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, mkLazyArgStop, mkRhsStop, contIsRhs, contIsRhsOrArg,
19 getContArgs, interestingCallContext, interestingArgContext,
20 interestingArg, isStrictType
24 #include "HsVersions.h"
27 import DynFlags ( SimplifierSwitch(..), SimplifierMode(..),
28 DynFlags, DynFlag(..), dopt )
29 import StaticFlags ( opt_UF_UpdateInPlace, opt_SimplNoPreInlining,
32 import CoreFVs ( exprFreeVars )
33 import CoreUtils ( cheapEqExpr, exprType, exprIsTrivial,
34 etaExpand, exprEtaExpandArity, bindNonRec, mkCoerce2,
35 findDefault, exprOkForSpeculation, exprIsHNF, mergeAlts
37 import Literal ( mkStringLit )
38 import CoreUnfold ( smallEnoughToInline )
39 import MkId ( eRROR_ID )
40 import Id ( Id, idType, isDataConWorkId, idOccInfo, isDictId,
41 isDeadBinder, idNewDemandInfo, isExportedId,
42 idUnfolding, idNewStrictness, idInlinePragma, idHasRules
44 import NewDemand ( isStrictDmd, isBotRes, splitStrictSig )
46 import Type ( Type, splitFunTys, dropForAlls, isStrictType,
47 splitTyConApp_maybe, tyConAppArgs
49 import TyCon ( tyConDataCons_maybe )
50 import DataCon ( dataConRepArity )
52 import BasicTypes ( TopLevelFlag(..), isNotTopLevel, OccInfo(..), isLoopBreaker, isOneOcc,
53 Activation, isAlwaysActive, isActive )
54 import Util ( lengthExceeds )
59 %************************************************************************
61 \subsection{The continuation data type}
63 %************************************************************************
66 data SimplCont -- Strict contexts
67 = Stop OutType -- Type of the result
69 Bool -- True <=> There is something interesting about
70 -- the context, and hence the inliner
71 -- should be a bit keener (see interestingCallContext)
73 -- (a) This is the RHS of a thunk whose type suggests
74 -- that update-in-place would be possible
75 -- (b) This is an argument of a function that has RULES
76 -- Inlining the call might allow the rule to fire
78 | CoerceIt OutType -- The To-type, simplified
82 InExpr SimplEnv -- The argument, as yet unsimplified,
83 SimplCont -- and its environment
86 InId [InAlt] SimplEnv -- The case binder, alts, and subst-env
89 | ArgOf LetRhsFlag -- An arbitrary strict context: the argument
90 -- of a strict function, or a primitive-arg fn
92 -- No DupFlag, because we never duplicate it
93 OutType -- arg_ty: type of the argument itself
94 OutType -- cont_ty: the type of the expression being sought by the context
95 -- f (error "foo") ==> coerce t (error "foo")
97 -- We need to know the type t, to which to coerce.
99 (SimplEnv -> OutExpr -> SimplM FloatsWithExpr) -- What to do with the result
100 -- The result expression in the OutExprStuff has type cont_ty
102 data LetRhsFlag = AnArg -- It's just an argument not a let RHS
103 | AnRhs -- It's the RHS of a let (so please float lets out of big lambdas)
105 instance Outputable LetRhsFlag where
106 ppr AnArg = ptext SLIT("arg")
107 ppr AnRhs = ptext SLIT("rhs")
109 instance Outputable SimplCont where
110 ppr (Stop ty is_rhs _) = ptext SLIT("Stop") <> brackets (ppr is_rhs) <+> ppr ty
111 ppr (ApplyTo dup arg se cont) = (ptext SLIT("ApplyTo") <+> ppr dup <+> ppr arg) $$ ppr cont
112 ppr (ArgOf _ _ _ _) = ptext SLIT("ArgOf...")
113 ppr (Select dup bndr alts se cont) = (ptext SLIT("Select") <+> ppr dup <+> ppr bndr) $$
114 (nest 4 (ppr alts)) $$ ppr cont
115 ppr (CoerceIt ty cont) = (ptext SLIT("CoerceIt") <+> ppr ty) $$ ppr cont
117 data DupFlag = OkToDup | NoDup
119 instance Outputable DupFlag where
120 ppr OkToDup = ptext SLIT("ok")
121 ppr NoDup = ptext SLIT("nodup")
125 mkBoringStop :: OutType -> SimplCont
126 mkBoringStop ty = Stop ty AnArg False
128 mkLazyArgStop :: OutType -> Bool -> SimplCont
129 mkLazyArgStop ty has_rules = Stop ty AnArg (canUpdateInPlace ty || has_rules)
131 mkRhsStop :: OutType -> SimplCont
132 mkRhsStop ty = Stop ty AnRhs (canUpdateInPlace ty)
134 contIsRhs :: SimplCont -> Bool
135 contIsRhs (Stop _ AnRhs _) = True
136 contIsRhs (ArgOf AnRhs _ _ _) = True
137 contIsRhs other = False
139 contIsRhsOrArg (Stop _ _ _) = True
140 contIsRhsOrArg (ArgOf _ _ _ _) = True
141 contIsRhsOrArg other = False
144 contIsDupable :: SimplCont -> Bool
145 contIsDupable (Stop _ _ _) = True
146 contIsDupable (ApplyTo OkToDup _ _ _) = True
147 contIsDupable (Select OkToDup _ _ _ _) = True
148 contIsDupable (CoerceIt _ cont) = contIsDupable cont
149 contIsDupable other = False
152 discardableCont :: SimplCont -> Bool
153 discardableCont (Stop _ _ _) = False
154 discardableCont (CoerceIt _ cont) = discardableCont cont
155 discardableCont other = True
157 discardCont :: SimplCont -- A continuation, expecting
158 -> SimplCont -- Replace the continuation with a suitable coerce
159 discardCont cont = case cont of
160 Stop to_ty is_rhs _ -> cont
161 other -> CoerceIt to_ty (mkBoringStop to_ty)
163 to_ty = contResultType cont
166 contResultType :: SimplCont -> OutType
167 contResultType (Stop to_ty _ _) = to_ty
168 contResultType (ArgOf _ _ to_ty _) = to_ty
169 contResultType (ApplyTo _ _ _ cont) = contResultType cont
170 contResultType (CoerceIt _ cont) = contResultType cont
171 contResultType (Select _ _ _ _ cont) = contResultType cont
174 countValArgs :: SimplCont -> Int
175 countValArgs (ApplyTo _ (Type ty) se cont) = countValArgs cont
176 countValArgs (ApplyTo _ val_arg se cont) = 1 + countValArgs cont
177 countValArgs other = 0
179 countArgs :: SimplCont -> Int
180 countArgs (ApplyTo _ arg se cont) = 1 + countArgs cont
184 pushContArgs :: SimplEnv -> [OutArg] -> SimplCont -> SimplCont
185 -- Pushes args with the specified environment
186 pushContArgs env [] cont = cont
187 pushContArgs env (arg : args) cont = ApplyTo NoDup arg env (pushContArgs env args cont)
192 getContArgs :: SwitchChecker
193 -> OutId -> SimplCont
194 -> ([(InExpr, SimplEnv, Bool)], -- Arguments; the Bool is true for strict args
195 SimplCont) -- Remaining continuation
196 -- getContArgs id k = (args, k', inl)
197 -- args are the leading ApplyTo items in k
198 -- (i.e. outermost comes first)
199 -- augmented with demand info from the functionn
200 getContArgs chkr fun orig_cont
202 -- Ignore strictness info if the no-case-of-case
203 -- flag is on. Strictness changes evaluation order
204 -- and that can change full laziness
205 stricts | switchIsOn chkr NoCaseOfCase = vanilla_stricts
206 | otherwise = computed_stricts
208 go [] stricts orig_cont
210 ----------------------------
213 go acc ss (ApplyTo _ arg@(Type _) se cont)
214 = go ((arg,se,False) : acc) ss cont
215 -- NB: don't bother to instantiate the function type
218 go acc (s:ss) (ApplyTo _ arg se cont)
219 = go ((arg,se,s) : acc) ss cont
221 -- We're run out of arguments, or else we've run out of demands
222 -- The latter only happens if the result is guaranteed bottom
223 -- This is the case for
224 -- * case (error "hello") of { ... }
225 -- * (error "Hello") arg
226 -- * f (error "Hello") where f is strict
228 -- Then, especially in the first of these cases, we'd like to discard
229 -- the continuation, leaving just the bottoming expression. But the
230 -- type might not be right, so we may have to add a coerce.
