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 CoreExpr -- The argument
83 (Maybe SimplEnv) -- (Just se) => the arg is un-simplified and this is its subst-env
84 -- Nothing => the arg is already simplified; don't repeatedly simplify it!
85 SimplCont -- and its environment
88 InId [InAlt] SimplEnv -- The case binder, alts, and subst-env
91 | ArgOf LetRhsFlag -- An arbitrary strict context: the argument
92 -- of a strict function, or a primitive-arg fn
94 -- No DupFlag, because we never duplicate it
95 OutType -- arg_ty: type of the argument itself
96 OutType -- cont_ty: the type of the expression being sought by the context
97 -- f (error "foo") ==> coerce t (error "foo")
99 -- We need to know the type t, to which to coerce.
101 (SimplEnv -> OutExpr -> SimplM FloatsWithExpr) -- What to do with the result
102 -- The result expression in the OutExprStuff has type cont_ty
104 data LetRhsFlag = AnArg -- It's just an argument not a let RHS
105 | AnRhs -- It's the RHS of a let (so please float lets out of big lambdas)
107 instance Outputable LetRhsFlag where
108 ppr AnArg = ptext SLIT("arg")
109 ppr AnRhs = ptext SLIT("rhs")
111 instance Outputable SimplCont where
112 ppr (Stop ty is_rhs _) = ptext SLIT("Stop") <> brackets (ppr is_rhs) <+> ppr ty
113 ppr (ApplyTo dup arg se cont) = (ptext SLIT("ApplyTo") <+> ppr dup <+> ppr arg) $$ ppr cont
114 ppr (ArgOf _ _ _ _) = ptext SLIT("ArgOf...")
115 ppr (Select dup bndr alts se cont) = (ptext SLIT("Select") <+> ppr dup <+> ppr bndr) $$
116 (nest 4 (ppr alts)) $$ ppr cont
117 ppr (CoerceIt ty cont) = (ptext SLIT("CoerceIt") <+> ppr ty) $$ ppr cont
119 data DupFlag = OkToDup | NoDup
121 instance Outputable DupFlag where
122 ppr OkToDup = ptext SLIT("ok")
123 ppr NoDup = ptext SLIT("nodup")
127 mkBoringStop :: OutType -> SimplCont
128 mkBoringStop ty = Stop ty AnArg False
130 mkLazyArgStop :: OutType -> Bool -> SimplCont
131 mkLazyArgStop ty has_rules = Stop ty AnArg (canUpdateInPlace ty || has_rules)
133 mkRhsStop :: OutType -> SimplCont
134 mkRhsStop ty = Stop ty AnRhs (canUpdateInPlace ty)
136 contIsRhs :: SimplCont -> Bool
137 contIsRhs (Stop _ AnRhs _) = True
138 contIsRhs (ArgOf AnRhs _ _ _) = True
139 contIsRhs other = False
141 contIsRhsOrArg (Stop _ _ _) = True
142 contIsRhsOrArg (ArgOf _ _ _ _) = True
143 contIsRhsOrArg other = False
146 contIsDupable :: SimplCont -> Bool
147 contIsDupable (Stop _ _ _) = True
148 contIsDupable (ApplyTo OkToDup _ _ _) = True
149 contIsDupable (Select OkToDup _ _ _ _) = True
150 contIsDupable (CoerceIt _ cont) = contIsDupable cont
151 contIsDupable other = False
154 discardableCont :: SimplCont -> Bool
155 discardableCont (Stop _ _ _) = False
156 discardableCont (CoerceIt _ cont) = discardableCont cont
157 discardableCont other = True
159 discardCont :: SimplCont -- A continuation, expecting
160 -> SimplCont -- Replace the continuation with a suitable coerce
161 discardCont cont = case cont of
162 Stop to_ty is_rhs _ -> cont
163 other -> CoerceIt to_ty (mkBoringStop to_ty)
165 to_ty = contResultType cont
168 contResultType :: SimplCont -> OutType
169 contResultType (Stop to_ty _ _) = to_ty
170 contResultType (ArgOf _ _ to_ty _) = to_ty
171 contResultType (ApplyTo _ _ _ cont) = contResultType cont
172 contResultType (CoerceIt _ cont) = contResultType cont
173 contResultType (Select _ _ _ _ cont) = contResultType cont
176 countValArgs :: SimplCont -> Int
177 countValArgs (ApplyTo _ (Type ty) se cont) = countValArgs cont
178 countValArgs (ApplyTo _ val_arg se cont) = 1 + countValArgs cont
179 countValArgs other = 0
181 countArgs :: SimplCont -> Int
182 countArgs (ApplyTo _ arg se cont) = 1 + countArgs cont
186 pushContArgs ::[OutArg] -> SimplCont -> SimplCont
187 -- Pushes args with the specified environment
188 pushContArgs [] cont = cont
189 pushContArgs (arg : args) cont = ApplyTo NoDup arg Nothing (pushContArgs args cont)
194 getContArgs :: SwitchChecker
195 -> OutId -> SimplCont
196 -> ([(InExpr, Maybe SimplEnv, Bool)], -- Arguments; the Bool is true for strict args
197 SimplCont) -- Remaining continuation
198 -- getContArgs id k = (args, k', inl)
199 -- args are the leading ApplyTo items in k
200 -- (i.e. outermost comes first)
201 -- augmented with demand info from the functionn
202 getContArgs chkr fun orig_cont
204 -- Ignore strictness info if the no-case-of-case
205 -- flag is on. Strictness changes evaluation order
206 -- and that can change full laziness
207 stricts | switchIsOn chkr NoCaseOfCase = vanilla_stricts
208 | otherwise = computed_stricts
210 go [] stricts orig_cont
212 ----------------------------
215 go acc ss (ApplyTo _ arg@(Type _) se cont)
216 = go ((arg,se,False) : acc) ss cont
217 -- NB: don't bother to instantiate the function type
220 go acc (s:ss) (ApplyTo _ arg se cont)
221 = go ((arg,se,s) : acc) ss cont
223 -- We're run out of arguments, or else we've run out of demands
224 -- The latter only happens if the result is guaranteed bottom
225 -- This is the case for
226 -- * case (error "hello") of { ... }
227 -- * (error "Hello") arg
228 -- * f (error "Hello") where f is strict
230 -- Then, especially in the first of these cases, we'd like to discard
231 -- the continuation, leaving just the bottoming expression. But the
232 -- type might not be right, so we may have to add a coerce.
