Simon Jakobi pushed to branch wip/sjakobi/T25450-march-native at Glasgow Haskell Compiler / GHC Commits: ab044331 by Simon Jakobi at 2026-06-02T10:02:32+02:00 Implement -march=native for x86/x86_64 Add a -march=native flag (native compiler only) that probes the host CPU at parse time via an in-process CPUID/XGETBV helper and enables the matching CPU-feature DynFlags, so the effect applies to both the NCG and LLVM backends. The flag handler only records a marker, since flag parsing is pure; the probe and feature application run in parseDynamicFlagsFull, guarded against non-x86 targets and cross-compilation. Detected SSE/AVX and BMI levels are collapsed to their maximum and folded additively into the existing feature flags, so explicit -m... options are never disabled. The probe is memoized for the lifetime of the process. Includes documentation, a changelog entry, driver tests (baseline, superset, additive, and arch-guard) verified via --print-enabled-cpu-features, and an NCG asm test gated on the host having SSE4.2. Assisted-by: Claude Opus 4.8 - - - - - 14 changed files: - + changelog.d/march-native - + compiler/GHC/Driver/CpuFeatures.hs - compiler/GHC/Driver/DynFlags.hs - compiler/GHC/Driver/Session.hs - + compiler/cbits/cpu_features_x86.c - compiler/ghc.cabal.in - docs/users_guide/using.rst - testsuite/tests/codeGen/should_gen_asm/all.T - + testsuite/tests/codeGen/should_gen_asm/march-native-enables-popcnt.asm - + testsuite/tests/codeGen/should_gen_asm/march-native-enables-popcnt.hs - testsuite/tests/driver/all.T - + testsuite/tests/driver/march_native.stdout - + testsuite/tests/driver/march_native_additive.stdout - + testsuite/tests/driver/march_native_unsupported_arch.stderr Changes: ===================================== changelog.d/march-native ===================================== @@ -0,0 +1,11 @@ +section: compiler +synopsis: Add -march=native flag +issues: #25450 + +description: + GHC now supports ``-march=native`` on x86 and x86_64. It probes the CPU of the + machine running GHC and enables all of the corresponding ``-m...`` CPU-feature + options automatically (such as ``-msse4.2``, ``-mavx2``, ``-mbmi2`` and + ``-mfma``), for both the native code generator and the LLVM backend. The + detected features are enabled in addition to any explicitly requested feature + flags. The flag is rejected for non-x86 targets and when cross-compiling. ===================================== compiler/GHC/Driver/CpuFeatures.hs ===================================== @@ -0,0 +1,88 @@ +{-# LANGUAGE CPP #-} + +module GHC.Driver.CpuFeatures + ( X86CpuFeature(..) + , cachedX86CpuFeatures + ) where + +import GHC.Prelude + +import Data.Bits (testBit) +import Data.Word (Word64) +import System.IO.Unsafe (unsafePerformIO) + +-- | x86 CPU features understood by GHC's native CPU feature probe. +data X86CpuFeature + = SSE2 + | SSE3 + | SSSE3 + | SSE4_1 + | SSE4_2 + | AVX + | AVX2 + | AVX512F + | AVX512BW + | AVX512CD + | AVX512DQ + | AVX512VL + | BMI1 + | BMI2 + | FMA + | GFNI + deriving (Eq, Ord, Show) + +-- | Decode the bitmask returned by 'ghc_detect_x86_cpu_features'. +-- +-- NOTE: Bit positions must match the enum in @compiler/cbits/cpu_features_x86.c@. +decodeX86CpuFeatureMask :: Word64 -> [X86CpuFeature] +decodeX86CpuFeatureMask mask = + [ feat + | (bit_ix, feat) <- cpuFeatureBitLayout + , testBit mask bit_ix + ] + +-- | Low-level FFI access to the C probe. +detectX86CpuFeatureMask :: IO Word64 +#if