X-Git-Url: http://git.megacz.com/?a=blobdiff_plain;f=ghc%2Fcompiler%2FsimplCore%2FSimplUtils.lhs;h=e8a6433279a1aa7213b680e9ce878bc8943afc01;hb=849b7bca043a521fc60e18393cc311c754f2d9fe;hp=9c5c64743d1b4c4d21712a06f38712b9c7371293;hpb=937b23b94b458172442ac583f8d5b6f5a093a24b;p=ghc-hetmet.git diff --git a/ghc/compiler/simplCore/SimplUtils.lhs b/ghc/compiler/simplCore/SimplUtils.lhs index 9c5c647..e8a6433 100644 --- a/ghc/compiler/simplCore/SimplUtils.lhs +++ b/ghc/compiler/simplCore/SimplUtils.lhs @@ -6,108 +6,479 @@ \begin{code} module SimplUtils ( simplBinder, simplBinders, simplIds, - mkRhsTyLam, - etaCoreExpr, - etaExpandCount, - mkCase, findAlt, findDefault + transformRhs, + mkCase, findAlt, findDefault, + + -- The continuation type + SimplCont(..), DupFlag(..), contIsDupable, contResultType, + countValArgs, countArgs, mkRhsStop, mkStop, + getContArgs, interestingCallContext, interestingArg, isStrictType, discardInline + ) where #include "HsVersions.h" -import BinderInfo -import CmdLineOpts ( opt_DoEtaReduction, switchIsOn, SimplifierSwitch(..) ) -import CoreSyn -import CoreUtils ( exprIsCheap, exprIsTrivial, exprFreeVars, cheapEqExpr, - FormSummary(..), - substId, substIds +import CmdLineOpts ( switchIsOn, SimplifierSwitch(..), + opt_SimplDoLambdaEtaExpansion, opt_SimplCaseMerge, opt_DictsStrict, + opt_UF_UpdateInPlace ) -import Id ( Id, idType, getIdArity, isId, idName, - getInlinePragma, setInlinePragma, - getIdDemandInfo +import CoreSyn +import CoreUtils ( exprIsTrivial, cheapEqExpr, exprType, exprIsCheap, exprEtaExpandArity, bindNonRec ) +import Subst ( InScopeSet, mkSubst, substBndrs, substBndr, substIds, substExpr ) +import Id ( idType, idName, + idUnfolding, idStrictness, + mkVanillaId, idInfo ) -import IdInfo ( arityLowerBound, InlinePragInfo(..) ) -import Maybes ( maybeToBool ) -import Const ( Con(..) ) -import Name ( isLocalName ) +import IdInfo ( StrictnessInfo(..), ArityInfo, atLeastArity ) +import Maybes ( maybeToBool, catMaybes ) +import Name ( setNameUnique ) +import Demand ( isStrict ) import SimplMonad -import Type ( Type, tyVarsOfType, tyVarsOfTypes, mkForAllTys, - splitTyConApp_maybe, substTyVar, mkTyVarTys +import Type ( Type, mkForAllTys, seqType, repType, + splitTyConApp_maybe, mkTyVarTys, splitFunTys, + isDictTy, isDataType, isUnLiftedType, + splitRepFunTys ) -import Var ( setVarUnique ) -import VarSet -import UniqSupply ( splitUniqSupply, uniqFromSupply ) -import Util ( zipWithEqual, mapAccumL ) +import TyCon ( tyConDataConsIfAvailable ) +import DataCon ( dataConRepArity ) +import VarEnv ( SubstEnv ) +import Util ( lengthExceeds ) import Outputable \end{code} %************************************************************************ %* * -\section{Dealing with a single binder} +\subsection{The continuation data type} %* * %************************************************************************ -When we hit a binder we may need to - (a) apply the the type envt (if non-empty) to its type - (b) apply the type envt and id envt to its SpecEnv (if it has one) - (c) give it a new unique to avoid name clashes +\begin{code} +data SimplCont -- Strict contexts + = Stop OutType -- Type of the result + Bool -- True => This is the RHS of a thunk whose type suggests + -- that update-in-place would be possible + -- (This makes the inliner a little keener.) + + | CoerceIt OutType -- The To-type, simplified + SimplCont + + | InlinePlease -- This continuation makes a function very + SimplCont -- keen to inline itelf + + | ApplyTo DupFlag + InExpr SubstEnv -- The argument, as yet unsimplified, + SimplCont -- and its subst-env + + | Select DupFlag + InId [InAlt] SubstEnv -- The case binder, alts, and subst-env + SimplCont + + | ArgOf DupFlag -- An arbitrary strict context: the argument + -- of a strict function, or a primitive-arg fn + -- or a PrimOp + OutType -- cont_ty: the type of the expression being sought by the context + -- f (error "foo") ==> coerce t (error "foo") + -- when f is strict + -- We need to know the type t, to which to coerce. + (OutExpr -> SimplM OutExprStuff) -- What to do with the result + -- The result expression in the OutExprStuff has type cont_ty + +instance Outputable SimplCont where + ppr (Stop _ _) = ptext SLIT("Stop") + ppr (ApplyTo dup arg se cont) = (ptext SLIT("ApplyTo") <+> ppr dup <+> ppr arg) $$ ppr cont + ppr (ArgOf dup _ _) = ptext SLIT("ArgOf...") <+> ppr dup + ppr (Select dup bndr alts se cont) = (ptext SLIT("Select") <+> ppr dup <+> ppr bndr) $$ + (nest 