2 % (c) The AQUA Project, Glasgow University, 1993-1998
4 \section[SimplUtils]{The simplifier utilities}
9 mkLam, mkCase, prepareAlts, bindCaseBndr,
12 preInlineUnconditionally, postInlineUnconditionally,
13 activeInline, activeRule, inlineMode,
15 -- The continuation type
16 SimplCont(..), DupFlag(..), LetRhsFlag(..),
17 contIsDupable, contResultType, contIsTrivial, contArgs, dropArgs,
18 countValArgs, countArgs,
19 mkBoringStop, mkLazyArgStop, mkRhsStop, contIsRhsOrArg,
20 interestingCallContext, interestingArgContext,
22 interestingArg, mkArgInfo,
27 #include "HsVersions.h"
33 import qualified CoreSubst
42 import Var ( isCoVar )
48 import Unify ( dataConCannotMatch )
57 %************************************************************************
61 %************************************************************************
63 A SimplCont allows the simplifier to traverse the expression in a
64 zipper-like fashion. The SimplCont represents the rest of the expression,
65 "above" the point of interest.
67 You can also think of a SimplCont as an "evaluation context", using
68 that term in the way it is used for operational semantics. This is the
69 way I usually think of it, For example you'll often see a syntax for
70 evaluation context looking like
71 C ::= [] | C e | case C of alts | C `cast` co
72 That's the kind of thing we are doing here, and I use that syntax in
77 * A SimplCont describes a *strict* context (just like
78 evaluation contexts do). E.g. Just [] is not a SimplCont
80 * A SimplCont describes a context that *does not* bind
81 any variables. E.g. \x. [] is not a SimplCont
85 = Stop -- An empty context, or hole, []
86 OutType -- Type of the result
88 Bool -- True <=> There is something interesting about
89 -- the context, and hence the inliner
90 -- should be a bit keener (see interestingCallContext)
92 -- (a) This is the RHS of a thunk whose type suggests
93 -- that update-in-place would be possible
94 -- (b) This is an argument of a function that has RULES
95 -- Inlining the call might allow the rule to fire
97 | CoerceIt -- C `cast` co
98 OutCoercion -- The coercion simplified
103 InExpr SimplEnv -- The argument and its static env
106 | Select -- case C of alts
108 InId [InAlt] SimplEnv -- The case binder, alts, and subst-env
111 -- The two strict forms have no DupFlag, because we never duplicate them
112 | StrictBind -- (\x* \xs. e) C
113 InId [InBndr] -- let x* = [] in e
114 InExpr SimplEnv -- is a special case
118 OutExpr OutType -- e and its type
119 (Bool,[Bool]) -- Whether the function at the head of e has rules,
120 SimplCont -- plus strictness flags for further args
122 data LetRhsFlag = AnArg -- It's just an argument not a let RHS
123 | AnRhs -- It's the RHS of a let (so please float lets out of big lambdas)
125 instance Outputable LetRhsFlag where
126 ppr AnArg = ptext SLIT("arg")
127 ppr AnRhs = ptext SLIT("rhs")
129 instance Outputable SimplCont where
130 ppr (Stop ty is_rhs _) = ptext SLIT("Stop") <> brackets (ppr is_rhs) <+> ppr ty
131 ppr (ApplyTo dup arg se cont) = ((ptext SLIT("ApplyTo") <+> ppr dup <+> pprParendExpr arg)
132 {- $$ nest 2 (pprSimplEnv se) -}) $$ ppr cont
133 ppr (StrictBind b _ _ _ cont) = (ptext SLIT("StrictBind") <+> ppr b) $$ ppr cont
134 ppr (StrictArg f _ _ cont) = (ptext SLIT("StrictArg") <+> ppr f) $$ ppr cont
135 ppr (Select dup bndr alts se cont) = (ptext SLIT("Select") <+> ppr dup <+> ppr bndr) $$
136 (nest 4 (ppr alts)) $$ ppr cont
137 ppr (CoerceIt co cont) = (ptext SLIT("CoerceIt") <+> ppr co) $$ ppr cont
139 data DupFlag = OkToDup | NoDup
141 instance Outputable DupFlag where
142 ppr OkToDup = ptext SLIT("ok")
143 ppr NoDup = ptext SLIT("nodup")
148 mkBoringStop :: OutType -> SimplCont
149 mkBoringStop ty = Stop ty AnArg False
151 mkLazyArgStop :: OutType -> Bool -> SimplCont
152 mkLazyArgStop ty has_rules = Stop ty AnArg (canUpdateInPlace ty || has_rules)
154 mkRhsStop :: OutType -> SimplCont
155 mkRhsStop ty = Stop ty AnRhs (canUpdateInPlace ty)
157 contIsRhsOrArg (Stop {}) = True
158 contIsRhsOrArg (StrictBind {}) = True
159 contIsRhsOrArg (StrictArg {}) = True
160 contIsRhsOrArg other = False
163 contIsDupable :: SimplCont -> Bool
164 contIsDupable (Stop {}) = True
165 contIsDupable (ApplyTo OkToDup _ _ _) = True
166 contIsDupable (Select OkToDup _ _ _ _) = True
167 contIsDupable (CoerceIt _ cont) = contIsDupable cont
168 contIsDupable other = False
171 contIsTrivial :: SimplCont -> Bool
172 contIsTrivial (Stop {}) = True
173 contIsTrivial (ApplyTo _ (Type _) _ cont) = contIsTrivial cont
174 contIsTrivial (CoerceIt _ cont) = contIsTrivial cont
175 contIsTrivial other = False
178 contResultType :: SimplCont -> OutType
179 contResultType (Stop to_ty _ _) = to_ty
180 contResultType (StrictArg _ _ _ cont) = contResultType cont
181 contResultType (StrictBind _ _ _ _ cont) = contResultType cont
182 contResultType (ApplyTo _ _ _ cont) = contResultType cont
183 contResultType (CoerceIt _ cont) = contResultType cont
184 contResultType (Select _ _ _ _ cont) = contResultType cont
187 countValArgs :: SimplCont -> Int
188 countValArgs (ApplyTo _ (Type ty) se cont) = countValArgs cont
189 countValArgs (ApplyTo _ val_arg se cont) = 1 + countValArgs cont
190 countValArgs other = 0
192 countArgs :: SimplCont -> Int
193 countArgs (ApplyTo _ arg se cont) = 1 + countArgs cont
196 contArgs :: SimplCont -> ([OutExpr], SimplCont)
197 -- Uses substitution to turn each arg into an OutExpr
198 contArgs cont = go [] cont
200 go args (ApplyTo _ arg se cont) = go (substExpr se arg : args) cont
201 go args cont = (reverse args, cont)
203 dropArgs :: Int -> SimplCont -> SimplCont
204 dropArgs 0 cont = cont
205 dropArgs n (ApplyTo _ _ _ cont) = dropArgs (n-1) cont
206 dropArgs n other = pprPanic "dropArgs" (ppr n <+> ppr other)
211 interestingArg :: OutExpr -> Bool
212 -- An argument is interesting if it has *some* structure
213 -- We are here trying to avoid unfolding a function that
214 -- is applied only to variables that have no unfolding
215 -- (i.e. they are probably lambda bound): f x y z
216 -- There is little point in inlining f here.