defined(javascript_HOST_ARCH) +detectX86CpuFeatureMask = pure 0 +#else +detectX86CpuFeatureMask = c_ghc_detect_x86_cpu_features +#endif + +-- | Probe host x86 CPU features and decode them into an ordered feature list. +detectX86CpuFeatures :: IO [X86CpuFeature] +detectX86CpuFeatures = decodeX86CpuFeatureMask <$> detectX86CpuFeatureMask + +-- | The host's x86 CPU features, probed once and memoized. +-- +-- CPUID results are constant for the lifetime of the process, so probing more +-- than once (e.g. once per @-march=native@ in a command line or file pragma) +-- is wasteful. This is referentially transparent despite the FFI call. +cachedX86CpuFeatures :: [X86CpuFeature] +cachedX86CpuFeatures = unsafePerformIO detectX86CpuFeatures +{-# NOINLINE cachedX86CpuFeatures #-} + +cpuFeatureBitLayout :: [(Int, X86CpuFeature)] +cpuFeatureBitLayout = + [ (0, SSE2) + , (1, SSE3) + , (2, SSSE3) + , (3, SSE4_1) + , (4, SSE4_2) + , (5, AVX) + , (6, AVX2) + , (7, AVX512F) + , (8, AVX512BW) + , (9, AVX512CD) + , (10, AVX512DQ) + , (11, AVX512VL) + , (12, BMI1) + , (13, BMI2) + , (14, FMA) + , (15, GFNI) + ] + +#if !defined(javascript_HOST_ARCH) +foreign import ccall unsafe "ghc_detect_x86_cpu_features" + c_ghc_detect_x86_cpu_features :: IO Word64 +#endif ===================================== compiler/GHC/Driver/DynFlags.hs ===================================== @@ -470,6 +470,8 @@ data DynFlags = DynFlags { fma :: Bool, -- ^ Enable FMA instructions. gfni :: Bool, -- ^ Enable GFNI Instructions. la664 :: Bool, -- ^ Enable LA664 instructions + marchNative :: Bool, -- ^ @-march=native@ was requested; the host + -- CPU features are applied during flag parsing. -- Constants used to control the amount of optimization done. @@ -760,6 +762,7 @@ defaultDynFlags mySettings = gfni = False, -- For LoongArch, la464 is used by default. la664 = False, + marchNative = False, maxInlineAllocSize = 128, maxInlineMemcpyInsns = 32, ===================================== compiler/GHC/Driver/Session.hs ===================================== @@ -245,6 +245,8 @@ import GHC.Platform import GHC.Platform.Ways import GHC.Platform.Profile import GHC.Platform.ArchOS +import GHC.Platform.Host (hostPlatformArch) +import qualified GHC.Driver.CpuFeatures as Cpu import GHC.Unit.Types import GHC.Unit.Parser @@ -906,8 +908,12 @@ parseDynamicFlagsFull activeFlags cmdline logger dflags0 args = do unless (null errs) $ liftIO $ throwGhcExceptionIO $ errorsToGhcException $ map ((rdr . ppr . getLoc &&& unLoc) . errMsg) $ errs + -- Apply -march=native: probe the host CPU and enable the matching feature + -- flags. This needs IO (CPUID), so it cannot live in the pure flag handlers. + dflags1' <- applyMarchNative dflags1 + -- check for disabled flags in safe haskell - let (dflags2, sh_warns) = safeFlagCheck cmdline dflags1 + let (dflags2, sh_warns) = safeFlagCheck cmdline dflags1' theWays = ways dflags2 unless (allowed_combination theWays) $ liftIO $ @@ -1743,6 +1749,7 @@ dynamic_flags_deps = [ , make_ord_flag defGhcFlag "mavx512vl" (noArg (\d -> d { avx512vl = True })) , make_ord_flag defGhcFlag "mfma" (noArg (\d -> d { fma = True })) , make_ord_flag defGhcFlag "mgfni" (noArg (\d -> d { gfni = True })) + , make_ord_flag defGhcFlag "march=native" (noArg (\d -> d { marchNative = True })) , make_ord_flag defGhcFlag "mla664" (noArg (\d -> d { la664 = True })) @@ -3806,6 +3813,59 @@ x86AsMFlags dflags = gfniFlags = [ "-mgfni" | gfni dflags ] +-- | Apply a requested @-march=native@ by probing the host CPU and enabling the +-- matching CPU-feature flags. +-- +-- This runs in 'parseDynamicFlagsFull' rather than in a flag handler because the +-- CPUID probe needs 'IO', whereas flag handlers are pure. The detected features +-- are folded into the existing feature 'DynFlags' so that 'makeDynFlagsConsistent' +-- and the backends treat them exactly like the corresponding @-m...