2 % (c) The AQUA Project, Glasgow University, 1993-1998
4 \section[SimplUtils]{The simplifier utilities}
8 -- The above warning supression flag is a temporary kludge.
9 -- While working on this module you are encouraged to remove it and fix
10 -- any warnings in the module. See
11 -- http://hackage.haskell.org/trac/ghc/wiki/Commentary/CodingStyle#Warnings
16 mkLam, mkCase, prepareAlts, bindCaseBndr,
19 preInlineUnconditionally, postInlineUnconditionally,
20 activeInline, activeRule, inlineMode,
22 -- The continuation type
23 SimplCont(..), DupFlag(..), ArgInfo(..),
24 contIsDupable, contResultType, contIsTrivial, contArgs, dropArgs,
25 countValArgs, countArgs, splitInlineCont,
26 mkBoringStop, mkLazyArgStop, mkRhsStop, contIsRhsOrArg,
27 interestingCallContext, interestingArgContext,
29 interestingArg, mkArgInfo,
34 #include "HsVersions.h"
40 import qualified CoreSubst
49 import Var ( isCoVar )
52 import Type hiding( substTy )
55 import Unify ( dataConCannotMatch )
66 %************************************************************************
70 %************************************************************************
72 A SimplCont allows the simplifier to traverse the expression in a
73 zipper-like fashion. The SimplCont represents the rest of the expression,
74 "above" the point of interest.
76 You can also think of a SimplCont as an "evaluation context", using
77 that term in the way it is used for operational semantics. This is the
78 way I usually think of it, For example you'll often see a syntax for
79 evaluation context looking like
80 C ::= [] | C e | case C of alts | C `cast` co
81 That's the kind of thing we are doing here, and I use that syntax in
86 * A SimplCont describes a *strict* context (just like
87 evaluation contexts do). E.g. Just [] is not a SimplCont
89 * A SimplCont describes a context that *does not* bind
90 any variables. E.g. \x. [] is not a SimplCont
94 = Stop -- An empty context, or hole, []
95 OutType -- Type of the result
96 CallCtxt -- True <=> There is something interesting about
97 -- the context, and hence the inliner
98 -- should be a bit keener (see interestingCallContext)
100 -- This is an argument of a function that has RULES
101 -- Inlining the call might allow the rule to fire
103 | CoerceIt -- C `cast` co
104 OutCoercion -- The coercion simplified
109 InExpr SimplEnv -- The argument and its static env
112 | Select -- case C of alts
114 InId [InAlt] SimplEnv -- The case binder, alts, and subst-env
117 -- The two strict forms have no DupFlag, because we never duplicate them
118 | StrictBind -- (\x* \xs. e) C
119 InId [InBndr] -- let x* = [] in e
120 InExpr SimplEnv -- is a special case
124 OutExpr OutType -- e and its type
125 CallCtxt -- Whether *this* argument position is interesting
126 ArgInfo -- Whether the function at the head of e has rules, etc
127 SimplCont -- plus strictness flags for *further* args
131 ai_rules :: Bool, -- Function has rules (recursively)
132 -- => be keener to inline in all args
133 ai_strs :: [Bool], -- Strictness of arguments
134 -- Usually infinite, but if it is finite it guarantees
135 -- that the function diverges after being given
136 -- that number of args
137 ai_discs :: [Int] -- Discounts for arguments; non-zero => be keener to inline
141 instance Outputable SimplCont where
142 ppr (Stop ty _) = ptext SLIT("Stop") <+> ppr ty
143 ppr (ApplyTo dup arg se cont) = ((ptext SLIT("ApplyTo") <+> ppr dup <+> pprParendExpr arg)
144 {- $$ nest 2 (pprSimplEnv se) -}) $$ ppr cont
145 ppr (StrictBind b _ _ _ cont) = (ptext SLIT("StrictBind") <+> ppr b) $$ ppr cont
146 ppr (StrictArg f _ _ _ cont) = (ptext SLIT("StrictArg") <+> ppr f) $$ ppr cont
147 ppr (Select dup bndr alts se cont) = (ptext SLIT("Select") <+> ppr dup <+> ppr bndr) $$
148 (nest 4 (ppr alts)) $$ ppr cont
149 ppr (CoerceIt co cont) = (ptext SLIT("CoerceIt") <+> ppr co) $$ ppr cont
151 data DupFlag = OkToDup | NoDup
153 instance Outputable DupFlag where
154 ppr OkToDup = ptext SLIT("ok")
155 ppr NoDup = ptext SLIT("nodup")
160 mkBoringStop :: OutType -> SimplCont
161 mkBoringStop ty = Stop ty BoringCtxt
163 mkLazyArgStop :: OutType -> CallCtxt -> SimplCont
164 mkLazyArgStop ty cci = Stop ty cci
166 mkRhsStop :: OutType -> SimplCont
167 mkRhsStop ty = Stop ty BoringCtxt
170 contIsRhsOrArg (Stop {}) = True
171 contIsRhsOrArg (StrictBind {}) = True
172 contIsRhsOrArg (StrictArg {}) = True
173 contIsRhsOrArg other = False
176 contIsDupable :: SimplCont -> Bool
177 contIsDupable (Stop {}) = True
178 contIsDupable (ApplyTo OkToDup _ _ _) = True
179 contIsDupable (Select OkToDup _ _ _ _) = True
180 contIsDupable (CoerceIt _ cont) = contIsDupable cont
181 contIsDupable other = False
184 contIsTrivial :: SimplCont -> Bool
185 contIsTrivial (Stop {}) = True
186 contIsTrivial (ApplyTo _ (Type _) _ cont) = contIsTrivial cont
187 contIsTrivial (CoerceIt _ cont) = contIsTrivial cont
188 contIsTrivial other = False
191 contResultType :: SimplCont -> OutType
192 contResultType (Stop to_ty _) = to_ty
193 contResultType (StrictArg _ _ _ _ cont) = contResultType cont
194 contResultType (StrictBind _ _ _ _ cont) = contResultType cont
195 contResultType (ApplyTo _ _ _ cont) = contResultType cont
196 contResultType (CoerceIt _ cont) = contResultType cont
197 contResultType (Select _ _ _ _ cont) = contResultType cont
200 countValArgs :: SimplCont -> Int
201 countValArgs (ApplyTo _ (Type ty) se cont) = countValArgs cont
202 countValArgs (ApplyTo _ val_arg se cont) = 1 + countValArgs cont
203 countValArgs other = 0
205 countArgs :: SimplCont -> Int
206 countArgs (ApplyTo _ arg se cont) = 1 + countArgs cont
209 contArgs :: SimplCont -> ([OutExpr], SimplCont)
210 -- Uses substitution to turn each arg into an OutExpr
211 contArgs cont = go [] cont
213 go args (ApplyTo _ arg se cont) = go (substExpr se arg : args) cont
214 go args cont = (reverse args, cont)
216 dropArgs :: Int -> SimplCont -> SimplCont
217 dropArgs 0 cont = cont
218 dropArgs n (ApplyTo _ _ _ cont) = dropArgs (n-1) cont
219 dropArgs n other = pprPanic "dropArgs" (ppr n <+> ppr other)
222 splitInlineCont :: SimplCont -> Maybe (SimplCont, SimplCont)
223 -- Returns Nothing if the continuation should dissolve an InlineMe Note
224 -- Return Just (c1,c2) otherwise,
225 -- where c1 is the continuation to put inside the InlineMe
228 -- Example: (__inline_me__ (/\a. e)) ty
229 -- Here we want to do the beta-redex without dissolving the InlineMe
230 -- See test simpl017 (and Trac #1627) for a good example of why this is important
232 splitInlineCont (ApplyTo dup (Type ty) se c)
233 | Just (c1, c2) <- splitInlineCont c = Just (ApplyTo dup (Type ty) se c1, c2)
234 splitInlineCont cont@(Stop ty _) = Just (mkBoringStop ty, cont)
235 splitInlineCont cont@(StrictBind bndr _ _ se _) = Just (mkBoringStop (substTy se (idType bndr)), cont)
236 splitInlineCont cont@(StrictArg _ fun_ty _ _ _) = Just (mkBoringStop (funArgTy fun_ty), cont)
237 splitInlineCont other = Nothing
238 -- NB: the calculation of the type for mkBoringStop is an annoying
239 -- duplication of the same calucation in mkDupableCont
244 interestingArg :: OutExpr -> Bool
245 -- An argument is interesting if it has *some* structure
246 -- We are here trying to avoid unfolding a function that
247 -- is applied only to variables that have no unfolding
248 -- (i.e. they are probably lambda bound): f x y z
249 -- There is little point in inlining f here.
