X-Git-Url: http://git.megacz.com/?a=blobdiff_plain;f=ghc%2Fcompiler%2FcoreSyn%2FCoreUtils.lhs;h=24b1f35cf2a5fec6ce9a8cc9161417d743d7b19b;hb=fda5605a79bc7bc5f0ef5bbaf241f89d951b65ce;hp=7c1b62ab4189f6956af4d3ba00f6855a1f01803d;hpb=647eb48674623156f7f5b699e4ecee9410ff585f;p=ghc-hetmet.git diff --git a/ghc/compiler/coreSyn/CoreUtils.lhs b/ghc/compiler/coreSyn/CoreUtils.lhs index 7c1b62a..24b1f35 100644 --- a/ghc/compiler/coreSyn/CoreUtils.lhs +++ b/ghc/compiler/coreSyn/CoreUtils.lhs @@ -1,53 +1,73 @@ % -% (c) The GRASP/AQUA Project, Glasgow University, 1992-1996 +% (c) The GRASP/AQUA Project, Glasgow University, 1992-1998 % \section[CoreUtils]{Utility functions on @Core@ syntax} \begin{code} module CoreUtils ( - coreExprType, coreAltsType, coreExprCc, + -- Construction + mkNote, mkInlineMe, mkSCC, mkCoerce, + bindNonRec, mkIfThenElse, mkAltExpr, + mkPiType, - mkCoreIfThenElse, - argToExpr, - unTagBinders, unTagBindersAlts, - - maybeErrorApp, - nonErrorRHSs, - squashableDictishCcExpr, - idSpecVars + -- Properties of expressions + exprType, coreAltsType, + exprIsBottom, exprIsDupable, exprIsTrivial, exprIsCheap, + exprIsValue,exprOkForSpeculation, exprIsBig, + exprIsConApp_maybe, + idAppIsBottom, idAppIsCheap, + + -- Expr transformation + etaReduceExpr, exprEtaExpandArity, + + -- Size + coreBindsSize, + + -- Hashing + hashExpr, + + -- Equality + cheapEqExpr, eqExpr, applyTypeToArgs ) where #include "HsVersions.h" -import CoreSyn -import CostCentre ( isDictCC, CostCentre, noCostCentre ) -import MkId ( mkSysLocal ) -import Id ( idType, isBottomingId, getIdSpecialisation, - dataConRepType, - Id - ) -import Literal ( literalType, Literal(..) ) -import Maybes ( catMaybes, maybeToBool ) -import PprCore -import PrimOp ( primOpType, PrimOp(..) ) -import SpecEnv ( specEnvValues ) -import SrcLoc ( noSrcLoc ) -import Type ( mkFunTy, mkForAllTy, mkTyVarTy, - splitFunTy_maybe, applyTys, isUnpointedType, - splitSigmaTy, splitFunTys, - Type +import GlaExts -- For `xori` + +import CoreSyn +import CoreFVs ( exprFreeVars ) +import PprCore ( pprCoreExpr ) +import Var ( Var, isId, isTyVar ) +import VarSet +import VarEnv +import Name ( isLocallyDefined, hashName ) +import Literal ( Literal, hashLiteral, literalType, litIsDupable ) +import DataCon ( DataCon, dataConRepArity ) +import PrimOp ( primOpOkForSpeculation, primOpIsCheap, + primOpIsDupable ) +import Id ( Id, idType, idFlavour, idStrictness, idLBVarInfo, + mkWildId, idArity, idName, idUnfolding, idInfo, + isDataConId_maybe, isPrimOpId_maybe ) -import TysWiredIn ( trueDataCon, falseDataCon ) -import BasicTypes ( Unused ) -import UniqSupply ( returnUs, thenUs, - mapAndUnzipUs, getUnique, - UniqSM +import IdInfo ( arityLowerBound, InlinePragInfo(..), + LBVarInfo(..), + IdFlavour(..), + megaSeqIdInfo ) +import Demand ( appIsBottom ) +import Type ( Type, mkFunTy, mkForAllTy, + splitFunTy_maybe, tyVarsOfType, tyVarsOfTypes, + isNotUsgTy, mkUsgTy, unUsgTy, UsageAnn(..), + applyTys, isUnLiftedType, seqType ) -import Outputable ( assertPanic, pprPanic, ppr, vcat, panic ) - +import TysWiredIn ( boolTy, stringTy, trueDataCon, falseDataCon ) +import CostCentre ( CostCentre ) +import Maybes ( maybeToBool ) +import Outputable +import TysPrim ( alphaTy ) -- Debugging only \end{code} + %************************************************************************ %* * \subsection{Find the type of a Core atom/expression} @@ -55,371 +75,682 @@ import Outputable ( assertPanic, pprPanic, ppr, vcat, panic ) %************************************************************************ \begin{code} -coreExprType :: CoreExpr -> Type +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 (TermUsg u) e) = mkUsgTy u (unUsgTy (exprType 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 (exprType fun) args + +exprType other = pprTrace "exprType" (pprCoreExpr other) alphaTy + +coreAltsType :: [CoreAlt] -> Type +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. -coreExprType (Var var) = idType var -coreExprType (Lit lit) = literalType lit +\begin{code} +mkPiType :: Var -> Type -> Type -- The more polymorphic version doesn't work... +mkPiType v ty | isId v = (case idLBVarInfo v of + IsOneShotLambda -> mkUsgTy UsOnce + otherwise -> id) $ + mkFunTy (idType v) ty + | isTyVar v = mkForAllTy v ty +\end{code} -coreExprType (Let _ body) = coreExprType body -coreExprType (Case _ alts) = coreAltsType alts +\begin{code} +-- The first argument is just for debugging +applyTypeToArgs :: CoreExpr -> Type -> [CoreExpr] -> Type +applyTypeToArgs e op_ty [] = op_ty -coreExprType (Note (Coerce ty _) e) = ty -coreExprType (Note other_note e) = coreExprType e +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 + where + (tys, rest_args) = go [ty] args + go tys (Type ty : args) = go (ty:tys) args + go tys rest_args = (reverse tys, rest_args) --- a Con is a fully-saturated application of a data constructor --- a Prim is of a PrimOp +applyTypeToArgs e op_ty (other_arg : args) + = case (splitFunTy_maybe op_ty) of + Just (_, res_ty) -> applyTypeToArgs e res_ty args + Nothing -> pprPanic "applyTypeToArgs" (pprCoreExpr e) +\end{code} -coreExprType (Con con args) = --- pprTrace "appTyArgs" (hsep [ppr con, semi, --- ppr con_ty, semi, --- ppr args]) $ - applyTypeToArgs con_ty args - where - con_ty = dataConRepType con -coreExprType (Prim op args) = applyTypeToArgs (primOpType op) args -coreExprType (Lam (ValBinder binder) expr) - = idType binder `mkFunTy` coreExprType expr +%************************************************************************ +%* * +\subsection{Attaching notes} +%* * +%************************************************************************ -coreExprType (Lam (TyBinder tyvar) expr) - = mkForAllTy tyvar (coreExprType expr) +mkNote removes redundant coercions, and SCCs where possible -coreExprType (App expr (TyArg ty)) - = -- Gather type args; more efficient to instantiate the type all at once - go expr [ty] - where - go (App expr (TyArg ty)) tys = go expr (ty:tys) - go expr tys = applyTys (coreExprType expr) tys - -coreExprType (App expr val_arg) - = ASSERT(isValArg val_arg) - let - fun_ty = coreExprType expr - in - case (splitFunTy_maybe fun_ty) of - Just (_, result_ty) -> result_ty -#ifdef DEBUG - Nothing -> pprPanic "coreExprType:\n" - (vcat [ppr fun_ty, ppr (App expr val_arg)]) -#endif +\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. + \begin{code} -coreAltsType :: CoreCaseAlts -> Type +mkInlineMe e | exprIsTrivial e = e + | otherwise = Note InlineMe e +\end{code} -coreAltsType (AlgAlts [] deflt) = default_ty deflt -coreAltsType (AlgAlts ((_,_,rhs1):_) _) = coreExprType rhs1 -coreAltsType (PrimAlts [] deflt) = default_ty deflt -coreAltsType (PrimAlts ((_,rhs1):_) _) = coreExprType rhs1 -default_ty NoDefault = panic "coreExprType:Case:default_ty" -default_ty (BindDefault _ rhs) = coreExprType rhs +\begin{code} +mkCoerce :: Type -> Type -> CoreExpr -> CoreExpr + +mkCoerce to_ty from_ty (Note (Coerce to_ty2 from_ty2) expr) + = ASSERT( from_ty == to_ty2 ) + mkCoerce to_ty from_ty2 expr + +mkCoerce to_ty from_ty expr + | to_ty == from_ty = expr + | otherwise = ASSERT( from_ty == exprType expr ) + Note (Coerce to_ty from_ty) expr \end{code} \begin{code} -applyTypeToArgs op_ty (TyArg ty : args) - = -- Accumulate type arguments so we can instantiate all at once - applyTypeToArgs (applyTys op_ty tys) rest_args - where - (tys, rest_args) = go [ty] args - go tys (TyArg ty : args) = go (ty:tys) args - go tys rest_args = (reverse tys, rest_args) - -applyTypeToArgs op_ty (val_or_lit_arg:args) - = case (splitFunTy_maybe op_ty) of - Just (_, res_ty) -> applyTypeToArgs res_ty args +mkSCC :: CostCentre -> Expr b -> Expr b + -- Note: Nested SCC's *are* preserved for the benefit