2 % (c) The AQUA Project, Glasgow University, 1993-1998
4 \section[SimplUtils]{The simplifier utilities}
8 simplBinder, simplBinders, simplRecBndrs, simplLetBndr,
9 simplLamBndrs, simplTopBndrs,
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, isLocalId,
34 mkSysLocal, hasNoBinding, 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, isDataTyCon )
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 simplTopBndrs, simplLamBndrs, simplRecBndrs
451 :: SimplEnv -> [InBinder] -> SimplM (SimplEnv, [OutBinder])
452 simplTopBndrs = simplBndrs simplTopBinder
453 simplRecBndrs = simplBndrs Subst.simplLetId
454 simplLamBndrs = simplBndrs Subst.simplLamBndr
456 -- For top-level binders, don't use simplLetId for GlobalIds.
457 -- There are some of these, notably consructor wrappers, and we don't
458 -- want to clone them or fiddle with them at all.
459 -- Rather tiresomely, the specialiser may float a use of a constructor
460 -- wrapper to before its definition (which shouldn't really matter)
461 -- because it doesn't see the constructor wrapper as free in the binding
462 -- it is floating (because it's a GlobalId).
463 -- Then the simplifier brings all top level Ids into scope at the
464 -- beginning, and we don't want to lose the IdInfo on the constructor
465 -- wrappers. It would also be Bad to clone it!
466 simplTopBinder subst bndr
467 | isLocalId bndr = Subst.simplLetId subst bndr
468 | otherwise = (subst, bndr)
470 simplBndrs simpl_bndr env bndrs
472 (subst', bndrs') = mapAccumL simpl_bndr (getSubst env) bndrs
474 seqBndrs bndrs' `seq`
475 returnSmpl (setSubst env subst', bndrs')
478 seqBndrs (b:bs) = seqBndr b `seq` seqBndrs bs
480 seqBndr b | isTyVar b = b `seq` ()
481 | otherwise = seqType (idType b) `seq`
488 newId :: UserFS -> Type -> SimplM Id
489 newId fs ty = getUniqueSmpl `thenSmpl` \ uniq ->
490 returnSmpl (mkSysLocal fs uniq ty)
494 %************************************************************************
496 \subsection{Rebuilding a lambda}
498 %************************************************************************
501 mkLam :: SimplEnv -> [OutBinder] -> OutExpr -> SimplCont -> SimplM FloatsWithExpr
505 a) eta reduction, if that gives a trivial expression
506 b) eta expansion [only if there are some value lambdas]
507 c) floating lets out through big lambdas
508 [only if all tyvar lambdas, and only if this lambda
512 mkLam env bndrs body cont
513 | opt_SimplDoEtaReduction,
514 Just etad_lam <- tryEtaReduce bndrs body
515 = tick (EtaReduction (head bndrs)) `thenSmpl_`
516 returnSmpl (emptyFloats env, etad_lam)
518 | opt_SimplDoLambdaEtaExpansion,
519 any isRuntimeVar bndrs
520 = tryEtaExpansion body `thenSmpl` \ body' ->
521 returnSmpl (emptyFloats env, mkLams bndrs body')
523 {- Sept 01: I'm experimenting with getting the
524 full laziness pass to float out past big lambdsa
525 | all isTyVar bndrs, -- Only for big lambdas
526 contIsRhs cont -- Only try the rhs type-lambda floating
527 -- if this is indeed a right-hand side; otherwise
528 -- we end up floating the thing out, only for float-in
529 -- to float it right back in again!
530 = tryRhsTyLam env bndrs body `thenSmpl` \ (floats, body') ->
531 returnSmpl (floats, mkLams bndrs body')
535 = returnSmpl (emptyFloats env, mkLams bndrs body)
539 %************************************************************************
541 \subsection{Eta expansion and reduction}
543 %************************************************************************
545 We try for eta reduction here, but *only* if we get all the
546 way to an exprIsTrivial expression.
