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 TcType ( isDictLikeTy )
49 import Type hiding( substTy )
50 import Coercion ( coercionKind )
52 import Unify ( dataConCannotMatch )
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 (CoerceIt _ cont) = contIsTrivial cont
212 contIsTrivial _ = False
215 contResultType :: SimplEnv -> OutType -> SimplCont -> OutType
216 contResultType env ty cont
219 subst_ty se ty = substTy (se `setInScope` env) ty
222 go (CoerceIt co cont) _ = go cont (snd (coercionKind co))
223 go (StrictBind _ bs body se cont) _ = go cont (subst_ty se (exprType (mkLams bs body)))
224 go (StrictArg ai _ cont) _ = go cont (funResultTy (argInfoResultTy ai))
225 go (Select _ _ alts se cont) _ = go cont (subst_ty se (coreAltsType alts))
226 go (ApplyTo _ arg se cont) ty = go cont (apply_to_arg ty arg se)
228 apply_to_arg ty (Type ty_arg) se = applyTy ty (subst_ty se ty_arg)
229 apply_to_arg ty _ _ = funResultTy ty
231 argInfoResultTy :: ArgInfo -> OutType
232 argInfoResultTy (ArgInfo { ai_fun = fun, ai_args = args })
233 = foldr (\arg fn_ty -> applyTypeToArg fn_ty arg) (idType fun) args
236 countValArgs :: SimplCont -> Int
237 countValArgs (ApplyTo _ (Type _) _ cont) = countValArgs cont
238 countValArgs (ApplyTo _ _ _ cont) = 1 + countValArgs cont
241 countArgs :: SimplCont -> Int
242 countArgs (ApplyTo _ _ _ cont) = 1 + countArgs cont
245 contArgs :: SimplCont -> (Bool, [ArgSummary], SimplCont)
246 -- Uses substitution to turn each arg into an OutExpr
247 contArgs cont@(ApplyTo {})
248 = case go [] cont of { (args, cont') -> (False, args, cont') }
250 go args (ApplyTo _ arg se cont)
251 | isTypeArg arg = go args cont
252 | otherwise = go (is_interesting arg se : args) cont
253 go args cont = (reverse args, cont)
255 is_interesting arg se = interestingArg (substExpr (text "contArgs") se arg)
256 -- Do *not* use short-cutting substitution here
257 -- because we want to get as much IdInfo as possible
259 contArgs cont = (True, [], cont)
261 pushSimplifiedArgs :: SimplEnv -> [CoreExpr] -> SimplCont -> SimplCont
262 pushSimplifiedArgs _env [] cont = cont
263 pushSimplifiedArgs env (arg:args) cont = ApplyTo Simplified arg env (pushSimplifiedArgs env args cont)
264 -- The env has an empty SubstEnv
266 dropArgs :: Int -> SimplCont -> SimplCont
267 dropArgs 0 cont = cont
268 dropArgs n (ApplyTo _ _ _ cont) = dropArgs (n-1) cont
269 dropArgs n other = pprPanic "dropArgs" (ppr n <+> ppr other)
273 Note [Interesting call context]
274 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
275 We want to avoid inlining an expression where there can't possibly be
276 any gain, such as in an argument position. Hence, if the continuation
277 is interesting (eg. a case scrutinee, application etc.) then we
278 inline, otherwise we don't.
280 Previously some_benefit used to return True only if the variable was
281 applied to some value arguments. This didn't work:
283 let x = _coerce_ (T Int) Int (I# 3) in
284 case _coerce_ Int (T Int) x of
287 we want to inline x, but can't see that it's a constructor in a case
288 scrutinee position, and some_benefit is False.
292 dMonadST = _/\_ t -> :Monad (g1 _@_ t, g2 _@_ t, g3 _@_ t)
294 .... case dMonadST _@_ x0 of (a,b,c) -> ....
296 we'd really like to inline dMonadST here, but we *don't* want to
297 inline if the case expression is just
299 case x of y { DEFAULT -> ... }
301 since we can just eliminate this case instead (x is in WHNF). Similar
302 applies when x is bound to a lambda expression. Hence
303 contIsInteresting looks for case expressions with just a single
308 interestingCallContext :: SimplCont -> CallCtxt
309 -- See Note [Interesting call context]
310 interestingCallContext cont
313 interesting (Select _ bndr _ _ _)
314 | isDeadBinder bndr = CaseCtxt
315 | otherwise = ArgCtxt False -- If the binder is used, this
316 -- is like a strict let
317 -- See Note [RHS of lets] in CoreUnfold
319 interesting (ApplyTo _ arg _ cont)
320 | isTypeArg arg = interesting cont
321 | otherwise = ValAppCtxt -- Can happen if we have (f Int |> co) y
322 -- If f has an INLINE prag we need to give it some
323 -- motivation to inline. See Note [Cast then apply]
326 interesting (StrictArg _ cci _) = cci
327 interesting (StrictBind {}) = BoringCtxt
328 interesting (Stop cci) = cci
329 interesting (CoerceIt _ cont) = interesting cont
330 -- If this call is the arg of a strict function, the context
331 -- is a bit interesting. If we inline here, we may get useful
332 -- evaluation information to avoid repeated evals: e.g.
334 -- Here the contIsInteresting makes the '*' keener to inline,
335 -- which in turn exposes a constructor which makes the '+' inline.
336 -- Assuming that +,* aren't small enough to inline regardless.
338 -- It's also very important to inline in a strict context for things
341 -- Here, the context of (f x) is strict, and if f's unfolding is
342 -- a build it's *great* to inline it here. So we must ensure that
343 -- the context for (f x) is not totally uninteresting.
348 -> [CoreRule] -- Rules for function
349 -> Int -- Number of value args
350 -> SimplCont -- Context of the call
353 mkArgInfo fun rules n_val_args call_cont
354 | n_val_args < idArity fun -- Note [Unsaturated functions]
355 = ArgInfo { ai_fun = fun, ai_args = [], ai_rules = rules
357 , ai_strs = vanilla_stricts
358 , ai_discs = vanilla_discounts }
360 = ArgInfo { ai_fun = fun, ai_args = [], ai_rules = rules
361 , ai_encl = interestingArgContext rules call_cont
362 , ai_strs = add_type_str (idType fun) arg_stricts
363 , ai_discs = arg_discounts }
365 vanilla_discounts, arg_discounts :: [Int]
366 vanilla_discounts = repeat 0
367 arg_discounts = case idUnfolding fun of
368 CoreUnfolding {uf_guidance = UnfIfGoodArgs {ug_args = discounts}}
369 -> discounts ++ vanilla_discounts
370 _ -> vanilla_discounts
372 vanilla_stricts, arg_stricts :: [Bool]
373 vanilla_stricts = repeat False
376 = case splitStrictSig (idStrictness fun) of
377 (demands, result_info)
378 | not (demands `lengthExceeds` n_val_args)
379 -> -- Enough args, use the strictness given.
380 -- For bottoming functions we used to pretend that the arg
381 -- is lazy, so that we don't treat the arg as an
382 -- interesting context. This avoids substituting
383 -- top-level bindings for (say) strings into
384 -- calls to error. But now we are more careful about
385 -- inlining lone variables, so its ok (see SimplUtils.analyseCont)
386 if isBotRes result_info then
387 map isStrictDmd demands -- Finite => result is bottom
389 map isStrictDmd demands ++ vanilla_stricts
391 -> WARN( True, text "More demands than arity" <+> ppr fun <+> ppr (idArity fun)
392 <+> ppr n_val_args <+> ppr demands )
393 vanilla_stricts -- Not enough args, or no strictness
395 add_type_str :: Type -> [Bool] -> [Bool]
396 -- If the function arg types are strict, record that in the 'strictness bits'
397 -- No need to instantiate because unboxed types (which dominate the strict
398 -- types) can't instantiate type variables.