232 | null ss && discardableCont cont = (reverse acc, discardCont cont)
233 | otherwise = (reverse acc, cont)
235 ----------------------------
236 vanilla_stricts, computed_stricts :: [Bool]
237 vanilla_stricts = repeat False
238 computed_stricts = zipWith (||) fun_stricts arg_stricts
240 ----------------------------
241 (val_arg_tys, _) = splitFunTys (dropForAlls (idType fun))
242 arg_stricts = map isStrictType val_arg_tys ++ repeat False
243 -- These argument types are used as a cheap and cheerful way to find
244 -- unboxed arguments, which must be strict. But it's an InType
245 -- and so there might be a type variable where we expect a function
246 -- type (the substitution hasn't happened yet). And we don't bother
247 -- doing the type applications for a polymorphic function.
248 -- Hence the splitFunTys*IgnoringForAlls*
250 ----------------------------
251 -- If fun_stricts is finite, it means the function returns bottom
252 -- after that number of value args have been consumed
253 -- Otherwise it's infinite, extended with False
255 = case splitStrictSig (idNewStrictness fun) of
256 (demands, result_info)
257 | not (demands `lengthExceeds` countValArgs orig_cont)
258 -> -- Enough args, use the strictness given.
259 -- For bottoming functions we used to pretend that the arg
260 -- is lazy, so that we don't treat the arg as an
261 -- interesting context. This avoids substituting
262 -- top-level bindings for (say) strings into
263 -- calls to error. But now we are more careful about
264 -- inlining lone variables, so its ok (see SimplUtils.analyseCont)
265 if isBotRes result_info then
266 map isStrictDmd demands -- Finite => result is bottom
268 map isStrictDmd demands ++ vanilla_stricts
270 other -> vanilla_stricts -- Not enough args, or no strictness
273 interestingArg :: OutExpr -> Bool
274 -- An argument is interesting if it has *some* structure
275 -- We are here trying to avoid unfolding a function that
276 -- is applied only to variables that have no unfolding
277 -- (i.e. they are probably lambda bound): f x y z
278 -- There is little point in inlining f here.
279 interestingArg (Var v) = hasSomeUnfolding (idUnfolding v)
280 -- Was: isValueUnfolding (idUnfolding v')
281 -- But that seems over-pessimistic
283 -- This accounts for an argument like
284 -- () or [], which is definitely interesting
285 interestingArg (Type _) = False
286 interestingArg (App fn (Type _)) = interestingArg fn
287 interestingArg (Note _ a) = interestingArg a
288 interestingArg other = True
289 -- Consider let x = 3 in f x
290 -- The substitution will contain (x -> ContEx 3), and we want to
291 -- to say that x is an interesting argument.
292 -- But consider also (\x. f x y) y
293 -- The substitution will contain (x -> ContEx y), and we want to say
294 -- that x is not interesting (assuming y has no unfolding)
297 Comment about interestingCallContext
298 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
299 We want to avoid inlining an expression where there can't possibly be
300 any gain, such as in an argument position. Hence, if the continuation
301 is interesting (eg. a case scrutinee, application etc.) then we
302 inline, otherwise we don't.
304 Previously some_benefit used to return True only if the variable was
305 applied to some value arguments. This didn't work:
307 let x = _coerce_ (T Int) Int (I# 3) in
308 case _coerce_ Int (T Int) x of
311 we want to inline x, but can't see that it's a constructor in a case
312 scrutinee position, and some_benefit is False.
316 dMonadST = _/\_ t -> :Monad (g1 _@_ t, g2 _@_ t, g3 _@_ t)
318 .... case dMonadST _@_ x0 of (a,b,c) -> ....
320 we'd really like to inline dMonadST here, but we *don't* want to
321 inline if the case expression is just
323 case x of y { DEFAULT -> ... }
325 since we can just eliminate this case instead (x is in WHNF). Similar
326 applies when x is bound to a lambda expression. Hence
327 contIsInteresting looks for case expressions with just a single
331 interestingCallContext :: Bool -- False <=> no args at all
332 -> Bool -- False <=> no value args
334 -- The "lone-variable" case is important. I spent ages
335 -- messing about with unsatisfactory varaints, but this is nice.
336 -- The idea is that if a variable appear all alone
337 -- as an arg of lazy fn, or rhs Stop
338 -- as scrutinee of a case Select
339 -- as arg of a strict fn ArgOf
340 -- then we should not inline it (unless there is some other reason,
341 -- e.g. is is the sole occurrence). We achieve this by making
342 -- interestingCallContext return False for a lone variable.
344 -- Why? At least in the case-scrutinee situation, turning
345 -- let x = (a,b) in case x of y -> ...
347 -- let x = (a,b) in case (a,b) of y -> ...
349 -- let x = (a,b) in let y = (a,b) in ...
350 -- is bad if the binding for x will remain.
352 -- Another example: I discovered that strings
353 -- were getting inlined straight back into applications of 'error'
354 -- because the latter is strict.