4 (ppr alts)) $$ ppr cont + ppr (CoerceIt ty cont) = (ptext SLIT("CoerceIt") <+> ppr ty) $$ ppr cont + ppr (InlinePlease cont) = ptext SLIT("InlinePlease") $$ ppr cont + +data DupFlag = OkToDup | NoDup + +instance Outputable DupFlag where + ppr OkToDup = ptext SLIT("ok") + ppr NoDup = ptext SLIT("nodup") + + +------------------- +mkRhsStop, mkStop :: OutType -> SimplCont +mkStop ty = Stop ty False +mkRhsStop ty = Stop ty (canUpdateInPlace ty) + + +------------------- +contIsDupable :: SimplCont -> Bool +contIsDupable (Stop _ _) = True +contIsDupable (ApplyTo OkToDup _ _ _) = True +contIsDupable (ArgOf OkToDup _ _) = True +contIsDupable (Select OkToDup _ _ _ _) = True +contIsDupable (CoerceIt _ cont) = contIsDupable cont +contIsDupable (InlinePlease cont) = contIsDupable cont +contIsDupable other = False + +------------------- +discardInline :: SimplCont -> SimplCont +discardInline (InlinePlease cont) = cont +discardInline (ApplyTo d e s cont) = ApplyTo d e s (discardInline cont) +discardInline cont = cont + +------------------- +discardableCont :: SimplCont -> Bool +discardableCont (Stop _ _) = False +discardableCont (CoerceIt _ cont) = discardableCont cont +discardableCont (InlinePlease cont) = discardableCont cont +discardableCont other = True + +discardCont :: SimplCont -- A continuation, expecting + -> SimplCont -- Replace the continuation with a suitable coerce +discardCont cont = case cont of + Stop to_ty _ -> cont + other -> CoerceIt to_ty (mkStop to_ty) + where + to_ty = contResultType cont + +------------------- +contResultType :: SimplCont -> OutType +contResultType (Stop to_ty _) = to_ty +contResultType (ArgOf _ to_ty _) = to_ty +contResultType (ApplyTo _ _ _ cont) = contResultType cont +contResultType (CoerceIt _ cont) = contResultType cont +contResultType (InlinePlease cont) = contResultType cont +contResultType (Select _ _ _ _ cont) = contResultType cont + +------------------- +countValArgs :: SimplCont -> Int +countValArgs (ApplyTo _ (Type ty) se cont) = countValArgs cont +countValArgs (ApplyTo _ val_arg se cont) = 1 + countValArgs cont +countValArgs other = 0 + +countArgs :: SimplCont -> Int +countArgs (ApplyTo _ arg se cont) = 1 + countArgs cont +countArgs other = 0 +\end{code} + + +\begin{code} +getContArgs :: OutId -> SimplCont + -> SimplM ([(InExpr, SubstEnv, Bool)], -- Arguments; the Bool is true for strict args + SimplCont, -- Remaining continuation + Bool) -- Whether we came across an InlineCall +-- getContArgs id k = (args, k', inl) +-- args are the leading ApplyTo items in k +-- (i.e. outermost comes first) +-- augmented with demand info from the functionn +getContArgs fun orig_cont + = getSwitchChecker `thenSmpl` \ chkr -> + let + -- Ignore strictness info if the no-case-of-case + -- flag is on. Strictness changes evaluation order + -- and that can change full laziness + stricts | switchIsOn chkr NoCaseOfCase = vanilla_stricts + | otherwise = computed_stricts + in + go [] stricts False orig_cont + where + ---------------------------- + + -- Type argument + go acc ss inl (ApplyTo _ arg@(Type _) se cont) + = go ((arg,se,False) : acc) ss inl cont + -- NB: don't bother to instantiate the function type + + -- Value argument + go acc (s:ss) inl (ApplyTo _ arg se cont) + = go ((arg,se,s) : acc) ss inl cont + + -- An Inline continuation + go acc ss inl (InlinePlease cont) + = go acc ss True cont + + -- We're run out of arguments, or else we've run out of demands + -- The latter only happens if the result is guaranteed bottom + -- This is the case for + -- * case (error "hello") of { ... } + -- * (error "Hello") arg + -- * f (error "Hello") where f is strict + -- etc + go acc ss inl cont + | null ss && discardableCont cont = tick BottomFound `thenSmpl_` + returnSmpl (reverse acc, discardCont cont, inl) + | otherwise = returnSmpl (reverse acc, cont, inl) + + ---------------------------- + vanilla_stricts, computed_stricts :: [Bool] + vanilla_stricts = repeat False + computed_stricts = zipWith (||) fun_stricts arg_stricts + + ---------------------------- + (val_arg_tys, _) = splitRepFunTys (idType fun) + arg_stricts = map isStrictType val_arg_tys ++ repeat False + -- These argument types are used as a cheap and cheerful way to find + -- unboxed arguments, which must be strict. But it's an InType + -- and so there might be a type variable where we expect a function + -- type (the substitution hasn't happened yet). And we don't bother + -- doing the type applications for a polymorphic