2 % (c) The AQUA Project, Glasgow University, 1993-1998
4 \section[SimplUtils]{The simplifier utilities}
8 simplBinder, simplBinders, simplRecBndrs,
9 simplLetBndr, simplLamBndrs,
10 newId, mkLam, prepareAlts, mkCase,
12 -- The continuation type
13 SimplCont(..), DupFlag(..), LetRhsFlag(..),
14 contIsDupable, contResultType,
15 countValArgs, countArgs, pushContArgs,
16 mkBoringStop, mkStop, contIsRhs, contIsRhsOrArg,
17 getContArgs, interestingCallContext, interestingArg, isStrictType
21 #include "HsVersions.h"
23 import CmdLineOpts ( SimplifierSwitch(..),
24 opt_SimplDoLambdaEtaExpansion, opt_SimplDoEtaReduction,
25 opt_SimplCaseMerge, opt_UF_UpdateInPlace
28 import CoreFVs ( exprFreeVars )
29 import CoreUtils ( cheapEqExpr, exprType, exprIsTrivial,
30 etaExpand, exprEtaExpandArity, bindNonRec, mkCoerce2,
31 findDefault, exprOkForSpeculation, exprIsValue
33 import qualified Subst ( simplBndrs, simplBndr, simplLetId, simplLamBndr )
34 import Id ( Id, idType, idInfo, isDataConWorkId,
35 mkSysLocal, isDeadBinder, idNewDemandInfo,
36 idUnfolding, idNewStrictness
38 import NewDemand ( isStrictDmd, isBotRes, splitStrictSig )
40 import Type ( Type, seqType, splitFunTys, dropForAlls, isStrictType,
41 splitTyConApp_maybe, tyConAppArgs, mkTyVarTys
43 import TcType ( isDictTy )
44 import OccName ( EncodedFS )
45 import TyCon ( tyConDataCons_maybe, isAlgTyCon, isNewTyCon )
46 import DataCon ( dataConRepArity, dataConExistentialTyVars, dataConArgTys )
47 import Var ( mkSysTyVar, tyVarKind )
49 import Util ( lengthExceeds, mapAccumL )
54 %************************************************************************
56 \subsection{The continuation data type}
58 %************************************************************************
61 data SimplCont -- Strict contexts
62 = Stop OutType -- Type of the result
64 Bool -- True <=> This is the RHS of a thunk whose type suggests
65 -- that update-in-place would be possible
66 -- (This makes the inliner a little keener.)
68 | CoerceIt OutType -- The To-type, simplified
71 | InlinePlease -- This continuation makes a function very
72 SimplCont -- keen to inline itelf
75 InExpr SimplEnv -- The argument, as yet unsimplified,
76 SimplCont -- and its environment
79 InId [InAlt] SimplEnv -- The case binder, alts, and subst-env
82 | ArgOf LetRhsFlag -- An arbitrary strict context: the argument
83 -- of a strict function, or a primitive-arg fn
85 -- No DupFlag because we never duplicate it
86 OutType -- arg_ty: type of the argument itself
87 OutType -- cont_ty: the type of the expression being sought by the context
88 -- f (error "foo") ==> coerce t (error "foo")
90 -- We need to know the type t, to which to coerce.
92 (SimplEnv -> OutExpr -> SimplM FloatsWithExpr) -- What to do with the result
93 -- The result expression in the OutExprStuff has type cont_ty
95 data LetRhsFlag = AnArg -- It's just an argument not a let RHS
96 | AnRhs -- It's the RHS of a let (so please float lets out of big lambdas)
98 instance Outputable LetRhsFlag where
99 ppr AnArg = ptext SLIT("arg")
100 ppr AnRhs = ptext SLIT("rhs")
102 instance Outputable SimplCont where
103 ppr (Stop _ is_rhs _) = ptext SLIT("Stop") <> brackets (ppr is_rhs)
104 ppr (ApplyTo dup arg se cont) = (ptext SLIT("ApplyTo") <+> ppr dup <+> ppr arg) $$ ppr cont
105 ppr (ArgOf _ _ _ _) = ptext SLIT("ArgOf...")
