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
102 -- Invariant: never an identity coercion
106 DupFlag -- See Note [DupFlag invariants]
107 InExpr StaticEnv -- The argument and its static env
110 | Select -- case C of alts
111 DupFlag -- See Note [DupFlag invariants]
112 InId [InAlt] StaticEnv -- The case binder, alts, and subst-env
115 -- The two strict forms have no DupFlag, because we never duplicate them
116 | StrictBind -- (\x* \xs. e) C
117 InId [InBndr] -- let x* = [] in e
118 InExpr StaticEnv -- is a special case
121 | StrictArg -- f e1 ..en C
122 ArgInfo -- Specifies f, e1..en, Whether f has rules, etc
123 -- plus strictness flags for *further* args
124 CallCtxt -- Whether *this* argument position is interesting
129 ai_fun :: Id, -- The function
130 ai_args :: [OutExpr], -- ...applied to these args (which are in *reverse* order)
131 ai_rules :: [CoreRule], -- Rules for this function
133 ai_encl :: Bool, -- Flag saying whether this function
134 -- or an enclosing one has rules (recursively)
135 -- True => be keener to inline in all args
137 ai_strs :: [Bool], -- Strictness of remaining arguments
138 -- Usually infinite, but if it is finite it guarantees
139 -- that the function diverges after being given
140 -- that number of args
141 ai_discs :: [Int] -- Discounts for remaining arguments; non-zero => be keener to inline
145 addArgTo :: ArgInfo -> OutExpr -> ArgInfo
146 addArgTo ai arg = ai { ai_args = arg : ai_args ai }
148 instance Outputable SimplCont where
149 ppr (Stop interesting) = ptext (sLit "Stop") <> brackets (ppr interesting)
150 ppr (ApplyTo dup arg _ cont) = ((ptext (sLit "ApplyTo") <+> ppr dup <+> pprParendExpr arg)
151 {- $$ nest 2 (pprSimplEnv se) -}) $$ ppr cont
152 ppr (StrictBind b _ _ _ cont) = (ptext (sLit "StrictBind") <+> ppr b) $$ ppr cont
153 ppr (StrictArg ai _ cont) = (ptext (sLit "StrictArg") <+> ppr (ai_fun ai)) $$ ppr cont
154 ppr (Select dup bndr alts se cont) = (ptext (sLit "Select") <+> ppr dup <+> ppr bndr) $$
155 (nest 2 $ vcat [ppr (seTvSubst se), ppr alts]) $$ ppr cont
156 ppr (CoerceIt co cont) = (ptext (sLit "CoerceIt") <+> ppr co) $$ ppr cont
158 data DupFlag = NoDup -- Unsimplified, might be big
159 | Simplified -- Simplified
160 | OkToDup -- Simplified and small
162 isSimplified :: DupFlag -> Bool
163 isSimplified NoDup = False
164 isSimplified _ = True -- Invariant: the subst-env is empty
166 instance Outputable DupFlag where
167 ppr OkToDup = ptext (sLit "ok")
168 ppr NoDup = ptext (sLit "nodup")
169 ppr Simplified = ptext (sLit "simpl")
172 Note [DupFlag invariants]
173 ~~~~~~~~~~~~~~~~~~~~~~~~~
174 In both (ApplyTo dup _ env k)
175 and (Select dup _ _ env k)
176 the following invariants hold
178 (a) if dup = OkToDup, then continuation k is also ok-to-dup
179 (b) if dup = OkToDup or Simplified, the subst-env is empty
180 (and and hence no need to re-simplify)
184 mkBoringStop :: SimplCont
185 mkBoringStop = Stop BoringCtxt
187 mkRhsStop :: SimplCont -- See Note [RHS of lets] in CoreUnfold
188 mkRhsStop = Stop (ArgCtxt False)
190 mkLazyArgStop :: CallCtxt -> SimplCont
191 mkLazyArgStop cci = Stop cci
194 contIsRhsOrArg :: SimplCont -> Bool
195 contIsRhsOrArg (Stop {}) = True
196 contIsRhsOrArg (StrictBind {}) = True
197 contIsRhsOrArg (StrictArg {}) = True
198 contIsRhsOrArg _ = False
201 contIsDupable :: SimplCont -> Bool
202 contIsDupable (Stop {}) = True
203 contIsDupable (ApplyTo OkToDup _ _ _) = True -- See Note [DupFlag invariants]
204 contIsDupable (Select OkToDup _ _ _ _) = True -- ...ditto...
