2 % (c) The AQUA Project, Glasgow University, 1993-1998
4 \section[SimplUtils]{The simplifier utilities}
11 preInlineUnconditionally, postInlineUnconditionally, activeInline, activeRule,
14 -- The continuation type
15 SimplCont(..), DupFlag(..), LetRhsFlag(..),
16 contIsDupable, contResultType,
17 countValArgs, countArgs, pushContArgs,
18 mkBoringStop, mkRhsStop, contIsRhs, contIsRhsOrArg,
19 getContArgs, interestingCallContext, interestingArg, isStrictType
23 #include "HsVersions.h"
26 import DynFlags ( SimplifierSwitch(..), SimplifierMode(..),
28 import StaticFlags ( opt_UF_UpdateInPlace, opt_SimplNoPreInlining,
31 import CoreFVs ( exprFreeVars )
32 import CoreUtils ( cheapEqExpr, exprType, exprIsTrivial, exprIsCheap,
33 etaExpand, exprEtaExpandArity, bindNonRec, mkCoerce2,
34 findDefault, exprOkForSpeculation, exprIsHNF, mergeAlts
36 import Literal ( mkStringLit )
37 import CoreUnfold ( smallEnoughToInline )
38 import MkId ( eRROR_ID )
39 import Id ( idType, isDataConWorkId, idOccInfo, isDictId,
40 isDeadBinder, idNewDemandInfo, isExportedId,
41 idUnfolding, idNewStrictness, idInlinePragma,
43 import NewDemand ( isStrictDmd, isBotRes, splitStrictSig )
45 import Type ( Type, splitFunTys, dropForAlls, isStrictType,
46 splitTyConApp_maybe, tyConAppArgs
48 import TyCon ( tyConDataCons_maybe )
49 import DataCon ( dataConRepArity )
51 import BasicTypes ( TopLevelFlag(..), isNotTopLevel, OccInfo(..), isLoopBreaker, isOneOcc,
52 Activation, isAlwaysActive, isActive )
53 import Util ( lengthExceeds )
58 %************************************************************************
60 \subsection{The continuation data type}
62 %************************************************************************
65 data SimplCont -- Strict contexts
66 = Stop OutType -- Type of the result
68 Bool -- True <=> This is the RHS of a thunk whose type suggests
69 -- that update-in-place would be possible
70 -- (This makes the inliner a little keener.)
72 | CoerceIt OutType -- The To-type, simplified
75 | InlinePlease -- This continuation makes a function very
76 SimplCont -- keen to inline itelf
79 InExpr SimplEnv -- The argument, as yet unsimplified,
80 SimplCont -- and its environment
83 InId [InAlt] SimplEnv -- The case binder, alts, and subst-env
86 | ArgOf LetRhsFlag -- An arbitrary strict context: the argument
87 -- of a strict function, or a primitive-arg fn
89 -- No DupFlag because we never duplicate it
90 OutType -- arg_ty: type of the argument itself
91 OutType -- cont_ty: the type of the expression being sought by the context
92 -- f (error "foo") ==> coerce t (error "foo")
94 -- We need to know the type t, to which to coerce.
96 (SimplEnv -> OutExpr -> SimplM FloatsWithExpr) -- What to do with the result
97 -- The result expression in the OutExprStuff has type cont_ty
99 data LetRhsFlag = AnArg -- It's just an argument not a let RHS
100 | AnRhs -- It's the RHS of a let (so please float lets out of big lambdas)
102 instance Outputable LetRhsFlag where
103 ppr AnArg = ptext SLIT("arg")
104 ppr AnRhs = ptext SLIT("rhs")
106 instance Outputable SimplCont where
107 ppr (Stop ty is_rhs _) = ptext SLIT("Stop") <> brackets (ppr is_rhs) <+> ppr ty
108 ppr (ApplyTo dup arg se cont) = (ptext SLIT("ApplyTo") <+> ppr dup <+> ppr arg) $$ ppr cont
109 ppr (ArgOf _ _ _ _) = ptext SLIT("ArgOf...")
110 ppr (Select dup bndr alts se cont) = (ptext SLIT("Select") <+> ppr dup <+> ppr bndr) $$
111 (nest 4 (ppr alts)) $$ ppr cont
112 ppr (CoerceIt ty cont) = (ptext SLIT("CoerceIt") <+> ppr ty) $$ ppr cont
113 ppr (InlinePlease cont) = ptext SLIT("InlinePlease") $$ ppr cont
115 data DupFlag = OkToDup | NoDup
117 instance Outputable DupFlag where
118 ppr OkToDup = ptext SLIT("ok")
119 ppr NoDup = ptext SLIT("nodup")
123 mkBoringStop, mkRhsStop :: OutType -> SimplCont
124 mkBoringStop ty = Stop ty AnArg (canUpdateInPlace ty)
125 mkRhsStop ty = Stop ty AnRhs (canUpdateInPlace ty)
127 contIsRhs :: SimplCont -> Bool
128 contIsRhs (Stop _ AnRhs _) = True
129 contIsRhs (ArgOf AnRhs _ _ _) = True
130 contIsRhs other = False
132 contIsRhsOrArg (Stop _ _ _) = True
133 contIsRhsOrArg (ArgOf _ _ _ _) = True
134 contIsRhsOrArg other = False
137 contIsDupable :: SimplCont -> Bool
138 contIsDupable (Stop _ _ _) = True
139 contIsDupable (ApplyTo OkToDup _ _ _) = True
140 contIsDupable (Select OkToDup _ _ _ _) = True
141 contIsDupable (CoerceIt _ cont) = contIsDupable cont
142 contIsDupable (InlinePlease cont) = contIsDupable cont
143 contIsDupable other = False
146 discardableCont :: SimplCont -> Bool
147 discardableCont (Stop _ _ _) = False
148 discardableCont (CoerceIt _ cont) = discardableCont cont
149 discardableCont (InlinePlease cont) = discardableCont cont
150 discardableCont other = True
152 discardCont :: SimplCont -- A continuation, expecting
153 -> SimplCont -- Replace the continuation with a suitable coerce
154 discardCont cont = case cont of
155 Stop to_ty is_rhs _ -> cont
156 other -> CoerceIt to_ty (mkBoringStop to_ty)
158 to_ty = contResultType cont
161 contResultType :: SimplCont -> OutType
162 contResultType (Stop to_ty _ _) = to_ty
163 contResultType (ArgOf _ _ to_ty _) = to_ty
164 contResultType (ApplyTo _ _ _ cont) = contResultType cont
165 contResultType (CoerceIt _ cont) = contResultType cont
166 contResultType (InlinePlease cont) = contResultType cont
167 contResultType (Select _ _ _ _ cont) = contResultType cont
170 countValArgs :: SimplCont -> Int
171 countValArgs (ApplyTo _ (Type ty) se cont) = countValArgs cont
172 countValArgs (ApplyTo _ val_arg se cont) = 1 + countValArgs cont
173 countValArgs other = 0
175 countArgs :: SimplCont -> Int
176 countArgs (ApplyTo _ arg se cont) = 1 + countArgs cont
180 pushContArgs :: SimplEnv -> [OutArg] -> SimplCont -> SimplCont
181 -- Pushes args with the specified environment
182 pushContArgs env [] cont = cont
