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
250 interestingArg :: OutExpr -> Bool
251 -- An argument is interesting if it has *some* structure
252 -- We are here trying to avoid unfolding a function that
253 -- is applied only to variables that have no unfolding
254 -- (i.e. they are probably lambda bound): f x y z
255 -- There is little point in inlining f here.
256 interestingArg (Var v) = hasSomeUnfolding (idUnfolding v)
257 -- Was: isValueUnfolding (idUnfolding v')
258 -- But that seems over-pessimistic
260 -- This accounts for an argument like
261 -- () or [], which is definitely interesting
262 interestingArg (Type _) = False
263 interestingArg (App fn (Type _)) = interestingArg fn
264 interestingArg (Note _ a) = interestingArg a
266 -- Idea (from Sam B); I'm not sure if it's a good idea, so commented out for now
267 -- interestingArg expr | isUnLiftedType (exprType expr)
268 -- -- Unlifted args are only ever interesting if we know what they are
273 interestingArg _ = True
274 -- Consider let x = 3 in f x
275 -- The substitution will contain (x -> ContEx 3), and we want to
276 -- to say that x is an interesting argument.
277 -- But consider also (\x. f x y) y
278 -- The substitution will contain (x -> ContEx y), and we want to say
279 -- that x is not interesting (assuming y has no unfolding)
283 Comment about interestingCallContext
284 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
285 We want to avoid inlining an expression where there can't possibly be
286 any gain, such as in an argument position. Hence, if the continuation
287 is interesting (eg. a case scrutinee, application etc.) then we
288 inline, otherwise we don't.
290 Previously some_benefit used to return True only if the variable was
291 applied to some value arguments. This didn't work:
293 let x = _coerce_ (T Int) Int (I# 3) in
294 case _coerce_ Int (T Int) x of
297 we want to inline x, but can't see that it's a constructor in a case
298 scrutinee position, and some_benefit is False.
302 dMonadST = _/\_ t -> :Monad (g1 _@_ t, g2 _@_ t, g3 _@_ t)
304 .... case dMonadST _@_ x0 of (a,b,c) -> ....
306 we'd really like to inline dMonadST here, but we *don't* want to
307 inline if the case expression is just
309 case x of y { DEFAULT -> ... }
311 since we can just eliminate this case instead (x is in WHNF). Similar
312 applies when x is bound to a lambda expression. Hence
313 contIsInteresting looks for case expressions with just a single
318 interestingCallContext :: SimplCont -> CallCtxt
319 interestingCallContext cont
322 interesting (Select _ bndr _ _ _)
323 | isDeadBinder bndr = CaseCtxt
324 | otherwise = ArgCtxt False 2 -- If the binder is used, this
325 -- is like a strict let
327 interesting (ApplyTo _ arg _ cont)
328 | isTypeArg arg = interesting cont
329 | otherwise = ValAppCtxt -- Can happen if we have (f Int |> co) y
330 -- If f has an INLINE prag we need to give it some
331 -- motivation to inline. See Note [Cast then apply]
334 interesting (StrictArg _ cci _ _) = cci
335 interesting (StrictBind {}) = BoringCtxt
336 interesting (Stop cci) = cci
337 interesting (CoerceIt _ cont) = interesting cont
338 -- If this call is the arg of a strict function, the context
339 -- is a bit interesting. If we inline here, we may get useful
340 -- evaluation information to avoid repeated evals: e.g.
342 -- Here the contIsInteresting makes the '*' keener to inline,
343 -- which in turn exposes a constructor which makes the '+' inline.
344 -- Assuming that +,* aren't small enough to inline regardless.
346 -- It's also very important to inline in a strict context for things
349 -- Here, the context of (f x) is strict, and if f's unfolding is
350 -- a build it's *great* to inline it here. So we must ensure that
351 -- the context for (f x) is not totally uninteresting.
356 -> Int -- Number of value args
357 -> SimplCont -- Context of the cal
360 mkArgInfo fun n_val_args call_cont
361 | n_val_args < idArity fun -- Note [Unsaturated functions]
362 = ArgInfo { ai_rules = False
363 , ai_strs = vanilla_stricts
364 , ai_discs = vanilla_discounts }
366 = ArgInfo { ai_rules = interestingArgContext fun call_cont
367 , ai_strs = add_type_str (idType fun) arg_stricts
368 , ai_discs = arg_discounts }
370 vanilla_discounts, arg_discounts :: [Int]
371 vanilla_discounts = repeat 0
372 arg_discounts = case idUnfolding fun of
373 CoreUnfolding _ _ _ _ _ (UnfoldIfGoodArgs _ discounts _ _)
374 -> discounts ++ vanilla_discounts
375 _ -> vanilla_discounts
377 vanilla_stricts, arg_stricts :: [Bool]
378 vanilla_stricts = repeat False
381 = case splitStrictSig (idNewStrictness fun) of
382 (demands, result_info)
383 | not (demands `lengthExceeds` n_val_args)
384 -> -- Enough args, use the strictness given.
385 -- For bottoming functions we used to pretend that the arg
386 -- is lazy, so that we don't treat the arg as an
387 -- interesting context. This avoids substituting
388 -- top-level bindings for (say) strings into
389 -- calls to error. But now we are more careful about
390 -- inlining lone variables, so its ok (see SimplUtils.analyseCont)
391 if isBotRes result_info then
392 map isStrictDmd demands -- Finite => result is bottom
394 map isStrictDmd demands ++ vanilla_stricts
396 -> WARN( True, text "More demands than arity" <+> ppr fun <+> ppr (idArity fun)
397 <+> ppr n_val_args <+> ppr demands )
398 vanilla_stricts -- Not enough args, or no strictness
400 add_type_str :: Type -> [Bool] -> [Bool]
401 -- If the function arg types are strict, record that in the 'strictness bits'
402 -- No need to instantiate because unboxed types (which dominate the strict
403 -- types) can't instantiate type variables.
