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
39 import DataCon ( dataConCannotMatch )
49 import Type hiding( substTy )
50 import Coercion hiding( substCo )
64 %************************************************************************
68 %************************************************************************
70 A SimplCont allows the simplifier to traverse the expression in a
71 zipper-like fashion. The SimplCont represents the rest of the expression,
72 "above" the point of interest.
74 You can also think of a SimplCont as an "evaluation context", using
75 that term in the way it is used for operational semantics. This is the
76 way I usually think of it, For example you'll often see a syntax for
77 evaluation context looking like
78 C ::= [] | C e | case C of alts | C `cast` co
79 That's the kind of thing we are doing here, and I use that syntax in
84 * A SimplCont describes a *strict* context (just like
85 evaluation contexts do). E.g. Just [] is not a SimplCont
87 * A SimplCont describes a context that *does not* bind
88 any variables. E.g. \x. [] is not a SimplCont
92 = Stop -- An empty context, or hole, []
93 CallCtxt -- True <=> There is something interesting about
94 -- the context, and hence the inliner
95 -- should be a bit keener (see interestingCallContext)
97 -- This is an argument of a function that has RULES
98 -- Inlining the call might allow the rule to fire
100 | CoerceIt -- C `cast` co
101 OutCoercion -- The coercion simplified
105 DupFlag -- See Note [DupFlag invariants]
106 InExpr StaticEnv -- The argument and its static env
109 | Select -- case C of alts
110 DupFlag -- See Note [DupFlag invariants]
111 InId [InAlt] StaticEnv -- The case binder, alts, and subst-env
114 -- The two strict forms have no DupFlag, because we never duplicate them
115 | StrictBind -- (\x* \xs. e) C
116 InId [InBndr] -- let x* = [] in e
117 InExpr StaticEnv -- is a special case
120 | StrictArg -- f e1 ..en C
121 ArgInfo -- Specifies f, e1..en, Whether f has rules, etc
122 -- plus strictness flags for *further* args
123 CallCtxt -- Whether *this* argument position is interesting
128 ai_fun :: Id, -- The function
129 ai_args :: [OutExpr], -- ...applied to these args (which are in *reverse* order)
130 ai_rules :: [CoreRule], -- Rules for this function
132 ai_encl :: Bool, -- Flag saying whether this function
133 -- or an enclosing one has rules (recursively)
134 -- True => be keener to inline in all args
136 ai_strs :: [Bool], -- Strictness of remaining arguments
137 -- Usually infinite, but if it is finite it guarantees
138 -- that the function diverges after being given
139 -- that number of args
140 ai_discs :: [Int] -- Discounts for remaining arguments; non-zero => be keener to inline
144 addArgTo :: ArgInfo -> OutExpr -> ArgInfo
145 addArgTo ai arg = ai { ai_args = arg : ai_args ai }
147 instance Outputable SimplCont where
148 ppr (Stop interesting) = ptext (sLit "Stop") <> brackets (ppr interesting)
149 ppr (ApplyTo dup arg _ cont) = ((ptext (sLit "ApplyTo") <+> ppr dup <+> pprParendExpr arg)
150 {- $$ nest 2 (pprSimplEnv se) -}) $$ ppr cont
151 ppr (StrictBind b _ _ _ cont) = (ptext (sLit "StrictBind") <+> ppr b) $$ ppr cont
152 ppr (StrictArg ai _ cont) = (ptext (sLit "StrictArg") <+> ppr (ai_fun ai)) $$ ppr cont
153 ppr (Select dup bndr alts se cont) = (ptext (sLit "Select") <+> ppr dup <+> ppr bndr) $$
154 (nest 2 $ vcat [ppr (seTvSubst se), ppr alts]) $$ ppr cont
155 ppr (CoerceIt co cont) = (ptext (sLit "CoerceIt") <+> ppr co) $$ ppr cont
157 data DupFlag = NoDup -- Unsimplified, might be big
158 | Simplified -- Simplified
159 | OkToDup -- Simplified and small
161 isSimplified :: DupFlag -> Bool
162 isSimplified NoDup = False
163 isSimplified _ = True -- Invariant: the subst-env is empty
165 instance Outputable DupFlag where
166 ppr OkToDup = ptext (sLit "ok")
167 ppr NoDup = ptext (sLit "nodup")
168 ppr Simplified = ptext (sLit "simpl")
171 Note [DupFlag invariants]
172 ~~~~~~~~~~~~~~~~~~~~~~~~~
173 In both (ApplyTo dup _ env k)
174 and (Select dup _ _ env k)
175 the following invariants hold
177 (a) if dup = OkToDup, then continuation k is also ok-to-dup
178 (b) if dup = OkToDup or Simplified, the subst-env is empty
179 (and and hence no need to re-simplify)
183 mkBoringStop :: SimplCont
184 mkBoringStop = Stop BoringCtxt
186 mkRhsStop :: SimplCont -- See Note [RHS of lets] in CoreUnfold
187 mkRhsStop = Stop (ArgCtxt False)
189 mkLazyArgStop :: CallCtxt -> SimplCont
190 mkLazyArgStop cci = Stop cci
193 contIsRhsOrArg :: SimplCont -> Bool
194 contIsRhsOrArg (Stop {}) = True
195 contIsRhsOrArg (StrictBind {}) = True
196 contIsRhsOrArg (StrictArg {}) = True
197 contIsRhsOrArg _ = False
200 contIsDupable :: SimplCont -> Bool
201 contIsDupable (Stop {}) = True
202 contIsDupable (ApplyTo OkToDup _ _ _) = True -- See Note [DupFlag invariants]
203 contIsDupable (Select OkToDup _ _ _ _) = True -- ...ditto...
204 contIsDupable (CoerceIt _ cont) = contIsDupable cont
205 contIsDupable _ = False
208 contIsTrivial :: SimplCont -> Bool
209 contIsTrivial (Stop {}) = True
210 contIsTrivial (ApplyTo _ (Type _) _ cont) = contIsTrivial cont
211 contIsTrivial (ApplyTo _ (Coercion _) _ cont) = contIsTrivial cont
212 contIsTrivial (CoerceIt _ cont) = contIsTrivial cont
213 contIsTrivial _ = False
216 contResultType :: SimplEnv -> OutType -> SimplCont -> OutType
217 contResultType env ty cont
220 subst_ty se ty = SimplEnv.substTy (se `setInScope` env) ty
221 subst_co se co = SimplEnv.substCo (se `setInScope` env) co
224 go (CoerceIt co cont) _ = go cont (pSnd (coercionKind co))
225 go (StrictBind _ bs body se cont) _ = go cont (subst_ty se (exprType (mkLams bs body)))
226 go (StrictArg ai _ cont) _ = go cont (funResultTy (argInfoResultTy ai))
227 go (Select _ _ alts se cont) _ = go cont (subst_ty se (coreAltsType alts))
228 go (ApplyTo _ arg se cont) ty = go cont (apply_to_arg ty arg se)
230 apply_to_arg ty (Type ty_arg) se = applyTy ty (subst_ty se ty_arg)
231 apply_to_arg ty (Coercion co_arg) se = applyCo ty (subst_co se co_arg)
232 apply_to_arg ty _ _ = funResultTy ty
234 argInfoResultTy :: ArgInfo -> OutType
235 argInfoResultTy (ArgInfo { ai_fun = fun, ai_args = args })
236 = foldr (\arg fn_ty -> applyTypeToArg fn_ty arg) (idType fun) args
239 countValArgs :: SimplCont -> Int
240 countValArgs (ApplyTo _ (Type _) _ cont) = countValArgs cont
241 countValArgs (ApplyTo _ (Coercion _) _ cont) = countValArgs cont
242 countValArgs (ApplyTo _ _ _ cont) = 1 + countValArgs cont
245 countArgs :: SimplCont -> Int
246 countArgs (ApplyTo _ _ _ cont) = 1 + countArgs cont
249 contArgs :: SimplCont -> (Bool, [ArgSummary], SimplCont)
250 -- Uses substitution to turn each arg into an OutExpr
251 contArgs cont@(ApplyTo {})
252 = case go [] cont of { (args, cont') -> (False, args, cont') }
254 go args (ApplyTo _ arg se cont)
255 | isTypeArg arg = go args cont
256 | otherwise = go (is_interesting arg se : args) cont
257 go args cont = (reverse args, cont)
259 is_interesting arg se = interestingArg (substExpr (text "contArgs") se arg)
260 -- Do *not* use short-cutting substitution here
261 -- because we want to get as much IdInfo as possible
263 contArgs cont = (True, [], cont)
265 pushSimplifiedArgs :: SimplEnv -> [CoreExpr] -> SimplCont -> SimplCont
266 pushSimplifiedArgs _env [] cont = cont
267 pushSimplifiedArgs env (arg:args) cont = ApplyTo Simplified arg env (pushSimplifiedArgs env args cont)
268 -- The env has an empty SubstEnv
270 dropArgs :: Int -> SimplCont -> SimplCont
271 dropArgs 0 cont = cont
272 dropArgs n (ApplyTo _ _ _ cont) = dropArgs (n-1) cont
273 dropArgs n other = pprPanic "dropArgs" (ppr n <+> ppr other)
277 Note [Interesting call context]
278 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
279 We want to avoid inlining an expression where there can't possibly be
280 any gain, such as in an argument position. Hence, if the continuation
281 is interesting (eg. a case scrutinee, application etc.) then we
282 inline, otherwise we don't.