106 ppr (Select dup bndr alts se cont) = (ptext SLIT("Select") <+> ppr dup <+> ppr bndr) $$
107 (nest 4 (ppr alts)) $$ ppr cont
108 ppr (CoerceIt ty cont) = (ptext SLIT("CoerceIt") <+> ppr ty) $$ ppr cont
109 ppr (InlinePlease cont) = ptext SLIT("InlinePlease") $$ ppr cont
111 data DupFlag = OkToDup | NoDup
113 instance Outputable DupFlag where
114 ppr OkToDup = ptext SLIT("ok")
115 ppr NoDup = ptext SLIT("nodup")
119 mkBoringStop :: OutType -> SimplCont
120 mkBoringStop ty = Stop ty AnArg (canUpdateInPlace ty)
122 mkStop :: OutType -> LetRhsFlag -> SimplCont
123 mkStop ty is_rhs = Stop ty is_rhs (canUpdateInPlace ty)
125 contIsRhs :: SimplCont -> Bool
126 contIsRhs (Stop _ AnRhs _) = True
127 contIsRhs (ArgOf AnRhs _ _ _) = True
128 contIsRhs other = False
130 contIsRhsOrArg (Stop _ _ _) = True
131 contIsRhsOrArg (ArgOf _ _ _ _) = True
132 contIsRhsOrArg other = False
135 contIsDupable :: SimplCont -> Bool
136 contIsDupable (Stop _ _ _) = True
137 contIsDupable (ApplyTo OkToDup _ _ _) = True
138 contIsDupable (Select OkToDup _ _ _ _) = True
139 contIsDupable (CoerceIt _ cont) = contIsDupable cont
140 contIsDupable (InlinePlease cont) = contIsDupable cont
141 contIsDupable other = False
144 discardableCont :: SimplCont -> Bool
145 discardableCont (Stop _ _ _) = False
146 discardableCont (CoerceIt _ cont) = discardableCont cont
147 discardableCont (InlinePlease cont) = discardableCont cont
148 discardableCont other = True
150 discardCont :: SimplCont -- A continuation, expecting
151 -> SimplCont -- Replace the continuation with a suitable coerce
152 discardCont cont = case cont of
153 Stop to_ty is_rhs _ -> cont
154 other -> CoerceIt to_ty (mkBoringStop to_ty)
156 to_ty = contResultType cont
159 contResultType :: SimplCont -> OutType
160 contResultType (Stop to_ty _ _) = to_ty
161 contResultType (ArgOf _ _ to_ty _) = to_ty
162 contResultType (ApplyTo _ _ _ cont) = contResultType cont
163 contResultType (CoerceIt _ cont) = contResultType cont
164 contResultType (InlinePlease cont) = contResultType cont
165 contResultType (Select _ _ _ _ cont) = contResultType cont
168 countValArgs :: SimplCont -> Int
169 countValArgs (ApplyTo _ (Type ty) se cont) = countValArgs cont
170 countValArgs (ApplyTo _ val_arg se cont) = 1 + countValArgs cont
171 countValArgs other = 0
173 countArgs :: SimplCont -> Int
174 countArgs (ApplyTo _ arg se cont) = 1 + countArgs cont
178 pushContArgs :: SimplEnv -> [OutArg] -> SimplCont -> SimplCont
179 -- Pushes args with the specified environment
180 pushContArgs env [] cont = cont
181 pushContArgs env (arg : args) cont = ApplyTo NoDup arg env (pushContArgs env args cont)
186 getContArgs :: SwitchChecker
187 -> OutId -> SimplCont
188 -> ([(InExpr, SimplEnv, Bool)], -- Arguments; the Bool is true for strict args
189 SimplCont, -- Remaining continuation
190 Bool) -- Whether we came across an InlineCall
191 -- getContArgs id k = (args, k', inl)
192 -- args are the leading ApplyTo items in k
193 -- (i.e. outermost comes first)
194 -- augmented with demand info from the functionn
195 getContArgs chkr fun orig_cont
197 -- Ignore strictness info if the no-case-of-case
198 -- flag is on. Strictness changes evaluation order
199 -- and that can change full laziness
200 stricts | switchIsOn chkr NoCaseOfCase = vanilla_stricts
201 | otherwise = computed_stricts
203 go [] stricts False orig_cont
205 ----------------------------
208 go acc ss inl (ApplyTo _ arg@(Type _) se cont)
209 = go ((arg,se,False) : acc) ss inl cont
210 -- NB: don't bother to instantiate the function type
213 go acc (s:ss) inl (ApplyTo _ arg se cont)
214 = go ((arg,se,s) : acc) ss inl cont
216 -- An Inline continuation
217 go acc ss inl (InlinePlease cont)
218 = go acc ss True cont
220 -- We're run out of arguments, or else we've run out of demands
221 -- The latter only happens if the result is guaranteed bottom
222 -- This is the case for
223 -- * case (error "hello") of { ... }
224 -- * (error "Hello") arg
225 -- * f (error "Hello") where f is strict
227 -- Then, especially in the first of these cases, we'd like to discard
228 -- the continuation, leaving just the bottoming expression. But the
229 -- type might not be right, so we may have to add a coerce.
231 | null ss && discardableCont cont = (reverse acc, discardCont cont, inl)
232 | otherwise = (reverse acc, cont, inl)
234 ----------------------------
235 vanilla_stricts, computed_stricts :: [Bool]
236 vanilla_stricts = repeat False
237 computed_stricts = zipWith (||) fun_stricts arg_stricts
239 ----------------------------
240 (val_arg_tys, _) = splitFunTys (dropForAlls (idType fun))
241 arg_stricts = map isStrictType val_arg_tys ++ repeat False
242 -- These argument types are used as a cheap and cheerful way to find
243 -- unboxed arguments, which must be strict. But it's an InType
244 -- and so there might be a type variable where we expect a function
245 -- type (the substitution hasn't happened yet). And we don't bother
246 -- doing the type applications for a polymorphic function.
247 -- Hence the splitFunTys*IgnoringForAlls*
249 ----------------------------
250 -- If fun_stricts is finite, it means the function returns bottom
251 -- after that number of value args have been consumed
252 -- Otherwise it's infinite, extended with False
254 = case splitStrictSig (idNewStrictness fun) of
255 (demands, result_info)
256 | not (demands `lengthExceeds` countValArgs orig_cont)
257 -> -- Enough args, use the strictness given.
258 -- For bottoming functions we used to pretend that the arg
259 -- is lazy, so that we don't treat the arg as an
260 -- interesting context. This avoids substituting
261 -- top-level bindings for (say) strings into
262 -- calls to error. But now we are more careful about
263 -- inlining lone variables, so its ok (see SimplUtils.analyseCont)
264 if isBotRes result_info then
265 map isStrictDmd demands -- Finite => result is bottom
267 map isStrictDmd demands ++ vanilla_stricts
269 other -> vanilla_stricts -- Not enough args, or no strictness
272 interestingArg :: OutExpr -> Bool
273 -- An argument is interesting if it has *some* structure
274 -- We are here trying to avoid unfolding a function that
275 -- is applied only to variables that have no unfolding
276 -- (i.e. they are probably lambda bound): f x y z
277 -- There is little point in inlining f here.