399 -- add_type_str is done repeatedly (for each call); might be better
400 -- once-for-all in the function
401 -- But beware primops/datacons with no strictness
402 add_type_str _ [] = []
403 add_type_str fun_ty strs -- Look through foralls
404 | Just (_, fun_ty') <- splitForAllTy_maybe fun_ty -- Includes coercions
405 = add_type_str fun_ty' strs
406 add_type_str fun_ty (str:strs) -- Add strict-type info
407 | Just (arg_ty, fun_ty') <- splitFunTy_maybe fun_ty
408 = (str || isStrictType arg_ty) : add_type_str fun_ty' strs
412 {- Note [Unsaturated functions]
413 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
414 Consider (test eyeball/inline4)
417 where f has arity 2. Then we do not want to inline 'x', because
418 it'll just be floated out again. Even if f has lots of discounts
419 on its first argument -- it must be saturated for these to kick in
422 interestingArgContext :: [CoreRule] -> SimplCont -> Bool
423 -- If the argument has form (f x y), where x,y are boring,
424 -- and f is marked INLINE, then we don't want to inline f.
425 -- But if the context of the argument is
427 -- where g has rules, then we *do* want to inline f, in case it
428 -- exposes a rule that might fire. Similarly, if the context is
430 -- where h has rules, then we do want to inline f; hence the
431 -- call_cont argument to interestingArgContext
433 -- The ai-rules flag makes this happen; if it's
434 -- set, the inliner gets just enough keener to inline f
435 -- regardless of how boring f's arguments are, if it's marked INLINE
437 -- The alternative would be to *always* inline an INLINE function,
438 -- regardless of how boring its context is; but that seems overkill
439 -- For example, it'd mean that wrapper functions were always inlined
440 interestingArgContext rules call_cont
441 = notNull rules || enclosing_fn_has_rules
443 enclosing_fn_has_rules = go call_cont
445 go (Select {}) = False
446 go (ApplyTo {}) = False
447 go (StrictArg _ cci _) = interesting cci
448 go (StrictBind {}) = False -- ??
449 go (CoerceIt _ c) = go c
450 go (Stop cci) = interesting cci
452 interesting (ArgCtxt rules) = rules
453 interesting _ = False
457 %************************************************************************
461 %************************************************************************
463 The SimplifierMode controls several switches; see its definition in
465 sm_rules :: Bool -- Whether RULES are enabled
466 sm_inline :: Bool -- Whether inlining is enabled
467 sm_case_case :: Bool -- Whether case-of-case is enabled
468 sm_eta_expand :: Bool -- Whether eta-expansion is enabled
471 simplEnvForGHCi :: SimplEnv
472 simplEnvForGHCi = mkSimplEnv $
473 SimplMode { sm_names = ["GHCi"]
474 , sm_phase = InitialPhase
475 , sm_rules = True, sm_inline = False
476 , sm_eta_expand = False, sm_case_case = True }
477 -- Do not do any inlining, in case we expose some unboxed
478 -- tuple stuff that confuses the bytecode interpreter
480 updModeForInlineRules :: Activation -> SimplifierMode -> SimplifierMode
481 -- See Note [Simplifying inside InlineRules]
482 updModeForInlineRules inline_rule_act current_mode
483 = current_mode { sm_phase = phaseFromActivation inline_rule_act
486 , sm_eta_expand = False }
488 phaseFromActivation (ActiveAfter n) = Phase n
489 phaseFromActivation _ = InitialPhase
492 Note [Inlining in gentle mode]
493 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
494 Something is inlined if
495 (i) the sm_inline flag is on, AND
496 (ii) the thing has an INLINE pragma, AND
497 (iii) the thing is inlinable in the earliest phase.
499 Example of why (iii) is important:
500 {-# INLINE [~1] g #-}
506 If we were to inline g into f's inlining, then an importing module would
508 f e --> g (g e) ---> RULE fires
509 because the InlineRule for f has had g inlined into it.
511 On the other hand, it is bad not to do ANY inlining into an
512 InlineRule, because then recursive knots in instance declarations
513 don't get unravelled.
515 However, *sometimes* SimplGently must do no call-site inlining at all
516 (hence sm_inline = False). Before full laziness we must be careful
517 not to inline wrappers, because doing so inhibits floating
518 e.g. ...(case f x of ...)...
519 ==> ...(case (case x of I# x# -> fw x#) of ...)...
520 ==> ...(case x of I# x# -> case fw x# of ...)...
521 and now the redex (f x) isn't floatable any more.
523 The no-inlining thing is also important for Template Haskell. You might be
524 compiling in one-shot mode with -O2; but when TH compiles a splice before
525 running it, we don't want to use -O2. Indeed, we don't want to inline
526 anything, because the byte-code interpreter might get confused about
527 unboxed tuples and suchlike.
529 Note [Simplifying inside InlineRules]
530 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
531 We must take care with simplification inside InlineRules (which come from
534 First, consider the following example
539 in ...g...g...g...g...g...
540 Now, if that's the ONLY occurrence of f, it might be inlined inside g,
541 and thence copied multiple times when g is inlined. HENCE we treat
542 any occurrence in an InlineRule as a multiple occurrence, not a single
543 one; see OccurAnal.addRuleUsage.
545 Second, we do want *do* to some modest rules/inlining stuff in InlineRules,
546 partly to eliminate senseless crap, and partly to break the recursive knots
547 generated by instance declarations.
549 However, suppose we have
550 {-# INLINE <act> f #-}
552 meaning "inline f in phases p where activation <act>(p) holds".
553 Then what inlinings/rules can we apply to the copy of <rhs> captured in
554 f's InlineRule? Our model is that literally <rhs> is substituted for
555 f when it is inlined. So our conservative plan (implemented by
556 updModeForInlineRules) is this:
558 -------------------------------------------------------------
559 When simplifying the RHS of an InlineRule, set the phase to the
560 phase in which the InlineRule first becomes active
561 -------------------------------------------------------------
565 a) Rules/inlinings that *cease* being active before p will
566 not apply to the InlineRule rhs, consistent with it being
567 inlined in its *original* form in phase p.
569 b) Rules/inlinings that only become active *after* p will
570 not apply to the InlineRule rhs, again to be consistent with
571 inlining the *original* rhs in phase p.
