[Support] Fix xxh3_128bits for Win32 builds after #95863

[MaskRay]

__emulu is used without including intrin.h. Actually, it's better to
rely on compiler optimizations. In this LLVM copy, we try to eliminate
unneceeded workarounds for old compilers.

[llvmbot]

@llvm/pr-subscribers-llvm-support

Author: Fangrui Song (MaskRay)

Changes

__emulu is used without including intrin.h. Actually, it's better to
rely on compiler optimizations. In this LLVM copy, we try to eliminate
unneceeded workarounds for old compilers.

---

Full diff: https://github.com/llvm/llvm-project/pull/96931.diff

1 Files Affected:

• (modified) llvm/lib/Support/xxhash.cpp (-11)

diff --git a/llvm/lib/Support/xxhash.cpp b/llvm/lib/Support/xxhash.cpp
index a0803297555ce..607789b391381 100644
--- a/llvm/lib/Support/xxhash.cpp
+++ b/llvm/lib/Support/xxhash.cpp
@@ -453,18 +453,7 @@ uint64_t llvm::xxh3_64bits(ArrayRef<uint8_t> data) {
 #define XXH_rotl64(x, r) (((x) << (r)) | ((x) >> (64 - (r))))
 #endif
 
-#if defined(_MSC_VER) && defined(_M_IX86)
-#define XXH_mult32to64(x, y) __emulu((unsigned)(x), (unsigned)(y))
-#else
-/*
- * Downcast + upcast is usually better than masking on older compilers like
- * GCC 4.2 (especially 32-bit ones), all without affecting newer compilers.
- *
- * The other method, (x & 0xFFFFFFFF) * (y & 0xFFFFFFFF), will AND both operands
- * and perform a full 64x64 multiply -- entirely redundant on 32-bit.
- */
 #define XXH_mult32to64(x, y) ((uint64_t)(uint32_t)(x) * (uint64_t)(uint32_t)(y))
-#endif
 
 /*!
  * @brief Calculates a 64->128-bit long multiply.

[dstutt]

Seems reasonable to me

[dukebw]

Thank you for the fix! For my education: is there a workflow I can monitor that would catch this regression?

[MaskRay]

Thank you for the fix! For my education: is there a workflow I can monitor that would catch this regression?

IIUC there is no bot for 32-bit Windows. In addition, Windows 11 has removed 32-bit support.
Therefore, the condition #if defined(_MSC_VER) && defined(_M_IX86) makes the failure unlikely caught by build bots. I am actually curious how you noticed the breakage due to xxh3_128bits :)