2 % (c) The AQUA Project, Glasgow University, 1993-1998
4 \section[SimplUtils]{The simplifier utilities}
8 simplBinder, simplBinders, simplRecBndrs,
9 simplLetBndr, simplLamBndrs,
10 newId, mkLam, prepareAlts, mkCase,
12 -- The continuation type
13 SimplCont(..), DupFlag(..), LetRhsFlag(..),
14 contIsDupable, contResultType,
15 countValArgs, countArgs, pushContArgs,
16 mkBoringStop, mkStop, contIsRhs, contIsRhsOrArg,
17 getContArgs, interestingCallContext, interestingArg, isStrictType
21 #include "HsVersions.h"
23 import CmdLineOpts ( SimplifierSwitch(..), opt_UF_UpdateInPlace,
26 import CoreFVs ( exprFreeVars )
27 import CoreUtils ( cheapEqExpr, exprType, exprIsTrivial,
28 etaExpand, exprEtaExpandArity, bindNonRec, mkCoerce2,
29 findDefault, exprOkForSpeculation, exprIsValue
31 import qualified Subst ( simplBndrs, simplBndr, simplLetId, simplLamBndr )
32 import Id ( Id, idType, idInfo, isDataConWorkId,
33 mkSysLocal, isDeadBinder, idNewDemandInfo,
34 idUnfolding, idNewStrictness
36 import NewDemand ( isStrictDmd, isBotRes, splitStrictSig )
38 import Type ( Type, seqType, splitFunTys, dropForAlls, isStrictType,
39 splitTyConApp_maybe, tyConAppArgs, mkTyVarTys
41 import TcType ( isDictTy )
42 import OccName ( EncodedFS )
43 import TyCon ( tyConDataCons_maybe, isAlgTyCon, isNewTyCon )
44 import DataCon ( dataConRepArity, dataConExistentialTyVars, dataConArgTys )
45 import Var ( mkSysTyVar, tyVarKind )
47 import Util ( lengthExceeds, mapAccumL )
52 %************************************************************************
54 \subsection{The continuation data type}
56 %************************************************************************
59 data SimplCont -- Strict contexts
60 = Stop OutType -- Type of the result
62 Bool -- True <=> This is the RHS of a thunk whose type suggests
63 -- that update-in-place would be possible
64 -- (This makes the inliner a little keener.)
66 | CoerceIt OutType -- The To-type, simplified
69 | InlinePlease -- This continuation makes a function very
70 SimplCont -- keen to inline itelf
73 InExpr SimplEnv -- The argument, as yet unsimplified,
74 SimplCont -- and its environment
77 InId [InAlt] SimplEnv -- The case binder, alts, and subst-env
80 | ArgOf LetRhsFlag -- An arbitrary strict context: the argument
81 -- of a strict function, or a primitive-arg fn
83 -- No DupFlag because we never duplicate it
84 OutType -- arg_ty: type of the argument itself
85 OutType -- cont_ty: the type of the expression being sought by the context
86 -- f (error "foo") ==> coerce t (error "foo")
88 -- We need to know the type t, to which to coerce.
90 (SimplEnv -> OutExpr -> SimplM FloatsWithExpr) -- What to do with the result
91 -- The result expression in the OutExprStuff has type cont_ty
93 data LetRhsFlag = AnArg -- It's just an argument not a let RHS
94 | AnRhs -- It's the RHS of a let (so please float lets out of big lambdas)
96 instance Outputable LetRhsFlag where
97 ppr AnArg = ptext SLIT("arg")
98 ppr AnRhs = ptext SLIT("rhs")
100 instance Outputable SimplCont where
101 ppr (Stop _ is_rhs _) = ptext SLIT("Stop") <> brackets (ppr is_rhs)
102 ppr (ApplyTo dup arg se cont) = (ptext SLIT("ApplyTo") <+> ppr dup <+> ppr arg) $$ ppr cont
103 ppr (ArgOf _ _ _ _) = ptext SLIT("ArgOf...")
104 ppr (Select dup bndr alts se cont) = (ptext SLIT("Select") <+> ppr dup <+> ppr bndr) $$
105 (nest 4 (ppr alts)) $$ ppr cont
106 ppr (CoerceIt ty cont) = (ptext SLIT("CoerceIt") <+> ppr ty) $$ ppr cont
107 ppr (InlinePlease cont) = ptext SLIT("InlinePlease") $$ ppr cont
109 data DupFlag = OkToDup | NoDup
111 instance Outputable DupFlag where
112 ppr OkToDup = ptext SLIT("ok")
113 ppr NoDup = ptext SLIT("nodup")
117 mkBoringStop :: OutType -> SimplCont
118 mkBoringStop ty = Stop ty AnArg (canUpdateInPlace ty)
120 mkStop :: OutType -> LetRhsFlag -> SimplCont
121 mkStop ty is_rhs = Stop ty is_rhs (canUpdateInPlace ty)
123 contIsRhs :: SimplCont -> Bool
124 contIsRhs (Stop _ AnRhs _) = True
125 contIsRhs (ArgOf AnRhs _ _ _) = True
126 contIsRhs other = False
128 contIsRhsOrArg (Stop _ _ _) = True
129 contIsRhsOrArg (ArgOf _ _ _ _) = True
130 contIsRhsOrArg other = False
133 contIsDupable :: SimplCont -> Bool
134 contIsDupable (Stop _ _ _) = True
135 contIsDupable (ApplyTo OkToDup _ _ _) = True
136 contIsDupable (Select OkToDup _ _ _ _) = True
137 contIsDupable (CoerceIt _ cont) = contIsDupable cont
138 contIsDupable (InlinePlease cont) = contIsDupable cont
139 contIsDupable other = False
142 discardableCont :: SimplCont -> Bool
143 discardableCont (Stop _ _ _) = False
144 discardableCont (CoerceIt _ cont) = discardableCont cont
145 discardableCont (InlinePlease cont) = discardableCont cont
146 discardableCont other = True
148 discardCont :: SimplCont -- A continuation, expecting
149 -> SimplCont -- Replace the continuation with a suitable coerce
150 discardCont cont = case cont of
151 Stop to_ty is_rhs _ -> cont
152 other -> CoerceIt to_ty (mkBoringStop to_ty)
154 to_ty = contResultType cont
157 contResultType :: SimplCont -> OutType
158 contResultType (Stop to_ty _ _) = to_ty
