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(..), LetRhsFlag(..),
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 )
64 %************************************************************************
68 %************************************************************************
70 A SimplCont allows the simplifier to traverse the expression in a
71 zipper-like fashion. The SimplCont represents the rest of the expression,
72 "above" the point of interest.
74 You can also think of a SimplCont as an "evaluation context", using
75 that term in the way it is used for operational semantics. This is the
76 way I usually think of it, For example you'll often see a syntax for
77 evaluation context looking like
78 C ::= [] | C e | case C of alts | C `cast` co
79 That's the kind of thing we are doing here, and I use that syntax in
84 * A SimplCont describes a *strict* context (just like
85 evaluation contexts do). E.g. Just [] is not a SimplCont
87 * A SimplCont describes a context that *does not* bind
88 any variables. E.g. \x. [] is not a SimplCont
92 = Stop -- An empty context, or hole, []
93 OutType -- Type of the result
95 Bool -- True <=> There is something interesting about
96 -- the context, and hence the inliner
97 -- should be a bit keener (see interestingCallContext)
99 -- (a) This is the RHS of a thunk whose type suggests
100 -- that update-in-place would be possible
101 -- (b) 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 (Bool,[Bool]) -- Whether the function at the head of e has rules,
127 SimplCont -- plus strictness flags for further args
129 data LetRhsFlag = AnArg -- It's just an argument not a let RHS
130 | AnRhs -- It's the RHS of a let (so please float lets out of big lambdas)
132 instance Outputable LetRhsFlag where
133 ppr AnArg = ptext SLIT("arg")
134 ppr AnRhs = ptext SLIT("rhs")
136 instance Outputable SimplCont where
137 ppr (Stop ty is_rhs _) = ptext SLIT("Stop") <> brackets (ppr is_rhs) <+> ppr ty
138 ppr (ApplyTo dup arg se cont) = ((ptext SLIT("ApplyTo") <+> ppr dup <+> pprParendExpr arg)
139 {- $$ nest 2 (pprSimplEnv se) -}) $$ ppr cont
140 ppr (StrictBind b _ _ _ cont) = (ptext SLIT("StrictBind") <+> ppr b) $$ ppr cont
141 ppr (StrictArg f _ _ cont) = (ptext SLIT("StrictArg") <+> ppr f) $$ ppr cont
142 ppr (Select dup bndr alts se cont) = (ptext SLIT("Select") <+> ppr dup <+> ppr bndr) $$
143 (nest 4 (ppr alts)) $$ ppr cont
144 ppr (CoerceIt co cont) = (ptext SLIT("CoerceIt") <+> ppr co) $$ ppr cont
146 data DupFlag = OkToDup | NoDup
148 instance Outputable DupFlag where
149 ppr OkToDup = ptext SLIT("ok")
150 ppr NoDup = ptext SLIT("nodup")
155 mkBoringStop :: OutType -> SimplCont
156 mkBoringStop ty = Stop ty AnArg False
158 mkLazyArgStop :: OutType -> Bool -> SimplCont
159 mkLazyArgStop ty has_rules = Stop ty AnArg (canUpdateInPlace ty || has_rules)
161 mkRhsStop :: OutType -> SimplCont
162 mkRhsStop ty = Stop ty AnRhs (canUpdateInPlace ty)
165 contIsRhsOrArg (Stop {}) = True
166 contIsRhsOrArg (StrictBind {}) = True
167 contIsRhsOrArg (StrictArg {}) = True
168 contIsRhsOrArg other = False
171 contIsDupable :: SimplCont -> Bool
172 contIsDupable (Stop {}) = True
173 contIsDupable (ApplyTo OkToDup _ _ _) = True
174 contIsDupable (Select OkToDup _ _ _ _) = True
175 contIsDupable (CoerceIt _ cont) = contIsDupable cont
176 contIsDupable other = False
179 contIsTrivial :: SimplCont -> Bool
180 contIsTrivial (Stop {}) = True
181 contIsTrivial (ApplyTo _ (Type _) _ cont) = contIsTrivial cont
182 contIsTrivial (CoerceIt _ cont) = contIsTrivial cont
183 contIsTrivial other = False
186 contResultType :: SimplCont -> OutType
187 contResultType (Stop to_ty _ _) = to_ty
188 contResultType (StrictArg _ _ _ cont) = contResultType cont
189 contResultType (StrictBind _ _ _ _ cont) = contResultType cont
190 contResultType (ApplyTo _ _ _ cont) = contResultType cont
191 contResultType (CoerceIt _ cont) = contResultType cont
192 contResultType (Select _ _ _ _ cont) = contResultType cont
195 countValArgs :: SimplCont -> Int
196 countValArgs (ApplyTo _ (Type ty) se cont) = countValArgs cont
197 countValArgs (ApplyTo _ val_arg se cont) = 1 + countValArgs cont
198 countValArgs other = 0
200 countArgs :: SimplCont -> Int
201 countArgs (ApplyTo _ arg se cont) = 1 + countArgs cont
204 contArgs :: SimplCont -> ([OutExpr], SimplCont)
205 -- Uses substitution to turn each arg into an OutExpr
206 contArgs cont = go [] cont
208 go args (ApplyTo _ arg se cont) = go (substExpr se arg : args) cont
209 go args cont = (reverse args, cont)
211 dropArgs :: Int -> SimplCont -> SimplCont
212 dropArgs 0 cont = cont
213 dropArgs n (ApplyTo _ _ _ cont) = dropArgs (n-1) cont
214 dropArgs n other = pprPanic "dropArgs" (ppr n <+> ppr other)
217 splitInlineCont :: SimplCont -> Maybe (SimplCont, SimplCont)
218 -- Returns Nothing if the continuation should dissolve an InlineMe Note
219 -- Return Just (c1,c2) otherwise,
220 -- where c1 is the continuation to put inside the InlineMe
223 -- Example: (__inline_me__ (/\a. e)) ty
224 -- Here we want to do the beta-redex without dissolving the InlineMe
225 -- See test simpl017 (and Trac #1627) for a good example of why this is important
227 splitInlineCont (ApplyTo dup (Type ty) se c)
228 | Just (c1, c2) <- splitInlineCont c = Just (ApplyTo dup (Type ty) se c1, c2)
229 splitInlineCont cont@(Stop ty _ _) = Just (mkBoringStop ty, cont)
230 splitInlineCont cont@(StrictBind bndr _ _ se _) = Just (mkBoringStop (substTy se (idType bndr)), cont)
231 splitInlineCont cont@(StrictArg _ fun_ty _ _) = Just (mkBoringStop (funArgTy fun_ty), cont)
232 splitInlineCont other = Nothing
233 -- NB: the calculation of the type for mkBoringStop is an annoying
234 -- duplication of the same calucation in mkDupableCont
239 interestingArg :: OutExpr -> Bool
240 -- An argument is interesting if it has *some* structure
241 -- We are here trying to avoid unfolding a function that
242 -- is applied only to variables that have no unfolding
243 -- (i.e. they are probably lambda bound): f x y z
244 -- There is little point in inlining f here.