234 | null ss && discardableCont cont = (reverse acc, discardCont cont)
235 | otherwise = (reverse acc, cont)
237 ----------------------------
238 vanilla_stricts, computed_stricts :: [Bool]
239 vanilla_stricts = repeat False
240 computed_stricts = zipWith (||) fun_stricts arg_stricts
242 ----------------------------
243 (val_arg_tys, _) = splitFunTys (dropForAlls (idType fun))
244 arg_stricts = map isStrictType val_arg_tys ++ repeat False
245 -- These argument types are used as a cheap and cheerful way to find
246 -- unboxed arguments, which must be strict. But it's an InType
247 -- and so there might be a type variable where we expect a function
248 -- type (the substitution hasn't happened yet). And we don't bother
249 -- doing the type applications for a polymorphic function.
250 -- Hence the splitFunTys*IgnoringForAlls*
252 ----------------------------
253 -- If fun_stricts is finite, it means the function returns bottom
254 -- after that number of value args have been consumed
255 -- Otherwise it's infinite, extended with False
257 = case splitStrictSig (idNewStrictness fun) of
258 (demands, result_info)
259 | not (demands `lengthExceeds` countValArgs orig_cont)
260 -> -- Enough args, use the strictness given.
261 -- For bottoming functions we used to pretend that the arg
262 -- is lazy, so that we don't treat the arg as an
263 -- interesting context. This avoids substituting
264 -- top-level bindings for (say) strings into
265 -- calls to error. But now we are more careful about
266 -- inlining lone variables, so its ok (see SimplUtils.analyseCont)
267 if isBotRes result_info then
268 map isStrictDmd demands -- Finite => result is bottom
270 map isStrictDmd demands ++ vanilla_stricts
272 other -> vanilla_stricts -- Not enough args, or no strictness
275 interestingArg :: OutExpr -> Bool
276 -- An argument is interesting if it has *some* structure
277 -- We are here trying to avoid unfolding a function that
278 -- is applied only to variables that have no unfolding
279 -- (i.e. they are probably lambda bound): f x y z
280 -- There is little point in inlining f here.
281 interestingArg (Var v) = hasSomeUnfolding (idUnfolding v)
282 -- Was: isValueUnfolding (idUnfolding v')
283 -- But that seems over-pessimistic
285 -- This accounts for an argument like
286 -- () or [], which is definitely interesting
287 interestingArg (Type _) = False
288 interestingArg (App fn (Type _)) = interestingArg fn
289 interestingArg (Note _ a) = interestingArg a
290 interestingArg other = True
291 -- Consider let x = 3 in f x
292 -- The substitution will contain (x -> ContEx 3), and we want to
293 -- to say that x is an interesting argument.
294 -- But consider also (\x. f x y) y
295 -- The substitution will contain (x -> ContEx y), and we want to say
296 -- that x is not interesting (assuming y has no unfolding)
299 Comment about interestingCallContext
300 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
301 We want to avoid inlining an expression where there can't possibly be
302 any gain, such as in an argument position. Hence, if the continuation
303 is interesting (eg. a case scrutinee, application etc.) then we
304 inline, otherwise we don't.
306 Previously some_benefit used to return True only if the variable was
307 applied to some value arguments. This didn't work:
309 let x = _coerce_ (T Int) Int (I# 3) in
310 case _coerce_ Int (T Int) x of
313 we want to inline x, but can't see that it's a constructor in a case
314 scrutinee position, and some_benefit is False.
318 dMonadST = _/\_ t -> :Monad (g1 _@_ t, g2 _@_ t, g3 _@_ t)
320 .... case dMonadST _@_ x0 of (a,b,c) -> ....
322 we'd really like to inline dMonadST here, but we *don't* want to
323 inline if the case expression is just
325 case x of y { DEFAULT -> ... }
327 since we can just eliminate this case instead (x is in WHNF). Similar
328 applies when x is bound to a lambda expression. Hence
329 contIsInteresting looks for case expressions with just a single
333 interestingCallContext :: Bool -- False <=> no args at all
334 -> Bool -- False <=> no value args
336 -- The "lone-variable" case is important. I spent ages
337 -- messing about with unsatisfactory varaints, but this is nice.
338 -- The idea is that if a variable appear all alone
339 -- as an arg of lazy fn, or rhs Stop
340 -- as scrutinee of a case Select
341 -- as arg of a strict fn ArgOf
342 -- then we should not inline it (unless there is some other reason,
343 -- e.g. is is the sole occurrence). We achieve this by making
344 -- interestingCallContext return False for a lone variable.
346 -- Why? At least in the case-scrutinee situation, turning
347 -- let x = (a,b) in case x of y -> ...
349 -- let x = (a,b) in case (a,b) of y -> ...
351 -- let x = (a,b) in let y = (a,b) in ...
352 -- is bad if the binding for x will remain.
354 -- Another example: I discovered that strings
355 -- were getting inlined straight back into applications of 'error'
356 -- because the latter is strict.
358 -- f = \x -> ...(error s)...
360 -- Fundamentally such contexts should not ecourage inlining because
361 -- the context can ``see'' the unfolding of the variable (e.g. case or a RULE)
362 -- so there's no gain.
364 -- However, even a type application or coercion isn't a lone variable.
366 -- case $fMonadST @ RealWorld of { :DMonad a b c -> c }
367 -- We had better inline that sucker! The case won't see through it.
369 -- For now, I'm treating treating a variable applied to types
370 -- in a *lazy* context "lone". The motivating example was
372 -- g = /\a. \y. h (f a)
373 -- There's no advantage in inlining f here, and perhaps
374 -- a significant disadvantage. Hence some_val_args in the Stop case
376 interestingCallContext some_args some_val_args cont
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 _ interesting) = some_val_args && interesting
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 interestingArgContext :: Id -> SimplCont -> Bool
405 -- If the argument has form (f x y), where x,y are boring,
406 -- and f is marked INLINE, then we don't want to inline f.
407 -- But if the context of the argument is
409 -- where g has rules, then we *do* want to inline f, in case it
410 -- exposes a rule that might fire. Similarly, if the context is
412 -- where h has rules, then we do want to inline f.