278 interestingArg (Var v) = hasSomeUnfolding (idUnfolding v)
279 -- Was: isValueUnfolding (idUnfolding v')
280 -- But that seems over-pessimistic
282 -- This accounts for an argument like
283 -- () or [], which is definitely interesting
284 interestingArg (Type _) = False
285 interestingArg (App fn (Type _)) = interestingArg fn
286 interestingArg (Note _ a) = interestingArg a
287 interestingArg other = True
288 -- Consider let x = 3 in f x
289 -- The substitution will contain (x -> ContEx 3), and we want to
290 -- to say that x is an interesting argument.
291 -- But consider also (\x. f x y) y
292 -- The substitution will contain (x -> ContEx y), and we want to say
293 -- that x is not interesting (assuming y has no unfolding)
296 Comment about interestingCallContext
297 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
298 We want to avoid inlining an expression where there can't possibly be
299 any gain, such as in an argument position. Hence, if the continuation
300 is interesting (eg. a case scrutinee, application etc.) then we
301 inline, otherwise we don't.
303 Previously some_benefit used to return True only if the variable was
304 applied to some value arguments. This didn't work:
306 let x = _coerce_ (T Int) Int (I# 3) in
307 case _coerce_ Int (T Int) x of
310 we want to inline x, but can't see that it's a constructor in a case
311 scrutinee position, and some_benefit is False.
315 dMonadST = _/\_ t -> :Monad (g1 _@_ t, g2 _@_ t, g3 _@_ t)
317 .... case dMonadST _@_ x0 of (a,b,c) -> ....
319 we'd really like to inline dMonadST here, but we *don't* want to
320 inline if the case expression is just
322 case x of y { DEFAULT -> ... }
324 since we can just eliminate this case instead (x is in WHNF). Similar
325 applies when x is bound to a lambda expression. Hence
326 contIsInteresting looks for case expressions with just a single
330 interestingCallContext :: Bool -- False <=> no args at all
331 -> Bool -- False <=> no value args
333 -- The "lone-variable" case is important. I spent ages
334 -- messing about with unsatisfactory varaints, but this is nice.
335 -- The idea is that if a variable appear all alone
336 -- as an arg of lazy fn, or rhs Stop
337 -- as scrutinee of a case Select
338 -- as arg of a strict fn ArgOf
339 -- then we should not inline it (unless there is some other reason,
340 -- e.g. is is the sole occurrence). We achieve this by making
341 -- interestingCallContext return False for a lone variable.
343 -- Why? At least in the case-scrutinee situation, turning
344 -- let x = (a,b) in case x of y -> ...
346 -- let x = (a,b) in case (a,b) of y -> ...
348 -- let x = (a,b) in let y = (a,b) in ...
349 -- is bad if the binding for x will remain.
351 -- Another example: I discovered that strings
352 -- were getting inlined straight back into applications of 'error'
353 -- because the latter is strict.
355 -- f = \x -> ...(error s)...
357 -- Fundamentally such contexts should not ecourage inlining because
358 -- the context can ``see'' the unfolding of the variable (e.g. case or a RULE)
359 -- so there's no gain.
361 -- However, even a type application or coercion isn't a lone variable.
363 -- case $fMonadST @ RealWorld of { :DMonad a b c -> c }
364 -- We had better inline that sucker! The case won't see through it.
366 -- For now, I'm treating treating a variable applied to types
367 -- in a *lazy* context "lone". The motivating example was
369 -- g = /\a. \y. h (f a)
370 -- There's no advantage in inlining f here, and perhaps
371 -- a significant disadvantage. Hence some_val_args in the Stop case
373 interestingCallContext some_args some_val_args cont
376 interesting (InlinePlease _) = True
377 interesting (Select _ _ _ _ _) = some_args
378 interesting (ApplyTo _ _ _ _) = True -- Can happen if we have (coerce t (f x)) y
379 -- Perhaps True is a bit over-keen, but I've
380 -- seen (coerce f) x, where f has an INLINE prag,
381 -- So we have to give some motivaiton for inlining it
382 interesting (ArgOf _ _ _ _) = some_val_args
383 interesting (Stop ty _ upd_in_place) = some_val_args && upd_in_place
384 interesting (CoerceIt _ cont) = interesting cont
385 -- If this call is the arg of a strict function, the context
386 -- is a bit interesting. If we inline here, we may get useful
387 -- evaluation information to avoid repeated evals: e.g.
389 -- Here the contIsInteresting makes the '*' keener to inline,
390 -- which in turn exposes a constructor which makes the '+' inline.
391 -- Assuming that +,* aren't small enough to inline regardless.
393 -- It's also very important to inline in a strict context for things
396 -- Here, the context of (f x) is strict, and if f's unfolding is
397 -- a build it's *great* to inline it here. So we must ensure that
398 -- the context for (f x) is not totally uninteresting.
402 canUpdateInPlace :: Type -> Bool
403 -- Consider let x = <wurble> in ...
404 -- If <wurble> returns an explicit constructor, we might be able
405 -- to do update in place. So we treat even a thunk RHS context
406 -- as interesting if update in place is possible. We approximate
407 -- this by seeing if the type has a single constructor with a
408 -- small arity. But arity zero isn't good -- we share the single copy
409 -- for that case, so no point in sharing.