577 {-# NOINLINE [1] g #-}
580 {-# RULE h g = ... #-}
581 Here we must not inline g into f's RHS, even when we get to phase 0,
582 because when f is later inlined into some other module we want the
590 and suppose that there are auto-generated specialisations and a strictness
591 wrapper for g. The specialisations get activation AlwaysActive, and the
592 strictness wrapper get activation (ActiveAfter 0). So the strictness
593 wrepper fails the test and won't be inlined into f's InlineRule. That
594 means f can inline, expose the specialised call to g, so the specialisation
597 A note about wrappers
598 ~~~~~~~~~~~~~~~~~~~~~
599 It's also important not to inline a worker back into a wrapper.
601 wraper = inline_me (\x -> ...worker... )
602 Normally, the inline_me prevents the worker getting inlined into
603 the wrapper (initially, the worker's only call site!). But,
604 if the wrapper is sure to be called, the strictness analyser will
605 mark it 'demanded', so when the RHS is simplified, it'll get an ArgOf
609 activeUnfolding :: SimplEnv -> Id -> Bool
611 | not (sm_inline mode) = active_unfolding_minimal
612 | otherwise = case sm_phase mode of
613 InitialPhase -> active_unfolding_gentle
614 Phase n -> active_unfolding n
618 getUnfoldingInRuleMatch :: SimplEnv -> IdUnfoldingFun
619 -- When matching in RULE, we want to "look through" an unfolding
620 -- (to see a constructor) if *rules* are on, even if *inlinings*
621 -- are not. A notable example is DFuns, which really we want to
622 -- match in rules like (op dfun) in gentle mode. Another example
623 -- is 'otherwise' which we want exprIsConApp_maybe to be able to
625 getUnfoldingInRuleMatch env id
626 | unf_is_active = idUnfolding id
627 | otherwise = NoUnfolding
631 | not (sm_rules mode) = active_unfolding_minimal id
632 | otherwise = isActive (sm_phase mode) (idInlineActivation id)
634 active_unfolding_minimal :: Id -> Bool
635 -- Compuslory unfoldings only
636 -- Ignore SimplGently, because we want to inline regardless;
637 -- the Id has no top-level binding at all
639 -- NB: we used to have a second exception, for data con wrappers.
640 -- On the grounds that we use gentle mode for rule LHSs, and
641 -- they match better when data con wrappers are inlined.
642 -- But that only really applies to the trivial wrappers (like (:)),
643 -- and they are now constructed as Compulsory unfoldings (in MkId)
644 -- so they'll happen anyway.
645 active_unfolding_minimal id = isCompulsoryUnfolding (realIdUnfolding id)
647 active_unfolding :: PhaseNum -> Id -> Bool
648 active_unfolding n id = isActiveIn n (idInlineActivation id)
650 active_unfolding_gentle :: Id -> Bool
651 -- Anything that is early-active
652 -- See Note [Gentle mode]
653 active_unfolding_gentle id
654 = isInlinePragma prag
655 && isEarlyActive (inlinePragmaActivation prag)
656 -- NB: wrappers are not early-active
658 prag = idInlinePragma id
660 ----------------------
661 activeRule :: DynFlags -> SimplEnv -> Maybe (Activation -> Bool)
662 -- Nothing => No rules at all
663 activeRule _dflags env
664 | not (sm_rules mode) = Nothing -- Rewriting is off
665 | otherwise = Just (isActive (sm_phase mode))
672 %************************************************************************
674 preInlineUnconditionally
676 %************************************************************************
678 preInlineUnconditionally
679 ~~~~~~~~~~~~~~~~~~~~~~~~
680 @preInlineUnconditionally@ examines a bndr to see if it is used just
681 once in a completely safe way, so that it is safe to discard the
682 binding inline its RHS at the (unique) usage site, REGARDLESS of how
683 big the RHS might be. If this is the case we don't simplify the RHS
684 first, but just inline it un-simplified.
686 This is much better than first simplifying a perhaps-huge RHS and then
687 inlining and re-simplifying it. Indeed, it can be at least quadratically
696 We may end up simplifying e1 N times, e2 N-1 times, e3 N-3 times etc.
697 This can happen with cascades of functions too:
704 THE MAIN INVARIANT is this:
706 ---- preInlineUnconditionally invariant -----
707 IF preInlineUnconditionally chooses to inline x = <rhs>
708 THEN doing the inlining should not change the occurrence
709 info for the free vars of <rhs>
710 ----------------------------------------------
712 For example, it's tempting to look at trivial binding like
714 and inline it unconditionally. But suppose x is used many times,
715 but this is the unique occurrence of y. Then inlining x would change
716 y's occurrence info, which breaks the invariant. It matters: y
717 might have a BIG rhs, which will now be dup'd at every occurrenc of x.
720 Even RHSs labelled InlineMe aren't caught here, because there might be
721 no benefit from inlining at the call site.
723 [Sept 01] Don't unconditionally inline a top-level thing, because that
724 can simply make a static thing into something built dynamically. E.g.
728 [Remember that we treat \s as a one-shot lambda.] No point in
729 inlining x unless there is something interesting about the call site.
731 But watch out: if you aren't careful, some useful foldr/build fusion
732 can be lost (most notably in spectral/hartel/parstof) because the
733 foldr didn't see the build. Doing the dynamic allocation isn't a big
734 deal, in fact, but losing the fusion can be. But the right thing here
735 seems to be to do a callSiteInline based on the fact that there is
736 something interesting about the call site (it's strict). Hmm. That
739 Conclusion: inline top level things gaily until Phase 0 (the last
740 phase), at which point don't.
742 Note [pre/postInlineUnconditionally in gentle mode]
743 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
744 Even in gentle mode we want to do preInlineUnconditionally. The
745 reason is that too little clean-up happens if you don't inline
746 use-once things. Also a bit of inlining is *good* for full laziness;
747 it can expose constant sub-expressions. Example in
748 spectral/mandel/Mandel.hs, where the mandelset function gets a useful
749 let-float if you inline windowToViewport
751 However, as usual for Gentle mode, do not inline things that are
752 inactive in the intial stages. See Note [Gentle mode].
754 Note [InlineRule and preInlineUnconditionally]
755 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
756 Surprisingly, do not pre-inline-unconditionally Ids with INLINE pragmas!
766 ...fInt...fInt...fInt...
768 Here f occurs just once, in the RHS of f1. But if we inline it there
769 we'll lose the opportunity to inline at each of fInt's call sites.
770 The INLINE pragma will only inline when the application is saturated
771 for exactly this reason; and we don't want PreInlineUnconditionally
772 to second-guess it. A live example is Trac #3736.
773 c.f. Note [InlineRule and postInlineUnconditionally]
775 Note [Top-level botomming Ids]
776 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
777 Don't inline top-level Ids that are bottoming, even if they are used just
778 once, because FloatOut has gone to some trouble to extract them out.
779 Inlining them won't make the program run faster!
782 preInlineUnconditionally :: SimplEnv -> TopLevelFlag -> InId -> InExpr -> Bool
783 preInlineUnconditionally env top_lvl bndr rhs
785 | isStableUnfolding (idUnfolding bndr) = False -- Note [InlineRule and preInlineUnconditionally]
786 | isTopLevel top_lvl && isBottomingId bndr = False -- Note [Top-level bottoming Ids]
787 | opt_SimplNoPreInlining = False
788 | otherwise = case idOccInfo bndr of
789 IAmDead -> True -- Happens in ((\x.1) v)
790 OneOcc in_lam True int_cxt -> try_once in_lam int_cxt
794 active = isActive (sm_phase mode) act
795 -- See Note [pre/postInlineUnconditionally in gentle mode]
796 act = idInlineActivation bndr
797 try_once in_lam int_cxt -- There's one textual occurrence
798 | not in_lam = isNotTopLevel top_lvl || early_phase
799 | otherwise = int_cxt && canInlineInLam rhs
801 -- Be very careful before inlining inside a lambda, because (a) we must not
802 -- invalidate occurrence information, and (b) we want to avoid pushing a
803 -- single allocation (here) into multiple allocations (inside lambda).
804 -- Inlining a *function* with a single *saturated* call would be ok, mind you.
