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, mkRhsStop, 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, dataConTyVars, dataConArgTys, isVanillaDataCon )
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, mkRhsStop :: OutType -> SimplCont
119 mkBoringStop ty = Stop ty AnArg (canUpdateInPlace ty)
120 mkRhsStop ty = Stop ty AnRhs (canUpdateInPlace ty)
122 contIsRhs :: SimplCont -> Bool
123 contIsRhs (Stop _ AnRhs _) = True
124 contIsRhs (ArgOf AnRhs _ _ _) = True
125 contIsRhs other = False
127 contIsRhsOrArg (Stop _ _ _) = True
128 contIsRhsOrArg (ArgOf _ _ _ _) = True
129 contIsRhsOrArg other = False
132 contIsDupable :: SimplCont -> Bool
133 contIsDupable (Stop _ _ _) = True
134 contIsDupable (ApplyTo OkToDup _ _ _) = True
135 contIsDupable (Select OkToDup _ _ _ _) = True
136 contIsDupable (CoerceIt _ cont) = contIsDupable cont
137 contIsDupable (InlinePlease cont) = contIsDupable cont
138 contIsDupable other = False
141 discardableCont :: SimplCont -> Bool
142 discardableCont (Stop _ _ _) = False
143 discardableCont (CoerceIt _ cont) = discardableCont cont
144 discardableCont (InlinePlease cont) = discardableCont cont
145 discardableCont other = True
147 discardCont :: SimplCont -- A continuation, expecting
148 -> SimplCont -- Replace the continuation with a suitable coerce
149 discardCont cont = case cont of
150 Stop to_ty is_rhs _ -> cont
151 other -> CoerceIt to_ty (mkBoringStop to_ty)
153 to_ty = contResultType cont
156 contResultType :: SimplCont -> OutType
157 contResultType (Stop to_ty _ _) = to_ty
158 contResultType (ArgOf _ _ to_ty _) = to_ty
159 contResultType (ApplyTo _ _ _ cont) = contResultType cont
160 contResultType (CoerceIt _ cont) = contResultType cont
161 contResultType (InlinePlease cont) = contResultType cont
162 contResultType (Select _ _ _ _ cont) = contResultType cont
165 countValArgs :: SimplCont -> Int
166 countValArgs (ApplyTo _ (Type ty) se cont) = countValArgs cont
167 countValArgs (ApplyTo _ val_arg se cont) = 1 + countValArgs cont
168 countValArgs other = 0
170 countArgs :: SimplCont -> Int
171 countArgs (ApplyTo _ arg se cont) = 1 + countArgs cont
175 pushContArgs :: SimplEnv -> [OutArg] -> SimplCont -> SimplCont
176 -- Pushes args with the specified environment
177 pushContArgs env [] cont = cont
178 pushContArgs env (arg : args) cont = ApplyTo NoDup arg env (pushContArgs env args cont)
183 getContArgs :: SwitchChecker
184 -> OutId -> SimplCont
185 -> ([(InExpr, SimplEnv, Bool)], -- Arguments; the Bool is true for strict args
186 SimplCont, -- Remaining continuation
187 Bool) -- Whether we came across an InlineCall
188 -- getContArgs id k = (args, k', inl)
189 -- args are the leading ApplyTo items in k
190 -- (i.e. outermost comes first)
191 -- augmented with demand info from the functionn
192 getContArgs chkr fun orig_cont
194 -- Ignore strictness info if the no-case-of-case
195 -- flag is on. Strictness changes evaluation order
196 -- and that can change full laziness
197 stricts | switchIsOn chkr NoCaseOfCase = vanilla_stricts
198 | otherwise = computed_stricts
200 go [] stricts False orig_cont
202 ----------------------------
205 go acc ss inl (ApplyTo _ arg@(Type _) se cont)
206 = go ((arg,se,False) : acc) ss inl cont
207 -- NB: don't bother to instantiate the function type
210 go acc (s:ss) inl (ApplyTo _ arg se cont)
211 = go ((arg,se,s) : acc) ss inl cont
213 -- An Inline continuation
214 go acc ss inl (InlinePlease cont)
215 = go acc ss True cont
217 -- We're run out of arguments, or else we've run out of demands
218 -- The latter only happens if the result is guaranteed bottom
219 -- This is the case for
220 -- * case (error "hello") of { ... }
221 -- * (error "Hello") arg
222 -- * f (error "Hello") where f is strict
224 -- Then, especially in the first of these cases, we'd like to discard
225 -- the continuation, leaving just the bottoming expression. But the
226 -- type might not be right, so we may have to add a coerce.