404 -- add_type_str is done repeatedly (for each call); might be better
405 -- once-for-all in the function
406 -- But beware primops/datacons with no strictness
407 add_type_str _ [] = []
408 add_type_str fun_ty strs -- Look through foralls
409 | Just (_, fun_ty') <- splitForAllTy_maybe fun_ty -- Includes coercions
410 = add_type_str fun_ty' strs
411 add_type_str fun_ty (str:strs) -- Add strict-type info
412 | Just (arg_ty, fun_ty') <- splitFunTy_maybe fun_ty
413 = (str || isStrictType arg_ty) : add_type_str fun_ty' strs
417 {- Note [Unsaturated functions]
418 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
419 Consider (test eyeball/inline4)
422 where f has arity 2. Then we do not want to inline 'x', because
423 it'll just be floated out again. Even if f has lots of discounts
424 on its first argument -- it must be saturated for these to kick in
427 interestingArgContext :: Id -> SimplCont -> Bool
428 -- If the argument has form (f x y), where x,y are boring,
429 -- and f is marked INLINE, then we don't want to inline f.
430 -- But if the context of the argument is
432 -- where g has rules, then we *do* want to inline f, in case it
433 -- exposes a rule that might fire. Similarly, if the context is
435 -- where h has rules, then we do want to inline f; hence the
436 -- call_cont argument to interestingArgContext
438 -- The interesting_arg_ctxt flag makes this happen; if it's
439 -- set, the inliner gets just enough keener to inline f
440 -- regardless of how boring f's arguments are, if it's marked INLINE
442 -- The alternative would be to *always* inline an INLINE function,
443 -- regardless of how boring its context is; but that seems overkill
444 -- For example, it'd mean that wrapper functions were always inlined
445 interestingArgContext fn call_cont
446 = idHasRules fn || go call_cont
448 go (Select {}) = False
449 go (ApplyTo {}) = False
450 go (StrictArg _ cci _ _) = interesting cci
451 go (StrictBind {}) = False -- ??
452 go (CoerceIt _ c) = go c
453 go (Stop cci) = interesting cci
455 interesting (ArgCtxt rules _) = rules
456 interesting _ = False
461 %************************************************************************
463 \subsection{Decisions about inlining}
465 %************************************************************************
467 Inlining is controlled partly by the SimplifierMode switch. This has two
470 SimplGently (a) Simplifying before specialiser/full laziness
471 (b) Simplifiying inside INLINE pragma
472 (c) Simplifying the LHS of a rule
473 (d) Simplifying a GHCi expression or Template
476 SimplPhase n _ Used at all other times
478 The key thing about SimplGently is that it does no call-site inlining.
479 Before full laziness we must be careful not to inline wrappers,
480 because doing so inhibits floating
481 e.g. ...(case f x of ...)...
482 ==> ...(case (case x of I# x# -> fw x#) of ...)...
483 ==> ...(case x of I# x# -> case fw x# of ...)...
484 and now the redex (f x) isn't floatable any more.
486 The no-inlining thing is also important for Template Haskell. You might be
487 compiling in one-shot mode with -O2; but when TH compiles a splice before
488 running it, we don't want to use -O2. Indeed, we don't want to inline
489 anything, because the byte-code interpreter might get confused about
490 unboxed tuples and suchlike.
494 SimplGently is also used as the mode to simplify inside an InlineMe note.
497 inlineMode :: SimplifierMode
498 inlineMode = SimplGently
501 It really is important to switch off inlinings inside such
502 expressions. Consider the following example
508 in ...g...g...g...g...g...
510 Now, if that's the ONLY occurrence of f, it will be inlined inside g,
511 and thence copied multiple times when g is inlined.
514 This function may be inlinined in other modules, so we
515 don't want to remove (by inlining) calls to functions that have
516 specialisations, or that may have transformation rules in an importing
519 E.g. {-# INLINE f #-}
522 and suppose that g is strict *and* has specialisations. If we inline
523 g's wrapper, we deny f the chance of getting the specialised version
524 of g when f is inlined at some call site (perhaps in some other
527 It's also important not to inline a worker back into a wrapper.
529 wraper = inline_me (\x -> ...worker... )
530 Normally, the inline_me prevents the worker getting inlined into
531 the wrapper (initially, the worker's only call site!). But,
532 if the wrapper is sure to be called, the strictness analyser will
533 mark it 'demanded', so when the RHS is simplified, it'll get an ArgOf
534 continuation. That's why the keep_inline predicate returns True for
535 ArgOf continuations. It shouldn't do any harm not to dissolve the
536 inline-me note under these circumstances.
538 Note that the result is that we do very little simplification
541 all xs = foldr (&&) True xs
542 any p = all . map p {-# INLINE any #-}
544 Problem: any won't get deforested, and so if it's exported and the
545 importer doesn't use the inlining, (eg passes it as an arg) then we
546 won't get deforestation at all. We havn't solved this problem yet!
549 preInlineUnconditionally
550 ~~~~~~~~~~~~~~~~~~~~~~~~
551 @preInlineUnconditionally@ examines a bndr to see if it is used just
552 once in a completely safe way, so that it is safe to discard the
553 binding inline its RHS at the (unique) usage site, REGARDLESS of how
554 big the RHS might be. If this is the case we don't simplify the RHS
555 first, but just inline it un-simplified.
557 This is much better than first simplifying a perhaps-huge RHS and then
558 inlining and re-simplifying it. Indeed, it can be at least quadratically
567 We may end up simplifying e1 N times, e2 N-1 times, e3 N-3 times etc.
568 This can happen with cascades of functions too:
575 THE MAIN INVARIANT is this:
577 ---- preInlineUnconditionally invariant -----
578 IF preInlineUnconditionally chooses to inline x = <rhs>
579 THEN doing the inlining should not change the occurrence
580 info for the free vars of <rhs>
581 ----------------------------------------------
583 For example, it's tempting to look at trivial binding like
585 and inline it unconditionally. But suppose x is used many times,
586 but this is the unique occurrence of y. Then inlining x would change
587 y's occurrence info, which breaks the invariant. It matters: y
588 might have a BIG rhs, which will now be dup'd at every occurrenc of x.
591 Even RHSs labelled InlineMe aren't caught here, because there might be
592 no benefit from inlining at the call site.