805 -- || (if is_cheap && not (canInlineInLam rhs) then pprTrace "preinline" (ppr bndr <+> ppr rhs) ok else ok)
807 -- is_cheap = exprIsCheap rhs
808 -- ok = is_cheap && int_cxt
810 -- int_cxt The context isn't totally boring
811 -- E.g. let f = \ab.BIG in \y. map f xs
812 -- Don't want to substitute for f, because then we allocate
813 -- its closure every time the \y is called
814 -- But: let f = \ab.BIG in \y. map (f y) xs
815 -- Now we do want to substitute for f, even though it's not
816 -- saturated, because we're going to allocate a closure for
817 -- (f y) every time round the loop anyhow.
819 -- canInlineInLam => free vars of rhs are (Once in_lam) or Many,
820 -- so substituting rhs inside a lambda doesn't change the occ info.
821 -- Sadly, not quite the same as exprIsHNF.
822 canInlineInLam (Lit _) = True
823 canInlineInLam (Lam b e) = isRuntimeVar b || canInlineInLam e
824 canInlineInLam (Note _ e) = canInlineInLam e
825 canInlineInLam _ = False
827 early_phase = case sm_phase mode of
830 -- If we don't have this early_phase test, consider
831 -- x = length [1,2,3]
832 -- The full laziness pass carefully floats all the cons cells to
833 -- top level, and preInlineUnconditionally floats them all back in.
834 -- Result is (a) static allocation replaced by dynamic allocation
835 -- (b) many simplifier iterations because this tickles
836 -- a related problem; only one inlining per pass
838 -- On the other hand, I have seen cases where top-level fusion is
839 -- lost if we don't inline top level thing (e.g. string constants)
840 -- Hence the test for phase zero (which is the phase for all the final
841 -- simplifications). Until phase zero we take no special notice of
842 -- top level things, but then we become more leery about inlining
847 %************************************************************************
849 postInlineUnconditionally
851 %************************************************************************
853 postInlineUnconditionally
854 ~~~~~~~~~~~~~~~~~~~~~~~~~
855 @postInlineUnconditionally@ decides whether to unconditionally inline
856 a thing based on the form of its RHS; in particular if it has a
857 trivial RHS. If so, we can inline and discard the binding altogether.
859 NB: a loop breaker has must_keep_binding = True and non-loop-breakers
860 only have *forward* references Hence, it's safe to discard the binding
862 NOTE: This isn't our last opportunity to inline. We're at the binding
863 site right now, and we'll get another opportunity when we get to the
866 Note that we do this unconditional inlining only for trival RHSs.
867 Don't inline even WHNFs inside lambdas; doing so may simply increase
868 allocation when the function is called. This isn't the last chance; see
871 NB: Even inline pragmas (e.g. IMustBeINLINEd) are ignored here Why?
872 Because we don't even want to inline them into the RHS of constructor
873 arguments. See NOTE above
875 NB: At one time even NOINLINE was ignored here: if the rhs is trivial
876 it's best to inline it anyway. We often get a=E; b=a from desugaring,
877 with both a and b marked NOINLINE. But that seems incompatible with
878 our new view that inlining is like a RULE, so I'm sticking to the 'active'
882 postInlineUnconditionally
883 :: SimplEnv -> TopLevelFlag
884 -> OutId -- The binder (an InId would be fine too)
885 -> OccInfo -- From the InId
889 postInlineUnconditionally env top_lvl bndr occ_info rhs unfolding
891 | isLoopBreaker occ_info = False -- If it's a loop-breaker of any kind, don't inline
892 -- because it might be referred to "earlier"
893 | isExportedId bndr = False
894 | isStableUnfolding unfolding = False -- Note [InlineRule and postInlineUnconditionally]
895 | exprIsTrivial rhs = True
896 | isTopLevel top_lvl = False -- Note [Top level and postInlineUnconditionally]
899 -- The point of examining occ_info here is that for *non-values*
900 -- that occur outside a lambda, the call-site inliner won't have
901 -- a chance (becuase it doesn't know that the thing
902 -- only occurs once). The pre-inliner won't have gotten
903 -- it either, if the thing occurs in more than one branch
904 -- So the main target is things like
907 -- True -> case x of ...
908 -- False -> case x of ...
909 -- This is very important in practice; e.g. wheel-seive1 doubles
910 -- in allocation if you miss this out
911 OneOcc in_lam _one_br int_cxt -- OneOcc => no code-duplication issue
912 -> smallEnoughToInline unfolding -- Small enough to dup
913 -- ToDo: consider discount on smallEnoughToInline if int_cxt is true
915 -- NB: Do NOT inline arbitrarily big things, even if one_br is True
916 -- Reason: doing so risks exponential behaviour. We simplify a big
917 -- expression, inline it, and simplify it again. But if the
918 -- very same thing happens in the big expression, we get
920 -- PRINCIPLE: when we've already simplified an expression once,
921 -- make sure that we only inline it if it's reasonably small.
924 -- Outside a lambda, we want to be reasonably aggressive
925 -- about inlining into multiple branches of case
926 -- e.g. let x = <non-value>
927 -- in case y of { C1 -> ..x..; C2 -> ..x..; C3 -> ... }
928 -- Inlining can be a big win if C3 is the hot-spot, even if
929 -- the uses in C1, C2 are not 'interesting'
930 -- An example that gets worse if you add int_cxt here is 'clausify'
932 (isCheapUnfolding unfolding && int_cxt))
933 -- isCheap => acceptable work duplication; in_lam may be true
934 -- int_cxt to prevent us inlining inside a lambda without some
935 -- good reason. See the notes on int_cxt in preInlineUnconditionally
937 IAmDead -> True -- This happens; for example, the case_bndr during case of
938 -- known constructor: case (a,b) of x { (p,q) -> ... }
939 -- Here x isn't mentioned in the RHS, so we don't want to
940 -- create the (dead) let-binding let x = (a,b) in ...
944 -- Here's an example that we don't handle well:
945 -- let f = if b then Left (\x.BIG) else Right (\y.BIG)
946 -- in \y. ....case f of {...} ....
947 -- Here f is used just once, and duplicating the case work is fine (exprIsCheap).
949 -- - We can't preInlineUnconditionally because that woud invalidate
950 -- the occ info for b.
951 -- - We can't postInlineUnconditionally because the RHS is big, and
952 -- that risks exponential behaviour
953 -- - We can't call-site inline, because the rhs is big
957 active = isActive (sm_phase (getMode env)) (idInlineActivation bndr)
958 -- See Note [pre/postInlineUnconditionally in gentle mode]
961 Note [Top level and postInlineUnconditionally]
962 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
963 We don't do postInlineUnconditionally for top-level things (exept ones that
965 * There is no point, because the main goal is to get rid of local
966 bindings used in multiple case branches.
967 * Doing so will inline top-level error expressions that have been
968 carefully floated out by FloatOut. More generally, it might
969 replace static allocation with dynamic.
971 Note [InlineRule and postInlineUnconditionally]
972 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
973 Do not do postInlineUnconditionally if the Id has an InlineRule, otherwise
974 we lose the unfolding. Example
976 -- f has InlineRule with rhs (e |> co)
980 Then there's a danger we'll optimise to
985 and now postInlineUnconditionally, losing the InlineRule on f. Now f'
986 won't inline because 'e' is too big.
