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 DupFlag -- An arbitrary strict context: the argument
81 -- of a strict function, or a primitive-arg fn
84 OutType -- cont_ty: the type of the expression being sought by the context
85 -- f (error "foo") ==> coerce t (error "foo")
87 -- We need to know the type t, to which to coerce.
88 (SimplEnv -> OutExpr -> SimplM FloatsWithExpr) -- What to do with the result
89 -- The result expression in the OutExprStuff has type cont_ty
91 data LetRhsFlag = AnArg -- It's just an argument not a let RHS
92 | AnRhs -- It's the RHS of a let (so please float lets out of big lambdas)
94 instance Outputable LetRhsFlag where
95 ppr AnArg = ptext SLIT("arg")
96 ppr AnRhs = ptext SLIT("rhs")
98 instance Outputable SimplCont where
99 ppr (Stop _ is_rhs _) = ptext SLIT("Stop") <> brackets (ppr is_rhs)
100 ppr (ApplyTo dup arg se cont) = (ptext SLIT("ApplyTo") <+> ppr dup <+> ppr arg) $$ ppr cont
101 ppr (ArgOf dup _ _ _) = ptext SLIT("ArgOf...") <+> ppr dup
102 ppr (Select dup bndr alts se cont) = (ptext SLIT("Select") <+> ppr dup <+> ppr bndr) $$
103 (nest 4 (ppr alts)) $$ ppr cont
104 ppr (CoerceIt ty cont) = (ptext SLIT("CoerceIt") <+> ppr ty) $$ ppr cont
105 ppr (InlinePlease cont) = ptext SLIT("InlinePlease") $$ ppr cont
107 data DupFlag = OkToDup | NoDup
109 instance Outputable DupFlag where
110 ppr OkToDup = ptext SLIT("ok")
111 ppr NoDup = ptext SLIT("nodup")
115 mkBoringStop :: OutType -> SimplCont
116 mkBoringStop ty = Stop ty AnArg (canUpdateInPlace ty)
118 mkStop :: OutType -> LetRhsFlag -> SimplCont
119 mkStop ty is_rhs = Stop ty is_rhs (canUpdateInPlace ty)
121 contIsRhs :: SimplCont -> Bool
122 contIsRhs (Stop _ AnRhs _) = True
123 contIsRhs (ArgOf _ AnRhs _ _) = True
124 contIsRhs other = False
126 contIsRhsOrArg (Stop _ _ _) = True
127 contIsRhsOrArg (ArgOf _ _ _ _) = True
128 contIsRhsOrArg other = False
131 contIsDupable :: SimplCont -> Bool
132 contIsDupable (Stop _ _ _) = True
133 contIsDupable (ApplyTo OkToDup _ _ _) = True
134 contIsDupable (ArgOf OkToDup _ _ _) = True
135 contIsDupable (Select OkToDup _ _ _ _) = True
136 contIsDupable (CoerceIt _ cont) = contIsDupable cont
137 contIsDupable (InlinePlease cont) = contIsDupable cont
138 contIsDupable other = False
141 discardableCont :: SimplCont -> Bool
142 discardableCont (Stop _ _ _) = False
143 discardableCont (CoerceIt _ cont) = discardableCont cont
144 discardableCont (InlinePlease cont) = discardableCont cont
145 discardableCont other = True
147 discardCont :: SimplCont -- A continuation, expecting
148 -> SimplCont -- Replace the continuation with a suitable coerce
149 discardCont cont = case cont of
150 Stop to_ty is_rhs _ -> cont
151 other -> CoerceIt to_ty (mkBoringStop to_ty)
153 to_ty = contResultType cont
156 contResultType :: SimplCont -> OutType
157 contResultType (Stop to_ty _ _) = to_ty
158 contResultType (ArgOf _ _ to_ty _) = to_ty
159 contResultType (ApplyTo _ _ _ cont) = contResultType cont
160 contResultType (CoerceIt _ cont) = contResultType cont
161 contResultType (InlinePlease cont) = contResultType cont
162 contResultType (Select _ _ _ _ cont) = contResultType cont
165 countValArgs :: SimplCont -> Int
166 countValArgs (ApplyTo _ (Type ty) se cont) = countValArgs cont
167 countValArgs (ApplyTo _ val_arg se cont) = 1 + countValArgs cont
168 countValArgs other = 0
170 countArgs :: SimplCont -> Int
171 countArgs (ApplyTo _ arg se cont) = 1 + countArgs cont
175 pushContArgs :: SimplEnv -> [OutArg] -> SimplCont -> SimplCont
176 -- Pushes args with the specified environment
177 pushContArgs env [] cont = cont
178 pushContArgs env (arg : args) cont = ApplyTo NoDup arg env (pushContArgs env args cont)
183 getContArgs :: SwitchChecker
184 -> OutId -> SimplCont
185 -> ([(InExpr, SimplEnv, Bool)], -- Arguments; the Bool is true for strict args
186 SimplCont, -- Remaining continuation
187 Bool) -- Whether we came across an InlineCall
188 -- getContArgs id k = (args, k', inl)
189 -- args are the leading ApplyTo items in k
190 -- (i.e. outermost comes first)
191 -- augmented with demand info from the functionn
192 getContArgs chkr fun orig_cont
194 -- Ignore strictness info if the no-case-of-case
195 -- flag is on. Strictness changes evaluation order
