2 % (c) The AQUA Project, Glasgow University, 1993-1998
4 \section[SimplUtils]{The simplifier utilities}
9 mkLam, mkCase, prepareAlts, bindCaseBndr,
12 preInlineUnconditionally, postInlineUnconditionally,
13 activeInline, activeRule, inlineMode,
15 -- The continuation type
16 SimplCont(..), DupFlag(..), LetRhsFlag(..),
17 contIsDupable, contResultType, contIsTrivial, contArgs, dropArgs,
18 countValArgs, countArgs, splitInlineCont,
19 mkBoringStop, mkLazyArgStop, mkRhsStop, contIsRhsOrArg,
20 interestingCallContext, interestingArgContext,
22 interestingArg, mkArgInfo,
27 #include "HsVersions.h"
33 import qualified CoreSubst
42 import Var ( isCoVar )
45 import Type ( Type, funArgTy, mkForAllTys, mkTyVarTys,
46 splitTyConApp_maybe, tyConAppArgs )
49 import Unify ( dataConCannotMatch )
58 %************************************************************************
62 %************************************************************************
64 A SimplCont allows the simplifier to traverse the expression in a
65 zipper-like fashion. The SimplCont represents the rest of the expression,
66 "above" the point of interest.
68 You can also think of a SimplCont as an "evaluation context", using
69 that term in the way it is used for operational semantics. This is the
70 way I usually think of it, For example you'll often see a syntax for
71 evaluation context looking like
72 C ::= [] | C e | case C of alts | C `cast` co
73 That's the kind of thing we are doing here, and I use that syntax in
78 * A SimplCont describes a *strict* context (just like
79 evaluation contexts do). E.g. Just [] is not a SimplCont
81 * A SimplCont describes a context that *does not* bind
82 any variables. E.g. \x. [] is not a SimplCont
86 = Stop -- An empty context, or hole, []
87 OutType -- Type of the result
89 Bool -- True <=> There is something interesting about
90 -- the context, and hence the inliner
91 -- should be a bit keener (see interestingCallContext)
93 -- (a) This is the RHS of a thunk whose type suggests
94 -- that update-in-place would be possible
95 -- (b) This is an argument of a function that has RULES
96 -- Inlining the call might allow the rule to fire
98 | CoerceIt -- C `cast` co
99 OutCoercion -- The coercion simplified
104 InExpr SimplEnv -- The argument and its static env
107 | Select -- case C of alts
109 InId [InAlt] SimplEnv -- The case binder, alts, and subst-env
112 -- The two strict forms have no DupFlag, because we never duplicate them
113 | StrictBind -- (\x* \xs. e) C
114 InId [InBndr] -- let x* = [] in e
115 InExpr SimplEnv -- is a special case
119 OutExpr OutType -- e and its type
120 (Bool,[Bool]) -- Whether the function at the head of e has rules,
121 SimplCont -- plus strictness flags for further args
123 data LetRhsFlag = AnArg -- It's just an argument not a let RHS
124 | AnRhs -- It's the RHS of a let (so please float lets out of big lambdas)
126 instance Outputable LetRhsFlag where
127 ppr AnArg = ptext SLIT("arg")
128 ppr AnRhs = ptext SLIT("rhs")
130 instance Outputable SimplCont where
131 ppr (Stop ty is_rhs _) = ptext SLIT("Stop") <> brackets (ppr is_rhs) <+> ppr ty
132 ppr (ApplyTo dup arg se cont) = ((ptext SLIT("ApplyTo") <+> ppr dup <+> pprParendExpr arg)
133 {- $$ nest 2 (pprSimplEnv se) -}) $$ ppr cont
134 ppr (StrictBind b _ _ _ cont) = (ptext SLIT("StrictBind") <+> ppr b) $$ ppr cont
135 ppr (StrictArg f _ _ cont) = (ptext SLIT("StrictArg") <+> ppr f) $$ ppr cont
136 ppr (Select dup bndr alts se cont) = (ptext SLIT("Select") <+> ppr dup <+> ppr bndr) $$
137 (nest 4 (ppr alts)) $$ ppr cont
138 ppr (CoerceIt co cont) = (ptext SLIT("CoerceIt") <+> ppr co) $$ ppr cont
140 data DupFlag = OkToDup | NoDup
142 instance Outputable DupFlag where
143 ppr OkToDup = ptext SLIT("ok")
144 ppr NoDup = ptext SLIT("nodup")
149 mkBoringStop :: OutType -> SimplCont
150 mkBoringStop ty = Stop ty AnArg False
152 mkLazyArgStop :: OutType -> Bool -> SimplCont
153 mkLazyArgStop ty has_rules = Stop ty AnArg (canUpdateInPlace ty || has_rules)
155 mkRhsStop :: OutType -> SimplCont
156 mkRhsStop ty = Stop ty AnRhs (canUpdateInPlace ty)
159 contIsRhsOrArg (Stop {}) = True
160 contIsRhsOrArg (StrictBind {}) = True
161 contIsRhsOrArg (StrictArg {}) = True
162 contIsRhsOrArg other = False
165 contIsDupable :: SimplCont -> Bool
166 contIsDupable (Stop {}) = True
167 contIsDupable (ApplyTo OkToDup _ _ _) = True
168 contIsDupable (Select OkToDup _ _ _ _) = True
169 contIsDupable (CoerceIt _ cont) = contIsDupable cont
170 contIsDupable other = False
173 contIsTrivial :: SimplCont -> Bool
174 contIsTrivial (Stop {}) = True
175 contIsTrivial (ApplyTo _ (Type _) _ cont) = contIsTrivial cont
176 contIsTrivial (CoerceIt _ cont) = contIsTrivial cont
177 contIsTrivial other = False
180 contResultType :: SimplCont -> OutType
181 contResultType (Stop to_ty _ _) = to_ty
182 contResultType (StrictArg _ _ _ cont) = contResultType cont
183 contResultType (StrictBind _ _ _ _ cont) = contResultType cont
184 contResultType (ApplyTo _ _ _ cont) = contResultType cont
185 contResultType (CoerceIt _ cont) = contResultType cont
186 contResultType (Select _ _ _ _ cont) = contResultType cont
189 countValArgs :: SimplCont -> Int
190 countValArgs (ApplyTo _ (Type ty) se cont) = countValArgs cont
191 countValArgs (ApplyTo _ val_arg se cont) = 1 + countValArgs cont
192 countValArgs other = 0
194 countArgs :: SimplCont -> Int
195 countArgs (ApplyTo _ arg se cont) = 1 + countArgs cont
198 contArgs :: SimplCont -> ([OutExpr], SimplCont)
199 -- Uses substitution to turn each arg into an OutExpr
200 contArgs cont = go [] cont
202 go args (ApplyTo _ arg se cont) = go (substExpr se arg : args) cont
203 go args cont = (reverse args, cont)
205 dropArgs :: Int -> SimplCont -> SimplCont
206 dropArgs 0 cont = cont
207 dropArgs n (ApplyTo _ _ _ cont) = dropArgs (n-1) cont
208 dropArgs n other = pprPanic "dropArgs" (ppr n <+> ppr other)
211 splitInlineCont :: SimplCont -> Maybe (SimplCont, SimplCont)
212 -- Returns Nothing if the continuation should dissolve an InlineMe Note
213 -- Return Just (c1,c2) otherwise,
214 -- where c1 is the continuation to put inside the InlineMe
217 -- Example: (__inline_me__ (/\a. e)) ty
218 -- Here we want to do the beta-redex without dissolving the InlineMe
219 -- See test simpl017 (and Trac #1627) for a good example of why this is important
221 splitInlineCont (ApplyTo dup (Type ty) se c)
222 | Just (c1, c2) <- splitInlineCont c = Just (ApplyTo dup (Type ty) se c1, c2)
223 splitInlineCont cont@(Stop ty _ _) = Just (mkBoringStop ty, cont)
224 splitInlineCont cont@(StrictBind bndr _ _ se _) = Just (mkBoringStop (substTy se (idType bndr)), cont)
225 splitInlineCont cont@(StrictArg _ fun_ty _ _) = Just (mkBoringStop (funArgTy fun_ty), cont)
226 splitInlineCont other = Nothing
227 -- NB: the calculation of the type for mkBoringStop is an annoying
228 -- duplication of the same calucation in mkDupableCont
233 interestingArg :: OutExpr -> Bool
234 -- An argument is interesting if it has *some* structure
235 -- We are here trying to avoid unfolding a function that
236 -- is applied only to variables that have no unfolding
237 -- (i.e. they are probably lambda bound): f x y z
238 -- There is little point in inlining f here.
