGN-1)) != 0);
+	pteval = pte_set_home(pteval, initial_heap_home());
+	*(pte_t *)pmd = pteval;
+	if (page_table != (pte_t *)pmd_page_vaddr(*pmd))
+		BUG();
+}
+
+#ifdef __tilegx__
+
+static inline pmd_t *alloc_pmd(void)
+{
+	return __alloc_bootmem(L1_KERNEL_PGTABLE_SIZE, HV_PAGE_TABLE_ALIGN, 0);
+}
+
+static inline void assign_pmd(pud_t *pud, pmd_t *pmd)
+{
+	assign_pte((pmd_t *)pud, (pte_t *)pmd);
+}
+
+#endif /* __tilegx__ */
+
+/* Replace the given pmd with a full PTE table. */
+void __init shatter_pmd(pmd_t *pmd)
+{
+	pte_t *pte = get_prealloc_pte(pte_pfn(*(pte_t *)pmd));
+	assign_pte(pmd, pte);
+}
+
+#ifdef __tilegx__
+static pmd_t *__init get_pmd(pgd_t pgtables[], unsigned long va)
+{
+	pud_t *pud = pud_offset(&pgtables[pgd_index(va)], va);
+	if (pud_none(*pud))
+		assign_pmd(pud, alloc_pmd());
+	return pmd_offset(pud, va);
+}
+#else
+static pmd_t *__init get_pmd(pgd_t pgtables[], unsigned long va)
+{
+	return pmd_offset(pud_offset(&pgtables[pgd_index(va)], va), va);
+}
+#endif
+
+/*
+ * This function initializes a certain range of kernel virtual memory
+ * with new bootmem page tables, everywhere page tables are missing in
+ * the given range.
+ */
+
+/*
+ * NOTE: The pagetables are allocated contiguous on the physical space
+ * so we can cache the place of the first one and move around without
+ * checking the pgd every time.
+ */
+static void __init page_table_range_init(unsigned long start,
+					 unsigned long end, pgd_t *pgd)
+{
+	unsigned long vaddr;
+	start = round_down(start, PMD_SIZE);
+	end = round_up(end, PMD_SIZE);
+	for (vaddr = start; vaddr < end; vaddr += PMD_SIZE) {
+		pmd_t *pmd = get_pmd(pgd, vaddr);
+		if (pmd_none(*pmd))
+			assign_pte(pmd, alloc_pte());
+	}
+}
+
+
+static int __initdata ktext_hash = 1;  /* .text pages */
+static int __initdata kdata_hash = 1;  /* .data and .bss pages */
+int __write_once hash_default = 1;     /* kernel allocator pages */
+EXPORT_SYMBOL(hash_default);
+int __write_once kstack_hash = 1;      /* if no homecaching, use h4h */
+
+/*
+ * CPUs to use to for striping the pages of kernel data.  If hash-for-home
+ * is available, this is only relevant if kcache_hash sets up the
+ * .data and .bss to be page-homed, and we don't want the default mode
+ * of using the full set of kernel cpus for the striping.
+ */
+static __initdata struct cpumask kdata_mask;
+static __initdata int kdata_arg_seen;
+
+int __write_once kdata_huge;       /* if no homecaching, small pages */
+
+
+/* Combine a generic pgprot_t with cache home to get a cache-aware pgprot. */
+static pgprot_t __init construct_pgprot(pgprot_t prot, int home)
+{
+	prot = pte_set_home(prot, home);
+	if (home == PAGE_HOME_IMMUTABLE) {
+		if (ktext_hash)
+			prot = hv_pte_set_mode(prot, HV_PTE_MODE_CACHE_HASH_L3);
+		else
+			prot = hv_pte_set_mode(prot, HV_PTE_MODE_CACHE_NO_L3);
+	}
+	return prot;
+}
+
+/*
+ * For a given kernel data VA, how should it be cached?
+ * We return the complete pgprot_t with caching bits set.
+ */
+static pgprot_t __init init_pgprot(ulong address)
+{
+	int cpu;
+	unsigned long page;
+	enum { CODE_DELTA = MEM_SV_START - PAGE_OFFSET };
+
+	/* For kdata=huge, everything is just hash-for-home. */
+	if (kdata_huge)
+		return construct_pgprot(PAGE_KERNEL, PAGE_HOME_HASH);
+
+	/*
+	 * We map the aliased pages of permanent text so we can
+	 * update them if necessary, for ftrace, etc.
+	 */
+	if (address < (ulong) _sinittext - CODE_DELTA)
+		return construct_pgprot(PAGE_KERNEL, PAGE_HOME_HASH);
+
+	/* We map read-only data non-coherent for performance. */
+	if ((address >= (ulong) __start_rodata &&
+	     address < (ulong) __end_rodata) ||
+	    address == (ulong) empty_zero_page) {
+		return construct_pgprot(PAGE_KERNEL_RO, PAGE_HOME_IMMUTABLE);
+	}
+
+#ifndef __tilegx__
+	/* Force the atomic_locks[] array page to be hash-for-home. */
+	if (address == (ulong) atomic_locks)
+		return construct_pgprot(PAGE_KERNEL, PAGE_HOME_HASH);
+#endif
+
+	/*
+	 * Everything else that isn't data or bss is heap, so mark it
+	 * with the initial heap home (hash-for-home, or this cpu).  This
+	 * includes any addresses after the loaded image and any address before
+	 * __init_end, since we already captured the case of text before
+	 * _sinittext, and __pa(einittext) is approximately __pa(__init_begin).
+	 *
+	 * All the LOWMEM pages that we mark this way will get their
+	 * struct page homecache properly marked later, in set_page_homes().
+	 * The HIGHMEM pages we leave with a default zero for their
+	 * homes, but with a zero free_time we don't have to actually
+	 * do a flush action the first time we use them, either.
+	 */
+	if (address >= (ulong) _end || address < (ulong) __init_end)
+		return construct_pgprot(PAGE_KERNEL, initial_heap_home());
+
+	/* Use hash-for-home if requested for data/bss. */
+	if (kdata_hash)
+		return construct_pgprot(PAGE_KERNEL, PAGE_HOME_HASH);
+
+	/*
+	 * Otherwise we just hand out consecutive cpus.  To avoid
+	 * requiring this function to hold state, we just walk forward from
+	 * __end_rodata by PAGE_SIZE, skipping the readonly and init data, to
+	 * reach the requested address, while walking cpu home around
+	 * kdata_mask. This is typically no more than a dozen or so iterations.
+	 */
+	page = (((ulong)__end_rodata) + PAGE_SIZE - 1) & PAGE_MASK;
+	BUG_ON(address < page || address >= (ulong)_end);
+	cpu = cpumask_first(&kdata_mask);
+	for (; page < address; page += PAGE_SIZE) {
+		if (page >= (ulong)&init_thread_union &&
+		    page < (ulong)&init_thread_union + THREAD_SIZE)
+			continue;
+		if (page == (ulong)empty_zero_page)
+			continue;
+#ifndef __tilegx__
+		if (page == (ulong)atomic_locks)
+			continue;
+#endif
+		cpu = cpumask_next(cpu, &kdata_mask);
+		if (cpu == NR_CPUS)
+			cpu = cpumask_first(&kdata_mask);
+	}
+	return construct_pgprot(PAGE_KERNEL, cpu);
+}
+
+/*
+ * This function sets up how we cache the kernel text.  If we have
+ * hash-for-home support, normally that is used instead (see the
+ * kcache_hash boot flag for more information).  But if we end up
+ * using a page-based caching technique, this option sets up the
+ * details of that.  In addition, the "ktext=nocache" option may
+ * always be used to disable local caching of text pages, if desired.
+ */
+
+static int __initdata ktext_arg_seen;
+static int __initdata ktext_small;
+static int __initdata ktext_local;
+static int __initdata ktext_all;
+static int __initdata ktext_nondataplane;
+static int __initdata ktext_nocache;
+static struct cpumask __initdata ktext_mask;
+
+static int __init setup_ktext(char *str)
+{
+	if (str == NULL)
+		return -EINVAL;
+
+	/* If you have a leading "nocache", turn off ktext caching */
+	if (strncmp(str, "nocache", 7) == 0) {
+		ktext_nocache = 1;
+		pr_info("ktext: disabling local caching of kernel text\n");
+		str += 7;
+		if (*str == ',')
+			++str;
+		if (*str == '\0')
+			return 0;
+	}
+
+	ktext_arg_seen = 1;
+
+	/* Default setting: use a huge page */
+	if (strcmp(str, "huge") == 0)
+		pr_info("ktext: using one huge locally cached page\n");
+
+	/* Pay TLB cost but get no cache benefit: cache small pages locally */
+	else if (strcmp(str, "local") == 0) {
+		ktext_small = 1;
+		ktext_local = 1;
+		pr_info("ktext: using small pages with local caching\n");
+	}
+
+	/* Neighborhood cache ktext pages on all cpus. */
+	else if (strcmp(str, "all") == 0) {
+		ktext_small = 1;
+		ktext_all = 1;
+		pr_info("ktext: using maximal caching neighborhood\n");
+	}
+
+
+	/* Neighborhood ktext pages on specified mask */
+	else if (cpulist_parse(str, &ktext_mask) == 0) {
+		if (cpumask_weight(&ktext_mask) > 1) {
+			ktext_small = 1;
+			pr_info("ktext: using caching neighborhood %*pbl with small pages\n",
+				cpumask_pr_args(&ktext_mask));
+		} else {
+			pr_info("ktext: caching on cpu %*pbl with one huge page\n",
+				cpumask_pr_args(&ktext_mask));
+		}
+	}
+
+	else if (*str)
+		return -EINVAL;
+
+	return 0;
+}
+
+early_param("ktext", setup_ktext);
+
+
+static inline pgprot_t ktext_set_nocache(pgprot_t prot)
+{
+	if (!ktext_nocache)
+		prot = hv_pte_set_nc(prot);
+	else
+		prot = hv_pte_set_no_alloc_l2(prot);
+	return prot;
+}
+
+/* Temporary page table we use for staging. */
+static pgd_t pgtables[PTRS_PER_PGD]
+ __attribute__((aligned(HV_PAGE_TABLE_ALIGN)));
+
+/*
+ * This maps the physical memory to kernel virtual address space, a total
+ * of max_low_pfn pages, by creating page tables starting from address
+ * PAGE_OFFSET.
