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(..), opt_UF_UpdateInPlace,
26 import CoreFVs ( exprFreeVars )
27 import CoreUtils ( cheapEqExpr, exprType, exprIsTrivial,
28 etaExpand, exprEtaExpandArity, bindNonRec, mkCoerce2,
29 findDefault, exprOkForSpeculation, exprIsValue
31 import qualified Subst ( simplBndrs, simplBndr, simplLetId, simplLamBndr )
32 import Id ( Id, idType, idInfo, isDataConWorkId,
33 mkSysLocal, isDeadBinder, idNewDemandInfo,
34 idUnfolding, idNewStrictness
36 import NewDemand ( isStrictDmd, isBotRes, splitStrictSig )
38 import Type ( Type, seqType, splitFunTys, dropForAlls, isStrictType,
39 splitTyConApp_maybe, tyConAppArgs, mkTyVarTys
41 import TcType ( isDictTy )
42 import Name ( mkSysTvName )
43 import OccName ( EncodedFS )
44 import TyCon ( tyConDataCons_maybe, isAlgTyCon, isNewTyCon )
45 import DataCon ( dataConRepArity, dataConExistentialTyVars, dataConArgTys )
46 import Var ( tyVarKind, mkTyVar )
48 import Util ( lengthExceeds, mapAccumL )
53 %************************************************************************
55 \subsection{The continuation data type}
57 %************************************************************************
60 data SimplCont -- Strict contexts
61 = Stop OutType -- Type of the result
63 Bool -- True <=> This is the RHS of a thunk whose type suggests
64 -- that update-in-place would be possible
65 -- (This makes the inliner a little keener.)
67 | CoerceIt OutType -- The To-type, simplified
70 | InlinePlease -- This continuation makes a function very
71 SimplCont -- keen to inline itelf
74 InExpr SimplEnv -- The argument, as yet unsimplified,
75 SimplCont -- and its environment
78 InId [InAlt] SimplEnv -- The case binder, alts, and subst-env
81 | ArgOf LetRhsFlag -- An arbitrary strict context: the argument
82 -- of a strict function, or a primitive-arg fn
84 -- No DupFlag because we never duplicate it
85 OutType -- arg_ty: type of the argument itself
86 OutType -- cont_ty: the type of the expression being sought by the context
87 -- f (error "foo") ==> coerce t (error "foo")
89 -- We need to know the type t, to which to coerce.
91 (SimplEnv -> OutExpr -> SimplM FloatsWithExpr) -- What to do with the result
92 -- The result expression in the OutExprStuff has type cont_ty
94 data LetRhsFlag = AnArg -- It's just an argument not a let RHS
95 | AnRhs -- It's the RHS of a let (so please float lets out of big lambdas)
97 instance Outputable LetRhsFlag where
98 ppr AnArg = ptext SLIT("arg")
99 ppr AnRhs = ptext SLIT("rhs")
101 instance Outputable SimplCont where
102 ppr (Stop _ is_rhs _) = ptext SLIT("Stop") <> brackets (ppr is_rhs)
103 ppr (ApplyTo dup arg se cont) = (ptext SLIT("ApplyTo") <+> ppr dup <+> ppr arg) $$ ppr cont
104 ppr (ArgOf _ _ _ _) = ptext SLIT("ArgOf...")
105 ppr (Select dup bndr alts se cont) = (ptext SLIT("Select") <+> ppr dup <+> ppr bndr) $$
106 (nest 4 (ppr alts)) $$ ppr cont
107 ppr (CoerceIt ty cont) = (ptext SLIT("CoerceIt") <+> ppr ty) $$ ppr cont
108 ppr (InlinePlease cont) = ptext SLIT("InlinePlease") $$ ppr cont
110 data DupFlag = OkToDup | NoDup
112 instance Outputable DupFlag where
113 ppr OkToDup = ptext SLIT("ok")
114 ppr NoDup = ptext SLIT("nodup")
118 mkBoringStop :: OutType -> SimplCont
119 mkBoringStop ty = Stop ty AnArg (canUpdateInPlace ty)
121 mkStop :: OutType -> LetRhsFlag -> SimplCont
122 mkStop ty is_rhs = Stop ty is_rhs (canUpdateInPlace ty)
124 contIsRhs :: SimplCont -> Bool
125 contIsRhs (Stop _ AnRhs _) = True
126 contIsRhs (ArgOf AnRhs _ _ _) = True
127 contIsRhs other = False
129 contIsRhsOrArg (Stop _ _ _) = True
130 contIsRhsOrArg (ArgOf _ _ _ _) = True
131 contIsRhsOrArg other = False
134 contIsDupable :: SimplCont -> Bool
135 contIsDupable (Stop _ _ _) = True
136 contIsDupable (ApplyTo OkToDup _ _ _) = True
137 contIsDupable (Select OkToDup _ _ _ _) = True
138 contIsDupable (CoerceIt _ cont) = contIsDupable cont
139 contIsDupable (InlinePlease cont) = contIsDupable cont
140 contIsDupable other = False
143 discardableCont :: SimplCont -> Bool
144 discardableCont (Stop _ _ _) = False
145 discardableCont (CoerceIt _ cont) = discardableCont cont
146 discardableCont (InlinePlease cont) = discardableCont cont
147 discardableCont other = True
149 discardCont :: SimplCont -- A continuation, expecting
150 -> SimplCont -- Replace the continuation with a suitable coerce
151 discardCont cont = case cont of
152 Stop to_ty is_rhs _ -> cont
153 other -> CoerceIt to_ty (mkBoringStop to_ty)
155 to_ty = contResultType cont
158 contResultType :: SimplCont -> OutType
159 contResultType (Stop to_ty _ _) = to_ty
160 contResultType (ArgOf _ _ to_ty _) = to_ty
161 contResultType (ApplyTo _ _ _ cont) = contResultType cont
162 contResultType (CoerceIt _ cont) = contResultType cont
163 contResultType (InlinePlease cont) = contResultType cont
164 contResultType (Select _ _ _ _ cont) = contResultType cont
167 countValArgs :: SimplCont -> Int
168 countValArgs (ApplyTo _ (Type ty) se cont) = countValArgs cont
169 countValArgs (ApplyTo _ val_arg se cont) = 1 + countValArgs cont
170 countValArgs other = 0
172 countArgs :: SimplCont -> Int
