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 )
67 %************************************************************************
71 %************************************************************************
73 A SimplCont allows the simplifier to traverse the expression in a
74 zipper-like fashion. The SimplCont represents the rest of the expression,
75 "above" the point of interest.
77 You can also think of a SimplCont as an "evaluation context", using
78 that term in the way it is used for operational semantics. This is the
79 way I usually think of it, For example you'll often see a syntax for
80 evaluation context looking like
81 C ::= [] | C e | case C of alts | C `cast` co
82 That's the kind of thing we are doing here, and I use that syntax in
87 * A SimplCont describes a *strict* context (just like
88 evaluation contexts do). E.g. Just [] is not a SimplCont
90 * A SimplCont describes a context that *does not* bind
91 any variables. E.g. \x. [] is not a SimplCont
95 = Stop -- An empty context, or hole, []
96 OutType -- Type of the result
97 CallCtxt -- True <=> There is something interesting about
98 -- the context, and hence the inliner
99 -- should be a bit keener (see interestingCallContext)
101 -- This is an argument of a function that has RULES
102 -- Inlining the call might allow the rule to fire
104 | CoerceIt -- C `cast` co
105 OutCoercion -- The coercion simplified
110 InExpr SimplEnv -- The argument and its static env
113 | Select -- case C of alts
115 InId [InAlt] SimplEnv -- The case binder, alts, and subst-env
118 -- The two strict forms have no DupFlag, because we never duplicate them
119 | StrictBind -- (\x* \xs. e) C
120 InId [InBndr] -- let x* = [] in e
121 InExpr SimplEnv -- is a special case
125 OutExpr OutType -- e and its type
126 CallCtxt -- Whether *this* argument position is interesting
127 ArgInfo -- Whether the function at the head of e has rules, etc
128 SimplCont -- plus strictness flags for *further* args
132 ai_rules :: Bool, -- Function has rules (recursively)
133 -- => be keener to inline in all args
134 ai_strs :: [Bool], -- Strictness of arguments
135 -- Usually infinite, but if it is finite it guarantees
136 -- that the function diverges after being given
137 -- that number of args
138 ai_discs :: [Int] -- Discounts for arguments; non-zero => be keener to inline
142 instance Outputable SimplCont where
143 ppr (Stop ty _) = ptext SLIT("Stop") <+> ppr ty
144 ppr (ApplyTo dup arg se cont) = ((ptext SLIT("ApplyTo") <+> ppr dup <+> pprParendExpr arg)
145 {- $$ nest 2 (pprSimplEnv se) -}) $$ ppr cont
146 ppr (StrictBind b _ _ _ cont) = (ptext SLIT("StrictBind") <+> ppr b) $$ ppr cont
147 ppr (StrictArg f _ _ _ cont) = (ptext SLIT("StrictArg") <+> ppr f) $$ ppr cont
148 ppr (Select dup bndr alts se cont) = (ptext SLIT("Select") <+> ppr dup <+> ppr bndr) $$
149 (nest 4 (ppr alts)) $$ ppr cont
150 ppr (CoerceIt co cont) = (ptext SLIT("CoerceIt") <+> ppr co) $$ ppr cont
152 data DupFlag = OkToDup | NoDup
154 instance Outputable DupFlag where
155 ppr OkToDup = ptext SLIT("ok")
156 ppr NoDup = ptext SLIT("nodup")
161 mkBoringStop :: OutType -> SimplCont
162 mkBoringStop ty = Stop ty BoringCtxt
164 mkLazyArgStop :: OutType -> CallCtxt -> SimplCont
165 mkLazyArgStop ty cci = Stop ty cci
167 mkRhsStop :: OutType -> SimplCont
168 mkRhsStop ty = Stop ty BoringCtxt
171 contIsRhsOrArg (Stop {}) = True
172 contIsRhsOrArg (StrictBind {}) = True
173 contIsRhsOrArg (StrictArg {}) = True
174 contIsRhsOrArg other = False
177 contIsDupable :: SimplCont -> Bool
178 contIsDupable (Stop {}) = True
179 contIsDupable (ApplyTo OkToDup _ _ _) = True
180 contIsDupable (Select OkToDup _ _ _ _) = True
181 contIsDupable (CoerceIt _ cont) = contIsDupable cont
182 contIsDupable other = False
185 contIsTrivial :: SimplCont -> Bool
186 contIsTrivial (Stop {}) = True
187 contIsTrivial (ApplyTo _ (Type _) _ cont) = contIsTrivial cont
188 contIsTrivial (CoerceIt _ cont) = contIsTrivial cont
189 contIsTrivial other = False
192 contResultType :: SimplCont -> OutType
193 contResultType (Stop to_ty _) = to_ty
194 contResultType (StrictArg _ _ _ _ cont) = contResultType cont
195 contResultType (StrictBind _ _ _ _ cont) = contResultType cont
196 contResultType (ApplyTo _ _ _ cont) = contResultType cont
197 contResultType (CoerceIt _ cont) = contResultType cont
198 contResultType (Select _ _ _ _ cont) = contResultType cont
201 countValArgs :: SimplCont -> Int
202 countValArgs (ApplyTo _ (Type ty) se cont) = countValArgs cont
203 countValArgs (ApplyTo _ val_arg se cont) = 1 + countValArgs cont
204 countValArgs other = 0
206 countArgs :: SimplCont -> Int
207 countArgs (ApplyTo _ arg se cont) = 1 + countArgs cont
210 contArgs :: SimplCont -> ([OutExpr], SimplCont)
211 -- Uses substitution to turn each arg into an OutExpr
212 contArgs cont = go [] cont
214 go args (ApplyTo _ arg se cont) = go (substExpr se arg : args) cont
215 go args cont = (reverse args, cont)
217 dropArgs :: Int -> SimplCont -> SimplCont
218 dropArgs 0 cont = cont
219 dropArgs n (ApplyTo _ _ _ cont) = dropArgs (n-1) cont
220 dropArgs n other = pprPanic "dropArgs" (ppr n <+> ppr other)
223 splitInlineCont :: SimplCont -> Maybe (SimplCont, SimplCont)
224 -- Returns Nothing if the continuation should dissolve an InlineMe Note
225 -- Return Just (c1,c2) otherwise,
226 -- where c1 is the continuation to put inside the InlineMe
229 -- Example: (__inline_me__ (/\a. e)) ty
230 -- Here we want to do the beta-redex without dissolving the InlineMe
231 -- See test simpl017 (and Trac #1627) for a good example of why this is important
233 splitInlineCont (ApplyTo dup (Type ty) se c)
234 | Just (c1, c2) <- splitInlineCont c = Just (ApplyTo dup (Type ty) se c1, c2)
235 splitInlineCont cont@(Stop ty _) = Just (mkBoringStop ty, cont)
236 splitInlineCont cont@(StrictBind bndr _ _ se _) = Just (mkBoringStop (substTy se (idType bndr)), cont)
237 splitInlineCont cont@(StrictArg _ fun_ty _ _ _) = Just (mkBoringStop (funArgTy fun_ty), cont)
238 splitInlineCont other = Nothing
239 -- NB: the calculation of the type for mkBoringStop is an annoying
240 -- duplication of the same calucation in mkDupableCont
245 interestingArg :: OutExpr -> Bool
246 -- An argument is interesting if it has *some* structure
247 -- We are here trying to avoid unfolding a function that
248 -- is applied only to variables that have no unfolding
249 -- (i.e. they are probably lambda bound): f x y z
250 -- There is little point in inlining f here.