412 | not opt_UF_UpdateInPlace = False
414 = case splitTyConApp_maybe ty of
416 Just (tycon, _) -> case tyConDataCons_maybe tycon of
417 Just [dc] -> arity == 1 || arity == 2
419 arity = dataConRepArity dc
425 %************************************************************************
427 \section{Dealing with a single binder}
429 %************************************************************************
431 These functions are in the monad only so that they can be made strict via seq.
434 simplBinders :: SimplEnv -> [InBinder] -> SimplM (SimplEnv, [OutBinder])
435 simplBinders env bndrs
437 (subst', bndrs') = Subst.simplBndrs (getSubst env) bndrs
439 seqBndrs bndrs' `seq`
440 returnSmpl (setSubst env subst', bndrs')
442 simplBinder :: SimplEnv -> InBinder -> SimplM (SimplEnv, OutBinder)
445 (subst', bndr') = Subst.simplBndr (getSubst env) bndr
448 returnSmpl (setSubst env subst', bndr')
451 simplLetBndr :: SimplEnv -> InBinder -> SimplM (SimplEnv, OutBinder)
454 (subst', id') = Subst.simplLetId (getSubst env) id
457 returnSmpl (setSubst env subst', id')
459 simplLamBndrs, simplRecBndrs
460 :: SimplEnv -> [InBinder] -> SimplM (SimplEnv, [OutBinder])
461 simplRecBndrs = simplBndrs Subst.simplLetId
462 simplLamBndrs = simplBndrs Subst.simplLamBndr
464 simplBndrs simpl_bndr env bndrs
466 (subst', bndrs') = mapAccumL simpl_bndr (getSubst env) bndrs
468 seqBndrs bndrs' `seq`
469 returnSmpl (setSubst env subst', bndrs')
472 seqBndrs (b:bs) = seqBndr b `seq` seqBndrs bs
474 seqBndr b | isTyVar b = b `seq` ()
475 | otherwise = seqType (idType b) `seq`
482 newId :: EncodedFS -> Type -> SimplM Id
483 newId fs ty = getUniqueSmpl `thenSmpl` \ uniq ->
484 returnSmpl (mkSysLocal fs uniq ty)
488 %************************************************************************
490 \subsection{Rebuilding a lambda}
492 %************************************************************************
495 mkLam :: SimplEnv -> [OutBinder] -> OutExpr -> SimplCont -> SimplM FloatsWithExpr
499 a) eta reduction, if that gives a trivial expression
500 b) eta expansion [only if there are some value lambdas]
501 c) floating lets out through big lambdas
502 [only if all tyvar lambdas, and only if this lambda
506 mkLam env bndrs body cont
507 | opt_SimplDoEtaReduction,
508 Just etad_lam <- tryEtaReduce bndrs body
509 = tick (EtaReduction (head bndrs)) `thenSmpl_`
510 returnSmpl (emptyFloats env, etad_lam)
512 | opt_SimplDoLambdaEtaExpansion,
513 any isRuntimeVar bndrs
514 = tryEtaExpansion body `thenSmpl` \ body' ->
515 returnSmpl (emptyFloats env, mkLams bndrs body')
517 {- Sept 01: I'm experimenting with getting the
518 full laziness pass to float out past big lambdsa
519 | all isTyVar bndrs, -- Only for big lambdas
520 contIsRhs cont -- Only try the rhs type-lambda floating
521 -- if this is indeed a right-hand side; otherwise
522 -- we end up floating the thing out, only for float-in
523 -- to float it right back in again!
524 = tryRhsTyLam env bndrs body `thenSmpl` \ (floats, body') ->
525 returnSmpl (floats, mkLams bndrs body')
529 = returnSmpl (emptyFloats env, mkLams bndrs body)
533 %************************************************************************
535 \subsection{Eta expansion and reduction}
537 %************************************************************************
539 We try for eta reduction here, but *only* if we get all the
540 way to an exprIsTrivial expression.
541 We don't want to remove extra lambdas unless we are going
542 to avoid allocating this thing altogether
545 tryEtaReduce :: [OutBinder] -> OutExpr -> Maybe OutExpr
546 tryEtaReduce bndrs body
547 -- We don't use CoreUtils.etaReduce, because we can be more
549 -- (a) we already have the binders
550 -- (b) we can do the triviality test before computing the free vars
551 = go (reverse bndrs) body
553 go (b : bs) (App fun arg) | ok_arg b arg = go bs fun -- Loop round
554 go [] fun | ok_fun fun = Just fun -- Success!
555 go _ _ = Nothing -- Failure!
557 ok_fun fun = exprIsTrivial fun
558 && not (any (`elemVarSet` (exprFreeVars fun)) bndrs)
559 && (exprIsValue fun || all ok_lam bndrs)
560 ok_lam v = isTyVar v || isDictTy (idType v)
561 -- The exprIsValue is because eta reduction is not
562 -- valid in general: \x. bot /= bot
563 -- So we need to be sure that the "fun" is a value.
565 -- However, we always want to reduce (/\a -> f a) to f
566 -- This came up in a RULE: foldr (build (/\a -> g a))
567 -- did not match foldr (build (/\b -> ...something complex...))
568 -- The type checker can insert these eta-expanded versions,
569 -- with both type and dictionary lambdas; hence the slightly
572 ok_arg b arg = varToCoreExpr b `cheapEqExpr` arg
576 Try eta expansion for RHSs
579 f = \x1..xn -> N ==> f = \x1..xn y1..ym -> N y1..ym
582 where (in both cases)
584 * The xi can include type variables
586 * The yi are all value variables
588 * N is a NORMAL FORM (i.e. no redexes anywhere)
589 wanting a suitable number of extra args.
591 We may have to sandwich some coerces between the lambdas
592 to make the types work. exprEtaExpandArity looks through coerces
593 when computing arity; and etaExpand adds the coerces as necessary when
594 actually computing the expansion.
597 tryEtaExpansion :: OutExpr -> SimplM OutExpr
598 -- There is at least one runtime binder in the binders
600 = getUniquesSmpl `thenSmpl` \ us ->
601 returnSmpl (etaExpand fun_arity us body (exprType body))
603 fun_arity = exprEtaExpandArity body
607 %************************************************************************
609 \subsection{Floating lets out of big lambdas}
611 %************************************************************************
613 tryRhsTyLam tries this transformation, when the big lambda appears as
614 the RHS of a let(rec) binding:
616 /\abc -> let(rec) x = e in b
618 let(rec) x' = /\abc -> let x = x' a b c in e
620 /\abc -> let x = x' a b c in b
622 This is good because it can turn things like:
624 let f = /\a -> letrec g = ... g ... in g
626 letrec g' = /\a -> ... g' a ...