function. + -- Hence the split*Rep*FunTys + + ---------------------------- + -- If fun_stricts is finite, it means the function returns bottom + -- after that number of value args have been consumed + -- Otherwise it's infinite, extended with False + fun_stricts + = case idStrictness fun of + StrictnessInfo demands result_bot + | not (demands `lengthExceeds` countValArgs orig_cont) + -> -- Enough args, use the strictness given. + -- For bottoming functions we used to pretend that the arg + -- is lazy, so that we don't treat the arg as an + -- interesting context. This avoids substituting + -- top-level bindings for (say) strings into + -- calls to error. But now we are more careful about + -- inlining lone variables, so its ok (see SimplUtils.analyseCont) + if result_bot then + map isStrict demands -- Finite => result is bottom + else + map isStrict demands ++ vanilla_stricts + + other -> vanilla_stricts -- Not enough args, or no strictness + + +------------------- +isStrictType :: Type -> Bool + -- isStrictType computes whether an argument (or let RHS) should + -- be computed strictly or lazily, based only on its type +isStrictType ty + | isUnLiftedType ty = True + | opt_DictsStrict && isDictTy ty && isDataType ty = True + | otherwise = False + -- Return true only for dictionary types where the dictionary + -- has more than one component (else we risk poking on the component + -- of a newtype dictionary) + +------------------- +interestingArg :: InScopeSet -> InExpr -> SubstEnv -> Bool + -- An argument is interesting if it has *some* structure + -- We are here trying to avoid unfolding a function that + -- is applied only to variables that have no unfolding + -- (i.e. they are probably lambda bound): f x y z + -- There is little point in inlining f here. +interestingArg in_scope arg subst + = analyse (substExpr (mkSubst in_scope subst) arg) + -- 'analyse' only looks at the top part of the result + -- and substExpr is lazy, so this isn't nearly as brutal + -- as it looks. + where + analyse (Var v) = hasSomeUnfolding (idUnfolding v) + -- Was: isValueUnfolding (idUnfolding v') + -- But that seems over-pessimistic + analyse (Type _) = False + analyse (App fn (Type _)) = analyse fn + analyse (Note _ a) = analyse a + analyse other = True + -- Consider let x = 3 in f x + -- The substitution will contain (x -> ContEx 3), and we want to + -- to say that x is an interesting argument. + -- But consider also (\x. f x y) y + -- The substitution will contain (x -> ContEx y), and we want to say + -- that x is not interesting (assuming y has no unfolding) +\end{code} + +Comment about interestingCallContext +~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ +We want to avoid inlining an expression where there can't possibly be +any gain, such as in an argument position. Hence, if the continuation +is interesting (eg. a case scrutinee, application etc.) then we +inline, otherwise we don't. + +Previously some_benefit used to return True only if the variable was +applied to some value arguments. This didn't work: + + let x = _coerce_ (T Int) Int (I# 3) in + case _coerce_ Int (T Int) x of + I# y -> .... + +we want to inline x, but can't see that it's a constructor in a case +scrutinee position, and some_benefit is False. + +Another example: + +dMonadST = _/\_ t -> :Monad (g1 _@_ t, g2 _@_ t, g3 _@_ t) + +.... case dMonadST _@_ x0 of (a,b,c) -> .... + +we'd really like to inline dMonadST here, but we *don't* want to +inline if the case expression is just + + case x of y { DEFAULT -> ... } + +since we can just eliminate this case instead (x is in WHNF). Similar +applies when x is bound to a lambda expression. Hence +contIsInteresting looks for case expressions with just a single +default case. + +\begin{code} +interestingCallContext :: Bool -- False <=> no args at all + -> Bool -- False <=> no value args + -> SimplCont -> Bool + -- The "lone-variable" case is important. I spent ages + -- messing about with unsatisfactory varaints, but this is nice. + -- The idea is that if a variable appear all alone + -- as an arg of lazy fn, or rhs Stop + -- as scrutinee of a case Select + -- as arg of a strict fn ArgOf + -- then we should not inline it (unless there is some other reason, + -- e.g. is is the sole occurrence). We achieve this by making + -- interestingCallContext return False for a lone variable. + -- + -- Why? At least in the case-scrutinee situation, turning + -- let x = (a,b) in case x of y -> ... + -- into + -- let x = (a,b) in case (a,b) of y -> ... + -- and thence to + -- let x = (a,b) in let y = (a,b) in ... + -- is bad if the binding for x will remain. + -- + -- Another example: I discovered that strings + -- were getting inlined straight back into applications of 'error' + -- because the latter is strict. + -- s = "foo" + -- f = \x -> ...