205 contIsDupable (CoerceIt _ cont) = contIsDupable cont
206 contIsDupable _ = False
209 contIsTrivial :: SimplCont -> Bool
210 contIsTrivial (Stop {}) = True
211 contIsTrivial (ApplyTo _ (Type _) _ cont) = contIsTrivial cont
212 contIsTrivial (ApplyTo _ (Coercion _) _ cont) = contIsTrivial cont
213 contIsTrivial (CoerceIt _ cont) = contIsTrivial cont
214 contIsTrivial _ = False
217 contResultType :: SimplEnv -> OutType -> SimplCont -> OutType
218 contResultType env ty cont
221 subst_ty se ty = SimplEnv.substTy (se `setInScope` env) ty
222 subst_co se co = SimplEnv.substCo (se `setInScope` env) co
225 go (CoerceIt co cont) _ = go cont (pSnd (coercionKind co))
226 go (StrictBind _ bs body se cont) _ = go cont (subst_ty se (exprType (mkLams bs body)))
227 go (StrictArg ai _ cont) _ = go cont (funResultTy (argInfoResultTy ai))
228 go (Select _ _ alts se cont) _ = go cont (subst_ty se (coreAltsType alts))
229 go (ApplyTo _ arg se cont) ty = go cont (apply_to_arg ty arg se)
231 apply_to_arg ty (Type ty_arg) se = applyTy ty (subst_ty se ty_arg)
232 apply_to_arg ty (Coercion co_arg) se = applyCo ty (subst_co se co_arg)
233 apply_to_arg ty _ _ = funResultTy ty
235 argInfoResultTy :: ArgInfo -> OutType
236 argInfoResultTy (ArgInfo { ai_fun = fun, ai_args = args })
237 = foldr (\arg fn_ty -> applyTypeToArg fn_ty arg) (idType fun) args
240 countValArgs :: SimplCont -> Int
241 countValArgs (ApplyTo _ (Type _) _ cont) = countValArgs cont
242 countValArgs (ApplyTo _ (Coercion _) _ cont) = countValArgs cont
243 countValArgs (ApplyTo _ _ _ cont) = 1 + countValArgs cont
246 countArgs :: SimplCont -> Int
247 countArgs (ApplyTo _ _ _ cont) = 1 + countArgs cont
250 contArgs :: SimplCont -> (Bool, [ArgSummary], SimplCont)
251 -- Uses substitution to turn each arg into an OutExpr
252 contArgs cont@(ApplyTo {})
253 = case go [] cont of { (args, cont') -> (False, args, cont') }
255 go args (ApplyTo _ arg se cont)
256 | isTypeArg arg = go args cont
257 | otherwise = go (is_interesting arg se : args) cont
258 go args cont = (reverse args, cont)
260 is_interesting arg se = interestingArg (substExpr (text "contArgs") se arg)
261 -- Do *not* use short-cutting substitution here
262 -- because we want to get as much IdInfo as possible
264 contArgs cont = (True, [], cont)
266 pushSimplifiedArgs :: SimplEnv -> [CoreExpr] -> SimplCont -> SimplCont
267 pushSimplifiedArgs _env [] cont = cont
268 pushSimplifiedArgs env (arg:args) cont = ApplyTo Simplified arg env (pushSimplifiedArgs env args cont)
269 -- The env has an empty SubstEnv
271 dropArgs :: Int -> SimplCont -> SimplCont
272 dropArgs 0 cont = cont
273 dropArgs n (ApplyTo _ _ _ cont) = dropArgs (n-1) cont
274 dropArgs n other = pprPanic "dropArgs" (ppr n <+> ppr other)
278 Note [Interesting call context]
279 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
280 We want to avoid inlining an expression where there can't possibly be
281 any gain, such as in an argument position. Hence, if the continuation
282 is interesting (eg. a case scrutinee, application etc.) then we
283 inline, otherwise we don't.
285 Previously some_benefit used to return True only if the variable was
286 applied to some value arguments. This didn't work:
288 let x = _coerce_ (T Int) Int (I# 3) in
289 case _coerce_ Int (T Int) x of
292 we want to inline x, but can't see that it's a constructor in a case
293 scrutinee position, and some_benefit is False.