196 -- and that can change full laziness
197 stricts | switchIsOn chkr NoCaseOfCase = vanilla_stricts
198 | otherwise = computed_stricts
200 go [] stricts False orig_cont
202 ----------------------------
205 go acc ss inl (ApplyTo _ arg@(Type _) se cont)
206 = go ((arg,se,False) : acc) ss inl cont
207 -- NB: don't bother to instantiate the function type
210 go acc (s:ss) inl (ApplyTo _ arg se cont)
211 = go ((arg,se,s) : acc) ss inl cont
213 -- An Inline continuation
214 go acc ss inl (InlinePlease cont)
215 = go acc ss True cont
217 -- We're run out of arguments, or else we've run out of demands
218 -- The latter only happens if the result is guaranteed bottom
219 -- This is the case for
220 -- * case (error "hello") of { ... }
221 -- * (error "Hello") arg
222 -- * f (error "Hello") where f is strict
225 | null ss && discardableCont cont = (reverse acc, discardCont cont, inl)
226 | otherwise = (reverse acc, cont, inl)
228 ----------------------------
229 vanilla_stricts, computed_stricts :: [Bool]
230 vanilla_stricts = repeat False
231 computed_stricts = zipWith (||) fun_stricts arg_stricts
233 ----------------------------
234 (val_arg_tys, _) = splitRepFunTys (idType fun)
235 arg_stricts = map isStrictType val_arg_tys ++ repeat False
236 -- These argument types are used as a cheap and cheerful way to find
237 -- unboxed arguments, which must be strict. But it's an InType
238 -- and so there might be a type variable where we expect a function
239 -- type (the substitution hasn't happened yet). And we don't bother
240 -- doing the type applications for a polymorphic function.
241 -- Hence the split*Rep*FunTys
243 ----------------------------
244 -- If fun_stricts is finite, it means the function returns bottom
245 -- after that number of value args have been consumed
246 -- Otherwise it's infinite, extended with False
248 = case splitStrictSig (idNewStrictness fun) of
249 (demands, result_info)
250 | not (demands `lengthExceeds` countValArgs orig_cont)
251 -> -- Enough args, use the strictness given.
252 -- For bottoming functions we used to pretend that the arg
253 -- is lazy, so that we don't treat the arg as an
254 -- interesting context. This avoids substituting
255 -- top-level bindings for (say) strings into
256 -- calls to error. But now we are more careful about
257 -- inlining lone variables, so its ok (see SimplUtils.analyseCont)
258 if isBotRes result_info then
259 map isStrictDmd demands -- Finite => result is bottom
261 map isStrictDmd demands ++ vanilla_stricts
263 other -> vanilla_stricts -- Not enough args, or no strictness
266 interestingArg :: OutExpr -> Bool
267 -- An argument is interesting if it has *some* structure
268 -- We are here trying to avoid unfolding a function that
269 -- is applied only to variables that have no unfolding
270 -- (i.e. they are probably lambda bound): f x y z
271 -- There is little point in inlining f here.
272 interestingArg (Var v) = hasSomeUnfolding (idUnfolding v)
273 -- Was: isValueUnfolding (idUnfolding v')
274 -- But that seems over-pessimistic
275 interestingArg (Type _) = False
276 interestingArg (App fn (Type _)) = interestingArg fn
277 interestingArg (Note _ a) = interestingArg a
278 interestingArg other = True
279 -- Consider let x = 3 in f x
280 -- The substitution will contain (x -> ContEx 3), and we want to
281 -- to say that x is an interesting argument.
282 -- But consider also (\x. f x y) y
283 -- The substitution will contain (x -> ContEx y), and we want to say
284 -- that x is not interesting (assuming y has no unfolding)
287 Comment about interestingCallContext
288 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
289 We want to avoid inlining an expression where there can't possibly be
290 any gain, such as in an argument position. Hence, if the continuation
291 is interesting (eg. a case scrutinee, application etc.) then we
292 inline, otherwise we don't.
294 Previously some_benefit used to return True only if the variable was
295 applied to some value arguments. This didn't work:
297 let x = _coerce_ (T Int) Int (I# 3) in
298 case _coerce_ Int (T Int) x of
301 we want to inline x, but can't see that it's a constructor in a case
302 scrutinee position, and some_benefit is False.
