X-Git-Url: http://git.megacz.com/?a=blobdiff_plain;f=ghc%2Fcompiler%2FcoreSyn%2FCoreUtils.lhs;h=d7a91a00be06dab9ec6c2aeee6f9da1fa9cea908;hb=979947f545d70c63edb7ca96f6e47008ac90e3bf;hp=9b9b03c85ee71d9cdc89b24e325ef92ab44567be;hpb=9d38678ea60ff32f756390a30c659daa22c98c93;p=ghc-hetmet.git diff --git a/ghc/compiler/coreSyn/CoreUtils.lhs b/ghc/compiler/coreSyn/CoreUtils.lhs index 9b9b03c..d7a91a0 100644 --- a/ghc/compiler/coreSyn/CoreUtils.lhs +++ b/ghc/compiler/coreSyn/CoreUtils.lhs @@ -5,48 +5,71 @@ \begin{code} module CoreUtils ( - coreExprType, coreAltsType, + -- Construction + mkNote, mkInlineMe, mkSCC, mkCoerce, + bindNonRec, needsCaseBinding, + mkIfThenElse, mkAltExpr, mkPiType, mkPiTypes, + -- Taking expressions apart + findDefault, findAlt, hasDefault, + + -- Properties of expressions + exprType, coreAltsType, exprIsBottom, exprIsDupable, exprIsTrivial, exprIsCheap, - exprIsValue, - exprOkForSpeculation, exprIsBig, hashExpr, - exprArity, exprGenerousArity, - cheapEqExpr, eqExpr, applyTypeToArgs + exprIsValue,exprOkForSpeculation, exprIsBig, + exprIsConApp_maybe, exprIsAtom, + idAppIsBottom, idAppIsCheap, + + + -- Arity and eta expansion + manifestArity, exprArity, + exprEtaExpandArity, etaExpand, + + -- Size + coreBindsSize, + + -- Hashing + hashExpr, + + -- Equality + cheapEqExpr, eqExpr, applyTypeToArgs, applyTypeToArg ) where #include "HsVersions.h" -import {-# SOURCE #-} CoreUnfold ( isEvaldUnfolding ) - import GlaExts -- For `xori` import CoreSyn import PprCore ( pprCoreExpr ) -import Var ( IdOrTyVar, isId, isTyVar ) -import VarSet +import Var ( Var, isId, isTyVar ) import VarEnv -import Name ( isLocallyDefined, hashName ) -import Const ( Con, isWHNFCon, conIsTrivial, conIsCheap, conIsDupable, - conType, conOkForSpeculation, conStrictness, hashCon +import Name ( hashName ) +import Literal ( hashLiteral, literalType, litIsDupable, isZeroLit ) +import DataCon ( DataCon, dataConRepArity, dataConArgTys, isExistentialDataCon, dataConTyCon ) +import PrimOp ( PrimOp(..), primOpOkForSpeculation, primOpIsCheap ) +import Id ( Id, idType, globalIdDetails, idNewStrictness, idLBVarInfo, + mkWildId, idArity, idName, idUnfolding, idInfo, isOneShotLambda, + isDataConId_maybe, mkSysLocal, isDataConId, isBottomingId ) -import Id ( Id, idType, setIdType, idUnique, idAppIsBottom, - getIdArity, idName, - getIdSpecialisation, setIdSpecialisation, - getInlinePragma, setInlinePragma, - getIdUnfolding, setIdUnfolding, idInfo +import IdInfo ( LBVarInfo(..), + GlobalIdDetails(..), + megaSeqIdInfo ) +import NewDemand ( appIsBottom ) +import Type ( Type, mkFunTy, mkForAllTy, splitFunTy_maybe, splitFunTy, + applyTys, isUnLiftedType, seqType, mkTyVarTy, + splitForAllTy_maybe, isForAllTy, splitNewType_maybe, + splitTyConApp_maybe, eqType, funResultTy, applyTy, + funResultTy, applyTy ) -import IdInfo ( arityLowerBound, InlinePragInfo(..), lbvarInfo, LBVarInfo(..) ) -import Type ( Type, mkFunTy, mkForAllTy, - splitFunTy_maybe, tyVarsOfType, tyVarsOfTypes, - isNotUsgTy, mkUsgTy, unUsgTy, UsageAnn(..), - tidyTyVar, applyTys, isUnLiftedType - ) -import Demand ( isPrim, isLazy ) -import Unique ( buildIdKey, augmentIdKey ) -import Util ( zipWithEqual, mapAccumL ) +import TyCon ( tyConArity ) +import TysWiredIn ( boolTy, trueDataCon, falseDataCon ) +import CostCentre ( CostCentre ) +import BasicTypes ( Arity ) +import Unique ( Unique ) import Outputable import TysPrim ( alphaTy ) -- Debugging only +import Util ( equalLength, lengthAtLeast ) \end{code} @@ -57,46 +80,60 @@ import TysPrim ( alphaTy ) -- Debugging only %************************************************************************ \begin{code} -coreExprType :: CoreExpr -> Type - -coreExprType (Var var) = idType var -coreExprType (Let _ body) = coreExprType body -coreExprType (Case _ _ alts) = coreAltsType alts -coreExprType (Note (Coerce ty _) e) = ty -coreExprType (Note (TermUsg u) e) = mkUsgTy u (unUsgTy (coreExprType e)) -coreExprType (Note other_note e) = coreExprType e -coreExprType e@(Con con args) = applyTypeToArgs e (conType con) args - -coreExprType (Lam binder expr) - | isId binder = (case (lbvarInfo . idInfo) binder of - IsOneShotLambda -> mkUsgTy UsOnce - otherwise -> id) $ - idType binder `mkFunTy` coreExprType expr - | isTyVar binder = mkForAllTy binder (coreExprType expr) - -coreExprType e@(App _ _) +exprType :: CoreExpr -> Type + +exprType (Var var) = idType var +exprType (Lit lit) = literalType lit +exprType (Let _ body) = exprType body +exprType (Case _ _ alts) = coreAltsType alts +exprType (Note (Coerce ty _) e) = ty -- **! should take usage from e +exprType (Note other_note e) = exprType e +exprType (Lam binder expr) = mkPiType binder (exprType expr) +exprType e@(App _ _) = case collectArgs e of - (fun, args) -> applyTypeToArgs e (coreExprType fun) args + (fun, args) -> applyTypeToArgs e (exprType fun) args -coreExprType other = pprTrace "coreExprType" (pprCoreExpr other) alphaTy +exprType other = pprTrace "exprType" (pprCoreExpr other) alphaTy coreAltsType :: [CoreAlt] -> Type -coreAltsType ((_,_,rhs) : _) = coreExprType rhs +coreAltsType ((_,_,rhs) : _) = exprType rhs \end{code} +@mkPiType@ makes a (->) type or a forall type, depending on whether +it is given a type variable or a term variable. We cleverly use the +lbvarinfo field to figure out the right annotation for the arrove in +case of a term variable. + \begin{code} --- The first argument is just for debugging +mkPiType :: Var -> Type -> Type -- The more polymorphic version +mkPiTypes :: [Var] -> Type -> Type -- doesn't work... + +mkPiTypes vs ty = foldr mkPiType ty vs + +mkPiType v ty + | isId v = mkFunTy (idType v) ty + | otherwise = mkForAllTy v ty +\end{code} + +\begin{code} +applyTypeToArg :: Type -> CoreExpr -> Type +applyTypeToArg fun_ty (Type arg_ty) = applyTy fun_ty arg_ty +applyTypeToArg fun_ty other_arg = funResultTy fun_ty + applyTypeToArgs :: CoreExpr -> Type -> [CoreExpr] -> Type +-- A more efficient version of applyTypeToArg +-- when we have several args +-- The first argument is just for debugging applyTypeToArgs e op_ty [] = op_ty applyTypeToArgs e op_ty (Type ty : args) = -- Accumulate type arguments so we can instantiate all at once - ASSERT2( all isNotUsgTy tys, ppr e <+> text "of" <+> ppr op_ty <+> text "to" <+> ppr (Type ty : args) <+> text "i.e." <+> ppr tys ) - applyTypeToArgs e (applyTys op_ty tys) rest_args + go [ty] args where - (tys, rest_args) = go [ty] args - go tys (Type ty : args) = go (ty:tys) args - go tys rest_args = (reverse tys, rest_args) + go rev_tys (Type ty : args) = go (ty:rev_tys) args + go rev_tys rest_args = applyTypeToArgs e op_ty' rest_args + where + op_ty' = applyTys op_ty (reverse rev_tys) applyTypeToArgs e op_ty (other_arg : args) = case (splitFunTy_maybe op_ty) of @@ -105,27 +142,203 @@ applyTypeToArgs e op_ty (other_arg : args) \end{code} + +%************************************************************************ +%* * +\subsection{Attaching notes} +%* * +%************************************************************************ + +mkNote removes redundant coercions, and SCCs where possible + +\begin{code} +mkNote :: Note -> CoreExpr -> CoreExpr +mkNote (Coerce to_ty from_ty) expr = mkCoerce to_ty from_ty expr +mkNote (SCC cc) expr = mkSCC cc expr +mkNote InlineMe expr = mkInlineMe expr +mkNote note expr = Note note expr + +-- Slide InlineCall in around the function +-- No longer necessary I think (SLPJ Apr 99) +-- mkNote InlineCall (App f a) = App (mkNote InlineCall f) a +-- mkNote InlineCall (Var v) = Note InlineCall (Var v) +-- mkNote InlineCall expr = expr +\end{code} + +Drop trivial InlineMe's. This is somewhat important, because if we have an unfolding +that looks like (Note InlineMe (Var v)), the InlineMe doesn't go away because it may +not be *applied* to anything. + +We don't use exprIsTrivial here, though, because we sometimes generate worker/wrapper +bindings like + fw = ... + f = inline_me (coerce t fw) +As usual, the inline_me prevents the worker from getting inlined back into the wrapper. +We want the split, so that the coerces can cancel at the call site. + +However, we can get left with tiresome type applications. Notably, consider + f = /\ a -> let t = e in (t, w) +Then lifting the let out of the big lambda gives + t' = /\a -> e + f = /\ a -> let t = inline_me (t' a) in (t, w) +The inline_me is to stop the simplifier inlining t' right back +into t's RHS. In the next phase we'll substitute for t (since +its rhs is trivial) and *then* we could get rid of the inline_me. +But it hardly seems worth it, so I don't bother. + +\begin{code} +mkInlineMe (Var v) = Var v +mkInlineMe e = Note InlineMe e +\end{code} + + + +\begin{code} +mkCoerce :: Type -> Type -> CoreExpr -> CoreExpr + +mkCoerce to_ty from_ty (Note (Coerce to_ty2 from_ty2) expr) + = ASSERT( from_ty `eqType` to_ty2 ) + mkCoerce to_ty from_ty2 expr + +mkCoerce to_ty from_ty expr + | to_ty `eqType` from_ty = expr + | otherwise = ASSERT( from_ty `eqType` exprType expr ) + Note (Coerce to_ty from_ty) expr +\end{code} + +\begin{code} +mkSCC :: CostCentre -> Expr b -> Expr b + -- Note: Nested SCC's *are* preserved for the benefit of + -- cost centre stack profiling +mkSCC cc (Lit lit) = Lit lit +mkSCC cc (Lam x e) = Lam x (mkSCC cc e) -- Move _scc_ inside lambda +mkSCC cc (Note (SCC cc') e) = Note (SCC cc) (Note (SCC cc') e) +mkSCC cc (Note n e) = Note n (mkSCC cc e) -- Move _scc_ inside notes +mkSCC cc expr = Note (SCC cc) expr +\end{code} + + +%************************************************************************ +%* * +\subsection{Other expression construction} +%* * +%************************************************************************ + +\begin{code} +bindNonRec :: Id -> CoreExpr -> CoreExpr -> CoreExpr +-- (bindNonRec x r b) produces either +-- let x = r in b +-- or +-- case r of x { _DEFAULT_ -> b } +-- +-- depending on whether x is unlifted or not +-- It's used by the desugarer to avoid building bindings +-- that give Core Lint a heart attack. Actually the simplifier +-- deals with them perfectly well. +bindNonRec bndr rhs body + | needsCaseBinding (idType bndr) rhs = Case rhs bndr [(DEFAULT,[],body)] + | otherwise = Let (NonRec bndr rhs) body + +needsCaseBinding ty rhs = isUnLiftedType ty && not (exprOkForSpeculation