From: "Rémi Bernon" <rbernon@codeweavers.com>
Subject: [PATCH 1/4] msvcrt: Improve memset performance using overlapping stores.
Message-Id: <20210913122341.2109469-1-rbernon@codeweavers.com>
Date: Mon, 13 Sep 2021 14:23:38 +0200

For n larger than 16 we store 16 bytes on each end of the buffer,
eventually overlapping, and then 16 additional bytes for n > 32.

Then we can find a 32-byte aligned range overlapping the remaining part
of the destination buffer, which is filled 32 bytes at a time in a loop.

Signed-off-by: Rémi Bernon <rbernon@codeweavers.com>
---

So, this is what I was thinking instead of having fully specialized
assembly versions.

Overall I believe the performance should be better than SSE2 for very
small sizes and very large sizes (when ERMS kicks in), but a bit worse
for 128 <= n <= 1024.

I also don't even think the last patch is really useful, it only helps
improving performance for the intermediate sizes, and I would think
instead that the ERMS path could cover them instead once it is good
enough on most CPUs for that range of size.

 dlls/msvcrt/string.c | 60 +++++++++++++++++++++++++++++++++++++++++---
 1 file changed, 56 insertions(+), 4 deletions(-)

diff --git a/dlls/msvcrt/string.c b/dlls/msvcrt/string.c
index 4d09405094d..3a7312572ab 100644
--- a/dlls/msvcrt/string.c
+++ b/dlls/msvcrt/string.c
@@ -2855,13 +2855,65 @@ void * __cdecl memcpy(void *dst, const void *src, size_t n)
     return memmove(dst, src, n);
 }
 
+static void memset_aligned_32(unsigned char *d, uint64_t v, size_t n)
+{
+    while (n >= 32)
+    {
+        *(uint64_t*)(d + n - 32) = v;
+        *(uint64_t*)(d + n - 24) = v;
+        *(uint64_t*)(d + n - 16) = v;
+        *(uint64_t*)(d + n -  8) = v;
+        n -= 32;
+    }
+}
+
 /*********************************************************************
  *		    memset (MSVCRT.@)
  */
-void* __cdecl memset(void *dst, int c, size_t n)
-{
-    volatile unsigned char *d = dst;  /* avoid gcc optimizations */
-    while (n--) *d++ = c;
+void *__cdecl memset(void *dst, int c, size_t n)
+{
+    uint64_t v = 0x101010101010101ull * (unsigned char)c;
+    unsigned char *d = (unsigned char *)dst;
+    size_t a = 0x20 - ((uintptr_t)d & 0x1f);
+
+    if (n >= 16)
+    {
+        *(uint64_t *)(d + 0) = v;
+        *(uint64_t *)(d + 8) = v;
+        *(uint64_t *)(d + n - 16) = v;
+        *(uint64_t *)(d + n - 8) = v;
+        if (n <= 32)  return dst;
+        *(uint64_t *)(d + 16) = v;
+        *(uint64_t *)(d + 24) = v;
+        *(uint64_t *)(d + n - 32) = v;
+        *(uint64_t *)(d + n - 24) = v;
+        if (n <= 64)  return dst;
+        memset_aligned_32(d + a, v, (n - a) & ~0x1f);
+        return dst;
+    }
+    if (n >= 8)
+    {
+        *(uint64_t *)d = v;
+        *(uint64_t *)(d + n - 8) = v;
+        return dst;
+    }
+    if (n >= 4)
+    {
+        *(uint32_t *)d = v;
+        *(uint32_t *)(d + n - 4) = v;
+        return dst;
+    }
+    if (n >= 2)
+    {
+        *(uint16_t *)d = v;
+        *(uint16_t *)(d + n - 2) = v;
+        return dst;
+    }
+    if (n >= 1)
+    {
+        *(uint8_t *)d = v;
+        return dst;
+    }
     return dst;
 }
 

-- 
2.33.0