X-Git-Url: http://git.megacz.com/?a=blobdiff_plain;f=ghc%2Fcompiler%2FsimplCore%2FSimplUtils.lhs;h=9e616b5df199dc04f9d257d4f9b905698c66b097;hb=2231d6544ba75e0fe41061339a23e1c90a0409bd;hp=5f8c77f47ca1bf7e659f29ad8edbd5f34b7ff640;hpb=372893551628ae83ebf9a40ff2e8ed39012a9271;p=ghc-hetmet.git diff --git a/ghc/compiler/simplCore/SimplUtils.lhs b/ghc/compiler/simplCore/SimplUtils.lhs index 5f8c77f..9e616b5 100644 --- a/ghc/compiler/simplCore/SimplUtils.lhs +++ b/ghc/compiler/simplCore/SimplUtils.lhs @@ -5,43 +5,53 @@ \begin{code} module SimplUtils ( - simplBinder, simplBinders, simplIds, - tryRhsTyLam, tryEtaExpansion, - mkCase, findAlt, findDefault, + mkLam, prepareAlts, mkCase, + + -- Inlining, + preInlineUnconditionally, postInlineUnconditionally, activeInline, activeRule, + inlineMode, -- The continuation type - SimplCont(..), DupFlag(..), contIsDupable, contResultType, - countValArgs, countArgs, mkRhsStop, mkStop, - getContArgs, interestingCallContext, interestingArg, isStrictType, discardInline + SimplCont(..), DupFlag(..), LetRhsFlag(..), + contIsDupable, contResultType, + countValArgs, countArgs, pushContArgs, + mkBoringStop, mkRhsStop, contIsRhs, contIsRhsOrArg, + getContArgs, interestingCallContext, interestingArg, isStrictType ) where #include "HsVersions.h" -import CmdLineOpts ( switchIsOn, SimplifierSwitch(..), - opt_SimplDoLambdaEtaExpansion, opt_SimplCaseMerge, opt_DictsStrict, - opt_UF_UpdateInPlace - ) +import SimplEnv +import DynFlags ( SimplifierSwitch(..), SimplifierMode(..), + DynFlag(..), dopt ) +import StaticFlags ( opt_UF_UpdateInPlace, opt_SimplNoPreInlining, + opt_RulesOff ) import CoreSyn -import CoreUtils ( exprIsTrivial, cheapEqExpr, exprType, exprIsCheap, etaExpand, exprEtaExpandArity, bindNonRec ) -import Subst ( InScopeSet, mkSubst, substBndrs, substBndr, substIds, substExpr ) -import Id ( idType, idName, - idUnfolding, idStrictness, - mkVanillaId, idInfo +import CoreFVs ( exprFreeVars ) +import CoreUtils ( cheapEqExpr, exprType, exprIsTrivial, exprIsCheap, + etaExpand, exprEtaExpandArity, bindNonRec, mkCoerce2, + findDefault, exprOkForSpeculation, exprIsHNF + ) +import Literal ( mkStringLit ) +import CoreUnfold ( smallEnoughToInline ) +import MkId ( eRROR_ID ) +import Id ( idType, isDataConWorkId, idOccInfo, isDictId, + mkSysLocal, isDeadBinder, idNewDemandInfo, isExportedId, + idUnfolding, idNewStrictness, idInlinePragma, ) -import IdInfo ( StrictnessInfo(..) ) -import Maybes ( maybeToBool, catMaybes ) -import Name ( setNameUnique ) -import Demand ( isStrict ) +import NewDemand ( isStrictDmd, isBotRes, splitStrictSig ) import SimplMonad -import Type ( Type, mkForAllTys, seqType, repType, - splitTyConApp_maybe, tyConAppArgs, mkTyVarTys, - isDictTy, isDataType, isUnLiftedType, - splitRepFunTys +import Type ( Type, splitFunTys, dropForAlls, isStrictType, + splitTyConApp_maybe, tyConAppArgs, mkTyVarTys ) -import TyCon ( tyConDataConsIfAvailable ) -import DataCon ( dataConRepArity ) -import VarEnv ( SubstEnv ) +import Name ( mkSysTvName ) +import TyCon ( tyConDataCons_maybe, isAlgTyCon, isNewTyCon ) +import DataCon ( dataConRepArity, dataConTyVars, dataConInstArgTys, isVanillaDataCon ) +import Var ( tyVarKind, mkTyVar ) +import VarSet +import BasicTypes ( TopLevelFlag(..), isNotTopLevel, OccInfo(..), isLoopBreaker, isOneOcc, + Activation, isAlwaysActive, isActive ) import Util ( lengthExceeds ) import Outputable \end{code} @@ -56,9 +66,10 @@ import Outputable \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.) + LetRhsFlag + 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 @@ -67,27 +78,37 @@ data SimplCont -- Strict contexts SimplCont -- keen to inline itelf | ApplyTo DupFlag - InExpr SubstEnv -- The argument, as yet unsimplified, - SimplCont -- and its subst-env + InExpr SimplEnv -- The argument, as yet unsimplified, + SimplCont -- and its environment | Select DupFlag - InId [InAlt] SubstEnv -- The case binder, alts, and subst-env + InId [InAlt] SimplEnv -- The case binder, alts, and subst-env SimplCont - | ArgOf DupFlag -- An arbitrary strict context: the argument + | ArgOf LetRhsFlag -- An arbitrary strict context: the argument -- of a strict function, or a primitive-arg fn -- or a PrimOp + -- No DupFlag because we never duplicate it + OutType -- arg_ty: type of the argument itself 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 + + (SimplEnv -> OutExpr -> SimplM FloatsWithExpr) -- What to do with the result -- The result expression in the OutExprStuff has type cont_ty +data LetRhsFlag = AnArg -- It's just an argument not a let RHS + | AnRhs -- It's the RHS of a let (so please float lets out of big lambdas) + +instance Outputable LetRhsFlag where + ppr AnArg = ptext SLIT("arg") + ppr AnRhs = ptext SLIT("rhs") + instance Outputable SimplCont where - ppr (Stop _ _) = ptext SLIT("Stop") + ppr (Stop ty is_rhs _) = ptext SLIT("Stop") <> brackets (ppr is_rhs) <+> ppr ty ppr (ApplyTo dup arg se cont) = (ptext SLIT("ApplyTo") <+> ppr dup <+> ppr arg) $$ ppr cont - ppr (ArgOf dup _ _) = ptext SLIT("ArgOf...") <+> ppr dup + ppr (ArgOf _ _ _ _) = ptext SLIT("ArgOf...") ppr (Select dup bndr alts se cont) = (ptext SLIT("Select") <+> ppr dup <+> ppr bndr) $$ (nest 4 (ppr alts)) $$ ppr cont ppr (CoerceIt ty cont) = (ptext SLIT("CoerceIt") <+> ppr ty) $$ ppr cont @@ -101,30 +122,31 @@ instance Outputable DupFlag where ------------------- -mkRhsStop, mkStop :: OutType -> SimplCont -mkStop ty = Stop ty False -mkRhsStop ty = Stop ty (canUpdateInPlace ty) +mkBoringStop, mkRhsStop :: OutType -> SimplCont +mkBoringStop ty = Stop ty AnArg (canUpdateInPlace ty) +mkRhsStop ty = Stop ty AnRhs (canUpdateInPlace ty) +contIsRhs :: SimplCont -> Bool +contIsRhs (Stop _ AnRhs _) = True +contIsRhs (ArgOf AnRhs _ _ _) = True +contIsRhs other = False + +contIsRhsOrArg (Stop _ _ _) = True +contIsRhsOrArg (ArgOf _ _ _ _) = True +contIsRhsOrArg other = False ------------------- contIsDupable :: SimplCont -> Bool -contIsDupable (Stop _ _) = True +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 +contIsDupable (InlinePlease cont) = contIsDupable cont +contIsDupable other = False ------------------- discardableCont :: SimplCont -> Bool -discardableCont (Stop _ _) = False +discardableCont (Stop _ _ _) = False discardableCont (CoerceIt _ cont) = discardableCont cont discardableCont (InlinePlease cont) = discardableCont cont discardableCont other = True @@ -132,15 +154,15 @@ 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) + Stop to_ty is_rhs _ -> cont + other -> CoerceIt to_ty (mkBoringStop to_ty) where to_ty = contResultType cont ------------------- contResultType :: SimplCont -> OutType -contResultType (Stop to_ty _) = to_ty -contResultType (ArgOf _ to_ty _) = to_ty +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 @@ -155,21 +177,27 @@ countValArgs other = 0 countArgs :: SimplCont -> Int countArgs (ApplyTo _ arg se cont) = 1 + countArgs cont countArgs other = 0 + +------------------- +pushContArgs :: SimplEnv -> [OutArg] -> SimplCont -> SimplCont +-- Pushes args with the specified environment +pushContArgs env [] cont = cont +pushContArgs env (arg : args) cont = ApplyTo NoDup arg env (pushContArgs env args cont) \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 :: SwitchChecker + -> OutId -> SimplCont + -> ([(InExpr, SimplEnv, 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 +getContArgs chkr fun orig_cont + = let -- Ignore strictness info if the no-case-of-case -- flag is on. Strictness changes evaluation order -- and that can change full laziness @@ -200,10 +228,12 @@ getContArgs fun orig_cont -- * (error "Hello") arg -- * f (error "Hello") where f is strict -- etc + -- Then, especially in the first of these cases, we'd like to discard + -- the continuation, leaving just the bottoming expression. But the + -- type might not be right, so we may have to add a coerce. go acc ss inl cont - | null ss && discardableCont cont = tick BottomFound `thenSmpl_` - returnSmpl (reverse acc, discardCont cont, inl) - | otherwise = returnSmpl (reverse acc, cont, inl) + | null ss && discardableCont cont = (reverse acc, discardCont cont, inl) + | otherwise = (reverse acc, cont, inl) ---------------------------- vanilla_stricts, computed_stricts :: [Bool] @@ -211,22 +241,22 @@ getContArgs fun orig_cont computed_stricts = zipWith (||) fun_stricts arg_stricts ---------------------------- - (val_arg_tys, _) = splitRepFunTys (idType fun) + (val_arg_tys, _) = splitFunTys (dropForAlls (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 + -- Hence the splitFunTys*IgnoringForAlls* ---------------------------- -- 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 + = case splitStrictSig (idNewStrictness fun) of + (demands, result_info) | not (demands `lengthExceeds` countValArgs orig_cont) -> -- Enough args, use the strictness given. -- For bottoming functions we used to pretend that the arg @@ -235,46 +265,30 @@ getContArgs fun orig_cont -- 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 + if isBotRes result_info then + map isStrictDmd demands -- Finite => result is bottom else - map isStrict demands ++ vanilla_stricts + map isStrictDmd 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 +interestingArg :: OutExpr -> 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 +interestingArg (Var v) = hasSomeUnfolding (idUnfolding v) + -- Was: isValueUnfolding (idUnfolding v') + -- But that seems over-pessimistic + || isDataConWorkId v + -- This accounts for an argument like + -- () or [], which is definitely interesting +interestingArg (Type _) = False +interestingArg (App fn (Type _)) = interestingArg fn +interestingArg (Note _ a) = interestingArg a +interestingArg 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. @@ -344,7 +358,7 @@ interestingCallContext :: Bool -- False <=> no args at all -- s = "foo" -- f = \x -> ...(error s)... - -- Fundamentally such contexts should not ecourage inlining becuase + -- Fundamentally such contexts should not ecourage inlining because -- the context can ``see'' the unfolding of the variable (e.g. case or a RULE) -- so there's no gain. -- @@ -363,12 +377,15 @@ interestingCallContext :: Bool -- False <=> no args at all 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 + interesting (InlinePlease _) = True + interesting (Select _ _ _ _ _) = some_args + interesting (ApplyTo _ _ _ _) = True -- Can happen if we have (coerce t (f x)) y + -- Perhaps True is a bit over-keen, but I've + -- seen (coerce f) x, where f has an INLINE prag, + -- So we have to give some motivaiton for inlining it + 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. @@ -395,80 +412,490 @@ canUpdateInPlace :: Type -> Bool -- 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 - } + = case splitTyConApp_maybe ty of + Nothing -> False + Just (tycon, _) -> case tyConDataCons_maybe tycon of + Just [dc] -> arity == 1 || arity == 2 + where + arity = dataConRepArity dc + other -> False \end{code} %************************************************************************ %* * -\section{Dealing with a single binder} +\subsection{Decisions about inlining} %* * %************************************************************************ +Inlining is controlled partly by the SimplifierMode switch. This has two +settings: + + SimplGently (a) Simplifying before specialiser/full laziness + (b) Simplifiying inside INLINE pragma + (c) Simplifying the LHS of a rule + (d) Simplifying a GHCi expression or Template + Haskell splice + + SimplPhase n Used at all other times + +The key thing about SimplGently is that it does no call-site inlining. +Before full laziness we must be careful not to inline wrappers, +because doing so inhibits floating + e.g. ...(case f x of ...)... + ==> ...(case (case x of I# x# -> fw x#) of ...)... + ==> ...(case x of I# x# -> case fw x# of ...)... +and now the redex (f x) isn't floatable any more. + +The no-inling thing is also important for Template Haskell. You might be +compiling in one-shot mode with -O2; but when TH compiles a splice before +running it, we don't want to use -O2. Indeed, we don't want to inline +anything, because the byte-code interpreter might get confused about +unboxed tuples and suchlike. + +INLINE pragmas +~~~~~~~~~~~~~~ +SimplGently is also used as the mode to simplify inside an InlineMe note. + \begin{code} -simplBinders :: [InBinder] -> ([OutBinder] -> SimplM a) -> SimplM a -simplBinders bndrs thing_inside - = getSubst `thenSmpl` \ subst -> - let - (subst', bndrs') = substBndrs subst bndrs - in - seqBndrs bndrs' `seq` - setSubst subst' (thing_inside bndrs') - -simplBinder :: InBinder -> (OutBinder -> SimplM a) -> SimplM a -simplBinder bndr thing_inside - = getSubst `thenSmpl` \ subst -> - let - (subst', bndr') = substBndr subst bndr - in - 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 - = getSubst `thenSmpl` \ subst -> - let - (subst', bndrs') = substIds subst ids - in - seqBndrs bndrs' `seq` - setSubst subst' (thing_inside bndrs') +inlineMode :: SimplifierMode +inlineMode = SimplGently +\end{code} -seqBndrs [] = () -seqBndrs (b:bs) = seqBndr b `seq` seqBndrs bs +It really is important to switch off inlinings inside such +expressions. Consider the following example + + let f = \pq -> BIG + in + let g = \y -> f y y + {-# INLINE g #-} + in ...g...g...g...g...g... + +Now, if that's the ONLY occurrence of f, it will be inlined inside g, +and thence copied multiple times when g is inlined. + + +This function may be inlinined in other modules, so we +don't want to remove (by inlining) calls to functions that have +specialisations, or that may have transformation rules in an importing +scope. + +E.g. {-# INLINE f #-} + f x = ...g... + +and suppose that g is strict *and* has specialisations. If we inline +g's wrapper, we