413 -- The interesting_arg_ctxt flag makes this happen; if it's
414 -- set, the inliner gets just enough keener to inline f
415 -- regardless of how boring f's arguments are, if it's marked INLINE
417 -- The alternative would be to *always* inline an INLINE function,
418 -- regardless of how boring its context is; but that seems overkill
419 -- For example, it'd mean that wrapper functions were always inlined
420 interestingArgContext fn cont
421 = idHasRules fn || go cont
423 go (Select {}) = False
424 go (ApplyTo {}) = False
426 go (CoerceIt _ c) = go c
427 go (Stop _ _ interesting) = interesting
430 canUpdateInPlace :: Type -> Bool
431 -- Consider let x = <wurble> in ...
432 -- If <wurble> returns an explicit constructor, we might be able
433 -- to do update in place. So we treat even a thunk RHS context
434 -- as interesting if update in place is possible. We approximate
435 -- this by seeing if the type has a single constructor with a
436 -- small arity. But arity zero isn't good -- we share the single copy
437 -- for that case, so no point in sharing.
440 | not opt_UF_UpdateInPlace = False
442 = case splitTyConApp_maybe ty of
444 Just (tycon, _) -> case tyConDataCons_maybe tycon of
445 Just [dc] -> arity == 1 || arity == 2
447 arity = dataConRepArity dc
453 %************************************************************************
455 \subsection{Decisions about inlining}
457 %************************************************************************
459 Inlining is controlled partly by the SimplifierMode switch. This has two
462 SimplGently (a) Simplifying before specialiser/full laziness
463 (b) Simplifiying inside INLINE pragma
464 (c) Simplifying the LHS of a rule
465 (d) Simplifying a GHCi expression or Template
468 SimplPhase n Used at all other times
470 The key thing about SimplGently is that it does no call-site inlining.
471 Before full laziness we must be careful not to inline wrappers,
472 because doing so inhibits floating
473 e.g. ...(case f x of ...)...
474 ==> ...(case (case x of I# x# -> fw x#) of ...)...
475 ==> ...(case x of I# x# -> case fw x# of ...)...
476 and now the redex (f x) isn't floatable any more.
478 The no-inlining thing is also important for Template Haskell. You might be
479 compiling in one-shot mode with -O2; but when TH compiles a splice before
480 running it, we don't want to use -O2. Indeed, we don't want to inline
481 anything, because the byte-code interpreter might get confused about
482 unboxed tuples and suchlike.
486 SimplGently is also used as the mode to simplify inside an InlineMe note.
489 inlineMode :: SimplifierMode
490 inlineMode = SimplGently
493 It really is important to switch off inlinings inside such
494 expressions. Consider the following example
500 in ...g...g...g...g...g...
502 Now, if that's the ONLY occurrence of f, it will be inlined inside g,
503 and thence copied multiple times when g is inlined.
506 This function may be inlinined in other modules, so we
507 don't want to remove (by inlining) calls to functions that have
508 specialisations, or that may have transformation rules in an importing
511 E.g. {-# INLINE f #-}
514 and suppose that g is strict *and* has specialisations. If we inline
515 g's wrapper, we deny f the chance of getting the specialised version
516 of g when f is inlined at some call site (perhaps in some other
519 It's also important not to inline a worker back into a wrapper.
521 wraper = inline_me (\x -> ...worker... )
522 Normally, the inline_me prevents the worker getting inlined into
523 the wrapper (initially, the worker's only call site!). But,
524 if the wrapper is sure to be called, the strictness analyser will
525 mark it 'demanded', so when the RHS is simplified, it'll get an ArgOf
526 continuation. That's why the keep_inline predicate returns True for
527 ArgOf continuations. It shouldn't do any harm not to dissolve the
528 inline-me note under these circumstances.
530 Note that the result is that we do very little simplification
533 all xs = foldr (&&) True xs
534 any p = all . map p {-# INLINE any #-}
536 Problem: any won't get deforested, and so if it's exported and the
537 importer doesn't use the inlining, (eg passes it as an arg) then we
538 won't get deforestation at all. We havn't solved this problem yet!
541 preInlineUnconditionally
542 ~~~~~~~~~~~~~~~~~~~~~~~~
543 @preInlineUnconditionally@ examines a bndr to see if it is used just
544 once in a completely safe way, so that it is safe to discard the
545 binding inline its RHS at the (unique) usage site, REGARDLESS of how
546 big the RHS might be. If this is the case we don't simplify the RHS
547 first, but just inline it un-simplified.
549 This is much better than first simplifying a perhaps-huge RHS and then
550 inlining and re-simplifying it. Indeed, it can be at least quadratically
559 We may end up simplifying e1 N times, e2 N-1 times, e3 N-3 times etc.
560 This can happen with cascades of functions too:
567 THE MAIN INVARIANT is this:
569 ---- preInlineUnconditionally invariant -----
570 IF preInlineUnconditionally chooses to inline x = <rhs>
571 THEN doing the inlining should not change the occurrence
572 info for the free vars of <rhs>
573 ----------------------------------------------
575 For example, it's tempting to look at trivial binding like
577 and inline it unconditionally. But suppose x is used many times,
578 but this is the unique occurrence of y. Then inlining x would change
579 y's occurrence info, which breaks the invariant. It matters: y
580 might have a BIG rhs, which will now be dup'd at every occurrenc of x.
583 Evne RHSs labelled InlineMe aren't caught here, because there might be
584 no benefit from inlining at the call site.
586 [Sept 01] Don't unconditionally inline a top-level thing, because that
587 can simply make a static thing into something built dynamically. E.g.
591 [Remember that we treat \s as a one-shot lambda.] No point in
592 inlining x unless there is something interesting about the call site.
594 But watch out: if you aren't careful, some useful foldr/build fusion
595 can be lost (most notably in spectral/hartel/parstof) because the
596 foldr didn't see the build. Doing the dynamic allocation isn't a big
597 deal, in fact, but losing the fusion can be. But the right thing here
598 seems to be to do a callSiteInline based on the fact that there is
599 something interesting about the call site (it's strict). Hmm. That
602 Conclusion: inline top level things gaily until Phase 0 (the last
603 phase), at which point don't.