217 interestingArg (Var v) = hasSomeUnfolding (idUnfolding v)
218 -- Was: isValueUnfolding (idUnfolding v')
219 -- But that seems over-pessimistic
221 -- This accounts for an argument like
222 -- () or [], which is definitely interesting
223 interestingArg (Type _) = False
224 interestingArg (App fn (Type _)) = interestingArg fn
225 interestingArg (Note _ a) = interestingArg a
227 -- Idea (from Sam B); I'm not sure if it's a good idea, so commented out for now
228 -- interestingArg expr | isUnLiftedType (exprType expr)
229 -- -- Unlifted args are only ever interesting if we know what they are
234 interestingArg other = True
235 -- Consider let x = 3 in f x
236 -- The substitution will contain (x -> ContEx 3), and we want to
237 -- to say that x is an interesting argument.
238 -- But consider also (\x. f x y) y
239 -- The substitution will contain (x -> ContEx y), and we want to say
240 -- that x is not interesting (assuming y has no unfolding)
244 Comment about interestingCallContext
245 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
246 We want to avoid inlining an expression where there can't possibly be
247 any gain, such as in an argument position. Hence, if the continuation
248 is interesting (eg. a case scrutinee, application etc.) then we
249 inline, otherwise we don't.
251 Previously some_benefit used to return True only if the variable was
252 applied to some value arguments. This didn't work:
254 let x = _coerce_ (T Int) Int (I# 3) in
255 case _coerce_ Int (T Int) x of
258 we want to inline x, but can't see that it's a constructor in a case
259 scrutinee position, and some_benefit is False.
263 dMonadST = _/\_ t -> :Monad (g1 _@_ t, g2 _@_ t, g3 _@_ t)
265 .... case dMonadST _@_ x0 of (a,b,c) -> ....
267 we'd really like to inline dMonadST here, but we *don't* want to
268 inline if the case expression is just
270 case x of y { DEFAULT -> ... }
272 since we can just eliminate this case instead (x is in WHNF). Similar
273 applies when x is bound to a lambda expression. Hence
274 contIsInteresting looks for case expressions with just a single
278 interestingCallContext :: Bool -- False <=> no args at all
279 -> Bool -- False <=> no value args
281 -- The "lone-variable" case is important. I spent ages
282 -- messing about with unsatisfactory varaints, but this is nice.
283 -- The idea is that if a variable appear all alone
284 -- as an arg of lazy fn, or rhs Stop
285 -- as scrutinee of a case Select
286 -- as arg of a strict fn ArgOf
287 -- then we should not inline it (unless there is some other reason,
288 -- e.g. is is the sole occurrence). We achieve this by making
289 -- interestingCallContext return False for a lone variable.
291 -- Why? At least in the case-scrutinee situation, turning
292 -- let x = (a,b) in case x of y -> ...
294 -- let x = (a,b) in case (a,b) of y -> ...
296 -- let x = (a,b) in let y = (a,b) in ...
297 -- is bad if the binding for x will remain.
299 -- Another example: I discovered that strings
300 -- were getting inlined straight back into applications of 'error'
301 -- because the latter is strict.
303 -- f = \x -> ...(error s)...
305 -- Fundamentally such contexts should not ecourage inlining because
306 -- the context can ``see'' the unfolding of the variable (e.g. case or a RULE)
307 -- so there's no gain.
309 -- However, even a type application or coercion isn't a lone variable.
311 -- case $fMonadST @ RealWorld of { :DMonad a b c -> c }
312 -- We had better inline that sucker! The case won't see through it.
314 -- For now, I'm treating treating a variable applied to types
315 -- in a *lazy* context "lone". The motivating example was
317 -- g = /\a. \y. h (f a)
318 -- There's no advantage in inlining f here, and perhaps
319 -- a significant disadvantage. Hence some_val_args in the Stop case
321 interestingCallContext some_args some_val_args cont
324 interesting (Select {}) = some_args
325 interesting (ApplyTo {}) = True -- Can happen if we have (coerce t (f x)) y
326 -- Perhaps True is a bit over-keen, but I've
327 -- seen (coerce f) x, where f has an INLINE prag,
328 -- So we have to give some motivaiton for inlining it
329 interesting (StrictArg {}) = some_val_args
330 interesting (StrictBind {}) = some_val_args -- ??
331 interesting (Stop ty _ interesting) = some_val_args && interesting
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
353 -> (Bool, [Bool]) -- Arg info
354 -- The arg info consists of
355 -- * A Bool indicating if the function has rules (recursively)
356 -- * A [Bool] indicating strictness for each arg
357 -- The [Bool] is usually infinite, but if it is finite it
358 -- guarantees that the function diverges after being given
359 -- that number of args
361 mkArgInfo fun n_val_args call_cont
362 = (interestingArgContext fun call_cont, fun_stricts)
364 vanilla_stricts, fun_stricts :: [Bool]
365 vanilla_stricts = repeat False
368 = case splitStrictSig (idNewStrictness fun) of
369 (demands, result_info)
370 | not (demands `lengthExceeds` n_val_args)
371 -> -- Enough args, use the strictness given.
372 -- For bottoming functions we used to pretend that the arg
373 -- is lazy, so that we don't treat the arg as an
374 -- interesting context. This avoids substituting
375 -- top-level bindings for (say) strings into
376 -- calls to error. But now we are more careful about
377 -- inlining lone variables, so its ok (see SimplUtils.analyseCont)
378 if isBotRes result_info then
379 map isStrictDmd demands -- Finite => result is bottom
381 map isStrictDmd demands ++ vanilla_stricts
383 other -> vanilla_stricts -- Not enough args, or no strictness
385 interestingArgContext :: Id -> SimplCont -> Bool
386 -- If the argument has form (f x y), where x,y are boring,
387 -- and f is marked INLINE, then we don't want to inline f.
388 -- But if the context of the argument is
390 -- where g has rules, then we *do* want to inline f, in case it
391 -- exposes a rule that might fire. Similarly, if the context is
393 -- where h has rules, then we do want to inline f.