245 interestingArg (Var v) = hasSomeUnfolding (idUnfolding v)
246 -- Was: isValueUnfolding (idUnfolding v')
247 -- But that seems over-pessimistic
249 -- This accounts for an argument like
250 -- () or [], which is definitely interesting
251 interestingArg (Type _) = False
252 interestingArg (App fn (Type _)) = interestingArg fn
253 interestingArg (Note _ a) = interestingArg a
255 -- Idea (from Sam B); I'm not sure if it's a good idea, so commented out for now
256 -- interestingArg expr | isUnLiftedType (exprType expr)
257 -- -- Unlifted args are only ever interesting if we know what they are
262 interestingArg other = True
263 -- Consider let x = 3 in f x
264 -- The substitution will contain (x -> ContEx 3), and we want to
265 -- to say that x is an interesting argument.
266 -- But consider also (\x. f x y) y
267 -- The substitution will contain (x -> ContEx y), and we want to say
268 -- that x is not interesting (assuming y has no unfolding)
272 Comment about interestingCallContext
273 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
274 We want to avoid inlining an expression where there can't possibly be
275 any gain, such as in an argument position. Hence, if the continuation
276 is interesting (eg. a case scrutinee, application etc.) then we
277 inline, otherwise we don't.
279 Previously some_benefit used to return True only if the variable was
280 applied to some value arguments. This didn't work:
282 let x = _coerce_ (T Int) Int (I# 3) in
283 case _coerce_ Int (T Int) x of
286 we want to inline x, but can't see that it's a constructor in a case
287 scrutinee position, and some_benefit is False.
291 dMonadST = _/\_ t -> :Monad (g1 _@_ t, g2 _@_ t, g3 _@_ t)
293 .... case dMonadST _@_ x0 of (a,b,c) -> ....
295 we'd really like to inline dMonadST here, but we *don't* want to
296 inline if the case expression is just
298 case x of y { DEFAULT -> ... }
300 since we can just eliminate this case instead (x is in WHNF). Similar
301 applies when x is bound to a lambda expression. Hence
302 contIsInteresting looks for case expressions with just a single
307 interestingCallContext :: SimplCont -> CallContInfo
308 interestingCallContext cont
311 interesting (Select _ bndr _ _ _)
312 | isDeadBinder bndr = CaseCont
313 | otherwise = InterestingCont
315 interesting (ApplyTo {}) = InterestingCont
316 -- Can happen if we have (coerce t (f x)) y
317 -- Perhaps True is a bit over-keen, but I've
318 -- seen (coerce f) x, where f has an INLINE prag,
319 -- So we have to give some motivation for inlining it
320 interesting (StrictArg {}) = InterestingCont
321 interesting (StrictBind {}) = InterestingCont
322 interesting (Stop ty _ yes) = if yes then InterestingCont else BoringCont
323 interesting (CoerceIt _ cont) = interesting cont
324 -- If this call is the arg of a strict function, the context
325 -- is a bit interesting. If we inline here, we may get useful
326 -- evaluation information to avoid repeated evals: e.g.
328 -- Here the contIsInteresting makes the '*' keener to inline,
329 -- which in turn exposes a constructor which makes the '+' inline.
330 -- Assuming that +,* aren't small enough to inline regardless.
332 -- It's also very important to inline in a strict context for things
335 -- Here, the context of (f x) is strict, and if f's unfolding is
336 -- a build it's *great* to inline it here. So we must ensure that
337 -- the context for (f x) is not totally uninteresting.
342 -> Int -- Number of value args
343 -> SimplCont -- Context of the cal
344 -> (Bool, [Bool]) -- Arg info
345 -- The arg info consists of
346 -- * A Bool indicating if the function has rules (recursively)
347 -- * A [Bool] indicating strictness for each arg
348 -- The [Bool] is usually infinite, but if it is finite it
349 -- guarantees that the function diverges after being given
350 -- that number of args
352 mkArgInfo fun n_val_args call_cont
353 = (interestingArgContext fun call_cont, fun_stricts)
355 vanilla_stricts, fun_stricts :: [Bool]
356 vanilla_stricts = repeat False
359 = case splitStrictSig (idNewStrictness fun) of
360 (demands, result_info)
361 | not (demands `lengthExceeds` n_val_args)
362 -> -- Enough args, use the strictness given.
363 -- For bottoming functions we used to pretend that the arg
364 -- is lazy, so that we don't treat the arg as an
365 -- interesting context. This avoids substituting
366 -- top-level bindings for (say) strings into
367 -- calls to error. But now we are more careful about
368 -- inlining lone variables, so its ok (see SimplUtils.analyseCont)
369 if isBotRes result_info then
370 map isStrictDmd demands -- Finite => result is bottom
372 map isStrictDmd demands ++ vanilla_stricts
374 other -> vanilla_stricts -- Not enough args, or no strictness
376 interestingArgContext :: Id -> SimplCont -> Bool
377 -- If the argument has form (f x y), where x,y are boring,
378 -- and f is marked INLINE, then we don't want to inline f.
379 -- But if the context of the argument is
381 -- where g has rules, then we *do* want to inline f, in case it
382 -- exposes a rule that might fire. Similarly, if the context is
384 -- where h has rules, then we do want to inline f; hence the
385 -- call_cont argument to interestingArgContext
387 -- The interesting_arg_ctxt flag makes this happen; if it's
388 -- set, the inliner gets just enough keener to inline f
389 -- regardless of how boring f's arguments are, if it's marked INLINE
391 -- The alternative would be to *always* inline an INLINE function,
392 -- regardless of how boring its context is; but that seems overkill
393 -- For example, it'd mean that wrapper functions were always inlined
394 interestingArgContext fn call_cont
395 = idHasRules fn || go call_cont
397 go (Select {}) = False
398 go (ApplyTo {}) = False
399 go (StrictArg {}) = True
400 go (StrictBind {}) = False -- ??
401 go (CoerceIt _ c) = go c
402 go (Stop _ _ interesting) = interesting
405 canUpdateInPlace :: Type -> Bool
406 -- Consider let x = <wurble> in ...
407 -- If <wurble> returns an explicit constructor, we might be able
408 -- to do update in place. So we treat even a thunk RHS context
409 -- as interesting if update in place is possible. We approximate
410 -- this by seeing if the type has a single constructor with a
411 -- small arity. But arity zero isn't good -- we share the single copy
412 -- for that case, so no point in sharing.