988 c.f. Note [InlineRule and preInlineUnconditionally]
991 %************************************************************************
995 %************************************************************************
998 mkLam :: SimplEnv -> [OutBndr] -> OutExpr -> SimplM OutExpr
999 -- mkLam tries three things
1000 -- a) eta reduction, if that gives a trivial expression
1001 -- b) eta expansion [only if there are some value lambdas]
1005 mkLam _env bndrs body
1006 = do { dflags <- getDOptsSmpl
1007 ; mkLam' dflags bndrs body }
1009 mkLam' :: DynFlags -> [OutBndr] -> OutExpr -> SimplM OutExpr
1010 mkLam' dflags bndrs (Cast body co)
1011 | not (any bad bndrs)
1012 -- Note [Casts and lambdas]
1013 = do { lam <- mkLam' dflags bndrs body
1014 ; return (mkCoerce (mkPiTypes bndrs co) lam) }
1016 co_vars = tyVarsOfType co
1017 bad bndr = isCoVar bndr && bndr `elemVarSet` co_vars
1019 mkLam' dflags bndrs body@(Lam {})
1020 = mkLam' dflags (bndrs ++ bndrs1) body1
1022 (bndrs1, body1) = collectBinders body
1024 mkLam' dflags bndrs body
1025 | dopt Opt_DoEtaReduction dflags
1026 , Just etad_lam <- tryEtaReduce bndrs body
1027 = do { tick (EtaReduction (head bndrs))
1031 = return (mkLams bndrs body)
1035 Note [Casts and lambdas]
1036 ~~~~~~~~~~~~~~~~~~~~~~~~
1038 (\x. (\y. e) `cast` g1) `cast` g2
1039 There is a danger here that the two lambdas look separated, and the
1040 full laziness pass might float an expression to between the two.
1042 So this equation in mkLam' floats the g1 out, thus:
1043 (\x. e `cast` g1) --> (\x.e) `cast` (tx -> g1)
1046 In general, this floats casts outside lambdas, where (I hope) they
1047 might meet and cancel with some other cast:
1048 \x. e `cast` co ===> (\x. e) `cast` (tx -> co)
1049 /\a. e `cast` co ===> (/\a. e) `cast` (/\a. co)
1050 /\g. e `cast` co ===> (/\g. e) `cast` (/\g. co)
1051 (if not (g `in` co))
1053 Notice that it works regardless of 'e'. Originally it worked only
1054 if 'e' was itself a lambda, but in some cases that resulted in
1055 fruitless iteration in the simplifier. A good example was when
1056 compiling Text.ParserCombinators.ReadPrec, where we had a definition
1057 like (\x. Get `cast` g)
1058 where Get is a constructor with nonzero arity. Then mkLam eta-expanded
1059 the Get, and the next iteration eta-reduced it, and then eta-expanded
1062 Note also the side condition for the case of coercion binders.
1063 It does not make sense to transform
1064 /\g. e `cast` g ==> (/\g.e) `cast` (/\g.g)
1065 because the latter is not well-kinded.
1067 %************************************************************************
1071 %************************************************************************
1073 When we meet a let-binding we try eta-expansion. To find the
1074 arity of the RHS we use a little fixpoint analysis; see Note [Arity analysis]
1077 tryEtaExpand :: SimplEnv -> OutId -> OutExpr -> SimplM (Arity, OutExpr)
1078 -- See Note [Eta-expanding at let bindings]
1079 tryEtaExpand env bndr rhs
1080 = do { dflags <- getDOptsSmpl
1081 ; (new_arity, new_rhs) <- try_expand dflags
1083 ; WARN( new_arity < old_arity || new_arity < _dmd_arity,
1084 (ptext (sLit "Arity decrease:") <+> (ppr bndr <+> ppr old_arity
1085 <+> ppr new_arity <+> ppr _dmd_arity) $$ ppr new_rhs) )
1086 -- Note [Arity decrease]
1087 return (new_arity, new_rhs) }
1090 | sm_eta_expand (getMode env) -- Provided eta-expansion is on
1091 , not (exprIsTrivial rhs)
1092 , let dicts_cheap = dopt Opt_DictsCheap dflags
1093 new_arity = findArity dicts_cheap bndr rhs old_arity
1094 , new_arity > rhs_arity
1095 = do { tick (EtaExpansion bndr)
1096 ; return (new_arity, etaExpand new_arity rhs) }
1098 = return (rhs_arity, rhs)
1100 rhs_arity = exprArity rhs
1101 old_arity = idArity bndr
1102 _dmd_arity = length $ fst $ splitStrictSig $ idStrictness bndr
1104 findArity :: Bool -> Id -> CoreExpr -> Arity -> Arity
1105 -- This implements the fixpoint loop for arity analysis
1106 -- See Note [Arity analysis]
1107 findArity dicts_cheap bndr rhs old_arity
1108 = go (exprEtaExpandArity (mk_cheap_fn dicts_cheap init_cheap_app) rhs)
1109 -- We always call exprEtaExpandArity once, but usually
1110 -- that produces a result equal to old_arity, and then
1111 -- we stop right away (since arities should not decrease)
1112 -- Result: the common case is that there is just one iteration
1114 go :: Arity -> Arity
1116 | cur_arity <= old_arity = cur_arity
1117 | new_arity == cur_arity = cur_arity
1118 | otherwise = ASSERT( new_arity < cur_arity )
1119 pprTrace "Exciting arity"
1120 (vcat [ ppr bndr <+> ppr cur_arity <+> ppr new_arity
1124 new_arity = exprEtaExpandArity (mk_cheap_fn dicts_cheap cheap_app) rhs
1126 cheap_app :: CheapAppFun
1127 cheap_app fn n_val_args
1128 | fn == bndr = n_val_args < cur_arity
1129 | otherwise = isCheapApp fn n_val_args
1131 init_cheap_app :: CheapAppFun
1132 init_cheap_app fn n_val_args
1134 | otherwise = isCheapApp fn n_val_args
1136 mk_cheap_fn :: Bool -> CheapAppFun -> CheapFun
1137 mk_cheap_fn dicts_cheap cheap_app
1139 = \e _ -> exprIsCheap' cheap_app e
1141 = \e mb_ty -> exprIsCheap' cheap_app e
1144 Just ty -> isDictLikeTy ty
1145 -- If the experimental -fdicts-cheap flag is on, we eta-expand through
1146 -- dictionary bindings. This improves arities. Thereby, it also
1147 -- means that full laziness is less prone to floating out the
1148 -- application of a function to its dictionary arguments, which
1149 -- can thereby lose opportunities for fusion. Example:
1150 -- foo :: Ord a => a -> ...
1151 -- foo = /\a \(d:Ord a). let d' = ...d... in \(x:a). ....
1152 -- -- So foo has arity 1
1154 -- f = \x. foo dInt $ bar x
1156 -- The (foo DInt) is floated out, and makes ineffective a RULE
1157 -- foo (bar x) = ...
1159 -- One could go further and make exprIsCheap reply True to any
1160 -- dictionary-typed expression, but that's more work.
1162 -- See Note [Dictionary-like types] in TcType.lhs for why we use
1163 -- isDictLikeTy here rather than isDictTy
1166 Note [Eta-expanding at let bindings]
1167 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
1168 We now eta expand at let-bindings, which is where the payoff
1171 One useful consequence is this example:
1172 genMap :: C a => ...
1173 {-# INLINE genMap #-}
1177 {-# INLINE myMap #-}
1180 Notice that 'genMap' should only inline if applied to two arguments.
1181 In the InlineRule for myMap we'll have the unfolding
1182 (\d -> genMap Int (..d..))
1183 We do not want to eta-expand to
1184 (\d f xs -> genMap Int (..d..) f xs)
1185 because then 'genMap' will inline, and it really shouldn't: at least
1186 as far as the programmer is concerned, it's not applied to two
1189 Note [Arity analysis]
1190 ~~~~~~~~~~~~~~~~~~~~~
1191 The motivating example for arity analysis is this:
1193 f = \x. let g = f (x+1)
1196 What arity does f have? Really it should have arity 2, but a naive
1197 look at the RHS won't see that. You need a fixpoint analysis which
1198 says it has arity "infinity" the first time round.