183 pushContArgs env (arg : args) cont = ApplyTo NoDup arg env (pushContArgs env args cont)
188 getContArgs :: SwitchChecker
189 -> OutId -> SimplCont
190 -> ([(InExpr, SimplEnv, Bool)], -- Arguments; the Bool is true for strict args
191 SimplCont, -- Remaining continuation
192 Bool) -- Whether we came across an InlineCall
193 -- getContArgs id k = (args, k', inl)
194 -- args are the leading ApplyTo items in k
195 -- (i.e. outermost comes first)
196 -- augmented with demand info from the functionn
197 getContArgs chkr fun orig_cont
199 -- Ignore strictness info if the no-case-of-case
200 -- flag is on. Strictness changes evaluation order
201 -- and that can change full laziness
202 stricts | switchIsOn chkr NoCaseOfCase = vanilla_stricts
203 | otherwise = computed_stricts
205 go [] stricts False orig_cont
207 ----------------------------
210 go acc ss inl (ApplyTo _ arg@(Type _) se cont)
211 = go ((arg,se,False) : acc) ss inl cont
212 -- NB: don't bother to instantiate the function type
215 go acc (s:ss) inl (ApplyTo _ arg se cont)
216 = go ((arg,se,s) : acc) ss inl cont
218 -- An Inline continuation
219 go acc ss inl (InlinePlease cont)
220 = go acc ss True cont
222 -- We're run out of arguments, or else we've run out of demands
223 -- The latter only happens if the result is guaranteed bottom
224 -- This is the case for
225 -- * case (error "hello") of { ... }
226 -- * (error "Hello") arg
227 -- * f (error "Hello") where f is strict
229 -- Then, especially in the first of these cases, we'd like to discard
230 -- the continuation, leaving just the bottoming expression. But the
231 -- type might not be right, so we may have to add a coerce.
233 | null ss && discardableCont cont = (reverse acc, discardCont cont, inl)
234 | otherwise = (reverse acc, cont, inl)
236 ----------------------------
237 vanilla_stricts, computed_stricts :: [Bool]
238 vanilla_stricts = repeat False
239 computed_stricts = zipWith (||) fun_stricts arg_stricts
241 ----------------------------
242 (val_arg_tys, _) = splitFunTys (dropForAlls (idType fun))
243 arg_stricts = map isStrictType val_arg_tys ++ repeat False
244 -- These argument types are used as a cheap and cheerful way to find
245 -- unboxed arguments, which must be strict. But it's an InType
246 -- and so there might be a type variable where we expect a function
247 -- type (the substitution hasn't happened yet). And we don't bother
248 -- doing the type applications for a polymorphic function.
249 -- Hence the splitFunTys*IgnoringForAlls*
251 ----------------------------
252 -- If fun_stricts is finite, it means the function returns bottom
253 -- after that number of value args have been consumed
254 -- Otherwise it's infinite, extended with False
256 = case splitStrictSig (idNewStrictness fun) of
257 (demands, result_info)
258 | not (demands `lengthExceeds` countValArgs orig_cont)
259 -> -- Enough args, use the strictness given.
260 -- For bottoming functions we used to pretend that the arg
261 -- is lazy, so that we don't treat the arg as an
262 -- interesting context. This avoids substituting
263 -- top-level bindings for (say) strings into
264 -- calls to error. But now we are more careful about
265 -- inlining lone variables, so its ok (see SimplUtils.analyseCont)
266 if isBotRes result_info then
267 map isStrictDmd demands -- Finite => result is bottom
269 map isStrictDmd demands ++ vanilla_stricts
271 other -> vanilla_stricts -- Not enough args, or no strictness
274 interestingArg :: OutExpr -> Bool
275 -- An argument is interesting if it has *some* structure
276 -- We are here trying to avoid unfolding a function that
277 -- is applied only to variables that have no unfolding
278 -- (i.e. they are probably lambda bound): f x y z
279 -- There is little point in inlining f here.
280 interestingArg (Var v) = hasSomeUnfolding (idUnfolding v)
281 -- Was: isValueUnfolding (idUnfolding v')
282 -- But that seems over-pessimistic
284 -- This accounts for an argument like
285 -- () or [], which is definitely interesting
286 interestingArg (Type _) = False
287 interestingArg (App fn (Type _)) = interestingArg fn
288 interestingArg (Note _ a) = interestingArg a
289 interestingArg other = True
290 -- Consider let x = 3 in f x
291 -- The substitution will contain (x -> ContEx 3), and we want to
292 -- to say that x is an interesting argument.
293 -- But consider also (\x. f x y) y
294 -- The substitution will contain (x -> ContEx y), and we want to say
295 -- that x is not interesting (assuming y has no unfolding)
298 Comment about interestingCallContext
299 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
300 We want to avoid inlining an expression where there can't possibly be
301 any gain, such as in an argument position. Hence, if the continuation
302 is interesting (eg. a case scrutinee, application etc.) then we
303 inline, otherwise we don't.
305 Previously some_benefit used to return True only if the variable was
306 applied to some value arguments. This didn't work:
308 let x = _coerce_ (T Int) Int (I# 3) in
309 case _coerce_ Int (T Int) x of
312 we want to inline x, but can't see that it's a constructor in a case
313 scrutinee position, and some_benefit is False.
317 dMonadST = _/\_ t -> :Monad (g1 _@_ t, g2 _@_ t, g3 _@_ t)
319 .... case dMonadST _@_ x0 of (a,b,c) -> ....
321 we'd really like to inline dMonadST here, but we *don't* want to
322 inline if the case expression is just
324 case x of y { DEFAULT -> ... }
326 since we can just eliminate this case instead (x is in WHNF). Similar
327 applies when x is bound to a lambda expression. Hence
328 contIsInteresting looks for case expressions with just a single
332 interestingCallContext :: Bool -- False <=> no args at all
333 -> Bool -- False <=> no value args
335 -- The "lone-variable" case is important. I spent ages
336 -- messing about with unsatisfactory varaints, but this is nice.
