2 % (c) The AQUA Project, Glasgow University, 1993-1998
4 \section[SimplUtils]{The simplifier utilities}
8 simplBinder, simplBinders, simplRecIds, simplLetId, simplLamBinders,
11 -- The continuation type
12 SimplCont(..), DupFlag(..), LetRhsFlag(..),
13 contIsDupable, contResultType,
14 countValArgs, countArgs, pushContArgs,
15 mkBoringStop, mkStop, contIsRhs, contIsRhsOrArg,
16 getContArgs, interestingCallContext, interestingArg, isStrictType
20 #include "HsVersions.h"
22 import CmdLineOpts ( SimplifierSwitch(..),
23 opt_SimplDoLambdaEtaExpansion, opt_SimplDoEtaReduction,
24 opt_SimplCaseMerge, opt_UF_UpdateInPlace
27 import CoreUtils ( cheapEqExpr, exprType,
28 etaExpand, exprEtaExpandArity, bindNonRec, mkCoerce,
29 findDefault, exprOkForSpeculation, exprIsValue
31 import qualified Subst ( simplBndrs, simplBndr, simplLetId, simplLamBndr )
32 import Id ( Id, idType, idInfo,
33 mkSysLocal, hasNoBinding, isDeadBinder, idNewDemandInfo,
34 idUnfolding, idNewStrictness
36 import NewDemand ( isStrictDmd, isBotRes, splitStrictSig )
38 import Type ( Type, seqType,
39 splitTyConApp_maybe, tyConAppArgs, mkTyVarTys,
40 splitRepFunTys, isStrictType
42 import OccName ( UserFS )
43 import TyCon ( tyConDataConsIfAvailable, isDataTyCon )
44 import DataCon ( dataConRepArity, dataConSig, dataConArgTys )
45 import Var ( mkSysTyVar, tyVarKind )
46 import Util ( lengthExceeds, mapAccumL )
51 %************************************************************************
53 \subsection{The continuation data type}
55 %************************************************************************
58 data SimplCont -- Strict contexts
59 = Stop OutType -- Type of the result
61 Bool -- True <=> This is the RHS of a thunk whose type suggests
62 -- that update-in-place would be possible
63 -- (This makes the inliner a little keener.)
65 | CoerceIt OutType -- The To-type, simplified
68 | InlinePlease -- This continuation makes a function very
69 SimplCont -- keen to inline itelf
72 InExpr SimplEnv -- The argument, as yet unsimplified,
73 SimplCont -- and its environment
76 InId [InAlt] SimplEnv -- The case binder, alts, and subst-env
79 | ArgOf DupFlag -- An arbitrary strict context: the argument
80 -- of a strict function, or a primitive-arg fn
83 OutType -- cont_ty: the type of the expression being sought by the context
84 -- f (error "foo") ==> coerce t (error "foo")
86 -- We need to know the type t, to which to coerce.
87 (SimplEnv -> OutExpr -> SimplM FloatsWithExpr) -- What to do with the result
88 -- The result expression in the OutExprStuff has type cont_ty
90 data LetRhsFlag = AnArg -- It's just an argument not a let RHS
91 | AnRhs -- It's the RHS of a let (so please float lets out of big lambdas)
93 instance Outputable LetRhsFlag where
94 ppr AnArg = ptext SLIT("arg")
95 ppr AnRhs = ptext SLIT("rhs")
97 instance Outputable SimplCont where
98 ppr (Stop _ is_rhs _) = ptext SLIT("Stop") <> brackets (ppr is_rhs)
99 ppr (ApplyTo dup arg se cont) = (ptext SLIT("ApplyTo") <+> ppr dup <+> ppr arg) $$ ppr cont
100 ppr (ArgOf dup _ _ _) = ptext SLIT("ArgOf...") <+> ppr dup
101 ppr (Select dup bndr alts se cont) = (ptext SLIT("Select") <+> ppr dup <+> ppr bndr) $$
102 (nest 4 (ppr alts)) $$ ppr cont
103 ppr (CoerceIt ty cont) = (ptext SLIT("CoerceIt") <+> ppr ty) $$ ppr cont
104 ppr (InlinePlease cont) = ptext SLIT("InlinePlease") $$ ppr cont
106 data DupFlag = OkToDup | NoDup
108 instance Outputable DupFlag where
109 ppr OkToDup = ptext SLIT("ok")
110 ppr NoDup = ptext SLIT("nodup")
114 mkBoringStop :: OutType -> SimplCont
115 mkBoringStop ty = Stop ty AnArg (canUpdateInPlace ty)
117 mkStop :: OutType -> LetRhsFlag -> SimplCont
118 mkStop ty is_rhs = Stop ty is_rhs (canUpdateInPlace ty)
120 contIsRhs :: SimplCont -> Bool
121 contIsRhs (Stop _ AnRhs _) = True
122 contIsRhs (ArgOf _ AnRhs _ _) = True
123 contIsRhs other = False
125 contIsRhsOrArg (Stop _ _ _) = True
126 contIsRhsOrArg (ArgOf _ _ _ _) = True
127 contIsRhsOrArg other = False
130 contIsDupable :: SimplCont -> Bool
131 contIsDupable (Stop _ _ _) = True
132 contIsDupable (ApplyTo OkToDup _ _ _) = True
133 contIsDupable (ArgOf OkToDup _ _ _) = True
134 contIsDupable (Select OkToDup _ _ _ _) = True