+ *
+ * This routine transitions us from using a set of compiled-in large
+ * pages to using some more precise caching, including removing access
+ * to code pages mapped at PAGE_OFFSET (executed only at MEM_SV_START)
+ * marking read-only data as locally cacheable, striping the remaining
+ * .data and .bss across all the available tiles, and removing access
+ * to pages above the top of RAM (thus ensuring a page fault from a bad
+ * virtual address rather than a hypervisor shoot down for accessing
+ * memory outside the assigned limits).
+ */
+static void __init kernel_physical_mapping_init(pgd_t *pgd_base)
+{
+	unsigned long long irqmask;
+	unsigned long address, pfn;
+	pmd_t *pmd;
+	pte_t *pte;
+	int pte_ofs;
+	const struct cpumask *my_cpu_mask = cpumask_of(smp_processor_id());
+	struct cpumask kstripe_mask;
+	int rc, i;
+
+	if (ktext_arg_seen && ktext_hash) {
+		pr_warn("warning: \"ktext\" boot argument ignored if \"kcache_hash\" sets up text hash-for-home\n");
+		ktext_small = 0;
+	}
+
+	if (kdata_arg_seen && kdata_hash) {
+		pr_warn("warning: \"kdata\" boot argument ignored if \"kcache_hash\" sets up data hash-for-home\n");
+	}
+
+	if (kdata_huge && !hash_default) {
+		pr_warn("warning: disabling \"kdata=huge\"; requires kcache_hash=all or =allbutstack\n");
+		kdata_huge = 0;
+	}
+
+	/*
+	 * Set up a mask for cpus to use for kernel striping.
+	 * This is normally all cpus, but minus dataplane cpus if any.
+	 * If the dataplane covers the whole chip, we stripe over
+	 * the whole chip too.
+	 */
+	cpumask_copy(&kstripe_mask, cpu_possible_mask);
+	if (!kdata_arg_seen)
+		kdata_mask = kstripe_mask;
+
+	/* Allocate and fill in L2 page tables */
+	for (i = 0; i < MAX_NUMNODES; ++i) {
+#ifdef CONFIG_HIGHMEM
+		unsigned long end_pfn = node_lowmem_end_pfn[i];
+#else
+		unsigned long end_pfn = node_end_pfn[i];
+#endif
+		unsigned long end_huge_pfn = 0;
+
+		/* Pre-shatter the last huge page to allow per-cpu pages. */
+		if (kdata_huge)
+			end_huge_pfn = end_pfn - (HPAGE_SIZE >> PAGE_SHIFT);
+
+		pfn = node_start_pfn[i];
+
+		/* Allocate enough memory to hold L2 page tables for node. */
+		init_prealloc_ptes(i, end_pfn - pfn);
+
+		address = (unsigned long) pfn_to_kaddr(pfn);
+		while (pfn < end_pfn) {
+			BUG_ON(address & (HPAGE_SIZE-1));
+			pmd = get_pmd(pgtables, address);
+			pte = get_prealloc_pte(pfn);
+			if (pfn < end_huge_pfn) {
+				pgprot_t prot = init_pgprot(address);
+				*(pte_t *)pmd = pte_mkhuge(pfn_pte(pfn, prot));
+				for (pte_ofs = 0; pte_ofs < PTRS_PER_PTE;
+				     pfn++, pte_ofs++, address += PAGE_SIZE)
+					pte[pte_ofs] = pfn_pte(pfn, prot);
+			} else {
+				if (kdata_huge)
+					printk(KERN_DEBUG "pre-shattered huge page at %#lx\n",
+					       address);
+				for (pte_ofs = 0; pte_ofs < PTRS_PER_PTE;
+				     pfn++, pte_ofs++, address += PAGE_SIZE) {
+					pgprot_t prot = init_pgprot(address);
+					pte[pte_ofs] = pfn_pte(pfn, prot);
+				}
+				assign_pte(pmd, pte);
+			}
+		}
+	}
+
+	/*
+	 * Set or check ktext_map now that we have cpu_possible_mask
+	 * and kstripe_mask to work with.
+	 */
+	if (ktext_all)
+		cpumask_copy(&ktext_mask, cpu_possible_mask);
+	else if (ktext_nondataplane)
+		ktext_mask = kstripe_mask;
+	else if (!cpumask_empty(&ktext_mask)) {
+		/* Sanity-check any mask that was requested */
+		struct cpumask bad;
+		cpumask_andnot(&bad, &ktext_mask, cpu_possible_mask);
+		cpumask_and(&ktext_mask, &ktext_mask, cpu_possible_mask);
+		if (!cpumask_empty(&bad))
+			pr_info("ktext: not using unavailable cpus %*pbl\n",
+				cpumask_pr_args(&bad));
+		if (cpumask_empty(&ktext_mask)) {
+			pr_warn("ktext: no valid cpus; caching on %d\n",
+				smp_processor_id());
+			cpumask_copy(&ktext_mask,
+				     cpumask_of(smp_processor_id()));
+		}
+	}
+
+	address = MEM_SV_START;
+	pmd = get_pmd(pgtables, address);
+	pfn = 0;  /* code starts at PA 0 */
+	if (ktext_small) {
+		/* Allocate an L2 PTE for the kernel text */
+		int cpu = 0;
+		pgprot_t prot = construct_pgprot(PAGE_KERNEL_EXEC,
+						 PAGE_HOME_IMMUTABLE);
+
+		if (ktext_local) {
+			if (ktext_nocache)
+				prot = hv_pte_set_mode(prot,
+						       HV_PTE_MODE_UNCACHED);
+			else
+				prot = hv_pte_set_mode(prot,
+						       HV_PTE_MODE_CACHE_NO_L3);
+		} else {
+			prot = hv_pte_set_mode(prot,
+					       HV_PTE_MODE_CACHE_TILE_L3);
+			cpu = cpumask_first(&ktext_mask);
+
+			prot = ktext_set_nocache(prot);
+		}
+
+		BUG_ON(address != (unsigned long)_text);
+		pte = NULL;
+		for (; address < (unsigned long)_einittext;
+		     pfn++, address += PAGE_SIZE) {
+			pte_ofs = pte_index(address);
+			if (pte_ofs == 0) {
+				if (pte)
+					assign_pte(pmd++, pte);
+				pte = alloc_pte();
+			}
+			if (!ktext_local) {
+				prot = set_remote_cache_cpu(prot, cpu);
+				cpu = cpumask_next(cpu, &ktext_mask);
+				if (cpu == NR_CPUS)
+					cpu = cpumask_first(&ktext_mask);
+			}
+			pte[pte_ofs] = pfn_pte(pfn, prot);
+		}
+		if (pte)
+			assign_pte(pmd, pte);
+	} else {
+		pte_t pteval = pfn_pte(0, PAGE_KERNEL_EXEC);
+		pteval = pte_mkhuge(pteval);
+		if (ktext_hash) {
+			pteval = hv_pte_set_mode(pteval,
+						 HV_PTE_MODE_CACHE_HASH_L3);
+			pteval = ktext_set_nocache(pteval);
+		} else
+		if (cpumask_weight(&ktext_mask) == 1) {
+			pteval = set_remote_cache_cpu(pteval,
+					      cpumask_first(&ktext_mask));
+			pteval = hv_pte_set_mode(pteval,
+						 HV_PTE_MODE_CACHE_TILE_L3);
+			pteval = ktext_set_nocache(pteval);
+		} else if (ktext_nocache)
+			pteval = hv_pte_set_mode(pteval,
+						 HV_PTE_MODE_UNCACHED);
+		else
+			pteval = hv_pte_set_mode(pteval,
+						 HV_PTE_MODE_CACHE_NO_L3);
+		for (; address < (unsigned long)_einittext;
+		     pfn += PFN_DOWN(HPAGE_SIZE), address += HPAGE_SIZE)
+			*(pte_t *)(pmd++) = pfn_pte(pfn, pteval);
+	}
+
+	/* Set swapper_pgprot here so it is flushed to memory right away. */
+	swapper_pgprot = init_pgprot((unsigned long)swapper_pg_dir);
+
+	/*
+	 * Since we may be changing the caching of the stack and page
+	 * table itself, we invoke an assembly helper to do the
+	 * following steps:
+	 *
+	 *  - flush the cache so we start with an empty slate
+	 *  - install pgtables[] as the real page table
+	 *  - flush the TLB so the new page table takes effect
+	 */
+	irqmask = interrupt_mask_save_mask();
+	interrupt_mask_set_mask(-1ULL);
+	rc = flush_and_install_context(__pa(pgtables),
+				       init_pgprot((unsigned long)pgtables),
+				       __this_cpu_read(current_asid),
+				       cpumask_bits(my_cpu_mask));
+	interrupt_mask_restore_mask(irqmask);
+	BUG_ON(rc != 0);
+
+	/* Copy the page table back to the normal swapper_pg_dir. */
+	memcpy(pgd_base, pgtables, sizeof(pgtables));
+	__install_page_table(pgd_base, __this_cpu_read(current_asid),
+			     swapper_pgprot);
+
+	/*
+	 * We just read swapper_pgprot and thus brought it into the cache,
+	 * with its new home & caching mode.  When we start the other CPUs,
+	 * they're going to reference swapper_pgprot via their initial fake
+	 * VA-is-PA mappings, which cache everything locally.  At that
+	 * time, if it's in our cache with a conflicting home, the
+	 * simulator's coherence checker will complain.  So, flush it out
+	 * of our cache; we're not going to ever use it again anyway.