594 [Sept 01] Don't unconditionally inline a top-level thing, because that
595 can simply make a static thing into something built dynamically. E.g.
599 [Remember that we treat \s as a one-shot lambda.] No point in
600 inlining x unless there is something interesting about the call site.
602 But watch out: if you aren't careful, some useful foldr/build fusion
603 can be lost (most notably in spectral/hartel/parstof) because the
604 foldr didn't see the build. Doing the dynamic allocation isn't a big
605 deal, in fact, but losing the fusion can be. But the right thing here
606 seems to be to do a callSiteInline based on the fact that there is
607 something interesting about the call site (it's strict). Hmm. That
610 Conclusion: inline top level things gaily until Phase 0 (the last
611 phase), at which point don't.
614 preInlineUnconditionally :: SimplEnv -> TopLevelFlag -> InId -> InExpr -> Bool
615 preInlineUnconditionally env top_lvl bndr rhs
617 | opt_SimplNoPreInlining = False
618 | otherwise = case idOccInfo bndr of
619 IAmDead -> True -- Happens in ((\x.1) v)
620 OneOcc in_lam True int_cxt -> try_once in_lam int_cxt
624 active = case phase of
625 SimplGently -> isAlwaysActive act
626 SimplPhase n _ -> isActive n act
627 act = idInlineActivation bndr
629 try_once in_lam int_cxt -- There's one textual occurrence
630 | not in_lam = isNotTopLevel top_lvl || early_phase
631 | otherwise = int_cxt && canInlineInLam rhs
633 -- Be very careful before inlining inside a lambda, becuase (a) we must not
634 -- invalidate occurrence information, and (b) we want to avoid pushing a
635 -- single allocation (here) into multiple allocations (inside lambda).
636 -- Inlining a *function* with a single *saturated* call would be ok, mind you.
637 -- || (if is_cheap && not (canInlineInLam rhs) then pprTrace "preinline" (ppr bndr <+> ppr rhs) ok else ok)
639 -- is_cheap = exprIsCheap rhs
640 -- ok = is_cheap && int_cxt
642 -- int_cxt The context isn't totally boring
643 -- E.g. let f = \ab.BIG in \y. map f xs
644 -- Don't want to substitute for f, because then we allocate
645 -- its closure every time the \y is called
646 -- But: let f = \ab.BIG in \y. map (f y) xs
647 -- Now we do want to substitute for f, even though it's not
648 -- saturated, because we're going to allocate a closure for
649 -- (f y) every time round the loop anyhow.
651 -- canInlineInLam => free vars of rhs are (Once in_lam) or Many,
652 -- so substituting rhs inside a lambda doesn't change the occ info.
653 -- Sadly, not quite the same as exprIsHNF.
654 canInlineInLam (Lit _) = True
655 canInlineInLam (Lam b e) = isRuntimeVar b || canInlineInLam e
656 canInlineInLam (Note _ e) = canInlineInLam e
657 canInlineInLam _ = False
659 early_phase = case phase of
660 SimplPhase 0 _ -> False
662 -- If we don't have this early_phase test, consider
663 -- x = length [1,2,3]
664 -- The full laziness pass carefully floats all the cons cells to
665 -- top level, and preInlineUnconditionally floats them all back in.
666 -- Result is (a) static allocation replaced by dynamic allocation
667 -- (b) many simplifier iterations because this tickles
668 -- a related problem; only one inlining per pass
670 -- On the other hand, I have seen cases where top-level fusion is
671 -- lost if we don't inline top level thing (e.g. string constants)
672 -- Hence the test for phase zero (which is the phase for all the final
673 -- simplifications). Until phase zero we take no special notice of
674 -- top level things, but then we become more leery about inlining
679 postInlineUnconditionally
680 ~~~~~~~~~~~~~~~~~~~~~~~~~
681 @postInlineUnconditionally@ decides whether to unconditionally inline
682 a thing based on the form of its RHS; in particular if it has a
683 trivial RHS. If so, we can inline and discard the binding altogether.
685 NB: a loop breaker has must_keep_binding = True and non-loop-breakers
686 only have *forward* references Hence, it's safe to discard the binding
688 NOTE: This isn't our last opportunity to inline. We're at the binding
689 site right now, and we'll get another opportunity when we get to the
692 Note that we do this unconditional inlining only for trival RHSs.
693 Don't inline even WHNFs inside lambdas; doing so may simply increase
694 allocation when the function is called. This isn't the last chance; see
697 NB: Even inline pragmas (e.g. IMustBeINLINEd) are ignored here Why?
698 Because we don't even want to inline them into the RHS of constructor
699 arguments. See NOTE above
701 NB: At one time even NOINLINE was ignored here: if the rhs is trivial
702 it's best to inline it anyway. We often get a=E; b=a from desugaring,
703 with both a and b marked NOINLINE. But that seems incompatible with
704 our new view that inlining is like a RULE, so I'm sticking to the 'active'
708 postInlineUnconditionally
709 :: SimplEnv -> TopLevelFlag
710 -> InId -- The binder (an OutId would be fine too)
711 -> OccInfo -- From the InId
715 postInlineUnconditionally env top_lvl bndr occ_info rhs unfolding
717 | isLoopBreaker occ_info = False -- If it's a loop-breaker of any kind, don't inline
718 -- because it might be referred to "earlier"
719 | isExportedId bndr = False
720 | exprIsTrivial rhs = True
723 -- The point of examining occ_info here is that for *non-values*
724 -- that occur outside a lambda, the call-site inliner won't have
725 -- a chance (becuase it doesn't know that the thing
726 -- only occurs once). The pre-inliner won't have gotten
727 -- it either, if the thing occurs in more than one branch
728 -- So the main target is things like
731 -- True -> case x of ...
732 -- False -> case x of ...
733 -- I'm not sure how important this is in practice
734 OneOcc in_lam _one_br int_cxt -- OneOcc => no code-duplication issue
735 -> smallEnoughToInline unfolding -- Small enough to dup
736 -- ToDo: consider discount on smallEnoughToInline if int_cxt is true
738 -- NB: Do NOT inline arbitrarily big things, even if one_br is True
739 -- Reason: doing so risks exponential behaviour. We simplify a big
740 -- expression, inline it, and simplify it again. But if the
741 -- very same thing happens in the big expression, we get
743 -- PRINCIPLE: when we've already simplified an expression once,
744 -- make sure that we only inline it if it's reasonably small.
