2 % (c) The AQUA Project, Glasgow University, 1993-1998
4 \section[SimplUtils]{The simplifier utilities}
9 mkLam, mkCase, prepareAlts, tryEtaExpand,
12 preInlineUnconditionally, postInlineUnconditionally,
13 activeUnfolding, activeRule,
14 getUnfoldingInRuleMatch,
15 simplEnvForGHCi, updModeForInlineRules,
17 -- The continuation type
18 SimplCont(..), DupFlag(..), ArgInfo(..),
20 contIsDupable, contResultType, contIsTrivial, contArgs, dropArgs,
21 pushSimplifiedArgs, countValArgs, countArgs, addArgTo,
22 mkBoringStop, mkRhsStop, mkLazyArgStop, contIsRhsOrArg,
23 interestingCallContext,
25 interestingArg, mkArgInfo,
30 #include "HsVersions.h"
33 import CoreMonad ( SimplifierMode(..), Tick(..) )
37 import qualified CoreSubst
48 import Type hiding( substTy )
49 import Coercion ( coercionKind )
51 import Unify ( dataConCannotMatch )
63 %************************************************************************
67 %************************************************************************
69 A SimplCont allows the simplifier to traverse the expression in a
70 zipper-like fashion. The SimplCont represents the rest of the expression,
71 "above" the point of interest.
73 You can also think of a SimplCont as an "evaluation context", using
74 that term in the way it is used for operational semantics. This is the
75 way I usually think of it, For example you'll often see a syntax for
76 evaluation context looking like
77 C ::= [] | C e | case C of alts | C `cast` co
78 That's the kind of thing we are doing here, and I use that syntax in
83 * A SimplCont describes a *strict* context (just like
84 evaluation contexts do). E.g. Just [] is not a SimplCont
86 * A SimplCont describes a context that *does not* bind
87 any variables. E.g. \x. [] is not a SimplCont
91 = Stop -- An empty context, or hole, []
92 CallCtxt -- True <=> There is something interesting about
93 -- the context, and hence the inliner
94 -- should be a bit keener (see interestingCallContext)
96 -- This is an argument of a function that has RULES
97 -- Inlining the call might allow the rule to fire
99 | CoerceIt -- C `cast` co
100 OutCoercion -- The coercion simplified
104 DupFlag -- See Note [DupFlag invariants]
105 InExpr StaticEnv -- The argument and its static env
108 | Select -- case C of alts
109 DupFlag -- See Note [DupFlag invariants]
110 InId [InAlt] StaticEnv -- The case binder, alts, and subst-env
113 -- The two strict forms have no DupFlag, because we never duplicate them
114 | StrictBind -- (\x* \xs. e) C
115 InId [InBndr] -- let x* = [] in e
116 InExpr StaticEnv -- is a special case
119 | StrictArg -- f e1 ..en C
120 ArgInfo -- Specifies f, e1..en, Whether f has rules, etc
121 -- plus strictness flags for *further* args
122 CallCtxt -- Whether *this* argument position is interesting
127 ai_fun :: Id, -- The function
128 ai_args :: [OutExpr], -- ...applied to these args (which are in *reverse* order)
129 ai_rules :: [CoreRule], -- Rules for this function
131 ai_encl :: Bool, -- Flag saying whether this function
132 -- or an enclosing one has rules (recursively)
133 -- True => be keener to inline in all args
135 ai_strs :: [Bool], -- Strictness of remaining arguments
136 -- Usually infinite, but if it is finite it guarantees
137 -- that the function diverges after being given
138 -- that number of args
139 ai_discs :: [Int] -- Discounts for remaining arguments; non-zero => be keener to inline
143 addArgTo :: ArgInfo -> OutExpr -> ArgInfo
144 addArgTo ai arg = ai { ai_args = arg : ai_args ai }
146 instance Outputable SimplCont where
147 ppr (Stop interesting) = ptext (sLit "Stop") <> brackets (ppr interesting)
148 ppr (ApplyTo dup arg _ cont) = ((ptext (sLit "ApplyTo") <+> ppr dup <+> pprParendExpr arg)
149 {- $$ nest 2 (pprSimplEnv se) -}) $$ ppr cont
150 ppr (StrictBind b _ _ _ cont) = (ptext (sLit "StrictBind") <+> ppr b) $$ ppr cont
151 ppr (StrictArg ai _ cont) = (ptext (sLit "StrictArg") <+> ppr (ai_fun ai)) $$ ppr cont
152 ppr (Select dup bndr alts se cont) = (ptext (sLit "Select") <+> ppr dup <+> ppr bndr) $$
153 (nest 2 $ vcat [ppr (seTvSubst se), ppr alts]) $$ ppr cont
154 ppr (CoerceIt co cont) = (ptext (sLit "CoerceIt") <+> ppr co) $$ ppr cont
156 data DupFlag = NoDup -- Unsimplified, might be big
157 | Simplified -- Simplified
158 | OkToDup -- Simplified and small
160 isSimplified :: DupFlag -> Bool
161 isSimplified NoDup = False
162 isSimplified _ = True -- Invariant: the subst-env is empty
164 instance Outputable DupFlag where
165 ppr OkToDup = ptext (sLit "ok")
166 ppr NoDup = ptext (sLit "nodup")
167 ppr Simplified = ptext (sLit "simpl")
170 Note [DupFlag invariants]
171 ~~~~~~~~~~~~~~~~~~~~~~~~~
172 In both (ApplyTo dup _ env k)
173 and (Select dup _ _ env k)
174 the following invariants hold
176 (a) if dup = OkToDup, then continuation k is also ok-to-dup
177 (b) if dup = OkToDup or Simplified, the subst-env is empty
178 (and and hence no need to re-simplify)
182 mkBoringStop :: SimplCont
183 mkBoringStop = Stop BoringCtxt
185 mkRhsStop :: SimplCont -- See Note [RHS of lets] in CoreUnfold
186 mkRhsStop = Stop (ArgCtxt False)
188 mkLazyArgStop :: CallCtxt -> SimplCont
189 mkLazyArgStop cci = Stop cci
192 contIsRhsOrArg :: SimplCont -> Bool
193 contIsRhsOrArg (Stop {}) = True
194 contIsRhsOrArg (StrictBind {}) = True
195 contIsRhsOrArg (StrictArg {}) = True
196 contIsRhsOrArg _ = False
199 contIsDupable :: SimplCont -> Bool
200 contIsDupable (Stop {}) = True
201 contIsDupable (ApplyTo OkToDup _ _ _) = True -- See Note [DupFlag invariants]
202 contIsDupable (Select OkToDup _ _ _ _) = True -- ...ditto...