394 -- The interesting_arg_ctxt flag makes this happen; if it's
395 -- set, the inliner gets just enough keener to inline f
396 -- regardless of how boring f's arguments are, if it's marked INLINE
398 -- The alternative would be to *always* inline an INLINE function,
399 -- regardless of how boring its context is; but that seems overkill
400 -- For example, it'd mean that wrapper functions were always inlined
401 interestingArgContext fn cont
402 = idHasRules fn || go cont
404 go (Select {}) = False
405 go (ApplyTo {}) = False
406 go (StrictArg {}) = True
407 go (StrictBind {}) = False -- ??
408 go (CoerceIt _ c) = go c
409 go (Stop _ _ interesting) = interesting
412 canUpdateInPlace :: Type -> Bool
413 -- Consider let x = <wurble> in ...
414 -- If <wurble> returns an explicit constructor, we might be able
415 -- to do update in place. So we treat even a thunk RHS context
416 -- as interesting if update in place is possible. We approximate
417 -- this by seeing if the type has a single constructor with a
418 -- small arity. But arity zero isn't good -- we share the single copy
419 -- for that case, so no point in sharing.
422 | not opt_UF_UpdateInPlace = False
424 = case splitTyConApp_maybe ty of
426 Just (tycon, _) -> case tyConDataCons_maybe tycon of
427 Just [dc] -> arity == 1 || arity == 2
429 arity = dataConRepArity dc
435 %************************************************************************
437 \subsection{Decisions about inlining}
439 %************************************************************************
441 Inlining is controlled partly by the SimplifierMode switch. This has two
444 SimplGently (a) Simplifying before specialiser/full laziness
445 (b) Simplifiying inside INLINE pragma
446 (c) Simplifying the LHS of a rule
447 (d) Simplifying a GHCi expression or Template
450 SimplPhase n Used at all other times
452 The key thing about SimplGently is that it does no call-site inlining.
453 Before full laziness we must be careful not to inline wrappers,
454 because doing so inhibits floating
455 e.g. ...(case f x of ...)...
456 ==> ...(case (case x of I# x# -> fw x#) of ...)...
457 ==> ...(case x of I# x# -> case fw x# of ...)...
458 and now the redex (f x) isn't floatable any more.
460 The no-inlining thing is also important for Template Haskell. You might be
461 compiling in one-shot mode with -O2; but when TH compiles a splice before
462 running it, we don't want to use -O2. Indeed, we don't want to inline
463 anything, because the byte-code interpreter might get confused about
464 unboxed tuples and suchlike.
468 SimplGently is also used as the mode to simplify inside an InlineMe note.
471 inlineMode :: SimplifierMode
472 inlineMode = SimplGently
475 It really is important to switch off inlinings inside such
476 expressions. Consider the following example
482 in ...g...g...g...g...g...
484 Now, if that's the ONLY occurrence of f, it will be inlined inside g,
485 and thence copied multiple times when g is inlined.
488 This function may be inlinined in other modules, so we
489 don't want to remove (by inlining) calls to functions that have
490 specialisations, or that may have transformation rules in an importing
493 E.g. {-# INLINE f #-}
496 and suppose that g is strict *and* has specialisations. If we inline
497 g's wrapper, we deny f the chance of getting the specialised version
498 of g when f is inlined at some call site (perhaps in some other
501 It's also important not to inline a worker back into a wrapper.
503 wraper = inline_me (\x -> ...worker... )
504 Normally, the inline_me prevents the worker getting inlined into
505 the wrapper (initially, the worker's only call site!). But,
506 if the wrapper is sure to be called, the strictness analyser will
507 mark it 'demanded', so when the RHS is simplified, it'll get an ArgOf
508 continuation. That's why the keep_inline predicate returns True for
509 ArgOf continuations. It shouldn't do any harm not to dissolve the
510 inline-me note under these circumstances.
512 Note that the result is that we do very little simplification
515 all xs = foldr (&&) True xs
516 any p = all . map p {-# INLINE any #-}
518 Problem: any won't get deforested, and so if it's exported and the
519 importer doesn't use the inlining, (eg passes it as an arg) then we
520 won't get deforestation at all. We havn't solved this problem yet!
523 preInlineUnconditionally
524 ~~~~~~~~~~~~~~~~~~~~~~~~
525 @preInlineUnconditionally@ examines a bndr to see if it is used just
526 once in a completely safe way, so that it is safe to discard the
527 binding inline its RHS at the (unique) usage site, REGARDLESS of how
528 big the RHS might be. If this is the case we don't simplify the RHS
529 first, but just inline it un-simplified.
531 This is much better than first simplifying a perhaps-huge RHS and then
532 inlining and re-simplifying it. Indeed, it can be at least quadratically
541 We may end up simplifying e1 N times, e2 N-1 times, e3 N-3 times etc.
542 This can happen with cascades of functions too:
549 THE MAIN INVARIANT is this:
551 ---- preInlineUnconditionally invariant -----
552 IF preInlineUnconditionally chooses to inline x = <rhs>
553 THEN doing the inlining should not change the occurrence
554 info for the free vars of <rhs>
555 ----------------------------------------------
557 For example, it's tempting to look at trivial binding like
559 and inline it unconditionally. But suppose x is used many times,
560 but this is the unique occurrence of y. Then inlining x would change
561 y's occurrence info, which breaks the invariant. It matters: y
562 might have a BIG rhs, which will now be dup'd at every occurrenc of x.
565 Evne RHSs labelled InlineMe aren't caught here, because there might be
566 no benefit from inlining at the call site.
568 [Sept 01] Don't unconditionally inline a top-level thing, because that
569 can simply make a static thing into something built dynamically. E.g.
573 [Remember that we treat \s as a one-shot lambda.] No point in
574 inlining x unless there is something interesting about the call site.
576 But watch out: if you aren't careful, some useful foldr/build fusion
577 can be lost (most notably in spectral/hartel/parstof) because the
578 foldr didn't see the build. Doing the dynamic allocation isn't a big
579 deal, in fact, but losing the fusion can be. But the right thing here
580 seems to be to do a callSiteInline based on the fact that there is
581 something interesting about the call site (it's strict). Hmm. That
584 Conclusion: inline top level things gaily until Phase 0 (the last
585 phase), at which point don't.