250 interestingArg (Var v) = hasSomeUnfolding (idUnfolding v)
251 -- Was: isValueUnfolding (idUnfolding v')
252 -- But that seems over-pessimistic
254 -- This accounts for an argument like
255 -- () or [], which is definitely interesting
256 interestingArg (Type _) = False
257 interestingArg (App fn (Type _)) = interestingArg fn
258 interestingArg (Note _ a) = interestingArg a
260 -- Idea (from Sam B); I'm not sure if it's a good idea, so commented out for now
261 -- interestingArg expr | isUnLiftedType (exprType expr)
262 -- -- Unlifted args are only ever interesting if we know what they are
267 interestingArg other = True
268 -- Consider let x = 3 in f x
269 -- The substitution will contain (x -> ContEx 3), and we want to
270 -- to say that x is an interesting argument.
271 -- But consider also (\x. f x y) y
272 -- The substitution will contain (x -> ContEx y), and we want to say
273 -- that x is not interesting (assuming y has no unfolding)
277 Comment about interestingCallContext
278 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
279 We want to avoid inlining an expression where there can't possibly be
280 any gain, such as in an argument position. Hence, if the continuation
281 is interesting (eg. a case scrutinee, application etc.) then we
282 inline, otherwise we don't.
284 Previously some_benefit used to return True only if the variable was
285 applied to some value arguments. This didn't work:
287 let x = _coerce_ (T Int) Int (I# 3) in
288 case _coerce_ Int (T Int) x of
291 we want to inline x, but can't see that it's a constructor in a case
292 scrutinee position, and some_benefit is False.
296 dMonadST = _/\_ t -> :Monad (g1 _@_ t, g2 _@_ t, g3 _@_ t)
298 .... case dMonadST _@_ x0 of (a,b,c) -> ....
300 we'd really like to inline dMonadST here, but we *don't* want to
301 inline if the case expression is just
303 case x of y { DEFAULT -> ... }
305 since we can just eliminate this case instead (x is in WHNF). Similar
306 applies when x is bound to a lambda expression. Hence
307 contIsInteresting looks for case expressions with just a single
312 interestingCallContext :: SimplCont -> CallCtxt
313 interestingCallContext cont
316 interestingCtxt = ArgCtxt False 2 -- Give *some* incentive!
318 interesting (Select _ bndr _ _ _)
319 | isDeadBinder bndr = CaseCtxt
320 | otherwise = interestingCtxt
322 interesting (ApplyTo {}) = interestingCtxt
323 -- Can happen if we have (coerce t (f x)) y
324 -- Perhaps interestingCtxt is a bit over-keen, but I've
325 -- seen (coerce f) x, where f has an INLINE prag,
326 -- So we have to give some motivation for inlining it
328 interesting (StrictArg _ _ cci _ _) = cci
329 interesting (StrictBind {}) = BoringCtxt
330 interesting (Stop ty cci) = cci
331 interesting (CoerceIt _ cont) = interesting cont
332 -- If this call is the arg of a strict function, the context
333 -- is a bit interesting. If we inline here, we may get useful
334 -- evaluation information to avoid repeated evals: e.g.
336 -- Here the contIsInteresting makes the '*' keener to inline,
337 -- which in turn exposes a constructor which makes the '+' inline.
338 -- Assuming that +,* aren't small enough to inline regardless.
340 -- It's also very important to inline in a strict context for things
343 -- Here, the context of (f x) is strict, and if f's unfolding is
344 -- a build it's *great* to inline it here. So we must ensure that
345 -- the context for (f x) is not totally uninteresting.
350 -> Int -- Number of value args
351 -> SimplCont -- Context of the cal
354 mkArgInfo fun n_val_args call_cont
355 = ArgInfo { ai_rules = interestingArgContext fun call_cont
356 , ai_strs = arg_stricts
357 , ai_discs = arg_discounts }
359 vanilla_discounts, arg_discounts :: [Int]
360 vanilla_discounts = repeat 0
361 arg_discounts = case idUnfolding fun of
362 CoreUnfolding _ _ _ _ (UnfoldIfGoodArgs _ discounts _ _)
363 -> discounts ++ vanilla_discounts
364 other -> vanilla_discounts
366 vanilla_stricts, arg_stricts :: [Bool]
367 vanilla_stricts = repeat False
370 = case splitStrictSig (idNewStrictness fun) of
371 (demands, result_info)
372 | not (demands `lengthExceeds` n_val_args)
373 -> -- Enough args, use the strictness given.
374 -- For bottoming functions we used to pretend that the arg
375 -- is lazy, so that we don't treat the arg as an
376 -- interesting context. This avoids substituting
377 -- top-level bindings for (say) strings into
378 -- calls to error. But now we are more careful about
379 -- inlining lone variables, so its ok (see SimplUtils.analyseCont)
380 if isBotRes result_info then
381 map isStrictDmd demands -- Finite => result is bottom
383 map isStrictDmd demands ++ vanilla_stricts
385 other -> vanilla_stricts -- Not enough args, or no strictness
387 interestingArgContext :: Id -> SimplCont -> Bool
388 -- If the argument has form (f x y), where x,y are boring,
389 -- and f is marked INLINE, then we don't want to inline f.