284 Previously some_benefit used to return True only if the variable was
285 applied to some value arguments. This didn't work:
287 let x = _coerce_ (T Int) Int (I# 3) in
288 case _coerce_ Int (T Int) x of
291 we want to inline x, but can't see that it's a constructor in a case
292 scrutinee position, and some_benefit is False.
296 dMonadST = _/\_ t -> :Monad (g1 _@_ t, g2 _@_ t, g3 _@_ t)
298 .... case dMonadST _@_ x0 of (a,b,c) -> ....
300 we'd really like to inline dMonadST here, but we *don't* want to
301 inline if the case expression is just
303 case x of y { DEFAULT -> ... }
305 since we can just eliminate this case instead (x is in WHNF). Similar
306 applies when x is bound to a lambda expression. Hence
307 contIsInteresting looks for case expressions with just a single
312 interestingCallContext :: SimplCont -> CallCtxt
313 -- See Note [Interesting call context]
314 interestingCallContext cont
317 interesting (Select _ bndr _ _ _)
318 | isDeadBinder bndr = CaseCtxt
319 | otherwise = ArgCtxt False -- If the binder is used, this
320 -- is like a strict let
321 -- See Note [RHS of lets] in CoreUnfold
323 interesting (ApplyTo _ arg _ cont)
324 | isTypeArg arg = interesting cont
325 | otherwise = ValAppCtxt -- Can happen if we have (f Int |> co) y
326 -- If f has an INLINE prag we need to give it some
327 -- motivation to inline. See Note [Cast then apply]
330 interesting (StrictArg _ cci _) = cci
331 interesting (StrictBind {}) = BoringCtxt
332 interesting (Stop cci) = cci
333 interesting (CoerceIt _ cont) = interesting cont
334 -- If this call is the arg of a strict function, the context
335 -- is a bit interesting. If we inline here, we may get useful
336 -- evaluation information to avoid repeated evals: e.g.
338 -- Here the contIsInteresting makes the '*' keener to inline,
339 -- which in turn exposes a constructor which makes the '+' inline.
340 -- Assuming that +,* aren't small enough to inline regardless.
342 -- It's also very important to inline in a strict context for things
345 -- Here, the context of (f x) is strict, and if f's unfolding is
346 -- a build it's *great* to inline it here. So we must ensure that
347 -- the context for (f x) is not totally uninteresting.
352 -> [CoreRule] -- Rules for function
353 -> Int -- Number of value args
354 -> SimplCont -- Context of the call
357 mkArgInfo fun rules n_val_args call_cont
358 | n_val_args < idArity fun -- Note [Unsaturated functions]
359 = ArgInfo { ai_fun = fun, ai_args = [], ai_rules = rules
361 , ai_strs = vanilla_stricts
362 , ai_discs = vanilla_discounts }
364 = ArgInfo { ai_fun = fun, ai_args = [], ai_rules = rules
365 , ai_encl = interestingArgContext rules call_cont
366 , ai_strs = add_type_str (idType fun) arg_stricts
367 , ai_discs = arg_discounts }
369 vanilla_discounts, arg_discounts :: [Int]
370 vanilla_discounts = repeat 0
371 arg_discounts = case idUnfolding fun of
372 CoreUnfolding {uf_guidance = UnfIfGoodArgs {ug_args = discounts}}
373 -> discounts ++ vanilla_discounts
374 _ -> vanilla_discounts
376 vanilla_stricts, arg_stricts :: [Bool]
377 vanilla_stricts = repeat False
380 = case splitStrictSig (idStrictness fun) of
381 (demands, result_info)
382 | not (demands `lengthExceeds` n_val_args)
383 -> -- Enough args, use the strictness given.
384 -- For bottoming functions we used to pretend that the arg
385 -- is lazy, so that we don't treat the arg as an
386 -- interesting context. This avoids substituting
387 -- top-level bindings for (say) strings into
388 -- calls to error. But now we are more careful about
389 -- inlining lone variables, so its ok (see SimplUtils.analyseCont)
390 if isBotRes result_info then
391 map isStrictDmd demands -- Finite => result is bottom
393 map isStrictDmd demands ++ vanilla_stricts
395 -> WARN( True, text "More demands than arity" <+> ppr fun <+> ppr (idArity fun)
396 <+> ppr n_val_args <+> ppr demands )
397 vanilla_stricts -- Not enough args, or no strictness
399 add_type_str :: Type -> [Bool] -> [Bool]
400 -- If the function arg types are strict, record that in the 'strictness bits'
401 -- No need to instantiate because unboxed types (which dominate the strict
402 -- types) can't instantiate type variables.
403 -- add_type_str is done repeatedly (for each call); might be better
404 -- once-for-all in the function
405 -- But beware primops/datacons with no strictness
406 add_type_str _ [] = []
407 add_type_str fun_ty strs -- Look through foralls
408 | Just (_, fun_ty') <- splitForAllTy_maybe fun_ty -- Includes coercions
409 = add_type_str fun_ty' strs
410 add_type_str fun_ty (str:strs) -- Add strict-type info
411 | Just (arg_ty, fun_ty') <- splitFunTy_maybe fun_ty
412 = (str || isStrictType arg_ty) : add_type_str fun_ty' strs
416 {- Note [Unsaturated functions]
417 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
418 Consider (test eyeball/inline4)
421 where f has arity 2. Then we do not want to inline 'x', because
422 it'll just be floated out again. Even if f has lots of discounts
423 on its first argument -- it must be saturated for these to kick in
426 interestingArgContext :: [CoreRule] -> SimplCont -> Bool
427 -- If the argument has form (f x y), where x,y are boring,
428 -- and f is marked INLINE, then we don't want to inline f.
429 -- But if the context of the argument is
431 -- where g has rules, then we *do* want to inline f, in case it
432 -- exposes a rule that might fire. Similarly, if the context is
434 -- where h has rules, then we do want to inline f; hence the
435 -- call_cont argument to interestingArgContext
437 -- The ai-rules flag makes this happen; if it's
438 -- set, the inliner gets just enough keener to inline f
439 -- regardless of how boring f's arguments are, if it's marked INLINE
441 -- The alternative would be to *always* inline an INLINE function,
442 -- regardless of how boring its context is; but that seems overkill
443 -- For example, it'd mean that wrapper functions were always inlined
444 interestingArgContext rules call_cont
445 = notNull rules || enclosing_fn_has_rules
447 enclosing_fn_has_rules = go call_cont
449 go (Select {}) = False
450 go (ApplyTo {}) = False
451 go (StrictArg _ cci _) = interesting cci
452 go (StrictBind {}) = False -- ??
