2 % (c) The AQUA Project, Glasgow University, 1993-1998
4 \section[SimplUtils]{The simplifier utilities}
8 simplBinder, simplBinders, simplRecBndrs,
9 simplLetBndr, simplLamBndrs,
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(..),
24 opt_SimplDoLambdaEtaExpansion, opt_SimplDoEtaReduction,
25 opt_SimplCaseMerge, opt_UF_UpdateInPlace
28 import CoreUtils ( cheapEqExpr, exprType,
29 etaExpand, exprEtaExpandArity, bindNonRec, mkCoerce,
30 findDefault, exprOkForSpeculation, exprIsValue
32 import qualified Subst ( simplBndrs, simplBndr, simplLetId, simplLamBndr )
33 import Id ( Id, idType, idInfo,
34 mkSysLocal, isDeadBinder, idNewDemandInfo,
35 idUnfolding, idNewStrictness
37 import NewDemand ( isStrictDmd, isBotRes, splitStrictSig )
39 import Type ( Type, seqType,
40 splitTyConApp_maybe, tyConAppArgs, mkTyVarTys,
41 splitRepFunTys, isStrictType
43 import OccName ( UserFS )
44 import TyCon ( tyConDataConsIfAvailable, isAlgTyCon, isNewTyCon )
45 import DataCon ( dataConRepArity, dataConSig, dataConArgTys )
46 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
226 | null ss && discardableCont cont = (reverse acc, discardCont cont, inl)
227 | otherwise = (reverse acc, cont, inl)
229 ----------------------------
230 vanilla_stricts, computed_stricts :: [Bool]
231 vanilla_stricts = repeat False
232 computed_stricts = zipWith (||) fun_stricts arg_stricts
234 ----------------------------
235 (val_arg_tys, _) = splitRepFunTys (idType fun)
236 arg_stricts = map isStrictType val_arg_tys ++ repeat False
237 -- These argument types are used as a cheap and cheerful way to find
238 -- unboxed arguments, which must be strict. But it's an InType
239 -- and so there might be a type variable where we expect a function
240 -- type (the substitution hasn't happened yet). And we don't bother
241 -- doing the type applications for a polymorphic function.
242 -- Hence the split*Rep*FunTys
244 ----------------------------
245 -- If fun_stricts is finite, it means the function returns bottom
246 -- after that number of value args have been consumed
247 -- Otherwise it's infinite, extended with False
249 = case splitStrictSig (idNewStrictness fun) of
250 (demands, result_info)
251 | not (demands `lengthExceeds` countValArgs orig_cont)
252 -> -- Enough args, use the strictness given.
253 -- For bottoming functions we used to pretend that the arg
254 -- is lazy, so that we don't treat the arg as an
255 -- interesting context. This avoids substituting
256 -- top-level bindings for (say) strings into
257 -- calls to error. But now we are more careful about
258 -- inlining lone variables, so its ok (see SimplUtils.analyseCont)
259 if isBotRes result_info then
260 map isStrictDmd demands -- Finite => result is bottom
262 map isStrictDmd demands ++ vanilla_stricts
264 other -> vanilla_stricts -- Not enough args, or no strictness
267 interestingArg :: OutExpr -> Bool
268 -- An argument is interesting if it has *some* structure
269 -- We are here trying to avoid unfolding a function that
270 -- is applied only to variables that have no unfolding
271 -- (i.e. they are probably lambda bound): f x y z
272 -- There is little point in inlining f here.
273 interestingArg (Var v) = hasSomeUnfolding (idUnfolding v)
274 -- Was: isValueUnfolding (idUnfolding v')
275 -- But that seems over-pessimistic
276 interestingArg (Type _) = False
277 interestingArg (App fn (Type _)) = interestingArg fn
278 interestingArg (Note _ a) = interestingArg a
279 interestingArg other = True
280 -- Consider let x = 3 in f x
281 -- The substitution will contain (x -> ContEx 3), and we want to
282 -- to say that x is an interesting argument.