135 contIsDupable (CoerceIt _ cont) = contIsDupable cont
136 contIsDupable (InlinePlease cont) = contIsDupable cont
137 contIsDupable other = False
140 discardableCont :: SimplCont -> Bool
141 discardableCont (Stop _ _ _) = False
142 discardableCont (CoerceIt _ cont) = discardableCont cont
143 discardableCont (InlinePlease cont) = discardableCont cont
144 discardableCont other = True
146 discardCont :: SimplCont -- A continuation, expecting
147 -> SimplCont -- Replace the continuation with a suitable coerce
148 discardCont cont = case cont of
149 Stop to_ty is_rhs _ -> cont
150 other -> CoerceIt to_ty (mkBoringStop to_ty)
152 to_ty = contResultType cont
155 contResultType :: SimplCont -> OutType
156 contResultType (Stop to_ty _ _) = to_ty
157 contResultType (ArgOf _ _ to_ty _) = to_ty
158 contResultType (ApplyTo _ _ _ cont) = contResultType cont
159 contResultType (CoerceIt _ cont) = contResultType cont
160 contResultType (InlinePlease cont) = contResultType cont
161 contResultType (Select _ _ _ _ cont) = contResultType cont
164 countValArgs :: SimplCont -> Int
165 countValArgs (ApplyTo _ (Type ty) se cont) = countValArgs cont
166 countValArgs (ApplyTo _ val_arg se cont) = 1 + countValArgs cont
167 countValArgs other = 0
169 countArgs :: SimplCont -> Int
170 countArgs (ApplyTo _ arg se cont) = 1 + countArgs cont
174 pushContArgs :: SimplEnv -> [OutArg] -> SimplCont -> SimplCont
175 -- Pushes args with the specified environment
176 pushContArgs env [] cont = cont
177 pushContArgs env (arg : args) cont = ApplyTo NoDup arg env (pushContArgs env args cont)
182 getContArgs :: SwitchChecker
183 -> OutId -> SimplCont
184 -> ([(InExpr, SimplEnv, Bool)], -- Arguments; the Bool is true for strict args
185 SimplCont, -- Remaining continuation
186 Bool) -- Whether we came across an InlineCall
187 -- getContArgs id k = (args, k', inl)
188 -- args are the leading ApplyTo items in k
189 -- (i.e. outermost comes first)
190 -- augmented with demand info from the functionn
191 getContArgs chkr fun orig_cont
193 -- Ignore strictness info if the no-case-of-case
194 -- flag is on. Strictness changes evaluation order
195 -- and that can change full laziness
196 stricts | switchIsOn chkr NoCaseOfCase = vanilla_stricts
197 | otherwise = computed_stricts
199 go [] stricts False orig_cont
201 ----------------------------
204 go acc ss inl (ApplyTo _ arg@(Type _) se cont)
205 = go ((arg,se,False) : acc) ss inl cont
206 -- NB: don't bother to instantiate the function type
209 go acc (s:ss) inl (ApplyTo _ arg se cont)
210 = go ((arg,se,s) : acc) ss inl cont
212 -- An Inline continuation
213 go acc ss inl (InlinePlease cont)
214 = go acc ss True cont
216 -- We're run out of arguments, or else we've run out of demands
217 -- The latter only happens if the result is guaranteed bottom
218 -- This is the case for
219 -- * case (error "hello") of { ... }
220 -- * (error "Hello") arg
221 -- * f (error "Hello") where f is strict
224 | null ss && discardableCont cont = (reverse acc, discardCont cont, inl)
225 | otherwise = (reverse acc, cont, inl)
227 ----------------------------
228 vanilla_stricts, computed_stricts :: [Bool]
229 vanilla_stricts = repeat False
230 computed_stricts = zipWith (||) fun_stricts arg_stricts
232 ----------------------------
233 (val_arg_tys, _) = splitRepFunTys (idType fun)
234 arg_stricts = map isStrictType val_arg_tys ++ repeat False
235 -- These argument types are used as a cheap and cheerful way to find
236 -- unboxed arguments, which must be strict. But it's an InType
237 -- and so there might be a type variable where we expect a function
238 -- type (the substitution hasn't happened yet). And we don't bother
239 -- doing the type applications for a polymorphic function.
240 -- Hence the split*Rep*FunTys
242 ----------------------------
243 -- If fun_stricts is finite, it means the function returns bottom
244 -- after that number of value args have been consumed
245 -- Otherwise it's infinite, extended with False
247 = case splitStrictSig (idNewStrictness fun) of
248 (demands, result_info)
249 | not (demands `lengthExceeds` countValArgs orig_cont)
250 -> -- Enough args, use the strictness given.