251 interestingArg (Var v) = hasSomeUnfolding (idUnfolding v)
252 -- Was: isValueUnfolding (idUnfolding v')
253 -- But that seems over-pessimistic
255 -- This accounts for an argument like
256 -- () or [], which is definitely interesting
257 interestingArg (Type _) = False
258 interestingArg (App fn (Type _)) = interestingArg fn
259 interestingArg (Note _ a) = interestingArg a
261 -- Idea (from Sam B); I'm not sure if it's a good idea, so commented out for now
262 -- interestingArg expr | isUnLiftedType (exprType expr)
263 -- -- Unlifted args are only ever interesting if we know what they are
268 interestingArg other = True
269 -- Consider let x = 3 in f x
270 -- The substitution will contain (x -> ContEx 3), and we want to
271 -- to say that x is an interesting argument.
272 -- But consider also (\x. f x y) y
273 -- The substitution will contain (x -> ContEx y), and we want to say
274 -- that x is not interesting (assuming y has no unfolding)
278 Comment about interestingCallContext
279 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
280 We want to avoid inlining an expression where there can't possibly be
281 any gain, such as in an argument position. Hence, if the continuation
282 is interesting (eg. a case scrutinee, application etc.) then we
283 inline, otherwise we don't.
285 Previously some_benefit used to return True only if the variable was
286 applied to some value arguments. This didn't work:
288 let x = _coerce_ (T Int) Int (I# 3) in
289 case _coerce_ Int (T Int) x of
292 we want to inline x, but can't see that it's a constructor in a case
293 scrutinee position, and some_benefit is False.
297 dMonadST = _/\_ t -> :Monad (g1 _@_ t, g2 _@_ t, g3 _@_ t)
299 .... case dMonadST _@_ x0 of (a,b,c) -> ....
301 we'd really like to inline dMonadST here, but we *don't* want to
302 inline if the case expression is just
304 case x of y { DEFAULT -> ... }
306 since we can just eliminate this case instead (x is in WHNF). Similar
307 applies when x is bound to a lambda expression. Hence
308 contIsInteresting looks for case expressions with just a single
313 interestingCallContext :: SimplCont -> CallCtxt
314 interestingCallContext cont
317 interestingCtxt = ArgCtxt False 2 -- Give *some* incentive!
319 interesting (Select _ bndr _ _ _)
320 | isDeadBinder bndr = CaseCtxt
321 | otherwise = interestingCtxt
323 interesting (ApplyTo {}) = interestingCtxt
324 -- Can happen if we have (coerce t (f x)) y
325 -- Perhaps interestingCtxt is a bit over-keen, but I've
326 -- seen (coerce f) x, where f has an INLINE prag,
327 -- So we have to give some motivation for inlining it
329 interesting (StrictArg _ _ cci _ _) = cci
330 interesting (StrictBind {}) = BoringCtxt
331 interesting (Stop ty cci) = cci
332 interesting (CoerceIt _ cont) = interesting cont
333 -- If this call is the arg of a strict function, the context
334 -- is a bit interesting. If we inline here, we may get useful
335 -- evaluation information to avoid repeated evals: e.g.
337 -- Here the contIsInteresting makes the '*' keener to inline,
338 -- which in turn exposes a constructor which makes the '+' inline.
339 -- Assuming that +,* aren't small enough to inline regardless.
341 -- It's also very important to inline in a strict context for things
344 -- Here, the context of (f x) is strict, and if f's unfolding is
345 -- a build it's *great* to inline it here. So we must ensure that
346 -- the context for (f x) is not totally uninteresting.
351 -> Int -- Number of value args
352 -> SimplCont -- Context of the cal
355 mkArgInfo fun n_val_args call_cont
356 | n_val_args < idArity fun -- Note [Unsaturated functions]
357 = ArgInfo { ai_rules = False
358 , ai_strs = vanilla_stricts
359 , ai_discs = vanilla_discounts }
361 = ArgInfo { ai_rules = interestingArgContext fun call_cont
362 , ai_strs = arg_stricts
363 , ai_discs = arg_discounts }
365 vanilla_discounts, arg_discounts :: [Int]
366 vanilla_discounts = repeat 0
367 arg_discounts = case idUnfolding fun of
368 CoreUnfolding _ _ _ _ (UnfoldIfGoodArgs _ discounts _ _)
369 -> discounts ++ vanilla_discounts
370 other -> vanilla_discounts
372 vanilla_stricts, arg_stricts :: [Bool]
373 vanilla_stricts = repeat False
376 = case splitStrictSig (idNewStrictness fun) of
377 (demands, result_info)
378 | not (demands `lengthExceeds` n_val_args)
379 -> -- Enough args, use the strictness given.
380 -- For bottoming functions we used to pretend that the arg
381 -- is lazy, so that we don't treat the arg as an
382 -- interesting context. This avoids substituting
383 -- top-level bindings for (say) strings into
384 -- calls to error. But now we are more careful about
385 -- inlining lone variables, so its ok (see SimplUtils.analyseCont)
386 if isBotRes result_info then
387 map isStrictDmd demands -- Finite => result is bottom
389 map isStrictDmd demands ++ vanilla_stricts
392 -> WARN( True, text "More demands than arity" <+> ppr fun <+> ppr (idArity fun)
393 <+> ppr n_val_args <+> ppr demands )
394 vanilla_stricts -- Not enough args, or no strictness
396 {- Note [Unsaturated functions]
397 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
398 Consider (test eyeball/inline4)
401 where f has arity 2. Then we do not want to inline 'x', because
402 it'll just be floated out again. Even if f has lots of discounts
403 on its first argument -- it must be saturated for these to kick in
406 interestingArgContext :: Id -> SimplCont -> Bool
407 -- If the argument has form (f x y), where x,y are boring,
408 -- and f is marked INLINE, then we don't want to inline f.