173 countArgs (ApplyTo _ arg se cont) = 1 + countArgs cont
177 pushContArgs :: SimplEnv -> [OutArg] -> SimplCont -> SimplCont
178 -- Pushes args with the specified environment
179 pushContArgs env [] cont = cont
180 pushContArgs env (arg : args) cont = ApplyTo NoDup arg env (pushContArgs env args cont)
185 getContArgs :: SwitchChecker
186 -> OutId -> SimplCont
187 -> ([(InExpr, SimplEnv, Bool)], -- Arguments; the Bool is true for strict args
188 SimplCont, -- Remaining continuation
189 Bool) -- Whether we came across an InlineCall
190 -- getContArgs id k = (args, k', inl)
191 -- args are the leading ApplyTo items in k
192 -- (i.e. outermost comes first)
193 -- augmented with demand info from the functionn
194 getContArgs chkr fun orig_cont
196 -- Ignore strictness info if the no-case-of-case
197 -- flag is on. Strictness changes evaluation order
198 -- and that can change full laziness
199 stricts | switchIsOn chkr NoCaseOfCase = vanilla_stricts
200 | otherwise = computed_stricts
202 go [] stricts False orig_cont
204 ----------------------------
207 go acc ss inl (ApplyTo _ arg@(Type _) se cont)
208 = go ((arg,se,False) : acc) ss inl cont
209 -- NB: don't bother to instantiate the function type
212 go acc (s:ss) inl (ApplyTo _ arg se cont)
213 = go ((arg,se,s) : acc) ss inl cont
215 -- An Inline continuation
216 go acc ss inl (InlinePlease cont)
217 = go acc ss True cont
219 -- We're run out of arguments, or else we've run out of demands
220 -- The latter only happens if the result is guaranteed bottom
221 -- This is the case for
222 -- * case (error "hello") of { ... }
223 -- * (error "Hello") arg
224 -- * f (error "Hello") where f is strict
226 -- Then, especially in the first of these cases, we'd like to discard
227 -- the continuation, leaving just the bottoming expression. But the
228 -- type might not be right, so we may have to add a coerce.
230 | null ss && discardableCont cont = (reverse acc, discardCont cont, inl)
231 | otherwise = (reverse acc, cont, inl)
233 ----------------------------
234 vanilla_stricts, computed_stricts :: [Bool]
235 vanilla_stricts = repeat False
236 computed_stricts = zipWith (||) fun_stricts arg_stricts
238 ----------------------------
239 (val_arg_tys, _) = splitFunTys (dropForAlls (idType fun))
240 arg_stricts = map isStrictType val_arg_tys ++ repeat False
241 -- These argument types are used as a cheap and cheerful way to find
242 -- unboxed arguments, which must be strict. But it's an InType
243 -- and so there might be a type variable where we expect a function
244 -- type (the substitution hasn't happened yet). And we don't bother
245 -- doing the type applications for a polymorphic function.
246 -- Hence the splitFunTys*IgnoringForAlls*
248 ----------------------------
249 -- If fun_stricts is finite, it means the function returns bottom
250 -- after that number of value args have been consumed
251 -- Otherwise it's infinite, extended with False
253 = case splitStrictSig (idNewStrictness fun) of
254 (demands, result_info)
255 | not (demands `lengthExceeds` countValArgs orig_cont)
256 -> -- Enough args, use the strictness given.
257 -- For bottoming functions we used to pretend that the arg
258 -- is lazy, so that we don't treat the arg as an
259 -- interesting context. This avoids substituting
260 -- top-level bindings for (say) strings into
261 -- calls to error. But now we are more careful about
262 -- inlining lone variables, so its ok (see SimplUtils.analyseCont)
263 if isBotRes result_info then
264 map isStrictDmd demands -- Finite => result is bottom
266 map isStrictDmd demands ++ vanilla_stricts
268 other -> vanilla_stricts -- Not enough args, or no strictness
271 interestingArg :: OutExpr -> Bool
272 -- An argument is interesting if it has *some* structure
273 -- We are here trying to avoid unfolding a function that
274 -- is applied only to variables that have no unfolding
275 -- (i.e. they are probably lambda bound): f x y z
276 -- There is little point in inlining f here.
277 interestingArg (Var v) = hasSomeUnfolding (idUnfolding v)
278 -- Was: isValueUnfolding (idUnfolding v')
279 -- But that seems over-pessimistic
281 -- This accounts for an argument like
282 -- () or [], which is definitely interesting
283 interestingArg (Type _) = False
284 interestingArg (App fn (Type _)) = interestingArg fn
285 interestingArg (Note _ a) = interestingArg a
286 interestingArg other = True
287 -- Consider let x = 3 in f x
288 -- The substitution will contain (x -> ContEx 3), and we want to
289 -- to say that x is an interesting argument.
290 -- But consider also (\x. f x y) y
291 -- The substitution will contain (x -> ContEx y), and we want to say
292 -- that x is not interesting (assuming y has no unfolding)
295 Comment about interestingCallContext
296 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
297 We want to avoid inlining an expression where there can't possibly be
298 any gain, such as in an argument position. Hence, if the continuation
299 is interesting (eg. a case scrutinee, application etc.) then we
300 inline, otherwise we don't.