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 = getDOptsSmpl `thenSmpl` \dflags ->
505 mkLam' dflags env bndrs body cont
507 mkLam' dflags env bndrs body cont
508 | dopt Opt_DoEtaReduction dflags,
509 Just etad_lam <- tryEtaReduce bndrs body
510 = tick (EtaReduction (head bndrs)) `thenSmpl_`
511 returnSmpl (emptyFloats env, etad_lam)
513 | dopt Opt_DoLambdaEtaExpansion dflags,
514 any isRuntimeVar bndrs
515 = tryEtaExpansion body `thenSmpl` \ body' ->
516 returnSmpl (emptyFloats env, mkLams bndrs body')
518 {- Sept 01: I'm experimenting with getting the
519 full laziness pass to float out past big lambdsa
520 | all isTyVar bndrs, -- Only for big lambdas
521 contIsRhs cont -- Only try the rhs type-lambda floating
522 -- if this is indeed a right-hand side; otherwise
523 -- we end up floating the thing out, only for float-in
524 -- to float it right back in again!
525 = tryRhsTyLam env bndrs body `thenSmpl` \ (floats, body') ->
526 returnSmpl (floats, mkLams bndrs body')
530 = returnSmpl (emptyFloats env, mkLams bndrs body)
534 %************************************************************************
536 \subsection{Eta expansion and reduction}
538 %************************************************************************
540 We try for eta reduction here, but *only* if we get all the
541 way to an exprIsTrivial expression.
542 We don't want to remove extra lambdas unless we are going
543 to avoid allocating this thing altogether
546 tryEtaReduce :: [OutBinder] -> OutExpr -> Maybe OutExpr
547 tryEtaReduce bndrs body
548 -- We don't use CoreUtils.etaReduce, because we can be more
550 -- (a) we already have the binders
551 -- (b) we can do the triviality test before computing the free vars
552 = go (reverse bndrs) body
554 go (b : bs) (App fun arg) | ok_arg b arg = go bs fun -- Loop round
555 go [] fun | ok_fun fun = Just fun -- Success!
556 go _ _ = Nothing -- Failure!
558 ok_fun fun = exprIsTrivial fun
559 && not (any (`elemVarSet` (exprFreeVars fun)) bndrs)
560 && (exprIsValue fun || all ok_lam bndrs)
561 ok_lam v = isTyVar v || isDictTy (idType v)
562 -- The exprIsValue is because eta reduction is not
563 -- valid in general: \x. bot /= bot
564 -- So we need to be sure that the "fun" is a value.
566 -- However, we always want to reduce (/\a -> f a) to f
567 -- This came up in a RULE: foldr (build (/\a -> g a))
568 -- did not match foldr (build (/\b -> ...something complex...))
569 -- The type checker can insert these eta-expanded versions,
570 -- with both type and dictionary lambdas; hence the slightly
573 ok_arg b arg = varToCoreExpr b `cheapEqExpr` arg
577 Try eta expansion for RHSs
580 f = \x1..xn -> N ==> f = \x1..xn y1..ym -> N y1..ym
583 where (in both cases)
585 * The xi can include type variables
587 * The yi are all value variables
589 * N is a NORMAL FORM (i.e. no redexes anywhere)
590 wanting a suitable number of extra args.
592 We may have to sandwich some coerces between the lambdas
593 to make the types work. exprEtaExpandArity looks through coerces
594 when computing arity; and etaExpand adds the coerces as necessary when
595 actually computing the expansion.