rhs) + -- Make a case expression instead of a let + -- These can arise either from the desugarer, + -- or from beta reductions: (\x.e) (x +# y) +\end{code} + +\begin{code} +mkAltExpr :: AltCon -> [CoreBndr] -> [Type] -> CoreExpr + -- This guy constructs the value that the scrutinee must have + -- when you are in one particular branch of a case +mkAltExpr (DataAlt con) args inst_tys + = mkConApp con (map Type inst_tys ++ map varToCoreExpr args) +mkAltExpr (LitAlt lit) [] [] + = Lit lit + +mkIfThenElse :: CoreExpr -> CoreExpr -> CoreExpr -> CoreExpr +mkIfThenElse guard then_expr else_expr + = Case guard (mkWildId boolTy) + [ (DataAlt trueDataCon, [], then_expr), + (DataAlt falseDataCon, [], else_expr) ] +\end{code} + + +%************************************************************************ +%* * +\subsection{Taking expressions apart} +%* * +%************************************************************************ + +The default alternative must be first, if it exists at all. +This makes it easy to find, though it makes matching marginally harder. + +\begin{code} +hasDefault :: [CoreAlt] -> Bool +hasDefault ((DEFAULT,_,_) : alts) = True +hasDefault _ = False + +findDefault :: [CoreAlt] -> ([CoreAlt], Maybe CoreExpr) +findDefault ((DEFAULT,args,rhs) : alts) = ASSERT( null args ) (alts, Just rhs) +findDefault alts = (alts, Nothing) + +findAlt :: AltCon -> [CoreAlt] -> CoreAlt +findAlt con alts + = case alts of + (deflt@(DEFAULT,_,_):alts) -> go alts deflt + other -> go alts panic_deflt + + where + panic_deflt = pprPanic "Missing alternative" (ppr con $$ vcat (map ppr alts)) + + go [] deflt = deflt + go (alt@(con1,_,_) : alts) deflt | con == con1 = alt + | otherwise = ASSERT( not (con1 == DEFAULT) ) + go alts deflt +\end{code} + + %************************************************************************ %* * \subsection{Figuring out things about expressions} %* * %************************************************************************ -@exprIsTrivial@ is true of expressions we are unconditionally - happy to duplicate; simple variables and constants, - and type applications. +@exprIsTrivial@ is true of expressions we are unconditionally happy to + duplicate; simple variables and constants, and type + applications. Note that primop Ids aren't considered + trivial unless @exprIsBottom@ is true of expressions that are guaranteed to diverge +There used to be a gruesome test for (hasNoBinding v) in the +Var case: + exprIsTrivial (Var v) | hasNoBinding v = idArity v == 0 +The idea here is that a constructor worker, like $wJust, is +really short for (\x -> $wJust x), becuase $wJust has no binding. +So it should be treated like a lambda. Ditto unsaturated primops. +But now constructor workers are not "have-no-binding" Ids. And +completely un-applied primops and foreign-call Ids are sufficiently +rare that I plan to allow them to be duplicated and put up with +saturating them. + \begin{code} -exprIsTrivial (Type _) = True -exprIsTrivial (Var v) = True -exprIsTrivial (App e arg) = isTypeArg arg && exprIsTrivial e -exprIsTrivial (Note _ e) = exprIsTrivial e -exprIsTrivial (Con con args) = conIsTrivial con && all isTypeArg args -exprIsTrivial (Lam b body) | isTyVar b = exprIsTrivial body -exprIsTrivial other = False +exprIsTrivial (Var v) = True -- See notes above +exprIsTrivial (Type _) = True +exprIsTrivial (Lit lit) = True +exprIsTrivial (App e arg) = not (isRuntimeArg arg) && exprIsTrivial e +exprIsTrivial (Note _ e) = exprIsTrivial e +exprIsTrivial (Lam b body) = not (isRuntimeVar b) && exprIsTrivial body +exprIsTrivial other = False + +exprIsAtom :: CoreExpr -> Bool +-- Used to decide whether to let-binding an STG argument +-- when compiling to ILX => type applications are not allowed +exprIsAtom (Var v) = True -- primOpIsDupable? +exprIsAtom (Lit lit) = True +exprIsAtom (Type ty) = True +exprIsAtom (Note (SCC _) e) = False +exprIsAtom (Note _ e) = exprIsAtom e +exprIsAtom other = False \end{code} @@ -141,15 +354,19 @@ exprIsTrivial other = False \begin{code} -exprIsDupable (Type _) = True -exprIsDupable (Con con args) = conIsDupable con && - all exprIsDupable args && - valArgCount args <= dupAppSize - -exprIsDupable (Note _ e) = exprIsDupable e -exprIsDupable expr = case collectArgs expr of - (Var f, args) -> valArgCount args <= dupAppSize - other -> False +exprIsDupable (Type _) = True +exprIsDupable (Var v) = True +exprIsDupable (Lit lit) = litIsDupable lit +exprIsDupable (Note InlineMe e) = True +exprIsDupable (Note _ e) = exprIsDupable e +exprIsDupable expr + = go expr 0 + where + go (Var v) n_args = True + go (App f a) n_args = n_args < dupAppSize + && exprIsDupable a + && go f (n_args+1) + go other n_args = False dupAppSize :: Int dupAppSize = 4 -- Size of application we are prepared to duplicate @@ -169,53 +386,74 @@ shared. The main examples of things which aren't WHNF but are * case e of pi -> ei + (where e, and all the ei are cheap) - where e, and all the ei are cheap; and - - * let x = e - in b - - where e and b are cheap; and + * let x = e in b + (where e and b are cheap) * op x1 ... xn - - where op is a cheap primitive operator + (where