283 -- But consider also (\x. f x y) y
284 -- The substitution will contain (x -> ContEx y), and we want to say
285 -- that x is not interesting (assuming y has no unfolding)
288 Comment about interestingCallContext
289 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
290 We want to avoid inlining an expression where there can't possibly be
291 any gain, such as in an argument position. Hence, if the continuation
292 is interesting (eg. a case scrutinee, application etc.) then we
293 inline, otherwise we don't.
295 Previously some_benefit used to return True only if the variable was
296 applied to some value arguments. This didn't work:
298 let x = _coerce_ (T Int) Int (I# 3) in
299 case _coerce_ Int (T Int) x of
302 we want to inline x, but can't see that it's a constructor in a case
303 scrutinee position, and some_benefit is False.
307 dMonadST = _/\_ t -> :Monad (g1 _@_ t, g2 _@_ t, g3 _@_ t)
309 .... case dMonadST _@_ x0 of (a,b,c) -> ....
311 we'd really like to inline dMonadST here, but we *don't* want to
312 inline if the case expression is just
314 case x of y { DEFAULT -> ... }
316 since we can just eliminate this case instead (x is in WHNF). Similar
317 applies when x is bound to a lambda expression. Hence
318 contIsInteresting looks for case expressions with just a single
322 interestingCallContext :: Bool -- False <=> no args at all
323 -> Bool -- False <=> no value args
325 -- The "lone-variable" case is important. I spent ages
326 -- messing about with unsatisfactory varaints, but this is nice.
327 -- The idea is that if a variable appear all alone
328 -- as an arg of lazy fn, or rhs Stop
329 -- as scrutinee of a case Select
330 -- as arg of a strict fn ArgOf
331 -- then we should not inline it (unless there is some other reason,
332 -- e.g. is is the sole occurrence). We achieve this by making
333 -- interestingCallContext return False for a lone variable.
335 -- Why? At least in the case-scrutinee situation, turning
336 -- let x = (a,b) in case x of y -> ...
338 -- let x = (a,b) in case (a,b) of y -> ...
340 -- let x = (a,b) in let y = (a,b) in ...
341 -- is bad if the binding for x will remain.
343 -- Another example: I discovered that strings
344 -- were getting inlined straight back into applications of 'error'
345 -- because the latter is strict.
347 -- f = \x -> ...(error s)...
349 -- Fundamentally such contexts should not ecourage inlining because
350 -- the context can ``see'' the unfolding of the variable (e.g. case or a RULE)
351 -- so there's no gain.
353 -- However, even a type application or coercion isn't a lone variable.
355 -- case $fMonadST @ RealWorld of { :DMonad a b c -> c }
356 -- We had better inline that sucker! The case won't see through it.
358 -- For now, I'm treating treating a variable applied to types
359 -- in a *lazy* context "lone". The motivating example was
361 -- g = /\a. \y. h (f a)
362 -- There's no advantage in inlining f here, and perhaps
363 -- a significant disadvantage. Hence some_val_args in the Stop case
365 interestingCallContext some_args some_val_args cont
368 interesting (InlinePlease _) = True
369 interesting (Select _ _ _ _ _) = some_args
370 interesting (ApplyTo _ _ _ _) = True -- Can happen if we have (coerce t (f x)) y
371 -- Perhaps True is a bit over-keen, but I've
372 -- seen (coerce f) x, where f has an INLINE prag,
373 -- So we have to give some motivaiton for inlining it
374 interesting (ArgOf _ _ _ _) = some_val_args
375 interesting (Stop ty _ upd_in_place) = some_val_args && upd_in_place
376 interesting (CoerceIt _ cont) = interesting cont
377 -- If this call is the arg of a strict function, the context
378 -- is a bit interesting. If we inline here, we may get useful
379 -- evaluation information to avoid repeated evals: e.g.