598 tryEtaExpansion :: OutExpr -> SimplM OutExpr
599 -- There is at least one runtime binder in the binders
601 = getUniquesSmpl `thenSmpl` \ us ->
602 returnSmpl (etaExpand fun_arity us body (exprType body))
604 fun_arity = exprEtaExpandArity body
608 %************************************************************************
610 \subsection{Floating lets out of big lambdas}
612 %************************************************************************
614 tryRhsTyLam tries this transformation, when the big lambda appears as
615 the RHS of a let(rec) binding:
617 /\abc -> let(rec) x = e in b
619 let(rec) x' = /\abc -> let x = x' a b c in e
621 /\abc -> let x = x' a b c in b
623 This is good because it can turn things like:
625 let f = /\a -> letrec g = ... g ... in g
627 letrec g' = /\a -> ... g' a ...
631 which is better. In effect, it means that big lambdas don't impede
634 This optimisation is CRUCIAL in eliminating the junk introduced by
635 desugaring mutually recursive definitions. Don't eliminate it lightly!
637 So far as the implementation is concerned:
639 Invariant: go F e = /\tvs -> F e
643 = Let x' = /\tvs -> F e
647 G = F . Let x = x' tvs
649 go F (Letrec xi=ei in b)
650 = Letrec {xi' = /\tvs -> G ei}
654 G = F . Let {xi = xi' tvs}
656 [May 1999] If we do this transformation *regardless* then we can
657 end up with some pretty silly stuff. For example,
660 st = /\ s -> let { x1=r1 ; x2=r2 } in ...
665 st = /\s -> ...[y1 s/x1, y2 s/x2]
668 Unless the "..." is a WHNF there is really no point in doing this.
669 Indeed it can make things worse. Suppose x1 is used strictly,
672 x1* = case f y of { (a,b) -> e }
674 If we abstract this wrt the tyvar we then can't do the case inline
675 as we would normally do.
679 {- Trying to do this in full laziness
681 tryRhsTyLam :: SimplEnv -> [OutTyVar] -> OutExpr -> SimplM FloatsWithExpr
682 -- Call ensures that all the binders are type variables
684 tryRhsTyLam env tyvars body -- Only does something if there's a let
685 | not (all isTyVar tyvars)
686 || not (worth_it body) -- inside a type lambda,
687 = returnSmpl (emptyFloats env, body) -- and a WHNF inside that
690 = go env (\x -> x) body
693 worth_it e@(Let _ _) = whnf_in_middle e
696 whnf_in_middle (Let (NonRec x rhs) e) | isUnLiftedType (idType x) = False
697 whnf_in_middle (Let _ e) = whnf_in_middle e
698 whnf_in_middle e = exprIsCheap e
700 main_tyvar_set = mkVarSet tyvars
702 go env fn (Let bind@(NonRec var rhs) body)
704 = go env (fn . Let bind) body
706 go env fn (Let (NonRec var rhs) body)
707 = mk_poly tyvars_here var `thenSmpl` \ (var', rhs') ->
708 addAuxiliaryBind env (NonRec var' (mkLams tyvars_here (fn rhs))) $ \ env ->
709 go env (fn . Let (mk_silly_bind var rhs')) body
713 tyvars_here = varSetElems (main_tyvar_set `intersectVarSet` exprSomeFreeVars isTyVar rhs)
714 -- Abstract only over the type variables free in the rhs
715 -- wrt which the new binding is abstracted. But the naive
716 -- approach of abstract wrt the tyvars free in the Id's type
718 -- /\ a b -> let t :: (a,b) = (e1, e2)
721 -- Here, b isn't free in x's type, but we must nevertheless
722 -- abstract wrt b as well, because t's type mentions b.