337 -- The idea is that if a variable appear all alone
338 -- as an arg of lazy fn, or rhs Stop
339 -- as scrutinee of a case Select
340 -- as arg of a strict fn ArgOf
341 -- then we should not inline it (unless there is some other reason,
342 -- e.g. is is the sole occurrence). We achieve this by making
343 -- interestingCallContext return False for a lone variable.
345 -- Why? At least in the case-scrutinee situation, turning
346 -- let x = (a,b) in case x of y -> ...
348 -- let x = (a,b) in case (a,b) of y -> ...
350 -- let x = (a,b) in let y = (a,b) in ...
351 -- is bad if the binding for x will remain.
353 -- Another example: I discovered that strings
354 -- were getting inlined straight back into applications of 'error'
355 -- because the latter is strict.
357 -- f = \x -> ...(error s)...
359 -- Fundamentally such contexts should not ecourage inlining because
360 -- the context can ``see'' the unfolding of the variable (e.g. case or a RULE)
361 -- so there's no gain.
363 -- However, even a type application or coercion isn't a lone variable.
365 -- case $fMonadST @ RealWorld of { :DMonad a b c -> c }
366 -- We had better inline that sucker! The case won't see through it.
368 -- For now, I'm treating treating a variable applied to types
369 -- in a *lazy* context "lone". The motivating example was
371 -- g = /\a. \y. h (f a)
372 -- There's no advantage in inlining f here, and perhaps
373 -- a significant disadvantage. Hence some_val_args in the Stop case
375 interestingCallContext some_args some_val_args cont
378 interesting (InlinePlease _) = True
379 interesting (Select _ _ _ _ _) = some_args
380 interesting (ApplyTo _ _ _ _) = True -- Can happen if we have (coerce t (f x)) y
381 -- Perhaps True is a bit over-keen, but I've
382 -- seen (coerce f) x, where f has an INLINE prag,
383 -- So we have to give some motivaiton for inlining it
384 interesting (ArgOf _ _ _ _) = some_val_args
385 interesting (Stop ty _ upd_in_place) = some_val_args && upd_in_place
386 interesting (CoerceIt _ cont) = interesting cont
387 -- If this call is the arg of a strict function, the context
388 -- is a bit interesting. If we inline here, we may get useful
389 -- evaluation information to avoid repeated evals: e.g.
391 -- Here the contIsInteresting makes the '*' keener to inline,
392 -- which in turn exposes a constructor which makes the '+' inline.
393 -- Assuming that +,* aren't small enough to inline regardless.
395 -- It's also very important to inline in a strict context for things
398 -- Here, the context of (f x) is strict, and if f's unfolding is
399 -- a build it's *great* to inline it here. So we must ensure that
400 -- the context for (f x) is not totally uninteresting.
404 canUpdateInPlace :: Type -> Bool
405 -- Consider let x = <wurble> in ...
406 -- If <wurble> returns an explicit constructor, we might be able
407 -- to do update in place. So we treat even a thunk RHS context
408 -- as interesting if update in place is possible. We approximate
409 -- this by seeing if the type has a single constructor with a
410 -- small arity. But arity zero isn't good -- we share the single copy
411 -- for that case, so no point in sharing.
414 | not opt_UF_UpdateInPlace = False
416 = case splitTyConApp_maybe ty of
418 Just (tycon, _) -> case tyConDataCons_maybe tycon of
419 Just [dc] -> arity == 1 || arity == 2
421 arity = dataConRepArity dc
427 %************************************************************************
429 \subsection{Decisions about inlining}
431 %************************************************************************
433 Inlining is controlled partly by the SimplifierMode switch. This has two
436 SimplGently (a) Simplifying before specialiser/full laziness
437 (b) Simplifiying inside INLINE pragma
438 (c) Simplifying the LHS of a rule
439 (d) Simplifying a GHCi expression or Template
442 SimplPhase n Used at all other times
444 The key thing about SimplGently is that it does no call-site inlining.
445 Before full laziness we must be careful not to inline wrappers,
446 because doing so inhibits floating
447 e.g. ...(case f x of ...)...
448 ==> ...(case (case x of I# x# -> fw x#) of ...)...
449 ==> ...(case x of I# x# -> case fw x# of ...)...
450 and now the redex (f x) isn't floatable any more.
452 The no-inling thing is also important for Template Haskell. You might be
453 compiling in one-shot mode with -O2; but when TH compiles a splice before
454 running it, we don't want to use -O2. Indeed, we don't want to inline
455 anything, because the byte-code interpreter might get confused about
456 unboxed tuples and suchlike.
460 SimplGently is also used as the mode to simplify inside an InlineMe note.
463 inlineMode :: SimplifierMode
464 inlineMode = SimplGently
467 It really is important to switch off inlinings inside such
468 expressions. Consider the following example
474 in ...g...g...g...g...g...
476 Now, if that's the ONLY occurrence of f, it will be inlined inside g,
477 and thence copied multiple times when g is inlined.
480 This function may be inlinined in other modules, so we
481 don't want to remove (by inlining) calls to functions that have
482 specialisations, or that may have transformation rules in an importing
485 E.g. {-# INLINE f #-}
488 and suppose that g is strict *and* has specialisations. If we inline
489 g's wrapper, we deny f the chance of getting the specialised version
490 of g when f is inlined at some call site (perhaps in some other
493 It's also important not to inline a worker back into a wrapper.
495 wraper = inline_me (\x -> ...worker... )
496 Normally, the inline_me prevents the worker getting inlined into
497 the wrapper (initially, the worker's only call site!). But,
498 if the wrapper is sure to be called, the strictness analyser will
499 mark it 'demanded', so when the RHS is simplified, it'll get an ArgOf
500 continuation. That's why the keep_inline predicate returns True for
501 ArgOf continuations. It shouldn't do any harm not to dissolve the
502 inline-me note under these circumstances.
504 Note that the result is that we do very little simplification
507 all xs = foldr (&&) True xs
508 any p = all . map p {-# INLINE any #-}
510 Problem: any won't get deforested, and so if it's exported and the
511 importer doesn't use the inlining, (eg passes it as an arg) then we
512 won't get deforestation at all. We havn't solved this problem yet!
515 preInlineUnconditionally
516 ~~~~~~~~~~~~~~~~~~~~~~~~
517 @preInlineUnconditionally@ examines a bndr to see if it is used just
518 once in a completely safe way, so that it is safe to discard the
519 binding inline its RHS at the (unique) usage site, REGARDLESS of how
520 big the RHS might be. If this is the case we don't simplify the RHS
521 first, but just inline it un-simplified.
