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, isDataConId,
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
228 | null ss && discardableCont cont = (reverse acc, discardCont cont, inl)
229 | otherwise = (reverse acc, cont, inl)
231 ----------------------------
232 vanilla_stricts, computed_stricts :: [Bool]
233 vanilla_stricts = repeat False
234 computed_stricts = zipWith (||) fun_stricts arg_stricts
236 ----------------------------
237 (val_arg_tys, _) = splitFunTys (dropForAlls (idType fun))
238 arg_stricts = map isStrictType val_arg_tys ++ repeat False
239 -- These argument types are used as a cheap and cheerful way to find
240 -- unboxed arguments, which must be strict. But it's an InType
241 -- and so there might be a type variable where we expect a function
242 -- type (the substitution hasn't happened yet). And we don't bother
243 -- doing the type applications for a polymorphic function.
244 -- Hence the splitFunTys*IgnoringForAlls*
246 ----------------------------
247 -- If fun_stricts is finite, it means the function returns bottom
248 -- after that number of value args have been consumed
249 -- Otherwise it's infinite, extended with False
251 = case splitStrictSig (idNewStrictness fun) of
252 (demands, result_info)
253 | not (demands `lengthExceeds` countValArgs orig_cont)
254 -> -- Enough args, use the strictness given.
255 -- For bottoming functions we used to pretend that the arg
256 -- is lazy, so that we don't treat the arg as an
257 -- interesting context. This avoids substituting
258 -- top-level bindings for (say) strings into
259 -- calls to error. But now we are more careful about
260 -- inlining lone variables, so its ok (see SimplUtils.analyseCont)
261 if isBotRes result_info then
262 map isStrictDmd demands -- Finite => result is bottom
264 map isStrictDmd demands ++ vanilla_stricts
266 other -> vanilla_stricts -- Not enough args, or no strictness
269 interestingArg :: OutExpr -> Bool
270 -- An argument is interesting if it has *some* structure
271 -- We are here trying to avoid unfolding a function that
272 -- is applied only to variables that have no unfolding
273 -- (i.e. they are probably lambda bound): f x y z
274 -- There is little point in inlining f here.
275 interestingArg (Var v) = hasSomeUnfolding (idUnfolding v)
276 -- Was: isValueUnfolding (idUnfolding v')
277 -- But that seems over-pessimistic
279 -- This accounts for an argument like
280 -- () or [], which is definitely interesting
281 interestingArg (Type _) = False
282 interestingArg (App fn (Type _)) = interestingArg fn
283 interestingArg (Note _ a) = interestingArg a
284 interestingArg other = True
285 -- Consider let x = 3 in f x
286 -- The substitution will contain (x -> ContEx 3), and we want to
287 -- to say that x is an interesting argument.
288 -- But consider also (\x. f x y) y
289 -- The substitution will contain (x -> ContEx y), and we want to say
290 -- that x is not interesting (assuming y has no unfolding)
293 Comment about interestingCallContext
294 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
295 We want to avoid inlining an expression where there can't possibly be
296 any gain, such as in an argument position. Hence, if the continuation
297 is interesting (eg. a case scrutinee, application etc.) then we
298 inline, otherwise we don't.
300 Previously some_benefit used to return True only if the variable was
301 applied to some value arguments. This didn't work:
303 let x = _coerce_ (T Int) Int (I# 3) in
304 case _coerce_ Int (T Int) x of
307 we want to inline x, but can't see that it's a constructor in a case
308 scrutinee position, and some_benefit is False.
312 dMonadST = _/\_ t -> :Monad (g1 _@_ t, g2 _@_ t, g3 _@_ t)
314 .... case dMonadST _@_ x0 of (a,b,c) -> ....
316 we'd really like to inline dMonadST here, but we *don't* want to
317 inline if the case expression is just
319 case x of y { DEFAULT -> ... }
321 since we can just eliminate this case instead (x is in WHNF). Similar
322 applies when x is bound to a lambda expression. Hence
323 contIsInteresting looks for case expressions with just a single
327 interestingCallContext :: Bool -- False <=> no args at all
328 -> Bool -- False <=> no value args
330 -- The "lone-variable" case is important. I spent ages
331 -- messing about with unsatisfactory varaints, but this is nice.
332 -- The idea is that if a variable appear all alone
333 -- as an arg of lazy fn, or rhs Stop
334 -- as scrutinee of a case Select
335 -- as arg of a strict fn ArgOf
336 -- then we should not inline it (unless there is some other reason,
337 -- e.g. is is the sole occurrence). We achieve this by making
338 -- interestingCallContext return False for a lone variable.