159 contResultType (ArgOf _ _ to_ty _) = to_ty
160 contResultType (ApplyTo _ _ _ cont) = contResultType cont
161 contResultType (CoerceIt _ cont) = contResultType cont
162 contResultType (InlinePlease cont) = contResultType cont
163 contResultType (Select _ _ _ _ cont) = contResultType cont
166 countValArgs :: SimplCont -> Int
167 countValArgs (ApplyTo _ (Type ty) se cont) = countValArgs cont
168 countValArgs (ApplyTo _ val_arg se cont) = 1 + countValArgs cont
169 countValArgs other = 0
171 countArgs :: SimplCont -> Int
172 countArgs (ApplyTo _ arg se cont) = 1 + countArgs cont
176 pushContArgs :: SimplEnv -> [OutArg] -> SimplCont -> SimplCont
177 -- Pushes args with the specified environment
178 pushContArgs env [] cont = cont
179 pushContArgs env (arg : args) cont = ApplyTo NoDup arg env (pushContArgs env args cont)
184 getContArgs :: SwitchChecker
185 -> OutId -> SimplCont
186 -> ([(InExpr, SimplEnv, Bool)], -- Arguments; the Bool is true for strict args
187 SimplCont, -- Remaining continuation
188 Bool) -- Whether we came across an InlineCall
189 -- getContArgs id k = (args, k', inl)
190 -- args are the leading ApplyTo items in k
191 -- (i.e. outermost comes first)
192 -- augmented with demand info from the functionn
193 getContArgs chkr fun orig_cont
195 -- Ignore strictness info if the no-case-of-case
196 -- flag is on. Strictness changes evaluation order
197 -- and that can change full laziness
198 stricts | switchIsOn chkr NoCaseOfCase = vanilla_stricts
199 | otherwise = computed_stricts
201 go [] stricts False orig_cont
203 ----------------------------
206 go acc ss inl (ApplyTo _ arg@(Type _) se cont)
207 = go ((arg,se,False) : acc) ss inl cont
208 -- NB: don't bother to instantiate the function type
211 go acc (s:ss) inl (ApplyTo _ arg se cont)
212 = go ((arg,se,s) : acc) ss inl cont
214 -- An Inline continuation
215 go acc ss inl (InlinePlease cont)
216 = go acc ss True cont
218 -- We're run out of arguments, or else we've run out of demands
219 -- The latter only happens if the result is guaranteed bottom
220 -- This is the case for
221 -- * case (error "hello") of { ... }
222 -- * (error "Hello") arg
223 -- * f (error "Hello") where f is strict
225 -- Then, especially in the first of these cases, we'd like to discard
226 -- the continuation, leaving just the bottoming expression. But the
227 -- type might not be right, so we may have to add a coerce.
229 | null ss && discardableCont cont = (reverse acc, discardCont cont, inl)
230 | otherwise = (reverse acc, cont, inl)
232 ----------------------------
233 vanilla_stricts, computed_stricts :: [Bool]
234 vanilla_stricts = repeat False
235 computed_stricts = zipWith (||) fun_stricts arg_stricts
237 ----------------------------
238 (val_arg_tys, _) = splitFunTys (dropForAlls (idType fun))
239 arg_stricts = map isStrictType val_arg_tys ++ repeat False
240 -- These argument types are used as a cheap and cheerful way to find
241 -- unboxed arguments, which must be strict. But it's an InType
242 -- and so there might be a type variable where we expect a function
243 -- type (the substitution hasn't happened yet). And we don't bother
244 -- doing the type applications for a polymorphic function.
245 -- Hence the splitFunTys*IgnoringForAlls*
247 ----------------------------
248 -- If fun_stricts is finite, it means the function returns bottom
249 -- after that number of value args have been consumed
250 -- Otherwise it's infinite, extended with False
252 = case splitStrictSig (idNewStrictness fun) of
253 (demands, result_info)
254 | not (demands `lengthExceeds` countValArgs orig_cont)
255 -> -- Enough args, use the strictness given.
256 -- For bottoming functions we used to pretend that the arg
257 -- is lazy, so that we don't treat the arg as an
258 -- interesting context. This avoids substituting
259 -- top-level bindings for (say) strings into
260 -- calls to error. But now we are more careful about
261 -- inlining lone variables, so its ok (see SimplUtils.analyseCont)
262 if isBotRes result_info then
263 map isStrictDmd demands -- Finite => result is bottom
265 map isStrictDmd demands ++ vanilla_stricts
267 other -> vanilla_stricts -- Not enough args, or no strictness
270 interestingArg :: OutExpr -> Bool
271 -- An argument is interesting if it has *some* structure
272 -- We are here trying to avoid unfolding a function that
273 -- is applied only to variables that have no unfolding
274 -- (i.e. they are probably lambda bound): f x y z
275 -- There is little point in inlining f here.
276 interestingArg (Var v) = hasSomeUnfolding (idUnfolding v)
277 -- Was: isValueUnfolding (idUnfolding v')
278 -- But that seems over-pessimistic
280 -- This accounts for an argument like
281 -- () or [], which is definitely interesting
282 interestingArg (Type _) = False
283 interestingArg (App fn (Type _)) = interestingArg fn
284 interestingArg (Note _ a) = interestingArg a
285 interestingArg other = True
286 -- Consider let x = 3 in f x
287 -- The substitution will contain (x -> ContEx 3), and we want to
288 -- to say that x is an interesting argument.