746 && ((isNotTopLevel top_lvl && not in_lam) ||
747 -- But outside a lambda, we want to be reasonably aggressive
748 -- about inlining into multiple branches of case
749 -- e.g. let x = <non-value>
750 -- in case y of { C1 -> ..x..; C2 -> ..x..; C3 -> ... }
751 -- Inlining can be a big win if C3 is the hot-spot, even if
752 -- the uses in C1, C2 are not 'interesting'
753 -- An example that gets worse if you add int_cxt here is 'clausify'
755 (isCheapUnfolding unfolding && int_cxt))
756 -- isCheap => acceptable work duplication; in_lam may be true
757 -- int_cxt to prevent us inlining inside a lambda without some
758 -- good reason. See the notes on int_cxt in preInlineUnconditionally
760 IAmDead -> True -- This happens; for example, the case_bndr during case of
761 -- known constructor: case (a,b) of x { (p,q) -> ... }
762 -- Here x isn't mentioned in the RHS, so we don't want to
763 -- create the (dead) let-binding let x = (a,b) in ...
767 -- Here's an example that we don't handle well:
768 -- let f = if b then Left (\x.BIG) else Right (\y.BIG)
769 -- in \y. ....case f of {...} ....
770 -- Here f is used just once, and duplicating the case work is fine (exprIsCheap).
772 -- - We can't preInlineUnconditionally because that woud invalidate
773 -- the occ info for b.
774 -- - We can't postInlineUnconditionally because the RHS is big, and
775 -- that risks exponential behaviour
776 -- - We can't call-site inline, because the rhs is big
780 active = case getMode env of
781 SimplGently -> isAlwaysActive act
782 SimplPhase n _ -> isActive n act
783 act = idInlineActivation bndr
785 activeInline :: SimplEnv -> OutId -> Bool
787 = case getMode env of
789 -- No inlining at all when doing gentle stuff,
790 -- except for local things that occur once (pre/postInlineUnconditionally)
791 -- The reason is that too little clean-up happens if you
792 -- don't inline use-once things. Also a bit of inlining is *good* for
793 -- full laziness; it can expose constant sub-expressions.
794 -- Example in spectral/mandel/Mandel.hs, where the mandelset
795 -- function gets a useful let-float if you inline windowToViewport
797 -- NB: we used to have a second exception, for data con wrappers.
798 -- On the grounds that we use gentle mode for rule LHSs, and
799 -- they match better when data con wrappers are inlined.
800 -- But that only really applies to the trivial wrappers (like (:)),
801 -- and they are now constructed as Compulsory unfoldings (in MkId)
802 -- so they'll happen anyway.
804 SimplPhase n _ -> isActive n act
806 act = idInlineActivation id
808 activeRule :: DynFlags -> SimplEnv -> Maybe (Activation -> Bool)
809 -- Nothing => No rules at all
810 activeRule dflags env
811 | not (dopt Opt_EnableRewriteRules dflags)
812 = Nothing -- Rewriting is off
814 = case getMode env of
815 SimplGently -> Just isAlwaysActive
816 -- Used to be Nothing (no rules in gentle mode)
817 -- Main motivation for changing is that I wanted
818 -- lift String ===> ...
819 -- to work in Template Haskell when simplifying
820 -- splices, so we get simpler code for literal strings
821 SimplPhase n _ -> Just (isActive n)
825 %************************************************************************
829 %************************************************************************
832 mkLam :: SimplEnv -> [OutBndr] -> OutExpr -> SimplM OutExpr
833 -- mkLam tries three things
834 -- a) eta reduction, if that gives a trivial expression
835 -- b) eta expansion [only if there are some value lambdas]
839 mkLam _env bndrs body
840 = do { dflags <- getDOptsSmpl
841 ; mkLam' dflags bndrs body }
843 mkLam' :: DynFlags -> [OutBndr] -> OutExpr -> SimplM OutExpr
844 mkLam' dflags bndrs (Cast body co)
845 | not (any bad bndrs)
846 -- Note [Casts and lambdas]
847 = do { lam <- mkLam' dflags bndrs body
848 ; return (mkCoerce (mkPiTypes bndrs co) lam) }
850 co_vars = tyVarsOfType co
851 bad bndr = isCoVar bndr && bndr `elemVarSet` co_vars
853 mkLam' dflags bndrs body
854 | dopt Opt_DoEtaReduction dflags,
855 Just etad_lam <- tryEtaReduce bndrs body
856 = do { tick (EtaReduction (head bndrs))
859 | dopt Opt_DoLambdaEtaExpansion dflags,
860 any isRuntimeVar bndrs
861 = do { let body' = tryEtaExpansion dflags body
862 ; return (mkLams bndrs body') }
865 = return (mkLams bndrs body)
868 Note [Casts and lambdas]
869 ~~~~~~~~~~~~~~~~~~~~~~~~
871 (\x. (\y. e) `cast` g1) `cast` g2
872 There is a danger here that the two lambdas look separated, and the
873 full laziness pass might float an expression to between the two.
875 So this equation in mkLam' floats the g1 out, thus:
876 (\x. e `cast` g1) --> (\x.e) `cast` (tx -> g1)
879 In general, this floats casts outside lambdas, where (I hope) they
880 might meet and cancel with some other cast:
881 \x. e `cast` co ===> (\x. e) `cast` (tx -> co)
882 /\a. e `cast` co ===> (/\a. e) `cast` (/\a. co)
883 /\g. e `cast` co ===> (/\g. e) `cast` (/\g. co)
886 Notice that it works regardless of 'e'. Originally it worked only
887 if 'e' was itself a lambda, but in some cases that resulted in
888 fruitless iteration in the simplifier. A good example was when
889 compiling Text.ParserCombinators.ReadPrec, where we had a definition
890 like (\x. Get `cast` g)
891 where Get is a constructor with nonzero arity. Then mkLam eta-expanded
892 the Get, and the next iteration eta-reduced it, and then eta-expanded
895 Note also the side condition for the case of coercion binders.
896 It does not make sense to transform
897 /\g. e `cast` g ==> (/\g.e) `cast` (/\g.g)
898 because the latter is not well-kinded.