of + -- cost centre stack profiling (Durham) -applyTypeToArgs op_ty [] = op_ty +mkSCC cc (Lit lit) = Lit lit +mkSCC cc (Lam x e) = Lam x (mkSCC cc e) -- Move _scc_ inside lambda +mkSCC cc expr = Note (SCC cc) expr \end{code} -coreExprCc gets the cost centre enclosing an expression, if any. -It looks inside lambdas because (scc "foo" \x.e) = \x.scc "foo" e + +%************************************************************************ +%* * +\subsection{Other expression construction} +%* * +%************************************************************************ \begin{code} -coreExprCc :: GenCoreExpr val_bdr val_occ flexi -> CostCentre -coreExprCc (Note (SCC cc) e) = cc -coreExprCc (Note other_note e) = coreExprCc e -coreExprCc (Lam _ e) = coreExprCc e -coreExprCc other = noCostCentre +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 + | isUnLiftedType (idType bndr) = Case rhs bndr [(DEFAULT,[],body)] + | otherwise = Let (NonRec bndr rhs) body +\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{Routines to manufacture bits of @CoreExpr@} +\subsection{Figuring out things about expressions} %* * %************************************************************************ +@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 + + \begin{code} -mkCoreIfThenElse (Var bool) then_expr else_expr - | bool == trueDataCon = then_expr - | bool == falseDataCon = else_expr - -mkCoreIfThenElse guard then_expr else_expr - = Case guard - (AlgAlts [ (trueDataCon, [], then_expr), - (falseDataCon, [], else_expr) ] - NoDefault ) +exprIsTrivial (Var v) + | Just op <- isPrimOpId_maybe v = primOpIsDupable op + | otherwise = True +exprIsTrivial (Type _) = True +exprIsTrivial (Lit lit) = True +exprIsTrivial (App e arg) = isTypeArg arg && exprIsTrivial e +exprIsTrivial (Note _ e) = exprIsTrivial e +exprIsTrivial (Lam b body) | isTyVar b = exprIsTrivial body +exprIsTrivial other = False \end{code} -For making @Apps@ and @Lets@, we must take appropriate evasive -action if the thing being bound has unboxed type. @mkCoApp@ requires -a name supply to do its work. -@mkCoApps@, @mkCoCon@ and @mkCoPrim@ also handle the -arguments-must-be-atoms constraint. +@exprIsDupable@ is true of expressions that can be duplicated at a modest + cost in code size. This will only happen in different case + branches, so there's no issue about duplicating work. + + That is, exprIsDupable returns True of (f x) even if + f is very very expensive to call. + + Its only purpose is to avoid fruitless let-binding + and then inlining of case join points + \begin{code} -data CoreArgOrExpr - = AnArg CoreArg - | AnExpr CoreExpr +exprIsDupable (Type _) = True +exprIsDupable (Var v) = True +exprIsDupable (Lit lit) = litIsDupable lit +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 +\end{code} + +@exprIsCheap@ looks at a Core expression and returns \tr{True} if +it is obviously in weak head normal form, or is cheap to get to WHNF. +[Note that that's not the same as exprIsDupable; an expression might be +big, and hence not dupable, but still cheap.] + +By ``cheap'' we mean a computation we're willing to: + push inside a lambda, or + inline at more than one place +That might mean it gets evaluated more than once, instead of being +shared. The main examples of things which aren't WHNF but are +``cheap'' are: + + * case e of + pi -> ei + (where e, and all the ei are cheap) + + * let x = e in b + (where e and b are cheap) -mkCoApps :: CoreExpr -> [CoreArgOrExpr] -> UniqSM CoreExpr -mkCoCon :: Id -> [CoreArgOrExpr] -> UniqSM CoreExpr -mkCoPrim :: PrimOp -> [CoreArgOrExpr] -> UniqSM CoreExpr + * op x1 ... xn + (where op is a cheap primitive operator) -mkCoApps fun args = co_thing (mkGenApp fun) args -mkCoCon con args = co_thing (Con con) args -mkCoPrim op args = co_thing (Prim op) args + * error "foo" + (because we are happy to substitute