203 contIsDupable (CoerceIt _ cont) = contIsDupable cont
204 contIsDupable _ = False
207 contIsTrivial :: SimplCont -> Bool
208 contIsTrivial (Stop {}) = True
209 contIsTrivial (ApplyTo _ (Type _) _ cont) = contIsTrivial cont
210 contIsTrivial (CoerceIt _ cont) = contIsTrivial cont
211 contIsTrivial _ = False
214 contResultType :: SimplEnv -> OutType -> SimplCont -> OutType
215 contResultType env ty cont
218 subst_ty se ty = substTy (se `setInScope` env) ty
221 go (CoerceIt co cont) _ = go cont (snd (coercionKind co))
222 go (StrictBind _ bs body se cont) _ = go cont (subst_ty se (exprType (mkLams bs body)))
223 go (StrictArg ai _ cont) _ = go cont (funResultTy (argInfoResultTy ai))
224 go (Select _ _ alts se cont) _ = go cont (subst_ty se (coreAltsType alts))
225 go (ApplyTo _ arg se cont) ty = go cont (apply_to_arg ty arg se)
227 apply_to_arg ty (Type ty_arg) se = applyTy ty (subst_ty se ty_arg)
228 apply_to_arg ty _ _ = funResultTy ty
230 argInfoResultTy :: ArgInfo -> OutType
231 argInfoResultTy (ArgInfo { ai_fun = fun, ai_args = args })
232 = foldr (\arg fn_ty -> applyTypeToArg fn_ty arg) (idType fun) args
235 countValArgs :: SimplCont -> Int
236 countValArgs (ApplyTo _ (Type _) _ cont) = countValArgs cont
237 countValArgs (ApplyTo _ _ _ cont) = 1 + countValArgs cont
240 countArgs :: SimplCont -> Int
241 countArgs (ApplyTo _ _ _ cont) = 1 + countArgs cont
244 contArgs :: SimplCont -> (Bool, [ArgSummary], SimplCont)
245 -- Uses substitution to turn each arg into an OutExpr
246 contArgs cont@(ApplyTo {})
247 = case go [] cont of { (args, cont') -> (False, args, cont') }
249 go args (ApplyTo _ arg se cont)
250 | isTypeArg arg = go args cont
251 | otherwise = go (is_interesting arg se : args) cont
252 go args cont = (reverse args, cont)
254 is_interesting arg se = interestingArg (substExpr (text "contArgs") se arg)
255 -- Do *not* use short-cutting substitution here
256 -- because we want to get as much IdInfo as possible
258 contArgs cont = (True, [], cont)
260 pushSimplifiedArgs :: SimplEnv -> [CoreExpr] -> SimplCont -> SimplCont
261 pushSimplifiedArgs _env [] cont = cont
262 pushSimplifiedArgs env (arg:args) cont = ApplyTo Simplified arg env (pushSimplifiedArgs env args cont)
263 -- The env has an empty SubstEnv
265 dropArgs :: Int -> SimplCont -> SimplCont
266 dropArgs 0 cont = cont
267 dropArgs n (ApplyTo _ _ _ cont) = dropArgs (n-1) cont
268 dropArgs n other = pprPanic "dropArgs" (ppr n <+> ppr other)
272 Note [Interesting call context]
273 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
274 We want to avoid inlining an expression where there can't possibly be
275 any gain, such as in an argument position. Hence, if the continuation
276 is interesting (eg. a case scrutinee, application etc.) then we
277 inline, otherwise we don't.
279 Previously some_benefit used to return True only if the variable was
280 applied to some value arguments. This didn't work:
282 let x = _coerce_ (T Int) Int (I# 3) in
283 case _coerce_ Int (T Int) x of
286 we want to inline x, but can't see that it's a constructor in a case
287 scrutinee position, and some_benefit is False.
291 dMonadST = _/\_ t -> :Monad (g1 _@_ t, g2 _@_ t, g3 _@_ t)
293 .... case dMonadST _@_ x0 of (a,b,c) -> ....
295 we'd really like to inline dMonadST here, but we *don't* want to
296 inline if the case expression is just
298 case x of y { DEFAULT -> ... }
300 since we can just eliminate this case instead (x is in WHNF). Similar
301 applies when x is bound to a lambda expression. Hence
302 contIsInteresting looks for case expressions with just a single
307 interestingCallContext :: SimplCont -> CallCtxt
308 -- See Note [Interesting call context]
309 interestingCallContext cont
312 interesting (Select _ bndr _ _ _)
313 | isDeadBinder bndr = CaseCtxt
314 | otherwise = ArgCtxt False -- If the binder is used, this
315 -- is like a strict let
316 -- See Note [RHS of lets] in CoreUnfold
318 interesting (ApplyTo _ arg _ cont)
319 | isTypeArg arg = interesting cont
320 | otherwise = ValAppCtxt -- Can happen if we have (f Int |> co) y
321 -- If f has an INLINE prag we need to give it some
322 -- motivation to inline. See Note [Cast then apply]
325 interesting (StrictArg _ cci _) = cci
326 interesting (StrictBind {}) = BoringCtxt
327 interesting (Stop cci) = cci
328 interesting (CoerceIt _ cont) = interesting cont
329 -- If this call is the arg of a strict function, the context
330 -- is a bit interesting. If we inline here, we may get useful
331 -- evaluation information to avoid repeated evals: e.g.
333 -- Here the contIsInteresting makes the '*' keener to inline,
334 -- which in turn exposes a constructor which makes the '+' inline.
335 -- Assuming that +,* aren't small enough to inline regardless.
337 -- It's also very important to inline in a strict context for things
340 -- Here, the context of (f x) is strict, and if f's unfolding is
341 -- a build it's *great* to inline it here. So we must ensure that
342 -- the context for (f x) is not totally uninteresting.
347 -> [CoreRule] -- Rules for function
348 -> Int -- Number of value args
349 -> SimplCont -- Context of the call
352 mkArgInfo fun rules n_val_args call_cont
353 | n_val_args < idArity fun -- Note [Unsaturated functions]
354 = ArgInfo { ai_fun = fun, ai_args = [], ai_rules = rules
356 , ai_strs = vanilla_stricts
357 , ai_discs = vanilla_discounts }
359 = ArgInfo { ai_fun = fun, ai_args = [], ai_rules = rules
360 , ai_encl = interestingArgContext rules call_cont
361 , ai_strs = add_type_str (idType fun) arg_stricts
362 , ai_discs = arg_discounts }
364 vanilla_discounts, arg_discounts :: [Int]
365 vanilla_discounts = repeat 0
366 arg_discounts = case idUnfolding fun of
367 CoreUnfolding {uf_guidance = UnfIfGoodArgs {ug_args = discounts}}
368 -> discounts ++ vanilla_discounts
369 _ -> vanilla_discounts
371 vanilla_stricts, arg_stricts :: [Bool]
372 vanilla_stricts = repeat False
375 = case splitStrictSig (idStrictness fun) of
376 (demands, result_info)
377 | not (demands `lengthExceeds` n_val_args)
378 -> -- Enough args, use the strictness given.
379 -- For bottoming functions we used to pretend that the arg
380 -- is lazy, so that we don't treat the arg as an
381 -- interesting context. This avoids substituting
382 -- top-level bindings for (say) strings into
383 -- calls to error. But now we are more careful about
384 -- inlining lone variables, so its ok (see SimplUtils.analyseCont)
385 if isBotRes result_info then
386 map isStrictDmd demands -- Finite => result is bottom
388 map isStrictDmd demands ++ vanilla_stricts
390 -> WARN( True, text "More demands than arity" <+> ppr fun <+> ppr (idArity fun)
391 <+> ppr n_val_args <+> ppr demands )
392 vanilla_stricts -- Not enough args, or no strictness
394 add_type_str :: Type -> [Bool] -> [Bool]
395 -- If the function arg types are strict, record that in the 'strictness bits'
396 -- No need to instantiate because unboxed types (which dominate the strict
397 -- types) can't instantiate type variables.
398 -- add_type_str is done repeatedly (for each call); might be better
399 -- once-for-all in the function
400 -- But beware primops/datacons with no strictness
401 add_type_str _ [] = []
402 add_type_str fun_ty strs -- Look through foralls
403 | Just (_, fun_ty') <- splitForAllTy_maybe fun_ty -- Includes coercions
404 = add_type_str fun_ty' strs
405 add_type_str fun_ty (str:strs) -- Add strict-type info
406 | Just (arg_ty, fun_ty') <- splitFunTy_maybe fun_ty
407 = (str || isStrictType arg_ty) : add_type_str fun_ty' strs
411 {- Note [Unsaturated functions]
412 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
413 Consider (test eyeball/inline4)
416 where f has arity 2. Then we do not want to inline 'x', because
417 it'll just be floated out again. Even if f has lots of discounts
418 on its first argument -- it must be saturated for these to kick in
421 interestingArgContext :: [CoreRule] -> SimplCont -> Bool
422 -- If the argument has form (f x y), where x,y are boring,
423 -- and f is marked INLINE, then we don't want to inline f.
