[compiler-rt] [compiler-rt][AArch64][FMV] Use the hw.optional.arm.caps fast path (PR #95275)

Jon Roelofs via llvm-commits llvm-commits at lists.llvm.org
Wed Jun 12 10:38:13 PDT 2024


https://github.com/jroelofs created https://github.com/llvm/llvm-project/pull/95275

MacOS 15.0 and iOS 18.0 added a new sysctl to fetch a bitvector of all the hw.optional.arm.FEAT_*'s in one go.  Using this has a perf advantage over doing multiple round-trips to the kernel and back, but since it's not present in older oses, we still need the slow fallback.

>From 9235a34b9110f881a52ef05c937f1b869fa9d71c Mon Sep 17 00:00:00 2001
From: Jon Roelofs <jonathan_roelofs at apple.com>
Date: Wed, 12 Jun 2024 10:29:50 -0700
Subject: [PATCH] [compiler-rt][AArch64][FMV] Use the hw.optional.arm.caps fast
 path when available.

MacOS 15.0 and iOS 18.0 added a new sysctl to fetch a bitvector of all the
hw.optional.arm.FEAT_*'s in one go.  Using this has a perf advantage over doing
multiple round-trips to the kernel and back, but since it's not present in
older oses, we still need the slow fallback.
---
 .../builtins/cpu_model/aarch64/fmv/apple.inc  | 62 +++++++++++++++++++
 1 file changed, 62 insertions(+)

diff --git a/compiler-rt/lib/builtins/cpu_model/aarch64/fmv/apple.inc b/compiler-rt/lib/builtins/cpu_model/aarch64/fmv/apple.inc
index 6fef109567b61..6c6a71eed49d0 100644
--- a/compiler-rt/lib/builtins/cpu_model/aarch64/fmv/apple.inc
+++ b/compiler-rt/lib/builtins/cpu_model/aarch64/fmv/apple.inc
@@ -2,6 +2,11 @@
 #if TARGET_OS_OSX || TARGET_OS_IPHONE
 #include <sys/sysctl.h>
 
+#if __has_include(<arm/cpu_capabilities_public.h>)
+#  include <arm/cpu_capabilities_public.h>
+#  define HAS_CPU_CAPABILITIES_PUBLIC_H 1
+#endif
+
 static bool isKnownAndSupported(const char *name) {
   int32_t val = 0;
   size_t size = sizeof(val);
@@ -21,6 +26,63 @@ void __init_cpu_features_resolver(void) {
 
   uint64_t features = 0;
 
+#ifdef HAS_CPU_CAPABILITIES_PUBLIC_H
+  uint8_t feats_bitvec[(CAP_BIT_NB + 7) / 8] = {0};
+  size_t len = sizeof(feats_bitvec);
+  // When hw.optional.arm.feats is available (macOS 15.0+, iOS 18.0+), use the
+  // fast path to get all the feature bits, otherwise fall back to the slow
+  // ~20-something sysctls path.
+  if (!sysctlbyname("hw.optional.arm.caps", &feats_bitvec, &len, 0, 0)) {
+
+    #define CHECK_BIT(FROM, TO) do { \
+      _Static_assert(FROM < CAP_BIT_NB, BROKEN_ABI_MESSAGE); \
+      if (feats_bitvec[FROM / 8] & (1u << ((FROM) & 7))) { \
+        features |= (1ULL << TO); \
+      } \
+    } while(0)
+
+    CHECK_BIT(CAP_BIT_FEAT_FlagM, FEAT_FLAGM);
+    CHECK_BIT(CAP_BIT_FEAT_FlagM2, FEAT_FLAGM2);
+    CHECK_BIT(CAP_BIT_FEAT_FHM, FEAT_FP16FML);
+    CHECK_BIT(CAP_BIT_FEAT_DotProd, FEAT_DOTPROD);
+    CHECK_BIT(CAP_BIT_FEAT_SHA3, FEAT_SHA3);
+    CHECK_BIT(CAP_BIT_FEAT_RDM, FEAT_RDM);
+    CHECK_BIT(CAP_BIT_FEAT_LSE, FEAT_LSE);
+    CHECK_BIT(CAP_BIT_FEAT_SHA256, FEAT_SHA2);
+    CHECK_BIT(CAP_BIT_FEAT_SHA1, FEAT_SHA1);
+    CHECK_BIT(CAP_BIT_FEAT_AES, FEAT_AES);
+    CHECK_BIT(CAP_BIT_FEAT_PMULL, FEAT_PMULL);
+    CHECK_BIT(CAP_BIT_FEAT_SPECRES, FEAT_PREDRES);
+    CHECK_BIT(CAP_BIT_FEAT_SB, FEAT_SB);
+    CHECK_BIT(CAP_BIT_FEAT_FRINTTS, FEAT_FRINTTS);
+    CHECK_BIT(CAP_BIT_FEAT_LRCPC, FEAT_RCPC);
+    CHECK_BIT(CAP_BIT_FEAT_LRCPC2, FEAT_RCPC2);
+    CHECK_BIT(CAP_BIT_FEAT_FCMA, FEAT_FCMA);
+    CHECK_BIT(CAP_BIT_FEAT_JSCVT, FEAT_JSCVT);
+    CHECK_BIT(CAP_BIT_FEAT_DPB, FEAT_DPB);
+    CHECK_BIT(CAP_BIT_FEAT_DPB2, FEAT_DPB2);
+    CHECK_BIT(CAP_BIT_FEAT_BF16, FEAT_BF16);
+    CHECK_BIT(CAP_BIT_FEAT_I8MM, FEAT_I8MM);
+    CHECK_BIT(CAP_BIT_FEAT_DIT, FEAT_DIT);
+    CHECK_BIT(CAP_BIT_FEAT_FP16, FEAT_FP16);
+    CHECK_BIT(CAP_BIT_FEAT_SSBS, FEAT_SSBS2);
+    CHECK_BIT(CAP_BIT_FEAT_BTI, FEAT_BTI);
+    features |= (1ULL << FEAT_FP);
+    CHECK_BIT(CAP_BIT_AdvSIMD, FEAT_SIMD);
+    CHECK_BIT(CAP_BIT_CRC32, FEAT_CRC);
+    CHECK_BIT(CAP_BIT_FEAT_SME, FEAT_SME);
+    CHECK_BIT(CAP_BIT_FEAT_SME2, FEAT_SME2);
+    CHECK_BIT(CAP_BIT_FEAT_SME_F64F64, FEAT_SME_F64);
+    CHECK_BIT(CAP_BIT_FEAT_SME_I16I64, FEAT_SME_I64);
+
+    features |= (1ULL << FEAT_INIT);
+
+    __atomic_store(&__aarch64_cpu_features.features, &features,
+                    __ATOMIC_RELAXED);
+    return;
+  }
+#endif
+
   // https://developer.apple.com/documentation/kernel/1387446-sysctlbyname/determining_instruction_set_characteristics
   static const struct {
     const char *sysctl_name;



More information about the llvm-commits mailing list