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

[James Y Knight via cfe-commits]

================
@@ -2108,9 +2106,8 @@ static __inline__ __m128i __DEFAULT_FN_ATTRS _mm_add_epi32(__m128i __a,
 /// \param __b
 ///    A 64-bit integer.
 /// \returns A 64-bit integer containing the sum of both parameters.
-static __inline__ __m64 __DEFAULT_FN_ATTRS_MMX _mm_add_si64(__m64 __a,
-                                                            __m64 __b) {
-  return (__m64)__builtin_ia32_paddq((__v1di)__a, (__v1di)__b);
+static __inline__ __m64 __DEFAULT_FN_ATTRS _mm_add_si64(__m64 __a, __m64 __b) {
+  return (__m64)(((unsigned long long)__a) + ((unsigned long long)__b));
----------------
jyknight wrote:

This must use an unsigned add, not a signed add, because wraparound must be defined-behavior.

The same rationale applies to the other addition/subtraction cases you've noted in various places below, so I'm going to just mark those conversations as resolved instead of responding to each separately.

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