340 -- Why? At least in the case-scrutinee situation, turning
341 -- let x = (a,b) in case x of y -> ...
343 -- let x = (a,b) in case (a,b) of y -> ...
345 -- let x = (a,b) in let y = (a,b) in ...
346 -- is bad if the binding for x will remain.
348 -- Another example: I discovered that strings
349 -- were getting inlined straight back into applications of 'error'
350 -- because the latter is strict.
352 -- f = \x -> ...(error s)...
354 -- Fundamentally such contexts should not ecourage inlining because
355 -- the context can ``see'' the unfolding of the variable (e.g. case or a RULE)
356 -- so there's no gain.
358 -- However, even a type application or coercion isn't a lone variable.
360 -- case $fMonadST @ RealWorld of { :DMonad a b c -> c }
361 -- We had better inline that sucker! The case won't see through it.
363 -- For now, I'm treating treating a variable applied to types
364 -- in a *lazy* context "lone". The motivating example was
366 -- g = /\a. \y. h (f a)
367 -- There's no advantage in inlining f here, and perhaps
368 -- a significant disadvantage. Hence some_val_args in the Stop case
370 interestingCallContext some_args some_val_args cont
373 interesting (InlinePlease _) = True
374 interesting (Select _ _ _ _ _) = some_args
375 interesting (ApplyTo _ _ _ _) = True -- Can happen if we have (coerce t (f x)) y
376 -- Perhaps True is a bit over-keen, but I've
377 -- seen (coerce f) x, where f has an INLINE prag,
378 -- So we have to give some motivaiton for inlining it
379 interesting (ArgOf _ _ _ _) = some_val_args
380 interesting (Stop ty _ upd_in_place) = some_val_args && upd_in_place
381 interesting (CoerceIt _ cont) = interesting cont
382 -- If this call is the arg of a strict function, the context
383 -- is a bit interesting. If we inline here, we may get useful
384 -- evaluation information to avoid repeated evals: e.g.
386 -- Here the contIsInteresting makes the '*' keener to inline,
387 -- which in turn exposes a constructor which makes the '+' inline.
388 -- Assuming that +,* aren't small enough to inline regardless.
390 -- It's also very important to inline in a strict context for things
393 -- Here, the context of (f x) is strict, and if f's unfolding is
394 -- a build it's *great* to inline it here. So we must ensure that
395 -- the context for (f x) is not totally uninteresting.
399 canUpdateInPlace :: Type -> Bool
400 -- Consider let x = <wurble> in ...
401 -- If <wurble> returns an explicit constructor, we might be able
402 -- to do update in place. So we treat even a thunk RHS context
403 -- as interesting if update in place is possible. We approximate
404 -- this by seeing if the type has a single constructor with a
405 -- small arity. But arity zero isn't good -- we share the single copy
406 -- for that case, so no point in sharing.
409 | not opt_UF_UpdateInPlace = False
411 = case splitTyConApp_maybe ty of
413 Just (tycon, _) -> case tyConDataCons_maybe tycon of
414 Just [dc] -> arity == 1 || arity == 2
416 arity = dataConRepArity dc
422 %************************************************************************
424 \section{Dealing with a single binder}
426 %************************************************************************
428 These functions are in the monad only so that they can be made strict via seq.
431 simplBinders :: SimplEnv -> [InBinder] -> SimplM (SimplEnv, [OutBinder])
432 simplBinders env bndrs
434 (subst', bndrs') = Subst.simplBndrs (getSubst env) bndrs
436 seqBndrs bndrs' `seq`
437 returnSmpl (setSubst env subst', bndrs')
439 simplBinder :: SimplEnv -> InBinder -> SimplM (SimplEnv, OutBinder)
442 (subst', bndr') = Subst.simplBndr (getSubst env) bndr
445 returnSmpl (setSubst env subst', bndr')
448 simplLetBndr :: SimplEnv -> InBinder -> SimplM (SimplEnv, OutBinder)
451 (subst', id') = Subst.simplLetId (getSubst env) id
454 returnSmpl (setSubst env subst', id')
456 simplLamBndrs, simplRecBndrs
457 :: SimplEnv -> [InBinder] -> SimplM (SimplEnv, [OutBinder])
458 simplRecBndrs = simplBndrs Subst.simplLetId
459 simplLamBndrs = simplBndrs Subst.simplLamBndr
461 simplBndrs simpl_bndr env bndrs
463 (subst', bndrs') = mapAccumL simpl_bndr (getSubst env) bndrs
465 seqBndrs bndrs' `seq`
466 returnSmpl (setSubst env subst', bndrs')
469 seqBndrs (b:bs) = seqBndr b `seq` seqBndrs bs
471 seqBndr b | isTyVar b = b `seq` ()
472 | otherwise = seqType (idType b) `seq`
479 newId :: EncodedFS -> Type -> SimplM Id