302 Previously some_benefit used to return True only if the variable was
303 applied to some value arguments. This didn't work:
305 let x = _coerce_ (T Int) Int (I# 3) in
306 case _coerce_ Int (T Int) x of
309 we want to inline x, but can't see that it's a constructor in a case
310 scrutinee position, and some_benefit is False.
314 dMonadST = _/\_ t -> :Monad (g1 _@_ t, g2 _@_ t, g3 _@_ t)
316 .... case dMonadST _@_ x0 of (a,b,c) -> ....
318 we'd really like to inline dMonadST here, but we *don't* want to
319 inline if the case expression is just
321 case x of y { DEFAULT -> ... }
323 since we can just eliminate this case instead (x is in WHNF). Similar
324 applies when x is bound to a lambda expression. Hence
325 contIsInteresting looks for case expressions with just a single
329 interestingCallContext :: Bool -- False <=> no args at all
330 -> Bool -- False <=> no value args
332 -- The "lone-variable" case is important. I spent ages
333 -- messing about with unsatisfactory varaints, but this is nice.
334 -- The idea is that if a variable appear all alone
335 -- as an arg of lazy fn, or rhs Stop
336 -- as scrutinee of a case Select
337 -- as arg of a strict fn ArgOf
338 -- then we should not inline it (unless there is some other reason,
339 -- e.g. is is the sole occurrence). We achieve this by making
340 -- interestingCallContext return False for a lone variable.
342 -- Why? At least in the case-scrutinee situation, turning
343 -- let x = (a,b) in case x of y -> ...
345 -- let x = (a,b) in case (a,b) of y -> ...
347 -- let x = (a,b) in let y = (a,b) in ...
348 -- is bad if the binding for x will remain.
350 -- Another example: I discovered that strings
351 -- were getting inlined straight back into applications of 'error'
352 -- because the latter is strict.
354 -- f = \x -> ...(error s)...
356 -- Fundamentally such contexts should not ecourage inlining because
357 -- the context can ``see'' the unfolding of the variable (e.g. case or a RULE)
358 -- so there's no gain.
360 -- However, even a type application or coercion isn't a lone variable.
362 -- case $fMonadST @ RealWorld of { :DMonad a b c -> c }
363 -- We had better inline that sucker! The case won't see through it.
365 -- For now, I'm treating treating a variable applied to types
366 -- in a *lazy* context "lone". The motivating example was
368 -- g = /\a. \y. h (f a)
369 -- There's no advantage in inlining f here, and perhaps
370 -- a significant disadvantage. Hence some_val_args in the Stop case
372 interestingCallContext some_args some_val_args cont
375 interesting (InlinePlease _) = True
376 interesting (Select _ _ _ _ _) = some_args
377 interesting (ApplyTo _ _ _ _) = True -- Can happen if we have (coerce t (f x)) y
378 -- Perhaps True is a bit over-keen, but I've
379 -- seen (coerce f) x, where f has an INLINE prag,
380 -- So we have to give some motivaiton for inlining it
381 interesting (ArgOf _ _ _ _) = some_val_args
382 interesting (Stop ty _ upd_in_place) = some_val_args && upd_in_place
383 interesting (CoerceIt _ cont) = interesting cont
384 -- If this call is the arg of a strict function, the context
385 -- is a bit interesting. If we inline here, we may get useful
386 -- evaluation information to avoid repeated evals: e.g.
388 -- Here the contIsInteresting makes the '*' keener to inline,
389 -- which in turn exposes a constructor which makes the '+' inline.
390 -- Assuming that +,* aren't small enough to inline regardless.
392 -- It's also very important to inline in a strict context for things
395 -- Here, the context of (f x) is strict, and if f's unfolding is
396 -- a build it's *great* to inline it here. So we must ensure that
397 -- the context for (f x) is not totally uninteresting.
401 canUpdateInPlace :: Type -> Bool
402 -- Consider let x = <wurble> in ...
403 -- If <wurble> returns an explicit constructor, we might be able
404 -- to do update in place. So we treat even a thunk RHS context
405 -- as interesting if update in place is possible. We approximate
406 -- this by seeing if the type has a single constructor with a
407 -- small arity. But arity zero isn't good -- we share the single copy
408 -- for that case, so no point in sharing.