356 -- f = \x -> ...(error s)...
358 -- Fundamentally such contexts should not ecourage inlining because
359 -- the context can ``see'' the unfolding of the variable (e.g. case or a RULE)
360 -- so there's no gain.
362 -- However, even a type application or coercion isn't a lone variable.
364 -- case $fMonadST @ RealWorld of { :DMonad a b c -> c }
365 -- We had better inline that sucker! The case won't see through it.
367 -- For now, I'm treating treating a variable applied to types
368 -- in a *lazy* context "lone". The motivating example was
370 -- g = /\a. \y. h (f a)
371 -- There's no advantage in inlining f here, and perhaps
372 -- a significant disadvantage. Hence some_val_args in the Stop case
374 interestingCallContext some_args some_val_args cont
377 interesting (Select _ _ _ _ _) = some_args
378 interesting (ApplyTo _ _ _ _) = True -- Can happen if we have (coerce t (f x)) y
379 -- Perhaps True is a bit over-keen, but I've
380 -- seen (coerce f) x, where f has an INLINE prag,
381 -- So we have to give some motivaiton for inlining it
382 interesting (ArgOf _ _ _ _) = some_val_args
383 interesting (Stop ty _ interesting) = some_val_args && interesting
384 interesting (CoerceIt _ cont) = interesting cont
385 -- If this call is the arg of a strict function, the context
386 -- is a bit interesting. If we inline here, we may get useful
387 -- evaluation information to avoid repeated evals: e.g.
389 -- Here the contIsInteresting makes the '*' keener to inline,
390 -- which in turn exposes a constructor which makes the '+' inline.
391 -- Assuming that +,* aren't small enough to inline regardless.
393 -- It's also very important to inline in a strict context for things
396 -- Here, the context of (f x) is strict, and if f's unfolding is
397 -- a build it's *great* to inline it here. So we must ensure that
398 -- the context for (f x) is not totally uninteresting.
402 interestingArgContext :: Id -> SimplCont -> Bool
403 -- If the argument has form (f x y), where x,y are boring,
404 -- and f is marked INLINE, then we don't want to inline f.
405 -- But if the context of the argument is
407 -- where g has rules, then we *do* want to inline f, in case it
408 -- exposes a rule that might fire. Similarly, if the context is
410 -- where h has rules, then we do want to inline f.
411 -- The interesting_arg_ctxt flag makes this happen; if it's
412 -- set, the inliner gets just enough keener to inline f
413 -- regardless of how boring f's arguments are, if it's marked INLINE
415 -- The alternative would be to *always* inline an INLINE function,
416 -- regardless of how boring its context is; but that seems overkill
417 -- For example, it'd mean that wrapper functions were always inlined
418 interestingArgContext fn cont
419 = idHasRules fn || go cont
421 go (Select {}) = False
422 go (ApplyTo {}) = False
424 go (CoerceIt _ c) = go c
425 go (Stop _ _ interesting) = interesting
428 canUpdateInPlace :: Type -> Bool
429 -- Consider let x = <wurble> in ...
430 -- If <wurble> returns an explicit constructor, we might be able
431 -- to do update in place. So we treat even a thunk RHS context
432 -- as interesting if update in place is possible. We approximate
433 -- this by seeing if the type has a single constructor with a
434 -- small arity. But arity zero isn't good -- we share the single copy
435 -- for that case, so no point in sharing.
438 | not opt_UF_UpdateInPlace = False
440 = case splitTyConApp_maybe ty of
442 Just (tycon, _) -> case tyConDataCons_maybe tycon of
443 Just [dc] -> arity == 1 || arity == 2
445 arity = dataConRepArity dc
451 %************************************************************************
453 \subsection{Decisions about inlining}
455 %************************************************************************
457 Inlining is controlled partly by the SimplifierMode switch. This has two
460 SimplGently (a) Simplifying before specialiser/full laziness
461 (b) Simplifiying inside INLINE pragma
462 (c) Simplifying the LHS of a rule
463 (d) Simplifying a GHCi expression or Template
466 SimplPhase n Used at all other times
468 The key thing about SimplGently is that it does no call-site inlining.
469 Before full laziness we must be careful not to inline wrappers,
470 because doing so inhibits floating
471 e.g. ...(case f x of ...)...
472 ==> ...(case (case x of I# x# -> fw x#) of ...)...
473 ==> ...(case x of I# x# -> case fw x# of ...)...
474 and now the redex (f x) isn't floatable any more.
476 The no-inlining thing is also important for Template Haskell. You might be
477 compiling in one-shot mode with -O2; but when TH compiles a splice before
478 running it, we don't want to use -O2. Indeed, we don't want to inline
479 anything, because the byte-code interpreter might get confused about
480 unboxed tuples and suchlike.
484 SimplGently is also used as the mode to simplify inside an InlineMe note.
487 inlineMode :: SimplifierMode
488 inlineMode = SimplGently
491 It really is important to switch off inlinings inside such
492 expressions. Consider the following example
498 in ...g...g...g...g...g...
500 Now, if that's the ONLY occurrence of f, it will be inlined inside g,
501 and thence copied multiple times when g is inlined.
504 This function may be inlinined in other modules, so we
505 don't want to remove (by inlining) calls to functions that have
506 specialisations, or that may have transformation rules in an importing
509 E.g. {-# INLINE f #-}
512 and suppose that g is strict *and* has specialisations. If we inline
513 g's wrapper, we deny f the chance of getting the specialised version
514 of g when f is inlined at some call site (perhaps in some other
517 It's also important not to inline a worker back into a wrapper.
519 wraper = inline_me (\x -> ...worker... )
520 Normally, the inline_me prevents the worker getting inlined into
521 the wrapper (initially, the worker's only call site!). But,
522 if the wrapper is sure to be called, the strictness analyser will
523 mark it 'demanded', so when the RHS is simplified, it'll get an ArgOf
524 continuation. That's why the keep_inline predicate returns True for
525 ArgOf continuations. It shouldn't do any harm not to dissolve the
526 inline-me note under these circumstances.
528 Note that the result is that we do very little simplification
531 all xs = foldr (&&) True xs
532 any p = all . map p {-# INLINE any #-}
534 Problem: any won't get deforested, and so if it's exported and the
535 importer doesn't use the inlining, (eg passes it as an arg) then we
536 won't get deforestation at all. We havn't solved this problem yet!
539 preInlineUnconditionally
540 ~~~~~~~~~~~~~~~~~~~~~~~~
541 @preInlineUnconditionally@ examines a bndr to see if it is used just
542 once in a completely safe way, so that it is safe to discard the
543 binding inline its RHS at the (unique) usage site, REGARDLESS of how
544 big the RHS might be. If this is the case we don't simplify the RHS
545 first, but just inline it un-simplified.