723 -- Since t is floated too, we'd end up with the bogus:
724 -- poly_t = /\ a b -> (e1, e2)
725 -- poly_x = /\ a -> fst (poly_t a *b*)
726 -- So for now we adopt the even more naive approach of
727 -- abstracting wrt *all* the tyvars. We'll see if that
728 -- gives rise to problems. SLPJ June 98
730 go env fn (Let (Rec prs) body)
731 = mapAndUnzipSmpl (mk_poly tyvars_here) vars `thenSmpl` \ (vars', rhss') ->
733 gn body = fn (foldr Let body (zipWith mk_silly_bind vars rhss'))
734 pairs = vars' `zip` [mkLams tyvars_here (gn rhs) | rhs <- rhss]
736 addAuxiliaryBind env (Rec pairs) $ \ env ->
739 (vars,rhss) = unzip prs
740 tyvars_here = varSetElems (main_tyvar_set `intersectVarSet` exprsSomeFreeVars isTyVar (map snd prs))
741 -- See notes with tyvars_here above
743 go env fn body = returnSmpl (emptyFloats env, fn body)
745 mk_poly tyvars_here var
746 = getUniqueSmpl `thenSmpl` \ uniq ->
748 poly_name = setNameUnique (idName var) uniq -- Keep same name
749 poly_ty = mkForAllTys tyvars_here (idType var) -- But new type of course
750 poly_id = mkLocalId poly_name poly_ty
752 -- In the olden days, it was crucial to copy the occInfo of the original var,
753 -- because we were looking at occurrence-analysed but as yet unsimplified code!
754 -- In particular, we mustn't lose the loop breakers. BUT NOW we are looking
755 -- at already simplified code, so it doesn't matter
757 -- It's even right to retain single-occurrence or dead-var info:
758 -- Suppose we started with /\a -> let x = E in B
759 -- where x occurs once in B. Then we transform to:
760 -- let x' = /\a -> E in /\a -> let x* = x' a in B
761 -- where x* has an INLINE prag on it. Now, once x* is inlined,
762 -- the occurrences of x' will be just the occurrences originally
765 returnSmpl (poly_id, mkTyApps (Var poly_id) (mkTyVarTys tyvars_here))
767 mk_silly_bind var rhs = NonRec var (Note InlineMe rhs)
768 -- Suppose we start with:
770 -- x = /\ a -> let g = G in E
772 -- Then we'll float to get
774 -- x = let poly_g = /\ a -> G
775 -- in /\ a -> let g = poly_g a in E
777 -- But now the occurrence analyser will see just one occurrence
778 -- of poly_g, not inside a lambda, so the simplifier will
779 -- PreInlineUnconditionally poly_g back into g! Badk to square 1!
780 -- (I used to think that the "don't inline lone occurrences" stuff
781 -- would stop this happening, but since it's the *only* occurrence,
782 -- PreInlineUnconditionally kicks in first!)
784 -- Solution: put an INLINE note on g's RHS, so that poly_g seems
785 -- to appear many times. (NB: mkInlineMe eliminates
786 -- such notes on trivial RHSs, so do it manually.)
790 %************************************************************************
792 \subsection{Case alternative filtering
794 %************************************************************************
796 prepareAlts does two things:
798 1. Eliminate alternatives that cannot match, including the
801 2. If the DEFAULT alternative can match only one possible constructor,
802 then make that constructor explicit.
804 case e of x { DEFAULT -> rhs }
806 case e of x { (a,b) -> rhs }
807 where the type is a single constructor type. This gives better code
808 when rhs also scrutinises x or e.
810 It's a good idea do do this stuff before simplifying the alternatives, to
811 avoid simplifying alternatives we know can't happen, and to come up with
812 the list of constructors that are handled, to put into the IdInfo of the
813 case binder, for use when simplifying the alternatives.
815 Eliminating the default alternative in (1) isn't so obvious, but it can
818 data Colour = Red | Green | Blue
827 DEFAULT -> [ case y of ... ]
829 If we inline h into f, the default case of the inlined h can't happen.
830 If we don't notice this, we may end up filtering out *all* the cases
831 of the inner case y, which give us nowhere to go!
