X-Git-Url: http://git.megacz.com/?p=ghc-hetmet.git;a=blobdiff_plain;f=compiler%2Fghci%2FRtClosureInspect.hs;h=e2a4f8e6088882806ac9d7b456530e181c552a69;hp=f289b14adedc6941c9a76ea00b0073cbb27fbfe9;hb=7fc749a43b4b6b85d234fa95d4928648259584f4;hpb=5c04842774b5ca60292762a9c89c23263496a556 diff --git a/compiler/ghci/RtClosureInspect.hs b/compiler/ghci/RtClosureInspect.hs index f289b14..e2a4f8e 100644 --- a/compiler/ghci/RtClosureInspect.hs +++ b/compiler/ghci/RtClosureInspect.hs @@ -6,11 +6,21 @@ -- ----------------------------------------------------------------------------- +{-# OPTIONS -w #-} +-- The above warning supression flag is a temporary kludge. +-- While working on this module you are encouraged to remove it and fix +-- any warnings in the module. See +-- http://hackage.haskell.org/trac/ghc/wiki/Commentary/CodingStyle#Warnings +-- for details + module RtClosureInspect( - cvObtainTerm, -- :: HscEnv -> Bool -> Maybe Type -> HValue -> IO Term + cvObtainTerm, -- :: HscEnv -> Int -> Bool -> Maybe Type -> HValue -> IO Term Term(..), + isTerm, + isSuspension, + isPrim, pprTerm, cPprTerm, cPprTermBase, @@ -23,55 +33,62 @@ module RtClosureInspect( isPointed, isFullyEvaluatedTerm, mapTermType, - termTyVars + termTyVars, -- unsafeDeepSeq, + cvReconstructType, + computeRTTIsubst, + sigmaType, + Closure(..), + getClosureData, + ClosureType(..), + isConstr, + isIndirection ) where #include "HsVersions.h" import ByteCodeItbls ( StgInfoTable ) import qualified ByteCodeItbls as BCI( StgInfoTable(..) ) -import ByteCodeLink ( HValue ) import HscTypes ( HscEnv ) +import Linker import DataCon import Type -import TcRnMonad ( TcM, initTcPrintErrors, ioToTcRn, recoverM, writeMutVar ) +import TcRnMonad ( TcM, initTc, ioToTcRn, + tryTcErrs) import TcType import TcMType import TcUnify import TcGadt +import TcEnv +import DriverPhases import TyCon -import Var import Name import VarEnv -import OccName +import Util import VarSet -import {-#SOURCE#-} TcRnDriver ( tcRnRecoverDataCon ) import TysPrim import PrelNames import TysWiredIn -import Constants ( wORD_SIZE ) +import Constants import Outputable import Maybes import Panic -import FiniteMap import GHC.Arr ( Array(..) ) -import GHC.Ptr ( Ptr(..), castPtr ) -import GHC.Exts -import GHC.Int ( Int32(..), Int64(..) ) -import GHC.Word ( Word32(..), Word64(..) ) +import GHC.Exts import Control.Monad import Data.Maybe import Data.Array.Base -import Data.List ( partition, nub ) -import Foreign.Storable - -import IO +import Data.List ( partition ) +import qualified Data.Sequence as Seq +import Data.Monoid +import Data.Sequence hiding (null, length, index, take, drop, splitAt, reverse) +import Foreign +import System.IO.Unsafe --------------------------------------------- -- * A representation of semi evaluated Terms @@ -89,12 +106,16 @@ import IO -} data Term = Term { ty :: Type - , dc :: DataCon + , dc :: Either String DataCon + -- The heap datacon. If ty is a newtype, + -- this is NOT the newtype datacon. + -- Empty if the datacon aint exported by the .hi + -- (private constructors in -O0 libraries) , val :: HValue , subTerms :: [Term] } | Prim { ty :: Type - , value :: String } + , value :: [Word] } | Suspension { ctype :: ClosureType , mb_ty :: Maybe Type @@ -102,6 +123,7 @@ data Term = Term { ty :: Type , bound_to :: Maybe Name -- Useful for printing } +isTerm, isSuspension, isPrim :: Term -> Bool isTerm Term{} = True isTerm _ = False isSuspension Suspension{} = True @@ -109,6 +131,7 @@ isSuspension _ = False isPrim Prim{} = True isPrim _ = False +termType :: Term -> Maybe Type termType t@(Suspension {}) = mb_ty t termType t = Just$ ty t @@ -138,7 +161,7 @@ data Closure = Closure { tipe :: ClosureType , infoPtr :: Ptr () , infoTable :: StgInfoTable , ptrs :: Array Int HValue - , nonPtrs :: ByteArray# + , nonPtrs :: [Word] } instance Outputable ClosureType where @@ -157,9 +180,13 @@ getClosureData a = (# iptr, ptrs, nptrs #) -> do itbl <- peek (Ptr iptr) let tipe = readCType (BCI.tipe itbl) - elems = BCI.ptrs itbl - ptrsList = Array 0 (fromIntegral$ elems) ptrs - ptrsList `seq` return (Closure tipe (Ptr iptr) itbl ptrsList nptrs) + elems = fromIntegral (BCI.ptrs itbl) + ptrsList = Array 0 (elems - 1) elems ptrs + nptrs_data = [W# (indexWordArray# nptrs i) + | I# i <- [0.. fromIntegral (BCI.nptrs itbl)] ] + ASSERT(fromIntegral elems >= 0) return () + ptrsList `seq` + return (Closure tipe (Ptr iptr) itbl ptrsList nptrs_data) readCType :: Integral a => a -> ClosureType readCType i @@ -174,7 +201,7 @@ readCType i | fromIntegral i == pAP_CODE = PAP | otherwise = Other (fromIntegral i) -isConstr, isIndirection :: ClosureType -> Bool +isConstr, isIndirection, isThunk :: ClosureType -> Bool isConstr Constr = True isConstr _ = False @@ -196,16 +223,16 @@ isFullyEvaluated a = do otherwise -> return False where amapM f = sequence . amap' f -amap' f (Array i0 i arr#) = map (\(I# i#) -> case indexArray# arr# i# of - (# e #) -> f e) - [0 .. i - i0] +amap' f (Array i0 i _ arr#) = map g [0 .. i - i0] + where g (I# i#) = case indexArray# arr# i# of + (# e #) -> f e -- TODO: Fix it. Probably the otherwise case is failing, trace/debug it {- unsafeDeepSeq :: a -> b -> b unsafeDeepSeq = unsafeDeepSeq1 2 where unsafeDeepSeq1 0 a b = seq a $! b - unsafeDeepSeq1 i a b -- 1st case avoids infinite loops for non reducible thunks + unsafeDeepSeq1 i a b -- 1st case avoids infinite loops for non reducible thunks | not (isConstr tipe) = seq a $! unsafeDeepSeq1 (i-1) a b -- | unsafePerformIO (isFullyEvaluated a) = b | otherwise = case unsafePerformIO (getClosureData a) of @@ -213,59 +240,31 @@ unsafeDeepSeq = unsafeDeepSeq1 2 where tipe = unsafePerformIO (getClosureType a) -} isPointed :: Type -> Bool -isPointed t | Just (t, _) <- splitTyConApp_maybe t = not$ isUnliftedTypeKind (tyConKind t) +isPointed t | Just (t, _) <- splitTyConApp_maybe t + = not$ isUnliftedTypeKind (tyConKind t) isPointed _ = True -#define MKDECODER(offset,cons,builder) (offset, show$ cons (builder addr 0#)) - -extractUnboxed :: [Type] -> ByteArray# -> [String] -extractUnboxed tt ba = helper tt (byteArrayContents# ba) - where helper :: [Type] -> Addr# -> [String] - helper (t:tt) addr - | Just ( tycon,_) <- splitTyConApp_maybe t - = let (offset, txt) = decode tycon addr - (I# word_offset) = offset*wORD_SIZE - in txt : helper tt (plusAddr# addr word_offset) - | otherwise - = -- ["extractUnboxed.helper: Urk. I got a " ++ showSDoc (ppr t)] - panic$ "extractUnboxed.helper: Urk. I got a " ++ showSDoc (ppr t) - helper [] addr = [] - decode :: TyCon -> Addr# -> (Int, String) - decode t addr - | t == charPrimTyCon = MKDECODER(1,C#,indexCharOffAddr#) - | t == intPrimTyCon = MKDECODER(1,I#,indexIntOffAddr#) - | t == wordPrimTyCon = MKDECODER(1,W#,indexWordOffAddr#) - | t == floatPrimTyCon = MKDECODER(1,F#,indexFloatOffAddr#) - | t == doublePrimTyCon = MKDECODER(2,D#,indexDoubleOffAddr#) - | t == int32PrimTyCon = MKDECODER(1,I32#,indexInt32OffAddr#) - | t == word32PrimTyCon = MKDECODER(1,W32#,indexWord32OffAddr#) - | t == int64PrimTyCon = MKDECODER(2,I64#,indexInt64OffAddr#) - | t == word64PrimTyCon = MKDECODER(2,W64#,indexWord64OffAddr#) - | t == addrPrimTyCon = MKDECODER(1,I#,(\x off-> addr2Int# (indexAddrOffAddr# x off))) --OPT Improve the presentation of addresses - | t == stablePtrPrimTyCon = (1, "") - | t == stableNamePrimTyCon = (1, "") - | t == statePrimTyCon = (1, "") - | t == realWorldTyCon = (1, "") - | t == threadIdPrimTyCon = (1, "") - | t == weakPrimTyCon = (1, "") - | t == arrayPrimTyCon = (1,"") - | t == byteArrayPrimTyCon = (1,"") - | t == mutableArrayPrimTyCon = (1, "") - | t == mutableByteArrayPrimTyCon = (1, "") - | t == mutVarPrimTyCon= (1, "") - | t == mVarPrimTyCon = (1, "") - | t == tVarPrimTyCon = (1, "") - | otherwise = (1, showSDoc (char '<' <> ppr t <> char '>')) - -- We cannot know the right offset in the otherwise case, so 1 is just a wild dangerous guess! - -- TODO: Improve the offset handling in decode (make it machine dependant) +extractUnboxed :: [Type] -> Closure -> [[Word]] +extractUnboxed tt clos = go tt (nonPtrs clos) + where sizeofType t + | Just (tycon,_) <- splitTyConApp_maybe t + = ASSERT (isPrimTyCon tycon) sizeofTyCon tycon + | otherwise = pprPanic "Expected a TcTyCon" (ppr t) + go [] _ = [] + go (t:tt) xx + | (x, rest) <- splitAt ((sizeofType t + wORD_SIZE - 1) `div` wORD_SIZE) xx + = x : go tt rest + +sizeofTyCon = sizeofPrimRep . tyConPrimRep ----------------------------------- -- * Traversals for Terms ----------------------------------- -data TermFold a = TermFold { fTerm :: Type -> DataCon -> HValue -> [a] -> a - , fPrim :: Type -> String -> a - , fSuspension :: ClosureType -> Maybe Type -> HValue -> Maybe Name -> a +data TermFold a = TermFold { fTerm :: Type -> Either String DataCon -> HValue -> [a] -> a + , fPrim :: Type -> [Word] -> a + , fSuspension :: ClosureType -> Maybe Type -> HValue + -> Maybe Name -> a } foldTerm :: TermFold a -> Term -> a @@ -286,11 +285,13 @@ idTermFoldM = TermFold { fSuspension = (((return.).).). Suspension } +mapTermType :: (Type -> Type) -> Term -> Term mapTermType f = foldTerm idTermFold { fTerm = \ty dc hval tt -> Term (f ty) dc hval tt, fSuspension = \ct mb_ty hval n -> Suspension ct (fmap f mb_ty) hval n } +termTyVars :: Term -> TyVarSet termTyVars = foldTerm TermFold { fTerm = \ty _ _ tt -> tyVarsOfType ty `plusVarEnv` concatVarEnv tt, @@ -302,73 +303,89 @@ termTyVars = foldTerm TermFold { -- Pretty printing of terms ---------------------------------- -app_prec::Int +app_prec,cons_prec ::Int app_prec = 10 +cons_prec = 5 -- TODO Extract this info from GHC itself -pprTerm :: Int -> Term -> SDoc -pprTerm p Term{dc=dc, subTerms=tt} -{- | dataConIsInfix dc, (t1:t2:tt') <- tt +pprTerm y p t | Just doc <- pprTermM y p t = doc + +pprTermM :: Monad m => (Int -> Term -> m SDoc) -> Int -> Term -> m SDoc +pprTermM y p t@Term{dc=Left dc_tag, subTerms=tt, ty=ty} = do + tt_docs <- mapM (y app_prec) tt + return$ cparen (not(null tt) && p >= app_prec) (text dc_tag <+> sep tt_docs) + +pprTermM y p t@Term{dc=Right dc, subTerms=tt, ty=ty} +{- | dataConIsInfix dc, (t1:t2:tt') <- tt --TODO fixity = parens (pprTerm1 True t1 <+> ppr dc <+> pprTerm1 True ppr t2) <+> hsep (map (pprTerm1 True) tt) --} - | null tt = ppr dc - | otherwise = cparen (p >= app_prec) - (ppr dc <+> sep (map (pprTerm app_prec) tt)) - - where fixity = undefined - -pprTerm _ t = pprTerm1 t - -pprTerm1 Prim{value=value} = text value -pprTerm1 t@Term{} = pprTerm 0 t -pprTerm1 Suspension{bound_to=Nothing} = char '_' -- <> ppr ct <> char '_' -pprTerm1 Suspension{mb_ty=Just ty, bound_to=Just n} - | Just _ <- splitFunTy_maybe ty = ptext SLIT("") - | otherwise = parens$ ppr n <> text "::" <> ppr ty - - -cPprTerm :: forall m. Monad m => ((Int->Term->m SDoc)->[Int->Term->m (Maybe SDoc)]) -> Term -> m SDoc +-} -- TODO Printing infix constructors properly + | null tt = return$ ppr dc + | Just (tc,_) <- splitNewTyConApp_maybe ty + , isNewTyCon tc + , Just new_dc <- maybeTyConSingleCon tc = do + real_value <- y 10 t{ty=repType ty} + return$ cparen (p >= app_prec) (ppr new_dc <+> real_value) + | otherwise = do + tt_docs <- mapM (y app_prec) tt + return$ cparen (p >= app_prec) (ppr dc <+> sep tt_docs) + +pprTermM y _ t = pprTermM1 y t +pprTermM1 _ Prim{value=words, ty=ty} = + return$ text$ repPrim (tyConAppTyCon ty) words +pprTermM1 y t@Term{} = panic "pprTermM1 - unreachable" +pprTermM1 _ Suspension{bound_to=Nothing} = return$ char '_' +pprTermM1 _ Suspension{mb_ty=Just ty, bound_to=Just n} + | Just _ <- splitFunTy_maybe ty = return$ ptext SLIT("") + | otherwise = return$ parens$ ppr n <> text "::" <> ppr ty + +-- Takes a list of custom printers with a explicit recursion knot and a term, +-- and returns the output of the first succesful printer, or the default printer +cPprTerm :: forall m. Monad m => + ((Int->Term->m SDoc)->[Int->Term->m (Maybe SDoc)]) -> Term -> m SDoc cPprTerm custom = go 0 where - go prec t@Term{subTerms=tt, dc=dc} = do - let mb_customDocs = map (($t) . ($prec)) (custom go) :: [m (Maybe SDoc)] - first_success <- firstJustM mb_customDocs - case first_success of - Just doc -> return$ cparen (prec>app_prec+1) doc --- | dataConIsInfix dc, (t1:t2:tt') <- tt = - Nothing -> do pprSubterms <- mapM (go (app_prec+1)) tt - return$ cparen (prec >= app_prec) - (ppr dc <+> sep pprSubterms) - go _ t = return$ pprTerm1 t + go prec t@Term{} = do + let default_ prec t = Just `liftM` pprTermM go prec t + mb_customDocs = [pp prec t | pp <- custom go ++ [default_]] + Just doc <- firstJustM mb_customDocs + return$ cparen (prec>app_prec+1) doc + go _ t = pprTermM1 go t firstJustM (mb:mbs) = mb >>= maybe (firstJustM mbs) (return . Just) firstJustM [] = return Nothing +-- Default set of custom printers. Note that the recursion knot is explicit cPprTermBase :: Monad m => (Int->Term-> m SDoc)->[Int->Term->m (Maybe SDoc)] -cPprTermBase pprP = +cPprTermBase y = [ - ifTerm isTupleDC (\_ -> liftM (parens . hcat . punctuate comma) - . mapM (pprP (-1)) . subTerms) - , ifTerm (isDC consDataCon) (\ p Term{subTerms=[h,t]} -> doList p h t) - , ifTerm (isDC intDataCon) (coerceShow$ \(a::Int)->a) - , ifTerm (isDC charDataCon) (coerceShow$ \(a::Char)->a) --- , ifTerm (isDC wordDataCon) (coerceShow$ \(a::Word)->a) - , ifTerm (isDC floatDataCon) (coerceShow$ \(a::Float)->a) - , ifTerm (isDC doubleDataCon) (coerceShow$ \(a::Double)->a) - , ifTerm isIntegerDC (coerceShow$ \(a::Integer)->a) + ifTerm isTupleTy (\_ -> liftM (parens . hcat . punctuate comma) + . mapM (y (-1)) . subTerms) + , ifTerm (\t -> isTyCon listTyCon t && subTerms t `lengthIs` 2) + (\ p Term{subTerms=[h,t]} -> doList p h t) + , ifTerm (isTyCon intTyCon) (coerceShow$ \(a::Int)->a) + , ifTerm (isTyCon charTyCon) (coerceShow$ \(a::Char)->a) +-- , ifTerm (isTyCon wordTyCon) (coerceShow$ \(a::Word)->a) + , ifTerm (isTyCon floatTyCon) (coerceShow$ \(a::Float)->a) + , ifTerm (isTyCon doubleTyCon) (coerceShow$ \(a::Double)->a) + , ifTerm isIntegerTy (coerceShow$ \(a::Integer)->a) ] - where ifTerm pred f p t = if pred t then liftM Just (f p t) else return Nothing - isIntegerDC Term{dc=dc} = - dataConName dc `elem` [ smallIntegerDataConName - , largeIntegerDataConName] - isTupleDC Term{dc=dc} = dc `elem` snd (unzip (elems boxedTupleArr)) - isDC a_dc Term{dc=dc} = a_dc == dc + where ifTerm pred f p t@Term{} | pred t = liftM Just (f p t) + ifTerm _ _ _ _ = return Nothing + isIntegerTy Term{ty=ty} = fromMaybe False $ do + (tc,_) <- splitTyConApp_maybe ty + return (tyConName tc == integerTyConName) + isTupleTy Term{ty=ty} = fromMaybe False $ do + (tc,_) <- splitTyConApp_maybe ty + return (tc `elem` (fst.unzip.elems) boxedTupleArr) + isTyCon a_tc Term{ty=ty} = fromMaybe False $ do + (tc,_) <- splitTyConApp_maybe ty + return (a_tc == tc) coerceShow f _ = return . text . show . f . unsafeCoerce# . val --TODO pprinting of list terms is not lazy doList p h t = do let elems = h : getListTerms t isConsLast = termType(last elems) /= termType h - print_elems <- mapM (pprP 5) elems + print_elems <- mapM (y cons_prec) elems return$ if isConsLast - then cparen (p >= 5) . hsep . punctuate (space<>colon) + then cparen (p >= cons_prec) . hsep . punctuate (space<>colon) $ print_elems else brackets (hcat$ punctuate comma print_elems) @@ -379,105 +396,81 @@ cPprTermBase pprP = getListTerms t@Suspension{} = [t] getListTerms t = pprPanic "getListTerms" (ppr t) + +repPrim :: TyCon -> [Word] -> String +repPrim t = rep where + rep x + | t == charPrimTyCon = show (build x :: Char) + | t == intPrimTyCon = show (build x :: Int) + | t == wordPrimTyCon = show (build x :: Word) + | t == floatPrimTyCon = show (build x :: Float) + | t == doublePrimTyCon = show (build x :: Double) + | t == int32PrimTyCon = show (build x :: Int32) + | t == word32PrimTyCon = show (build x :: Word32) + | t == int64PrimTyCon = show (build x :: Int64) + | t == word64PrimTyCon = show (build x :: Word64) + | t == addrPrimTyCon = show (nullPtr `plusPtr` build x) + | t == stablePtrPrimTyCon = "" + | t == stableNamePrimTyCon = "" + | t == statePrimTyCon = "" + | t == realWorldTyCon = "" + | t == threadIdPrimTyCon = "" + | t == weakPrimTyCon = "" + | t == arrayPrimTyCon = "" + | t == byteArrayPrimTyCon = "" + | t == mutableArrayPrimTyCon = "" + | t == mutableByteArrayPrimTyCon = "" + | t == mutVarPrimTyCon= "" + | t == mVarPrimTyCon = "" + | t == tVarPrimTyCon = "" + | otherwise = showSDoc (char '<' <> ppr t <> char '>') + where build ww = unsafePerformIO $ withArray ww (peek . castPtr) +-- This ^^^ relies on the representation of Haskell heap values being +-- the same as in a C array. + ----------------------------------- -- Type Reconstruction ----------------------------------- +{- +Type Reconstruction is type inference done on heap closures. +The algorithm walks the heap generating a set of equations, which +are solved with syntactic unification. +A type reconstruction equation looks like: + + = + +The full equation set is generated by traversing all the subterms, starting +from a given term. + +The only difficult part is that newtypes are only found in the lhs of equations. +Right hand sides are missing them. We can either (a) drop them from the lhs, or +(b) reconstruct them in the rhs when possible. + +The function congruenceNewtypes takes a shot at (b) +-} -- The Type Reconstruction monad type TR a = TcM a -runTR :: HscEnv -> TR Term -> IO Term +runTR :: HscEnv -> TR a -> IO a runTR hsc_env c = do - mb_term <- initTcPrintErrors hsc_env iNTERACTIVE c + mb_term <- runTR_maybe hsc_env c case mb_term of Nothing -> panic "Can't unify" - Just term -> return term + Just x -> return x + +runTR_maybe :: HscEnv -> TR a -> IO (Maybe a) +runTR_maybe hsc_env = fmap snd . initTc hsc_env HsSrcFile False iNTERACTIVE trIO :: IO