409 -- But if the context of the argument is
411 -- where g has rules, then we *do* want to inline f, in case it
412 -- exposes a rule that might fire. Similarly, if the context is
414 -- where h has rules, then we do want to inline f; hence the
415 -- call_cont argument to interestingArgContext
417 -- The interesting_arg_ctxt flag makes this happen; if it's
418 -- set, the inliner gets just enough keener to inline f
419 -- regardless of how boring f's arguments are, if it's marked INLINE
421 -- The alternative would be to *always* inline an INLINE function,
422 -- regardless of how boring its context is; but that seems overkill
423 -- For example, it'd mean that wrapper functions were always inlined
424 interestingArgContext fn call_cont
425 = idHasRules fn || go call_cont
427 go (Select {}) = False
428 go (ApplyTo {}) = False
429 go (StrictArg _ _ cci _ _) = interesting cci
430 go (StrictBind {}) = False -- ??
431 go (CoerceIt _ c) = go c
432 go (Stop _ cci) = interesting cci
434 interesting (ArgCtxt rules _) = rules
435 interesting other = False
440 %************************************************************************
442 \subsection{Decisions about inlining}
444 %************************************************************************
446 Inlining is controlled partly by the SimplifierMode switch. This has two
449 SimplGently (a) Simplifying before specialiser/full laziness
450 (b) Simplifiying inside INLINE pragma
451 (c) Simplifying the LHS of a rule
452 (d) Simplifying a GHCi expression or Template
455 SimplPhase n _ Used at all other times
457 The key thing about SimplGently is that it does no call-site inlining.
458 Before full laziness we must be careful not to inline wrappers,
459 because doing so inhibits floating
460 e.g. ...(case f x of ...)...
461 ==> ...(case (case x of I# x# -> fw x#) of ...)...
462 ==> ...(case x of I# x# -> case fw x# of ...)...
463 and now the redex (f x) isn't floatable any more.
465 The no-inlining thing is also important for Template Haskell. You might be
466 compiling in one-shot mode with -O2; but when TH compiles a splice before
467 running it, we don't want to use -O2. Indeed, we don't want to inline
468 anything, because the byte-code interpreter might get confused about
469 unboxed tuples and suchlike.
473 SimplGently is also used as the mode to simplify inside an InlineMe note.
476 inlineMode :: SimplifierMode
477 inlineMode = SimplGently
480 It really is important to switch off inlinings inside such
481 expressions. Consider the following example
487 in ...g...g...g...g...g...
489 Now, if that's the ONLY occurrence of f, it will be inlined inside g,
490 and thence copied multiple times when g is inlined.
493 This function may be inlinined in other modules, so we
494 don't want to remove (by inlining) calls to functions that have
495 specialisations, or that may have transformation rules in an importing
498 E.g. {-# INLINE f #-}
501 and suppose that g is strict *and* has specialisations. If we inline
502 g's wrapper, we deny f the chance of getting the specialised version
503 of g when f is inlined at some call site (perhaps in some other
506 It's also important not to inline a worker back into a wrapper.
508 wraper = inline_me (\x -> ...worker... )
509 Normally, the inline_me prevents the worker getting inlined into
510 the wrapper (initially, the worker's only call site!). But,
511 if the wrapper is sure to be called, the strictness analyser will
512 mark it 'demanded', so when the RHS is simplified, it'll get an ArgOf
513 continuation. That's why the keep_inline predicate returns True for
514 ArgOf continuations. It shouldn't do any harm not to dissolve the
515 inline-me note under these circumstances.
517 Note that the result is that we do very little simplification
520 all xs = foldr (&&) True xs
521 any p = all . map p {-# INLINE any #-}
523 Problem: any won't get deforested, and so if it's exported and the
524 importer doesn't use the inlining, (eg passes it as an arg) then we
525 won't get deforestation at all. We havn't solved this problem yet!
528 preInlineUnconditionally
529 ~~~~~~~~~~~~~~~~~~~~~~~~
530 @preInlineUnconditionally@ examines a bndr to see if it is used just
531 once in a completely safe way, so that it is safe to discard the
532 binding inline its RHS at the (unique) usage site, REGARDLESS of how
533 big the RHS might be. If this is the case we don't simplify the RHS
534 first, but just inline it un-simplified.
536 This is much better than first simplifying a perhaps-huge RHS and then
537 inlining and re-simplifying it. Indeed, it can be at least quadratically
546 We may end up simplifying e1 N times, e2 N-1 times, e3 N-3 times etc.
547 This can happen with cascades of functions too:
554 THE MAIN INVARIANT is this:
556 ---- preInlineUnconditionally invariant -----
557 IF preInlineUnconditionally chooses to inline x = <rhs>
558 THEN doing the inlining should not change the occurrence
559 info for the free vars of <rhs>
560 ----------------------------------------------
562 For example, it's tempting to look at trivial binding like
564 and inline it unconditionally. But suppose x is used many times,
565 but this is the unique occurrence of y. Then inlining x would change
566 y's occurrence info, which breaks the invariant. It matters: y
567 might have a BIG rhs, which will now be dup'd at every occurrenc of x.
570 Even RHSs labelled InlineMe aren't caught here, because there might be
571 no benefit from inlining at the call site.
573 [Sept 01] Don't unconditionally inline a top-level thing, because that
574 can simply make a static thing into something built dynamically. E.g.
578 [Remember that we treat \s as a one-shot lambda.] No point in
579 inlining x unless there is something interesting about the call site.
581 But watch out: if you aren't careful, some useful foldr/build fusion
582 can be lost (most notably in spectral/hartel/parstof) because the
583 foldr didn't see the build. Doing the dynamic allocation isn't a big
584 deal, in fact, but losing the fusion can be. But the right thing here
585 seems to be to do a callSiteInline based on the fact that there is
586 something interesting about the call site (it's strict). Hmm. That
589 Conclusion: inline top level things gaily until Phase 0 (the last
590 phase), at which point don't.
593 preInlineUnconditionally :: SimplEnv -> TopLevelFlag -> InId -> InExpr -> Bool
594 preInlineUnconditionally env top_lvl bndr rhs
596 | opt_SimplNoPreInlining = False
597 | otherwise = case idOccInfo bndr of
598 IAmDead -> True -- Happens in ((\x.1) v)
599 OneOcc in_lam True int_cxt -> try_once in_lam int_cxt
603 active = case phase of
604 SimplGently -> isAlwaysActive prag
605 SimplPhase n _ -> isActive n prag
606 prag = idInlinePragma bndr
608 try_once in_lam int_cxt -- There's one textual occurrence
609 | not in_lam = isNotTopLevel top_lvl || early_phase
610 | otherwise = int_cxt && canInlineInLam rhs
612 -- Be very careful before inlining inside a lambda, becuase (a) we must not
613 -- invalidate occurrence information, and (b) we want to avoid pushing a
614 -- single allocation (here) into multiple allocations (inside lambda).
615 -- Inlining a *function* with a single *saturated* call would be ok, mind you.