297 dMonadST = _/\_ t -> :Monad (g1 _@_ t, g2 _@_ t, g3 _@_ t)
299 .... case dMonadST _@_ x0 of (a,b,c) -> ....
301 we'd really like to inline dMonadST here, but we *don't* want to
302 inline if the case expression is just
304 case x of y { DEFAULT -> ... }
306 since we can just eliminate this case instead (x is in WHNF). Similar
307 applies when x is bound to a lambda expression. Hence
308 contIsInteresting looks for case expressions with just a single
313 interestingCallContext :: SimplCont -> CallCtxt
314 -- See Note [Interesting call context]
315 interestingCallContext cont
318 interesting (Select _ bndr _ _ _)
319 | isDeadBinder bndr = CaseCtxt
320 | otherwise = ArgCtxt False -- If the binder is used, this
321 -- is like a strict let
322 -- See Note [RHS of lets] in CoreUnfold
324 interesting (ApplyTo _ arg _ cont)
325 | isTypeArg arg = interesting cont
326 | otherwise = ValAppCtxt -- Can happen if we have (f Int |> co) y
327 -- If f has an INLINE prag we need to give it some
328 -- motivation to inline. See Note [Cast then apply]
331 interesting (StrictArg _ cci _) = cci
332 interesting (StrictBind {}) = BoringCtxt
333 interesting (Stop cci) = cci
334 interesting (CoerceIt _ cont) = interesting cont
335 -- If this call is the arg of a strict function, the context
336 -- is a bit interesting. If we inline here, we may get useful
337 -- evaluation information to avoid repeated evals: e.g.
339 -- Here the contIsInteresting makes the '*' keener to inline,
340 -- which in turn exposes a constructor which makes the '+' inline.
341 -- Assuming that +,* aren't small enough to inline regardless.
343 -- It's also very important to inline in a strict context for things
346 -- Here, the context of (f x) is strict, and if f's unfolding is
347 -- a build it's *great* to inline it here. So we must ensure that
348 -- the context for (f x) is not totally uninteresting.
353 -> [CoreRule] -- Rules for function
354 -> Int -- Number of value args
355 -> SimplCont -- Context of the call
358 mkArgInfo fun rules n_val_args call_cont
359 | n_val_args < idArity fun -- Note [Unsaturated functions]
360 = ArgInfo { ai_fun = fun, ai_args = [], ai_rules = rules
362 , ai_strs = vanilla_stricts
363 , ai_discs = vanilla_discounts }
365 = ArgInfo { ai_fun = fun, ai_args = [], ai_rules = rules
366 , ai_encl = interestingArgContext rules call_cont
367 , ai_strs = add_type_str (idType fun) arg_stricts
368 , ai_discs = arg_discounts }
370 vanilla_discounts, arg_discounts :: [Int]
371 vanilla_discounts = repeat 0
372 arg_discounts = case idUnfolding fun of
373 CoreUnfolding {uf_guidance = UnfIfGoodArgs {ug_args = discounts}}
374 -> discounts ++ vanilla_discounts
375 _ -> vanilla_discounts
377 vanilla_stricts, arg_stricts :: [Bool]
378 vanilla_stricts = repeat False
381 = case splitStrictSig (idStrictness fun) of
382 (demands, result_info)
383 | not (demands `lengthExceeds` n_val_args)
384 -> -- Enough args, use the strictness given.
385 -- For bottoming functions we used to pretend that the arg
386 -- is lazy, so that we don't treat the arg as an
387 -- interesting context. This avoids substituting
388 -- top-level bindings for (say) strings into
389 -- calls to error. But now we are more careful about
390 -- inlining lone variables, so its ok (see SimplUtils.analyseCont)
391 if isBotRes result_info then
392 map isStrictDmd demands -- Finite => result is bottom
394 map isStrictDmd demands ++ vanilla_stricts
396 -> WARN( True, text "More demands than arity" <+> ppr fun <+> ppr (idArity fun)
397 <+> ppr n_val_args <+> ppr demands )
398 vanilla_stricts -- Not enough args, or no strictness
400 add_type_str :: Type -> [Bool] -> [Bool]
401 -- If the function arg types are strict, record that in the 'strictness bits'
402 -- No need to instantiate because unboxed types (which dominate the strict
403 -- types) can't instantiate type variables.
