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(..), ArgInfo(..),
17 contIsDupable, contResultType, contIsTrivial, contArgs, dropArgs,
18 countValArgs, countArgs, splitInlineCont,
19 mkBoringStop, mkLazyArgStop, contIsRhsOrArg,
20 interestingCallContext, interestingArgContext,
22 interestingArg, mkArgInfo,
27 #include "HsVersions.h"
33 import qualified CoreSubst
37 import CoreArity ( etaExpand, exprEtaExpandArity )
41 import Var ( isCoVar )
44 import Type hiding( substTy )
45 import Coercion ( coercionKind )
47 import Unify ( dataConCannotMatch )
59 %************************************************************************
63 %************************************************************************
65 A SimplCont allows the simplifier to traverse the expression in a
66 zipper-like fashion. The SimplCont represents the rest of the expression,
67 "above" the point of interest.
69 You can also think of a SimplCont as an "evaluation context", using
70 that term in the way it is used for operational semantics. This is the
71 way I usually think of it, For example you'll often see a syntax for
72 evaluation context looking like
73 C ::= [] | C e | case C of alts | C `cast` co
74 That's the kind of thing we are doing here, and I use that syntax in
79 * A SimplCont describes a *strict* context (just like
80 evaluation contexts do). E.g. Just [] is not a SimplCont
82 * A SimplCont describes a context that *does not* bind
83 any variables. E.g. \x. [] is not a SimplCont
87 = Stop -- An empty context, or hole, []
88 CallCtxt -- True <=> There is something interesting about
89 -- the context, and hence the inliner
90 -- should be a bit keener (see interestingCallContext)
92 -- This is an argument of a function that has RULES
93 -- Inlining the call might allow the rule to fire
95 | CoerceIt -- C `cast` co
96 OutCoercion -- The coercion simplified
101 InExpr SimplEnv -- The argument and its static env
104 | Select -- case C of alts
106 InId [InAlt] SimplEnv -- The case binder, alts, and subst-env
109 -- The two strict forms have no DupFlag, because we never duplicate them
110 | StrictBind -- (\x* \xs. e) C
111 InId [InBndr] -- let x* = [] in e
112 InExpr SimplEnv -- is a special case
116 OutExpr -- e; *always* of form (Var v `App1` e1 .. `App` en)
117 CallCtxt -- Whether *this* argument position is interesting
118 ArgInfo -- Whether the function at the head of e has rules, etc
119 SimplCont -- plus strictness flags for *further* args
123 ai_rules :: Bool, -- Function has rules (recursively)
124 -- => be keener to inline in all args
125 ai_strs :: [Bool], -- Strictness of arguments
126 -- Usually infinite, but if it is finite it guarantees
127 -- that the function diverges after being given
128 -- that number of args
129 ai_discs :: [Int] -- Discounts for arguments; non-zero => be keener to inline
133 instance Outputable SimplCont where
134 ppr (Stop interesting) = ptext (sLit "Stop") <> brackets (ppr interesting)
135 ppr (ApplyTo dup arg _ cont) = ((ptext (sLit "ApplyTo") <+> ppr dup <+> pprParendExpr arg)
136 {- $$ nest 2 (pprSimplEnv se) -}) $$ ppr cont
137 ppr (StrictBind b _ _ _ cont) = (ptext (sLit "StrictBind") <+> ppr b) $$ ppr cont
138 ppr (StrictArg f _ _ cont) = (ptext (sLit "StrictArg") <+> ppr f) $$ ppr cont
139 ppr (Select dup bndr alts _ cont) = (ptext (sLit "Select") <+> ppr dup <+> ppr bndr) $$
140 (nest 4 (ppr alts)) $$ ppr cont
141 ppr (CoerceIt co cont) = (ptext (sLit "CoerceIt") <+> ppr co) $$ ppr cont
143 data DupFlag = OkToDup | NoDup
145 instance Outputable DupFlag where
146 ppr OkToDup = ptext (sLit "ok")
147 ppr NoDup = ptext (sLit "nodup")
152 mkBoringStop :: SimplCont
153 mkBoringStop = Stop BoringCtxt
155 mkLazyArgStop :: CallCtxt -> SimplCont
156 mkLazyArgStop cci = Stop cci
159 contIsRhsOrArg :: SimplCont -> Bool
160 contIsRhsOrArg (Stop {}) = True
161 contIsRhsOrArg (StrictBind {}) = True
162 contIsRhsOrArg (StrictArg {}) = True