424 -- But if the context of the argument is
426 -- where g has rules, then we *do* want to inline f, in case it
427 -- exposes a rule that might fire. Similarly, if the context is
429 -- where h has rules, then we do want to inline f; hence the
430 -- call_cont argument to interestingArgContext
432 -- The ai-rules flag makes this happen; if it's
433 -- set, the inliner gets just enough keener to inline f
434 -- regardless of how boring f's arguments are, if it's marked INLINE
436 -- The alternative would be to *always* inline an INLINE function,
437 -- regardless of how boring its context is; but that seems overkill
438 -- For example, it'd mean that wrapper functions were always inlined
439 interestingArgContext rules call_cont
440 = notNull rules || enclosing_fn_has_rules
442 enclosing_fn_has_rules = go call_cont
444 go (Select {}) = False
445 go (ApplyTo {}) = False
446 go (StrictArg _ cci _) = interesting cci
447 go (StrictBind {}) = False -- ??
448 go (CoerceIt _ c) = go c
449 go (Stop cci) = interesting cci
451 interesting (ArgCtxt rules) = rules
452 interesting _ = False
456 %************************************************************************
460 %************************************************************************
462 The SimplifierMode controls several switches; see its definition in
464 sm_rules :: Bool -- Whether RULES are enabled
465 sm_inline :: Bool -- Whether inlining is enabled
466 sm_case_case :: Bool -- Whether case-of-case is enabled
467 sm_eta_expand :: Bool -- Whether eta-expansion is enabled
470 simplEnvForGHCi :: SimplEnv
471 simplEnvForGHCi = mkSimplEnv $
472 SimplMode { sm_names = ["GHCi"]
473 , sm_phase = InitialPhase
474 , sm_rules = True, sm_inline = False
475 , sm_eta_expand = False, sm_case_case = True }
476 -- Do not do any inlining, in case we expose some unboxed
477 -- tuple stuff that confuses the bytecode interpreter
479 updModeForInlineRules :: Activation -> SimplifierMode -> SimplifierMode
480 -- See Note [Simplifying inside InlineRules]
481 updModeForInlineRules inline_rule_act current_mode
482 = current_mode { sm_phase = phaseFromActivation inline_rule_act
485 , sm_eta_expand = False }
487 phaseFromActivation (ActiveAfter n) = Phase n
488 phaseFromActivation _ = InitialPhase
491 Note [Inlining in gentle mode]
492 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
493 Something is inlined if
494 (i) the sm_inline flag is on, AND
495 (ii) the thing has an INLINE pragma, AND
496 (iii) the thing is inlinable in the earliest phase.
498 Example of why (iii) is important:
499 {-# INLINE [~1] g #-}
505 If we were to inline g into f's inlining, then an importing module would
507 f e --> g (g e) ---> RULE fires
508 because the InlineRule for f has had g inlined into it.
510 On the other hand, it is bad not to do ANY inlining into an
511 InlineRule, because then recursive knots in instance declarations
512 don't get unravelled.
514 However, *sometimes* SimplGently must do no call-site inlining at all
515 (hence sm_inline = False). Before full laziness we must be careful
516 not to inline wrappers, because doing so inhibits floating
517 e.g. ...(case f x of ...)...
518 ==> ...(case (case x of I# x# -> fw x#) of ...)...
519 ==> ...(case x of I# x# -> case fw x# of ...)...
520 and now the redex (f x) isn't floatable any more.
522 The no-inlining thing is also important for Template Haskell. You might be
523 compiling in one-shot mode with -O2; but when TH compiles a splice before
524 running it, we don't want to use -O2. Indeed, we don't want to inline
525 anything, because the byte-code interpreter might get confused about
526 unboxed tuples and suchlike.
528 Note [Simplifying inside InlineRules]
529 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
530 We must take care with simplification inside InlineRules (which come from
533 First, consider the following example
538 in ...g...g...g...g...g...
539 Now, if that's the ONLY occurrence of f, it might be inlined inside g,
540 and thence copied multiple times when g is inlined. HENCE we treat
541 any occurrence in an InlineRule as a multiple occurrence, not a single
542 one; see OccurAnal.addRuleUsage.
544 Second, we do want *do* to some modest rules/inlining stuff in InlineRules,
545 partly to eliminate senseless crap, and partly to break the recursive knots
546 generated by instance declarations.
548 However, suppose we have
549 {-# INLINE <act> f #-}
551 meaning "inline f in phases p where activation <act>(p) holds".
552 Then what inlinings/rules can we apply to the copy of <rhs> captured in
553 f's InlineRule? Our model is that literally <rhs> is substituted for
554 f when it is inlined. So our conservative plan (implemented by
555 updModeForInlineRules) is this:
557 -------------------------------------------------------------
558 When simplifying the RHS of an InlineRule, set the phase to the
559 phase in which the InlineRule first becomes active
560 -------------------------------------------------------------
564 a) Rules/inlinings that *cease* being active before p will
565 not apply to the InlineRule rhs, consistent with it being
566 inlined in its *original* form in phase p.
568 b) Rules/inlinings that only become active *after* p will
569 not apply to the InlineRule rhs, again to be consistent with
570 inlining the *original* rhs in phase p.
576 {-# NOINLINE [1] g #-}
579 {-# RULE h g = ... #-}
580 Here we must not inline g into f's RHS, even when we get to phase 0,
581 because when f is later inlined into some other module we want the
589 and suppose that there are auto-generated specialisations and a strictness
590 wrapper for g. The specialisations get activation AlwaysActive, and the
591 strictness wrapper get activation (ActiveAfter 0). So the strictness
592 wrepper fails the test and won't be inlined into f's InlineRule. That
593 means f can inline, expose the specialised call to g, so the specialisation
596 A note about wrappers
597 ~~~~~~~~~~~~~~~~~~~~~
598 It's also important not to inline a worker back into a wrapper.
600 wraper = inline_me (\x -> ...worker... )
601 Normally, the inline_me prevents the worker getting inlined into
602 the wrapper (initially, the worker's only call site!). But,
603 if the wrapper is sure to be called, the strictness analyser will
604 mark it 'demanded', so when the RHS is simplified, it'll get an ArgOf
608 activeUnfolding :: SimplEnv -> Id -> Bool
610 | not (sm_inline mode) = active_unfolding_minimal
611 | otherwise = case sm_phase mode of
612 InitialPhase -> active_unfolding_gentle
613 Phase n -> active_unfolding n
617 getUnfoldingInRuleMatch :: SimplEnv -> IdUnfoldingFun
618 -- When matching in RULE, we want to "look through" an unfolding
619 -- (to see a constructor) if *rules* are on, even if *inlinings*
620 -- are not. A notable example is DFuns, which really we want to
621 -- match in rules like (op dfun) in gentle mode. Another example
622 -- is 'otherwise' which we want exprIsConApp_maybe to be able to
624 getUnfoldingInRuleMatch env id
625 | unf_is_active = idUnfolding id
626 | otherwise = NoUnfolding
630 | not (sm_rules mode) = active_unfolding_minimal id
631 | otherwise = isActive (sm_phase mode) (idInlineActivation id)
633 active_unfolding_minimal :: Id -> Bool
634 -- Compuslory unfoldings only
635 -- Ignore SimplGently, because we want to inline regardless;
636 -- the Id has no top-level binding at all
638 -- NB: we used to have a second exception, for data con wrappers.
639 -- On the grounds that we use gentle mode for rule LHSs, and
640 -- they match better when data con wrappers are inlined.
641 -- But that only really applies to the trivial wrappers (like (:)),
642 -- and they are now constructed as Compulsory unfoldings (in MkId)
643 -- so they'll happen anyway.