306 dMonadST = _/\_ t -> :Monad (g1 _@_ t, g2 _@_ t, g3 _@_ t)
308 .... case dMonadST _@_ x0 of (a,b,c) -> ....
310 we'd really like to inline dMonadST here, but we *don't* want to
311 inline if the case expression is just
313 case x of y { DEFAULT -> ... }
315 since we can just eliminate this case instead (x is in WHNF). Similar
316 applies when x is bound to a lambda expression. Hence
317 contIsInteresting looks for case expressions with just a single
321 interestingCallContext :: Bool -- False <=> no args at all
322 -> Bool -- False <=> no value args
324 -- The "lone-variable" case is important. I spent ages
325 -- messing about with unsatisfactory varaints, but this is nice.
326 -- The idea is that if a variable appear all alone
327 -- as an arg of lazy fn, or rhs Stop
328 -- as scrutinee of a case Select
329 -- as arg of a strict fn ArgOf
330 -- then we should not inline it (unless there is some other reason,
331 -- e.g. is is the sole occurrence). We achieve this by making
332 -- interestingCallContext return False for a lone variable.
334 -- Why? At least in the case-scrutinee situation, turning
335 -- let x = (a,b) in case x of y -> ...
337 -- let x = (a,b) in case (a,b) of y -> ...
339 -- let x = (a,b) in let y = (a,b) in ...
340 -- is bad if the binding for x will remain.
342 -- Another example: I discovered that strings
343 -- were getting inlined straight back into applications of 'error'
344 -- because the latter is strict.
346 -- f = \x -> ...(error s)...
348 -- Fundamentally such contexts should not ecourage inlining because
349 -- the context can ``see'' the unfolding of the variable (e.g. case or a RULE)
350 -- so there's no gain.
352 -- However, even a type application or coercion isn't a lone variable.
354 -- case $fMonadST @ RealWorld of { :DMonad a b c -> c }
355 -- We had better inline that sucker! The case won't see through it.
357 -- For now, I'm treating treating a variable applied to types
358 -- in a *lazy* context "lone". The motivating example was
360 -- g = /\a. \y. h (f a)
361 -- There's no advantage in inlining f here, and perhaps
362 -- a significant disadvantage. Hence some_val_args in the Stop case
364 interestingCallContext some_args some_val_args cont
367 interesting (InlinePlease _) = True
368 interesting (Select _ _ _ _ _) = some_args
369 interesting (ApplyTo _ _ _ _) = True -- Can happen if we have (coerce t (f x)) y
370 -- Perhaps True is a bit over-keen, but I've
371 -- seen (coerce f) x, where f has an INLINE prag,
372 -- So we have to give some motivaiton for inlining it
373 interesting (ArgOf _ _ _ _) = some_val_args
374 interesting (Stop ty _ upd_in_place) = some_val_args && upd_in_place
375 interesting (CoerceIt _ cont) = interesting cont
376 -- If this call is the arg of a strict function, the context
377 -- is a bit interesting. If we inline here, we may get useful
378 -- evaluation information to avoid repeated evals: e.g.
380 -- Here the contIsInteresting makes the '*' keener to inline,
381 -- which in turn exposes a constructor which makes the '+' inline.
382 -- Assuming that +,* aren't small enough to inline regardless.
384 -- It's also very important to inline in a strict context for things
387 -- Here, the context of (f x) is strict, and if f's unfolding is
388 -- a build it's *great* to inline it here. So we must ensure that
389 -- the context for (f x) is not totally uninteresting.
393 canUpdateInPlace :: Type -> Bool
394 -- Consider let x = <wurble> in ...
395 -- If <wurble> returns an explicit constructor, we might be able
396 -- to do update in place. So we treat even a thunk RHS context
397 -- as interesting if update in place is possible. We approximate
398 -- this by seeing if the type has a single constructor with a
399 -- small arity. But arity zero isn't good -- we share the single copy
400 -- for that case, so no point in sharing.