381 -- Here the contIsInteresting makes the '*' keener to inline,
382 -- which in turn exposes a constructor which makes the '+' inline.
383 -- Assuming that +,* aren't small enough to inline regardless.
385 -- It's also very important to inline in a strict context for things
388 -- Here, the context of (f x) is strict, and if f's unfolding is
389 -- a build it's *great* to inline it here. So we must ensure that
390 -- the context for (f x) is not totally uninteresting.
394 canUpdateInPlace :: Type -> Bool
395 -- Consider let x = <wurble> in ...
396 -- If <wurble> returns an explicit constructor, we might be able
397 -- to do update in place. So we treat even a thunk RHS context
398 -- as interesting if update in place is possible. We approximate
399 -- this by seeing if the type has a single constructor with a
400 -- small arity. But arity zero isn't good -- we share the single copy
401 -- for that case, so no point in sharing.
404 | not opt_UF_UpdateInPlace = False
406 = case splitTyConApp_maybe ty of
408 Just (tycon, _) -> case tyConDataConsIfAvailable tycon of
409 [dc] -> arity == 1 || arity == 2
411 arity = dataConRepArity dc
417 %************************************************************************
419 \section{Dealing with a single binder}
421 %************************************************************************
423 These functions are in the monad only so that they can be made strict via seq.
426 simplBinders :: SimplEnv -> [InBinder] -> SimplM (SimplEnv, [OutBinder])
427 simplBinders env bndrs
429 (subst', bndrs') = Subst.simplBndrs (getSubst env) bndrs
431 seqBndrs bndrs' `seq`
432 returnSmpl (setSubst env subst', bndrs')
434 simplBinder :: SimplEnv -> InBinder -> SimplM (SimplEnv, OutBinder)
437 (subst', bndr') = Subst.simplBndr (getSubst env) bndr
440 returnSmpl (setSubst env subst', bndr')
443 simplLetBndr :: SimplEnv -> InBinder -> SimplM (SimplEnv, OutBinder)
446 (subst', id') = Subst.simplLetId (getSubst env) id
449 returnSmpl (setSubst env subst', id')
451 simplLamBndrs, simplRecBndrs
452 :: SimplEnv -> [InBinder] -> SimplM (SimplEnv, [OutBinder])
453 simplRecBndrs = simplBndrs Subst.simplLetId
454 simplLamBndrs = simplBndrs Subst.simplLamBndr
456 simplBndrs simpl_bndr env bndrs
458 (subst', bndrs') = mapAccumL simpl_bndr (getSubst env) bndrs
460 seqBndrs bndrs' `seq`
461 returnSmpl (setSubst env subst', bndrs')
464 seqBndrs (b:bs) = seqBndr b `seq` seqBndrs bs
466 seqBndr b | isTyVar b = b `seq` ()
467 | otherwise = seqType (idType b) `seq`
474 newId :: UserFS -> Type -> SimplM Id
475 newId fs ty = getUniqueSmpl `thenSmpl` \ uniq ->
476 returnSmpl (mkSysLocal fs uniq ty)
480 %************************************************************************
482 \subsection{Rebuilding a lambda}
484 %************************************************************************
487 mkLam :: SimplEnv -> [OutBinder] -> OutExpr -> SimplCont -> SimplM FloatsWithExpr
491 a) eta reduction, if that gives a trivial expression
492 b) eta expansion [only if there are some value lambdas]
493 c) floating lets out through big lambdas
494 [only if all tyvar lambdas, and only if this lambda
498 mkLam env bndrs body cont
499 | opt_SimplDoEtaReduction,
500 Just etad_lam <- tryEtaReduce bndrs body
501 = tick (EtaReduction (head bndrs)) `thenSmpl_`
502 returnSmpl (emptyFloats env, etad_lam)
504 | opt_SimplDoLambdaEtaExpansion,
505 any isRuntimeVar bndrs
506 = tryEtaExpansion body `thenSmpl` \ body' ->
507 returnSmpl (emptyFloats env, mkLams bndrs body')
509 {- Sept 01: I'm experimenting with getting the
510 full laziness pass to float out past big lambdsa
511 | all isTyVar bndrs, -- Only for big lambdas
512 contIsRhs cont -- Only try the rhs type-lambda floating
513 -- if this is indeed a right-hand side; otherwise
514 -- we end up floating the thing out, only for float-in
515 -- to float it right back in again!