411 | not opt_UF_UpdateInPlace = False
413 = case splitTyConApp_maybe ty of
415 Just (tycon, _) -> case tyConDataCons_maybe tycon of
416 Just [dc] -> arity == 1 || arity == 2
418 arity = dataConRepArity dc
424 %************************************************************************
426 \section{Dealing with a single binder}
428 %************************************************************************
430 These functions are in the monad only so that they can be made strict via seq.
433 simplBinders :: SimplEnv -> [InBinder] -> SimplM (SimplEnv, [OutBinder])
434 simplBinders env bndrs
436 (subst', bndrs') = Subst.simplBndrs (getSubst env) bndrs
438 seqBndrs bndrs' `seq`
439 returnSmpl (setSubst env subst', bndrs')
441 simplBinder :: SimplEnv -> InBinder -> SimplM (SimplEnv, OutBinder)
444 (subst', bndr') = Subst.simplBndr (getSubst env) bndr
447 returnSmpl (setSubst env subst', bndr')
450 simplLetBndr :: SimplEnv -> InBinder -> SimplM (SimplEnv, OutBinder)
453 (subst', id') = Subst.simplLetId (getSubst env) id
456 returnSmpl (setSubst env subst', id')
458 simplLamBndrs, simplRecBndrs
459 :: SimplEnv -> [InBinder] -> SimplM (SimplEnv, [OutBinder])
460 simplRecBndrs = simplBndrs Subst.simplLetId
461 simplLamBndrs = simplBndrs Subst.simplLamBndr
463 simplBndrs simpl_bndr env bndrs
465 (subst', bndrs') = mapAccumL simpl_bndr (getSubst env) bndrs
467 seqBndrs bndrs' `seq`
468 returnSmpl (setSubst env subst', bndrs')
471 seqBndrs (b:bs) = seqBndr b `seq` seqBndrs bs
473 seqBndr b | isTyVar b = b `seq` ()
474 | otherwise = seqType (idType b) `seq`
481 newId :: EncodedFS -> Type -> SimplM Id
482 newId fs ty = getUniqueSmpl `thenSmpl` \ uniq ->
483 returnSmpl (mkSysLocal fs uniq ty)
487 %************************************************************************
489 \subsection{Rebuilding a lambda}
491 %************************************************************************
494 mkLam :: SimplEnv -> [OutBinder] -> OutExpr -> SimplCont -> SimplM FloatsWithExpr
498 a) eta reduction, if that gives a trivial expression
499 b) eta expansion [only if there are some value lambdas]
500 c) floating lets out through big lambdas
501 [only if all tyvar lambdas, and only if this lambda
505 mkLam env bndrs body cont
506 = getDOptsSmpl `thenSmpl` \dflags ->
507 mkLam' dflags env bndrs body cont
509 mkLam' dflags env bndrs body cont
510 | dopt Opt_DoEtaReduction dflags,
511 Just etad_lam <- tryEtaReduce bndrs body
512 = tick (EtaReduction (head bndrs)) `thenSmpl_`
513 returnSmpl (emptyFloats env, etad_lam)
515 | dopt Opt_DoLambdaEtaExpansion dflags,
516 any isRuntimeVar bndrs
517 = tryEtaExpansion body `thenSmpl` \ body' ->
518 returnSmpl (emptyFloats env, mkLams bndrs body')
520 {- Sept 01: I'm experimenting with getting the
521 full laziness pass to float out past big lambdsa
522 | all isTyVar bndrs, -- Only for big lambdas
523 contIsRhs cont -- Only try the rhs type-lambda floating
524 -- if this is indeed a right-hand side; otherwise
525 -- we end up floating the thing out, only for float-in
526 -- to float it right back in again!
527 = tryRhsTyLam env bndrs body `thenSmpl` \ (floats, body') ->
528 returnSmpl (floats, mkLams bndrs body')
532 = returnSmpl (emptyFloats env, mkLams bndrs body)
536 %************************************************************************
538 \subsection{Eta expansion and reduction}
540 %************************************************************************
542 We try for eta reduction here, but *only* if we get all the
543 way to an exprIsTrivial expression.
544 We don't want to remove extra lambdas unless we are going
545 to avoid allocating this thing altogether
548 tryEtaReduce :: [OutBinder] -> OutExpr -> Maybe OutExpr
549 tryEtaReduce bndrs body
550 -- We don't use CoreUtils.etaReduce, because we can be more
552 -- (a) we already have the binders
553 -- (b) we can do the triviality test before computing the free vars
554 = go (reverse bndrs) body
556 go (b : bs) (App fun arg) | ok_arg b arg = go bs fun -- Loop round
557 go [] fun | ok_fun fun = Just fun -- Success!
558 go _ _ = Nothing -- Failure!
560 ok_fun fun = exprIsTrivial fun
561 && not (any (`elemVarSet` (exprFreeVars fun)) bndrs)
562 && (exprIsValue fun || all ok_lam bndrs)
563 ok_lam v = isTyVar v || isDictTy (idType v)
564 -- The exprIsValue is because eta reduction is not
565 -- valid in general: \x. bot /= bot
566 -- So we need to be sure that the "fun" is a value.
568 -- However, we always want to reduce (/\a -> f a) to f
569 -- This came up in a RULE: foldr (build (/\a -> g a))
570 -- did not match foldr (build (/\b -> ...something complex...))
571 -- The type checker can insert these eta-expanded versions,
572 -- with both type and dictionary lambdas; hence the slightly
575 ok_arg b arg = varToCoreExpr b `cheapEqExpr` arg
579 Try eta expansion for RHSs
582 f = \x1..xn -> N ==> f = \x1..xn y1..ym -> N y1..ym
585 where (in both cases)
587 * The xi can include type variables
589 * The yi are all value variables
591 * N is a NORMAL FORM (i.e. no redexes anywhere)
592 wanting a suitable number of extra args.
594 We may have to sandwich some coerces between the lambdas
595 to make the types work. exprEtaExpandArity looks through coerces
596 when computing arity; and etaExpand adds the coerces as necessary when
597 actually computing the expansion.
600 tryEtaExpansion :: OutExpr -> SimplM OutExpr
601 -- There is at least one runtime binder in the binders
603 = getUniquesSmpl `thenSmpl` \ us ->
604 returnSmpl (etaExpand fun_arity us body (exprType body))
606 fun_arity = exprEtaExpandArity body
610 %************************************************************************
612 \subsection{Floating lets out of big lambdas}
614 %************************************************************************
616 tryRhsTyLam tries this transformation, when the big lambda appears as
617 the RHS of a let(rec) binding:
619 /\abc -> let(rec) x = e in b
621 let(rec) x' = /\abc -> let x = x' a b c in e
623 /\abc -> let x = x' a b c in b
625 This is good because it can turn things like:
627 let f = /\a -> letrec g = ... g ... in g
629 letrec g' = /\a -> ... g' a ...