390 -- But if the context of the argument is
392 -- where g has rules, then we *do* want to inline f, in case it
393 -- exposes a rule that might fire. Similarly, if the context is
395 -- where h has rules, then we do want to inline f; hence the
396 -- call_cont argument to interestingArgContext
398 -- The interesting_arg_ctxt flag makes this happen; if it's
399 -- set, the inliner gets just enough keener to inline f
400 -- regardless of how boring f's arguments are, if it's marked INLINE
402 -- The alternative would be to *always* inline an INLINE function,
403 -- regardless of how boring its context is; but that seems overkill
404 -- For example, it'd mean that wrapper functions were always inlined
405 interestingArgContext fn call_cont
406 = idHasRules fn || go call_cont
408 go (Select {}) = False
409 go (ApplyTo {}) = False
410 go (StrictArg _ _ cci _ _) = interesting cci
411 go (StrictBind {}) = False -- ??
412 go (CoerceIt _ c) = go c
413 go (Stop _ cci) = interesting cci
415 interesting (ArgCtxt rules _) = rules
416 interesting other = False
421 %************************************************************************
423 \subsection{Decisions about inlining}
425 %************************************************************************
427 Inlining is controlled partly by the SimplifierMode switch. This has two
430 SimplGently (a) Simplifying before specialiser/full laziness
431 (b) Simplifiying inside INLINE pragma
432 (c) Simplifying the LHS of a rule
433 (d) Simplifying a GHCi expression or Template
436 SimplPhase n Used at all other times
438 The key thing about SimplGently is that it does no call-site inlining.
439 Before full laziness we must be careful not to inline wrappers,
440 because doing so inhibits floating
441 e.g. ...(case f x of ...)...
442 ==> ...(case (case x of I# x# -> fw x#) of ...)...
443 ==> ...(case x of I# x# -> case fw x# of ...)...
444 and now the redex (f x) isn't floatable any more.
446 The no-inlining thing is also important for Template Haskell. You might be
447 compiling in one-shot mode with -O2; but when TH compiles a splice before
448 running it, we don't want to use -O2. Indeed, we don't want to inline
449 anything, because the byte-code interpreter might get confused about
450 unboxed tuples and suchlike.
454 SimplGently is also used as the mode to simplify inside an InlineMe note.
457 inlineMode :: SimplifierMode
458 inlineMode = SimplGently
461 It really is important to switch off inlinings inside such
462 expressions. Consider the following example
468 in ...g...g...g...g...g...
470 Now, if that's the ONLY occurrence of f, it will be inlined inside g,
471 and thence copied multiple times when g is inlined.
474 This function may be inlinined in other modules, so we
475 don't want to remove (by inlining) calls to functions that have
476 specialisations, or that may have transformation rules in an importing
479 E.g. {-# INLINE f #-}
482 and suppose that g is strict *and* has specialisations. If we inline
483 g's wrapper, we deny f the chance of getting the specialised version
484 of g when f is inlined at some call site (perhaps in some other
487 It's also important not to inline a worker back into a wrapper.
489 wraper = inline_me (\x -> ...worker... )
490 Normally, the inline_me prevents the worker getting inlined into
491 the wrapper (initially, the worker's only call site!). But,
492 if the wrapper is sure to be called, the strictness analyser will
493 mark it 'demanded', so when the RHS is simplified, it'll get an ArgOf
494 continuation. That's why the keep_inline predicate returns True for
495 ArgOf continuations. It shouldn't do any harm not to dissolve the
496 inline-me note under these circumstances.
498 Note that the result is that we do very little simplification
501 all xs = foldr (&&) True xs
502 any p = all . map p {-# INLINE any #-}
504 Problem: any won't get deforested, and so if it's exported and the
505 importer doesn't use the inlining, (eg passes it as an arg) then we
506 won't get deforestation at all. We havn't solved this problem yet!
509 preInlineUnconditionally
510 ~~~~~~~~~~~~~~~~~~~~~~~~
511 @preInlineUnconditionally@ examines a bndr to see if it is used just
512 once in a completely safe way, so that it is safe to discard the
513 binding inline its RHS at the (unique) usage site, REGARDLESS of how
514 big the RHS might be. If this is the case we don't simplify the RHS
515 first, but just inline it un-simplified.
517 This is much better than first simplifying a perhaps-huge RHS and then
518 inlining and re-simplifying it. Indeed, it can be at least quadratically
527 We may end up simplifying e1 N times, e2 N-1 times, e3 N-3 times etc.
528 This can happen with cascades of functions too:
535 THE MAIN INVARIANT is this:
537 ---- preInlineUnconditionally invariant -----
538 IF preInlineUnconditionally chooses to inline x = <rhs>
539 THEN doing the inlining should not change the occurrence
540 info for the free vars of <rhs>
541 ----------------------------------------------
543 For example, it's tempting to look at trivial binding like
545 and inline it unconditionally. But suppose x is used many times,
546 but this is the unique occurrence of y. Then inlining x would change
547 y's occurrence info, which breaks the invariant. It matters: y
548 might have a BIG rhs, which will now be dup'd at every occurrenc of x.
551 Even RHSs labelled InlineMe aren't caught here, because there might be
552 no benefit from inlining at the call site.
554 [Sept 01] Don't unconditionally inline a top-level thing, because that
555 can simply make a static thing into something built dynamically. E.g.
559 [Remember that we treat \s as a one-shot lambda.] No point in
560 inlining x unless there is something interesting about the call site.
562 But watch out: if you aren't careful, some useful foldr/build fusion
563 can be lost (most notably in spectral/hartel/parstof) because the
564 foldr didn't see the build. Doing the dynamic allocation isn't a big
565 deal, in fact, but losing the fusion can be. But the right thing here
566 seems to be to do a callSiteInline based on the fact that there is
567 something interesting about the call site (it's strict). Hmm. That
570 Conclusion: inline top level things gaily until Phase 0 (the last
571 phase), at which point don't.
574 preInlineUnconditionally :: SimplEnv -> TopLevelFlag -> InId -> InExpr -> Bool
575 preInlineUnconditionally env top_lvl bndr rhs
577 | opt_SimplNoPreInlining = False
578 | otherwise = case idOccInfo bndr of
579 IAmDead -> True -- Happens in ((\x.1) v)
580 OneOcc in_lam True int_cxt -> try_once in_lam int_cxt
584 active = case phase of
585 SimplGently -> isAlwaysActive prag
586 SimplPhase n -> isActive n prag
587 prag = idInlinePragma bndr
589 try_once in_lam int_cxt -- There's one textual occurrence
590 | not in_lam = isNotTopLevel top_lvl || early_phase
591 | otherwise = int_cxt && canInlineInLam rhs
593 -- Be very careful before inlining inside a lambda, becuase (a) we must not
594 -- invalidate occurrence information, and (b) we want to avoid pushing a
595 -- single allocation (here) into multiple allocations (inside lambda).