415 | not opt_UF_UpdateInPlace = False
417 = case splitTyConApp_maybe ty of
419 Just (tycon, _) -> case tyConDataCons_maybe tycon of
420 Just [dc] -> arity == 1 || arity == 2
422 arity = dataConRepArity dc
428 %************************************************************************
430 \subsection{Decisions about inlining}
432 %************************************************************************
434 Inlining is controlled partly by the SimplifierMode switch. This has two
437 SimplGently (a) Simplifying before specialiser/full laziness
438 (b) Simplifiying inside INLINE pragma
439 (c) Simplifying the LHS of a rule
440 (d) Simplifying a GHCi expression or Template
443 SimplPhase n Used at all other times
445 The key thing about SimplGently is that it does no call-site inlining.
446 Before full laziness we must be careful not to inline wrappers,
447 because doing so inhibits floating
448 e.g. ...(case f x of ...)...
449 ==> ...(case (case x of I# x# -> fw x#) of ...)...
450 ==> ...(case x of I# x# -> case fw x# of ...)...
451 and now the redex (f x) isn't floatable any more.
453 The no-inlining thing is also important for Template Haskell. You might be
454 compiling in one-shot mode with -O2; but when TH compiles a splice before
455 running it, we don't want to use -O2. Indeed, we don't want to inline
456 anything, because the byte-code interpreter might get confused about
457 unboxed tuples and suchlike.
461 SimplGently is also used as the mode to simplify inside an InlineMe note.
464 inlineMode :: SimplifierMode
465 inlineMode = SimplGently
468 It really is important to switch off inlinings inside such
469 expressions. Consider the following example
475 in ...g...g...g...g...g...
477 Now, if that's the ONLY occurrence of f, it will be inlined inside g,
478 and thence copied multiple times when g is inlined.
481 This function may be inlinined in other modules, so we
482 don't want to remove (by inlining) calls to functions that have
483 specialisations, or that may have transformation rules in an importing
486 E.g. {-# INLINE f #-}
489 and suppose that g is strict *and* has specialisations. If we inline
490 g's wrapper, we deny f the chance of getting the specialised version
491 of g when f is inlined at some call site (perhaps in some other
494 It's also important not to inline a worker back into a wrapper.
496 wraper = inline_me (\x -> ...worker... )
497 Normally, the inline_me prevents the worker getting inlined into
498 the wrapper (initially, the worker's only call site!). But,
499 if the wrapper is sure to be called, the strictness analyser will
500 mark it 'demanded', so when the RHS is simplified, it'll get an ArgOf
501 continuation. That's why the keep_inline predicate returns True for
502 ArgOf continuations. It shouldn't do any harm not to dissolve the
503 inline-me note under these circumstances.
505 Note that the result is that we do very little simplification
508 all xs = foldr (&&) True xs
509 any p = all . map p {-# INLINE any #-}
511 Problem: any won't get deforested, and so if it's exported and the
512 importer doesn't use the inlining, (eg passes it as an arg) then we
513 won't get deforestation at all. We havn't solved this problem yet!
516 preInlineUnconditionally
517 ~~~~~~~~~~~~~~~~~~~~~~~~
518 @preInlineUnconditionally@ examines a bndr to see if it is used just
519 once in a completely safe way, so that it is safe to discard the
520 binding inline its RHS at the (unique) usage site, REGARDLESS of how
521 big the RHS might be. If this is the case we don't simplify the RHS
522 first, but just inline it un-simplified.
524 This is much better than first simplifying a perhaps-huge RHS and then
525 inlining and re-simplifying it. Indeed, it can be at least quadratically
534 We may end up simplifying e1 N times, e2 N-1 times, e3 N-3 times etc.
535 This can happen with cascades of functions too:
542 THE MAIN INVARIANT is this:
544 ---- preInlineUnconditionally invariant -----
545 IF preInlineUnconditionally chooses to inline x = <rhs>
546 THEN doing the inlining should not change the occurrence
547 info for the free vars of <rhs>
548 ----------------------------------------------
550 For example, it's tempting to look at trivial binding like
552 and inline it unconditionally. But suppose x is used many times,
553 but this is the unique occurrence of y. Then inlining x would change
554 y's occurrence info, which breaks the invariant. It matters: y
555 might have a BIG rhs, which will now be dup'd at every occurrenc of x.
558 Even RHSs labelled InlineMe aren't caught here, because there might be
559 no benefit from inlining at the call site.
561 [Sept 01] Don't unconditionally inline a top-level thing, because that
562 can simply make a static thing into something built dynamically. E.g.
566 [Remember that we treat \s as a one-shot lambda.] No point in
567 inlining x unless there is something interesting about the call site.
569 But watch out: if you aren't careful, some useful foldr/build fusion
570 can be lost (most notably in spectral/hartel/parstof) because the
571 foldr didn't see the build. Doing the dynamic allocation isn't a big
572 deal, in fact, but losing the fusion can be. But the right thing here
573 seems to be to do a callSiteInline based on the fact that there is
574 something interesting about the call site (it's strict). Hmm. That
577 Conclusion: inline top level things gaily until Phase 0 (the last
578 phase), at which point don't.
581 preInlineUnconditionally :: SimplEnv -> TopLevelFlag -> InId -> InExpr -> Bool
582 preInlineUnconditionally env top_lvl bndr rhs
584 | opt_SimplNoPreInlining = False
585 | otherwise = case idOccInfo bndr of
586 IAmDead -> True -- Happens in ((\x.1) v)
587 OneOcc in_lam True int_cxt -> try_once in_lam int_cxt
591 active = case phase of
592 SimplGently -> isAlwaysActive prag
593 SimplPhase n -> isActive n prag
594 prag = idInlinePragma bndr
596 try_once in_lam int_cxt -- There's one textual occurrence
597 | not in_lam = isNotTopLevel top_lvl || early_phase
598 | otherwise = int_cxt && canInlineInLam rhs
600 -- Be very careful before inlining inside a lambda, becuase (a) we must not
601 -- invalidate occurrence information, and (b) we want to avoid pushing a
602 -- single allocation (here) into multiple allocations (inside lambda).
603 -- Inlining a *function* with a single *saturated* call would be ok, mind you.
604 -- || (if is_cheap && not (canInlineInLam rhs) then pprTrace "preinline" (ppr bndr <+> ppr rhs) ok else ok)
606 -- is_cheap = exprIsCheap rhs
607 -- ok = is_cheap && int_cxt
609 -- int_cxt The context isn't totally boring
610 -- E.g. let f = \ab.BIG in \y. map f xs
611 -- Don't want to substitute for f, because then we allocate
612 -- its closure every time the \y is called
613 -- But: let f = \ab.BIG in \y. map (f y) xs
614 -- Now we do want to substitute for f, even though it's not
615 -- saturated, because we're going to allocate a closure for
616 -- (f y) every time round the loop anyhow.