251 -- For bottoming functions we used to pretend that the arg
252 -- is lazy, so that we don't treat the arg as an
253 -- interesting context. This avoids substituting
254 -- top-level bindings for (say) strings into
255 -- calls to error. But now we are more careful about
256 -- inlining lone variables, so its ok (see SimplUtils.analyseCont)
257 if isBotRes result_info then
258 map isStrictDmd demands -- Finite => result is bottom
260 map isStrictDmd demands ++ vanilla_stricts
262 other -> vanilla_stricts -- Not enough args, or no strictness
265 interestingArg :: OutExpr -> Bool
266 -- An argument is interesting if it has *some* structure
267 -- We are here trying to avoid unfolding a function that
268 -- is applied only to variables that have no unfolding
269 -- (i.e. they are probably lambda bound): f x y z
270 -- There is little point in inlining f here.
271 interestingArg (Var v) = hasSomeUnfolding (idUnfolding v)
272 -- Was: isValueUnfolding (idUnfolding v')
273 -- But that seems over-pessimistic
274 interestingArg (Type _) = False
275 interestingArg (App fn (Type _)) = interestingArg fn
276 interestingArg (Note _ a) = interestingArg a
277 interestingArg other = True
278 -- Consider let x = 3 in f x
279 -- The substitution will contain (x -> ContEx 3), and we want to
280 -- to say that x is an interesting argument.
281 -- But consider also (\x. f x y) y
282 -- The substitution will contain (x -> ContEx y), and we want to say
283 -- that x is not interesting (assuming y has no unfolding)
286 Comment about interestingCallContext
287 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
288 We want to avoid inlining an expression where there can't possibly be
289 any gain, such as in an argument position. Hence, if the continuation
290 is interesting (eg. a case scrutinee, application etc.) then we
291 inline, otherwise we don't.
293 Previously some_benefit used to return True only if the variable was
294 applied to some value arguments. This didn't work:
296 let x = _coerce_ (T Int) Int (I# 3) in
297 case _coerce_ Int (T Int) x of
300 we want to inline x, but can't see that it's a constructor in a case
301 scrutinee position, and some_benefit is False.
305 dMonadST = _/\_ t -> :Monad (g1 _@_ t, g2 _@_ t, g3 _@_ t)
307 .... case dMonadST _@_ x0 of (a,b,c) -> ....
309 we'd really like to inline dMonadST here, but we *don't* want to
310 inline if the case expression is just
312 case x of y { DEFAULT -> ... }
314 since we can just eliminate this case instead (x is in WHNF). Similar
315 applies when x is bound to a lambda expression. Hence
316 contIsInteresting looks for case expressions with just a single
320 interestingCallContext :: Bool -- False <=> no args at all
321 -> Bool -- False <=> no value args
323 -- The "lone-variable" case is important. I spent ages
324 -- messing about with unsatisfactory varaints, but this is nice.
325 -- The idea is that if a variable appear all alone
326 -- as an arg of lazy fn, or rhs Stop
327 -- as scrutinee of a case Select
328 -- as arg of a strict fn ArgOf
329 -- then we should not inline it (unless there is some other reason,
330 -- e.g. is is the sole occurrence). We achieve this by making
331 -- interestingCallContext return False for a lone variable.
333 -- Why? At least in the case-scrutinee situation, turning
334 -- let x = (a,b) in case x of y -> ...
336 -- let x = (a,b) in case (a,b) of y -> ...
338 -- let x = (a,b) in let y = (a,b) in ...
339 -- is bad if the binding for x will remain.
341 -- Another example: I discovered that strings
342 -- were getting inlined straight back into applications of 'error'
343 -- because the latter is strict.
345 -- f = \x -> ...(error s)...
347 -- Fundamentally such contexts should not ecourage inlining because
348 -- the context can ``see'' the unfolding of the variable (e.g. case or a RULE)
349 -- so there's no gain.
351 -- However, even a type application or coercion isn't a lone variable.
353 -- case $fMonadST @ RealWorld of { :DMonad a b c -> c }
354 -- We had better inline that sucker! The case won't see through it.
356 -- For now, I'm treating treating a variable applied to types
357 -- in a *lazy* context "lone". The motivating example was
359 -- g = /\a. \y. h (f a)
360 -- There's no advantage in inlining f here, and perhaps
361 -- a significant disadvantage. Hence some_val_args in the Stop case
363 interestingCallContext some_args some_val_args cont
366 interesting (InlinePlease _) = True
367 interesting (Select _ _ _ _ _) = some_args
368 interesting (ApplyTo _ _ _ _) = True -- Can happen if we have (coerce t (f x)) y
369 -- Perhaps True is a bit over-keen, but I've
370 -- seen (coerce f) x, where f has an INLINE prag,
371 -- So we have to give some motivaiton for inlining it
372 interesting (ArgOf _ _ _ _) = some_val_args
373 interesting (Stop ty _ upd_in_place) = some_val_args && upd_in_place
374 interesting (CoerceIt _ cont) = interesting cont
375 -- If this call is the arg of a strict function, the context
376 -- is a bit interesting. If we inline here, we may get useful
377 -- evaluation information to avoid repeated evals: e.g.