453 go (CoerceIt _ c) = go c
454 go (Stop cci) = interesting cci
456 interesting (ArgCtxt rules) = rules
457 interesting _ = False
461 %************************************************************************
465 %************************************************************************
467 The SimplifierMode controls several switches; see its definition in
469 sm_rules :: Bool -- Whether RULES are enabled
470 sm_inline :: Bool -- Whether inlining is enabled
471 sm_case_case :: Bool -- Whether case-of-case is enabled
472 sm_eta_expand :: Bool -- Whether eta-expansion is enabled
475 simplEnvForGHCi :: DynFlags -> SimplEnv
476 simplEnvForGHCi dflags
477 = mkSimplEnv $ SimplMode { sm_names = ["GHCi"]
478 , sm_phase = InitialPhase
479 , sm_rules = rules_on
481 , sm_eta_expand = eta_expand_on
482 , sm_case_case = True }
484 rules_on = dopt Opt_EnableRewriteRules dflags
485 eta_expand_on = dopt Opt_DoLambdaEtaExpansion dflags
486 -- Do not do any inlining, in case we expose some unboxed
487 -- tuple stuff that confuses the bytecode interpreter
489 updModeForInlineRules :: Activation -> SimplifierMode -> SimplifierMode
490 -- See Note [Simplifying inside InlineRules]
491 updModeForInlineRules inline_rule_act current_mode
492 = current_mode { sm_phase = phaseFromActivation inline_rule_act
494 , sm_eta_expand = False }
495 -- For sm_rules, just inherit; sm_rules might be "off"
496 -- becuase of -fno-enable-rewrite-rules
498 phaseFromActivation (ActiveAfter n) = Phase n
499 phaseFromActivation _ = InitialPhase
502 Note [Inlining in gentle mode]
503 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
504 Something is inlined if
505 (i) the sm_inline flag is on, AND
506 (ii) the thing has an INLINE pragma, AND
507 (iii) the thing is inlinable in the earliest phase.
509 Example of why (iii) is important:
510 {-# INLINE [~1] g #-}
516 If we were to inline g into f's inlining, then an importing module would
518 f e --> g (g e) ---> RULE fires
519 because the InlineRule for f has had g inlined into it.
521 On the other hand, it is bad not to do ANY inlining into an
522 InlineRule, because then recursive knots in instance declarations
523 don't get unravelled.
525 However, *sometimes* SimplGently must do no call-site inlining at all
526 (hence sm_inline = False). Before full laziness we must be careful
527 not to inline wrappers, because doing so inhibits floating
528 e.g. ...(case f x of ...)...
529 ==> ...(case (case x of I# x# -> fw x#) of ...)...
530 ==> ...(case x of I# x# -> case fw x# of ...)...
531 and now the redex (f x) isn't floatable any more.
533 The no-inlining thing is also important for Template Haskell. You might be
534 compiling in one-shot mode with -O2; but when TH compiles a splice before
535 running it, we don't want to use -O2. Indeed, we don't want to inline
536 anything, because the byte-code interpreter might get confused about
537 unboxed tuples and suchlike.
539 Note [Simplifying inside InlineRules]
540 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
541 We must take care with simplification inside InlineRules (which come from
544 First, consider the following example
549 in ...g...g...g...g...g...
550 Now, if that's the ONLY occurrence of f, it might be inlined inside g,
551 and thence copied multiple times when g is inlined. HENCE we treat
552 any occurrence in an InlineRule as a multiple occurrence, not a single
553 one; see OccurAnal.addRuleUsage.
555 Second, we do want *do* to some modest rules/inlining stuff in InlineRules,
556 partly to eliminate senseless crap, and partly to break the recursive knots
557 generated by instance declarations.
559 However, suppose we have
560 {-# INLINE <act> f #-}
562 meaning "inline f in phases p where activation <act>(p) holds".
563 Then what inlinings/rules can we apply to the copy of <rhs> captured in
564 f's InlineRule? Our model is that literally <rhs> is substituted for
565 f when it is inlined. So our conservative plan (implemented by
566 updModeForInlineRules) is this:
568 -------------------------------------------------------------
569 When simplifying the RHS of an InlineRule, set the phase to the
570 phase in which the InlineRule first becomes active
571 -------------------------------------------------------------
575 a) Rules/inlinings that *cease* being active before p will
576 not apply to the InlineRule rhs, consistent with it being
577 inlined in its *original* form in phase p.
579 b) Rules/inlinings that only become active *after* p will
580 not apply to the InlineRule rhs, again to be consistent with
581 inlining the *original* rhs in phase p.
587 {-# NOINLINE [1] g #-}
590 {-# RULE h g = ... #-}
591 Here we must not inline g into f's RHS, even when we get to phase 0,
592 because when f is later inlined into some other module we want the
600 and suppose that there are auto-generated specialisations and a strictness
601 wrapper for g. The specialisations get activation AlwaysActive, and the
602 strictness wrapper get activation (ActiveAfter 0). So the strictness
603 wrepper fails the test and won't be inlined into f's InlineRule. That
604 means f can inline, expose the specialised call to g, so the specialisation
607 A note about wrappers
608 ~~~~~~~~~~~~~~~~~~~~~
609 It's also important not to inline a worker back into a wrapper.
611 wraper = inline_me (\x -> ...worker... )
612 Normally, the inline_me prevents the worker getting inlined into
613 the wrapper (initially, the worker's only call site!). But,
614 if the wrapper is sure to be called, the strictness analyser will
615 mark it 'demanded', so when the RHS is simplified, it'll get an ArgOf
619 activeUnfolding :: SimplEnv -> Id -> Bool
621 | not (sm_inline mode) = active_unfolding_minimal
622 | otherwise = case sm_phase mode of
623 InitialPhase -> active_unfolding_gentle
624 Phase n -> active_unfolding n
628 getUnfoldingInRuleMatch :: SimplEnv -> IdUnfoldingFun
629 -- When matching in RULE, we want to "look through" an unfolding
630 -- (to see a constructor) if *rules* are on, even if *inlinings*
631 -- are not. A notable example is DFuns, which really we want to
632 -- match in rules like (op dfun) in gentle mode. Another example
633 -- is 'otherwise' which we want exprIsConApp_maybe to be able to
635 getUnfoldingInRuleMatch env id
636 | unf_is_active = idUnfolding id
637 | otherwise = NoUnfolding
641 | not (sm_rules mode) = active_unfolding_minimal id
642 | otherwise = isActive (sm_phase mode) (idInlineActivation id)
644 active_unfolding_minimal :: Id -> Bool
645 -- Compuslory unfoldings only
646 -- Ignore SimplGently, because we want to inline regardless;
647 -- the Id has no top-level binding at all
649 -- NB: we used to have a second exception, for data con wrappers.
650 -- On the grounds that we use gentle mode for rule LHSs, and
651 -- they match better when data con wrappers are inlined.
652 -- But that only really applies to the trivial wrappers (like (:)),
653 -- and they are now constructed as Compulsory unfoldings (in MkId)
654 -- so they'll happen anyway.
655 active_unfolding_minimal id = isCompulsoryUnfolding (realIdUnfolding id)
657 active_unfolding :: PhaseNum -> Id -> Bool
658 active_unfolding n id = isActiveIn n (idInlineActivation id)
660 active_unfolding_gentle :: Id -> Bool
661 -- Anything that is early-active
662 -- See Note [Gentle mode]
663 active_unfolding_gentle id
664 = isInlinePragma prag
665 && isEarlyActive (inlinePragmaActivation prag)
666 -- NB: wrappers are not early-active
668 prag = idInlinePragma id
670 ----------------------
671 activeRule :: DynFlags -> SimplEnv -> Maybe (Activation -> Bool)
672 -- Nothing => No rules at all
673 activeRule _dflags env
674 | not (sm_rules mode) = Nothing -- Rewriting is off
675 | otherwise = Just (isActive (sm_phase mode))
682 %************************************************************************
684 preInlineUnconditionally
686 %************************************************************************
688 preInlineUnconditionally
689 ~~~~~~~~~~~~~~~~~~~~~~~~
690 @preInlineUnconditionally@ examines a bndr to see if it is used just
691 once in a completely safe way, so that it is safe to discard the
692 binding inline its RHS at the (unique) usage site, REGARDLESS of how
693 big the RHS might be. If this is the case we don't simplify the RHS
694 first, but just inline it un-simplified.
696 This is much better than first simplifying a perhaps-huge RHS and then
697 inlining and re-simplifying it. Indeed, it can be at least quadratically
706 We may end up simplifying e1 N times, e2 N-1 times, e3 N-3 times etc.
707 This can happen with cascades of functions too:
714 THE MAIN INVARIANT is this:
716 ---- preInlineUnconditionally invariant -----
717 IF preInlineUnconditionally chooses to inline x = <rhs>
718 THEN doing the inlining should not change the occurrence
719 info for the free vars of <rhs>
720 ----------------------------------------------
722 For example, it's tempting to look at trivial binding like
724 and inline it unconditionally. But suppose x is used many times,
725 but this is the unique occurrence of y. Then inlining x would change
726 y's occurrence info, which breaks the invariant. It matters: y
727 might have a BIG rhs, which will now be dup'd at every occurrenc of x.