it inside a lambda) -co_thing :: ([CoreArg] -> CoreExpr) - -> [CoreArgOrExpr] - -> UniqSM CoreExpr +Notice that a variable is considered 'cheap': we can push it inside a lambda, +because sharing will make sure it is only evaluated once. -co_thing thing arg_exprs - = mapAndUnzipUs expr_to_arg arg_exprs `thenUs` \ (args, maybe_binds) -> - returnUs (mkCoLetsUnboxedToCase (catMaybes maybe_binds) (thing args)) +\begin{code} +exprIsCheap :: CoreExpr -> Bool +exprIsCheap (Lit lit) = True +exprIsCheap (Type _) = True +exprIsCheap (Var _) = True +exprIsCheap (Note _ e) = exprIsCheap e +exprIsCheap (Lam x e) = if isId x then True else 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 - expr_to_arg :: CoreArgOrExpr - -> UniqSM (CoreArg, Maybe CoreBinding) - - expr_to_arg (AnArg arg) = returnUs (arg, Nothing) - expr_to_arg (AnExpr (Var v)) = returnUs (VarArg v, Nothing) - expr_to_arg (AnExpr (Lit l)) = returnUs (LitArg l, Nothing) - expr_to_arg (AnExpr other_expr) - = let - e_ty = coreExprType other_expr - in - getUnique `thenUs` \ uniq -> - let - new_var = mkSysLocal SLIT("a") uniq e_ty noSrcLoc - in - returnUs (VarArg new_var, Just (NonRec new_var other_expr)) + 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 + | isTypeArg 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 + | otherwise = case idFlavour 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 + + * safe to evaluate even if normal order eval might not + evaluate the expression at all, or + + * safe *not* to evaluate even if normal order would do so + +It returns True iff + + the expression guarantees to terminate, + soon, + 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# } + in E +==> + case y# +# 1# of { r# -> + let x = I# r# + in E + } + +We can only do this if the (y+1) is ok for speculation: it has no +side effects, and can't diverge or raise an exception. + \begin{code} -argToExpr :: - GenCoreArg val_occ flexi -> GenCoreExpr val_bdr val_occ flexi +exprOkForSpeculation :: CoreExpr -> Bool +exprOkForSpeculation (Lit _) = True +exprOkForSpeculation (Var v) = isUnLiftedType (idType v) +exprOkForSpeculation (Note _ e) = exprOkForSpeculation e +exprOkForSpeculation other_expr + = go other_expr 0 True + where + go (Var f) n_args args_ok + = case idFlavour f of + DataConId _ -> True -- The strictness of the constructor has already + -- been expressed by its "wrapper", so we don't need + -- to take the arguments into account + + PrimOpId op -> primOpOkForSpeculation op && args_ok + -- A bit conservative: we don't really need + -- to care about lazy arguments, but this is easy -argToExpr (VarArg v) = Var v -argToExpr (LitArg lit) = Lit lit + other -> False + + go (App f a) n_args args_ok + | isTypeArg a = go f n_args args_ok + | otherwise = go f (n_args + 1) (exprOkForSpeculation a && args_ok) + + go other n_args args_ok = False \end{code} -All the following functions operate on binders, perform a uniform -transformation on them; ie. the function @(\ x -> (x,False))@ -annotates all binders with False. \begin{code} -unTagBinders :: GenCoreExpr (Id,tag) bdee flexi -> GenCoreExpr Id bdee flexi -unTagBinders expr = bop_expr fst expr - -unTagBindersAlts :: GenCoreCaseAlts (Id,tag) bdee flexi -> GenCoreCaseAlts Id bdee flexi -unTagBindersAlts alts = bop_alts fst alts +exprIsBottom :: CoreExpr -> Bool -- True => definitely bottom +exprIsBottom e = go 0 e + where + -- n is the number of args + go n (Note _ e) = go n e + go n (Let _ e) = go n 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 (Lit _) = False + go n (Lam _ _) = False + +idAppIsBottom :: Id -> Int -> Bool +idAppIsBottom id n_val_args = appIsBottom (idStrictness 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 + 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 + \begin{code} -bop_expr :: (a -> b) -> GenCoreExpr a bdee flexi -> GenCoreExpr b bdee