239 interestingArg (Var v) = hasSomeUnfolding (idUnfolding v)
240 -- Was: isValueUnfolding (idUnfolding v')
241 -- But that seems over-pessimistic
243 -- This accounts for an argument like
244 -- () or [], which is definitely interesting
245 interestingArg (Type _) = False
246 interestingArg (App fn (Type _)) = interestingArg fn
247 interestingArg (Note _ a) = interestingArg a
249 -- Idea (from Sam B); I'm not sure if it's a good idea, so commented out for now
250 -- interestingArg expr | isUnLiftedType (exprType expr)
251 -- -- Unlifted args are only ever interesting if we know what they are
256 interestingArg other = True
257 -- Consider let x = 3 in f x
258 -- The substitution will contain (x -> ContEx 3), and we want to
259 -- to say that x is an interesting argument.
260 -- But consider also (\x. f x y) y
261 -- The substitution will contain (x -> ContEx y), and we want to say
262 -- that x is not interesting (assuming y has no unfolding)
266 Comment about interestingCallContext
267 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
268 We want to avoid inlining an expression where there can't possibly be
269 any gain, such as in an argument position. Hence, if the continuation
270 is interesting (eg. a case scrutinee, application etc.) then we
271 inline, otherwise we don't.
273 Previously some_benefit used to return True only if the variable was
274 applied to some value arguments. This didn't work:
276 let x = _coerce_ (T Int) Int (I# 3) in
277 case _coerce_ Int (T Int) x of
280 we want to inline x, but can't see that it's a constructor in a case
281 scrutinee position, and some_benefit is False.
285 dMonadST = _/\_ t -> :Monad (g1 _@_ t, g2 _@_ t, g3 _@_ t)
287 .... case dMonadST _@_ x0 of (a,b,c) -> ....
289 we'd really like to inline dMonadST here, but we *don't* want to
290 inline if the case expression is just
292 case x of y { DEFAULT -> ... }
294 since we can just eliminate this case instead (x is in WHNF). Similar
295 applies when x is bound to a lambda expression. Hence
296 contIsInteresting looks for case expressions with just a single
300 interestingCallContext :: Bool -- False <=> no args at all
301 -> Bool -- False <=> no value args
303 -- The "lone-variable" case is important. I spent ages
304 -- messing about with unsatisfactory varaints, but this is nice.
305 -- The idea is that if a variable appear all alone
306 -- as an arg of lazy fn, or rhs Stop
307 -- as scrutinee of a case Select
308 -- as arg of a strict fn ArgOf
309 -- then we should not inline it (unless there is some other reason,
310 -- e.g. is is the sole occurrence). We achieve this by making
311 -- interestingCallContext return False for a lone variable.
313 -- Why? At least in the case-scrutinee situation, turning
314 -- let x = (a,b) in case x of y -> ...
316 -- let x = (a,b) in case (a,b) of y -> ...
318 -- let x = (a,b) in let y = (a,b) in ...
319 -- is bad if the binding for x will remain.
321 -- Another example: I discovered that strings
322 -- were getting inlined straight back into applications of 'error'
323 -- because the latter is strict.
325 -- f = \x -> ...(error s)...
327 -- Fundamentally such contexts should not ecourage inlining because
328 -- the context can ``see'' the unfolding of the variable (e.g. case or a RULE)
329 -- so there's no gain.
331 -- However, even a type application or coercion isn't a lone variable.
333 -- case $fMonadST @ RealWorld of { :DMonad a b c -> c }
334 -- We had better inline that sucker! The case won't see through it.
336 -- For now, I'm treating treating a variable applied to types
337 -- in a *lazy* context "lone". The motivating example was
339 -- g = /\a. \y. h (f a)
340 -- There's no advantage in inlining f here, and perhaps
341 -- a significant disadvantage. Hence some_val_args in the Stop case
343 interestingCallContext some_args some_val_args cont
346 interesting (Select {}) = some_args
347 interesting (ApplyTo {}) = True -- Can happen if we have (coerce t (f x)) y
348 -- Perhaps True is a bit over-keen, but I've
349 -- seen (coerce f) x, where f has an INLINE prag,
350 -- So we have to give some motivaiton for inlining it
351 interesting (StrictArg {}) = some_val_args
352 interesting (StrictBind {}) = some_val_args -- ??
353 interesting (Stop ty _ interesting) = some_val_args && interesting
354 interesting (CoerceIt _ cont) = interesting cont
355 -- If this call is the arg of a strict function, the context
356 -- is a bit interesting. If we inline here, we may get useful
357 -- evaluation information to avoid repeated evals: e.g.
359 -- Here the contIsInteresting makes the '*' keener to inline,
360 -- which in turn exposes a constructor which makes the '+' inline.
361 -- Assuming that +,* aren't small enough to inline regardless.
363 -- It's also very important to inline in a strict context for things
366 -- Here, the context of (f x) is strict, and if f's unfolding is
367 -- a build it's *great* to inline it here. So we must ensure that
368 -- the context for (f x) is not totally uninteresting.