900 -- c) floating lets out through big lambdas
901 -- [only if all tyvar lambdas, and only if this lambda
902 -- is the RHS of a let]
904 {- Sept 01: I'm experimenting with getting the
905 full laziness pass to float out past big lambdsa
906 | all isTyVar bndrs, -- Only for big lambdas
907 contIsRhs cont -- Only try the rhs type-lambda floating
908 -- if this is indeed a right-hand side; otherwise
909 -- we end up floating the thing out, only for float-in
910 -- to float it right back in again!
911 = do (floats, body') <- tryRhsTyLam env bndrs body
912 return (floats, mkLams bndrs body')
916 %************************************************************************
920 %************************************************************************
922 Note [Eta reduction conditions]
923 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
924 We try for eta reduction here, but *only* if we get all the way to an
925 trivial expression. We don't want to remove extra lambdas unless we
926 are going to avoid allocating this thing altogether.
928 There are some particularly delicate points here:
930 * Eta reduction is not valid in general:
932 This matters, partly for old-fashioned correctness reasons but,
933 worse, getting it wrong can yield a seg fault. Consider
935 h y = case (case y of { True -> f `seq` True; False -> False }) of
936 True -> ...; False -> ...
938 If we (unsoundly) eta-reduce f to get f=f, the strictness analyser
939 says f=bottom, and replaces the (f `seq` True) with just
940 (f `cast` unsafe-co). BUT, as thing stand, 'f' got arity 1, and it
941 *keeps* arity 1 (perhaps also wrongly). So CorePrep eta-expands
942 the definition again, so that it does not termninate after all.
943 Result: seg-fault because the boolean case actually gets a function value.
946 So it's important to to the right thing.
948 * Note [Arity care]: we need to be careful if we just look at f's
949 arity. Currently (Dec07), f's arity is visible in its own RHS (see
950 Note [Arity robustness] in SimplEnv) so we must *not* trust the
951 arity when checking that 'f' is a value. Otherwise we will
956 Which might change a terminiating program (think (f `seq` e)) to a
957 non-terminating one. So we check for being a loop breaker first.
959 However for GlobalIds we can look at the arity; and for primops we
960 must, since they have no unfolding.
962 * Regardless of whether 'f' is a value, we always want to
963 reduce (/\a -> f a) to f
964 This came up in a RULE: foldr (build (/\a -> g a))
965 did not match foldr (build (/\b -> ...something complex...))
966 The type checker can insert these eta-expanded versions,
967 with both type and dictionary lambdas; hence the slightly
970 * Never *reduce* arity. For example
972 Then if h has arity 1 we don't want to eta-reduce because then
973 f's arity would decrease, and that is bad
975 These delicacies are why we don't use exprIsTrivial and exprIsHNF here.
979 tryEtaReduce :: [OutBndr] -> OutExpr -> Maybe OutExpr
980 tryEtaReduce bndrs body
981 = go (reverse bndrs) body
983 incoming_arity = count isId bndrs
985 go (b : bs) (App fun arg) | ok_arg b arg = go bs fun -- Loop round
986 go [] fun | ok_fun fun = Just fun -- Success!
987 go _ _ = Nothing -- Failure!
989 -- Note [Eta reduction conditions]
990 ok_fun (App fun (Type ty))
991 | not (any (`elemVarSet` tyVarsOfType ty) bndrs)
994 = not (fun_id `elem` bndrs)
995 && (ok_fun_id fun_id || all ok_lam bndrs)
998 ok_fun_id fun = fun_arity fun >= incoming_arity
1000 fun_arity fun -- See Note [Arity care]
1001 | isLocalId fun && isLoopBreaker (idOccInfo fun) = 0
1002 | otherwise = idArity fun
1004 ok_lam v = isTyVar v || isDictId v
1006 ok_arg b arg = varToCoreExpr b `cheapEqExpr` arg
1010 %************************************************************************
1014 %************************************************************************
1018 f = \x1..xn -> N ==> f = \x1..xn y1..ym -> N y1..ym
1021 where (in both cases)
1023 * The xi can include type variables
1025 * The yi are all value variables
1027 * N is a NORMAL FORM (i.e. no redexes anywhere)
1028 wanting a suitable number of extra args.
1030 The biggest reason for doing this is for cases like
1036 Here we want to get the lambdas together. A good exmaple is the nofib
1037 program fibheaps, which gets 25% more allocation if you don't do this
1040 We may have to sandwich some coerces between the lambdas
1041 to make the types work. exprEtaExpandArity looks through coerces
1042 when computing arity; and etaExpand adds the coerces as necessary when
1043 actually computing the expansion.
1046 tryEtaExpansion :: DynFlags -> OutExpr -> OutExpr
1047 -- There is at least one runtime binder in the binders
1048 tryEtaExpansion dflags body
1049 = etaExpand fun_arity body
1051 fun_arity = exprEtaExpandArity dflags body
1055 %************************************************************************
1057 \subsection{Floating lets out of big lambdas}
1059 %************************************************************************
1061 Note [Floating and type abstraction]
1062 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
1065 We'd like to float this to
1068 x = /\a. C (y1 a) (y2 a)
1069 for the usual reasons: we want to inline x rather vigorously.
1071 You may think that this kind of thing is rare. But in some programs it is
1072 common. For example, if you do closure conversion you might get:
1074 data a :-> b = forall e. (e -> a -> b) :$ e
1076 f_cc :: forall a. a :-> a
1077 f_cc = /\a. (\e. id a) :$ ()
1079 Now we really want to inline that f_cc thing so that the
1080 construction of the closure goes away.
1082 So I have elaborated simplLazyBind to understand right-hand sides that look
1086 and treat them specially. The real work is done in SimplUtils.abstractFloats,
1087 but there is quite a bit of plumbing in simplLazyBind as well.