630 which is better. In effect, it means that big lambdas don't impede
633 This optimisation is CRUCIAL in eliminating the junk introduced by
634 desugaring mutually recursive definitions. Don't eliminate it lightly!
636 So far as the implementation is concerned:
638 Invariant: go F e = /\tvs -> F e
642 = Let x' = /\tvs -> F e
646 G = F . Let x = x' tvs
648 go F (Letrec xi=ei in b)
649 = Letrec {xi' = /\tvs -> G ei}
653 G = F . Let {xi = xi' tvs}
655 [May 1999] If we do this transformation *regardless* then we can
656 end up with some pretty silly stuff. For example,
659 st = /\ s -> let { x1=r1 ; x2=r2 } in ...
664 st = /\s -> ...[y1 s/x1, y2 s/x2]
667 Unless the "..." is a WHNF there is really no point in doing this.
668 Indeed it can make things worse. Suppose x1 is used strictly,
671 x1* = case f y of { (a,b) -> e }
673 If we abstract this wrt the tyvar we then can't do the case inline
674 as we would normally do.
678 {- Trying to do this in full laziness
680 tryRhsTyLam :: SimplEnv -> [OutTyVar] -> OutExpr -> SimplM FloatsWithExpr
681 -- Call ensures that all the binders are type variables
683 tryRhsTyLam env tyvars body -- Only does something if there's a let
684 | not (all isTyVar tyvars)
685 || not (worth_it body) -- inside a type lambda,
686 = returnSmpl (emptyFloats env, body) -- and a WHNF inside that
689 = go env (\x -> x) body
692 worth_it e@(Let _ _) = whnf_in_middle e
695 whnf_in_middle (Let (NonRec x rhs) e) | isUnLiftedType (idType x) = False
696 whnf_in_middle (Let _ e) = whnf_in_middle e
697 whnf_in_middle e = exprIsCheap e
699 main_tyvar_set = mkVarSet tyvars
701 go env fn (Let bind@(NonRec var rhs) body)
703 = go env (fn . Let bind) body
705 go env fn (Let (NonRec var rhs) body)
706 = mk_poly tyvars_here var `thenSmpl` \ (var', rhs') ->
707 addAuxiliaryBind env (NonRec var' (mkLams tyvars_here (fn rhs))) $ \ env ->
708 go env (fn . Let (mk_silly_bind var rhs')) body
712 tyvars_here = varSetElems (main_tyvar_set `intersectVarSet` exprSomeFreeVars isTyVar rhs)
713 -- Abstract only over the type variables free in the rhs
714 -- wrt which the new binding is abstracted. But the naive
715 -- approach of abstract wrt the tyvars free in the Id's type
717 -- /\ a b -> let t :: (a,b) = (e1, e2)
720 -- Here, b isn't free in x's type, but we must nevertheless
721 -- abstract wrt b as well, because t's type mentions b.
722 -- Since t is floated too, we'd end up with the bogus:
723 -- poly_t = /\ a b -> (e1, e2)
724 -- poly_x = /\ a -> fst (poly_t a *b*)
725 -- So for now we adopt the even more naive approach of
726 -- abstracting wrt *all* the tyvars. We'll see if that
727 -- gives rise to problems. SLPJ June 98
729 go env fn (Let (Rec prs) body)
730 = mapAndUnzipSmpl (mk_poly tyvars_here) vars `thenSmpl` \ (vars', rhss') ->
732 gn body = fn (foldr Let body (zipWith mk_silly_bind vars rhss'))
733 pairs = vars' `zip` [mkLams tyvars_here (gn rhs) | rhs <- rhss]
735 addAuxiliaryBind env (Rec pairs) $ \ env ->
738 (vars,rhss) = unzip prs
739 tyvars_here = varSetElems (main_tyvar_set `intersectVarSet` exprsSomeFreeVars isTyVar (map snd prs))
740 -- See notes with tyvars_here above
742 go env fn body = returnSmpl (emptyFloats env, fn body)
744 mk_poly tyvars_here var
745 = getUniqueSmpl `thenSmpl` \ uniq ->
747 poly_name = setNameUnique (idName var) uniq -- Keep same name
748 poly_ty = mkForAllTys tyvars_here (idType var) -- But new type of course
749 poly_id = mkLocalId poly_name poly_ty
751 -- In the olden days, it was crucial to copy the occInfo of the original var,
752 -- because we were looking at occurrence-analysed but as yet unsimplified code!
753 -- In particular, we mustn't lose the loop breakers. BUT NOW we are looking
754 -- at already simplified code, so it doesn't matter
756 -- It's even right to retain single-occurrence or dead-var info:
757 -- Suppose we started with /\a -> let x = E in B
758 -- where x occurs once in B. Then we transform to:
759 -- let x' = /\a -> E in /\a -> let x* = x' a in B
760 -- where x* has an INLINE prag on it. Now, once x* is inlined,
761 -- the occurrences of x' will be just the occurrences originally
764 returnSmpl (poly_id, mkTyApps (Var poly_id) (mkTyVarTys tyvars_here))
766 mk_silly_bind var rhs = NonRec var (Note InlineMe rhs)
767 -- Suppose we start with:
769 -- x = /\ a -> let g = G in E
771 -- Then we'll float to get
773 -- x = let poly_g = /\ a -> G
774 -- in /\ a -> let g = poly_g a in E
776 -- But now the occurrence analyser will see just one occurrence
777 -- of poly_g, not inside a lambda, so the simplifier will
778 -- PreInlineUnconditionally poly_g back into g! Badk to square 1!