516 = tryRhsTyLam env bndrs body `thenSmpl` \ (floats, body') ->
517 returnSmpl (floats, mkLams bndrs body')
521 = returnSmpl (emptyFloats env, mkLams bndrs body)
525 %************************************************************************
527 \subsection{Eta expansion and reduction}
529 %************************************************************************
531 We try for eta reduction here, but *only* if we get all the
532 way to an exprIsTrivial expression.
533 We don't want to remove extra lambdas unless we are going
534 to avoid allocating this thing altogether
537 tryEtaReduce :: [OutBinder] -> OutExpr -> Maybe OutExpr
538 tryEtaReduce bndrs body
539 -- We don't use CoreUtils.etaReduce, because we can be more
541 -- (a) we already have the binders
542 -- (b) we can do the triviality test before computing the free vars
543 -- [in fact I take the simple path and look for just a variable]
544 = go (reverse bndrs) body
546 go (b : bs) (App fun arg) | ok_arg b arg = go bs fun -- Loop round
547 go [] (Var fun) | ok_fun fun = Just (Var fun) -- Success!
548 go _ _ = Nothing -- Failure!
550 ok_fun fun = not (fun `elem` bndrs) &&
551 isEvaldUnfolding (idUnfolding fun)
552 -- The exprIsValue is because eta reduction is not
553 -- valid in general: \x. bot /= bot
554 -- So we need to be sure that the "fun" is a value.
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 -> [AltCon] -> OutId -> [OutAlt] -> SimplM OutExpr
784 mkCase scrut handled_cons case_bndr alts
785 = mkAlts scrut handled_cons 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 handled_cons 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 handled_cons case_bndr alts
876 | (alts_no_deflt, Just rhs) <- findDefault alts,
877 -- There is a DEFAULT case
879 Just (tycon, inst_tys) <- splitTyConApp_maybe (idType case_bndr),
880 isAlgTyCon tycon, -- It's a data type, tuple, or unboxed tuples.
881 not (isNewTyCon tycon), -- We can have a newtype, if we are just doing an eval:
882 -- case x of { DEFAULT -> e }
883 -- and we don't want to fill in a default for them!
885 [missing_con] <- [con | con <- tyConDataConsIfAvailable tycon,
886 not (con `elem` handled_data_cons)]
887 -- There is just one missing constructor!
889 = tick (FillInCaseDefault case_bndr) `thenSmpl_`
890 getUniquesSmpl `thenSmpl` \ tv_uniqs ->
891 getUniquesSmpl `thenSmpl` \ id_uniqs ->
893 (_,_,ex_tyvars,_,_,_) = dataConSig missing_con
894 ex_tyvars' = zipWith mk tv_uniqs ex_tyvars
895 mk uniq tv = mkSysTyVar uniq (tyVarKind tv)
896 arg_ids = zipWith (mkSysLocal SLIT("a")) id_uniqs arg_tys
897 arg_tys = dataConArgTys missing_con (inst_tys ++ mkTyVarTys ex_tyvars')
898 better_alts = (DataAlt missing_con, ex_tyvars' ++ arg_ids, rhs) : alts_no_deflt
900 returnSmpl better_alts
902 handled_data_cons = [data_con | DataAlt data_con <- handled_cons]
904 --------------------------------------------------
905 -- 3. Merge nested cases
906 --------------------------------------------------
908 mkAlts scrut handled_cons outer_bndr outer_alts
909 | opt_SimplCaseMerge,
910 (outer_alts_without_deflt, maybe_outer_deflt) <- findDefault outer_alts,
911 Just (Case (Var scrut_var) inner_bndr inner_alts) <- maybe_outer_deflt,
912 scruting_same_var scrut_var
914 = let -- Eliminate any inner alts which are shadowed by the outer ones
915 outer_cons = [con | (con,_,_) <- outer_alts_without_deflt]
917 munged_inner_alts = [ (con, args, munge_rhs rhs)
918 | (con, args, rhs) <- inner_alts,
919 not (con `elem` outer_cons) -- Eliminate shadowed inner alts
921 munge_rhs rhs = bindCaseBndr inner_bndr (Var outer_bndr) rhs
923 (inner_con_alts, maybe_inner_default) = findDefault munged_inner_alts
925 new_alts = add_default maybe_inner_default
926 (outer_alts_without_deflt ++ inner_con_alts)
928 tick (CaseMerge outer_bndr) `thenSmpl_`
930 -- Warning: don't call mkAlts recursively!