547 We don't want to remove extra lambdas unless we are going
548 to avoid allocating this thing altogether
551 tryEtaReduce :: [OutBinder] -> OutExpr -> Maybe OutExpr
552 tryEtaReduce bndrs body
553 -- We don't use CoreUtils.etaReduce, because we can be more
555 -- (a) we already have the binders
556 -- (b) we can do the triviality test before computing the free vars
557 -- [in fact I take the simple path and look for just a variable]
558 = go (reverse bndrs) body
560 go (b : bs) (App fun arg) | ok_arg b arg = go bs fun -- Loop round
561 go [] (Var fun) | ok_fun fun = Just (Var fun) -- Success!
562 go _ _ = Nothing -- Failure!
564 ok_fun fun = not (fun `elem` bndrs) && not (hasNoBinding fun)
565 -- Note the awkward "hasNoBinding" test
566 -- Details with exprIsTrivial
567 ok_arg b arg = varToCoreExpr b `cheapEqExpr` arg
571 Try eta expansion for RHSs
574 f = \x1..xn -> N ==> f = \x1..xn y1..ym -> N y1..ym
577 where (in both cases)
579 * The xi can include type variables
581 * The yi are all value variables
583 * N is a NORMAL FORM (i.e. no redexes anywhere)
584 wanting a suitable number of extra args.
586 We may have to sandwich some coerces between the lambdas
587 to make the types work. exprEtaExpandArity looks through coerces
588 when computing arity; and etaExpand adds the coerces as necessary when
589 actually computing the expansion.
592 tryEtaExpansion :: OutExpr -> SimplM OutExpr
593 -- There is at least one runtime binder in the binders
595 = getUniquesSmpl `thenSmpl` \ us ->
596 returnSmpl (etaExpand fun_arity us body (exprType body))
598 fun_arity = exprEtaExpandArity body
602 %************************************************************************
604 \subsection{Floating lets out of big lambdas}
606 %************************************************************************
608 tryRhsTyLam tries this transformation, when the big lambda appears as
609 the RHS of a let(rec) binding:
611 /\abc -> let(rec) x = e in b
613 let(rec) x' = /\abc -> let x = x' a b c in e
615 /\abc -> let x = x' a b c in b
617 This is good because it can turn things like:
619 let f = /\a -> letrec g = ... g ... in g
621 letrec g' = /\a -> ... g' a ...
625 which is better. In effect, it means that big lambdas don't impede
628 This optimisation is CRUCIAL in eliminating the junk introduced by
629 desugaring mutually recursive definitions. Don't eliminate it lightly!
631 So far as the implementation is concerned:
633 Invariant: go F e = /\tvs -> F e
637 = Let x' = /\tvs -> F e
641 G = F . Let x = x' tvs
643 go F (Letrec xi=ei in b)
644 = Letrec {xi' = /\tvs -> G ei}
648 G = F . Let {xi = xi' tvs}
650 [May 1999] If we do this transformation *regardless* then we can
651 end up with some pretty silly stuff. For example,
654 st = /\ s -> let { x1=r1 ; x2=r2 } in ...
659 st = /\s -> ...[y1 s/x1, y2 s/x2]
662 Unless the "..." is a WHNF there is really no point in doing this.
663 Indeed it can make things worse. Suppose x1 is used strictly,
666 x1* = case f y of { (a,b) -> e }
668 If we abstract this wrt the tyvar we then can't do the case inline
669 as we would normally do.
