X-Git-Url: http://git.megacz.com/?p=ghc-hetmet.git;a=blobdiff_plain;f=compiler%2Fghci%2FRtClosureInspect.hs;h=b4068a7aa76ecb92908c8f1c9a15e5eb2a3db80a;hp=e8157ac734606ea1bb7111dbd1bc81429d82b500;hb=35a1ec430a5e44a9bc79d385b997422c20cb427b;hpb=cb429c8ac482f3b294f709b5ba50423fdf1f35b0 diff --git a/compiler/ghci/RtClosureInspect.hs b/compiler/ghci/RtClosureInspect.hs index e8157ac..b4068a7 100644 --- a/compiler/ghci/RtClosureInspect.hs +++ b/compiler/ghci/RtClosureInspect.hs @@ -7,122 +7,131 @@ ----------------------------------------------------------------------------- module RtClosureInspect( - - cvObtainTerm, -- :: HscEnv -> Bool -> Maybe Type -> HValue -> IO Term - - ClosureType(..), - getClosureData, -- :: a -> IO Closure - Closure ( tipe, infoPtr, ptrs, nonPtrs ), - isConstr, -- :: ClosureType -> Bool - isIndirection, -- :: ClosureType -> Bool - - Term(..), - pprTerm, - cPprTerm, - cPprTermBase, - termType, - foldTerm, - TermFold(..), - idTermFold, - idTermFoldM, - isFullyEvaluated, - isPointed, - isFullyEvaluatedTerm, --- unsafeDeepSeq, - ) where + cvObtainTerm, -- :: HscEnv -> Int -> Bool -> Maybe Type -> HValue -> IO Term + cvReconstructType, + improveRTTIType, + + Term(..), + isTerm, isSuspension, isPrim, isFun, isFunLike, isNewtypeWrap, + isFullyEvaluated, isFullyEvaluatedTerm, + termType, mapTermType, termTyVars, + foldTerm, TermFold(..), foldTermM, TermFoldM(..), idTermFold, + pprTerm, cPprTerm, cPprTermBase, CustomTermPrinter, + +-- unsafeDeepSeq, + + 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 HscTypes +import Linker -import DataCon -import Type -import TcRnMonad ( TcM, initTcPrintErrors, ioToTcRn, recoverM, writeMutVar ) +import DataCon +import Type +import qualified Unify as U +import TypeRep -- I know I know, this is cheating +import Var +import TcRnMonad import TcType import TcMType import TcUnify -import TcGadt -import TyCon -import Var -import Name +import TcEnv + +import TyCon +import Name import VarEnv -import OccName +import Util +import ListSetOps import VarSet -import {-#SOURCE#-} TcRnDriver ( tcRnRecoverDataCon ) - -import TysPrim +import TysPrim import PrelNames import TysWiredIn +import DynFlags +import Outputable +import FastString +-- import Panic import Constants ( wORD_SIZE ) -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 GHC.IO ( IO(..) ) import Control.Monad import Data.Maybe import Data.Array.Base -import Data.List ( partition ) -import Foreign.Storable - -import IO +import Data.Ix +import Data.List +import qualified Data.Sequence as Seq +import Data.Monoid +import Data.Sequence (viewl, ViewL(..)) +import Foreign hiding (unsafePerformIO) +import System.IO.Unsafe --------------------------------------------- -- * A representation of semi evaluated Terms --------------------------------------------- -{- - A few examples in this representation: - > Just 10 = Term Data.Maybe Data.Maybe.Just (Just 10) [Term Int I# (10) "10"] - - > (('a',_,_),_,('b',_,_)) = - Term ((Char,b,c),d,(Char,e,f)) (,,) (('a',_,_),_,('b',_,_)) - [ Term (Char, b, c) (,,) ('a',_,_) [Term Char C# "a", Suspension, Suspension] - , Suspension - , Term (Char, e, f) (,,) ('b',_,_) [Term Char C# "b", Suspension, Suspension]] --} - -data Term = Term { ty :: Type - , dc :: DataCon +data Term = Term { ty :: RttiType + , dc :: Either String DataCon + -- Carries a text representation if the datacon is + -- not exported by the .hi file, which is the case + -- for private constructors in -O0 compiled libraries , val :: HValue , subTerms :: [Term] } - | Prim { ty :: Type - , value :: String } + | Prim { ty :: RttiType + , value :: [Word] } | Suspension { ctype :: ClosureType - , mb_ty :: Maybe Type + , ty :: RttiType , val :: HValue , bound_to :: Maybe Name -- Useful for printing } - + | NewtypeWrap{ -- At runtime there are no newtypes, and hence no + -- newtype constructors. A NewtypeWrap is just a + -- made-up tag saying "heads up, there used to be + -- a newtype constructor here". + ty :: RttiType + , dc :: Either String DataCon + , wrapped_term :: Term } + | RefWrap { -- The contents of a reference + ty :: RttiType + , wrapped_term :: Term } + +isTerm, isSuspension, isPrim, isFun, isFunLike, isNewtypeWrap :: Term -> Bool isTerm Term{} = True isTerm _ = False isSuspension Suspension{} = True isSuspension _ = False isPrim Prim{} = True isPrim _ = False +isNewtypeWrap NewtypeWrap{} = True +isNewtypeWrap _ = False + +isFun Suspension{ctype=Fun} = True +isFun _ = False + +isFunLike s@Suspension{ty=ty} = isFun s || isFunTy ty +isFunLike _ = False -termType t@(Suspension {}) = mb_ty t -termType t = Just$ ty t +termType :: Term -> RttiType +termType t = ty t isFullyEvaluatedTerm :: Term -> Bool isFullyEvaluatedTerm Term {subTerms=tt} = all isFullyEvaluatedTerm tt -isFullyEvaluatedTerm Suspension {} = False isFullyEvaluatedTerm Prim {} = True +isFullyEvaluatedTerm NewtypeWrap{wrapped_term=t} = isFullyEvaluatedTerm t +isFullyEvaluatedTerm RefWrap{wrapped_term=t} = isFullyEvaluatedTerm t +isFullyEvaluatedTerm _ = False instance Outputable (Term) where - ppr = head . cPprTerm cPprTermBase + ppr t | Just doc <- cPprTerm cPprTermBase t = doc + | otherwise = panic "Outputable Term instance" ------------------------------------------------------------------------- -- Runtime Closure Datatype and functions for retrieving closure related stuff @@ -135,21 +144,24 @@ data ClosureType = Constr | AP | PAP | Indirection Int - | Other Int + | MutVar Int + | MVar Int + | Other Int deriving (Show, Eq) data Closure = Closure { tipe :: ClosureType , infoPtr :: Ptr () , infoTable :: StgInfoTable , ptrs :: Array Int HValue - , nonPtrs :: ByteArray# + , nonPtrs :: [Word] } instance Outputable ClosureType where ppr = text . show -#include "../includes/ClosureTypes.h" +#include "../includes/rts/storage/ClosureTypes.h" +aP_CODE, pAP_CODE :: Int aP_CODE = AP pAP_CODE = PAP #undef AP @@ -159,474 +171,1058 @@ getClosureData :: a -> IO Closure getClosureData a = case unpackClosure# a of (# iptr, ptrs, nptrs #) -> do - itbl <- peek (Ptr iptr) + let iptr' + | ghciTablesNextToCode = + Ptr iptr + | otherwise = + -- the info pointer we get back from unpackClosure# + -- is to the beginning of the standard info table, + -- but the Storable instance for info tables takes + -- into account the extra entry pointer when + -- !ghciTablesNextToCode, so we must adjust here: + Ptr iptr `plusPtr` negate wORD_SIZE + itbl <- peek 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(elems >= 0) return () + ptrsList `seq` + return (Closure tipe (Ptr iptr) itbl ptrsList nptrs_data) readCType :: Integral a => a -> ClosureType -readCType i +readCType i | i >= CONSTR && i <= CONSTR_NOCAF_STATIC = Constr | i >= FUN && i <= FUN_STATIC = Fun - | i >= THUNK && i < THUNK_SELECTOR = Thunk (fromIntegral i) + | i >= THUNK && i < THUNK_SELECTOR = Thunk i' | i == THUNK_SELECTOR = ThunkSelector | i == BLACKHOLE = Blackhole - | i >= IND && i <= IND_STATIC = Indirection (fromIntegral i) - | fromIntegral i == aP_CODE = AP - | fromIntegral i == pAP_CODE = PAP - | otherwise = Other (fromIntegral i) - -isConstr, isIndirection :: ClosureType -> Bool + | i >= IND && i <= IND_STATIC = Indirection i' + | i' == aP_CODE = AP + | i == AP_STACK = AP + | i' == pAP_CODE = PAP + | i == MUT_VAR_CLEAN || i == MUT_VAR_DIRTY= MutVar i' + | i == MVAR_CLEAN || i == MVAR_DIRTY = MVar i' + | otherwise = Other i' + where i' = fromIntegral i + +isConstr, isIndirection, isThunk :: ClosureType -> Bool isConstr Constr = True isConstr _ = False isIndirection (Indirection _) = True ---isIndirection ThunkSelector = True isIndirection _ = False +isThunk (Thunk _) = True +isThunk ThunkSelector = True +isThunk AP = True +isThunk _ = False + isFullyEvaluated :: a -> IO Bool isFullyEvaluated a = do closure <- getClosureData a case tipe closure of Constr -> do are_subs_evaluated <- amapM isFullyEvaluated (ptrs closure) return$ and are_subs_evaluated - otherwise -> return False + _ -> 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] - -- 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 $! unsafeDeepSeq1 (i-1) a b -- | unsafePerformIO (isFullyEvaluated a) = b | otherwise = case unsafePerformIO (getClosureData a) of closure -> foldl' (flip unsafeDeepSeq) b (ptrs closure) where tipe = unsafePerformIO (getClosureType a) -} -isPointed :: Type -> Bool -isPointed t | Just (t, _) <- splitTyConApp_maybe t = not$ 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) ----------------------------------- -- * Traversals for Terms ----------------------------------- +type TermProcessor a b = RttiType -> Either String DataCon -> HValue -> [a] -> b + +data TermFold a = TermFold { fTerm :: TermProcessor a a + , fPrim :: RttiType -> [Word] -> a + , fSuspension :: ClosureType -> RttiType -> HValue + -> Maybe Name -> a + , fNewtypeWrap :: RttiType -> Either String DataCon + -> a -> a + , fRefWrap :: RttiType -> a -> a + } + -data TermFold a = TermFold { fTerm :: Type -> DataCon -> HValue -> [a] -> a - , fPrim :: Type -> String -> a - , fSuspension :: ClosureType -> Maybe Type -> HValue -> Maybe Name -> a +data TermFoldM m a = + TermFoldM {fTermM :: TermProcessor a (m a) + , fPrimM :: RttiType -> [Word] -> m a + , fSuspensionM :: ClosureType -> RttiType -> HValue + -> Maybe Name -> m a + , fNewtypeWrapM :: RttiType -> Either String DataCon + -> a -> m a + , fRefWrapM :: RttiType -> a -> m a } foldTerm :: TermFold a -> Term -> a foldTerm tf (Term ty dc v tt) = fTerm tf ty dc v (map (foldTerm tf) tt) foldTerm tf (Prim ty v ) = fPrim tf ty v foldTerm tf (Suspension ct ty v b) = fSuspension tf ct ty v b +foldTerm tf (NewtypeWrap ty dc t) = fNewtypeWrap tf ty dc (foldTerm tf t) +foldTerm tf (RefWrap ty t) = fRefWrap tf ty (foldTerm tf t) + + +foldTermM :: Monad m => TermFoldM m a -> Term -> m a +foldTermM tf (Term ty dc v tt) = mapM (foldTermM tf) tt >>= fTermM tf ty dc v +foldTermM tf (Prim ty v ) = fPrimM tf ty v +foldTermM tf (Suspension ct ty v b) = fSuspensionM tf ct ty v b +foldTermM tf (NewtypeWrap ty dc t) = foldTermM tf t >>= fNewtypeWrapM tf ty dc +foldTermM tf (RefWrap ty t) = foldTermM tf t >>= fRefWrapM tf ty idTermFold :: TermFold Term idTermFold = TermFold { fTerm = Term, fPrim = Prim, - fSuspension = Suspension + fSuspension = Suspension, + fNewtypeWrap = NewtypeWrap, + fRefWrap = RefWrap } -idTermFoldM :: Monad m => TermFold (m Term) -idTermFoldM = TermFold { - fTerm = \ty dc v tt -> sequence tt >>= return . Term ty dc v, - fPrim = (return.). Prim, - fSuspension = (((return.).).). Suspension - } + +mapTermType :: (RttiType -> Type) -> Term -> Term +mapTermType f = foldTerm idTermFold { + fTerm = \ty dc hval tt -> Term (f ty) dc hval tt, + fSuspension = \ct ty hval n -> + Suspension ct (f ty) hval n, + fNewtypeWrap= \ty dc t -> NewtypeWrap (f ty) dc t, + fRefWrap = \ty t -> RefWrap (f ty) t} + +mapTermTypeM :: Monad m => (RttiType -> m Type) -> Term -> m Term +mapTermTypeM f = foldTermM TermFoldM { + fTermM = \ty dc hval tt -> f ty >>= \ty' -> return $ Term ty' dc hval tt, + fPrimM = (return.) . Prim, + fSuspensionM = \ct ty hval n -> + f ty >>= \ty' -> return $ Suspension ct ty' hval n, + fNewtypeWrapM= \ty dc t -> f ty >>= \ty' -> return $ NewtypeWrap ty' dc t, + fRefWrapM = \ty t -> f ty >>= \ty' -> return $ RefWrap ty' t} + +termTyVars :: Term -> TyVarSet +termTyVars = foldTerm TermFold { + fTerm = \ty _ _ tt -> + tyVarsOfType ty `plusVarEnv` concatVarEnv tt, + fSuspension = \_ ty _ _ -> tyVarsOfType ty, + fPrim = \ _ _ -> emptyVarEnv, + fNewtypeWrap= \ty _ t -> tyVarsOfType ty `plusVarEnv` t, + fRefWrap = \ty t -> tyVarsOfType ty `plusVarEnv` t} + where concatVarEnv = foldr plusVarEnv emptyVarEnv ---------------------------------- -- Pretty printing of terms ---------------------------------- -app_prec::Int -app_prec = 10 +type Precedence = Int +type TermPrinter = Precedence -> Term -> SDoc +type TermPrinterM m = Precedence -> Term -> m SDoc + +app_prec,cons_prec, max_prec ::Int +max_prec = 10 +app_prec = max_prec +cons_prec = 5 -- TODO Extract this info from GHC itself + +pprTerm :: TermPrinter -> TermPrinter +pprTerm y p t | Just doc <- pprTermM (\p -> Just . y p) p t = doc +pprTerm _ _ _ = panic "pprTerm" + +pprTermM, ppr_termM, pprNewtypeWrap :: Monad m => TermPrinterM m -> TermPrinterM m +pprTermM y p t = pprDeeper `liftM` ppr_termM y p t + +ppr_termM y p Term{dc=Left dc_tag, subTerms=tt} = do + tt_docs <- mapM (y app_prec) tt + return$ cparen (not(null tt) && p >= app_prec) (text dc_tag <+> pprDeeperList fsep tt_docs) + +ppr_termM y p Term{dc=Right dc, subTerms=tt} +{- | dataConIsInfix dc, (t1:t2:tt') <- tt --TODO fixity + = parens (ppr_term1 True t1 <+> ppr dc <+> ppr_term1 True ppr t2) + <+> hsep (map (ppr_term1 True) tt) +-} -- TODO Printing infix constructors properly + | null tt = return$ ppr dc + | otherwise = do + tt_docs <- mapM (y app_prec) tt + return$ cparen (p >= app_prec) (ppr dc <+> pprDeeperList fsep tt_docs) + +ppr_termM y p t@NewtypeWrap{} = pprNewtypeWrap y p t +ppr_termM y p RefWrap{wrapped_term=t} = do + contents <- y app_prec t + return$ cparen (p >= app_prec) (text "GHC.Prim.MutVar#" <+> contents) + -- The constructor name is wired in here ^^^ for the sake of simplicity. + -- I don't think mutvars are going to change in a near future. + -- In any case this is solely a presentation matter: MutVar# is + -- a datatype with no constructors, implemented by the RTS + -- (hence there is no way to obtain a datacon and print it). +ppr_termM _ _ t = ppr_termM1 t + + +ppr_termM1 :: Monad m => Term -> m SDoc +ppr_termM1 Prim{value=words, ty=ty} = + return$ text$ repPrim (tyConAppTyCon ty) words +ppr_termM1 Suspension{ty=ty, bound_to=Nothing} = + return (char '_' <+> ifPprDebug (text "::" <> ppr ty)) +ppr_termM1 Suspension{ty=ty, bound_to=Just n} +-- | Just _ <- splitFunTy_maybe ty = return$ ptext (sLit("") + | otherwise = return$ parens$ ppr n <> text "::" <> ppr ty +ppr_termM1 Term{} = panic "ppr_termM1 - Term" +ppr_termM1 RefWrap{} = panic "ppr_termM1 - RefWrap" +ppr_termM1 NewtypeWrap{} = panic "ppr_termM1 - NewtypeWrap" + +pprNewtypeWrap y p NewtypeWrap{ty=ty, wrapped_term=t} + | Just (tc,_) <- tcSplitTyConApp_maybe ty + , ASSERT(isNewTyCon tc) True + , Just new_dc <- tyConSingleDataCon_maybe tc = do + real_term <- y max_prec t + return $ cparen (p >= app_prec) (ppr new_dc <+> real_term) +pprNewtypeWrap _ _ _ = panic "pprNewtypeWrap" + +------------------------------------------------------- +-- Custom Term Pretty Printers +------------------------------------------------------- + +-- We can want to customize the representation of a +-- term depending on its type. +-- However, note that custom printers have to work with +-- type representations, instead of directly with types. +-- We cannot use type classes here, unless we employ some +-- typerep trickery (e.g. Weirich's RepLib tricks), +-- which I didn't. Therefore, this code replicates a lot +-- of what type classes provide for free. + +type CustomTermPrinter m = TermPrinterM m + -> [Precedence -> Term -> (m (Maybe SDoc))] + +-- | 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 :: Monad m => CustomTermPrinter m -> Term -> m SDoc +cPprTerm printers_ = go 0 where + printers = printers_ go + go prec t = do + let default_ = Just `liftM` pprTermM go prec t + mb_customDocs = [pp prec t | pp <- printers] ++ [default_] + Just doc <- firstJustM mb_customDocs + return$ cparen (prec>app_prec+1) doc -pprTerm :: Int -> Term -> SDoc -pprTerm p Term{dc=dc, subTerms=tt} -{- | dataConIsInfix dc, (t1:t2:tt') <- tt - = 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 -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 firstJustM (mb:mbs) = mb >>= maybe (firstJustM mbs) (return . Just) firstJustM [] = return Nothing -cPprTermBase :: Monad m => (Int->Term-> m SDoc)->[Int->Term->m (Maybe SDoc)] -cPprTermBase pprP = - [ - 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) - ] - 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 - 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 +-- Default set of custom printers. Note that the recursion knot is explicit +cPprTermBase :: Monad m => CustomTermPrinter m +cPprTermBase y = + [ ifTerm (isTupleTy.ty) (\_p -> liftM (parens . hcat . punctuate comma) + . mapM (y (-1)) + . subTerms) + , ifTerm (\t -> isTyCon listTyCon (ty t) && subTerms t `lengthIs` 2) + (\ p t -> doList p t) + , ifTerm (isTyCon intTyCon . ty) (coerceShow$ \(a::Int)->a) + , ifTerm (isTyCon charTyCon . ty) (coerceShow$ \(a::Char)->a) + , ifTerm (isTyCon floatTyCon . ty) (coerceShow$ \(a::Float)->a) + , ifTerm (isTyCon doubleTyCon . ty) (coerceShow$ \(a::Double)->a) + , ifTerm (isIntegerTy . ty) (coerceShow$ \(a::Integer)->a) + ] + where ifTerm pred f prec t@Term{} + | pred t = Just `liftM` f prec t + ifTerm _ _ _ _ = return Nothing + + isTupleTy ty = fromMaybe False $ do + (tc,_) <- tcSplitTyConApp_maybe ty + return (isBoxedTupleTyCon tc) + + isTyCon a_tc ty = fromMaybe False $ do + (tc,_) <- tcSplitTyConApp_maybe ty + return (a_tc == tc) + + isIntegerTy ty = fromMaybe False $ do + (tc,_) <- tcSplitTyConApp_maybe ty + return (tyConName tc == integerTyConName) + + coerceShow f _p = return . text . show . f . unsafeCoerce# . val + + --Note pprinting of list terms is not lazy + doList p (Term{subTerms=[h,t]}) = do + let elems = h : getListTerms t + isConsLast = not(termType(last elems) `coreEqType` termType h) + print_elems <- mapM (y cons_prec) elems return$ if isConsLast - then cparen (p >= 5) . hsep . punctuate (space<>colon) - $ print_elems - else brackets (hcat$ punctuate comma print_elems) - - where Just a /= Just b = not (a `coreEqType` b) - _ /= _ = True - getListTerms Term{subTerms=[h,t]} = h : getListTerms t - getListTerms t@Term{subTerms=[]} = [] + then cparen (p >= cons_prec) + . pprDeeperList fsep + . punctuate (space<>colon) + $ print_elems + else brackets (pprDeeperList fcat$ + punctuate comma print_elems) + + where getListTerms Term{subTerms=[h,t]} = h : getListTerms t + getListTerms Term{subTerms=[]} = [] getListTerms t@Suspension{} = [t] getListTerms t = pprPanic "getListTerms" (ppr t) + doList _ _ = panic "doList" + + +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 Type Reconstruction monad -type TR a = TcM a + = -runTR :: HscEnv -> TR Term -> IO Term -runTR hsc_env c = do - mb_term <- initTcPrintErrors hsc_env iNTERACTIVE (c >>= zonkTerm) - case mb_term of - Nothing -> panic "Can't unify" - Just term -> return term +The full equation set is generated by traversing all the subterms, starting +from a given term. -trIO :: IO a -> TR a -trIO = liftTcM . ioToTcRn +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. -addConstraint :: TcType -> TcType -> TR () -addConstraint t1 t2 = congruenceNewtypes t1 t2 >>= uncurry unifyType +The function congruenceNewtypes takes a shot at (b) +-} -{- - 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 +-- A (non-mutable) tau type containing +-- existentially quantified tyvars. +-- (since GHC type language currently does not support +-- existentials, we leave these variables unquantified) +type RttiType = Type + +-- An incomplete type as stored in GHCi: +-- no polymorphism: no quantifiers & all tyvars are skolem. +type GhciType = Type + + +-- The Type Reconstruction monad +-------------------------------- +type TR a = TcM a - 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'') +runTR :: HscEnv -> TR a -> IO a +runTR hsc_env thing = do + mb_val <- runTR_maybe hsc_env thing + case mb_val of + Nothing -> error "unable to :print the term" + Just x -> return x - | otherwise = return (lhs,rhs) +runTR_maybe :: HscEnv -> TR a -> IO (Maybe a) +runTR_maybe hsc_env = fmap snd . initTc hsc_env HsSrcFile False iNTERACTIVE - 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" +traceTR :: SDoc -> TR () +traceTR = liftTcM . traceOptTcRn Opt_D_dump_rtti -newVar :: Kind -> TR TcTyVar -newVar = liftTcM . newFlexiTyVar +-- Semantically different to recoverM in TcRnMonad +-- recoverM retains the errors in the first action, +-- whereas recoverTc here does not +recoverTR :: TR a -> TR a -> TR a +recoverTR recover thing = do + (_,mb_res) <- tryTcErrs thing + case mb_res of + Nothing -> recover + Just res -> return res + +trIO :: IO a -> TR a +trIO = liftTcM . liftIO + +liftTcM :: TcM a -> TR a liftTcM = id --- | Returns the instantiated type scheme ty', and the substitution sigma --- such that sigma(ty') = ty -instScheme :: Type -> TR (TcType, TvSubst) -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 a = do - -- Obtain the term and tidy the type before returning it - term <- cvObtainTerm1 hsc_env force mb_ty a - return $ tidyTypes term - where - tidyTypes = foldTerm idTermFold { - fTerm = \ty dc hval tt -> Term (tidy ty) dc hval tt, - fSuspension = \ct mb_ty hval n -> - Suspension ct (fmap tidy mb_ty) hval n - } - tidy ty = tidyType (emptyTidyOccEnv, tidyVarEnv ty) ty - tidyVarEnv ty = mkVarEnv$ - [ (v, setTyVarName v (tyVarName tv)) - | (tv,v) <- zip alphaTyVars vars] - where vars = varSetElems$ tyVarsOfType ty - -cvObtainTerm1 :: HscEnv -> Bool -> Maybe Type -> HValue -> IO Term -cvObtainTerm1 hsc_env force mb_ty hval = runTR hsc_env $ do - tv <- liftM mkTyVarTy (newVar argTypeKind) - case mb_ty of - Nothing -> go tv tv hval - Just ty | isMonomorphic ty -> go ty ty hval - Just ty -> do - (ty',rev_subst) <- instScheme (sigmaType$ fromJust mb_ty) - addConstraint tv ty' - term <- go tv tv hval - --restore original Tyvars - return$ flip foldTerm term idTermFold { - fTerm = \ty dc hval tt -> Term (substTy rev_subst ty) dc hval tt, - fSuspension = \ct mb_ty hval n -> - Suspension ct (substTy rev_subst `fmap` mb_ty) hval n} +newVar :: Kind -> TR TcType +newVar = liftTcM . newFlexiTyVarTy + +type RttiInstantiation = [(TcTyVar, TyVar)] + -- Associates the typechecker-world meta type variables + -- (which are mutable and may be refined), to their + -- debugger-world RuntimeUnk counterparts. + -- If the TcTyVar has not been refined by the runtime type + -- elaboration, then we want to turn it back into the + -- original RuntimeUnk + +-- | Returns the instantiated type scheme ty', and the +-- mapping from new (instantiated) -to- old (skolem) type variables +instScheme :: QuantifiedType -> TR (TcType, RttiInstantiation) +instScheme (tvs, ty) + = liftTcM $ do { (tvs', _, subst) <- tcInstTyVars tvs + ; let rtti_inst = [(tv',tv) | (tv',tv) <- tvs' `zip` tvs] + ; return (substTy subst ty, rtti_inst) } + +applyRevSubst :: RttiInstantiation -> TR () +-- Apply the *reverse* substitution in-place to any un-filled-in +-- meta tyvars. This recovers the original debugger-world variable +-- unless it has been refined by new information from the heap +applyRevSubst pairs = liftTcM (mapM_ do_pair pairs) + where + do_pair (tc_tv, rtti_tv) + = do { tc_ty <- zonkTcTyVar tc_tv + ; case tcGetTyVar_maybe tc_ty of + Just tv | isMetaTyVar tv -> writeMetaTyVar tv (mkTyVarTy rtti_tv) + _ -> return () } + +-- 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 actual expected = do + traceTR (text "add constraint:" <+> fsep [ppr actual, equals, ppr expected]) + recoverTR (traceTR $ fsep [text "Failed to unify", ppr actual, + text "with", ppr expected]) $ + do { (ty1, ty2) <- congruenceNewtypes actual expected + ; _ <- captureConstraints $ unifyType ty1 ty2 + ; return () } + -- TOMDO: what about the coercion? + -- we should consider family instances + + +-- Type & Term reconstruction +------------------------------ +cvObtainTerm :: HscEnv -> Int -> Bool -> RttiType -> HValue -> IO Term +cvObtainTerm hsc_env max_depth force old_ty hval = runTR hsc_env $ do + -- we quantify existential tyvars as universal, + -- as this is needed to be able to manipulate + -- them