616 -- || (if is_cheap && not (canInlineInLam rhs) then pprTrace "preinline" (ppr bndr <+> ppr rhs) ok else ok)
618 -- is_cheap = exprIsCheap rhs
619 -- ok = is_cheap && int_cxt
621 -- int_cxt The context isn't totally boring
622 -- E.g. let f = \ab.BIG in \y. map f xs
623 -- Don't want to substitute for f, because then we allocate
624 -- its closure every time the \y is called
625 -- But: let f = \ab.BIG in \y. map (f y) xs
626 -- Now we do want to substitute for f, even though it's not
627 -- saturated, because we're going to allocate a closure for
628 -- (f y) every time round the loop anyhow.
630 -- canInlineInLam => free vars of rhs are (Once in_lam) or Many,
631 -- so substituting rhs inside a lambda doesn't change the occ info.
632 -- Sadly, not quite the same as exprIsHNF.
633 canInlineInLam (Lit l) = True
634 canInlineInLam (Lam b e) = isRuntimeVar b || canInlineInLam e
635 canInlineInLam (Note _ e) = canInlineInLam e
636 canInlineInLam _ = False
638 early_phase = case phase of
639 SimplPhase 0 _ -> False
641 -- If we don't have this early_phase test, consider
642 -- x = length [1,2,3]
643 -- The full laziness pass carefully floats all the cons cells to
644 -- top level, and preInlineUnconditionally floats them all back in.
645 -- Result is (a) static allocation replaced by dynamic allocation
646 -- (b) many simplifier iterations because this tickles
647 -- a related problem; only one inlining per pass
649 -- On the other hand, I have seen cases where top-level fusion is
650 -- lost if we don't inline top level thing (e.g. string constants)
651 -- Hence the test for phase zero (which is the phase for all the final
652 -- simplifications). Until phase zero we take no special notice of
653 -- top level things, but then we become more leery about inlining
658 postInlineUnconditionally
659 ~~~~~~~~~~~~~~~~~~~~~~~~~
660 @postInlineUnconditionally@ decides whether to unconditionally inline
661 a thing based on the form of its RHS; in particular if it has a
662 trivial RHS. If so, we can inline and discard the binding altogether.
664 NB: a loop breaker has must_keep_binding = True and non-loop-breakers
665 only have *forward* references Hence, it's safe to discard the binding
667 NOTE: This isn't our last opportunity to inline. We're at the binding
668 site right now, and we'll get another opportunity when we get to the
671 Note that we do this unconditional inlining only for trival RHSs.
672 Don't inline even WHNFs inside lambdas; doing so may simply increase
673 allocation when the function is called. This isn't the last chance; see
676 NB: Even inline pragmas (e.g. IMustBeINLINEd) are ignored here Why?
677 Because we don't even want to inline them into the RHS of constructor
678 arguments. See NOTE above
680 NB: At one time even NOINLINE was ignored here: if the rhs is trivial
681 it's best to inline it anyway. We often get a=E; b=a from desugaring,
682 with both a and b marked NOINLINE. But that seems incompatible with
683 our new view that inlining is like a RULE, so I'm sticking to the 'active'
687 postInlineUnconditionally
688 :: SimplEnv -> TopLevelFlag
689 -> InId -- The binder (an OutId would be fine too)
690 -> OccInfo -- From the InId
694 postInlineUnconditionally env top_lvl bndr occ_info rhs unfolding
696 | isLoopBreaker occ_info = False -- If it's a loop-breaker of any kind, dont' inline
697 -- because it might be referred to "earlier"
698 | isExportedId bndr = False
699 | exprIsTrivial rhs = True
702 -- The point of examining occ_info here is that for *non-values*
703 -- that occur outside a lambda, the call-site inliner won't have
704 -- a chance (becuase it doesn't know that the thing
705 -- only occurs once). The pre-inliner won't have gotten
706 -- it either, if the thing occurs in more than one branch
707 -- So the main target is things like
710 -- True -> case x of ...
711 -- False -> case x of ...
712 -- I'm not sure how important this is in practice
713 OneOcc in_lam one_br int_cxt -- OneOcc => no code-duplication issue
714 -> smallEnoughToInline unfolding -- Small enough to dup
715 -- ToDo: consider discount on smallEnoughToInline if int_cxt is true
717 -- NB: Do NOT inline arbitrarily big things, even if one_br is True
718 -- Reason: doing so risks exponential behaviour. We simplify a big
719 -- expression, inline it, and simplify it again. But if the
720 -- very same thing happens in the big expression, we get
722 -- PRINCIPLE: when we've already simplified an expression once,
723 -- make sure that we only inline it if it's reasonably small.
725 && ((isNotTopLevel top_lvl && not in_lam) ||
726 -- But outside a lambda, we want to be reasonably aggressive
727 -- about inlining into multiple branches of case
728 -- e.g. let x = <non-value>
729 -- in case y of { C1 -> ..x..; C2 -> ..x..; C3 -> ... }
730 -- Inlining can be a big win if C3 is the hot-spot, even if
731 -- the uses in C1, C2 are not 'interesting'
732 -- An example that gets worse if you add int_cxt here is 'clausify'
734 (isCheapUnfolding unfolding && int_cxt))
735 -- isCheap => acceptable work duplication; in_lam may be true
736 -- int_cxt to prevent us inlining inside a lambda without some
737 -- good reason. See the notes on int_cxt in preInlineUnconditionally
739 IAmDead -> True -- This happens; for example, the case_bndr during case of
740 -- known constructor: case (a,b) of x { (p,q) -> ... }
741 -- Here x isn't mentioned in the RHS, so we don't want to
742 -- create the (dead) let-binding let x = (a,b) in ...
746 -- Here's an example that we don't handle well:
747 -- let f = if b then Left (\x.BIG) else Right (\y.BIG)
748 -- in \y. ....case f of {...} ....
749 -- Here f is used just once, and duplicating the case work is fine (exprIsCheap).
751 -- * We can't preInlineUnconditionally because that woud invalidate
752 -- the occ info for b.
753 -- * We can't postInlineUnconditionally because the RHS is big, and
754 -- that risks exponential behaviour
755 -- * We can't call-site inline, because the rhs is big
759 active = case getMode env of
760 SimplGently -> isAlwaysActive prag
761 SimplPhase n _ -> isActive n prag
762 prag = idInlinePragma bndr
764 activeInline :: SimplEnv -> OutId -> Bool
766 = case getMode env of
768 -- No inlining at all when doing gentle stuff,
769 -- except for local things that occur once (pre/postInlineUnconditionally)
770 -- The reason is that too little clean-up happens if you
771 -- don't inline use-once things. Also a bit of inlining is *good* for
772 -- full laziness; it can expose constant sub-expressions.
773 -- Example in spectral/mandel/Mandel.hs, where the mandelset
774 -- function gets a useful let-float if you inline windowToViewport
776 -- NB: we used to have a second exception, for data con wrappers.
777 -- On the grounds that we use gentle mode for rule LHSs, and
778 -- they match better when data con wrappers are inlined.
