Simon Jakobi pushed to branch wip/sjakobi/T25450-march-native at Glasgow Haskell Compiler / GHC Commits: 35de03e0 by Simon Jakobi at 2026-06-03T11:41:43+02:00 Detect AVX-512 via sysctl on x86_64 macOS On x86_64 macOS the kernel enables AVX-512 XSAVE state lazily: XCR0 reads back with the opmask/ZMM bits clear until a process first faults on an AVX-512 instruction. The XCR0-based check therefore reports a false negative on AVX-512-capable Macs, so -march=native failed to enable -mavx512f and friends there. Query hw.optional.avx512f via sysctlbyname on Apple instead. Checking AVX512F alone suffices; the sub-features (BW/CD/DQ/VL) are still decoded from CPUID leaf 7, which is unaffected by the lazy-state quirk. AVX/AVX2/ FMA detection is left on the XCR0 path, which is correct on macOS (YMM state is enabled eagerly). Reported by @aratamizuki in !16126. Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com> - - - - - 1 changed file: - compiler/cbits/cpu_features_x86.c Changes: ===================================== compiler/cbits/cpu_features_x86.c ===================================== @@ -10,6 +10,10 @@ #include <cpuid.h> #endif +#if defined(__APPLE__) && (defined(__i386__) || defined(__x86_64__)) +#include <sys/sysctl.h> +#endif + enum { GHC_X86_FEAT_SSE2 = 0, GHC_X86_FEAT_SSE3, @@ -70,6 +74,19 @@ static uint64_t ghc_xgetbv0(void) #endif } +#if defined(__APPLE__) && (defined(__i386__) || defined(__x86_64__)) +/* Query a macOS CPU-capability sysctl, e.g. "hw.optional.avx512f". */ +static int ghc_macos_sysctl_flag(const char *name) +{ + int result = 0; + size_t len = sizeof(result); + if (sysctlbyname(name, &result, &len, NULL, 0) != 0) { + return 0; + } + return result != 0; +} +#endif + HsWord64 ghc_detect_x86_cpu_features(void) { HsWord64 feats = 0; @@ -107,6 +124,20 @@ HsWord64 ghc_detect_x86_cpu_features(void) avx512_usable = ((xcr0 & 0xE6u) == 0xE6u); /* XMM+YMM+opmask+ZMM */ } +#if defined(__APPLE__) + /* On x86_64 macOS the kernel enables AVX-512 XSAVE state lazily: XCR0 + reads back with the opmask/ZMM bits clear until a process first faults + on an AVX-512 instruction, so the XCR0 check above is a false negative + on AVX-512-capable Macs. Use the OS feature query instead. Checking + AVX512F alone suffices here; the AVX-512 sub-features (BW/CD/DQ/VL) are + still decoded from CPUID leaf 7 below. + + Refs: + https://zenn.dev/mod_poppo/articles/detect-processor-features-x86?locale=en#... + https://github.com/minoki/haskell-cpu-features */ + avx512_usable = ghc_macos_sysctl_flag("hw.optional.avx512f"); +#endif + if (has_sse2) { SET_FEAT(feats, GHC_X86_FEAT_SSE2); } View it on GitLab: https://gitlab.haskell.org/ghc/ghc/-/commit/35de03e08ed4ac369fe0406194438098... -- View it on GitLab: https://gitlab.haskell.org/ghc/ghc/-/commit/35de03e08ed4ac369fe0406194438098... You're receiving this email because of your account on gitlab.haskell.org.