588 preInlineUnconditionally :: SimplEnv -> TopLevelFlag -> InId -> InExpr -> Bool
589 preInlineUnconditionally env top_lvl bndr rhs
591 | opt_SimplNoPreInlining = False
592 | otherwise = case idOccInfo bndr of
593 IAmDead -> True -- Happens in ((\x.1) v)
594 OneOcc in_lam True int_cxt -> try_once in_lam int_cxt
598 active = case phase of
599 SimplGently -> isAlwaysActive prag
600 SimplPhase n -> isActive n prag
601 prag = idInlinePragma bndr
603 try_once in_lam int_cxt -- There's one textual occurrence
604 | not in_lam = isNotTopLevel top_lvl || early_phase
605 | otherwise = int_cxt && canInlineInLam rhs
607 -- Be very careful before inlining inside a lambda, becuase (a) we must not
608 -- invalidate occurrence information, and (b) we want to avoid pushing a
609 -- single allocation (here) into multiple allocations (inside lambda).
610 -- Inlining a *function* with a single *saturated* call would be ok, mind you.
611 -- || (if is_cheap && not (canInlineInLam rhs) then pprTrace "preinline" (ppr bndr <+> ppr rhs) ok else ok)
613 -- is_cheap = exprIsCheap rhs
614 -- ok = is_cheap && int_cxt
616 -- int_cxt The context isn't totally boring
617 -- E.g. let f = \ab.BIG in \y. map f xs
618 -- Don't want to substitute for f, because then we allocate
619 -- its closure every time the \y is called
620 -- But: let f = \ab.BIG in \y. map (f y) xs
621 -- Now we do want to substitute for f, even though it's not
622 -- saturated, because we're going to allocate a closure for
623 -- (f y) every time round the loop anyhow.
625 -- canInlineInLam => free vars of rhs are (Once in_lam) or Many,
626 -- so substituting rhs inside a lambda doesn't change the occ info.
627 -- Sadly, not quite the same as exprIsHNF.
628 canInlineInLam (Lit l) = True
629 canInlineInLam (Lam b e) = isRuntimeVar b || canInlineInLam e
630 canInlineInLam (Note _ e) = canInlineInLam e
631 canInlineInLam _ = False
633 early_phase = case phase of
634 SimplPhase 0 -> False
636 -- If we don't have this early_phase test, consider
637 -- x = length [1,2,3]
638 -- The full laziness pass carefully floats all the cons cells to
639 -- top level, and preInlineUnconditionally floats them all back in.
640 -- Result is (a) static allocation replaced by dynamic allocation
641 -- (b) many simplifier iterations because this tickles
642 -- a related problem; only one inlining per pass
644 -- On the other hand, I have seen cases where top-level fusion is
645 -- lost if we don't inline top level thing (e.g. string constants)
646 -- Hence the test for phase zero (which is the phase for all the final
647 -- simplifications). Until phase zero we take no special notice of
648 -- top level things, but then we become more leery about inlining
653 postInlineUnconditionally
654 ~~~~~~~~~~~~~~~~~~~~~~~~~
655 @postInlineUnconditionally@ decides whether to unconditionally inline
656 a thing based on the form of its RHS; in particular if it has a
657 trivial RHS. If so, we can inline and discard the binding altogether.
659 NB: a loop breaker has must_keep_binding = True and non-loop-breakers
660 only have *forward* references Hence, it's safe to discard the binding
662 NOTE: This isn't our last opportunity to inline. We're at the binding
663 site right now, and we'll get another opportunity when we get to the
666 Note that we do this unconditional inlining only for trival RHSs.
667 Don't inline even WHNFs inside lambdas; doing so may simply increase
668 allocation when the function is called. This isn't the last chance; see
671 NB: Even inline pragmas (e.g. IMustBeINLINEd) are ignored here Why?
672 Because we don't even want to inline them into the RHS of constructor
673 arguments. See NOTE above
675 NB: At one time even NOINLINE was ignored here: if the rhs is trivial
676 it's best to inline it anyway. We often get a=E; b=a from desugaring,
677 with both a and b marked NOINLINE. But that seems incompatible with
678 our new view that inlining is like a RULE, so I'm sticking to the 'active'
682 postInlineUnconditionally
683 :: SimplEnv -> TopLevelFlag
684 -> InId -- The binder (an OutId would be fine too)
685 -> OccInfo -- From the InId
689 postInlineUnconditionally env top_lvl bndr occ_info rhs unfolding
691 | isLoopBreaker occ_info = False -- If it's a loop-breaker of any kind, dont' inline
692 -- because it might be referred to "earlier"
693 | isExportedId bndr = False
694 | exprIsTrivial rhs = True
697 -- The point of examining occ_info here is that for *non-values*
698 -- that occur outside a lambda, the call-site inliner won't have
699 -- a chance (becuase it doesn't know that the thing
700 -- only occurs once). The pre-inliner won't have gotten
701 -- it either, if the thing occurs in more than one branch
702 -- So the main target is things like
705 -- True -> case x of ...
706 -- False -> case x of ...
707 -- I'm not sure how important this is in practice
708 OneOcc in_lam one_br int_cxt -- OneOcc => no code-duplication issue
709 -> smallEnoughToInline unfolding -- Small enough to dup
710 -- ToDo: consider discount on smallEnoughToInline if int_cxt is true
712 -- NB: Do NOT inline arbitrarily big things, even if one_br is True
713 -- Reason: doing so risks exponential behaviour. We simplify a big
714 -- expression, inline it, and simplify it again. But if the
715 -- very same thing happens in the big expression, we get
717 -- PRINCIPLE: when we've already simplified an expression once,
718 -- make sure that we only inline it if it's reasonably small.
720 && ((isNotTopLevel top_lvl && not in_lam) ||
721 -- But outside a lambda, we want to be reasonably aggressive
722 -- about inlining into multiple branches of case
723 -- e.g. let x = <non-value>
724 -- in case y of { C1 -> ..x..; C2 -> ..x..; C3 -> ... }
725 -- Inlining can be a big win if C3 is the hot-spot, even if
726 -- the uses in C1, C2 are not 'interesting'
727 -- An example that gets worse if you add int_cxt here is 'clausify'
729 (isCheapUnfolding unfolding && int_cxt))
730 -- isCheap => acceptable work duplication; in_lam may be true
731 -- int_cxt to prevent us inlining inside a lambda without some
732 -- good reason. See the notes on int_cxt in preInlineUnconditionally
734 IAmDead -> True -- This happens; for example, the case_bndr during case of
735 -- known constructor: case (a,b) of x { (p,q) -> ... }
736 -- Here x isn't mentioned in the RHS, so we don't want to
737 -- create the (dead) let-binding let x = (a,b) in ...
741 -- Here's an example that we don't handle well:
742 -- let f = if b then Left (\x.BIG) else Right (\y.BIG)
743 -- in \y. ....case f of {...} ....