931 -- Firstly, there's no point, because inner alts have already had
932 -- mkCase applied to them, so they won't have a case in their default
933 -- Secondly, if you do, you get an infinite loop, because the bindCaseBndr
934 -- in munge_rhs may put a case into the DEFAULT branch!
936 -- We are scrutinising the same variable if it's
937 -- the outer case-binder, or if the outer case scrutinises a variable
938 -- (and it's the same). Testing both allows us not to replace the
939 -- outer scrut-var with the outer case-binder (Simplify.simplCaseBinder).
940 scruting_same_var = case scrut of
941 Var outer_scrut -> \ v -> v == outer_bndr || v == outer_scrut
942 other -> \ v -> v == outer_bndr
944 add_default (Just rhs) alts = (DEFAULT,[],rhs) : alts
945 add_default Nothing alts = alts
948 --------------------------------------------------
950 --------------------------------------------------
952 mkAlts scrut handled_cons case_bndr other_alts = returnSmpl other_alts
957 =================================================================================
959 mkCase1 tries these things
961 1. Eliminate the case altogether if possible
972 Start with a simple situation:
974 case x# of ===> e[x#/y#]
977 (when x#, y# are of primitive type, of course). We can't (in general)
978 do this for algebraic cases, because we might turn bottom into
981 Actually, we generalise this idea to look for a case where we're
982 scrutinising a variable, and we know that only the default case can
987 other -> ...(case x of
991 Here the inner case can be eliminated. This really only shows up in
992 eliminating error-checking code.
994 We also make sure that we deal with this very common case:
999 Here we are using the case as a strict let; if x is used only once
1000 then we want to inline it. We have to be careful that this doesn't
1001 make the program terminate when it would have diverged before, so we
1003 - x is used strictly, or
1004 - e is already evaluated (it may so if e is a variable)
1006 Lastly, we generalise the transformation to handle this:
1012 We only do this for very cheaply compared r's (constructors, literals
1013 and variables). If pedantic bottoms is on, we only do it when the
1014 scrutinee is a PrimOp which can't fail.
1016 We do it *here*, looking at un-simplified alternatives, because we
1017 have to check that r doesn't mention the variables bound by the
1018 pattern in each alternative, so the binder-info is rather useful.
1020 So the case-elimination algorithm is:
1022 1. Eliminate alternatives which can't match
1024 2. Check whether all the remaining alternatives
1025 (a) do not mention in their rhs any of the variables bound in their pattern
1026 and (b) have equal rhss
1028 3. Check we can safely ditch the case:
1029 * PedanticBottoms is off,
1030 or * the scrutinee is an already-evaluated variable
1031 or * the scrutinee is a primop which is ok for speculation
1032 -- ie we want to preserve divide-by-zero errors, and
1033 -- calls to error itself!
