Use SSE instructions for prefetch when __builtin_prefetch is unavailable

This notably gets prefetch working on MSVC

Implementation note:
https://docs.microsoft.com/en-us/windows/win32/api/winnt/nf-winnt-prefetchcacheline
MSVC does have PreFetchCacheLine, but that would require including <windows.h>
in a header

PiperOrigin-RevId: 449602543
Change-Id: I5e6ca4b7c3d287779aa03c2fd348b41fb65c3680

absl/base/config.h

@@ -837,6 +837,21 @@ static_assert(ABSL_INTERNAL_INLINE_NAMESPACE_STR[0] != 'h' ||
 #define ABSL_INTERNAL_HAS_RTTI 1
 #endif  // !defined(__GNUC__) || defined(__GXX_RTTI)
 
+// ABSL_INTERNAL_HAVE_SSE is used for compile-time detection of SSE support.
+// See https://gcc.gnu.org/onlinedocs/gcc/x86-Options.html for an overview of
+// which architectures support the various x86 instruction sets.
+#ifdef ABSL_INTERNAL_HAVE_SSE
+#error ABSL_INTERNAL_HAVE_SSE cannot be directly set
+#elif defined(__SSE__)
+#define ABSL_INTERNAL_HAVE_SSE 1
+#elif defined(_M_X64) || (defined(_M_IX86_FP) && _M_IX86_FP >= 1)
+// MSVC only defines _M_IX86_FP for x86 32-bit code, and _M_IX86_FP >= 1
+// indicates that at least SSE was targeted with the /arch:SSE option.
+// All x86-64 processors support SSE, so support can be assumed.
+// https://docs.microsoft.com/en-us/cpp/preprocessor/predefined-macros
+#define ABSL_INTERNAL_HAVE_SSE 1
+#endif
+
 // ABSL_INTERNAL_HAVE_SSE2 is used for compile-time detection of SSE2 support.
 // See https://gcc.gnu.org/onlinedocs/gcc/x86-Options.html for an overview of
 // which architectures support the various x86 instruction sets.

absl/base/internal/prefetch.h

@@ -17,6 +17,15 @@
 
 #include "absl/base/config.h"
 
+#ifdef __SSE__
+#include <xmmintrin.h>
+#endif
+
+#if defined(_MSC_VER) && defined(ABSL_INTERNAL_HAVE_SSE)
+#include <intrin.h>
+#pragma intrinsic(_mm_prefetch)
+#endif
+
 // Compatibility wrappers around __builtin_prefetch, to prefetch data
 // for read if supported by the toolchain.
 
@@ -97,6 +106,24 @@ inline void PrefetchNta(const void* addr) {
   // Note: this uses prefetchtnta on Intel.
   __builtin_prefetch(addr, 0, 0);
 }
+
+#elif defined(ABSL_INTERNAL_HAVE_SSE)
+
+#define ABSL_INTERNAL_HAVE_PREFETCH 1
+
+inline void PrefetchT0(const void* addr) {
+  _mm_prefetch(reinterpret_cast<const char*>(addr), _MM_HINT_T0);
+}
+inline void PrefetchT1(const void* addr) {
+  _mm_prefetch(reinterpret_cast<const char*>(addr), _MM_HINT_T1);
+}
+inline void PrefetchT2(const void* addr) {
+  _mm_prefetch(reinterpret_cast<const char*>(addr), _MM_HINT_T2);
+}
+inline void PrefetchNta(const void* addr) {
+  _mm_prefetch(reinterpret_cast<const char*>(addr), _MM_HINT_NTA);
+}
+
 #else
 inline void PrefetchT0(const void*) {}
 inline void PrefetchT1(const void*) {}