@ flags. +applyMarchNative :: MonadIO m => DynFlags -> m DynFlags +applyMarchNative dflags + | not (marchNative dflags) = return dflags + | otherwise = do + let arch = platformArch (targetPlatform dflags) + unless (arch == ArchX86 || arch == ArchX86_64) $ liftIO $ + throwGhcExceptionIO $ CmdLineError + "-march=native is not supported on this target yet (only x86 and x86_64 are supported so far)" + unless (arch == hostPlatformArch) $ liftIO $ + throwGhcExceptionIO $ CmdLineError + "-march=native is not supported when cross-compiling" + return (applyX86CpuFeatures Cpu.cachedX86CpuFeatures dflags) + +-- | Enable the 'DynFlags' CPU-feature fields corresponding to a probed set of +-- host x86 features. SSE/AVX and BMI levels are collapsed to their maximum, +-- since 'sseAvxVersion' and 'bmiVersion' each record a single level. +applyX86CpuFeatures :: [Cpu.X86CpuFeature] -> DynFlags -> DynFlags +applyX86CpuFeatures feats dflags = dflags + { sseAvxVersion = foldr (max . Just) (sseAvxVersion dflags) sseLevels + , bmiVersion = foldr (max . Just) (bmiVersion dflags) bmiLevels + , avx512f = avx512f dflags || has Cpu.AVX512F + , avx512bw = avx512bw dflags || has Cpu.AVX512BW + , avx512cd = avx512cd dflags || has Cpu.AVX512CD + , avx512dq = avx512dq dflags || has Cpu.AVX512DQ + , avx512vl = avx512vl dflags || has Cpu.AVX512VL + , fma = fma dflags || has Cpu.FMA + , gfni = gfni dflags || has Cpu.GFNI + } + where + has feat = feat `elem` feats + sseLevels = [ lvl | feat <- feats, Just lvl <- [sseLevelOf feat] ] + bmiLevels = [ lvl | feat <- feats, Just lvl <- [bmiLevelOf feat] ] + sseLevelOf feat = case feat of + Cpu.SSE2 -> Just SSE2 + Cpu.SSE3 -> Just SSE3 + Cpu.SSSE3 -> Just SSSE3 + Cpu.SSE4_1 -> Just SSE4 + Cpu.SSE4_2 -> Just SSE42 + Cpu.AVX -> Just AVX1 + Cpu.AVX2 -> Just AVX2 + _ -> Nothing + bmiLevelOf feat = case feat of + Cpu.BMI1 -> Just BMI1 + Cpu.BMI2 -> Just BMI2 + _ -> Nothing + -- | Query if the target RTS has the given 'Ways'. It's computed from -- the @"RTS ways"@ field in the settings file. targetHasRTSWays :: DynFlags -> Ways -> Bool ===================================== compiler/cbits/cpu_features_x86.c ===================================== @@ -0,0 +1,177 @@ +#include +#include + +#if defined(_MSC_VER) && (defined(_M_IX86) || defined(_M_X64)) +#include +#include +#endif + +#if !defined(_MSC_VER) && (defined(__i386__) || defined(__x86_64__)) +#include +#endif + +enum { + GHC_X86_FEAT_SSE2 = 0, + GHC_X86_FEAT_SSE3, + GHC_X86_FEAT_SSSE3, + GHC_X86_FEAT_SSE4_1, + GHC_X86_FEAT_SSE4_2, + GHC_X86_FEAT_AVX, + GHC_X86_FEAT_AVX2, + GHC_X86_FEAT_AVX512F, + GHC_X86_FEAT_AVX512BW, + GHC_X86_FEAT_AVX512CD, + GHC_X86_FEAT_AVX512DQ, + GHC_X86_FEAT_AVX512VL, + GHC_X86_FEAT_BMI1, + GHC_X86_FEAT_BMI2, + GHC_X86_FEAT_FMA, + GHC_X86_FEAT_GFNI +}; + +#define SET_FEAT(mask, bit) ((mask) |= ((HsWord64)1ULL << (bit))) + +static int ghc_cpuid_count(uint32_t leaf, uint32_t subleaf, + uint32_t *a, uint32_t *b, uint32_t *c, uint32_t *d) +{ +#if defined(_MSC_VER) && (defined(_M_IX86) || defined(_M_X64)) + int