403 | not opt_UF_UpdateInPlace = False
405 = case splitTyConApp_maybe ty of
407 Just (tycon, _) -> case tyConDataConsIfAvailable tycon of
408 [dc] -> arity == 1 || arity == 2
410 arity = dataConRepArity dc
416 %************************************************************************
418 \section{Dealing with a single binder}
420 %************************************************************************
422 These functions are in the monad only so that they can be made strict via seq.
425 simplBinders :: SimplEnv -> [InBinder] -> SimplM (SimplEnv, [OutBinder])
426 simplBinders env bndrs
428 (subst', bndrs') = Subst.simplBndrs (getSubst env) bndrs
430 seqBndrs bndrs' `seq`
431 returnSmpl (setSubst env subst', bndrs')
433 simplBinder :: SimplEnv -> InBinder -> SimplM (SimplEnv, OutBinder)
436 (subst', bndr') = Subst.simplBndr (getSubst env) bndr
439 returnSmpl (setSubst env subst', bndr')
442 simplLetBndr :: SimplEnv -> InBinder -> SimplM (SimplEnv, OutBinder)
445 (subst', id') = Subst.simplLetId (getSubst env) id
448 returnSmpl (setSubst env subst', id')
450 simplLamBndrs, simplRecBndrs
451 :: SimplEnv -> [InBinder] -> SimplM (SimplEnv, [OutBinder])
452 simplRecBndrs = simplBndrs Subst.simplLetId
453 simplLamBndrs = simplBndrs Subst.simplLamBndr
455 simplBndrs simpl_bndr env bndrs
457 (subst', bndrs') = mapAccumL simpl_bndr (getSubst env) bndrs
459 seqBndrs bndrs' `seq`
460 returnSmpl (setSubst env subst', bndrs')
463 seqBndrs (b:bs) = seqBndr b `seq` seqBndrs bs
465 seqBndr b | isTyVar b = b `seq` ()
466 | otherwise = seqType (idType b) `seq`
473 newId :: UserFS -> Type -> SimplM Id
474 newId fs ty = getUniqueSmpl `thenSmpl` \ uniq ->
475 returnSmpl (mkSysLocal fs uniq ty)
479 %************************************************************************
481 \subsection{Rebuilding a lambda}
483 %************************************************************************
486 mkLam :: SimplEnv -> [OutBinder] -> OutExpr -> SimplCont -> SimplM FloatsWithExpr
490 a) eta reduction, if that gives a trivial expression
491 b) eta expansion [only if there are some value lambdas]
492 c) floating lets out through big lambdas
493 [only if all tyvar lambdas, and only if this lambda
497 mkLam env bndrs body cont
498 | opt_SimplDoEtaReduction,
499 Just etad_lam <- tryEtaReduce bndrs body
500 = tick (EtaReduction (head bndrs)) `thenSmpl_`
501 returnSmpl (emptyFloats env, etad_lam)
503 | opt_SimplDoLambdaEtaExpansion,
504 any isRuntimeVar bndrs
505 = tryEtaExpansion body `thenSmpl` \ body' ->
506 returnSmpl (emptyFloats env, mkLams bndrs body')
508 {- Sept 01: I'm experimenting with getting the
509 full laziness pass to float out past big lambdsa
510 | all isTyVar bndrs, -- Only for big lambdas
511 contIsRhs cont -- Only try the rhs type-lambda floating
512 -- if this is indeed a right-hand side; otherwise
513 -- we end up floating the thing out, only for float-in
514 -- to float it right back in again!
515 = tryRhsTyLam env bndrs body `thenSmpl` \ (floats, body') ->
516 returnSmpl (floats, mkLams bndrs body')
520 = returnSmpl (emptyFloats env, mkLams bndrs body)
524 %************************************************************************
526 \subsection{Eta expansion and reduction}
528 %************************************************************************
530 We try for eta reduction here, but *only* if we get all the
531 way to an exprIsTrivial expression.
532 We don't want to remove extra lambdas unless we are going
533 to avoid allocating this thing altogether
536 tryEtaReduce :: [OutBinder] -> OutExpr -> Maybe OutExpr
537 tryEtaReduce bndrs body
538 -- We don't use CoreUtils.etaReduce, because we can be more
540 -- (a) we already have the binders
541 -- (b) we can do the triviality test before computing the free vars
542 -- [in fact I take the simple path and look for just a variable]
543 = go (reverse bndrs) body
545 go (b : bs) (App fun arg) | ok_arg b arg = go bs fun -- Loop round
546 go [] (Var fun) | ok_fun fun = Just (Var fun) -- Success!
547 go _ _ = Nothing -- Failure!
549 ok_fun fun = not (fun `elem` bndrs)
550 ok_arg b arg = varToCoreExpr b `cheapEqExpr` arg
554 Try eta expansion for RHSs
557 f = \x1..xn -> N ==> f = \x1..xn y1..ym -> N y1..ym
560 where (in both cases)
562 * The xi can include type variables
564 * The yi are all value variables
566 * N is a NORMAL FORM (i.e. no redexes anywhere)
567 wanting a suitable number of extra args.
569 We may have to sandwich some coerces between the lambdas
570 to make the types work. exprEtaExpandArity looks through coerces
571 when computing arity; and etaExpand adds the coerces as necessary when
572 actually computing the expansion.