163 contIsRhsOrArg _ = False
166 contIsDupable :: SimplCont -> Bool
167 contIsDupable (Stop {}) = True
168 contIsDupable (ApplyTo OkToDup _ _ _) = True
169 contIsDupable (Select OkToDup _ _ _ _) = True
170 contIsDupable (CoerceIt _ cont) = contIsDupable cont
171 contIsDupable _ = False
174 contIsTrivial :: SimplCont -> Bool
175 contIsTrivial (Stop {}) = True
176 contIsTrivial (ApplyTo _ (Type _) _ cont) = contIsTrivial cont
177 contIsTrivial (CoerceIt _ cont) = contIsTrivial cont
178 contIsTrivial _ = False
181 contResultType :: SimplEnv -> OutType -> SimplCont -> OutType
182 contResultType env ty cont
185 subst_ty se ty = substTy (se `setInScope` env) ty
188 go (CoerceIt co cont) _ = go cont (snd (coercionKind co))
189 go (StrictBind _ bs body se cont) _ = go cont (subst_ty se (exprType (mkLams bs body)))
190 go (StrictArg fn _ _ cont) _ = go cont (funResultTy (exprType fn))
191 go (Select _ _ alts se cont) _ = go cont (subst_ty se (coreAltsType alts))
192 go (ApplyTo _ arg se cont) ty = go cont (apply_to_arg ty arg se)
194 apply_to_arg ty (Type ty_arg) se = applyTy ty (subst_ty se ty_arg)
195 apply_to_arg ty _ _ = funResultTy ty
198 countValArgs :: SimplCont -> Int
199 countValArgs (ApplyTo _ (Type _) _ cont) = countValArgs cont
200 countValArgs (ApplyTo _ _ _ cont) = 1 + countValArgs cont
203 countArgs :: SimplCont -> Int
204 countArgs (ApplyTo _ _ _ cont) = 1 + countArgs cont
207 contArgs :: SimplCont -> ([OutExpr], SimplCont)
208 -- Uses substitution to turn each arg into an OutExpr
209 contArgs cont = go [] cont
211 go args (ApplyTo _ arg se cont) = go (substExpr se arg : args) cont
212 go args cont = (reverse args, cont)
214 dropArgs :: Int -> SimplCont -> SimplCont
215 dropArgs 0 cont = cont
216 dropArgs n (ApplyTo _ _ _ cont) = dropArgs (n-1) cont
217 dropArgs n other = pprPanic "dropArgs" (ppr n <+> ppr other)
220 splitInlineCont :: SimplCont -> Maybe (SimplCont, SimplCont)
221 -- Returns Nothing if the continuation should dissolve an InlineMe Note
222 -- Return Just (c1,c2) otherwise,
223 -- where c1 is the continuation to put inside the InlineMe
226 -- Example: (__inline_me__ (/\a. e)) ty
227 -- Here we want to do the beta-redex without dissolving the InlineMe
228 -- See test simpl017 (and Trac #1627) for a good example of why this is important
230 splitInlineCont (ApplyTo dup (Type ty) se c)
231 | Just (c1, c2) <- splitInlineCont c = Just (ApplyTo dup (Type ty) se c1, c2)
232 splitInlineCont cont@(Stop {}) = Just (mkBoringStop, cont)
233 splitInlineCont cont@(StrictBind {}) = Just (mkBoringStop, cont)
234 splitInlineCont _ = Nothing
235 -- NB: we dissolve an InlineMe in any strict context,
236 -- not just function aplication.
237 -- E.g. foldr k z (__inline_me (case x of p -> build ...))
238 -- Here we want to get rid of the __inline_me__ so we
239 -- can float the case, and see foldr/build
241 -- However *not* in a strict RHS, else we get
242 -- let f = __inline_me__ (\x. e) in ...f...
243 -- Now if f is guaranteed to be called, hence a strict binding
244 -- we don't thereby want to dissolve the __inline_me__; for
245 -- example, 'f' might be a wrapper, so we'd inline the worker
249 Note [Interesting call context]
250 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
251 We want to avoid inlining an expression where there can't possibly be
252 any gain, such as in an argument position. Hence, if the continuation
253 is interesting (eg. a case scrutinee, application etc.) then we
254 inline, otherwise we don't.
256 Previously some_benefit used to return True only if the variable was
257 applied to some value arguments. This didn't work:
259 let x = _coerce_ (T Int) Int (I# 3) in
260 case _coerce_ Int (T Int) x of
263 we want to inline x, but can't see that it's a constructor in a case
264 scrutinee position, and some_benefit is False.
268 dMonadST = _/\_ t -> :Monad (g1 _@_ t, g2 _@_ t, g3 _@_ t)
270 .... case dMonadST _@_ x0 of (a,b,c) -> ....