618 -- canInlineInLam => free vars of rhs are (Once in_lam) or Many,
619 -- so substituting rhs inside a lambda doesn't change the occ info.
620 -- Sadly, not quite the same as exprIsHNF.
621 canInlineInLam (Lit l) = True
622 canInlineInLam (Lam b e) = isRuntimeVar b || canInlineInLam e
623 canInlineInLam (Note _ e) = canInlineInLam e
624 canInlineInLam _ = False
626 early_phase = case phase of
627 SimplPhase 0 -> False
629 -- If we don't have this early_phase test, consider
630 -- x = length [1,2,3]
631 -- The full laziness pass carefully floats all the cons cells to
632 -- top level, and preInlineUnconditionally floats them all back in.
633 -- Result is (a) static allocation replaced by dynamic allocation
634 -- (b) many simplifier iterations because this tickles
635 -- a related problem; only one inlining per pass
637 -- On the other hand, I have seen cases where top-level fusion is
638 -- lost if we don't inline top level thing (e.g. string constants)
639 -- Hence the test for phase zero (which is the phase for all the final
640 -- simplifications). Until phase zero we take no special notice of
641 -- top level things, but then we become more leery about inlining
646 postInlineUnconditionally
647 ~~~~~~~~~~~~~~~~~~~~~~~~~
648 @postInlineUnconditionally@ decides whether to unconditionally inline
649 a thing based on the form of its RHS; in particular if it has a
650 trivial RHS. If so, we can inline and discard the binding altogether.
652 NB: a loop breaker has must_keep_binding = True and non-loop-breakers
653 only have *forward* references Hence, it's safe to discard the binding
655 NOTE: This isn't our last opportunity to inline. We're at the binding
656 site right now, and we'll get another opportunity when we get to the
659 Note that we do this unconditional inlining only for trival RHSs.
660 Don't inline even WHNFs inside lambdas; doing so may simply increase
661 allocation when the function is called. This isn't the last chance; see
664 NB: Even inline pragmas (e.g. IMustBeINLINEd) are ignored here Why?
665 Because we don't even want to inline them into the RHS of constructor
666 arguments. See NOTE above
668 NB: At one time even NOINLINE was ignored here: if the rhs is trivial
669 it's best to inline it anyway. We often get a=E; b=a from desugaring,
670 with both a and b marked NOINLINE. But that seems incompatible with
671 our new view that inlining is like a RULE, so I'm sticking to the 'active'
675 postInlineUnconditionally
676 :: SimplEnv -> TopLevelFlag
677 -> InId -- The binder (an OutId would be fine too)
678 -> OccInfo -- From the InId
682 postInlineUnconditionally env top_lvl bndr occ_info rhs unfolding
684 | isLoopBreaker occ_info = False -- If it's a loop-breaker of any kind, dont' inline
685 -- because it might be referred to "earlier"
686 | isExportedId bndr = False
687 | exprIsTrivial rhs = True
690 -- The point of examining occ_info here is that for *non-values*
691 -- that occur outside a lambda, the call-site inliner won't have
692 -- a chance (becuase it doesn't know that the thing
693 -- only occurs once). The pre-inliner won't have gotten
694 -- it either, if the thing occurs in more than one branch
695 -- So the main target is things like
698 -- True -> case x of ...
699 -- False -> case x of ...
700 -- I'm not sure how important this is in practice
701 OneOcc in_lam one_br int_cxt -- OneOcc => no code-duplication issue
702 -> smallEnoughToInline unfolding -- Small enough to dup
703 -- ToDo: consider discount on smallEnoughToInline if int_cxt is true
705 -- NB: Do NOT inline arbitrarily big things, even if one_br is True
706 -- Reason: doing so risks exponential behaviour. We simplify a big
707 -- expression, inline it, and simplify it again. But if the
708 -- very same thing happens in the big expression, we get
710 -- PRINCIPLE: when we've already simplified an expression once,
711 -- make sure that we only inline it if it's reasonably small.
713 && ((isNotTopLevel top_lvl && not in_lam) ||
714 -- But outside a lambda, we want to be reasonably aggressive
715 -- about inlining into multiple branches of case
716 -- e.g. let x = <non-value>
717 -- in case y of { C1 -> ..x..; C2 -> ..x..; C3 -> ... }
718 -- Inlining can be a big win if C3 is the hot-spot, even if
719 -- the uses in C1, C2 are not 'interesting'
720 -- An example that gets worse if you add int_cxt here is 'clausify'
722 (isCheapUnfolding unfolding && int_cxt))
723 -- isCheap => acceptable work duplication; in_lam may be true
724 -- int_cxt to prevent us inlining inside a lambda without some
725 -- good reason. See the notes on int_cxt in preInlineUnconditionally
727 IAmDead -> True -- This happens; for example, the case_bndr during case of
728 -- known constructor: case (a,b) of x { (p,q) -> ... }
729 -- Here x isn't mentioned in the RHS, so we don't want to
730 -- create the (dead) let-binding let x = (a,b) in ...
734 -- Here's an example that we don't handle well:
735 -- let f = if b then Left (\x.BIG) else Right (\y.BIG)
736 -- in \y. ....case f of {...} ....
737 -- Here f is used just once, and duplicating the case work is fine (exprIsCheap).
739 -- * We can't preInlineUnconditionally because that woud invalidate
740 -- the occ info for b.
741 -- * We can't postInlineUnconditionally because the RHS is big, and
742 -- that risks exponential behaviour
743 -- * We can't call-site inline, because the rhs is big
747 active = case getMode env of
748 SimplGently -> isAlwaysActive prag
749 SimplPhase n -> isActive n prag
750 prag = idInlinePragma bndr
752 activeInline :: SimplEnv -> OutId -> Bool
754 = case getMode env of
756 -- No inlining at all when doing gentle stuff,
757 -- except for local things that occur once
758 -- The reason is that too little clean-up happens if you
759 -- don't inline use-once things. Also a bit of inlining is *good* for
760 -- full laziness; it can expose constant sub-expressions.
761 -- Example in spectral/mandel/Mandel.hs, where the mandelset
762 -- function gets a useful let-float if you inline windowToViewport
764 -- NB: we used to have a second exception, for data con wrappers.
765 -- On the grounds that we use gentle mode for rule LHSs, and
766 -- they match better when data con wrappers are inlined.
767 -- But that only really applies to the trivial wrappers (like (:)),
768 -- and they are now constructed as Compulsory unfoldings (in MkId)
769 -- so they'll happen anyway.
771 SimplPhase n -> isActive n prag
773 prag = idInlinePragma id
775 activeRule :: DynFlags -> SimplEnv -> Maybe (Activation -> Bool)
776 -- Nothing => No rules at all
777 activeRule dflags env
778 | not (dopt Opt_RewriteRules dflags)
779 = Nothing -- Rewriting is off
781 = case getMode env of
782 SimplGently -> Just isAlwaysActive
783 -- Used to be Nothing (no rules in gentle mode)
784 -- Main motivation for changing is that I wanted
785 -- lift String ===> ...