730 Even RHSs labelled InlineMe aren't caught here, because there might be
731 no benefit from inlining at the call site.
733 [Sept 01] Don't unconditionally inline a top-level thing, because that
734 can simply make a static thing into something built dynamically. E.g.
738 [Remember that we treat \s as a one-shot lambda.] No point in
739 inlining x unless there is something interesting about the call site.
741 But watch out: if you aren't careful, some useful foldr/build fusion
742 can be lost (most notably in spectral/hartel/parstof) because the
743 foldr didn't see the build. Doing the dynamic allocation isn't a big
744 deal, in fact, but losing the fusion can be. But the right thing here
745 seems to be to do a callSiteInline based on the fact that there is
746 something interesting about the call site (it's strict). Hmm. That
749 Conclusion: inline top level things gaily until Phase 0 (the last
750 phase), at which point don't.
752 Note [pre/postInlineUnconditionally in gentle mode]
753 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
754 Even in gentle mode we want to do preInlineUnconditionally. The
755 reason is that too little clean-up happens if you don't inline
756 use-once things. Also a bit of inlining is *good* for full laziness;
757 it can expose constant sub-expressions. Example in
758 spectral/mandel/Mandel.hs, where the mandelset function gets a useful
759 let-float if you inline windowToViewport
761 However, as usual for Gentle mode, do not inline things that are
762 inactive in the intial stages. See Note [Gentle mode].
764 Note [InlineRule and preInlineUnconditionally]
765 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
766 Surprisingly, do not pre-inline-unconditionally Ids with INLINE pragmas!
776 ...fInt...fInt...fInt...
778 Here f occurs just once, in the RHS of f1. But if we inline it there
779 we'll lose the opportunity to inline at each of fInt's call sites.
780 The INLINE pragma will only inline when the application is saturated
781 for exactly this reason; and we don't want PreInlineUnconditionally
782 to second-guess it. A live example is Trac #3736.
783 c.f. Note [InlineRule and postInlineUnconditionally]
785 Note [Top-level botomming Ids]
786 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
787 Don't inline top-level Ids that are bottoming, even if they are used just
788 once, because FloatOut has gone to some trouble to extract them out.
789 Inlining them won't make the program run faster!
792 preInlineUnconditionally :: SimplEnv -> TopLevelFlag -> InId -> InExpr -> Bool
793 preInlineUnconditionally env top_lvl bndr rhs
795 | isStableUnfolding (idUnfolding bndr) = False -- Note [InlineRule and preInlineUnconditionally]
796 | isTopLevel top_lvl && isBottomingId bndr = False -- Note [Top-level bottoming Ids]
797 | opt_SimplNoPreInlining = False
798 | otherwise = case idOccInfo bndr of
799 IAmDead -> True -- Happens in ((\x.1) v)
800 OneOcc in_lam True int_cxt -> try_once in_lam int_cxt
804 active = isActive (sm_phase mode) act
805 -- See Note [pre/postInlineUnconditionally in gentle mode]
806 act = idInlineActivation bndr
807 try_once in_lam int_cxt -- There's one textual occurrence
808 | not in_lam = isNotTopLevel top_lvl || early_phase
809 | otherwise = int_cxt && canInlineInLam rhs
811 -- Be very careful before inlining inside a lambda, because (a) we must not
812 -- invalidate occurrence information, and (b) we want to avoid pushing a
813 -- single allocation (here) into multiple allocations (inside lambda).
814 -- Inlining a *function* with a single *saturated* call would be ok, mind you.
815 -- || (if is_cheap && not (canInlineInLam rhs) then pprTrace "preinline" (ppr bndr <+> ppr rhs) ok else ok)
817 -- is_cheap = exprIsCheap rhs
818 -- ok = is_cheap && int_cxt
820 -- int_cxt The context isn't totally boring
821 -- E.g. let f = \ab.BIG in \y. map f xs
822 -- Don't want to substitute for f, because then we allocate
823 -- its closure every time the \y is called
824 -- But: let f = \ab.BIG in \y. map (f y) xs
825 -- Now we do want to substitute for f, even though it's not
826 -- saturated, because we're going to allocate a closure for
827 -- (f y) every time round the loop anyhow.
829 -- canInlineInLam => free vars of rhs are (Once in_lam) or Many,
830 -- so substituting rhs inside a lambda doesn't change the occ info.
831 -- Sadly, not quite the same as exprIsHNF.
832 canInlineInLam (Lit _) = True
833 canInlineInLam (Lam b e) = isRuntimeVar b || canInlineInLam e
834 canInlineInLam (Note _ e) = canInlineInLam e
835 canInlineInLam _ = False
837 early_phase = case sm_phase mode of
840 -- If we don't have this early_phase test, consider
841 -- x = length [1,2,3]
842 -- The full laziness pass carefully floats all the cons cells to
843 -- top level, and preInlineUnconditionally floats them all back in.
844 -- Result is (a) static allocation replaced by dynamic allocation
845 -- (b) many simplifier iterations because this tickles
846 -- a related problem; only one inlining per pass
848 -- On the other hand, I have seen cases where top-level fusion is
849 -- lost if we don't inline top level thing (e.g. string constants)
850 -- Hence the test for phase zero (which is the phase for all the final
851 -- simplifications). Until phase zero we take no special notice of
852 -- top level things, but then we become more leery about inlining
857 %************************************************************************
859 postInlineUnconditionally
861 %************************************************************************
863 postInlineUnconditionally
864 ~~~~~~~~~~~~~~~~~~~~~~~~~
865 @postInlineUnconditionally@ decides whether to unconditionally inline
866 a thing based on the form of its RHS; in particular if it has a
867 trivial RHS. If so, we can inline and discard the binding altogether.
869 NB: a loop breaker has must_keep_binding = True and non-loop-breakers
870 only have *forward* references. Hence, it's safe to discard the binding
872 NOTE: This isn't our last opportunity to inline. We're at the binding
873 site right now, and we'll get another opportunity when we get to the
876 Note that we do this unconditional inlining only for trival RHSs.
877 Don't inline even WHNFs inside lambdas; doing so may simply increase
878 allocation when the function is called. This isn't the last chance; see
881 NB: Even inline pragmas (e.g. IMustBeINLINEd) are ignored here Why?
882 Because we don't even want to inline them into the RHS of constructor
883 arguments. See NOTE above
885 NB: At one time even NOINLINE was ignored here: if the rhs is trivial
886 it's best to inline it anyway. We often get a=E; b=a from desugaring,
887 with both a and b marked NOINLINE. But that seems incompatible with
888 our new view that inlining is like a RULE, so I'm sticking to the 'active'
892 postInlineUnconditionally
893 :: SimplEnv -> TopLevelFlag
894 -> OutId -- The binder (an InId would be fine too)
895 -> OccInfo -- From the InId
899 postInlineUnconditionally env top_lvl bndr occ_info rhs unfolding
901 | isLoopBreaker occ_info = False -- If it's a loop-breaker of any kind, don't inline
902 -- because it might be referred to "earlier"
903 | isExportedId bndr = False
904 | isStableUnfolding unfolding = False -- Note [InlineRule and postInlineUnconditionally]
905 | isTopLevel top_lvl = False -- Note [Top level and postInlineUnconditionally]
906 | exprIsTrivial rhs = True
909 -- The point of examining occ_info here is that for *non-values*
910 -- that occur outside a lambda, the call-site inliner won't have
911 -- a chance (becuase it doesn't know that the thing
912 -- only occurs once). The pre-inliner won't have gotten
913 -- it either, if the thing occurs in more than one branch
914 -- So the main target is things like
917 -- True -> case x of ...
918 -- False -> case x of ...
919 -- This is very important in practice; e.g. wheel-seive1 doubles
920 -- in allocation if you miss this out
921 OneOcc in_lam _one_br int_cxt -- OneOcc => no code-duplication issue
922 -> smallEnoughToInline unfolding -- Small enough to dup
923 -- ToDo: consider discount on smallEnoughToInline if int_cxt is true
925 -- NB: Do NOT inline arbitrarily big things, even if one_br is True
926 -- Reason: doing so risks exponential behaviour. We simplify a big
927 -- expression, inline it, and simplify it again. But if the
928 -- very same thing happens in the big expression, we get
930 -- PRINCIPLE: when we've already simplified an expression once,
931 -- make sure that we only inline it if it's reasonably small.