deny f the chance of getting the specialised version +of g when f is inlined at some call site (perhaps in some other +module). + +It's also important not to inline a worker back into a wrapper. +A wrapper looks like + wraper = inline_me (\x -> ...worker... ) +Normally, the inline_me prevents the worker getting inlined into +the wrapper (initially, the worker's only call site!). But, +if the wrapper is sure to be called, the strictness analyser will +mark it 'demanded', so when the RHS is simplified, it'll get an ArgOf +continuation. That's why the keep_inline predicate returns True for +ArgOf continuations. It shouldn't do any harm not to dissolve the +inline-me note under these circumstances. + +Note that the result is that we do very little simplification +inside an InlineMe. + + all xs = foldr (&&) True xs + any p = all . map p {-# INLINE any #-} + +Problem: any won't get deforested, and so if it's exported and the +importer doesn't use the inlining, (eg passes it as an arg) then we +won't get deforestation at all. We havn't solved this problem yet! + + +preInlineUnconditionally +~~~~~~~~~~~~~~~~~~~~~~~~ +@preInlineUnconditionally@ examines a bndr to see if it is used just +once in a completely safe way, so that it is safe to discard the +binding inline its RHS at the (unique) usage site, REGARDLESS of how +big the RHS might be. If this is the case we don't simplify the RHS +first, but just inline it un-simplified. + +This is much better than first simplifying a perhaps-huge RHS and then +inlining and re-simplifying it. Indeed, it can be at least quadratically +better. Consider + + x1 = e1 + x2 = e2[x1] + x3 = e3[x2] + ...etc... + xN = eN[xN-1] + +We may end up simplifying e1 N times, e2 N-1 times, e3 N-3 times etc. +This can happen with cascades of functions too: + + f1 = \x1.e1 + f2 = \xs.e2[f1] + f3 = \xs.e3[f3] + ...etc... + +THE MAIN INVARIANT is this: + + ---- preInlineUnconditionally invariant ----- + IF preInlineUnconditionally chooses to inline x = + THEN doing the inlining should not change the occurrence + info for the free vars of + ---------------------------------------------- + +For example, it's tempting to look at trivial binding like + x = y +and inline it unconditionally. But suppose x is used many times, +but this is the unique occurrence of y. Then inlining x would change +y's occurrence info, which breaks the invariant. It matters: y +might have a BIG rhs, which will now be dup'd at every occurrenc of x. + + +Evne RHSs labelled InlineMe aren't caught here, because there might be +no benefit from inlining at the call site. + +[Sept 01] Don't unconditionally inline a top-level thing, because that +can simply make a static thing into something built dynamically. E.g. + x = (a,b) + main = \s -> h x + +[Remember that we treat \s as a one-shot lambda.] No point in +inlining x unless there is something interesting about the call site. + +But watch out: if you aren't careful, some useful foldr/build fusion +can be lost (most notably in spectral/hartel/parstof) because the +foldr didn't see the build. Doing the dynamic allocation isn't a big +deal, in fact, but losing the fusion can be. But the right thing here +seems to be to do a callSiteInline based on the fact that there is +something interesting about the call site (it's strict). Hmm. That +seems a bit fragile. + +Conclusion: inline top level things gaily until Phase 0 (the last +phase), at which point don't. -seqBndr b | isTyVar b = b `seq` () - | otherwise = seqType (idType b) `seq` - idInfo b `seq` - () +\begin{code} +preInlineUnconditionally :: SimplEnv -> TopLevelFlag -> InId -> InExpr -> Bool +preInlineUnconditionally env top_lvl bndr rhs + | not active = False + | opt_SimplNoPreInlining = False + | otherwise = case idOccInfo bndr of + IAmDead -> True -- Happens in ((\x.1) v) + OneOcc in_lam True int_cxt -> try_once in_lam int_cxt + other -> False + where + phase = getMode env + active = case phase of + SimplGently -> isAlwaysActive prag + SimplPhase n -> isActive n prag + prag = idInlinePragma bndr + + try_once in_lam int_cxt -- There's one textual occurrence + | not in_lam = isNotTopLevel top_lvl || early_phase + | otherwise = int_cxt && canInlineInLam rhs + +-- Be very careful before inlining inside a lambda, becuase (a) we must not +-- invalidate occurrence information, and (b) we want to avoid pushing a +-- single allocation (here) into multiple allocations (inside lambda). +-- Inlining a *function* with a single *saturated* call would be ok, mind you. +-- || (if is_cheap && not (canInlineInLam rhs) then pprTrace "preinline" (ppr bndr <+> ppr rhs) ok else ok) +-- where +-- is_cheap = exprIsCheap rhs +-- ok = is_cheap && int_cxt + + -- int_cxt The context isn't totally boring + -- E.g. let f = \ab.BIG in \y. map f xs + -- Don't want to substitute for f, because then we allocate + -- its closure every time the \y is called + -- But: let f = \ab.BIG in \y. map (f y) xs + -- Now we do want to substitute for f, even though it's not + -- saturated, because we're going to allocate a closure for + -- (f y) every time round the loop anyhow. + + -- canInlineInLam => free vars of rhs are (Once in_lam) or Many, + -- so substituting rhs inside a lambda doesn't change the occ info. + -- Sadly, not quite the same as exprIsHNF. + canInlineInLam (Lit l) = True + canInlineInLam (Lam b e) = isRuntimeVar b || canInlineInLam e + canInlineInLam (Note _ e) = canInlineInLam e + canInlineInLam _ = False + + early_phase = case phase of + SimplPhase 0 -> False + other -> True +-- If we don't have this early_phase test, consider +-- x = length [1,2,3] +-- The full laziness pass carefully floats all the cons cells to +-- top level, and preInlineUnconditionally floats them all back in. +-- Result is (a) static allocation replaced by dynamic allocation +-- (b) many simplifier iterations because this tickles +-- a related problem; only one inlining per pass +-- +-- On the other hand, I have seen cases where top-level fusion is +-- lost if we don't inline top level thing (e.g. string constants) +-- Hence the test for phase zero (which is the phase for all the final +-- simplifications). Until phase zero we take no special notice of +-- top level things, but then we become more leery about inlining +-- them. + +\end{code} + +postInlineUnconditionally +~~~~~~~~~~~~~~~~~~~~~~~~~ +@postInlineUnconditionally@ decides whether to unconditionally inline +a thing based on the form of its RHS; in particular if it has a +trivial RHS. If so, we can inline and discard the binding altogether. + +NB: a loop breaker has must_keep_binding = True and non-loop-breakers +only have *forward* references Hence, it's safe to discard the binding + +NOTE: This isn't our last opportunity to inline. We're at the binding +site right now, and we'll get another opportunity when we get to the +ocurrence(s) + +Note that we do this