786 -- to work in Template Haskell when simplifying
787 -- splices, so we get simpler code for literal strings
788 SimplPhase n -> Just (isActive n)
792 %************************************************************************
796 %************************************************************************
799 mkLam :: [OutBndr] -> OutExpr -> SimplM OutExpr
800 -- mkLam tries three things
801 -- a) eta reduction, if that gives a trivial expression
802 -- b) eta expansion [only if there are some value lambdas]
807 = do { dflags <- getDOptsSmpl
808 ; mkLam' dflags bndrs body }
810 mkLam' :: DynFlags -> [OutBndr] -> OutExpr -> SimplM OutExpr
811 mkLam' dflags bndrs (Cast body co)
812 | not (any bad bndrs)
813 -- Note [Casts and lambdas]
814 = do { lam <- mkLam' dflags bndrs body
815 ; return (mkCoerce (mkPiTypes bndrs co) lam) }
817 co_vars = tyVarsOfType co
818 bad bndr = isCoVar bndr && bndr `elemVarSet` co_vars
820 mkLam' dflags bndrs body
821 | dopt Opt_DoEtaReduction dflags,
822 Just etad_lam <- tryEtaReduce bndrs body
823 = do { tick (EtaReduction (head bndrs))
826 | dopt Opt_DoLambdaEtaExpansion dflags,
827 any isRuntimeVar bndrs
828 = do { body' <- tryEtaExpansion dflags body
829 ; return (mkLams bndrs body') }
832 = returnSmpl (mkLams bndrs body)
835 Note [Casts and lambdas]
836 ~~~~~~~~~~~~~~~~~~~~~~~~
838 (\x. (\y. e) `cast` g1) `cast` g2
839 There is a danger here that the two lambdas look separated, and the
840 full laziness pass might float an expression to between the two.
842 So this equation in mkLam' floats the g1 out, thus:
843 (\x. e `cast` g1) --> (\x.e) `cast` (tx -> g1)
846 In general, this floats casts outside lambdas, where (I hope) they
847 might meet and cancel with some other cast:
848 \x. e `cast` co ===> (\x. e) `cast` (tx -> co)
849 /\a. e `cast` co ===> (/\a. e) `cast` (/\a. co)
850 /\g. e `cast` co ===> (/\g. e) `cast` (/\g. co)
853 Notice that it works regardless of 'e'. Originally it worked only
854 if 'e' was itself a lambda, but in some cases that resulted in
855 fruitless iteration in the simplifier. A good example was when
856 compiling Text.ParserCombinators.ReadPrec, where we had a definition
857 like (\x. Get `cast` g)
858 where Get is a constructor with nonzero arity. Then mkLam eta-expanded
859 the Get, and the next iteration eta-reduced it, and then eta-expanded
862 Note also the side condition for the case of coercion binders.
863 It does not make sense to transform
864 /\g. e `cast` g ==> (/\g.e) `cast` (/\g.g)
865 because the latter is not well-kinded.
867 -- c) floating lets out through big lambdas
868 -- [only if all tyvar lambdas, and only if this lambda
869 -- is the RHS of a let]
871 {- Sept 01: I'm experimenting with getting the
872 full laziness pass to float out past big lambdsa
873 | all isTyVar bndrs, -- Only for big lambdas
874 contIsRhs cont -- Only try the rhs type-lambda floating
875 -- if this is indeed a right-hand side; otherwise
876 -- we end up floating the thing out, only for float-in
877 -- to float it right back in again!
878 = tryRhsTyLam env bndrs body `thenSmpl` \ (floats, body') ->
879 returnSmpl (floats, mkLams bndrs body')
883 %************************************************************************
887 %************************************************************************
889 Note [Eta reduction conditions]
890 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
891 We try for eta reduction here, but *only* if we get all the way to an
892 trivial expression. We don't want to remove extra lambdas unless we
893 are going to avoid allocating this thing altogether.
895 There are some particularly delicate points here:
897 * Eta reduction is not valid in general:
899 This matters, partly for old-fashioned correctness reasons but,
900 worse, getting it wrong can yield a seg fault. Consider
902 h y = case (case y of { True -> f `seq` True; False -> False }) of
903 True -> ...; False -> ...
905 If we (unsoundly) eta-reduce f to get f=f, the strictness analyser
906 says f=bottom, and replaces the (f `seq` True) with just
907 (f `cast` unsafe-co). BUT, as thing stand, 'f' got arity 1, and it
908 *keeps* arity 1 (perhaps also wrongly). So CorePrep eta-expands
909 the definition again, so that it does not termninate after all.
910 Result: seg-fault because the boolean case actually gets a function value.
913 So it's important to to the right thing.
915 * We need to be careful if we just look at f's arity. Currently (Dec07),
916 f's arity is visible in its own RHS (see Note [Arity robustness] in
917 SimplEnv) so we must *not* trust the arity when checking that 'f' is
918 a value. Instead, look at the unfolding.
920 However for GlobalIds we can look at the arity; and for primops we
921 must, since they have no unfolding.
923 * Regardless of whether 'f' is a vlaue, we always want to
924 reduce (/\a -> f a) to f
925 This came up in a RULE: foldr (build (/\a -> g a))
926 did not match foldr (build (/\b -> ...something complex...))
927 The type checker can insert these eta-expanded versions,
928 with both type and dictionary lambdas; hence the slightly
931 These delicacies are why we don't use exprIsTrivial and exprIsHNF here.
935 tryEtaReduce :: [OutBndr] -> OutExpr -> Maybe OutExpr
936 tryEtaReduce bndrs body
937 = go (reverse bndrs) body
939 go (b : bs) (App fun arg) | ok_arg b arg = go bs fun -- Loop round
940 go [] fun | ok_fun fun = Just fun -- Success!
941 go _ _ = Nothing -- Failure!
943 -- Note [Eta reduction conditions]
944 ok_fun (App fun (Type ty))
945 | not (any (`elemVarSet` tyVarsOfType ty) bndrs)
948 = not (fun_id `elem` bndrs)
949 && (ok_fun_id fun_id || all ok_lam bndrs)
953 | isLocalId fun = isEvaldUnfolding (idUnfolding fun)
954 | isDataConWorkId fun = True
955 | isGlobalId fun = idArity fun > 0
957 ok_lam v = isTyVar v || isDictId v
959 ok_arg b arg = varToCoreExpr b `cheapEqExpr` arg
963 %************************************************************************
967 %************************************************************************
971 f = \x1..xn -> N ==> f = \x1..xn y1..ym -> N y1..ym
974 where (in both cases)
976 * The xi can include type variables
978 * The yi are all value variables
980 * N is a NORMAL FORM (i.e. no redexes anywhere)
981 wanting a suitable number of extra args.