835 prepareAlts :: OutExpr -- Scrutinee
836 -> InId -- Case binder
838 -> SimplM ([InAlt], -- Better alternatives
839 [AltCon]) -- These cases are handled
841 prepareAlts scrut case_bndr alts
843 (alts_wo_default, maybe_deflt) = findDefault alts
845 impossible_cons = case scrut of
846 Var v -> otherCons (idUnfolding v)
849 -- Filter out alternatives that can't possibly match
850 better_alts | null impossible_cons = alts_wo_default
851 | otherwise = [alt | alt@(con,_,_) <- alts_wo_default,
852 not (con `elem` impossible_cons)]
854 -- "handled_cons" are handled either by the context,
855 -- or by a branch in this case expression
856 -- (Don't add DEFAULT to the handled_cons!!)
857 handled_cons = impossible_cons ++ [con | (con,_,_) <- better_alts]
859 -- Filter out the default, if it can't happen,
860 -- or replace it with "proper" alternative if there
861 -- is only one constructor left
862 prepareDefault case_bndr handled_cons maybe_deflt `thenSmpl` \ deflt_alt ->
864 returnSmpl (deflt_alt ++ better_alts, handled_cons)
866 prepareDefault case_bndr handled_cons (Just rhs)
867 | Just (tycon, inst_tys) <- splitTyConApp_maybe (idType case_bndr),
868 isAlgTyCon tycon, -- It's a data type, tuple, or unboxed tuples.
869 not (isNewTyCon tycon), -- We can have a newtype, if we are just doing an eval:
870 -- case x of { DEFAULT -> e }
871 -- and we don't want to fill in a default for them!
872 Just all_cons <- tyConDataCons_maybe tycon,
873 not (null all_cons), -- This is a tricky corner case. If the data type has no constructors,
874 -- which GHC allows, then the case expression will have at most a default
875 -- alternative. We don't want to eliminate that alternative, because the
876 -- invariant is that there's always one alternative. It's more convenient
878 -- case x of { DEFAULT -> e }
879 -- as it is, rather than transform it to
880 -- error "case cant match"
881 -- which would be quite legitmate. But it's a really obscure corner, and
882 -- not worth wasting code on.
883 let handled_data_cons = [data_con | DataAlt data_con <- handled_cons],
884 let missing_cons = [con | con <- all_cons,
885 not (con `elem` handled_data_cons)]
886 = case missing_cons of
887 [] -> returnSmpl [] -- Eliminate the default alternative
890 [con] -> -- It matches exactly one constructor, so fill it in
891 tick (FillInCaseDefault case_bndr) `thenSmpl_`
892 mk_args con inst_tys `thenSmpl` \ args ->
893 returnSmpl [(DataAlt con, args, rhs)]
895 two_or_more -> returnSmpl [(DEFAULT, [], rhs)]
898 = returnSmpl [(DEFAULT, [], rhs)]
900 prepareDefault case_bndr handled_cons Nothing
903 mk_args missing_con inst_tys
904 = mk_tv_bndrs missing_con inst_tys `thenSmpl` \ (tv_bndrs, inst_tys') ->
905 getUniquesSmpl `thenSmpl` \ id_uniqs ->
906 let arg_tys = dataConArgTys missing_con inst_tys'
907 arg_ids = zipWith (mkSysLocal FSLIT("a")) id_uniqs arg_tys
909 returnSmpl (tv_bndrs ++ arg_ids)
911 mk_tv_bndrs missing_con inst_tys
912 | isVanillaDataCon missing_con
913 = returnSmpl ([], inst_tys)
915 = getUniquesSmpl `thenSmpl` \ tv_uniqs ->
916 let new_tvs = zipWith mk tv_uniqs (dataConTyVars missing_con)
917 mk uniq tv = mkTyVar (mkSysTvName uniq FSLIT("t")) (tyVarKind tv)
919 returnSmpl (new_tvs, mkTyVarTys new_tvs)
923 %************************************************************************
925 \subsection{Case absorption and identity-case elimination}
927 %************************************************************************
929 mkCase puts a case expression back together, trying various transformations first.