523 This is much better than first simplifying a perhaps-huge RHS and then
524 inlining and re-simplifying it. Indeed, it can be at least quadratically
533 We may end up simplifying e1 N times, e2 N-1 times, e3 N-3 times etc.
534 This can happen with cascades of functions too:
541 THE MAIN INVARIANT is this:
543 ---- preInlineUnconditionally invariant -----
544 IF preInlineUnconditionally chooses to inline x = <rhs>
545 THEN doing the inlining should not change the occurrence
546 info for the free vars of <rhs>
547 ----------------------------------------------
549 For example, it's tempting to look at trivial binding like
551 and inline it unconditionally. But suppose x is used many times,
552 but this is the unique occurrence of y. Then inlining x would change
553 y's occurrence info, which breaks the invariant. It matters: y
554 might have a BIG rhs, which will now be dup'd at every occurrenc of x.
557 Evne RHSs labelled InlineMe aren't caught here, because there might be
558 no benefit from inlining at the call site.
560 [Sept 01] Don't unconditionally inline a top-level thing, because that
561 can simply make a static thing into something built dynamically. E.g.
565 [Remember that we treat \s as a one-shot lambda.] No point in
566 inlining x unless there is something interesting about the call site.
568 But watch out: if you aren't careful, some useful foldr/build fusion
569 can be lost (most notably in spectral/hartel/parstof) because the
570 foldr didn't see the build. Doing the dynamic allocation isn't a big
571 deal, in fact, but losing the fusion can be. But the right thing here
572 seems to be to do a callSiteInline based on the fact that there is
573 something interesting about the call site (it's strict). Hmm. That
576 Conclusion: inline top level things gaily until Phase 0 (the last
577 phase), at which point don't.
580 preInlineUnconditionally :: SimplEnv -> TopLevelFlag -> InId -> InExpr -> Bool
581 preInlineUnconditionally env top_lvl bndr rhs
583 | opt_SimplNoPreInlining = False
584 | otherwise = case idOccInfo bndr of
585 IAmDead -> True -- Happens in ((\x.1) v)
586 OneOcc in_lam True int_cxt -> try_once in_lam int_cxt
590 active = case phase of
591 SimplGently -> isAlwaysActive prag
592 SimplPhase n -> isActive n prag
593 prag = idInlinePragma bndr
595 try_once in_lam int_cxt -- There's one textual occurrence
596 | not in_lam = isNotTopLevel top_lvl || early_phase
597 | otherwise = int_cxt && canInlineInLam rhs
599 -- Be very careful before inlining inside a lambda, becuase (a) we must not
600 -- invalidate occurrence information, and (b) we want to avoid pushing a
601 -- single allocation (here) into multiple allocations (inside lambda).
602 -- Inlining a *function* with a single *saturated* call would be ok, mind you.
603 -- || (if is_cheap && not (canInlineInLam rhs) then pprTrace "preinline" (ppr bndr <+> ppr rhs) ok else ok)
605 -- is_cheap = exprIsCheap rhs
606 -- ok = is_cheap && int_cxt
608 -- int_cxt The context isn't totally boring
609 -- E.g. let f = \ab.BIG in \y. map f xs
610 -- Don't want to substitute for f, because then we allocate
611 -- its closure every time the \y is called
612 -- But: let f = \ab.BIG in \y. map (f y) xs
613 -- Now we do want to substitute for f, even though it's not
614 -- saturated, because we're going to allocate a closure for
615 -- (f y) every time round the loop anyhow.
617 -- canInlineInLam => free vars of rhs are (Once in_lam) or Many,
618 -- so substituting rhs inside a lambda doesn't change the occ info.
619 -- Sadly, not quite the same as exprIsHNF.
620 canInlineInLam (Lit l) = True
621 canInlineInLam (Lam b e) = isRuntimeVar b || canInlineInLam e
622 canInlineInLam (Note _ e) = canInlineInLam e
623 canInlineInLam _ = False
625 early_phase = case phase of
626 SimplPhase 0 -> False
628 -- If we don't have this early_phase test, consider
629 -- x = length [1,2,3]
630 -- The full laziness pass carefully floats all the cons cells to
631 -- top level, and preInlineUnconditionally floats them all back in.
632 -- Result is (a) static allocation replaced by dynamic allocation
633 -- (b) many simplifier iterations because this tickles
634 -- a related problem; only one inlining per pass
636 -- On the other hand, I have seen cases where top-level fusion is
637 -- lost if we don't inline top level thing (e.g. string constants)
638 -- Hence the test for phase zero (which is the phase for all the final
639 -- simplifications). Until phase zero we take no special notice of
640 -- top level things, but then we become more leery about inlining
645 postInlineUnconditionally
646 ~~~~~~~~~~~~~~~~~~~~~~~~~
647 @postInlineUnconditionally@ decides whether to unconditionally inline
648 a thing based on the form of its RHS; in particular if it has a
649 trivial RHS. If so, we can inline and discard the binding altogether.
651 NB: a loop breaker has must_keep_binding = True and non-loop-breakers
652 only have *forward* references Hence, it's safe to discard the binding
654 NOTE: This isn't our last opportunity to inline. We're at the binding
655 site right now, and we'll get another opportunity when we get to the
658 Note that we do this unconditional inlining only for trival RHSs.
659 Don't inline even WHNFs inside lambdas; doing so may simply increase
660 allocation when the function is called. This isn't the last chance; see
663 NB: Even inline pragmas (e.g. IMustBeINLINEd) are ignored here Why?
664 Because we don't even want to inline them into the RHS of constructor
665 arguments. See NOTE above
667 NB: At one time even NOINLINE was ignored here: if the rhs is trivial
668 it's best to inline it anyway. We often get a=E; b=a from desugaring,
669 with both a and b marked NOINLINE. But that seems incompatible with
670 our new view that inlining is like a RULE, so I'm sticking to the 'active'
674 postInlineUnconditionally :: SimplEnv -> TopLevelFlag -> OutId -> OccInfo -> OutExpr -> Unfolding -> Bool
675 postInlineUnconditionally env top_lvl bndr occ_info rhs unfolding
677 | isLoopBreaker occ_info = False
678 | isExportedId bndr = False
679 | exprIsTrivial rhs = True
682 OneOcc in_lam one_br int_cxt
683 -> (one_br || smallEnoughToInline unfolding) -- Small enough to dup
684 -- ToDo: consider discount on smallEnoughToInline if int_cxt is true
686 -- NB: Do we want to inline arbitrarily big things becuase
687 -- one_br is True? that can lead to inline cascades. But
688 -- preInlineUnconditionlly has dealt with all the common cases
689 -- so perhaps it's worth the risk. Here's an example
690 -- let f = if b then Left (\x.BIG) else Right (\y.BIG)
692 -- We can't preInlineUnconditionally because that woud invalidate
693 -- the occ info for b. Yet f is used just once, and duplicating
694 -- the case work is fine (exprIsCheap).