289 -- But consider also (\x. f x y) y
290 -- The substitution will contain (x -> ContEx y), and we want to say
291 -- that x is not interesting (assuming y has no unfolding)
294 Comment about interestingCallContext
295 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
296 We want to avoid inlining an expression where there can't possibly be
297 any gain, such as in an argument position. Hence, if the continuation
298 is interesting (eg. a case scrutinee, application etc.) then we
299 inline, otherwise we don't.
301 Previously some_benefit used to return True only if the variable was
302 applied to some value arguments. This didn't work:
304 let x = _coerce_ (T Int) Int (I# 3) in
305 case _coerce_ Int (T Int) x of
308 we want to inline x, but can't see that it's a constructor in a case
309 scrutinee position, and some_benefit is False.
313 dMonadST = _/\_ t -> :Monad (g1 _@_ t, g2 _@_ t, g3 _@_ t)
315 .... case dMonadST _@_ x0 of (a,b,c) -> ....
317 we'd really like to inline dMonadST here, but we *don't* want to
318 inline if the case expression is just
320 case x of y { DEFAULT -> ... }
322 since we can just eliminate this case instead (x is in WHNF). Similar
323 applies when x is bound to a lambda expression. Hence
324 contIsInteresting looks for case expressions with just a single
328 interestingCallContext :: Bool -- False <=> no args at all
329 -> Bool -- False <=> no value args
331 -- The "lone-variable" case is important. I spent ages
332 -- messing about with unsatisfactory varaints, but this is nice.
333 -- The idea is that if a variable appear all alone
334 -- as an arg of lazy fn, or rhs Stop
335 -- as scrutinee of a case Select
336 -- as arg of a strict fn ArgOf
337 -- then we should not inline it (unless there is some other reason,
338 -- e.g. is is the sole occurrence). We achieve this by making
339 -- interestingCallContext return False for a lone variable.
341 -- Why? At least in the case-scrutinee situation, turning
342 -- let x = (a,b) in case x of y -> ...
344 -- let x = (a,b) in case (a,b) of y -> ...
346 -- let x = (a,b) in let y = (a,b) in ...
347 -- is bad if the binding for x will remain.
349 -- Another example: I discovered that strings
350 -- were getting inlined straight back into applications of 'error'
351 -- because the latter is strict.
353 -- f = \x -> ...(error s)...
355 -- Fundamentally such contexts should not ecourage inlining because
356 -- the context can ``see'' the unfolding of the variable (e.g. case or a RULE)
357 -- so there's no gain.
359 -- However, even a type application or coercion isn't a lone variable.
361 -- case $fMonadST @ RealWorld of { :DMonad a b c -> c }
362 -- We had better inline that sucker! The case won't see through it.
364 -- For now, I'm treating treating a variable applied to types
365 -- in a *lazy* context "lone". The motivating example was
367 -- g = /\a. \y. h (f a)
368 -- There's no advantage in inlining f here, and perhaps
369 -- a significant disadvantage. Hence some_val_args in the Stop case
371 interestingCallContext some_args some_val_args cont
374 interesting (InlinePlease _) = True
375 interesting (Select _ _ _ _ _) = some_args
376 interesting (ApplyTo _ _ _ _) = True -- Can happen if we have (coerce t (f x)) y
377 -- Perhaps True is a bit over-keen, but I've
378 -- seen (coerce f) x, where f has an INLINE prag,
379 -- So we have to give some motivaiton for inlining it
380 interesting (ArgOf _ _ _ _) = some_val_args
381 interesting (Stop ty _ upd_in_place) = some_val_args && upd_in_place
382 interesting (CoerceIt _ cont) = interesting cont
383 -- If this call is the arg of a strict function, the context
384 -- is a bit interesting. If we inline here, we may get useful
385 -- evaluation information to avoid repeated evals: e.g.
387 -- Here the contIsInteresting makes the '*' keener to inline,
388 -- which in turn exposes a constructor which makes the '+' inline.
389 -- Assuming that +,* aren't small enough to inline regardless.
391 -- It's also very important to inline in a strict context for things
394 -- Here, the context of (f x) is strict, and if f's unfolding is
395 -- a build it's *great* to inline it here. So we must ensure that
396 -- the context for (f x) is not totally uninteresting.
400 canUpdateInPlace :: Type -> Bool
401 -- Consider let x = <wurble> in ...
402 -- If <wurble> returns an explicit constructor, we might be able
403 -- to do update in place. So we treat even a thunk RHS context
404 -- as interesting if update in place is possible. We approximate
405 -- this by seeing if the type has a single constructor with a
406 -- small arity. But arity zero isn't good -- we share the single copy
407 -- for that case, so no point in sharing.