379 -- Here the contIsInteresting makes the '*' keener to inline,
380 -- which in turn exposes a constructor which makes the '+' inline.
381 -- Assuming that +,* aren't small enough to inline regardless.
383 -- It's also very important to inline in a strict context for things
386 -- Here, the context of (f x) is strict, and if f's unfolding is
387 -- a build it's *great* to inline it here. So we must ensure that
388 -- the context for (f x) is not totally uninteresting.
392 canUpdateInPlace :: Type -> Bool
393 -- Consider let x = <wurble> in ...
394 -- If <wurble> returns an explicit constructor, we might be able
395 -- to do update in place. So we treat even a thunk RHS context
396 -- as interesting if update in place is possible. We approximate
397 -- this by seeing if the type has a single constructor with a
398 -- small arity. But arity zero isn't good -- we share the single copy
399 -- for that case, so no point in sharing.
402 | not opt_UF_UpdateInPlace = False
404 = case splitTyConApp_maybe ty of
406 Just (tycon, _) -> case tyConDataConsIfAvailable tycon of
407 [dc] -> arity == 1 || arity == 2
409 arity = dataConRepArity dc
415 %************************************************************************
417 \section{Dealing with a single binder}
419 %************************************************************************
421 These functions are in the monad only so that they can be made strict via seq.
424 simplBinders :: SimplEnv -> [InBinder] -> SimplM (SimplEnv, [OutBinder])
425 simplBinders env bndrs
427 (subst', bndrs') = Subst.simplBndrs (getSubst env) bndrs
429 seqBndrs bndrs' `seq`
430 returnSmpl (setSubst env subst', bndrs')
432 simplBinder :: SimplEnv -> InBinder -> SimplM (SimplEnv, OutBinder)
435 (subst', bndr') = Subst.simplBndr (getSubst env) bndr
438 returnSmpl (setSubst env subst', bndr')
441 simplLamBinders :: SimplEnv -> [InBinder] -> SimplM (SimplEnv, [OutBinder])
442 simplLamBinders env bndrs
444 (subst', bndrs') = mapAccumL Subst.simplLamBndr (getSubst env) bndrs
446 seqBndrs bndrs' `seq`
447 returnSmpl (setSubst env subst', bndrs')
449 simplRecIds :: SimplEnv -> [InBinder] -> SimplM (SimplEnv, [OutBinder])
452 (subst', ids') = mapAccumL Subst.simplLetId (getSubst env) ids
455 returnSmpl (setSubst env subst', ids')
457 simplLetId :: SimplEnv -> InBinder -> SimplM (SimplEnv, OutBinder)
460 (subst', id') = Subst.simplLetId (getSubst env) id
463 returnSmpl (setSubst env subst', id')
466 seqBndrs (b:bs) = seqBndr b `seq` seqBndrs bs
468 seqBndr b | isTyVar b = b `seq` ()
469 | otherwise = seqType (idType b) `seq`
476 newId :: UserFS -> Type -> SimplM Id
477 newId fs ty = getUniqueSmpl `thenSmpl` \ uniq ->
478 returnSmpl (mkSysLocal fs uniq ty)
482 %************************************************************************
484 \subsection{Rebuilding a lambda}
486 %************************************************************************
489 mkLam :: SimplEnv -> [OutBinder] -> OutExpr -> SimplCont -> SimplM FloatsWithExpr
493 a) eta reduction, if that gives a trivial expression
494 b) eta expansion [only if there are some value lambdas]
495 c) floating lets out through big lambdas
496 [only if all tyvar lambdas, and only if this lambda
500 mkLam env bndrs body cont
501 | opt_SimplDoEtaReduction,
502 Just etad_lam <- tryEtaReduce bndrs body
503 = tick (EtaReduction (head bndrs)) `thenSmpl_`
504 returnSmpl (emptyFloats env, etad_lam)
506 | opt_SimplDoLambdaEtaExpansion,
507 any isRuntimeVar bndrs
508 = tryEtaExpansion body `thenSmpl` \ body' ->
509 returnSmpl (emptyFloats env, mkLams bndrs body')
511 {- Sept 01: I'm experimenting with getting the
512 full laziness pass to float out past big lambdsa
513 | all isTyVar bndrs, -- Only for big lambdas
514 contIsRhs cont -- Only try the rhs type-lambda floating
515 -- if this is indeed a right-hand side; otherwise
516 -- we end up floating the thing out, only for float-in
517 -- to float it right back in again!
518 = tryRhsTyLam env bndrs body `thenSmpl` \ (floats, body') ->
519 returnSmpl (floats, mkLams bndrs body')
523 = returnSmpl (emptyFloats env, mkLams bndrs body)
527 %************************************************************************
529 \subsection{Eta expansion and reduction}
531 %************************************************************************
533 We try for eta reduction here, but *only* if we get all the
534 way to an exprIsTrivial expression.