373 -> Int -- Number of value args
374 -> SimplCont -- Context of the cal
375 -> (Bool, [Bool]) -- Arg info
376 -- The arg info consists of
377 -- * A Bool indicating if the function has rules (recursively)
378 -- * A [Bool] indicating strictness for each arg
379 -- The [Bool] is usually infinite, but if it is finite it
380 -- guarantees that the function diverges after being given
381 -- that number of args
383 mkArgInfo fun n_val_args call_cont
384 = (interestingArgContext fun call_cont, fun_stricts)
386 vanilla_stricts, fun_stricts :: [Bool]
387 vanilla_stricts = repeat False
390 = case splitStrictSig (idNewStrictness fun) of
391 (demands, result_info)
392 | not (demands `lengthExceeds` n_val_args)
393 -> -- Enough args, use the strictness given.
394 -- For bottoming functions we used to pretend that the arg
395 -- is lazy, so that we don't treat the arg as an
396 -- interesting context. This avoids substituting
397 -- top-level bindings for (say) strings into
398 -- calls to error. But now we are more careful about
399 -- inlining lone variables, so its ok (see SimplUtils.analyseCont)
400 if isBotRes result_info then
401 map isStrictDmd demands -- Finite => result is bottom
403 map isStrictDmd demands ++ vanilla_stricts
405 other -> vanilla_stricts -- Not enough args, or no strictness
407 interestingArgContext :: Id -> SimplCont -> Bool
408 -- If the argument has form (f x y), where x,y are boring,
409 -- and f is marked INLINE, then we don't want to inline f.
410 -- But if the context of the argument is
412 -- where g has rules, then we *do* want to inline f, in case it
413 -- exposes a rule that might fire. Similarly, if the context is
415 -- where h has rules, then we do want to inline f.
416 -- The interesting_arg_ctxt flag makes this happen; if it's
417 -- set, the inliner gets just enough keener to inline f
418 -- regardless of how boring f's arguments are, if it's marked INLINE
420 -- The alternative would be to *always* inline an INLINE function,
421 -- regardless of how boring its context is; but that seems overkill
422 -- For example, it'd mean that wrapper functions were always inlined
423 interestingArgContext fn cont
424 = idHasRules fn || go cont
426 go (Select {}) = False
427 go (ApplyTo {}) = False
428 go (StrictArg {}) = True
429 go (StrictBind {}) = False -- ??
430 go (CoerceIt _ c) = go c
431 go (Stop _ _ interesting) = interesting
434 canUpdateInPlace :: Type -> Bool
435 -- Consider let x = <wurble> in ...
436 -- If <wurble> returns an explicit constructor, we might be able
437 -- to do update in place. So we treat even a thunk RHS context
438 -- as interesting if update in place is possible. We approximate
439 -- this by seeing if the type has a single constructor with a
440 -- small arity. But arity zero isn't good -- we share the single copy
441 -- for that case, so no point in sharing.
444 | not opt_UF_UpdateInPlace = False
446 = case splitTyConApp_maybe ty of
448 Just (tycon, _) -> case tyConDataCons_maybe tycon of
449 Just [dc] -> arity == 1 || arity == 2
451 arity = dataConRepArity dc
457 %************************************************************************
459 \subsection{Decisions about inlining}
461 %************************************************************************
463 Inlining is controlled partly by the SimplifierMode switch. This has two
466 SimplGently (a) Simplifying before specialiser/full laziness
467 (b) Simplifiying inside INLINE pragma
468 (c) Simplifying the LHS of a rule
469 (d) Simplifying a GHCi expression or Template
472 SimplPhase n Used at all other times
474 The key thing about SimplGently is that it does no call-site inlining.
475 Before full laziness we must be careful not to inline wrappers,
476 because doing so inhibits floating
477 e.g. ...(case f x of ...)...
478 ==> ...(case (case x of I# x# -> fw x#) of ...)...
479 ==> ...(case x of I# x# -> case fw x# of ...)...
480 and now the redex (f x) isn't floatable any more.
482 The no-inlining thing is also important for Template Haskell. You might be
483 compiling in one-shot mode with -O2; but when TH compiles a splice before
484 running it, we don't want to use -O2. Indeed, we don't want to inline
485 anything, because the byte-code interpreter might get confused about
486 unboxed tuples and suchlike.
490 SimplGently is also used as the mode to simplify inside an InlineMe note.
493 inlineMode :: SimplifierMode
494 inlineMode = SimplGently
497 It really is important to switch off inlinings inside such
498 expressions. Consider the following example
504 in ...g...g...g...g...g...
506 Now, if that's the ONLY occurrence of f, it will be inlined inside g,
507 and thence copied multiple times when g is inlined.
510 This function may be inlinined in other modules, so we
511 don't want to remove (by inlining) calls to functions that have
512 specialisations, or that may have transformation rules in an importing
515 E.g. {-# INLINE f #-}
518 and suppose that g is strict *and* has specialisations. If we inline
519 g's wrapper, we deny f the chance of getting the specialised version
520 of g when f is inlined at some call site (perhaps in some other
523 It's also important not to inline a worker back into a wrapper.
525 wraper = inline_me (\x -> ...worker... )
526 Normally, the inline_me prevents the worker getting inlined into
527 the wrapper (initially, the worker's only call site!). But,
528 if the wrapper is sure to be called, the strictness analyser will
529 mark it 'demanded', so when the RHS is simplified, it'll get an ArgOf
530 continuation. That's why the keep_inline predicate returns True for
531 ArgOf continuations. It shouldn't do any harm not to dissolve the
532 inline-me note under these circumstances.
534 Note that the result is that we do very little simplification
537 all xs = foldr (&&) True xs
538 any p = all . map p {-# INLINE any #-}
540 Problem: any won't get deforested, and so if it's exported and the
541 importer doesn't use the inlining, (eg passes it as an arg) then we
542 won't get deforestation at all. We havn't solved this problem yet!
545 preInlineUnconditionally
546 ~~~~~~~~~~~~~~~~~~~~~~~~
547 @preInlineUnconditionally@ examines a bndr to see if it is used just
548 once in a completely safe way, so that it is safe to discard the
549 binding inline its RHS at the (unique) usage site, REGARDLESS of how
550 big the RHS might be. If this is the case we don't simplify the RHS
551 first, but just inline it un-simplified.
553 This is much better than first simplifying a perhaps-huge RHS and then
554 inlining and re-simplifying it. Indeed, it can be at least quadratically
563 We may end up simplifying e1 N times, e2 N-1 times, e3 N-3 times etc.
564 This can happen with cascades of functions too:
571 THE MAIN INVARIANT is this:
573 ---- preInlineUnconditionally invariant -----
574 IF preInlineUnconditionally chooses to inline x = <rhs>
575 THEN doing the inlining should not change the occurrence
576 info for the free vars of <rhs>
577 ----------------------------------------------
579 For example, it's tempting to look at trivial binding like
581 and inline it unconditionally. But suppose x is used many times,
582 but this is the unique occurrence of y. Then inlining x would change
583 y's occurrence info, which breaks the invariant. It matters: y
584 might have a BIG rhs, which will now be dup'd at every occurrenc of x.