575 tryEtaExpansion :: OutExpr -> SimplM OutExpr
576 -- There is at least one runtime binder in the binders
578 = getUniquesSmpl `thenSmpl` \ us ->
579 returnSmpl (etaExpand fun_arity us body (exprType body))
581 fun_arity = exprEtaExpandArity body
585 %************************************************************************
587 \subsection{Floating lets out of big lambdas}
589 %************************************************************************
591 tryRhsTyLam tries this transformation, when the big lambda appears as
592 the RHS of a let(rec) binding:
594 /\abc -> let(rec) x = e in b
596 let(rec) x' = /\abc -> let x = x' a b c in e
598 /\abc -> let x = x' a b c in b
600 This is good because it can turn things like:
602 let f = /\a -> letrec g = ... g ... in g
604 letrec g' = /\a -> ... g' a ...
608 which is better. In effect, it means that big lambdas don't impede
611 This optimisation is CRUCIAL in eliminating the junk introduced by
612 desugaring mutually recursive definitions. Don't eliminate it lightly!
614 So far as the implementation is concerned:
616 Invariant: go F e = /\tvs -> F e
620 = Let x' = /\tvs -> F e
624 G = F . Let x = x' tvs
626 go F (Letrec xi=ei in b)
627 = Letrec {xi' = /\tvs -> G ei}
631 G = F . Let {xi = xi' tvs}
633 [May 1999] If we do this transformation *regardless* then we can
634 end up with some pretty silly stuff. For example,
637 st = /\ s -> let { x1=r1 ; x2=r2 } in ...
642 st = /\s -> ...[y1 s/x1, y2 s/x2]
645 Unless the "..." is a WHNF there is really no point in doing this.
646 Indeed it can make things worse. Suppose x1 is used strictly,
649 x1* = case f y of { (a,b) -> e }
651 If we abstract this wrt the tyvar we then can't do the case inline
652 as we would normally do.
656 {- Trying to do this in full laziness
658 tryRhsTyLam :: SimplEnv -> [OutTyVar] -> OutExpr -> SimplM FloatsWithExpr
659 -- Call ensures that all the binders are type variables
661 tryRhsTyLam env tyvars body -- Only does something if there's a let
662 | not (all isTyVar tyvars)
663 || not (worth_it body) -- inside a type lambda,
664 = returnSmpl (emptyFloats env, body) -- and a WHNF inside that
667 = go env (\x -> x) body
670 worth_it e@(Let _ _) = whnf_in_middle e
673 whnf_in_middle (Let (NonRec x rhs) e) | isUnLiftedType (idType x) = False
674 whnf_in_middle (Let _ e) = whnf_in_middle e
675 whnf_in_middle e = exprIsCheap e
677 main_tyvar_set = mkVarSet tyvars
679 go env fn (Let bind@(NonRec var rhs) body)
681 = go env (fn . Let bind) body
683 go env fn (Let (NonRec var rhs) body)
684 = mk_poly tyvars_here var `thenSmpl` \ (var', rhs') ->
685 addAuxiliaryBind env (NonRec var' (mkLams tyvars_here (fn rhs))) $ \ env ->
686 go env (fn . Let (mk_silly_bind var rhs')) body
690 tyvars_here = varSetElems (main_tyvar_set `intersectVarSet` exprSomeFreeVars isTyVar rhs)
691 -- Abstract only over the type variables free in the rhs
692 -- wrt which the new binding is abstracted. But the naive
693 -- approach of abstract wrt the tyvars free in the Id's type
695 -- /\ a b -> let t :: (a,b) = (e1, e2)
698 -- Here, b isn't free in x's type, but we must nevertheless
699 -- abstract wrt b as well, because t's type mentions b.
700 -- Since t is floated too, we'd end up with the bogus:
701 -- poly_t = /\ a b -> (e1, e2)
702 -- poly_x = /\ a -> fst (poly_t a *b*)
703 -- So for now we adopt the even more naive approach of
704 -- abstracting wrt *all* the tyvars. We'll see if that
705 -- gives rise to problems. SLPJ June 98
707 go env fn (Let (Rec prs) body)
708 = mapAndUnzipSmpl (mk_poly tyvars_here) vars `thenSmpl` \ (vars', rhss') ->
710 gn body = fn (foldr Let body (zipWith mk_silly_bind vars rhss'))
711 pairs = vars' `zip` [mkLams tyvars_here (gn rhs) | rhs <- rhss]
713 addAuxiliaryBind env (Rec pairs) $ \ env ->
716 (vars,rhss) = unzip prs
717 tyvars_here = varSetElems (main_tyvar_set `intersectVarSet` exprsSomeFreeVars isTyVar (map snd prs))
718 -- See notes with tyvars_here above
720 go env fn body = returnSmpl (emptyFloats env, fn body)
722 mk_poly tyvars_here var
723 = getUniqueSmpl `thenSmpl` \ uniq ->
725 poly_name = setNameUnique (idName var) uniq -- Keep same name
726 poly_ty = mkForAllTys tyvars_here (idType var) -- But new type of course
727 poly_id = mkLocalId poly_name poly_ty
729 -- In the olden days, it was crucial to copy the occInfo of the original var,
730 -- because we were looking at occurrence-analysed but as yet unsimplified code!