673 {- Trying to do this in full laziness
675 tryRhsTyLam :: SimplEnv -> [OutTyVar] -> OutExpr -> SimplM FloatsWithExpr
676 -- Call ensures that all the binders are type variables
678 tryRhsTyLam env tyvars body -- Only does something if there's a let
679 | not (all isTyVar tyvars)
680 || not (worth_it body) -- inside a type lambda,
681 = returnSmpl (emptyFloats env, body) -- and a WHNF inside that
684 = go env (\x -> x) body
687 worth_it e@(Let _ _) = whnf_in_middle e
690 whnf_in_middle (Let (NonRec x rhs) e) | isUnLiftedType (idType x) = False
691 whnf_in_middle (Let _ e) = whnf_in_middle e
692 whnf_in_middle e = exprIsCheap e
694 main_tyvar_set = mkVarSet tyvars
696 go env fn (Let bind@(NonRec var rhs) body)
698 = go env (fn . Let bind) body
700 go env fn (Let (NonRec var rhs) body)
701 = mk_poly tyvars_here var `thenSmpl` \ (var', rhs') ->
702 addAuxiliaryBind env (NonRec var' (mkLams tyvars_here (fn rhs))) $ \ env ->
703 go env (fn . Let (mk_silly_bind var rhs')) body
707 tyvars_here = varSetElems (main_tyvar_set `intersectVarSet` exprSomeFreeVars isTyVar rhs)
708 -- Abstract only over the type variables free in the rhs
709 -- wrt which the new binding is abstracted. But the naive
710 -- approach of abstract wrt the tyvars free in the Id's type
712 -- /\ a b -> let t :: (a,b) = (e1, e2)
715 -- Here, b isn't free in x's type, but we must nevertheless
716 -- abstract wrt b as well, because t's type mentions b.
717 -- Since t is floated too, we'd end up with the bogus:
718 -- poly_t = /\ a b -> (e1, e2)
719 -- poly_x = /\ a -> fst (poly_t a *b*)
720 -- So for now we adopt the even more naive approach of
721 -- abstracting wrt *all* the tyvars. We'll see if that
722 -- gives rise to problems. SLPJ June 98
724 go env fn (Let (Rec prs) body)
725 = mapAndUnzipSmpl (mk_poly tyvars_here) vars `thenSmpl` \ (vars', rhss') ->
727 gn body = fn (foldr Let body (zipWith mk_silly_bind vars rhss'))
728 pairs = vars' `zip` [mkLams tyvars_here (gn rhs) | rhs <- rhss]
730 addAuxiliaryBind env (Rec pairs) $ \ env ->
733 (vars,rhss) = unzip prs
734 tyvars_here = varSetElems (main_tyvar_set `intersectVarSet` exprsSomeFreeVars isTyVar (map snd prs))
735 -- See notes with tyvars_here above
737 go env fn body = returnSmpl (emptyFloats env, fn body)
739 mk_poly tyvars_here var
740 = getUniqueSmpl `thenSmpl` \ uniq ->
742 poly_name = setNameUnique (idName var) uniq -- Keep same name
743 poly_ty = mkForAllTys tyvars_here (idType var) -- But new type of course
744 poly_id = mkLocalId poly_name poly_ty
746 -- In the olden days, it was crucial to copy the occInfo of the original var,
747 -- because we were looking at occurrence-analysed but as yet unsimplified code!
748 -- In particular, we mustn't lose the loop breakers. BUT NOW we are looking
749 -- at already simplified code, so it doesn't matter
751 -- It's even right to retain single-occurrence or dead-var info:
752 -- Suppose we started with /\a -> let x = E in B
753 -- where x occurs once in B. Then we transform to:
754 -- let x' = /\a -> E in /\a -> let x* = x' a in B
755 -- where x* has an INLINE prag on it. Now, once x* is inlined,
756 -- the occurrences of x' will be just the occurrences originally
759 returnSmpl (poly_id, mkTyApps (Var poly_id) (mkTyVarTys tyvars_here))
761 mk_silly_bind var rhs = NonRec var (Note InlineMe rhs)
762 -- Suppose we start with:
764 -- x = /\ a -> let g = G in E
766 -- Then we'll float to get
768 -- x = let poly_g = /\ a -> G
769 -- in /\ a -> let g = poly_g a in E
771 -- But now the occurrence analyser will see just one occurrence
772 -- of poly_g, not inside a lambda, so the simplifier will
773 -- PreInlineUnconditionally poly_g back into g! Badk to square 1!
774 -- (I used to think that the "don't inline lone occurrences" stuff
775 -- would stop this happening, but since it's the *only* occurrence,
776 -- PreInlineUnconditionally kicks in first!)
778 -- Solution: put an INLINE note on g's RHS, so that poly_g seems
779 -- to appear many times. (NB: mkInlineMe eliminates
780 -- such notes on trivial RHSs, so do it manually.)
785 %************************************************************************
787 \subsection{Case absorption and identity-case elimination}
789 %************************************************************************
791 mkCase puts a case expression back together, trying various transformations first.