934 -- Outside a lambda, we want to be reasonably aggressive
935 -- about inlining into multiple branches of case
936 -- e.g. let x = <non-value>
937 -- in case y of { C1 -> ..x..; C2 -> ..x..; C3 -> ... }
938 -- Inlining can be a big win if C3 is the hot-spot, even if
939 -- the uses in C1, C2 are not 'interesting'
940 -- An example that gets worse if you add int_cxt here is 'clausify'
942 (isCheapUnfolding unfolding && int_cxt))
943 -- isCheap => acceptable work duplication; in_lam may be true
944 -- int_cxt to prevent us inlining inside a lambda without some
945 -- good reason. See the notes on int_cxt in preInlineUnconditionally
947 IAmDead -> True -- This happens; for example, the case_bndr during case of
948 -- known constructor: case (a,b) of x { (p,q) -> ... }
949 -- Here x isn't mentioned in the RHS, so we don't want to
950 -- create the (dead) let-binding let x = (a,b) in ...
954 -- Here's an example that we don't handle well:
955 -- let f = if b then Left (\x.BIG) else Right (\y.BIG)
956 -- in \y. ....case f of {...} ....
957 -- Here f is used just once, and duplicating the case work is fine (exprIsCheap).
959 -- - We can't preInlineUnconditionally because that woud invalidate
960 -- the occ info for b.
961 -- - We can't postInlineUnconditionally because the RHS is big, and
962 -- that risks exponential behaviour
963 -- - We can't call-site inline, because the rhs is big
967 active = isActive (sm_phase (getMode env)) (idInlineActivation bndr)
968 -- See Note [pre/postInlineUnconditionally in gentle mode]
971 Note [Top level and postInlineUnconditionally]
972 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
973 We don't do postInlineUnconditionally for top-level things (even for
974 ones that are trivial):
976 * Doing so will inline top-level error expressions that have been
977 carefully floated out by FloatOut. More generally, it might
978 replace static allocation with dynamic.
980 * Even for trivial expressions there's a problem. Consider
981 {-# RULE "foo" forall (xs::[T]). reverse xs = ruggle xs #-}
984 In one simplifier pass we might fire the rule, getting
986 but in *that* simplifier pass we must not do postInlineUnconditionally
987 on 'ruggle' because then we'll have an unbound occurrence of 'ruggle'
989 If the rhs is trivial it'll be inlined by callSiteInline, and then
990 the binding will be dead and discarded by the next use of OccurAnal
992 * There is less point, because the main goal is to get rid of local
993 bindings used in multiple case branches.
996 Note [InlineRule and postInlineUnconditionally]
997 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
998 Do not do postInlineUnconditionally if the Id has an InlineRule, otherwise
999 we lose the unfolding. Example
1001 -- f has InlineRule with rhs (e |> co)
1005 Then there's a danger we'll optimise to
1010 and now postInlineUnconditionally, losing the InlineRule on f. Now f'
1011 won't inline because 'e' is too big.
1013 c.f. Note [InlineRule and preInlineUnconditionally]
1016 %************************************************************************
1020 %************************************************************************
1023 mkLam :: SimplEnv -> [OutBndr] -> OutExpr -> SimplM OutExpr
1024 -- mkLam tries three things
1025 -- a) eta reduction, if that gives a trivial expression
1026 -- b) eta expansion [only if there are some value lambdas]
1030 mkLam _env bndrs body
1031 = do { dflags <- getDOptsSmpl
1032 ; mkLam' dflags bndrs body }
1034 mkLam' :: DynFlags -> [OutBndr] -> OutExpr -> SimplM OutExpr
1035 mkLam' dflags bndrs (Cast body co)
1036 | not (any bad bndrs)
1037 -- Note [Casts and lambdas]
1038 = do { lam <- mkLam' dflags bndrs body
1039 ; return (mkCoerce (mkPiCos bndrs co) lam) }
1041 co_vars = tyCoVarsOfCo co
1042 bad bndr = isCoVar bndr && bndr `elemVarSet` co_vars
1044 mkLam' dflags bndrs body@(Lam {})
1045 = mkLam' dflags (bndrs ++ bndrs1) body1
1047 (bndrs1, body1) = collectBinders body
1049 mkLam' dflags bndrs body
1050 | dopt Opt_DoEtaReduction dflags
1051 , Just etad_lam <- tryEtaReduce bndrs body
1052 = do { tick (EtaReduction (head bndrs))
1056 = return (mkLams bndrs body)
1060 Note [Casts and lambdas]
1061 ~~~~~~~~~~~~~~~~~~~~~~~~
1063 (\x. (\y. e) `cast` g1) `cast` g2
1064 There is a danger here that the two lambdas look separated, and the
1065 full laziness pass might float an expression to between the two.
1067 So this equation in mkLam' floats the g1 out, thus:
1068 (\x. e `cast` g1) --> (\x.e) `cast` (tx -> g1)
1071 In general, this floats casts outside lambdas, where (I hope) they
1072 might meet and cancel with some other cast:
1073 \x. e `cast` co ===> (\x. e) `cast` (tx -> co)
1074 /\a. e `cast` co ===> (/\a. e) `cast` (/\a. co)
1075 /\g. e `cast` co ===> (/\g. e) `cast` (/\g. co)
1076 (if not (g `in` co))
1078 Notice that it works regardless of 'e'. Originally it worked only
1079 if 'e' was itself a lambda, but in some cases that resulted in
1080 fruitless iteration in the simplifier. A good example was when
1081 compiling Text.ParserCombinators.ReadPrec, where we had a definition
1082 like (\x. Get `cast` g)
1083 where Get is a constructor with nonzero arity. Then mkLam eta-expanded
1084 the Get, and the next iteration eta-reduced it, and then eta-expanded
1087 Note also the side condition for the case of coercion binders.
1088 It does not make sense to transform
1089 /\g. e `cast` g ==> (/\g.e) `cast` (/\g.g)
1090 because the latter is not well-kinded.
1092 %************************************************************************
1096 %************************************************************************
1099 tryEtaExpand :: SimplEnv -> OutId -> OutExpr -> SimplM (Arity, OutExpr)
1100 -- See Note [Eta-expanding at let bindings]
1101 tryEtaExpand env bndr rhs
1102 = do { dflags <- getDOptsSmpl
1103 ; (new_arity, new_rhs) <- try_expand dflags
1105 ; WARN( new_arity < old_arity || new_arity < _dmd_arity,
1106 (ptext (sLit "Arity decrease:") <+> (ppr bndr <+> ppr old_arity
1107 <+> ppr new_arity <+> ppr _dmd_arity) $$ ppr new_rhs) )
1108 -- Note [Arity decrease]
1109 return (new_arity, new_rhs) }
1112 | sm_eta_expand (getMode env) -- Provided eta-expansion is on
1113 , not (exprIsTrivial rhs)
1114 , let dicts_cheap = dopt Opt_DictsCheap dflags
1115 new_arity = findArity dicts_cheap bndr rhs old_arity
1116 , new_arity > rhs_arity
1117 = do { tick (EtaExpansion bndr)
1118 ; return (new_arity, etaExpand new_arity rhs) }
1120 = return (rhs_arity, rhs)
1122 rhs_arity = exprArity rhs
1123 old_arity = idArity bndr
1124 _dmd_arity = length $ fst $ splitStrictSig $ idStrictness bndr
1126 findArity :: Bool -> Id -> CoreExpr -> Arity -> Arity
1127 -- This implements the fixpoint loop for arity analysis
1128 -- See Note [Arity analysis]
1129 findArity dicts_cheap bndr rhs old_arity
1130 = go (exprEtaExpandArity (mk_cheap_fn dicts_cheap init_cheap_app) rhs)
1131 -- We always call exprEtaExpandArity once, but usually
1132 -- that produces a result equal to old_arity, and then
1133 -- we stop right away (since arities should not decrease)
1134 -- Result: the common case is that there is just one iteration
1136 go :: Arity -> Arity
1138 | cur_arity <= old_arity = cur_arity
1139 | new_arity == cur_arity = cur_arity
1140 | otherwise = ASSERT( new_arity < cur_arity )
1141 pprTrace "Exciting arity"
1142 (vcat [ ppr bndr <+> ppr cur_arity <+> ppr new_arity
1146 new_arity = exprEtaExpandArity (mk_cheap_fn dicts_cheap cheap_app) rhs
1148 cheap_app :: CheapAppFun
1149 cheap_app fn n_val_args
1150 | fn == bndr = n_val_args < cur_arity
1151 | otherwise = isCheapApp fn n_val_args
1153 init_cheap_app :: CheapAppFun
1154 init_cheap_app fn n_val_args
1156 | otherwise = isCheapApp fn n_val_args
1158 mk_cheap_fn :: Bool -> CheapAppFun -> CheapFun
1159 mk_cheap_fn dicts_cheap cheap_app
1161 = \e _ -> exprIsCheap' cheap_app e
1163 = \e mb_ty -> exprIsCheap' cheap_app e
1166 Just ty -> isDictLikeTy ty
1167 -- If the experimental -fdicts-cheap flag is on, we eta-expand through
1168 -- dictionary bindings. This improves arities. Thereby, it also
1169 -- means that full laziness is less prone to floating out the
1170 -- application of a function to its dictionary arguments, which
1171 -- can thereby lose opportunities for fusion. Example:
1172 -- foo :: Ord a => a -> ...