606 preInlineUnconditionally :: SimplEnv -> TopLevelFlag -> InId -> InExpr -> Bool
607 preInlineUnconditionally env top_lvl bndr rhs
609 | opt_SimplNoPreInlining = False
610 | otherwise = case idOccInfo bndr of
611 IAmDead -> True -- Happens in ((\x.1) v)
612 OneOcc in_lam True int_cxt -> try_once in_lam int_cxt
616 active = case phase of
617 SimplGently -> isAlwaysActive prag
618 SimplPhase n -> isActive n prag
619 prag = idInlinePragma bndr
621 try_once in_lam int_cxt -- There's one textual occurrence
622 | not in_lam = isNotTopLevel top_lvl || early_phase
623 | otherwise = int_cxt && canInlineInLam rhs
625 -- Be very careful before inlining inside a lambda, becuase (a) we must not
626 -- invalidate occurrence information, and (b) we want to avoid pushing a
627 -- single allocation (here) into multiple allocations (inside lambda).
628 -- Inlining a *function* with a single *saturated* call would be ok, mind you.
629 -- || (if is_cheap && not (canInlineInLam rhs) then pprTrace "preinline" (ppr bndr <+> ppr rhs) ok else ok)
631 -- is_cheap = exprIsCheap rhs
632 -- ok = is_cheap && int_cxt
634 -- int_cxt The context isn't totally boring
635 -- E.g. let f = \ab.BIG in \y. map f xs
636 -- Don't want to substitute for f, because then we allocate
637 -- its closure every time the \y is called
638 -- But: let f = \ab.BIG in \y. map (f y) xs
639 -- Now we do want to substitute for f, even though it's not
640 -- saturated, because we're going to allocate a closure for
641 -- (f y) every time round the loop anyhow.
643 -- canInlineInLam => free vars of rhs are (Once in_lam) or Many,
644 -- so substituting rhs inside a lambda doesn't change the occ info.
645 -- Sadly, not quite the same as exprIsHNF.
646 canInlineInLam (Lit l) = True
647 canInlineInLam (Lam b e) = isRuntimeVar b || canInlineInLam e
648 canInlineInLam (Note _ e) = canInlineInLam e
649 canInlineInLam _ = False
651 early_phase = case phase of
652 SimplPhase 0 -> False
654 -- If we don't have this early_phase test, consider
655 -- x = length [1,2,3]
656 -- The full laziness pass carefully floats all the cons cells to
657 -- top level, and preInlineUnconditionally floats them all back in.
658 -- Result is (a) static allocation replaced by dynamic allocation
659 -- (b) many simplifier iterations because this tickles
660 -- a related problem; only one inlining per pass
662 -- On the other hand, I have seen cases where top-level fusion is
663 -- lost if we don't inline top level thing (e.g. string constants)
664 -- Hence the test for phase zero (which is the phase for all the final
665 -- simplifications). Until phase zero we take no special notice of
666 -- top level things, but then we become more leery about inlining
671 postInlineUnconditionally
672 ~~~~~~~~~~~~~~~~~~~~~~~~~
673 @postInlineUnconditionally@ decides whether to unconditionally inline
674 a thing based on the form of its RHS; in particular if it has a
675 trivial RHS. If so, we can inline and discard the binding altogether.
677 NB: a loop breaker has must_keep_binding = True and non-loop-breakers
678 only have *forward* references Hence, it's safe to discard the binding
680 NOTE: This isn't our last opportunity to inline. We're at the binding
681 site right now, and we'll get another opportunity when we get to the
684 Note that we do this unconditional inlining only for trival RHSs.
685 Don't inline even WHNFs inside lambdas; doing so may simply increase
686 allocation when the function is called. This isn't the last chance; see
689 NB: Even inline pragmas (e.g. IMustBeINLINEd) are ignored here Why?
690 Because we don't even want to inline them into the RHS of constructor
691 arguments. See NOTE above
693 NB: At one time even NOINLINE was ignored here: if the rhs is trivial
694 it's best to inline it anyway. We often get a=E; b=a from desugaring,
695 with both a and b marked NOINLINE. But that seems incompatible with
696 our new view that inlining is like a RULE, so I'm sticking to the 'active'
700 postInlineUnconditionally
701 :: SimplEnv -> TopLevelFlag
702 -> InId -- The binder (an OutId would be fine too)
703 -> OccInfo -- From the InId
707 postInlineUnconditionally env top_lvl bndr occ_info rhs unfolding
709 | isLoopBreaker occ_info = False
710 | isExportedId bndr = False
711 | exprIsTrivial rhs = True
714 -- The point of examining occ_info here is that for *non-values*
715 -- that occur outside a lambda, the call-site inliner won't have
716 -- a chance (becuase it doesn't know that the thing
717 -- only occurs once). The pre-inliner won't have gotten
718 -- it either, if the thing occurs in more than one branch
719 -- So the main target is things like
722 -- True -> case x of ...
723 -- False -> case x of ...
724 -- I'm not sure how important this is in practice
725 OneOcc in_lam one_br int_cxt -- OneOcc => no work-duplication issue
726 -> smallEnoughToInline unfolding -- Small enough to dup
727 -- ToDo: consider discount on smallEnoughToInline if int_cxt is true
729 -- NB: Do NOT inline arbitrarily big things, even if one_br is True
730 -- Reason: doing so risks exponential behaviour. We simplify a big
731 -- expression, inline it, and simplify it again. But if the
732 -- very same thing happens in the big expression, we get
734 -- PRINCIPLE: when we've already simplified an expression once,
735 -- make sure that we only inline it if it's reasonably small.
737 && ((isNotTopLevel top_lvl && not in_lam) ||
738 -- But outside a lambda, we want to be reasonably aggressive
739 -- about inlining into multiple branches of case
740 -- e.g. let x = <non-value>
741 -- in case y of { C1 -> ..x..; C2 -> ..x..; C3 -> ... }
742 -- Inlining can be a big win if C3 is the hot-spot, even if
743 -- the uses in C1, C2 are not 'interesting'
744 -- An example that gets worse if you add int_cxt here is 'clausify'
746 (isCheapUnfolding unfolding && int_cxt))
747 -- isCheap => acceptable work duplication; in_lam may be true
748 -- int_cxt to prevent us inlining inside a lambda without some
749 -- good reason. See the notes on int_cxt in preInlineUnconditionally
751 IAmDead -> True -- This happens; for example, the case_bndr during case of
752 -- known constructor: case (a,b) of x { (p,q) -> ... }
753 -- Here x isn't mentioned in the RHS, so we don't want to
754 -- create the (dead) let-binding let x = (a,b) in ...