932 mkCase :: OutExpr -> OutId -> OutType -> [OutAlt] -> SimplM OutExpr
934 mkCase scrut case_bndr ty alts
935 = mkAlts scrut case_bndr alts `thenSmpl` \ better_alts ->
936 mkCase1 scrut case_bndr ty better_alts
940 mkAlts tries these things:
942 1. If several alternatives are identical, merge them into
943 a single DEFAULT alternative. I've occasionally seen this
944 making a big difference:
946 case e of =====> case e of
947 C _ -> f x D v -> ....v....
948 D v -> ....v.... DEFAULT -> f x
951 The point is that we merge common RHSs, at least for the DEFAULT case.
952 [One could do something more elaborate but I've never seen it needed.]
953 To avoid an expensive test, we just merge branches equal to the *first*
954 alternative; this picks up the common cases
955 a) all branches equal
956 b) some branches equal to the DEFAULT (which occurs first)
959 case e of b { ==> case e of b {
960 p1 -> rhs1 p1 -> rhs1
962 pm -> rhsm pm -> rhsm
963 _ -> case b of b' { pn -> let b'=b in rhsn
965 ... po -> let b'=b in rhso
966 po -> rhso _ -> let b'=b in rhsd
970 which merges two cases in one case when -- the default alternative of
971 the outer case scrutises the same variable as the outer case This
972 transformation is called Case Merging. It avoids that the same
973 variable is scrutinised multiple times.
976 The case where transformation (1) showed up was like this (lib/std/PrelCError.lhs):
982 where @is@ was something like
984 p `is` n = p /= (-1) && p == n
986 This gave rise to a horrible sequence of cases
993 and similarly in cascade for all the join points!
998 --------------------------------------------------
999 -- 1. Merge identical branches
1000 --------------------------------------------------
1001 mkAlts scrut case_bndr alts@((con1,bndrs1,rhs1) : con_alts)
1002 | all isDeadBinder bndrs1, -- Remember the default
1003 length filtered_alts < length con_alts -- alternative comes first
1004 = tick (AltMerge case_bndr) `thenSmpl_`
1005 returnSmpl better_alts
1007 filtered_alts = filter keep con_alts
1008 keep (con,bndrs,rhs) = not (all isDeadBinder bndrs && rhs `cheapEqExpr` rhs1)
1009 better_alts = (DEFAULT, [], rhs1) : filtered_alts
1012 --------------------------------------------------
1013 -- 2. Merge nested cases
1014 --------------------------------------------------
1016 mkAlts scrut outer_bndr outer_alts
1017 = getDOptsSmpl `thenSmpl` \dflags ->
1018 mkAlts' dflags scrut outer_bndr outer_alts
1020 mkAlts' dflags scrut outer_bndr outer_alts
1021 | dopt Opt_CaseMerge dflags,
1022 (outer_alts_without_deflt, maybe_outer_deflt) <- findDefault outer_alts,
1024 Just (Case (Var scrut_var) inner_bndr _ inner_alts) <- maybe_outer_deflt,
1025 scruting_same_var scrut_var
1027 = let -- Eliminate any inner alts which are shadowed by the outer ones
1028 outer_cons = [con | (con,_,_) <- outer_alts_without_deflt]
1030 munged_inner_alts = [ (con, args, munge_rhs rhs)
1031 | (con, args, rhs) <- inner_alts,
1032 not (con `elem` outer_cons) -- Eliminate shadowed inner alts
1034 munge_rhs rhs = bindCaseBndr inner_bndr (Var outer_bndr) rhs
1036 (inner_con_alts, maybe_inner_default) = findDefault munged_inner_alts
1038 new_alts = add_default maybe_inner_default
1039 (outer_alts_without_deflt ++ inner_con_alts)
1041 tick (CaseMerge outer_bndr) `thenSmpl_`
1043 -- Warning: don't call mkAlts recursively!
1044 -- Firstly, there's no point, because inner alts have already had
1045 -- mkCase applied to them, so they won't have a case in their default
1046 -- Secondly, if you do, you get an infinite loop, because the bindCaseBndr
1047 -- in munge_rhs may put a case into the DEFAULT branch!