794 mkCase :: OutExpr -> OutId -> [OutAlt] -> SimplM OutExpr
796 mkCase scrut case_bndr alts
797 = mkAlts scrut case_bndr alts `thenSmpl` \ better_alts ->
798 mkCase1 scrut case_bndr better_alts
802 mkAlts tries these things:
804 1. If several alternatives are identical, merge them into
805 a single DEFAULT alternative. I've occasionally seen this
806 making a big difference:
808 case e of =====> case e of
809 C _ -> f x D v -> ....v....
810 D v -> ....v.... DEFAULT -> f x
813 The point is that we merge common RHSs, at least for the DEFAULT case.
814 [One could do something more elaborate but I've never seen it needed.]
815 To avoid an expensive test, we just merge branches equal to the *first*
816 alternative; this picks up the common cases
817 a) all branches equal
818 b) some branches equal to the DEFAULT (which occurs first)
820 2. If the DEFAULT alternative can match only one possible constructor,
821 then make that constructor explicit.
823 case e of x { DEFAULT -> rhs }
825 case e of x { (a,b) -> rhs }
826 where the type is a single constructor type. This gives better code
827 when rhs also scrutinises x or e.
830 case e of b { ==> case e of b {
831 p1 -> rhs1 p1 -> rhs1
833 pm -> rhsm pm -> rhsm
834 _ -> case b of b' { pn -> let b'=b in rhsn
836 ... po -> let b'=b in rhso
837 po -> rhso _ -> let b'=b in rhsd
841 which merges two cases in one case when -- the default alternative of
842 the outer case scrutises the same variable as the outer case This
843 transformation is called Case Merging. It avoids that the same
844 variable is scrutinised multiple times.
847 The case where transformation (1) showed up was like this (lib/std/PrelCError.lhs):
853 where @is@ was something like
855 p `is` n = p /= (-1) && p == n
857 This gave rise to a horrible sequence of cases
864 and similarly in cascade for all the join points!
869 --------------------------------------------------
870 -- 1. Merge identical branches
871 --------------------------------------------------
872 mkAlts scrut case_bndr alts@((con1,bndrs1,rhs1) : con_alts)
873 | all isDeadBinder bndrs1, -- Remember the default
874 length filtered_alts < length con_alts -- alternative comes first
875 = tick (AltMerge case_bndr) `thenSmpl_`
876 returnSmpl better_alts
878 filtered_alts = filter keep con_alts
879 keep (con,bndrs,rhs) = not (all isDeadBinder bndrs && rhs `cheapEqExpr` rhs1)
880 better_alts = (DEFAULT, [], rhs1) : filtered_alts
883 --------------------------------------------------
884 -- 2. Fill in missing constructor
885 --------------------------------------------------
887 mkAlts scrut case_bndr alts
888 | Just (tycon, inst_tys) <- splitTyConApp_maybe (idType case_bndr),
889 isDataTyCon tycon, -- It's a data type
890 (alts_no_deflt, Just rhs) <- findDefault alts,
891 -- There is a DEFAULT case
892 [missing_con] <- filter is_missing (tyConDataConsIfAvailable tycon)
893 -- There is just one missing constructor!