unconditional inlining only for trival RHSs. +Don't inline even WHNFs inside lambdas; doing so may simply increase +allocation when the function is called. This isn't the last chance; see +NOTE above. + +NB: Even inline pragmas (e.g. IMustBeINLINEd) are ignored here Why? +Because we don't even want to inline them into the RHS of constructor +arguments. See NOTE above + +NB: At one time even NOINLINE was ignored here: if the rhs is trivial +it's best to inline it anyway. We often get a=E; b=a from desugaring, +with both a and b marked NOINLINE. But that seems incompatible with +our new view that inlining is like a RULE, so I'm sticking to the 'active' +story for now. + +\begin{code} +postInlineUnconditionally :: SimplEnv -> TopLevelFlag -> OutId -> OccInfo -> OutExpr -> Unfolding -> Bool +postInlineUnconditionally env top_lvl bndr occ_info rhs unfolding + | not active = False + | isLoopBreaker occ_info = False + | isExportedId bndr = False + | exprIsTrivial rhs = True + | otherwise + = case occ_info of + OneOcc in_lam one_br int_cxt + -> (one_br || smallEnoughToInline unfolding) -- Small enough to dup + -- ToDo: consider discount on smallEnoughToInline if int_cxt is true + -- + -- NB: Do we want to inline arbitrarily big things becuase + -- one_br is True? that can lead to inline cascades. But + -- preInlineUnconditionlly has dealt with all the common cases + -- so perhaps it's worth the risk. Here's an example + -- let f = if b then Left (\x.BIG) else Right (\y.BIG) + -- in \y. ....f.... + -- We can't preInlineUnconditionally because that woud invalidate + -- the occ info for b. Yet f is used just once, and duplicating + -- the case work is fine (exprIsCheap). + + && ((isNotTopLevel top_lvl && not in_lam) || + -- But outside a lambda, we want to be reasonably aggressive + -- about inlining into multiple branches of case + -- e.g. let x = + -- in case y of { C1 -> ..x..; C2 -> ..x..; C3 -> ... } + -- Inlining can be a big win if C3 is the hot-spot, even if + -- the uses in C1, C2 are not 'interesting' + -- An example that gets worse if you add int_cxt here is 'clausify' + + (isCheapUnfolding unfolding && int_cxt)) + -- isCheap => acceptable work duplication; in_lam may be true + -- int_cxt to prevent us inlining inside a lambda without some + -- good reason. See the notes on int_cxt in preInlineUnconditionally + + other -> False + -- The point here is that for *non-values* that occur + -- outside a lambda, the call-site inliner won't have + -- a chance (becuase it doesn't know that the thing + -- only occurs once). The pre-inliner won't have gotten + -- it either, if the thing occurs in more than one branch + -- So the main target is things like + -- let x = f y in + -- case v of + -- True -> case x of ... + -- False -> case x of ... + -- I'm not sure how important this is in practice + where + active = case getMode env of + SimplGently -> isAlwaysActive prag + SimplPhase n -> isActive n prag + prag = idInlinePragma bndr + +activeInline :: SimplEnv -> OutId -> OccInfo -> Bool +activeInline env id occ + = case getMode env of + SimplGently -> isOneOcc occ && isAlwaysActive prag + -- No inlining at all when doing gentle stuff, + -- except for local things that occur once + -- The reason is that too little clean-up happens if you + -- don't inline use-once things. Also a bit of inlining is *good* for + -- full laziness; it can expose constant sub-expressions. + -- Example in spectral/mandel/Mandel.hs, where the mandelset + -- function gets a useful let-float if you inline windowToViewport + + -- NB: we used to have a second exception, for data con wrappers. + -- On the grounds that we use gentle mode for rule LHSs, and + -- they match better when data con wrappers are inlined. + -- But that only really applies to the trivial wrappers (like (:)), + -- and they are now constructed as Compulsory unfoldings (in MkId) + -- so they'll happen anyway. + + SimplPhase n -> isActive n prag + where + prag = idInlinePragma id + +activeRule :: SimplEnv -> Maybe (Activation -> Bool) +-- Nothing => No rules at all +activeRule env + | opt_RulesOff = Nothing + | otherwise + = case getMode env of + SimplGently -> Just isAlwaysActive + -- Used to be Nothing (no rules in gentle mode) + -- Main motivation for changing is that I wanted + -- lift String ===> ... + -- to work in Template Haskell when simplifying + -- splices, so we get simpler code for literal strings + SimplPhase n -> Just (isActive n) +\end{code} + + +%************************************************************************ +%* * +\subsection{Rebuilding a lambda} +%* * +%************************************************************************ + +\begin{code} +mkLam :: SimplEnv -> [OutBinder] -> OutExpr -> SimplCont -> SimplM FloatsWithExpr +\end{code} + +Try three things + a) eta reduction, if that gives a trivial expression + b) eta expansion [only if there are some value lambdas] + c) floating lets out through big lambdas + [only if all tyvar lambdas, and only if this lambda + is the RHS of a let] + +\begin{code} +mkLam env bndrs body cont + = getDOptsSmpl `thenSmpl` \dflags -> + mkLam' dflags env bndrs body cont + where + mkLam' dflags env bndrs body cont + | dopt Opt_DoEtaReduction dflags, + Just etad_lam <- tryEtaReduce bndrs body + = tick (EtaReduction (head bndrs)) `thenSmpl_` + returnSmpl (emptyFloats env, etad_lam) + + | dopt Opt_DoLambdaEtaExpansion dflags, + any isRuntimeVar bndrs + = tryEtaExpansion body `thenSmpl` \ body' -> + returnSmpl (emptyFloats env, mkLams bndrs body') + +{- Sept 01: I'm experimenting with getting the + full laziness pass to float out past big lambdsa + | all isTyVar bndrs, -- Only for big lambdas + contIsRhs cont -- Only try the rhs type-lambda floating + -- if this is indeed a right-hand side; otherwise + -- we end up floating the thing out, only for float-in + -- to float it right back in again! + = tryRhsTyLam env bndrs body `thenSmpl` \ (floats, body') -> + returnSmpl (floats, mkLams bndrs body') +-} + + | otherwise + = returnSmpl (emptyFloats env, mkLams bndrs body) +\end{code} + + +%************************************************************************ +%* * +\subsection{Eta expansion and reduction} +%* * +%************************************************************************ + +We try for eta reduction here, but *only* if we get all the +way to an exprIsTrivial expression. +We don't want to remove extra lambdas unless we are going +to avoid allocating this thing altogether + +\begin{code} +tryEtaReduce :: [OutBinder] -> OutExpr -> Maybe OutExpr +tryEtaReduce bndrs body + -- We don't use CoreUtils.etaReduce, because we can be more + -- efficient here: + -- (a) we already have the binders + -- (b) we can do the triviality test before computing the free vars + = go (reverse bndrs) body + where + go (b : bs) (App fun arg) | ok_arg b arg = go bs fun -- Loop