272 we'd really like to inline dMonadST here, but we *don't* want to
273 inline if the case expression is just
275 case x of y { DEFAULT -> ... }
277 since we can just eliminate this case instead (x is in WHNF). Similar
278 applies when x is bound to a lambda expression. Hence
279 contIsInteresting looks for case expressions with just a single
284 interestingCallContext :: SimplCont -> CallCtxt
285 -- See Note [Interesting call context]
286 interestingCallContext cont
289 interesting (Select _ bndr _ _ _)
290 | isDeadBinder bndr = CaseCtxt
291 | otherwise = ArgCtxt False 2 -- If the binder is used, this
292 -- is like a strict let
294 interesting (ApplyTo _ arg _ cont)
295 | isTypeArg arg = interesting cont
296 | otherwise = ValAppCtxt -- Can happen if we have (f Int |> co) y
297 -- If f has an INLINE prag we need to give it some
298 -- motivation to inline. See Note [Cast then apply]
301 interesting (StrictArg _ cci _ _) = cci
302 interesting (StrictBind {}) = BoringCtxt
303 interesting (Stop cci) = cci
304 interesting (CoerceIt _ cont) = interesting cont
305 -- If this call is the arg of a strict function, the context
306 -- is a bit interesting. If we inline here, we may get useful
307 -- evaluation information to avoid repeated evals: e.g.
309 -- Here the contIsInteresting makes the '*' keener to inline,
310 -- which in turn exposes a constructor which makes the '+' inline.
311 -- Assuming that +,* aren't small enough to inline regardless.
313 -- It's also very important to inline in a strict context for things
316 -- Here, the context of (f x) is strict, and if f's unfolding is
317 -- a build it's *great* to inline it here. So we must ensure that
318 -- the context for (f x) is not totally uninteresting.
323 -> Int -- Number of value args
324 -> SimplCont -- Context of the call
327 mkArgInfo fun n_val_args call_cont
328 | n_val_args < idArity fun -- Note [Unsaturated functions]
329 = ArgInfo { ai_rules = False
330 , ai_strs = vanilla_stricts
331 , ai_discs = vanilla_discounts }
333 = ArgInfo { ai_rules = interestingArgContext fun call_cont
334 , ai_strs = add_type_str (idType fun) arg_stricts
335 , ai_discs = arg_discounts }
337 vanilla_discounts, arg_discounts :: [Int]
338 vanilla_discounts = repeat 0
339 arg_discounts = case idUnfolding fun of
340 CoreUnfolding _ _ _ _ _ (UnfoldIfGoodArgs _ discounts _ _)
341 -> discounts ++ vanilla_discounts
342 _ -> vanilla_discounts
344 vanilla_stricts, arg_stricts :: [Bool]
345 vanilla_stricts = repeat False
348 = case splitStrictSig (idNewStrictness fun) of
349 (demands, result_info)
350 | not (demands `lengthExceeds` n_val_args)
351 -> -- Enough args, use the strictness given.
352 -- For bottoming functions we used to pretend that the arg
353 -- is lazy, so that we don't treat the arg as an
354 -- interesting context. This avoids substituting
355 -- top-level bindings for (say) strings into
356 -- calls to error. But now we are more careful about
357 -- inlining lone variables, so its ok (see SimplUtils.analyseCont)
358 if isBotRes result_info then
359 map isStrictDmd demands -- Finite => result is bottom
361 map isStrictDmd demands ++ vanilla_stricts
363 -> WARN( True, text "More demands than arity" <+> ppr fun <+> ppr (idArity fun)
364 <+> ppr n_val_args <+> ppr demands )
365 vanilla_stricts -- Not enough args, or no strictness
367 add_type_str :: Type -> [Bool] -> [Bool]
368 -- If the function arg types are strict, record that in the 'strictness bits'
369 -- No need to instantiate because unboxed types (which dominate the strict
370 -- types) can't instantiate type variables.
371 -- add_type_str is done repeatedly (for each call); might be better
372 -- once-for-all in the function
373 -- But beware primops/datacons with no strictness
374 add_type_str _ [] = []
375 add_type_str fun_ty strs -- Look through foralls
376 | Just (_, fun_ty') <- splitForAllTy_maybe fun_ty -- Includes coercions
377 = add_type_str fun_ty' strs
378 add_type_str fun_ty (str:strs) -- Add strict-type info
379 | Just (arg_ty, fun_ty') <- splitFunTy_maybe fun_ty
380 = (str || isStrictType arg_ty) : add_type_str fun_ty' strs
384 {- Note [Unsaturated functions]
385 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
386 Consider (test eyeball/inline4)
389 where f has arity 2. Then we do not want to inline 'x', because
390 it'll just be floated out again. Even if f has lots of discounts
391 on its first argument -- it must be saturated for these to kick in
394 interestingArgContext :: Id -> SimplCont -> Bool
395 -- If the argument has form (f x y), where x,y are boring,
396 -- and f is marked INLINE, then we don't want to inline f.