(error s)... + + -- Fundamentally such contexts should not ecourage inlining becuase + -- the context can ``see'' the unfolding of the variable (e.g. case or a RULE) + -- so there's no gain. + -- + -- However, even a type application or coercion isn't a lone variable. + -- Consider + -- case $fMonadST @ RealWorld of { :DMonad a b c -> c } + -- We had better inline that sucker! The case won't see through it. + -- + -- For now, I'm treating treating a variable applied to types + -- in a *lazy* context "lone". The motivating example was + -- f = /\a. \x. BIG + -- g = /\a. \y. h (f a) + -- There's no advantage in inlining f here, and perhaps + -- a significant disadvantage. Hence some_val_args in the Stop case + +interestingCallContext some_args some_val_args cont + = interesting cont + where + interesting (InlinePlease _) = True + interesting (Select _ _ _ _ _) = some_args + interesting (ApplyTo _ _ _ _) = some_args -- Can happen if we have (coerce t (f x)) y + interesting (ArgOf _ _ _) = some_val_args + interesting (Stop ty upd_in_place) = some_val_args && upd_in_place + interesting (CoerceIt _ cont) = interesting cont + -- If this call is the arg of a strict function, the context + -- is a bit interesting. If we inline here, we may get useful + -- evaluation information to avoid repeated evals: e.g. + -- x + (y * z) + -- Here the contIsInteresting makes the '*' keener to inline, + -- which in turn exposes a constructor which makes the '+' inline. + -- Assuming that +,* aren't small enough to inline regardless. + -- + -- It's also very important to inline in a strict context for things + -- like + -- foldr k z (f x) + -- Here, the context of (f x) is strict, and if f's unfolding is + -- a build it's *great* to inline it here. So we must ensure that + -- the context for (f x) is not totally uninteresting. + + +------------------- +canUpdateInPlace :: Type -> Bool +-- Consider let x = in ... +-- If returns an explicit constructor, we might be able +-- to do update in place. So we treat even a thunk RHS context +-- as interesting if update in place is possible. We approximate +-- this by seeing if the type has a single constructor with a +-- small arity. But arity zero isn't good -- we share the single copy +-- for that case, so no point in sharing. + +-- Note the repType: we want to look through newtypes for this purpose + +canUpdateInPlace ty + | not opt_UF_UpdateInPlace = False + | otherwise + = case splitTyConApp_maybe (repType ty) of { + Nothing -> False ; + Just (tycon, _) -> + + case tyConDataConsIfAvailable tycon of + [dc] -> arity == 1 || arity == 2 + where + arity = dataConRepArity dc + other -> False + } +\end{code} + + + +%************************************************************************ +%* * +\section{Dealing with a single binder} +%* * +%************************************************************************ \begin{code} simplBinders :: [InBinder] -> ([OutBinder] -> SimplM a) -> SimplM a simplBinders bndrs thing_inside - = getSwitchChecker `thenSmpl` \ sw_chkr -> - getSimplBinderStuff `thenSmpl` \ stuff -> + = getSubst `thenSmpl` \ subst -> let - must_clone = switchIsOn sw_chkr SimplPleaseClone - (stuff', bndrs') = mapAccumL (subst_binder must_clone) stuff bndrs + (subst', bndrs') = substBndrs subst bndrs in - setSimplBinderStuff stuff' $ - thing_inside bndrs' + seqBndrs bndrs' `seq` + setSubst subst' (thing_inside bndrs') simplBinder :: InBinder -> (OutBinder -> SimplM a) -> SimplM a simplBinder bndr thing_inside - = getSwitchChecker `thenSmpl` \ sw_chkr -> - getSimplBinderStuff `thenSmpl` \ stuff -> + = getSubst `thenSmpl` \ subst -> let - must_clone = switchIsOn sw_chkr SimplPleaseClone - (stuff', bndr') = subst_binder must_clone stuff bndr + (subst', bndr') = substBndr subst bndr in - setSimplBinderStuff stuff' $ - thing_inside bndr' + seqBndr bndr' `seq` + setSubst subst' (thing_inside bndr') + -- Same semantics as simplBinders, but a little less -- plumbing and hence a little more efficient. -- Maybe not worth the candle? simplIds :: [InBinder] -> ([OutBinder] -> SimplM a) -> SimplM a simplIds ids thing_inside - = getSwitchChecker `thenSmpl` \ sw_chkr -> - getSimplBinderStuff `thenSmpl` \ (ty_subst, id_subst, in_scope, us) -> + = getSubst `thenSmpl` \ subst -> let - must_clone = switchIsOn sw_chkr SimplPleaseClone - (id_subst', in_scope', us', ids') = substIds (simpl_clone_fn must_clone) - ty_subst id_subst in_scope us ids + (subst', bndrs') = substIds subst ids in - setSimplBinderStuff (ty_subst, id_subst', in_scope', us') $ - thing_inside ids' + seqBndrs bndrs' `seq` + setSubst subst' (thing_inside bndrs') -subst_binder must_clone (ty_subst, id_subst, in_scope, us) bndr - | isTyVar bndr - = case substTyVar ty_subst in_scope bndr of - (ty_subst', in_scope', bndr') -> ((ty_subst', id_subst, in_scope', us), bndr') +seqBndrs [] = () +seqBndrs (b:bs) = seqBndr b `seq` seqBndrs bs - | otherwise - = case substId (simpl_clone_fn must_clone) ty_subst id_subst in_scope us bndr of - (id_subst', in_scope', us', bndr') - -> ((ty_subst, id_subst', in_scope', us'), bndr') - -simpl_clone_fn must_clone in_scope us id - | (must_clone && isLocalName (idName id)) - || id `elemVarSet` in_scope - = case splitUniqSupply us of - (us1, us2) -> Just (us1, setVarUnique id (uniqFromSupply us2)) - - | otherwise - = Nothing +seqBndr b | isTyVar b = b `seq` () + | otherwise = seqType (idType b) `seq` + idInfo b `seq` + () +\end{code} + + +%************************************************************************ +%* * +\subsection{Transform a RHS} +%* * +%************************************************************************ + +Try (a) eta expansion + (b) type-lambda swizzling + +\begin{code} +transformRhs :: OutExpr + -> (ArityInfo -> OutExpr -> SimplM (OutStuff a)) + -> SimplM (OutStuff a) + +transformRhs rhs thing_inside + = tryRhsTyLam rhs $ \ rhs1 -> + tryEtaExpansion rhs1 thing_inside \end{code} @@ -140,7 +511,7 @@ let-floating. This optimisation is CRUCIAL in eliminating the junk introduced by desugaring mutually recursive definitions. Don't eliminate it lightly! -So far as the implemtation is concerned: +So far as the implementation is concerned: Invariant: go F e = /\tvs -> F e @@ -159,29 +530,58 @@ So far as the implemtation is concerned: where G = F . Let {xi = xi' tvs} -\begin{code} -mkRhsTyLam (Lam b e) - | isTyVar b = case collectTyBinders e of - (bs,body) -> mkRhsTyLam_help (b:bs) body +[May 1999] If we do this transformation *regardless* then we can +end up with some pretty silly stuff. For example, + + let + st = /\ s -> let { x1=r1 ; x2=r2 } in ... + in .. +becomes + let y1 = /\s -> r1 + y2 = /\s -> r2 + st = /\s -> ...[y1 s/x1, y2 s/x2] + in .. + +Unless the "..." is a WHNF there is really no point in doing this. +Indeed it can make things worse. Suppose x1 is used strictly, +and is of the form + + x1* = case f y of { (a,b) -> e } + +If we abstract this wrt the tyvar we then can't do the case inline +as we would normally do. -mkRhsTyLam other_expr -- No-op if not a type lambda - = returnSmpl other_expr +\begin{code} +tryRhsTyLam rhs thing_inside -- Only does something if there's a let + | null tyvars || not (worth_it body) -- inside a type lambda, and a WHNF inside that + = thing_inside rhs + | otherwise + = go (\x -> x) body $ \ body' -> + thing_inside (mkLams tyvars body') -mkRhsTyLam_help tyvars body - = go (\x -> x) body where - main_tyvar_set = mkVarSet tyvars + (tyvars, body) = collectTyBinders rhs + + worth_it (Let _ e) = whnf_in_middle e + worth_it other = False + whnf_in_middle (Let _ e) = whnf_in_middle e + whnf_in_middle e = exprIsCheap e + + + go fn (Let bind@(NonRec var rhs) body) thing_inside + | exprIsTrivial rhs + = go (fn . Let bind) body thing_inside - go fn (Let bind@(NonRec var rhs) body) | exprIsTrivial rhs - = go (fn . Let bind) body + go fn (Let bind@(NonRec var rhs) body) thing_inside + = mk_poly tyvars_here var `thenSmpl` \ (var', rhs') -> + addAuxiliaryBind (NonRec var' (mkLams tyvars_here (fn rhs))) $ + go (fn . Let (mk_silly_bind var rhs')) body thing_inside - go fn (Let bind@(NonRec var rhs) body) - = mk_poly tyvars_here var `thenSmpl` \ (var', rhs') -> - go (fn . Let (mk_silly_bind var rhs')) body `thenSmpl` \ body' -> - returnSmpl (Let (NonRec var' (mkLams tyvars_here (fn rhs))) body') where tyvars_here = tyvars + -- main_tyvar_set = mkVarSet tyvars + -- var_ty = idType var -- varSetElems (main_tyvar_set `intersectVarSet` tyVarsOfType var_ty) -- tyvars_here was an attempt to reduce the number of tyvars -- wrt which the new binding is abstracted. But the naive @@ -190,7 +590,7 @@ mkRhsTyLam_help tyvars body -- /\ a b -> let t :: (a,b) = (e1, e2) -- x :: a = fst t -- in ... - -- Here, b isn't free in a's type, but we must nevertheless + -- Here, b isn't free in x's type, but we must nevertheless -- abstract wrt b as well, because t's type mentions b. -- Since t is floated too, we'd end up with the bogus: -- poly_t = /\ a b -> (e1, e2) @@ -199,181 +599,176 @@ mkRhsTyLam_help tyvars body -- abstracting wrt *all* the tyvars. We'll see if that -- gives rise to problems. SLPJ June 98 - var_ty = idType var - - go fn (Let (Rec prs) body) + go fn (Let (Rec prs) body) thing_inside = mapAndUnzipSmpl (mk_poly tyvars_here) vars `thenSmpl` \ (vars', rhss') -> let - gn body = fn $ foldr Let body (zipWith mk_silly_bind vars rhss') + gn body = fn (foldr Let body (zipWith mk_silly_bind vars rhss')) in - go gn body `thenSmpl` \ body' -> - returnSmpl (Let (Rec (vars' `zip` [mkLams tyvars_here (gn rhs) | rhs <- rhss])) body') + addAuxiliaryBind (Rec (vars' `zip` [mkLams tyvars_here (gn rhs) | rhs <- rhss])) $ + go gn body thing_inside where (vars,rhss) = unzip prs tyvars_here = tyvars -- varSetElems (main_tyvar_set `intersectVarSet` tyVarsOfTypes var_tys) + -- var_tys = map idType vars -- See notes with tyvars_here above - var_tys = map idType vars - go fn body = returnSmpl (mkLams tyvars (fn body)) + go fn body thing_inside = thing_inside (fn body) mk_poly tyvars_here var - = newId (mkForAllTys tyvars_here (idType var)) $ \ poly_id -> + = getUniqueSmpl `thenSmpl` \ uniq -> let - -- It's crucial to copy the inline-prag of the original var, because - -- we're looking at occurrence-analysed but as yet unsimplified code! - -- In particular, we mustn't lose the loop breakers. + poly_name = setNameUnique (idName var) uniq -- Keep same name + poly_ty = mkForAllTys tyvars_here (idType var) -- But new type of course + poly_id = mkVanillaId poly_name poly_ty + + -- In the olden days, it was crucial to copy the occInfo of the original var, + -- because we were looking at occurrence-analysed but as yet unsimplified code! + -- In particular, we mustn't lose the loop breakers. BUT NOW we are looking + -- at already simplified code, so it doesn't matter -- - -- *However* we don't want to retain a single-occurrence or dead-var info - -- because we're adding a load of "silly bindings" of the form - -- var _U_ = poly_var t1 t2 - -- with a must-inline pragma on the silly binding to prevent the - -- poly-var from being inlined right back in. Since poly_var now - -- occurs inside an INLINE binding, it should be given a ManyOcc, - -- else it may get inlined unconditionally - poly_inline_prag = case getInlinePragma var of - ICanSafelyBeINLINEd _ _ -> NoInlinePragInfo - IAmDead -> NoInlinePragInfo - var_inline_prag -> var_inline_prag - - poly_id' = setInlinePragma poly_id poly_inline_prag + -- It's even right to retain single-occurrence or dead-var info: + -- Suppose we started with /\a -> let x = E in B + -- where x occurs once in B. Then we transform to: + -- let x' = /\a -> E in /\a -> let x* = x' a in B + -- where x* has an INLINE prag on it. Now, once x* is inlined, + -- the occurrences of x' will be just the occurrences originally + -- pinned on x. in - returnSmpl (poly_id', mkTyApps (Var poly_id') (mkTyVarTys tyvars_here)) + returnSmpl (poly_id, mkTyApps (Var poly_id) (mkTyVarTys tyvars_here)) - mk_silly_bind var rhs = NonRec (setInlinePragma var IWantToBeINLINEd) rhs - -- The addInlinePragma is really important! If we don't say - -- INLINE on these silly little bindings then look what happens! + mk_silly_bind var rhs = NonRec var (Note InlineMe rhs) -- Suppose we start with: -- - -- x = let g = /\a -> \x -> f x x - -- in - -- /\ b -> let g* = g b in E + -- x = /\ a -> let g = G in E + -- + -- Then we'll float to get + -- + -- x = let poly_g = /\ a -> G + -- in /\ a -> let g = poly_g a in E -- - -- Then: * the binding for g gets floated out - -- * but then it gets inlined into the rhs of g* - -- * then the binding for g* is floated out of the /\b - -- * so we're back to square one - -- The silly binding for g* must be INLINE, so that no inlining - -- will happen in its RHS. - -- PS: Jun 98: actually this isn't important any more; - -- inlineUnconditionally will catch the type applicn - -- and inline it unconditionally, without ever trying - -- to simplify the RHS + -- But now the occurrence analyser will see just one occurrence + -- of poly_g, not inside a lambda, so the simplifier will + -- PreInlineUnconditionally poly_g back into g! Badk