894 = tick (FillInCaseDefault case_bndr) `thenSmpl_`
895 getUniquesSmpl `thenSmpl` \ tv_uniqs ->
896 getUniquesSmpl `thenSmpl` \ id_uniqs ->
898 (_,_,ex_tyvars,_,_,_) = dataConSig missing_con
899 ex_tyvars' = zipWith mk tv_uniqs ex_tyvars
900 mk uniq tv = mkSysTyVar uniq (tyVarKind tv)
901 arg_ids = zipWith (mkSysLocal SLIT("a")) id_uniqs arg_tys
902 arg_tys = dataConArgTys missing_con (inst_tys ++ mkTyVarTys ex_tyvars')
903 better_alts = (DataAlt missing_con, ex_tyvars' ++ arg_ids, rhs) : alts_no_deflt
905 returnSmpl better_alts
907 impossible_cons = otherCons (idUnfolding case_bndr)
908 handled_data_cons = [data_con | DataAlt data_con <- impossible_cons] ++
909 [data_con | (DataAlt data_con, _, _) <- alts]
910 is_missing con = not (con `elem` handled_data_cons)
912 --------------------------------------------------
913 -- 3. Merge nested cases
914 --------------------------------------------------
916 mkAlts scrut outer_bndr outer_alts
917 | opt_SimplCaseMerge,
918 (outer_alts_without_deflt, maybe_outer_deflt) <- findDefault outer_alts,
919 Just (Case (Var scrut_var) inner_bndr inner_alts) <- maybe_outer_deflt,
920 scruting_same_var scrut_var
922 = let -- Eliminate any inner alts which are shadowed by the outer ones
923 outer_cons = [con | (con,_,_) <- outer_alts_without_deflt]
925 munged_inner_alts = [ (con, args, munge_rhs rhs)
926 | (con, args, rhs) <- inner_alts,
927 not (con `elem` outer_cons) -- Eliminate shadowed inner alts
929 munge_rhs rhs = bindCaseBndr inner_bndr (Var outer_bndr) rhs
931 (inner_con_alts, maybe_inner_default) = findDefault munged_inner_alts
933 new_alts = add_default maybe_inner_default
934 (outer_alts_without_deflt ++ inner_con_alts)
936 tick (CaseMerge outer_bndr) `thenSmpl_`
938 -- Warning: don't call mkAlts recursively!
939 -- Firstly, there's no point, because inner alts have already had
940 -- mkCase applied to them, so they won't have a case in their default
941 -- Secondly, if you do, you get an infinite loop, because the bindCaseBndr
942 -- in munge_rhs may put a case into the DEFAULT branch!
944 -- We are scrutinising the same variable if it's
945 -- the outer case-binder, or if the outer case scrutinises a variable
946 -- (and it's the same). Testing both allows us not to replace the
947 -- outer scrut-var with the outer case-binder (Simplify.simplCaseBinder).
948 scruting_same_var = case scrut of
949 Var outer_scrut -> \ v -> v == outer_bndr || v == outer_scrut
950 other -> \ v -> v == outer_bndr
952 add_default (Just rhs) alts = (DEFAULT,[],rhs) : alts
953 add_default Nothing alts = alts
956 --------------------------------------------------
958 --------------------------------------------------
960 mkAlts scrut case_bndr other_alts = returnSmpl other_alts
965 =================================================================================
967 mkCase1 tries these things
969 1. Eliminate the case altogether if possible
980 Start with a simple situation:
982 case x# of ===> e[x#/y#]
985 (when x#, y# are of primitive type, of course). We can't (in general)
986 do this for algebraic cases, because we might turn bottom into
989 Actually, we generalise this idea to look for a case where we're
990 scrutinising a variable, and we know that only the default case can
995 other -> ...(case x of
999 Here the inner case can be eliminated. This really only shows up in
1000 eliminating error-checking code.
1002 We also make sure that we deal with this very common case:
1007 Here we are using the case as a strict let; if x is used only once
1008 then we want to inline it. We have to be careful that this doesn't
1009 make the program terminate when it would have diverged before, so we
1011 - x is used strictly, or
1012 - e is already evaluated (it may so if e is a variable)
1014 Lastly, we generalise the transformation to handle this:
1020 We only do this for very cheaply compared r's (constructors, literals
1021 and variables). If pedantic bottoms is on, we only do it when the
1022 scrutinee is a PrimOp which can't fail.
1024 We do it *here*, looking at un-simplified alternatives, because we
1025 have to check that r doesn't mention the variables bound by the
1026 pattern in each alternative, so the binder-info is rather useful.
1028 So the case-elimination algorithm is:
1030 1. Eliminate alternatives which can't match
1032 2. Check whether all the remaining alternatives
1033 (a) do not mention in their rhs any of the variables bound in their pattern
1034 and (b) have equal rhss
1036 3. Check we can safely ditch the case:
1037 * PedanticBottoms is off,
1038 or * the scrutinee is an already-evaluated variable
1039 or * the scrutinee is a primop which is ok for speculation
1040 -- ie we want to preserve divide-by-zero errors, and
1041 -- calls to error itself!
