recursion-ninja pushed to branch wip/fix-26065 at Glasgow Haskell Compiler / GHC

Commits:

4 changed files:

Changes:

  • compiler/GHC/CmmToLlvm/CodeGen.hs
    ... ... @@ -240,12 +240,22 @@ genCall (PrimTarget op@(MO_BRev w)) [dst] args =
    240 240 
         genCallSimpleCast w op dst args
    241 241 
     genCall (PrimTarget op@(MO_BSwap w)) [dst] args =
    242 242 
         genCallSimpleCast w op dst args
    243 
    -genCall (PrimTarget op@(MO_Pdep w)) [dst] args =
    244 
    -    genCallSimpleCast w op dst args
    245 
    -genCall (PrimTarget op@(MO_Pext w)) [dst] args =
    246 
    -    genCallSimpleCast w op dst args
    247 243 
     genCall (PrimTarget op@(MO_PopCnt w)) [dst] args =
    248 244 
         genCallSimpleCast w op dst args
    245 
    +-- Check if the Intel BMI are enabled, and if the bit-width is less than 'W32'.
    246 
    +-- If so, we truncate to the 'W32' call the 32-bit intrinsic operation because
    247 
    +-- LLVM does not expose a call to 'PDep' and 'PExt' operations for bit-widths
    248 
    +-- of 'W8 and 'W16'.
    249 
    +genCall (PrimTarget op@(MO_Pdep w)) [dst] args = do
    250 
    +    cfg <- getConfig
    251 
    +    if   llvmCgBmiVersion cfg >= Just BMI2
    252 
    +    then genCallMinimumTruncationCast W32 w op dst args
    253 
    +    else genCallSimpleCast w op dst args
    254 
    +genCall (PrimTarget op@(MO_Pext w)) [dst] args = do
    255 
    +    cfg <- getConfig
    256 
    +    if   llvmCgBmiVersion cfg >= Just BMI2
    257 
    +    then genCallMinimumTruncationCast W32 w op dst args
    258 
    +    else genCallSimpleCast w op dst args
    249 259
    250 260 
     genCall (PrimTarget (MO_AtomicRMW width amop)) [dst] [addr, n] = runStmtsDecls $ do
    251 261 
         addrVar <- exprToVarW addr
    ... ... @@ -641,8 +651,12 @@ genCallExtract _ _ _ _ =
    641 651 
     -- from i32 to i8 explicitly as LLVM is strict about types.
    642 652 
     genCallSimpleCast :: Width -> CallishMachOp -> CmmFormal -> [CmmActual]
    643 653 
                       -> LlvmM StmtData
    644 
    -genCallSimpleCast specW op dst args = do
    645 
    -    let width   = widthToLlvmInt specW
    654 
    +genCallSimpleCast w = genCallMinimumTruncationCast w w
    655 
    +
    656 
    +genCallMinimumTruncationCast :: Width -> Width -> CallishMachOp -> CmmFormal
    657 
    +                             -> [CmmActual] -> LlvmM StmtData
    658 
    +genCallMinimumTruncationCast minW specW op dst args = do
    659 
    +    let width   = widthToLlvmInt $ max minW specW
    646 660 
             argsW   = const width <$> args
    647 661 
             dstType = cmmToLlvmType $ localRegType dst
    648 662 
             signage = cmmPrimOpRetValSignage op
    ... ... @@ -945,17 +959,24 @@ cmmPrimOpFunctions mop = do
    945 959 
           W256 -> fsLit "llvm.cttz.i256"
    946 960 
           W512 -> fsLit "llvm.cttz.i512"
    947 961 
         MO_Pdep w
    962 
    +      -- If the Intel BMI are enabled, then we will be calling the intrinsic operation
    963 
    +      -- through the LLVM binding, unless however the bit-width is 'W8' or 'W16'.
    964 
    +      -- In these cases, we truncate to the 'W32' bit-width and /directly/ call the
    965 
    +      -- 32-bit BMI operation. This is necessary because the LLVM does not expose a
    966 
    +      -- call to the 'PDep' and 'PExt' operation for bit-diths of 'W8 and 'W16'.
    967 
    +      -- Hence the necessity to to call the BMI intrinsic operation directlky from
    968 
    +      -- outside the LLVM.
    948 969 
           | isBmi2Enabled -> case w of
    949 
    -          W8   -> fsLit "llvm.x86.bmi.pdep.8"
    950 
    -          W16  -> fsLit "llvm.x86.bmi.pdep.16"
    970 
    +          W8   -> fsLit "llvm.x86.bmi.pdep.32"
    971 
    +          W16  -> fsLit "llvm.x86.bmi.pdep.32"
    951 972 
               W32  -> fsLit "llvm.x86.bmi.pdep.32"
    952 973 
               W64  -> fsLit "llvm.x86.bmi.pdep.64"
    953 974 
               W128 -> fsLit "llvm.x86.bmi.pdep.128"
    954 975 
               W256 -> fsLit "llvm.x86.bmi.pdep.256"
    955 976 
               W512 -> fsLit "llvm.x86.bmi.pdep.512"
    956 977 
           | otherwise -> case w of
    957 
    -          W8   -> fsLit "hs_pdep8"
    958 
    -          W16  -> fsLit "hs_pdep16"
    978 
    +          W8   -> fsLit "hs_pdep32"
    979 
    +          W16  -> fsLit "hs_pdep32"
    959 980 
               W32  -> fsLit "hs_pdep32"
    960 981 
               W64  -> fsLit "hs_pdep64"
    961 982 
               W128 -> fsLit "hs_pdep128"
    ... ... @@ -963,8 +984,10 @@ cmmPrimOpFunctions mop = do
    963 984 
               W512 -> fsLit "hs_pdep512"
    964 985 
         MO_Pext w
    965 986 
           | isBmi2Enabled -> case w of
    966 
    -          W8   -> fsLit "llvm.x86.bmi.pext.8"
    967 
    -          W16  -> fsLit "llvm.x86.bmi.pext.16"
    987 
    +          -- See the 'Mo_Pdep' commentary above as to why we call 'pext.32'
    988 
    +          -- instead of calling 'pext.8' or 'pext.16' operations.
    989 
    +          W8   -> fsLit "llvm.x86.bmi.pext.32"
    990 
    +          W16  -> fsLit "llvm.x86.bmi.pext.32"
    968 991 
               W32  -> fsLit "llvm.x86.bmi.pext.32"
    969 992 
               W64  -> fsLit "llvm.x86.bmi.pext.64"
    970 993 
               W128 -> fsLit "llvm.x86.bmi.pext.128"