758 -- Here's an example that we don't handle well:
759 -- let f = if b then Left (\x.BIG) else Right (\y.BIG)
760 -- in \y. ....case f of {...} ....
761 -- Here f is used just once, and duplicating the case work is fine (exprIsCheap).
763 -- * We can't preInlineUnconditionally because that woud invalidate
764 -- the occ info for b.
765 -- * We can't postInlineUnconditionally because the RHS is big, and
766 -- that risks exponential behaviour
767 -- * We can't call-site inline, because the rhs is big
771 active = case getMode env of
772 SimplGently -> isAlwaysActive prag
773 SimplPhase n -> isActive n prag
774 prag = idInlinePragma bndr
776 activeInline :: SimplEnv -> OutId -> OccInfo -> Bool
777 activeInline env id occ
778 = case getMode env of
779 SimplGently -> isOneOcc occ && isAlwaysActive prag
780 -- No inlining at all when doing gentle stuff,
781 -- except for local things that occur once
782 -- The reason is that too little clean-up happens if you
783 -- don't inline use-once things. Also a bit of inlining is *good* for
784 -- full laziness; it can expose constant sub-expressions.
785 -- Example in spectral/mandel/Mandel.hs, where the mandelset
786 -- function gets a useful let-float if you inline windowToViewport
788 -- NB: we used to have a second exception, for data con wrappers.
789 -- On the grounds that we use gentle mode for rule LHSs, and
790 -- they match better when data con wrappers are inlined.
791 -- But that only really applies to the trivial wrappers (like (:)),
792 -- and they are now constructed as Compulsory unfoldings (in MkId)
793 -- so they'll happen anyway.
795 SimplPhase n -> isActive n prag
797 prag = idInlinePragma id
799 activeRule :: SimplEnv -> Maybe (Activation -> Bool)
800 -- Nothing => No rules at all
802 | opt_RulesOff = Nothing
804 = case getMode env of
805 SimplGently -> Just isAlwaysActive
806 -- Used to be Nothing (no rules in gentle mode)
807 -- Main motivation for changing is that I wanted
808 -- lift String ===> ...
809 -- to work in Template Haskell when simplifying
810 -- splices, so we get simpler code for literal strings
811 SimplPhase n -> Just (isActive n)
815 %************************************************************************
817 \subsection{Rebuilding a lambda}
819 %************************************************************************
822 mkLam :: SimplEnv -> [OutBinder] -> OutExpr -> SimplCont -> SimplM FloatsWithExpr
826 a) eta reduction, if that gives a trivial expression
827 b) eta expansion [only if there are some value lambdas]
828 c) floating lets out through big lambdas
829 [only if all tyvar lambdas, and only if this lambda
833 mkLam env bndrs body cont
834 = getDOptsSmpl `thenSmpl` \dflags ->
835 mkLam' dflags env bndrs body cont
837 mkLam' dflags env bndrs body cont
838 | dopt Opt_DoEtaReduction dflags,
839 Just etad_lam <- tryEtaReduce bndrs body
840 = tick (EtaReduction (head bndrs)) `thenSmpl_`
841 returnSmpl (emptyFloats env, etad_lam)
843 | dopt Opt_DoLambdaEtaExpansion dflags,
844 any isRuntimeVar bndrs
845 = tryEtaExpansion dflags body `thenSmpl` \ body' ->
846 returnSmpl (emptyFloats env, mkLams bndrs body')
848 {- Sept 01: I'm experimenting with getting the
849 full laziness pass to float out past big lambdsa
850 | all isTyVar bndrs, -- Only for big lambdas
851 contIsRhs cont -- Only try the rhs type-lambda floating
852 -- if this is indeed a right-hand side; otherwise
853 -- we end up floating the thing out, only for float-in
854 -- to float it right back in again!
855 = tryRhsTyLam env bndrs body `thenSmpl` \ (floats, body') ->
856 returnSmpl (floats, mkLams bndrs body')
860 = returnSmpl (emptyFloats env, mkLams bndrs body)
864 %************************************************************************
866 \subsection{Eta expansion and reduction}
868 %************************************************************************
870 We try for eta reduction here, but *only* if we get all the
871 way to an exprIsTrivial expression.
872 We don't want to remove extra lambdas unless we are going
873 to avoid allocating this thing altogether
876 tryEtaReduce :: [OutBinder] -> OutExpr -> Maybe OutExpr
877 tryEtaReduce bndrs body
878 -- We don't use CoreUtils.etaReduce, because we can be more
880 -- (a) we already have the binders
881 -- (b) we can do the triviality test before computing the free vars
882 = go (reverse bndrs) body
884 go (b : bs) (App fun arg) | ok_arg b arg = go bs fun -- Loop round
885 go [] fun | ok_fun fun = Just fun -- Success!
886 go _ _ = Nothing -- Failure!
888 ok_fun fun = exprIsTrivial fun
889 && not (any (`elemVarSet` (exprFreeVars fun)) bndrs)
890 && (exprIsHNF fun || all ok_lam bndrs)
891 ok_lam v = isTyVar v || isDictId v
892 -- The exprIsHNF is because eta reduction is not
893 -- valid in general: \x. bot /= bot
894 -- So we need to be sure that the "fun" is a value.
896 -- However, we always want to reduce (/\a -> f a) to f
897 -- This came up in a RULE: foldr (build (/\a -> g a))
898 -- did not match foldr (build (/\b -> ...something complex...))
899 -- The type checker can insert these eta-expanded versions,
900 -- with both type and dictionary lambdas; hence the slightly
903 ok_arg b arg = varToCoreExpr b `cheapEqExpr` arg
907 Try eta expansion for RHSs
910 f = \x1..xn -> N ==> f = \x1..xn y1..ym -> N y1..ym
913 where (in both cases)
915 * The xi can include type variables
917 * The yi are all value variables
919 * N is a NORMAL FORM (i.e. no redexes anywhere)
920 wanting a suitable number of extra args.
922 We may have to sandwich some coerces between the lambdas
923 to make the types work. exprEtaExpandArity looks through coerces
924 when computing arity; and etaExpand adds the coerces as necessary when
925 actually computing the expansion.