644 active_unfolding_minimal id = isCompulsoryUnfolding (realIdUnfolding id)
646 active_unfolding :: PhaseNum -> Id -> Bool
647 active_unfolding n id = isActiveIn n (idInlineActivation id)
649 active_unfolding_gentle :: Id -> Bool
650 -- Anything that is early-active
651 -- See Note [Gentle mode]
652 active_unfolding_gentle id
653 = isInlinePragma prag
654 && isEarlyActive (inlinePragmaActivation prag)
655 -- NB: wrappers are not early-active
657 prag = idInlinePragma id
659 ----------------------
660 activeRule :: DynFlags -> SimplEnv -> Maybe (Activation -> Bool)
661 -- Nothing => No rules at all
662 activeRule _dflags env
663 | not (sm_rules mode) = Nothing -- Rewriting is off
664 | otherwise = Just (isActive (sm_phase mode))
671 %************************************************************************
673 preInlineUnconditionally
675 %************************************************************************
677 preInlineUnconditionally
678 ~~~~~~~~~~~~~~~~~~~~~~~~
679 @preInlineUnconditionally@ examines a bndr to see if it is used just
680 once in a completely safe way, so that it is safe to discard the
681 binding inline its RHS at the (unique) usage site, REGARDLESS of how
682 big the RHS might be. If this is the case we don't simplify the RHS
683 first, but just inline it un-simplified.
685 This is much better than first simplifying a perhaps-huge RHS and then
686 inlining and re-simplifying it. Indeed, it can be at least quadratically
695 We may end up simplifying e1 N times, e2 N-1 times, e3 N-3 times etc.
696 This can happen with cascades of functions too:
703 THE MAIN INVARIANT is this:
705 ---- preInlineUnconditionally invariant -----
706 IF preInlineUnconditionally chooses to inline x = <rhs>
707 THEN doing the inlining should not change the occurrence
708 info for the free vars of <rhs>
709 ----------------------------------------------
711 For example, it's tempting to look at trivial binding like
713 and inline it unconditionally. But suppose x is used many times,
714 but this is the unique occurrence of y. Then inlining x would change
715 y's occurrence info, which breaks the invariant. It matters: y
716 might have a BIG rhs, which will now be dup'd at every occurrenc of x.
719 Even RHSs labelled InlineMe aren't caught here, because there might be
720 no benefit from inlining at the call site.
722 [Sept 01] Don't unconditionally inline a top-level thing, because that
723 can simply make a static thing into something built dynamically. E.g.
727 [Remember that we treat \s as a one-shot lambda.] No point in
728 inlining x unless there is something interesting about the call site.
730 But watch out: if you aren't careful, some useful foldr/build fusion
731 can be lost (most notably in spectral/hartel/parstof) because the
732 foldr didn't see the build. Doing the dynamic allocation isn't a big
733 deal, in fact, but losing the fusion can be. But the right thing here
734 seems to be to do a callSiteInline based on the fact that there is
735 something interesting about the call site (it's strict). Hmm. That
738 Conclusion: inline top level things gaily until Phase 0 (the last
739 phase), at which point don't.
741 Note [pre/postInlineUnconditionally in gentle mode]
742 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
743 Even in gentle mode we want to do preInlineUnconditionally. The
744 reason is that too little clean-up happens if you don't inline
745 use-once things. Also a bit of inlining is *good* for full laziness;
746 it can expose constant sub-expressions. Example in
747 spectral/mandel/Mandel.hs, where the mandelset function gets a useful
748 let-float if you inline windowToViewport
750 However, as usual for Gentle mode, do not inline things that are
751 inactive in the intial stages. See Note [Gentle mode].
753 Note [InlineRule and preInlineUnconditionally]
754 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
755 Surprisingly, do not pre-inline-unconditionally Ids with INLINE pragmas!
765 ...fInt...fInt...fInt...
767 Here f occurs just once, in the RHS of f1. But if we inline it there
768 we'll lose the opportunity to inline at each of fInt's call sites.
769 The INLINE pragma will only inline when the application is saturated
770 for exactly this reason; and we don't want PreInlineUnconditionally
771 to second-guess it. A live example is Trac #3736.
772 c.f. Note [InlineRule and postInlineUnconditionally]
774 Note [Top-level botomming Ids]
775 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
776 Don't inline top-level Ids that are bottoming, even if they are used just
777 once, because FloatOut has gone to some trouble to extract them out.
778 Inlining them won't make the program run faster!
781 preInlineUnconditionally :: SimplEnv -> TopLevelFlag -> InId -> InExpr -> Bool
782 preInlineUnconditionally env top_lvl bndr rhs
784 | isStableUnfolding (idUnfolding bndr) = False -- Note [InlineRule and preInlineUnconditionally]
785 | isTopLevel top_lvl && isBottomingId bndr = False -- Note [Top-level bottoming Ids]
786 | opt_SimplNoPreInlining = False
787 | otherwise = case idOccInfo bndr of
788 IAmDead -> True -- Happens in ((\x.1) v)
789 OneOcc in_lam True int_cxt -> try_once in_lam int_cxt
793 active = isActive (sm_phase mode) act
794 -- See Note [pre/postInlineUnconditionally in gentle mode]
795 act = idInlineActivation bndr
796 try_once in_lam int_cxt -- There's one textual occurrence
797 | not in_lam = isNotTopLevel top_lvl || early_phase
798 | otherwise = int_cxt && canInlineInLam rhs
800 -- Be very careful before inlining inside a lambda, because (a) we must not
801 -- invalidate occurrence information, and (b) we want to avoid pushing a
802 -- single allocation (here) into multiple allocations (inside lambda).
803 -- Inlining a *function* with a single *saturated* call would be ok, mind you.
804 -- || (if is_cheap && not (canInlineInLam rhs) then pprTrace "preinline" (ppr bndr <+> ppr rhs) ok else ok)
806 -- is_cheap = exprIsCheap rhs
807 -- ok = is_cheap && int_cxt
809 -- int_cxt The context isn't totally boring
810 -- E.g. let f = \ab.BIG in \y. map f xs
811 -- Don't want to substitute for f, because then we allocate
812 -- its closure every time the \y is called
813 -- But: let f = \ab.BIG in \y. map (f y) xs
814 -- Now we do want to substitute for f, even though it's not
815 -- saturated, because we're going to allocate a closure for
816 -- (f y) every time round the loop anyhow.
818 -- canInlineInLam => free vars of rhs are (Once in_lam) or Many,
819 -- so substituting rhs inside a lambda doesn't change the occ info.
820 -- Sadly, not quite the same as exprIsHNF.
821 canInlineInLam (Lit _) = True
822 canInlineInLam (Lam b e) = isRuntimeVar b || canInlineInLam e
823 canInlineInLam (Note _ e) = canInlineInLam e
824 canInlineInLam _ = False
826 early_phase = case sm_phase mode of
829 -- If we don't have this early_phase test, consider
830 -- x = length [1,2,3]
831 -- The full laziness pass carefully floats all the cons cells to
832 -- top level, and preInlineUnconditionally floats them all back in.
833 -- Result is (a) static allocation replaced by dynamic allocation
834 -- (b) many simplifier iterations because this tickles
835 -- a related problem; only one inlining per pass
837 -- On the other hand, I have seen cases where top-level fusion is
838 -- lost if we don't inline top level thing (e.g. string constants)
839 -- Hence the test for phase zero (which is the phase for all the final
840 -- simplifications). Until phase zero we take no special notice of
841 -- top level things, but then we become more leery about inlining
846 %************************************************************************
848 postInlineUnconditionally
850 %************************************************************************
852 postInlineUnconditionally
853 ~~~~~~~~~~~~~~~~~~~~~~~~~
854 @postInlineUnconditionally@ decides whether to unconditionally inline
855 a thing based on the form of its RHS; in particular if it has a
856 trivial RHS. If so, we can inline and discard the binding altogether.