op is a cheap primitive operator) * error "foo" + (because we are happy to substitute it inside a lambda) Notice that a variable is considered 'cheap': we can push it inside a lambda, because sharing will make sure it is only evaluated once. \begin{code} exprIsCheap :: CoreExpr -> Bool -exprIsCheap (Type _) = True -exprIsCheap (Var _) = True -exprIsCheap (Con con args) = conIsCheap con && all exprIsCheap args -exprIsCheap (Note _ e) = exprIsCheap e -exprIsCheap (Lam x e) = if isId x then True else exprIsCheap e -exprIsCheap other_expr -- look for manifest partial application - = case collectArgs other_expr of - (f, args) -> isPap f (valArgCount args) && all exprIsCheap args -\end{code} - -\begin{code} -isPap :: CoreExpr -- Function - -> Int -- Number of value args - -> Bool -isPap (Var f) n_val_args - = idAppIsBottom f n_val_args - -- Application of a function which - -- always gives bottom; we treat this as - -- a WHNF, because it certainly doesn't - -- need to be shared! - - || n_val_args == 0 -- Just a type application of +exprIsCheap (Lit lit) = True +exprIsCheap (Type _) = True +exprIsCheap (Var _) = True +exprIsCheap (Note InlineMe e) = True +exprIsCheap (Note _ e) = exprIsCheap e +exprIsCheap (Lam x e) = isRuntimeVar x || exprIsCheap e +exprIsCheap (Case e _ alts) = exprIsCheap e && + and [exprIsCheap rhs | (_,_,rhs) <- alts] + -- Experimentally, treat (case x of ...) as cheap + -- (and case __coerce x etc.) + -- This improves arities of overloaded functions where + -- there is only dictionary selection (no construction) involved +exprIsCheap (Let (NonRec x _) e) + | isUnLiftedType (idType x) = exprIsCheap e + | otherwise = False + -- strict lets always have cheap right hand sides, and + -- do no allocation. + +exprIsCheap other_expr + = go other_expr 0 True + where + go (Var f) n_args args_cheap + = (idAppIsCheap f n_args && args_cheap) + -- A constructor, cheap primop, or partial application + + || idAppIsBottom f n_args + -- Application of a function which + -- always gives bottom; we treat this as cheap + -- because it certainly doesn't need to be shared! + + go (App f a) n_args args_cheap + | not (isRuntimeArg a) = go f n_args args_cheap + | otherwise = go f (n_args + 1) (exprIsCheap a && args_cheap) + + go other n_args args_cheap = False + +idAppIsCheap :: Id -> Int -> Bool +idAppIsCheap id n_val_args + | n_val_args == 0 = True -- Just a type application of -- a variable (f t1 t2 t3) -- counts as WHNF - - || n_val_args < arityLowerBound (getIdArity f) - -isPap fun n_val_args = False + | otherwise = case globalIdDetails id of + DataConId _ -> True + RecordSelId _ -> True -- I'm experimenting with making record selection + -- look cheap, so we will substitute it inside a + -- lambda. Particularly for dictionary field selection + + PrimOpId op -> primOpIsCheap op -- In principle we should worry about primops + -- that return a type variable, since the result + -- might be applied to something, but I'm not going + -- to bother to check the number of args + other -> n_val_args < idArity id \end{code} exprOkForSpeculation returns True of an expression that it is @@ -229,7 +467,8 @@ It returns True iff the expression guarantees to terminate, soon, - without raising an exceptoin + without raising an exception, + without causing a side effect (e.g. writing a mutable variable) E.G. let x = case y# +# 1# of { r# -> I# r# } @@ -245,17 +484,51 @@ side effects, and can't diverge or raise an exception. \begin{code} exprOkForSpeculation :: CoreExpr -> Bool -exprOkForSpeculation (Var v) = isUnLiftedType (idType v) -exprOkForSpeculation (Note _ e) = exprOkForSpeculation e - -exprOkForSpeculation (Con con args) - = conOkForSpeculation con && - and (zipWith ok (filter isValArg args) (fst (conStrictness con))) +exprOkForSpeculation (Lit _) = True +exprOkForSpeculation (Type _) = True +exprOkForSpeculation (Var v) = isUnLiftedType (idType v) +exprOkForSpeculation (Note _ e) = exprOkForSpeculation e +exprOkForSpeculation other_expr + = case collectArgs other_expr of + (Var f, args) -> spec_ok (globalIdDetails f) args + other -> False + where - ok arg demand | isLazy demand = True - | otherwise = exprOkForSpeculation arg - -exprOkForSpeculation other = False -- Conservative + spec_ok (DataConId _) args + = True -- The strictness of the constructor has already + -- been expressed by its "wrapper", so we don't need + -- to take the arguments into account + + spec_ok (PrimOpId op) args + | isDivOp op, -- Special case for dividing operations that fail + [arg1, Lit lit] <- args -- only if the divisor is zero + = not (isZeroLit lit) && exprOkForSpeculation arg1 + -- Often there is a literal divisor, and this + -- can get rid of a thunk in an inner looop + + | otherwise + = primOpOkForSpeculation op && + all exprOkForSpeculation args + -- A bit conservative: we don't really need + -- to care about lazy arguments, but this is easy + + spec_ok other args = False + +isDivOp :: PrimOp -> Bool +-- True of dyadic operators that can fail +-- only if the second arg is zero +-- This function probably belongs in PrimOp, or even in +-- an automagically