diff --git a/llvm-general.cabal b/llvm-general.cabal --- a/llvm-general.cabal +++ b/llvm-general.cabal @@ -1,5 +1,5 @@ name: llvm-general -version: 3.4.3.1 +version: 3.4.4.0 license: BSD3 license-file: LICENSE author: Benjamin S.Scarlet @@ -16,7 +16,7 @@ handles almost all of the stateful complexities of using the LLVM API to build IR; and it supports moving IR not only from Haskell into LLVM C++ objects, but the other direction - from LLVM C++ into Haskell. . - For haddock, see . + For haddock, see . extra-source-files: src/LLVM/General/Internal/FFI/Analysis.h src/LLVM/General/Internal/FFI/BinaryOperator.h @@ -40,7 +40,7 @@ type: git location: git://github.com/bscarlet/llvm-general.git branch: llvm-3.4 - tag: v3.4.3.1 + tag: v3.4.4.0 flag shared-llvm description: link against llvm shared rather than static library @@ -64,7 +64,7 @@ parsec >= 3.1.3, array >= 0.4.0.0, setenv >= 0.1.0, - llvm-general-pure == 3.4.3.1 + llvm-general-pure == 3.4.4.0 extra-libraries: stdc++ hs-source-dirs: src extensions: @@ -90,6 +90,7 @@ LLVM.General.Target LLVM.General.Target.LibraryFunction LLVM.General.Target.Options + LLVM.General.Threading LLVM.General.Transforms other-modules: @@ -128,6 +129,7 @@ LLVM.General.Internal.RMWOperation LLVM.General.Internal.String LLVM.General.Internal.Target + LLVM.General.Internal.Threading LLVM.General.Internal.Type LLVM.General.Internal.Value LLVM.General.Internal.FFI.Analysis @@ -160,6 +162,7 @@ LLVM.General.Internal.FFI.RawOStream LLVM.General.Internal.FFI.SMDiagnostic LLVM.General.Internal.FFI.Target + LLVM.General.Internal.FFI.Threading LLVM.General.Internal.FFI.Transforms LLVM.General.Internal.FFI.Type LLVM.General.Internal.FFI.User @@ -199,8 +202,8 @@ HUnit >= 1.2.4.2, test-framework-quickcheck2 >= 0.3.0.1, QuickCheck >= 2.5.1.1, - llvm-general == 3.4.3.1, - llvm-general-pure == 3.4.3.1, + llvm-general == 3.4.4.0, + llvm-general-pure == 3.4.4.0, containers >= 0.4.2.1, mtl >= 2.0.1.0 hs-source-dirs: test diff --git a/src/LLVM/General/Internal/FFI/PassManager.hs b/src/LLVM/General/Internal/FFI/PassManager.hs --- a/src/LLVM/General/Internal/FFI/PassManager.hs +++ b/src/LLVM/General/Internal/FFI/PassManager.hs @@ -129,3 +129,9 @@ foreign import ccall unsafe "LLVM_General_PassManagerBuilderSetLibraryInfo" passManagerBuilderSetLibraryInfo :: Ptr PassManagerBuilder -> Ptr TargetLibraryInfo -> IO () + +foreign import ccall unsafe "LLVM_General_PassManagerBuilderSetLoopVectorize" passManagerBuilderSetLoopVectorize :: + Ptr PassManagerBuilder -> LLVMBool -> IO () + +foreign import ccall unsafe "LLVM_General_PassManagerBuilderSetSuperwordLevelParallelismVectorize" passManagerBuilderSetSuperwordLevelParallelismVectorize :: + Ptr PassManagerBuilder -> LLVMBool -> IO () diff --git a/src/LLVM/General/Internal/FFI/PassManagerC.cpp b/src/LLVM/General/Internal/FFI/PassManagerC.cpp --- a/src/LLVM/General/Internal/FFI/PassManagerC.cpp +++ b/src/LLVM/General/Internal/FFI/PassManagerC.cpp @@ -246,14 +246,34 @@ unwrap(PM)->add(createDebugIRPass()); } -void -LLVM_General_PassManagerBuilderSetLibraryInfo( - LLVMPassManagerBuilderRef PMB, - LLVMTargetLibraryInfoRef l +void LLVM_General_AddLoopVectorizePass( + LLVMPassManagerRef PM, + LLVMBool noUnrolling ) { - // The PassManager frees the TargetLibraryInfo when done, - // but we also free our ref, so give it a new copy. - unwrap(PMB)->LibraryInfo = new TargetLibraryInfo(*unwrap(l)); + unwrap(PM)->add(createLoopVectorizePass(noUnrolling)); +} + +void LLVM_General_PassManagerBuilderSetLibraryInfo( + LLVMPassManagerBuilderRef PMB, + LLVMTargetLibraryInfoRef l +) { + // The PassManager frees the TargetLibraryInfo