1200 This example happens a lot; it first showed up in Andy Gill's thesis,
1201 fifteen years ago! It also shows up in the code for 'rnf' on lists
1204 The analysis is easy to achieve because exprEtaExpandArity takes an
1206 type CheapFun = CoreExpr -> Maybe Type -> Bool
1207 used to decide if an expression is cheap enough to push inside a
1208 lambda. And exprIsCheap' in turn takes an argument
1209 type CheapAppFun = Id -> Int -> Bool
1210 which tells when an application is cheap. This makes it easy to
1211 write the analysis loop.
1213 The analysis is cheap-and-cheerful because it doesn't deal with
1214 mutual recursion. But the self-recursive case is the important one.
1217 %************************************************************************
1219 \subsection{Floating lets out of big lambdas}
1221 %************************************************************************
1223 Note [Floating and type abstraction]
1224 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
1227 We'd like to float this to
1230 x = /\a. C (y1 a) (y2 a)
1231 for the usual reasons: we want to inline x rather vigorously.
1233 You may think that this kind of thing is rare. But in some programs it is
1234 common. For example, if you do closure conversion you might get:
1236 data a :-> b = forall e. (e -> a -> b) :$ e
1238 f_cc :: forall a. a :-> a
1239 f_cc = /\a. (\e. id a) :$ ()
1241 Now we really want to inline that f_cc thing so that the
1242 construction of the closure goes away.
1244 So I have elaborated simplLazyBind to understand right-hand sides that look
1248 and treat them specially. The real work is done in SimplUtils.abstractFloats,
1249 but there is quite a bit of plumbing in simplLazyBind as well.
1251 The same transformation is good when there are lets in the body:
1253 /\abc -> let(rec) x = e in b
1255 let(rec) x' = /\abc -> let x = x' a b c in e
1257 /\abc -> let x = x' a b c in b
1259 This is good because it can turn things like:
1261 let f = /\a -> letrec g = ... g ... in g
1263 letrec g' = /\a -> ... g' a ...
1265 let f = /\ a -> g' a
1267 which is better. In effect, it means that big lambdas don't impede
1270 This optimisation is CRUCIAL in eliminating the junk introduced by
1271 desugaring mutually recursive definitions. Don't eliminate it lightly!
1273 [May 1999] If we do this transformation *regardless* then we can
1274 end up with some pretty silly stuff. For example,
1277 st = /\ s -> let { x1=r1 ; x2=r2 } in ...
1282 st = /\s -> ...[y1 s/x1, y2 s/x2]
1285 Unless the "..." is a WHNF there is really no point in doing this.
1286 Indeed it can make things worse. Suppose x1 is used strictly,
1289 x1* = case f y of { (a,b) -> e }
1291 If we abstract this wrt the tyvar we then can't do the case inline
1292 as we would normally do.
1294 That's why the whole transformation is part of the same process that
1295 floats let-bindings and constructor arguments out of RHSs. In particular,
1296 it is guarded by the doFloatFromRhs call in simplLazyBind.
1300 abstractFloats :: [OutTyVar] -> SimplEnv -> OutExpr -> SimplM ([OutBind], OutExpr)
1301 abstractFloats main_tvs body_env body
1302 = ASSERT( notNull body_floats )
1303 do { (subst, float_binds) <- mapAccumLM abstract empty_subst body_floats
1304 ; return (float_binds, CoreSubst.substExpr (text "abstract_floats1") subst body) }
1306 main_tv_set = mkVarSet main_tvs
1307 body_floats = getFloats body_env
1308 empty_subst = CoreSubst.mkEmptySubst (seInScope body_env)
1310 abstract :: CoreSubst.Subst -> OutBind -> SimplM (CoreSubst.Subst, OutBind)
1311 abstract subst (NonRec id rhs)
1312 = do { (poly_id, poly_app) <- mk_poly tvs_here id
1313 ; let poly_rhs = mkLams tvs_here rhs'
1314 subst' = CoreSubst.extendIdSubst subst id poly_app
1315 ; return (subst', (NonRec poly_id poly_rhs)) }
1317 rhs' = CoreSubst.substExpr (text "abstract_floats2") subst rhs
1318 tvs_here | any isCoVar main_tvs = main_tvs -- Note [Abstract over coercions]
1320 = varSetElems (main_tv_set `intersectVarSet` exprSomeFreeVars isTyCoVar rhs')
1322 -- Abstract only over the type variables free in the rhs
1323 -- wrt which the new binding is abstracted. But the naive
1324 -- approach of abstract wrt the tyvars free in the Id's type
1326 -- /\ a b -> let t :: (a,b) = (e1, e2)
1329 -- Here, b isn't free in x's type, but we must nevertheless
1330 -- abstract wrt b as well, because t's type mentions b.
1331 -- Since t is floated too, we'd end up with the bogus:
1332 -- poly_t = /\ a b -> (e1, e2)
1333 -- poly_x = /\ a -> fst (poly_t a *b*)
1334 -- So for now we adopt the even more naive approach of
1335 -- abstracting wrt *all* the tyvars. We'll see if that
1336 -- gives rise to problems. SLPJ June 98
1338 abstract subst (Rec prs)
1339 = do { (poly_ids, poly_apps) <- mapAndUnzipM (mk_poly tvs_here) ids
1340 ; let subst' = CoreSubst.extendSubstList subst (ids `zip` poly_apps)
1341 poly_rhss = [mkLams tvs_here (CoreSubst.substExpr (text "abstract_floats3") subst' rhs)
1343 ; return (subst', Rec (poly_ids `zip` poly_rhss)) }
1345 (ids,rhss) = unzip prs
1346 -- For a recursive group, it's a bit of a pain to work out the minimal
1347 -- set of tyvars over which to abstract:
1348 -- /\ a b c. let x = ...a... in
1349 -- letrec { p = ...x...q...
1350 -- q = .....p...b... } in
1352 -- Since 'x' is abstracted over 'a', the {p,q} group must be abstracted
1353 -- over 'a' (because x is replaced by (poly_x a)) as well as 'b'.
1354 -- Since it's a pain, we just use the whole set, which is always safe
1356 -- If you ever want to be more selective, remember this bizarre case too:
1358 -- Here, we must abstract 'x' over 'a'.
1361 mk_poly tvs_here var
1362 = do { uniq <- getUniqueM
1363 ; let poly_name = setNameUnique (idName var) uniq -- Keep same name
1364 poly_ty = mkForAllTys tvs_here (idType var) -- But new type of course
1365 poly_id = transferPolyIdInfo var tvs_here $ -- Note [transferPolyIdInfo] in Id.lhs
1366 mkLocalId poly_name poly_ty
1367 ; return (poly_id, mkTyApps (Var poly_id) (mkTyVarTys tvs_here)) }
1368 -- In the olden days, it was crucial to copy the occInfo of the original var,
1369 -- because we were looking at occurrence-analysed but as yet unsimplified code!