779 -- (I used to think that the "don't inline lone occurrences" stuff
780 -- would stop this happening, but since it's the *only* occurrence,
781 -- PreInlineUnconditionally kicks in first!)
783 -- Solution: put an INLINE note on g's RHS, so that poly_g seems
784 -- to appear many times. (NB: mkInlineMe eliminates
785 -- such notes on trivial RHSs, so do it manually.)
789 %************************************************************************
791 \subsection{Case alternative filtering
793 %************************************************************************
795 prepareAlts does two things:
797 1. Eliminate alternatives that cannot match, including the
800 2. If the DEFAULT alternative can match only one possible constructor,
801 then make that constructor explicit.
803 case e of x { DEFAULT -> rhs }
805 case e of x { (a,b) -> rhs }
806 where the type is a single constructor type. This gives better code
807 when rhs also scrutinises x or e.
809 It's a good idea do do this stuff before simplifying the alternatives, to
810 avoid simplifying alternatives we know can't happen, and to come up with
811 the list of constructors that are handled, to put into the IdInfo of the
812 case binder, for use when simplifying the alternatives.
814 Eliminating the default alternative in (1) isn't so obvious, but it can
817 data Colour = Red | Green | Blue
826 DEFAULT -> [ case y of ... ]
828 If we inline h into f, the default case of the inlined h can't happen.
829 If we don't notice this, we may end up filtering out *all* the cases
830 of the inner case y, which give us nowhere to go!
834 prepareAlts :: OutExpr -- Scrutinee
835 -> InId -- Case binder
837 -> SimplM ([InAlt], -- Better alternatives
838 [AltCon]) -- These cases are handled
840 prepareAlts scrut case_bndr alts
842 (alts_wo_default, maybe_deflt) = findDefault alts
844 impossible_cons = case scrut of
845 Var v -> otherCons (idUnfolding v)
848 -- Filter out alternatives that can't possibly match
849 better_alts | null impossible_cons = alts_wo_default
850 | otherwise = [alt | alt@(con,_,_) <- alts_wo_default,
851 not (con `elem` impossible_cons)]
853 -- "handled_cons" are handled either by the context,
854 -- or by a branch in this case expression
855 -- (Don't add DEFAULT to the handled_cons!!)
856 handled_cons = impossible_cons ++ [con | (con,_,_) <- better_alts]
858 -- Filter out the default, if it can't happen,
859 -- or replace it with "proper" alternative if there
860 -- is only one constructor left
861 prepareDefault case_bndr handled_cons maybe_deflt `thenSmpl` \ deflt_alt ->
863 returnSmpl (deflt_alt ++ better_alts, handled_cons)
865 prepareDefault case_bndr handled_cons (Just rhs)
866 | Just (tycon, inst_tys) <- splitTyConApp_maybe (idType case_bndr),
867 isAlgTyCon tycon, -- It's a data type, tuple, or unboxed tuples.
868 not (isNewTyCon tycon), -- We can have a newtype, if we are just doing an eval:
869 -- case x of { DEFAULT -> e }
870 -- and we don't want to fill in a default for them!
871 Just all_cons <- tyConDataCons_maybe tycon,
872 not (null all_cons), -- This is a tricky corner case. If the data type has no constructors,
873 -- which GHC allows, then the case expression will have at most a default
874 -- alternative. We don't want to eliminate that alternative, because the
875 -- invariant is that there's always one alternative. It's more convenient
877 -- case x of { DEFAULT -> e }
878 -- as it is, rather than transform it to
879 -- error "case cant match"
880 -- which would be quite legitmate. But it's a really obscure corner, and
881 -- not worth wasting code on.
882 let handled_data_cons = [data_con | DataAlt data_con <- handled_cons],
883 let missing_cons = [con | con <- all_cons,
884 not (con `elem` handled_data_cons)]
885 = case missing_cons of
886 [] -> returnSmpl [] -- Eliminate the default alternative
889 [con] -> -- It matches exactly one constructor, so fill it in
890 tick (FillInCaseDefault case_bndr) `thenSmpl_`
891 mk_args con inst_tys `thenSmpl` \ args ->
892 returnSmpl [(DataAlt con, args, rhs)]
894 two_or_more -> returnSmpl [(DEFAULT, [], rhs)]
897 = returnSmpl [(DEFAULT, [], rhs)]
899 prepareDefault case_bndr handled_cons Nothing
902 mk_args missing_con inst_tys
903 = getUniquesSmpl `thenSmpl` \ tv_uniqs ->
904 getUniquesSmpl `thenSmpl` \ id_uniqs ->
906 ex_tyvars = dataConExistentialTyVars missing_con
907 ex_tyvars' = zipWith mk tv_uniqs ex_tyvars
908 mk uniq tv = mkSysTyVar uniq (tyVarKind tv)
909 arg_tys = dataConArgTys missing_con (inst_tys ++ mkTyVarTys ex_tyvars')
910 arg_ids = zipWith (mkSysLocal FSLIT("a")) id_uniqs arg_tys
912 returnSmpl (ex_tyvars' ++ arg_ids)
916 %************************************************************************
918 \subsection{Case absorption and identity-case elimination}
920 %************************************************************************
922 mkCase puts a case expression back together, trying various transformations first.
925 mkCase :: OutExpr -> OutId -> [OutAlt] -> SimplM OutExpr
927 mkCase scrut case_bndr alts
928 = mkAlts scrut case_bndr alts `thenSmpl` \ better_alts ->
929 mkCase1 scrut case_bndr better_alts
933 mkAlts tries these things:
935 1. If several alternatives are identical, merge them into
936 a single DEFAULT alternative. I've occasionally seen this
937 making a big difference:
939 case e of =====> case e of
940 C _ -> f x D v -> ....v....
941 D v -> ....v.... DEFAULT -> f x
944 The point is that we merge common RHSs, at least for the DEFAULT case.
945 [One could do something more elaborate but I've never seen it needed.]