633 which is better. In effect, it means that big lambdas don't impede
636 This optimisation is CRUCIAL in eliminating the junk introduced by
637 desugaring mutually recursive definitions. Don't eliminate it lightly!
639 So far as the implementation is concerned:
641 Invariant: go F e = /\tvs -> F e
645 = Let x' = /\tvs -> F e
649 G = F . Let x = x' tvs
651 go F (Letrec xi=ei in b)
652 = Letrec {xi' = /\tvs -> G ei}
656 G = F . Let {xi = xi' tvs}
658 [May 1999] If we do this transformation *regardless* then we can
659 end up with some pretty silly stuff. For example,
662 st = /\ s -> let { x1=r1 ; x2=r2 } in ...
667 st = /\s -> ...[y1 s/x1, y2 s/x2]
670 Unless the "..." is a WHNF there is really no point in doing this.
671 Indeed it can make things worse. Suppose x1 is used strictly,
674 x1* = case f y of { (a,b) -> e }
676 If we abstract this wrt the tyvar we then can't do the case inline
677 as we would normally do.
681 {- Trying to do this in full laziness
683 tryRhsTyLam :: SimplEnv -> [OutTyVar] -> OutExpr -> SimplM FloatsWithExpr
684 -- Call ensures that all the binders are type variables
686 tryRhsTyLam env tyvars body -- Only does something if there's a let
687 | not (all isTyVar tyvars)
688 || not (worth_it body) -- inside a type lambda,
689 = returnSmpl (emptyFloats env, body) -- and a WHNF inside that
692 = go env (\x -> x) body
695 worth_it e@(Let _ _) = whnf_in_middle e
698 whnf_in_middle (Let (NonRec x rhs) e) | isUnLiftedType (idType x) = False
699 whnf_in_middle (Let _ e) = whnf_in_middle e
700 whnf_in_middle e = exprIsCheap e
702 main_tyvar_set = mkVarSet tyvars
704 go env fn (Let bind@(NonRec var rhs) body)
706 = go env (fn . Let bind) body
708 go env fn (Let (NonRec var rhs) body)
709 = mk_poly tyvars_here var `thenSmpl` \ (var', rhs') ->
710 addAuxiliaryBind env (NonRec var' (mkLams tyvars_here (fn rhs))) $ \ env ->
711 go env (fn . Let (mk_silly_bind var rhs')) body
715 tyvars_here = varSetElems (main_tyvar_set `intersectVarSet` exprSomeFreeVars isTyVar rhs)
716 -- Abstract only over the type variables free in the rhs
717 -- wrt which the new binding is abstracted. But the naive
718 -- approach of abstract wrt the tyvars free in the Id's type
720 -- /\ a b -> let t :: (a,b) = (e1, e2)
723 -- Here, b isn't free in x's type, but we must nevertheless
724 -- abstract wrt b as well, because t's type mentions b.
725 -- Since t is floated too, we'd end up with the bogus:
726 -- poly_t = /\ a b -> (e1, e2)
727 -- poly_x = /\ a -> fst (poly_t a *b*)
728 -- So for now we adopt the even more naive approach of
729 -- abstracting wrt *all* the tyvars. We'll see if that
730 -- gives rise to problems. SLPJ June 98
732 go env fn (Let (Rec prs) body)
733 = mapAndUnzipSmpl (mk_poly tyvars_here) vars `thenSmpl` \ (vars', rhss') ->
735 gn body = fn (foldr Let body (zipWith mk_silly_bind vars rhss'))
736 pairs = vars' `zip` [mkLams tyvars_here (gn rhs) | rhs <- rhss]
738 addAuxiliaryBind env (Rec pairs) $ \ env ->
741 (vars,rhss) = unzip prs
742 tyvars_here = varSetElems (main_tyvar_set `intersectVarSet` exprsSomeFreeVars isTyVar (map snd prs))
743 -- See notes with tyvars_here above
745 go env fn body = returnSmpl (emptyFloats env, fn body)
747 mk_poly tyvars_here var
748 = getUniqueSmpl `thenSmpl` \ uniq ->
750 poly_name = setNameUnique (idName var) uniq -- Keep same name
751 poly_ty = mkForAllTys tyvars_here (idType var) -- But new type of course
752 poly_id = mkLocalId poly_name poly_ty
754 -- In the olden days, it was crucial to copy the occInfo of the original var,
755 -- because we were looking at occurrence-analysed but as yet unsimplified code!
756 -- In particular, we mustn't lose the loop breakers. BUT NOW we are looking
757 -- at already simplified code, so it doesn't matter
759 -- It's even right to retain single-occurrence or dead-var info:
760 -- Suppose we started with /\a -> let x = E in B
761 -- where x occurs once in B. Then we transform to:
762 -- let x' = /\a -> E in /\a -> let x* = x' a in B
763 -- where x* has an INLINE prag on it. Now, once x* is inlined,
764 -- the occurrences of x' will be just the occurrences originally
767 returnSmpl (poly_id, mkTyApps (Var poly_id) (mkTyVarTys tyvars_here))
769 mk_silly_bind var rhs = NonRec var (Note InlineMe rhs)
770 -- Suppose we start with:
772 -- x = /\ a -> let g = G in E
774 -- Then we'll float to get
776 -- x = let poly_g = /\ a -> G
777 -- in /\ a -> let g = poly_g a in E
779 -- But now the occurrence analyser will see just one occurrence
780 -- of poly_g, not inside a lambda, so the simplifier will
781 -- PreInlineUnconditionally poly_g back into g! Badk to square 1!
782 -- (I used to think that the "don't inline lone occurrences" stuff
783 -- would stop this happening, but since it's the *only* occurrence,
784 -- PreInlineUnconditionally kicks in first!)