  • testsuite/tests/llvm/should_run/T26065.hs
    1 
    +{-# LANGUAGE MagicHash #-}
    2 
    +import GHC.Exts
    3 
    +import GHC.Word
    4 
    +
    5 
    +pdep8 :: Word8 -> Word8 -> Word8
    6 
    +pdep8 (W8# a) (W8# b) = W8# (wordToWord8# (pdep8# (word8ToWord# a) (word8ToWord# b)))
    7 
    +{-# NOINLINE pdep8 #-}
    8 
    +
    9 
    +pdep16 :: Word16 -> Word16 -> Word16
    10 
    +pdep16 (W16# a) (W16# b) = W16# (wordToWord16# (pdep16# (word16ToWord# a) (word16ToWord# b)))
    11 
    +{-# NOINLINE pdep16 #-}
    12 
    +
    13 
    +pdep32 :: Word32 -> Word32 -> Word32
    14 
    +pdep32 (W32# a) (W32# b) = W32# (wordToWord32# (pdep32# (word32ToWord# a) (word32ToWord# b)))
    15 
    +{-# NOINLINE pdep32 #-}
    16 
    +
    17 
    +pdep64 :: Word64 -> Word64 -> Word64
    18 
    +pdep64 (W64# a) (W64# b) = W64# (pdep64# a b)
    19 
    +{-# NOINLINE pdep64 #-}
    20 
    +
    21 
    +pext8 :: Word8 -> Word8 -> Word8
    22 
    +pext8 (W8# a) (W8# b) = W8# (wordToWord8# (pext8# (word8ToWord# a) (word8ToWord# b)))
    23 
    +{-# NOINLINE pext8 #-}
    24 
    +
    25 
    +pext16 :: Word16 -> Word16 -> Word16
    26 
    +pext16 (W16# a) (W16# b) = W16# (wordToWord16# (pext16# (word16ToWord# a) (word16ToWord# b)))
    27 
    +{-# NOINLINE pext16 #-}
    28 
    +
    29 
    +pext32 :: Word32 -> Word32 -> Word32
    30 
    +pext32 (W32# a) (W32# b) = W32# (wordToWord32# (pext32# (word32ToWord# a) (word32ToWord# b)))
    31 
    +{-# NOINLINE pext32 #-}
    32 
    +
    33 
    +pext64 :: Word64 -> Word64 -> Word64
    34 
    +pext64 (W64# a) (W64# b) = W64# (pext64# a b)
    35 
    +{-# NOINLINE pext64 #-}
    36 
    +
    37 
    +main :: IO ()
    38 
    +main = do
    39 
    +  putStr "pdep8:\t"  *> print (pdep8  7 3)
    40 
    +  putStr "pdep16:\t" *> print (pdep16 7 3)
    41 
    +  putStr "pdep32:\t" *> print (pdep32 7 3)
    42 
    +  putStr "pdep64:\t" *> print (pdep64 7 3)
    43 
    +  putStr "pext8:\t"  *> print (pext8  7 3)
    44 
    +  putStr "pext16:\t" *> print (pext16 7 3)
    45 
    +  putStr "pext32:\t" *> print (pext32 7 3)
    46 
    +  putStr "pext64:\t" *> print (pext64 7 3)

  • testsuite/tests/llvm/should_run/T26065.stdout
    1 
    +pdep8:	3
    2 
    +pdep16:	3
    3 
    +pdep32:	3
    4 
    +pdep64:	3
    5 
    +pext8:	3
    6 
    +pext16:	3
    7 
    +pext32:	3
    8 
    +pext64:	3

  • testsuite/tests/llvm/should_run/all.T
    ... ... @@ -18,3 +18,8 @@ test('T22033', [normal, normalise_errmsg_fun(ignore_llvm_and_vortex)], compile_a
    18 18 
     test('T25730', [req_c, unless(arch('x86_64'), skip), normalise_errmsg_fun(ignore_llvm_and_vortex)], compile_and_run, ['T25730C.c'])
    19 19 
       # T25730C.c contains Intel instrinsics, so only run this test on x86
    20 20 
     test('T20645', [normal, normalise_errmsg_fun(ignore_llvm_and_vortex), when(have_llvm(), extra_ways(["optllvm"]))], compile_and_run, [''])
    21 
    +# T26065.c tests LLVM linking of Intel instrinsics, so only run this test on x86
    22 
    +test('T26065', [normal, normalise_errmsg_fun(ignore_llvm_and_vortex), when(have_llvm(), extra_ways(["optllvm"])),
    23 
    +                unless((arch('x86_64') or arch('i386')) and have_cpu_feature('bmi2'),skip)],
    24 
    +                compile_and_run, ['-mbmi2'])
    25 
    +

    • —
    View it on GitLab.
    You're receiving this email because of your account on gitlab.haskell.org. Manage all notifications · Help