731 -- In particular, we mustn't lose the loop breakers. BUT NOW we are looking
732 -- at already simplified code, so it doesn't matter
734 -- It's even right to retain single-occurrence or dead-var info:
735 -- Suppose we started with /\a -> let x = E in B
736 -- where x occurs once in B. Then we transform to:
737 -- let x' = /\a -> E in /\a -> let x* = x' a in B
738 -- where x* has an INLINE prag on it. Now, once x* is inlined,
739 -- the occurrences of x' will be just the occurrences originally
742 returnSmpl (poly_id, mkTyApps (Var poly_id) (mkTyVarTys tyvars_here))
744 mk_silly_bind var rhs = NonRec var (Note InlineMe rhs)
745 -- Suppose we start with:
747 -- x = /\ a -> let g = G in E
749 -- Then we'll float to get
751 -- x = let poly_g = /\ a -> G
752 -- in /\ a -> let g = poly_g a in E
754 -- But now the occurrence analyser will see just one occurrence
755 -- of poly_g, not inside a lambda, so the simplifier will
756 -- PreInlineUnconditionally poly_g back into g! Badk to square 1!
757 -- (I used to think that the "don't inline lone occurrences" stuff
758 -- would stop this happening, but since it's the *only* occurrence,
759 -- PreInlineUnconditionally kicks in first!)
761 -- Solution: put an INLINE note on g's RHS, so that poly_g seems
762 -- to appear many times. (NB: mkInlineMe eliminates
763 -- such notes on trivial RHSs, so do it manually.)
768 %************************************************************************
770 \subsection{Case absorption and identity-case elimination}
772 %************************************************************************
774 mkCase puts a case expression back together, trying various transformations first.
777 mkCase :: OutExpr -> [AltCon] -> OutId -> [OutAlt] -> SimplM OutExpr
779 mkCase scrut handled_cons case_bndr alts
780 = mkAlts scrut handled_cons case_bndr alts `thenSmpl` \ better_alts ->
781 mkCase1 scrut case_bndr better_alts
785 mkAlts tries these things:
787 1. If several alternatives are identical, merge them into
788 a single DEFAULT alternative. I've occasionally seen this
789 making a big difference:
791 case e of =====> case e of
792 C _ -> f x D v -> ....v....
793 D v -> ....v.... DEFAULT -> f x
796 The point is that we merge common RHSs, at least for the DEFAULT case.
797 [One could do something more elaborate but I've never seen it needed.]
798 To avoid an expensive test, we just merge branches equal to the *first*
799 alternative; this picks up the common cases
800 a) all branches equal
801 b) some branches equal to the DEFAULT (which occurs first)
803 2. If the DEFAULT alternative can match only one possible constructor,
804 then make that constructor explicit.
806 case e of x { DEFAULT -> rhs }
808 case e of x { (a,b) -> rhs }
809 where the type is a single constructor type. This gives better code
810 when rhs also scrutinises x or e.
813 case e of b { ==> case e of b {
814 p1 -> rhs1 p1 -> rhs1
816 pm -> rhsm pm -> rhsm
817 _ -> case b of b' { pn -> let b'=b in rhsn
819 ... po -> let b'=b in rhso
820 po -> rhso _ -> let b'=b in rhsd
824 which merges two cases in one case when -- the default alternative of
825 the outer case scrutises the same variable as the outer case This
826 transformation is called Case Merging. It avoids that the same
827 variable is scrutinised multiple times.
830 The case where transformation (1) showed up was like this (lib/std/PrelCError.lhs):
836 where @is@ was something like
838 p `is` n = p /= (-1) && p == n
840 This gave rise to a horrible sequence of cases
847 and similarly in cascade for all the join points!
852 --------------------------------------------------
853 -- 1. Merge identical branches
854 --------------------------------------------------
855 mkAlts scrut handled_cons case_bndr alts@((con1,bndrs1,rhs1) : con_alts)
856 | all isDeadBinder bndrs1, -- Remember the default
857 length filtered_alts < length con_alts -- alternative comes first
858 = tick (AltMerge case_bndr) `thenSmpl_`
859 returnSmpl better_alts
861 filtered_alts = filter keep con_alts
862 keep (con,bndrs,rhs) = not (all isDeadBinder bndrs && rhs `cheapEqExpr` rhs1)
863 better_alts = (DEFAULT, [], rhs1) : filtered_alts
866 --------------------------------------------------
867 -- 2. Fill in missing constructor
868 --------------------------------------------------
870 mkAlts scrut handled_cons case_bndr alts
871 | (alts_no_deflt, Just rhs) <- findDefault alts,
872 -- There is a DEFAULT case
874 Just (tycon, inst_tys) <- splitTyConApp_maybe (idType case_bndr),
875 isAlgTyCon tycon, -- It's a data type, tuple, or unboxed tuples.