397 -- But if the context of the argument is
399 -- where g has rules, then we *do* want to inline f, in case it
400 -- exposes a rule that might fire. Similarly, if the context is
402 -- where h has rules, then we do want to inline f; hence the
403 -- call_cont argument to interestingArgContext
405 -- The interesting_arg_ctxt flag makes this happen; if it's
406 -- set, the inliner gets just enough keener to inline f
407 -- regardless of how boring f's arguments are, if it's marked INLINE
409 -- The alternative would be to *always* inline an INLINE function,
410 -- regardless of how boring its context is; but that seems overkill
411 -- For example, it'd mean that wrapper functions were always inlined
412 interestingArgContext fn call_cont
413 = idHasRules fn || go call_cont
415 go (Select {}) = False
416 go (ApplyTo {}) = False
417 go (StrictArg _ cci _ _) = interesting cci
418 go (StrictBind {}) = False -- ??
419 go (CoerceIt _ c) = go c
420 go (Stop cci) = interesting cci
422 interesting (ArgCtxt rules _) = rules
423 interesting _ = False
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 act
593 SimplPhase n _ -> isActive n act
594 act = idInlineActivation 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 _) = 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, don't 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 act
749 SimplPhase n _ -> isActive n act
750 act = idInlineActivation 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 (pre/postInlineUnconditionally)
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 act
773 act = idInlineActivation id
775 activeRule :: DynFlags -> SimplEnv -> Maybe (Activation -> Bool)
776 -- Nothing => No rules at all
777 activeRule dflags env
778 | not (dopt Opt_EnableRewriteRules 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 :: SimplEnv -> [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]
806 mkLam _env bndrs body
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 { let body' = tryEtaExpansion dflags body
829 ; return (mkLams bndrs body') }
832 = return (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 = do (floats, body') <- tryRhsTyLam env bndrs body
879 return (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 * Note [Arity care]: we need to be careful if we just look at f's
916 arity. Currently (Dec07), f's arity is visible in its own RHS (see
917 Note [Arity robustness] in SimplEnv) so we must *not* trust the
918 arity when checking that 'f' is a value. Otherwise we will
923 Which might change a terminiating program (think (f `seq` e)) to a
924 non-terminating one. So we check for being a loop breaker first.
926 However for GlobalIds we can look at the arity; and for primops we
927 must, since they have no unfolding.
929 * Regardless of whether 'f' is a value, we always want to
930 reduce (/\a -> f a) to f
931 This came up in a RULE: foldr (build (/\a -> g a))
932 did not match foldr (build (/\b -> ...something complex...))
933 The type checker can insert these eta-expanded versions,
934 with both type and dictionary lambdas; hence the slightly
937 * Never *reduce* arity. For example
939 Then if h has arity 1 we don't want to eta-reduce because then
940 f's arity would decrease, and that is bad
942 These delicacies are why we don't use exprIsTrivial and exprIsHNF here.
946 tryEtaReduce :: [OutBndr] -> OutExpr -> Maybe OutExpr
947 tryEtaReduce bndrs body
948 = go (reverse bndrs) body
950 incoming_arity = count isId bndrs
952 go (b : bs) (App fun arg) | ok_arg b arg = go bs fun -- Loop round
953 go [] fun | ok_fun fun = Just fun -- Success!
954 go _ _ = Nothing -- Failure!
956 -- Note [Eta reduction conditions]
957 ok_fun (App fun (Type ty))
958 | not (any (`elemVarSet` tyVarsOfType ty) bndrs)
961 = not (fun_id `elem` bndrs)
962 && (ok_fun_id fun_id || all ok_lam bndrs)
965 ok_fun_id fun = fun_arity fun >= incoming_arity
967 fun_arity fun -- See Note [Arity care]
968 | isLocalId fun && isLoopBreaker (idOccInfo fun) = 0
969 | otherwise = idArity fun
971 ok_lam v = isTyVar v || isDictId v
973 ok_arg b arg = varToCoreExpr b `cheapEqExpr` arg
977 %************************************************************************
981 %************************************************************************
985 f = \x1..xn -> N ==> f = \x1..xn y1..ym -> N y1..ym
988 where (in both cases)
990 * The xi can include type variables
992 * The yi are all value variables
994 * N is a NORMAL FORM (i.e. no redexes anywhere)
995 wanting a suitable number of extra args.
997 The biggest reason for doing this is for cases like
1003 Here we want to get the lambdas together. A good exmaple is the nofib
1004 program fibheaps, which gets 25% more allocation if you don't do this
1007 We may have to sandwich some coerces between the lambdas
1008 to make the types work. exprEtaExpandArity looks through coerces
1009 when computing arity; and etaExpand adds the coerces as necessary when
1010 actually computing the expansion.
1013 tryEtaExpansion :: DynFlags -> OutExpr -> OutExpr
1014 -- There is at least one runtime binder in the binders
1015 tryEtaExpansion dflags body
1016 = etaExpand fun_arity body
1018 fun_arity = exprEtaExpandArity dflags body
1022 %************************************************************************
1024 \subsection{Floating lets out of big lambdas}
1026 %************************************************************************
1028 Note [Floating and type abstraction]
1029 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
1032 We'd like to float this to
1035 x = /\a. C (y1 a) (y2 a)
1036 for the usual reasons: we want to inline x rather vigorously.