744 -- Here f is used just once, and duplicating the case work is fine (exprIsCheap).
746 -- * We can't preInlineUnconditionally because that woud invalidate
747 -- the occ info for b.
748 -- * We can't postInlineUnconditionally because the RHS is big, and
749 -- that risks exponential behaviour
750 -- * We can't call-site inline, because the rhs is big
754 active = case getMode env of
755 SimplGently -> isAlwaysActive prag
756 SimplPhase n -> isActive n prag
757 prag = idInlinePragma bndr
759 activeInline :: SimplEnv -> OutId -> Bool
761 = case getMode env of
763 -- No inlining at all when doing gentle stuff,
764 -- except for local things that occur once
765 -- The reason is that too little clean-up happens if you
766 -- don't inline use-once things. Also a bit of inlining is *good* for
767 -- full laziness; it can expose constant sub-expressions.
768 -- Example in spectral/mandel/Mandel.hs, where the mandelset
769 -- function gets a useful let-float if you inline windowToViewport
771 -- NB: we used to have a second exception, for data con wrappers.
772 -- On the grounds that we use gentle mode for rule LHSs, and
773 -- they match better when data con wrappers are inlined.
774 -- But that only really applies to the trivial wrappers (like (:)),
775 -- and they are now constructed as Compulsory unfoldings (in MkId)
776 -- so they'll happen anyway.
778 SimplPhase n -> isActive n prag
780 prag = idInlinePragma id
782 activeRule :: DynFlags -> SimplEnv -> Maybe (Activation -> Bool)
783 -- Nothing => No rules at all
784 activeRule dflags env
785 | not (dopt Opt_RewriteRules dflags)
786 = Nothing -- Rewriting is off
788 = case getMode env of
789 SimplGently -> Just isAlwaysActive
790 -- Used to be Nothing (no rules in gentle mode)
791 -- Main motivation for changing is that I wanted
792 -- lift String ===> ...
793 -- to work in Template Haskell when simplifying
794 -- splices, so we get simpler code for literal strings
795 SimplPhase n -> Just (isActive n)
799 %************************************************************************
803 %************************************************************************
806 mkLam :: [OutBndr] -> OutExpr -> SimplM OutExpr
807 -- mkLam tries three things
808 -- a) eta reduction, if that gives a trivial expression
809 -- b) eta expansion [only if there are some value lambdas]
814 = do { dflags <- getDOptsSmpl
815 ; mkLam' dflags bndrs body }
817 mkLam' :: DynFlags -> [OutBndr] -> OutExpr -> SimplM OutExpr
818 mkLam' dflags bndrs (Cast body@(Lam _ _) co)
819 -- Note [Casts and lambdas]
820 = do { lam <- mkLam' dflags (bndrs ++ bndrs') body'
821 ; return (mkCoerce (mkPiTypes bndrs co) lam) }
823 (bndrs',body') = collectBinders body
825 mkLam' dflags bndrs body
826 | dopt Opt_DoEtaReduction dflags,
827 Just etad_lam <- tryEtaReduce bndrs body
828 = do { tick (EtaReduction (head bndrs))
831 | dopt Opt_DoLambdaEtaExpansion dflags,
832 any isRuntimeVar bndrs
833 = do { body' <- tryEtaExpansion dflags body
834 ; return (mkLams bndrs body') }
837 = returnSmpl (mkLams bndrs body)
840 Note [Casts and lambdas]
841 ~~~~~~~~~~~~~~~~~~~~~~~~
843 (\x. (\y. e) `cast` g1) `cast` g2
844 There is a danger here that the two lambdas look separated, and the
845 full laziness pass might float an expression to between the two.
847 So this equation in mkLam' floats the g1 out, thus:
848 (\x. e `cast` g1) --> (\x.e) `cast` (tx -> g1)
851 In general, this floats casts outside lambdas, where (I hope) they might meet
852 and cancel with some other cast.
855 -- c) floating lets out through big lambdas
856 -- [only if all tyvar lambdas, and only if this lambda
857 -- is the RHS of a let]
859 {- Sept 01: I'm experimenting with getting the
860 full laziness pass to float out past big lambdsa
861 | all isTyVar bndrs, -- Only for big lambdas
862 contIsRhs cont -- Only try the rhs type-lambda floating
863 -- if this is indeed a right-hand side; otherwise
864 -- we end up floating the thing out, only for float-in
865 -- to float it right back in again!
866 = tryRhsTyLam env bndrs body `thenSmpl` \ (floats, body') ->
867 returnSmpl (floats, mkLams bndrs body')
871 %************************************************************************
873 \subsection{Eta expansion and reduction}
875 %************************************************************************
877 We try for eta reduction here, but *only* if we get all the
878 way to an exprIsTrivial expression.
879 We don't want to remove extra lambdas unless we are going
880 to avoid allocating this thing altogether
883 tryEtaReduce :: [OutBndr] -> OutExpr -> Maybe OutExpr
884 tryEtaReduce bndrs body
885 -- We don't use CoreUtils.etaReduce, because we can be more
887 -- (a) we already have the binders
888 -- (b) we can do the triviality test before computing the free vars
889 = go (reverse bndrs) body
891 go (b : bs) (App fun arg) | ok_arg b arg = go bs fun -- Loop round
892 go [] fun | ok_fun fun = Just fun -- Success!
893 go _ _ = Nothing -- Failure!
895 ok_fun fun = exprIsTrivial fun
896 && not (any (`elemVarSet` (exprFreeVars fun)) bndrs)
897 && (exprIsHNF fun || all ok_lam bndrs)
898 ok_lam v = isTyVar v || isDictId v
899 -- The exprIsHNF is because eta reduction is not
900 -- valid in general: \x. bot /= bot
901 -- So we need to be sure that the "fun" is a value.
903 -- However, we always want to reduce (/\a -> f a) to f
904 -- This came up in a RULE: foldr (build (/\a -> g a))
905 -- did not match foldr (build (/\b -> ...something complex...))
906 -- The type checker can insert these eta-expanded versions,
907 -- with both type and dictionary lambdas; hence the slightly
910 ok_arg b arg = varToCoreExpr b `cheapEqExpr` arg
914 Try eta expansion for RHSs
917 f = \x1..xn -> N ==> f = \x1..xn y1..ym -> N y1..ym
920 where (in both cases)
922 * The xi can include type variables
924 * The yi are all value variables
926 * N is a NORMAL FORM (i.e. no redexes anywhere)
927 wanting a suitable number of extra args.
929 We may have to sandwich some coerces between the lambdas
930 to make the types work. exprEtaExpandArity looks through coerces
931 when computing arity; and etaExpand adds the coerces as necessary when
932 actually computing the expansion.