876 not (isNewTyCon tycon), -- We can have a newtype, if we are just doing an eval:
877 -- case x of { DEFAULT -> e }
878 -- and we don't want to fill in a default for them!
880 [missing_con] <- [con | con <- tyConDataConsIfAvailable tycon,
881 not (con `elem` handled_data_cons)]
882 -- There is just one missing constructor!
884 = tick (FillInCaseDefault case_bndr) `thenSmpl_`
885 getUniquesSmpl `thenSmpl` \ tv_uniqs ->
886 getUniquesSmpl `thenSmpl` \ id_uniqs ->
888 (_,_,ex_tyvars,_,_,_) = dataConSig missing_con
889 ex_tyvars' = zipWith mk tv_uniqs ex_tyvars
890 mk uniq tv = mkSysTyVar uniq (tyVarKind tv)
891 arg_ids = zipWith (mkSysLocal SLIT("a")) id_uniqs arg_tys
892 arg_tys = dataConArgTys missing_con (inst_tys ++ mkTyVarTys ex_tyvars')
893 better_alts = (DataAlt missing_con, ex_tyvars' ++ arg_ids, rhs) : alts_no_deflt
895 returnSmpl better_alts
897 handled_data_cons = [data_con | DataAlt data_con <- handled_cons]
899 --------------------------------------------------
900 -- 3. Merge nested cases
901 --------------------------------------------------
903 mkAlts scrut handled_cons outer_bndr outer_alts
904 | opt_SimplCaseMerge,
905 (outer_alts_without_deflt, maybe_outer_deflt) <- findDefault outer_alts,
906 Just (Case (Var scrut_var) inner_bndr inner_alts) <- maybe_outer_deflt,
907 scruting_same_var scrut_var
909 = let -- Eliminate any inner alts which are shadowed by the outer ones
910 outer_cons = [con | (con,_,_) <- outer_alts_without_deflt]
912 munged_inner_alts = [ (con, args, munge_rhs rhs)
913 | (con, args, rhs) <- inner_alts,
914 not (con `elem` outer_cons) -- Eliminate shadowed inner alts
916 munge_rhs rhs = bindCaseBndr inner_bndr (Var outer_bndr) rhs
918 (inner_con_alts, maybe_inner_default) = findDefault munged_inner_alts
920 new_alts = add_default maybe_inner_default
921 (outer_alts_without_deflt ++ inner_con_alts)
923 tick (CaseMerge outer_bndr) `thenSmpl_`
925 -- Warning: don't call mkAlts recursively!
926 -- Firstly, there's no point, because inner alts have already had
927 -- mkCase applied to them, so they won't have a case in their default
928 -- Secondly, if you do, you get an infinite loop, because the bindCaseBndr
929 -- in munge_rhs may put a case into the DEFAULT branch!
931 -- We are scrutinising the same variable if it's
932 -- the outer case-binder, or if the outer case scrutinises a variable
933 -- (and it's the same). Testing both allows us not to replace the
934 -- outer scrut-var with the outer case-binder (Simplify.simplCaseBinder).
935 scruting_same_var = case scrut of
936 Var outer_scrut -> \ v -> v == outer_bndr || v == outer_scrut
937 other -> \ v -> v == outer_bndr
939 add_default (Just rhs) alts = (DEFAULT,[],rhs) : alts
940 add_default Nothing alts = alts
943 --------------------------------------------------
945 --------------------------------------------------
947 mkAlts scrut handled_cons case_bndr other_alts = returnSmpl other_alts
952 =================================================================================
954 mkCase1 tries these things
956 1. Eliminate the case altogether if possible
967 Start with a simple situation:
969 case x# of ===> e[x#/y#]
972 (when x#, y# are of primitive type, of course). We can't (in general)
973 do this for algebraic cases, because we might turn bottom into
976 Actually, we generalise this idea to look for a case where we're
977 scrutinising a variable, and we know that only the default case can
982 other -> ...(case x of
986 Here the inner case can be eliminated. This really only shows up in
987 eliminating error-checking code.
989 We also make sure that we deal with this very common case:
994 Here we are using the case as a strict let; if x is used only once
995 then we want to inline it. We have to be careful that this doesn't
996 make the program terminate when it would have diverged before, so we
998 - x is used strictly, or
999 - e is already evaluated (it may so if e is a variable)
1001 Lastly, we generalise the transformation to handle this:
1007 We only do this for very cheaply compared r's (constructors, literals
1008 and variables). If pedantic bottoms is on, we only do it when the
1009 scrutinee is a PrimOp which can't fail.