858 NB: a loop breaker has must_keep_binding = True and non-loop-breakers
859 only have *forward* references Hence, it's safe to discard the binding
861 NOTE: This isn't our last opportunity to inline. We're at the binding
862 site right now, and we'll get another opportunity when we get to the
865 Note that we do this unconditional inlining only for trival RHSs.
866 Don't inline even WHNFs inside lambdas; doing so may simply increase
867 allocation when the function is called. This isn't the last chance; see
870 NB: Even inline pragmas (e.g. IMustBeINLINEd) are ignored here Why?
871 Because we don't even want to inline them into the RHS of constructor
872 arguments. See NOTE above
874 NB: At one time even NOINLINE was ignored here: if the rhs is trivial
875 it's best to inline it anyway. We often get a=E; b=a from desugaring,
876 with both a and b marked NOINLINE. But that seems incompatible with
877 our new view that inlining is like a RULE, so I'm sticking to the 'active'
881 postInlineUnconditionally
882 :: SimplEnv -> TopLevelFlag
883 -> OutId -- The binder (an InId would be fine too)
884 -> OccInfo -- From the InId
888 postInlineUnconditionally env top_lvl bndr occ_info rhs unfolding
890 | isLoopBreaker occ_info = False -- If it's a loop-breaker of any kind, don't inline
891 -- because it might be referred to "earlier"
892 | isExportedId bndr = False
893 | isStableUnfolding unfolding = False -- Note [InlineRule and postInlineUnconditionally]
894 | exprIsTrivial rhs = True
895 | isTopLevel top_lvl = False -- Note [Top level and postInlineUnconditionally]
898 -- The point of examining occ_info here is that for *non-values*
899 -- that occur outside a lambda, the call-site inliner won't have
900 -- a chance (becuase it doesn't know that the thing
901 -- only occurs once). The pre-inliner won't have gotten
902 -- it either, if the thing occurs in more than one branch
903 -- So the main target is things like
906 -- True -> case x of ...
907 -- False -> case x of ...
908 -- This is very important in practice; e.g. wheel-seive1 doubles
909 -- in allocation if you miss this out
910 OneOcc in_lam _one_br int_cxt -- OneOcc => no code-duplication issue
911 -> smallEnoughToInline unfolding -- Small enough to dup
912 -- ToDo: consider discount on smallEnoughToInline if int_cxt is true
914 -- NB: Do NOT inline arbitrarily big things, even if one_br is True
915 -- Reason: doing so risks exponential behaviour. We simplify a big
916 -- expression, inline it, and simplify it again. But if the
917 -- very same thing happens in the big expression, we get
919 -- PRINCIPLE: when we've already simplified an expression once,
920 -- make sure that we only inline it if it's reasonably small.
923 -- Outside a lambda, we want to be reasonably aggressive
924 -- about inlining into multiple branches of case
925 -- e.g. let x = <non-value>
926 -- in case y of { C1 -> ..x..; C2 -> ..x..; C3 -> ... }
927 -- Inlining can be a big win if C3 is the hot-spot, even if
928 -- the uses in C1, C2 are not 'interesting'
929 -- An example that gets worse if you add int_cxt here is 'clausify'
931 (isCheapUnfolding unfolding && int_cxt))
932 -- isCheap => acceptable work duplication; in_lam may be true
933 -- int_cxt to prevent us inlining inside a lambda without some
934 -- good reason. See the notes on int_cxt in preInlineUnconditionally
936 IAmDead -> True -- This happens; for example, the case_bndr during case of
937 -- known constructor: case (a,b) of x { (p,q) -> ... }
938 -- Here x isn't mentioned in the RHS, so we don't want to
939 -- create the (dead) let-binding let x = (a,b) in ...
943 -- Here's an example that we don't handle well:
944 -- let f = if b then Left (\x.BIG) else Right (\y.BIG)
945 -- in \y. ....case f of {...} ....
946 -- Here f is used just once, and duplicating the case work is fine (exprIsCheap).
948 -- - We can't preInlineUnconditionally because that woud invalidate
949 -- the occ info for b.
950 -- - We can't postInlineUnconditionally because the RHS is big, and
951 -- that risks exponential behaviour
952 -- - We can't call-site inline, because the rhs is big
956 active = isActive (sm_phase (getMode env)) (idInlineActivation bndr)
957 -- See Note [pre/postInlineUnconditionally in gentle mode]
960 Note [Top level and postInlineUnconditionally]
961 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
962 We don't do postInlineUnconditionally for top-level things (exept ones that
964 * There is no point, because the main goal is to get rid of local
965 bindings used in multiple case branches.
966 * Doing so will inline top-level error expressions that have been
967 carefully floated out by FloatOut. More generally, it might
968 replace static allocation with dynamic.
970 Note [InlineRule and postInlineUnconditionally]
971 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
972 Do not do postInlineUnconditionally if the Id has an InlineRule, otherwise
973 we lose the unfolding. Example
975 -- f has InlineRule with rhs (e |> co)
979 Then there's a danger we'll optimise to
984 and now postInlineUnconditionally, losing the InlineRule on f. Now f'
985 won't inline because 'e' is too big.
987 c.f. Note [InlineRule and preInlineUnconditionally]
990 %************************************************************************
994 %************************************************************************
997 mkLam :: SimplEnv -> [OutBndr] -> OutExpr -> SimplM OutExpr
998 -- mkLam tries three things
999 -- a) eta reduction, if that gives a trivial expression
1000 -- b) eta expansion [only if there are some value lambdas]
1004 mkLam _env bndrs body
1005 = do { dflags <- getDOptsSmpl
1006 ; mkLam' dflags bndrs body }
1008 mkLam' :: DynFlags -> [OutBndr] -> OutExpr -> SimplM OutExpr
1009 mkLam' dflags bndrs (Cast body co)
1010 | not (any bad bndrs)
1011 -- Note [Casts and lambdas]
1012 = do { lam <- mkLam' dflags bndrs body
1013 ; return (mkCoerce (mkPiTypes bndrs co) lam) }
1015 co_vars = tyVarsOfType co
1016 bad bndr = isCoVar bndr && bndr `elemVarSet` co_vars
1018 mkLam' dflags bndrs body@(Lam {})
1019 = mkLam' dflags (bndrs ++ bndrs1) body1
1021 (bndrs1, body1) = collectBinders body
1023 mkLam' dflags bndrs body
1024 | dopt Opt_DoEtaReduction dflags
1025 , Just etad_lam <- tryEtaReduce bndrs body
1026 = do { tick (EtaReduction (head bndrs))
1030 = return (mkLams bndrs body)
1034 Note [Casts and lambdas]
1035 ~~~~~~~~~~~~~~~~~~~~~~~~
1037 (\x. (\y. e) `cast` g1) `cast` g2
1038 There is a danger here that the two lambdas look separated, and the
1039 full laziness pass might float an expression to between the two.
1041 So this equation in mkLam' floats the g1 out, thus:
1042 (\x. e `cast` g1) --> (\x.e) `cast` (tx -> g1)
1045 In general, this floats casts outside lambdas, where (I hope) they
1046 might meet and cancel with some other cast:
1047 \x. e `cast` co ===> (\x. e) `cast` (tx -> co)
1048 /\a. e `cast` co ===> (/\a. e) `cast` (/\a. co)
1049 /\g. e `cast` co ===> (/\g. e) `cast` (/\g. co)
1050 (if not (g `in` co))
1052 Notice that it works regardless of 'e'. Originally it worked only
1053 if 'e' was itself a lambda, but in some cases that resulted in
1054 fruitless iteration in the simplifier. A good example was when
1055 compiling Text.ParserCombinators.ReadPrec, where we had a definition
1056 like (\x. Get `cast` g)
1057 where Get is a constructor with nonzero arity. Then mkLam eta-expanded
1058 the Get, and the next iteration eta-reduced it, and then eta-expanded
1061 Note also the side condition for the case of coercion binders.
1062 It does not make sense to transform
1063 /\g. e `cast` g ==> (/\g.e) `cast` (/\g.g)
1064 because the latter is not well-kinded.