1370 -- In particular, we mustn't lose the loop breakers. BUT NOW we are looking
1371 -- at already simplified code, so it doesn't matter
1373 -- It's even right to retain single-occurrence or dead-var info:
1374 -- Suppose we started with /\a -> let x = E in B
1375 -- where x occurs once in B. Then we transform to:
1376 -- let x' = /\a -> E in /\a -> let x* = x' a in B
1377 -- where x* has an INLINE prag on it. Now, once x* is inlined,
1378 -- the occurrences of x' will be just the occurrences originally
1382 Note [Abstract over coercions]
1383 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
1384 If a coercion variable (g :: a ~ Int) is free in the RHS, then so is the
1385 type variable a. Rather than sort this mess out, we simply bale out and abstract
1386 wrt all the type variables if any of them are coercion variables.
1389 Historical note: if you use let-bindings instead of a substitution, beware of this:
1391 -- Suppose we start with:
1393 -- x = /\ a -> let g = G in E
1395 -- Then we'll float to get
1397 -- x = let poly_g = /\ a -> G
1398 -- in /\ a -> let g = poly_g a in E
1400 -- But now the occurrence analyser will see just one occurrence
1401 -- of poly_g, not inside a lambda, so the simplifier will
1402 -- PreInlineUnconditionally poly_g back into g! Badk to square 1!
1403 -- (I used to think that the "don't inline lone occurrences" stuff
1404 -- would stop this happening, but since it's the *only* occurrence,
1405 -- PreInlineUnconditionally kicks in first!)
1407 -- Solution: put an INLINE note on g's RHS, so that poly_g seems
1408 -- to appear many times. (NB: mkInlineMe eliminates
1409 -- such notes on trivial RHSs, so do it manually.)
1411 %************************************************************************
1415 %************************************************************************
1417 prepareAlts tries these things:
1419 1. Eliminate alternatives that cannot match, including the
1420 DEFAULT alternative.
1422 2. If the DEFAULT alternative can match only one possible constructor,
1423 then make that constructor explicit.
1425 case e of x { DEFAULT -> rhs }
1427 case e of x { (a,b) -> rhs }
1428 where the type is a single constructor type. This gives better code
1429 when rhs also scrutinises x or e.
1431 3. Returns a list of the constructors that cannot holds in the
1432 DEFAULT alternative (if there is one)
1434 Here "cannot match" includes knowledge from GADTs
1436 It's a good idea do do this stuff before simplifying the alternatives, to
1437 avoid simplifying alternatives we know can't happen, and to come up with
1438 the list of constructors that are handled, to put into the IdInfo of the
1439 case binder, for use when simplifying the alternatives.
1441 Eliminating the default alternative in (1) isn't so obvious, but it can
1444 data Colour = Red | Green | Blue
1453 DEFAULT -> [ case y of ... ]
1455 If we inline h into f, the default case of the inlined h can't happen.
1456 If we don't notice this, we may end up filtering out *all* the cases
1457 of the inner case y, which give us nowhere to go!
1460 prepareAlts :: OutExpr -> OutId -> [InAlt] -> SimplM ([AltCon], [InAlt])
1461 prepareAlts scrut case_bndr' alts
1462 = do { let (alts_wo_default, maybe_deflt) = findDefault alts
1463 alt_cons = [con | (con,_,_) <- alts_wo_default]
1464 imposs_deflt_cons = nub (imposs_cons ++ alt_cons)
1465 -- "imposs_deflt_cons" are handled
1466 -- EITHER by the context,
1467 -- OR by a non-DEFAULT branch in this case expression.
1469 ; default_alts <- prepareDefault case_bndr' mb_tc_app
1470 imposs_deflt_cons maybe_deflt
1472 ; let trimmed_alts = filterOut impossible_alt alts_wo_default
1473 merged_alts = mergeAlts trimmed_alts default_alts
1474 -- We need the mergeAlts in case the new default_alt
1475 -- has turned into a constructor alternative.
1476 -- The merge keeps the inner DEFAULT at the front, if there is one
1477 -- and interleaves the alternatives in the right order
1479 ; return (imposs_deflt_cons, merged_alts) }
1481 mb_tc_app = splitTyConApp_maybe (idType case_bndr')
1482 Just (_, inst_tys) = mb_tc_app
1484 imposs_cons = case scrut of
1485 Var v -> otherCons (idUnfolding v)
1488 impossible_alt :: CoreAlt -> Bool
1489 impossible_alt (con, _, _) | con `elem` imposs_cons = True
1490 impossible_alt (DataAlt con, _, _) = dataConCannotMatch inst_tys con
1491 impossible_alt _ = False
1494 prepareDefault :: OutId -- Case binder; need just for its type. Note that as an
1495 -- OutId, it has maximum information; this is important.
1496 -- Test simpl013 is an example
1497 -> Maybe (TyCon, [Type]) -- Type of scrutinee, decomposed
1498 -> [AltCon] -- These cons can't happen when matching the default
1499 -> Maybe InExpr -- Rhs
1500 -> SimplM [InAlt] -- Still unsimplified
1501 -- We use a list because it's what mergeAlts expects,
1503 --------- Fill in known constructor -----------
1504 prepareDefault case_bndr (Just (tycon, inst_tys)) imposs_cons (Just deflt_rhs)
1505 | -- This branch handles the case where we are
1506 -- scrutinisng an algebraic data type
1507 isAlgTyCon tycon -- It's a data type, tuple, or unboxed tuples.
1508 , not (isNewTyCon tycon) -- We can have a newtype, if we are just doing an eval:
1509 -- case x of { DEFAULT -> e }
1510 -- and we don't want to fill in a default for them!
1511 , Just all_cons <- tyConDataCons_maybe tycon
1512 , not (null all_cons)
1513 -- This is a tricky corner case. If the data type has no constructors,
1514 -- which GHC allows, then the case expression will have at most a default
1515 -- alternative. We don't want to eliminate that alternative, because the
1516 -- invariant is that there's always one alternative. It's more convenient
1518 -- case x of { DEFAULT -> e }
1519 -- as it is, rather than transform it to
1520 -- error "case cant match"
1521 -- which would be quite legitmate. But it's a really obscure corner, and
1522 -- not worth wasting code on.
1523 , let imposs_data_cons = [con | DataAlt con <- imposs_cons] -- We now know it's a data type
1524 impossible con = con `elem` imposs_data_cons || dataConCannotMatch inst_tys con
1525 = case filterOut impossible all_cons of
1526 [] -> return [] -- Eliminate the default alternative
1527 -- altogether if it can't match
1529 [con] -> -- It matches exactly one constructor, so fill it in
1530 do { tick (FillInCaseDefault case_bndr)
1532 ; let (ex_tvs, co_tvs, arg_ids) =
1533 dataConRepInstPat us con inst_tys
1534 ; return [(DataAlt con, ex_tvs ++ co_tvs ++ arg_ids, deflt_rhs)] }
1536 _ -> return [(DEFAULT, [], deflt_rhs)]
1538 | debugIsOn, isAlgTyCon tycon
1539 , null (tyConDataCons tycon)
1540 , not (isFamilyTyCon tycon || isAbstractTyCon tycon)
1541 -- Check for no data constructors
1542 -- This can legitimately happen for abstract types and type families,
1543 -- so don't report that
1544 = pprTrace "prepareDefault" (ppr case_bndr <+> ppr tycon)
1545 $ return [(DEFAULT, [], deflt_rhs)]
1547 --------- Catch-all cases -----------
1548 prepareDefault _case_bndr _bndr_ty _imposs_cons (Just deflt_rhs)
1549 = return [(DEFAULT, [], deflt_rhs)]
1551 prepareDefault _case_bndr _bndr_ty _imposs_cons Nothing
1552 = return [] -- No default branch
1557 %************************************************************************
1561 %************************************************************************
1563 mkCase tries these things
1565 1. Merge Nested Cases
1567 case e of b { ==> case e of b {
1568 p1 -> rhs1 p1 -> rhs1
1570 pm -> rhsm pm -> rhsm
1571 _ -> case b of b' { pn -> let b'=b in rhsn
1573 ... po -> let b'=b in rhso
1574 po -> rhso _ -> let b'=b in rhsd
1578 which merges two cases in one case when -- the default alternative of
1579 the outer case scrutises the same variable as the outer case. This
1580 transformation is called Case Merging. It avoids that the same
1581 variable is scrutinised multiple times.