flexi - -bop_expr f (Var b) = Var b -bop_expr f (Lit lit) = Lit lit -bop_expr f (Con con args) = Con con args -bop_expr f (Prim op args) = Prim op args -bop_expr f (Lam binder expr) = Lam (bop_binder f binder) (bop_expr f expr) -bop_expr f (App expr arg) = App (bop_expr f expr) arg -bop_expr f (Note note expr) = Note note (bop_expr f expr) -bop_expr f (Let bind expr) = Let (bop_bind f bind) (bop_expr f expr) -bop_expr f (Case expr alts) = Case (bop_expr f expr) (bop_alts f alts) - -bop_binder f (ValBinder v) = ValBinder (f v) -bop_binder f (TyBinder t) = TyBinder t - -bop_bind f (NonRec b e) = NonRec (f b) (bop_expr f e) -bop_bind f (Rec pairs) = Rec [(f b, bop_expr f e) | (b, e) <- pairs] - -bop_alts f (AlgAlts alts deflt) - = AlgAlts [ (con, [f b | b <- binders], bop_expr f e) - | (con, binders, e) <- alts ] - (bop_deflt f deflt) - -bop_alts f (PrimAlts alts deflt) - = PrimAlts [ (lit, bop_expr f e) | (lit, e) <- alts ] - (bop_deflt f deflt) - -bop_deflt f (NoDefault) = NoDefault -bop_deflt f (BindDefault b expr) = BindDefault (f b) (bop_expr f expr) +exprIsValue :: CoreExpr -> Bool -- True => Value-lambda, constructor, PAP +exprIsValue (Type ty) = True -- Types are honorary Values; we don't mind + -- copying them +exprIsValue (Lit l) = True +exprIsValue (Lam b e) = isId b || exprIsValue e +exprIsValue (Note _ e) = exprIsValue e +exprIsValue other_expr + = go other_expr 0 + where + go (Var f) n_args = idAppIsValue f n_args + + go (App f a) n_args + | isTypeArg a = go f n_args + | otherwise = go f (n_args + 1) + + go (Note _ f) n_args = go f n_args + + go other n_args = False + +idAppIsValue :: Id -> Int -> Bool +idAppIsValue id n_val_args + = case idFlavour id of + DataConId _ -> True + PrimOpId _ -> n_val_args < idArity id + other | n_val_args == 0 -> isEvaldUnfolding (idUnfolding id) + | otherwise -> n_val_args < idArity id + -- A worry: what if an Id's unfolding is just itself: + -- then we could get an infinite loop... \end{code} -OLD (but left here because of the nice example): @singleAlt@ checks -whether a bunch of case alternatives is actually just one alternative. -It specifically {\em ignores} alternatives which consist of just a -call to @error@, because they won't result in any code duplication. +\begin{code} +exprIsConApp_maybe :: CoreExpr -> Maybe (DataCon, [CoreExpr]) +exprIsConApp_maybe expr + = analyse (collectArgs expr) + where + analyse (Var fun, args) + | maybeToBool maybe_con_app = maybe_con_app + where + maybe_con_app = case isDataConId_maybe fun of + Just con | length args >= dataConRepArity con + -- Might be > because the arity excludes type args + -> Just (con, args) + other -> Nothing -Example: -\begin{verbatim} - case (case of - True -> - False -> error "Foo") of - + analyse (Var fun, []) + = case maybeUnfoldingTemplate (idUnfolding fun) of + Nothing -> Nothing + Just unf -> exprIsConApp_maybe unf -===> + analyse other = Nothing +\end{code} - case of - True -> case of - - False -> case error "Foo" of - -===> +%************************************************************************ +%* * +\subsection{Eta reduction and expansion} +%* * +%************************************************************************ - case of - True -> case of - - False -> error "Foo" -\end{verbatim} -Notice that the \tr{} don't get duplicated. +@etaReduceExpr@ trys an eta reduction at the top level of a Core Expr. -\begin{code} -nonErrorRHSs :: GenCoreCaseAlts a Id Unused -> [GenCoreExpr a Id Unused] +e.g. \ x y -> f x y ===> f -nonErrorRHSs alts - = filter not_error_app (find_rhss alts) - where - find_rhss (AlgAlts as deflt) = [rhs | (_,_,rhs) <- as] ++ deflt_rhs deflt - find_rhss (PrimAlts as deflt) = [rhs | (_,rhs) <- as] ++ deflt_rhs deflt +But we only do this if it gets rid of a whole lambda, not part. +The idea is that lambdas are often quite helpful: they indicate +head