1173 -- foo = /\a \(d:Ord a). let d' = ...d... in \(x:a). ....
1174 -- -- So foo has arity 1
1176 -- f = \x. foo dInt $ bar x
1178 -- The (foo DInt) is floated out, and makes ineffective a RULE
1179 -- foo (bar x) = ...
1181 -- One could go further and make exprIsCheap reply True to any
1182 -- dictionary-typed expression, but that's more work.
1184 -- See Note [Dictionary-like types] in TcType.lhs for why we use
1185 -- isDictLikeTy here rather than isDictTy
1188 Note [Eta-expanding at let bindings]
1189 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
1190 We now eta expand at let-bindings, which is where the payoff
1193 One useful consequence is this example:
1194 genMap :: C a => ...
1195 {-# INLINE genMap #-}
1199 {-# INLINE myMap #-}
1202 Notice that 'genMap' should only inline if applied to two arguments.
1203 In the InlineRule for myMap we'll have the unfolding
1204 (\d -> genMap Int (..d..))
1205 We do not want to eta-expand to
1206 (\d f xs -> genMap Int (..d..) f xs)
1207 because then 'genMap' will inline, and it really shouldn't: at least
1208 as far as the programmer is concerned, it's not applied to two
1211 Note [Arity analysis]
1212 ~~~~~~~~~~~~~~~~~~~~~
1213 The motivating example for arity analysis is this:
1215 f = \x. let g = f (x+1)
1218 What arity does f have? Really it should have arity 2, but a naive
1219 look at the RHS won't see that. You need a fixpoint analysis which
1220 says it has arity "infinity" the first time round.
1222 This example happens a lot; it first showed up in Andy Gill's thesis,
1223 fifteen years ago! It also shows up in the code for 'rnf' on lists
1226 The analysis is easy to achieve because exprEtaExpandArity takes an
1228 type CheapFun = CoreExpr -> Maybe Type -> Bool
1229 used to decide if an expression is cheap enough to push inside a
1230 lambda. And exprIsCheap' in turn takes an argument
1231 type CheapAppFun = Id -> Int -> Bool
1232 which tells when an application is cheap. This makes it easy to
1233 write the analysis loop.
1235 The analysis is cheap-and-cheerful because it doesn't deal with
1236 mutual recursion. But the self-recursive case is the important one.
1239 %************************************************************************
1241 \subsection{Floating lets out of big lambdas}
1243 %************************************************************************
1245 Note [Floating and type abstraction]
1246 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
1249 We'd like to float this to
1252 x = /\a. C (y1 a) (y2 a)
1253 for the usual reasons: we want to inline x rather vigorously.
1255 You may think that this kind of thing is rare. But in some programs it is
1256 common. For example, if you do closure conversion you might get:
1258 data a :-> b = forall e. (e -> a -> b) :$ e
1260 f_cc :: forall a. a :-> a
1261 f_cc = /\a. (\e. id a) :$ ()
1263 Now we really want to inline that f_cc thing so that the
1264 construction of the closure goes away.
1266 So I have elaborated simplLazyBind to understand right-hand sides that look
1270 and treat them specially. The real work is done in SimplUtils.abstractFloats,
1271 but there is quite a bit of plumbing in simplLazyBind as well.
1273 The same transformation is good when there are lets in the body:
1275 /\abc -> let(rec) x = e in b
1277 let(rec) x' = /\abc -> let x = x' a b c in e
1279 /\abc -> let x = x' a b c in b
1281 This is good because it can turn things like:
1283 let f = /\a -> letrec g = ... g ... in g
1285 letrec g' = /\a -> ... g' a ...
1287 let f = /\ a -> g' a
1289 which is better. In effect, it means that big lambdas don't impede
1292 This optimisation is CRUCIAL in eliminating the junk introduced by
1293 desugaring mutually recursive definitions. Don't eliminate it lightly!
1295 [May 1999] If we do this transformation *regardless* then we can
1296 end up with some pretty silly stuff. For example,
1299 st = /\ s -> let { x1=r1 ; x2=r2 } in ...
1304 st = /\s -> ...[y1 s/x1, y2 s/x2]
1307 Unless the "..." is a WHNF there is really no point in doing this.
1308 Indeed it can make things worse. Suppose x1 is used strictly,
1311 x1* = case f y of { (a,b) -> e }
1313 If we abstract this wrt the tyvar we then can't do the case inline
1314 as we would normally do.
1316 That's why the whole transformation is part of the same process that
1317 floats let-bindings and constructor arguments out of RHSs. In particular,
1318 it is guarded by the doFloatFromRhs call in simplLazyBind.
1322 abstractFloats :: [OutTyVar] -> SimplEnv -> OutExpr -> SimplM ([OutBind], OutExpr)
1323 abstractFloats main_tvs body_env body
1324 = ASSERT( notNull body_floats )
1325 do { (subst, float_binds) <- mapAccumLM abstract empty_subst body_floats
1326 ; return (float_binds, CoreSubst.substExpr (text "abstract_floats1") subst body) }
1328 main_tv_set = mkVarSet main_tvs
1329 body_floats = getFloats body_env
1330 empty_subst = CoreSubst.mkEmptySubst (seInScope body_env)
1332 abstract :: CoreSubst.Subst -> OutBind -> SimplM (CoreSubst.Subst, OutBind)
1333 abstract subst (NonRec id rhs)
1334 = do { (poly_id, poly_app) <- mk_poly tvs_here id
1335 ; let poly_rhs = mkLams tvs_here rhs'
1336 subst' = CoreSubst.extendIdSubst subst id poly_app
1337 ; return (subst', (NonRec poly_id poly_rhs)) }
1339 rhs' = CoreSubst.substExpr (text "abstract_floats2") subst rhs
1340 tvs_here = varSetElems (main_tv_set `intersectVarSet` exprSomeFreeVars isTyVar rhs')
1342 -- Abstract only over the type variables free in the rhs
1343 -- wrt which the new binding is abstracted. But the naive
1344 -- approach of abstract wrt the tyvars free in the Id's type
1346 -- /\ a b -> let t :: (a,b) = (e1, e2)
1349 -- Here, b isn't free in x's type, but we must nevertheless
1350 -- abstract wrt b as well, because t's type mentions b.
1351 -- Since t is floated too, we'd end up with the bogus:
1352 -- poly_t = /\ a b -> (e1, e2)
1353 -- poly_x = /\ a -> fst (poly_t a *b*)
1354 -- So for now we adopt the even more naive approach of
1355 -- abstracting wrt *all* the tyvars. We'll see if that
1356 -- gives rise to problems. SLPJ June 98
1358 abstract subst (Rec prs)
1359 = do { (poly_ids, poly_apps) <- mapAndUnzipM (mk_poly tvs_here) ids
1360 ; let subst' = CoreSubst.extendSubstList subst (ids `zip` poly_apps)
1361 poly_rhss = [mkLams tvs_here (CoreSubst.substExpr (text "abstract_floats3") subst' rhs)
1363 ; return (subst', Rec (poly_ids `zip` poly_rhss)) }
1365 (ids,rhss) = unzip prs
1366 -- For a recursive group, it's a bit of a pain to work out the minimal
1367 -- set of tyvars over which to abstract:
1368 -- /\ a b c. let x = ...a... in
1369 -- letrec { p = ...x...q...
