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

[James Y Knight via cfe-commits]

================
@@ -124,10 +143,11 @@ _mm_cvtm64_si64(__m64 __m)
 ///    written to the upper 32 bits of the result.
 /// \returns A 64-bit integer vector of [8 x i8] containing the converted
 ///    values.
-static __inline__ __m64 __DEFAULT_FN_ATTRS
+static __inline__ __m64 __DEFAULT_FN_ATTRS_SSE2
 _mm_packs_pi16(__m64 __m1, __m64 __m2)
 {
-    return (__m64)__builtin_ia32_packsswb((__v4hi)__m1, (__v4hi)__m2);
+    return __extract2_32(__builtin_ia32_packsswb128((__v8hi)__anyext128(__m1),
----------------
jyknight wrote:

So, the current version assembles to:
```
       0: 66 0f 63 c1                   packsswb        %xmm1, %xmm0
       4: 66 0f 70 c0 e8                pshufd  $0xe8, %xmm0, %xmm0     # xmm0 = xmm0[0,2,2,3]
       9: c3                            retq
```

You're suggesting to instead shuffle the inputs, like:
```
return __trunc64(__builtin_ia32_packsswb128(
    __builtin_shufflevector(__m1, __m2, 0, 1), (__v8hi){}));
```

I agree, that's better. Saves 1 byte of code, and also less register pressure.
```
       0: 66 0f 6c c1                   punpcklqdq      %xmm1, %xmm0    # xmm0 = xmm0[0],xmm1[0]
       4: 66 0f 63 c0                   packsswb        %xmm0, %xmm0
       8: c3                            retq
```

Done -- eliminated all uses of `__extract2_32`.

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