983 The biggest reason for doing this is for cases like
989 Here we want to get the lambdas together. A good exmaple is the nofib
990 program fibheaps, which gets 25% more allocation if you don't do this
993 We may have to sandwich some coerces between the lambdas
994 to make the types work. exprEtaExpandArity looks through coerces
995 when computing arity; and etaExpand adds the coerces as necessary when
996 actually computing the expansion.
999 tryEtaExpansion :: DynFlags -> OutExpr -> SimplM OutExpr
1000 -- There is at least one runtime binder in the binders
1001 tryEtaExpansion dflags body
1002 = getUniquesSmpl `thenSmpl` \ us ->
1003 returnSmpl (etaExpand fun_arity us body (exprType body))
1005 fun_arity = exprEtaExpandArity dflags body
1009 %************************************************************************
1011 \subsection{Floating lets out of big lambdas}
1013 %************************************************************************
1015 Note [Floating and type abstraction]
1016 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
1019 We'd like to float this to
1022 x = /\a. C (y1 a) (y2 a)
1023 for the usual reasons: we want to inline x rather vigorously.
1025 You may think that this kind of thing is rare. But in some programs it is
1026 common. For example, if you do closure conversion you might get:
1028 data a :-> b = forall e. (e -> a -> b) :$ e
1030 f_cc :: forall a. a :-> a
1031 f_cc = /\a. (\e. id a) :$ ()
1033 Now we really want to inline that f_cc thing so that the
1034 construction of the closure goes away.
1036 So I have elaborated simplLazyBind to understand right-hand sides that look
1040 and treat them specially. The real work is done in SimplUtils.abstractFloats,
1041 but there is quite a bit of plumbing in simplLazyBind as well.
1043 The same transformation is good when there are lets in the body:
1045 /\abc -> let(rec) x = e in b
1047 let(rec) x' = /\abc -> let x = x' a b c in e
1049 /\abc -> let x = x' a b c in b
1051 This is good because it can turn things like:
1053 let f = /\a -> letrec g = ... g ... in g
1055 letrec g' = /\a -> ... g' a ...
1057 let f = /\ a -> g' a
1059 which is better. In effect, it means that big lambdas don't impede
1062 This optimisation is CRUCIAL in eliminating the junk introduced by
1063 desugaring mutually recursive definitions. Don't eliminate it lightly!
1065 [May 1999] If we do this transformation *regardless* then we can
1066 end up with some pretty silly stuff. For example,
1069 st = /\ s -> let { x1=r1 ; x2=r2 } in ...
1074 st = /\s -> ...[y1 s/x1, y2 s/x2]
1077 Unless the "..." is a WHNF there is really no point in doing this.
1078 Indeed it can make things worse. Suppose x1 is used strictly,
1081 x1* = case f y of { (a,b) -> e }
1083 If we abstract this wrt the tyvar we then can't do the case inline
1084 as we would normally do.
1086 That's why the whole transformation is part of the same process that
1087 floats let-bindings and constructor arguments out of RHSs. In particular,
1088 it is guarded by the doFloatFromRhs call in simplLazyBind.
1092 abstractFloats :: [OutTyVar] -> SimplEnv -> OutExpr -> SimplM ([OutBind], OutExpr)
1093 abstractFloats main_tvs body_env body
1094 = ASSERT( notNull body_floats )
1095 do { (subst, float_binds) <- mapAccumLSmpl abstract empty_subst body_floats
1096 ; return (float_binds, CoreSubst.substExpr subst body) }
1098 main_tv_set = mkVarSet main_tvs
1099 body_floats = getFloats body_env
1100 empty_subst = CoreSubst.mkEmptySubst (seInScope body_env)
1102 abstract :: CoreSubst.Subst -> OutBind -> SimplM (CoreSubst.Subst, OutBind)
1103 abstract subst (NonRec id rhs)
1104 = do { (poly_id, poly_app) <- mk_poly tvs_here id
1105 ; let poly_rhs = mkLams tvs_here rhs'
1106 subst' = CoreSubst.extendIdSubst subst id poly_app
1107 ; return (subst', (NonRec poly_id poly_rhs)) }
1109 rhs' = CoreSubst.substExpr subst rhs
1110 tvs_here | any isCoVar main_tvs = main_tvs -- Note [Abstract over coercions]
1112 = varSetElems (main_tv_set `intersectVarSet` exprSomeFreeVars isTyVar rhs')
1114 -- Abstract only over the type variables free in the rhs
1115 -- wrt which the new binding is abstracted. But the naive
1116 -- approach of abstract wrt the tyvars free in the Id's type
1118 -- /\ a b -> let t :: (a,b) = (e1, e2)
1121 -- Here, b isn't free in x's type, but we must nevertheless
1122 -- abstract wrt b as well, because t's type mentions b.
1123 -- Since t is floated too, we'd end up with the bogus:
1124 -- poly_t = /\ a b -> (e1, e2)
1125 -- poly_x = /\ a -> fst (poly_t a *b*)
1126 -- So for now we adopt the even more naive approach of
1127 -- abstracting wrt *all* the tyvars. We'll see if that
1128 -- gives rise to problems. SLPJ June 98
1130 abstract subst (Rec prs)
1131 = do { (poly_ids, poly_apps) <- mapAndUnzipSmpl (mk_poly tvs_here) ids
1132 ; let subst' = CoreSubst.extendSubstList subst (ids `zip` poly_apps)
1133 poly_rhss = [mkLams tvs_here (CoreSubst.substExpr subst' rhs) | rhs <- rhss]
1134 ; return (subst', Rec (poly_ids `zip` poly_rhss)) }
1136 (ids,rhss) = unzip prs
1137 -- For a recursive group, it's a bit of a pain to work out the minimal
1138 -- set of tyvars over which to abstract:
1139 -- /\ a b c. let x = ...a... in
1140 -- letrec { p = ...x...q...
1141 -- q = .....p...b... } in
1143 -- Since 'x' is abstracted over 'a', the {p,q} group must be abstracted
1144 -- over 'a' (because x is replaced by (poly_x a)) as well as 'b'.
1145 -- Since it's a pain, we just use the whole set, which is always safe
1147 -- If you ever want to be more selective, remember this bizarre case too:
1149 -- Here, we must abstract 'x' over 'a'.
1152 mk_poly tvs_here var
1153 = do { uniq <- getUniqueSmpl
1154 ; let poly_name = setNameUnique (idName var) uniq -- Keep same name
1155 poly_ty = mkForAllTys tvs_here (idType var) -- But new type of course
1156 poly_id = mkLocalId poly_name poly_ty
1157 ; return (poly_id, mkTyApps (Var poly_id) (mkTyVarTys tvs_here)) }
1158 -- In the olden days, it was crucial to copy the occInfo of the original var,
1159 -- because we were looking at occurrence-analysed but as yet unsimplified code!
