From 208c57f4da8c63f7e0f3b9c069af963e5972d854 Mon Sep 17 00:00:00 2001
From: arrow <130365147+merkalev@users.noreply.github.com>
Date: Tue, 25 Aug 2026 10:55:09 +0700
Subject: [PATCH] Fix macOS Intel wheel: AppleClang rejects '#pragma GCC
 target', use per-function target attribute

---
 .github/workflows/ci.yml |  5 ++++-
 src/v2_simd.hpp          |  2 +-
 src/v2_simd_avx2.cpp     | 28 ++++++++++++++--------------
 3 files changed, 19 insertions(+), 16 deletions(-)

diff --git a/.github/workflows/ci.yml b/.github/workflows/ci.yml
index 09f1118..486bf70 100644
--- a/.github/workflows/ci.yml
+++ b/.github/workflows/ci.yml
@@ -91,7 +91,10 @@ jobs:
     strategy:
       fail-fast: false
       matrix:
-        os: [ubuntu-latest, windows-latest, macos-latest]
+        # macos-15-intel compiles the AVX2 kernels with AppleClang, the only
+        # toolchain that rejects '#pragma GCC target' - it guards the exact
+        # wheel-build configuration used by the macOS Intel release job.
+        os: [ubuntu-latest, windows-latest, macos-latest, macos-15-intel]
     steps:
       - uses: actions/checkout@v7
       - name: Configure MSVC
diff --git a/src/v2_simd.hpp b/src/v2_simd.hpp
index d4bd868..d31b84b 100644
--- a/src/v2_simd.hpp
+++ b/src/v2_simd.hpp
@@ -4,7 +4,7 @@
 // Instruction-set-specific kernels are isolated in dedicated translation
 // units so vector code can never leak into baseline compilation:
 //
-//   v2_simd_avx2.cpp  x86-64 AVX2 kernels (in-source '#pragma GCC target',
+//   v2_simd_avx2.cpp  x86-64 AVX2 kernels (per-function target attribute,
 //                     or MSVC '/arch:AVX2' scoped to that file)
 //   v2_simd_neon.cpp  AArch64 NEON kernels (baseline ISA, no flags needed)
 //   v2_simd_crc.cpp   AArch64 CRC-32 extension (optional, runtime-probed)
diff --git a/src/v2_simd_avx2.cpp b/src/v2_simd_avx2.cpp
index 47c0132..64cd0fd 100644
--- a/src/v2_simd_avx2.cpp
+++ b/src/v2_simd_avx2.cpp
@@ -1,10 +1,13 @@
 // WIMF v2 AVX2 kernels.
 //
-// GCC/Clang enable AVX2 in-source via '#pragma GCC target' so no global
-// compiler flag is needed and other translation units stay baseline-portable.
-// MSVC has no per-function ISA selection, so the build system must compile
-// this file with '/arch:AVX2' (CMake scopes that flag to this file only).
-// The whole unit is skipped unless WIMF_SIMD_ENABLE_AVX2 is defined.
+// GCC, Clang, and AppleClang enable AVX2 per function via
+// '__attribute__((target("avx2")))' so no global compiler flag is needed and
+// other translation units stay baseline-portable. AppleClang rejects the
+// GCC-style '#pragma GCC target' form, which is why the attribute is applied
+// to each kernel directly. MSVC has no per-function ISA selection, so the
+// build system must compile this file with '/arch:AVX2' (CMake scopes that
+// flag to this file only). The whole unit is skipped unless
+// WIMF_SIMD_ENABLE_AVX2 is defined.
 
 #include "v2_simd.hpp"
 
@@ -14,16 +17,17 @@
 #error "WIMF_SIMD_ENABLE_AVX2 requires /arch:AVX2 when compiling with MSVC"
 #endif
 
-#if defined(__GNUC__) || defined(__clang__)
-#pragma GCC push_options
-#pragma GCC target("avx2")
+#if defined(_MSC_VER)
+#define WIMF_AVX2_TARGET
+#else
+#define WIMF_AVX2_TARGET __attribute__((target("avx2")))
 #endif
 
 #include <immintrin.h>
 
 namespace wimf::v2::simd::avx2 {
 
-uint64_t left_filter_cost(const uint8_t* row, size_t width) noexcept {
+WIMF_AVX2_TARGET uint64_t left_filter_cost(const uint8_t* row, size_t width) noexcept {
     if (width == 0) return 0;
     uint64_t cost = row[0] <= 128 ? row[0] : 256u - row[0];
     const __m256i zero = _mm256_setzero_si256();
@@ -50,7 +54,7 @@ uint64_t left_filter_cost(const uint8_t* row, size_t width) noexcept {
     return cost;
 }
 
-void left_filter_emit(const uint8_t* row, uint8_t* out, size_t width) noexcept {
+WIMF_AVX2_TARGET void left_filter_emit(const uint8_t* row, uint8_t* out, size_t width) noexcept {
     if (width == 0) return;
     out[0] = row[0];
     size_t x = 1;
@@ -64,8 +68,4 @@ void left_filter_emit(const uint8_t* row, uint8_t* out, size_t width) noexcept {
 
 }  // namespace wimf::v2::simd::avx2
 
-#if defined(__GNUC__) || defined(__clang__)
-#pragma GCC pop_options
-#endif
-
 #endif  // WIMF_AVX2_KERNELS