535 We don't want to remove extra lambdas unless we are going
536 to avoid allocating this thing altogether
539 tryEtaReduce :: [OutBinder] -> OutExpr -> Maybe OutExpr
540 tryEtaReduce bndrs body
541 -- We don't use CoreUtils.etaReduce, because we can be more
543 -- (a) we already have the binders
544 -- (b) we can do the triviality test before computing the free vars
545 -- [in fact I take the simple path and look for just a variable]
546 = go (reverse bndrs) body
548 go (b : bs) (App fun arg) | ok_arg b arg = go bs fun -- Loop round
549 go [] (Var fun) | ok_fun fun = Just (Var fun) -- Success!
550 go _ _ = Nothing -- Failure!
552 ok_fun fun = not (fun `elem` bndrs) && not (hasNoBinding fun)
553 -- Note the awkward "hasNoBinding" test
554 -- Details with exprIsTrivial
555 ok_arg b arg = varToCoreExpr b `cheapEqExpr` arg
559 Try eta expansion for RHSs
562 f = \x1..xn -> N ==> f = \x1..xn y1..ym -> N y1..ym
565 where (in both cases)
567 * The xi can include type variables
569 * The yi are all value variables
571 * N is a NORMAL FORM (i.e. no redexes anywhere)
572 wanting a suitable number of extra args.
574 We may have to sandwich some coerces between the lambdas
575 to make the types work. exprEtaExpandArity looks through coerces
576 when computing arity; and etaExpand adds the coerces as necessary when
577 actually computing the expansion.
580 tryEtaExpansion :: OutExpr -> SimplM OutExpr
581 -- There is at least one runtime binder in the binders
583 = getUniquesSmpl `thenSmpl` \ us ->
584 returnSmpl (etaExpand fun_arity us body (exprType body))
586 fun_arity = exprEtaExpandArity body
590 %************************************************************************
592 \subsection{Floating lets out of big lambdas}
594 %************************************************************************
596 tryRhsTyLam tries this transformation, when the big lambda appears as
597 the RHS of a let(rec) binding:
599 /\abc -> let(rec) x = e in b
601 let(rec) x' = /\abc -> let x = x' a b c in e
603 /\abc -> let x = x' a b c in b
605 This is good because it can turn things like:
607 let f = /\a -> letrec g = ... g ... in g
609 letrec g' = /\a -> ... g' a ...
613 which is better. In effect, it means that big lambdas don't impede
616 This optimisation is CRUCIAL in eliminating the junk introduced by
617 desugaring mutually recursive definitions. Don't eliminate it lightly!
619 So far as the implementation is concerned:
621 Invariant: go F e = /\tvs -> F e
625 = Let x' = /\tvs -> F e
629 G = F . Let x = x' tvs
631 go F (Letrec xi=ei in b)
632 = Letrec {xi' = /\tvs -> G ei}
636 G = F . Let {xi = xi' tvs}
638 [May 1999] If we do this transformation *regardless* then we can
639 end up with some pretty silly stuff. For example,
642 st = /\ s -> let { x1=r1 ; x2=r2 } in ...
647 st = /\s -> ...[y1 s/x1, y2 s/x2]
650 Unless the "..." is a WHNF there is really no point in doing this.
651 Indeed it can make things worse. Suppose x1 is used strictly,
654 x1* = case f y of { (a,b) -> e }
656 If we abstract this wrt the tyvar we then can't do the case inline
657 as we would normally do.
661 {- Trying to do this in full laziness
663 tryRhsTyLam :: SimplEnv -> [OutTyVar] -> OutExpr -> SimplM FloatsWithExpr
664 -- Call ensures that all the binders are type variables
666 tryRhsTyLam env tyvars body -- Only does something if there's a let
667 | not (all isTyVar tyvars)
668 || not (worth_it body) -- inside a type lambda,
669 = returnSmpl (emptyFloats env, body) -- and a WHNF inside that
672 = go env (\x -> x) body
675 worth_it e@(Let _ _) = whnf_in_middle e
678 whnf_in_middle (Let (NonRec x rhs) e) | isUnLiftedType (idType x) = False
679 whnf_in_middle (Let _ e) = whnf_in_middle e
680 whnf_in_middle e = exprIsCheap e
682 main_tyvar_set = mkVarSet tyvars
684 go env fn (Let bind@(NonRec var rhs) body)
686 = go env (fn . Let bind) body
688 go env fn (Let (NonRec var rhs) body)
689 = mk_poly tyvars_here var `thenSmpl` \ (var', rhs') ->
690 addAuxiliaryBind env (NonRec var' (mkLams tyvars_here (fn rhs))) $ \ env ->
691 go env (fn . Let (mk_silly_bind var rhs')) body
695 tyvars_here = varSetElems (main_tyvar_set `intersectVarSet` exprSomeFreeVars isTyVar rhs)
696 -- Abstract only over the type variables free in the rhs
697 -- wrt which the new binding is abstracted. But the naive
698 -- approach of abstract wrt the tyvars free in the Id's type
700 -- /\ a b -> let t :: (a,b) = (e1, e2)
703 -- Here, b isn't free in x's type, but we must nevertheless
704 -- abstract wrt b as well, because t's type mentions b.