480 newId fs ty = getUniqueSmpl `thenSmpl` \ uniq ->
481 returnSmpl (mkSysLocal fs uniq ty)
485 %************************************************************************
487 \subsection{Rebuilding a lambda}
489 %************************************************************************
492 mkLam :: SimplEnv -> [OutBinder] -> OutExpr -> SimplCont -> SimplM FloatsWithExpr
496 a) eta reduction, if that gives a trivial expression
497 b) eta expansion [only if there are some value lambdas]
498 c) floating lets out through big lambdas
499 [only if all tyvar lambdas, and only if this lambda
503 mkLam env bndrs body cont
504 | opt_SimplDoEtaReduction,
505 Just etad_lam <- tryEtaReduce bndrs body
506 = tick (EtaReduction (head bndrs)) `thenSmpl_`
507 returnSmpl (emptyFloats env, etad_lam)
509 | opt_SimplDoLambdaEtaExpansion,
510 any isRuntimeVar bndrs
511 = tryEtaExpansion body `thenSmpl` \ body' ->
512 returnSmpl (emptyFloats env, mkLams bndrs body')
514 {- Sept 01: I'm experimenting with getting the
515 full laziness pass to float out past big lambdsa
516 | all isTyVar bndrs, -- Only for big lambdas
517 contIsRhs cont -- Only try the rhs type-lambda floating
518 -- if this is indeed a right-hand side; otherwise
519 -- we end up floating the thing out, only for float-in
520 -- to float it right back in again!
521 = tryRhsTyLam env bndrs body `thenSmpl` \ (floats, body') ->
522 returnSmpl (floats, mkLams bndrs body')
526 = returnSmpl (emptyFloats env, mkLams bndrs body)
530 %************************************************************************
532 \subsection{Eta expansion and reduction}
534 %************************************************************************
536 We try for eta reduction here, but *only* if we get all the
537 way to an exprIsTrivial expression.
538 We don't want to remove extra lambdas unless we are going
539 to avoid allocating this thing altogether
542 tryEtaReduce :: [OutBinder] -> OutExpr -> Maybe OutExpr
543 tryEtaReduce bndrs body
544 -- We don't use CoreUtils.etaReduce, because we can be more
546 -- (a) we already have the binders
547 -- (b) we can do the triviality test before computing the free vars
548 = go (reverse bndrs) body
550 go (b : bs) (App fun arg) | ok_arg b arg = go bs fun -- Loop round
551 go [] fun | ok_fun fun = Just fun -- Success!
552 go _ _ = Nothing -- Failure!
554 ok_fun fun = exprIsTrivial fun
555 && not (any (`elemVarSet` (exprFreeVars fun)) bndrs)
556 && (exprIsValue fun || all ok_lam bndrs)
557 ok_lam v = isTyVar v || isDictTy (idType v)
558 -- The exprIsValue is because eta reduction is not
559 -- valid in general: \x. bot /= bot
560 -- So we need to be sure that the "fun" is a value.
562 -- However, we always want to reduce (/\a -> f a) to f
563 -- This came up in a RULE: foldr (build (/\a -> g a))
564 -- did not match foldr (build (/\b -> ...something complex...))
565 -- The type checker can insert these eta-expanded versions,
566 -- with both type and dictionary lambdas; hence the slightly
569 ok_arg b arg = varToCoreExpr b `cheapEqExpr` arg
573 Try eta expansion for RHSs
576 f = \x1..xn -> N ==> f = \x1..xn y1..ym -> N y1..ym
579 where (in both cases)
581 * The xi can include type variables
583 * The yi are all value variables
585 * N is a NORMAL FORM (i.e. no redexes anywhere)
586 wanting a suitable number of extra args.
588 We may have to sandwich some coerces between the lambdas
589 to make the types work. exprEtaExpandArity looks through coerces
590 when computing arity; and etaExpand adds the coerces as necessary when
591 actually computing the expansion.
594 tryEtaExpansion :: OutExpr -> SimplM OutExpr
595 -- There is at least one runtime binder in the binders
597 = getUniquesSmpl `thenSmpl` \ us ->
598 returnSmpl (etaExpand fun_arity us body (exprType body))
600 fun_arity = exprEtaExpandArity body
604 %************************************************************************
606 \subsection{Floating lets out of big lambdas}
608 %************************************************************************
610 tryRhsTyLam tries this transformation, when the big lambda appears as
611 the RHS of a let(rec) binding:
613 /\abc -> let(rec) x = e in b
615 let(rec) x' = /\abc -> let x = x' a b c in e
617 /\abc -> let x = x' a b c in b
619 This is good because it can turn things like:
621 let f = /\a -> letrec g = ... g ... in g
623 letrec g' = /\a -> ... g' a ...