404 -- add_type_str is done repeatedly (for each call); might be better
405 -- once-for-all in the function
406 -- But beware primops/datacons with no strictness
407 add_type_str _ [] = []
408 add_type_str fun_ty strs -- Look through foralls
409 | Just (_, fun_ty') <- splitForAllTy_maybe fun_ty -- Includes coercions
410 = add_type_str fun_ty' strs
411 add_type_str fun_ty (str:strs) -- Add strict-type info
412 | Just (arg_ty, fun_ty') <- splitFunTy_maybe fun_ty
413 = (str || isStrictType arg_ty) : add_type_str fun_ty' strs
417 {- Note [Unsaturated functions]
418 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
419 Consider (test eyeball/inline4)
422 where f has arity 2. Then we do not want to inline 'x', because
423 it'll just be floated out again. Even if f has lots of discounts
424 on its first argument -- it must be saturated for these to kick in
427 interestingArgContext :: [CoreRule] -> SimplCont -> Bool
428 -- If the argument has form (f x y), where x,y are boring,
429 -- and f is marked INLINE, then we don't want to inline f.
430 -- But if the context of the argument is
432 -- where g has rules, then we *do* want to inline f, in case it
433 -- exposes a rule that might fire. Similarly, if the context is
435 -- where h has rules, then we do want to inline f; hence the
436 -- call_cont argument to interestingArgContext
438 -- The ai-rules flag makes this happen; if it's
439 -- set, the inliner gets just enough keener to inline f
440 -- regardless of how boring f's arguments are, if it's marked INLINE
442 -- The alternative would be to *always* inline an INLINE function,
443 -- regardless of how boring its context is; but that seems overkill
444 -- For example, it'd mean that wrapper functions were always inlined
445 interestingArgContext rules call_cont
446 = notNull rules || enclosing_fn_has_rules
448 enclosing_fn_has_rules = go call_cont
450 go (Select {}) = False
451 go (ApplyTo {}) = False
452 go (StrictArg _ cci _) = interesting cci
453 go (StrictBind {}) = False -- ??
454 go (CoerceIt _ c) = go c
455 go (Stop cci) = interesting cci
457 interesting (ArgCtxt rules) = rules
458 interesting _ = False
462 %************************************************************************
466 %************************************************************************
468 The SimplifierMode controls several switches; see its definition in
470 sm_rules :: Bool -- Whether RULES are enabled
471 sm_inline :: Bool -- Whether inlining is enabled
472 sm_case_case :: Bool -- Whether case-of-case is enabled
473 sm_eta_expand :: Bool -- Whether eta-expansion is enabled
476 simplEnvForGHCi :: DynFlags -> SimplEnv
477 simplEnvForGHCi dflags
478 = mkSimplEnv $ SimplMode { sm_names = ["GHCi"]
479 , sm_phase = InitialPhase
480 , sm_rules = rules_on
482 , sm_eta_expand = eta_expand_on
483 , sm_case_case = True }
485 rules_on = dopt Opt_EnableRewriteRules dflags
486 eta_expand_on = dopt Opt_DoLambdaEtaExpansion dflags
487 -- Do not do any inlining, in case we expose some unboxed
488 -- tuple stuff that confuses the bytecode interpreter
490 updModeForInlineRules :: Activation -> SimplifierMode -> SimplifierMode
491 -- See Note [Simplifying inside InlineRules]
492 updModeForInlineRules inline_rule_act current_mode
493 = current_mode { sm_phase = phaseFromActivation inline_rule_act
495 , sm_eta_expand = False }
496 -- For sm_rules, just inherit; sm_rules might be "off"
497 -- becuase of -fno-enable-rewrite-rules
499 phaseFromActivation (ActiveAfter n) = Phase n
500 phaseFromActivation _ = InitialPhase
503 Note [Inlining in gentle mode]
504 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
505 Something is inlined if
506 (i) the sm_inline flag is on, AND
507 (ii) the thing has an INLINE pragma, AND
508 (iii) the thing is inlinable in the earliest phase.
510 Example of why (iii) is important:
511 {-# INLINE [~1] g #-}
517 If we were to inline g into f's inlining, then an importing module would
519 f e --> g (g e) ---> RULE fires
520 because the InlineRule for f has had g inlined into it.
