[clang] [llvm] Clang: convert `__m64` intrinsics to unconditionally use SSE2 instead of MMX. (PR #96540)

[James Y Knight via cfe-commits]

================
@@ -2502,10 +2509,25 @@ _mm_mulhi_pu16(__m64 __a, __m64 __b)
 ///    A pointer to a 64-bit memory location that will receive the conditionally
 ///    copied integer values. The address of the memory location does not have
 ///    to be aligned.
-static __inline__ void __DEFAULT_FN_ATTRS_MMX
+static __inline__ void __DEFAULT_FN_ATTRS_SSE2
 _mm_maskmove_si64(__m64 __d, __m64 __n, char *__p)
 {
-  __builtin_ia32_maskmovq((__v8qi)__d, (__v8qi)__n, __p);
+  // This is complex, because we need to support the case where __p is pointing
+  // within the last 15 to 8 bytes of a page. In that case, using a 128-bit
+  // write might cause a trap where a 64-bit maskmovq would not. (Memory
+  // locations not selected by the mask bits might still cause traps.)
+  __m128i __d128  = __anyext128(__d);
+  __m128i __n128  = __zext128(__n);
+  if (((__SIZE_TYPE__)__p & 0xfff) >= 4096-15 &&
+      ((__SIZE_TYPE__)__p & 0xfff) <= 4096-8) {
----------------
jyknight wrote:

I believe it's correct as written: we need to ensure that we cross a potential page-protection boundary in exactly the same situations as we would've originally.

Since we're now executing a 16-byte write instead of the specified 8-byte write, that means we need to back up by 8 bytes when we're at offsets 15, 14, 13, 12, 11, 10, 9, and 8 before the end of the page.

At 16 bytes before, we're guaranteed to be within the page for both writes, and at 7 bytes before, we're guaranteed to cross potential-boundary for both.

https://github.com/llvm/llvm-project/pull/96540