627 which is better. In effect, it means that big lambdas don't impede
630 This optimisation is CRUCIAL in eliminating the junk introduced by
631 desugaring mutually recursive definitions. Don't eliminate it lightly!
633 So far as the implementation is concerned:
635 Invariant: go F e = /\tvs -> F e
639 = Let x' = /\tvs -> F e
643 G = F . Let x = x' tvs
645 go F (Letrec xi=ei in b)
646 = Letrec {xi' = /\tvs -> G ei}
650 G = F . Let {xi = xi' tvs}
652 [May 1999] If we do this transformation *regardless* then we can
653 end up with some pretty silly stuff. For example,
656 st = /\ s -> let { x1=r1 ; x2=r2 } in ...
661 st = /\s -> ...[y1 s/x1, y2 s/x2]
664 Unless the "..." is a WHNF there is really no point in doing this.
665 Indeed it can make things worse. Suppose x1 is used strictly,
668 x1* = case f y of { (a,b) -> e }
670 If we abstract this wrt the tyvar we then can't do the case inline
671 as we would normally do.
675 {- Trying to do this in full laziness
677 tryRhsTyLam :: SimplEnv -> [OutTyVar] -> OutExpr -> SimplM FloatsWithExpr
678 -- Call ensures that all the binders are type variables
680 tryRhsTyLam env tyvars body -- Only does something if there's a let
681 | not (all isTyVar tyvars)
682 || not (worth_it body) -- inside a type lambda,
683 = returnSmpl (emptyFloats env, body) -- and a WHNF inside that
686 = go env (\x -> x) body
689 worth_it e@(Let _ _) = whnf_in_middle e
692 whnf_in_middle (Let (NonRec x rhs) e) | isUnLiftedType (idType x) = False
693 whnf_in_middle (Let _ e) = whnf_in_middle e
694 whnf_in_middle e = exprIsCheap e
696 main_tyvar_set = mkVarSet tyvars
698 go env fn (Let bind@(NonRec var rhs) body)
700 = go env (fn . Let bind) body
702 go env fn (Let (NonRec var rhs) body)
703 = mk_poly tyvars_here var `thenSmpl` \ (var', rhs') ->
704 addAuxiliaryBind env (NonRec var' (mkLams tyvars_here (fn rhs))) $ \ env ->
705 go env (fn . Let (mk_silly_bind var rhs')) body
709 tyvars_here = varSetElems (main_tyvar_set `intersectVarSet` exprSomeFreeVars isTyVar rhs)
710 -- Abstract only over the type variables free in the rhs
711 -- wrt which the new binding is abstracted. But the naive
712 -- approach of abstract wrt the tyvars free in the Id's type
714 -- /\ a b -> let t :: (a,b) = (e1, e2)
717 -- Here, b isn't free in x's type, but we must nevertheless
718 -- abstract wrt b as well, because t's type mentions b.
719 -- Since t is floated too, we'd end up with the bogus:
720 -- poly_t = /\ a b -> (e1, e2)
721 -- poly_x = /\ a -> fst (poly_t a *b*)
722 -- So for now we adopt the even more naive approach of
723 -- abstracting wrt *all* the tyvars. We'll see if that
724 -- gives rise to problems. SLPJ June 98
726 go env fn (Let (Rec prs) body)
727 = mapAndUnzipSmpl (mk_poly tyvars_here) vars `thenSmpl` \ (vars', rhss') ->
729 gn body = fn (foldr Let body (zipWith mk_silly_bind vars rhss'))
730 pairs = vars' `zip` [mkLams tyvars_here (gn rhs) | rhs <- rhss]
732 addAuxiliaryBind env (Rec pairs) $ \ env ->
735 (vars,rhss) = unzip prs
736 tyvars_here = varSetElems (main_tyvar_set `intersectVarSet` exprsSomeFreeVars isTyVar (map snd prs))
737 -- See notes with tyvars_here above
739 go env fn body = returnSmpl (emptyFloats env, fn body)
741 mk_poly tyvars_here var
742 = getUniqueSmpl `thenSmpl` \ uniq ->
744 poly_name = setNameUnique (idName var) uniq -- Keep same name
745 poly_ty = mkForAllTys tyvars_here (idType var) -- But new type of course
746 poly_id = mkLocalId poly_name poly_ty
748 -- In the olden days, it was crucial to copy the occInfo of the original var,
749 -- because we were looking at occurrence-analysed but as yet unsimplified code!