547 This is much better than first simplifying a perhaps-huge RHS and then
548 inlining and re-simplifying it. Indeed, it can be at least quadratically
557 We may end up simplifying e1 N times, e2 N-1 times, e3 N-3 times etc.
558 This can happen with cascades of functions too:
565 THE MAIN INVARIANT is this:
567 ---- preInlineUnconditionally invariant -----
568 IF preInlineUnconditionally chooses to inline x = <rhs>
569 THEN doing the inlining should not change the occurrence
570 info for the free vars of <rhs>
571 ----------------------------------------------
573 For example, it's tempting to look at trivial binding like
575 and inline it unconditionally. But suppose x is used many times,
576 but this is the unique occurrence of y. Then inlining x would change
577 y's occurrence info, which breaks the invariant. It matters: y
578 might have a BIG rhs, which will now be dup'd at every occurrenc of x.
581 Evne RHSs labelled InlineMe aren't caught here, because there might be
582 no benefit from inlining at the call site.
584 [Sept 01] Don't unconditionally inline a top-level thing, because that
585 can simply make a static thing into something built dynamically. E.g.
589 [Remember that we treat \s as a one-shot lambda.] No point in
590 inlining x unless there is something interesting about the call site.
592 But watch out: if you aren't careful, some useful foldr/build fusion
593 can be lost (most notably in spectral/hartel/parstof) because the
594 foldr didn't see the build. Doing the dynamic allocation isn't a big
595 deal, in fact, but losing the fusion can be. But the right thing here
596 seems to be to do a callSiteInline based on the fact that there is
597 something interesting about the call site (it's strict). Hmm. That
600 Conclusion: inline top level things gaily until Phase 0 (the last
601 phase), at which point don't.
604 preInlineUnconditionally :: SimplEnv -> TopLevelFlag -> InId -> InExpr -> Bool
605 preInlineUnconditionally env top_lvl bndr rhs
607 | opt_SimplNoPreInlining = False
608 | otherwise = case idOccInfo bndr of
609 IAmDead -> True -- Happens in ((\x.1) v)
610 OneOcc in_lam True int_cxt -> try_once in_lam int_cxt
614 active = case phase of
615 SimplGently -> isAlwaysActive prag
616 SimplPhase n -> isActive n prag
617 prag = idInlinePragma bndr
619 try_once in_lam int_cxt -- There's one textual occurrence
620 | not in_lam = isNotTopLevel top_lvl || early_phase
621 | otherwise = int_cxt && canInlineInLam rhs
623 -- Be very careful before inlining inside a lambda, becuase (a) we must not
624 -- invalidate occurrence information, and (b) we want to avoid pushing a
625 -- single allocation (here) into multiple allocations (inside lambda).
626 -- Inlining a *function* with a single *saturated* call would be ok, mind you.
627 -- || (if is_cheap && not (canInlineInLam rhs) then pprTrace "preinline" (ppr bndr <+> ppr rhs) ok else ok)
629 -- is_cheap = exprIsCheap rhs
630 -- ok = is_cheap && int_cxt
632 -- int_cxt The context isn't totally boring
633 -- E.g. let f = \ab.BIG in \y. map f xs
634 -- Don't want to substitute for f, because then we allocate
635 -- its closure every time the \y is called
636 -- But: let f = \ab.BIG in \y. map (f y) xs
637 -- Now we do want to substitute for f, even though it's not
638 -- saturated, because we're going to allocate a closure for
639 -- (f y) every time round the loop anyhow.
641 -- canInlineInLam => free vars of rhs are (Once in_lam) or Many,
642 -- so substituting rhs inside a lambda doesn't change the occ info.
643 -- Sadly, not quite the same as exprIsHNF.
644 canInlineInLam (Lit l) = True
645 canInlineInLam (Lam b e) = isRuntimeVar b || canInlineInLam e
646 canInlineInLam (Note _ e) = canInlineInLam e
647 canInlineInLam _ = False
649 early_phase = case phase of
650 SimplPhase 0 -> False
652 -- If we don't have this early_phase test, consider
653 -- x = length [1,2,3]
654 -- The full laziness pass carefully floats all the cons cells to
655 -- top level, and preInlineUnconditionally floats them all back in.
656 -- Result is (a) static allocation replaced by dynamic allocation
657 -- (b) many simplifier iterations because this tickles
658 -- a related problem; only one inlining per pass
660 -- On the other hand, I have seen cases where top-level fusion is
661 -- lost if we don't inline top level thing (e.g. string constants)
662 -- Hence the test for phase zero (which is the phase for all the final
663 -- simplifications). Until phase zero we take no special notice of
664 -- top level things, but then we become more leery about inlining
669 postInlineUnconditionally
670 ~~~~~~~~~~~~~~~~~~~~~~~~~
671 @postInlineUnconditionally@ decides whether to unconditionally inline
672 a thing based on the form of its RHS; in particular if it has a
673 trivial RHS. If so, we can inline and discard the binding altogether.
675 NB: a loop breaker has must_keep_binding = True and non-loop-breakers
676 only have *forward* references Hence, it's safe to discard the binding
678 NOTE: This isn't our last opportunity to inline. We're at the binding
679 site right now, and we'll get another opportunity when we get to the
682 Note that we do this unconditional inlining only for trival RHSs.
683 Don't inline even WHNFs inside lambdas; doing so may simply increase
684 allocation when the function is called. This isn't the last chance; see
687 NB: Even inline pragmas (e.g. IMustBeINLINEd) are ignored here Why?
688 Because we don't even want to inline them into the RHS of constructor
689 arguments. See NOTE above
691 NB: At one time even NOINLINE was ignored here: if the rhs is trivial
692 it's best to inline it anyway. We often get a=E; b=a from desugaring,
693 with both a and b marked NOINLINE. But that seems incompatible with
694 our new view that inlining is like a RULE, so I'm sticking to the 'active'
698 postInlineUnconditionally
699 :: SimplEnv -> TopLevelFlag
700 -> InId -- The binder (an OutId would be fine too)
701 -> OccInfo -- From the InId
705 postInlineUnconditionally env top_lvl bndr occ_info rhs unfolding
707 | isLoopBreaker occ_info = False
708 | isExportedId bndr = False
709 | exprIsTrivial rhs = True
712 -- The point of examining occ_info here is that for *non-values*
713 -- that occur outside a lambda, the call-site inliner won't have
714 -- a chance (becuase it doesn't know that the thing
715 -- only occurs once). The pre-inliner won't have gotten
716 -- it either, if the thing occurs in more than one branch
717 -- So the main target is things like
720 -- True -> case x of ...
721 -- False -> case x of ...