786 -- Solution: put an INLINE note on g's RHS, so that poly_g seems
787 -- to appear many times. (NB: mkInlineMe eliminates
788 -- such notes on trivial RHSs, so do it manually.)
792 %************************************************************************
794 \subsection{Case alternative filtering
796 %************************************************************************
798 prepareAlts does two things:
800 1. Eliminate alternatives that cannot match, including the
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.
812 It's a good idea do do this stuff before simplifying the alternatives, to
813 avoid simplifying alternatives we know can't happen, and to come up with
814 the list of constructors that are handled, to put into the IdInfo of the
815 case binder, for use when simplifying the alternatives.
817 Eliminating the default alternative in (1) isn't so obvious, but it can
820 data Colour = Red | Green | Blue
829 DEFAULT -> [ case y of ... ]
831 If we inline h into f, the default case of the inlined h can't happen.
832 If we don't notice this, we may end up filtering out *all* the cases
833 of the inner case y, which give us nowhere to go!
837 prepareAlts :: OutExpr -- Scrutinee
838 -> InId -- Case binder
840 -> SimplM ([InAlt], -- Better alternatives
841 [AltCon]) -- These cases are handled
843 prepareAlts scrut case_bndr alts
845 (alts_wo_default, maybe_deflt) = findDefault alts
847 impossible_cons = case scrut of
848 Var v -> otherCons (idUnfolding v)
851 -- Filter out alternatives that can't possibly match
852 better_alts | null impossible_cons = alts_wo_default
853 | otherwise = [alt | alt@(con,_,_) <- alts_wo_default,
854 not (con `elem` impossible_cons)]
856 -- "handled_cons" are handled either by the context,
857 -- or by a branch in this case expression
858 -- (Don't add DEFAULT to the handled_cons!!)
859 handled_cons = impossible_cons ++ [con | (con,_,_) <- better_alts]
861 -- Filter out the default, if it can't happen,
862 -- or replace it with "proper" alternative if there
863 -- is only one constructor left
864 prepareDefault case_bndr handled_cons maybe_deflt `thenSmpl` \ deflt_alt ->
866 returnSmpl (deflt_alt ++ better_alts, handled_cons)
868 prepareDefault case_bndr handled_cons (Just rhs)
869 | Just (tycon, inst_tys) <- splitTyConApp_maybe (idType case_bndr),
870 isAlgTyCon tycon, -- It's a data type, tuple, or unboxed tuples.
871 not (isNewTyCon tycon), -- We can have a newtype, if we are just doing an eval:
872 -- case x of { DEFAULT -> e }
873 -- and we don't want to fill in a default for them!
874 Just all_cons <- tyConDataCons_maybe tycon,
875 not (null all_cons), -- This is a tricky corner case. If the data type has no constructors,
876 -- which GHC allows, then the case expression will have at most a default
877 -- alternative. We don't want to eliminate that alternative, because the
878 -- invariant is that there's always one alternative. It's more convenient
880 -- case x of { DEFAULT -> e }
881 -- as it is, rather than transform it to
882 -- error "case cant match"
883 -- which would be quite legitmate. But it's a really obscure corner, and
884 -- not worth wasting code on.
885 let handled_data_cons = [data_con | DataAlt data_con <- handled_cons],
886 let missing_cons = [con | con <- all_cons,
887 not (con `elem` handled_data_cons)]
888 = case missing_cons of
889 [] -> returnSmpl [] -- Eliminate the default alternative
892 [con] -> -- It matches exactly one constructor, so fill it in
893 tick (FillInCaseDefault case_bndr) `thenSmpl_`
894 mk_args con inst_tys `thenSmpl` \ args ->
895 returnSmpl [(DataAlt con, args, rhs)]
897 two_or_more -> returnSmpl [(DEFAULT, [], rhs)]
900 = returnSmpl [(DEFAULT, [], rhs)]
902 prepareDefault case_bndr handled_cons Nothing
905 mk_args missing_con inst_tys
906 = getUniquesSmpl `thenSmpl` \ tv_uniqs ->
907 getUniquesSmpl `thenSmpl` \ id_uniqs ->
909 ex_tyvars = dataConExistentialTyVars missing_con
910 ex_tyvars' = zipWith mk tv_uniqs ex_tyvars
911 mk uniq tv = mkTyVar (mkSysTvName uniq FSLIT("t")) (tyVarKind tv)
912 arg_tys = dataConArgTys missing_con (inst_tys ++ mkTyVarTys ex_tyvars')
913 arg_ids = zipWith (mkSysLocal FSLIT("a")) id_uniqs arg_tys
915 returnSmpl (ex_tyvars' ++ arg_ids)
919 %************************************************************************
921 \subsection{Case absorption and identity-case elimination}
923 %************************************************************************
925 mkCase puts a case expression back together, trying various transformations first.
928 mkCase :: OutExpr -> OutId -> [OutAlt] -> SimplM OutExpr
930 mkCase scrut case_bndr alts
931 = mkAlts scrut case_bndr alts `thenSmpl` \ better_alts ->
932 mkCase1 scrut case_bndr better_alts
936 mkAlts tries these things:
938 1. If several alternatives are identical, merge them into
939 a single DEFAULT alternative. I've occasionally seen this
940 making a big difference:
942 case e of =====> case e of
943 C _ -> f x D v -> ....v....
944 D v -> ....v.... DEFAULT -> f x
947 The point is that we merge common RHSs, at least for the DEFAULT case.
948 [One could do something more elaborate but I've never seen it needed.]