587 Evne RHSs labelled InlineMe aren't caught here, because there might be
588 no benefit from inlining at the call site.
590 [Sept 01] Don't unconditionally inline a top-level thing, because that
591 can simply make a static thing into something built dynamically. E.g.
595 [Remember that we treat \s as a one-shot lambda.] No point in
596 inlining x unless there is something interesting about the call site.
598 But watch out: if you aren't careful, some useful foldr/build fusion
599 can be lost (most notably in spectral/hartel/parstof) because the
600 foldr didn't see the build. Doing the dynamic allocation isn't a big
601 deal, in fact, but losing the fusion can be. But the right thing here
602 seems to be to do a callSiteInline based on the fact that there is
603 something interesting about the call site (it's strict). Hmm. That
606 Conclusion: inline top level things gaily until Phase 0 (the last
607 phase), at which point don't.
610 preInlineUnconditionally :: SimplEnv -> TopLevelFlag -> InId -> InExpr -> Bool
611 preInlineUnconditionally env top_lvl bndr rhs
613 | opt_SimplNoPreInlining = False
614 | otherwise = case idOccInfo bndr of
615 IAmDead -> True -- Happens in ((\x.1) v)
616 OneOcc in_lam True int_cxt -> try_once in_lam int_cxt
620 active = case phase of
621 SimplGently -> isAlwaysActive prag
622 SimplPhase n -> isActive n prag
623 prag = idInlinePragma bndr
625 try_once in_lam int_cxt -- There's one textual occurrence
626 | not in_lam = isNotTopLevel top_lvl || early_phase
627 | otherwise = int_cxt && canInlineInLam rhs
629 -- Be very careful before inlining inside a lambda, becuase (a) we must not
630 -- invalidate occurrence information, and (b) we want to avoid pushing a
631 -- single allocation (here) into multiple allocations (inside lambda).
632 -- Inlining a *function* with a single *saturated* call would be ok, mind you.
633 -- || (if is_cheap && not (canInlineInLam rhs) then pprTrace "preinline" (ppr bndr <+> ppr rhs) ok else ok)
635 -- is_cheap = exprIsCheap rhs
636 -- ok = is_cheap && int_cxt
638 -- int_cxt The context isn't totally boring
639 -- E.g. let f = \ab.BIG in \y. map f xs
640 -- Don't want to substitute for f, because then we allocate
641 -- its closure every time the \y is called
642 -- But: let f = \ab.BIG in \y. map (f y) xs
643 -- Now we do want to substitute for f, even though it's not
644 -- saturated, because we're going to allocate a closure for
645 -- (f y) every time round the loop anyhow.
647 -- canInlineInLam => free vars of rhs are (Once in_lam) or Many,
648 -- so substituting rhs inside a lambda doesn't change the occ info.
649 -- Sadly, not quite the same as exprIsHNF.
650 canInlineInLam (Lit l) = True
651 canInlineInLam (Lam b e) = isRuntimeVar b || canInlineInLam e
652 canInlineInLam (Note _ e) = canInlineInLam e
653 canInlineInLam _ = False
655 early_phase = case phase of
656 SimplPhase 0 -> False
658 -- If we don't have this early_phase test, consider
659 -- x = length [1,2,3]
660 -- The full laziness pass carefully floats all the cons cells to
661 -- top level, and preInlineUnconditionally floats them all back in.
662 -- Result is (a) static allocation replaced by dynamic allocation
663 -- (b) many simplifier iterations because this tickles
664 -- a related problem; only one inlining per pass
666 -- On the other hand, I have seen cases where top-level fusion is
667 -- lost if we don't inline top level thing (e.g. string constants)
668 -- Hence the test for phase zero (which is the phase for all the final
669 -- simplifications). Until phase zero we take no special notice of
670 -- top level things, but then we become more leery about inlining
675 postInlineUnconditionally
676 ~~~~~~~~~~~~~~~~~~~~~~~~~
677 @postInlineUnconditionally@ decides whether to unconditionally inline
678 a thing based on the form of its RHS; in particular if it has a
679 trivial RHS. If so, we can inline and discard the binding altogether.
681 NB: a loop breaker has must_keep_binding = True and non-loop-breakers
682 only have *forward* references Hence, it's safe to discard the binding
684 NOTE: This isn't our last opportunity to inline. We're at the binding
685 site right now, and we'll get another opportunity when we get to the
688 Note that we do this unconditional inlining only for trival RHSs.
689 Don't inline even WHNFs inside lambdas; doing so may simply increase
690 allocation when the function is called. This isn't the last chance; see
693 NB: Even inline pragmas (e.g. IMustBeINLINEd) are ignored here Why?
694 Because we don't even want to inline them into the RHS of constructor
695 arguments. See NOTE above
697 NB: At one time even NOINLINE was ignored here: if the rhs is trivial
698 it's best to inline it anyway. We often get a=E; b=a from desugaring,
699 with both a and b marked NOINLINE. But that seems incompatible with
700 our new view that inlining is like a RULE, so I'm sticking to the 'active'
704 postInlineUnconditionally
705 :: SimplEnv -> TopLevelFlag
706 -> InId -- The binder (an OutId would be fine too)
707 -> OccInfo -- From the InId
711 postInlineUnconditionally env top_lvl bndr occ_info rhs unfolding
713 | isLoopBreaker occ_info = False -- If it's a loop-breaker of any kind, dont' inline
714 -- because it might be referred to "earlier"
715 | isExportedId bndr = False
716 | exprIsTrivial rhs = True
719 -- The point of examining occ_info here is that for *non-values*
720 -- that occur outside a lambda, the call-site inliner won't have
721 -- a chance (becuase it doesn't know that the thing
722 -- only occurs once). The pre-inliner won't have gotten
723 -- it either, if the thing occurs in more than one branch
724 -- So the main target is things like
727 -- True -> case x of ...
728 -- False -> case x of ...
729 -- I'm not sure how important this is in practice
730 OneOcc in_lam one_br int_cxt -- OneOcc => no code-duplication issue
731 -> smallEnoughToInline unfolding -- Small enough to dup
732 -- ToDo: consider discount on smallEnoughToInline if int_cxt is true
734 -- NB: Do NOT inline arbitrarily big things, even if one_br is True
735 -- Reason: doing so risks exponential behaviour. We simplify a big
736 -- expression, inline it, and simplify it again. But if the
737 -- very same thing happens in the big expression, we get
739 -- PRINCIPLE: when we've already simplified an expression once,
740 -- make sure that we only inline it if it's reasonably small.