410 | not opt_UF_UpdateInPlace = False
412 = case splitTyConApp_maybe ty of
414 Just (tycon, _) -> case tyConDataCons_maybe tycon of
415 Just [dc] -> arity == 1 || arity == 2
417 arity = dataConRepArity dc
423 %************************************************************************
425 \section{Dealing with a single binder}
427 %************************************************************************
429 These functions are in the monad only so that they can be made strict via seq.
432 simplBinders :: SimplEnv -> [InBinder] -> SimplM (SimplEnv, [OutBinder])
433 simplBinders env bndrs
435 (subst', bndrs') = Subst.simplBndrs (getSubst env) bndrs
437 seqBndrs bndrs' `seq`
438 returnSmpl (setSubst env subst', bndrs')
440 simplBinder :: SimplEnv -> InBinder -> SimplM (SimplEnv, OutBinder)
443 (subst', bndr') = Subst.simplBndr (getSubst env) bndr
446 returnSmpl (setSubst env subst', bndr')
449 simplLetBndr :: SimplEnv -> InBinder -> SimplM (SimplEnv, OutBinder)
452 (subst', id') = Subst.simplLetId (getSubst env) id
455 returnSmpl (setSubst env subst', id')
457 simplLamBndrs, simplRecBndrs
458 :: SimplEnv -> [InBinder] -> SimplM (SimplEnv, [OutBinder])
459 simplRecBndrs = simplBndrs Subst.simplLetId
460 simplLamBndrs = simplBndrs Subst.simplLamBndr
462 simplBndrs simpl_bndr env bndrs
464 (subst', bndrs') = mapAccumL simpl_bndr (getSubst env) bndrs
466 seqBndrs bndrs' `seq`
467 returnSmpl (setSubst env subst', bndrs')
470 seqBndrs (b:bs) = seqBndr b `seq` seqBndrs bs
472 seqBndr b | isTyVar b = b `seq` ()
473 | otherwise = seqType (idType b) `seq`
480 newId :: EncodedFS -> Type -> SimplM Id
481 newId fs ty = getUniqueSmpl `thenSmpl` \ uniq ->
482 returnSmpl (mkSysLocal fs uniq ty)
486 %************************************************************************
488 \subsection{Rebuilding a lambda}
490 %************************************************************************
493 mkLam :: SimplEnv -> [OutBinder] -> OutExpr -> SimplCont -> SimplM FloatsWithExpr
497 a) eta reduction, if that gives a trivial expression
498 b) eta expansion [only if there are some value lambdas]
499 c) floating lets out through big lambdas
500 [only if all tyvar lambdas, and only if this lambda
504 mkLam env bndrs body cont
505 = getDOptsSmpl `thenSmpl` \dflags ->
506 mkLam' dflags env bndrs body cont
508 mkLam' dflags env bndrs body cont
509 | dopt Opt_DoEtaReduction dflags,
510 Just etad_lam <- tryEtaReduce bndrs body
511 = tick (EtaReduction (head bndrs)) `thenSmpl_`
512 returnSmpl (emptyFloats env, etad_lam)
514 | dopt Opt_DoLambdaEtaExpansion dflags,
515 any isRuntimeVar bndrs
516 = tryEtaExpansion body `thenSmpl` \ body' ->
517 returnSmpl (emptyFloats env, mkLams bndrs body')
519 {- Sept 01: I'm experimenting with getting the
520 full laziness pass to float out past big lambdsa
521 | all isTyVar bndrs, -- Only for big lambdas
522 contIsRhs cont -- Only try the rhs type-lambda floating
523 -- if this is indeed a right-hand side; otherwise
524 -- we end up floating the thing out, only for float-in
525 -- to float it right back in again!
526 = tryRhsTyLam env bndrs body `thenSmpl` \ (floats, body') ->
527 returnSmpl (floats, mkLams bndrs body')
531 = returnSmpl (emptyFloats env, mkLams bndrs body)
535 %************************************************************************
537 \subsection{Eta expansion and reduction}
539 %************************************************************************
541 We try for eta reduction here, but *only* if we get all the
542 way to an exprIsTrivial expression.
543 We don't want to remove extra lambdas unless we are going
544 to avoid allocating this thing altogether
547 tryEtaReduce :: [OutBinder] -> OutExpr -> Maybe OutExpr
548 tryEtaReduce bndrs body
549 -- We don't use CoreUtils.etaReduce, because we can be more
551 -- (a) we already have the binders
552 -- (b) we can do the triviality test before computing the free vars
553 = go (reverse bndrs) body
555 go (b : bs) (App fun arg) | ok_arg b arg = go bs fun -- Loop round
556 go [] fun | ok_fun fun = Just fun -- Success!
557 go _ _ = Nothing -- Failure!
559 ok_fun fun = exprIsTrivial fun
560 && not (any (`elemVarSet` (exprFreeVars fun)) bndrs)
561 && (exprIsValue fun || all ok_lam bndrs)
562 ok_lam v = isTyVar v || isDictTy (idType v)
563 -- The exprIsValue is because eta reduction is not
564 -- valid in general: \x. bot /= bot
565 -- So we need to be sure that the "fun" is a value.
567 -- However, we always want to reduce (/\a -> f a) to f
568 -- This came up in a RULE: foldr (build (/\a -> g a))
569 -- did not match foldr (build (/\b -> ...something complex...))
570 -- The type checker can insert these eta-expanded versions,
571 -- with both type and dictionary lambdas; hence the slightly
574 ok_arg b arg = varToCoreExpr b `cheapEqExpr` arg
578 Try eta expansion for RHSs
581 f = \x1..xn -> N ==> f = \x1..xn y1..ym -> N y1..ym
584 where (in both cases)
586 * The xi can include type variables
588 * The yi are all value variables
590 * N is a NORMAL FORM (i.e. no redexes anywhere)
591 wanting a suitable number of extra args.
593 We may have to sandwich some coerces between the lambdas
594 to make the types work. exprEtaExpandArity looks through coerces
595 when computing arity; and etaExpand adds the coerces as necessary when
596 actually computing the expansion.
599 tryEtaExpansion :: OutExpr -> SimplM OutExpr
600 -- There is at least one runtime binder in the binders
602 = getUniquesSmpl `thenSmpl` \ us ->
603 returnSmpl (etaExpand fun_arity us body (exprType body))
605 fun_arity = exprEtaExpandArity body
609 %************************************************************************
611 \subsection{Floating lets out of big lambdas}
613 %************************************************************************
615 tryRhsTyLam tries this transformation, when the big lambda appears as
616 the RHS of a let(rec) binding:
618 /\abc -> let(rec) x = e in b
620 let(rec) x' = /\abc -> let x = x' a b c in e
622 /\abc -> let x = x' a b c in b
624 This is good because it can turn things like:
626 let f = /\a -> letrec g = ... g ... in g
628 letrec g' = /\a -> ... g' a ...