round + go [] fun | ok_fun fun = Just fun -- Success! + go _ _ = Nothing -- Failure! + + ok_fun fun = exprIsTrivial fun + && not (any (`elemVarSet` (exprFreeVars fun)) bndrs) + && (exprIsHNF fun || all ok_lam bndrs) + ok_lam v = isTyVar v || isDictId v + -- The exprIsHNF is because eta reduction is not + -- valid in general: \x. bot /= bot + -- So we need to be sure that the "fun" is a value. + -- + -- However, we always want to reduce (/\a -> f a) to f + -- This came up in a RULE: foldr (build (/\a -> g a)) + -- did not match foldr (build (/\b -> ...something complex...)) + -- The type checker can insert these eta-expanded versions, + -- with both type and dictionary lambdas; hence the slightly + -- ad-hoc isDictTy + + ok_arg b arg = varToCoreExpr b `cheapEqExpr` arg +\end{code} + + + Try eta expansion for RHSs + +We go for: + f = \x1..xn -> N ==> f = \x1..xn y1..ym -> N y1..ym + (n >= 0) + +where (in both cases) + + * The xi can include type variables + + * The yi are all value variables + + * N is a NORMAL FORM (i.e. no redexes anywhere) + wanting a suitable number of extra args. + +We may have to sandwich some coerces between the lambdas +to make the types work. exprEtaExpandArity looks through coerces +when computing arity; and etaExpand adds the coerces as necessary when +actually computing the expansion. + +\begin{code} +tryEtaExpansion :: OutExpr -> SimplM OutExpr +-- There is at least one runtime binder in the binders +tryEtaExpansion body + = getUniquesSmpl `thenSmpl` \ us -> + returnSmpl (etaExpand fun_arity us body (exprType body)) + where + fun_arity = exprEtaExpandArity body \end{code} %************************************************************************ %* * -\subsection{Local tyvar-lifting} +\subsection{Floating lets out of big lambdas} %* * %************************************************************************ -mkRhsTyLam tries this transformation, when the big lambda appears as +tryRhsTyLam tries this transformation, when the big lambda appears as the RHS of a let(rec) binding: /\abc -> let(rec) x = e in b @@ -533,42 +960,42 @@ as we would normally do. \begin{code} -tryRhsTyLam :: OutExpr -> SimplM ([OutBind], OutExpr) +{- Trying to do this in full laziness + +tryRhsTyLam :: SimplEnv -> [OutTyVar] -> OutExpr -> SimplM FloatsWithExpr +-- Call ensures that all the binders are type variables -tryRhsTyLam rhs -- Only does something if there's a let - | null tyvars || not (worth_it body) -- inside a type lambda, - = returnSmpl ([], rhs) -- and a WHNF inside that +tryRhsTyLam env tyvars body -- Only does something if there's a let + | not (all isTyVar tyvars) + || not (worth_it body) -- inside a type lambda, + = returnSmpl (emptyFloats env, body) -- and a WHNF inside that | otherwise - = go (\x -> x) body `thenSmpl` \ (binds, body') -> - returnSmpl (binds, mkLams tyvars body') + = go env (\x -> x) body where - (tyvars, body) = collectTyBinders rhs - - worth_it (Let (NonRec x rhs) e) | isUnLiftedType (exprType rhs) = False - worth_it (Let _ e) = whnf_in_middle e - worth_it other = False + worth_it e@(Let _ _) = whnf_in_middle e + worth_it e = False - whnf_in_middle (Let (NonRec x rhs) e) | isUnLiftedType (exprType rhs) = False + whnf_in_middle (Let (NonRec x rhs) e) | isUnLiftedType (idType x) = False whnf_in_middle (Let _ e) = whnf_in_middle e whnf_in_middle e = exprIsCheap e - go fn (Let bind@(NonRec var rhs) body) + main_tyvar_set = mkVarSet tyvars + + go env fn (Let bind@(NonRec var rhs) body) | exprIsTrivial rhs - = go (fn . Let bind) body + = go env (fn . Let bind) body - go fn (Let (NonRec var rhs) body) - = mk_poly tyvars_here var `thenSmpl` \ (var', rhs') -> - go (fn . Let (mk_silly_bind var rhs')) body `thenSmpl` \ (binds, body') -> - returnSmpl (NonRec var' (mkLams tyvars_here (fn rhs)) : binds, body') + go env fn (Let (NonRec var rhs) body) + = mk_poly tyvars_here var `thenSmpl` \ (var', rhs') -> + addAuxiliaryBind env (NonRec var' (mkLams tyvars_here (fn rhs))) $ \ env -> + go env (fn . Let (mk_silly_bind var 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 + + tyvars_here = varSetElems (main_tyvar_set `intersectVarSet` exprSomeFreeVars isTyVar rhs) + -- Abstract only over the type variables free in the rhs -- wrt which the new binding is abstracted. But the naive -- approach of abstract wrt the tyvars free in the Id's type -- fails. Consider: @@ -584,29 +1011,27 @@ tryRhsTyLam rhs -- Only does something if there's a let -- abstracting wrt *all* the tyvars. We'll see if that -- gives rise to problems. SLPJ June 98 - go fn (Let (Rec prs) body) + go env fn (Let (Rec prs) body) = mapAndUnzipSmpl (mk_poly tyvars_here) vars `thenSmpl` \ (vars', rhss') -> let - gn body = fn (foldr Let body (zipWith mk_silly_bind vars rhss')) - new_bind = Rec (vars' `zip` [mkLams tyvars_here (gn rhs) | rhs <- rhss]) + gn body = fn (foldr Let body (zipWith mk_silly_bind vars rhss')) + pairs = vars' `zip` [mkLams tyvars_here (gn rhs) | rhs <- rhss] in - go gn body `thenSmpl` \ (binds, body') -> - returnSmpl (new_bind : binds, body') + addAuxiliaryBind env (Rec pairs) $ \ env -> + go env gn body where (vars,rhss) = unzip prs - tyvars_here = tyvars - -- varSetElems (main_tyvar_set `intersectVarSet` tyVarsOfTypes var_tys) - -- var_tys = map idType vars + tyvars_here = varSetElems (main_tyvar_set `intersectVarSet` exprsSomeFreeVars isTyVar (map snd prs)) -- See notes with tyvars_here above - go fn body = returnSmpl ([], fn body) + go env fn body = returnSmpl (emptyFloats env, fn body) mk_poly tyvars_here var = getUniqueSmpl `thenSmpl` \ uniq -> let 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 + poly_id = mkLocalId 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! @@ -643,75 +1068,144 @@ tryRhsTyLam rhs -- Only does something if there's a let -- 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 expansion} +\subsection{Case alternative filtering %* * %************************************************************************ - Try eta expansion for RHSs +prepareAlts does two things: -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 +1. Eliminate alternatives that cannot match, including the + DEFAULT alternative. -where (in both cases) +2. If the DEFAULT alternative can match only one possible constructor, + then make that constructor explicit. + e.g. + case e of x { DEFAULT -> rhs } + ===> + case e of x { (a,b) -> rhs } + where the type is a single constructor type. This gives better code + when rhs also scrutinises x or e. - * The xi can include type variables +It's a good idea do do this stuff before simplifying the alternatives, to +avoid simplifying alternatives we know can't happen, and to come up with +the list of constructors that are handled, to put into the IdInfo of the +case binder, for use when simplifying the alternatives. - * The yi are all value variables +Eliminating the default alternative in (1) isn't