1583 2. Eliminate Identity Case
1589 and similar friends.
1591 3. Merge identical alternatives.
1592 If several alternatives are identical, merge them into
1593 a single DEFAULT alternative. I've occasionally seen this
1594 making a big difference:
1596 case e of =====> case e of
1597 C _ -> f x D v -> ....v....
1598 D v -> ....v.... DEFAULT -> f x
1601 The point is that we merge common RHSs, at least for the DEFAULT case.
1602 [One could do something more elaborate but I've never seen it needed.]
1603 To avoid an expensive test, we just merge branches equal to the *first*
1604 alternative; this picks up the common cases
1605 a) all branches equal
1606 b) some branches equal to the DEFAULT (which occurs first)
1608 The case where Merge Identical Alternatives transformation showed up
1609 was like this (base/Foreign/C/Err/Error.lhs):
1615 where @is@ was something like
1617 p `is` n = p /= (-1) && p == n
1619 This gave rise to a horrible sequence of cases
1626 and similarly in cascade for all the join points!
1630 mkCase, mkCase1, mkCase2
1633 -> [OutAlt] -- Alternatives in standard (increasing) order
1636 --------------------------------------------------
1637 -- 1. Merge Nested Cases
1638 --------------------------------------------------
1640 mkCase dflags scrut outer_bndr ((DEFAULT, _, deflt_rhs) : outer_alts)
1641 | dopt Opt_CaseMerge dflags
1642 , Case (Var inner_scrut_var) inner_bndr _ inner_alts <- deflt_rhs
1643 , inner_scrut_var == outer_bndr
1644 = do { tick (CaseMerge outer_bndr)
1646 ; let wrap_alt (con, args, rhs) = ASSERT( outer_bndr `notElem` args )
1647 (con, args, wrap_rhs rhs)
1648 -- Simplifier's no-shadowing invariant should ensure
1649 -- that outer_bndr is not shadowed by the inner patterns
1650 wrap_rhs rhs = Let (NonRec inner_bndr (Var outer_bndr)) rhs
1651 -- The let is OK even for unboxed binders,
1653 wrapped_alts | isDeadBinder inner_bndr = inner_alts
1654 | otherwise = map wrap_alt inner_alts
1656 merged_alts = mergeAlts outer_alts wrapped_alts
1657 -- NB: mergeAlts gives priority to the left
1660 -- DEFAULT -> case x of
1663 -- When we merge, we must ensure that e1 takes
1664 -- precedence over e2 as the value for A!
1666 ; mkCase1 dflags scrut outer_bndr merged_alts
1668 -- Warning: don't call mkCase recursively!
1669 -- Firstly, there's no point, because inner alts have already had
1670 -- mkCase applied to them, so they won't have a case in their default
1671 -- Secondly, if you do, you get an infinite loop, because the bindCaseBndr
1672 -- in munge_rhs may put a case into the DEFAULT branch!
1674 mkCase dflags scrut bndr alts = mkCase1 dflags scrut bndr alts
1676 --------------------------------------------------
1677 -- 2. Eliminate Identity Case
1678 --------------------------------------------------
1680 mkCase1 _dflags scrut case_bndr alts -- Identity case
1681 | all identity_alt alts
1682 = do { tick (CaseIdentity case_bndr)
1683 ; return (re_cast scrut) }
1685 identity_alt (con, args, rhs) = check_eq con args (de_cast rhs)
1687 check_eq DEFAULT _ (Var v) = v == case_bndr
1688 check_eq (LitAlt lit') _ (Lit lit) = lit == lit'
1689 check_eq (DataAlt con) args rhs = rhs `cheapEqExpr` mkConApp con (arg_tys ++ varsToCoreExprs args)
1690 || rhs `cheapEqExpr` Var case_bndr
1691 check_eq _ _ _ = False
1693 arg_tys = map Type (tyConAppArgs (idType case_bndr))
1696 -- case e of x { _ -> x `cast` c }
1697 -- And we definitely want to eliminate this case, to give
1699 -- So we throw away the cast from the RHS, and reconstruct
1700 -- it at the other end. All the RHS casts must be the same
1701 -- if (all identity_alt alts) holds.
1703 -- Don't worry about nested casts, because the simplifier combines them
1704 de_cast (Cast e _) = e
1707 re_cast scrut = case head alts of
1708 (_,_,Cast _ co) -> Cast scrut co
1711 --------------------------------------------------
1712 -- 3. Merge Identical Alternatives
1713 --------------------------------------------------
1714 mkCase1 dflags scrut case_bndr ((_con1,bndrs1,rhs1) : con_alts)
1715 | all isDeadBinder bndrs1 -- Remember the default
1716 , length filtered_alts < length con_alts -- alternative comes first
1717 -- Also Note [Dead binders]
1718 = do { tick (AltMerge case_bndr)
1719 ; mkCase2 dflags scrut case_bndr alts' }
1721 alts' = (DEFAULT, [], rhs1) : filtered_alts
1722 filtered_alts = filter keep con_alts
1723 keep (_con,bndrs,rhs) = not (all isDeadBinder bndrs && rhs `cheapEqExpr` rhs1)
1725 mkCase1 dflags scrut bndr alts = mkCase2 dflags scrut bndr alts
1727 --------------------------------------------------
1729 --------------------------------------------------
1730 mkCase2 _dflags scrut bndr alts
1731 = return (Case scrut bndr (coreAltsType alts) alts)
1735 ~~~~~~~~~~~~~~~~~~~~
1736 Note that dead-ness is maintained by the simplifier, so that it is
1737 accurate after simplification as well as before.
1740 Note [Cascading case merge]
1741 ~~~~~~~~~~~~~~~~~~~~~~~~~~~
1742 Case merging should cascade in one sweep, because it
1746 DEFAULT -> case a of b
1747 DEFAULT -> case b of c {
1754 DEFAULT -> case a of b
1755 DEFAULT -> let c = b in e
1756 A -> let c = b in ea
1761 DEFAULT -> let b = a in let c = b in e
1762 A -> let b = a in let c = b in ea
1763 B -> let b = a in eb
1767 However here's a tricky case that we still don't catch, and I don't
1768 see how to catch it in one pass:
1770 case x of c1 { I# a1 ->
1773 DEFAULT -> case x of c3 { I# a2 ->
1776 After occurrence analysis (and its binder-swap) we get this
1778 case x of c1 { I# a1 ->
1779 let x = c1 in -- Binder-swap addition
1782 DEFAULT -> case x of c3 { I# a2 ->
1785 When we simplify the inner case x, we'll see that
1786 x=c1=I# a1. So we'll bind a2 to a1, and get
1788 case x of c1 { I# a1 ->
1791 DEFAULT -> case a1 of ...
1793 This is corect, but we can't do a case merge in this sweep
1794 because c2 /= a1. Reason: the binding c1=I# a1 went inwards
1795 without getting changed to c1=I# c2.
1797 I don't think this is worth fixing, even if I knew how. It'll
1798 all come out in the next pass anyway.