632 which is better. In effect, it means that big lambdas don't impede
635 This optimisation is CRUCIAL in eliminating the junk introduced by
636 desugaring mutually recursive definitions. Don't eliminate it lightly!
638 So far as the implementation is concerned:
640 Invariant: go F e = /\tvs -> F e
644 = Let x' = /\tvs -> F e
648 G = F . Let x = x' tvs
650 go F (Letrec xi=ei in b)
651 = Letrec {xi' = /\tvs -> G ei}
655 G = F . Let {xi = xi' tvs}
657 [May 1999] If we do this transformation *regardless* then we can
658 end up with some pretty silly stuff. For example,
661 st = /\ s -> let { x1=r1 ; x2=r2 } in ...
666 st = /\s -> ...[y1 s/x1, y2 s/x2]
669 Unless the "..." is a WHNF there is really no point in doing this.
670 Indeed it can make things worse. Suppose x1 is used strictly,
673 x1* = case f y of { (a,b) -> e }
675 If we abstract this wrt the tyvar we then can't do the case inline
676 as we would normally do.
680 {- Trying to do this in full laziness
682 tryRhsTyLam :: SimplEnv -> [OutTyVar] -> OutExpr -> SimplM FloatsWithExpr
683 -- Call ensures that all the binders are type variables
685 tryRhsTyLam env tyvars body -- Only does something if there's a let
686 | not (all isTyVar tyvars)
687 || not (worth_it body) -- inside a type lambda,
688 = returnSmpl (emptyFloats env, body) -- and a WHNF inside that
691 = go env (\x -> x) body
694 worth_it e@(Let _ _) = whnf_in_middle e
697 whnf_in_middle (Let (NonRec x rhs) e) | isUnLiftedType (idType x) = False
698 whnf_in_middle (Let _ e) = whnf_in_middle e
699 whnf_in_middle e = exprIsCheap e
701 main_tyvar_set = mkVarSet tyvars
703 go env fn (Let bind@(NonRec var rhs) body)
705 = go env (fn . Let bind) body
707 go env fn (Let (NonRec var rhs) body)
708 = mk_poly tyvars_here var `thenSmpl` \ (var', rhs') ->
709 addAuxiliaryBind env (NonRec var' (mkLams tyvars_here (fn rhs))) $ \ env ->
710 go env (fn . Let (mk_silly_bind var rhs')) body
714 tyvars_here = varSetElems (main_tyvar_set `intersectVarSet` exprSomeFreeVars isTyVar rhs)
715 -- Abstract only over the type variables free in the rhs
716 -- wrt which the new binding is abstracted. But the naive
717 -- approach of abstract wrt the tyvars free in the Id's type
719 -- /\ a b -> let t :: (a,b) = (e1, e2)
722 -- Here, b isn't free in x's type, but we must nevertheless
723 -- abstract wrt b as well, because t's type mentions b.
724 -- Since t is floated too, we'd end up with the bogus:
725 -- poly_t = /\ a b -> (e1, e2)
726 -- poly_x = /\ a -> fst (poly_t a *b*)
727 -- So for now we adopt the even more naive approach of
728 -- abstracting wrt *all* the tyvars. We'll see if that
729 -- gives rise to problems. SLPJ June 98
731 go env fn (Let (Rec prs) body)
732 = mapAndUnzipSmpl (mk_poly tyvars_here) vars `thenSmpl` \ (vars', rhss') ->
734 gn body = fn (foldr Let body (zipWith mk_silly_bind vars rhss'))
735 pairs = vars' `zip` [mkLams tyvars_here (gn rhs) | rhs <- rhss]
737 addAuxiliaryBind env (Rec pairs) $ \ env ->
740 (vars,rhss) = unzip prs
741 tyvars_here = varSetElems (main_tyvar_set `intersectVarSet` exprsSomeFreeVars isTyVar (map snd prs))
742 -- See notes with tyvars_here above
744 go env fn body = returnSmpl (emptyFloats env, fn body)
746 mk_poly tyvars_here var
747 = getUniqueSmpl `thenSmpl` \ uniq ->
749 poly_name = setNameUnique (idName var) uniq -- Keep same name
750 poly_ty = mkForAllTys tyvars_here (idType var) -- But new type of course
751 poly_id = mkLocalId poly_name poly_ty
753 -- In the olden days, it was crucial to copy the occInfo of the original var,
754 -- because we were looking at occurrence-analysed but as yet unsimplified code!
755 -- In particular, we mustn't lose the loop breakers. BUT NOW we are looking
756 -- at already simplified code, so it doesn't matter
758 -- It's even right to retain single-occurrence or dead-var info:
759 -- Suppose we started with /\a -> let x = E in B
760 -- where x occurs once in B. Then we transform to:
761 -- let x' = /\a -> E in /\a -> let x* = x' a in B
762 -- where x* has an INLINE prag on it. Now, once x* is inlined,
763 -- the occurrences of x' will be just the occurrences originally
766 returnSmpl (poly_id, mkTyApps (Var poly_id) (mkTyVarTys tyvars_here))
768 mk_silly_bind var rhs = NonRec var (Note InlineMe rhs)
769 -- Suppose we start with:
771 -- x = /\ a -> let g = G in E
773 -- Then we'll float to get
775 -- x = let poly_g = /\ a -> G
776 -- in /\ a -> let g = poly_g a in E
778 -- But now the occurrence analyser will see just one occurrence
779 -- of poly_g, not inside a lambda, so the simplifier will
780 -- PreInlineUnconditionally poly_g back into g! Badk to square 1!