522 On the other hand, it is bad not to do ANY inlining into an
523 InlineRule, because then recursive knots in instance declarations
524 don't get unravelled.
526 However, *sometimes* SimplGently must do no call-site inlining at all
527 (hence sm_inline = False). Before full laziness we must be careful
528 not to inline wrappers, because doing so inhibits floating
529 e.g. ...(case f x of ...)...
530 ==> ...(case (case x of I# x# -> fw x#) of ...)...
531 ==> ...(case x of I# x# -> case fw x# of ...)...
532 and now the redex (f x) isn't floatable any more.
534 The no-inlining thing is also important for Template Haskell. You might be
535 compiling in one-shot mode with -O2; but when TH compiles a splice before
536 running it, we don't want to use -O2. Indeed, we don't want to inline
537 anything, because the byte-code interpreter might get confused about
538 unboxed tuples and suchlike.
540 Note [Simplifying inside InlineRules]
541 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
542 We must take care with simplification inside InlineRules (which come from
545 First, consider the following example
550 in ...g...g...g...g...g...
551 Now, if that's the ONLY occurrence of f, it might be inlined inside g,
552 and thence copied multiple times when g is inlined. HENCE we treat
553 any occurrence in an InlineRule as a multiple occurrence, not a single
554 one; see OccurAnal.addRuleUsage.
556 Second, we do want *do* to some modest rules/inlining stuff in InlineRules,
557 partly to eliminate senseless crap, and partly to break the recursive knots
558 generated by instance declarations.
560 However, suppose we have
561 {-# INLINE <act> f #-}
563 meaning "inline f in phases p where activation <act>(p) holds".
564 Then what inlinings/rules can we apply to the copy of <rhs> captured in
565 f's InlineRule? Our model is that literally <rhs> is substituted for
566 f when it is inlined. So our conservative plan (implemented by
567 updModeForInlineRules) is this:
569 -------------------------------------------------------------
570 When simplifying the RHS of an InlineRule, set the phase to the
571 phase in which the InlineRule first becomes active
572 -------------------------------------------------------------
576 a) Rules/inlinings that *cease* being active before p will
577 not apply to the InlineRule rhs, consistent with it being
578 inlined in its *original* form in phase p.
580 b) Rules/inlinings that only become active *after* p will
581 not apply to the InlineRule rhs, again to be consistent with
582 inlining the *original* rhs in phase p.
588 {-# NOINLINE [1] g #-}
591 {-# RULE h g = ... #-}
592 Here we must not inline g into f's RHS, even when we get to phase 0,
593 because when f is later inlined into some other module we want the
601 and suppose that there are auto-generated specialisations and a strictness
602 wrapper for g. The specialisations get activation AlwaysActive, and the
603 strictness wrapper get activation (ActiveAfter 0). So the strictness
604 wrepper fails the test and won't be inlined into f's InlineRule. That
605 means f can inline, expose the specialised call to g, so the specialisation
608 A note about wrappers
609 ~~~~~~~~~~~~~~~~~~~~~
610 It's also important not to inline a worker back into a wrapper.
612 wraper = inline_me (\x -> ...worker... )
613 Normally, the inline_me prevents the worker getting inlined into
614 the wrapper (initially, the worker's only call site!). But,
615 if the wrapper is sure to be called, the strictness analyser will
616 mark it 'demanded', so when the RHS is simplified, it'll get an ArgOf
620 activeUnfolding :: SimplEnv -> Id -> Bool
622 | not (sm_inline mode) = active_unfolding_minimal
623 | otherwise = case sm_phase mode of
624 InitialPhase -> active_unfolding_gentle
625 Phase n -> active_unfolding n
629 getUnfoldingInRuleMatch :: SimplEnv -> IdUnfoldingFun
630 -- When matching in RULE, we want to "look through" an unfolding
631 -- (to see a constructor) if *rules* are on, even if *inlinings*
632 -- are not. A notable example is DFuns, which really we want to
633 -- match in rules like (op dfun) in gentle mode. Another example
634 -- is 'otherwise' which we want exprIsConApp_maybe to be able to
636 getUnfoldingInRuleMatch env id
637 | unf_is_active = idUnfolding id
638 | otherwise = NoUnfolding
642 | not (sm_rules mode) = active_unfolding_minimal id
643 | otherwise = isActive (sm_phase mode) (idInlineActivation id)
645 active_unfolding_minimal :: Id -> Bool
646 -- Compuslory unfoldings only
647 -- Ignore SimplGently, because we want to inline regardless;
648 -- the Id has no top-level binding at all
650 -- NB: we used to have a second exception, for data con wrappers.