properly + let quant_old_ty@(old_tvs, old_tau) = quantifyType old_ty + sigma_old_ty = mkForAllTys old_tvs old_tau + traceTR (text "Term reconstruction started with initial type " <> ppr old_ty) + term <- + if null old_tvs + then do + term <- go max_depth sigma_old_ty sigma_old_ty hval + term' <- zonkTerm term + return $ fixFunDictionaries $ expandNewtypes term' + else do + (old_ty', rev_subst) <- instScheme quant_old_ty + my_ty <- newVar argTypeKind + when (check1 quant_old_ty) (traceTR (text "check1 passed") >> + addConstraint my_ty old_ty') + term <- go max_depth my_ty sigma_old_ty hval + new_ty <- zonkTcType (termType term) + if isMonomorphic new_ty || check2 (quantifyType new_ty) quant_old_ty + then do + traceTR (text "check2 passed") + addConstraint new_ty old_ty' + applyRevSubst rev_subst + zterm' <- zonkTerm term + return ((fixFunDictionaries . expandNewtypes) zterm') + else do + traceTR (text "check2 failed" <+> parens + (ppr term <+> text "::" <+> ppr new_ty)) + -- we have unsound types. Replace constructor types in + -- subterms with tyvars + zterm' <- mapTermTypeM + (\ty -> case tcSplitTyConApp_maybe ty of + Just (tc, _:_) | tc /= funTyCon + -> newVar argTypeKind + _ -> return ty) + term + zonkTerm zterm' + traceTR (text "Term reconstruction completed." $$ + text "Term obtained: " <> ppr term $$ + text "Type obtained: " <> ppr (termType term)) + return 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 :: Int -> Type -> Type -> HValue -> TcM Term + go max_depth _ _ _ | seq max_depth False = undefined + go 0 my_ty _old_ty a = do + traceTR (text "Gave up reconstructing a term after" <> + int max_depth <> text " steps") + clos <- trIO $ getClosureData a + return (Suspension (tipe clos) my_ty a Nothing) + go max_depth my_ty old_ty a = do + let monomorphic = not(isTyVarTy my_ty) + -- 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 - Thunk _ | force -> seq a $ go tv ty a --- We always follow indirections - Indirection _ -> go tv ty $! (ptrs clos ! 0) + t | isThunk t && force -> traceTR (text "Forcing a " <> text (show t)) >> + seq a (go (pred max_depth) my_ty old_ty a) +-- Blackholes are indirections iff the payload is not TSO or BLOCKING_QUEUE. So we +-- treat them like indirections; if the payload is TSO or BLOCKING_QUEUE, we'll end up +-- showing '_' which is what we want. + Blackhole -> do traceTR (text "Following a BLACKHOLE") + appArr (go max_depth my_ty old_ty) (ptrs clos) 0 +-- We always follow indirections + Indirection i -> do traceTR (text "Following an indirection" <> parens (int i) ) + go max_depth my_ty old_ty $! (ptrs clos ! 0) +-- We also follow references + MutVar _ | Just (tycon,[world,contents_ty]) <- tcSplitTyConApp_maybe old_ty + -> do + -- Deal with the MutVar# primitive + -- It does not have a constructor at all, + -- so we simulate the following one + -- MutVar# :: contents_ty -> MutVar# s contents_ty + traceTR (text "Following a MutVar") + contents_tv <- newVar liftedTypeKind + contents <- trIO$ IO$ \w -> readMutVar# (unsafeCoerce# a) w + ASSERT(isUnliftedTypeKind $ typeKind my_ty) return () + (mutvar_ty,_) <- instScheme $ quantifyType $ mkFunTy + contents_ty (mkTyConApp tycon [world,contents_ty]) + addConstraint (mkFunTy contents_tv my_ty) mutvar_ty + x <- go (pred max_depth) contents_tv contents_ty contents + return (RefWrap my_ty x) + -- 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" - Just dc -> do - 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) - | (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. + traceTR (text "entering a constructor " <> + if monomorphic + then parens (text "already monomorphic: " <> ppr my_ty) + else Outputable.empty) + 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 max_depth) tv tv) (ptrs clos) i + | (i, tv) <- zip [0..] vars] + return (Term my_ty (Left ('<' : tag ++ ">")) a subTerms) + Just dc -> do + let subTtypes = matchSubTypes dc old_ty + subTermTvs <- mapMif (not . isMonomorphic) + (\t -> newVar (typeKind t)) + subTtypes + let (subTermsP, subTermsNP) = partition (\(ty,_) -> isLifted ty + || isRefType ty) + (zip subTtypes subTermTvs) + (subTtypesP, subTermTvsP ) = unzip subTermsP + (subTtypesNP, _subTermTvsNP) = unzip subTermsNP + + -- When we already have all the information, avoid solving + -- unnecessary constraints. Propagation of type information + -- to subterms is already being done via matching. 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 - | (i,tv,t) <- zip3 [0..] subTermTvs subTtypesP] - let unboxeds = extractUnboxed subTtypesNP (nonPtrs clos) - subTermsNP = map (uncurry Prim) (zip subTtypesNP unboxeds) - subTerms = reOrderTerms subTermsP subTermsNP (drop extra_args subTtypes) - return (Term tv dc a subTerms) + let myType = mkFunTys subTermTvs my_ty + (signatureType,_) <- instScheme (mydataConType dc) + -- It is vital for newtype reconstruction that the unification step + -- is done right here, _before_ the subterms are RTTI reconstructed + addConstraint myType signatureType + subTermsP <- sequence + [ appArr (go (pred max_depth) tv t) (ptrs clos) i + | (i,tv,t) <- zip3 [0..] subTermTvsP subTtypesP] + let unboxeds = extractUnboxed subTtypesNP clos + subTermsNP = map (uncurry Prim) (zip subTtypesNP unboxeds) + subTerms = reOrderTerms subTermsP subTermsNP subTtypes + return (Term my_ty (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 my_ty a Nothing) matchSubTypes dc ty - | Just (_,ty_args) <- splitTyConApp_maybe (repType ty) - , null (dataConExTyVars dc) --TODO Handle the case of extra existential tyvars - = dataConInstArgTys dc ty_args - + | ty' <- repType ty -- look through newtypes + , Just (tc,ty_args) <- tcSplitTyConApp_maybe ty' + , dc `elem` tyConDataCons tc + -- It is necessary to check that dc is actually a constructor for tycon tc, + -- because it may be the case that tc is a recursive newtype and tcSplitTyConApp + -- has not removed it. In that case, we