[libc-commits] [libc] [libc] Fix compilation when AVX is not available (PR #211726)

[Roland McGrath via libc-commits]

https://github.com/frobtech created https://github.com/llvm/llvm-project/pull/211726

None

>From 763891b5d56eb0052fb3f83eb089fec371793e49 Mon Sep 17 00:00:00 2001
From: Roland McGrath <mcgrathr at google.com>
Date: Thu, 23 Jul 2026 22:28:46 -0700
Subject: [PATCH] [libc] Fix compilation when AVX is not available

---
 libc/src/string/memory_utils/op_x86.h         | 12 +++++-
 .../memory_utils/x86_64/inline_memcpy.h       | 38 +++++++++++--------
 2 files changed, 34 insertions(+), 16 deletions(-)

diff --git a/libc/src/string/memory_utils/op_x86.h b/libc/src/string/memory_utils/op_x86.h
index 3afcbd02bd525..ce9ad7dacc387 100644
--- a/libc/src/string/memory_utils/op_x86.h
+++ b/libc/src/string/memory_utils/op_x86.h
@@ -26,7 +26,10 @@
 
 #if defined(__AVX512BW__) || defined(__AVX512F__) || defined(__AVX2__) ||      \
     defined(__SSE2__)
+#define LIBC_TARGET_HAS_X86_INTRIN 1
 #include <immintrin.h>
+#else
+#define LIBC_TARGET_HAS_X86_INTRIN 0
 #endif
 
 // Define fake functions to prevent the compiler from failing on undefined
@@ -82,7 +85,14 @@ template <typename T> LIBC_INLINE void stream(Ptr dst, T value) {
   store<T>(dst, value);
 #endif
 }
-template <typename T> LIBC_INLINE void fence() { _mm_sfence(); }
+
+template <typename T> LIBC_INLINE void fence() {
+  // This won't be reached when nothing fancy is enabled, but it can be
+  // referenced in if constexpr context, so it's still defined, as a no-op.
+#if LIBC_TARGET_HAS_X86_INTRIN
+  _mm_sfence();
+#endif
+}
 
 ///////////////////////////////////////////////////////////////////////////////
 // Specializations for uint16_t
diff --git a/libc/src/string/memory_utils/x86_64/inline_memcpy.h b/libc/src/string/memory_utils/x86_64/inline_memcpy.h
index e569ecceb80cc..ff6903b8d1a31 100644
--- a/libc/src/string/memory_utils/x86_64/inline_memcpy.h
+++ b/libc/src/string/memory_utils/x86_64/inline_memcpy.h
@@ -8,6 +8,8 @@
 #ifndef LLVM_LIBC_SRC_STRING_MEMORY_UTILS_X86_64_INLINE_MEMCPY_H
 #define LLVM_LIBC_SRC_STRING_MEMORY_UTILS_X86_64_INLINE_MEMCPY_H
 
+#include <stddef.h> // size_t
+
 #include "hdr/stdint_proxy.h"                // SIZE_MAX
 #include "src/__support/macros/attributes.h" // LIBC_INLINE_VAR
 #include "src/__support/macros/is_defined.h"
@@ -16,8 +18,6 @@
 #include "src/string/memory_utils/op_x86.h"
 #include "src/string/memory_utils/utils.h"
 
-#include <stddef.h> // size_t
-
 #ifdef LLVM_LIBC_MEMCPY_X86_USE_ONLY_REPMOVSB
 #error LLVM_LIBC_MEMCPY_X86_USE_ONLY_REPMOVSB is deprecated use LIBC_COPT_MEMCPY_X86_USE_REPMOVSB_FROM_SIZE=0 instead.
 #endif // LLVM_LIBC_MEMCPY_X86_USE_ONLY_REPMOVSB
@@ -27,7 +27,6 @@
 #endif // LLVM_LIBC_MEMCPY_X86_USE_REPMOVSB_FROM_SIZE
 
 namespace LIBC_NAMESPACE_DECL {
-
 namespace x86 {
 
 LIBC_INLINE_VAR constexpr size_t K_ONE_CACHELINE = 64;
@@ -54,6 +53,19 @@ LIBC_INLINE_VAR constexpr size_t K_NTA_THRESHOLD = 0;
 LIBC_INLINE_VAR constexpr size_t K_REP_MOVSB_THRESHOLD =
     LIBC_COPT_MEMCPY_X86_USE_REPMOVSB_FROM_SIZE;
 
+// The code using __m256i is if constexpr unreachable, but not entirely #if'd
+// out to avoid bitrot when compiling all the permutations.  However, the
+// __m256i type doesn't exist at all without AVX and an unknown type name can't
+// appear even in an if constexpr unreachable context.  So define an alias or
+// stub type to use in the AVX-specific code below.
+#ifdef __AVX__
+using m256i = __m256i;
+#else
+struct alignas(32) m256i {
+  uint64_t x[256 / 64];
+};
+#endif
+
 } // namespace x86
 
 [[maybe_unused]] LIBC_INLINE void
@@ -159,12 +171,13 @@ inline_memcpy_x86_avx_ge64_sw_prefetching(Ptr __restrict dst,
     if (count >= x86::K_NTA_THRESHOLD) {
       while (offset + K_THREE_CACHELINES + 64 <= count) {
         for (size_t i = 0; i < 3; ++i, offset += K_ONE_CACHELINE) {
-          generic::stream(dst + offset, generic::load<__m256i>(src + offset));
+          generic::stream(dst + offset,
+                          generic::load<x86::m256i>(src + offset));
           generic::stream(dst + offset + 32,
-                          generic::load<__m256i>(src + offset + 32));
+                          generic::load<x86::m256i>(src + offset + 32));
         }
       }
-      generic::fence<__m256i>();
+      generic::fence<x86::m256i>();
       need_prefetch_run = false;
     }
   }
@@ -190,15 +203,10 @@ inline_memcpy_x86_avx_ge64_sw_prefetching(Ptr __restrict dst,
 
 [[maybe_unused]] LIBC_INLINE void
 inline_memcpy_x86(Ptr __restrict dst, CPtr __restrict src, size_t count) {
-#if defined(__AVX512F__)
-  constexpr size_t VECTOR_SIZE = 64;
-#elif defined(__AVX__)
-  constexpr size_t VECTOR_SIZE = 32;
-#elif defined(__SSE2__)
-  constexpr size_t VECTOR_SIZE = 16;
-#else
-  constexpr size_t VECTOR_SIZE = 8;
-#endif
+  constexpr size_t VECTOR_SIZE = x86::K_AVX512_F ? 64
+                                 : x86::K_AVX    ? 32
+                                 : x86::K_SSE2   ? 16
+                                                 : 8;
   if (count == 0)
     return;
   if (count == 1)