651 -- On the grounds that we use gentle mode for rule LHSs, and
652 -- they match better when data con wrappers are inlined.
653 -- But that only really applies to the trivial wrappers (like (:)),
654 -- and they are now constructed as Compulsory unfoldings (in MkId)
655 -- so they'll happen anyway.
656 active_unfolding_minimal id = isCompulsoryUnfolding (realIdUnfolding id)
658 active_unfolding :: PhaseNum -> Id -> Bool
659 active_unfolding n id = isActiveIn n (idInlineActivation id)
661 active_unfolding_gentle :: Id -> Bool
662 -- Anything that is early-active
663 -- See Note [Gentle mode]
664 active_unfolding_gentle id
665 = isInlinePragma prag
666 && isEarlyActive (inlinePragmaActivation prag)
667 -- NB: wrappers are not early-active
669 prag = idInlinePragma id
671 ----------------------
672 activeRule :: DynFlags -> SimplEnv -> Maybe (Activation -> Bool)
673 -- Nothing => No rules at all
674 activeRule _dflags env
675 | not (sm_rules mode) = Nothing -- Rewriting is off
676 | otherwise = Just (isActive (sm_phase mode))
683 %************************************************************************
685 preInlineUnconditionally
687 %************************************************************************
689 preInlineUnconditionally
690 ~~~~~~~~~~~~~~~~~~~~~~~~
691 @preInlineUnconditionally@ examines a bndr to see if it is used just
692 once in a completely safe way, so that it is safe to discard the
693 binding inline its RHS at the (unique) usage site, REGARDLESS of how
694 big the RHS might be. If this is the case we don't simplify the RHS
695 first, but just inline it un-simplified.
697 This is much better than first simplifying a perhaps-huge RHS and then
698 inlining and re-simplifying it. Indeed, it can be at least quadratically
707 We may end up simplifying e1 N times, e2 N-1 times, e3 N-3 times etc.
708 This can happen with cascades of functions too:
715 THE MAIN INVARIANT is this:
717 ---- preInlineUnconditionally invariant -----
718 IF preInlineUnconditionally chooses to inline x = <rhs>
719 THEN doing the inlining should not change the occurrence
720 info for the free vars of <rhs>
721 ----------------------------------------------
723 For example, it's tempting to look at trivial binding like
725 and inline it unconditionally. But suppose x is used many times,
726 but this is the unique occurrence of y. Then inlining x would change
727 y's occurrence info, which breaks the invariant. It matters: y
728 might have a BIG rhs, which will now be dup'd at every occurrenc of x.
731 Even RHSs labelled InlineMe aren't caught here, because there might be
732 no benefit from inlining at the call site.
734 [Sept 01] Don't unconditionally inline a top-level thing, because that
735 can simply make a static thing into something built dynamically. E.g.
739 [Remember that we treat \s as a one-shot lambda.] No point in
740 inlining x unless there is something interesting about the call site.
742 But watch out: if you aren't careful, some useful foldr/build fusion
743 can be lost (most notably in spectral/hartel/parstof) because the
744 foldr didn't see the build. Doing the dynamic allocation isn't a big
745 deal, in fact, but losing the fusion can be. But the right thing here
746 seems to be to do a callSiteInline based on the fact that there is
747 something interesting about the call site (it's strict). Hmm. That
750 Conclusion: inline top level things gaily until Phase 0 (the last
751 phase), at which point don't.
753 Note [pre/postInlineUnconditionally in gentle mode]
754 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
755 Even in gentle mode we want to do preInlineUnconditionally. The
756 reason is that too little clean-up happens if you don't inline
757 use-once things. Also a bit of inlining is *good* for full laziness;
758 it can expose constant sub-expressions. Example in
759 spectral/mandel/Mandel.hs, where the mandelset function gets a useful
760 let-float if you inline windowToViewport
762 However, as usual for Gentle mode, do not inline things that are
763 inactive in the intial stages. See Note [Gentle mode].