696 && ((isNotTopLevel top_lvl && not in_lam) ||
697 -- But outside a lambda, we want to be reasonably aggressive
698 -- about inlining into multiple branches of case
699 -- e.g. let x = <non-value>
700 -- in case y of { C1 -> ..x..; C2 -> ..x..; C3 -> ... }
701 -- Inlining can be a big win if C3 is the hot-spot, even if
702 -- the uses in C1, C2 are not 'interesting'
703 -- An example that gets worse if you add int_cxt here is 'clausify'
705 (isCheapUnfolding unfolding && int_cxt))
706 -- isCheap => acceptable work duplication; in_lam may be true
707 -- int_cxt to prevent us inlining inside a lambda without some
708 -- good reason. See the notes on int_cxt in preInlineUnconditionally
711 -- The point here is that for *non-values* that occur
712 -- outside a lambda, the call-site inliner won't have
713 -- a chance (becuase it doesn't know that the thing
714 -- only occurs once). The pre-inliner won't have gotten
715 -- it either, if the thing occurs in more than one branch
716 -- So the main target is things like
719 -- True -> case x of ...
720 -- False -> case x of ...
721 -- I'm not sure how important this is in practice
723 active = case getMode env of
724 SimplGently -> isAlwaysActive prag
725 SimplPhase n -> isActive n prag
726 prag = idInlinePragma bndr
728 activeInline :: SimplEnv -> OutId -> OccInfo -> Bool
729 activeInline env id occ
730 = case getMode env of
731 SimplGently -> isOneOcc occ && isAlwaysActive prag
732 -- No inlining at all when doing gentle stuff,
733 -- except for local things that occur once
734 -- The reason is that too little clean-up happens if you
735 -- don't inline use-once things. Also a bit of inlining is *good* for
736 -- full laziness; it can expose constant sub-expressions.
737 -- Example in spectral/mandel/Mandel.hs, where the mandelset
738 -- function gets a useful let-float if you inline windowToViewport
740 -- NB: we used to have a second exception, for data con wrappers.
741 -- On the grounds that we use gentle mode for rule LHSs, and
742 -- they match better when data con wrappers are inlined.
743 -- But that only really applies to the trivial wrappers (like (:)),
744 -- and they are now constructed as Compulsory unfoldings (in MkId)
745 -- so they'll happen anyway.
747 SimplPhase n -> isActive n prag
749 prag = idInlinePragma id
751 activeRule :: SimplEnv -> Maybe (Activation -> Bool)
752 -- Nothing => No rules at all
754 | opt_RulesOff = Nothing
756 = case getMode env of
757 SimplGently -> Just isAlwaysActive
758 -- Used to be Nothing (no rules in gentle mode)
759 -- Main motivation for changing is that I wanted
760 -- lift String ===> ...
761 -- to work in Template Haskell when simplifying
762 -- splices, so we get simpler code for literal strings
763 SimplPhase n -> Just (isActive n)
767 %************************************************************************
769 \subsection{Rebuilding a lambda}
771 %************************************************************************
774 mkLam :: SimplEnv -> [OutBinder] -> OutExpr -> SimplCont -> SimplM FloatsWithExpr
778 a) eta reduction, if that gives a trivial expression
779 b) eta expansion [only if there are some value lambdas]
780 c) floating lets out through big lambdas
781 [only if all tyvar lambdas, and only if this lambda
785 mkLam env bndrs body cont
786 = getDOptsSmpl `thenSmpl` \dflags ->
787 mkLam' dflags env bndrs body cont
789 mkLam' dflags env bndrs body cont
790 | dopt Opt_DoEtaReduction dflags,
791 Just etad_lam <- tryEtaReduce bndrs body
792 = tick (EtaReduction (head bndrs)) `thenSmpl_`
793 returnSmpl (emptyFloats env, etad_lam)
795 | dopt Opt_DoLambdaEtaExpansion dflags,
796 any isRuntimeVar bndrs
797 = tryEtaExpansion body `thenSmpl` \ body' ->
798 returnSmpl (emptyFloats env, mkLams bndrs body')
800 {- Sept 01: I'm experimenting with getting the
801 full laziness pass to float out past big lambdsa
802 | all isTyVar bndrs, -- Only for big lambdas
803 contIsRhs cont -- Only try the rhs type-lambda floating
804 -- if this is indeed a right-hand side; otherwise
805 -- we end up floating the thing out, only for float-in
806 -- to float it right back in again!
807 = tryRhsTyLam env bndrs body `thenSmpl` \ (floats, body') ->
808 returnSmpl (floats, mkLams bndrs body')
812 = returnSmpl (emptyFloats env, mkLams bndrs body)
816 %************************************************************************
818 \subsection{Eta expansion and reduction}
820 %************************************************************************
822 We try for eta reduction here, but *only* if we get all the
823 way to an exprIsTrivial expression.
824 We don't want to remove extra lambdas unless we are going
825 to avoid allocating this thing altogether
828 tryEtaReduce :: [OutBinder] -> OutExpr -> Maybe OutExpr
829 tryEtaReduce bndrs body
830 -- We don't use CoreUtils.etaReduce, because we can be more
832 -- (a) we already have the binders
833 -- (b) we can do the triviality test before computing the free vars
834 = go (reverse bndrs) body
836 go (b : bs) (App fun arg) | ok_arg b arg = go bs fun -- Loop round
837 go [] fun | ok_fun fun = Just fun -- Success!
838 go _ _ = Nothing -- Failure!
840 ok_fun fun = exprIsTrivial fun
841 && not (any (`elemVarSet` (exprFreeVars fun)) bndrs)
842 && (exprIsHNF fun || all ok_lam bndrs)
843 ok_lam v = isTyVar v || isDictId v
844 -- The exprIsHNF is because eta reduction is not
845 -- valid in general: \x. bot /= bot
846 -- So we need to be sure that the "fun" is a value.
848 -- However, we always want to reduce (/\a -> f a) to f
849 -- This came up in a RULE: foldr (build (/\a -> g a))
850 -- did not match foldr (build (/\b -> ...something complex...))