742 && ((isNotTopLevel top_lvl && not in_lam) ||
743 -- But outside a lambda, we want to be reasonably aggressive
744 -- about inlining into multiple branches of case
745 -- e.g. let x = <non-value>
746 -- in case y of { C1 -> ..x..; C2 -> ..x..; C3 -> ... }
747 -- Inlining can be a big win if C3 is the hot-spot, even if
748 -- the uses in C1, C2 are not 'interesting'
749 -- An example that gets worse if you add int_cxt here is 'clausify'
751 (isCheapUnfolding unfolding && int_cxt))
752 -- isCheap => acceptable work duplication; in_lam may be true
753 -- int_cxt to prevent us inlining inside a lambda without some
754 -- good reason. See the notes on int_cxt in preInlineUnconditionally
756 IAmDead -> True -- This happens; for example, the case_bndr during case of
757 -- known constructor: case (a,b) of x { (p,q) -> ... }
758 -- Here x isn't mentioned in the RHS, so we don't want to
759 -- create the (dead) let-binding let x = (a,b) in ...
763 -- Here's an example that we don't handle well:
764 -- let f = if b then Left (\x.BIG) else Right (\y.BIG)
765 -- in \y. ....case f of {...} ....
766 -- Here f is used just once, and duplicating the case work is fine (exprIsCheap).
768 -- * We can't preInlineUnconditionally because that woud invalidate
769 -- the occ info for b.
770 -- * We can't postInlineUnconditionally because the RHS is big, and
771 -- that risks exponential behaviour
772 -- * We can't call-site inline, because the rhs is big
776 active = case getMode env of
777 SimplGently -> isAlwaysActive prag
778 SimplPhase n -> isActive n prag
779 prag = idInlinePragma bndr
781 activeInline :: SimplEnv -> OutId -> Bool
783 = case getMode env of
785 -- No inlining at all when doing gentle stuff,
786 -- except for local things that occur once
787 -- The reason is that too little clean-up happens if you
788 -- don't inline use-once things. Also a bit of inlining is *good* for
789 -- full laziness; it can expose constant sub-expressions.
790 -- Example in spectral/mandel/Mandel.hs, where the mandelset
791 -- function gets a useful let-float if you inline windowToViewport
793 -- NB: we used to have a second exception, for data con wrappers.
794 -- On the grounds that we use gentle mode for rule LHSs, and
795 -- they match better when data con wrappers are inlined.
796 -- But that only really applies to the trivial wrappers (like (:)),
797 -- and they are now constructed as Compulsory unfoldings (in MkId)
798 -- so they'll happen anyway.
800 SimplPhase n -> isActive n prag
802 prag = idInlinePragma id
804 activeRule :: DynFlags -> SimplEnv -> Maybe (Activation -> Bool)
805 -- Nothing => No rules at all
806 activeRule dflags env
807 | not (dopt Opt_RewriteRules dflags)
808 = Nothing -- Rewriting is off
810 = case getMode env of
811 SimplGently -> Just isAlwaysActive
812 -- Used to be Nothing (no rules in gentle mode)
813 -- Main motivation for changing is that I wanted
814 -- lift String ===> ...
815 -- to work in Template Haskell when simplifying
816 -- splices, so we get simpler code for literal strings
817 SimplPhase n -> Just (isActive n)
821 %************************************************************************
825 %************************************************************************
828 mkLam :: [OutBndr] -> OutExpr -> SimplM OutExpr
829 -- mkLam tries three things
830 -- a) eta reduction, if that gives a trivial expression
831 -- b) eta expansion [only if there are some value lambdas]
836 = do { dflags <- getDOptsSmpl
837 ; mkLam' dflags bndrs body }
839 mkLam' :: DynFlags -> [OutBndr] -> OutExpr -> SimplM OutExpr
840 mkLam' dflags bndrs (Cast body@(Lam _ _) co)
841 -- Note [Casts and lambdas]
842 = do { lam <- mkLam' dflags (bndrs ++ bndrs') body'
843 ; return (mkCoerce (mkPiTypes bndrs co) lam) }
845 (bndrs',body') = collectBinders body
847 mkLam' dflags bndrs body
848 | dopt Opt_DoEtaReduction dflags,
849 Just etad_lam <- tryEtaReduce bndrs body
850 = do { tick (EtaReduction (head bndrs))
853 | dopt Opt_DoLambdaEtaExpansion dflags,
854 any isRuntimeVar bndrs
855 = do { body' <- tryEtaExpansion dflags body
856 ; return (mkLams bndrs body') }
859 = returnSmpl (mkLams bndrs body)
862 Note [Casts and lambdas]
863 ~~~~~~~~~~~~~~~~~~~~~~~~
865 (\x. (\y. e) `cast` g1) `cast` g2
866 There is a danger here that the two lambdas look separated, and the
867 full laziness pass might float an expression to between the two.
869 So this equation in mkLam' floats the g1 out, thus:
870 (\x. e `cast` g1) --> (\x.e) `cast` (tx -> g1)
873 In general, this floats casts outside lambdas, where (I hope) they might meet
874 and cancel with some other cast.
877 -- c) floating lets out through big lambdas
878 -- [only if all tyvar lambdas, and only if this lambda
879 -- is the RHS of a let]
881 {- Sept 01: I'm experimenting with getting the
882 full laziness pass to float out past big lambdsa
883 | all isTyVar bndrs, -- Only for big lambdas
884 contIsRhs cont -- Only try the rhs type-lambda floating
885 -- if this is indeed a right-hand side; otherwise
886 -- we end up floating the thing out, only for float-in
887 -- to float it right back in again!
888 = tryRhsTyLam env bndrs body `thenSmpl` \ (floats, body') ->
889 returnSmpl (floats, mkLams bndrs body')
893 %************************************************************************
895 \subsection{Eta expansion and reduction}
897 %************************************************************************
899 We try for eta reduction here, but *only* if we get all the
900 way to an exprIsTrivial expression.
901 We don't want to remove extra lambdas unless we are going
902 to avoid allocating this thing altogether
905 tryEtaReduce :: [OutBndr] -> OutExpr -> Maybe OutExpr
906 tryEtaReduce bndrs body
907 -- We don't use CoreUtils.etaReduce, because we can be more
909 -- (a) we already have the binders
910 -- (b) we can do the triviality test before computing the free vars
911 = go (reverse bndrs) body
913 go (b : bs) (App fun arg) | ok_arg b arg = go bs fun -- Loop round
914 go [] fun | ok_fun fun = Just fun -- Success!
915 go _ _ = Nothing -- Failure!
917 ok_fun fun = exprIsTrivial fun
918 && not (any (`elemVarSet` (exprFreeVars fun)) bndrs)
919 && (exprIsHNF fun || all ok_lam bndrs)
920 ok_lam v = isTyVar v || isDictId v
921 -- The exprIsHNF is because eta reduction is not
922 -- valid in general: \x. bot /= bot
923 -- So we need to be sure that the "fun" is a value.
925 -- However, we always want to reduce (/\a -> f a) to f
926 -- This came up in a RULE: foldr (build (/\a -> g a))
927 -- did not match foldr (build (/\b -> ...something complex...))