935 tryEtaExpansion :: DynFlags -> OutExpr -> SimplM OutExpr
936 -- There is at least one runtime binder in the binders
937 tryEtaExpansion dflags body
938 = getUniquesSmpl `thenSmpl` \ us ->
939 returnSmpl (etaExpand fun_arity us body (exprType body))
941 fun_arity = exprEtaExpandArity dflags body
945 %************************************************************************
947 \subsection{Floating lets out of big lambdas}
949 %************************************************************************
951 Note [Floating and type abstraction]
952 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
955 We'd like to float this to
958 x = /\a. C (y1 a) (y2 a)
959 for the usual reasons: we want to inline x rather vigorously.
961 You may think that this kind of thing is rare. But in some programs it is
962 common. For example, if you do closure conversion you might get:
964 data a :-> b = forall e. (e -> a -> b) :$ e
966 f_cc :: forall a. a :-> a
967 f_cc = /\a. (\e. id a) :$ ()
969 Now we really want to inline that f_cc thing so that the
970 construction of the closure goes away.
972 So I have elaborated simplLazyBind to understand right-hand sides that look
976 and treat them specially. The real work is done in SimplUtils.abstractFloats,
977 but there is quite a bit of plumbing in simplLazyBind as well.
979 The same transformation is good when there are lets in the body:
981 /\abc -> let(rec) x = e in b
983 let(rec) x' = /\abc -> let x = x' a b c in e
985 /\abc -> let x = x' a b c in b
987 This is good because it can turn things like:
989 let f = /\a -> letrec g = ... g ... in g
991 letrec g' = /\a -> ... g' a ...
995 which is better. In effect, it means that big lambdas don't impede
998 This optimisation is CRUCIAL in eliminating the junk introduced by
999 desugaring mutually recursive definitions. Don't eliminate it lightly!
1001 [May 1999] If we do this transformation *regardless* then we can
1002 end up with some pretty silly stuff. For example,
1005 st = /\ s -> let { x1=r1 ; x2=r2 } in ...
1010 st = /\s -> ...[y1 s/x1, y2 s/x2]
1013 Unless the "..." is a WHNF there is really no point in doing this.
1014 Indeed it can make things worse. Suppose x1 is used strictly,
1017 x1* = case f y of { (a,b) -> e }
1019 If we abstract this wrt the tyvar we then can't do the case inline
1020 as we would normally do.
1022 That's why the whole transformation is part of the same process that
1023 floats let-bindings and constructor arguments out of RHSs. In particular,
1024 it is guarded by the doFloatFromRhs call in simplLazyBind.
1028 abstractFloats :: [OutTyVar] -> SimplEnv -> OutExpr -> SimplM ([OutBind], OutExpr)
1029 abstractFloats main_tvs body_env body
1030 = ASSERT( notNull body_floats )
1031 do { (subst, float_binds) <- mapAccumLSmpl abstract empty_subst body_floats
1032 ; return (float_binds, CoreSubst.substExpr subst body) }
1034 main_tv_set = mkVarSet main_tvs
1035 body_floats = getFloats body_env
1036 empty_subst = CoreSubst.mkEmptySubst (seInScope body_env)
1038 abstract :: CoreSubst.Subst -> OutBind -> SimplM (CoreSubst.Subst, OutBind)
1039 abstract subst (NonRec id rhs)
1040 = do { (poly_id, poly_app) <- mk_poly tvs_here id
1041 ; let poly_rhs = mkLams tvs_here rhs'
1042 subst' = CoreSubst.extendIdSubst subst id poly_app
1043 ; return (subst', (NonRec poly_id poly_rhs)) }
1045 rhs' = CoreSubst.substExpr subst rhs
1046 tvs_here | any isCoVar main_tvs = main_tvs -- Note [Abstract over coercions]
1048 = varSetElems (main_tv_set `intersectVarSet` exprSomeFreeVars isTyVar rhs')
1050 -- Abstract only over the type variables free in the rhs
1051 -- wrt which the new binding is abstracted. But the naive
1052 -- approach of abstract wrt the tyvars free in the Id's type
1054 -- /\ a b -> let t :: (a,b) = (e1, e2)
1057 -- Here, b isn't free in x's type, but we must nevertheless
1058 -- abstract wrt b as well, because t's type mentions b.
1059 -- Since t is floated too, we'd end up with the bogus:
1060 -- poly_t = /\ a b -> (e1, e2)
1061 -- poly_x = /\ a -> fst (poly_t a *b*)
1062 -- So for now we adopt the even more naive approach of
1063 -- abstracting wrt *all* the tyvars. We'll see if that
1064 -- gives rise to problems. SLPJ June 98
1066 abstract subst (Rec prs)
1067 = do { (poly_ids, poly_apps) <- mapAndUnzipSmpl (mk_poly tvs_here) ids
1068 ; let subst' = CoreSubst.extendSubstList subst (ids `zip` poly_apps)
1069 poly_rhss = [mkLams tvs_here (CoreSubst.substExpr subst' rhs) | rhs <- rhss]
1070 ; return (subst', Rec (poly_ids `zip` poly_rhss)) }
1072 (ids,rhss) = unzip prs
1073 -- For a recursive group, it's a bit of a pain to work out the minimal
1074 -- set of tyvars over which to abstract:
1075 -- /\ a b c. let x = ...a... in
1076 -- letrec { p = ...x...q...
1077 -- q = .....p...b... } in
1079 -- Since 'x' is abstracted over 'a', the {p,q} group must be abstracted
1080 -- over 'a' (because x is replaced by (poly_x a)) as well as 'b'.
1081 -- Since it's a pain, we just use the whole set, which is always safe
1083 -- If you ever want to be more selective, remember this bizarre case too:
1085 -- Here, we must abstract 'x' over 'a'.
1088 mk_poly tvs_here var
1089 = do { uniq <- getUniqueSmpl
1090 ; let poly_name = setNameUnique (idName var) uniq -- Keep same name
1091 poly_ty = mkForAllTys tvs_here (idType var) -- But new type of course
1092 poly_id = mkLocalId poly_name poly_ty
1093 ; return (poly_id, mkTyApps (Var poly_id) (mkTyVarTys tvs_here)) }
1094 -- In the olden days, it was crucial to copy the occInfo of the original var,
1095 -- because we were looking at occurrence-analysed but as yet unsimplified code!