705 -- Since t is floated too, we'd end up with the bogus:
706 -- poly_t = /\ a b -> (e1, e2)
707 -- poly_x = /\ a -> fst (poly_t a *b*)
708 -- So for now we adopt the even more naive approach of
709 -- abstracting wrt *all* the tyvars. We'll see if that
710 -- gives rise to problems. SLPJ June 98
712 go env fn (Let (Rec prs) body)
713 = mapAndUnzipSmpl (mk_poly tyvars_here) vars `thenSmpl` \ (vars', rhss') ->
715 gn body = fn (foldr Let body (zipWith mk_silly_bind vars rhss'))
716 pairs = vars' `zip` [mkLams tyvars_here (gn rhs) | rhs <- rhss]
718 addAuxiliaryBind env (Rec pairs) $ \ env ->
721 (vars,rhss) = unzip prs
722 tyvars_here = varSetElems (main_tyvar_set `intersectVarSet` exprsSomeFreeVars isTyVar (map snd prs))
723 -- See notes with tyvars_here above
725 go env fn body = returnSmpl (emptyFloats env, fn body)
727 mk_poly tyvars_here var
728 = getUniqueSmpl `thenSmpl` \ uniq ->
730 poly_name = setNameUnique (idName var) uniq -- Keep same name
731 poly_ty = mkForAllTys tyvars_here (idType var) -- But new type of course
732 poly_id = mkLocalId poly_name poly_ty
734 -- In the olden days, it was crucial to copy the occInfo of the original var,
735 -- because we were looking at occurrence-analysed but as yet unsimplified code!
736 -- In particular, we mustn't lose the loop breakers. BUT NOW we are looking
737 -- at already simplified code, so it doesn't matter
739 -- It's even right to retain single-occurrence or dead-var info:
740 -- Suppose we started with /\a -> let x = E in B
741 -- where x occurs once in B. Then we transform to:
742 -- let x' = /\a -> E in /\a -> let x* = x' a in B
743 -- where x* has an INLINE prag on it. Now, once x* is inlined,
744 -- the occurrences of x' will be just the occurrences originally
747 returnSmpl (poly_id, mkTyApps (Var poly_id) (mkTyVarTys tyvars_here))
749 mk_silly_bind var rhs = NonRec var (Note InlineMe rhs)
750 -- Suppose we start with:
752 -- x = /\ a -> let g = G in E
754 -- Then we'll float to get
756 -- x = let poly_g = /\ a -> G
757 -- in /\ a -> let g = poly_g a in E
759 -- But now the occurrence analyser will see just one occurrence
760 -- of poly_g, not inside a lambda, so the simplifier will
761 -- PreInlineUnconditionally poly_g back into g! Badk to square 1!
762 -- (I used to think that the "don't inline lone occurrences" stuff
763 -- would stop this happening, but since it's the *only* occurrence,
764 -- PreInlineUnconditionally kicks in first!)
766 -- Solution: put an INLINE note on g's RHS, so that poly_g seems
767 -- to appear many times. (NB: mkInlineMe eliminates
768 -- such notes on trivial RHSs, so do it manually.)
773 %************************************************************************
775 \subsection{Case absorption and identity-case elimination}
777 %************************************************************************
779 mkCase puts a case expression back together, trying various transformations first.
782 mkCase :: OutExpr -> OutId -> [OutAlt] -> SimplM OutExpr
784 mkCase scrut case_bndr alts
785 = mkAlts scrut case_bndr alts `thenSmpl` \ better_alts ->
786 mkCase1 scrut case_bndr better_alts
790 mkAlts tries these things:
792 1. If several alternatives are identical, merge them into
793 a single DEFAULT alternative. I've occasionally seen this
794 making a big difference:
796 case e of =====> case e of
797 C _ -> f x D v -> ....v....
798 D v -> ....v.... DEFAULT -> f x
801 The point is that we merge common RHSs, at least for the DEFAULT case.
802 [One could do something more elaborate but I've never seen it needed.]
803 To avoid an expensive test, we just merge branches equal to the *first*
804 alternative; this picks up the common cases
805 a) all branches equal
806 b) some branches equal to the DEFAULT (which occurs first)
808 2. If the DEFAULT alternative can match only one possible constructor,
809 then make that constructor explicit.
811 case e of x { DEFAULT -> rhs }
813 case e of x { (a,b) -> rhs }
814 where the type is a single constructor type. This gives better code
815 when rhs also scrutinises x or e.
818 case e of b { ==> case e of b {
819 p1 -> rhs1 p1 -> rhs1
821 pm -> rhsm pm -> rhsm
822 _ -> case b of b' { pn -> let b'=b in rhsn
824 ... po -> let b'=b in rhso
825 po -> rhso _ -> let b'=b in rhsd
829 which merges two cases in one case when -- the default alternative of
830 the outer case scrutises the same variable as the outer case This
831 transformation is called Case Merging. It avoids that the same
832 variable is scrutinised multiple times.
835 The case where transformation (1) showed up was like this (lib/std/PrelCError.lhs):
841 where @is@ was something like
843 p `is` n = p /= (-1) && p == n
845 This gave rise to a horrible sequence of cases
852 and similarly in cascade for all the join points!