1370 -- q = .....p...b... } in
1372 -- Since 'x' is abstracted over 'a', the {p,q} group must be abstracted
1373 -- over 'a' (because x is replaced by (poly_x a)) as well as 'b'.
1374 -- Since it's a pain, we just use the whole set, which is always safe
1376 -- If you ever want to be more selective, remember this bizarre case too:
1378 -- Here, we must abstract 'x' over 'a'.
1381 mk_poly tvs_here var
1382 = do { uniq <- getUniqueM
1383 ; let poly_name = setNameUnique (idName var) uniq -- Keep same name
1384 poly_ty = mkForAllTys tvs_here (idType var) -- But new type of course
1385 poly_id = transferPolyIdInfo var tvs_here $ -- Note [transferPolyIdInfo] in Id.lhs
1386 mkLocalId poly_name poly_ty
1387 ; return (poly_id, mkTyApps (Var poly_id) (mkTyVarTys tvs_here)) }
1388 -- In the olden days, it was crucial to copy the occInfo of the original var,
1389 -- because we were looking at occurrence-analysed but as yet unsimplified code!
1390 -- In particular, we mustn't lose the loop breakers. BUT NOW we are looking
1391 -- at already simplified code, so it doesn't matter
1393 -- It's even right to retain single-occurrence or dead-var info:
1394 -- Suppose we started with /\a -> let x = E in B
1395 -- where x occurs once in B. Then we transform to:
1396 -- let x' = /\a -> E in /\a -> let x* = x' a in B
1397 -- where x* has an INLINE prag on it. Now, once x* is inlined,
1398 -- the occurrences of x' will be just the occurrences originally
1402 Note [Abstract over coercions]
1403 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
1404 If a coercion variable (g :: a ~ Int) is free in the RHS, then so is the
1405 type variable a. Rather than sort this mess out, we simply bale out and abstract
1406 wrt all the type variables if any of them are coercion variables.
1409 Historical note: if you use let-bindings instead of a substitution, beware of this:
1411 -- Suppose we start with:
1413 -- x = /\ a -> let g = G in E
1415 -- Then we'll float to get
1417 -- x = let poly_g = /\ a -> G
1418 -- in /\ a -> let g = poly_g a in E
1420 -- But now the occurrence analyser will see just one occurrence
1421 -- of poly_g, not inside a lambda, so the simplifier will
1422 -- PreInlineUnconditionally poly_g back into g! Badk to square 1!
1423 -- (I used to think that the "don't inline lone occurrences" stuff
1424 -- would stop this happening, but since it's the *only* occurrence,
1425 -- PreInlineUnconditionally kicks in first!)
1427 -- Solution: put an INLINE note on g's RHS, so that poly_g seems
1428 -- to appear many times. (NB: mkInlineMe eliminates
1429 -- such notes on trivial RHSs, so do it manually.)
1431 %************************************************************************
1435 %************************************************************************
1437 prepareAlts tries these things:
1439 1. Eliminate alternatives that cannot match, including the
1440 DEFAULT alternative.
1442 2. If the DEFAULT alternative can match only one possible constructor,
1443 then make that constructor explicit.
1445 case e of x { DEFAULT -> rhs }
1447 case e of x { (a,b) -> rhs }
1448 where the type is a single constructor type. This gives better code
1449 when rhs also scrutinises x or e.
1451 3. Returns a list of the constructors that cannot holds in the
1452 DEFAULT alternative (if there is one)
1454 Here "cannot match" includes knowledge from GADTs
1456 It's a good idea do do this stuff before simplifying the alternatives, to
1457 avoid simplifying alternatives we know can't happen, and to come up with
1458 the list of constructors that are handled, to put into the IdInfo of the
1459 case binder, for use when simplifying the alternatives.
1461 Eliminating the default alternative in (1) isn't so obvious, but it can
1464 data Colour = Red | Green | Blue
1473 DEFAULT -> [ case y of ... ]
1475 If we inline h into f, the default case of the inlined h can't happen.
1476 If we don't notice this, we may end up filtering out *all* the cases
1477 of the inner case y, which give us nowhere to go!
1480 prepareAlts :: OutExpr -> OutId -> [InAlt] -> SimplM ([AltCon], [InAlt])
1481 prepareAlts scrut case_bndr' alts
1482 = do { let (alts_wo_default, maybe_deflt) = findDefault alts
1483 alt_cons = [con | (con,_,_) <- alts_wo_default]
1484 imposs_deflt_cons = nub (imposs_cons ++ alt_cons)
1485 -- "imposs_deflt_cons" are handled
1486 -- EITHER by the context,
1487 -- OR by a non-DEFAULT branch in this case expression.
1489 ; default_alts <- prepareDefault case_bndr' mb_tc_app
1490 imposs_deflt_cons maybe_deflt
1492 ; let trimmed_alts = filterOut impossible_alt alts_wo_default
1493 merged_alts = mergeAlts trimmed_alts default_alts
1494 -- We need the mergeAlts in case the new default_alt
1495 -- has turned into a constructor alternative.
1496 -- The merge keeps the inner DEFAULT at the front, if there is one
1497 -- and interleaves the alternatives in the right order
1499 ; return (imposs_deflt_cons, merged_alts) }
1501 mb_tc_app = splitTyConApp_maybe (idType case_bndr')
1502 Just (_, inst_tys) = mb_tc_app
1504 imposs_cons = case scrut of
1505 Var v -> otherCons (idUnfolding v)
1508 impossible_alt :: CoreAlt -> Bool
1509 impossible_alt (con, _, _) | con `elem` imposs_cons = True
1510 impossible_alt (DataAlt con, _, _) = dataConCannotMatch inst_tys con
1511 impossible_alt _ = False
1514 prepareDefault :: OutId -- Case binder; need just for its type. Note that as an
1515 -- OutId, it has maximum information; this is important.
1516 -- Test simpl013 is an example
1517 -> Maybe (TyCon, [Type]) -- Type of scrutinee, decomposed
1518 -> [AltCon] -- These cons can't happen when matching the default
1519 -> Maybe InExpr -- Rhs
1520 -> SimplM [InAlt] -- Still unsimplified
1521 -- We use a list because it's what mergeAlts expects,
1523 --------- Fill in known constructor -----------
1524 prepareDefault case_bndr (Just (tycon, inst_tys)) imposs_cons (Just deflt_rhs)
1525 | -- This branch handles the case where we are
1526 -- scrutinisng an algebraic data type
1527 isAlgTyCon tycon -- It's a data type, tuple, or unboxed tuples.
1528 , not (isNewTyCon tycon) -- We can have a newtype, if we are just doing an eval:
1529 -- case x of { DEFAULT -> e }
1530 -- and we don't want to fill in a default for them!
1531 , Just all_cons <- tyConDataCons_maybe tycon
1532 , not (null all_cons)
1533 -- This is a tricky corner case. If the data type has no constructors,
1534 -- which GHC allows, then the case expression will have at most a default
1535 -- alternative. We don't want to eliminate that alternative, because the
1536 -- invariant is that there's always one alternative. It's more convenient
1538 -- case x of { DEFAULT -> e }
1539 -- as it is, rather than transform it to
1540 -- error "case cant match"
1541 -- which would be quite legitmate. But it's a really obscure corner, and
1542 -- not worth wasting code on.