when done, + // but we also free our ref, so give it a new copy. + unwrap(PMB)->LibraryInfo = new TargetLibraryInfo(*unwrap(l)); +} + +void LLVM_General_PassManagerBuilderSetLoopVectorize( + LLVMPassManagerBuilderRef PMB, + LLVMBool runLoopVectorization +) { + unwrap(PMB)->LoopVectorize = runLoopVectorization; +} + +void LLVM_General_PassManagerBuilderSetSuperwordLevelParallelismVectorize( + LLVMPassManagerBuilderRef PMB, + LLVMBool runSLPVectorization +) { + unwrap(PMB)->SLPVectorize = runSLPVectorization; } } diff --git a/src/LLVM/General/Internal/FFI/Threading.hs b/src/LLVM/General/Internal/FFI/Threading.hs new file mode 100644 --- /dev/null +++ b/src/LLVM/General/Internal/FFI/Threading.hs @@ -0,0 +1,17 @@ +{-# LANGUAGE + ForeignFunctionInterface + #-} +module LLVM.General.Internal.FFI.Threading where + +import Foreign.C + +import LLVM.General.Internal.FFI.LLVMCTypes + +foreign import ccall unsafe "LLVMStartMultithreaded" startMultithreaded :: + IO LLVMBool + +foreign import ccall unsafe "LLVMStopMultithreaded" stopMultithreaded :: + IO () + +foreign import ccall unsafe "LLVMIsMultithreaded" isMultithreaded :: + IO LLVMBool diff --git a/src/LLVM/General/Internal/FFI/Transforms.hs b/src/LLVM/General/Internal/FFI/Transforms.hs --- a/src/LLVM/General/Internal/FFI/Transforms.hs +++ b/src/LLVM/General/Internal/FFI/Transforms.hs @@ -33,6 +33,7 @@ "OldScalarReplacementOfAggregates" -> "ScalarReplAggregates" "SimplifyControlFlowGraph" -> "CFGSimplification" "SparseConditionalConstantPropagation" -> "SCCP" + "SuperwordLevelParallelismVectorize" -> "SLPVectorize" h -> h patchImpls = [ "AddressSanitizer", @@ -54,6 +55,7 @@ "LoopClosedSingleStaticAssignment", "LoopInstructionSimplify", "LoopStrengthReduce", + "LoopVectorize", "LowerAtomic", "LowerInvoke", "LowerSwitch", diff --git a/src/LLVM/General/Internal/PassManager.hs b/src/LLVM/General/Internal/PassManager.hs --- a/src/LLVM/General/Internal/PassManager.hs +++ b/src/LLVM/General/Internal/PassManager.hs @@ -54,8 +54,12 @@ sizeLevel :: Maybe Word, unitAtATime :: Maybe Bool, simplifyLibCalls :: Maybe Bool, + loopVectorize :: Maybe Bool, + superwordLevelParallelismVectorize :: Maybe Bool, useInlinerWithThreshold :: Maybe Word, - curatedTargetLibraryInfo :: Maybe TargetLibraryInfo + dataLayout :: Maybe DataLayout, + targetLibraryInfo :: Maybe TargetLibraryInfo, + targetMachine :: Maybe TargetMachine } -- | Helper to make a curated 'PassSetSpec' @@ -64,8 +68,12 @@ sizeLevel = Nothing, unitAtATime = Nothing, simplifyLibCalls = Nothing, + loopVectorize = Nothing, + superwordLevelParallelismVectorize = Nothing, useInlinerWithThreshold = Nothing, - curatedTargetLibraryInfo = Nothing + dataLayout = Nothing, + targetLibraryInfo = Nothing, + targetMachine = Nothing } -- | an empty 'PassSetSpec' @@ -82,6 +90,10 @@ createPassManager :: PassSetSpec -> IO (Ptr FFI.PassManager) createPassManager pss = flip runAnyContT return $ do pm <- liftIO $ FFI.createPassManager + forM_ (dataLayout pss) $ \dl -> liftIO $ withFFIDataLayout dl $ FFI.addDataLayoutPass pm + forM_ (targetLibraryInfo pss) $ \(TargetLibraryInfo tli) -> do + liftIO $ FFI.addTargetLibraryInfoPass pm tli + forM_ (targetMachine pss) $ \(TargetMachine tm) -> liftIO $ FFI.addAnalysisPasses tm pm case pss of s@CuratedPassSetSpec {} -> liftIO $ do bracket FFI.passManagerBuilderCreate FFI.passManagerBuilderDispose $ \b -> do @@ -91,16 +103,11 @@ handleOption FFI.passManagerBuilderSetDisableUnitAtATime (liftM not . unitAtATime) handleOption FFI.passManagerBuilderSetDisableSimplifyLibCalls (liftM not . simplifyLibCalls) handleOption FFI.passManagerBuilderUseInlinerWithThreshold