a -> TR a trIO = liftTcM . ioToTcRn -addConstraint :: TcType -> TcType -> TR () -addConstraint t1 t2 = congruenceNewtypes t1 t2 >>= uncurry unifyType - -{- - A parallel fold over two Type values, - compensating for missing newtypes on both sides. - This is necessary because newtypes are not present - in runtime, but since sometimes there is evidence - available we do our best to reconstruct them. - Evidence can come from DataCon signatures or - from compile-time type inference. - I am using the words congruence and rewriting - because what we are doing here is an approximation - of unification modulo a set of equations, which would - come from newtype definitions. These should be the - equality coercions seen in System Fc. Rewriting - is performed, taking those equations as rules, - before launching unification. - - It doesn't make sense to rewrite everywhere, - or we would end up with all newtypes. So we rewrite - only in presence of evidence. - The lhs comes from the heap structure of ptrs,nptrs. - The rhs comes from a DataCon type signature. - Rewriting in the rhs is restricted to the result type. - - Note that it is very tricky to make this 'rewriting' - work with the unification implemented by TcM, where - substitutions are 'inlined'. The order in which - constraints are unified is vital for this (or I am - using TcM wrongly). --} -congruenceNewtypes :: TcType -> TcType -> TcM (TcType,TcType) -congruenceNewtypes = go True - where - go rewriteRHS lhs rhs - -- TyVar lhs inductive case - | Just tv <- getTyVar_maybe lhs - = recoverM (return (lhs,rhs)) $ do - Indirect ty_v <- readMetaTyVar tv - (lhs', rhs') <- go rewriteRHS ty_v rhs - writeMutVar (metaTvRef tv) (Indirect lhs') - return (lhs, rhs') - -- TyVar rhs inductive case - | Just tv <- getTyVar_maybe rhs - = recoverM (return (lhs,rhs)) $ do - Indirect ty_v <- readMetaTyVar tv - (lhs', rhs') <- go rewriteRHS lhs ty_v - writeMutVar (metaTvRef tv) (Indirect rhs') - return (lhs', rhs) --- FunTy inductive case - | Just (l1,l2) <- splitFunTy_maybe lhs - , Just (r1,r2) <- splitFunTy_maybe rhs - = do (l2',r2') <- go True l2 r2 - (l1',r1') <- go False l1 r1 - return (mkFunTy l1' l2', mkFunTy r1' r2') --- TyconApp Inductive case; this is the interesting bit. - | Just (tycon_l, args_l) <- splitNewTyConApp_maybe lhs - , Just (tycon_r, args_r) <- splitNewTyConApp_maybe rhs = do - - let (tycon_l',args_l') = if isNewTyCon tycon_r && not(isNewTyCon tycon_l) - then (tycon_r, rewrite tycon_r lhs) - else (tycon_l, args_l) - (tycon_r',args_r') = if rewriteRHS && isNewTyCon tycon_l && not(isNewTyCon tycon_r) - then (tycon_l, rewrite tycon_l rhs) - else (tycon_r, args_r) - (args_l'', args_r'') <- unzip `liftM` zipWithM (go rewriteRHS) args_l' args_r' - return (mkTyConApp tycon_l' args_l'', mkTyConApp tycon_r' args_r'') - - | otherwise = return (lhs,rhs) - - where rewrite newtyped_tc lame_tipe - | (tvs, tipe) <- newTyConRep newtyped_tc - = case tcUnifyTys (const BindMe) [tipe] [lame_tipe] of - Just subst -> substTys subst (map mkTyVarTy tvs) - otherwise -> panic "congruenceNewtypes: Can't unify a newtype" - -newVar :: Kind -> TR TcTyVar -newVar = liftTcM . newFlexiTyVar - +liftTcM :: TcM a -> TR a liftTcM = id +newVar :: Kind -> TR TcType +newVar = liftTcM . fmap mkTyVarTy . newFlexiTyVar + -- | Returns the instantiated type scheme ty', and the substitution sigma -- such that sigma(ty') = ty instScheme :: Type -> TR (TcType, TvSubst) @@ -485,70 +478,100 @@ instScheme ty | (tvs, rho) <- tcSplitForAllTys ty = liftTcM$ do (tvs',theta,ty') <- tcInstType (mapM tcInstTyVar) ty return (ty', zipTopTvSubst tvs' (mkTyVarTys tvs)) -cvObtainTerm :: HscEnv -> Bool -> Maybe Type -> HValue -> IO Term -cvObtainTerm hsc_env force mb_ty hval = runTR hsc_env $ do - tv <- liftM mkTyVarTy (newVar argTypeKind) +-- Adds a constraint of the form t1 == t2 +-- t1 is expected to come from walking the heap +-- t2 is expected to come from a datacon signature +-- Before unification, congruenceNewtypes needs to +-- do its magic. +addConstraint :: TcType -> TcType -> TR () +addConstraint t1 t2 = congruenceNewtypes t1 t2 >>= uncurry unifyType + >> return () -- TOMDO: what about the coercion? + -- we should consider family instances + +-- Type & Term reconstruction +cvObtainTerm :: HscEnv -> Int -> Bool -> Maybe Type -> HValue -> IO Term +cvObtainTerm hsc_env bound force mb_ty hval = runTR hsc_env $ do + tv <- newVar argTypeKind case mb_ty of - Nothing -> go tv tv hval >>= zonkTerm - Just ty | isMonomorphic ty -> go ty ty hval >>= zonkTerm + Nothing -> go bound tv tv hval >>= zonkTerm + Just ty | isMonomorphic ty -> go bound ty ty hval >>= zonkTerm Just ty -> do (ty',rev_subst) <- instScheme (sigmaType ty) addConstraint tv ty' - term <- go tv tv hval >>= zonkTerm + term <- go bound tv tv hval >>= zonkTerm --restore original Tyvars return$ mapTermType (substTy rev_subst) term where - go tv ty a = do - let monomorphic = not(isTyVarTy tv) -- This is a convention. The ancestor tests for - -- monomorphism and passes a type instead of a tv + go bound _ _ _ | seq bound False = undefined + go 0 tv ty a = do + clos <- trIO $ getClosureData a + return (Suspension (tipe clos) (Just tv) a Nothing) + go bound tv ty a = do + let monomorphic = not(isTyVarTy tv) + -- This ^^^ is a convention. The ancestor tests for + -- monomorphism and passes a type instead of a tv clos <- trIO $ getClosureData a case tipe clos of -- Thunks we may want to force -- NB. this won't attempt to force a BLACKHOLE. Even with :force, we never -- force blackholes, because it would almost certainly result in deadlock, -- and showing the '_' is more useful. - t | isThunk t && force -> seq a $ go tv ty a + t | isThunk t && force -> seq a $ go (pred bound) tv ty a -- We always follow indirections - Indirection _ -> go tv ty $! (ptrs clos ! 0) + Indirection _ -> go (pred bound) tv ty $! (ptrs clos ! 0) -- The interesting case Constr -> do - m_dc <- trIO$ tcRnRecoverDataCon hsc_env (infoPtr clos) - case m_dc of - Nothing -> panic "Can't find the DataCon for a term" + Right dcname <- dataConInfoPtrToName (infoPtr clos) + (_,mb_dc) <- tryTcErrs (tcLookupDataCon dcname) + case mb_dc of + Nothing -> do -- This can happen for private constructors compiled -O0 + -- where the .hi descriptor does not export them + -- In such case, we return a best approximation: + -- ignore the unpointed args, and recover the pointeds + -- This preserves laziness, and should be safe. + let tag = showSDoc (ppr dcname) + vars <- replicateM (length$ elems$ ptrs clos) + (newVar (liftedTypeKind)) + subTerms <- sequence [appArr (go (pred bound) tv tv) (ptrs clos) i + | (i, tv) <- zip [0..] vars] + return (Term tv (Left ('<' : tag ++ ">")) a subTerms) Just dc -> do - let extra_args = length(dataConRepArgTys dc) - length(dataConOrigArgTys dc) + let extra_args = length(dataConRepArgTys dc) - + length(dataConOrigArgTys dc) subTtypes = matchSubTypes dc ty (subTtypesP, subTtypesNP) = partition isPointed subTtypes subTermTvs <- sequence - [ if isMonomorphic t then return t else (mkTyVarTy `fmap` newVar k) + [ if isMonomorphic t then return t + else (newVar k) | (t,k) <- zip subTtypesP (map typeKind subTtypesP)] - -- It is vital for newtype reconstruction that the unification step is done - -- right here, _before_ the subterms are RTTI reconstructed. + -- It is vital for newtype reconstruction that the unification step + -- is done right here, _before_ the subterms are RTTI reconstructed when (not monomorphic) $ do - let myType = mkFunTys (reOrderTerms subTermTvs subTtypesNP subTtypes) tv - instScheme(dataConRepType dc) >>= addConstraint myType . fst - subTermsP <- sequence $ drop extra_args -- all extra arguments are pointed - [ appArr (go tv t) (ptrs clos) i + let myType = mkFunTys (reOrderTerms subTermTvs + subTtypesNP + subTtypes) + tv + (signatureType,_) <- instScheme(dataConRepType dc) + addConstraint myType signatureType + subTermsP <- sequence $ drop extra_args + -- ^^^ all extra arguments are pointed + [ appArr (go (pred bound) tv t) (ptrs clos) i | (i,tv,t) <- zip3 [0..] subTermTvs subTtypesP] - let unboxeds = extractUnboxed subTtypesNP (nonPtrs clos) + let unboxeds = extractUnboxed subTtypesNP clos subTermsNP = map (uncurry Prim) (zip subTtypesNP unboxeds) - subTerms = reOrderTerms subTermsP subTermsNP (drop extra_args subTtypes) - return (Term tv dc a subTerms) + subTerms = reOrderTerms subTermsP subTermsNP + (drop extra_args subTtypes) + return (Term tv (Right dc) a subTerms) -- The otherwise case: can be a Thunk,AP,PAP,etc. - otherwise -> - return (Suspension (tipe clos) (Just tv) a Nothing) - --- Access the array of pointers and recurse down. Needs to be done with --- care of no introducing a thunk! or go will fail to do its job - appArr f arr (I# i#) = case arr of - (Array _ _ ptrs#) -> case indexArray# ptrs# i# of - (# e #) -> f e + tipe_clos -> + return (Suspension tipe_clos (Just tv) a Nothing) +-- matchSubTypes dc ty | pprTrace "matchSubtypes" (ppr dc <+> ppr ty) False = undefined matchSubTypes dc ty | Just (_,ty_args) <- splitTyConApp_maybe (repType ty) - , null (dataConExTyVars dc) --TODO Handle the case of extra existential tyvars +-- assumption: ^^^ looks through newtypes + , isVanillaDataCon dc --TODO non-vanilla case = dataConInstArgTys dc ty_args - | otherwise = dataConRepArgTys dc -- This is used to put together pointed and nonpointed subterms in the @@ -556,14 +579,182 @@ cvObtainTerm hsc_env force mb_ty hval = runTR hsc_env $ do reOrderTerms _ _ [] = [] reOrderTerms pointed unpointed (ty:tys) | isPointed ty = ASSERT2(not(null pointed) - , ptext SLIT("reOrderTerms") $$ (ppr pointed $$ ppr unpointed)) + , ptext SLIT("reOrderTerms") $$ + (ppr pointed $$ ppr unpointed)) head pointed : reOrderTerms (tail pointed) unpointed tys | otherwise = ASSERT2(not(null unpointed) - , ptext SLIT("reOrderTerms") $$ (ppr pointed $$ ppr unpointed)) + , ptext SLIT("reOrderTerms") $$ + (ppr pointed $$ ppr unpointed)) head unpointed : reOrderTerms pointed (tail unpointed) tys -isMonomorphic ty | isForAllTy ty = False -isMonomorphic ty = (isEmptyVarSet . tyVarsOfType) ty + + +-- Fast, breadth-first Type reconstruction +max_depth = 10 :: Int +cvReconstructType :: HscEnv -> Bool -> Maybe Type -> HValue -> IO (Maybe Type) +cvReconstructType hsc_env force mb_ty hval = runTR_maybe hsc_env $ do + tv <- newVar argTypeKind + case mb_ty of + Nothing -> do search (isMonomorphic `fmap` zonkTcType tv) + (uncurry go) + (Seq.singleton (tv, hval)) + max_depth + zonkTcType tv -- TODO untested! + Just ty | isMonomorphic ty -> return ty + Just ty -> do + (ty',rev_subst) <- instScheme (sigmaType ty) + addConstraint tv ty' + search (isMonomorphic `fmap` zonkTcType tv) + (\(ty,a) -> go ty a) + (Seq.singleton (tv, hval)) + max_depth + substTy rev_subst `fmap` zonkTcType tv + where +-- search :: m Bool -> ([a] -> [a] -> [a]) -> [a] -> m () + search stop expand l depth | Seq.null l = return () + search stop expand x 0 = fail$ "Failed to reconstruct a type after " ++ + show max_depth ++ " steps" + search stop expand l d | x :< xx <- viewl l = unlessM stop $ do + new <- expand x + search stop expand (xx `mappend` Seq.fromList new) $! (pred d) + + -- returns unification tasks,since we are going to want a breadth-first search + go :: Type -> HValue -> TR [(Type, HValue)] + go tv a = do + clos <- trIO $ getClosureData a + case tipe clos of + Indirection _ -> go tv $! (ptrs clos ! 0) + Constr -> do + Right dcname <- dataConInfoPtrToName (infoPtr clos) + (_,mb_dc) <- tryTcErrs (tcLookupDataCon dcname) + case mb_dc of + Nothing-> do + -- TODO: Check this case + vars <- replicateM (length$ elems$ ptrs clos) + (newVar (liftedTypeKind)) + subTerms <- sequence [ appArr (go tv) (ptrs clos) i + | (i, tv) <- zip [0..] vars] + forM [0..length (elems $ ptrs clos)] $ \i -> do + tv <- newVar liftedTypeKind + return$ appArr (\e->(tv,e)) (ptrs clos) i + + Just dc -> do + let extra_args = length(dataConRepArgTys dc) - + length(dataConOrigArgTys dc) + subTtypes <- mapMif (not . isMonomorphic) + (\t -> newVar (typeKind t)) + (dataConRepArgTys dc) + + -- It is vital for newtype reconstruction that the unification step + -- is done right here, _before_ the subterms are RTTI reconstructed + let myType = mkFunTys subTtypes tv + (signatureType,_) <- instScheme(dataConRepType dc) + addConstraint myType signatureType + return $ [ appArr (\e->(t,e)) (ptrs clos) i + | (i,t) <- drop extra_args $ + zip [0..] (filter isPointed subTtypes)] + otherwise -> return [] + + -- This helper computes the difference between a base type t and the + -- improved rtti_t computed by RTTI + -- The main difference between RTTI types and their normal counterparts + -- is that the former are _not_ polymorphic, thus polymorphism must + -- be stripped. Syntactically, forall's must be stripped +computeRTTIsubst ty rtti_ty = + -- In addition, we strip newtypes too, since the reconstructed type might + -- not have recovered them all + tcUnifyTys (const BindMe) + [repType' $ dropForAlls$ ty] + [repType' $ rtti_ty] +-- TODO stripping newtypes shouldn't be necessary, test + + +-- Dealing with newtypes +{- + A parallel fold over two Type values, + compensating for missing newtypes on both sides. + This is necessary because newtypes are not present + in runtime, but since sometimes there