so obvious, but it can +happen: - * N is a NORMAL FORM (i.e. no redexes anywhere) - wanting a suitable number of extra args. +data Colour = Red | Green | Blue + +f x = case x of + Red -> .. + Green -> .. + DEFAULT -> h x - * the Ei must not have unlifted type +h y = case y of + Blue -> .. + DEFAULT -> [ case y of ... ] + +If we inline h into f, the default case of the inlined h can't happen. +If we don't notice this, we may end up filtering out *all* the cases +of the inner case y, which give us nowhere to go! -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. \begin{code} -tryEtaExpansion :: OutExpr -> OutType -> SimplM ([OutBind], OutExpr) -tryEtaExpansion rhs rhs_ty - | not opt_SimplDoLambdaEtaExpansion -- Not if switched off - || exprIsTrivial rhs -- Not if RHS is trivial - || final_arity == 0 -- Not if arity is zero - = returnSmpl ([], rhs) - - | n_val_args == 0 && not arity_is_manifest - = -- Some lambdas but not enough: case 1 - getUniqSupplySmpl `thenSmpl` \ us -> - returnSmpl ([], etaExpand final_arity us rhs rhs_ty) - - | n_val_args > 0 && not (any cant_bind arg_infos) - = -- Partial application: case 2 - mapAndUnzipSmpl bind_z_arg arg_infos `thenSmpl` \ (maybe_z_binds, z_args) -> - getUniqSupplySmpl `thenSmpl` \ us -> - returnSmpl (catMaybes maybe_z_binds, - etaExpand final_arity us (mkApps fun z_args) rhs_ty) +prepareAlts :: OutExpr -- Scrutinee + -> InId -- Case binder (passed only to use in statistics) + -> [InAlt] -- Increasing order + -> SimplM ([InAlt], -- Better alternatives, still incresaing order + [AltCon]) -- These cases are handled + +prepareAlts scrut case_bndr alts + = let + (alts_wo_default, maybe_deflt) = findDefault alts + + impossible_cons = case scrut of + Var v -> otherCons (idUnfolding v) + other -> [] + + -- Filter out alternatives that can't possibly match + better_alts | null impossible_cons = alts_wo_default + | otherwise = [alt | alt@(con,_,_) <- alts_wo_default, + not (con `elem` impossible_cons)] + + -- "handled_cons" are handled either by the context, + -- or by a branch in this case expression + -- (Don't add DEFAULT to the handled_cons!!) + handled_cons = impossible_cons ++ [con | (con,_,_) <- better_alts] + in + -- Filter out the default, if it can't happen, + -- or replace it with "proper" alternative if there + -- is only one constructor left + prepareDefault scrut case_bndr handled_cons maybe_deflt `thenSmpl` \ deflt_alt -> + + returnSmpl (mergeAlts better_alts deflt_alt, handled_cons) + -- We need the mergeAlts in case the new default_alt + -- has turned into a constructor alternative. + +prepareDefault scrut case_bndr handled_cons (Just rhs) + | Just (tycon, inst_tys) <- splitTyConApp_maybe (exprType scrut), + -- Use exprType scrut here, rather than idType case_bndr, because + -- case_bndr is an InId, so exprType scrut may have more information + -- Test simpl013 is an example + isAlgTyCon tycon, -- It's a data type, tuple, or unboxed tuples. + not (isNewTyCon tycon), -- We can have a newtype, if we are just doing an eval: + -- case x of { DEFAULT -> e } + -- and we don't want to fill in a default for them! + Just all_cons <- tyConDataCons_maybe tycon, + not (null all_cons), -- This is a tricky corner case. If the data type has no constructors, + -- which GHC allows, then the case expression will have at most a default + -- alternative. We don't want to eliminate that alternative, because the + -- invariant is that there's always one alternative. It's more convenient + -- to leave + -- case x of { DEFAULT -> e } + -- as it is, rather than transform it to + -- error "case cant match" + -- which would be quite legitmate. But it's a really obscure corner, and + -- not worth wasting code on. + let handled_data_cons = [data_con | DataAlt data_con <- handled_cons], + let missing_cons = [con | con <- all_cons, + not (con `elem` handled_data_cons)] + = case missing_cons of + [] -> returnSmpl [] -- Eliminate the default alternative + -- if it can't match + + [con] -> -- It matches exactly one constructor, so fill it in + tick (FillInCaseDefault case_bndr) `thenSmpl_` + mk_args con inst_tys `thenSmpl` \ args -> + returnSmpl [(DataAlt con, args, rhs)] + + two_or_more -> returnSmpl [(DEFAULT, [], rhs)] | otherwise - = returnSmpl ([], rhs) - where - (fun, args) = collectArgs rhs - n_val_args = valArgCount args - (fun_arity, arity_is_manifest) = exprEtaExpandArity fun - final_arity = 0 `max` (fun_arity - n_val_args) - arg_infos = [(arg, exprType arg, exprIsTrivial arg) | arg <- args] - cant_bind (_, ty, triv) = not triv && isUnLiftedType ty - - 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) + = returnSmpl [(DEFAULT, [], rhs)] + +prepareDefault scrut case_bndr handled_cons Nothing + = returnSmpl [] + +mk_args missing_con inst_tys + = mk_tv_bndrs missing_con inst_tys `thenSmpl` \ (tv_bndrs, inst_tys') -> + getUniquesSmpl `thenSmpl` \ id_uniqs -> + let arg_tys = dataConInstArgTys missing_con inst_tys' + arg_ids = zipWith (mkSysLocal FSLIT("a")) id_uniqs arg_tys + in + returnSmpl (tv_bndrs ++ arg_ids) + +mk_tv_bndrs missing_con inst_tys + | isVanillaDataCon missing_con + = returnSmpl ([], inst_tys) + | otherwise + = getUniquesSmpl `thenSmpl` \ tv_uniqs -> + let new_tvs = zipWith mk tv_uniqs (dataConTyVars missing_con) + mk uniq tv = mkTyVar (mkSysTvName uniq FSLIT("t")) (tyVarKind tv) + in + returnSmpl (new_tvs, mkTyVarTys new_tvs) \end{code} @@ -721,109 +1215,378 @@ tryEtaExpansion rhs rhs_ty %* * %************************************************************************ +mkCase puts a case expression back together, trying various transformations first. + \begin{code} -mkCase :: OutExpr -> OutId -> [OutAlt] -> SimplM OutExpr +mkCase :: OutExpr -> OutId -> OutType + -> [OutAlt] -- Increasing order + -> SimplM OutExpr + +mkCase scrut case_bndr ty alts + = getDOptsSmpl `thenSmpl` \dflags -> + mkAlts dflags scrut case_bndr alts `thenSmpl` \ better_alts -> + mkCase1 scrut case_bndr ty better_alts \end{code} -@mkCase@ tries the following transformation (if possible): - -case e of b { ==> case e of b { - p1 -> rhs1 p1 -> rhs1 - ... ... - pm -> rhsm pm -> rhsm - _ -> case b of b' { pn -> rhsn[b/b'] {or (alg) let b=b' in rhsn} - {or (prim) case b of b' { _ -> rhsn}} - pn -> rhsn ... - ... po -> rhso[b/b'] - po -> rhso _ -> rhsd[b/b'] {or let b'=b in rhsd} - _ -> rhsd -} - -which merges two cases in one case when -- the default alternative of -the outer case scrutises the same variable as the outer case This -transformation is called Case Merging. It avoids that the same -variable is scrutinised multiple times. + +mkAlts tries these things: + +1. If several alternatives are identical, merge them into + a single DEFAULT alternative. I've occasionally seen this + making a big difference: + + case e of =====> case e of + C _ -> f x D v -> ....v.... + D v -> ....v.... DEFAULT -> f x + DEFAULT -> f x + + The point is