+	 */
+	__insn_finv(&swapper_pgprot);
+}
+
+/*
+ * devmem_is_allowed() checks to see if /dev/mem access to a certain address
+ * is valid. The argument is a physical page number.
+ *
+ * On Tile, the only valid things for which we can just hand out unchecked
+ * PTEs are the kernel code and data.  Anything else might change its
+ * homing with time, and we wouldn't know to adjust the /dev/mem PTEs.
+ * Note that init_thread_union is released to heap soon after boot,
+ * so we include it in the init data.
+ *
+ * For TILE-Gx, we might want to consider allowing access to PA
+ * regions corresponding to PCI space, etc.
+ */
+int devmem_is_allowed(unsigned long pagenr)
+{
+	return pagenr < kaddr_to_pfn(_end) &&
+		!(pagenr >= kaddr_to_pfn(&init_thread_union) ||
+		  pagenr < kaddr_to_pfn(__init_end)) &&
+		!(pagenr >= kaddr_to_pfn(_sinittext) ||
+		  pagenr <= kaddr_to_pfn(_einittext-1));
+}
+
+#ifdef CONFIG_HIGHMEM
+static void __init permanent_kmaps_init(pgd_t *pgd_base)
+{
+	pgd_t *pgd;
+	pud_t *pud;
+	pmd_t *pmd;
+	pte_t *pte;
+	unsigned long vaddr;
+
+	vaddr = PKMAP_BASE;
+	page_table_range_init(vaddr, vaddr + PAGE_SIZE*LAST_PKMAP, pgd_base);
+
+	pgd = swapper_pg_dir + pgd_index(vaddr);
+	pud = pud_offset(pgd, vaddr);
+	pmd = pmd_offset(pud, vaddr);
+	pte = pte_offset_kernel(pmd, vaddr);
+	pkmap_page_table = pte;
+}
+#endif /* CONFIG_HIGHMEM */
+
+
+#ifndef CONFIG_64BIT
+static void __init init_free_pfn_range(unsigned long start, unsigned long end)
+{
+	unsigned long pfn;
+	struct page *page = pfn_to_page(start);
+
+	for (pfn = start; pfn < end; ) {
+		/* Optimize by freeing pages in large batches */
+		int order = __ffs(pfn);
+		int count, i;
+		struct page *p;
+
+		if (order >= MAX_ORDER)
+			order = MAX_ORDER-1;
+		count = 1 << order;
+		while (pfn + count > end) {
+			count >>= 1;
+			--order;
+		}
+		for (p = page, i = 0; i < count; ++i, ++p) {
+			__ClearPageReserved(p);
+			/*
+			 * Hacky direct set to avoid unnecessary
+			 * lock take/release for EVERY page here.
+			 */
+			p->_count.counter = 0;
+			p->_mapcount.counter = -1;
+		}
+		init_page_count(page);
+		__free_pages(page, order);
+		adjust_managed_page_count(page, count);
+
+		page += count;
+		pfn += count;
+	}
+}
+
+static void __init set_non_bootmem_pages_init(void)
+{
+	struct zone *z;
+	for_each_zone(z) {
+		unsigned long start, end;
+		int nid = z->zone_pgdat->node_id;
+#ifdef CONFIG_HIGHMEM
+		int idx = zone_idx(z);
+#endif
+
+		start = z->zone_start_pfn;
+		end = start + z->spanned_pages;
+		start = max(start, node_free_pfn[nid]);
+		start = max(start, max_low_pfn);
+
+#ifdef CONFIG_HIGHMEM
+		if (idx == ZONE_HIGHMEM)
+			totalhigh_pages += z->spanned_pages;
+#endif
+		if (kdata_huge) {
+			unsigned long percpu_pfn = node_percpu_pfn[nid];
+			if (start < percpu_pfn && end > percpu_pfn)
+				end = percpu_pfn;
+		}
+#ifdef CONFIG_PCI
+		if (start <= pci_reserve_start_pfn &&
+		    end > pci_reserve_start_pfn) {
+			if (end > pci_reserve_end_pfn)
+				init_free_pfn_range(pci_reserve_end_pfn, end);
+			end = pci_reserve_start_pfn;
+		}
+#endif
+		init_free_pfn_range(start, end);
+	}
+}
+#endif
+
+/*
+ * paging_init() sets up the page tables - note that all of lowmem is
+ * already mapped by head.S.
+ */
+void __init paging_init(void)
+{
+#ifdef __tilegx__
+	pud_t *pud;
+#endif
+	pgd_t *pgd_base = swapper_pg_dir;
+
+	kernel_physical_mapping_init(pgd_base);
+
+	/* Fixed mappings, only the page table structure has to be created. */
+	page_table_range_init(fix_to_virt(__end_of_fixed_addresses - 1),
+			      FIXADDR_TOP, pgd_base);
+
+#ifdef CONFIG_HIGHMEM
+	permanent_kmaps_init(pgd_base);
+#endif
+
+#ifdef __tilegx__
+	/*
+	 * Since GX allocates just one pmd_t array worth of vmalloc space,
+	 * we go ahead and allocate it statically here, then share it
+	 * globally.  As a result we don't have to worry about any task
+	 * changing init_mm once we get up and running, and there's no
+	 * need for e.g. vmalloc_sync_all().
+	 */
+	BUILD_BUG_ON(pgd_index(VMALLOC_START) != pgd_index(VMALLOC_END - 1));
+	pud = pud_offset(pgd_base + pgd_index(VMALLOC_START), VMALLOC_START);
+	assign_pmd(pud, alloc_pmd());
+#endif
+}
+
+
+/*
+ * Walk the kernel page tables and derive the page_home() from
+ * the PTEs, so that set_pte() can properly validate the caching
+ * of all PTEs it sees.
+ */
+void __init set_page_homes(void)
+{
+}
+
+static void __init set_max_mapnr_init(void)
+{
+#ifdef CONFIG_FLATMEM
+	max_mapnr = max_low_pfn;
+#endif
+}
+
+void __init mem_init(void)
+{
+	int i;
+#ifndef __tilegx__
+	void *last;
+#endif
+
+#ifdef CONFIG_FLATMEM
+	BUG_ON(!mem_map);
+#endif
+
+#ifdef CONFIG_HIGHMEM
+	/* check that fixmap and pkmap do not overlap */
+	if (PKMAP_ADDR(LAST_PKMAP-1) >= FIXADDR_START) {
+		pr_err("fixmap and kmap areas overlap - this will crash\n");
+		pr_err("pkstart: %lxh pkend: %lxh fixstart %lxh\n",
+		       PKMAP_BASE, PKMAP_ADDR(LAST_PKMAP-1), FIXADDR_START);
+		BUG();
+	}
+#endif
+
+	set_max_mapnr_init();
+
+	/* this will put all bootmem onto the freelists */
+	free_all_bootmem();
+
+#ifndef CONFIG_64BIT
+	/* count all remaining LOWMEM and give all HIGHMEM to page allocator */
+	set_non_bootmem_pages_init();
+#endif
+
+	mem_init_print_info(NULL);
+
+	/*
+	 * In debug mode, dump some interesting memory mappings.