779 -- But that only really applies to the trivial wrappers (like (:)),
780 -- and they are now constructed as Compulsory unfoldings (in MkId)
781 -- so they'll happen anyway.
783 SimplPhase n _ -> isActive n prag
785 prag = idInlinePragma id
787 activeRule :: DynFlags -> SimplEnv -> Maybe (Activation -> Bool)
788 -- Nothing => No rules at all
789 activeRule dflags env
790 | not (dopt Opt_RewriteRules dflags)
791 = Nothing -- Rewriting is off
793 = case getMode env of
794 SimplGently -> Just isAlwaysActive
795 -- Used to be Nothing (no rules in gentle mode)
796 -- Main motivation for changing is that I wanted
797 -- lift String ===> ...
798 -- to work in Template Haskell when simplifying
799 -- splices, so we get simpler code for literal strings
800 SimplPhase n _ -> Just (isActive n)
804 %************************************************************************
808 %************************************************************************
811 mkLam :: [OutBndr] -> OutExpr -> SimplM OutExpr
812 -- mkLam tries three things
813 -- a) eta reduction, if that gives a trivial expression
814 -- b) eta expansion [only if there are some value lambdas]
819 = do { dflags <- getDOptsSmpl
820 ; mkLam' dflags bndrs body }
822 mkLam' :: DynFlags -> [OutBndr] -> OutExpr -> SimplM OutExpr
823 mkLam' dflags bndrs (Cast body co)
824 | not (any bad bndrs)
825 -- Note [Casts and lambdas]
826 = do { lam <- mkLam' dflags bndrs body
827 ; return (mkCoerce (mkPiTypes bndrs co) lam) }
829 co_vars = tyVarsOfType co
830 bad bndr = isCoVar bndr && bndr `elemVarSet` co_vars
832 mkLam' dflags bndrs body
833 | dopt Opt_DoEtaReduction dflags,
834 Just etad_lam <- tryEtaReduce bndrs body
835 = do { tick (EtaReduction (head bndrs))
838 | dopt Opt_DoLambdaEtaExpansion dflags,
839 any isRuntimeVar bndrs
840 = do { body' <- tryEtaExpansion dflags body
841 ; return (mkLams bndrs body') }
844 = return (mkLams bndrs body)
847 Note [Casts and lambdas]
848 ~~~~~~~~~~~~~~~~~~~~~~~~
850 (\x. (\y. e) `cast` g1) `cast` g2
851 There is a danger here that the two lambdas look separated, and the
852 full laziness pass might float an expression to between the two.
854 So this equation in mkLam' floats the g1 out, thus:
855 (\x. e `cast` g1) --> (\x.e) `cast` (tx -> g1)
858 In general, this floats casts outside lambdas, where (I hope) they
859 might meet and cancel with some other cast:
860 \x. e `cast` co ===> (\x. e) `cast` (tx -> co)
861 /\a. e `cast` co ===> (/\a. e) `cast` (/\a. co)
862 /\g. e `cast` co ===> (/\g. e) `cast` (/\g. co)
865 Notice that it works regardless of 'e'. Originally it worked only
866 if 'e' was itself a lambda, but in some cases that resulted in
867 fruitless iteration in the simplifier. A good example was when
868 compiling Text.ParserCombinators.ReadPrec, where we had a definition
869 like (\x. Get `cast` g)
870 where Get is a constructor with nonzero arity. Then mkLam eta-expanded
871 the Get, and the next iteration eta-reduced it, and then eta-expanded
874 Note also the side condition for the case of coercion binders.
875 It does not make sense to transform
876 /\g. e `cast` g ==> (/\g.e) `cast` (/\g.g)
877 because the latter is not well-kinded.
879 -- c) floating lets out through big lambdas
880 -- [only if all tyvar lambdas, and only if this lambda
881 -- is the RHS of a let]
883 {- Sept 01: I'm experimenting with getting the
884 full laziness pass to float out past big lambdsa
885 | all isTyVar bndrs, -- Only for big lambdas
886 contIsRhs cont -- Only try the rhs type-lambda floating
887 -- if this is indeed a right-hand side; otherwise
888 -- we end up floating the thing out, only for float-in
889 -- to float it right back in again!
890 = do (floats, body') <- tryRhsTyLam env bndrs body
891 return (floats, mkLams bndrs body')
895 %************************************************************************
899 %************************************************************************
901 Note [Eta reduction conditions]
902 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
903 We try for eta reduction here, but *only* if we get all the way to an
904 trivial expression. We don't want to remove extra lambdas unless we
905 are going to avoid allocating this thing altogether.
907 There are some particularly delicate points here:
909 * Eta reduction is not valid in general:
911 This matters, partly for old-fashioned correctness reasons but,
912 worse, getting it wrong can yield a seg fault. Consider
914 h y = case (case y of { True -> f `seq` True; False -> False }) of
915 True -> ...; False -> ...
917 If we (unsoundly) eta-reduce f to get f=f, the strictness analyser
918 says f=bottom, and replaces the (f `seq` True) with just
919 (f `cast` unsafe-co). BUT, as thing stand, 'f' got arity 1, and it
920 *keeps* arity 1 (perhaps also wrongly). So CorePrep eta-expands
921 the definition again, so that it does not termninate after all.
922 Result: seg-fault because the boolean case actually gets a function value.
925 So it's important to to the right thing.
927 * We need to be careful if we just look at f's arity. Currently (Dec07),
928 f's arity is visible in its own RHS (see Note [Arity robustness] in
929 SimplEnv) so we must *not* trust the arity when checking that 'f' is
930 a value. Instead, look at the unfolding.
932 However for GlobalIds we can look at the arity; and for primops we
933 must, since they have no unfolding.
935 * Regardless of whether 'f' is a vlaue, we always want to
936 reduce (/\a -> f a) to f
937 This came up in a RULE: foldr (build (/\a -> g a))
938 did not match foldr (build (/\b -> ...something complex...))
939 The type checker can insert these eta-expanded versions,
940 with both type and dictionary lambdas; hence the slightly
943 These delicacies are why we don't use exprIsTrivial and exprIsHNF here.
947 tryEtaReduce :: [OutBndr] -> OutExpr -> Maybe OutExpr
948 tryEtaReduce bndrs body
949 = go (reverse bndrs) body
951 go (b : bs) (App fun arg) | ok_arg b arg = go bs fun -- Loop round
952 go [] fun | ok_fun fun = Just fun -- Success!
953 go _ _ = Nothing -- Failure!
