GitLab

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#notes-on-detecting-avx-512-on-macos
 +         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);
+     }