generated file.. but it's such a +-- special case I thought I'd leave it here for now. +isDivOp IntQuotOp = True +isDivOp IntRemOp = True +isDivOp WordQuotOp = True +isDivOp WordRemOp = True +isDivOp IntegerQuotRemOp = True +isDivOp IntegerDivModOp = True +isDivOp FloatDivOp = True +isDivOp DoubleDivOp = True +isDivOp other = False \end{code} @@ -269,47 +542,146 @@ exprIsBottom e = go 0 e go n (Case e _ _) = go 0 e -- Just check the scrut go n (App e _) = go (n+1) e go n (Var v) = idAppIsBottom v n - go n (Con _ _) = False + go n (Lit _) = False go n (Lam _ _) = False + +idAppIsBottom :: Id -> Int -> Bool +idAppIsBottom id n_val_args = appIsBottom (idNewStrictness id) n_val_args \end{code} @exprIsValue@ returns true for expressions that are certainly *already* -evaluated to WHNF. This is used to decide wether it's ok to change +evaluated to *head* normal form. This is used to decide whether it's ok +to change + case x of _ -> e ===> e and to decide whether it's safe to discard a `seq` -So, it does *not* treat variables as evaluated, unless they say they are +So, it does *not* treat variables as evaluated, unless they say they are. + +But it *does* treat partial applications and constructor applications +as values, even if their arguments are non-trivial, provided the argument +type is lifted; + e.g. (:) (f x) (map f xs) is a value + map (...redex...) is a value +Because `seq` on such things completes immediately + +For unlifted argument types, we have to be careful: + C (f x :: Int#) +Suppose (f x) diverges; then C (f x) is not a value. True, but +this form is illegal (see the invariants in CoreSyn). Args of unboxed +type must be ok-for-speculation (or trivial). \begin{code} exprIsValue :: CoreExpr -> Bool -- True => Value-lambda, constructor, PAP exprIsValue (Type ty) = True -- Types are honorary Values; we don't mind -- copying them -exprIsValue (Var v) = isEvaldUnfolding (getIdUnfolding v) -exprIsValue (Lam b e) = isId b || exprIsValue e +exprIsValue (Lit l) = True +exprIsValue (Lam b e) = isRuntimeVar b || exprIsValue e exprIsValue (Note _ e) = exprIsValue e -exprIsValue (Let _ e) = False -exprIsValue (Case _ _ _) = False -exprIsValue (Con con _) = isWHNFCon con -exprIsValue e@(App _ _) = case collectArgs e of - (Var v, args) -> fun_arity > valArgCount args - where - fun_arity = arityLowerBound (getIdArity v) - _ -> False +exprIsValue (Var v) = idArity v > 0 || isEvaldUnfolding (idUnfolding v) + -- The idArity case catches data cons and primops that + -- don't have unfoldings + -- A worry: what if an Id's unfolding is just itself: + -- then we could get an infinite loop... +exprIsValue other_expr + | (Var fun, args) <- collectArgs other_expr, + isDataConId fun || valArgCount args < idArity fun + = check (idType fun) args + | otherwise + = False + where + -- 'check' checks that unlifted-type args are in + -- fact guaranteed non-divergent + check fun_ty [] = True + check fun_ty (Type _ : args) = case splitForAllTy_maybe fun_ty of + Just (_, ty) -> check ty args + check fun_ty (arg : args) + | isUnLiftedType arg_ty = exprOkForSpeculation arg + | otherwise = check res_ty args + where + (arg_ty, res_ty) = splitFunTy fun_ty \end{code} \begin{code} -exprArity :: CoreExpr -> Int -- How many value lambdas are at the top -exprArity (Lam b e) | isTyVar b = exprArity e - | otherwise = 1 + exprArity e -exprArity (Note note e) | ok_note note = exprArity e -exprArity other = 0 +exprIsConApp_maybe :: CoreExpr -> Maybe (DataCon, [CoreExpr]) +exprIsConApp_maybe (Note (Coerce to_ty from_ty) expr) + = -- Maybe this is over the top, but here we try to turn + -- coerce (S,T) ( x, y ) + -- effectively into + -- ( coerce S x, coerce T y ) + -- This happens in anger in PrelArrExts which has a coerce + -- case coerce memcpy a b of + -- (# r, s #) -> ... + -- where the memcpy is in the IO monad, but the call is in + -- the (ST s) monad + case exprIsConApp_maybe expr of { + Nothing -> Nothing ; + Just (dc, args) -> + + case splitTyConApp_maybe to_ty of { + Nothing -> Nothing ; + Just (tc, tc_arg_tys) | tc /= dataConTyCon dc -> Nothing + | isExistentialDataCon dc -> Nothing + | otherwise -> + -- Type constructor must match + -- We knock out existentials to keep matters simple(r) + let + arity = tyConArity tc + val_args = drop arity args + to_arg_tys = dataConArgTys dc tc_arg_tys + mk_coerce ty arg = mkCoerce ty (exprType arg) arg + new_val_args = zipWith mk_coerce to_arg_tys val_args + in + ASSERT( all isTypeArg (take arity args) ) + ASSERT( equalLength val_args to_arg_tys ) + Just (dc, map Type tc_arg_tys ++ new_val_args) + }} + +exprIsConApp_maybe (Note _ expr) + = exprIsConApp_maybe expr + -- We ignore InlineMe notes in case we have + -- x = __inline_me__ (a,b) + -- All part of making sure that INLINE pragmas never hurt + -- Marcin tripped on this one when making dictionaries more inlinable + -- + -- In fact, we ignore all notes. For example, + -- case _scc_ "foo" (C a b) of + -- C a b -> e + -- should be optimised away, but it will be only if we look + -- through the SCC note. + +exprIsConApp_maybe expr = analyse (collectArgs expr) + where + analyse (Var