928 -- The type checker can insert these eta-expanded versions,
929 -- with both type and dictionary lambdas; hence the slightly
932 ok_arg b arg = varToCoreExpr b `cheapEqExpr` arg
936 Try eta expansion for RHSs
939 f = \x1..xn -> N ==> f = \x1..xn y1..ym -> N y1..ym
942 where (in both cases)
944 * The xi can include type variables
946 * The yi are all value variables
948 * N is a NORMAL FORM (i.e. no redexes anywhere)
949 wanting a suitable number of extra args.
951 We may have to sandwich some coerces between the lambdas
952 to make the types work. exprEtaExpandArity looks through coerces
953 when computing arity; and etaExpand adds the coerces as necessary when
954 actually computing the expansion.
957 tryEtaExpansion :: DynFlags -> OutExpr -> SimplM OutExpr
958 -- There is at least one runtime binder in the binders
959 tryEtaExpansion dflags body
960 = getUniquesSmpl `thenSmpl` \ us ->
961 returnSmpl (etaExpand fun_arity us body (exprType body))
963 fun_arity = exprEtaExpandArity dflags body
967 %************************************************************************
969 \subsection{Floating lets out of big lambdas}
971 %************************************************************************
973 Note [Floating and type abstraction]
974 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
977 We'd like to float this to
980 x = /\a. C (y1 a) (y2 a)
981 for the usual reasons: we want to inline x rather vigorously.
983 You may think that this kind of thing is rare. But in some programs it is
984 common. For example, if you do closure conversion you might get:
986 data a :-> b = forall e. (e -> a -> b) :$ e
988 f_cc :: forall a. a :-> a
989 f_cc = /\a. (\e. id a) :$ ()
991 Now we really want to inline that f_cc thing so that the
992 construction of the closure goes away.
994 So I have elaborated simplLazyBind to understand right-hand sides that look
998 and treat them specially. The real work is done in SimplUtils.abstractFloats,
999 but there is quite a bit of plumbing in simplLazyBind as well.
1001 The same transformation is good when there are lets in the body:
1003 /\abc -> let(rec) x = e in b
1005 let(rec) x' = /\abc -> let x = x' a b c in e
1007 /\abc -> let x = x' a b c in b
1009 This is good because it can turn things like:
1011 let f = /\a -> letrec g = ... g ... in g
1013 letrec g' = /\a -> ... g' a ...
1015 let f = /\ a -> g' a
1017 which is better. In effect, it means that big lambdas don't impede
1020 This optimisation is CRUCIAL in eliminating the junk introduced by
1021 desugaring mutually recursive definitions. Don't eliminate it lightly!
1023 [May 1999] If we do this transformation *regardless* then we can
1024 end up with some pretty silly stuff. For example,
1027 st = /\ s -> let { x1=r1 ; x2=r2 } in ...
1032 st = /\s -> ...[y1 s/x1, y2 s/x2]
1035 Unless the "..." is a WHNF there is really no point in doing this.
1036 Indeed it can make things worse. Suppose x1 is used strictly,
1039 x1* = case f y of { (a,b) -> e }
1041 If we abstract this wrt the tyvar we then can't do the case inline
1042 as we would normally do.
1044 That's why the whole transformation is part of the same process that
1045 floats let-bindings and constructor arguments out of RHSs. In particular,
1046 it is guarded by the doFloatFromRhs call in simplLazyBind.
1050 abstractFloats :: [OutTyVar] -> SimplEnv -> OutExpr -> SimplM ([OutBind], OutExpr)
1051 abstractFloats main_tvs body_env body
1052 = ASSERT( notNull body_floats )
1053 do { (subst, float_binds) <- mapAccumLSmpl abstract empty_subst body_floats
1054 ; return (float_binds, CoreSubst.substExpr subst body) }
1056 main_tv_set = mkVarSet main_tvs
1057 body_floats = getFloats body_env
1058 empty_subst = CoreSubst.mkEmptySubst (seInScope body_env)
1060 abstract :: CoreSubst.Subst -> OutBind -> SimplM (CoreSubst.Subst, OutBind)
1061 abstract subst (NonRec id rhs)
1062 = do { (poly_id, poly_app) <- mk_poly tvs_here id
1063 ; let poly_rhs = mkLams tvs_here rhs'
1064 subst' = CoreSubst.extendIdSubst subst id poly_app
1065 ; return (subst', (NonRec poly_id poly_rhs)) }
1067 rhs' = CoreSubst.substExpr subst rhs
1068 tvs_here | any isCoVar main_tvs = main_tvs -- Note [Abstract over coercions]
1070 = varSetElems (main_tv_set `intersectVarSet` exprSomeFreeVars isTyVar rhs')
1072 -- Abstract only over the type variables free in the rhs
1073 -- wrt which the new binding is abstracted. But the naive
1074 -- approach of abstract wrt the tyvars free in the Id's type
1076 -- /\ a b -> let t :: (a,b) = (e1, e2)
1079 -- Here, b isn't free in x's type, but we must nevertheless
1080 -- abstract wrt b as well, because t's type mentions b.
1081 -- Since t is floated too, we'd end up with the bogus:
1082 -- poly_t = /\ a b -> (e1, e2)
1083 -- poly_x = /\ a -> fst (poly_t a *b*)
1084 -- So for now we adopt the even more naive approach of
1085 -- abstracting wrt *all* the tyvars. We'll see if that
1086 -- gives rise to problems. SLPJ June 98
1088 abstract subst (Rec prs)
1089 = do { (poly_ids, poly_apps) <- mapAndUnzipSmpl (mk_poly tvs_here) ids
1090 ; let subst' = CoreSubst.extendSubstList subst (ids `zip` poly_apps)
1091 poly_rhss = [mkLams tvs_here (CoreSubst.substExpr subst' rhs) | rhs <- rhss]
1092 ; return (subst', Rec (poly_ids `zip` poly_rhss)) }
1094 (ids,rhss) = unzip prs
1095 -- For a recursive group, it's a bit of a pain to work out the minimal
1096 -- set of tyvars over which to abstract:
1097 -- /\ a b c. let x = ...a... in
1098 -- letrec { p = ...x...q...
1099 -- q = .....p...b... } in
1101 -- Since 'x' is abstracted over 'a', the {p,q} group must be abstracted
1102 -- over 'a' (because x is replaced by (poly_x a)) as well as 'b'.
1103 -- Since it's a pain, we just use the whole set, which is always safe
1105 -- If you ever want to be more selective, remember this bizarre case too:
1107 -- Here, we must abstract 'x' over 'a'.