+	 */
+#ifdef CONFIG_HIGHMEM
+	printk(KERN_DEBUG "  KMAP    %#lx - %#lx\n",
+	       FIXADDR_START, FIXADDR_TOP + PAGE_SIZE - 1);
+	printk(KERN_DEBUG "  PKMAP   %#lx - %#lx\n",
+	       PKMAP_BASE, PKMAP_ADDR(LAST_PKMAP) - 1);
+#endif
+	printk(KERN_DEBUG "  VMALLOC %#lx - %#lx\n",
+	       _VMALLOC_START, _VMALLOC_END - 1);
+#ifdef __tilegx__
+	for (i = MAX_NUMNODES-1; i >= 0; --i) {
+		struct pglist_data *node = &node_data[i];
+		if (node->node_present_pages) {
+			unsigned long start = (unsigned long)
+				pfn_to_kaddr(node->node_start_pfn);
+			unsigned long end = start +
+				(node->node_present_pages << PAGE_SHIFT);
+			printk(KERN_DEBUG "  MEM%d    %#lx - %#lx\n",
+			       i, start, end - 1);
+		}
+	}
+#else
+	last = high_memory;
+	for (i = MAX_NUMNODES-1; i >= 0; --i) {
+		if ((unsigned long)vbase_map[i] != -1UL) {
+			printk(KERN_DEBUG "  LOWMEM%d %#lx - %#lx\n",
+			       i, (unsigned long) (vbase_map[i]),
+			       (unsigned long) (last-1));
+			last = vbase_map[i];
+		}
+	}
+#endif
+
+#ifndef __tilegx__
+	/*
+	 * Convert from using one lock for all atomic operations to
+	 * one per cpu.
+	 */
+	__init_atomic_per_cpu();
+#endif
+}
+
+/*
+ * this is for the non-NUMA, single node SMP system case.
+ * Specifically, in the case of x86, we will always add
+ * memory to the highmem for now.
+ */
+#ifndef CONFIG_NEED_MULTIPLE_NODES
+int arch_add_memory(u64 start, u64 size)
+{
+	struct pglist_data *pgdata = &contig_page_data;
+	struct zone *zone = pgdata->node_zones + MAX_NR_ZONES-1;
+	unsigned long start_pfn = start >> PAGE_SHIFT;
+	unsigned long nr_pages = size >> PAGE_SHIFT;
+
+	return __add_pages(zone, start_pfn, nr_pages);
+}
+
+int remove_memory(u64 start, u64 size)
+{
+	return -EINVAL;
+}
+
+#ifdef CONFIG_MEMORY_HOTREMOVE
+int arch_remove_memory(u64 start, u64 size)
+{
+	/* TODO */
+	return -EBUSY;
+}
+#endif
+#endif
+
+struct kmem_cache *pgd_cache;
+
+void __init pgtable_cache_init(void)
+{
+	pgd_cache = kmem_cache_create("pgd", SIZEOF_PGD, SIZEOF_PGD, 0, NULL);
+	if (!pgd_cache)
+		panic("pgtable_cache_init(): Cannot create pgd cache");
+}
+
+#ifdef CONFIG_DEBUG_PAGEALLOC
+static long __write_once initfree;
+#else
+static long __write_once initfree = 1;
+#endif
+
+/* Select whether to free (1) or mark unusable (0) the __init pages. */
+static int __init set_initfree(char *str)
+{
+	long val;
+	if (kstrtol(str, 0, &val) == 0) {
+		initfree = val;
+		pr_info("initfree: %s free init pages\n",
+			initfree ? "will" : "won't");
+	}
+	return 1;
+}
+__setup("initfree=", set_initfree);
+
+static void free_init_pages(char *what, unsigned long begin, unsigned long end)
+{
+	unsigned long addr = (unsigned long) begin;
+
+	if (kdata_huge && !initfree) {
+		pr_warn("Warning: ignoring initfree=0: incompatible with kdata=huge\n");
+		initfree = 1;
+	}
+	end = (end + PAGE_SIZE - 1) & PAGE_MASK;
+	local_flush_tlb_pages(NULL, begin, PAGE_SIZE, end - begin);
+	for (addr = begin; addr < end; addr += PAGE_SIZE) {
+		/*
+		 * Note we just reset the home here directly in the
+		 * page table.  We know this is safe because our caller
+		 * just flushed the caches on all the other cpus,
+		 * and they won't be touching any of these pages.
+		 */
+		int pfn = kaddr_to_pfn((void *)addr);
+		struct page *page = pfn_to_page(pfn);
+		pte_t *ptep = virt_to_kpte(addr);
+		if (!initfree) {
+			/*
+			 * If debugging page accesses then do not free
+			 * this memory but mark them not present - any
+			 * buggy init-section access will create a
+			 * kernel page fault:
+			 */
+			pte_clear(&init_mm, addr, ptep);
+			continue;
+		}
+		if (pte_huge(*ptep))
+			BUG_ON(!kdata_huge);
+		else
+			set_pte_at(&init_mm, addr, ptep,
+				   pfn_pte(pfn, PAGE_KERNEL));
+		memset((void *)addr, POISON_FREE_INITMEM, PAGE_SIZE);
+		free_reserved_page(page);
+	}
+	pr_info("Freeing %s: %ldk freed\n", what, (end - begin) >> 10);
+}
+
+void free_initmem(void)
+{
+	const unsigned long text_delta = MEM_SV_START - PAGE_OFFSET;
+
+	/*
+	 * Evict the cache on all cores to avoid incoherence.
+	 * We are guaranteed that no one will touch the init pages any more.
+	 */
+	homecache_evict(&cpu_cacheable_map);
+
+	/* Free the data pages that we won't use again after init. */
+	free_init_pages("unused kernel data",
+			(unsigned long)__init_begin,
+			(unsigned long)__init_end);
+
+	/*
+	 * Free the pages mapped from 0xc0000000 that correspond to code
+	 * pages from MEM_SV_START that we won't use again after init.
+	 */
+	free_init_pages("unused kernel text",
+			(unsigned long)_sinittext - text_delta,
+			(unsigned long)_einittext - text_delta);
+	/* Do a global TLB flush so everyone sees the changes. */
+	flush_tlb_all();
+}
diff --git a/kernel/arch/tile/mm/migrate.h b/kernel/arch/tile/mm/migrate.h
new file mode 100644
index 000000000..91683d979
--- /dev/null
+++ b/kernel/arch/tile/mm/migrate.h
@@ -0,0 +1,56 @@
+/*
+ * Copyright 2010 Tilera Corporation. All Rights Reserved.
+ *
+ *   This program is free software; you can redistribute it and/or
+ *   modify it under the terms of the GNU General Public License
+ *   as published by the Free Software Foundation, version 2.
+ *
+ *   This program is distributed in the hope that it will be useful, but
+ *   WITHOUT ANY WARRANTY; without even the implied warranty of
+ *   MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
+ *   NON INFRINGEMENT.  See the GNU General Public License for
+ *   more details.
+ *
+ * Structure definitions for migration, exposed here for use by
+ * arch/tile/kernel/asm-offsets.c.
+ */
+
+#ifndef MM_MIGRATE_H
+#define MM_MIGRATE_H
+
+#include <linux/cpumask.h>
+#include <hv/hypervisor.h>
+
+/*
+ * This function is used as a helper when setting up the initial
+ * page table (swapper_pg_dir).
+ *
+ * You must mask ALL interrupts prior to invoking this code, since
+ * you can't legally touch the stack during the cache flush.
+ */
+extern int flush_and_install_context(HV_PhysAddr page_table, HV_PTE access,
+				     HV_ASID asid,
+				     const unsigned long *cpumask);
+
+/*
+ * This function supports migration as a "helper" as follows:
+ *
+ *  - Set the stack PTE itself to "migrating".
+ *  - Do a global TLB flush for (va,length) and the specified ASIDs.
+ *  - Do a cache-evict on all necessary cpus.
+ *  - Write the new stack PTE.
+ *
+ * Note that any non-NULL pointers must not point to the page that
+ * is handled by the stack_pte itself.
+ *
+ * You must mask ALL interrupts prior to invoking this code, since
+ * you can't legally touch the stack during the cache flush.
+ */
+extern int homecache_migrate_stack_and_flush(pte_t stack_pte, unsigned long va,
+				     size_t length, pte_t *stack_ptep,
+				     const struct cpumask *cache_cpumask,
+				     const struct cpumask *tlb_cpumask,
+				     HV_Remote_ASID *asids,
+				     int asidcount);
+
+#endif /* MM_MIGRATE_H */
diff --git a/kernel/arch/tile/mm/migrate_32.S b/kernel/arch/tile/mm/migrate_32.S
new file mode 100644
index 000000000..772085491
--- /dev/null
+++ b/kernel/arch/tile/mm/migrate_32.S
@@ -0,0 +1,192 @@
+/*
+ * Copyright 2010 Tilera Corporation. All Rights Reserved.
+ *
+ *   This program is free software; you can redistribute it and/or
+ *   modify it under the terms of the GNU General Public License
+ *   as published by the Free Software Foundation, version 2.
+ *
+ *   This program is distributed in the hope that it will be useful, but
+ *   WITHOUT ANY WARRANTY; without even the implied warranty of
+ *   MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
+ *   NON INFRINGEMENT.  See the GNU General Public License for
+ *   more details.
+ *
+ * This routine is a helper for migrating the home of a set of pages to
+ * a new cpu.  See the documentation in homecache.c for more information.
+ */
+
+#include <linux/linkage.h>
+#include <linux/threads.h>
+#include <asm/page.h>
+#include <asm/thread_info.h>
+#include <asm/types.h>
+#include <asm/asm-offsets.h>
+#include <hv/hypervisor.h>
+
+	.text
+
+/*
+ * First, some definitions that apply to all the code in the file.