normal forms, so we don't want to chuck them away lightly. - deflt_rhs NoDefault = [] - deflt_rhs (BindDefault _ rhs) = [rhs] +\begin{code} +etaReduceExpr :: CoreExpr -> CoreExpr + -- ToDo: we should really check that we don't turn a non-bottom + -- lambda into a bottom variable. Sigh - not_error_app rhs - = case (maybeErrorApp rhs Nothing) of - Just _ -> False - Nothing -> True -\end{code} +etaReduceExpr expr@(Lam bndr body) + = check (reverse binders) body + where + (binders, body) = collectBinders expr -maybeErrorApp checks whether an expression is of the form + check [] body + | not (any (`elemVarSet` body_fvs) binders) + = body -- Success! + where + body_fvs = exprFreeVars body - error ty args + check (b : bs) (App fun arg) + | (varToCoreExpr b `cheapEqExpr` arg) + = check bs fun -If so, it returns + check _ _ = expr -- Bale out - Just (error ty' args) +etaReduceExpr expr = expr -- The common case +\end{code} + -where ty' is supplied as an argument to maybeErrorApp. +\begin{code} +exprEtaExpandArity :: CoreExpr -> Int -- The number of args the thing can be applied to + -- without doing much work +-- This is used when eta expanding +-- e ==> \xy -> e x y +-- +-- It returns 1 (or more) to: +-- 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 + +exprEtaExpandArity e + = go e `max` 0 -- Never go -ve! + where + go (Var v) = idArity v + go (App f (Type _)) = go f + go (App f a) | exprIsCheap a = go f - 1 + go (Lam x e) | isId x = go e + 1 + | otherwise = go e + go (Note n e) | ok_note n = go e + go (Case scrut _ alts) + | exprIsCheap scrut = min_zero [go rhs | (_,_,rhs) <- alts] + go (Let b e) + | all exprIsCheap (rhssOfBind b) = go e + + go other = 0 + + ok_note (Coerce _ _) = True + ok_note InlineCall = True + ok_note other = False + -- Notice that we do not look through __inline_me__ + -- 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 + +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 -Here's where it is useful: +\end{code} - case (error ty "Foo" e1 e2) of - ===> - error ty' "Foo" -where ty' is the type of any of the alternatives. You might think -this never occurs, but see the comments on the definition of -@singleAlt@. +%************************************************************************ +%* * +\subsection{Equality} +%* * +%************************************************************************ -Note: we *avoid* the case where ty' might end up as a primitive type: -this is very uncool (totally wrong). +@cheapEqExpr@ is a cheap equality test which bales out fast! + True => definitely equal + False => may or may not be equal -NOTICE: in the example above we threw away e1 and e2, but not the -string "Foo". How did we know to do that? +\begin{code} +cheapEqExpr :: Expr b -> Expr b -> Bool -Answer: for now anyway, we only handle the case of a function whose -type is of form +cheapEqExpr (Var v1) (Var v2) = v1==v2 +cheapEqExpr (Lit lit1) (Lit lit2) = lit1 == lit2 +cheapEqExpr (Type t1) (Type t2) = t1 == t2 - bottomingFn :: forall a. t1 -> ... -> tn -> a - ^---------------------^ NB! +cheapEqExpr (App f1 a1) (App f2 a2) + = f1 `cheapEqExpr` f2 && a1 `cheapEqExpr` a2 -Furthermore, we only count a bottomingApp if the function is applied -to more than n args. If so, we transform: +cheapEqExpr _ _ = False - bottomingFn ty e1 ... en en+1 ... em -to - bottomingFn ty' e1 ... en +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 other = True +\end{code} -That is, we discard en+1 .. em \begin{code} -maybeErrorApp - :: GenCoreExpr a Id Unused -- Expr to look at - -> Maybe Type -- Just ty => a result type *already cloned*; - -- Nothing => don't know result ty; we - -- *pretend* that the result ty won't be - -- primitive -- somebody later must - -- ensure this. - -> Maybe (GenCoreExpr b Id Unused) - -maybeErrorApp expr result_ty_maybe - = case (collectArgs expr) of - (Var