happily give up and don't match + = myDataConInstArgTys dc ty_args | otherwise = dataConRepArgTys dc --- This is used to put together pointed and nonpointed subterms in the --- correct order. + -- put together pointed and nonpointed subterms in the + -- correct order. reOrderTerms _ _ [] = [] reOrderTerms pointed unpointed (ty:tys) - | isPointed ty = ASSERT2(not(null pointed) - , ptext SLIT("reOrderTerms") $$ (ppr pointed $$ ppr unpointed)) - head pointed : reOrderTerms (tail pointed) unpointed tys + | isLifted ty || isRefType ty + = ASSERT2(not(null pointed) + , ptext (sLit "reOrderTerms") $$ + (ppr pointed $$ ppr unpointed)) + let (t:tt) = pointed in t : reOrderTerms tt unpointed tys | otherwise = ASSERT2(not(null unpointed) - , ptext SLIT("reOrderTerms") $$ (ppr pointed $$ ppr unpointed)) - head unpointed : reOrderTerms pointed (tail unpointed) tys - -isMonomorphic = isEmptyVarSet . tyVarsOfType - -zonkTerm :: Term -> TcM Term -zonkTerm = foldTerm idTermFoldM { - fTerm = \ty dc v tt -> sequence tt >>= \tt -> - zonkTcType ty >>= \ty' -> - return (Term ty' dc v tt) - ,fSuspension = \ct ty v b -> fmapMMaybe zonkTcType ty >>= \ty -> - return (Suspension ct ty v b)} - - --- Is this defined elsewhere? --- Generalize the type: find all free tyvars and wrap in the appropiate ForAll. -sigmaType ty = mkForAllTys (varSetElems$ tyVarsOfType (dropForAlls ty)) ty + , ptext (sLit "reOrderTerms") $$ + (ppr pointed $$ ppr unpointed)) + let (t:tt) = unpointed in t : reOrderTerms pointed tt tys + + -- insert NewtypeWraps around newtypes + expandNewtypes = foldTerm idTermFold { fTerm = worker } where + worker ty dc hval tt + | Just (tc, args) <- tcSplitTyConApp_maybe ty + , isNewTyCon tc + , wrapped_type <- newTyConInstRhs tc args + , Just dc' <- tyConSingleDataCon_maybe tc + , t' <- worker wrapped_type dc hval tt + = NewtypeWrap ty (Right dc') t' + | otherwise = Term ty dc hval tt + + + -- Avoid returning types where predicates have been expanded to dictionaries. + fixFunDictionaries = foldTerm idTermFold {fSuspension = worker} where + worker ct ty hval n | isFunTy ty = Suspension ct (dictsView ty) hval n + | otherwise = Suspension ct ty hval n + + +-- Fast, breadth-first Type reconstruction +------------------------------------------ +cvReconstructType :: HscEnv -> Int -> GhciType -> HValue -> IO (Maybe Type) +cvReconstructType hsc_env max_depth old_ty hval = runTR_maybe hsc_env $ do + traceTR (text "RTTI started with initial type " <> ppr old_ty) + let sigma_old_ty@(old_tvs, _) = quantifyType old_ty + new_ty <- + if null old_tvs + then return old_ty + else do + (old_ty', rev_subst) <- instScheme sigma_old_ty + my_ty <- newVar argTypeKind + when (check1 sigma_old_ty) (traceTR (text "check1 passed") >> + addConstraint my_ty old_ty') + search (isMonomorphic `fmap` zonkTcType my_ty) + (\(ty,a) -> go ty a) + (Seq.singleton (my_ty, hval)) + max_depth + new_ty <- zonkTcType my_ty + if isMonomorphic new_ty || check2 (quantifyType new_ty) sigma_old_ty + then do + traceTR (text "check2 passed" <+> ppr old_ty $$ ppr new_ty) + addConstraint my_ty old_ty' + applyRevSubst rev_subst + zonkRttiType new_ty + else traceTR (text "check2 failed" <+> parens (ppr new_ty)) >> + return old_ty + traceTR (text "RTTI completed. Type obtained:" <+> ppr new_ty) + return new_ty + where +-- search :: m Bool -> ([a] -> [a] -> [a]) -> [a] -> m () + search _ _ _ 0 = traceTR (text "Failed to reconstruct a type after " <> + int max_depth <> text " steps") + search stop expand l d = + case viewl l of + EmptyL -> return () + x :< xx -> 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 my_ty a = do + clos <- trIO $ getClosureData a + case tipe clos of + Blackhole -> appArr (go my_ty) (ptrs clos) 0 -- carefully, don't eval the TSO + Indirection _ -> go my_ty $! (ptrs clos ! 0) + MutVar _ -> do + contents <- trIO$ IO$ \w -> readMutVar# (unsafeCoerce# a) w + tv' <- newVar liftedTypeKind + world <- newVar liftedTypeKind + addConstraint my_ty (mkTyConApp mutVarPrimTyCon [world,tv']) + return [(tv', contents)] + Constr -> do + Right dcname <- dataConInfoPtrToName (infoPtr clos) + (_,mb_dc) <- tryTcErrs (tcLookupDataCon dcname) + case mb_dc of + Nothing-> do + -- TODO: Check this case + forM [0..length (elems $ ptrs clos)] $ \i -> do + tv <- newVar liftedTypeKind + return$ appArr (\e->(tv,e)) (ptrs clos) i + + Just dc -> do + 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 my_ty + (signatureType,_) <- instScheme (mydataConType dc) + addConstraint myType signatureType + return $ [ appArr (\e->(t,e)) (ptrs clos) i + | (i,t) <- zip [0..] (filter (isLifted |.| isRefType) subTtypes)] + _ -> return [] + +-- Compute the difference between a base type and the type found by RTTI +-- improveType +-- The types can contain skolem type variables, which need to be treated as normal vars. +-- In particular, we want them to unify with things. +improveRTTIType :: HscEnv -> RttiType -> RttiType -> Maybe TvSubst +improveRTTIType _ base_ty new_ty + = U.tcUnifyTys (const U.BindMe) [base_ty] [new_ty] + +myDataConInstArgTys :: DataCon -> [Type] -> [Type] +myDataConInstArgTys dc args + | null (dataConExTyVars dc) && null (dataConEqTheta dc) = dataConInstArgTys dc args + | otherwise = dataConRepArgTys dc +mydataConType :: DataCon -> QuantifiedType +-- ^ Custom version of DataCon.dataConUserType where we +-- - remove the equality constraints +-- - use the representation types for arguments, including dictionaries +-- - keep the original result type +mydataConType dc + = ( (univ_tvs `minusList` map fst eq_spec) ++ ex_tvs + , mkFunTys arg_tys res_ty ) + where univ_tvs = dataConUnivTyVars dc + ex_tvs = dataConExTyVars dc + eq_spec = dataConEqSpec dc + arg_tys = [case a of + PredTy p -> predTypeRep p + _ -> a + | a <- dataConRepArgTys dc] + res_ty = dataConOrigResTy dc + +isRefType :: Type -> Bool +isRefType ty + | Just (tc, _) <- tcSplitTyConApp_maybe ty' = isRefTyCon