----------------------------------------------------------------------------- -- -- GHC Interactive support for inspecting arbitrary closures at runtime -- -- Pepe Iborra (supported by Google SoC) 2006 -- ----------------------------------------------------------------------------- {-# OPTIONS -fno-warn-tabs #-} -- The above warning supression flag is a temporary kludge. -- While working on this module you are encouraged to remove it and -- detab the module (please do the detabbing in a separate patch). See -- http://hackage.haskell.org/trac/ghc/wiki/Commentary/CodingStyle#TabsvsSpaces -- for details module RtClosureInspect( 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 DebuggerUtils import ByteCodeItbls ( StgInfoTable ) import qualified ByteCodeItbls as BCI( StgInfoTable(..) ) import HscTypes import Linker import DataCon import Type import qualified Unify as U import Var import TcRnMonad import TcType import TcMType import TcHsSyn ( mkZonkTcTyVar ) import TcUnify import TcEnv import TyCon import Name import VarEnv import Util import VarSet import TysPrim import PrelNames import TysWiredIn import DynFlags import Outputable as Ppr import FastString import Constants ( wORD_SIZE ) import GHC.Arr ( Array(..) ) import GHC.Exts import GHC.IO ( IO(..) ) import StaticFlags( opt_PprStyle_Debug ) import Control.Monad import Data.Maybe import Data.Array.Base import Data.Ix import Data.List import qualified Data.Sequence as Seq import Data.Monoid import Data.Sequence (viewl, ViewL(..)) import Foreign.Safe import System.IO.Unsafe --------------------------------------------- -- * A representation of semi evaluated Terms --------------------------------------------- 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 :: RttiType , value :: [Word] } | Suspension { ctype :: ClosureType , 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 :: Term -> RttiType termType t = ty t isFullyEvaluatedTerm :: Term -> Bool isFullyEvaluatedTerm Term {subTerms=tt} = all isFullyEvaluatedTerm tt isFullyEvaluatedTerm Prim {} = True isFullyEvaluatedTerm NewtypeWrap{wrapped_term=t} = isFullyEvaluatedTerm t isFullyEvaluatedTerm RefWrap{wrapped_term=t} = isFullyEvaluatedTerm t isFullyEvaluatedTerm _ = False instance Outputable (Term) where ppr t | Just doc <- cPprTerm cPprTermBase t = doc | otherwise = panic "Outputable Term instance" ------------------------------------------------------------------------- -- Runtime Closure Datatype and functions for retrieving closure related stuff ------------------------------------------------------------------------- data ClosureType = Constr | Fun | Thunk Int | ThunkSelector | Blackhole | AP | PAP | Indirection Int | MutVar Int | MVar Int | Other Int deriving (Show, Eq) data Closure = Closure { tipe :: ClosureType , infoPtr :: Ptr () , infoTable :: StgInfoTable , ptrs :: Array Int HValue , nonPtrs :: [Word] } instance Outputable ClosureType where ppr = text . show #include "../includes/rts/storage/ClosureTypes.h" aP_CODE, pAP_CODE :: Int aP_CODE = AP pAP_CODE = PAP #undef AP #undef PAP getClosureData :: a -> IO Closure getClosureData a = case unpackClosure# a of (# iptr, ptrs, nptrs #) -> do 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 = fromIntegral (BCI.ptrs itbl) ptrsList = Array 0 (elems - 1) elems ptrs nptrs_data = [W# (indexWordArray# nptrs i) | I# i <- [0.. fromIntegral (BCI.nptrs itbl)-1] ] ASSERT(elems >= 0) return () ptrsList `seq` return (Closure tipe (Ptr iptr) itbl ptrsList nptrs_data) readCType :: Integral a => a -> ClosureType readCType i | i >= CONSTR && i <= CONSTR_NOCAF_STATIC = Constr | i >= FUN && i <= FUN_STATIC = Fun | i >= THUNK && i < THUNK_SELECTOR = Thunk i' | i == THUNK_SELECTOR = ThunkSelector | i == BLACKHOLE = Blackhole | 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 _ = 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 _ -> return False where amapM f = sequence . amap' f -- 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 | 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) -} ----------------------------------- -- * 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 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, fNewtypeWrap = NewtypeWrap, fRefWrap = RefWrap } 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 ---------------------------------- 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 sub_terms_to_show = return (ppr dc) | otherwise = do { tt_docs <- mapM (y app_prec) sub_terms_to_show ; return $ cparen (p >= app_prec) $ sep [ppr dc, nest 2 (pprDeeperList fsep tt_docs)] } where sub_terms_to_show -- Don't show the dictionary arguments to -- constructors unless -dppr-debug is on | opt_PprStyle_Debug = tt | otherwise = dropList (dataConTheta dc) tt 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 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 :: forall m. 