fun, [ty], other_args) - | isBottomingId fun - && maybeToBool result_ty_maybe -- we *know* the result type - -- (otherwise: live a fairy-tale existence...) - && not (isUnpointedType result_ty) -> - - case (splitSigmaTy (idType fun)) of - ([tyvar], [], tau_ty) -> - case (splitFunTys tau_ty) of { (arg_tys, res_ty) -> - let - n_args_to_keep = length arg_tys - args_to_keep = take n_args_to_keep other_args - in - if (res_ty == mkTyVarTy tyvar) - && n_args_to_keep <= length other_args - then - -- Phew! We're in business - Just (mkGenApp (Var fun) (TyArg result_ty : args_to_keep)) - else - Nothing - } - - other -> Nothing -- Function type wrong shape - other -> Nothing +eqExpr :: CoreExpr -> CoreExpr -> Bool + -- Works ok at more general type, but only needed at CoreExpr +eqExpr e1 e2 + = eq emptyVarEnv e1 e2 where - Just result_ty = result_ty_maybe + -- The "env" maps variables in e1 to variables in ty2 + -- So when comparing lambdas etc, + -- we in effect substitute v2 for v1 in e1 before continuing + eq env (Var v1) (Var v2) = case lookupVarEnv env v1 of + Just v1' -> v1' == v2 + Nothing -> v1 == v2 + + 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 && + and (zipWith eq_rhs ps1 ps2) && + eq env' e1 e2 + where + env' = extendVarEnvList env [(v1,v2) | ((v1,_),(v2,_)) <- zip ps1 ps2] + 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 && + 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 e1 e2 = False + + eq_list env [] [] = True + eq_list env (e1:es1) (e2:es2) = eq env e1 e2 && eq_list env es1 es2 + eq_list env es1 es2 = False + + eq_alt env (c1,vs1,r1) (c2,vs2,r2) = c1==c2 && + eq (extendVarEnvList env (vs1 `zip` vs2)) r1 r2 + + eq_note env (SCC cc1) (SCC cc2) = cc1 == cc2 + eq_note env (Coerce t1 f1) (Coerce t2 f2) = t1==t2 && f1==f2 + eq_note env InlineCall InlineCall = True + eq_note env other1 other2 = False \end{code} -\begin{code} -squashableDictishCcExpr :: CostCentre -> GenCoreExpr a b c -> Bool -squashableDictishCcExpr cc expr - = if not (isDictCC cc) then - False -- that was easy... - else - squashable expr -- note: quite like the "atomic_rhs" stuff in simplifier - where - squashable (Var _) = True - squashable (Con _ _) = True -- I think so... WDP 94/09 - squashable (Prim _ _) = True -- ditto - squashable (App f a) - | notValArg a = squashable f - squashable other = False +%************************************************************************ +%* * +\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) = varSize v +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 +noteSize (TermUsg usg) = usg `seq` 1 + +exprsSize = foldr ((+) . exprSize) 0 + +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} -Given an Id, idSpecVars returns all its specialisations. -We extract these from its SpecEnv. -This is used by the occurrence analyser and free-var finder; -we regard an Id's specialisations as free in the Id's definition. +%************************************************************************ +%* * +\subsection{Hashing} +%* * +%************************************************************************ \begin{code} -idSpecVars :: Id -> [Id] -idSpecVars id - = map get_spec (specEnvValues (getIdSpecialisation id)) - where - -- get_spec is another cheapo function like dictRhsFVs - -- It knows what these specialisation temlates look like, - -- and just goes for the jugular - get_spec (App f _) = get_spec f - get_spec (Lam _ b) = get_spec b - get_spec (Var v) = v +hashExpr :: CoreExpr -> Int +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 (Var v) = hashId v +hash_expr (Lit lit) = hashLiteral lit +hash_expr (Lam b _) = hashId b +hash_expr (Type t) = trace "hash_expr: type" 1 -- Shouldn't happen + +fast_hash_expr (Var v) = hashId v +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 = 1 + +hashId :: Id -> Int +hashId id = hashName (idName id) \end{code}