summaryrefslogtreecommitdiffstats
path: root/qemu/roms/qemu-palcode/memcpy.c
diff options
context:
space:
mode:
authorYang Zhang <yang.z.zhang@intel.com>2015-08-28 09:58:54 +0800
committerYang Zhang <yang.z.zhang@intel.com>2015-09-01 12:44:00 +0800
commite44e3482bdb4d0ebde2d8b41830ac2cdb07948fb (patch)
tree66b09f592c55df2878107a468a91d21506104d3f /qemu/roms/qemu-palcode/memcpy.c
parent9ca8dbcc65cfc63d6f5ef3312a33184e1d726e00 (diff)
Add qemu 2.4.0
Change-Id: Ic99cbad4b61f8b127b7dc74d04576c0bcbaaf4f5 Signed-off-by: Yang Zhang <yang.z.zhang@intel.com>
Diffstat (limited to 'qemu/roms/qemu-palcode/memcpy.c')
-rw-r--r--qemu/roms/qemu-palcode/memcpy.c160
1 files changed, 160 insertions, 0 deletions
diff --git a/qemu/roms/qemu-palcode/memcpy.c b/qemu/roms/qemu-palcode/memcpy.c
new file mode 100644
index 000000000..b6bbb7434
--- /dev/null
+++ b/qemu/roms/qemu-palcode/memcpy.c
@@ -0,0 +1,160 @@
+/*
+ * linux/arch/alpha/lib/memcpy.c
+ *
+ * Copyright (C) 1995 Linus Torvalds
+ */
+
+/*
+ * This is a reasonably optimized memcpy() routine.
+ */
+
+#include <string.h>
+
+/*
+ * Note that the C code is written to be optimized into good assembly. However,
+ * at this point gcc is unable to sanely compile "if (n >= 0)", resulting in a
+ * explicit compare against 0 (instead of just using the proper "blt reg, xx" or
+ * "bge reg, xx"). I hope alpha-gcc will be fixed to notice this eventually..
+ */
+
+/*
+ * This should be done in one go with ldq_u*2/mask/stq_u. Do it
+ * with a macro so that we can fix it up later..
+ */
+#define ALIGN_DEST_TO8_UP(d,s,n) \
+ while (d & 7) { \
+ if (n <= 0) return; \
+ n--; \
+ *(char *) d = *(char *) s; \
+ d++; s++; \
+ }
+#define ALIGN_DEST_TO8_DN(d,s,n) \
+ while (d & 7) { \
+ if (n <= 0) return; \
+ n--; \
+ d--; s--; \
+ *(char *) d = *(char *) s; \
+ }
+
+/*
+ * This should similarly be done with ldq_u*2/mask/stq. The destination
+ * is aligned, but we don't fill in a full quad-word
+ */
+#define DO_REST_UP(d,s,n) \
+ while (n > 0) { \
+ n--; \
+ *(char *) d = *(char *) s; \
+ d++; s++; \
+ }
+#define DO_REST_DN(d,s,n) \
+ while (n > 0) { \
+ n--; \
+ d--; s--; \
+ *(char *) d = *(char *) s; \
+ }
+
+/*
+ * This should be done with ldq/mask/stq. The source and destination are
+ * aligned, but we don't fill in a full quad-word
+ */
+#define DO_REST_ALIGNED_UP(d,s,n) DO_REST_UP(d,s,n)
+#define DO_REST_ALIGNED_DN(d,s,n) DO_REST_DN(d,s,n)
+
+/*
+ * This does unaligned memory copies. We want to avoid storing to
+ * an unaligned address, as that would do a read-modify-write cycle.
+ * We also want to avoid double-reading the unaligned reads.
+ *
+ * Note the ordering to try to avoid load (and address generation) latencies.
+ */
+static inline void __memcpy_unaligned_up (unsigned long d, unsigned long s,
+ long n)
+{
+ ALIGN_DEST_TO8_UP(d,s,n);
+ n -= 8; /* to avoid compare against 8 in the loop */
+ if (n >= 0) {
+ unsigned long low_word, high_word;
+ __asm__("ldq_u %0,%1":"=r" (low_word):"m" (*(unsigned long *) s));
+ do {
+ unsigned long tmp;
+ __asm__("ldq_u %0,%1":"=r" (high_word):"m" (*(unsigned long *)(s+8)));
+ n -= 8;
+ __asm__("extql %1,%2,%0"
+ :"=r" (low_word)
+ :"r" (low_word), "r" (s));
+ __asm__("extqh %1,%2,%0"
+ :"=r" (tmp)
+ :"r" (high_word), "r" (s));
+ s += 8;
+ *(unsigned long *) d = low_word | tmp;
+ d += 8;
+ low_word = high_word;
+ } while (n >= 0);
+ }
+ n += 8;
+ DO_REST_UP(d,s,n);
+}
+
+static inline void __memcpy_unaligned_dn (unsigned long d, unsigned long s,
+ long n)
+{
+ /* I don't understand AXP assembler well enough for this. -Tim */
+ s += n;
+ d += n;
+ while (n--)
+ * (char *) --d = * (char *) --s;
+}
+
+/*
+ * Hmm.. Strange. The __asm__ here is there to make gcc use an integer register
+ * for the load-store. I don't know why, but it would seem that using a floating
+ * point register for the move seems to slow things down (very small difference,
+ * though).
+ *
+ * Note the ordering to try to avoid load (and address generation) latencies.
+ */
+static inline void __memcpy_aligned_up (unsigned long d, unsigned long s,
+ long n)
+{
+ ALIGN_DEST_TO8_UP(d,s,n);
+ n -= 8;
+ while (n >= 0) {
+ unsigned long tmp;
+ __asm__("ldq %0,%1":"=r" (tmp):"m" (*(unsigned long *) s));
+ n -= 8;
+ s += 8;
+ *(unsigned long *) d = tmp;
+ d += 8;
+ }
+ n += 8;
+ DO_REST_ALIGNED_UP(d,s,n);
+}
+static inline void __memcpy_aligned_dn (unsigned long d, unsigned long s,
+ long n)
+{
+ s += n;
+ d += n;
+ ALIGN_DEST_TO8_DN(d,s,n);
+ n -= 8;
+ while (n >= 0) {
+ unsigned long tmp;
+ s -= 8;
+ __asm__("ldq %0,%1":"=r" (tmp):"m" (*(unsigned long *) s));
+ n -= 8;
+ d -= 8;
+ *(unsigned long *) d = tmp;
+ }
+ n += 8;
+ DO_REST_ALIGNED_DN(d,s,n);
+}
+
+void * memcpy(void * dest, const void *src, size_t n)
+{
+ if (!(((unsigned long) dest ^ (unsigned long) src) & 7)) {
+ __memcpy_aligned_up ((unsigned long) dest, (unsigned long) src,
+ n);
+ return dest;
+ }
+ __memcpy_unaligned_up ((unsigned long) dest, (unsigned long) src, n);
+ return dest;
+}