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) ppr_list , ifTerm (isTyCon intTyCon . ty) ppr_int , ifTerm (isTyCon charTyCon . ty) ppr_char , ifTerm (isTyCon floatTyCon . ty) ppr_float , ifTerm (isTyCon doubleTyCon . ty) ppr_double , ifTerm (isIntegerTy . ty) ppr_integer ] where ifTerm :: (Term -> Bool) -> (Precedence -> Term -> m SDoc) -> Precedence -> Term -> m (Maybe SDoc) 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) ppr_int, ppr_char, ppr_float, ppr_double, ppr_integer :: Precedence -> Term -> m SDoc ppr_int _ v = return (Ppr.int (unsafeCoerce# (val v))) ppr_char _ v = return (Ppr.char '\'' <> Ppr.char (unsafeCoerce# (val v)) <> Ppr.char '\'') ppr_float _ v = return (Ppr.float (unsafeCoerce# (val v))) ppr_double _ v = return (Ppr.double (unsafeCoerce# (val v))) ppr_integer _ v = return (Ppr.integer (unsafeCoerce# (val v))) --Note pprinting of list terms is not lazy ppr_list :: Precedence -> Term -> m SDoc ppr_list p (Term{subTerms=[h,t]}) = do let elems = h : getListTerms t isConsLast = not(termType(last elems) `eqType` termType h) is_string = all (isCharTy . ty) elems print_elems <- mapM (y cons_prec) elems if is_string then return (Ppr.doubleQuotes (Ppr.text (unsafeCoerce# (map val elems)))) else if isConsLast then return $ cparen (p >= cons_prec) $ pprDeeperList fsep $ punctuate (space<>colon) print_elems else return $ 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) ppr_list _ _ = 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 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) -} -- 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 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 runTR_maybe :: HscEnv -> TR a -> IO (Maybe a) runTR_maybe hsc_env = fmap snd . initTc hsc_env HsSrcFile False iNTERACTIVE traceTR :: SDoc -> TR () traceTR = liftTcM . traceOptTcRn Opt_D_dump_rtti -- 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 newVar :: Kind -> TR TcType newVar = liftTcM . newFlexiTyVarTy instTyVars :: [TyVar] -> TR ([TcTyVar], [TcType], TvSubst) -- Instantiate fresh mutable type variables from some TyVars -- This function preserves the print-name, which helps error messages instTyVars = liftTcM . tcInstTyVars 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 :: Int -> Type -> Type -> HValue -> TcM Term -- [SPJ May 11] I don't understand the difference between my_ty and old_ty 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 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 traceTR (text "entering a constructor " <> if monomorphic then parens (text "already monomorphic: " <> ppr my_ty) else Ppr.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. traceTR (text "Nothing" <+> ppr dcname) 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 traceTR (text "Just" <+> ppr dc) subTtypes <- getDataConArgTys dc my_ty let (subTtypesP, subTtypesNP) = partition isPtrType subTtypes subTermsP <- sequence [ appArr (go (pred max_depth) ty ty) (ptrs clos) i | (i,ty) <- zip [0..] subTtypesP] let unboxeds = extractUnboxed subTtypesNP clos subTermsNP = zipWith Prim 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. tipe_clos -> return (Suspension tipe_clos my_ty a Nothing) -- put together pointed and nonpointed subterms in the -- correct order. reOrderTerms _ _ [] = [] reOrderTerms pointed unpointed (ty:tys) | isPtrType 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)) 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 traceTR (text "go" <+> ppr my_ty) 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) traceTR (text "Constr1" <+> ppr dcname) (_,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 arg_tys <- getDataConArgTys dc my_ty traceTR (text "Constr2" <+> ppr dcname <+> ppr arg_tys) return $ [ appArr (\e-> (ty,e)) (ptrs clos) i | (i,ty) <- zip [0..] (filter isPtrType arg_tys)] _ -> 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] getDataConArgTys :: DataCon -> Type -> TR [Type] -- Given the result type ty of a constructor application (D a b c :: ty) -- return the types of the arguments. This is RTTI-land, so 'ty' might -- not be fully known. Moreover, the arg types might involve existentials; -- if so, make up fresh RTTI type variables for them getDataConArgTys dc con_app_ty = do { (_, ex_tys, _) <- instTyVars ex_tvs ; let rep_con_app_ty = repType con_app_ty ; ty_args <- case tcSplitTyConApp_maybe rep_con_app_ty of Just (tc, ty_args) | dataConTyCon dc == tc -> ASSERT( univ_tvs `equalLength` ty_args) return ty_args _ -> do { (_, ty_args, subst) <- instTyVars univ_tvs ; let res_ty = substTy subst (dataConOrigResTy dc) ; addConstraint rep_con_app_ty res_ty ; return ty_args } -- It is necessary to check dataConTyCon dc == 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 ; let subst = zipTopTvSubst (univ_tvs ++ ex_tvs) (ty_args ++ ex_tys) ; return (substTys subst (dataConRepArgTys dc)) } where univ_tvs = dataConUnivTyVars dc ex_tvs = dataConExTyVars dc isPtrType :: Type -> Bool isPtrType ty = case typePrimRep ty of PtrRep -> True _ -> False -- Soundness checks -------------------- {- This is not formalized anywhere, so hold to your seats! RTTI in the presence of newtypes can be a tricky and unsound business. 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 -} 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. 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, _) <- instTyVars (tyConTyVars new_tycon) let ty' = mkTyConApp new_tycon vars _ <- liftTcM (unifyType ty (repType ty')) -- assumes that reptype doesn't ^^^^ touch tyconApp args return ty' 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 mkTyVarTy) 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 -- 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) unlessM :: Monad m => m Bool -> m () -> m () unlessM condM acc = condM >>= \c -> unless c acc -- 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 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 extractUnboxed :: [Type] -> Closure -> [[Word]] extractUnboxed tt clos = go tt (nonPtrs clos) where sizeofType t = primRepSizeW (typePrimRep t) go [] _ = [] go (t:tt) xx | (x, rest) <- splitAt (sizeofType t) xx = x : go tt rest