that we merge common RHSs, at least for the DEFAULT case. + [One could do something more elaborate but I've never seen it needed.] + To avoid an expensive test, we just merge branches equal to the *first* + alternative; this picks up the common cases + a) all branches equal + b) some branches equal to the DEFAULT (which occurs first) + +2. Case merging: + case e of b { ==> case e of b { + p1 -> rhs1 p1 -> rhs1 + ... ... + pm -> rhsm pm -> rhsm + _ -> case b of b' { pn -> let b'=b in rhsn + pn -> rhsn ... + ... po -> let b'=b in rhso + po -> rhso _ -> let b'=b in rhsd + _ -> rhsd + } + + which merges two cases in one case when -- the default alternative of + the outer case scrutises the same variable as the outer case This + transformation is called Case Merging. It avoids that the same + variable is scrutinised multiple times. + + +The case where transformation (1) showed up was like this (lib/std/PrelCError.lhs): + + x | p `is` 1 -> e1 + | p `is` 2 -> e2 + ...etc... + +where @is@ was something like + + p `is` n = p /= (-1) && p == n + +This gave rise to a horrible sequence of cases + + case p of + (-1) -> $j p + 1 -> e1 + DEFAULT -> $j p + +and similarly in cascade for all the join points! + + \begin{code} -mkCase scrut outer_bndr outer_alts - | opt_SimplCaseMerge - && maybeToBool maybe_case_in_default - - = 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 - -- mkCase applied to them, so they won't have a case in their default - -- Secondly, if you do, you get an infinite loop, because the bindNonRec - -- in munge_rhs puts a case into the DEFAULT branch! +-------------------------------------------------- +-- 1. Merge identical branches +-------------------------------------------------- +mkAlts dflags scrut case_bndr alts@((con1,bndrs1,rhs1) : con_alts) + | all isDeadBinder bndrs1, -- Remember the default + length filtered_alts < length con_alts -- alternative comes first + = tick (AltMerge case_bndr) `thenSmpl_` + returnSmpl better_alts + where + filtered_alts = filter keep con_alts + keep (con,bndrs,rhs) = not (all isDeadBinder bndrs && rhs `cheapEqExpr` rhs1) + better_alts = (DEFAULT, [], rhs1) : filtered_alts + + +-------------------------------------------------- +-- 2. Merge nested cases +-------------------------------------------------- + +mkAlts dflags scrut outer_bndr outer_alts + | dopt Opt_CaseMerge dflags, + (outer_alts_without_deflt, maybe_outer_deflt) <- findDefault outer_alts, + Just (Case (Var scrut_var) inner_bndr _ inner_alts) <- maybe_outer_deflt, + scruting_same_var scrut_var + = let + munged_inner_alts = [(con, args, munge_rhs rhs) | (con, args, rhs) <- inner_alts] + munge_rhs rhs = bindCaseBndr inner_bndr (Var outer_bndr) rhs + + new_alts = mergeAlts outer_alts_without_deflt munged_inner_alts + -- The merge keeps the inner DEFAULT at the front, if there is one + -- and eliminates any inner_alts that are shadowed by the outer_alts + in + tick (CaseMerge outer_bndr) `thenSmpl_` + returnSmpl new_alts + -- Warning: don't call mkAlts recursively! + -- Firstly, there's no point, because inner alts have already had + -- mkCase applied to them, so they won't have a case in their default + -- Secondly, if you do, you get an infinite loop, because the bindCaseBndr + -- in munge_rhs may put a case into the DEFAULT branch! where - new_alts = outer_alts_without_deflt ++ munged_inner_alts - maybe_case_in_default = case findDefault outer_alts of - (outer_alts_without_default, - Just (Case (Var scrut_var) inner_bndr inner_alts)) - - | outer_bndr == scrut_var - -> Just (outer_alts_without_default, inner_bndr, inner_alts) - other -> Nothing - - Just (outer_alts_without_deflt, inner_bndr, inner_alts) = maybe_case_in_default - - -- Eliminate any inner alts which are shadowed by the outer ones - outer_cons = [con | (con,_,_) <- outer_alts_without_deflt] - - munged_inner_alts = [ (con, args, munge_rhs rhs) - | (con, args, rhs) <- inner_alts, - not (con `elem` outer_cons) -- Eliminate shadowed inner alts - ] - munge_rhs rhs = bindNonRec inner_bndr (Var outer_bndr) rhs + -- We are scrutinising the same variable if it's + -- the outer case-binder, or if the outer case scrutinises a variable + -- (and it's the same). Testing both allows us not to replace the + -- outer scrut-var with the outer case-binder (Simplify.simplCaseBinder). + scruting_same_var = case scrut of + Var outer_scrut -> \ v -> v == outer_bndr || v == outer_scrut + other -> \ v -> v == outer_bndr + +------------------------------------------------ +-- Catch-all +------------------------------------------------ + +mkAlts dflags scrut case_bndr other_alts = returnSmpl other_alts + + +--------------------------------- +mergeAlts :: [OutAlt] -> [OutAlt] -> [OutAlt] +-- Merge preserving order; alternatives in the first arg +-- shadow ones in the second +mergeAlts [] as2 = as2 +mergeAlts as1 [] = as1 +mergeAlts (a1:as1) (a2:as2) + = case a1 `cmpAlt` a2 of + LT -> a1 : mergeAlts as1 (a2:as2) + EQ -> a1 : mergeAlts as1 as2 -- Discard a2 + GT -> a2 : mergeAlts (a1:as1) as2 \end{code} -Now the identity-case transformation: + + +================================================================================= + +mkCase1 tries these things + +1. Eliminate the case altogether if possible + +2. Case-identity: case e of ===> e - True -> True; + True -> True; False -> False -and similar friends. + and similar friends. + + +Start with a simple situation: + + case x# of ===> e[x#/y#] + y# -> e + +(when x#, y# are of primitive type, of course). We can't (in general) +do this for algebraic cases, because we might turn bottom into +non-bottom! + +Actually, we generalise this idea to look for a case where we're +scrutinising a variable, and we know that only the default case can +match. For example: +\begin{verbatim} + case x of + 0# -> ... + other -> ...(case x of + 0# -> ... + other -> ...) ... +\end{code} +Here the inner case can be eliminated. This really only shows up in +eliminating error-checking code. + +We also make sure that we deal with this very common case: + + case e of + x -> ...x... + +Here we are using the case as a strict let; if x is used only once +then we want to inline it. We have to be careful that this doesn't +make the program terminate when it would have diverged before, so we +check that + - x is used strictly, or + - e is already evaluated (it may so if e is a variable) + +Lastly, we generalise the transformation to handle this: + + case e of ===> r + True -> r + False -> r + +We only do this for very cheaply compared r's (constructors, literals +and variables). If pedantic bottoms is on, we only do it when the +scrutinee is a PrimOp which can't fail. + +We do it *here*, looking at un-simplified alternatives, because we +have to check that r doesn't mention the variables bound by the +pattern in each alternative, so the binder-info is rather useful. + +So the