928 tryEtaExpansion :: DynFlags -> OutExpr -> SimplM OutExpr
929 -- There is at least one runtime binder in the binders
930 tryEtaExpansion dflags body
931 = getUniquesSmpl `thenSmpl` \ us ->
932 returnSmpl (etaExpand fun_arity us body (exprType body))
934 fun_arity = exprEtaExpandArity dflags body
938 %************************************************************************
940 \subsection{Floating lets out of big lambdas}
942 %************************************************************************
944 tryRhsTyLam tries this transformation, when the big lambda appears as
945 the RHS of a let(rec) binding:
947 /\abc -> let(rec) x = e in b
949 let(rec) x' = /\abc -> let x = x' a b c in e
951 /\abc -> let x = x' a b c in b
953 This is good because it can turn things like:
955 let f = /\a -> letrec g = ... g ... in g
957 letrec g' = /\a -> ... g' a ...
961 which is better. In effect, it means that big lambdas don't impede
964 This optimisation is CRUCIAL in eliminating the junk introduced by
965 desugaring mutually recursive definitions. Don't eliminate it lightly!
967 So far as the implementation is concerned:
969 Invariant: go F e = /\tvs -> F e
973 = Let x' = /\tvs -> F e
977 G = F . Let x = x' tvs
979 go F (Letrec xi=ei in b)
980 = Letrec {xi' = /\tvs -> G ei}
984 G = F . Let {xi = xi' tvs}
986 [May 1999] If we do this transformation *regardless* then we can
987 end up with some pretty silly stuff. For example,
990 st = /\ s -> let { x1=r1 ; x2=r2 } in ...
995 st = /\s -> ...[y1 s/x1, y2 s/x2]
998 Unless the "..." is a WHNF there is really no point in doing this.
999 Indeed it can make things worse. Suppose x1 is used strictly,
1002 x1* = case f y of { (a,b) -> e }
1004 If we abstract this wrt the tyvar we then can't do the case inline
1005 as we would normally do.
1009 {- Trying to do this in full laziness
1011 tryRhsTyLam :: SimplEnv -> [OutTyVar] -> OutExpr -> SimplM FloatsWithExpr
1012 -- Call ensures that all the binders are type variables
1014 tryRhsTyLam env tyvars body -- Only does something if there's a let
1015 | not (all isTyVar tyvars)
1016 || not (worth_it body) -- inside a type lambda,
1017 = returnSmpl (emptyFloats env, body) -- and a WHNF inside that
1020 = go env (\x -> x) body
1023 worth_it e@(Let _ _) = whnf_in_middle e
1026 whnf_in_middle (Let (NonRec x rhs) e) | isUnLiftedType (idType x) = False
1027 whnf_in_middle (Let _ e) = whnf_in_middle e
1028 whnf_in_middle e = exprIsCheap e
1030 main_tyvar_set = mkVarSet tyvars
1032 go env fn (Let bind@(NonRec var rhs) body)
1034 = go env (fn . Let bind) body
1036 go env fn (Let (NonRec var rhs) body)
1037 = mk_poly tyvars_here var `thenSmpl` \ (var', rhs') ->
1038 addAuxiliaryBind env (NonRec var' (mkLams tyvars_here (fn rhs))) $ \ env ->
1039 go env (fn . Let (mk_silly_bind var rhs')) body
1043 tyvars_here = varSetElems (main_tyvar_set `intersectVarSet` exprSomeFreeVars isTyVar rhs)
1044 -- Abstract only over the type variables free in the rhs
1045 -- wrt which the new binding is abstracted. But the naive
1046 -- approach of abstract wrt the tyvars free in the Id's type
1048 -- /\ a b -> let t :: (a,b) = (e1, e2)
1051 -- Here, b isn't free in x's type, but we must nevertheless
1052 -- abstract wrt b as well, because t's type mentions b.
1053 -- Since t is floated too, we'd end up with the bogus:
1054 -- poly_t = /\ a b -> (e1, e2)
1055 -- poly_x = /\ a -> fst (poly_t a *b*)
1056 -- So for now we adopt the even more naive approach of
1057 -- abstracting wrt *all* the tyvars. We'll see if that
1058 -- gives rise to problems. SLPJ June 98
1060 go env fn (Let (Rec prs) body)
1061 = mapAndUnzipSmpl (mk_poly tyvars_here) vars `thenSmpl` \ (vars', rhss') ->
1063 gn body = fn (foldr Let body (zipWith mk_silly_bind vars rhss'))
1064 pairs = vars' `zip` [mkLams tyvars_here (gn rhs) | rhs <- rhss]
1066 addAuxiliaryBind env (Rec pairs) $ \ env ->
1069 (vars,rhss) = unzip prs
1070 tyvars_here = varSetElems (main_tyvar_set `intersectVarSet` exprsSomeFreeVars isTyVar (map snd prs))
1071 -- See notes with tyvars_here above
1073 go env fn body = returnSmpl (emptyFloats env, fn body)
1075 mk_poly tyvars_here var
1076 = getUniqueSmpl `thenSmpl` \ uniq ->
1078 poly_name = setNameUnique (idName var) uniq -- Keep same name
1079 poly_ty = mkForAllTys tyvars_here (idType var) -- But new type of course
1080 poly_id = mkLocalId poly_name poly_ty
1082 -- In the olden days, it was crucial to copy the occInfo of the original var,
1083 -- because we were looking at occurrence-analysed but as yet unsimplified code!
1084 -- In particular, we mustn't lose the loop breakers. BUT NOW we are looking
1085 -- at already simplified code, so it doesn't matter
1087 -- It's even right to retain single-occurrence or dead-var info:
1088 -- Suppose we started with /\a -> let x = E in B
1089 -- where x occurs once in B. Then we transform to:
1090 -- let x' = /\a -> E in /\a -> let x* = x' a in B
1091 -- where x* has an INLINE prag on it. Now, once x* is inlined,
1092 -- the occurrences of x' will be just the occurrences originally
1095 returnSmpl (poly_id, mkTyApps (Var poly_id) (mkTyVarTys tyvars_here))
1097 mk_silly_bind var rhs = NonRec var (Note InlineMe rhs)
1098 -- Suppose we start with:
1100 -- x = /\ a -> let g = G in E
1102 -- Then we'll float to get
1104 -- x = let poly_g = /\ a -> G
1105 -- in /\ a -> let g = poly_g a in E
1107 -- But now the occurrence analyser will see just one occurrence
1108 -- of poly_g, not inside a lambda, so the simplifier will
1109 -- PreInlineUnconditionally poly_g back into g! Badk to square 1!
