Simon Jakobi pushed to branch wip/sjakobi/T24813 at Glasgow Haskell Compiler / GHC Commits: 4b7193ab by Simon Jakobi at 2026-09-06T02:26:01+02:00 X86 NCG: use 32-bit POPCNT for sub-word popcounts For MO_PopCnt at W8/W16, use 32-bit POPCNT instead of the 16-bit form. A 32-bit write zero-extends the result to the full register, so the separate zero-extension of the result is no longer needed. Also skip zero-extending the argument when it is visibly already zero-extended, as in popCount on Word8, which calls popCnt8# on word8ToWord# x. For the byte-wise popcount loop from #24813 this shrinks the popcount sequence from 5 instructions to 3. Assisted-by: Claude Fable 5.1 - - - - - 5 changed files: - + changelog.d/ncg-x86-subword-popcnt - compiler/GHC/CmmToAsm/X86/CodeGen.hs - + testsuite/tests/codeGen/should_gen_asm/T24813.asm - + testsuite/tests/codeGen/should_gen_asm/T24813.cmm - testsuite/tests/codeGen/should_gen_asm/all.T Changes: ===================================== changelog.d/ncg-x86-subword-popcnt ===================================== @@ -0,0 +1,10 @@ +section: compiler +synopsis: The x86 native code generator now emits shorter code for ``popCnt8#`` + and ``popCnt16#`` with ``-msse4.2`` +description: + Sub-word popcounts now use the 32-bit ``popcnt`` instruction, which + zero-extends its result for free, and skip zero-extending an argument that + is already zero-extended, as in ``popCount`` on ``Word8`` or ``Word16``. + The popcount sequence shrinks from up to five instructions to one to three. +mrs: !16640 +issues: #24813 ===================================== compiler/GHC/CmmToAsm/X86/CodeGen.hs ===================================== @@ -6739,20 +6739,23 @@ genPopCnt bid width dst src = do True -> do code_src <- getAnyReg src - src_r <- getNewRegNat format + -- Sub-word popcounts use 32-bit POPCNT, whose write zero-extends the + -- result to the full register. + src_r <- getNewRegNat (max II32 format) let dst_r = getRegisterReg platform (CmmLocal dst) + -- A sub-word argument is a full Word# whose upper bits may be + -- dirty, so zero-extend it unless it visibly already is. + zx_src + | CmmMachOp (MO_UU_Conv from _) [_] <- src, from <= width + = nilOL + | otherwise + = unitOL (MOVZxL format (OpReg src_r) (OpReg src_r)) return $ code_src src_r `appOL` - (if width == W8 then - -- The POPCNT instruction doesn't take a r/m8 - unitOL (MOVZxL II8 (OpReg src_r) (OpReg src_r)) `appOL` - unitOL (POPCNT II16 (OpReg src_r) dst_r) - else - unitOL (POPCNT format (OpReg src_r) dst_r)) `appOL` (if width == W8 || width == W16 then - -- We used a 16-bit destination register above, - -- so zero-extend - unitOL (MOVZxL II16 (OpReg dst_r) (OpReg dst_r)) - else nilOL) + zx_src `appOL` + unitOL (POPCNT II32 (OpReg src_r) dst_r) + else + unitOL (POPCNT format (OpReg src_r) dst_r)) False -> -- generate C call to hs_popcntN in ghc-prim ===================================== testsuite/tests/codeGen/should_gen_asm/T24813.asm ===================================== @@ -0,0 +1,10 @@ +popcnt8: + movzbl + popcnt + movq + jmp +popcnt16: + movzwl + popcnt + movq + jmp ===================================== testsuite/tests/codeGen/should_gen_asm/T24813.cmm ===================================== @@ -0,0 +1,16 @@ +#include "Cmm.h" + +// Popcount of a zero-extended sub-word value should be a single +// zero-extending load followed by one POPCNT (#24813). + +popcnt8 (W_ p) { + W_ r; + (r) = prim %popcnt8(%zx64(bits8[p])); + return (r); +} + +popcnt16 (W_ p) { + W_ r; + (r) = prim %popcnt16(%zx64(bits16[p])); + return (r); +} ===================================== testsuite/tests/codeGen/should_gen_asm/all.T ===================================== @@ -27,6 +27,7 @@ test('T25233', [unless(arch('x86_64'), skip), when(unregisterised(), skip)], compile_grep_asm, ['hs', True, '-O']) test('T25233b', [unless(arch('x86_64'), skip), when(unregisterised(), skip)], compile_grep_asm, ['cmm', True, '']) +test('T24813', is_amd64_codegen, compile_cmp_asm, ['cmm', '-msse4.2']) is_aarch64_codegen = [ unless(arch('aarch64'), skip), when(unregisterised(), skip), View it on GitLab: https://gitlab.haskell.org/ghc/ghc/-/commit/4b7193ab6a22557a55650cdb8efd1ccc... -- View it on GitLab: https://gitlab.haskell.org/ghc/ghc/-/commit/4b7193ab6a22557a55650cdb8efd1ccc... You're receiving this email because of your account on gitlab.haskell.org. Manage all notifications: https://gitlab.haskell.org/-/profile/notifications | Help: https://gitlab.haskell.org/help