946 To avoid an expensive test, we just merge branches equal to the *first*
947 alternative; this picks up the common cases
948 a) all branches equal
949 b) some branches equal to the DEFAULT (which occurs first)
952 case e of b { ==> case e of b {
953 p1 -> rhs1 p1 -> rhs1
955 pm -> rhsm pm -> rhsm
956 _ -> case b of b' { pn -> let b'=b in rhsn
958 ... po -> let b'=b in rhso
959 po -> rhso _ -> let b'=b in rhsd
963 which merges two cases in one case when -- the default alternative of
964 the outer case scrutises the same variable as the outer case This
965 transformation is called Case Merging. It avoids that the same
966 variable is scrutinised multiple times.
969 The case where transformation (1) showed up was like this (lib/std/PrelCError.lhs):
975 where @is@ was something like
977 p `is` n = p /= (-1) && p == n
979 This gave rise to a horrible sequence of cases
986 and similarly in cascade for all the join points!
991 --------------------------------------------------
992 -- 1. Merge identical branches
993 --------------------------------------------------
994 mkAlts scrut case_bndr alts@((con1,bndrs1,rhs1) : con_alts)
995 | all isDeadBinder bndrs1, -- Remember the default
996 length filtered_alts < length con_alts -- alternative comes first
997 = tick (AltMerge case_bndr) `thenSmpl_`
998 returnSmpl better_alts
1000 filtered_alts = filter keep con_alts
1001 keep (con,bndrs,rhs) = not (all isDeadBinder bndrs && rhs `cheapEqExpr` rhs1)
1002 better_alts = (DEFAULT, [], rhs1) : filtered_alts
1005 --------------------------------------------------
1006 -- 2. Merge nested cases
1007 --------------------------------------------------
1009 mkAlts scrut outer_bndr outer_alts
1010 | opt_SimplCaseMerge,
1011 (outer_alts_without_deflt, maybe_outer_deflt) <- findDefault outer_alts,
1012 Just (Case (Var scrut_var) inner_bndr inner_alts) <- maybe_outer_deflt,
1013 scruting_same_var scrut_var
1015 = let -- Eliminate any inner alts which are shadowed by the outer ones
1016 outer_cons = [con | (con,_,_) <- outer_alts_without_deflt]
1018 munged_inner_alts = [ (con, args, munge_rhs rhs)
1019 | (con, args, rhs) <- inner_alts,
1020 not (con `elem` outer_cons) -- Eliminate shadowed inner alts
1022 munge_rhs rhs = bindCaseBndr inner_bndr (Var outer_bndr) rhs
1024 (inner_con_alts, maybe_inner_default) = findDefault munged_inner_alts
1026 new_alts = add_default maybe_inner_default
1027 (outer_alts_without_deflt ++ inner_con_alts)
1029 tick (CaseMerge outer_bndr) `thenSmpl_`
1031 -- Warning: don't call mkAlts recursively!
1032 -- Firstly, there's no point, because inner alts have already had
1033 -- mkCase applied to them, so they won't have a case in their default
1034 -- Secondly, if you do, you get an infinite loop, because the bindCaseBndr
1035 -- in munge_rhs may put a case into the DEFAULT branch!
1037 -- We are scrutinising the same variable if it's
1038 -- the outer case-binder, or if the outer case scrutinises a variable
1039 -- (and it's the same). Testing both allows us not to replace the
1040 -- outer scrut-var with the outer case-binder (Simplify.simplCaseBinder).
1041 scruting_same_var = case scrut of
1042 Var outer_scrut -> \ v -> v == outer_bndr || v == outer_scrut
1043 other -> \ v -> v == outer_bndr
1045 add_default (Just rhs) alts = (DEFAULT,[],rhs) : alts
1046 add_default Nothing alts = alts
1049 --------------------------------------------------
1051 --------------------------------------------------
1053 mkAlts scrut case_bndr other_alts = returnSmpl other_alts
1058 =================================================================================
1060 mkCase1 tries these things
1062 1. Eliminate the case altogether if possible
1070 and similar friends.
1073 Start with a simple situation:
1075 case x# of ===> e[x#/y#]
1078 (when x#, y# are of primitive type, of course). We can't (in general)
1079 do this for algebraic cases, because we might turn bottom into
1082 Actually, we generalise this idea to look for a case where we're
1083 scrutinising a variable, and we know that only the default case can
1088 other -> ...(case x of
1092 Here the inner case can be eliminated. This really only shows up in
1093 eliminating error-checking code.
1095 We also make sure that we deal with this very common case:
1100 Here we are using the case as a strict let; if x is used only once
1101 then we want to inline it. We have to be careful that this doesn't
1102 make the program terminate when it would have diverged before, so we
1104 - x is used strictly, or
1105 - e is already evaluated (it may so if e is a variable)
1107 Lastly, we generalise the transformation to handle this:
1113 We only do this for very cheaply compared r's (constructors, literals
1114 and variables). If pedantic bottoms is on, we only do it when the
1115 scrutinee is a PrimOp which can't fail.
1117 We do it *here*, looking at un-simplified alternatives, because we
1118 have to check that r doesn't mention the variables bound by the
1119 pattern in each alternative, so the binder-info is rather useful.
1121 So the case-elimination algorithm is:
1123 1. Eliminate alternatives which can't match
1125 2. Check whether all the remaining alternatives
1126 (a) do not mention in their rhs any of the variables bound in their pattern
1127 and (b) have equal rhss
1129 3. Check we can safely ditch the case:
1130 * PedanticBottoms is off,
1131 or * the scrutinee is an already-evaluated variable
1132 or * the scrutinee is a primop which is ok for speculation
1133 -- ie we want to preserve divide-by-zero errors, and
1134 -- calls to error itself!