949 To avoid an expensive test, we just merge branches equal to the *first*
950 alternative; this picks up the common cases
951 a) all branches equal
952 b) some branches equal to the DEFAULT (which occurs first)
955 case e of b { ==> case e of b {
956 p1 -> rhs1 p1 -> rhs1
958 pm -> rhsm pm -> rhsm
959 _ -> case b of b' { pn -> let b'=b in rhsn
961 ... po -> let b'=b in rhso
962 po -> rhso _ -> let b'=b in rhsd
966 which merges two cases in one case when -- the default alternative of
967 the outer case scrutises the same variable as the outer case This
968 transformation is called Case Merging. It avoids that the same
969 variable is scrutinised multiple times.
972 The case where transformation (1) showed up was like this (lib/std/PrelCError.lhs):
978 where @is@ was something like
980 p `is` n = p /= (-1) && p == n
982 This gave rise to a horrible sequence of cases
989 and similarly in cascade for all the join points!
994 --------------------------------------------------
995 -- 1. Merge identical branches
996 --------------------------------------------------
997 mkAlts scrut case_bndr alts@((con1,bndrs1,rhs1) : con_alts)
998 | all isDeadBinder bndrs1, -- Remember the default
999 length filtered_alts < length con_alts -- alternative comes first
1000 = tick (AltMerge case_bndr) `thenSmpl_`
1001 returnSmpl better_alts
1003 filtered_alts = filter keep con_alts
1004 keep (con,bndrs,rhs) = not (all isDeadBinder bndrs && rhs `cheapEqExpr` rhs1)
1005 better_alts = (DEFAULT, [], rhs1) : filtered_alts
1008 --------------------------------------------------
1009 -- 2. Merge nested cases
1010 --------------------------------------------------
1012 mkAlts scrut outer_bndr outer_alts
1013 = getDOptsSmpl `thenSmpl` \dflags ->
1014 mkAlts' dflags scrut outer_bndr outer_alts
1016 mkAlts' dflags scrut outer_bndr outer_alts
1017 | dopt Opt_CaseMerge dflags,
1018 (outer_alts_without_deflt, maybe_outer_deflt) <- findDefault outer_alts,
1019 Just (Case (Var scrut_var) inner_bndr inner_alts) <- maybe_outer_deflt,
1020 scruting_same_var scrut_var
1022 = let -- Eliminate any inner alts which are shadowed by the outer ones
1023 outer_cons = [con | (con,_,_) <- outer_alts_without_deflt]
1025 munged_inner_alts = [ (con, args, munge_rhs rhs)
1026 | (con, args, rhs) <- inner_alts,
1027 not (con `elem` outer_cons) -- Eliminate shadowed inner alts
1029 munge_rhs rhs = bindCaseBndr inner_bndr (Var outer_bndr) rhs
1031 (inner_con_alts, maybe_inner_default) = findDefault munged_inner_alts
1033 new_alts = add_default maybe_inner_default
1034 (outer_alts_without_deflt ++ inner_con_alts)
1036 tick (CaseMerge outer_bndr) `thenSmpl_`
1038 -- Warning: don't call mkAlts recursively!
1039 -- Firstly, there's no point, because inner alts have already had
1040 -- mkCase applied to them, so they won't have a case in their default
1041 -- Secondly, if you do, you get an infinite loop, because the bindCaseBndr
1042 -- in munge_rhs may put a case into the DEFAULT branch!
1044 -- We are scrutinising the same variable if it's
1045 -- the outer case-binder, or if the outer case scrutinises a variable
1046 -- (and it's the same). Testing both allows us not to replace the
1047 -- outer scrut-var with the outer case-binder (Simplify.simplCaseBinder).
1048 scruting_same_var = case scrut of
1049 Var outer_scrut -> \ v -> v == outer_bndr || v == outer_scrut
1050 other -> \ v -> v == outer_bndr
1052 add_default (Just rhs) alts = (DEFAULT,[],rhs) : alts
1053 add_default Nothing alts = alts
1056 --------------------------------------------------
1058 --------------------------------------------------
1060 mkAlts' dflags scrut case_bndr other_alts = returnSmpl other_alts
1065 =================================================================================
1067 mkCase1 tries these things
1069 1. Eliminate the case altogether if possible
1077 and similar friends.
1080 Start with a simple situation:
1082 case x# of ===> e[x#/y#]
1085 (when x#, y# are of primitive type, of course). We can't (in general)
1086 do this for algebraic cases, because we might turn bottom into
1089 Actually, we generalise this idea to look for a case where we're
1090 scrutinising a variable, and we know that only the default case can
1095 other -> ...(case x of
1099 Here the inner case can be eliminated. This really only shows up in
1100 eliminating error-checking code.
1102 We also make sure that we deal with this very common case:
1107 Here we are using the case as a strict let; if x is used only once
1108 then we want to inline it. We have to be careful that this doesn't
1109 make the program terminate when it would have diverged before, so we
1111 - x is used strictly, or
1112 - e is already evaluated (it may so if e is a variable)
1114 Lastly, we generalise the transformation to handle this:
1120 We only do this for very cheaply compared r's (constructors, literals
1121 and variables). If pedantic bottoms is on, we only do it when the
1122 scrutinee is a PrimOp which can't fail.
1124 We do it *here*, looking at un-simplified alternatives, because we
1125 have to check that r doesn't mention the variables bound by the
1126 pattern in each alternative, so the binder-info is rather useful.
1128 So the case-elimination algorithm is:
1130 1. Eliminate alternatives which can't match
1132 2. Check whether all the remaining alternatives
1133 (a) do not mention in their rhs any of the variables bound in their pattern
1134 and (b) have equal rhss
1136 3. Check we can safely ditch the case:
1137 * PedanticBottoms is off,
1138 or * the scrutinee is an already-evaluated variable
1139 or * the scrutinee is a primop which is ok for speculation
1140 -- ie we want to preserve divide-by-zero errors, and
1141 -- calls to error itself!