955 -- Note [Eta reduction conditions]
956 ok_fun (App fun (Type ty))
957 | not (any (`elemVarSet` tyVarsOfType ty) bndrs)
960 = not (fun_id `elem` bndrs)
961 && (ok_fun_id fun_id || all ok_lam bndrs)
965 | isLocalId fun = isEvaldUnfolding (idUnfolding fun)
966 | isDataConWorkId fun = True
967 | isGlobalId fun = idArity fun > 0
969 ok_lam v = isTyVar v || isDictId v
971 ok_arg b arg = varToCoreExpr b `cheapEqExpr` arg
975 %************************************************************************
979 %************************************************************************
983 f = \x1..xn -> N ==> f = \x1..xn y1..ym -> N y1..ym
986 where (in both cases)
988 * The xi can include type variables
990 * The yi are all value variables
992 * N is a NORMAL FORM (i.e. no redexes anywhere)
993 wanting a suitable number of extra args.
995 The biggest reason for doing this is for cases like
1001 Here we want to get the lambdas together. A good exmaple is the nofib
1002 program fibheaps, which gets 25% more allocation if you don't do this
1005 We may have to sandwich some coerces between the lambdas
1006 to make the types work. exprEtaExpandArity looks through coerces
1007 when computing arity; and etaExpand adds the coerces as necessary when
1008 actually computing the expansion.
1011 tryEtaExpansion :: DynFlags -> OutExpr -> SimplM OutExpr
1012 -- There is at least one runtime binder in the binders
1013 tryEtaExpansion dflags body = do
1015 return (etaExpand fun_arity us body (exprType body))
1017 fun_arity = exprEtaExpandArity dflags body
1021 %************************************************************************
1023 \subsection{Floating lets out of big lambdas}
1025 %************************************************************************
1027 Note [Floating and type abstraction]
1028 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
1031 We'd like to float this to
1034 x = /\a. C (y1 a) (y2 a)
1035 for the usual reasons: we want to inline x rather vigorously.
1037 You may think that this kind of thing is rare. But in some programs it is
1038 common. For example, if you do closure conversion you might get:
1040 data a :-> b = forall e. (e -> a -> b) :$ e
1042 f_cc :: forall a. a :-> a
1043 f_cc = /\a. (\e. id a) :$ ()
1045 Now we really want to inline that f_cc thing so that the
1046 construction of the closure goes away.
1048 So I have elaborated simplLazyBind to understand right-hand sides that look
1052 and treat them specially. The real work is done in SimplUtils.abstractFloats,
1053 but there is quite a bit of plumbing in simplLazyBind as well.
1055 The same transformation is good when there are lets in the body:
1057 /\abc -> let(rec) x = e in b
1059 let(rec) x' = /\abc -> let x = x' a b c in e
1061 /\abc -> let x = x' a b c in b
1063 This is good because it can turn things like:
1065 let f = /\a -> letrec g = ... g ... in g
1067 letrec g' = /\a -> ... g' a ...
1069 let f = /\ a -> g' a
1071 which is better. In effect, it means that big lambdas don't impede
1074 This optimisation is CRUCIAL in eliminating the junk introduced by
1075 desugaring mutually recursive definitions. Don't eliminate it lightly!
1077 [May 1999] If we do this transformation *regardless* then we can
1078 end up with some pretty silly stuff. For example,
1081 st = /\ s -> let { x1=r1 ; x2=r2 } in ...
1086 st = /\s -> ...[y1 s/x1, y2 s/x2]
1089 Unless the "..." is a WHNF there is really no point in doing this.
1090 Indeed it can make things worse. Suppose x1 is used strictly,
1093 x1* = case f y of { (a,b) -> e }
1095 If we abstract this wrt the tyvar we then can't do the case inline
1096 as we would normally do.
1098 That's why the whole transformation is part of the same process that
1099 floats let-bindings and constructor arguments out of RHSs. In particular,
1100 it is guarded by the doFloatFromRhs call in simplLazyBind.
1104 abstractFloats :: [OutTyVar] -> SimplEnv -> OutExpr -> SimplM ([OutBind], OutExpr)
1105 abstractFloats main_tvs body_env body
1106 = ASSERT( notNull body_floats )
1107 do { (subst, float_binds) <- mapAccumLM abstract empty_subst body_floats
1108 ; return (float_binds, CoreSubst.substExpr subst body) }
1110 main_tv_set = mkVarSet main_tvs
1111 body_floats = getFloats body_env
1112 empty_subst = CoreSubst.mkEmptySubst (seInScope body_env)
1114 abstract :: CoreSubst.Subst -> OutBind -> SimplM (CoreSubst.Subst, OutBind)
1115 abstract subst (NonRec id rhs)
1116 = do { (poly_id, poly_app) <- mk_poly tvs_here id
1117 ; let poly_rhs = mkLams tvs_here rhs'
1118 subst' = CoreSubst.extendIdSubst subst id poly_app
1119 ; return (subst', (NonRec poly_id poly_rhs)) }
1121 rhs' = CoreSubst.substExpr subst rhs
1122 tvs_here | any isCoVar main_tvs = main_tvs -- Note [Abstract over coercions]
1124 = varSetElems (main_tv_set `intersectVarSet` exprSomeFreeVars isTyVar rhs')
1126 -- Abstract only over the type variables free in the rhs
1127 -- wrt which the new binding is abstracted. But the naive
1128 -- approach of abstract wrt the tyvars free in the Id's type
1130 -- /\ a b -> let t :: (a,b) = (e1, e2)
1133 -- Here, b isn't free in x's type, but we must nevertheless
1134 -- abstract wrt b as well, because t's type mentions b.
1135 -- Since t is floated too, we'd end up with the bogus:
1136 -- poly_t = /\ a b -> (e1, e2)
1137 -- poly_x = /\ a -> fst (poly_t a *b*)
1138 -- So for now we adopt the even more naive approach of
1139 -- abstracting wrt *all* the tyvars. We'll see if that
1140 -- gives rise to problems. SLPJ June 98
1142 abstract subst (Rec prs)
1143 = do { (poly_ids, poly_apps) <- mapAndUnzipM (mk_poly tvs_here) ids
1144 ; let subst' = CoreSubst.extendSubstList subst (ids `zip` poly_apps)
1145 poly_rhss = [mkLams tvs_here (CoreSubst.substExpr subst' rhs) | rhs <- rhss]
1146 ; return (subst', Rec (poly_ids `zip` poly_rhss)) }
1148 (ids,rhss) = unzip prs
1149 -- For a recursive group, it's a bit of a pain to work out the minimal
1150 -- set of tyvars over which to abstract:
1151 -- /\ a b c. let x = ...a... in
1152 -- letrec { p = ...x...q...
1153 -- q = .....p...b... } in
1155 -- Since 'x' is abstracted over 'a', the {p,q} group must be abstracted
1156 -- over 'a' (because x is replaced by (poly_x a)) as well as 'b'.
1157 -- Since it's a pain, we just use the whole set, which is always safe
1159 -- If you ever want to be more selective, remember this bizarre case too:
1161 -- Here, we must abstract 'x' over 'a'.