596 -- Inlining a *function* with a single *saturated* call would be ok, mind you.
597 -- || (if is_cheap && not (canInlineInLam rhs) then pprTrace "preinline" (ppr bndr <+> ppr rhs) ok else ok)
599 -- is_cheap = exprIsCheap rhs
600 -- ok = is_cheap && int_cxt
602 -- int_cxt The context isn't totally boring
603 -- E.g. let f = \ab.BIG in \y. map f xs
604 -- Don't want to substitute for f, because then we allocate
605 -- its closure every time the \y is called
606 -- But: let f = \ab.BIG in \y. map (f y) xs
607 -- Now we do want to substitute for f, even though it's not
608 -- saturated, because we're going to allocate a closure for
609 -- (f y) every time round the loop anyhow.
611 -- canInlineInLam => free vars of rhs are (Once in_lam) or Many,
612 -- so substituting rhs inside a lambda doesn't change the occ info.
613 -- Sadly, not quite the same as exprIsHNF.
614 canInlineInLam (Lit l) = True
615 canInlineInLam (Lam b e) = isRuntimeVar b || canInlineInLam e
616 canInlineInLam (Note _ e) = canInlineInLam e
617 canInlineInLam _ = False
619 early_phase = case phase of
620 SimplPhase 0 -> False
622 -- If we don't have this early_phase test, consider
623 -- x = length [1,2,3]
624 -- The full laziness pass carefully floats all the cons cells to
625 -- top level, and preInlineUnconditionally floats them all back in.
626 -- Result is (a) static allocation replaced by dynamic allocation
627 -- (b) many simplifier iterations because this tickles
628 -- a related problem; only one inlining per pass
630 -- On the other hand, I have seen cases where top-level fusion is
631 -- lost if we don't inline top level thing (e.g. string constants)
632 -- Hence the test for phase zero (which is the phase for all the final
633 -- simplifications). Until phase zero we take no special notice of
634 -- top level things, but then we become more leery about inlining
639 postInlineUnconditionally
640 ~~~~~~~~~~~~~~~~~~~~~~~~~
641 @postInlineUnconditionally@ decides whether to unconditionally inline
642 a thing based on the form of its RHS; in particular if it has a
643 trivial RHS. If so, we can inline and discard the binding altogether.
645 NB: a loop breaker has must_keep_binding = True and non-loop-breakers
646 only have *forward* references Hence, it's safe to discard the binding
648 NOTE: This isn't our last opportunity to inline. We're at the binding
649 site right now, and we'll get another opportunity when we get to the
652 Note that we do this unconditional inlining only for trival RHSs.
653 Don't inline even WHNFs inside lambdas; doing so may simply increase
654 allocation when the function is called. This isn't the last chance; see
657 NB: Even inline pragmas (e.g. IMustBeINLINEd) are ignored here Why?
658 Because we don't even want to inline them into the RHS of constructor
659 arguments. See NOTE above
661 NB: At one time even NOINLINE was ignored here: if the rhs is trivial
662 it's best to inline it anyway. We often get a=E; b=a from desugaring,
663 with both a and b marked NOINLINE. But that seems incompatible with
664 our new view that inlining is like a RULE, so I'm sticking to the 'active'
668 postInlineUnconditionally
669 :: SimplEnv -> TopLevelFlag
670 -> InId -- The binder (an OutId would be fine too)
671 -> OccInfo -- From the InId
675 postInlineUnconditionally env top_lvl bndr occ_info rhs unfolding
677 | isLoopBreaker occ_info = False -- If it's a loop-breaker of any kind, dont' inline
678 -- because it might be referred to "earlier"
679 | isExportedId bndr = False
680 | exprIsTrivial rhs = True
683 -- The point of examining occ_info here is that for *non-values*
684 -- that occur outside a lambda, the call-site inliner won't have
685 -- a chance (becuase it doesn't know that the thing
686 -- only occurs once). The pre-inliner won't have gotten
687 -- it either, if the thing occurs in more than one branch
688 -- So the main target is things like
691 -- True -> case x of ...
692 -- False -> case x of ...
693 -- I'm not sure how important this is in practice
694 OneOcc in_lam one_br int_cxt -- OneOcc => no code-duplication issue
695 -> smallEnoughToInline unfolding -- Small enough to dup
696 -- ToDo: consider discount on smallEnoughToInline if int_cxt is true
698 -- NB: Do NOT inline arbitrarily big things, even if one_br is True
699 -- Reason: doing so risks exponential behaviour. We simplify a big
700 -- expression, inline it, and simplify it again. But if the
701 -- very same thing happens in the big expression, we get
703 -- PRINCIPLE: when we've already simplified an expression once,
704 -- make sure that we only inline it if it's reasonably small.
706 && ((isNotTopLevel top_lvl && not in_lam) ||
707 -- But outside a lambda, we want to be reasonably aggressive
708 -- about inlining into multiple branches of case
709 -- e.g. let x = <non-value>
710 -- in case y of { C1 -> ..x..; C2 -> ..x..; C3 -> ... }
711 -- Inlining can be a big win if C3 is the hot-spot, even if
712 -- the uses in C1, C2 are not 'interesting'
713 -- An example that gets worse if you add int_cxt here is 'clausify'
715 (isCheapUnfolding unfolding && int_cxt))
716 -- isCheap => acceptable work duplication; in_lam may be true
717 -- int_cxt to prevent us inlining inside a lambda without some
718 -- good reason. See the notes on int_cxt in preInlineUnconditionally
720 IAmDead -> True -- This happens; for example, the case_bndr during case of
721 -- known constructor: case (a,b) of x { (p,q) -> ... }
722 -- Here x isn't mentioned in the RHS, so we don't want to
723 -- create the (dead) let-binding let x = (a,b) in ...
727 -- Here's an example that we don't handle well:
728 -- let f = if b then Left (\x.BIG) else Right (\y.BIG)
729 -- in \y. ....case f of {...} ....
730 -- Here f is used just once, and duplicating the case work is fine (exprIsCheap).
732 -- * We can't preInlineUnconditionally because that woud invalidate
733 -- the occ info for b.
734 -- * We can't postInlineUnconditionally because the RHS is big, and
735 -- that risks exponential behaviour
736 -- * We can't call-site inline, because the rhs is big
740 active = case getMode env of
741 SimplGently -> isAlwaysActive prag
742 SimplPhase n -> isActive n prag
743 prag = idInlinePragma bndr
745 activeInline :: SimplEnv -> OutId -> Bool
747 = case getMode env of
749 -- No inlining at all when doing gentle stuff,
750 -- except for local things that occur once
751 -- The reason is that too little clean-up happens if you
752 -- don't inline use-once things. Also a bit of inlining is *good* for
753 -- full laziness; it can expose constant sub-expressions.
754 -- Example in spectral/mandel/Mandel.hs, where the mandelset
755 -- function gets a useful let-float if you inline windowToViewport
757 -- NB: we used to have a second exception, for data con wrappers.