1011 We do it *here*, looking at un-simplified alternatives, because we
1012 have to check that r doesn't mention the variables bound by the
1013 pattern in each alternative, so the binder-info is rather useful.
1015 So the case-elimination algorithm is:
1017 1. Eliminate alternatives which can't match
1019 2. Check whether all the remaining alternatives
1020 (a) do not mention in their rhs any of the variables bound in their pattern
1021 and (b) have equal rhss
1023 3. Check we can safely ditch the case:
1024 * PedanticBottoms is off,
1025 or * the scrutinee is an already-evaluated variable
1026 or * the scrutinee is a primop which is ok for speculation
1027 -- ie we want to preserve divide-by-zero errors, and
1028 -- calls to error itself!
1030 or * [Prim cases] the scrutinee is a primitive variable
1032 or * [Alg cases] the scrutinee is a variable and
1033 either * the rhs is the same variable
1034 (eg case x of C a b -> x ===> x)
1035 or * there is only one alternative, the default alternative,
1036 and the binder is used strictly in its scope.
1037 [NB this is helped by the "use default binder where
1038 possible" transformation; see below.]
1041 If so, then we can replace the case with one of the rhss.
1045 --------------------------------------------------
1046 -- 1. Eliminate the case altogether if poss
1047 --------------------------------------------------
1049 mkCase1 scrut case_bndr [(con,bndrs,rhs)]
1050 -- See if we can get rid of the case altogether
1051 -- See the extensive notes on case-elimination above
1052 -- mkCase made sure that if all the alternatives are equal,
1053 -- then there is now only one (DEFAULT) rhs
1054 | all isDeadBinder bndrs,
1056 -- Check that the scrutinee can be let-bound instead of case-bound
1057 exprOkForSpeculation scrut
1058 -- OK not to evaluate it
1059 -- This includes things like (==# a# b#)::Bool
1060 -- so that we simplify
1061 -- case ==# a# b# of { True -> x; False -> x }
1064 -- This particular example shows up in default methods for
1065 -- comparision operations (e.g. in (>=) for Int.Int32)
1066 || exprIsValue scrut -- It's already evaluated
1067 || var_demanded_later scrut -- It'll be demanded later
1069 -- || not opt_SimplPedanticBottoms) -- Or we don't care!
1070 -- We used to allow improving termination by discarding cases, unless -fpedantic-bottoms was on,
1071 -- but that breaks badly for the dataToTag# primop, which relies on a case to evaluate
1072 -- its argument: case x of { y -> dataToTag# y }
1073 -- Here we must *not* discard the case, because dataToTag# just fetches the tag from
1074 -- the info pointer. So we'll be pedantic all the time, and see if that gives any
1076 = tick (CaseElim case_bndr) `thenSmpl_`
1077 returnSmpl (bindCaseBndr case_bndr scrut rhs)
1080 -- The case binder is going to be evaluated later,
1081 -- and the scrutinee is a simple variable
1082 var_demanded_later (Var v) = isStrictDmd (idNewDemandInfo case_bndr)
1083 var_demanded_later other = False
1086 --------------------------------------------------
1088 --------------------------------------------------
1090 mkCase1 scrut case_bndr alts -- Identity case
1091 | all identity_alt alts
1092 = tick (CaseIdentity case_bndr) `thenSmpl_`
1093 returnSmpl (re_note scrut)
1095 identity_alt (con, args, rhs) = de_note rhs `cheapEqExpr` identity_rhs con args
1097 identity_rhs (DataAlt con) args = mkConApp con (arg_tys ++ map varToCoreExpr args)
1098 identity_rhs (LitAlt lit) _ = Lit lit
1099 identity_rhs DEFAULT _ = Var case_bndr
1101 arg_tys = map Type (tyConAppArgs (idType case_bndr))
1104 -- case coerce T e of x { _ -> coerce T' x }
1105 -- And we definitely want to eliminate this case!
1106 -- So we throw away notes from the RHS, and reconstruct
1107 -- (at least an approximation) at the other end
1108 de_note (Note _ e) = de_note e
1111 -- re_note wraps a coerce if it might be necessary
1112 re_note scrut = case head alts of
1113 (_,_,rhs1@(Note _ _)) -> mkCoerce (exprType rhs1) (idType case_bndr) scrut
1117 --------------------------------------------------
1119 --------------------------------------------------
1120 mkCase1 scrut bndr alts = returnSmpl (Case scrut bndr alts)
1124 When adding auxiliary bindings for the case binder, it's worth checking if
1125 its dead, because it often is, and occasionally these mkCase transformations
1126 cascade rather nicely.
1129 bindCaseBndr bndr rhs body
1130 | isDeadBinder bndr = body
1131 | otherwise = bindNonRec bndr rhs body