1089 The same transformation is good when there are lets in the body:
1091 /\abc -> let(rec) x = e in b
1093 let(rec) x' = /\abc -> let x = x' a b c in e
1095 /\abc -> let x = x' a b c in b
1097 This is good because it can turn things like:
1099 let f = /\a -> letrec g = ... g ... in g
1101 letrec g' = /\a -> ... g' a ...
1103 let f = /\ a -> g' a
1105 which is better. In effect, it means that big lambdas don't impede
1108 This optimisation is CRUCIAL in eliminating the junk introduced by
1109 desugaring mutually recursive definitions. Don't eliminate it lightly!
1111 [May 1999] If we do this transformation *regardless* then we can
1112 end up with some pretty silly stuff. For example,
1115 st = /\ s -> let { x1=r1 ; x2=r2 } in ...
1120 st = /\s -> ...[y1 s/x1, y2 s/x2]
1123 Unless the "..." is a WHNF there is really no point in doing this.
1124 Indeed it can make things worse. Suppose x1 is used strictly,
1127 x1* = case f y of { (a,b) -> e }
1129 If we abstract this wrt the tyvar we then can't do the case inline
1130 as we would normally do.
1132 That's why the whole transformation is part of the same process that
1133 floats let-bindings and constructor arguments out of RHSs. In particular,
1134 it is guarded by the doFloatFromRhs call in simplLazyBind.
1138 abstractFloats :: [OutTyVar] -> SimplEnv -> OutExpr -> SimplM ([OutBind], OutExpr)
1139 abstractFloats main_tvs body_env body
1140 = ASSERT( notNull body_floats )
1141 do { (subst, float_binds) <- mapAccumLM abstract empty_subst body_floats
1142 ; return (float_binds, CoreSubst.substExpr subst body) }
1144 main_tv_set = mkVarSet main_tvs
1145 body_floats = getFloats body_env
1146 empty_subst = CoreSubst.mkEmptySubst (seInScope body_env)
1148 abstract :: CoreSubst.Subst -> OutBind -> SimplM (CoreSubst.Subst, OutBind)
1149 abstract subst (NonRec id rhs)
1150 = do { (poly_id, poly_app) <- mk_poly tvs_here id
1151 ; let poly_rhs = mkLams tvs_here rhs'
1152 subst' = CoreSubst.extendIdSubst subst id poly_app
1153 ; return (subst', (NonRec poly_id poly_rhs)) }
1155 rhs' = CoreSubst.substExpr subst rhs
1156 tvs_here | any isCoVar main_tvs = main_tvs -- Note [Abstract over coercions]
1158 = varSetElems (main_tv_set `intersectVarSet` exprSomeFreeVars isTyVar rhs')
1160 -- Abstract only over the type variables free in the rhs
1161 -- wrt which the new binding is abstracted. But the naive
1162 -- approach of abstract wrt the tyvars free in the Id's type
1164 -- /\ a b -> let t :: (a,b) = (e1, e2)
1167 -- Here, b isn't free in x's type, but we must nevertheless
1168 -- abstract wrt b as well, because t's type mentions b.
1169 -- Since t is floated too, we'd end up with the bogus:
1170 -- poly_t = /\ a b -> (e1, e2)
1171 -- poly_x = /\ a -> fst (poly_t a *b*)
1172 -- So for now we adopt the even more naive approach of
1173 -- abstracting wrt *all* the tyvars. We'll see if that
1174 -- gives rise to problems. SLPJ June 98
1176 abstract subst (Rec prs)
1177 = do { (poly_ids, poly_apps) <- mapAndUnzipM (mk_poly tvs_here) ids
1178 ; let subst' = CoreSubst.extendSubstList subst (ids `zip` poly_apps)
1179 poly_rhss = [mkLams tvs_here (CoreSubst.substExpr subst' rhs) | rhs <- rhss]
1180 ; return (subst', Rec (poly_ids `zip` poly_rhss)) }
1182 (ids,rhss) = unzip prs
1183 -- For a recursive group, it's a bit of a pain to work out the minimal
1184 -- set of tyvars over which to abstract:
1185 -- /\ a b c. let x = ...a... in
1186 -- letrec { p = ...x...q...
1187 -- q = .....p...b... } in
1189 -- Since 'x' is abstracted over 'a', the {p,q} group must be abstracted
1190 -- over 'a' (because x is replaced by (poly_x a)) as well as 'b'.
1191 -- Since it's a pain, we just use the whole set, which is always safe
1193 -- If you ever want to be more selective, remember this bizarre case too:
1195 -- Here, we must abstract 'x' over 'a'.
1198 mk_poly tvs_here var
1199 = do { uniq <- getUniqueM
1200 ; let poly_name = setNameUnique (idName var) uniq -- Keep same name
1201 poly_ty = mkForAllTys tvs_here (idType var) -- But new type of course
1202 poly_id = transferPolyIdInfo var tvs_here $ -- Note [transferPolyIdInfo] in Id.lhs
1203 mkLocalId poly_name poly_ty
1204 ; return (poly_id, mkTyApps (Var poly_id) (mkTyVarTys tvs_here)) }
1205 -- In the olden days, it was crucial to copy the occInfo of the original var,
1206 -- because we were looking at occurrence-analysed but as yet unsimplified code!
1207 -- In particular, we mustn't lose the loop breakers. BUT NOW we are looking
1208 -- at already simplified code, so it doesn't matter
1210 -- It's even right to retain single-occurrence or dead-var info:
1211 -- Suppose we started with /\a -> let x = E in B
1212 -- where x occurs once in B. Then we transform to:
1213 -- let x' = /\a -> E in /\a -> let x* = x' a in B
1214 -- where x* has an INLINE prag on it. Now, once x* is inlined,
1215 -- the occurrences of x' will be just the occurrences originally
1219 Note [Abstract over coercions]
1220 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
1221 If a coercion variable (g :: a ~ Int) is free in the RHS, then so is the
1222 type variable a. Rather than sort this mess out, we simply bale out and abstract
1223 wrt all the type variables if any of them are coercion variables.