722 -- I'm not sure how important this is in practice
723 OneOcc in_lam one_br int_cxt -- OneOcc => no work-duplication issue
724 -> smallEnoughToInline unfolding -- Small enough to dup
725 -- ToDo: consider discount on smallEnoughToInline if int_cxt is true
727 -- NB: Do NOT inline arbitrarily big things, even if one_br is True
728 -- Reason: doing so risks exponential behaviour. We simplify a big
729 -- expression, inline it, and simplify it again. But if the
730 -- very same thing happens in the big expression, we get
732 -- PRINCIPLE: when we've already simplified an expression once,
733 -- make sure that we only inline it if it's reasonably small.
735 && ((isNotTopLevel top_lvl && not in_lam) ||
736 -- But outside a lambda, we want to be reasonably aggressive
737 -- about inlining into multiple branches of case
738 -- e.g. let x = <non-value>
739 -- in case y of { C1 -> ..x..; C2 -> ..x..; C3 -> ... }
740 -- Inlining can be a big win if C3 is the hot-spot, even if
741 -- the uses in C1, C2 are not 'interesting'
742 -- An example that gets worse if you add int_cxt here is 'clausify'
744 (isCheapUnfolding unfolding && int_cxt))
745 -- isCheap => acceptable work duplication; in_lam may be true
746 -- int_cxt to prevent us inlining inside a lambda without some
747 -- good reason. See the notes on int_cxt in preInlineUnconditionally
751 -- Here's an example that we don't handle well:
752 -- let f = if b then Left (\x.BIG) else Right (\y.BIG)
753 -- in \y. ....case f of {...} ....
754 -- Here f is used just once, and duplicating the case work is fine (exprIsCheap).
756 -- * We can't preInlineUnconditionally because that woud invalidate
757 -- the occ info for b.
758 -- * We can't postInlineUnconditionally because the RHS is big, and
759 -- that risks exponential behaviour
760 -- * We can't call-site inline, because the rhs is big
764 active = case getMode env of
765 SimplGently -> isAlwaysActive prag
766 SimplPhase n -> isActive n prag
767 prag = idInlinePragma bndr
769 activeInline :: SimplEnv -> OutId -> OccInfo -> Bool
770 activeInline env id occ
771 = case getMode env of
772 SimplGently -> isOneOcc occ && isAlwaysActive prag
773 -- No inlining at all when doing gentle stuff,
774 -- except for local things that occur once
775 -- The reason is that too little clean-up happens if you
776 -- don't inline use-once things. Also a bit of inlining is *good* for
777 -- full laziness; it can expose constant sub-expressions.
778 -- Example in spectral/mandel/Mandel.hs, where the mandelset
779 -- function gets a useful let-float if you inline windowToViewport
781 -- NB: we used to have a second exception, for data con wrappers.
782 -- On the grounds that we use gentle mode for rule LHSs, and
783 -- they match better when data con wrappers are inlined.
784 -- But that only really applies to the trivial wrappers (like (:)),
785 -- and they are now constructed as Compulsory unfoldings (in MkId)
786 -- so they'll happen anyway.
788 SimplPhase n -> isActive n prag
790 prag = idInlinePragma id
792 activeRule :: SimplEnv -> Maybe (Activation -> Bool)
793 -- Nothing => No rules at all
795 | opt_RulesOff = Nothing
797 = case getMode env of
798 SimplGently -> Just isAlwaysActive
799 -- Used to be Nothing (no rules in gentle mode)
800 -- Main motivation for changing is that I wanted
801 -- lift String ===> ...
802 -- to work in Template Haskell when simplifying
803 -- splices, so we get simpler code for literal strings
804 SimplPhase n -> Just (isActive n)
808 %************************************************************************
810 \subsection{Rebuilding a lambda}
812 %************************************************************************
815 mkLam :: SimplEnv -> [OutBinder] -> OutExpr -> SimplCont -> SimplM FloatsWithExpr
819 a) eta reduction, if that gives a trivial expression
820 b) eta expansion [only if there are some value lambdas]
821 c) floating lets out through big lambdas
822 [only if all tyvar lambdas, and only if this lambda
826 mkLam env bndrs body cont
827 = getDOptsSmpl `thenSmpl` \dflags ->
828 mkLam' dflags env bndrs body cont
830 mkLam' dflags env bndrs body cont
831 | dopt Opt_DoEtaReduction dflags,
832 Just etad_lam <- tryEtaReduce bndrs body
833 = tick (EtaReduction (head bndrs)) `thenSmpl_`
834 returnSmpl (emptyFloats env, etad_lam)
836 | dopt Opt_DoLambdaEtaExpansion dflags,
837 any isRuntimeVar bndrs
838 = tryEtaExpansion dflags body `thenSmpl` \ body' ->
839 returnSmpl (emptyFloats env, mkLams bndrs body')
841 {- Sept 01: I'm experimenting with getting the
842 full laziness pass to float out past big lambdsa
843 | all isTyVar bndrs, -- Only for big lambdas
844 contIsRhs cont -- Only try the rhs type-lambda floating
845 -- if this is indeed a right-hand side; otherwise
846 -- we end up floating the thing out, only for float-in
847 -- to float it right back in again!
848 = tryRhsTyLam env bndrs body `thenSmpl` \ (floats, body') ->
849 returnSmpl (floats, mkLams bndrs body')
853 = returnSmpl (emptyFloats env, mkLams bndrs body)
857 %************************************************************************
859 \subsection{Eta expansion and reduction}
861 %************************************************************************
863 We try for eta reduction here, but *only* if we get all the
864 way to an exprIsTrivial expression.
865 We don't want to remove extra lambdas unless we are going
866 to avoid allocating this thing altogether
869 tryEtaReduce :: [OutBinder] -> OutExpr -> Maybe OutExpr
870 tryEtaReduce bndrs body
871 -- We don't use CoreUtils.etaReduce, because we can be more
873 -- (a) we already have the binders
874 -- (b) we can do the triviality test before computing the free vars
875 = go (reverse bndrs) body
877 go (b : bs) (App fun arg) | ok_arg b arg = go bs fun -- Loop round
878 go [] fun | ok_fun fun = Just fun -- Success!
879 go _ _ = Nothing -- Failure!
881 ok_fun fun = exprIsTrivial fun
882 && not (any (`elemVarSet` (exprFreeVars fun)) bndrs)
883 && (exprIsHNF fun || all ok_lam bndrs)
884 ok_lam v = isTyVar v || isDictId v
885 -- The exprIsHNF is because eta reduction is not
886 -- valid in general: \x. bot /= bot
887 -- So we need to be sure that the "fun" is a value.
889 -- However, we always want to reduce (/\a -> f a) to f
890 -- This came up in a RULE: foldr (build (/\a -> g a))
891 -- did not match foldr (build (/\b -> ...something complex...))