1066 %************************************************************************
1070 %************************************************************************
1073 tryEtaExpand :: SimplEnv -> OutId -> OutExpr -> SimplM (Arity, OutExpr)
1074 -- See Note [Eta-expanding at let bindings]
1075 tryEtaExpand env bndr rhs
1076 = do { dflags <- getDOptsSmpl
1077 ; (new_arity, new_rhs) <- try_expand dflags
1079 ; WARN( new_arity < old_arity || new_arity < _dmd_arity,
1080 (ptext (sLit "Arity decrease:") <+> (ppr bndr <+> ppr old_arity
1081 <+> ppr new_arity <+> ppr _dmd_arity) $$ ppr new_rhs) )
1082 -- Note [Arity decrease]
1083 return (new_arity, new_rhs) }
1086 | sm_eta_expand (getMode env) -- Provided eta-expansion is on
1087 , not (exprIsTrivial rhs)
1088 , let new_arity = exprEtaExpandArity dflags rhs
1089 , new_arity > rhs_arity
1090 = do { tick (EtaExpansion bndr)
1091 ; return (new_arity, etaExpand new_arity rhs) }
1093 = return (rhs_arity, rhs)
1095 rhs_arity = exprArity rhs
1096 old_arity = idArity bndr
1097 _dmd_arity = length $ fst $ splitStrictSig $ idStrictness bndr
1100 Note [Eta-expanding at let bindings]
1101 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
1102 We now eta expand at let-bindings, which is where the payoff
1105 One useful consequence is this example:
1106 genMap :: C a => ...
1107 {-# INLINE genMap #-}
1111 {-# INLINE myMap #-}
1114 Notice that 'genMap' should only inline if applied to two arguments.
1115 In the InlineRule for myMap we'll have the unfolding
1116 (\d -> genMap Int (..d..))
1117 We do not want to eta-expand to
1118 (\d f xs -> genMap Int (..d..) f xs)
1119 because then 'genMap' will inline, and it really shouldn't: at least
1120 as far as the programmer is concerned, it's not applied to two
1124 %************************************************************************
1126 \subsection{Floating lets out of big lambdas}
1128 %************************************************************************
1130 Note [Floating and type abstraction]
1131 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
1134 We'd like to float this to
1137 x = /\a. C (y1 a) (y2 a)
1138 for the usual reasons: we want to inline x rather vigorously.
1140 You may think that this kind of thing is rare. But in some programs it is
1141 common. For example, if you do closure conversion you might get:
1143 data a :-> b = forall e. (e -> a -> b) :$ e
1145 f_cc :: forall a. a :-> a
1146 f_cc = /\a. (\e. id a) :$ ()
1148 Now we really want to inline that f_cc thing so that the
1149 construction of the closure goes away.
1151 So I have elaborated simplLazyBind to understand right-hand sides that look
1155 and treat them specially. The real work is done in SimplUtils.abstractFloats,
1156 but there is quite a bit of plumbing in simplLazyBind as well.
1158 The same transformation is good when there are lets in the body:
1160 /\abc -> let(rec) x = e in b
1162 let(rec) x' = /\abc -> let x = x' a b c in e
1164 /\abc -> let x = x' a b c in b
1166 This is good because it can turn things like:
1168 let f = /\a -> letrec g = ... g ... in g
1170 letrec g' = /\a -> ... g' a ...
1172 let f = /\ a -> g' a
1174 which is better. In effect, it means that big lambdas don't impede
1177 This optimisation is CRUCIAL in eliminating the junk introduced by
1178 desugaring mutually recursive definitions. Don't eliminate it lightly!
1180 [May 1999] If we do this transformation *regardless* then we can
1181 end up with some pretty silly stuff. For example,
1184 st = /\ s -> let { x1=r1 ; x2=r2 } in ...
1189 st = /\s -> ...[y1 s/x1, y2 s/x2]
1192 Unless the "..." is a WHNF there is really no point in doing this.
1193 Indeed it can make things worse. Suppose x1 is used strictly,
1196 x1* = case f y of { (a,b) -> e }
1198 If we abstract this wrt the tyvar we then can't do the case inline
1199 as we would normally do.
1201 That's why the whole transformation is part of the same process that
1202 floats let-bindings and constructor arguments out of RHSs. In particular,
1203 it is guarded by the doFloatFromRhs call in simplLazyBind.
1207 abstractFloats :: [OutTyVar] -> SimplEnv -> OutExpr -> SimplM ([OutBind], OutExpr)
1208 abstractFloats main_tvs body_env body
1209 = ASSERT( notNull body_floats )
1210 do { (subst, float_binds) <- mapAccumLM abstract empty_subst body_floats
1211 ; return (float_binds, CoreSubst.substExpr (text "abstract_floats1") subst body) }
1213 main_tv_set = mkVarSet main_tvs
1214 body_floats = getFloats body_env
1215 empty_subst = CoreSubst.mkEmptySubst (seInScope body_env)
1217 abstract :: CoreSubst.Subst -> OutBind -> SimplM (CoreSubst.Subst, OutBind)
1218 abstract subst (NonRec id rhs)
1219 = do { (poly_id, poly_app) <- mk_poly tvs_here id
1220 ; let poly_rhs = mkLams tvs_here rhs'
1221 subst' = CoreSubst.extendIdSubst subst id poly_app
1222 ; return (subst', (NonRec poly_id poly_rhs)) }
1224 rhs' = CoreSubst.substExpr (text "abstract_floats2") subst rhs
1225 tvs_here | any isCoVar main_tvs = main_tvs -- Note [Abstract over coercions]
1227 = varSetElems (main_tv_set `intersectVarSet` exprSomeFreeVars isTyCoVar rhs')
1229 -- Abstract only over the type variables free in the rhs
1230 -- wrt which the new binding is abstracted. But the naive
1231 -- approach of abstract wrt the tyvars free in the Id's type
1233 -- /\ a b -> let t :: (a,b) = (e1, e2)
1236 -- Here, b isn't free in x's type, but we must nevertheless
1237 -- abstract wrt b as well, because t's type mentions b.
1238 -- Since t is floated too, we'd end up with the bogus:
1239 -- poly_t = /\ a b -> (e1, e2)
1240 -- poly_x = /\ a -> fst (poly_t a *b*)
1241 -- So for now we adopt the even more naive approach of
1242 -- abstracting wrt *all* the tyvars. We'll see if that
1243 -- gives rise to problems. SLPJ June 98
1245 abstract subst (Rec prs)
1246 = do { (poly_ids, poly_apps) <- mapAndUnzipM (mk_poly tvs_here) ids
1247 ; let subst' = CoreSubst.extendSubstList subst (ids `zip` poly_apps)
1248 poly_rhss = [mkLams tvs_here (CoreSubst.substExpr (text "abstract_floats3") subst' rhs)
1250 ; return (subst', Rec (poly_ids `zip` poly_rhss)) }
1252 (ids,rhss) = unzip prs
1253 -- For a recursive group, it's a bit of a pain to work out the minimal
1254 -- set of tyvars over which to abstract:
1255 -- /\ a b c. let x = ...a... in
1256 -- letrec { p = ...x...q...
1257 -- q = .....p...b... } in
1259 -- Since 'x' is abstracted over 'a', the {p,q} group must be abstracted
1260 -- over 'a' (because x is replaced by (poly_x a)) as well as 'b'.
1261 -- Since it's a pain, we just use the whole set, which is always safe
1263 -- If you ever want to be more selective, remember this bizarre case too:
1265 -- Here, we must abstract 'x' over 'a'.