regs[4]; + __cpuidex(regs, (int)leaf, (int)subleaf); + *a = (uint32_t)regs[0]; + *b = (uint32_t)regs[1]; + *c = (uint32_t)regs[2]; + *d = (uint32_t)regs[3]; + return 1; +#elif defined(__i386__) || defined(__x86_64__) + return __get_cpuid_count(leaf, subleaf, a, b, c, d); +#else + (void)leaf; + (void)subleaf; + (void)a; + (void)b; + (void)c; + (void)d; + return 0; +#endif +} + +static uint64_t ghc_xgetbv0(void) +{ +#if defined(_MSC_VER) && (defined(_M_IX86) || defined(_M_X64)) + return (uint64_t)_xgetbv(0); +#elif defined(__i386__) || defined(__x86_64__) + uint32_t eax, edx; + __asm__ volatile(".byte 0x0f, 0x01, 0xd0" /* xgetbv */ + : "=a"(eax), "=d"(edx) + : "c"(0)); + return ((uint64_t)edx << 32) | (uint64_t)eax; +#else + return 0; +#endif +} + +HsWord64 ghc_detect_x86_cpu_features(void) +{ + HsWord64 feats = 0; + +#if defined(_M_IX86) || defined(_M_X64) || defined(__i386__) || defined(__x86_64__) + uint32_t a, b, c, d; + uint32_t max_basic = 0; + + if (!ghc_cpuid_count(0, 0, &a, &b, &c, &d)) { + return 0; + } + max_basic = a; + if (max_basic < 1) { + return 0; + } + + ghc_cpuid_count(1, 0, &a, &b, &c, &d); + + { + int has_sse2 = !!(d & (1u << 26)); + int has_sse3 = !!(c & (1u << 0)); + int has_ssse3 = !!(c & (1u << 9)); + int has_sse4_1 = !!(c & (1u << 19)); + int has_sse4_2 = !!(c & (1u << 20)); + int has_fma_hw = !!(c & (1u << 12)); + int has_avx_hw = !!(c & (1u << 28)); + int has_osxsave = !!(c & (1u << 27)); + + int avx_usable = 0; + int avx512_usable = 0; + + if (has_osxsave) { + uint64_t xcr0 = ghc_xgetbv0(); + avx_usable = ((xcr0 & 0x6u) == 0x6u); /* XMM + YMM state */ + avx512_usable = ((xcr0 & 0xE6u) == 0xE6u); /* XMM+YMM+opmask+ZMM */ + } + + if (has_sse2) { + SET_FEAT(feats, GHC_X86_FEAT_SSE2); + } + if (has_sse3) { + SET_FEAT(feats, GHC_X86_FEAT_SSE3); + } + if (has_ssse3) { + SET_FEAT(feats, GHC_X86_FEAT_SSSE3); + } + if (has_sse4_1) { + SET_FEAT(feats, GHC_X86_FEAT_SSE4_1); + } + if (has_sse4_2) { + SET_FEAT(feats, GHC_X86_FEAT_SSE4_2); + } + if (has_avx_hw && avx_usable) { + SET_FEAT(feats, GHC_X86_FEAT_AVX); + } + if (has_fma_hw && avx_usable) { + SET_FEAT(feats, GHC_X86_FEAT_FMA); + } + + if (max_basic >= 7 && ghc_cpuid_count(7, 0, &a, &b, &c, &d)) { + int has_bmi1 = !!(b & (1u << 3)); + int has_avx2_hw = !!(b & (1u << 5)); + int has_bmi2 = !!(b & (1u << 8)); + int has_avx512f = !!(b & (1u << 16)); + int has_avx512dq = !!(b & (1u << 17)); + int has_avx512cd = !!(b & (1u << 28)); + int has_avx512bw = !!(b & (1u << 30)); + int has_avx512vl = !!(b & (1u << 31)); + int has_gfni = !!(c & (1u << 8)); + + if (has_bmi1) { + SET_FEAT(feats, GHC_X86_FEAT_BMI1); + } + if (has_bmi2) { + SET_FEAT(feats, GHC_X86_FEAT_BMI2); + } + if (avx_usable && has_avx2_hw) { + SET_FEAT(feats, GHC_X86_FEAT_AVX2); + } + + if (avx512_usable && has_avx512f) { + SET_FEAT(feats, GHC_X86_FEAT_AVX512F); + if (has_avx512bw) { + SET_FEAT(feats, GHC_X86_FEAT_AVX512BW); + } + if (has_avx512cd) { + SET_FEAT(feats, GHC_X86_FEAT_AVX512CD); + } + if (has_avx512dq) { + SET_FEAT(feats, GHC_X86_FEAT_AVX512DQ); + } + if (has_avx512vl) { + SET_FEAT(feats, GHC_X86_FEAT_AVX512VL); + } + } + + if (has_gfni) { + SET_FEAT(feats, GHC_X86_FEAT_GFNI); + } + } + } +#endif + + return feats; +} ===================================== compiler/ghc.cabal.in ===================================== @@ -187,6 +187,7 @@ Library else c-sources: cbits/cutils.c + cbits/cpu_features_x86.c cbits/genSym.c