892 -- The type checker can insert these eta-expanded versions,
893 -- with both type and dictionary lambdas; hence the slightly
896 ok_arg b arg = varToCoreExpr b `cheapEqExpr` arg
900 Try eta expansion for RHSs
903 f = \x1..xn -> N ==> f = \x1..xn y1..ym -> N y1..ym
906 where (in both cases)
908 * The xi can include type variables
910 * The yi are all value variables
912 * N is a NORMAL FORM (i.e. no redexes anywhere)
913 wanting a suitable number of extra args.
915 We may have to sandwich some coerces between the lambdas
916 to make the types work. exprEtaExpandArity looks through coerces
917 when computing arity; and etaExpand adds the coerces as necessary when
918 actually computing the expansion.
921 tryEtaExpansion :: DynFlags -> OutExpr -> SimplM OutExpr
922 -- There is at least one runtime binder in the binders
923 tryEtaExpansion dflags body
924 = getUniquesSmpl `thenSmpl` \ us ->
925 returnSmpl (etaExpand fun_arity us body (exprType body))
927 fun_arity = exprEtaExpandArity dflags body
931 %************************************************************************
933 \subsection{Floating lets out of big lambdas}
935 %************************************************************************
937 tryRhsTyLam tries this transformation, when the big lambda appears as
938 the RHS of a let(rec) binding:
940 /\abc -> let(rec) x = e in b
942 let(rec) x' = /\abc -> let x = x' a b c in e
944 /\abc -> let x = x' a b c in b
946 This is good because it can turn things like:
948 let f = /\a -> letrec g = ... g ... in g
950 letrec g' = /\a -> ... g' a ...
954 which is better. In effect, it means that big lambdas don't impede
957 This optimisation is CRUCIAL in eliminating the junk introduced by
958 desugaring mutually recursive definitions. Don't eliminate it lightly!
960 So far as the implementation is concerned:
962 Invariant: go F e = /\tvs -> F e
966 = Let x' = /\tvs -> F e
970 G = F . Let x = x' tvs
972 go F (Letrec xi=ei in b)
973 = Letrec {xi' = /\tvs -> G ei}
977 G = F . Let {xi = xi' tvs}
979 [May 1999] If we do this transformation *regardless* then we can
980 end up with some pretty silly stuff. For example,
983 st = /\ s -> let { x1=r1 ; x2=r2 } in ...
988 st = /\s -> ...[y1 s/x1, y2 s/x2]
991 Unless the "..." is a WHNF there is really no point in doing this.
992 Indeed it can make things worse. Suppose x1 is used strictly,
995 x1* = case f y of { (a,b) -> e }
997 If we abstract this wrt the tyvar we then can't do the case inline
998 as we would normally do.
1002 {- Trying to do this in full laziness
1004 tryRhsTyLam :: SimplEnv -> [OutTyVar] -> OutExpr -> SimplM FloatsWithExpr
1005 -- Call ensures that all the binders are type variables
1007 tryRhsTyLam env tyvars body -- Only does something if there's a let
1008 | not (all isTyVar tyvars)
1009 || not (worth_it body) -- inside a type lambda,
1010 = returnSmpl (emptyFloats env, body) -- and a WHNF inside that
1013 = go env (\x -> x) body
1016 worth_it e@(Let _ _) = whnf_in_middle e
1019 whnf_in_middle (Let (NonRec x rhs) e) | isUnLiftedType (idType x) = False
1020 whnf_in_middle (Let _ e) = whnf_in_middle e
1021 whnf_in_middle e = exprIsCheap e
1023 main_tyvar_set = mkVarSet tyvars
1025 go env fn (Let bind@(NonRec var rhs) body)
1027 = go env (fn . Let bind) body
1029 go env fn (Let (NonRec var rhs) body)
1030 = mk_poly tyvars_here var `thenSmpl` \ (var', rhs') ->
1031 addAuxiliaryBind env (NonRec var' (mkLams tyvars_here (fn rhs))) $ \ env ->
1032 go env (fn . Let (mk_silly_bind var rhs')) body
1036 tyvars_here = varSetElems (main_tyvar_set `intersectVarSet` exprSomeFreeVars isTyVar rhs)
1037 -- Abstract only over the type variables free in the rhs
1038 -- wrt which the new binding is abstracted. But the naive
1039 -- approach of abstract wrt the tyvars free in the Id's type
1041 -- /\ a b -> let t :: (a,b) = (e1, e2)
1044 -- Here, b isn't free in x's type, but we must nevertheless
1045 -- abstract wrt b as well, because t's type mentions b.
1046 -- Since t is floated too, we'd end up with the bogus:
1047 -- poly_t = /\ a b -> (e1, e2)
1048 -- poly_x = /\ a -> fst (poly_t a *b*)
1049 -- So for now we adopt the even more naive approach of
1050 -- abstracting wrt *all* the tyvars. We'll see if that
1051 -- gives rise to problems. SLPJ June 98
1053 go env fn (Let (Rec prs) body)
1054 = mapAndUnzipSmpl (mk_poly tyvars_here) vars `thenSmpl` \ (vars', rhss') ->
1056 gn body = fn (foldr Let body (zipWith mk_silly_bind vars rhss'))
1057 pairs = vars' `zip` [mkLams tyvars_here (gn rhs) | rhs <- rhss]
1059 addAuxiliaryBind env (Rec pairs) $ \ env ->
1062 (vars,rhss) = unzip prs
1063 tyvars_here = varSetElems (main_tyvar_set `intersectVarSet` exprsSomeFreeVars isTyVar (map snd prs))
1064 -- See notes with tyvars_here above
1066 go env fn body = returnSmpl (emptyFloats env, fn body)
1068 mk_poly tyvars_here var
1069 = getUniqueSmpl `thenSmpl` \ uniq ->
1071 poly_name = setNameUnique (idName var) uniq -- Keep same name
1072 poly_ty = mkForAllTys tyvars_here (idType var) -- But new type of course
1073 poly_id = mkLocalId poly_name poly_ty
1075 -- In the olden days, it was crucial to copy the occInfo of the original var,
1076 -- because we were looking at occurrence-analysed but as yet unsimplified code!