1160 -- In particular, we mustn't lose the loop breakers. BUT NOW we are looking
1161 -- at already simplified code, so it doesn't matter
1163 -- It's even right to retain single-occurrence or dead-var info:
1164 -- Suppose we started with /\a -> let x = E in B
1165 -- where x occurs once in B. Then we transform to:
1166 -- let x' = /\a -> E in /\a -> let x* = x' a in B
1167 -- where x* has an INLINE prag on it. Now, once x* is inlined,
1168 -- the occurrences of x' will be just the occurrences originally
1172 Note [Abstract over coercions]
1173 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
1174 If a coercion variable (g :: a ~ Int) is free in the RHS, then so is the
1175 type variable a. Rather than sort this mess out, we simply bale out and abstract
1176 wrt all the type variables if any of them are coercion variables.
1179 Historical note: if you use let-bindings instead of a substitution, beware of this:
1181 -- Suppose we start with:
1183 -- x = /\ a -> let g = G in E
1185 -- Then we'll float to get
1187 -- x = let poly_g = /\ a -> G
1188 -- in /\ a -> let g = poly_g a in E
1190 -- But now the occurrence analyser will see just one occurrence
1191 -- of poly_g, not inside a lambda, so the simplifier will
1192 -- PreInlineUnconditionally poly_g back into g! Badk to square 1!
1193 -- (I used to think that the "don't inline lone occurrences" stuff
1194 -- would stop this happening, but since it's the *only* occurrence,
1195 -- PreInlineUnconditionally kicks in first!)
1197 -- Solution: put an INLINE note on g's RHS, so that poly_g seems
1198 -- to appear many times. (NB: mkInlineMe eliminates
1199 -- such notes on trivial RHSs, so do it manually.)
1201 %************************************************************************
1205 %************************************************************************
1207 prepareAlts tries these things:
1209 1. If several alternatives are identical, merge them into
1210 a single DEFAULT alternative. I've occasionally seen this
1211 making a big difference:
1213 case e of =====> case e of
1214 C _ -> f x D v -> ....v....
1215 D v -> ....v.... DEFAULT -> f x
1218 The point is that we merge common RHSs, at least for the DEFAULT case.
1219 [One could do something more elaborate but I've never seen it needed.]
1220 To avoid an expensive test, we just merge branches equal to the *first*
1221 alternative; this picks up the common cases
1222 a) all branches equal
1223 b) some branches equal to the DEFAULT (which occurs first)
1226 case e of b { ==> case e of b {
1227 p1 -> rhs1 p1 -> rhs1
1229 pm -> rhsm pm -> rhsm
1230 _ -> case b of b' { pn -> let b'=b in rhsn
1232 ... po -> let b'=b in rhso
1233 po -> rhso _ -> let b'=b in rhsd
1237 which merges two cases in one case when -- the default alternative of
1238 the outer case scrutises the same variable as the outer case This
1239 transformation is called Case Merging. It avoids that the same
1240 variable is scrutinised multiple times.
1243 The case where transformation (1) showed up was like this (lib/std/PrelCError.lhs):
1249 where @is@ was something like
1251 p `is` n = p /= (-1) && p == n
1253 This gave rise to a horrible sequence of cases
1260 and similarly in cascade for all the join points!
1263 ~~~~~~~~~~~~~~~~~~~~
1264 We do this *here*, looking at un-simplified alternatives, because we
1265 have to check that r doesn't mention the variables bound by the
1266 pattern in each alternative, so the binder-info is rather useful.
1269 prepareAlts :: SimplEnv -> OutExpr -> OutId -> [InAlt] -> SimplM ([AltCon], [InAlt])
1270 prepareAlts env scrut case_bndr' alts
1271 = do { dflags <- getDOptsSmpl
1272 ; alts <- combineIdenticalAlts case_bndr' alts
1274 ; let (alts_wo_default, maybe_deflt) = findDefault alts
1275 alt_cons = [con | (con,_,_) <- alts_wo_default]
1276 imposs_deflt_cons = nub (imposs_cons ++ alt_cons)
1277 -- "imposs_deflt_cons" are handled
1278 -- EITHER by the context,
1279 -- OR by a non-DEFAULT branch in this case expression.
1281 ; default_alts <- prepareDefault dflags env case_bndr' mb_tc_app
1282 imposs_deflt_cons maybe_deflt
1284 ; let trimmed_alts = filterOut impossible_alt alts_wo_default
1285 merged_alts = mergeAlts trimmed_alts default_alts
1286 -- We need the mergeAlts in case the new default_alt
1287 -- has turned into a constructor alternative.
1288 -- The merge keeps the inner DEFAULT at the front, if there is one
1289 -- and interleaves the alternatives in the right order
1291 ; return (imposs_deflt_cons, merged_alts) }
1293 mb_tc_app = splitTyConApp_maybe (idType case_bndr')
1294 Just (_, inst_tys) = mb_tc_app
1296 imposs_cons = case scrut of
1297 Var v -> otherCons (idUnfolding v)
1300 impossible_alt :: CoreAlt -> Bool
1301 impossible_alt (con, _, _) | con `elem` imposs_cons = True
1302 impossible_alt (DataAlt con, _, _) = dataConCannotMatch inst_tys con
1303 impossible_alt alt = False
1306 --------------------------------------------------
1307 -- 1. Merge identical branches
1308 --------------------------------------------------
1309 combineIdenticalAlts :: OutId -> [InAlt] -> SimplM [InAlt]
1311 combineIdenticalAlts case_bndr alts@((con1,bndrs1,rhs1) : con_alts)
1312 | all isDeadBinder bndrs1, -- Remember the default
1313 length filtered_alts < length con_alts -- alternative comes first
1314 -- Also Note [Dead binders]
1315 = do { tick (AltMerge case_bndr)
1316 ; return ((DEFAULT, [], rhs1) : filtered_alts) }
1318 filtered_alts = filter keep con_alts
1319 keep (con,bndrs,rhs) = not (all isDeadBinder bndrs && rhs `cheapEqExpr` rhs1)
1321 combineIdenticalAlts case_bndr alts = return alts
1323 -------------------------------------------------------------------------
1324 -- Prepare the default alternative
1325 -------------------------------------------------------------------------
1326 prepareDefault :: DynFlags
1328 -> OutId -- Case binder; need just for its type. Note that as an
1329 -- OutId, it has maximum information; this is important.