+ */
+
+/* Locals (caller-save) */
+#define r_tmp		r10
+#define r_save_sp	r11
+
+/* What we save where in the stack frame; must include all callee-saves. */
+#define FRAME_SP	4
+#define FRAME_R30	8
+#define FRAME_R31	12
+#define FRAME_R32	16
+#define FRAME_R33	20
+#define FRAME_R34	24
+#define FRAME_SIZE	28
+
+
+
+
+/*
+ * On entry:
+ *
+ *   r0 low word of the new context PA to install (moved to r_context_lo)
+ *   r1 high word of the new context PA to install (moved to r_context_hi)
+ *   r2 low word of PTE to use for context access (moved to r_access_lo)
+ *   r3 high word of PTE to use for context access (moved to r_access_lo)
+ *   r4 ASID to use for new context (moved to r_asid)
+ *   r5 pointer to cpumask with just this cpu set in it (r_my_cpumask)
+ */
+
+/* Arguments (caller-save) */
+#define r_context_lo_in	r0
+#define r_context_hi_in	r1
+#define r_access_lo_in	r2
+#define r_access_hi_in	r3
+#define r_asid_in	r4
+#define r_my_cpumask	r5
+
+/* Locals (callee-save); must not be more than FRAME_xxx above. */
+#define r_context_lo	r30
+#define r_context_hi	r31
+#define r_access_lo	r32
+#define r_access_hi	r33
+#define r_asid		r34
+
+STD_ENTRY(flush_and_install_context)
+	/*
+	 * Create a stack frame; we can't touch it once we flush the
+	 * cache until we install the new page table and flush the TLB.
+	 */
+	{
+	 move r_save_sp, sp
+	 sw sp, lr
+	 addi sp, sp, -FRAME_SIZE
+	}
+	addi r_tmp, sp, FRAME_SP
+	{
+	 sw r_tmp, r_save_sp
+	 addi r_tmp, sp, FRAME_R30
+	}
+	{
+	 sw r_tmp, r30
+	 addi r_tmp, sp, FRAME_R31
+	}
+	{
+	 sw r_tmp, r31
+	 addi r_tmp, sp, FRAME_R32
+	}
+	{
+	 sw r_tmp, r32
+	 addi r_tmp, sp, FRAME_R33
+	}
+	{
+	 sw r_tmp, r33
+	 addi r_tmp, sp, FRAME_R34
+	}
+	sw r_tmp, r34
+
+	/* Move some arguments to callee-save registers. */
+	{
+	 move r_context_lo, r_context_lo_in
+	 move r_context_hi, r_context_hi_in
+	}
+	{
+	 move r_access_lo, r_access_lo_in
+	 move r_access_hi, r_access_hi_in
+	}
+	move r_asid, r_asid_in
+
+	/* First, flush our L2 cache. */
+	{
+	 move r0, zero  /* cache_pa */
+	 move r1, zero
+	}
+	{
+	 auli r2, zero, ha16(HV_FLUSH_EVICT_L2)  /* cache_control */
+	 move r3, r_my_cpumask  /* cache_cpumask */
+	}
+	{
+	 move r4, zero  /* tlb_va */
+	 move r5, zero  /* tlb_length */
+	}
+	{
+	 move r6, zero  /* tlb_pgsize */
+	 move r7, zero  /* tlb_cpumask */
+	}
+	{
+	 move r8, zero  /* asids */
+	 move r9, zero  /* asidcount */
+	}
+	jal _hv_flush_remote
+	bnz r0, .Ldone
+
+	/* Now install the new page table. */
+	{
+	 move r0, r_context_lo
+	 move r1, r_context_hi
+	}
+	{
+	 move r2, r_access_lo
+	 move r3, r_access_hi
+	}
+	{
+	 move r4, r_asid
+	 moveli r5, HV_CTX_DIRECTIO | CTX_PAGE_FLAG
+	}
+	jal _hv_install_context
+	bnz r0, .Ldone
+
+	/* Finally, flush the TLB. */
+	{
+	 movei r0, 0   /* preserve_global */
+	 jal hv_flush_all
+	}
+
+.Ldone:
+	/* Restore the callee-saved registers and return. */
+	addli lr, sp, FRAME_SIZE
+	{
+	 lw lr, lr
+	 addli r_tmp, sp, FRAME_R30
+	}
+	{
+	 lw r30, r_tmp
+	 addli r_tmp, sp, FRAME_R31
+	}
+	{
+	 lw r31, r_tmp
+	 addli r_tmp, sp, FRAME_R32
+	}
+	{
+	 lw r32, r_tmp
+	 addli r_tmp, sp, FRAME_R33
+	}
+	{
+	 lw r33, r_tmp
+	 addli r_tmp, sp, FRAME_R34
+	}
+	{
+	 lw r34, r_tmp
+	 addi sp, sp, FRAME_SIZE
+	}
+	jrp lr
+	STD_ENDPROC(flush_and_install_context)
diff --git a/kernel/arch/tile/mm/migrate_64.S b/kernel/arch/tile/mm/migrate_64.S
new file mode 100644
index 000000000..a49eee38f
--- /dev/null
+++ b/kernel/arch/tile/mm/migrate_64.S
@@ -0,0 +1,167 @@
+/*
+ * Copyright 2011 Tilera Corporation. All Rights Reserved.
+ *
+ *   This program is free software; you can redistribute it and/or
+ *   modify it under the terms of the GNU General Public License
+ *   as published by the Free Software Foundation, version 2.
+ *
+ *   This program is distributed in the hope that it will be useful, but
+ *   WITHOUT ANY WARRANTY; without even the implied warranty of
+ *   MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
+ *   NON INFRINGEMENT.  See the GNU General Public License for
+ *   more details.
+ *
+ * This routine is a helper for migrating the home of a set of pages to
+ * a new cpu.  See the documentation in homecache.c for more information.
+ */
+
+#include <linux/linkage.h>
+#include <linux/threads.h>
+#include <asm/page.h>
+#include <asm/thread_info.h>
+#include <asm/types.h>
+#include <asm/asm-offsets.h>
+#include <hv/hypervisor.h>
+
+	.text
+
+/*
+ * First, some definitions that apply to all the code in the file.
+ */
+
+/* Locals (caller-save) */
+#define r_tmp		r10
+#define r_save_sp	r11
+
+/* What we save where in the stack frame; must include all callee-saves. */
+#define FRAME_SP	8
+#define FRAME_R30	16
+#define FRAME_R31	24
+#define FRAME_R32	32
+#define FRAME_SIZE	40
+
+
+
+
+/*
+ * On entry:
+ *
+ *   r0 the new context PA to install (moved to r_context)
+ *   r1 PTE to use for context access (moved to r_access)
+ *   r2 ASID to use for new context (moved to r_asid)
+ *   r3 pointer to cpumask with just this cpu set in it (r_my_cpumask)
+ */
+
+/* Arguments (caller-save) */
+#define r_context_in	r0
+#define r_access_in	r1
+#define r_asid_in	r2
+#define r_my_cpumask	r3
+
+/* Locals (callee-save); must not be more than FRAME_xxx above. */
+#define r_context	r30
+#define r_access	r31
+#define r_asid		r32
+
+/*
+ * Caller-save locals and frame constants are the same as
+ * for homecache_migrate_stack_and_flush.
+ */
+
+STD_ENTRY(flush_and_install_context)
+	/*
+	 * Create a stack frame; we can't touch it once we flush the
+	 * cache until we install the new page table and flush the TLB.