tc + | otherwise = False + where ty'= repType ty + +isRefTyCon :: TyCon -> Bool +isRefTyCon tc = tc `elem` [mutVarPrimTyCon, mVarPrimTyCon, tVarPrimTyCon] + +-- Soundness checks +-------------------- {- -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: +This is not formalized anywhere, so hold to your seats! +RTTI in the presence of newtypes can be a tricky and unsound business. -t = Opaque (map Just [[1,1],[2,2]]) +Example: +~~~~~~~~~ +Suppose we are doing RTTI for a partially evaluated +closure t, the real type of which is t :: MkT Int, for + + newtype MkT a = MkT [Maybe a] + +The table below shows the results of RTTI and the improvement +calculated for different combinations of evaluatedness and :type t. +Regard the two first columns as input and the next two as output. + + # | t | :type t | rtti(t) | improv. | result + ------------------------------------------------------------ + 1 | _ | t b | a | none | OK + 2 | _ | MkT b | a | none | OK + 3 | _ | t Int | a | none | OK + + If t is not evaluated at *all*, we are safe. + + 4 | (_ : _) | t b | [a] | t = [] | UNSOUND + 5 | (_ : _) | MkT b | MkT a | none | OK (compensating for the missing newtype) + 6 | (_ : _) | t Int | [Int] | t = [] | UNSOUND + + If a is a minimal whnf, we run into trouble. Note that + row 5 above does newtype enrichment on the ty_rtty parameter. + + 7 | (Just _:_)| t b |[Maybe a] | t = [], | UNSOUND + | | | b = Maybe a| + + 8 | (Just _:_)| MkT b | MkT a | none | OK + 9 | (Just _:_)| t Int | FAIL | none | OK + + And if t is any more evaluated than whnf, we are still in trouble. + Because constraints are solved in top-down order, when we reach the + Maybe subterm what we got is already unsound. This explains why the + row 9 fails to complete. + + 10 | (Just _:_)| t Int | [Maybe a] | FAIL | OK + 11 | (Just 1:_)| t Int | [Maybe Int] | FAIL | OK + + We can undo the failure in row 9 by leaving out the constraint + coming from the type signature of t (i.e., the 2nd column). + Note that this type information is still used + to calculate the improvement. But we fail + when trying to calculate the improvement, as there is no unifier for + t Int = [Maybe a] or t Int = [Maybe Int]. + + + Another set of examples with t :: [MkT (Maybe Int)] \equiv [[Maybe (Maybe Int)]] + + # | t | :type t | rtti(t) | improvement | result + --------------------------------------------------------------------- + 1 |(Just _:_) | [t (Maybe a)] | [[Maybe b]] | t = [] | + | | | | b = Maybe a | + +The checks: +~~~~~~~~~~~ +Consider a function obtainType that takes a value and a type and produces +the Term representation and a substitution (the improvement). +Assume an auxiliar rtti' function which does the actual job if recovering +the type, but which may produce a false type. + +In pseudocode: + + rtti' :: a -> IO Type -- Does not use the static type information + + obtainType :: a -> Type -> IO (Maybe (Term, Improvement)) + obtainType v old_ty = do + rtti_ty <- rtti' v + if monomorphic rtti_ty || (check rtti_ty old_ty) + then ... + else return Nothing + where check rtti_ty old_ty = check1 rtti_ty && + check2 rtti_ty old_ty + + check1 :: Type -> Bool + check2 :: Type -> Type -> Bool + +Now, if rtti' returns a monomorphic type, we are safe. +If that is not the case, then we consider two conditions. + + +1. To prevent the class of unsoundness displayed by + rows 4 and 7 in the example: no higher kind tyvars + accepted. + + check1 (t a) = NO + check1 (t Int) = NO + check1 ([] a) = YES + +2. To prevent the class of unsoundness shown by row 6, + the rtti type should be structurally more + defined than the old type we are comparing it to. + check2 :: NewType -> OldType -> Bool + check2 a _ = True + check2 [a] a = True + check2 [a] (t Int) = False + check2 [a] (t a) = False -- By check1 we never reach this equation + check2 [Int] a = True + check2 [Int] (t Int) = True + check2 [Maybe a] (t Int) = False + check2 [Maybe Int] (t Int) = True + check2 (Maybe [a]) (m [Int]) = False + check2 (Maybe [Int]) (m [Int]) = True -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]) ) +check1 :: QuantifiedType -> Bool +check1 (tvs, _) = not $ any isHigherKind (map tyVarKind tvs) + where + isHigherKind = not . null . fst . splitKindFunTys + +check2 :: QuantifiedType -> QuantifiedType -> Bool +check2 (_, rtti_ty) (_, old_ty) + | Just (_, rttis) <- tcSplitTyConApp_maybe rtti_ty + = case () of + _ | Just (_,olds) <- tcSplitTyConApp_maybe old_ty + -> and$ zipWith check2 (map quantifyType rttis) (map quantifyType olds) + _ | Just _ <- splitAppTy_maybe old_ty + -> isMonomorphicOnNonPhantomArgs rtti_ty + _ -> True + | otherwise = True + +-- Dealing with newtypes +-------------------------- +{- + congruenceNewtypes does 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 sometimes there is evidence available. + Evidence can come from DataCon signatures or + from compile-time type inference. + What we are doing here is an approximation + of unification modulo a set of equations derived + from newtype definitions. These equations should be the + same as the equality coercions generated for newtypes + in System Fc. The idea is to perform a sort of rewriting, + taking those equations as rules, before launching unification. + + The caller must ensure the following. + The 1st type (lhs) comes from the heap structure of ptrs,nptrs. + The 2nd type (rhs) comes from a DataCon type signature. + Rewriting (i.e. adding/removing a newtype wrapper) can happen + in both types, but in the rhs it is restricted to the result type. -seq _3 () + Note that it is very tricky to make this 'rewriting' + work with the unification implemented by TcM, where + substitutions are operationally inlined. The order in which + constraints are unified is vital as we cannot modify + anything that has been touched by a previous unification step. +Therefore, congruenceNewtypes is sound only if the types +recovered