1049 -- We are scrutinising the same variable if it's
1050 -- the outer case-binder, or if the outer case scrutinises a variable
1051 -- (and it's the same). Testing both allows us not to replace the
1052 -- outer scrut-var with the outer case-binder (Simplify.simplCaseBinder).
1053 scruting_same_var = case scrut of
1054 Var outer_scrut -> \ v -> v == outer_bndr || v == outer_scrut
1055 other -> \ v -> v == outer_bndr
1057 add_default (Just rhs) alts = (DEFAULT,[],rhs) : alts
1058 add_default Nothing alts = alts
1061 --------------------------------------------------
1063 --------------------------------------------------
1065 mkAlts' dflags scrut case_bndr other_alts = returnSmpl other_alts
1070 =================================================================================
1072 mkCase1 tries these things
1074 1. Eliminate the case altogether if possible
1082 and similar friends.
1085 Start with a simple situation:
1087 case x# of ===> e[x#/y#]
1090 (when x#, y# are of primitive type, of course). We can't (in general)
1091 do this for algebraic cases, because we might turn bottom into
1094 Actually, we generalise this idea to look for a case where we're
1095 scrutinising a variable, and we know that only the default case can
1100 other -> ...(case x of
1104 Here the inner case can be eliminated. This really only shows up in
1105 eliminating error-checking code.
1107 We also make sure that we deal with this very common case:
1112 Here we are using the case as a strict let; if x is used only once
1113 then we want to inline it. We have to be careful that this doesn't
1114 make the program terminate when it would have diverged before, so we
1116 - x is used strictly, or
1117 - e is already evaluated (it may so if e is a variable)
1119 Lastly, we generalise the transformation to handle this:
1125 We only do this for very cheaply compared r's (constructors, literals
1126 and variables). If pedantic bottoms is on, we only do it when the
1127 scrutinee is a PrimOp which can't fail.
1129 We do it *here*, looking at un-simplified alternatives, because we
1130 have to check that r doesn't mention the variables bound by the
1131 pattern in each alternative, so the binder-info is rather useful.
1133 So the case-elimination algorithm is:
1135 1. Eliminate alternatives which can't match
1137 2. Check whether all the remaining alternatives
1138 (a) do not mention in their rhs any of the variables bound in their pattern
1139 and (b) have equal rhss
1141 3. Check we can safely ditch the case:
1142 * PedanticBottoms is off,
1143 or * the scrutinee is an already-evaluated variable
1144 or * the scrutinee is a primop which is ok for speculation
1145 -- ie we want to preserve divide-by-zero errors, and
1146 -- calls to error itself!
1148 or * [Prim cases] the scrutinee is a primitive variable
1150 or * [Alg cases] the scrutinee is a variable and
1151 either * the rhs is the same variable
1152 (eg case x of C a b -> x ===> x)
1153 or * there is only one alternative, the default alternative,
1154 and the binder is used strictly in its scope.
1155 [NB this is helped by the "use default binder where
1156 possible" transformation; see below.]
1159 If so, then we can replace the case with one of the rhss.
1161 Further notes about case elimination
1162 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
1163 Consider: test :: Integer -> IO ()
1166 Turns out that this compiles to:
1169 eta1 :: State# RealWorld ->
1170 case PrelNum.< eta PrelNum.zeroInteger of wild { __DEFAULT ->
1172 (PrelNum.jtos eta ($w[] @ Char))
1174 of wild1 { (# new_s, a4 #) -> PrelIO.lvl23 new_s }}
1176 Notice the strange '<' which has no effect at all. This is a funny one.
1177 It started like this:
1179 f x y = if x < 0 then jtos x
1180 else if y==0 then "" else jtos x
1182 At a particular call site we have (f v 1). So we inline to get
1184 if v < 0 then jtos x
1185 else if 1==0 then "" else jtos x
1187 Now simplify the 1==0 conditional:
1189 if v<0 then jtos v else jtos v
1191 Now common-up the two branches of the case:
1193 case (v<0) of DEFAULT -> jtos v
1195 Why don't we drop the case? Because it's strict in v. It's technically
1196 wrong to drop even unnecessary evaluations, and in practice they
1197 may be a result of 'seq' so we *definitely* don't want to drop those.