1043 or * [Prim cases] the scrutinee is a primitive variable
1045 or * [Alg cases] the scrutinee is a variable and
1046 either * the rhs is the same variable
1047 (eg case x of C a b -> x ===> x)
1048 or * there is only one alternative, the default alternative,
1049 and the binder is used strictly in its scope.
1050 [NB this is helped by the "use default binder where
1051 possible" transformation; see below.]
1054 If so, then we can replace the case with one of the rhss.
1058 --------------------------------------------------
1059 -- 1. Eliminate the case altogether if poss
1060 --------------------------------------------------
1062 mkCase1 scrut case_bndr [(con,bndrs,rhs)]
1063 -- See if we can get rid of the case altogether
1064 -- See the extensive notes on case-elimination above
1065 -- mkCase made sure that if all the alternatives are equal,
1066 -- then there is now only one (DEFAULT) rhs
1067 | all isDeadBinder bndrs,
1069 -- Check that the scrutinee can be let-bound instead of case-bound
1070 exprOkForSpeculation scrut
1071 -- OK not to evaluate it
1072 -- This includes things like (==# a# b#)::Bool
1073 -- so that we simplify
1074 -- case ==# a# b# of { True -> x; False -> x }
1077 -- This particular example shows up in default methods for
1078 -- comparision operations (e.g. in (>=) for Int.Int32)
1079 || exprIsValue scrut -- It's already evaluated
1080 || var_demanded_later scrut -- It'll be demanded later
1082 -- || not opt_SimplPedanticBottoms) -- Or we don't care!
1083 -- We used to allow improving termination by discarding cases, unless -fpedantic-bottoms was on,
1084 -- but that breaks badly for the dataToTag# primop, which relies on a case to evaluate
1085 -- its argument: case x of { y -> dataToTag# y }
1086 -- Here we must *not* discard the case, because dataToTag# just fetches the tag from
1087 -- the info pointer. So we'll be pedantic all the time, and see if that gives any
1089 = tick (CaseElim case_bndr) `thenSmpl_`
1090 returnSmpl (bindCaseBndr case_bndr scrut rhs)
1093 -- The case binder is going to be evaluated later,
1094 -- and the scrutinee is a simple variable
1095 var_demanded_later (Var v) = isStrictDmd (idNewDemandInfo case_bndr)
1096 var_demanded_later other = False
1099 --------------------------------------------------
1101 --------------------------------------------------
1103 mkCase1 scrut case_bndr alts -- Identity case
1104 | all identity_alt alts
1105 = tick (CaseIdentity case_bndr) `thenSmpl_`
1106 returnSmpl (re_note scrut)
1108 identity_alt (con, args, rhs) = de_note rhs `cheapEqExpr` identity_rhs con args
1110 identity_rhs (DataAlt con) args = mkConApp con (arg_tys ++ map varToCoreExpr args)
1111 identity_rhs (LitAlt lit) _ = Lit lit
1112 identity_rhs DEFAULT _ = Var case_bndr
1114 arg_tys = map Type (tyConAppArgs (idType case_bndr))
1117 -- case coerce T e of x { _ -> coerce T' x }
1118 -- And we definitely want to eliminate this case!
1119 -- So we throw away notes from the RHS, and reconstruct
1120 -- (at least an approximation) at the other end
1121 de_note (Note _ e) = de_note e
1124 -- re_note wraps a coerce if it might be necessary
1125 re_note scrut = case head alts of
1126 (_,_,rhs1@(Note _ _)) -> mkCoerce (exprType rhs1) (idType case_bndr) scrut
1130 --------------------------------------------------
1132 --------------------------------------------------
1133 mkCase1 scrut bndr alts = returnSmpl (Case scrut bndr alts)
1137 When adding auxiliary bindings for the case binder, it's worth checking if
1138 its dead, because it often is, and occasionally these mkCase transformations
1139 cascade rather nicely.
1142 bindCaseBndr bndr rhs body
1143 | isDeadBinder bndr = body
1144 | otherwise = bindNonRec bndr rhs body