1268 mk_poly tvs_here var
1269 = do { uniq <- getUniqueM
1270 ; let poly_name = setNameUnique (idName var) uniq -- Keep same name
1271 poly_ty = mkForAllTys tvs_here (idType var) -- But new type of course
1272 poly_id = transferPolyIdInfo var tvs_here $ -- Note [transferPolyIdInfo] in Id.lhs
1273 mkLocalId poly_name poly_ty
1274 ; return (poly_id, mkTyApps (Var poly_id) (mkTyVarTys tvs_here)) }
1275 -- In the olden days, it was crucial to copy the occInfo of the original var,
1276 -- because we were looking at occurrence-analysed but as yet unsimplified code!
1277 -- In particular, we mustn't lose the loop breakers. BUT NOW we are looking
1278 -- at already simplified code, so it doesn't matter
1280 -- It's even right to retain single-occurrence or dead-var info:
1281 -- Suppose we started with /\a -> let x = E in B
1282 -- where x occurs once in B. Then we transform to:
1283 -- let x' = /\a -> E in /\a -> let x* = x' a in B
1284 -- where x* has an INLINE prag on it. Now, once x* is inlined,
1285 -- the occurrences of x' will be just the occurrences originally
1289 Note [Abstract over coercions]
1290 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
1291 If a coercion variable (g :: a ~ Int) is free in the RHS, then so is the
1292 type variable a. Rather than sort this mess out, we simply bale out and abstract
1293 wrt all the type variables if any of them are coercion variables.
1296 Historical note: if you use let-bindings instead of a substitution, beware of this:
1298 -- Suppose we start with:
1300 -- x = /\ a -> let g = G in E
1302 -- Then we'll float to get
1304 -- x = let poly_g = /\ a -> G
1305 -- in /\ a -> let g = poly_g a in E
1307 -- But now the occurrence analyser will see just one occurrence
1308 -- of poly_g, not inside a lambda, so the simplifier will
1309 -- PreInlineUnconditionally poly_g back into g! Badk to square 1!
1310 -- (I used to think that the "don't inline lone occurrences" stuff
1311 -- would stop this happening, but since it's the *only* occurrence,
1312 -- PreInlineUnconditionally kicks in first!)
1314 -- Solution: put an INLINE note on g's RHS, so that poly_g seems
1315 -- to appear many times. (NB: mkInlineMe eliminates
1316 -- such notes on trivial RHSs, so do it manually.)
1318 %************************************************************************
1322 %************************************************************************
1324 prepareAlts tries these things:
1326 1. Eliminate alternatives that cannot match, including the
1327 DEFAULT alternative.
1329 2. If the DEFAULT alternative can match only one possible constructor,
1330 then make that constructor explicit.
1332 case e of x { DEFAULT -> rhs }
1334 case e of x { (a,b) -> rhs }
1335 where the type is a single constructor type. This gives better code
1336 when rhs also scrutinises x or e.
1338 3. Returns a list of the constructors that cannot holds in the
1339 DEFAULT alternative (if there is one)
1341 Here "cannot match" includes knowledge from GADTs
1343 It's a good idea do do this stuff before simplifying the alternatives, to
1344 avoid simplifying alternatives we know can't happen, and to come up with
1345 the list of constructors that are handled, to put into the IdInfo of the
1346 case binder, for use when simplifying the alternatives.
1348 Eliminating the default alternative in (1) isn't so obvious, but it can
1351 data Colour = Red | Green | Blue
1360 DEFAULT -> [ case y of ... ]
1362 If we inline h into f, the default case of the inlined h can't happen.
1363 If we don't notice this, we may end up filtering out *all* the cases
1364 of the inner case y, which give us nowhere to go!
1367 prepareAlts :: OutExpr -> OutId -> [InAlt] -> SimplM ([AltCon], [InAlt])
1368 prepareAlts scrut case_bndr' alts
1369 = do { let (alts_wo_default, maybe_deflt) = findDefault alts
1370 alt_cons = [con | (con,_,_) <- alts_wo_default]
1371 imposs_deflt_cons = nub (imposs_cons ++ alt_cons)
1372 -- "imposs_deflt_cons" are handled
1373 -- EITHER by the context,
1374 -- OR by a non-DEFAULT branch in this case expression.
1376 ; default_alts <- prepareDefault case_bndr' mb_tc_app
1377 imposs_deflt_cons maybe_deflt
1379 ; let trimmed_alts = filterOut impossible_alt alts_wo_default
1380 merged_alts = mergeAlts trimmed_alts default_alts
1381 -- We need the mergeAlts in case the new default_alt
1382 -- has turned into a constructor alternative.
1383 -- The merge keeps the inner DEFAULT at the front, if there is one
1384 -- and interleaves the alternatives in the right order
1386 ; return (imposs_deflt_cons, merged_alts) }
1388 mb_tc_app = splitTyConApp_maybe (idType case_bndr')
1389 Just (_, inst_tys) = mb_tc_app
1391 imposs_cons = case scrut of
1392 Var v -> otherCons (idUnfolding v)
1395 impossible_alt :: CoreAlt -> Bool
1396 impossible_alt (con, _, _) | con `elem` imposs_cons = True
1397 impossible_alt (DataAlt con, _, _) = dataConCannotMatch inst_tys con
1398 impossible_alt _ = False
1401 prepareDefault :: OutId -- Case binder; need just for its type. Note that as an
1402 -- OutId, it has maximum information; this is important.
1403 -- Test simpl013 is an example
1404 -> Maybe (TyCon, [Type]) -- Type of scrutinee, decomposed
1405 -> [AltCon] -- These cons can't happen when matching the default
1406 -> Maybe InExpr -- Rhs
1407 -> SimplM [InAlt] -- Still unsimplified
1408 -- We use a list because it's what mergeAlts expects,
1410 --------- Fill in known constructor -----------
1411 prepareDefault case_bndr (Just (tycon, inst_tys)) imposs_cons (Just deflt_rhs)
1412 | -- This branch handles the case where we are
1413 -- scrutinisng an algebraic data type
1414 isAlgTyCon tycon -- It's a data type, tuple, or unboxed tuples.
1415 , not (isNewTyCon tycon) -- We can have a newtype, if we are just doing an eval:
1416 -- case x of { DEFAULT -> e }
1417 -- and we don't want to fill in a default for them!
1418 , Just all_cons <- tyConDataCons_maybe tycon
1419 , not (null all_cons)
1420 -- This is a tricky corner case. If the data type has no constructors,
1421 -- which GHC allows, then the case expression will have at most a default
1422 -- alternative. We don't want to eliminate that alternative, because the
1423 -- invariant is that there's always one alternative. It's more convenient
1425 -- case x of { DEFAULT -> e }
1426 -- as it is, rather than transform it to
1427 -- error "case cant match"
1428 -- which would be quite legitmate. But it's a really obscure corner, and
1429 -- not worth wasting code on.
1430 , let imposs_data_cons = [con | DataAlt con <- imposs_cons] -- We now know it's a data type
1431 impossible con = con `elem` imposs_data_cons || dataConCannotMatch inst_tys con
1432 = case filterOut impossible all_cons of
1433 [] -> return [] -- Eliminate the default alternative
1434 -- altogether if it can't match
1436 [con] -> -- It matches exactly one constructor, so fill it in
1437 do { tick (FillInCaseDefault case_bndr)
1439 ; let (ex_tvs, co_tvs, arg_ids) =
1440 dataConRepInstPat us con inst_tys
1441 ; return [(DataAlt con, ex_tvs ++ co_tvs ++ arg_ids, deflt_rhs)] }
1443 _ -> return [(DEFAULT, [], deflt_rhs)]
1445 | debugIsOn, isAlgTyCon tycon
1446 , null (tyConDataCons tycon)
1447 , not (isFamilyTyCon tycon || isAbstractTyCon tycon)
1448 -- Check for no data constructors
1449 -- This can legitimately happen for abstract types and type families,
1450 -- so don't report that
1451 = pprTrace "prepareDefault" (ppr case_bndr <+> ppr tycon)
1452 $ return [(DEFAULT, [], deflt_rhs)]
1454 --------- Catch-all cases -----------
1455 prepareDefault _case_bndr _bndr_ty _imposs_cons (Just deflt_rhs)
1456 = return [(DEFAULT, [], deflt_rhs)]
1458 prepareDefault _case_bndr _bndr_ty _imposs_cons Nothing
1459 = return [] -- No default branch
1464 %************************************************************************
1468 %************************************************************************
1470 mkCase tries these things
1472 1. Merge Nested Cases
1474 case e of b { ==> case e of b {
1475 p1 -> rhs1 p1 -> rhs1
1477 pm -> rhsm pm -> rhsm
1478 _ -> case b of b' { pn -> let b'=b in rhsn
1480 ... po -> let b'=b in rhso
1481 po -> rhso _ -> let b'=b in rhsd
1485 which merges two cases in one case when -- the default alternative of
1486 the outer case scrutises the same variable as the outer case. This
1487 transformation is called Case Merging. It avoids that the same
1488 variable is scrutinised multiple times.