1164 mk_poly tvs_here var
1165 = do { uniq <- getUniqueM
1166 ; let poly_name = setNameUnique (idName var) uniq -- Keep same name
1167 poly_ty = mkForAllTys tvs_here (idType var) -- But new type of course
1168 poly_id = mkLocalId poly_name poly_ty
1169 ; return (poly_id, mkTyApps (Var poly_id) (mkTyVarTys tvs_here)) }
1170 -- In the olden days, it was crucial to copy the occInfo of the original var,
1171 -- because we were looking at occurrence-analysed but as yet unsimplified code!
1172 -- In particular, we mustn't lose the loop breakers. BUT NOW we are looking
1173 -- at already simplified code, so it doesn't matter
1175 -- It's even right to retain single-occurrence or dead-var info:
1176 -- Suppose we started with /\a -> let x = E in B
1177 -- where x occurs once in B. Then we transform to:
1178 -- let x' = /\a -> E in /\a -> let x* = x' a in B
1179 -- where x* has an INLINE prag on it. Now, once x* is inlined,
1180 -- the occurrences of x' will be just the occurrences originally
1184 Note [Abstract over coercions]
1185 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
1186 If a coercion variable (g :: a ~ Int) is free in the RHS, then so is the
1187 type variable a. Rather than sort this mess out, we simply bale out and abstract
1188 wrt all the type variables if any of them are coercion variables.
1191 Historical note: if you use let-bindings instead of a substitution, beware of this:
1193 -- Suppose we start with:
1195 -- x = /\ a -> let g = G in E
1197 -- Then we'll float to get
1199 -- x = let poly_g = /\ a -> G
1200 -- in /\ a -> let g = poly_g a in E
1202 -- But now the occurrence analyser will see just one occurrence
1203 -- of poly_g, not inside a lambda, so the simplifier will
1204 -- PreInlineUnconditionally poly_g back into g! Badk to square 1!
1205 -- (I used to think that the "don't inline lone occurrences" stuff
1206 -- would stop this happening, but since it's the *only* occurrence,
1207 -- PreInlineUnconditionally kicks in first!)
1209 -- Solution: put an INLINE note on g's RHS, so that poly_g seems
1210 -- to appear many times. (NB: mkInlineMe eliminates
1211 -- such notes on trivial RHSs, so do it manually.)
1213 %************************************************************************
1217 %************************************************************************
1219 prepareAlts tries these things:
1221 1. If several alternatives are identical, merge them into
1222 a single DEFAULT alternative. I've occasionally seen this
1223 making a big difference:
1225 case e of =====> case e of
1226 C _ -> f x D v -> ....v....
1227 D v -> ....v.... DEFAULT -> f x
1230 The point is that we merge common RHSs, at least for the DEFAULT case.
1231 [One could do something more elaborate but I've never seen it needed.]
1232 To avoid an expensive test, we just merge branches equal to the *first*
1233 alternative; this picks up the common cases
1234 a) all branches equal
1235 b) some branches equal to the DEFAULT (which occurs first)
1238 case e of b { ==> case e of b {
1239 p1 -> rhs1 p1 -> rhs1
1241 pm -> rhsm pm -> rhsm
1242 _ -> case b of b' { pn -> let b'=b in rhsn
1244 ... po -> let b'=b in rhso
1245 po -> rhso _ -> let b'=b in rhsd
1249 which merges two cases in one case when -- the default alternative of
1250 the outer case scrutises the same variable as the outer case This
1251 transformation is called Case Merging. It avoids that the same
1252 variable is scrutinised multiple times.
1255 The case where transformation (1) showed up was like this (lib/std/PrelCError.lhs):
1261 where @is@ was something like
1263 p `is` n = p /= (-1) && p == n
1265 This gave rise to a horrible sequence of cases
1272 and similarly in cascade for all the join points!
1275 ~~~~~~~~~~~~~~~~~~~~
1276 We do this *here*, looking at un-simplified alternatives, because we
1277 have to check that r doesn't mention the variables bound by the
1278 pattern in each alternative, so the binder-info is rather useful.
1281 prepareAlts :: SimplEnv -> OutExpr -> OutId -> [InAlt] -> SimplM ([AltCon], [InAlt])
1282 prepareAlts env scrut case_bndr' alts
1283 = do { dflags <- getDOptsSmpl
1284 ; alts <- combineIdenticalAlts case_bndr' alts
1286 ; let (alts_wo_default, maybe_deflt) = findDefault alts
1287 alt_cons = [con | (con,_,_) <- alts_wo_default]
1288 imposs_deflt_cons = nub (imposs_cons ++ alt_cons)
1289 -- "imposs_deflt_cons" are handled
1290 -- EITHER by the context,
1291 -- OR by a non-DEFAULT branch in this case expression.
1293 ; default_alts <- prepareDefault dflags env case_bndr' mb_tc_app
1294 imposs_deflt_cons maybe_deflt
1296 ; let trimmed_alts = filterOut impossible_alt alts_wo_default
1297 merged_alts = mergeAlts trimmed_alts default_alts
1298 -- We need the mergeAlts in case the new default_alt
1299 -- has turned into a constructor alternative.
1300 -- The merge keeps the inner DEFAULT at the front, if there is one
1301 -- and interleaves the alternatives in the right order
1303 ; return (imposs_deflt_cons, merged_alts) }
1305 mb_tc_app = splitTyConApp_maybe (idType case_bndr')
1306 Just (_, inst_tys) = mb_tc_app
1308 imposs_cons = case scrut of
1309 Var v -> otherCons (idUnfolding v)
1312 impossible_alt :: CoreAlt -> Bool
1313 impossible_alt (con, _, _) | con `elem` imposs_cons = True
1314 impossible_alt (DataAlt con, _, _) = dataConCannotMatch inst_tys con
1315 impossible_alt alt = False
1318 --------------------------------------------------
1319 -- 1. Merge identical branches
1320 --------------------------------------------------
1321 combineIdenticalAlts :: OutId -> [InAlt] -> SimplM [InAlt]
1323 combineIdenticalAlts case_bndr alts@((con1,bndrs1,rhs1) : con_alts)
1324 | all isDeadBinder bndrs1, -- Remember the default
1325 length filtered_alts < length con_alts -- alternative comes first
1326 -- Also Note [Dead binders]
1327 = do { tick (AltMerge case_bndr)
1328 ; return ((DEFAULT, [], rhs1) : filtered_alts) }
1330 filtered_alts = filter keep con_alts
1331 keep (con,bndrs,rhs) = not (all isDeadBinder bndrs && rhs `cheapEqExpr` rhs1)
1333 combineIdenticalAlts case_bndr alts = return alts
1335 -------------------------------------------------------------------------
1336 -- Prepare the default alternative
1337 -------------------------------------------------------------------------
1338 prepareDefault :: DynFlags
1340 -> OutId -- Case binder; need just for its type. Note that as an
1341 -- OutId, it has maximum information; this is important.