1038 You may think that this kind of thing is rare. But in some programs it is
1039 common. For example, if you do closure conversion you might get:
1041 data a :-> b = forall e. (e -> a -> b) :$ e
1043 f_cc :: forall a. a :-> a
1044 f_cc = /\a. (\e. id a) :$ ()
1046 Now we really want to inline that f_cc thing so that the
1047 construction of the closure goes away.
1049 So I have elaborated simplLazyBind to understand right-hand sides that look
1053 and treat them specially. The real work is done in SimplUtils.abstractFloats,
1054 but there is quite a bit of plumbing in simplLazyBind as well.
1056 The same transformation is good when there are lets in the body:
1058 /\abc -> let(rec) x = e in b
1060 let(rec) x' = /\abc -> let x = x' a b c in e
1062 /\abc -> let x = x' a b c in b
1064 This is good because it can turn things like:
1066 let f = /\a -> letrec g = ... g ... in g
1068 letrec g' = /\a -> ... g' a ...
1070 let f = /\ a -> g' a
1072 which is better. In effect, it means that big lambdas don't impede
1075 This optimisation is CRUCIAL in eliminating the junk introduced by
1076 desugaring mutually recursive definitions. Don't eliminate it lightly!
1078 [May 1999] If we do this transformation *regardless* then we can
1079 end up with some pretty silly stuff. For example,
1082 st = /\ s -> let { x1=r1 ; x2=r2 } in ...
1087 st = /\s -> ...[y1 s/x1, y2 s/x2]
1090 Unless the "..." is a WHNF there is really no point in doing this.
1091 Indeed it can make things worse. Suppose x1 is used strictly,
1094 x1* = case f y of { (a,b) -> e }
1096 If we abstract this wrt the tyvar we then can't do the case inline
1097 as we would normally do.
1099 That's why the whole transformation is part of the same process that
1100 floats let-bindings and constructor arguments out of RHSs. In particular,
1101 it is guarded by the doFloatFromRhs call in simplLazyBind.
1105 abstractFloats :: [OutTyVar] -> SimplEnv -> OutExpr -> SimplM ([OutBind], OutExpr)
1106 abstractFloats main_tvs body_env body
1107 = ASSERT( notNull body_floats )
1108 do { (subst, float_binds) <- mapAccumLM abstract empty_subst body_floats
1109 ; return (float_binds, CoreSubst.substExpr subst body) }
1111 main_tv_set = mkVarSet main_tvs
1112 body_floats = getFloats body_env
1113 empty_subst = CoreSubst.mkEmptySubst (seInScope body_env)
1115 abstract :: CoreSubst.Subst -> OutBind -> SimplM (CoreSubst.Subst, OutBind)
1116 abstract subst (NonRec id rhs)
1117 = do { (poly_id, poly_app) <- mk_poly tvs_here id
1118 ; let poly_rhs = mkLams tvs_here rhs'
1119 subst' = CoreSubst.extendIdSubst subst id poly_app
1120 ; return (subst', (NonRec poly_id poly_rhs)) }
1122 rhs' = CoreSubst.substExpr subst rhs
1123 tvs_here | any isCoVar main_tvs = main_tvs -- Note [Abstract over coercions]
1125 = varSetElems (main_tv_set `intersectVarSet` exprSomeFreeVars isTyVar rhs')
1127 -- Abstract only over the type variables free in the rhs
1128 -- wrt which the new binding is abstracted. But the naive
1129 -- approach of abstract wrt the tyvars free in the Id's type
1131 -- /\ a b -> let t :: (a,b) = (e1, e2)
1134 -- Here, b isn't free in x's type, but we must nevertheless
1135 -- abstract wrt b as well, because t's type mentions b.
1136 -- Since t is floated too, we'd end up with the bogus:
1137 -- poly_t = /\ a b -> (e1, e2)
1138 -- poly_x = /\ a -> fst (poly_t a *b*)
1139 -- So for now we adopt the even more naive approach of
1140 -- abstracting wrt *all* the tyvars. We'll see if that
1141 -- gives rise to problems. SLPJ June 98
1143 abstract subst (Rec prs)
1144 = do { (poly_ids, poly_apps) <- mapAndUnzipM (mk_poly tvs_here) ids
1145 ; let subst' = CoreSubst.extendSubstList subst (ids `zip` poly_apps)
1146 poly_rhss = [mkLams tvs_here (CoreSubst.substExpr subst' rhs) | rhs <- rhss]
1147 ; return (subst', Rec (poly_ids `zip` poly_rhss)) }
1149 (ids,rhss) = unzip prs
1150 -- For a recursive group, it's a bit of a pain to work out the minimal
1151 -- set of tyvars over which to abstract:
1152 -- /\ a b c. let x = ...a... in
1153 -- letrec { p = ...x...q...