758 -- On the grounds that we use gentle mode for rule LHSs, and
759 -- they match better when data con wrappers are inlined.
760 -- But that only really applies to the trivial wrappers (like (:)),
761 -- and they are now constructed as Compulsory unfoldings (in MkId)
762 -- so they'll happen anyway.
764 SimplPhase n -> isActive n prag
766 prag = idInlinePragma id
768 activeRule :: DynFlags -> SimplEnv -> Maybe (Activation -> Bool)
769 -- Nothing => No rules at all
770 activeRule dflags env
771 | not (dopt Opt_RewriteRules dflags)
772 = Nothing -- Rewriting is off
774 = case getMode env of
775 SimplGently -> Just isAlwaysActive
776 -- Used to be Nothing (no rules in gentle mode)
777 -- Main motivation for changing is that I wanted
778 -- lift String ===> ...
779 -- to work in Template Haskell when simplifying
780 -- splices, so we get simpler code for literal strings
781 SimplPhase n -> Just (isActive n)
785 %************************************************************************
789 %************************************************************************
792 mkLam :: [OutBndr] -> OutExpr -> SimplM OutExpr
793 -- mkLam tries three things
794 -- a) eta reduction, if that gives a trivial expression
795 -- b) eta expansion [only if there are some value lambdas]
800 = do { dflags <- getDOptsSmpl
801 ; mkLam' dflags bndrs body }
803 mkLam' :: DynFlags -> [OutBndr] -> OutExpr -> SimplM OutExpr
804 mkLam' dflags bndrs (Cast body co)
805 | not (any bad bndrs)
806 -- Note [Casts and lambdas]
807 = do { lam <- mkLam' dflags bndrs body
808 ; return (mkCoerce (mkPiTypes bndrs co) lam) }
810 co_vars = tyVarsOfType co
811 bad bndr = isCoVar bndr && bndr `elemVarSet` co_vars
813 mkLam' dflags bndrs body
814 | dopt Opt_DoEtaReduction dflags,
815 Just etad_lam <- tryEtaReduce bndrs body
816 = do { tick (EtaReduction (head bndrs))
819 | dopt Opt_DoLambdaEtaExpansion dflags,
820 any isRuntimeVar bndrs
821 = do { body' <- tryEtaExpansion dflags body
822 ; return (mkLams bndrs body') }
825 = return (mkLams bndrs body)
828 Note [Casts and lambdas]
829 ~~~~~~~~~~~~~~~~~~~~~~~~
831 (\x. (\y. e) `cast` g1) `cast` g2
832 There is a danger here that the two lambdas look separated, and the
833 full laziness pass might float an expression to between the two.
835 So this equation in mkLam' floats the g1 out, thus:
836 (\x. e `cast` g1) --> (\x.e) `cast` (tx -> g1)
839 In general, this floats casts outside lambdas, where (I hope) they
840 might meet and cancel with some other cast:
841 \x. e `cast` co ===> (\x. e) `cast` (tx -> co)
842 /\a. e `cast` co ===> (/\a. e) `cast` (/\a. co)
843 /\g. e `cast` co ===> (/\g. e) `cast` (/\g. co)
846 Notice that it works regardless of 'e'. Originally it worked only
847 if 'e' was itself a lambda, but in some cases that resulted in
848 fruitless iteration in the simplifier. A good example was when
849 compiling Text.ParserCombinators.ReadPrec, where we had a definition
850 like (\x. Get `cast` g)
851 where Get is a constructor with nonzero arity. Then mkLam eta-expanded
852 the Get, and the next iteration eta-reduced it, and then eta-expanded
855 Note also the side condition for the case of coercion binders.
856 It does not make sense to transform
857 /\g. e `cast` g ==> (/\g.e) `cast` (/\g.g)
858 because the latter is not well-kinded.
860 -- c) floating lets out through big lambdas
861 -- [only if all tyvar lambdas, and only if this lambda
862 -- is the RHS of a let]
864 {- Sept 01: I'm experimenting with getting the
865 full laziness pass to float out past big lambdsa
866 | all isTyVar bndrs, -- Only for big lambdas
867 contIsRhs cont -- Only try the rhs type-lambda floating
868 -- if this is indeed a right-hand side; otherwise
869 -- we end up floating the thing out, only for float-in
870 -- to float it right back in again!
871 = do (floats, body') <- tryRhsTyLam env bndrs body
872 return (floats, mkLams bndrs body')
876 %************************************************************************
880 %************************************************************************
882 Note [Eta reduction conditions]
883 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
884 We try for eta reduction here, but *only* if we get all the way to an
885 trivial expression. We don't want to remove extra lambdas unless we
886 are going to avoid allocating this thing altogether.
888 There are some particularly delicate points here:
890 * Eta reduction is not valid in general:
892 This matters, partly for old-fashioned correctness reasons but,
893 worse, getting it wrong can yield a seg fault. Consider
895 h y = case (case y of { True -> f `seq` True; False -> False }) of
896 True -> ...; False -> ...
898 If we (unsoundly) eta-reduce f to get f=f, the strictness analyser
899 says f=bottom, and replaces the (f `seq` True) with just
900 (f `cast` unsafe-co). BUT, as thing stand, 'f' got arity 1, and it
901 *keeps* arity 1 (perhaps also wrongly). So CorePrep eta-expands
902 the definition again, so that it does not termninate after all.
903 Result: seg-fault because the boolean case actually gets a function value.
906 So it's important to to the right thing.
908 * We need to be careful if we just look at f's arity. Currently (Dec07),
909 f's arity is visible in its own RHS (see Note [Arity robustness] in
910 SimplEnv) so we must *not* trust the arity when checking that 'f' is
911 a value. Instead, look at the unfolding.
913 However for GlobalIds we can look at the arity; and for primops we
914 must, since they have no unfolding.
916 * Regardless of whether 'f' is a vlaue, we always want to
917 reduce (/\a -> f a) to f
918 This came up in a RULE: foldr (build (/\a -> g a))
919 did not match foldr (build (/\b -> ...something complex...))
920 The type checker can insert these eta-expanded versions,
921 with both type and dictionary lambdas; hence the slightly
924 These delicacies are why we don't use exprIsTrivial and exprIsHNF here.
928 tryEtaReduce :: [OutBndr] -> OutExpr -> Maybe OutExpr
929 tryEtaReduce bndrs body
930 = go (reverse bndrs) body
932 go (b : bs) (App fun arg) | ok_arg b arg = go bs fun -- Loop round
933 go [] fun | ok_fun fun = Just fun -- Success!
934 go _ _ = Nothing -- Failure!