1096 -- In particular, we mustn't lose the loop breakers. BUT NOW we are looking
1097 -- at already simplified code, so it doesn't matter
1099 -- It's even right to retain single-occurrence or dead-var info:
1100 -- Suppose we started with /\a -> let x = E in B
1101 -- where x occurs once in B. Then we transform to:
1102 -- let x' = /\a -> E in /\a -> let x* = x' a in B
1103 -- where x* has an INLINE prag on it. Now, once x* is inlined,
1104 -- the occurrences of x' will be just the occurrences originally
1108 Note [Abstract over coercions]
1109 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
1110 If a coercion variable (g :: a ~ Int) is free in the RHS, then so is the
1111 type variable a. Rather than sort this mess out, we simply bale out and abstract
1112 wrt all the type variables if any of them are coercion variables.
1115 Historical note: if you use let-bindings instead of a substitution, beware of this:
1117 -- Suppose we start with:
1119 -- x = /\ a -> let g = G in E
1121 -- Then we'll float to get
1123 -- x = let poly_g = /\ a -> G
1124 -- in /\ a -> let g = poly_g a in E
1126 -- But now the occurrence analyser will see just one occurrence
1127 -- of poly_g, not inside a lambda, so the simplifier will
1128 -- PreInlineUnconditionally poly_g back into g! Badk to square 1!
1129 -- (I used to think that the "don't inline lone occurrences" stuff
1130 -- would stop this happening, but since it's the *only* occurrence,
1131 -- PreInlineUnconditionally kicks in first!)
1133 -- Solution: put an INLINE note on g's RHS, so that poly_g seems
1134 -- to appear many times. (NB: mkInlineMe eliminates
1135 -- such notes on trivial RHSs, so do it manually.)
1137 %************************************************************************
1141 %************************************************************************
1143 prepareAlts tries these things:
1145 1. If several alternatives are identical, merge them into
1146 a single DEFAULT alternative. I've occasionally seen this
1147 making a big difference:
1149 case e of =====> case e of
1150 C _ -> f x D v -> ....v....
1151 D v -> ....v.... DEFAULT -> f x
1154 The point is that we merge common RHSs, at least for the DEFAULT case.
1155 [One could do something more elaborate but I've never seen it needed.]
1156 To avoid an expensive test, we just merge branches equal to the *first*
1157 alternative; this picks up the common cases
1158 a) all branches equal
1159 b) some branches equal to the DEFAULT (which occurs first)
1162 case e of b { ==> case e of b {
1163 p1 -> rhs1 p1 -> rhs1
1165 pm -> rhsm pm -> rhsm
1166 _ -> case b of b' { pn -> let b'=b in rhsn
1168 ... po -> let b'=b in rhso
1169 po -> rhso _ -> let b'=b in rhsd
1173 which merges two cases in one case when -- the default alternative of
1174 the outer case scrutises the same variable as the outer case This
1175 transformation is called Case Merging. It avoids that the same
1176 variable is scrutinised multiple times.
1179 The case where transformation (1) showed up was like this (lib/std/PrelCError.lhs):
1185 where @is@ was something like
1187 p `is` n = p /= (-1) && p == n
1189 This gave rise to a horrible sequence of cases
1196 and similarly in cascade for all the join points!
1199 ~~~~~~~~~~~~~~~~~~~~
1200 We do this *here*, looking at un-simplified alternatives, because we
1201 have to check that r doesn't mention the variables bound by the
1202 pattern in each alternative, so the binder-info is rather useful.
1205 prepareAlts :: OutExpr -> OutId -> [InAlt] -> SimplM ([AltCon], [InAlt])
1206 prepareAlts scrut case_bndr' alts
1207 = do { dflags <- getDOptsSmpl
1208 ; alts <- combineIdenticalAlts case_bndr' alts
1210 ; let (alts_wo_default, maybe_deflt) = findDefault alts
1211 alt_cons = [con | (con,_,_) <- alts_wo_default]
1212 imposs_deflt_cons = nub (imposs_cons ++ alt_cons)
1213 -- "imposs_deflt_cons" are handled
1214 -- EITHER by the context,
1215 -- OR by a non-DEFAULT branch in this case expression.
1217 ; default_alts <- prepareDefault dflags scrut case_bndr' mb_tc_app
1218 imposs_deflt_cons maybe_deflt
1220 ; let trimmed_alts = filterOut impossible_alt alts_wo_default
1221 merged_alts = mergeAlts trimmed_alts default_alts
1222 -- We need the mergeAlts in case the new default_alt
1223 -- has turned into a constructor alternative.
1224 -- The merge keeps the inner DEFAULT at the front, if there is one
1225 -- and interleaves the alternatives in the right order
1227 ; return (imposs_deflt_cons, merged_alts) }
1229 mb_tc_app = splitTyConApp_maybe (idType case_bndr')
1230 Just (_, inst_tys) = mb_tc_app
1232 imposs_cons = case scrut of
1233 Var v -> otherCons (idUnfolding v)
1236 impossible_alt :: CoreAlt -> Bool
1237 impossible_alt (con, _, _) | con `elem` imposs_cons = True
1238 impossible_alt (DataAlt con, _, _) = dataConCannotMatch inst_tys con
1239 impossible_alt alt = False
1242 --------------------------------------------------
1243 -- 1. Merge identical branches
1244 --------------------------------------------------
1245 combineIdenticalAlts :: OutId -> [InAlt] -> SimplM [InAlt]
1247 combineIdenticalAlts case_bndr alts@((con1,bndrs1,rhs1) : con_alts)
1248 | all isDeadBinder bndrs1, -- Remember the default
1249 length filtered_alts < length con_alts -- alternative comes first
1250 -- Also Note [Dead binders]
1251 = do { tick (AltMerge case_bndr)
1252 ; return ((DEFAULT, [], rhs1) : filtered_alts) }
1254 filtered_alts = filter keep con_alts
1255 keep (con,bndrs,rhs) = not (all isDeadBinder bndrs && rhs `cheapEqExpr` rhs1)
1257 combineIdenticalAlts case_bndr alts = return alts
1259 -------------------------------------------------------------------------
1260 -- Prepare the default alternative
1261 -------------------------------------------------------------------------
1262 prepareDefault :: DynFlags
1263 -> OutExpr -- Scrutinee
1264 -> OutId -- Case binder; need just for its type. Note that as an
1265 -- OutId, it has maximum information; this is important.