useInlinerWithThreshold - case (curatedTargetLibraryInfo s) of - (Just (TargetLibraryInfo tl)) -> FFI.passManagerBuilderSetLibraryInfo b tl - Nothing -> return () + handleOption FFI.passManagerBuilderSetLoopVectorize loopVectorize + handleOption FFI.passManagerBuilderSetSuperwordLevelParallelismVectorize superwordLevelParallelismVectorize FFI.passManagerBuilderPopulateModulePassManager b pm PassSetSpec ps dl tli tm' -> do let tm = maybe nullPtr (\(TargetMachine tm) -> tm) tm' - forM_ tli $ \(TargetLibraryInfo tli) -> do - liftIO $ FFI.addTargetLibraryInfoPass pm tli - forM_ dl $ \dl -> liftIO $ withFFIDataLayout dl $ FFI.addDataLayoutPass pm - forM_ tm' $ \(TargetMachine tm) -> liftIO $ FFI.addAnalysisPasses tm pm forM_ ps $ \p -> $( do TH.TyConI (TH.DataD _ _ _ cons _) <- TH.reify ''Pass diff --git a/src/LLVM/General/Internal/Threading.hs b/src/LLVM/General/Internal/Threading.hs new file mode 100644 --- /dev/null +++ b/src/LLVM/General/Internal/Threading.hs @@ -0,0 +1,33 @@ +module LLVM.General.Internal.Threading ( + setMultithreaded, + isMultithreaded + ) where + +import Control.Concurrent.MVar +import System.IO.Unsafe +import Control.Monad + +import qualified LLVM.General.Internal.FFI.Threading as FFI + +import LLVM.General.Internal.Coding + +lock :: MVar () +{-# NOINLINE lock #-} +lock = unsafePerformIO $ newMVar () + +-- | Set the multithreading mode of LLVM. If it is disabled (as by default) locks are not enforced. +setMultithreaded :: Bool -> IO () +setMultithreaded s = do + takeMVar lock + multi <- isMultithreaded + case (s,multi) of + (True,False) -> do + success <- decodeM =<< FFI.startMultithreaded + when (not success) $ error "setMultithreaded: LLVM not built with threading support" + (False,True) -> FFI.stopMultithreaded + _ -> return () + putMVar lock () + +-- | Check if multithreading is enabled in LLVM +isMultithreaded :: IO Bool +isMultithreaded = FFI.isMultithreaded >>= decodeM diff --git a/src/LLVM/General/Threading.hs b/src/LLVM/General/Threading.hs new file mode 100644 --- /dev/null +++ b/src/LLVM/General/Threading.hs @@ -0,0 +1,7 @@ +-- | functionality necessary when running LLVM in multiple threads at the same time. +module LLVM.General.Threading ( + setMultithreaded, + isMultithreaded + ) where + +import LLVM.General.Internal.Threading diff --git a/src/LLVM/General/Transforms.hs b/src/LLVM/General/Transforms.hs --- a/src/LLVM/General/Transforms.hs +++ b/src/LLVM/General/Transforms.hs @@ -104,7 +104,10 @@ noMemoryOperationBoost :: Bool, fastDependencyAnalysis :: Bool } - | LoopVectorize + | LoopVectorize { + noUnrolling :: Bool + } + | SuperwordLevelParallelismVectorize -- here begin the instrumentation passes | GCOVProfiler { @@ -143,6 +146,11 @@ } | DebugExistingIR deriving (Eq, Ord, Read, Show, Typeable, Data) + +-- | Defaults for the 'LoopVectorize' pass +defaultLoopVectorize = LoopVectorize { + noUnrolling = False + } -- | Defaults for the 'BasicBlockVectorize' pass - copied from the C++ code to keep these defaults -- constant. (The C++ defaults are modifiable through global objects used for command-line processing, diff --git a/test/LLVM/General/Test/Optimization.hs b/test/LLVM/General/Test/Optimization.hs --- a/test/LLVM/General/Test/Optimization.hs +++ b/test/LLVM/General/Test/Optimization.hs @@ -251,7 +251,7 @@ withTargetOptions $ \targetOptions -> do withTargetMachine target triple "" Set.empty targetOptions R.Default CM.Default CGO.Default $ \tm -> do optimize (defaultPassSetSpec { - transforms = [ LoopVectorize ], + transforms = [ defaultLoopVectorize ], dataLayout = moduleDataLayout mIn, targetMachine = Just tm }) mIn