1543 , let imposs_data_cons = [con | DataAlt con <- imposs_cons] -- We now know it's a data type
1544 impossible con = con `elem` imposs_data_cons || dataConCannotMatch inst_tys con
1545 = case filterOut impossible all_cons of
1546 [] -> return [] -- Eliminate the default alternative
1547 -- altogether if it can't match
1549 [con] -> -- It matches exactly one constructor, so fill it in
1550 do { tick (FillInCaseDefault case_bndr)
1552 ; let (ex_tvs, arg_ids) = dataConRepInstPat us con inst_tys
1553 ; return [(DataAlt con, ex_tvs ++ arg_ids, deflt_rhs)] }
1555 _ -> return [(DEFAULT, [], deflt_rhs)]
1557 | debugIsOn, isAlgTyCon tycon
1558 , null (tyConDataCons tycon)
1559 , not (isFamilyTyCon tycon || isAbstractTyCon tycon)
1560 -- Check for no data constructors
1561 -- This can legitimately happen for abstract types and type families,
1562 -- so don't report that
1563 = pprTrace "prepareDefault" (ppr case_bndr <+> ppr tycon)
1564 $ return [(DEFAULT, [], deflt_rhs)]
1566 --------- Catch-all cases -----------
1567 prepareDefault _case_bndr _bndr_ty _imposs_cons (Just deflt_rhs)
1568 = return [(DEFAULT, [], deflt_rhs)]
1570 prepareDefault _case_bndr _bndr_ty _imposs_cons Nothing
1571 = return [] -- No default branch
1576 %************************************************************************
1580 %************************************************************************
1582 mkCase tries these things
1584 1. Merge Nested Cases
1586 case e of b { ==> case e of b {
1587 p1 -> rhs1 p1 -> rhs1
1589 pm -> rhsm pm -> rhsm
1590 _ -> case b of b' { pn -> let b'=b in rhsn
1592 ... po -> let b'=b in rhso
1593 po -> rhso _ -> let b'=b in rhsd
1597 which merges two cases in one case when -- the default alternative of
1598 the outer case scrutises the same variable as the outer case. This
1599 transformation is called Case Merging. It avoids that the same
1600 variable is scrutinised multiple times.
1602 2. Eliminate Identity Case
1608 and similar friends.
1610 3. Merge identical alternatives.
1611 If several alternatives are identical, merge them into
1612 a single DEFAULT alternative. I've occasionally seen this
1613 making a big difference:
1615 case e of =====> case e of
1616 C _ -> f x D v -> ....v....
1617 D v -> ....v.... DEFAULT -> f x
1620 The point is that we merge common RHSs, at least for the DEFAULT case.
1621 [One could do something more elaborate but I've never seen it needed.]
1622 To avoid an expensive test, we just merge branches equal to the *first*
1623 alternative; this picks up the common cases
1624 a) all branches equal
1625 b) some branches equal to the DEFAULT (which occurs first)
1627 The case where Merge Identical Alternatives transformation showed up
1628 was like this (base/Foreign/C/Err/Error.lhs):
1634 where @is@ was something like
1636 p `is` n = p /= (-1) && p == n
1638 This gave rise to a horrible sequence of cases
1645 and similarly in cascade for all the join points!
1649 mkCase, mkCase1, mkCase2
1652 -> [OutAlt] -- Alternatives in standard (increasing) order
1655 --------------------------------------------------
1656 -- 1. Merge Nested Cases
1657 --------------------------------------------------
1659 mkCase dflags scrut outer_bndr ((DEFAULT, _, deflt_rhs) : outer_alts)
1660 | dopt Opt_CaseMerge dflags
1661 , Case (Var inner_scrut_var) inner_bndr _ inner_alts <- deflt_rhs
1662 , inner_scrut_var == outer_bndr
1663 = do { tick (CaseMerge outer_bndr)
1665 ; let wrap_alt (con, args, rhs) = ASSERT( outer_bndr `notElem` args )
1666 (con, args, wrap_rhs rhs)
1667 -- Simplifier's no-shadowing invariant should ensure
1668 -- that outer_bndr is not shadowed by the inner patterns
1669 wrap_rhs rhs = Let (NonRec inner_bndr (Var outer_bndr)) rhs
1670 -- The let is OK even for unboxed binders,
1672 wrapped_alts | isDeadBinder inner_bndr = inner_alts
1673 | otherwise = map wrap_alt inner_alts
1675 merged_alts = mergeAlts outer_alts wrapped_alts
1676 -- NB: mergeAlts gives priority to the left
1679 -- DEFAULT -> case x of
1682 -- When we merge, we must ensure that e1 takes
1683 -- precedence over e2 as the value for A!
1685 ; mkCase1 dflags scrut outer_bndr merged_alts
1687 -- Warning: don't call mkCase recursively!
1688 -- Firstly, there's no point, because inner alts have already had
1689 -- mkCase applied to them, so they won't have a case in their default
1690 -- Secondly, if you do, you get an infinite loop, because the bindCaseBndr
1691 -- in munge_rhs may put a case into the DEFAULT branch!
1693 mkCase dflags scrut bndr alts = mkCase1 dflags scrut bndr alts
1695 --------------------------------------------------
1696 -- 2. Eliminate Identity Case
1697 --------------------------------------------------
1699 mkCase1 _dflags scrut case_bndr alts -- Identity case
1700 | all identity_alt alts
1701 = do { tick (CaseIdentity case_bndr)
1702 ; return (re_cast scrut) }
1704 identity_alt (con, args, rhs) = check_eq con args (de_cast rhs)
1706 check_eq DEFAULT _ (Var v) = v == case_bndr
1707 check_eq (LitAlt lit') _ (Lit lit) = lit == lit'
1708 check_eq (DataAlt con) args rhs = rhs `cheapEqExpr` mkConApp con (arg_tys ++ varsToCoreExprs args)
1709 || rhs `cheapEqExpr` Var case_bndr
1710 check_eq _ _ _ = False
1712 arg_tys = map Type (tyConAppArgs (idType case_bndr))
1715 -- case e of x { _ -> x `cast` c }
1716 -- And we definitely want to eliminate this case, to give
1718 -- So we throw away the cast from the RHS, and reconstruct
1719 -- it at the other end. All the RHS casts must be the same
1720 -- if (all identity_alt alts) holds.
1722 -- Don't worry about nested casts, because the simplifier combines them
1723 de_cast (Cast e _) = e
1726 re_cast scrut = case head alts of
1727 (_,_,Cast _ co) -> Cast scrut co
1730 --------------------------------------------------
1731 -- 3. Merge Identical Alternatives
1732 --------------------------------------------------
1733 mkCase1 dflags scrut case_bndr ((_con1,bndrs1,rhs1) : con_alts)
1734 | all isDeadBinder bndrs1 -- Remember the default
1735 , length filtered_alts < length con_alts -- alternative comes first
1736 -- Also Note [Dead binders]
1737 = do { tick (AltMerge case_bndr)
1738 ; mkCase2 dflags scrut case_bndr alts' }
1740 alts' = (DEFAULT, [], rhs1) : filtered_alts
1741 filtered_alts = filter keep con_alts
1742 keep (_con,bndrs,rhs) = not (all isDeadBinder bndrs && rhs `cheapEqExpr` rhs1)
1744 mkCase1 dflags scrut bndr alts = mkCase2 dflags scrut bndr alts
1746 --------------------------------------------------
1748 --------------------------------------------------
1749 mkCase2 _dflags scrut bndr alts
1750 = return (Case scrut bndr (coreAltsType alts) alts)
1754 ~~~~~~~~~~~~~~~~~~~~
1755 Note that dead-ness is maintained by the simplifier, so that it is
1756 accurate after simplification as well as before.
1759 Note [Cascading case merge]
1760 ~~~~~~~~~~~~~~~~~~~~~~~~~~~
1761 Case merging should cascade in one sweep, because it
1765 DEFAULT -> case a of b
1766 DEFAULT -> case b of c {
1773 DEFAULT -> case a of b
1774 DEFAULT -> let c = b in e
1775 A -> let c = b in ea
1780 DEFAULT -> let b = a in let c = b in e
1781 A -> let b = a in let c = b in ea
1782 B -> let b = a in eb
1786 However here's a tricky case that we still don't catch, and I don't
1787 see how to catch it in one pass:
1789 case x of c1 { I# a1 ->
1792 DEFAULT -> case x of c3 { I# a2 ->
1795 After occurrence analysis (and its binder-swap) we get this
1797 case x of c1 { I# a1 ->
1798 let x = c1 in -- Binder-swap addition
1801 DEFAULT -> case x of c3 { I# a2 ->
1804 When we simplify the inner case x, we'll see that
1805 x=c1=I# a1. So we'll bind a2 to a1, and get
1807 case x of c1 { I# a1 ->
1810 DEFAULT -> case a1 of ...
1812 This is corect, but we can't do a case merge in this sweep
1813 because c2 /= a1. Reason: the binding c1=I# a1 went inwards
1814 without getting changed to c1=I# c2.
1816 I don't think this is worth fixing, even if I knew how. It'll
1817 all come out in the next pass anyway.