1143 or * [Prim cases] the scrutinee is a primitive variable
1145 or * [Alg cases] the scrutinee is a variable and
1146 either * the rhs is the same variable
1147 (eg case x of C a b -> x ===> x)
1148 or * there is only one alternative, the default alternative,
1149 and the binder is used strictly in its scope.
1150 [NB this is helped by the "use default binder where
1151 possible" transformation; see below.]
1154 If so, then we can replace the case with one of the rhss.
1156 Further notes about case elimination
1157 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
1158 Consider: test :: Integer -> IO ()
1161 Turns out that this compiles to:
1164 eta1 :: State# RealWorld ->
1165 case PrelNum.< eta PrelNum.zeroInteger of wild { __DEFAULT ->
1167 (PrelNum.jtos eta ($w[] @ Char))
1169 of wild1 { (# new_s, a4 #) -> PrelIO.lvl23 new_s }}
1171 Notice the strange '<' which has no effect at all. This is a funny one.
1172 It started like this:
1174 f x y = if x < 0 then jtos x
1175 else if y==0 then "" else jtos x
1177 At a particular call site we have (f v 1). So we inline to get
1179 if v < 0 then jtos x
1180 else if 1==0 then "" else jtos x
1182 Now simplify the 1==0 conditional:
1184 if v<0 then jtos v else jtos v
1186 Now common-up the two branches of the case:
1188 case (v<0) of DEFAULT -> jtos v
1190 Why don't we drop the case? Because it's strict in v. It's technically
1191 wrong to drop even unnecessary evaluations, and in practice they
1192 may be a result of 'seq' so we *definitely* don't want to drop those.
1193 I don't really know how to improve this situation.
1197 --------------------------------------------------
1198 -- 0. Check for empty alternatives
1199 --------------------------------------------------
1202 mkCase1 scrut case_bndr []
1203 = pprTrace "mkCase1: null alts" (ppr case_bndr <+> ppr scrut) $
1207 --------------------------------------------------
1208 -- 1. Eliminate the case altogether if poss
1209 --------------------------------------------------
1211 mkCase1 scrut case_bndr [(con,bndrs,rhs)]
1212 -- See if we can get rid of the case altogether
1213 -- See the extensive notes on case-elimination above
1214 -- mkCase made sure that if all the alternatives are equal,
1215 -- then there is now only one (DEFAULT) rhs
1216 | all isDeadBinder bndrs,
1218 -- Check that the scrutinee can be let-bound instead of case-bound
1219 exprOkForSpeculation scrut
1220 -- OK not to evaluate it
1221 -- This includes things like (==# a# b#)::Bool
1222 -- so that we simplify
1223 -- case ==# a# b# of { True -> x; False -> x }
1226 -- This particular example shows up in default methods for
1227 -- comparision operations (e.g. in (>=) for Int.Int32)
1228 || exprIsValue scrut -- It's already evaluated
1229 || var_demanded_later scrut -- It'll be demanded later
1231 -- || not opt_SimplPedanticBottoms) -- Or we don't care!
1232 -- We used to allow improving termination by discarding cases, unless -fpedantic-bottoms was on,
1233 -- but that breaks badly for the dataToTag# primop, which relies on a case to evaluate
1234 -- its argument: case x of { y -> dataToTag# y }
1235 -- Here we must *not* discard the case, because dataToTag# just fetches the tag from
1236 -- the info pointer. So we'll be pedantic all the time, and see if that gives any
1238 -- Also we don't want to discard 'seq's
1239 = tick (CaseElim case_bndr) `thenSmpl_`
1240 returnSmpl (bindCaseBndr case_bndr scrut rhs)
1243 -- The case binder is going to be evaluated later,
1244 -- and the scrutinee is a simple variable
1245 var_demanded_later (Var v) = isStrictDmd (idNewDemandInfo case_bndr)
1246 var_demanded_later other = False
1249 --------------------------------------------------
1251 --------------------------------------------------
1253 mkCase1 scrut case_bndr alts -- Identity case
1254 | all identity_alt alts
1255 = tick (CaseIdentity case_bndr) `thenSmpl_`
1256 returnSmpl (re_note scrut)
1258 identity_alt (con, args, rhs) = de_note rhs `cheapEqExpr` identity_rhs con args
1260 identity_rhs (DataAlt con) args = mkConApp con (arg_tys ++ map varToCoreExpr args)
1261 identity_rhs (LitAlt lit) _ = Lit lit
1262 identity_rhs DEFAULT _ = Var case_bndr
1264 arg_tys = map Type (tyConAppArgs (idType case_bndr))
1267 -- case coerce T e of x { _ -> coerce T' x }
1268 -- And we definitely want to eliminate this case!
1269 -- So we throw away notes from the RHS, and reconstruct
1270 -- (at least an approximation) at the other end
1271 de_note (Note _ e) = de_note e
1274 -- re_note wraps a coerce if it might be necessary
1275 re_note scrut = case head alts of
1276 (_,_,rhs1@(Note _ _)) -> mkCoerce2 (exprType rhs1) (idType case_bndr) scrut
1280 --------------------------------------------------
1282 --------------------------------------------------
1283 mkCase1 scrut bndr alts = returnSmpl (Case scrut bndr alts)
1287 When adding auxiliary bindings for the case binder, it's worth checking if
1288 its dead, because it often is, and occasionally these mkCase transformations
1289 cascade rather nicely.
1292 bindCaseBndr bndr rhs body
1293 | isDeadBinder bndr = body
1294 | otherwise = bindNonRec bndr rhs body