936 -- Note [Eta reduction conditions]
937 ok_fun (App fun (Type ty))
938 | not (any (`elemVarSet` tyVarsOfType ty) bndrs)
941 = not (fun_id `elem` bndrs)
942 && (ok_fun_id fun_id || all ok_lam bndrs)
946 | isLocalId fun = isEvaldUnfolding (idUnfolding fun)
947 | isDataConWorkId fun = True
948 | isGlobalId fun = idArity fun > 0
950 ok_lam v = isTyVar v || isDictId v
952 ok_arg b arg = varToCoreExpr b `cheapEqExpr` arg
956 %************************************************************************
960 %************************************************************************
964 f = \x1..xn -> N ==> f = \x1..xn y1..ym -> N y1..ym
967 where (in both cases)
969 * The xi can include type variables
971 * The yi are all value variables
973 * N is a NORMAL FORM (i.e. no redexes anywhere)
974 wanting a suitable number of extra args.
976 The biggest reason for doing this is for cases like
982 Here we want to get the lambdas together. A good exmaple is the nofib
983 program fibheaps, which gets 25% more allocation if you don't do this
986 We may have to sandwich some coerces between the lambdas
987 to make the types work. exprEtaExpandArity looks through coerces
988 when computing arity; and etaExpand adds the coerces as necessary when
989 actually computing the expansion.
992 tryEtaExpansion :: DynFlags -> OutExpr -> SimplM OutExpr
993 -- There is at least one runtime binder in the binders
994 tryEtaExpansion dflags body = do
996 return (etaExpand fun_arity us body (exprType body))
998 fun_arity = exprEtaExpandArity dflags body
1002 %************************************************************************
1004 \subsection{Floating lets out of big lambdas}
1006 %************************************************************************
1008 Note [Floating and type abstraction]
1009 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
1012 We'd like to float this to
1015 x = /\a. C (y1 a) (y2 a)
1016 for the usual reasons: we want to inline x rather vigorously.
1018 You may think that this kind of thing is rare. But in some programs it is
1019 common. For example, if you do closure conversion you might get:
1021 data a :-> b = forall e. (e -> a -> b) :$ e
1023 f_cc :: forall a. a :-> a
1024 f_cc = /\a. (\e. id a) :$ ()
1026 Now we really want to inline that f_cc thing so that the
1027 construction of the closure goes away.
1029 So I have elaborated simplLazyBind to understand right-hand sides that look
1033 and treat them specially. The real work is done in SimplUtils.abstractFloats,
1034 but there is quite a bit of plumbing in simplLazyBind as well.
1036 The same transformation is good when there are lets in the body:
1038 /\abc -> let(rec) x = e in b
1040 let(rec) x' = /\abc -> let x = x' a b c in e
1042 /\abc -> let x = x' a b c in b
1044 This is good because it can turn things like:
1046 let f = /\a -> letrec g = ... g ... in g
1048 letrec g' = /\a -> ... g' a ...
1050 let f = /\ a -> g' a
1052 which is better. In effect, it means that big lambdas don't impede
1055 This optimisation is CRUCIAL in eliminating the junk introduced by
1056 desugaring mutually recursive definitions. Don't eliminate it lightly!
1058 [May 1999] If we do this transformation *regardless* then we can
1059 end up with some pretty silly stuff. For example,
1062 st = /\ s -> let { x1=r1 ; x2=r2 } in ...
1067 st = /\s -> ...[y1 s/x1, y2 s/x2]
1070 Unless the "..." is a WHNF there is really no point in doing this.
1071 Indeed it can make things worse. Suppose x1 is used strictly,
1074 x1* = case f y of { (a,b) -> e }
1076 If we abstract this wrt the tyvar we then can't do the case inline
1077 as we would normally do.
1079 That's why the whole transformation is part of the same process that
1080 floats let-bindings and constructor arguments out of RHSs. In particular,
1081 it is guarded by the doFloatFromRhs call in simplLazyBind.
1085 abstractFloats :: [OutTyVar] -> SimplEnv -> OutExpr -> SimplM ([OutBind], OutExpr)
1086 abstractFloats main_tvs body_env body
1087 = ASSERT( notNull body_floats )
1088 do { (subst, float_binds) <- mapAccumLM abstract empty_subst body_floats
1089 ; return (float_binds, CoreSubst.substExpr subst body) }
1091 main_tv_set = mkVarSet main_tvs
1092 body_floats = getFloats body_env
1093 empty_subst = CoreSubst.mkEmptySubst (seInScope body_env)
1095 abstract :: CoreSubst.Subst -> OutBind -> SimplM (CoreSubst.Subst, OutBind)
1096 abstract subst (NonRec id rhs)
1097 = do { (poly_id, poly_app) <- mk_poly tvs_here id
1098 ; let poly_rhs = mkLams tvs_here rhs'
1099 subst' = CoreSubst.extendIdSubst subst id poly_app
1100 ; return (subst', (NonRec poly_id poly_rhs)) }
1102 rhs' = CoreSubst.substExpr subst rhs
1103 tvs_here | any isCoVar main_tvs = main_tvs -- Note [Abstract over coercions]
1105 = varSetElems (main_tv_set `intersectVarSet` exprSomeFreeVars isTyVar rhs')
1107 -- Abstract only over the type variables free in the rhs
1108 -- wrt which the new binding is abstracted. But the naive
1109 -- approach of abstract wrt the tyvars free in the Id's type
1111 -- /\ a b -> let t :: (a,b) = (e1, e2)
1114 -- Here, b isn't free in x's type, but we must nevertheless
1115 -- abstract wrt b as well, because t's type mentions b.
1116 -- Since t is floated too, we'd end up with the bogus:
1117 -- poly_t = /\ a b -> (e1, e2)
1118 -- poly_x = /\ a -> fst (poly_t a *b*)
1119 -- So for now we adopt the even more naive approach of
1120 -- abstracting wrt *all* the tyvars. We'll see if that
1121 -- gives rise to problems. SLPJ June 98
1123 abstract subst (Rec prs)
1124 = do { (poly_ids, poly_apps) <- mapAndUnzipM (mk_poly tvs_here) ids
1125 ; let subst' = CoreSubst.extendSubstList subst (ids `zip` poly_apps)
1126 poly_rhss = [mkLams tvs_here (CoreSubst.substExpr subst' rhs) | rhs <- rhss]
1127 ; return (subst', Rec (poly_ids `zip` poly_rhss)) }
1129 (ids,rhss) = unzip prs
1130 -- For a recursive group, it's a bit of a pain to work out the minimal
1131 -- set of tyvars over which to abstract:
1132 -- /\ a b c. let x = ...a... in
1133 -- letrec { p = ...x...q...
1134 -- q = .....p...b... } in
1136 -- Since 'x' is abstracted over 'a', the {p,q} group must be abstracted
1137 -- over 'a' (because x is replaced by (poly_x a)) as well as 'b'.
1138 -- Since it's a pain, we just use the whole set, which is always safe
1140 -- If you ever want to be more selective, remember this bizarre case too:
1142 -- Here, we must abstract 'x' over 'a'.