by the RTTI mechanism are unified Top-Down. +-} +congruenceNewtypes :: TcType -> TcType -> TR (TcType,TcType) +congruenceNewtypes lhs rhs = go lhs rhs >>= \rhs' -> return (lhs,rhs') + where + go l r + -- TyVar lhs inductive case + | Just tv <- getTyVar_maybe l + , isTcTyVar tv + , isMetaTyVar tv + = recoverTR (return r) $ do + Indirect ty_v <- readMetaTyVar tv + traceTR $ fsep [text "(congruence) Following indirect tyvar:", + ppr tv, equals, ppr ty_v] + go ty_v r +-- FunTy inductive case + | Just (l1,l2) <- splitFunTy_maybe l + , Just (r1,r2) <- splitFunTy_maybe r + = do r2' <- go l2 r2 + r1' <- go l1 r1 + return (mkFunTy r1' r2') +-- TyconApp Inductive case; this is the interesting bit. + | Just (tycon_l, _) <- tcSplitTyConApp_maybe lhs + , Just (tycon_r, _) <- tcSplitTyConApp_maybe rhs + , tycon_l /= tycon_r + = upgrade tycon_l r + + | otherwise = return r + + where upgrade :: TyCon -> Type -> TR Type + upgrade new_tycon ty + | not (isNewTyCon new_tycon) = do + traceTR (text "(Upgrade) Not matching newtype evidence: " <> + ppr new_tycon <> text " for " <> ppr ty) + return ty + | otherwise = do + traceTR (text "(Upgrade) upgraded " <> ppr ty <> + text " in presence of newtype evidence " <> ppr new_tycon) + vars <- mapM (newVar . tyVarKind) (tyConTyVars new_tycon) + let ty' = mkTyConApp new_tycon vars + _ <- liftTcM (unifyType ty (repType ty')) + -- assumes that reptype doesn't ^^^^ touch tyconApp args + return ty' -t - O ( (Just (_5::c)) : (_4::[b]) ) -At this point, we know that b = (Maybe c) +zonkTerm :: Term -> TcM Term +zonkTerm = foldTermM (TermFoldM + { fTermM = \ty dc v tt -> zonkRttiType ty >>= \ty' -> + return (Term ty' dc v tt) + , fSuspensionM = \ct ty v b -> zonkRttiType ty >>= \ty -> + return (Suspension ct ty v b) + , fNewtypeWrapM = \ty dc t -> zonkRttiType ty >>= \ty' -> + return$ NewtypeWrap ty' dc t + , fRefWrapM = \ty t -> return RefWrap `ap` + zonkRttiType ty `ap` return t + , fPrimM = (return.) . Prim }) + +zonkRttiType :: TcType -> TcM Type +-- Zonk the type, replacing any unbound Meta tyvars +-- by skolems, safely out of Meta-tyvar-land +zonkRttiType = zonkType (mkZonkTcTyVar zonk_unbound_meta) + where + zonk_unbound_meta tv + = ASSERT( isTcTyVar tv ) + do { tv' <- skolemiseUnboundMetaTyVar tv RuntimeUnk + -- This is where RuntimeUnks are born: + -- otherwise-unconstrained unification variables are + -- turned into RuntimeUnks as they leave the + -- typechecker's monad + ; return (mkTyVarTy tv') } + +-------------------------------------------------------------------------------- +-- Restore Class predicates out of a representation type +dictsView :: Type -> Type +-- dictsView ty = ty +dictsView (FunTy (TyConApp tc_dict args) ty) + | Just c <- tyConClass_maybe tc_dict + = FunTy (PredTy (ClassP c args)) (dictsView ty) +dictsView ty + | Just (tc_fun, [TyConApp tc_dict args, ty2]) <- tcSplitTyConApp_maybe ty + , Just c <- tyConClass_maybe tc_dict + = mkTyConApp tc_fun [PredTy (ClassP c args), dictsView ty2] +dictsView ty = ty + + +-- Use only for RTTI types +isMonomorphic :: RttiType -> Bool +isMonomorphic ty = noExistentials && noUniversals + where (tvs, _, ty') = tcSplitSigmaTy ty + noExistentials = isEmptyVarSet (tyVarsOfType ty') + noUniversals = null tvs + +-- Use only for RTTI types +isMonomorphicOnNonPhantomArgs :: RttiType -> Bool +isMonomorphicOnNonPhantomArgs ty + | Just (tc, all_args) <- tcSplitTyConApp_maybe (repType ty) + , phantom_vars <- tyConPhantomTyVars tc + , concrete_args <- [ arg | (tyv,arg) <- tyConTyVars tc `zip` all_args + , tyv `notElem` phantom_vars] + = all isMonomorphicOnNonPhantomArgs concrete_args + | Just (ty1, ty2) <- splitFunTy_maybe ty + = all isMonomorphicOnNonPhantomArgs [ty1,ty2] + | otherwise = isMonomorphic ty + +tyConPhantomTyVars :: TyCon -> [TyVar] +tyConPhantomTyVars tc + | isAlgTyCon tc + , Just dcs <- tyConDataCons_maybe tc + , dc_vars <- concatMap dataConUnivTyVars dcs + = tyConTyVars tc \\ dc_vars +tyConPhantomTyVars _ = [] + +type QuantifiedType = ([TyVar], Type) -- Make the free type variables explicit + +quantifyType :: Type -> QuantifiedType +-- Generalize the type: find all free tyvars and wrap in the appropiate ForAll. +quantifyType ty = (varSetElems (tyVarsOfType ty), ty) -seq _5 () +mapMif :: Monad m => (a -> Bool) -> (a -> m a) -> [a] -> m [a] +mapMif pred f xx = sequence $ mapMif_ pred f xx + where + mapMif_ _ _ [] = [] + mapMif_ pred f (x:xx) = (if pred x then f x else return x) : mapMif_ pred f xx -t - O ( (Just ((_6::d) : (_7::[d]) )) : (_4::[b]) ) +unlessM :: Monad m => m Bool -> m () -> m () +unlessM condM acc = condM >>= \c -> unless c acc -At this point, we know that c = [d] -seq _6 () +-- Strict application of f at index i +appArr :: Ix i => (e -> a) -> Array i e -> Int -> a +appArr f a@(Array _ _ _ ptrs#) i@(I# i#) + = ASSERT2 (i < length(elems a), ppr(length$ elems a, i)) + case indexArray# ptrs# i# of + (# e #) -> f e -t - O ( (Just (1 : (_7::[d]) )) : (_4::[b]) ) +amap' :: (t -> b) -> Array Int t -> [b] +amap' f (Array i0 i _ arr#) = map g [0 .. i - i0] + where g (I# i#) = case indexArray# arr# i# of + (# e #) -> f e -At this point, we know that d = Integer -The fully reconstructed expressions, with propagation, would be: +isLifted :: Type -> Bool +isLifted = not . isUnLiftedType -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]]) ) +extractUnboxed :: [Type] -> Closure -> [[Word]] +extractUnboxed tt clos = go tt (nonPtrs clos) + where sizeofType t + | Just (tycon,_) <- tcSplitTyConApp_maybe t + = ASSERT (isPrimTyCon tycon) sizeofTyCon tycon + | otherwise = pprPanic "Expected a TcTyCon" (ppr t) + go [] _ = [] + go (t:tt) xx + | (x, rest) <- splitAt (sizeofType t) xx + = x : go tt rest +sizeofTyCon :: TyCon -> Int -- in *words* +sizeofTyCon = primRepSizeW . tyConPrimRep -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 --} +(|.|) :: (a -> Bool) -> (a -> Bool) -> a -> Bool +(f |.| g) x = f x || g x