750 -- In particular, we mustn't lose the loop breakers. BUT NOW we are looking
751 -- at already simplified code, so it doesn't matter
753 -- It's even right to retain single-occurrence or dead-var info:
754 -- Suppose we started with /\a -> let x = E in B
755 -- where x occurs once in B. Then we transform to:
756 -- let x' = /\a -> E in /\a -> let x* = x' a in B
757 -- where x* has an INLINE prag on it. Now, once x* is inlined,
758 -- the occurrences of x' will be just the occurrences originally
761 returnSmpl (poly_id, mkTyApps (Var poly_id) (mkTyVarTys tyvars_here))
763 mk_silly_bind var rhs = NonRec var (Note InlineMe rhs)
764 -- Suppose we start with:
766 -- x = /\ a -> let g = G in E
768 -- Then we'll float to get
770 -- x = let poly_g = /\ a -> G
771 -- in /\ a -> let g = poly_g a in E
773 -- But now the occurrence analyser will see just one occurrence
774 -- of poly_g, not inside a lambda, so the simplifier will
775 -- PreInlineUnconditionally poly_g back into g! Badk to square 1!
776 -- (I used to think that the "don't inline lone occurrences" stuff
777 -- would stop this happening, but since it's the *only* occurrence,
778 -- PreInlineUnconditionally kicks in first!)
780 -- Solution: put an INLINE note on g's RHS, so that poly_g seems
781 -- to appear many times. (NB: mkInlineMe eliminates
782 -- such notes on trivial RHSs, so do it manually.)
786 %************************************************************************
788 \subsection{Case alternative filtering
790 %************************************************************************
792 prepareAlts does two things:
794 1. Eliminate alternatives that cannot match, including the
797 2. If the DEFAULT alternative can match only one possible constructor,
798 then make that constructor explicit.
800 case e of x { DEFAULT -> rhs }
802 case e of x { (a,b) -> rhs }
803 where the type is a single constructor type. This gives better code
804 when rhs also scrutinises x or e.
806 It's a good idea do do this stuff before simplifying the alternatives, to
807 avoid simplifying alternatives we know can't happen, and to come up with
808 the list of constructors that are handled, to put into the IdInfo of the
809 case binder, for use when simplifying the alternatives.
811 Eliminating the default alternative in (1) isn't so obvious, but it can
814 data Colour = Red | Green | Blue
823 DEFAULT -> [ case y of ... ]
825 If we inline h into f, the default case of the inlined h can't happen.
826 If we don't notice this, we may end up filtering out *all* the cases
827 of the inner case y, which give us nowhere to go!
831 prepareAlts :: OutExpr -- Scrutinee
832 -> InId -- Case binder
834 -> SimplM ([InAlt], -- Better alternatives
835 [AltCon]) -- These cases are handled
837 prepareAlts scrut case_bndr alts
839 (alts_wo_default, maybe_deflt) = findDefault alts
841 impossible_cons = case scrut of
842 Var v -> otherCons (idUnfolding v)
845 -- Filter out alternatives that can't possibly match
846 better_alts | null impossible_cons = alts_wo_default
847 | otherwise = [alt | alt@(con,_,_) <- alts_wo_default,
848 not (con `elem` impossible_cons)]
850 -- "handled_cons" are handled either by the context,
851 -- or by a branch in this case expression
852 -- (Don't add DEFAULT to the handled_cons!!)
853 handled_cons = impossible_cons ++ [con | (con,_,_) <- better_alts]
855 -- Filter out the default, if it can't happen,
856 -- or replace it with "proper" alternative if there
857 -- is only one constructor left
858 prepareDefault case_bndr handled_cons maybe_deflt `thenSmpl` \ deflt_alt ->
860 returnSmpl (deflt_alt ++ better_alts, handled_cons)
862 prepareDefault case_bndr handled_cons (Just rhs)
863 | Just (tycon, inst_tys) <- splitTyConApp_maybe (idType case_bndr),
864 isAlgTyCon tycon, -- It's a data type, tuple, or unboxed tuples.
865 not (isNewTyCon tycon), -- We can have a newtype, if we are just doing an eval:
866 -- case x of { DEFAULT -> e }
867 -- and we don't want to fill in a default for them!
868 Just all_cons <- tyConDataCons_maybe tycon,
869 not (null all_cons), -- This is a tricky corner case. If the data type has no constructors,
870 -- which GHC allows, then the case expression will have at most a default
871 -- alternative. We don't want to eliminate that alternative, because the
872 -- invariant is that there's always one alternative. It's more convenient
874 -- case x of { DEFAULT -> e }
875 -- as it is, rather than transform it to
876 -- error "case cant match"
877 -- which would be quite legitmate. But it's a really obscure corner, and
878 -- not worth wasting code on.