781 -- (I used to think that the "don't inline lone occurrences" stuff
782 -- would stop this happening, but since it's the *only* occurrence,
783 -- PreInlineUnconditionally kicks in first!)
785 -- Solution: put an INLINE note on g's RHS, so that poly_g seems
786 -- to appear many times. (NB: mkInlineMe eliminates
787 -- such notes on trivial RHSs, so do it manually.)
791 %************************************************************************
793 \subsection{Case alternative filtering
795 %************************************************************************
797 prepareAlts does two things:
799 1. Eliminate alternatives that cannot match, including the
802 2. If the DEFAULT alternative can match only one possible constructor,
803 then make that constructor explicit.
805 case e of x { DEFAULT -> rhs }
807 case e of x { (a,b) -> rhs }
808 where the type is a single constructor type. This gives better code
809 when rhs also scrutinises x or e.
811 It's a good idea do do this stuff before simplifying the alternatives, to
812 avoid simplifying alternatives we know can't happen, and to come up with
813 the list of constructors that are handled, to put into the IdInfo of the
814 case binder, for use when simplifying the alternatives.
816 Eliminating the default alternative in (1) isn't so obvious, but it can
819 data Colour = Red | Green | Blue
828 DEFAULT -> [ case y of ... ]
830 If we inline h into f, the default case of the inlined h can't happen.
831 If we don't notice this, we may end up filtering out *all* the cases
832 of the inner case y, which give us nowhere to go!
836 prepareAlts :: OutExpr -- Scrutinee
837 -> InId -- Case binder
839 -> SimplM ([InAlt], -- Better alternatives
840 [AltCon]) -- These cases are handled
842 prepareAlts scrut case_bndr alts
844 (alts_wo_default, maybe_deflt) = findDefault alts
846 impossible_cons = case scrut of
847 Var v -> otherCons (idUnfolding v)
850 -- Filter out alternatives that can't possibly match
851 better_alts | null impossible_cons = alts_wo_default
852 | otherwise = [alt | alt@(con,_,_) <- alts_wo_default,
853 not (con `elem` impossible_cons)]
855 -- "handled_cons" are handled either by the context,
856 -- or by a branch in this case expression
857 -- (Don't add DEFAULT to the handled_cons!!)
858 handled_cons = impossible_cons ++ [con | (con,_,_) <- better_alts]
860 -- Filter out the default, if it can't happen,
861 -- or replace it with "proper" alternative if there
862 -- is only one constructor left
863 prepareDefault case_bndr handled_cons maybe_deflt `thenSmpl` \ deflt_alt ->
865 returnSmpl (deflt_alt ++ better_alts, handled_cons)
867 prepareDefault case_bndr handled_cons (Just rhs)
868 | Just (tycon, inst_tys) <- splitTyConApp_maybe (idType case_bndr),
869 isAlgTyCon tycon, -- It's a data type, tuple, or unboxed tuples.
870 not (isNewTyCon tycon), -- We can have a newtype, if we are just doing an eval:
871 -- case x of { DEFAULT -> e }
872 -- and we don't want to fill in a default for them!
873 Just all_cons <- tyConDataCons_maybe tycon,
874 not (null all_cons), -- This is a tricky corner case. If the data type has no constructors,
875 -- which GHC allows, then the case expression will have at most a default
876 -- alternative. We don't want to eliminate that alternative, because the
877 -- invariant is that there's always one alternative. It's more convenient
879 -- case x of { DEFAULT -> e }
880 -- as it is, rather than transform it to
881 -- error "case cant match"
882 -- which would be quite legitmate. But it's a really obscure corner, and
883 -- not worth wasting code on.
884 let handled_data_cons = [data_con | DataAlt data_con <- handled_cons],
885 let missing_cons = [con | con <- all_cons,
886 not (con `elem` handled_data_cons)]
887 = case missing_cons of
888 [] -> returnSmpl [] -- Eliminate the default alternative
891 [con] -> -- It matches exactly one constructor, so fill it in
892 tick (FillInCaseDefault case_bndr) `thenSmpl_`
893 mk_args con inst_tys `thenSmpl` \ args ->
894 returnSmpl [(DataAlt con, args, rhs)]
896 two_or_more -> returnSmpl [(DEFAULT, [], rhs)]
899 = returnSmpl [(DEFAULT, [], rhs)]
901 prepareDefault case_bndr handled_cons Nothing
904 mk_args missing_con inst_tys
905 = getUniquesSmpl `thenSmpl` \ tv_uniqs ->
906 getUniquesSmpl `thenSmpl` \ id_uniqs ->
908 ex_tyvars = dataConExistentialTyVars missing_con
909 ex_tyvars' = zipWith mk tv_uniqs ex_tyvars
910 mk uniq tv = mkSysTyVar uniq (tyVarKind tv)
911 arg_tys = dataConArgTys missing_con (inst_tys ++ mkTyVarTys ex_tyvars')
912 arg_ids = zipWith (mkSysLocal FSLIT("a")) id_uniqs arg_tys
914 returnSmpl (ex_tyvars' ++ arg_ids)
918 %************************************************************************
920 \subsection{Case absorption and identity-case elimination}
922 %************************************************************************
924 mkCase puts a case expression back together, trying various transformations first.
927 mkCase :: OutExpr -> OutId -> [OutAlt] -> SimplM OutExpr
929 mkCase scrut case_bndr alts
930 = mkAlts scrut case_bndr alts `thenSmpl` \ better_alts ->
931 mkCase1 scrut case_bndr better_alts
935 mkAlts tries these things:
937 1. If several alternatives are identical, merge them into
938 a single DEFAULT alternative. I've occasionally seen this
939 making a big difference:
941 case e of =====> case e of
942 C _ -> f x D v -> ....v....
943 D v -> ....v.... DEFAULT -> f x
946 The point is that we merge common RHSs, at least for the DEFAULT case.
947 [One could do something more elaborate but I've never seen it needed.]