851 -- The type checker can insert these eta-expanded versions,
852 -- with both type and dictionary lambdas; hence the slightly
855 ok_arg b arg = varToCoreExpr b `cheapEqExpr` arg
859 Try eta expansion for RHSs
862 f = \x1..xn -> N ==> f = \x1..xn y1..ym -> N y1..ym
865 where (in both cases)
867 * The xi can include type variables
869 * The yi are all value variables
871 * N is a NORMAL FORM (i.e. no redexes anywhere)
872 wanting a suitable number of extra args.
874 We may have to sandwich some coerces between the lambdas
875 to make the types work. exprEtaExpandArity looks through coerces
876 when computing arity; and etaExpand adds the coerces as necessary when
877 actually computing the expansion.
880 tryEtaExpansion :: OutExpr -> SimplM OutExpr
881 -- There is at least one runtime binder in the binders
883 = getUniquesSmpl `thenSmpl` \ us ->
884 returnSmpl (etaExpand fun_arity us body (exprType body))
886 fun_arity = exprEtaExpandArity body
890 %************************************************************************
892 \subsection{Floating lets out of big lambdas}
894 %************************************************************************
896 tryRhsTyLam tries this transformation, when the big lambda appears as
897 the RHS of a let(rec) binding:
899 /\abc -> let(rec) x = e in b
901 let(rec) x' = /\abc -> let x = x' a b c in e
903 /\abc -> let x = x' a b c in b
905 This is good because it can turn things like:
907 let f = /\a -> letrec g = ... g ... in g
909 letrec g' = /\a -> ... g' a ...
913 which is better. In effect, it means that big lambdas don't impede
916 This optimisation is CRUCIAL in eliminating the junk introduced by
917 desugaring mutually recursive definitions. Don't eliminate it lightly!
919 So far as the implementation is concerned:
921 Invariant: go F e = /\tvs -> F e
925 = Let x' = /\tvs -> F e
929 G = F . Let x = x' tvs
931 go F (Letrec xi=ei in b)
932 = Letrec {xi' = /\tvs -> G ei}
936 G = F . Let {xi = xi' tvs}
938 [May 1999] If we do this transformation *regardless* then we can
939 end up with some pretty silly stuff. For example,
942 st = /\ s -> let { x1=r1 ; x2=r2 } in ...
947 st = /\s -> ...[y1 s/x1, y2 s/x2]
950 Unless the "..." is a WHNF there is really no point in doing this.
951 Indeed it can make things worse. Suppose x1 is used strictly,
954 x1* = case f y of { (a,b) -> e }
956 If we abstract this wrt the tyvar we then can't do the case inline
957 as we would normally do.
961 {- Trying to do this in full laziness
963 tryRhsTyLam :: SimplEnv -> [OutTyVar] -> OutExpr -> SimplM FloatsWithExpr
964 -- Call ensures that all the binders are type variables
966 tryRhsTyLam env tyvars body -- Only does something if there's a let
967 | not (all isTyVar tyvars)
968 || not (worth_it body) -- inside a type lambda,
969 = returnSmpl (emptyFloats env, body) -- and a WHNF inside that
972 = go env (\x -> x) body
975 worth_it e@(Let _ _) = whnf_in_middle e
978 whnf_in_middle (Let (NonRec x rhs) e) | isUnLiftedType (idType x) = False
979 whnf_in_middle (Let _ e) = whnf_in_middle e
980 whnf_in_middle e = exprIsCheap e
982 main_tyvar_set = mkVarSet tyvars
984 go env fn (Let bind@(NonRec var rhs) body)
986 = go env (fn . Let bind) body
988 go env fn (Let (NonRec var rhs) body)
989 = mk_poly tyvars_here var `thenSmpl` \ (var', rhs') ->
990 addAuxiliaryBind env (NonRec var' (mkLams tyvars_here (fn rhs))) $ \ env ->
991 go env (fn . Let (mk_silly_bind var rhs')) body
995 tyvars_here = varSetElems (main_tyvar_set `intersectVarSet` exprSomeFreeVars isTyVar rhs)
996 -- Abstract only over the type variables free in the rhs
997 -- wrt which the new binding is abstracted. But the naive
998 -- approach of abstract wrt the tyvars free in the Id's type
1000 -- /\ a b -> let t :: (a,b) = (e1, e2)
1003 -- Here, b isn't free in x's type, but we must nevertheless
1004 -- abstract wrt b as well, because t's type mentions b.
1005 -- Since t is floated too, we'd end up with the bogus:
1006 -- poly_t = /\ a b -> (e1, e2)
1007 -- poly_x = /\ a -> fst (poly_t a *b*)
1008 -- So for now we adopt the even more naive approach of
1009 -- abstracting wrt *all* the tyvars. We'll see if that
1010 -- gives rise to problems. SLPJ June 98
1012 go env fn (Let (Rec prs) body)
1013 = mapAndUnzipSmpl (mk_poly tyvars_here) vars `thenSmpl` \ (vars', rhss') ->
1015 gn body = fn (foldr Let body (zipWith mk_silly_bind vars rhss'))
1016 pairs = vars' `zip` [mkLams tyvars_here (gn rhs) | rhs <- rhss]
1018 addAuxiliaryBind env (Rec pairs) $ \ env ->
1021 (vars,rhss) = unzip prs
1022 tyvars_here = varSetElems (main_tyvar_set `intersectVarSet` exprsSomeFreeVars isTyVar (map snd prs))
1023 -- See notes with tyvars_here above
1025 go env fn body = returnSmpl (emptyFloats env, fn body)
1027 mk_poly tyvars_here var
1028 = getUniqueSmpl `thenSmpl` \ uniq ->
1030 poly_name = setNameUnique (idName var) uniq -- Keep same name
1031 poly_ty = mkForAllTys tyvars_here (idType var) -- But new type of course
1032 poly_id = mkLocalId poly_name poly_ty
1034 -- In the olden days, it was crucial to copy the occInfo of the original var,
1035 -- because we were looking at occurrence-analysed but as yet unsimplified code!
1036 -- In particular, we mustn't lose the loop breakers. BUT NOW we are looking
1037 -- at already simplified code, so it doesn't matter
1039 -- It's even right to retain single-occurrence or dead-var info:
1040 -- Suppose we started with /\a -> let x = E in B
1041 -- where x occurs once in B. Then we transform to:
1042 -- let x' = /\a -> E in /\a -> let x* = x' a in B
1043 -- where x* has an INLINE prag on it. Now, once x* is inlined,
1044 -- the occurrences of x' will be just the occurrences originally
1047 returnSmpl (poly_id, mkTyApps (Var poly_id) (mkTyVarTys tyvars_here))
1049 mk_silly_bind var rhs = NonRec var (Note InlineMe rhs)
1050 -- Suppose we start with:
1052 -- x = /\ a -> let g = G in E
1054 -- Then we'll float to get
1056 -- x = let poly_g = /\ a -> G
1057 -- in /\ a -> let g = poly_g a in E
1059 -- But now the occurrence analyser will see just one occurrence
1060 -- of poly_g, not inside a lambda, so the simplifier will
1061 -- PreInlineUnconditionally poly_g back into g! Badk to square 1!