1136 or * [Prim cases] the scrutinee is a primitive variable
1138 or * [Alg cases] the scrutinee is a variable and
1139 either * the rhs is the same variable
1140 (eg case x of C a b -> x ===> x)
1141 or * there is only one alternative, the default alternative,
1142 and the binder is used strictly in its scope.
1143 [NB this is helped by the "use default binder where
1144 possible" transformation; see below.]
1147 If so, then we can replace the case with one of the rhss.
1149 Further notes about case elimination
1150 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
1151 Consider: test :: Integer -> IO ()
1154 Turns out that this compiles to:
1157 eta1 :: State# RealWorld ->
1158 case PrelNum.< eta PrelNum.zeroInteger of wild { __DEFAULT ->
1160 (PrelNum.jtos eta ($w[] @ Char))
1162 of wild1 { (# new_s, a4 #) -> PrelIO.lvl23 new_s }}
1164 Notice the strange '<' which has no effect at all. This is a funny one.
1165 It started like this:
1167 f x y = if x < 0 then jtos x
1168 else if y==0 then "" else jtos x
1170 At a particular call site we have (f v 1). So we inline to get
1172 if v < 0 then jtos x
1173 else if 1==0 then "" else jtos x
1175 Now simplify the 1==0 conditional:
1177 if v<0 then jtos v else jtos v
1179 Now common-up the two branches of the case:
1181 case (v<0) of DEFAULT -> jtos v
1183 Why don't we drop the case? Because it's strict in v. It's technically
1184 wrong to drop even unnecessary evaluations, and in practice they
1185 may be a result of 'seq' so we *definitely* don't want to drop those.
1186 I don't really know how to improve this situation.
1190 --------------------------------------------------
1191 -- 0. Check for empty alternatives
1192 --------------------------------------------------
1195 mkCase1 scrut case_bndr []
1196 = pprTrace "mkCase1: null alts" (ppr case_bndr <+> ppr scrut) $
1200 --------------------------------------------------
1201 -- 1. Eliminate the case altogether if poss
1202 --------------------------------------------------
1204 mkCase1 scrut case_bndr [(con,bndrs,rhs)]
1205 -- See if we can get rid of the case altogether
1206 -- See the extensive notes on case-elimination above
1207 -- mkCase made sure that if all the alternatives are equal,
1208 -- then there is now only one (DEFAULT) rhs
1209 | all isDeadBinder bndrs,
1211 -- Check that the scrutinee can be let-bound instead of case-bound
1212 exprOkForSpeculation scrut
1213 -- OK not to evaluate it
1214 -- This includes things like (==# a# b#)::Bool
1215 -- so that we simplify
1216 -- case ==# a# b# of { True -> x; False -> x }
1219 -- This particular example shows up in default methods for
1220 -- comparision operations (e.g. in (>=) for Int.Int32)
1221 || exprIsValue scrut -- It's already evaluated
1222 || var_demanded_later scrut -- It'll be demanded later
1224 -- || not opt_SimplPedanticBottoms) -- Or we don't care!
1225 -- We used to allow improving termination by discarding cases, unless -fpedantic-bottoms was on,
1226 -- but that breaks badly for the dataToTag# primop, which relies on a case to evaluate
1227 -- its argument: case x of { y -> dataToTag# y }
1228 -- Here we must *not* discard the case, because dataToTag# just fetches the tag from
1229 -- the info pointer. So we'll be pedantic all the time, and see if that gives any
1231 -- Also we don't want to discard 'seq's
1232 = tick (CaseElim case_bndr) `thenSmpl_`
1233 returnSmpl (bindCaseBndr case_bndr scrut rhs)
1236 -- The case binder is going to be evaluated later,
1237 -- and the scrutinee is a simple variable
1238 var_demanded_later (Var v) = isStrictDmd (idNewDemandInfo case_bndr)
1239 var_demanded_later other = False
1242 --------------------------------------------------
1244 --------------------------------------------------
1246 mkCase1 scrut case_bndr alts -- Identity case
1247 | all identity_alt alts
1248 = tick (CaseIdentity case_bndr) `thenSmpl_`
1249 returnSmpl (re_note scrut)
1251 identity_alt (con, args, rhs) = de_note rhs `cheapEqExpr` identity_rhs con args
1253 identity_rhs (DataAlt con) args = mkConApp con (arg_tys ++ map varToCoreExpr args)
1254 identity_rhs (LitAlt lit) _ = Lit lit
1255 identity_rhs DEFAULT _ = Var case_bndr
1257 arg_tys = map Type (tyConAppArgs (idType case_bndr))
1260 -- case coerce T e of x { _ -> coerce T' x }
1261 -- And we definitely want to eliminate this case!
1262 -- So we throw away notes from the RHS, and reconstruct
1263 -- (at least an approximation) at the other end
1264 de_note (Note _ e) = de_note e
1267 -- re_note wraps a coerce if it might be necessary
1268 re_note scrut = case head alts of
1269 (_,_,rhs1@(Note _ _)) -> mkCoerce2 (exprType rhs1) (idType case_bndr) scrut
1273 --------------------------------------------------
1275 --------------------------------------------------
1276 mkCase1 scrut bndr alts = returnSmpl (Case scrut bndr alts)
1280 When adding auxiliary bindings for the case binder, it's worth checking if
1281 its dead, because it often is, and occasionally these mkCase transformations
1282 cascade rather nicely.
1285 bindCaseBndr bndr rhs body
1286 | isDeadBinder bndr = body
1287 | otherwise = bindNonRec bndr rhs body