879 let handled_data_cons = [data_con | DataAlt data_con <- handled_cons],
880 let missing_cons = [con | con <- all_cons,
881 not (con `elem` handled_data_cons)]
882 = case missing_cons of
883 [] -> returnSmpl [] -- Eliminate the default alternative
886 [con] -> -- It matches exactly one constructor, so fill it in
887 tick (FillInCaseDefault case_bndr) `thenSmpl_`
888 mk_args con inst_tys `thenSmpl` \ args ->
889 returnSmpl [(DataAlt con, args, rhs)]
891 two_or_more -> returnSmpl [(DEFAULT, [], rhs)]
894 = returnSmpl [(DEFAULT, [], rhs)]
896 prepareDefault case_bndr handled_cons Nothing
899 mk_args missing_con inst_tys
900 = getUniquesSmpl `thenSmpl` \ tv_uniqs ->
901 getUniquesSmpl `thenSmpl` \ id_uniqs ->
903 ex_tyvars = dataConExistentialTyVars missing_con
904 ex_tyvars' = zipWith mk tv_uniqs ex_tyvars
905 mk uniq tv = mkSysTyVar uniq (tyVarKind tv)
906 arg_tys = dataConArgTys missing_con (inst_tys ++ mkTyVarTys ex_tyvars')
907 arg_ids = zipWith (mkSysLocal FSLIT("a")) id_uniqs arg_tys
909 returnSmpl (ex_tyvars' ++ arg_ids)
913 %************************************************************************
915 \subsection{Case absorption and identity-case elimination}
917 %************************************************************************
919 mkCase puts a case expression back together, trying various transformations first.
922 mkCase :: OutExpr -> OutId -> [OutAlt] -> SimplM OutExpr
924 mkCase scrut case_bndr alts
925 = mkAlts scrut case_bndr alts `thenSmpl` \ better_alts ->
926 mkCase1 scrut case_bndr better_alts
930 mkAlts tries these things:
932 1. If several alternatives are identical, merge them into
933 a single DEFAULT alternative. I've occasionally seen this
934 making a big difference:
936 case e of =====> case e of
937 C _ -> f x D v -> ....v....
938 D v -> ....v.... DEFAULT -> f x
941 The point is that we merge common RHSs, at least for the DEFAULT case.
942 [One could do something more elaborate but I've never seen it needed.]
943 To avoid an expensive test, we just merge branches equal to the *first*
944 alternative; this picks up the common cases
945 a) all branches equal
946 b) some branches equal to the DEFAULT (which occurs first)
949 case e of b { ==> case e of b {
950 p1 -> rhs1 p1 -> rhs1
952 pm -> rhsm pm -> rhsm
953 _ -> case b of b' { pn -> let b'=b in rhsn
955 ... po -> let b'=b in rhso
956 po -> rhso _ -> let b'=b in rhsd
960 which merges two cases in one case when -- the default alternative of
961 the outer case scrutises the same variable as the outer case This
962 transformation is called Case Merging. It avoids that the same
963 variable is scrutinised multiple times.
966 The case where transformation (1) showed up was like this (lib/std/PrelCError.lhs):
972 where @is@ was something like
974 p `is` n = p /= (-1) && p == n
976 This gave rise to a horrible sequence of cases
983 and similarly in cascade for all the join points!
988 --------------------------------------------------
989 -- 1. Merge identical branches
990 --------------------------------------------------
991 mkAlts scrut case_bndr alts@((con1,bndrs1,rhs1) : con_alts)
992 | all isDeadBinder bndrs1, -- Remember the default
993 length filtered_alts < length con_alts -- alternative comes first
994 = tick (AltMerge case_bndr) `thenSmpl_`
995 returnSmpl better_alts
997 filtered_alts = filter keep con_alts
998 keep (con,bndrs,rhs) = not (all isDeadBinder bndrs && rhs `cheapEqExpr` rhs1)
999 better_alts = (DEFAULT, [], rhs1) : filtered_alts
1002 --------------------------------------------------
1003 -- 2. Merge nested cases
1004 --------------------------------------------------
1006 mkAlts scrut outer_bndr outer_alts
1007 | opt_SimplCaseMerge,
1008 (outer_alts_without_deflt, maybe_outer_deflt) <- findDefault outer_alts,
1009 Just (Case (Var scrut_var) inner_bndr inner_alts) <- maybe_outer_deflt,
1010 scruting_same_var scrut_var
1012 = let -- Eliminate any inner alts which are shadowed by the outer ones
1013 outer_cons = [con | (con,_,_) <- outer_alts_without_deflt]
1015 munged_inner_alts = [ (con, args, munge_rhs rhs)
1016 | (con, args, rhs) <- inner_alts,
1017 not (con `elem` outer_cons) -- Eliminate shadowed inner alts
1019 munge_rhs rhs = bindCaseBndr inner_bndr (Var outer_bndr) rhs
1021 (inner_con_alts, maybe_inner_default) = findDefault munged_inner_alts
1023 new_alts = add_default maybe_inner_default
1024 (outer_alts_without_deflt ++ inner_con_alts)
1026 tick (CaseMerge outer_bndr) `thenSmpl_`
1028 -- Warning: don't call mkAlts recursively!