857 --------------------------------------------------
858 -- 1. Merge identical branches
859 --------------------------------------------------
860 mkAlts scrut case_bndr alts@((con1,bndrs1,rhs1) : con_alts)
861 | all isDeadBinder bndrs1, -- Remember the default
862 length filtered_alts < length con_alts -- alternative comes first
863 = tick (AltMerge case_bndr) `thenSmpl_`
864 returnSmpl better_alts
866 filtered_alts = filter keep con_alts
867 keep (con,bndrs,rhs) = not (all isDeadBinder bndrs && rhs `cheapEqExpr` rhs1)
868 better_alts = (DEFAULT, [], rhs1) : filtered_alts
871 --------------------------------------------------
872 -- 2. Fill in missing constructor
873 --------------------------------------------------
875 mkAlts scrut case_bndr alts
876 | Just (tycon, inst_tys) <- splitTyConApp_maybe (idType case_bndr),
877 isDataTyCon tycon, -- It's a data type
878 (alts_no_deflt, Just rhs) <- findDefault alts,
879 -- There is a DEFAULT case
880 [missing_con] <- filter is_missing (tyConDataConsIfAvailable tycon)
881 -- There is just one missing constructor!
882 = tick (FillInCaseDefault case_bndr) `thenSmpl_`
883 getUniquesSmpl `thenSmpl` \ tv_uniqs ->
884 getUniquesSmpl `thenSmpl` \ id_uniqs ->
886 (_,_,ex_tyvars,_,_,_) = dataConSig missing_con
887 ex_tyvars' = zipWith mk tv_uniqs ex_tyvars
888 mk uniq tv = mkSysTyVar uniq (tyVarKind tv)
889 arg_ids = zipWith (mkSysLocal SLIT("a")) id_uniqs arg_tys
890 arg_tys = dataConArgTys missing_con (inst_tys ++ mkTyVarTys ex_tyvars')
891 better_alts = (DataAlt missing_con, ex_tyvars' ++ arg_ids, rhs) : alts_no_deflt
893 returnSmpl better_alts
895 impossible_cons = otherCons (idUnfolding case_bndr)
896 handled_data_cons = [data_con | DataAlt data_con <- impossible_cons] ++
897 [data_con | (DataAlt data_con, _, _) <- alts]
898 is_missing con = not (con `elem` handled_data_cons)
900 --------------------------------------------------
901 -- 3. Merge nested cases
902 --------------------------------------------------
904 mkAlts scrut outer_bndr outer_alts
905 | opt_SimplCaseMerge,
906 (outer_alts_without_deflt, maybe_outer_deflt) <- findDefault outer_alts,
907 Just (Case (Var scrut_var) inner_bndr inner_alts) <- maybe_outer_deflt,
908 scruting_same_var scrut_var
910 = let -- Eliminate any inner alts which are shadowed by the outer ones
911 outer_cons = [con | (con,_,_) <- outer_alts_without_deflt]
913 munged_inner_alts = [ (con, args, munge_rhs rhs)
914 | (con, args, rhs) <- inner_alts,
915 not (con `elem` outer_cons) -- Eliminate shadowed inner alts
917 munge_rhs rhs = bindCaseBndr inner_bndr (Var outer_bndr) rhs
919 (inner_con_alts, maybe_inner_default) = findDefault munged_inner_alts
921 new_alts = add_default maybe_inner_default
922 (outer_alts_without_deflt ++ inner_con_alts)
924 tick (CaseMerge outer_bndr) `thenSmpl_`
926 -- Warning: don't call mkAlts recursively!
927 -- Firstly, there's no point, because inner alts have already had
928 -- mkCase applied to them, so they won't have a case in their default
929 -- Secondly, if you do, you get an infinite loop, because the bindCaseBndr
930 -- in munge_rhs may put a case into the DEFAULT branch!
932 -- We are scrutinising the same variable if it's
933 -- the outer case-binder, or if the outer case scrutinises a variable
934 -- (and it's the same). Testing both allows us not to replace the
935 -- outer scrut-var with the outer case-binder (Simplify.simplCaseBinder).
936 scruting_same_var = case scrut of
937 Var outer_scrut -> \ v -> v == outer_bndr || v == outer_scrut
938 other -> \ v -> v == outer_bndr
940 add_default (Just rhs) alts = (DEFAULT,[],rhs) : alts
941 add_default Nothing alts = alts
944 --------------------------------------------------
946 --------------------------------------------------
948 mkAlts scrut case_bndr other_alts = returnSmpl other_alts
953 =================================================================================
955 mkCase1 tries these things
957 1. Eliminate the case altogether if possible
968 Start with a simple situation:
970 case x# of ===> e[x#/y#]
973 (when x#, y# are of primitive type, of course). We can't (in general)
974 do this for algebraic cases, because we might turn bottom into
977 Actually, we generalise this idea to look for a case where we're
978 scrutinising a variable, and we know that only the default case can
983 other -> ...(case x of
987 Here the inner case can be eliminated. This really only shows up in
988 eliminating error-checking code.