to square 1! + -- (I used to think that the "don't inline lone occurrences" stuff + -- would stop this happening, but since it's the *only* occurrence, + -- PreInlineUnconditionally kicks in first!) + -- + -- Solution: put an INLINE note on g's RHS, so that poly_g seems + -- to appear many times. (NB: mkInlineMe eliminates + -- such notes on trivial RHSs, so do it manually.) \end{code} %************************************************************************ %* * -\subsection{Eta reduction} +\subsection{Eta expansion} %* * %************************************************************************ -@etaCoreExpr@ trys an eta reduction at the top level of a Core Expr. + Try eta expansion for RHSs -e.g. \ x y -> f x y ===> f +We go for: + Case 1 f = \x1..xn -> N ==> f = \x1..xn y1..ym -> N y1..ym + (n >= 0) + OR + Case 2 f = N E1..En ==> z1=E1 + (n > 0) .. + zn=En + f = \y1..ym -> N z1..zn y1..ym -It is used - a) Before constructing an Unfolding, to - try to make the unfolding smaller; - b) In tidyCoreExpr, which is done just before converting to STG. +where (in both cases) -But we only do this if - i) 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. + * The xi can include type variables - ii) It exposes a simple variable or a type application; in short - it exposes a "trivial" expression. (exprIsTrivial) + * The yi are all value variables -\begin{code} -etaCoreExpr :: CoreExpr -> CoreExpr - -- ToDo: we should really check that we don't turn a non-bottom - -- lambda into a bottom variable. Sigh - -etaCoreExpr expr@(Lam bndr body) - | opt_DoEtaReduction - = check (reverse binders) body - where - (binders, body) = collectBinders expr + * N is a NORMAL FORM (i.e. no redexes anywhere) + wanting a suitable number of extra args. - check [] body - | exprIsTrivial body && not (any (`elemVarSet` body_fvs) binders) - = body -- Success! - where - body_fvs = exprFreeVars body + * the Ei must not have unlifted type - check (b : bs) (App fun arg) - | (varToCoreExpr b `cheapEqExpr` arg) - = check bs fun +There is no point in looking for a combination of the two, because +that would leave use with some lets sandwiched between lambdas; that's +what the final test in the first equation is for. - check _ _ = expr -- Bale out +\begin{code} +tryEtaExpansion :: OutExpr + -> (ArityInfo -> OutExpr -> SimplM (OutStuff a)) + -> SimplM (OutStuff a) +tryEtaExpansion rhs thing_inside + | not opt_SimplDoLambdaEtaExpansion + || null y_tys -- No useful expansion + || not (is_case1 || is_case2) -- Neither case matches + = thing_inside final_arity rhs -- So, no eta expansion, but + -- return a good arity + + | is_case1 + = make_y_bndrs $ \ y_bndrs -> + thing_inside final_arity + (mkLams x_bndrs $ mkLams y_bndrs $ + mkApps body (map Var y_bndrs)) + + | otherwise -- Must be case 2 + = mapAndUnzipSmpl bind_z_arg arg_infos `thenSmpl` \ (maybe_z_binds, z_args) -> + addAuxiliaryBinds (catMaybes maybe_z_binds) $ + make_y_bndrs $ \ y_bndrs -> + thing_inside final_arity + (mkLams y_bndrs $ + mkApps (mkApps fun z_args) (map Var y_bndrs)) + where + all_trivial_args = all is_trivial arg_infos + is_case1 = all_trivial_args + is_case2 = null x_bndrs && not (any unlifted_non_trivial arg_infos) -etaCoreExpr expr = expr -- The common case -\end{code} - + (x_bndrs, body) = collectBinders rhs -- NB: x_bndrs can include type variables + x_arity = valBndrCount x_bndrs -%************************************************************************ -%* * -\subsection{Eta expansion} -%* * -%************************************************************************ + (fun, args) = collectArgs body + arg_infos = [(arg, exprType arg, exprIsTrivial arg) | arg <- args] -@etaExpandCount@ takes an expression, E, and returns an integer n, -such that + is_trivial (_, _, triv) = triv + unlifted_non_trivial (_, ty, triv) = not triv && isUnLiftedType ty - E ===> (\x1::t1 x1::t2 ... xn::tn -> E x1 x2 ... xn) + fun_arity = exprEtaExpandArity fun -is a safe transformation. In particular, the transformation should -not cause work to be duplicated, unless it is ``cheap'' (see -@manifestlyCheap@ below). + final_arity | all_trivial_args = atLeastArity (x_arity + extra_args_wanted) + | otherwise = atLeastArity x_arity + -- Arity can be more than the number of lambdas + -- because of coerces. E.g. \x -> coerce t (\y -> e) + -- will have arity at least 2 + -- The worker/wrapper pass will bring the coerce out to the top -@etaExpandCount@ errs on the conservative side. It is always