[PATCH v2 resent] msvcrt: Add an ermsb-based memset for x86/64.

[Guillaume Charifi]

Signed-off-by: Guillaume Charifi 
---
 dlls/msvcrt/string.c | 63 ++++++++++++++++++++++++++++++++++++++++++++
 1 file changed, 63 insertions(+)

diff --git a/dlls/msvcrt/string.c b/dlls/msvcrt/string.c
index 4d09405094d..6d1500cb194 100644
--- a/dlls/msvcrt/string.c
+++ b/dlls/msvcrt/string.c
@@ -2527,6 +2527,7 @@ int __cdecl memcmp(const void *ptr1, const void *ptr2, size_t n)
     __ASM_CFI(".cfi_adjust_cfa_offset -8\n\t") \
     "popq " SRC_REG "\n\t" \
     __ASM_CFI(".cfi_adjust_cfa_offset -8\n\t")
+
 #endif

void * __cdecl sse2_memmove(void *dst, const void *src, size_t n);
@@ -2732,6 +2733,64 @@ __ASM_GLOBAL_FUNC( sse2_memmove,
         MEMMOVE_CLEANUP
         "ret" )

+#undef DEST_REG
+#undef SRC_REG
+#undef LEN_REG
+#undef TMP_REG
+
+#ifdef __i386__
+
+#define DEST_REG "%edi"
+#define SRC_REG "%eax"
+#define LEN_REG "%ecx"
+#define TMP_REG
+
+#define MEMSET_INIT \
+    "pushl " DEST_REG "\n\t" \
+    __ASM_CFI(".cfi_adjust_cfa_offset 4\n\t") \
+    "movl 8(%esp), " DEST_REG "\n\t" \
+    "movl 12(%esp), " SRC_REG "\n\t" \
+    "movl 16(%esp), " LEN_REG "\n\t"
+
+#define MEMSET_CLEANUP \
+    "movl 8(%esp), %eax\n\t" \
+    "popl " DEST_REG "\n\t" \
+    __ASM_CFI(".cfi_adjust_cfa_offset -4\n\t")
+
+#else
+
+#define DEST_REG "%rdi"
+#define SRC_REG "%rax"
+#define LEN_REG "%rcx"
+#define TMP_REG "%r9"
+
+#define MEMSET_INIT \
+    "pushq " DEST_REG "\n\t" \
+    __ASM_CFI(".cfi_adjust_cfa_offset 8\n\t") \
+    "movq %rcx, " DEST_REG "\n\t" \
+    "movq %rdx, " SRC_REG "\n\t" \
+    "movq %r8, " LEN_REG "\n\t" \
+    "movq " DEST_REG ", " TMP_REG "\n\t"
+
+#define MEMSET_CLEANUP \
+    "movq " TMP_REG ", %rax\n\t" \
+    "popq " DEST_REG "\n\t" \
+    __ASM_CFI(".cfi_adjust_cfa_offset -8\n\t")
+
+#endif
+
+void * __cdecl ermsb_memset(void *dst, int c, size_t n);
+__ASM_GLOBAL_FUNC( ermsb_memset,
+        MEMSET_INIT
+        "rep stosb\n\t"
+        MEMSET_CLEANUP
+        "ret" )
+
+#undef DEST_REG
+#undef SRC_REG
+#undef LEN_REG
+#undef TMP_REG
+
 #endif

/*********************************************************************
@@ -2860,9 +2919,13 @@ void * __cdecl memcpy(void *dst, const void *src, size_t n)
  */
 void* __cdecl memset(void *dst, int c, size_t n)
 {
+#if defined(__i386__) || defined(__x86_64__)
+    return ermsb_memset(dst, c, n);
+#else
     volatile unsigned char *d = dst;  /* avoid gcc optimizations */
     while (n--) *d++ = c;
     return dst;
+#endif
 }

/*********************************************************************

[Piotr Caban]

Hi Guillaume,

The "rep stosb" can't be used unconditionally since it's very slow on 
some processors.

Rémi sent a similar patch that adds the checks, see:
https://source.winehq.org/patches/data/214107
He also improved the C version of the function:
https://source.winehq.org/patches/data/214106

Is the "improved C" version good enough for your use case?

Thanks,
Piotr