1062 -- (I used to think that the "don't inline lone occurrences" stuff
1063 -- would stop this happening, but since it's the *only* occurrence,
1064 -- PreInlineUnconditionally kicks in first!)
1066 -- Solution: put an INLINE note on g's RHS, so that poly_g seems
1067 -- to appear many times. (NB: mkInlineMe eliminates
1068 -- such notes on trivial RHSs, so do it manually.)
1072 %************************************************************************
1074 \subsection{Case absorption and identity-case elimination}
1076 %************************************************************************
1078 mkCase puts a case expression back together, trying various transformations first.
1081 mkCase :: OutExpr -> OutId -> OutType
1082 -> [OutAlt] -- Increasing order
1085 mkCase scrut case_bndr ty alts
1086 = getDOptsSmpl `thenSmpl` \dflags ->
1087 mkAlts dflags scrut case_bndr alts `thenSmpl` \ better_alts ->
1088 mkCase1 scrut case_bndr ty better_alts
1092 mkAlts tries these things:
1094 1. If several alternatives are identical, merge them into
1095 a single DEFAULT alternative. I've occasionally seen this
1096 making a big difference:
1098 case e of =====> case e of
1099 C _ -> f x D v -> ....v....
1100 D v -> ....v.... DEFAULT -> f x
1103 The point is that we merge common RHSs, at least for the DEFAULT case.
1104 [One could do something more elaborate but I've never seen it needed.]
1105 To avoid an expensive test, we just merge branches equal to the *first*
1106 alternative; this picks up the common cases
1107 a) all branches equal
1108 b) some branches equal to the DEFAULT (which occurs first)
1111 case e of b { ==> case e of b {
1112 p1 -> rhs1 p1 -> rhs1
1114 pm -> rhsm pm -> rhsm
1115 _ -> case b of b' { pn -> let b'=b in rhsn
1117 ... po -> let b'=b in rhso
1118 po -> rhso _ -> let b'=b in rhsd
1122 which merges two cases in one case when -- the default alternative of
1123 the outer case scrutises the same variable as the outer case This
1124 transformation is called Case Merging. It avoids that the same
1125 variable is scrutinised multiple times.
1128 The case where transformation (1) showed up was like this (lib/std/PrelCError.lhs):
1134 where @is@ was something like
1136 p `is` n = p /= (-1) && p == n
1138 This gave rise to a horrible sequence of cases
1145 and similarly in cascade for all the join points!
1150 --------------------------------------------------
1151 -- 1. Merge identical branches
1152 --------------------------------------------------
1153 mkAlts dflags scrut case_bndr alts@((con1,bndrs1,rhs1) : con_alts)
1154 | all isDeadBinder bndrs1, -- Remember the default
1155 length filtered_alts < length con_alts -- alternative comes first
1156 = tick (AltMerge case_bndr) `thenSmpl_`
1157 returnSmpl better_alts
1159 filtered_alts = filter keep con_alts
1160 keep (con,bndrs,rhs) = not (all isDeadBinder bndrs && rhs `cheapEqExpr` rhs1)
1161 better_alts = (DEFAULT, [], rhs1) : filtered_alts
1164 --------------------------------------------------
1165 -- 2. Merge nested cases
1166 --------------------------------------------------
1168 mkAlts dflags scrut outer_bndr outer_alts
1169 | dopt Opt_CaseMerge dflags,
1170 (outer_alts_without_deflt, maybe_outer_deflt) <- findDefault outer_alts,
1171 Just (Case (Var scrut_var) inner_bndr _ inner_alts) <- maybe_outer_deflt,
1172 scruting_same_var scrut_var
1174 munged_inner_alts = [(con, args, munge_rhs rhs) | (con, args, rhs) <- inner_alts]
1175 munge_rhs rhs = bindCaseBndr inner_bndr (Var outer_bndr) rhs
1177 new_alts = mergeAlts outer_alts_without_deflt munged_inner_alts
1178 -- The merge keeps the inner DEFAULT at the front, if there is one
1179 -- and eliminates any inner_alts that are shadowed by the outer_alts
1181 tick (CaseMerge outer_bndr) `thenSmpl_`
1183 -- Warning: don't call mkAlts recursively!
1184 -- Firstly, there's no point, because inner alts have already had
1185 -- mkCase applied to them, so they won't have a case in their default
1186 -- Secondly, if you do, you get an infinite loop, because the bindCaseBndr
1187 -- in munge_rhs may put a case into the DEFAULT branch!
1189 -- We are scrutinising the same variable if it's
1190 -- the outer case-binder, or if the outer case scrutinises a variable
1191 -- (and it's the same). Testing both allows us not to replace the
1192 -- outer scrut-var with the outer case-binder (Simplify.simplCaseBinder).
1193 scruting_same_var = case scrut of
1194 Var outer_scrut -> \ v -> v == outer_bndr || v == outer_scrut
1195 other -> \ v -> v == outer_bndr
1197 ------------------------------------------------
1199 ------------------------------------------------
1201 mkAlts dflags scrut case_bndr other_alts = returnSmpl other_alts
1206 =================================================================================
1208 mkCase1 tries these things
1210 1. Eliminate the case altogether if possible
1218 and similar friends.
1221 Start with a simple situation:
1223 case x# of ===> e[x#/y#]
1226 (when x#, y# are of primitive type, of course). We can't (in general)
1227 do this for algebraic cases, because we might turn bottom into
1230 Actually, we generalise this idea to look for a case where we're
1231 scrutinising a variable, and we know that only the default case can
1236 other -> ...(case x of
1240 Here the inner case can be eliminated. This really only shows up in
1241 eliminating error-checking code.
1243 We also make sure that we deal with this very common case:
1248 Here we are using the case as a strict let; if x is used only once
1249 then we want to inline it. We have to be careful that this doesn't
1250 make the program terminate when it would have diverged before, so we
1252 - x is used strictly, or
1253 - e is already evaluated (it may so if e is a variable)
1255 Lastly, we generalise the transformation to handle this:
1261 We only do this for very cheaply compared r's (constructors, literals
1262 and variables). If pedantic bottoms is on, we only do it when the
1263 scrutinee is a PrimOp which can't fail.