990 We also make sure that we deal with this very common case:
995 Here we are using the case as a strict let; if x is used only once
996 then we want to inline it. We have to be careful that this doesn't
997 make the program terminate when it would have diverged before, so we
999 - x is used strictly, or
1000 - e is already evaluated (it may so if e is a variable)
1002 Lastly, we generalise the transformation to handle this:
1008 We only do this for very cheaply compared r's (constructors, literals
1009 and variables). If pedantic bottoms is on, we only do it when the
1010 scrutinee is a PrimOp which can't fail.
1012 We do it *here*, looking at un-simplified alternatives, because we
1013 have to check that r doesn't mention the variables bound by the
1014 pattern in each alternative, so the binder-info is rather useful.
1016 So the case-elimination algorithm is:
1018 1. Eliminate alternatives which can't match
1020 2. Check whether all the remaining alternatives
1021 (a) do not mention in their rhs any of the variables bound in their pattern
1022 and (b) have equal rhss
1024 3. Check we can safely ditch the case:
1025 * PedanticBottoms is off,
1026 or * the scrutinee is an already-evaluated variable
1027 or * the scrutinee is a primop which is ok for speculation
1028 -- ie we want to preserve divide-by-zero errors, and
1029 -- calls to error itself!
1031 or * [Prim cases] the scrutinee is a primitive variable
1033 or * [Alg cases] the scrutinee is a variable and
1034 either * the rhs is the same variable
1035 (eg case x of C a b -> x ===> x)
1036 or * there is only one alternative, the default alternative,
1037 and the binder is used strictly in its scope.
1038 [NB this is helped by the "use default binder where
1039 possible" transformation; see below.]
1042 If so, then we can replace the case with one of the rhss.
1046 --------------------------------------------------
1047 -- 1. Eliminate the case altogether if poss
1048 --------------------------------------------------
1050 mkCase1 scrut case_bndr [(con,bndrs,rhs)]
1051 -- See if we can get rid of the case altogether
1052 -- See the extensive notes on case-elimination above
1053 -- mkCase made sure that if all the alternatives are equal,
1054 -- then there is now only one (DEFAULT) rhs
1055 | all isDeadBinder bndrs,
1057 -- Check that the scrutinee can be let-bound instead of case-bound
1058 exprOkForSpeculation scrut
1059 -- OK not to evaluate it
1060 -- This includes things like (==# a# b#)::Bool
1061 -- so that we simplify
1062 -- case ==# a# b# of { True -> x; False -> x }
1065 -- This particular example shows up in default methods for
1066 -- comparision operations (e.g. in (>=) for Int.Int32)
1067 || exprIsValue scrut -- It's already evaluated
1068 || var_demanded_later scrut -- It'll be demanded later
1070 -- || not opt_SimplPedanticBottoms) -- Or we don't care!
1071 -- We used to allow improving termination by discarding cases, unless -fpedantic-bottoms was on,
1072 -- but that breaks badly for the dataToTag# primop, which relies on a case to evaluate
1073 -- its argument: case x of { y -> dataToTag# y }
1074 -- Here we must *not* discard the case, because dataToTag# just fetches the tag from
1075 -- the info pointer. So we'll be pedantic all the time, and see if that gives any
1077 = tick (CaseElim case_bndr) `thenSmpl_`
1078 returnSmpl (bindCaseBndr case_bndr scrut rhs)
1081 -- The case binder is going to be evaluated later,
1082 -- and the scrutinee is a simple variable
1083 var_demanded_later (Var v) = isStrictDmd (idNewDemandInfo case_bndr)
1084 var_demanded_later other = False
1087 --------------------------------------------------
1089 --------------------------------------------------
1091 mkCase1 scrut case_bndr alts -- Identity case
1092 | all identity_alt alts
1093 = tick (CaseIdentity case_bndr) `thenSmpl_`
1094 returnSmpl (re_note scrut)
1096 identity_alt (con, args, rhs) = de_note rhs `cheapEqExpr` identity_rhs con args
1098 identity_rhs (DataAlt con) args = mkConApp con (arg_tys ++ map varToCoreExpr args)
1099 identity_rhs (LitAlt lit) _ = Lit lit
1100 identity_rhs DEFAULT _ = Var case_bndr
1102 arg_tys = map Type (tyConAppArgs (idType case_bndr))
1105 -- case coerce T e of x { _ -> coerce T' x }
1106 -- And we definitely want to eliminate this case!
1107 -- So we throw away notes from the RHS, and reconstruct
1108 -- (at least an approximation) at the other end
1109 de_note (Note _ e) = de_note e
1112 -- re_note wraps a coerce if it might be necessary
1113 re_note scrut = case head alts of
1114 (_,_,rhs1@(Note _ _)) -> mkCoerce (exprType rhs1) (idType case_bndr) scrut
1118 --------------------------------------------------
1120 --------------------------------------------------
1121 mkCase1 scrut bndr alts = returnSmpl (Case scrut bndr alts)
1125 When adding auxiliary bindings for the case binder, it's worth checking if
1126 its dead, because it often is, and occasionally these mkCase transformations
1127 cascade rather nicely.
1130 bindCaseBndr bndr rhs body
1131 | isDeadBinder bndr = body
1132 | otherwise = bindNonRec bndr rhs body