case-elimination algorithm is: + + 1. Eliminate alternatives which can't match + + 2. Check whether all the remaining alternatives + (a) do not mention in their rhs any of the variables bound in their pattern + and (b) have equal rhss + + 3. Check we can safely ditch the case: + * PedanticBottoms is off, + or * the scrutinee is an already-evaluated variable + or * the scrutinee is a primop which is ok for speculation + -- ie we want to preserve divide-by-zero errors, and + -- calls to error itself! + + or * [Prim cases] the scrutinee is a primitive variable + + or * [Alg cases] the scrutinee is a variable and + either * the rhs is the same variable + (eg case x of C a b -> x ===> x) + or * there is only one alternative, the default alternative, + and the binder is used strictly in its scope. + [NB this is helped by the "use default binder where + possible" transformation; see below.] + + +If so, then we can replace the case with one of the rhss. + +Further notes about case elimination +~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ +Consider: test :: Integer -> IO () + test = print + +Turns out that this compiles to: + Print.test + = \ eta :: Integer + eta1 :: State# RealWorld -> + case PrelNum.< eta PrelNum.zeroInteger of wild { __DEFAULT -> + case hPutStr stdout + (PrelNum.jtos eta ($w[] @ Char)) + eta1 + of wild1 { (# new_s, a4 #) -> PrelIO.lvl23 new_s }} + +Notice the strange '<' which has no effect at all. This is a funny one. +It started like this: + +f x y = if x < 0 then jtos x + else if y==0 then "" else jtos x + +At a particular call site we have (f v 1). So we inline to get + + if v < 0 then jtos x + else if 1==0 then "" else jtos x + +Now simplify the 1==0 conditional: + + if v<0 then jtos v else jtos v + +Now common-up the two branches of the case: + + case (v<0) of DEFAULT -> jtos v + +Why don't we drop the case? Because it's strict in v. It's technically +wrong to drop even unnecessary evaluations, and in practice they +may be a result of 'seq' so we *definitely* don't want to drop those. +I don't really know how to improve this situation. + \begin{code} -mkCase scrut case_bndr alts +-------------------------------------------------- +-- 0. Check for empty alternatives +-------------------------------------------------- + +-- This isn't strictly an error. It's possible that the simplifer might "see" +-- that an inner case has no accessible alternatives before it "sees" that the +-- entire branch of an outer case is inaccessible. So we simply +-- put an error case here insteadd +mkCase1 scrut case_bndr ty [] + = pprTrace "mkCase1: null alts" (ppr case_bndr <+> ppr scrut) $ + return (mkApps (Var eRROR_ID) + [Type ty, Lit (mkStringLit "Impossible alternative")]) + +-------------------------------------------------- +-- 1. Eliminate the case altogether if poss +-------------------------------------------------- + +mkCase1 scrut case_bndr ty [(con,bndrs,rhs)] + -- See if we can get rid of the case altogether + -- See the extensive notes on case-elimination above + -- mkCase made sure that if all the alternatives are equal, + -- then there is now only one (DEFAULT) rhs + | all isDeadBinder bndrs, + + -- Check that the scrutinee can be let-bound instead of case-bound + exprOkForSpeculation scrut + -- OK not to evaluate it + -- This includes things like (==# a# b#)::Bool + -- so that we simplify + -- case ==# a# b# of { True -> x; False -> x } + -- to just + -- x + -- This particular example shows up in default methods for + -- comparision operations (e.g. in (>=) for Int.Int32) + || exprIsHNF scrut -- It's already evaluated + || var_demanded_later scrut -- It'll be demanded later + +-- || not opt_SimplPedanticBottoms) -- Or we don't care! +-- We used to allow improving termination by discarding cases, unless -fpedantic-bottoms was on, +-- but that breaks badly for the dataToTag# primop, which relies on a case to evaluate +-- its argument: case x of { y -> dataToTag# y } +-- Here we must *not* discard the case, because dataToTag# just fetches the tag from +-- the info pointer. So we'll be pedantic all the time, and see if that gives any +-- other problems +-- Also we don't want to discard 'seq's + = tick (CaseElim case_bndr) `thenSmpl_` + returnSmpl (bindCaseBndr case_bndr scrut rhs) + + where + -- The case binder is going to be evaluated later, + -- and the scrutinee is a simple variable + var_demanded_later (Var v) = isStrictDmd (idNewDemandInfo case_bndr) + var_demanded_later other = False + + +-------------------------------------------------- +-- 2. Identity case +-------------------------------------------------- + +mkCase1 scrut case_bndr ty alts -- Identity case | all identity_alt alts = tick (CaseIdentity case_bndr) `thenSmpl_` - returnSmpl scrut + returnSmpl (re_note scrut) where - 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 + identity_alt (con, args, rhs) = de_note rhs `cheapEqExpr` identity_rhs con args - arg_tys = tyConAppArgs (idType case_bndr) -\end{code} + identity_rhs (DataAlt con) args = mkConApp con (arg_tys ++ map varToCoreExpr args) + identity_rhs (LitAlt lit) _ = Lit lit + identity_rhs DEFAULT _ = Var case_bndr -The catch-all case + arg_tys = map Type (tyConAppArgs (idType case_bndr)) -\begin{code} -mkCase other_scrut case_bndr other_alts - = returnSmpl (Case other_scrut case_bndr other_alts) + -- We've seen this: + -- case coerce T e of x { _ -> coerce T' x } + -- And we definitely want to eliminate this case! + -- So we throw away notes from the RHS, and reconstruct + -- (at least an approximation) at the other end + de_note (Note _ e) = de_note e + de_note e = e + + -- re_note wraps a coerce if it might be necessary + re_note scrut = case head alts of + (_,_,rhs1@(Note _ _)) -> mkCoerce2 (exprType rhs1) (idType case_bndr) scrut + other -> scrut + + +-------------------------------------------------- +-- Catch-all +-------------------------------------------------- +mkCase1 scrut bndr ty alts = returnSmpl (Case scrut bndr ty alts) \end{code} -\begin{code} -findDefault :: [CoreAlt] -> ([CoreAlt], Maybe CoreExpr) -findDefault [] = ([], Nothing) -findDefault ((DEFAULT,args,rhs) : alts) = ASSERT( null alts && null args ) - ([], Just rhs) -findDefault (alt : alts) = case findDefault alts of - (alts', deflt) -> (alt : alts', deflt) - -findAlt :: AltCon -> [CoreAlt] -> CoreAlt -findAlt con alts - = go alts - where - go [] = pprPanic "Missing alternative" (ppr con $$ vcat (map ppr alts)) - go (alt : alts) | matches alt = alt - | otherwise = go alts +When adding auxiliary bindings for the case binder, it's worth checking if +its dead, because it often is, and occasionally these mkCase transformations +cascade rather nicely. - matches (DEFAULT, _, _) = True - matches (con1, _, _) = con == con1 +\begin{code} +bindCaseBndr bndr rhs body + | isDeadBinder bndr = body + | otherwise = bindNonRec bndr rhs body \end{code}