1029 -- Firstly, there's no point, because inner alts have already had
1030 -- mkCase applied to them, so they won't have a case in their default
1031 -- Secondly, if you do, you get an infinite loop, because the bindCaseBndr
1032 -- in munge_rhs may put a case into the DEFAULT branch!
1034 -- We are scrutinising the same variable if it's
1035 -- the outer case-binder, or if the outer case scrutinises a variable
1036 -- (and it's the same). Testing both allows us not to replace the
1037 -- outer scrut-var with the outer case-binder (Simplify.simplCaseBinder).
1038 scruting_same_var = case scrut of
1039 Var outer_scrut -> \ v -> v == outer_bndr || v == outer_scrut
1040 other -> \ v -> v == outer_bndr
1042 add_default (Just rhs) alts = (DEFAULT,[],rhs) : alts
1043 add_default Nothing alts = alts
1046 --------------------------------------------------
1048 --------------------------------------------------
1050 mkAlts scrut case_bndr other_alts = returnSmpl other_alts
1055 =================================================================================
1057 mkCase1 tries these things
1059 1. Eliminate the case altogether if possible
1067 and similar friends.
1070 Start with a simple situation:
1072 case x# of ===> e[x#/y#]
1075 (when x#, y# are of primitive type, of course). We can't (in general)
1076 do this for algebraic cases, because we might turn bottom into
1079 Actually, we generalise this idea to look for a case where we're
1080 scrutinising a variable, and we know that only the default case can
1085 other -> ...(case x of
1089 Here the inner case can be eliminated. This really only shows up in
1090 eliminating error-checking code.
1092 We also make sure that we deal with this very common case:
1097 Here we are using the case as a strict let; if x is used only once
1098 then we want to inline it. We have to be careful that this doesn't
1099 make the program terminate when it would have diverged before, so we
1101 - x is used strictly, or
1102 - e is already evaluated (it may so if e is a variable)
1104 Lastly, we generalise the transformation to handle this:
1110 We only do this for very cheaply compared r's (constructors, literals
1111 and variables). If pedantic bottoms is on, we only do it when the
1112 scrutinee is a PrimOp which can't fail.
1114 We do it *here*, looking at un-simplified alternatives, because we
1115 have to check that r doesn't mention the variables bound by the
1116 pattern in each alternative, so the binder-info is rather useful.
1118 So the case-elimination algorithm is:
1120 1. Eliminate alternatives which can't match
1122 2. Check whether all the remaining alternatives
1123 (a) do not mention in their rhs any of the variables bound in their pattern
1124 and (b) have equal rhss
1126 3. Check we can safely ditch the case:
1127 * PedanticBottoms is off,
1128 or * the scrutinee is an already-evaluated variable
1129 or * the scrutinee is a primop which is ok for speculation
1130 -- ie we want to preserve divide-by-zero errors, and
1131 -- calls to error itself!
1133 or * [Prim cases] the scrutinee is a primitive variable
1135 or * [Alg cases] the scrutinee is a variable and
1136 either * the rhs is the same variable
1137 (eg case x of C a b -> x ===> x)
1138 or * there is only one alternative, the default alternative,
1139 and the binder is used strictly in its scope.
1140 [NB this is helped by the "use default binder where
1141 possible" transformation; see below.]