1342 -- Test simpl013 is an example
1343 -> Maybe (TyCon, [Type]) -- Type of scrutinee, decomposed
1344 -> [AltCon] -- These cons can't happen when matching the default
1345 -> Maybe InExpr -- Rhs
1346 -> SimplM [InAlt] -- Still unsimplified
1347 -- We use a list because it's what mergeAlts expects,
1348 -- And becuase case-merging can cause many to show up
1350 ------- Merge nested cases ----------
1351 prepareDefault dflags env outer_bndr bndr_ty imposs_cons (Just deflt_rhs)
1352 | dopt Opt_CaseMerge dflags
1353 , Case (Var inner_scrut_var) inner_bndr _ inner_alts <- deflt_rhs
1354 , DoneId inner_scrut_var' <- substId env inner_scrut_var
1355 -- Remember, inner_scrut_var is an InId, but outer_bndr is an OutId
1356 , inner_scrut_var' == outer_bndr
1357 -- NB: the substId means that if the outer scrutinee was a
1358 -- variable, and inner scrutinee is the same variable,
1359 -- then inner_scrut_var' will be outer_bndr
1360 -- via the magic of simplCaseBinder
1361 = do { tick (CaseMerge outer_bndr)
1363 ; let munge_rhs rhs = bindCaseBndr inner_bndr (Var outer_bndr) rhs
1364 ; return [(con, args, munge_rhs rhs) | (con, args, rhs) <- inner_alts,
1365 not (con `elem` imposs_cons) ]
1366 -- NB: filter out any imposs_cons. Example:
1369 -- DEFAULT -> case x of
1372 -- When we merge, we must ensure that e1 takes
1373 -- precedence over e2 as the value for A!
1375 -- Warning: don't call prepareAlts recursively!
1376 -- Firstly, there's no point, because inner alts have already had
1377 -- mkCase applied to them, so they won't have a case in their default
1378 -- Secondly, if you do, you get an infinite loop, because the bindCaseBndr
1379 -- in munge_rhs may put a case into the DEFAULT branch!
1382 --------- Fill in known constructor -----------
1383 prepareDefault dflags env case_bndr (Just (tycon, inst_tys)) imposs_cons (Just deflt_rhs)
1384 | -- This branch handles the case where we are
1385 -- scrutinisng an algebraic data type
1386 isAlgTyCon tycon -- It's a data type, tuple, or unboxed tuples.
1387 , not (isNewTyCon tycon) -- We can have a newtype, if we are just doing an eval:
1388 -- case x of { DEFAULT -> e }
1389 -- and we don't want to fill in a default for them!
1390 , Just all_cons <- tyConDataCons_maybe tycon
1391 , not (null all_cons) -- This is a tricky corner case. If the data type has no constructors,
1392 -- which GHC allows, then the case expression will have at most a default
1393 -- alternative. We don't want to eliminate that alternative, because the
1394 -- invariant is that there's always one alternative. It's more convenient
1396 -- case x of { DEFAULT -> e }
1397 -- as it is, rather than transform it to
1398 -- error "case cant match"
1399 -- which would be quite legitmate. But it's a really obscure corner, and
1400 -- not worth wasting code on.
1401 , let imposs_data_cons = [con | DataAlt con <- imposs_cons] -- We now know it's a data type
1402 impossible con = con `elem` imposs_data_cons || dataConCannotMatch inst_tys con
1403 = case filterOut impossible all_cons of
1404 [] -> return [] -- Eliminate the default alternative
1405 -- altogether if it can't match
1407 [con] -> -- It matches exactly one constructor, so fill it in
1408 do { tick (FillInCaseDefault case_bndr)
1410 ; let (ex_tvs, co_tvs, arg_ids) =
1411 dataConRepInstPat us con inst_tys
1412 ; return [(DataAlt con, ex_tvs ++ co_tvs ++ arg_ids, deflt_rhs)] }
1414 two_or_more -> return [(DEFAULT, [], deflt_rhs)]
1416 --------- Catch-all cases -----------
1417 prepareDefault dflags env case_bndr bndr_ty imposs_cons (Just deflt_rhs)
1418 = return [(DEFAULT, [], deflt_rhs)]
1420 prepareDefault dflags env case_bndr bndr_ty imposs_cons Nothing
1421 = return [] -- No default branch
1426 =================================================================================
1428 mkCase tries these things
1430 1. Eliminate the case altogether if possible
1438 and similar friends.
1442 mkCase :: OutExpr -> OutId -> OutType
1443 -> [OutAlt] -- Increasing order
1446 --------------------------------------------------
1447 -- 1. Check for empty alternatives
1448 --------------------------------------------------
1450 -- This isn't strictly an error. It's possible that the simplifer might "see"
1451 -- that an inner case has no accessible alternatives before it "sees" that the
1452 -- entire branch of an outer case is inaccessible. So we simply
1453 -- put an error case here insteadd
1454 mkCase scrut case_bndr ty []
1455 = pprTrace "mkCase: null alts" (ppr case_bndr <+> ppr scrut) $
1456 return (mkApps (Var rUNTIME_ERROR_ID)
1457 [Type ty, Lit (mkStringLit "Impossible alternative")])
1460 --------------------------------------------------
1462 --------------------------------------------------
1464 mkCase scrut case_bndr ty alts -- Identity case
1465 | all identity_alt alts
1466 = do tick (CaseIdentity case_bndr)
1467 return (re_cast scrut)
1469 identity_alt (con, args, rhs) = check_eq con args (de_cast rhs)
1471 check_eq DEFAULT _ (Var v) = v == case_bndr
1472 check_eq (LitAlt lit') _ (Lit lit) = lit == lit'
1473 check_eq (DataAlt con) args rhs = rhs `cheapEqExpr` mkConApp con (arg_tys ++ varsToCoreExprs args)
1474 || rhs `cheapEqExpr` Var case_bndr
1475 check_eq con args rhs = False
1477 arg_tys = map Type (tyConAppArgs (idType case_bndr))
1480 -- case e of x { _ -> x `cast` c }
1481 -- And we definitely want to eliminate this case, to give
1483 -- So we throw away the cast from the RHS, and reconstruct
1484 -- it at the other end. All the RHS casts must be the same
1485 -- if (all identity_alt alts) holds.
1487 -- Don't worry about nested casts, because the simplifier combines them
1488 de_cast (Cast e _) = e
1491 re_cast scrut = case head alts of
1492 (_,_,Cast _ co) -> Cast scrut co
1497 --------------------------------------------------
1499 --------------------------------------------------
1500 mkCase scrut bndr ty alts = return (Case scrut bndr ty alts)
1504 When adding auxiliary bindings for the case binder, it's worth checking if
1505 its dead, because it often is, and occasionally these mkCase transformations
1506 cascade rather nicely.
1509 bindCaseBndr bndr rhs body
1510 | isDeadBinder bndr = body
1511 | otherwise = bindNonRec bndr rhs body