1154 -- q = .....p...b... } in
1156 -- Since 'x' is abstracted over 'a', the {p,q} group must be abstracted
1157 -- over 'a' (because x is replaced by (poly_x a)) as well as 'b'.
1158 -- Since it's a pain, we just use the whole set, which is always safe
1160 -- If you ever want to be more selective, remember this bizarre case too:
1162 -- Here, we must abstract 'x' over 'a'.
1165 mk_poly tvs_here var
1166 = do { uniq <- getUniqueM
1167 ; let poly_name = setNameUnique (idName var) uniq -- Keep same name
1168 poly_ty = mkForAllTys tvs_here (idType var) -- But new type of course
1169 poly_id = transferPolyIdInfo var tvs_here $ -- Note [transferPolyIdInfo] in Id.lhs
1170 mkLocalId poly_name poly_ty
1171 ; return (poly_id, mkTyApps (Var poly_id) (mkTyVarTys tvs_here)) }
1172 -- In the olden days, it was crucial to copy the occInfo of the original var,
1173 -- because we were looking at occurrence-analysed but as yet unsimplified code!
1174 -- In particular, we mustn't lose the loop breakers. BUT NOW we are looking
1175 -- at already simplified code, so it doesn't matter
1177 -- It's even right to retain single-occurrence or dead-var info:
1178 -- Suppose we started with /\a -> let x = E in B
1179 -- where x occurs once in B. Then we transform to:
1180 -- let x' = /\a -> E in /\a -> let x* = x' a in B
1181 -- where x* has an INLINE prag on it. Now, once x* is inlined,
1182 -- the occurrences of x' will be just the occurrences originally
1186 Note [Abstract over coercions]
1187 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
1188 If a coercion variable (g :: a ~ Int) is free in the RHS, then so is the
1189 type variable a. Rather than sort this mess out, we simply bale out and abstract
1190 wrt all the type variables if any of them are coercion variables.
1193 Historical note: if you use let-bindings instead of a substitution, beware of this:
1195 -- Suppose we start with:
1197 -- x = /\ a -> let g = G in E
1199 -- Then we'll float to get
1201 -- x = let poly_g = /\ a -> G
1202 -- in /\ a -> let g = poly_g a in E
1204 -- But now the occurrence analyser will see just one occurrence
1205 -- of poly_g, not inside a lambda, so the simplifier will
1206 -- PreInlineUnconditionally poly_g back into g! Badk to square 1!
1207 -- (I used to think that the "don't inline lone occurrences" stuff
1208 -- would stop this happening, but since it's the *only* occurrence,
1209 -- PreInlineUnconditionally kicks in first!)
1211 -- Solution: put an INLINE note on g's RHS, so that poly_g seems
1212 -- to appear many times. (NB: mkInlineMe eliminates
1213 -- such notes on trivial RHSs, so do it manually.)
1215 %************************************************************************
1219 %************************************************************************
1221 prepareAlts tries these things:
1223 1. If several alternatives are identical, merge them into
1224 a single DEFAULT alternative. I've occasionally seen this
1225 making a big difference:
1227 case e of =====> case e of
1228 C _ -> f x D v -> ....v....
1229 D v -> ....v.... DEFAULT -> f x
1232 The point is that we merge common RHSs, at least for the DEFAULT case.
1233 [One could do something more elaborate but I've never seen it needed.]
1234 To avoid an expensive test, we just merge branches equal to the *first*
1235 alternative; this picks up the common cases
1236 a) all branches equal
1237 b) some branches equal to the DEFAULT (which occurs first)
1240 case e of b { ==> case e of b {
1241 p1 -> rhs1 p1 -> rhs1
1243 pm -> rhsm pm -> rhsm
1244 _ -> case b of b' { pn -> let b'=b in rhsn
1246 ... po -> let b'=b in rhso
1247 po -> rhso _ -> let b'=b in rhsd
1251 which merges two cases in one case when -- the default alternative of
1252 the outer case scrutises the same variable as the outer case This
1253 transformation is called Case Merging. It avoids that the same
1254 variable is scrutinised multiple times.
1257 The case where transformation (1) showed up was like this (lib/std/PrelCError.lhs):
1263 where @is@ was something like
1265 p `is` n = p /= (-1) && p == n
1267 This gave rise to a horrible sequence of cases
1274 and similarly in cascade for all the join points!
1277 ~~~~~~~~~~~~~~~~~~~~
1278 We do this *here*, looking at un-simplified alternatives, because we
1279 have to check that r doesn't mention the variables bound by the
1280 pattern in each alternative, so the binder-info is rather useful.