1110 mk_poly tvs_here var
1111 = do { uniq <- getUniqueSmpl
1112 ; let poly_name = setNameUnique (idName var) uniq -- Keep same name
1113 poly_ty = mkForAllTys tvs_here (idType var) -- But new type of course
1114 poly_id = mkLocalId poly_name poly_ty
1115 ; return (poly_id, mkTyApps (Var poly_id) (mkTyVarTys tvs_here)) }
1116 -- In the olden days, it was crucial to copy the occInfo of the original var,
1117 -- because we were looking at occurrence-analysed but as yet unsimplified code!
1118 -- In particular, we mustn't lose the loop breakers. BUT NOW we are looking
1119 -- at already simplified code, so it doesn't matter
1121 -- It's even right to retain single-occurrence or dead-var info:
1122 -- Suppose we started with /\a -> let x = E in B
1123 -- where x occurs once in B. Then we transform to:
1124 -- let x' = /\a -> E in /\a -> let x* = x' a in B
1125 -- where x* has an INLINE prag on it. Now, once x* is inlined,
1126 -- the occurrences of x' will be just the occurrences originally
1130 Note [Abstract over coercions]
1131 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
1132 If a coercion variable (g :: a ~ Int) is free in the RHS, then so is the
1133 type variable a. Rather than sort this mess out, we simply bale out and abstract
1134 wrt all the type variables if any of them are coercion variables.
1137 Historical note: if you use let-bindings instead of a substitution, beware of this:
1139 -- Suppose we start with:
1141 -- x = /\ a -> let g = G in E
1143 -- Then we'll float to get
1145 -- x = let poly_g = /\ a -> G
1146 -- in /\ a -> let g = poly_g a in E
1148 -- But now the occurrence analyser will see just one occurrence
1149 -- of poly_g, not inside a lambda, so the simplifier will
1150 -- PreInlineUnconditionally poly_g back into g! Badk to square 1!
1151 -- (I used to think that the "don't inline lone occurrences" stuff
1152 -- would stop this happening, but since it's the *only* occurrence,
1153 -- PreInlineUnconditionally kicks in first!)
1155 -- Solution: put an INLINE note on g's RHS, so that poly_g seems
1156 -- to appear many times. (NB: mkInlineMe eliminates
1157 -- such notes on trivial RHSs, so do it manually.)
1159 %************************************************************************
1163 %************************************************************************
1165 prepareAlts tries these things:
1167 1. If several alternatives are identical, merge them into
1168 a single DEFAULT alternative. I've occasionally seen this
1169 making a big difference:
1171 case e of =====> case e of
1172 C _ -> f x D v -> ....v....
1173 D v -> ....v.... DEFAULT -> f x
1176 The point is that we merge common RHSs, at least for the DEFAULT case.
1177 [One could do something more elaborate but I've never seen it needed.]
1178 To avoid an expensive test, we just merge branches equal to the *first*
1179 alternative; this picks up the common cases
1180 a) all branches equal
1181 b) some branches equal to the DEFAULT (which occurs first)
1184 case e of b { ==> case e of b {
1185 p1 -> rhs1 p1 -> rhs1
1187 pm -> rhsm pm -> rhsm
1188 _ -> case b of b' { pn -> let b'=b in rhsn
1190 ... po -> let b'=b in rhso
1191 po -> rhso _ -> let b'=b in rhsd
1195 which merges two cases in one case when -- the default alternative of
1196 the outer case scrutises the same variable as the outer case This
1197 transformation is called Case Merging. It avoids that the same
1198 variable is scrutinised multiple times.
1201 The case where transformation (1) showed up was like this (lib/std/PrelCError.lhs):
1207 where @is@ was something like
1209 p `is` n = p /= (-1) && p == n
1211 This gave rise to a horrible sequence of cases
1218 and similarly in cascade for all the join points!
1221 ~~~~~~~~~~~~~~~~~~~~
1222 We do this *here*, looking at un-simplified alternatives, because we
1223 have to check that r doesn't mention the variables bound by the
1224 pattern in each alternative, so the binder-info is rather useful.
1227 prepareAlts :: OutExpr -> OutId -> [InAlt] -> SimplM ([AltCon], [InAlt])
1228 prepareAlts scrut case_bndr' alts
1229 = do { dflags <- getDOptsSmpl
1230 ; alts <- combineIdenticalAlts case_bndr' alts
1232 ; let (alts_wo_default, maybe_deflt) = findDefault alts
1233 alt_cons = [con | (con,_,_) <- alts_wo_default]
1234 imposs_deflt_cons = nub (imposs_cons ++ alt_cons)
1235 -- "imposs_deflt_cons" are handled
1236 -- EITHER by the context,
1237 -- OR by a non-DEFAULT branch in this case expression.
1239 ; default_alts <- prepareDefault dflags scrut case_bndr' mb_tc_app
1240 imposs_deflt_cons maybe_deflt
1242 ; let trimmed_alts = filterOut impossible_alt alts_wo_default
1243 merged_alts = mergeAlts trimmed_alts default_alts
1244 -- We need the mergeAlts in case the new default_alt
1245 -- has turned into a constructor alternative.
1246 -- The merge keeps the inner DEFAULT at the front, if there is one
1247 -- and interleaves the alternatives in the right order
1249 ; return (imposs_deflt_cons, merged_alts) }
1251 mb_tc_app = splitTyConApp_maybe (idType case_bndr')
1252 Just (_, inst_tys) = mb_tc_app
1254 imposs_cons = case scrut of
1255 Var v -> otherCons (idUnfolding v)
1258 impossible_alt :: CoreAlt -> Bool
1259 impossible_alt (con, _, _) | con `elem` imposs_cons = True
1260 impossible_alt (DataAlt con, _, _) = dataConCannotMatch inst_tys con
1261 impossible_alt alt = False
1264 --------------------------------------------------
1265 -- 1. Merge identical branches
1266 --------------------------------------------------
1267 combineIdenticalAlts :: OutId -> [InAlt] -> SimplM [InAlt]
1269 combineIdenticalAlts case_bndr alts@((con1,bndrs1,rhs1) : con_alts)
1270 | all isDeadBinder bndrs1, -- Remember the default
1271 length filtered_alts < length con_alts -- alternative comes first
1272 -- Also Note [Dead binders]
1273 = do { tick (AltMerge case_bndr)
1274 ; return ((DEFAULT, [], rhs1) : filtered_alts) }
1276 filtered_alts = filter keep con_alts
1277 keep (con,bndrs,rhs) = not (all isDeadBinder bndrs && rhs `cheapEqExpr` rhs1)
1279 combineIdenticalAlts case_bndr alts = return alts
1281 -------------------------------------------------------------------------
1282 -- Prepare the default alternative
1283 -------------------------------------------------------------------------
1284 prepareDefault :: DynFlags
1285 -> OutExpr -- Scrutinee
1286 -> OutId -- Case binder; need just for its type. Note that as an
1287 -- OutId, it has maximum information; this is important.