safe to -return 0. + bind_z_arg (arg, arg_ty, trivial_arg) + | trivial_arg = returnSmpl (Nothing, arg) + | otherwise = newId SLIT("z") arg_ty $ \ z -> + returnSmpl (Just (NonRec z arg), Var z) -An application of @error@ is special, because it can absorb as many -arguments as you care to give it. For this special case we return -100, to represent "infinity", which is a bit of a hack. + make_y_bndrs thing_inside + = ASSERT( not (exprIsTrivial rhs) ) + newIds SLIT("y") y_tys $ \ y_bndrs -> + tick (EtaExpansion (head y_bndrs)) `thenSmpl_` + thing_inside y_bndrs -\begin{code} -etaExpandCount :: CoreExpr - -> Int -- Number of extra args you can safely abstract - -etaExpandCount (Lam b body) - | isId b - = 1 + etaExpandCount body - -etaExpandCount (Let bind body) - | all exprIsCheap (rhssOfBind bind) - = etaExpandCount body - -etaExpandCount (Case scrut _ alts) - | exprIsCheap scrut - = minimum [etaExpandCount rhs | (_,_,rhs) <- alts] - -etaExpandCount fun@(Var _) = eta_fun fun - -etaExpandCount (App fun (Type ty)) - = eta_fun fun -etaExpandCount (App fun arg) - | exprIsCheap arg = case etaExpandCount fun of - 0 -> 0 - n -> n-1 -- Knock off one - -etaExpandCount other = 0 -- Give up - -- Lit, Con, Prim, - -- non-val Lam, - -- Scc (pessimistic; ToDo), - -- Let with non-whnf rhs(s), - -- Case with non-whnf scrutinee - ------------------------------ -eta_fun :: CoreExpr -- The function - -> Int -- How many args it can safely be applied to - -eta_fun (App fun (Type ty)) = eta_fun fun -eta_fun (Var v) = arityLowerBound (getIdArity v) -eta_fun other = 0 -- Give up + (potential_extra_arg_tys, _) = splitFunTys (exprType body) + + y_tys :: [InType] + y_tys = take extra_args_wanted potential_extra_arg_tys + + extra_args_wanted :: Int -- Number of extra args we want + extra_args_wanted = 0 `max` (fun_arity - valArgCount args) + + -- We used to expand the arity to the previous arity fo the + -- function; but this is pretty dangerous. Consdier + -- f = \xy -> e + -- so that f has arity 2. Now float something into f's RHS: + -- f = let z = BIG in \xy -> e + -- The last thing we want to do now is to put some lambdas + -- outside, to get + -- f = \xy -> let z = BIG in e + -- + -- (bndr_arity - no_of_xs) `max` \end{code} @@ -384,7 +779,7 @@ eta_fun other = 0 -- Give up %************************************************************************ \begin{code} -mkCase :: SwitchChecker -> OutExpr -> OutId -> [OutAlt] -> SimplM OutExpr +mkCase :: OutExpr -> OutId -> [OutAlt] -> SimplM OutExpr \end{code} @mkCase@ tries the following transformation (if possible): @@ -407,11 +802,11 @@ transformation is called Case Merging. It avoids that the same variable is scrutinised multiple times. \begin{code} -mkCase sw_chkr scrut outer_bndr outer_alts - | switchIsOn sw_chkr SimplCaseMerge +mkCase scrut outer_bndr outer_alts + | opt_SimplCaseMerge && maybeToBool maybe_case_in_default - = tick CaseMerge `thenSmpl_` + = tick (CaseMerge outer_bndr) `thenSmpl_` returnSmpl (Case scrut outer_bndr new_alts) -- Warning: don't call mkCase recursively! -- Firstly, there's no point, because inner alts have already had @@ -449,18 +844,15 @@ Now the identity-case transformation: and similar friends. \begin{code} -mkCase sw_chkr scrut case_bndr alts +mkCase scrut case_bndr alts | all identity_alt alts - = tick CaseIdentity `thenSmpl_` + = tick (CaseIdentity case_bndr) `thenSmpl_` returnSmpl scrut where - identity_alt (DEFAULT, [], Var v) = v == case_bndr - identity_alt (con, args, Con con' args') = con == con' && - and (zipWithEqual "mkCase" - cheapEqExpr - (map Type arg_tys ++ map varToCoreExpr args) - args') - identity_alt other = False + identity_alt (DEFAULT, [], Var v) = v == case_bndr + identity_alt (DataAlt con, args, rhs) = cheapEqExpr rhs + (mkConApp con (map Type arg_tys ++ map varToCoreExpr args)) + identity_alt other = False arg_tys = case splitTyConApp_maybe (idType case_bndr) of Just (tycon, arg_tys) -> arg_tys @@ -469,7 +861,7 @@ mkCase sw_chkr scrut case_bndr alts The catch-all case \begin{code} -mkCase sw_chkr other_scrut case_bndr other_alts +mkCase other_scrut case_bndr other_alts = returnSmpl (Case other_scrut case_bndr other_alts) \end{code} @@ -482,7 +874,7 @@ findDefault ((DEFAULT,args,rhs) : alts) = ASSERT( null alts && null args ) findDefault (alt : alts) = case findDefault alts of (alts', deflt) -> (alt : alts', deflt) -findAlt :: Con -> [CoreAlt] -> CoreAlt +findAlt :: AltCon -> [CoreAlt] -> CoreAlt findAlt con alts = go alts where