+	 */
+	{
+	 move r_save_sp, sp
+	 st sp, lr
+	 addi sp, sp, -FRAME_SIZE
+	}
+	addi r_tmp, sp, FRAME_SP
+	{
+	 st r_tmp, r_save_sp
+	 addi r_tmp, sp, FRAME_R30
+	}
+	{
+	 st r_tmp, r30
+	 addi r_tmp, sp, FRAME_R31
+	}
+	{
+	 st r_tmp, r31
+	 addi r_tmp, sp, FRAME_R32
+	}
+	st r_tmp, r32
+
+	/* Move some arguments to callee-save registers. */
+	{
+	 move r_context, r_context_in
+	 move r_access, r_access_in
+	}
+	move r_asid, r_asid_in
+
+	/* First, flush our L2 cache. */
+	{
+	 move r0, zero  /* cache_pa */
+	 moveli r1, hw2_last(HV_FLUSH_EVICT_L2)  /* cache_control */
+	}
+	{
+	 shl16insli r1, r1, hw1(HV_FLUSH_EVICT_L2)
+	 move r2, r_my_cpumask  /* cache_cpumask */
+	}
+	{
+	 shl16insli r1, r1, hw0(HV_FLUSH_EVICT_L2)
+	 move r3, zero  /* tlb_va */
+	}
+	{
+	 move r4, zero  /* tlb_length */
+	 move r5, zero  /* tlb_pgsize */
+	}
+	{
+	 move r6, zero  /* tlb_cpumask */
+	 move r7, zero  /* asids */
+	}
+	{
+	 move r8, zero  /* asidcount */
+	 jal _hv_flush_remote
+	}
+	bnez r0, 1f
+
+	/* Now install the new page table. */
+	{
+	 move r0, r_context
+	 move r1, r_access
+	}
+	{
+	 move r2, r_asid
+	 moveli r3, HV_CTX_DIRECTIO | CTX_PAGE_FLAG
+	}
+	jal _hv_install_context
+	bnez r0, 1f
+
+	/* Finally, flush the TLB. */
+	{
+	 movei r0, 0   /* preserve_global */
+	 jal hv_flush_all
+	}
+
+1:	/* Restore the callee-saved registers and return. */
+	addli lr, sp, FRAME_SIZE
+	{
+	 ld lr, lr
+	 addli r_tmp, sp, FRAME_R30
+	}
+	{
+	 ld r30, r_tmp
+	 addli r_tmp, sp, FRAME_R31
+	}
+	{
+	 ld r31, r_tmp
+	 addli r_tmp, sp, FRAME_R32
+	}
+	{
+	 ld r32, r_tmp
+	 addi sp, sp, FRAME_SIZE
+	}
+	jrp lr
+	STD_ENDPROC(flush_and_install_context)
diff --git a/kernel/arch/tile/mm/mmap.c b/kernel/arch/tile/mm/mmap.c
new file mode 100644
index 000000000..851a94e6a
--- /dev/null
+++ b/kernel/arch/tile/mm/mmap.c
@@ -0,0 +1,93 @@
+/*
+ * Copyright 2010 Tilera Corporation. All Rights Reserved.
+ *
+ *   This program is free software; you can redistribute it and/or
+ *   modify it under the terms of the GNU General Public License
+ *   as published by the Free Software Foundation, version 2.
+ *
+ *   This program is distributed in the hope that it will be useful, but
+ *   WITHOUT ANY WARRANTY; without even the implied warranty of
+ *   MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
+ *   NON INFRINGEMENT.  See the GNU General Public License for
+ *   more details.
+ *
+ * Taken from the i386 architecture and simplified.
+ */
+
+#include <linux/mm.h>
+#include <linux/random.h>
+#include <linux/limits.h>
+#include <linux/sched.h>
+#include <linux/mman.h>
+#include <linux/compat.h>
+
+/*
+ * Top of mmap area (just below the process stack).
+ *
+ * Leave an at least ~128 MB hole.
+ */
+#define MIN_GAP (128*1024*1024)
+#define MAX_GAP (TASK_SIZE/6*5)
+
+static inline unsigned long mmap_base(struct mm_struct *mm)
+{
+	unsigned long gap = rlimit(RLIMIT_STACK);
+	unsigned long random_factor = 0;
+
+	if (current->flags & PF_RANDOMIZE)
+		random_factor = get_random_int() % (1024*1024);
+
+	if (gap < MIN_GAP)
+		gap = MIN_GAP;
+	else if (gap > MAX_GAP)
+		gap = MAX_GAP;
+
+	return PAGE_ALIGN(TASK_SIZE - gap - random_factor);
+}
+
+/*
+ * This function, called very early during the creation of a new
+ * process VM image, sets up which VM layout function to use:
+ */
+void arch_pick_mmap_layout(struct mm_struct *mm)
+{
+#if !defined(__tilegx__)
+	int is_32bit = 1;
+#elif defined(CONFIG_COMPAT)
+	int is_32bit = is_compat_task();
+#else
+	int is_32bit = 0;
+#endif
+	unsigned long random_factor = 0UL;
+
+	/*
+	 *  8 bits of randomness in 32bit mmaps, 24 address space bits
+	 * 12 bits of randomness in 64bit mmaps, 28 address space bits
+	 */
+	if (current->flags & PF_RANDOMIZE) {
+		if (is_32bit)
+			random_factor = get_random_int() % (1<<8);
+		else
+			random_factor = get_random_int() % (1<<12);
+
+		random_factor <<= PAGE_SHIFT;
+	}
+
+	/*
+	 * Use standard layout if the expected stack growth is unlimited
+	 * or we are running native 64 bits.
+	 */
+	if (rlimit(RLIMIT_STACK) == RLIM_INFINITY) {
+		mm->mmap_base = TASK_UNMAPPED_BASE + random_factor;
+		mm->get_unmapped_area = arch_get_unmapped_area;
+	} else {
+		mm->mmap_base = mmap_base(mm);
+		mm->get_unmapped_area = arch_get_unmapped_area_topdown;
+	}
+}
+
+unsigned long arch_randomize_brk(struct mm_struct *mm)
+{
+	unsigned long range_end = mm->brk + 0x02000000;
+	return randomize_range(mm->brk, range_end, 0) ? : mm->brk;
+}
diff --git a/kernel/arch/tile/mm/pgtable.c b/kernel/arch/tile/mm/pgtable.c
new file mode 100644
index 000000000..7bf2491a9
--- /dev/null
+++ b/kernel/arch/tile/mm/pgtable.c
@@ -0,0 +1,584 @@
+/*
+ * Copyright 2010 Tilera Corporation. All Rights Reserved.
+ *
+ *   This program is free software; you can redistribute it and/or
+ *   modify it under the terms of the GNU General Public License
+ *   as published by the Free Software Foundation, version 2.
+ *
+ *   This program is distributed in the hope that it will be useful, but
+ *   WITHOUT ANY WARRANTY; without even the implied warranty of
+ *   MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
+ *   NON INFRINGEMENT.  See the GNU General Public License for
+ *   more details.
+ */
+
+#include <linux/sched.h>
+#include <linux/kernel.h>
+#include <linux/errno.h>
+#include <linux/mm.h>
+#include <linux/swap.h>
+#include <linux/highmem.h>
+#include <linux/slab.h>
+#include <linux/pagemap.h>
+#include <linux/spinlock.h>
+#include <linux/cpumask.h>
+#include <linux/module.h>
+#include <linux/io.h>
+#include <linux/vmalloc.h>
+#include <linux/smp.h>
+
+#include <asm/pgtable.h>
+#include <asm/pgalloc.h>
+#include <asm/fixmap.h>
+#include <asm/tlb.h>
+#include <asm/tlbflush.h>
+#include <asm/homecache.h>
+
+#define K(x) ((x) << (PAGE_SHIFT-10))
+
+/*
+ * The normal show_free_areas() is too verbose on Tile, with dozens
+ * of processors and often four NUMA zones each with high and lowmem.
+ */
+void show_mem(unsigned int filter)
+{
+	struct zone *zone;
+
+	pr_err("Active:%lu inactive:%lu dirty:%lu writeback:%lu unstable:%lu free:%lu\n slab:%lu mapped:%lu pagetables:%lu bounce:%lu pagecache:%lu swap:%lu\n",
+	       (global_page_state(NR_ACTIVE_ANON) +
+		global_page_state(NR_ACTIVE_FILE)),
+	       (global_page_state(NR_INACTIVE_ANON) +
+		global_page_state(NR_INACTIVE_FILE)),
+	       global_page_state(NR_FILE_DIRTY),
+	       global_page_state(NR_WRITEBACK),
+	       global_page_state(NR_UNSTABLE_NFS),
+	       global_page_state(NR_FREE_PAGES),
+	       (global_page_state(NR_SLAB_RECLAIMABLE) +
+		global_page_state(NR_SLAB_UNRECLAIMABLE)),
+	       global_page_state(NR_FILE_MAPPED),
+	       global_page_state(NR_PAGETABLE),
+	       global_page_state(NR_BOUNCE),
+	       global_page_state(NR_FILE_PAGES),
+	       get_nr_swap_pages());
+
+	for_each_zone(zone) {
+		unsigned long flags, order, total = 0, largest_order = -1;
+
+		if (!populated_zone(zone))
+			continue;
+
+		spin_lock_irqsave(&zone->lock, flags);
+		for (order = 0; order < MAX_ORDER; order++) {
+			int nr = zone->free_area[order].nr_free;
+			total += nr << order;
+			if (nr)
+				largest_order = order;
+		}
+		spin_unlock_irqrestore(&zone->lock, flags);
+		pr_err("Node %d %7s: %lukB (largest %luKb)\n",
+		       zone_to_nid(zone), zone->name,
+		       K(total), largest_order ? K(1UL) << largest_order : 0);
+	}
+}
+
+/**
+ * shatter_huge_page() - ensure a given address is mapped by a small page.
+ *
+ * This function converts a huge PTE mapping kernel LOWMEM into a bunch
+ * of small PTEs with the same caching.  No cache flush required, but we
+ * must do a global TLB flush.
+ *
+ * Any caller that wishes to modify a kernel mapping that might
+ * have been made with a huge page should call this function,
+ * since doing so properly avoids race conditions with installing the
+ * newly-shattered page and then flushing all the TLB entries.
+ *
+ * @addr: Address at which to shatter any existing huge page.