1265 We do it *here*, looking at un-simplified alternatives, because we
1266 have to check that r doesn't mention the variables bound by the
1267 pattern in each alternative, so the binder-info is rather useful.
1269 So the case-elimination algorithm is:
1271 1. Eliminate alternatives which can't match
1273 2. Check whether all the remaining alternatives
1274 (a) do not mention in their rhs any of the variables bound in their pattern
1275 and (b) have equal rhss
1277 3. Check we can safely ditch the case:
1278 * PedanticBottoms is off,
1279 or * the scrutinee is an already-evaluated variable
1280 or * the scrutinee is a primop which is ok for speculation
1281 -- ie we want to preserve divide-by-zero errors, and
1282 -- calls to error itself!
1284 or * [Prim cases] the scrutinee is a primitive variable
1286 or * [Alg cases] the scrutinee is a variable and
1287 either * the rhs is the same variable
1288 (eg case x of C a b -> x ===> x)
1289 or * there is only one alternative, the default alternative,
1290 and the binder is used strictly in its scope.
1291 [NB this is helped by the "use default binder where
1292 possible" transformation; see below.]
1295 If so, then we can replace the case with one of the rhss.
1297 Further notes about case elimination
1298 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
1299 Consider: test :: Integer -> IO ()
1302 Turns out that this compiles to:
1305 eta1 :: State# RealWorld ->
1306 case PrelNum.< eta PrelNum.zeroInteger of wild { __DEFAULT ->
1308 (PrelNum.jtos eta ($w[] @ Char))
1310 of wild1 { (# new_s, a4 #) -> PrelIO.lvl23 new_s }}
1312 Notice the strange '<' which has no effect at all. This is a funny one.
1313 It started like this:
1315 f x y = if x < 0 then jtos x
1316 else if y==0 then "" else jtos x
1318 At a particular call site we have (f v 1). So we inline to get
1320 if v < 0 then jtos x
1321 else if 1==0 then "" else jtos x
1323 Now simplify the 1==0 conditional:
1325 if v<0 then jtos v else jtos v
1327 Now common-up the two branches of the case:
1329 case (v<0) of DEFAULT -> jtos v
1331 Why don't we drop the case? Because it's strict in v. It's technically
1332 wrong to drop even unnecessary evaluations, and in practice they
1333 may be a result of 'seq' so we *definitely* don't want to drop those.
1334 I don't really know how to improve this situation.
1338 --------------------------------------------------
1339 -- 0. Check for empty alternatives
1340 --------------------------------------------------
1342 -- This isn't strictly an error. It's possible that the simplifer might "see"
1343 -- that an inner case has no accessible alternatives before it "sees" that the
1344 -- entire branch of an outer case is inaccessible. So we simply
1345 -- put an error case here insteadd
1346 mkCase1 scrut case_bndr ty []
1347 = pprTrace "mkCase1: null alts" (ppr case_bndr <+> ppr scrut) $
1348 return (mkApps (Var eRROR_ID)
1349 [Type ty, Lit (mkStringLit "Impossible alternative")])
1351 --------------------------------------------------
1352 -- 1. Eliminate the case altogether if poss
1353 --------------------------------------------------
1355 mkCase1 scrut case_bndr ty [(con,bndrs,rhs)]
1356 -- See if we can get rid of the case altogether
1357 -- See the extensive notes on case-elimination above
1358 -- mkCase made sure that if all the alternatives are equal,
1359 -- then there is now only one (DEFAULT) rhs
1360 | all isDeadBinder bndrs,
1362 -- Check that the scrutinee can be let-bound instead of case-bound
1363 exprOkForSpeculation scrut
1364 -- OK not to evaluate it
1365 -- This includes things like (==# a# b#)::Bool
1366 -- so that we simplify
1367 -- case ==# a# b# of { True -> x; False -> x }
1370 -- This particular example shows up in default methods for
1371 -- comparision operations (e.g. in (>=) for Int.Int32)
1372 || exprIsHNF scrut -- It's already evaluated
1373 || var_demanded_later scrut -- It'll be demanded later
1375 -- || not opt_SimplPedanticBottoms) -- Or we don't care!
1376 -- We used to allow improving termination by discarding cases, unless -fpedantic-bottoms was on,
1377 -- but that breaks badly for the dataToTag# primop, which relies on a case to evaluate
1378 -- its argument: case x of { y -> dataToTag# y }
1379 -- Here we must *not* discard the case, because dataToTag# just fetches the tag from
1380 -- the info pointer. So we'll be pedantic all the time, and see if that gives any
1382 -- Also we don't want to discard 'seq's
1383 = tick (CaseElim case_bndr) `thenSmpl_`
1384 returnSmpl (bindCaseBndr case_bndr scrut rhs)
1387 -- The case binder is going to be evaluated later,
1388 -- and the scrutinee is a simple variable
1389 var_demanded_later (Var v) = isStrictDmd (idNewDemandInfo case_bndr)
1390 var_demanded_later other = False
1393 --------------------------------------------------
1395 --------------------------------------------------
1397 mkCase1 scrut case_bndr ty alts -- Identity case
1398 | all identity_alt alts
1399 = tick (CaseIdentity case_bndr) `thenSmpl_`
1400 returnSmpl (re_note scrut)
1402 identity_alt (con, args, rhs) = de_note rhs `cheapEqExpr` identity_rhs con args
1404 identity_rhs (DataAlt con) args = mkConApp con (arg_tys ++ map varToCoreExpr args)
1405 identity_rhs (LitAlt lit) _ = Lit lit
1406 identity_rhs DEFAULT _ = Var case_bndr
1408 arg_tys = map Type (tyConAppArgs (idType case_bndr))
1411 -- case coerce T e of x { _ -> coerce T' x }
1412 -- And we definitely want to eliminate this case!
1413 -- So we throw away notes from the RHS, and reconstruct
1414 -- (at least an approximation) at the other end
1415 de_note (Note _ e) = de_note e
1418 -- re_note wraps a coerce if it might be necessary
1419 re_note scrut = case head alts of
1420 (_,_,rhs1@(Note _ _)) -> mkCoerce2 (exprType rhs1) (idType case_bndr) scrut
1424 --------------------------------------------------
1426 --------------------------------------------------
1427 mkCase1 scrut bndr ty alts = returnSmpl (Case scrut bndr ty alts)
1431 When adding auxiliary bindings for the case binder, it's worth checking if
1432 its dead, because it often is, and occasionally these mkCase transformations
1433 cascade rather nicely.
1436 bindCaseBndr bndr rhs body
1437 | isDeadBinder bndr = body
1438 | otherwise = bindNonRec bndr rhs body