1110 -- (I used to think that the "don't inline lone occurrences" stuff
1111 -- would stop this happening, but since it's the *only* occurrence,
1112 -- PreInlineUnconditionally kicks in first!)
1114 -- Solution: put an INLINE note on g's RHS, so that poly_g seems
1115 -- to appear many times. (NB: mkInlineMe eliminates
1116 -- such notes on trivial RHSs, so do it manually.)
1120 %************************************************************************
1122 \subsection{Case absorption and identity-case elimination}
1124 %************************************************************************
1126 mkCase puts a case expression back together, trying various transformations first.
1129 mkCase :: OutExpr -> OutId -> OutType
1130 -> [OutAlt] -- Increasing order
1133 mkCase scrut case_bndr ty alts
1134 = getDOptsSmpl `thenSmpl` \dflags ->
1135 mkAlts dflags scrut case_bndr alts `thenSmpl` \ better_alts ->
1136 mkCase1 scrut case_bndr ty better_alts
1140 mkAlts tries these things:
1142 1. If several alternatives are identical, merge them into
1143 a single DEFAULT alternative. I've occasionally seen this
1144 making a big difference:
1146 case e of =====> case e of
1147 C _ -> f x D v -> ....v....
1148 D v -> ....v.... DEFAULT -> f x
1151 The point is that we merge common RHSs, at least for the DEFAULT case.
1152 [One could do something more elaborate but I've never seen it needed.]
1153 To avoid an expensive test, we just merge branches equal to the *first*
1154 alternative; this picks up the common cases
1155 a) all branches equal
1156 b) some branches equal to the DEFAULT (which occurs first)
1159 case e of b { ==> case e of b {
1160 p1 -> rhs1 p1 -> rhs1
1162 pm -> rhsm pm -> rhsm
1163 _ -> case b of b' { pn -> let b'=b in rhsn
1165 ... po -> let b'=b in rhso
1166 po -> rhso _ -> let b'=b in rhsd
1170 which merges two cases in one case when -- the default alternative of
1171 the outer case scrutises the same variable as the outer case This
1172 transformation is called Case Merging. It avoids that the same
1173 variable is scrutinised multiple times.
1176 The case where transformation (1) showed up was like this (lib/std/PrelCError.lhs):
1182 where @is@ was something like
1184 p `is` n = p /= (-1) && p == n
1186 This gave rise to a horrible sequence of cases
1193 and similarly in cascade for all the join points!
1198 --------------------------------------------------
1199 -- 1. Merge identical branches
1200 --------------------------------------------------
1201 mkAlts dflags scrut case_bndr alts@((con1,bndrs1,rhs1) : con_alts)
1202 | all isDeadBinder bndrs1, -- Remember the default
1203 length filtered_alts < length con_alts -- alternative comes first
1204 = tick (AltMerge case_bndr) `thenSmpl_`
1205 returnSmpl better_alts
1207 filtered_alts = filter keep con_alts
1208 keep (con,bndrs,rhs) = not (all isDeadBinder bndrs && rhs `cheapEqExpr` rhs1)
1209 better_alts = (DEFAULT, [], rhs1) : filtered_alts
1212 --------------------------------------------------
1213 -- 2. Merge nested cases
1214 --------------------------------------------------
1216 mkAlts dflags scrut outer_bndr outer_alts
1217 | dopt Opt_CaseMerge dflags,
1218 (outer_alts_without_deflt, maybe_outer_deflt) <- findDefault outer_alts,
1219 Just (Case (Var scrut_var) inner_bndr _ inner_alts) <- maybe_outer_deflt,
1220 scruting_same_var scrut_var
1222 munged_inner_alts = [(con, args, munge_rhs rhs) | (con, args, rhs) <- inner_alts]
1223 munge_rhs rhs = bindCaseBndr inner_bndr (Var outer_bndr) rhs
1225 new_alts = mergeAlts outer_alts_without_deflt munged_inner_alts
1226 -- The merge keeps the inner DEFAULT at the front, if there is one
1227 -- and eliminates any inner_alts that are shadowed by the outer_alts
1229 tick (CaseMerge outer_bndr) `thenSmpl_`
1231 -- Warning: don't call mkAlts recursively!
1232 -- Firstly, there's no point, because inner alts have already had
1233 -- mkCase applied to them, so they won't have a case in their default
1234 -- Secondly, if you do, you get an infinite loop, because the bindCaseBndr
1235 -- in munge_rhs may put a case into the DEFAULT branch!
1237 -- We are scrutinising the same variable if it's
1238 -- the outer case-binder, or if the outer case scrutinises a variable
1239 -- (and it's the same). Testing both allows us not to replace the
1240 -- outer scrut-var with the outer case-binder (Simplify.simplCaseBinder).
1241 scruting_same_var = case scrut of
1242 Var outer_scrut -> \ v -> v == outer_bndr || v == outer_scrut
1243 other -> \ v -> v == outer_bndr
1245 ------------------------------------------------
1247 ------------------------------------------------
1249 mkAlts dflags scrut case_bndr other_alts = returnSmpl other_alts
1254 =================================================================================
1256 mkCase1 tries these things
1258 1. Eliminate the case altogether if possible
1266 and similar friends.
1269 Start with a simple situation:
1271 case x# of ===> e[x#/y#]
1274 (when x#, y# are of primitive type, of course). We can't (in general)
1275 do this for algebraic cases, because we might turn bottom into
1278 Actually, we generalise this idea to look for a case where we're
1279 scrutinising a variable, and we know that only the default case can
1284 other -> ...(case x of
1288 Here the inner case can be eliminated. This really only shows up in
1289 eliminating error-checking code.
1291 We also make sure that we deal with this very common case:
1296 Here we are using the case as a strict let; if x is used only once
1297 then we want to inline it. We have to be careful that this doesn't
1298 make the program terminate when it would have diverged before, so we
1300 - x is used strictly, or
1301 - e is already evaluated (it may so if e is a variable)
1303 Lastly, we generalise the transformation to handle this:
1309 We only do this for very cheaply compared r's (constructors, literals
1310 and variables). If pedantic bottoms is on, we only do it when the
1311 scrutinee is a PrimOp which can't fail.
1313 We do it *here*, looking at un-simplified alternatives, because we
1314 have to check that r doesn't mention the variables bound by the
1315 pattern in each alternative, so the binder-info is rather useful.