+ */
+void shatter_huge_page(unsigned long addr)
+{
+	pgd_t *pgd;
+	pud_t *pud;
+	pmd_t *pmd;
+	unsigned long flags = 0;  /* happy compiler */
+#ifdef __PAGETABLE_PMD_FOLDED
+	struct list_head *pos;
+#endif
+
+	/* Get a pointer to the pmd entry that we need to change. */
+	addr &= HPAGE_MASK;
+	BUG_ON(pgd_addr_invalid(addr));
+	BUG_ON(addr < PAGE_OFFSET);  /* only for kernel LOWMEM */
+	pgd = swapper_pg_dir + pgd_index(addr);
+	pud = pud_offset(pgd, addr);
+	BUG_ON(!pud_present(*pud));
+	pmd = pmd_offset(pud, addr);
+	BUG_ON(!pmd_present(*pmd));
+	if (!pmd_huge_page(*pmd))
+		return;
+
+	spin_lock_irqsave(&init_mm.page_table_lock, flags);
+	if (!pmd_huge_page(*pmd)) {
+		/* Lost the race to convert the huge page. */
+		spin_unlock_irqrestore(&init_mm.page_table_lock, flags);
+		return;
+	}
+
+	/* Shatter the huge page into the preallocated L2 page table. */
+	pmd_populate_kernel(&init_mm, pmd, get_prealloc_pte(pmd_pfn(*pmd)));
+
+#ifdef __PAGETABLE_PMD_FOLDED
+	/* Walk every pgd on the system and update the pmd there. */
+	spin_lock(&pgd_lock);
+	list_for_each(pos, &pgd_list) {
+		pmd_t *copy_pmd;
+		pgd = list_to_pgd(pos) + pgd_index(addr);
+		pud = pud_offset(pgd, addr);
+		copy_pmd = pmd_offset(pud, addr);
+		__set_pmd(copy_pmd, *pmd);
+	}
+	spin_unlock(&pgd_lock);
+#endif
+
+	/* Tell every cpu to notice the change. */
+	flush_remote(0, 0, NULL, addr, HPAGE_SIZE, HPAGE_SIZE,
+		     cpu_possible_mask, NULL, 0);
+
+	/* Hold the lock until the TLB flush is finished to avoid races. */
+	spin_unlock_irqrestore(&init_mm.page_table_lock, flags);
+}
+
+/*
+ * List of all pgd's needed so it can invalidate entries in both cached
+ * and uncached pgd's. This is essentially codepath-based locking
+ * against pageattr.c; it is the unique case in which a valid change
+ * of kernel pagetables can't be lazily synchronized by vmalloc faults.
+ * vmalloc faults work because attached pagetables are never freed.
+ *
+ * The lock is always taken with interrupts disabled, unlike on x86
+ * and other platforms, because we need to take the lock in
+ * shatter_huge_page(), which may be called from an interrupt context.
+ * We are not at risk from the tlbflush IPI deadlock that was seen on
+ * x86, since we use the flush_remote() API to have the hypervisor do
+ * the TLB flushes regardless of irq disabling.
+ */
+DEFINE_SPINLOCK(pgd_lock);
+LIST_HEAD(pgd_list);
+
+static inline void pgd_list_add(pgd_t *pgd)
+{
+	list_add(pgd_to_list(pgd), &pgd_list);
+}
+
+static inline void pgd_list_del(pgd_t *pgd)
+{
+	list_del(pgd_to_list(pgd));
+}
+
+#define KERNEL_PGD_INDEX_START pgd_index(PAGE_OFFSET)
+#define KERNEL_PGD_PTRS (PTRS_PER_PGD - KERNEL_PGD_INDEX_START)
+
+static void pgd_ctor(pgd_t *pgd)
+{
+	unsigned long flags;
+
+	memset(pgd, 0, KERNEL_PGD_INDEX_START*sizeof(pgd_t));
+	spin_lock_irqsave(&pgd_lock, flags);
+
+#ifndef __tilegx__
+	/*
+	 * Check that the user interrupt vector has no L2.
+	 * It never should for the swapper, and new page tables
+	 * should always start with an empty user interrupt vector.
+	 */
+	BUG_ON(((u64 *)swapper_pg_dir)[pgd_index(MEM_USER_INTRPT)] != 0);
+#endif
+
+	memcpy(pgd + KERNEL_PGD_INDEX_START,
+	       swapper_pg_dir + KERNEL_PGD_INDEX_START,
+	       KERNEL_PGD_PTRS * sizeof(pgd_t));
+
+	pgd_list_add(pgd);
+	spin_unlock_irqrestore(&pgd_lock, flags);
+}
+
+static void pgd_dtor(pgd_t *pgd)
+{
+	unsigned long flags; /* can be called from interrupt context */
+
+	spin_lock_irqsave(&pgd_lock, flags);
+	pgd_list_del(pgd);
+	spin_unlock_irqrestore(&pgd_lock, flags);
+}
+
+pgd_t *pgd_alloc(struct mm_struct *mm)
+{
+	pgd_t *pgd = kmem_cache_alloc(pgd_cache, GFP_KERNEL);
+	if (pgd)
+		pgd_ctor(pgd);
+	return pgd;
+}
+
+void pgd_free(struct mm_struct *mm, pgd_t *pgd)
+{
+	pgd_dtor(pgd);
+	kmem_cache_free(pgd_cache, pgd);
+}
+
+
+#define L2_USER_PGTABLE_PAGES (1 << L2_USER_PGTABLE_ORDER)
+
+struct page *pgtable_alloc_one(struct mm_struct *mm, unsigned long address,
+			       int order)
+{
+	gfp_t flags = GFP_KERNEL|__GFP_REPEAT|__GFP_ZERO;
+	struct page *p;
+	int i;
+
+	p = alloc_pages(flags, L2_USER_PGTABLE_ORDER);
+	if (p == NULL)
+		return NULL;
+
+	if (!pgtable_page_ctor(p)) {
+		__free_pages(p, L2_USER_PGTABLE_ORDER);
+		return NULL;
+	}
+
+	/*
+	 * Make every page have a page_count() of one, not just the first.
+	 * We don't use __GFP_COMP since it doesn't look like it works
+	 * correctly with tlb_remove_page().
+	 */
+	for (i = 1; i < order; ++i) {
+		init_page_count(p+i);
+		inc_zone_page_state(p+i, NR_PAGETABLE);
+	}
+
+	return p;
+}
+
+/*
+ * Free page immediately (used in __pte_alloc if we raced with another
+ * process).  We have to correct whatever pte_alloc_one() did before
+ * returning the pages to the allocator.
+ */
+void pgtable_free(struct mm_struct *mm, struct page *p, int order)
+{
+	int i;
+
+	pgtable_page_dtor(p);
+	__free_page(p);
+
+	for (i = 1; i < order; ++i) {
+		__free_page(p+i);
+		dec_zone_page_state(p+i, NR_PAGETABLE);
+	}
+}
+
+void __pgtable_free_tlb(struct mmu_gather *tlb, struct page *pte,
+			unsigned long address, int order)
+{
+	int i;
+
+	pgtable_page_dtor(pte);
+	tlb_remove_page(tlb, pte);
+
+	for (i = 1; i < order; ++i) {
+		tlb_remove_page(tlb, pte + i);
+		dec_zone_page_state(pte + i, NR_PAGETABLE);
+	}
+}
+
+#ifndef __tilegx__
+
+/*
+ * FIXME: needs to be atomic vs hypervisor writes.  For now we make the
+ * window of vulnerability a bit smaller by doing an unlocked 8-bit update.
+ */
+int ptep_test_and_clear_young(struct vm_area_struct *vma,
+			      unsigned long addr, pte_t *ptep)
+{
+#if HV_PTE_INDEX_ACCESSED < 8 || HV_PTE_INDEX_ACCESSED >= 16
+# error Code assumes HV_PTE "accessed" bit in second byte
+#endif
+	u8 *tmp = (u8 *)ptep;
+	u8 second_byte = tmp[1];
+	if (!(second_byte & (1 << (HV_PTE_INDEX_ACCESSED - 8))))
+		return 0;
+	tmp[1] = second_byte & ~(1 << (HV_PTE_INDEX_ACCESSED - 8));
+	return 1;
+}
+
+/*
+ * This implementation is atomic vs hypervisor writes, since the hypervisor
+ * always writes the low word (where "accessed" and "dirty" are) and this
+ * routine only writes the high word.
+ */
+void ptep_set_wrprotect(struct mm_struct *mm,
+			unsigned long addr, pte_t *ptep)
+{
+#if HV_PTE_INDEX_WRITABLE < 32
+# error Code assumes HV_PTE "writable" bit in high word
+#endif
+	u32 *tmp = (u32 *)ptep;
+	tmp[1] = tmp[1] & ~(1 << (HV_PTE_INDEX_WRITABLE - 32));
+}
+
+#endif
+
+/*
+ * Return a pointer to the PTE that corresponds to the given
+ * address in the given page table.  A NULL page table just uses
+ * the standard kernel page table; the preferred API in this case
+ * is virt_to_kpte().
+ *
+ * The returned pointer can point to a huge page in other levels
+ * of the page table than the bottom, if the huge page is present
+ * in the page table.  For bottom-level PTEs, the returned pointer
+ * can point to a PTE that is either present or not.