1145 mk_poly tvs_here var
1146 = do { uniq <- getUniqueM
1147 ; let poly_name = setNameUnique (idName var) uniq -- Keep same name
1148 poly_ty = mkForAllTys tvs_here (idType var) -- But new type of course
1149 poly_id = mkLocalId poly_name poly_ty
1150 ; return (poly_id, mkTyApps (Var poly_id) (mkTyVarTys tvs_here)) }
1151 -- In the olden days, it was crucial to copy the occInfo of the original var,
1152 -- because we were looking at occurrence-analysed but as yet unsimplified code!
1153 -- In particular, we mustn't lose the loop breakers. BUT NOW we are looking
1154 -- at already simplified code, so it doesn't matter
1156 -- It's even right to retain single-occurrence or dead-var info:
1157 -- Suppose we started with /\a -> let x = E in B
1158 -- where x occurs once in B. Then we transform to:
1159 -- let x' = /\a -> E in /\a -> let x* = x' a in B
1160 -- where x* has an INLINE prag on it. Now, once x* is inlined,
1161 -- the occurrences of x' will be just the occurrences originally
1165 Note [Abstract over coercions]
1166 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
1167 If a coercion variable (g :: a ~ Int) is free in the RHS, then so is the
1168 type variable a. Rather than sort this mess out, we simply bale out and abstract
1169 wrt all the type variables if any of them are coercion variables.
1172 Historical note: if you use let-bindings instead of a substitution, beware of this:
1174 -- Suppose we start with:
1176 -- x = /\ a -> let g = G in E
1178 -- Then we'll float to get
1180 -- x = let poly_g = /\ a -> G
1181 -- in /\ a -> let g = poly_g a in E
1183 -- But now the occurrence analyser will see just one occurrence
1184 -- of poly_g, not inside a lambda, so the simplifier will
1185 -- PreInlineUnconditionally poly_g back into g! Badk to square 1!
1186 -- (I used to think that the "don't inline lone occurrences" stuff
1187 -- would stop this happening, but since it's the *only* occurrence,
1188 -- PreInlineUnconditionally kicks in first!)
1190 -- Solution: put an INLINE note on g's RHS, so that poly_g seems
1191 -- to appear many times. (NB: mkInlineMe eliminates
1192 -- such notes on trivial RHSs, so do it manually.)
1194 %************************************************************************
1198 %************************************************************************
1200 prepareAlts tries these things:
1202 1. If several alternatives are identical, merge them into
1203 a single DEFAULT alternative. I've occasionally seen this
1204 making a big difference:
1206 case e of =====> case e of
1207 C _ -> f x D v -> ....v....
1208 D v -> ....v.... DEFAULT -> f x
1211 The point is that we merge common RHSs, at least for the DEFAULT case.
1212 [One could do something more elaborate but I've never seen it needed.]
1213 To avoid an expensive test, we just merge branches equal to the *first*
1214 alternative; this picks up the common cases
1215 a) all branches equal
1216 b) some branches equal to the DEFAULT (which occurs first)
1219 case e of b { ==> case e of b {
1220 p1 -> rhs1 p1 -> rhs1
1222 pm -> rhsm pm -> rhsm
1223 _ -> case b of b' { pn -> let b'=b in rhsn
1225 ... po -> let b'=b in rhso
1226 po -> rhso _ -> let b'=b in rhsd
1230 which merges two cases in one case when -- the default alternative of
1231 the outer case scrutises the same variable as the outer case This
1232 transformation is called Case Merging. It avoids that the same
1233 variable is scrutinised multiple times.
1236 The case where transformation (1) showed up was like this (lib/std/PrelCError.lhs):
1242 where @is@ was something like
1244 p `is` n = p /= (-1) && p == n
1246 This gave rise to a horrible sequence of cases
1253 and similarly in cascade for all the join points!
1256 ~~~~~~~~~~~~~~~~~~~~
1257 We do this *here*, looking at un-simplified alternatives, because we
1258 have to check that r doesn't mention the variables bound by the
1259 pattern in each alternative, so the binder-info is rather useful.
1262 prepareAlts :: SimplEnv -> OutExpr -> OutId -> [InAlt] -> SimplM ([AltCon], [InAlt])
1263 prepareAlts env scrut case_bndr' alts
1264 = do { dflags <- getDOptsSmpl
1265 ; alts <- combineIdenticalAlts case_bndr' alts
1267 ; let (alts_wo_default, maybe_deflt) = findDefault alts
1268 alt_cons = [con | (con,_,_) <- alts_wo_default]
1269 imposs_deflt_cons = nub (imposs_cons ++ alt_cons)
1270 -- "imposs_deflt_cons" are handled
1271 -- EITHER by the context,
1272 -- OR by a non-DEFAULT branch in this case expression.
1274 ; default_alts <- prepareDefault dflags env case_bndr' mb_tc_app
1275 imposs_deflt_cons maybe_deflt
1277 ; let trimmed_alts = filterOut impossible_alt alts_wo_default
1278 merged_alts = mergeAlts trimmed_alts default_alts
1279 -- We need the mergeAlts in case the new default_alt
1280 -- has turned into a constructor alternative.
1281 -- The merge keeps the inner DEFAULT at the front, if there is one
1282 -- and interleaves the alternatives in the right order
1284 ; return (imposs_deflt_cons, merged_alts) }
1286 mb_tc_app = splitTyConApp_maybe (idType case_bndr')
1287 Just (_, inst_tys) = mb_tc_app
1289 imposs_cons = case scrut of
1290 Var v -> otherCons (idUnfolding v)
1293 impossible_alt :: CoreAlt -> Bool
1294 impossible_alt (con, _, _) | con `elem` imposs_cons = True
1295 impossible_alt (DataAlt con, _, _) = dataConCannotMatch inst_tys con
1296 impossible_alt alt = False
1299 --------------------------------------------------
1300 -- 1. Merge identical branches
1301 --------------------------------------------------
1302 combineIdenticalAlts :: OutId -> [InAlt] -> SimplM [InAlt]
1304 combineIdenticalAlts case_bndr alts@((con1,bndrs1,rhs1) : con_alts)
1305 | all isDeadBinder bndrs1, -- Remember the default
1306 length filtered_alts < length con_alts -- alternative comes first
1307 -- Also Note [Dead binders]
1308 = do { tick (AltMerge case_bndr)
1309 ; return ((DEFAULT, [], rhs1) : filtered_alts) }
1311 filtered_alts = filter keep con_alts
1312 keep (con,bndrs,rhs) = not (all isDeadBinder bndrs && rhs `cheapEqExpr` rhs1)
1314 combineIdenticalAlts case_bndr alts = return alts
1316 -------------------------------------------------------------------------
1317 -- Prepare the default alternative
1318 -------------------------------------------------------------------------
1319 prepareDefault :: DynFlags
1321 -> OutId -- Case binder; need just for its type. Note that as an
1322 -- OutId, it has maximum information; this is important.