1226 Historical note: if you use let-bindings instead of a substitution, beware of this:
1228 -- Suppose we start with:
1230 -- x = /\ a -> let g = G in E
1232 -- Then we'll float to get
1234 -- x = let poly_g = /\ a -> G
1235 -- in /\ a -> let g = poly_g a in E
1237 -- But now the occurrence analyser will see just one occurrence
1238 -- of poly_g, not inside a lambda, so the simplifier will
1239 -- PreInlineUnconditionally poly_g back into g! Badk to square 1!
1240 -- (I used to think that the "don't inline lone occurrences" stuff
1241 -- would stop this happening, but since it's the *only* occurrence,
1242 -- PreInlineUnconditionally kicks in first!)
1244 -- Solution: put an INLINE note on g's RHS, so that poly_g seems
1245 -- to appear many times. (NB: mkInlineMe eliminates
1246 -- such notes on trivial RHSs, so do it manually.)
1248 %************************************************************************
1252 %************************************************************************
1254 prepareAlts tries these things:
1256 1. If several alternatives are identical, merge them into
1257 a single DEFAULT alternative. I've occasionally seen this
1258 making a big difference:
1260 case e of =====> case e of
1261 C _ -> f x D v -> ....v....
1262 D v -> ....v.... DEFAULT -> f x
1265 The point is that we merge common RHSs, at least for the DEFAULT case.
1266 [One could do something more elaborate but I've never seen it needed.]
1267 To avoid an expensive test, we just merge branches equal to the *first*
1268 alternative; this picks up the common cases
1269 a) all branches equal
1270 b) some branches equal to the DEFAULT (which occurs first)
1273 case e of b { ==> case e of b {
1274 p1 -> rhs1 p1 -> rhs1
1276 pm -> rhsm pm -> rhsm
1277 _ -> case b of b' { pn -> let b'=b in rhsn
1279 ... po -> let b'=b in rhso
1280 po -> rhso _ -> let b'=b in rhsd
1284 which merges two cases in one case when -- the default alternative of
1285 the outer case scrutises the same variable as the outer case This
1286 transformation is called Case Merging. It avoids that the same
1287 variable is scrutinised multiple times.
1290 The case where transformation (1) showed up was like this (lib/std/PrelCError.lhs):
1296 where @is@ was something like
1298 p `is` n = p /= (-1) && p == n
1300 This gave rise to a horrible sequence of cases
1307 and similarly in cascade for all the join points!
1310 ~~~~~~~~~~~~~~~~~~~~
1311 We do this *here*, looking at un-simplified alternatives, because we
1312 have to check that r doesn't mention the variables bound by the
1313 pattern in each alternative, so the binder-info is rather useful.
1316 prepareAlts :: SimplEnv -> OutExpr -> OutId -> [InAlt] -> SimplM ([AltCon], [InAlt])
1317 prepareAlts env scrut case_bndr' alts
1318 = do { dflags <- getDOptsSmpl
1319 ; alts <- combineIdenticalAlts case_bndr' alts
1321 ; let (alts_wo_default, maybe_deflt) = findDefault alts
1322 alt_cons = [con | (con,_,_) <- alts_wo_default]
1323 imposs_deflt_cons = nub (imposs_cons ++ alt_cons)
1324 -- "imposs_deflt_cons" are handled
1325 -- EITHER by the context,
1326 -- OR by a non-DEFAULT branch in this case expression.
1328 ; default_alts <- prepareDefault dflags env case_bndr' mb_tc_app
1329 imposs_deflt_cons maybe_deflt
1331 ; let trimmed_alts = filterOut impossible_alt alts_wo_default
1332 merged_alts = mergeAlts trimmed_alts default_alts
1333 -- We need the mergeAlts in case the new default_alt
1334 -- has turned into a constructor alternative.
1335 -- The merge keeps the inner DEFAULT at the front, if there is one
1336 -- and interleaves the alternatives in the right order
1338 ; return (imposs_deflt_cons, merged_alts) }
1340 mb_tc_app = splitTyConApp_maybe (idType case_bndr')
1341 Just (_, inst_tys) = mb_tc_app
1343 imposs_cons = case scrut of
1344 Var v -> otherCons (idUnfolding v)
1347 impossible_alt :: CoreAlt -> Bool
1348 impossible_alt (con, _, _) | con `elem` imposs_cons = True
1349 impossible_alt (DataAlt con, _, _) = dataConCannotMatch inst_tys con
1350 impossible_alt _ = False
1353 --------------------------------------------------
1354 -- 1. Merge identical branches
1355 --------------------------------------------------
1356 combineIdenticalAlts :: OutId -> [InAlt] -> SimplM [InAlt]
1358 combineIdenticalAlts case_bndr ((_con1,bndrs1,rhs1) : con_alts)
1359 | all isDeadBinder bndrs1, -- Remember the default
1360 length filtered_alts < length con_alts -- alternative comes first
1361 -- Also Note [Dead binders]
1362 = do { tick (AltMerge case_bndr)
1363 ; return ((DEFAULT, [], rhs1) : filtered_alts) }
1365 filtered_alts = filter keep con_alts
1366 keep (_con,bndrs,rhs) = not (all isDeadBinder bndrs && rhs `cheapEqExpr` rhs1)
1368 combineIdenticalAlts _ alts = return alts
1370 -------------------------------------------------------------------------
1371 -- Prepare the default alternative
1372 -------------------------------------------------------------------------
1373 prepareDefault :: DynFlags
1375 -> OutId -- Case binder; need just for its type. Note that as an
1376 -- OutId, it has maximum information; this is important.