1077 -- In particular, we mustn't lose the loop breakers. BUT NOW we are looking
1078 -- at already simplified code, so it doesn't matter
1080 -- It's even right to retain single-occurrence or dead-var info:
1081 -- Suppose we started with /\a -> let x = E in B
1082 -- where x occurs once in B. Then we transform to:
1083 -- let x' = /\a -> E in /\a -> let x* = x' a in B
1084 -- where x* has an INLINE prag on it. Now, once x* is inlined,
1085 -- the occurrences of x' will be just the occurrences originally
1088 returnSmpl (poly_id, mkTyApps (Var poly_id) (mkTyVarTys tyvars_here))
1090 mk_silly_bind var rhs = NonRec var (Note InlineMe rhs)
1091 -- Suppose we start with:
1093 -- x = /\ a -> let g = G in E
1095 -- Then we'll float to get
1097 -- x = let poly_g = /\ a -> G
1098 -- in /\ a -> let g = poly_g a in E
1100 -- But now the occurrence analyser will see just one occurrence
1101 -- of poly_g, not inside a lambda, so the simplifier will
1102 -- PreInlineUnconditionally poly_g back into g! Badk to square 1!
1103 -- (I used to think that the "don't inline lone occurrences" stuff
1104 -- would stop this happening, but since it's the *only* occurrence,
1105 -- PreInlineUnconditionally kicks in first!)
1107 -- Solution: put an INLINE note on g's RHS, so that poly_g seems
1108 -- to appear many times. (NB: mkInlineMe eliminates
1109 -- such notes on trivial RHSs, so do it manually.)
1113 %************************************************************************
1115 \subsection{Case absorption and identity-case elimination}
1117 %************************************************************************
1119 mkCase puts a case expression back together, trying various transformations first.
1122 mkCase :: OutExpr -> OutId -> OutType
1123 -> [OutAlt] -- Increasing order
1126 mkCase scrut case_bndr ty alts
1127 = getDOptsSmpl `thenSmpl` \dflags ->
1128 mkAlts dflags scrut case_bndr alts `thenSmpl` \ better_alts ->
1129 mkCase1 scrut case_bndr ty better_alts
1133 mkAlts tries these things:
1135 1. If several alternatives are identical, merge them into
1136 a single DEFAULT alternative. I've occasionally seen this
1137 making a big difference:
1139 case e of =====> case e of
1140 C _ -> f x D v -> ....v....
1141 D v -> ....v.... DEFAULT -> f x
1144 The point is that we merge common RHSs, at least for the DEFAULT case.
1145 [One could do something more elaborate but I've never seen it needed.]
1146 To avoid an expensive test, we just merge branches equal to the *first*
1147 alternative; this picks up the common cases
1148 a) all branches equal
1149 b) some branches equal to the DEFAULT (which occurs first)
1152 case e of b { ==> case e of b {
1153 p1 -> rhs1 p1 -> rhs1
1155 pm -> rhsm pm -> rhsm
1156 _ -> case b of b' { pn -> let b'=b in rhsn
1158 ... po -> let b'=b in rhso
1159 po -> rhso _ -> let b'=b in rhsd
1163 which merges two cases in one case when -- the default alternative of
1164 the outer case scrutises the same variable as the outer case This
1165 transformation is called Case Merging. It avoids that the same
1166 variable is scrutinised multiple times.
1169 The case where transformation (1) showed up was like this (lib/std/PrelCError.lhs):
1175 where @is@ was something like
1177 p `is` n = p /= (-1) && p == n
1179 This gave rise to a horrible sequence of cases
1186 and similarly in cascade for all the join points!
1191 --------------------------------------------------
1192 -- 1. Merge identical branches
1193 --------------------------------------------------
1194 mkAlts dflags scrut case_bndr alts@((con1,bndrs1,rhs1) : con_alts)
1195 | all isDeadBinder bndrs1, -- Remember the default
1196 length filtered_alts < length con_alts -- alternative comes first
1197 = tick (AltMerge case_bndr) `thenSmpl_`
1198 returnSmpl better_alts
1200 filtered_alts = filter keep con_alts
1201 keep (con,bndrs,rhs) = not (all isDeadBinder bndrs && rhs `cheapEqExpr` rhs1)
1202 better_alts = (DEFAULT, [], rhs1) : filtered_alts
1205 --------------------------------------------------
1206 -- 2. Merge nested cases
1207 --------------------------------------------------
1209 mkAlts dflags scrut outer_bndr outer_alts
1210 | dopt Opt_CaseMerge dflags,
1211 (outer_alts_without_deflt, maybe_outer_deflt) <- findDefault outer_alts,
1212 Just (Case (Var scrut_var) inner_bndr _ inner_alts) <- maybe_outer_deflt,
1213 scruting_same_var scrut_var
1215 munged_inner_alts = [(con, args, munge_rhs rhs) | (con, args, rhs) <- inner_alts]
1216 munge_rhs rhs = bindCaseBndr inner_bndr (Var outer_bndr) rhs
1218 new_alts = mergeAlts outer_alts_without_deflt munged_inner_alts
1219 -- The merge keeps the inner DEFAULT at the front, if there is one
1220 -- and eliminates any inner_alts that are shadowed by the outer_alts
1222 tick (CaseMerge outer_bndr) `thenSmpl_`
1224 -- Warning: don't call mkAlts recursively!
1225 -- Firstly, there's no point, because inner alts have already had
1226 -- mkCase applied to them, so they won't have a case in their default
1227 -- Secondly, if you do, you get an infinite loop, because the bindCaseBndr
1228 -- in munge_rhs may put a case into the DEFAULT branch!
1230 -- We are scrutinising the same variable if it's
1231 -- the outer case-binder, or if the outer case scrutinises a variable
1232 -- (and it's the same). Testing both allows us not to replace the
1233 -- outer scrut-var with the outer case-binder (Simplify.simplCaseBinder).
1234 scruting_same_var = case scrut of
1235 Var outer_scrut -> \ v -> v == outer_bndr || v == outer_scrut
1236 other -> \ v -> v == outer_bndr
1238 ------------------------------------------------
1240 ------------------------------------------------
1242 mkAlts dflags scrut case_bndr other_alts = returnSmpl other_alts
1247 =================================================================================
1249 mkCase1 tries these things
1251 1. Eliminate the case altogether if possible
1259 and similar friends.
1262 Start with a simple situation:
1264 case x# of ===> e[x#/y#]
1267 (when x#, y# are of primitive type, of course). We can't (in general)
1268 do this for algebraic cases, because we might turn bottom into
1271 Actually, we generalise this idea to look for a case where we're
1272 scrutinising a variable, and we know that only the default case can
1277 other -> ...(case x of
1281 Here the inner case can be eliminated. This really only shows up in
1282 eliminating error-checking code.
1284 We also make sure that we deal with this very common case:
1289 Here we are using the case as a strict let; if x is used only once
1290 then we want to inline it. We have to be careful that this doesn't
1291 make the program terminate when it would have diverged before, so we
1293 - x is used strictly, or
1294 - e is already evaluated (it may so if e is a variable)
1296 Lastly, we generalise the transformation to handle this:
1302 We only do this for very cheaply compared r's (constructors, literals
1303 and variables). If pedantic bottoms is on, we only do it when the
1304 scrutinee is a PrimOp which can't fail.