1330 -- Test simpl013 is an example
1331 -> Maybe (TyCon, [Type]) -- Type of scrutinee, decomposed
1332 -> [AltCon] -- These cons can't happen when matching the default
1333 -> Maybe InExpr -- Rhs
1334 -> SimplM [InAlt] -- Still unsimplified
1335 -- We use a list because it's what mergeAlts expects,
1336 -- And becuase case-merging can cause many to show up
1338 ------- Merge nested cases ----------
1339 prepareDefault dflags env outer_bndr bndr_ty imposs_cons (Just deflt_rhs)
1340 | dopt Opt_CaseMerge dflags
1341 , Case (Var inner_scrut_var) inner_bndr _ inner_alts <- deflt_rhs
1342 , DoneId inner_scrut_var' <- substId env inner_scrut_var
1343 -- Remember, inner_scrut_var is an InId, but outer_bndr is an OutId
1344 , inner_scrut_var' == outer_bndr
1345 -- NB: the substId means that if the outer scrutinee was a
1346 -- variable, and inner scrutinee is the same variable,
1347 -- then inner_scrut_var' will be outer_bndr
1348 -- via the magic of simplCaseBinder
1349 = do { tick (CaseMerge outer_bndr)
1351 ; let munge_rhs rhs = bindCaseBndr inner_bndr (Var outer_bndr) rhs
1352 ; return [(con, args, munge_rhs rhs) | (con, args, rhs) <- inner_alts,
1353 not (con `elem` imposs_cons) ]
1354 -- NB: filter out any imposs_cons. Example:
1357 -- DEFAULT -> case x of
1360 -- When we merge, we must ensure that e1 takes
1361 -- precedence over e2 as the value for A!
1363 -- Warning: don't call prepareAlts recursively!
1364 -- Firstly, there's no point, because inner alts have already had
1365 -- mkCase applied to them, so they won't have a case in their default
1366 -- Secondly, if you do, you get an infinite loop, because the bindCaseBndr
1367 -- in munge_rhs may put a case into the DEFAULT branch!
1370 --------- Fill in known constructor -----------
1371 prepareDefault dflags env case_bndr (Just (tycon, inst_tys)) imposs_cons (Just deflt_rhs)
1372 | -- This branch handles the case where we are
1373 -- scrutinisng an algebraic data type
1374 isAlgTyCon tycon -- It's a data type, tuple, or unboxed tuples.
1375 , not (isNewTyCon tycon) -- We can have a newtype, if we are just doing an eval:
1376 -- case x of { DEFAULT -> e }
1377 -- and we don't want to fill in a default for them!
1378 , Just all_cons <- tyConDataCons_maybe tycon
1379 , not (null all_cons) -- This is a tricky corner case. If the data type has no constructors,
1380 -- which GHC allows, then the case expression will have at most a default
1381 -- alternative. We don't want to eliminate that alternative, because the
1382 -- invariant is that there's always one alternative. It's more convenient
1384 -- case x of { DEFAULT -> e }
1385 -- as it is, rather than transform it to
1386 -- error "case cant match"
1387 -- which would be quite legitmate. But it's a really obscure corner, and
1388 -- not worth wasting code on.
1389 , let imposs_data_cons = [con | DataAlt con <- imposs_cons] -- We now know it's a data type
1390 impossible con = con `elem` imposs_data_cons || dataConCannotMatch inst_tys con
1391 = case filterOut impossible all_cons of
1392 [] -> return [] -- Eliminate the default alternative
1393 -- altogether if it can't match
1395 [con] -> -- It matches exactly one constructor, so fill it in
1396 do { tick (FillInCaseDefault case_bndr)
1397 ; us <- getUniquesSmpl
1398 ; let (ex_tvs, co_tvs, arg_ids) =
1399 dataConRepInstPat us con inst_tys
1400 ; return [(DataAlt con, ex_tvs ++ co_tvs ++ arg_ids, deflt_rhs)] }
1402 two_or_more -> return [(DEFAULT, [], deflt_rhs)]
1404 --------- Catch-all cases -----------
1405 prepareDefault dflags env case_bndr bndr_ty imposs_cons (Just deflt_rhs)
1406 = return [(DEFAULT, [], deflt_rhs)]
1408 prepareDefault dflags env case_bndr bndr_ty imposs_cons Nothing
1409 = return [] -- No default branch
1414 =================================================================================
1416 mkCase tries these things
1418 1. Eliminate the case altogether if possible
1426 and similar friends.
1430 mkCase :: OutExpr -> OutId -> OutType
1431 -> [OutAlt] -- Increasing order
1434 --------------------------------------------------
1435 -- 1. Check for empty alternatives
1436 --------------------------------------------------
1438 -- This isn't strictly an error. It's possible that the simplifer might "see"
1439 -- that an inner case has no accessible alternatives before it "sees" that the
1440 -- entire branch of an outer case is inaccessible. So we simply
1441 -- put an error case here insteadd
1442 mkCase scrut case_bndr ty []
1443 = pprTrace "mkCase: null alts" (ppr case_bndr <+> ppr scrut) $
1444 return (mkApps (Var rUNTIME_ERROR_ID)
1445 [Type ty, Lit (mkStringLit "Impossible alternative")])
1448 --------------------------------------------------
1450 --------------------------------------------------
1452 mkCase scrut case_bndr ty alts -- Identity case
1453 | all identity_alt alts
1454 = tick (CaseIdentity case_bndr) `thenSmpl_`
1455 returnSmpl (re_cast scrut)
1457 identity_alt (con, args, rhs) = check_eq con args (de_cast rhs)
1459 check_eq DEFAULT _ (Var v) = v == case_bndr
1460 check_eq (LitAlt lit') _ (Lit lit) = lit == lit'
1461 check_eq (DataAlt con) args rhs = rhs `cheapEqExpr` mkConApp con (arg_tys ++ varsToCoreExprs args)
1462 || rhs `cheapEqExpr` Var case_bndr
1463 check_eq con args rhs = False
1465 arg_tys = map Type (tyConAppArgs (idType case_bndr))
1468 -- case e of x { _ -> x `cast` c }
1469 -- And we definitely want to eliminate this case, to give
1471 -- So we throw away the cast from the RHS, and reconstruct
1472 -- it at the other end. All the RHS casts must be the same
1473 -- if (all identity_alt alts) holds.
1475 -- Don't worry about nested casts, because the simplifier combines them
1476 de_cast (Cast e _) = e
1479 re_cast scrut = case head alts of
1480 (_,_,Cast _ co) -> Cast scrut co
1485 --------------------------------------------------
1487 --------------------------------------------------
1488 mkCase scrut bndr ty alts = returnSmpl (Case scrut bndr ty alts)
1492 When adding auxiliary bindings for the case binder, it's worth checking if
1493 its dead, because it often is, and occasionally these mkCase transformations
1494 cascade rather nicely.
1497 bindCaseBndr bndr rhs body
1498 | isDeadBinder bndr = body
1499 | otherwise = bindNonRec bndr rhs body