948 To avoid an expensive test, we just merge branches equal to the *first*
949 alternative; this picks up the common cases
950 a) all branches equal
951 b) some branches equal to the DEFAULT (which occurs first)
954 case e of b { ==> case e of b {
955 p1 -> rhs1 p1 -> rhs1
957 pm -> rhsm pm -> rhsm
958 _ -> case b of b' { pn -> let b'=b in rhsn
960 ... po -> let b'=b in rhso
961 po -> rhso _ -> let b'=b in rhsd
965 which merges two cases in one case when -- the default alternative of
966 the outer case scrutises the same variable as the outer case This
967 transformation is called Case Merging. It avoids that the same
968 variable is scrutinised multiple times.
971 The case where transformation (1) showed up was like this (lib/std/PrelCError.lhs):
977 where @is@ was something like
979 p `is` n = p /= (-1) && p == n
981 This gave rise to a horrible sequence of cases
988 and similarly in cascade for all the join points!
993 --------------------------------------------------
994 -- 1. Merge identical branches
995 --------------------------------------------------
996 mkAlts scrut case_bndr alts@((con1,bndrs1,rhs1) : con_alts)
997 | all isDeadBinder bndrs1, -- Remember the default
998 length filtered_alts < length con_alts -- alternative comes first
999 = tick (AltMerge case_bndr) `thenSmpl_`
1000 returnSmpl better_alts
1002 filtered_alts = filter keep con_alts
1003 keep (con,bndrs,rhs) = not (all isDeadBinder bndrs && rhs `cheapEqExpr` rhs1)
1004 better_alts = (DEFAULT, [], rhs1) : filtered_alts
1007 --------------------------------------------------
1008 -- 2. Merge nested cases
1009 --------------------------------------------------
1011 mkAlts scrut outer_bndr outer_alts
1012 = getDOptsSmpl `thenSmpl` \dflags ->
1013 mkAlts' dflags scrut outer_bndr outer_alts
1015 mkAlts' dflags scrut outer_bndr outer_alts
1016 | dopt Opt_CaseMerge dflags,
1017 (outer_alts_without_deflt, maybe_outer_deflt) <- findDefault outer_alts,
1018 Just (Case (Var scrut_var) inner_bndr inner_alts) <- maybe_outer_deflt,
1019 scruting_same_var scrut_var
1021 = let -- Eliminate any inner alts which are shadowed by the outer ones
1022 outer_cons = [con | (con,_,_) <- outer_alts_without_deflt]
1024 munged_inner_alts = [ (con, args, munge_rhs rhs)
1025 | (con, args, rhs) <- inner_alts,
1026 not (con `elem` outer_cons) -- Eliminate shadowed inner alts
1028 munge_rhs rhs = bindCaseBndr inner_bndr (Var outer_bndr) rhs
1030 (inner_con_alts, maybe_inner_default) = findDefault munged_inner_alts
1032 new_alts = add_default maybe_inner_default
1033 (outer_alts_without_deflt ++ inner_con_alts)
1035 tick (CaseMerge outer_bndr) `thenSmpl_`
1037 -- Warning: don't call mkAlts recursively!
1038 -- Firstly, there's no point, because inner alts have already had
1039 -- mkCase applied to them, so they won't have a case in their default
1040 -- Secondly, if you do, you get an infinite loop, because the bindCaseBndr
1041 -- in munge_rhs may put a case into the DEFAULT branch!
1043 -- We are scrutinising the same variable if it's
1044 -- the outer case-binder, or if the outer case scrutinises a variable
1045 -- (and it's the same). Testing both allows us not to replace the
1046 -- outer scrut-var with the outer case-binder (Simplify.simplCaseBinder).
1047 scruting_same_var = case scrut of
1048 Var outer_scrut -> \ v -> v == outer_bndr || v == outer_scrut
1049 other -> \ v -> v == outer_bndr
1051 add_default (Just rhs) alts = (DEFAULT,[],rhs) : alts
1052 add_default Nothing alts = alts
1055 --------------------------------------------------
1057 --------------------------------------------------
1059 mkAlts' dflags scrut case_bndr other_alts = returnSmpl other_alts
1064 =================================================================================
1066 mkCase1 tries these things
1068 1. Eliminate the case altogether if possible
1076 and similar friends.
1079 Start with a simple situation:
1081 case x# of ===> e[x#/y#]
1084 (when x#, y# are of primitive type, of course). We can't (in general)
1085 do this for algebraic cases, because we might turn bottom into
1088 Actually, we generalise this idea to look for a case where we're
1089 scrutinising a variable, and we know that only the default case can
1094 other -> ...(case x of
1098 Here the inner case can be eliminated. This really only shows up in
1099 eliminating error-checking code.
1101 We also make sure that we deal with this very common case:
1106 Here we are using the case as a strict let; if x is used only once
1107 then we want to inline it. We have to be careful that this doesn't
1108 make the program terminate when it would have diverged before, so we
1110 - x is used strictly, or
1111 - e is already evaluated (it may so if e is a variable)
1113 Lastly, we generalise the transformation to handle this:
1119 We only do this for very cheaply compared r's (constructors, literals
1120 and variables). If pedantic bottoms is on, we only do it when the
1121 scrutinee is a PrimOp which can't fail.
1123 We do it *here*, looking at un-simplified alternatives, because we
1124 have to check that r doesn't mention the variables bound by the
1125 pattern in each alternative, so the binder-info is rather useful.
1127 So the case-elimination algorithm is:
1129 1. Eliminate alternatives which can't match
1131 2. Check whether all the remaining alternatives
1132 (a) do not mention in their rhs any of the variables bound in their pattern
1133 and (b) have equal rhss
1135 3. Check we can safely ditch the case:
1136 * PedanticBottoms is off,
1137 or * the scrutinee is an already-evaluated variable
1138 or * the scrutinee is a primop which is ok for speculation
1139 -- ie we want to preserve divide-by-zero errors, and
1140 -- calls to error itself!