cbits/keepCAFsForGHCi.c @@ -514,6 +515,7 @@ Library GHC.Driver.Config.StgToCmm GHC.Driver.Config.Tidy GHC.Driver.Config.StgToJS + GHC.Driver.CpuFeatures GHC.Driver.DynFlags GHC.Driver.IncludeSpecs GHC.Driver.Downsweep ===================================== docs/users_guide/using.rst ===================================== @@ -1854,6 +1854,31 @@ Some flags only make sense for particular target platforms. so this flag has no effect when used with the :ref:`native code generator <native-code-gen>` or the :ref:`LLVM backend <llvm-code-gen>`. +.. ghc-flag:: -march=native + :shortdesc: (x86 only) Enable all CPU features supported by the host + :type: dynamic + :category: platform-options + + (x86/x86_64 only) Probe the CPU of the machine running GHC and enable all of + the corresponding ``-m...`` CPU-feature options automatically (for example + ``-msse4.2``, ``-mavx2``, ``-mbmi2``, ``-mfma``). The detected features apply + to both the :ref:`native code generator <native-code-gen>` and the + :ref:`LLVM backend <llvm-code-gen>`. + + The detected features are enabled *in addition* to any CPU-feature flags you + pass explicitly, regardless of their order on the command line; ``-march=native`` + never disables a feature. + + .. warning:: + + Code compiled with ``-march=native`` may use instructions that are not + available on other CPUs, and is therefore not portable to a different + machine. + + Only x86 and x86_64 targets are supported so far; the flag is rejected on + other targets. It is also rejected when cross-compiling, since the host CPU + is then unrelated to the target. + Haddock ------- ===================================== testsuite/tests/codeGen/should_gen_asm/all.T ===================================== @@ -17,6 +17,11 @@ test('msse-option-order', [unless(arch('x86_64') or arch('i386'), skip), when(unregisterised(), skip)], compile_grep_asm, ['hs', False, '-msse4.2 -msse2']) test('mavx-should-enable-popcnt', [unless(arch('x86_64') or arch('i386'), skip), when(unregisterised(), skip)], compile_grep_asm, ['hs', False, '-mavx']) +# -march=native probes the host CPU, so gate on the host actually having SSE4.2 +# (have_cpu_feature reports nothing under cross, skipping the test there too). +test('march-native-enables-popcnt', + [unless((arch('x86_64') or arch('i386')) and have_cpu_feature('sse4_2'), skip), + when(unregisterised(), skip)], compile_grep_asm, ['hs', False, '-march=native']) test('avx512-int64-mul', [unless(arch('x86_64'), skip), when(unregisterised(), skip)], compile_grep_asm, ['hs', True, '-mavx512dq -mavx512vl']) test('avx512-int64-minmax', [unless(arch('x86_64'), skip), ===================================== testsuite/tests/codeGen/should_gen_asm/march-native-enables-popcnt.asm ===================================== @@ -0,0 +1 @@ +popcnt(?![0-9]) \ No newline at end of file ===================================== testsuite/tests/codeGen/should_gen_asm/march-native-enables-popcnt.hs ===================================== @@ -0,0 +1,11 @@ +-- `-march=native` enables the host's CPU features. On a host with SSE4.2 +-- (gated in all.T via have_cpu_feature) this makes popCount compile to a +-- `popcnt` instruction rather than the SSE2-baseline software fallback. +import Data.Bits + +{-# NOINLINE foo #-} +foo :: Int -> Int +foo x = 1 + popCount x + +main :: IO () +main = print (foo 42) ===================================== testsuite/tests/driver/all.T ===================================== @@ -7,6 +7,12 @@ def normalise_unknown_flag(msg): m = re.search(r'unrecognised flag: \S+', msg) return m.group(0) + '\n' if m else msg +def normalise_march_native_error(msg): + # Keep only the stable '-march=native ...' diagnostic; the program-name + # prefix and any usage trailer vary across configurations. + m = re.search(r'-march=native is [^\n]+', msg) + return m.group(0) + '\n' if m else msg + test('driver011', [extra_files(['A011.hs'])], makefile_test, ['test011']) test('driver012', [extra_files(['A012.hs'])], makefile_test, ['test012']) @@ -265,6 +271,42 @@ test('print_enabled_cpu_features_unknown_flag', run_command, ['{compiler} -mavx22 --print-enabled-cpu-features']) +# -march=native enables at least the x86_64 baseline (SSE2). The full feature +# set is host-dependent, so we only assert the always-present baseline. +test('march_native', + [unless(arch('x86_64') or arch('i386'), skip)], + run_command, + ['{compiler} -march=native --print-enabled-cpu-features | grep -o SSE2']) + +# On non-x86 targets -march=native must be rejected. +test('march_native_unsupported_arch', + [when(arch('x86_64') or arch('i386'), skip), + normalise_fun(normalise_march_native_error), exit_code(1)], + run_command, + ['{compiler} -march=native --print-enabled-cpu-features']) + +# -march=native is additive: its feature set is a superset of the default set. +# We extract the "features" arrays with and without the flag and assert that no +# baseline feature is dropped (comm -23 prints baseline-only entries; expect none). +# This avoids hard-coding the host-specific feature set. +test('march_native_superset', + [unless(arch('x86_64') or arch('i386'), skip)], + run_command, + ['{compiler} --print-enabled-cpu-features | ' + 'sed \'s/.*"features":\\[//;s/].*//;s/"//g\' | tr \',\' \'\\n\' | sort > base.txt && ' + '{compiler} -march=native --print-enabled-cpu-features | ' + 'sed \'s/.*"features":\\[//;s/].*//;s/"//g\' | tr \',\' \'\\n\' | sort > native.txt && ' + 'comm -23 base.txt native.txt']) + +# -march=native is additive with explicit -m flags, regardless of order: an +# explicitly requested feature (here AVX2, forced on independent of the host) is +# still present whether the flag comes before or after -march=native. +test('march_native_additive', + [unless(arch('x86_64') or arch('i386'), skip)], + run_command, + ['{compiler} -mavx2 -march=native --print-enabled-cpu-features | grep -o AVX2 && ' + '{compiler} -march=native -mavx2 --print-enabled-cpu-features | grep -o AVX2']) + test('T10219', normal, run_command, # `-x hspp` in make mode should work. # Note: need to specify `-x hspp` before the filename. ===================================== testsuite/tests/driver/march_native.stdout ===================================== @@ -0,0 +1 @@ +SSE2 ===================================== testsuite/tests/driver/march_native_additive.stdout ===================================== @@ -0,0 +1,2 @@ +AVX2 +AVX2 ===================================== testsuite/tests/driver/march_native_unsupported_arch.stderr ===================================== @@ -0,0 +1 @@ +-march=native is not supported on this target yet (only x86 and x86_64 are supported so far) View it on GitLab: https://gitlab.haskell.org/ghc/ghc/-/commit/ab044331be6b319ed609fa0023eb3857... -- View it on GitLab: https://gitlab.haskell.org/ghc/ghc/-/commit/ab044331be6b319ed609fa0023eb3857... You're receiving this email because of your account on gitlab.haskell.org.