is evidence + available we do our best to reconstruct them. + Evidence can come from DataCon signatures or + from compile-time type inference. + I am using the words congruence and rewriting + because what we are doing here is an approximation + of unification modulo a set of equations, which would + come from newtype definitions. These should be the + equality coercions seen in System Fc. Rewriting + is performed, taking those equations as rules, + before launching unification. + + It doesn't make sense to rewrite everywhere, + or we would end up with all newtypes. So we rewrite + only in presence of evidence. + The lhs comes from the heap structure of ptrs,nptrs. + The rhs comes from a DataCon type signature. + Rewriting in the rhs is restricted to the result type. + + Note that it is very tricky to make this 'rewriting' + work with the unification implemented by TcM, where + substitutions are 'inlined'. The order in which + constraints are unified is vital for this (or I am + using TcM wrongly). +-} +congruenceNewtypes :: TcType -> TcType -> TcM (TcType,TcType) +congruenceNewtypes lhs rhs + -- TyVar lhs inductive case + | Just tv <- getTyVar_maybe lhs + = recoverTc (return (lhs,rhs)) $ do + Indirect ty_v <- readMetaTyVar tv + (lhs1, rhs1) <- congruenceNewtypes ty_v rhs + return (lhs, rhs1) +-- FunTy inductive case + | Just (l1,l2) <- splitFunTy_maybe lhs + , Just (r1,r2) <- splitFunTy_maybe rhs + = do (l2',r2') <- congruenceNewtypes l2 r2 + (l1',r1') <- congruenceNewtypes l1 r1 + return (mkFunTy l1' l2', mkFunTy r1' r2') +-- TyconApp Inductive case; this is the interesting bit. + | Just (tycon_l, args_l) <- splitNewTyConApp_maybe lhs + , Just (tycon_r, args_r) <- splitNewTyConApp_maybe rhs + , tycon_l /= tycon_r + = return (lhs, upgrade tycon_l rhs) + + | otherwise = return (lhs,rhs) + + where upgrade :: TyCon -> Type -> Type + upgrade new_tycon ty + | not (isNewTyCon new_tycon) = ty + | ty' <- mkTyConApp new_tycon (map mkTyVarTy $ tyConTyVars new_tycon) + , Just subst <- tcUnifyTys (const BindMe) [ty] [repType ty'] + = substTy subst ty' + -- assumes that reptype doesn't touch tyconApp args ^^^ + + +-------------------------------------------------------------------------------- +-- Semantically different to recoverM in TcRnMonad +-- recoverM retains the errors in the first action, +-- whereas recoverTc here does not +recoverTc recover thing = do + (_,mb_res) <- tryTcErrs thing + case mb_res of + Nothing -> recover + Just res -> return res + +isMonomorphic ty | (tvs, ty') <- splitForAllTys ty + = null tvs && (isEmptyVarSet . tyVarsOfType) ty' + +mapMif :: Monad m => (a -> Bool) -> (a -> m a) -> [a] -> m [a] +mapMif pred f xx = sequence $ mapMif_ pred f xx +mapMif_ pred f [] = [] +mapMif_ pred f (x:xx) = (if pred x then f x else return x) : mapMif_ pred f xx + +unlessM condM acc = condM >>= \c -> unless c acc + +-- Strict application of f at index i +appArr f a@(Array _ _ _ ptrs#) i@(I# i#) + = ASSERT (i < length(elems a)) + case indexArray# ptrs# i# of + (# e #) -> f e zonkTerm :: Term -> TcM Term zonkTerm = foldTerm idTermFoldM { @@ -578,53 +769,4 @@ zonkTerm = foldTerm idTermFoldM { -- Generalize the type: find all free tyvars and wrap in the appropiate ForAll. sigmaType ty = mkForAllTys (varSetElems$ tyVarsOfType (dropForAlls ty)) ty -{- -Example of Type Reconstruction --------------------------------- -Suppose we have an existential type such as - -data Opaque = forall a. Opaque a - -And we have a term built as: - -t = Opaque (map Just [[1,1],[2,2]]) -The type of t as far as the typechecker goes is t :: Opaque -If we seq the head of t, we obtain: - -t - O (_1::a) - -seq _1 () - -t - O ( (_3::b) : (_4::[b]) ) - -seq _3 () - -t - O ( (Just (_5::c)) : (_4::[b]) ) - -At this point, we know that b = (Maybe c) - -seq _5 () - -t - O ( (Just ((_6::d) : (_7::[d]) )) : (_4::[b]) ) - -At this point, we know that c = [d] - -seq _6 () - -t - O ( (Just (1 : (_7::[d]) )) : (_4::[b]) ) - -At this point, we know that d = Integer - -The fully reconstructed expressions, with propagation, would be: - -t - O ( (Just (_5::c)) : (_4::[Maybe c]) ) -t - O ( (Just ((_6::d) : (_7::[d]) )) : (_4::[Maybe [d]]) ) -t - O ( (Just (1 : (_7::[Integer]) )) : (_4::[Maybe [Integer]]) ) - - -For reference, the type of the thing inside the opaque is -map Just [[1,1],[2,2]] :: [Maybe [Integer]] - -NOTE: (Num t) contexts have been manually replaced by Integer for clarity --}