765 Note [InlineRule and preInlineUnconditionally]
766 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
767 Surprisingly, do not pre-inline-unconditionally Ids with INLINE pragmas!
777 ...fInt...fInt...fInt...
779 Here f occurs just once, in the RHS of f1. But if we inline it there
780 we'll lose the opportunity to inline at each of fInt's call sites.
781 The INLINE pragma will only inline when the application is saturated
782 for exactly this reason; and we don't want PreInlineUnconditionally
783 to second-guess it. A live example is Trac #3736.
784 c.f. Note [InlineRule and postInlineUnconditionally]
786 Note [Top-level botomming Ids]
787 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
788 Don't inline top-level Ids that are bottoming, even if they are used just
789 once, because FloatOut has gone to some trouble to extract them out.
790 Inlining them won't make the program run faster!
792 Note [Do not inline CoVars unconditionally]
793 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
794 Coercion variables appear inside coercions, and have a separate
795 substitution, so don't inline them via the IdSubst!
798 preInlineUnconditionally :: SimplEnv -> TopLevelFlag -> InId -> InExpr -> Bool
799 preInlineUnconditionally env top_lvl bndr rhs
801 | isStableUnfolding (idUnfolding bndr) = False -- Note [InlineRule and preInlineUnconditionally]
802 | isTopLevel top_lvl && isBottomingId bndr = False -- Note [Top-level bottoming Ids]
803 | opt_SimplNoPreInlining = False
804 | isCoVar bndr = False -- Note [Do not inline CoVars unconditionally]
805 | otherwise = case idOccInfo bndr of
806 IAmDead -> True -- Happens in ((\x.1) v)
807 OneOcc in_lam True int_cxt -> try_once in_lam int_cxt
811 active = isActive (sm_phase mode) act
812 -- See Note [pre/postInlineUnconditionally in gentle mode]
813 act = idInlineActivation bndr
814 try_once in_lam int_cxt -- There's one textual occurrence
815 | not in_lam = isNotTopLevel top_lvl || early_phase
816 | otherwise = int_cxt && canInlineInLam rhs
818 -- Be very careful before inlining inside a lambda, because (a) we must not
819 -- invalidate occurrence information, and (b) we want to avoid pushing a
820 -- single allocation (here) into multiple allocations (inside lambda).
821 -- Inlining a *function* with a single *saturated* call would be ok, mind you.
822 -- || (if is_cheap && not (canInlineInLam rhs) then pprTrace "preinline" (ppr bndr <+> ppr rhs) ok else ok)
824 -- is_cheap = exprIsCheap rhs
825 -- ok = is_cheap && int_cxt
827 -- int_cxt The context isn't totally boring
828 -- E.g. let f = \ab.BIG in \y. map f xs
829 -- Don't want to substitute for f, because then we allocate
830 -- its closure every time the \y is called
831 -- But: let f = \ab.BIG in \y. map (f y) xs
832 -- Now we do want to substitute for f, even though it's not
833 -- saturated, because we're going to allocate a closure for
834 -- (f y) every time round the loop anyhow.
836 -- canInlineInLam => free vars of rhs are (Once in_lam) or Many,
837 -- so substituting rhs inside a lambda doesn't change the occ info.
838 -- Sadly, not quite the same as exprIsHNF.
839 canInlineInLam (Lit _) = True
840 canInlineInLam (Lam b e) = isRuntimeVar b || canInlineInLam e
841 canInlineInLam (Note _ e) = canInlineInLam e
842 canInlineInLam _ = False
844 early_phase = case sm_phase mode of
847 -- If we don't have this early_phase test, consider
848 -- x = length [1,2,3]
849 -- The full laziness pass carefully floats all the cons cells to
850 -- top level, and preInlineUnconditionally floats them all back in.
851 -- Result is (a) static allocation replaced by dynamic allocation
852 -- (b) many simplifier iterations because this tickles
853 -- a related problem; only one inlining per pass
855 -- On the other hand, I have seen cases where top-level fusion is
856 -- lost if we don't inline top level thing (e.g. string constants)
857 -- Hence the test for phase zero (which is the phase for all the final
858 -- simplifications). Until phase zero we take no special notice of
859 -- top level things, but then we become more leery about inlining
864 %************************************************************************
866 postInlineUnconditionally
868 %************************************************************************
870 postInlineUnconditionally
871 ~~~~~~~~~~~~~~~~~~~~~~~~~
872 @postInlineUnconditionally@ decides whether to unconditionally inline
873 a thing based on the form of its RHS; in particular if it has a
874 trivial RHS. If so, we can inline and discard the binding altogether.