1142 or * [Prim cases] the scrutinee is a primitive variable
1144 or * [Alg cases] the scrutinee is a variable and
1145 either * the rhs is the same variable
1146 (eg case x of C a b -> x ===> x)
1147 or * there is only one alternative, the default alternative,
1148 and the binder is used strictly in its scope.
1149 [NB this is helped by the "use default binder where
1150 possible" transformation; see below.]
1153 If so, then we can replace the case with one of the rhss.
1155 Further notes about case elimination
1156 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
1157 Consider: test :: Integer -> IO ()
1160 Turns out that this compiles to:
1163 eta1 :: State# RealWorld ->
1164 case PrelNum.< eta PrelNum.zeroInteger of wild { __DEFAULT ->
1166 (PrelNum.jtos eta ($w[] @ Char))
1168 of wild1 { (# new_s, a4 #) -> PrelIO.lvl23 new_s }}
1170 Notice the strange '<' which has no effect at all. This is a funny one.
1171 It started like this:
1173 f x y = if x < 0 then jtos x
1174 else if y==0 then "" else jtos x
1176 At a particular call site we have (f v 1). So we inline to get
1178 if v < 0 then jtos x
1179 else if 1==0 then "" else jtos x
1181 Now simplify the 1==0 conditional:
1183 if v<0 then jtos v else jtos v
1185 Now common-up the two branches of the case:
1187 case (v<0) of DEFAULT -> jtos v
1189 Why don't we drop the case? Because it's strict in v. It's technically
1190 wrong to drop even unnecessary evaluations, and in practice they
1191 may be a result of 'seq' so we *definitely* don't want to drop those.
1192 I don't really know how to improve this situation.
1196 --------------------------------------------------
1197 -- 0. Check for empty alternatives
1198 --------------------------------------------------
1201 mkCase1 scrut case_bndr []
1202 = pprTrace "mkCase1: null alts" (ppr case_bndr <+> ppr scrut) $
1206 --------------------------------------------------
1207 -- 1. Eliminate the case altogether if poss
1208 --------------------------------------------------
1210 mkCase1 scrut case_bndr [(con,bndrs,rhs)]
1211 -- See if we can get rid of the case altogether
1212 -- See the extensive notes on case-elimination above
1213 -- mkCase made sure that if all the alternatives are equal,
1214 -- then there is now only one (DEFAULT) rhs
1215 | all isDeadBinder bndrs,
1217 -- Check that the scrutinee can be let-bound instead of case-bound
1218 exprOkForSpeculation scrut
1219 -- OK not to evaluate it
1220 -- This includes things like (==# a# b#)::Bool
1221 -- so that we simplify
1222 -- case ==# a# b# of { True -> x; False -> x }
1225 -- This particular example shows up in default methods for
1226 -- comparision operations (e.g. in (>=) for Int.Int32)
1227 || exprIsValue scrut -- It's already evaluated
1228 || var_demanded_later scrut -- It'll be demanded later
1230 -- || not opt_SimplPedanticBottoms) -- Or we don't care!
1231 -- We used to allow improving termination by discarding cases, unless -fpedantic-bottoms was on,
1232 -- but that breaks badly for the dataToTag# primop, which relies on a case to evaluate
1233 -- its argument: case x of { y -> dataToTag# y }
1234 -- Here we must *not* discard the case, because dataToTag# just fetches the tag from
1235 -- the info pointer. So we'll be pedantic all the time, and see if that gives any
1237 -- Also we don't want to discard 'seq's
1238 = tick (CaseElim case_bndr) `thenSmpl_`
1239 returnSmpl (bindCaseBndr case_bndr scrut rhs)
1242 -- The case binder is going to be evaluated later,
1243 -- and the scrutinee is a simple variable
1244 var_demanded_later (Var v) = isStrictDmd (idNewDemandInfo case_bndr)
1245 var_demanded_later other = False
1248 --------------------------------------------------
1250 --------------------------------------------------
1252 mkCase1 scrut case_bndr alts -- Identity case
1253 | all identity_alt alts
1254 = tick (CaseIdentity case_bndr) `thenSmpl_`
1255 returnSmpl (re_note scrut)
1257 identity_alt (con, args, rhs) = de_note rhs `cheapEqExpr` identity_rhs con args
1259 identity_rhs (DataAlt con) args = mkConApp con (arg_tys ++ map varToCoreExpr args)
1260 identity_rhs (LitAlt lit) _ = Lit lit
1261 identity_rhs DEFAULT _ = Var case_bndr
1263 arg_tys = map Type (tyConAppArgs (idType case_bndr))
1266 -- case coerce T e of x { _ -> coerce T' x }
1267 -- And we definitely want to eliminate this case!
1268 -- So we throw away notes from the RHS, and reconstruct
1269 -- (at least an approximation) at the other end
1270 de_note (Note _ e) = de_note e
1273 -- re_note wraps a coerce if it might be necessary
1274 re_note scrut = case head alts of
1275 (_,_,rhs1@(Note _ _)) -> mkCoerce2 (exprType rhs1) (idType case_bndr) scrut
1279 --------------------------------------------------
1281 --------------------------------------------------
1282 mkCase1 scrut bndr alts = returnSmpl (Case scrut bndr alts)
1286 When adding auxiliary bindings for the case binder, it's worth checking if
1287 its dead, because it often is, and occasionally these mkCase transformations
1288 cascade rather nicely.
1291 bindCaseBndr bndr rhs body
1292 | isDeadBinder bndr = body
1293 | otherwise = bindNonRec bndr rhs body