1144 If so, then we can replace the case with one of the rhss.
1146 Further notes about case elimination
1147 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
1148 Consider: test :: Integer -> IO ()
1151 Turns out that this compiles to:
1154 eta1 :: State# RealWorld ->
1155 case PrelNum.< eta PrelNum.zeroInteger of wild { __DEFAULT ->
1157 (PrelNum.jtos eta ($w[] @ Char))
1159 of wild1 { (# new_s, a4 #) -> PrelIO.lvl23 new_s }}
1161 Notice the strange '<' which has no effect at all. This is a funny one.
1162 It started like this:
1164 f x y = if x < 0 then jtos x
1165 else if y==0 then "" else jtos x
1167 At a particular call site we have (f v 1). So we inline to get
1169 if v < 0 then jtos x
1170 else if 1==0 then "" else jtos x
1172 Now simplify the 1==0 conditional:
1174 if v<0 then jtos v else jtos v
1176 Now common-up the two branches of the case:
1178 case (v<0) of DEFAULT -> jtos v
1180 Why don't we drop the case? Because it's strict in v. It's technically
1181 wrong to drop even unnecessary evaluations, and in practice they
1182 may be a result of 'seq' so we *definitely* don't want to drop those.
1183 I don't really know how to improve this situation.
1187 --------------------------------------------------
1188 -- 0. Check for empty alternatives
1189 --------------------------------------------------
1192 mkCase1 scrut case_bndr []
1193 = pprTrace "mkCase1: null alts" (ppr case_bndr <+> ppr scrut) $
1197 --------------------------------------------------
1198 -- 1. Eliminate the case altogether if poss
1199 --------------------------------------------------
1201 mkCase1 scrut case_bndr [(con,bndrs,rhs)]
1202 -- See if we can get rid of the case altogether
1203 -- See the extensive notes on case-elimination above
1204 -- mkCase made sure that if all the alternatives are equal,
1205 -- then there is now only one (DEFAULT) rhs
1206 | all isDeadBinder bndrs,
1208 -- Check that the scrutinee can be let-bound instead of case-bound
1209 exprOkForSpeculation scrut
1210 -- OK not to evaluate it
1211 -- This includes things like (==# a# b#)::Bool
1212 -- so that we simplify
1213 -- case ==# a# b# of { True -> x; False -> x }
1216 -- This particular example shows up in default methods for
1217 -- comparision operations (e.g. in (>=) for Int.Int32)
1218 || exprIsValue scrut -- It's already evaluated
1219 || var_demanded_later scrut -- It'll be demanded later
1221 -- || not opt_SimplPedanticBottoms) -- Or we don't care!
1222 -- We used to allow improving termination by discarding cases, unless -fpedantic-bottoms was on,
1223 -- but that breaks badly for the dataToTag# primop, which relies on a case to evaluate
1224 -- its argument: case x of { y -> dataToTag# y }
1225 -- Here we must *not* discard the case, because dataToTag# just fetches the tag from
1226 -- the info pointer. So we'll be pedantic all the time, and see if that gives any
1228 -- Also we don't want to discard 'seq's
1229 = tick (CaseElim case_bndr) `thenSmpl_`
1230 returnSmpl (bindCaseBndr case_bndr scrut rhs)
1233 -- The case binder is going to be evaluated later,
1234 -- and the scrutinee is a simple variable
1235 var_demanded_later (Var v) = isStrictDmd (idNewDemandInfo case_bndr)
1236 var_demanded_later other = False
1239 --------------------------------------------------
1241 --------------------------------------------------
1243 mkCase1 scrut case_bndr alts -- Identity case
1244 | all identity_alt alts
1245 = tick (CaseIdentity case_bndr) `thenSmpl_`
1246 returnSmpl (re_note scrut)
1248 identity_alt (con, args, rhs) = de_note rhs `cheapEqExpr` identity_rhs con args
1250 identity_rhs (DataAlt con) args = mkConApp con (arg_tys ++ map varToCoreExpr args)
1251 identity_rhs (LitAlt lit) _ = Lit lit
1252 identity_rhs DEFAULT _ = Var case_bndr
1254 arg_tys = map Type (tyConAppArgs (idType case_bndr))
1257 -- case coerce T e of x { _ -> coerce T' x }
1258 -- And we definitely want to eliminate this case!
1259 -- So we throw away notes from the RHS, and reconstruct
1260 -- (at least an approximation) at the other end
1261 de_note (Note _ e) = de_note e
1264 -- re_note wraps a coerce if it might be necessary
1265 re_note scrut = case head alts of
1266 (_,_,rhs1@(Note _ _)) -> mkCoerce2 (exprType rhs1) (idType case_bndr) scrut
1270 --------------------------------------------------
1272 --------------------------------------------------
1273 mkCase1 scrut bndr alts = returnSmpl (Case scrut bndr alts)
1277 When adding auxiliary bindings for the case binder, it's worth checking if
1278 its dead, because it often is, and occasionally these mkCase transformations
1279 cascade rather nicely.
1282 bindCaseBndr bndr rhs body
1283 | isDeadBinder bndr = body
1284 | otherwise = bindNonRec bndr rhs body