1283 prepareAlts :: SimplEnv -> OutExpr -> OutId -> [InAlt] -> SimplM ([AltCon], [InAlt])
1284 prepareAlts env scrut case_bndr' alts
1285 = do { dflags <- getDOptsSmpl
1286 ; alts <- combineIdenticalAlts case_bndr' alts
1288 ; let (alts_wo_default, maybe_deflt) = findDefault alts
1289 alt_cons = [con | (con,_,_) <- alts_wo_default]
1290 imposs_deflt_cons = nub (imposs_cons ++ alt_cons)
1291 -- "imposs_deflt_cons" are handled
1292 -- EITHER by the context,
1293 -- OR by a non-DEFAULT branch in this case expression.
1295 ; default_alts <- prepareDefault dflags env case_bndr' mb_tc_app
1296 imposs_deflt_cons maybe_deflt
1298 ; let trimmed_alts = filterOut impossible_alt alts_wo_default
1299 merged_alts = mergeAlts trimmed_alts default_alts
1300 -- We need the mergeAlts in case the new default_alt
1301 -- has turned into a constructor alternative.
1302 -- The merge keeps the inner DEFAULT at the front, if there is one
1303 -- and interleaves the alternatives in the right order
1305 ; return (imposs_deflt_cons, merged_alts) }
1307 mb_tc_app = splitTyConApp_maybe (idType case_bndr')
1308 Just (_, inst_tys) = mb_tc_app
1310 imposs_cons = case scrut of
1311 Var v -> otherCons (idUnfolding v)
1314 impossible_alt :: CoreAlt -> Bool
1315 impossible_alt (con, _, _) | con `elem` imposs_cons = True
1316 impossible_alt (DataAlt con, _, _) = dataConCannotMatch inst_tys con
1317 impossible_alt _ = False
1320 --------------------------------------------------
1321 -- 1. Merge identical branches
1322 --------------------------------------------------
1323 combineIdenticalAlts :: OutId -> [InAlt] -> SimplM [InAlt]
1325 combineIdenticalAlts case_bndr ((_con1,bndrs1,rhs1) : con_alts)
1326 | all isDeadBinder bndrs1, -- Remember the default
1327 length filtered_alts < length con_alts -- alternative comes first
1328 -- Also Note [Dead binders]
1329 = do { tick (AltMerge case_bndr)
1330 ; return ((DEFAULT, [], rhs1) : filtered_alts) }
1332 filtered_alts = filter keep con_alts
1333 keep (_con,bndrs,rhs) = not (all isDeadBinder bndrs && rhs `cheapEqExpr` rhs1)
1335 combineIdenticalAlts _ alts = return alts
1337 -------------------------------------------------------------------------
1338 -- Prepare the default alternative
1339 -------------------------------------------------------------------------
1340 prepareDefault :: DynFlags
1342 -> OutId -- Case binder; need just for its type. Note that as an
1343 -- OutId, it has maximum information; this is important.
1344 -- Test simpl013 is an example
1345 -> Maybe (TyCon, [Type]) -- Type of scrutinee, decomposed
1346 -> [AltCon] -- These cons can't happen when matching the default
1347 -> Maybe InExpr -- Rhs
1348 -> SimplM [InAlt] -- Still unsimplified
1349 -- We use a list because it's what mergeAlts expects,
1350 -- And becuase case-merging can cause many to show up
1352 ------- Merge nested cases ----------
1353 prepareDefault dflags env outer_bndr _bndr_ty imposs_cons (Just deflt_rhs)
1354 | dopt Opt_CaseMerge dflags
1355 , Case (Var inner_scrut_var) inner_bndr _ inner_alts <- deflt_rhs
1356 , DoneId inner_scrut_var' <- substId env inner_scrut_var
1357 -- Remember, inner_scrut_var is an InId, but outer_bndr is an OutId
1358 , inner_scrut_var' == outer_bndr
1359 -- NB: the substId means that if the outer scrutinee was a
1360 -- variable, and inner scrutinee is the same variable,
1361 -- then inner_scrut_var' will be outer_bndr
1362 -- via the magic of simplCaseBinder
1363 = do { tick (CaseMerge outer_bndr)
1365 ; let munge_rhs rhs = bindCaseBndr inner_bndr (Var outer_bndr) rhs
1366 ; return [(con, args, munge_rhs rhs) | (con, args, rhs) <- inner_alts,
1367 not (con `elem` imposs_cons) ]
1368 -- NB: filter out any imposs_cons. Example:
1371 -- DEFAULT -> case x of
1374 -- When we merge, we must ensure that e1 takes
1375 -- precedence over e2 as the value for A!
1377 -- Warning: don't call prepareAlts recursively!
1378 -- Firstly, there's no point, because inner alts have already had
1379 -- mkCase applied to them, so they won't have a case in their default
1380 -- Secondly, if you do, you get an infinite loop, because the bindCaseBndr
1381 -- in munge_rhs may put a case into the DEFAULT branch!