1490 2. Eliminate Identity Case
1496 and similar friends.
1498 3. Merge identical alternatives.
1499 If several alternatives are identical, merge them into
1500 a single DEFAULT alternative. I've occasionally seen this
1501 making a big difference:
1503 case e of =====> case e of
1504 C _ -> f x D v -> ....v....
1505 D v -> ....v.... DEFAULT -> f x
1508 The point is that we merge common RHSs, at least for the DEFAULT case.
1509 [One could do something more elaborate but I've never seen it needed.]
1510 To avoid an expensive test, we just merge branches equal to the *first*
1511 alternative; this picks up the common cases
1512 a) all branches equal
1513 b) some branches equal to the DEFAULT (which occurs first)
1515 The case where Merge Identical Alternatives transformation showed up
1516 was like this (base/Foreign/C/Err/Error.lhs):
1522 where @is@ was something like
1524 p `is` n = p /= (-1) && p == n
1526 This gave rise to a horrible sequence of cases
1533 and similarly in cascade for all the join points!
1537 mkCase, mkCase1, mkCase2
1540 -> [OutAlt] -- Alternatives in standard (increasing) order
1543 --------------------------------------------------
1544 -- 1. Merge Nested Cases
1545 --------------------------------------------------
1547 mkCase dflags scrut outer_bndr ((DEFAULT, _, deflt_rhs) : outer_alts)
1548 | dopt Opt_CaseMerge dflags
1549 , Case (Var inner_scrut_var) inner_bndr _ inner_alts <- deflt_rhs
1550 , inner_scrut_var == outer_bndr
1551 = do { tick (CaseMerge outer_bndr)
1553 ; let wrap_alt (con, args, rhs) = ASSERT( outer_bndr `notElem` args )
1554 (con, args, wrap_rhs rhs)
1555 -- Simplifier's no-shadowing invariant should ensure
1556 -- that outer_bndr is not shadowed by the inner patterns
1557 wrap_rhs rhs = Let (NonRec inner_bndr (Var outer_bndr)) rhs
1558 -- The let is OK even for unboxed binders,
1560 wrapped_alts | isDeadBinder inner_bndr = inner_alts
1561 | otherwise = map wrap_alt inner_alts
1563 merged_alts = mergeAlts outer_alts wrapped_alts
1564 -- NB: mergeAlts gives priority to the left
1567 -- DEFAULT -> case x of
1570 -- When we merge, we must ensure that e1 takes
1571 -- precedence over e2 as the value for A!
1573 ; mkCase1 dflags scrut outer_bndr merged_alts
1575 -- Warning: don't call mkCase recursively!
1576 -- Firstly, there's no point, because inner alts have already had
1577 -- mkCase applied to them, so they won't have a case in their default
1578 -- Secondly, if you do, you get an infinite loop, because the bindCaseBndr
1579 -- in munge_rhs may put a case into the DEFAULT branch!
1581 mkCase dflags scrut bndr alts = mkCase1 dflags scrut bndr alts
1583 --------------------------------------------------
1584 -- 2. Eliminate Identity Case
1585 --------------------------------------------------
1587 mkCase1 _dflags scrut case_bndr alts -- Identity case
1588 | all identity_alt alts
1589 = do { tick (CaseIdentity case_bndr)
1590 ; return (re_cast scrut) }
1592 identity_alt (con, args, rhs) = check_eq con args (de_cast rhs)
1594 check_eq DEFAULT _ (Var v) = v == case_bndr
1595 check_eq (LitAlt lit') _ (Lit lit) = lit == lit'
1596 check_eq (DataAlt con) args rhs = rhs `cheapEqExpr` mkConApp con (arg_tys ++ varsToCoreExprs args)
1597 || rhs `cheapEqExpr` Var case_bndr
1598 check_eq _ _ _ = False
1600 arg_tys = map Type (tyConAppArgs (idType case_bndr))
1603 -- case e of x { _ -> x `cast` c }
1604 -- And we definitely want to eliminate this case, to give
1606 -- So we throw away the cast from the RHS, and reconstruct
1607 -- it at the other end. All the RHS casts must be the same
1608 -- if (all identity_alt alts) holds.
1610 -- Don't worry about nested casts, because the simplifier combines them
1611 de_cast (Cast e _) = e
1614 re_cast scrut = case head alts of
1615 (_,_,Cast _ co) -> Cast scrut co
1618 --------------------------------------------------
1619 -- 3. Merge Identical Alternatives
1620 --------------------------------------------------
1621 mkCase1 dflags scrut case_bndr ((_con1,bndrs1,rhs1) : con_alts)
1622 | all isDeadBinder bndrs1 -- Remember the default
1623 , length filtered_alts < length con_alts -- alternative comes first
1624 -- Also Note [Dead binders]
1625 = do { tick (AltMerge case_bndr)
1626 ; mkCase2 dflags scrut case_bndr alts' }
1628 alts' = (DEFAULT, [], rhs1) : filtered_alts
1629 filtered_alts = filter keep con_alts
1630 keep (_con,bndrs,rhs) = not (all isDeadBinder bndrs && rhs `cheapEqExpr` rhs1)
1632 mkCase1 dflags scrut bndr alts = mkCase2 dflags scrut bndr alts
1634 --------------------------------------------------
1636 --------------------------------------------------
1637 mkCase2 _dflags scrut bndr alts
1638 = return (Case scrut bndr (coreAltsType alts) alts)
1642 ~~~~~~~~~~~~~~~~~~~~
1643 Note that dead-ness is maintained by the simplifier, so that it is
1644 accurate after simplification as well as before.
1647 Note [Cascading case merge]
1648 ~~~~~~~~~~~~~~~~~~~~~~~~~~~
1649 Case merging should cascade in one sweep, because it
1653 DEFAULT -> case a of b
1654 DEFAULT -> case b of c {
1661 DEFAULT -> case a of b
1662 DEFAULT -> let c = b in e
1663 A -> let c = b in ea
1668 DEFAULT -> let b = a in let c = b in e
1669 A -> let b = a in let c = b in ea
1670 B -> let b = a in eb
1674 However here's a tricky case that we still don't catch, and I don't
1675 see how to catch it in one pass:
1677 case x of c1 { I# a1 ->
1680 DEFAULT -> case x of c3 { I# a2 ->
1683 After occurrence analysis (and its binder-swap) we get this
1685 case x of c1 { I# a1 ->
1686 let x = c1 in -- Binder-swap addition
1689 DEFAULT -> case x of c3 { I# a2 ->
1692 When we simplify the inner case x, we'll see that
1693 x=c1=I# a1. So we'll bind a2 to a1, and get
1695 case x of c1 { I# a1 ->
1698 DEFAULT -> case a1 of ...
1700 This is corect, but we can't do a case merge in this sweep
1701 because c2 /= a1. Reason: the binding c1=I# a1 went inwards
1702 without getting changed to c1=I# c2.
1704 I don't think this is worth fixing, even if I knew how. It'll
1705 all come out in the next pass anyway.