+ */
+pte_t *virt_to_pte(struct mm_struct* mm, unsigned long addr)
+{
+	pgd_t *pgd;
+	pud_t *pud;
+	pmd_t *pmd;
+
+	if (pgd_addr_invalid(addr))
+		return NULL;
+
+	pgd = mm ? pgd_offset(mm, addr) : swapper_pg_dir + pgd_index(addr);
+	pud = pud_offset(pgd, addr);
+	if (!pud_present(*pud))
+		return NULL;
+	if (pud_huge_page(*pud))
+		return (pte_t *)pud;
+	pmd = pmd_offset(pud, addr);
+	if (!pmd_present(*pmd))
+		return NULL;
+	if (pmd_huge_page(*pmd))
+		return (pte_t *)pmd;
+	return pte_offset_kernel(pmd, addr);
+}
+EXPORT_SYMBOL(virt_to_pte);
+
+pte_t *virt_to_kpte(unsigned long kaddr)
+{
+	BUG_ON(kaddr < PAGE_OFFSET);
+	return virt_to_pte(NULL, kaddr);
+}
+EXPORT_SYMBOL(virt_to_kpte);
+
+pgprot_t set_remote_cache_cpu(pgprot_t prot, int cpu)
+{
+	unsigned int width = smp_width;
+	int x = cpu % width;
+	int y = cpu / width;
+	BUG_ON(y >= smp_height);
+	BUG_ON(hv_pte_get_mode(prot) != HV_PTE_MODE_CACHE_TILE_L3);
+	BUG_ON(cpu < 0 || cpu >= NR_CPUS);
+	BUG_ON(!cpu_is_valid_lotar(cpu));
+	return hv_pte_set_lotar(prot, HV_XY_TO_LOTAR(x, y));
+}
+
+int get_remote_cache_cpu(pgprot_t prot)
+{
+	HV_LOTAR lotar = hv_pte_get_lotar(prot);
+	int x = HV_LOTAR_X(lotar);
+	int y = HV_LOTAR_Y(lotar);
+	BUG_ON(hv_pte_get_mode(prot) != HV_PTE_MODE_CACHE_TILE_L3);
+	return x + y * smp_width;
+}
+
+/*
+ * Convert a kernel VA to a PA and homing information.
+ */
+int va_to_cpa_and_pte(void *va, unsigned long long *cpa, pte_t *pte)
+{
+	struct page *page = virt_to_page(va);
+	pte_t null_pte = { 0 };
+
+	*cpa = __pa(va);
+
+	/* Note that this is not writing a page table, just returning a pte. */
+	*pte = pte_set_home(null_pte, page_home(page));
+
+	return 0; /* return non-zero if not hfh? */
+}
+EXPORT_SYMBOL(va_to_cpa_and_pte);
+
+void __set_pte(pte_t *ptep, pte_t pte)
+{
+#ifdef __tilegx__
+	*ptep = pte;
+#else
+# if HV_PTE_INDEX_PRESENT >= 32 || HV_PTE_INDEX_MIGRATING >= 32
+#  error Must write the present and migrating bits last
+# endif
+	if (pte_present(pte)) {
+		((u32 *)ptep)[1] = (u32)(pte_val(pte) >> 32);
+		barrier();
+		((u32 *)ptep)[0] = (u32)(pte_val(pte));
+	} else {
+		((u32 *)ptep)[0] = (u32)(pte_val(pte));
+		barrier();
+		((u32 *)ptep)[1] = (u32)(pte_val(pte) >> 32);
+	}
+#endif /* __tilegx__ */
+}
+
+void set_pte(pte_t *ptep, pte_t pte)
+{
+	if (pte_present(pte) &&
+	    (!CHIP_HAS_MMIO() || hv_pte_get_mode(pte) != HV_PTE_MODE_MMIO)) {
+		/* The PTE actually references physical memory. */
+		unsigned long pfn = pte_pfn(pte);
+		if (pfn_valid(pfn)) {
+			/* Update the home of the PTE from the struct page. */
+			pte = pte_set_home(pte, page_home(pfn_to_page(pfn)));
+		} else if (hv_pte_get_mode(pte) == 0) {
+			/* remap_pfn_range(), etc, must supply PTE mode. */
+			panic("set_pte(): out-of-range PFN and mode 0\n");
+		}
+	}
+
+	__set_pte(ptep, pte);
+}
+
+/* Can this mm load a PTE with cached_priority set? */
+static inline int mm_is_priority_cached(struct mm_struct *mm)
+{
+	return mm->context.priority_cached != 0;
+}
+
+/*
+ * Add a priority mapping to an mm_context and
+ * notify the hypervisor if this is the first one.
+ */
+void start_mm_caching(struct mm_struct *mm)
+{
+	if (!mm_is_priority_cached(mm)) {
+		mm->context.priority_cached = -1UL;
+		hv_set_caching(-1UL);
+	}
+}
+
+/*
+ * Validate and return the priority_cached flag.  We know if it's zero
+ * that we don't need to scan, since we immediately set it non-zero
+ * when we first consider a MAP_CACHE_PRIORITY mapping.
+ *
+ * We only _try_ to acquire the mmap_sem semaphore; if we can't acquire it,
+ * since we're in an interrupt context (servicing switch_mm) we don't
+ * worry about it and don't unset the "priority_cached" field.
+ * Presumably we'll come back later and have more luck and clear
+ * the value then; for now we'll just keep the cache marked for priority.
+ */
+static unsigned long update_priority_cached(struct mm_struct *mm)
+{
+	if (mm->context.priority_cached && down_write_trylock(&mm->mmap_sem)) {
+		struct vm_area_struct *vm;
+		for (vm = mm->mmap; vm; vm = vm->vm_next) {
+			if (hv_pte_get_cached_priority(vm->vm_page_prot))
+				break;
+		}
+		if (vm == NULL)
+			mm->context.priority_cached = 0;
+		up_write(&mm->mmap_sem);
+	}
+	return mm->context.priority_cached;
+}
+
+/* Set caching correctly for an mm that we are switching to. */
+void check_mm_caching(struct mm_struct *prev, struct mm_struct *next)
+{
+	if (!mm_is_priority_cached(next)) {
+		/*
+		 * If the new mm doesn't use priority caching, just see if we
+		 * need the hv_set_caching(), or can assume it's already zero.
+		 */
+		if (mm_is_priority_cached(prev))
+			hv_set_caching(0);
+	} else {
+		hv_set_caching(update_priority_cached(next));
+	}
+}
+
+#if CHIP_HAS_MMIO()
+
+/* Map an arbitrary MMIO address, homed according to pgprot, into VA space. */
+void __iomem *ioremap_prot(resource_size_t phys_addr, unsigned long size,
+			   pgprot_t home)
+{
+	void *addr;
+	struct vm_struct *area;
+	unsigned long offset, last_addr;
+	pgprot_t pgprot;
+
+	/* Don't allow wraparound or zero size */
+	last_addr = phys_addr + size - 1;
+	if (!size || last_addr < phys_addr)
+		return NULL;
+
+	/* Create a read/write, MMIO VA mapping homed at the requested shim. */
+	pgprot = PAGE_KERNEL;
+	pgprot = hv_pte_set_mode(pgprot, HV_PTE_MODE_MMIO);
+	pgprot = hv_pte_set_lotar(pgprot, hv_pte_get_lotar(home));
+
+	/*
+	 * Mappings have to be page-aligned
+	 */
+	offset = phys_addr & ~PAGE_MASK;
+	phys_addr &= PAGE_MASK;
+	size = PAGE_ALIGN(last_addr+1) - phys_addr;
+
+	/*
+	 * Ok, go for it..
+	 */
+	area = get_vm_area(size, VM_IOREMAP /* | other flags? */);
+	if (!area)
+		return NULL;
+	area->phys_addr = phys_addr;
+	addr = area->addr;
+	if (ioremap_page_range((unsigned long)addr, (unsigned long)addr + size,
+			       phys_addr, pgprot)) {
+		free_vm_area(area);
+		return NULL;
+	}
+	return (__force void __iomem *) (offset + (char *)addr);
+}
+EXPORT_SYMBOL(ioremap_prot);
+
+/* Unmap an MMIO VA mapping. */
+void iounmap(volatile void __iomem *addr_in)
+{
+	volatile void __iomem *addr = (volatile void __iomem *)
+		(PAGE_MASK & (unsigned long __force)addr_in);
+#if 1
+	vunmap((void * __force)addr);
+#else
+	/* x86 uses this complicated flow instead of vunmap().  Is
+	 * there any particular reason we should do the same? */
+	struct vm_struct *p, *o;
+
+	/* Use the vm area unlocked, assuming the caller
+	   ensures there isn't another iounmap for the same address
+	   in parallel. Reuse of the virtual address is prevented by
+	   leaving it in the global lists until we're done with it.
+	   cpa takes care of the direct mappings. */
+	p = find_vm_area((void *)addr);
+
+	if (!p) {
+		pr_err("iounmap: bad address %p\n", addr);
+		dump_stack();
+		return;
+	}
+
+	/* Finally remove it */
+	o = remove_vm_area((void *)addr);
+	BUG_ON(p != o || o == NULL);
+	kfree(p);
+#endif
+}
+EXPORT_SYMBOL(iounmap);
+
+#endif /* CHIP_HAS_MMIO() */