1377 -- Test simpl013 is an example
1378 -> Maybe (TyCon, [Type]) -- Type of scrutinee, decomposed
1379 -> [AltCon] -- These cons can't happen when matching the default
1380 -> Maybe InExpr -- Rhs
1381 -> SimplM [InAlt] -- Still unsimplified
1382 -- We use a list because it's what mergeAlts expects,
1383 -- And becuase case-merging can cause many to show up
1385 ------- Merge nested cases ----------
1386 prepareDefault dflags env outer_bndr _bndr_ty imposs_cons (Just deflt_rhs)
1387 | dopt Opt_CaseMerge dflags
1388 , Case (Var inner_scrut_var) inner_bndr _ inner_alts <- deflt_rhs
1389 , DoneId inner_scrut_var' <- substId env inner_scrut_var
1390 -- Remember, inner_scrut_var is an InId, but outer_bndr is an OutId
1391 , inner_scrut_var' == outer_bndr
1392 -- NB: the substId means that if the outer scrutinee was a
1393 -- variable, and inner scrutinee is the same variable,
1394 -- then inner_scrut_var' will be outer_bndr
1395 -- via the magic of simplCaseBinder
1396 = do { tick (CaseMerge outer_bndr)
1398 ; let munge_rhs rhs = bindCaseBndr inner_bndr (Var outer_bndr) rhs
1399 ; return [(con, args, munge_rhs rhs) | (con, args, rhs) <- inner_alts,
1400 not (con `elem` imposs_cons) ]
1401 -- NB: filter out any imposs_cons. Example:
1404 -- DEFAULT -> case x of
1407 -- When we merge, we must ensure that e1 takes
1408 -- precedence over e2 as the value for A!
1410 -- Warning: don't call prepareAlts recursively!
1411 -- Firstly, there's no point, because inner alts have already had
1412 -- mkCase applied to them, so they won't have a case in their default
1413 -- Secondly, if you do, you get an infinite loop, because the bindCaseBndr
1414 -- in munge_rhs may put a case into the DEFAULT branch!
1417 --------- Fill in known constructor -----------
1418 prepareDefault _ _ case_bndr (Just (tycon, inst_tys)) imposs_cons (Just deflt_rhs)
1419 | -- This branch handles the case where we are
1420 -- scrutinisng an algebraic data type
1421 isAlgTyCon tycon -- It's a data type, tuple, or unboxed tuples.
1422 , not (isNewTyCon tycon) -- We can have a newtype, if we are just doing an eval:
1423 -- case x of { DEFAULT -> e }
1424 -- and we don't want to fill in a default for them!
1425 , Just all_cons <- tyConDataCons_maybe tycon
1426 , not (null all_cons) -- This is a tricky corner case. If the data type has no constructors,
1427 -- which GHC allows, then the case expression will have at most a default
1428 -- alternative. We don't want to eliminate that alternative, because the
1429 -- invariant is that there's always one alternative. It's more convenient
1431 -- case x of { DEFAULT -> e }
1432 -- as it is, rather than transform it to
1433 -- error "case cant match"
1434 -- which would be quite legitmate. But it's a really obscure corner, and
1435 -- not worth wasting code on.
1436 , let imposs_data_cons = [con | DataAlt con <- imposs_cons] -- We now know it's a data type
1437 impossible con = con `elem` imposs_data_cons || dataConCannotMatch inst_tys con
1438 = case filterOut impossible all_cons of
1439 [] -> return [] -- Eliminate the default alternative
1440 -- altogether if it can't match
1442 [con] -> -- It matches exactly one constructor, so fill it in
1443 do { tick (FillInCaseDefault case_bndr)
1445 ; let (ex_tvs, co_tvs, arg_ids) =
1446 dataConRepInstPat us con inst_tys
1447 ; return [(DataAlt con, ex_tvs ++ co_tvs ++ arg_ids, deflt_rhs)] }
1449 _ -> return [(DEFAULT, [], deflt_rhs)]
1451 | debugIsOn, isAlgTyCon tycon, not (isOpenTyCon tycon), null (tyConDataCons tycon)
1452 -- This can legitimately happen for type families, so don't report that
1453 = pprTrace "prepareDefault" (ppr case_bndr <+> ppr tycon)
1454 $ return [(DEFAULT, [], deflt_rhs)]
1456 --------- Catch-all cases -----------
1457 prepareDefault _dflags _env _case_bndr _bndr_ty _imposs_cons (Just deflt_rhs)
1458 = return [(DEFAULT, [], deflt_rhs)]
1460 prepareDefault _dflags _env _case_bndr _bndr_ty _imposs_cons Nothing
1461 = return [] -- No default branch
1466 =================================================================================
1468 mkCase tries these things
1470 1. Eliminate the case altogether if possible
1478 and similar friends.
1482 mkCase :: OutExpr -> OutId -> [OutAlt] -- Increasing order
1485 --------------------------------------------------
1487 --------------------------------------------------
1489 mkCase scrut case_bndr alts -- Identity case
1490 | all identity_alt alts
1491 = do tick (CaseIdentity case_bndr)
1492 return (re_cast scrut)
1494 identity_alt (con, args, rhs) = check_eq con args (de_cast rhs)
1496 check_eq DEFAULT _ (Var v) = v == case_bndr
1497 check_eq (LitAlt lit') _ (Lit lit) = lit == lit'
1498 check_eq (DataAlt con) args rhs = rhs `cheapEqExpr` mkConApp con (arg_tys ++ varsToCoreExprs args)
1499 || rhs `cheapEqExpr` Var case_bndr
1500 check_eq _ _ _ = False
1502 arg_tys = map Type (tyConAppArgs (idType case_bndr))
1505 -- case e of x { _ -> x `cast` c }
1506 -- And we definitely want to eliminate this case, to give
1508 -- So we throw away the cast from the RHS, and reconstruct
1509 -- it at the other end. All the RHS casts must be the same
1510 -- if (all identity_alt alts) holds.
1512 -- Don't worry about nested casts, because the simplifier combines them
1513 de_cast (Cast e _) = e
1516 re_cast scrut = case head alts of
1517 (_,_,Cast _ co) -> Cast scrut co
1522 --------------------------------------------------
1524 --------------------------------------------------
1525 mkCase scrut bndr alts = return (Case scrut bndr (coreAltsType alts) alts)
1529 When adding auxiliary bindings for the case binder, it's worth checking if
1530 its dead, because it often is, and occasionally these mkCase transformations
1531 cascade rather nicely.
1534 bindCaseBndr :: Id -> CoreExpr -> CoreExpr -> CoreExpr
1535 bindCaseBndr bndr rhs body
1536 | isDeadBinder bndr = body
1537 | otherwise = bindNonRec bndr rhs body