1198 I don't really know how to improve this situation.
1202 --------------------------------------------------
1203 -- 0. Check for empty alternatives
1204 --------------------------------------------------
1207 mkCase1 scrut case_bndr ty []
1208 = pprTrace "mkCase1: null alts" (ppr case_bndr <+> ppr scrut) $
1212 --------------------------------------------------
1213 -- 1. Eliminate the case altogether if poss
1214 --------------------------------------------------
1216 mkCase1 scrut case_bndr ty [(con,bndrs,rhs)]
1217 -- See if we can get rid of the case altogether
1218 -- See the extensive notes on case-elimination above
1219 -- mkCase made sure that if all the alternatives are equal,
1220 -- then there is now only one (DEFAULT) rhs
1221 | all isDeadBinder bndrs,
1223 -- Check that the scrutinee can be let-bound instead of case-bound
1224 exprOkForSpeculation scrut
1225 -- OK not to evaluate it
1226 -- This includes things like (==# a# b#)::Bool
1227 -- so that we simplify
1228 -- case ==# a# b# of { True -> x; False -> x }
1231 -- This particular example shows up in default methods for
1232 -- comparision operations (e.g. in (>=) for Int.Int32)
1233 || exprIsValue scrut -- It's already evaluated
1234 || var_demanded_later scrut -- It'll be demanded later
1236 -- || not opt_SimplPedanticBottoms) -- Or we don't care!
1237 -- We used to allow improving termination by discarding cases, unless -fpedantic-bottoms was on,
1238 -- but that breaks badly for the dataToTag# primop, which relies on a case to evaluate
1239 -- its argument: case x of { y -> dataToTag# y }
1240 -- Here we must *not* discard the case, because dataToTag# just fetches the tag from
1241 -- the info pointer. So we'll be pedantic all the time, and see if that gives any
1243 -- Also we don't want to discard 'seq's
1244 = tick (CaseElim case_bndr) `thenSmpl_`
1245 returnSmpl (bindCaseBndr case_bndr scrut rhs)
1248 -- The case binder is going to be evaluated later,
1249 -- and the scrutinee is a simple variable
1250 var_demanded_later (Var v) = isStrictDmd (idNewDemandInfo case_bndr)
1251 var_demanded_later other = False
1254 --------------------------------------------------
1256 --------------------------------------------------
1258 mkCase1 scrut case_bndr ty alts -- Identity case
1259 | all identity_alt alts
1260 = tick (CaseIdentity case_bndr) `thenSmpl_`
1261 returnSmpl (re_note scrut)
1263 identity_alt (con, args, rhs) = de_note rhs `cheapEqExpr` identity_rhs con args
1265 identity_rhs (DataAlt con) args = mkConApp con (arg_tys ++ map varToCoreExpr args)
1266 identity_rhs (LitAlt lit) _ = Lit lit
1267 identity_rhs DEFAULT _ = Var case_bndr
1269 arg_tys = map Type (tyConAppArgs (idType case_bndr))
1272 -- case coerce T e of x { _ -> coerce T' x }
1273 -- And we definitely want to eliminate this case!
1274 -- So we throw away notes from the RHS, and reconstruct
1275 -- (at least an approximation) at the other end
1276 de_note (Note _ e) = de_note e
1279 -- re_note wraps a coerce if it might be necessary
1280 re_note scrut = case head alts of
1281 (_,_,rhs1@(Note _ _)) -> mkCoerce2 (exprType rhs1) (idType case_bndr) scrut
1285 --------------------------------------------------
1287 --------------------------------------------------
1289 mkCase1 scrut bndr ty alts = returnSmpl (Case scrut bndr ty alts)
1293 When adding auxiliary bindings for the case binder, it's worth checking if
1294 its dead, because it often is, and occasionally these mkCase transformations
1295 cascade rather nicely.
1298 bindCaseBndr bndr rhs body
1299 | isDeadBinder bndr = body
1300 | otherwise = bindNonRec bndr rhs body