1323 -- Test simpl013 is an example
1324 -> Maybe (TyCon, [Type]) -- Type of scrutinee, decomposed
1325 -> [AltCon] -- These cons can't happen when matching the default
1326 -> Maybe InExpr -- Rhs
1327 -> SimplM [InAlt] -- Still unsimplified
1328 -- We use a list because it's what mergeAlts expects,
1329 -- And becuase case-merging can cause many to show up
1331 ------- Merge nested cases ----------
1332 prepareDefault dflags env outer_bndr bndr_ty imposs_cons (Just deflt_rhs)
1333 | dopt Opt_CaseMerge dflags
1334 , Case (Var inner_scrut_var) inner_bndr _ inner_alts <- deflt_rhs
1335 , DoneId inner_scrut_var' <- substId env inner_scrut_var
1336 -- Remember, inner_scrut_var is an InId, but outer_bndr is an OutId
1337 , inner_scrut_var' == outer_bndr
1338 -- NB: the substId means that if the outer scrutinee was a
1339 -- variable, and inner scrutinee is the same variable,
1340 -- then inner_scrut_var' will be outer_bndr
1341 -- via the magic of simplCaseBinder
1342 = do { tick (CaseMerge outer_bndr)
1344 ; let munge_rhs rhs = bindCaseBndr inner_bndr (Var outer_bndr) rhs
1345 ; return [(con, args, munge_rhs rhs) | (con, args, rhs) <- inner_alts,
1346 not (con `elem` imposs_cons) ]
1347 -- NB: filter out any imposs_cons. Example:
1350 -- DEFAULT -> case x of
1353 -- When we merge, we must ensure that e1 takes
1354 -- precedence over e2 as the value for A!
1356 -- Warning: don't call prepareAlts recursively!
1357 -- Firstly, there's no point, because inner alts have already had
1358 -- mkCase applied to them, so they won't have a case in their default
1359 -- Secondly, if you do, you get an infinite loop, because the bindCaseBndr
1360 -- in munge_rhs may put a case into the DEFAULT branch!
1363 --------- Fill in known constructor -----------
1364 prepareDefault dflags env case_bndr (Just (tycon, inst_tys)) imposs_cons (Just deflt_rhs)
1365 | -- This branch handles the case where we are
1366 -- scrutinisng an algebraic data type
1367 isAlgTyCon tycon -- It's a data type, tuple, or unboxed tuples.
1368 , not (isNewTyCon tycon) -- We can have a newtype, if we are just doing an eval:
1369 -- case x of { DEFAULT -> e }
1370 -- and we don't want to fill in a default for them!
1371 , Just all_cons <- tyConDataCons_maybe tycon
1372 , not (null all_cons) -- This is a tricky corner case. If the data type has no constructors,
1373 -- which GHC allows, then the case expression will have at most a default
1374 -- alternative. We don't want to eliminate that alternative, because the
1375 -- invariant is that there's always one alternative. It's more convenient
1377 -- case x of { DEFAULT -> e }
1378 -- as it is, rather than transform it to
1379 -- error "case cant match"
1380 -- which would be quite legitmate. But it's a really obscure corner, and
1381 -- not worth wasting code on.
1382 , let imposs_data_cons = [con | DataAlt con <- imposs_cons] -- We now know it's a data type
1383 impossible con = con `elem` imposs_data_cons || dataConCannotMatch inst_tys con
1384 = case filterOut impossible all_cons of
1385 [] -> return [] -- Eliminate the default alternative
1386 -- altogether if it can't match
1388 [con] -> -- It matches exactly one constructor, so fill it in
1389 do { tick (FillInCaseDefault case_bndr)
1391 ; let (ex_tvs, co_tvs, arg_ids) =
1392 dataConRepInstPat us con inst_tys
1393 ; return [(DataAlt con, ex_tvs ++ co_tvs ++ arg_ids, deflt_rhs)] }
1395 two_or_more -> return [(DEFAULT, [], deflt_rhs)]
1397 --------- Catch-all cases -----------
1398 prepareDefault dflags env case_bndr bndr_ty imposs_cons (Just deflt_rhs)
1399 = return [(DEFAULT, [], deflt_rhs)]
1401 prepareDefault dflags env case_bndr bndr_ty imposs_cons Nothing
1402 = return [] -- No default branch
1407 =================================================================================
1409 mkCase tries these things
1411 1. Eliminate the case altogether if possible
1419 and similar friends.
1423 mkCase :: OutExpr -> OutId -> OutType
1424 -> [OutAlt] -- Increasing order
1427 --------------------------------------------------
1428 -- 1. Check for empty alternatives
1429 --------------------------------------------------
1431 -- This isn't strictly an error. It's possible that the simplifer might "see"
1432 -- that an inner case has no accessible alternatives before it "sees" that the
1433 -- entire branch of an outer case is inaccessible. So we simply
1434 -- put an error case here insteadd
1435 mkCase scrut case_bndr ty []
1436 = pprTrace "mkCase: null alts" (ppr case_bndr <+> ppr scrut) $
1437 return (mkApps (Var rUNTIME_ERROR_ID)
1438 [Type ty, Lit (mkStringLit "Impossible alternative")])
1441 --------------------------------------------------
1443 --------------------------------------------------
1445 mkCase scrut case_bndr ty alts -- Identity case
1446 | all identity_alt alts
1447 = do tick (CaseIdentity case_bndr)
1448 return (re_cast scrut)
1450 identity_alt (con, args, rhs) = check_eq con args (de_cast rhs)
1452 check_eq DEFAULT _ (Var v) = v == case_bndr
1453 check_eq (LitAlt lit') _ (Lit lit) = lit == lit'
1454 check_eq (DataAlt con) args rhs = rhs `cheapEqExpr` mkConApp con (arg_tys ++ varsToCoreExprs args)
1455 || rhs `cheapEqExpr` Var case_bndr
1456 check_eq con args rhs = False
1458 arg_tys = map Type (tyConAppArgs (idType case_bndr))
1461 -- case e of x { _ -> x `cast` c }
1462 -- And we definitely want to eliminate this case, to give
1464 -- So we throw away the cast from the RHS, and reconstruct
1465 -- it at the other end. All the RHS casts must be the same
1466 -- if (all identity_alt alts) holds.
1468 -- Don't worry about nested casts, because the simplifier combines them
1469 de_cast (Cast e _) = e
1472 re_cast scrut = case head alts of
1473 (_,_,Cast _ co) -> Cast scrut co
1478 --------------------------------------------------
1480 --------------------------------------------------
1481 mkCase scrut bndr ty alts = return (Case scrut bndr ty alts)
1485 When adding auxiliary bindings for the case binder, it's worth checking if
1486 its dead, because it often is, and occasionally these mkCase transformations
1487 cascade rather nicely.
1490 bindCaseBndr bndr rhs body
1491 | isDeadBinder bndr = body
1492 | otherwise = bindNonRec bndr rhs body