1317 So the case-elimination algorithm is:
1319 1. Eliminate alternatives which can't match
1321 2. Check whether all the remaining alternatives
1322 (a) do not mention in their rhs any of the variables bound in their pattern
1323 and (b) have equal rhss
1325 3. Check we can safely ditch the case:
1326 * PedanticBottoms is off,
1327 or * the scrutinee is an already-evaluated variable
1328 or * the scrutinee is a primop which is ok for speculation
1329 -- ie we want to preserve divide-by-zero errors, and
1330 -- calls to error itself!
1332 or * [Prim cases] the scrutinee is a primitive variable
1334 or * [Alg cases] the scrutinee is a variable and
1335 either * the rhs is the same variable
1336 (eg case x of C a b -> x ===> x)
1337 or * there is only one alternative, the default alternative,
1338 and the binder is used strictly in its scope.
1339 [NB this is helped by the "use default binder where
1340 possible" transformation; see below.]
1343 If so, then we can replace the case with one of the rhss.
1345 Further notes about case elimination
1346 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
1347 Consider: test :: Integer -> IO ()
1350 Turns out that this compiles to:
1353 eta1 :: State# RealWorld ->
1354 case PrelNum.< eta PrelNum.zeroInteger of wild { __DEFAULT ->
1356 (PrelNum.jtos eta ($w[] @ Char))
1358 of wild1 { (# new_s, a4 #) -> PrelIO.lvl23 new_s }}
1360 Notice the strange '<' which has no effect at all. This is a funny one.
1361 It started like this:
1363 f x y = if x < 0 then jtos x
1364 else if y==0 then "" else jtos x
1366 At a particular call site we have (f v 1). So we inline to get
1368 if v < 0 then jtos x
1369 else if 1==0 then "" else jtos x
1371 Now simplify the 1==0 conditional:
1373 if v<0 then jtos v else jtos v
1375 Now common-up the two branches of the case:
1377 case (v<0) of DEFAULT -> jtos v
1379 Why don't we drop the case? Because it's strict in v. It's technically
1380 wrong to drop even unnecessary evaluations, and in practice they
1381 may be a result of 'seq' so we *definitely* don't want to drop those.
1382 I don't really know how to improve this situation.
1386 --------------------------------------------------
1387 -- 0. Check for empty alternatives
1388 --------------------------------------------------
1390 -- This isn't strictly an error. It's possible that the simplifer might "see"
1391 -- that an inner case has no accessible alternatives before it "sees" that the
1392 -- entire branch of an outer case is inaccessible. So we simply
1393 -- put an error case here insteadd
1394 mkCase1 scrut case_bndr ty []
1395 = pprTrace "mkCase1: null alts" (ppr case_bndr <+> ppr scrut) $
1396 return (mkApps (Var eRROR_ID)
1397 [Type ty, Lit (mkStringLit "Impossible alternative")])
1399 --------------------------------------------------
1400 -- 1. Eliminate the case altogether if poss
1401 --------------------------------------------------
1403 mkCase1 scrut case_bndr ty [(con,bndrs,rhs)]
1404 -- See if we can get rid of the case altogether
1405 -- See the extensive notes on case-elimination above
1406 -- mkCase made sure that if all the alternatives are equal,
1407 -- then there is now only one (DEFAULT) rhs
1408 | all isDeadBinder bndrs,
1410 -- Check that the scrutinee can be let-bound instead of case-bound
1411 exprOkForSpeculation scrut
1412 -- OK not to evaluate it
1413 -- This includes things like (==# a# b#)::Bool
1414 -- so that we simplify
1415 -- case ==# a# b# of { True -> x; False -> x }
1418 -- This particular example shows up in default methods for
1419 -- comparision operations (e.g. in (>=) for Int.Int32)
1420 || exprIsHNF scrut -- It's already evaluated
1421 || var_demanded_later scrut -- It'll be demanded later
1423 -- || not opt_SimplPedanticBottoms) -- Or we don't care!
1424 -- We used to allow improving termination by discarding cases, unless -fpedantic-bottoms was on,
1425 -- but that breaks badly for the dataToTag# primop, which relies on a case to evaluate
1426 -- its argument: case x of { y -> dataToTag# y }
1427 -- Here we must *not* discard the case, because dataToTag# just fetches the tag from
1428 -- the info pointer. So we'll be pedantic all the time, and see if that gives any
1430 -- Also we don't want to discard 'seq's
1431 = tick (CaseElim case_bndr) `thenSmpl_`
1432 returnSmpl (bindCaseBndr case_bndr scrut rhs)
1435 -- The case binder is going to be evaluated later,
1436 -- and the scrutinee is a simple variable
1437 var_demanded_later (Var v) = isStrictDmd (idNewDemandInfo case_bndr)
1438 var_demanded_later other = False
1441 --------------------------------------------------
1443 --------------------------------------------------
1445 mkCase1 scrut case_bndr ty alts -- Identity case
1446 | all identity_alt alts
1447 = tick (CaseIdentity case_bndr) `thenSmpl_`
1448 returnSmpl (re_note scrut)
1450 identity_alt (con, args, rhs) = de_note rhs `cheapEqExpr` identity_rhs con args
1452 identity_rhs (DataAlt con) args = mkConApp con (arg_tys ++ map varToCoreExpr args)
1453 identity_rhs (LitAlt lit) _ = Lit lit
1454 identity_rhs DEFAULT _ = Var case_bndr
1456 arg_tys = map Type (tyConAppArgs (idType case_bndr))
1459 -- case coerce T e of x { _ -> coerce T' x }
1460 -- And we definitely want to eliminate this case!
1461 -- So we throw away notes from the RHS, and reconstruct
1462 -- (at least an approximation) at the other end
1463 de_note (Note _ e) = de_note e
1466 -- re_note wraps a coerce if it might be necessary
1467 re_note scrut = case head alts of
1468 (_,_,rhs1@(Note _ _)) -> mkCoerce2 (exprType rhs1) (idType case_bndr) scrut
1472 --------------------------------------------------
1474 --------------------------------------------------
1475 mkCase1 scrut bndr ty alts = returnSmpl (Case scrut bndr ty alts)
1479 When adding auxiliary bindings for the case binder, it's worth checking if
1480 its dead, because it often is, and occasionally these mkCase transformations
1481 cascade rather nicely.
1484 bindCaseBndr bndr rhs body
1485 | isDeadBinder bndr = body
1486 | otherwise = bindNonRec bndr rhs body