fun, args) + | Just con <- isDataConId_maybe fun, + args `lengthAtLeast` dataConRepArity con + -- Might be > because the arity excludes type args + = Just (con,args) + + -- Look through unfoldings, but only cheap ones, because + -- we are effectively duplicating the unfolding + analyse (Var fun, []) + | let unf = idUnfolding fun, + isCheapUnfolding unf + = exprIsConApp_maybe (unfoldingTemplate unf) + + analyse other = Nothing \end{code} + +%************************************************************************ +%* * +\subsection{Eta reduction and expansion} +%* * +%************************************************************************ + \begin{code} -exprGenerousArity :: CoreExpr -> Int -- The number of args the thing can be applied to - -- without doing much work +exprEtaExpandArity :: CoreExpr -> Arity +-- The Int is number of value args the thing can be +-- applied to without doing much work +-- -- This is used when eta expanding -- e ==> \xy -> e x y -- @@ -317,48 +689,233 @@ exprGenerousArity :: CoreExpr -> Int -- The number of args the thing can be app -- case x of p -> \s -> ... -- because for I/O ish things we really want to get that \s to the top. -- We are prepared to evaluate x each time round the loop in order to get that --- Hence "generous" arity - -exprGenerousArity (Var v) = arityLowerBound (getIdArity v) -exprGenerousArity (Note note e) - | ok_note note = exprGenerousArity e -exprGenerousArity (Lam x e) - | isId x = 1 + exprGenerousArity e - | otherwise = exprGenerousArity e -exprGenerousArity (Let bind body) - | all exprIsCheap (rhssOfBind bind) = exprGenerousArity body -exprGenerousArity (Case scrut _ alts) - | exprIsCheap scrut = min_zero [exprGenerousArity rhs | (_,_,rhs) <- alts] -exprGenerousArity other = 0 -- Could do better for applications - -min_zero :: [Int] -> Int -- Find the minimum, but zero is the smallest -min_zero (x:xs) = go x xs - where - go 0 xs = 0 -- Nothing beats zero - go min [] = min - go min (x:xs) | x < min = go x xs - | otherwise = go min xs - -ok_note (SCC _) = False -- (Over?) conservative -ok_note (TermUsg _) = False -- Doesn't matter much - -ok_note (Coerce _ _) = True - -- We *do* look through coerces when getting arities. - -- Reason: arities are to do with *representation* and - -- work duplication. - -ok_note InlineCall = True -ok_note InlineMe = False - -- This one is a bit more surprising, but consider - -- f = _inline_me (\x -> e) - -- We DO NOT want to eta expand this to - -- f = \x -> (_inline_me (\x -> e)) x - -- because the _inline_me gets dropped now it is applied, - -- giving just - -- f = \x -> e - -- A Bad Idea + +-- It's all a bit more subtle than it looks. Consider one-shot lambdas +-- let x = expensive in \y z -> E +-- We want this to have arity 2 if the \y-abstraction is a 1-shot lambda +-- Hence the ArityType returned by arityType + +-- NB: this is particularly important/useful for IO state +-- transformers, where we often get +-- let x = E in \ s -> ... +-- and the \s is a real-world state token abstraction. Such +-- abstractions are almost invariably 1-shot, so we want to +-- pull the \s out, past the let x=E. +-- The hack is in Id.isOneShotLambda +-- +-- Consider also +-- f = \x -> error "foo" +-- Here, arity 1 is fine. But if it is +-- f = \x -> case e of +-- True -> error "foo" +-- False -> \y -> x+y +-- then we want to get arity 2. +-- Hence the ABot/ATop in ArityType + + +exprEtaExpandArity e = arityDepth (arityType e) + +-- A limited sort of function type +data ArityType = AFun Bool ArityType -- True <=> one-shot + | ATop -- Know nothing + | ABot -- Diverges + +arityDepth :: ArityType -> Arity +arityDepth (AFun _ ty) = 1 + arityDepth ty +arityDepth ty = 0 + +andArityType ABot at2 = at2 +andArityType ATop at2 = ATop +andArityType (AFun t1 at1) (AFun t2 at2) = AFun (t1 && t2) (andArityType at1 at2) +andArityType at1 at2 = andArityType at2 at1 + +arityType :: CoreExpr -> ArityType + -- (go1 e) = [b1,..,bn] + -- means expression can be rewritten \x_b1 -> ... \x_bn -> body + -- where bi is True <=> the lambda is one-shot + +arityType (Note n e) = arityType e +-- Not needed any more: etaExpand is cleverer +-- | ok_note n = arityType e +-- | otherwise = ATop + +arityType (Var v) + = mk (idArity v) + where + mk :: Arity -> ArityType + mk 0 | isBottomingId v = ABot + | otherwise = ATop + mk n = AFun False (mk (n-1)) + + -- When the type of the Id encodes one-shot-ness, + -- use the idinfo here + + -- Lambdas; increase arity +arityType (Lam x e) | isId x = AFun (isOneShotLambda x) (arityType e) + | otherwise = arityType e + + -- Applications; decrease arity +arityType (App f (Type _)) = arityType f +arityType (App f a) = case arityType f of + AFun one_shot xs | one_shot -> xs + | exprIsCheap a -> xs + other -> ATop + + -- Case/Let; keep arity if either the expression is cheap + -- or it's a 1-shot lambda +arityType (Case scrut _ alts) = case foldr1 andArityType [arityType rhs | (_,_,rhs) <- alts] of + xs@(AFun one_shot _) | one_shot -> xs + xs | exprIsCheap scrut -> xs + | otherwise -> ATop + +arityType (Let b e) = case arityType e of + xs@(AFun one_shot _) | one_shot -> xs + xs | all exprIsCheap (rhssOfBind b) -> xs + | otherwise -> ATop + +arityType other = ATop + +{- NOT NEEDED ANY MORE: etaExpand is cleverer +ok_note InlineMe = False +ok_note other = True + -- Notice that we do not look through __inline_me__ + -- This may seem surprising, but consider + -- f = _inline_me (\x -> e) + -- We DO NOT want to eta expand this to + -- f = \x -> (_inline_me (\x -> e)) x + -- because the _inline_me gets dropped now it is applied, + -- giving just + -- f = \x -> e + -- A Bad Idea +-} +\end{code} + + +\begin{code} +etaExpand :: Arity -- Result should have this number of value args + -> [Unique] + -> CoreExpr -> Type -- Expression and its type + -> CoreExpr +-- (etaExpand n us e ty) returns an expression with +-- the same meaning as 'e', but with arity 'n'. +-- +-- Given e' = etaExpand n us e ty +-- We should have +-- ty = exprType e = exprType e' + +etaExpand n us expr ty + | manifestArity expr >= n = expr -- The no-op case + | otherwise = eta_expand n us expr ty + where + +-- manifestArity sees how many leading value lambdas there are +manifestArity :: CoreExpr -> Arity +manifestArity (Lam v e) | isId v = 1 + manifestArity e + | otherwise = manifestArity e +manifestArity (Note _ e) = manifestArity e +manifestArity e = 0 + +-- etaExpand deals with for-alls. For example: +-- etaExpand 1 E +-- where E :: forall a. a -> a +-- would return +-- (/\b. \y::a -> E b y) +-- +-- It deals with coerces too, though they are now rare +-- so perhaps the extra code isn't worth it + +eta_expand n us expr ty + | n == 0 && + -- The ILX code generator requires eta expansion for type arguments + -- too, but alas the 'n' doesn't tell us how many of them there + -- may be. So we eagerly eta expand any big lambdas, and just + -- cross our fingers about possible loss of sharing in the + -- ILX case. + -- The Right Thing is probably to make 'arity' include + -- type variables throughout the compiler. (ToDo.) + not (isForAllTy ty) + -- Saturated, so nothing to do + = expr + +eta_expand n us (Note note@(Coerce _ ty) e) _ + = Note note (eta_expand n us e ty) + + -- Use mkNote so that _scc_s get pushed inside any lambdas that + -- are generated as part of the eta expansion. We rely on this + -- behaviour in CorePrep, when we eta expand an already-prepped RHS. +eta_expand n us (Note note e) ty + = mkNote note (eta_expand n us e ty) + + -- Short cut for the case where there already + -- is a lambda; no point in gratuitously adding more +eta_expand n us (Lam v body) ty + | isTyVar v + = Lam v (eta_expand n us body (applyTy ty (mkTyVarTy v))) + + | otherwise + = Lam v (eta_expand (n-1) us body (funResultTy ty)) + +eta_expand n us expr ty + = case splitForAllTy_maybe ty of { + Just (tv,ty') -> Lam tv (eta_expand n us (App expr (Type (mkTyVarTy tv))) ty') + + ; Nothing -> + + case splitFunTy_maybe ty of { + Just (arg_ty, res_ty) -> Lam arg1 (eta_expand (n-1) us2 (App expr (Var arg1)) res_ty) + where + arg1 = mkSysLocal SLIT("eta") uniq arg_ty + (uniq:us2) = us + + ; Nothing -> + + case splitNewType_maybe ty of { + Just ty' -> mkCoerce ty ty' (eta_expand n us (mkCoerce ty' ty expr) ty') ; + Nothing -> pprTrace "Bad eta expand" (ppr expr $$ ppr ty) expr + }}} \end{code} +exprArity is a cheap-and-cheerful version of exprEtaExpandArity. +It tells how many things the expression can be applied to before doing +any work. It doesn't look inside cases, lets, etc. The idea is that +exprEtaExpandArity will do the hard work, leaving something that's easy +for exprArity to grapple with. In particular, Simplify uses exprArity to +compute the ArityInfo for the Id. + +Originally I thought that it was enough just to look for top-level lambdas, but +it isn't. I've seen this + + foo = PrelBase.timesInt + +We want foo to get arity 2 even though the eta-expander will leave it +unchanged, in the expectation that it'll be inlined. But occasionally it +isn't, because foo is blacklisted (used in a rule). + +Similarly, see the ok_note check in exprEtaExpandArity. So + f = __inline_me (\x -> e) +won't be eta-expanded. + +And in any case it seems more robust to have exprArity be a bit more intelligent. +But note that (\x y z -> f x y z) +should have arity 3, regardless of f's arity. + +\begin{code} +exprArity :: CoreExpr -> Arity +exprArity e = go e + where + go (Var v) = idArity v + go (Lam x e) | isId x = go e + 1 + | otherwise = go e + go (Note n e) = go e + go (App e (Type t)) = go e + go (App f a) | exprIsCheap a = (go f - 1) `max` 0 + -- NB: exprIsCheap a! + -- f (fac x) does not have arity 2, + -- even if f has arity 3! + -- NB: `max 0`! (\x y -> f x) has arity 2, even if f is + -- unknown, hence arity 0 + go _ = 0 +\end{code} %************************************************************************ %* * @@ -373,24 +930,21 @@ ok_note InlineMe = False \begin{code} cheapEqExpr :: Expr b -> Expr b -> Bool -cheapEqExpr (Var v1) (Var v2) = v1==v2 -cheapEqExpr (Con con1 args1) (Con con2 args2) - = con1 == con2 && - and (zipWithEqual "cheapEqExpr" cheapEqExpr args1 args2) +cheapEqExpr (Var v1) (Var v2) = v1==v2 +cheapEqExpr (Lit lit1) (Lit lit2) = lit1 == lit2 +cheapEqExpr (Type t1) (Type t2) = t1 `eqType` t2 cheapEqExpr (App f1 a1) (App f2 a2) = f1 `cheapEqExpr` f2 && a1 `cheapEqExpr` a2 -cheapEqExpr (Type t1) (Type t2) = t1 == t2 - cheapEqExpr _ _ = False exprIsBig :: Expr b -> Bool -- Returns True of expressions that are too big to be compared by cheapEqExpr +exprIsBig (Lit _) = False exprIsBig (Var v) = False exprIsBig (Type t) = False exprIsBig (App f a) = exprIsBig f || exprIsBig a -exprIsBig (Con _ args) = any exprIsBig args exprIsBig other = True \end{code} @@ -398,6 +952,9 @@ exprIsBig other = True \begin{code} eqExpr :: CoreExpr -> CoreExpr -> Bool -- Works ok at more general type, but only needed at CoreExpr + -- Used in rule matching, so when we find a type we use + -- eqTcType, which doesn't look through newtypes + -- [And it doesn't risk falling into a black hole either.] eqExpr e1 e2 = eq emptyVarEnv e1 e2 where @@ -408,13 +965,13 @@ eqExpr e1 e2 Just v1' -> v1' == v2 Nothing -> v1 == v2 - eq env (Con c1 es1) (Con c2 es2) = c1 == c2 && eq_list env es1 es2 + eq env (Lit lit1) (Lit lit2) = lit1 == lit2 eq env (App f1 a1) (App f2 a2) = eq env f1 f2 && eq env a1 a2 eq env (Lam v1 e1) (Lam v2 e2) = eq (extendVarEnv env v1 v2) e1 e2 eq env (Let (NonRec v1 r1) e1) (Let (NonRec v2 r2) e2) = eq env r1 r2 && eq (extendVarEnv env v1 v2) e1 e2 eq env (Let (Rec ps1) e1) - (Let (Rec ps2) e2) = length ps1 == length ps2 && + (Let (Rec ps2) e2) = equalLength ps1 ps2 && and (zipWith eq_rhs ps1 ps2) && eq env' e1 e2 where @@ -422,13 +979,13 @@ eqExpr e1 e2 eq_rhs (_,r1) (_,r2) = eq env' r1 r2 eq env (Case e1 v1 a1) (Case e2 v2 a2) = eq env e1 e2 && - length a1 == length a2 && + equalLength a1 a2 && and (zipWith (eq_alt env') a1 a2) where env' = extendVarEnv env v1 v2 eq env (Note n1 e1) (Note n2 e2) = eq_note env n1 n2 && eq env e1 e2 - eq env (Type t1) (Type t2) = t1 == t2 + eq env (Type t1) (Type t2) = t1 `eqType` t2 eq env e1 e2 = False eq_list env [] [] = True @@ -439,11 +996,56 @@ eqExpr e1 e2 eq (extendVarEnvList env (vs1 `zip` vs2)) r1 r2 eq_note env (SCC cc1) (SCC cc2) = cc1 == cc2 - eq_note env (Coerce f1 t1) (Coerce f2 t2) = f1==f2 && t1==t2 + eq_note env (Coerce t1 f1) (Coerce t2 f2) = t1 `eqType` t2 && f1 `eqType` f2 eq_note env InlineCall InlineCall = True eq_note env other1 other2 = False \end{code} + +%************************************************************************ +%* * +\subsection{The size of an expression} +%* * +%************************************************************************ + +\begin{code} +coreBindsSize :: [CoreBind] -> Int +coreBindsSize bs = foldr ((+) . bindSize) 0 bs + +exprSize :: CoreExpr -> Int + -- A measure of the size of the expressions + -- It also forces the expression pretty drastically as a side effect +exprSize (Var v) = v `seq` 1 +exprSize (Lit lit) = lit `seq` 1 +exprSize (App f a) = exprSize f + exprSize a +exprSize (Lam b e) = varSize b + exprSize e +exprSize (Let b e) = bindSize b + exprSize e +exprSize (Case e b as) = exprSize e + varSize b + foldr ((+) . altSize) 0 as +exprSize (Note n e) = noteSize n + exprSize e +exprSize (Type t) = seqType t `seq` 1 + +noteSize (SCC cc) = cc `seq` 1 +noteSize (Coerce t1 t2) = seqType t1 `seq` seqType t2 `seq` 1 +noteSize InlineCall = 1 +noteSize InlineMe = 1 + +varSize :: Var -> Int +varSize b | isTyVar b = 1 + | otherwise = seqType (idType b) `seq` + megaSeqIdInfo (idInfo b) `seq` + 1 + +varsSize = foldr ((+) . varSize) 0 + +bindSize (NonRec b e) = varSize b + exprSize e +bindSize (Rec prs) = foldr ((+) . pairSize) 0 prs + +pairSize (b,e) = varSize b + exprSize e + +altSize (c,bs,e) = c `seq` varsSize bs + exprSize e +\end{code} + + %************************************************************************ %* * \subsection{Hashing} @@ -452,29 +1054,27 @@ eqExpr e1 e2 \begin{code} hashExpr :: CoreExpr -> Int -hashExpr e = abs (hash_expr e) - -- Negative numbers kill UniqFM +hashExpr e | hash < 0 = 77 -- Just in case we hit -maxInt + | otherwise = hash + where + hash = abs (hash_expr e) -- Negative numbers kill UniqFM hash_expr (Note _ e) = hash_expr e hash_expr (Let (NonRec b r) e) = hashId b hash_expr (Let (Rec ((b,r):_)) e) = hashId b hash_expr (Case _ b _) = hashId b -hash_expr (App f e) = hash_expr f + fast_hash_expr e +hash_expr (App f e) = hash_expr f * fast_hash_expr e hash_expr (Var v) = hashId v -hash_expr (Con con args) = foldr ((+) . fast_hash_expr) (hashCon con) args +hash_expr (Lit lit) = hashLiteral lit hash_expr (Lam b _) = hashId b -hash_expr (Type t) = trace "hash_expr: type" 0 -- Shouldn't happen +hash_expr (Type t) = trace "hash_expr: type" 1 -- Shouldn't happen fast_hash_expr (Var v) = hashId v -fast_hash_expr (Con con args) = fast_hash_args args con +fast_hash_expr (Lit lit) = hashLiteral lit fast_hash_expr (App f (Type _)) = fast_hash_expr f fast_hash_expr (App f a) = fast_hash_expr a fast_hash_expr (Lam b _) = hashId b -fast_hash_expr other = 0 - -fast_hash_args [] con = hashCon con -fast_hash_args (Type t : args) con = fast_hash_args args con -fast_hash_args (arg : args) con = fast_hash_expr arg +fast_hash_expr other = 1 hashId :: Id -> Int hashId id = hashName (idName id)