1384 --------- Fill in known constructor -----------
1385 prepareDefault _ _ case_bndr (Just (tycon, inst_tys)) imposs_cons (Just deflt_rhs)
1386 | -- This branch handles the case where we are
1387 -- scrutinisng an algebraic data type
1388 isAlgTyCon tycon -- It's a data type, tuple, or unboxed tuples.
1389 , not (isNewTyCon tycon) -- We can have a newtype, if we are just doing an eval:
1390 -- case x of { DEFAULT -> e }
1391 -- and we don't want to fill in a default for them!
1392 , Just all_cons <- tyConDataCons_maybe tycon
1393 , not (null all_cons) -- This is a tricky corner case. If the data type has no constructors,
1394 -- which GHC allows, then the case expression will have at most a default
1395 -- alternative. We don't want to eliminate that alternative, because the
1396 -- invariant is that there's always one alternative. It's more convenient
1398 -- case x of { DEFAULT -> e }
1399 -- as it is, rather than transform it to
1400 -- error "case cant match"
1401 -- which would be quite legitmate. But it's a really obscure corner, and
1402 -- not worth wasting code on.
1403 , let imposs_data_cons = [con | DataAlt con <- imposs_cons] -- We now know it's a data type
1404 impossible con = con `elem` imposs_data_cons || dataConCannotMatch inst_tys con
1405 = case filterOut impossible all_cons of
1406 [] -> return [] -- Eliminate the default alternative
1407 -- altogether if it can't match
1409 [con] -> -- It matches exactly one constructor, so fill it in
1410 do { tick (FillInCaseDefault case_bndr)
1412 ; let (ex_tvs, co_tvs, arg_ids) =
1413 dataConRepInstPat us con inst_tys
1414 ; return [(DataAlt con, ex_tvs ++ co_tvs ++ arg_ids, deflt_rhs)] }
1416 _ -> return [(DEFAULT, [], deflt_rhs)]
1418 | debugIsOn, isAlgTyCon tycon, not (isOpenTyCon tycon), null (tyConDataCons tycon)
1419 -- This can legitimately happen for type families, so don't report that
1420 = pprTrace "prepareDefault" (ppr case_bndr <+> ppr tycon)
1421 $ return [(DEFAULT, [], deflt_rhs)]
1423 --------- Catch-all cases -----------
1424 prepareDefault _dflags _env _case_bndr _bndr_ty _imposs_cons (Just deflt_rhs)
1425 = return [(DEFAULT, [], deflt_rhs)]
1427 prepareDefault _dflags _env _case_bndr _bndr_ty _imposs_cons Nothing
1428 = return [] -- No default branch
1433 =================================================================================
1435 mkCase tries these things
1437 1. Eliminate the case altogether if possible
1445 and similar friends.
1449 mkCase :: OutExpr -> OutId -> [OutAlt] -- Increasing order
1452 --------------------------------------------------
1454 --------------------------------------------------
1456 mkCase scrut case_bndr alts -- Identity case
1457 | all identity_alt alts
1458 = do tick (CaseIdentity case_bndr)
1459 return (re_cast scrut)
1461 identity_alt (con, args, rhs) = check_eq con args (de_cast rhs)
1463 check_eq DEFAULT _ (Var v) = v == case_bndr
1464 check_eq (LitAlt lit') _ (Lit lit) = lit == lit'
1465 check_eq (DataAlt con) args rhs = rhs `cheapEqExpr` mkConApp con (arg_tys ++ varsToCoreExprs args)
1466 || rhs `cheapEqExpr` Var case_bndr
1467 check_eq _ _ _ = False
1469 arg_tys = map Type (tyConAppArgs (idType case_bndr))
1472 -- case e of x { _ -> x `cast` c }
1473 -- And we definitely want to eliminate this case, to give
1475 -- So we throw away the cast from the RHS, and reconstruct
1476 -- it at the other end. All the RHS casts must be the same
1477 -- if (all identity_alt alts) holds.
1479 -- Don't worry about nested casts, because the simplifier combines them
1480 de_cast (Cast e _) = e
1483 re_cast scrut = case head alts of
1484 (_,_,Cast _ co) -> Cast scrut co
1489 --------------------------------------------------
1491 --------------------------------------------------
1492 mkCase scrut bndr alts = return (Case scrut bndr (coreAltsType alts) alts)
1496 When adding auxiliary bindings for the case binder, it's worth checking if
1497 its dead, because it often is, and occasionally these mkCase transformations
1498 cascade rather nicely.
1501 bindCaseBndr :: Id -> CoreExpr -> CoreExpr -> CoreExpr
1502 bindCaseBndr bndr rhs body
1503 | isDeadBinder bndr = body
1504 | otherwise = bindNonRec bndr rhs body