1266 -- Test simpl013 is an example
1267 -> Maybe (TyCon, [Type]) -- Type of scrutinee, decomposed
1268 -> [AltCon] -- These cons can't happen when matching the default
1269 -> Maybe InExpr -- Rhs
1270 -> SimplM [InAlt] -- Still unsimplified
1271 -- We use a list because it's what mergeAlts expects,
1272 -- And becuase case-merging can cause many to show up
1274 ------- Merge nested cases ----------
1275 prepareDefault dflags scrut outer_bndr bndr_ty imposs_cons (Just deflt_rhs)
1276 | dopt Opt_CaseMerge dflags
1277 , Case (Var scrut_var) inner_bndr _ inner_alts <- deflt_rhs
1278 , scruting_same_var scrut_var
1279 = do { tick (CaseMerge outer_bndr)
1281 ; let munge_rhs rhs = bindCaseBndr inner_bndr (Var outer_bndr) rhs
1282 ; return [(con, args, munge_rhs rhs) | (con, args, rhs) <- inner_alts,
1283 not (con `elem` imposs_cons) ]
1284 -- NB: filter out any imposs_cons. Example:
1287 -- DEFAULT -> case x of
1290 -- When we merge, we must ensure that e1 takes
1291 -- precedence over e2 as the value for A!
1293 -- Warning: don't call prepareAlts recursively!
1294 -- Firstly, there's no point, because inner alts have already had
1295 -- mkCase applied to them, so they won't have a case in their default
1296 -- Secondly, if you do, you get an infinite loop, because the bindCaseBndr
1297 -- in munge_rhs may put a case into the DEFAULT branch!
1299 -- We are scrutinising the same variable if it's
1300 -- the outer case-binder, or if the outer case scrutinises a variable
1301 -- (and it's the same). Testing both allows us not to replace the
1302 -- outer scrut-var with the outer case-binder (Simplify.simplCaseBinder).
1303 scruting_same_var = case scrut of
1304 Var outer_scrut -> \ v -> v == outer_bndr || v == outer_scrut
1305 other -> \ v -> v == outer_bndr
1307 --------- Fill in known constructor -----------
1308 prepareDefault dflags scrut case_bndr (Just (tycon, inst_tys)) imposs_cons (Just deflt_rhs)
1309 | -- This branch handles the case where we are
1310 -- scrutinisng an algebraic data type
1311 isAlgTyCon tycon -- It's a data type, tuple, or unboxed tuples.
1312 , not (isNewTyCon tycon) -- We can have a newtype, if we are just doing an eval:
1313 -- case x of { DEFAULT -> e }
1314 -- and we don't want to fill in a default for them!
1315 , Just all_cons <- tyConDataCons_maybe tycon
1316 , not (null all_cons) -- This is a tricky corner case. If the data type has no constructors,
1317 -- which GHC allows, then the case expression will have at most a default
1318 -- alternative. We don't want to eliminate that alternative, because the
1319 -- invariant is that there's always one alternative. It's more convenient
1321 -- case x of { DEFAULT -> e }
1322 -- as it is, rather than transform it to
1323 -- error "case cant match"
1324 -- which would be quite legitmate. But it's a really obscure corner, and
1325 -- not worth wasting code on.
1326 , let imposs_data_cons = [con | DataAlt con <- imposs_cons] -- We now know it's a data type
1327 impossible con = con `elem` imposs_data_cons || dataConCannotMatch inst_tys con
1328 = case filterOut impossible all_cons of
1329 [] -> return [] -- Eliminate the default alternative
1330 -- altogether if it can't match
1332 [con] -> -- It matches exactly one constructor, so fill it in
1333 do { tick (FillInCaseDefault case_bndr)
1334 ; us <- getUniquesSmpl
1335 ; let (ex_tvs, co_tvs, arg_ids) =
1336 dataConRepInstPat us con inst_tys
1337 ; return [(DataAlt con, ex_tvs ++ co_tvs ++ arg_ids, deflt_rhs)] }
1339 two_or_more -> return [(DEFAULT, [], deflt_rhs)]
1341 --------- Catch-all cases -----------
1342 prepareDefault dflags scrut case_bndr bndr_ty imposs_cons (Just deflt_rhs)
1343 = return [(DEFAULT, [], deflt_rhs)]
1345 prepareDefault dflags scrut case_bndr bndr_ty imposs_cons Nothing
1346 = return [] -- No default branch
1351 =================================================================================
1353 mkCase tries these things
1355 1. Eliminate the case altogether if possible
1363 and similar friends.
1367 mkCase :: OutExpr -> OutId -> OutType
1368 -> [OutAlt] -- Increasing order
1371 --------------------------------------------------
1372 -- 1. Check for empty alternatives
1373 --------------------------------------------------
1375 -- This isn't strictly an error. It's possible that the simplifer might "see"
1376 -- that an inner case has no accessible alternatives before it "sees" that the
1377 -- entire branch of an outer case is inaccessible. So we simply
1378 -- put an error case here insteadd
1379 mkCase scrut case_bndr ty []
1380 = pprTrace "mkCase: null alts" (ppr case_bndr <+> ppr scrut) $
1381 return (mkApps (Var rUNTIME_ERROR_ID)
1382 [Type ty, Lit (mkStringLit "Impossible alternative")])
1385 --------------------------------------------------
1387 --------------------------------------------------
1389 mkCase scrut case_bndr ty alts -- Identity case
1390 | all identity_alt alts
1391 = tick (CaseIdentity case_bndr) `thenSmpl_`
1392 returnSmpl (re_cast scrut)
1394 identity_alt (con, args, rhs) = check_eq con args (de_cast rhs)
1396 check_eq DEFAULT _ (Var v) = v == case_bndr
1397 check_eq (LitAlt lit') _ (Lit lit) = lit == lit'
1398 check_eq (DataAlt con) args rhs = rhs `cheapEqExpr` mkConApp con (arg_tys ++ varsToCoreExprs args)
1399 || rhs `cheapEqExpr` Var case_bndr
1400 check_eq con args rhs = False
1402 arg_tys = map Type (tyConAppArgs (idType case_bndr))
1405 -- case e of x { _ -> x `cast` c }
1406 -- And we definitely want to eliminate this case, to give
1408 -- So we throw away the cast from the RHS, and reconstruct
1409 -- it at the other end. All the RHS casts must be the same
1410 -- if (all identity_alt alts) holds.
1412 -- Don't worry about nested casts, because the simplifier combines them
1413 de_cast (Cast e _) = e
1416 re_cast scrut = case head alts of
1417 (_,_,Cast _ co) -> Cast scrut co
1422 --------------------------------------------------
1424 --------------------------------------------------
1425 mkCase scrut bndr ty alts = returnSmpl (Case scrut bndr ty alts)
1429 When adding auxiliary bindings for the case binder, it's worth checking if
1430 its dead, because it often is, and occasionally these mkCase transformations
1431 cascade rather nicely.
1434 bindCaseBndr bndr rhs body
1435 | isDeadBinder bndr = body
1436 | otherwise = bindNonRec bndr rhs body