1288 -- Test simpl013 is an example
1289 -> Maybe (TyCon, [Type]) -- Type of scrutinee, decomposed
1290 -> [AltCon] -- These cons can't happen when matching the default
1291 -> Maybe InExpr -- Rhs
1292 -> SimplM [InAlt] -- Still unsimplified
1293 -- We use a list because it's what mergeAlts expects,
1294 -- And becuase case-merging can cause many to show up
1296 ------- Merge nested cases ----------
1297 prepareDefault dflags scrut outer_bndr bndr_ty imposs_cons (Just deflt_rhs)
1298 | dopt Opt_CaseMerge dflags
1299 , Case (Var scrut_var) inner_bndr _ inner_alts <- deflt_rhs
1300 , scruting_same_var scrut_var
1301 = do { tick (CaseMerge outer_bndr)
1303 ; let munge_rhs rhs = bindCaseBndr inner_bndr (Var outer_bndr) rhs
1304 ; return [(con, args, munge_rhs rhs) | (con, args, rhs) <- inner_alts,
1305 not (con `elem` imposs_cons) ]
1306 -- NB: filter out any imposs_cons. Example:
1309 -- DEFAULT -> case x of
1312 -- When we merge, we must ensure that e1 takes
1313 -- precedence over e2 as the value for A!
1315 -- Warning: don't call prepareAlts recursively!
1316 -- Firstly, there's no point, because inner alts have already had
1317 -- mkCase applied to them, so they won't have a case in their default
1318 -- Secondly, if you do, you get an infinite loop, because the bindCaseBndr
1319 -- in munge_rhs may put a case into the DEFAULT branch!
1321 -- We are scrutinising the same variable if it's
1322 -- the outer case-binder, or if the outer case scrutinises a variable
1323 -- (and it's the same). Testing both allows us not to replace the
1324 -- outer scrut-var with the outer case-binder (Simplify.simplCaseBinder).
1325 scruting_same_var = case scrut of
1326 Var outer_scrut -> \ v -> v == outer_bndr || v == outer_scrut
1327 other -> \ v -> v == outer_bndr
1329 --------- Fill in known constructor -----------
1330 prepareDefault dflags scrut case_bndr (Just (tycon, inst_tys)) imposs_cons (Just deflt_rhs)
1331 | -- This branch handles the case where we are
1332 -- scrutinisng an algebraic data type
1333 isAlgTyCon tycon -- It's a data type, tuple, or unboxed tuples.
1334 , not (isNewTyCon tycon) -- We can have a newtype, if we are just doing an eval:
1335 -- case x of { DEFAULT -> e }
1336 -- and we don't want to fill in a default for them!
1337 , Just all_cons <- tyConDataCons_maybe tycon
1338 , not (null all_cons) -- This is a tricky corner case. If the data type has no constructors,
1339 -- which GHC allows, then the case expression will have at most a default
1340 -- alternative. We don't want to eliminate that alternative, because the
1341 -- invariant is that there's always one alternative. It's more convenient
1343 -- case x of { DEFAULT -> e }
1344 -- as it is, rather than transform it to
1345 -- error "case cant match"
1346 -- which would be quite legitmate. But it's a really obscure corner, and
1347 -- not worth wasting code on.
1348 , let imposs_data_cons = [con | DataAlt con <- imposs_cons] -- We now know it's a data type
1349 impossible con = con `elem` imposs_data_cons || dataConCannotMatch inst_tys con
1350 = case filterOut impossible all_cons of
1351 [] -> return [] -- Eliminate the default alternative
1352 -- altogether if it can't match
1354 [con] -> -- It matches exactly one constructor, so fill it in
1355 do { tick (FillInCaseDefault case_bndr)
1356 ; us <- getUniquesSmpl
1357 ; let (ex_tvs, co_tvs, arg_ids) =
1358 dataConRepInstPat us con inst_tys
1359 ; return [(DataAlt con, ex_tvs ++ co_tvs ++ arg_ids, deflt_rhs)] }
1361 two_or_more -> return [(DEFAULT, [], deflt_rhs)]
1363 --------- Catch-all cases -----------
1364 prepareDefault dflags scrut case_bndr bndr_ty imposs_cons (Just deflt_rhs)
1365 = return [(DEFAULT, [], deflt_rhs)]
1367 prepareDefault dflags scrut case_bndr bndr_ty imposs_cons Nothing
1368 = return [] -- No default branch
1373 =================================================================================
1375 mkCase tries these things
1377 1. Eliminate the case altogether if possible
1385 and similar friends.
1389 mkCase :: OutExpr -> OutId -> OutType
1390 -> [OutAlt] -- Increasing order
1393 --------------------------------------------------
1394 -- 1. Check for empty alternatives
1395 --------------------------------------------------
1397 -- This isn't strictly an error. It's possible that the simplifer might "see"
1398 -- that an inner case has no accessible alternatives before it "sees" that the
1399 -- entire branch of an outer case is inaccessible. So we simply
1400 -- put an error case here insteadd
1401 mkCase scrut case_bndr ty []
1402 = pprTrace "mkCase: null alts" (ppr case_bndr <+> ppr scrut) $
1403 return (mkApps (Var rUNTIME_ERROR_ID)
1404 [Type ty, Lit (mkStringLit "Impossible alternative")])
1407 --------------------------------------------------
1409 --------------------------------------------------
1411 mkCase scrut case_bndr ty alts -- Identity case
1412 | all identity_alt alts
1413 = tick (CaseIdentity case_bndr) `thenSmpl_`
1414 returnSmpl (re_cast scrut)
1416 identity_alt (con, args, rhs) = check_eq con args (de_cast rhs)
1418 check_eq DEFAULT _ (Var v) = v == case_bndr
1419 check_eq (LitAlt lit') _ (Lit lit) = lit == lit'
1420 check_eq (DataAlt con) args rhs = rhs `cheapEqExpr` mkConApp con (arg_tys ++ varsToCoreExprs args)
1421 || rhs `cheapEqExpr` Var case_bndr
1422 check_eq con args rhs = False
1424 arg_tys = map Type (tyConAppArgs (idType case_bndr))
1427 -- case e of x { _ -> x `cast` c }
1428 -- And we definitely want to eliminate this case, to give
1430 -- So we throw away the cast from the RHS, and reconstruct
1431 -- it at the other end. All the RHS casts must be the same
1432 -- if (all identity_alt alts) holds.
1434 -- Don't worry about nested casts, because the simplifier combines them
1435 de_cast (Cast e _) = e
1438 re_cast scrut = case head alts of
1439 (_,_,Cast _ co) -> Cast scrut co
1444 --------------------------------------------------
1446 --------------------------------------------------
1447 mkCase scrut bndr ty alts = returnSmpl (Case scrut bndr ty alts)
1451 When adding auxiliary bindings for the case binder, it's worth checking if
1452 its dead, because it often is, and occasionally these mkCase transformations
1453 cascade rather nicely.
1456 bindCaseBndr bndr rhs body
1457 | isDeadBinder bndr = body
1458 | otherwise = bindNonRec bndr rhs body