1306 We do it *here*, looking at un-simplified alternatives, because we
1307 have to check that r doesn't mention the variables bound by the
1308 pattern in each alternative, so the binder-info is rather useful.
1310 So the case-elimination algorithm is:
1312 1. Eliminate alternatives which can't match
1314 2. Check whether all the remaining alternatives
1315 (a) do not mention in their rhs any of the variables bound in their pattern
1316 and (b) have equal rhss
1318 3. Check we can safely ditch the case:
1319 * PedanticBottoms is off,
1320 or * the scrutinee is an already-evaluated variable
1321 or * the scrutinee is a primop which is ok for speculation
1322 -- ie we want to preserve divide-by-zero errors, and
1323 -- calls to error itself!
1325 or * [Prim cases] the scrutinee is a primitive variable
1327 or * [Alg cases] the scrutinee is a variable and
1328 either * the rhs is the same variable
1329 (eg case x of C a b -> x ===> x)
1330 or * there is only one alternative, the default alternative,
1331 and the binder is used strictly in its scope.
1332 [NB this is helped by the "use default binder where
1333 possible" transformation; see below.]
1336 If so, then we can replace the case with one of the rhss.
1338 Further notes about case elimination
1339 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
1340 Consider: test :: Integer -> IO ()
1343 Turns out that this compiles to:
1346 eta1 :: State# RealWorld ->
1347 case PrelNum.< eta PrelNum.zeroInteger of wild { __DEFAULT ->
1349 (PrelNum.jtos eta ($w[] @ Char))
1351 of wild1 { (# new_s, a4 #) -> PrelIO.lvl23 new_s }}
1353 Notice the strange '<' which has no effect at all. This is a funny one.
1354 It started like this:
1356 f x y = if x < 0 then jtos x
1357 else if y==0 then "" else jtos x
1359 At a particular call site we have (f v 1). So we inline to get
1361 if v < 0 then jtos x
1362 else if 1==0 then "" else jtos x
1364 Now simplify the 1==0 conditional:
1366 if v<0 then jtos v else jtos v
1368 Now common-up the two branches of the case:
1370 case (v<0) of DEFAULT -> jtos v
1372 Why don't we drop the case? Because it's strict in v. It's technically
1373 wrong to drop even unnecessary evaluations, and in practice they
1374 may be a result of 'seq' so we *definitely* don't want to drop those.
1375 I don't really know how to improve this situation.
1379 --------------------------------------------------
1380 -- 0. Check for empty alternatives
1381 --------------------------------------------------
1383 -- This isn't strictly an error. It's possible that the simplifer might "see"
1384 -- that an inner case has no accessible alternatives before it "sees" that the
1385 -- entire branch of an outer case is inaccessible. So we simply
1386 -- put an error case here insteadd
1387 mkCase1 scrut case_bndr ty []
1388 = pprTrace "mkCase1: null alts" (ppr case_bndr <+> ppr scrut) $
1389 return (mkApps (Var eRROR_ID)
1390 [Type ty, Lit (mkStringLit "Impossible alternative")])
1392 --------------------------------------------------
1393 -- 1. Eliminate the case altogether if poss
1394 --------------------------------------------------
1396 mkCase1 scrut case_bndr ty [(con,bndrs,rhs)]
1397 -- See if we can get rid of the case altogether
1398 -- See the extensive notes on case-elimination above
1399 -- mkCase made sure that if all the alternatives are equal,
1400 -- then there is now only one (DEFAULT) rhs
1401 | all isDeadBinder bndrs,
1403 -- Check that the scrutinee can be let-bound instead of case-bound
1404 exprOkForSpeculation scrut
1405 -- OK not to evaluate it
1406 -- This includes things like (==# a# b#)::Bool
1407 -- so that we simplify
1408 -- case ==# a# b# of { True -> x; False -> x }
1411 -- This particular example shows up in default methods for
1412 -- comparision operations (e.g. in (>=) for Int.Int32)
1413 || exprIsHNF scrut -- It's already evaluated
1414 || var_demanded_later scrut -- It'll be demanded later
1416 -- || not opt_SimplPedanticBottoms) -- Or we don't care!
1417 -- We used to allow improving termination by discarding cases, unless -fpedantic-bottoms was on,
1418 -- but that breaks badly for the dataToTag# primop, which relies on a case to evaluate
1419 -- its argument: case x of { y -> dataToTag# y }
1420 -- Here we must *not* discard the case, because dataToTag# just fetches the tag from
1421 -- the info pointer. So we'll be pedantic all the time, and see if that gives any
1423 -- Also we don't want to discard 'seq's
1424 = tick (CaseElim case_bndr) `thenSmpl_`
1425 returnSmpl (bindCaseBndr case_bndr scrut rhs)
1428 -- The case binder is going to be evaluated later,
1429 -- and the scrutinee is a simple variable
1430 var_demanded_later (Var v) = isStrictDmd (idNewDemandInfo case_bndr)
1431 var_demanded_later other = False
1434 --------------------------------------------------
1436 --------------------------------------------------
1438 mkCase1 scrut case_bndr ty alts -- Identity case
1439 | all identity_alt alts
1440 = tick (CaseIdentity case_bndr) `thenSmpl_`
1441 returnSmpl (re_note scrut)
1443 identity_alt (con, args, rhs) = de_note rhs `cheapEqExpr` identity_rhs con args
1445 identity_rhs (DataAlt con) args = mkConApp con (arg_tys ++ map varToCoreExpr args)
1446 identity_rhs (LitAlt lit) _ = Lit lit
1447 identity_rhs DEFAULT _ = Var case_bndr
1449 arg_tys = map Type (tyConAppArgs (idType case_bndr))
1452 -- case coerce T e of x { _ -> coerce T' x }
1453 -- And we definitely want to eliminate this case!
1454 -- So we throw away notes from the RHS, and reconstruct
1455 -- (at least an approximation) at the other end
1456 de_note (Note _ e) = de_note e
1459 -- re_note wraps a coerce if it might be necessary
1460 re_note scrut = case head alts of
1461 (_,_,rhs1@(Note _ _)) -> mkCoerce2 (exprType rhs1) (idType case_bndr) scrut
1465 --------------------------------------------------
1467 --------------------------------------------------
1468 mkCase1 scrut bndr ty alts = returnSmpl (Case scrut bndr ty alts)
1472 When adding auxiliary bindings for the case binder, it's worth checking if
1473 its dead, because it often is, and occasionally these mkCase transformations
1474 cascade rather nicely.
1477 bindCaseBndr bndr rhs body
1478 | isDeadBinder bndr = body
1479 | otherwise = bindNonRec bndr rhs body