1035 or * [Prim cases] the scrutinee is a primitive variable
1037 or * [Alg cases] the scrutinee is a variable and
1038 either * the rhs is the same variable
1039 (eg case x of C a b -> x ===> x)
1040 or * there is only one alternative, the default alternative,
1041 and the binder is used strictly in its scope.
1042 [NB this is helped by the "use default binder where
1043 possible" transformation; see below.]
1046 If so, then we can replace the case with one of the rhss.
1050 --------------------------------------------------
1051 -- 1. Eliminate the case altogether if poss
1052 --------------------------------------------------
1054 mkCase1 scrut case_bndr [(con,bndrs,rhs)]
1055 -- See if we can get rid of the case altogether
1056 -- See the extensive notes on case-elimination above
1057 -- mkCase made sure that if all the alternatives are equal,
1058 -- then there is now only one (DEFAULT) rhs
1059 | all isDeadBinder bndrs,
1061 -- Check that the scrutinee can be let-bound instead of case-bound
1062 exprOkForSpeculation scrut
1063 -- OK not to evaluate it
1064 -- This includes things like (==# a# b#)::Bool
1065 -- so that we simplify
1066 -- case ==# a# b# of { True -> x; False -> x }
1069 -- This particular example shows up in default methods for
1070 -- comparision operations (e.g. in (>=) for Int.Int32)
1071 || exprIsValue scrut -- It's already evaluated
1072 || var_demanded_later scrut -- It'll be demanded later
1074 -- || not opt_SimplPedanticBottoms) -- Or we don't care!
1075 -- We used to allow improving termination by discarding cases, unless -fpedantic-bottoms was on,
1076 -- but that breaks badly for the dataToTag# primop, which relies on a case to evaluate
1077 -- its argument: case x of { y -> dataToTag# y }
1078 -- Here we must *not* discard the case, because dataToTag# just fetches the tag from
1079 -- the info pointer. So we'll be pedantic all the time, and see if that gives any
1081 = tick (CaseElim case_bndr) `thenSmpl_`
1082 returnSmpl (bindCaseBndr case_bndr scrut rhs)
1085 -- The case binder is going to be evaluated later,
1086 -- and the scrutinee is a simple variable
1087 var_demanded_later (Var v) = isStrictDmd (idNewDemandInfo case_bndr)
1088 var_demanded_later other = False
1091 --------------------------------------------------
1093 --------------------------------------------------
1095 mkCase1 scrut case_bndr alts -- Identity case
1096 | all identity_alt alts
1097 = tick (CaseIdentity case_bndr) `thenSmpl_`
1098 returnSmpl (re_note scrut)
1100 identity_alt (con, args, rhs) = de_note rhs `cheapEqExpr` identity_rhs con args
1102 identity_rhs (DataAlt con) args = mkConApp con (arg_tys ++ map varToCoreExpr args)
1103 identity_rhs (LitAlt lit) _ = Lit lit
1104 identity_rhs DEFAULT _ = Var case_bndr
1106 arg_tys = map Type (tyConAppArgs (idType case_bndr))
1109 -- case coerce T e of x { _ -> coerce T' x }
1110 -- And we definitely want to eliminate this case!
1111 -- So we throw away notes from the RHS, and reconstruct
1112 -- (at least an approximation) at the other end
1113 de_note (Note _ e) = de_note e
1116 -- re_note wraps a coerce if it might be necessary
1117 re_note scrut = case head alts of
1118 (_,_,rhs1@(Note _ _)) -> mkCoerce (exprType rhs1) (idType case_bndr) scrut
1122 --------------------------------------------------
1124 --------------------------------------------------
1125 mkCase1 scrut bndr alts = returnSmpl (Case scrut bndr alts)
1129 When adding auxiliary bindings for the case binder, it's worth checking if
1130 its dead, because it often is, and occasionally these mkCase transformations
1131 cascade rather nicely.
1134 bindCaseBndr bndr rhs body
1135 | isDeadBinder bndr = body
1136 | otherwise = bindNonRec bndr rhs body