876 NB: a loop breaker has must_keep_binding = True and non-loop-breakers
877 only have *forward* references. Hence, it's safe to discard the binding
879 NOTE: This isn't our last opportunity to inline. We're at the binding
880 site right now, and we'll get another opportunity when we get to the
883 Note that we do this unconditional inlining only for trival RHSs.
884 Don't inline even WHNFs inside lambdas; doing so may simply increase
885 allocation when the function is called. This isn't the last chance; see
888 NB: Even inline pragmas (e.g. IMustBeINLINEd) are ignored here Why?
889 Because we don't even want to inline them into the RHS of constructor
890 arguments. See NOTE above
892 NB: At one time even NOINLINE was ignored here: if the rhs is trivial
893 it's best to inline it anyway. We often get a=E; b=a from desugaring,
894 with both a and b marked NOINLINE. But that seems incompatible with
895 our new view that inlining is like a RULE, so I'm sticking to the 'active'
899 postInlineUnconditionally
900 :: SimplEnv -> TopLevelFlag
901 -> OutId -- The binder (an InId would be fine too)
903 -> OccInfo -- From the InId
907 postInlineUnconditionally env top_lvl bndr occ_info rhs unfolding
909 | isLoopBreaker occ_info = False -- If it's a loop-breaker of any kind, don't inline
910 -- because it might be referred to "earlier"
911 | isExportedId bndr = False
912 | isStableUnfolding unfolding = False -- Note [InlineRule and postInlineUnconditionally]
913 | isTopLevel top_lvl = False -- Note [Top level and postInlineUnconditionally]
914 | exprIsTrivial rhs = True
917 -- The point of examining occ_info here is that for *non-values*
918 -- that occur outside a lambda, the call-site inliner won't have
919 -- a chance (becuase it doesn't know that the thing
920 -- only occurs once). The pre-inliner won't have gotten
921 -- it either, if the thing occurs in more than one branch
922 -- So the main target is things like
925 -- True -> case x of ...
926 -- False -> case x of ...
927 -- This is very important in practice; e.g. wheel-seive1 doubles
928 -- in allocation if you miss this out
929 OneOcc in_lam _one_br int_cxt -- OneOcc => no code-duplication issue
930 -> smallEnoughToInline unfolding -- Small enough to dup
931 -- ToDo: consider discount on smallEnoughToInline if int_cxt is true
933 -- NB: Do NOT inline arbitrarily big things, even if one_br is True
934 -- Reason: doing so risks exponential behaviour. We simplify a big
935 -- expression, inline it, and simplify it again. But if the
936 -- very same thing happens in the big expression, we get
938 -- PRINCIPLE: when we've already simplified an expression once,
939 -- make sure that we only inline it if it's reasonably small.
942 -- Outside a lambda, we want to be reasonably aggressive
943 -- about inlining into multiple branches of case
944 -- e.g. let x = <non-value>
945 -- in case y of { C1 -> ..x..; C2 -> ..x..; C3 -> ... }
946 -- Inlining can be a big win if C3 is the hot-spot, even if
947 -- the uses in C1, C2 are not 'interesting'
948 -- An example that gets worse if you add int_cxt here is 'clausify'
950 (isCheapUnfolding unfolding && int_cxt))
951 -- isCheap => acceptable work duplication; in_lam may be true
952 -- int_cxt to prevent us inlining inside a lambda without some
953 -- good reason. See the notes on int_cxt in preInlineUnconditionally
955 IAmDead -> True -- This happens; for example, the case_bndr during case of
956 -- known constructor: case (a,b) of x { (p,q) -> ... }
957 -- Here x isn't mentioned in the RHS, so we don't want to
958 -- create the (dead) let-binding let x = (a,b) in ...
962 -- Here's an example that we don't handle well:
963 -- let f = if b then Left (\x.BIG) else Right (\y.BIG)
964 -- in \y. ....case f of {...} ....
965 -- Here f is used just once, and duplicating the case work is fine (exprIsCheap).
967 -- - We can't preInlineUnconditionally because that woud invalidate
968 -- the occ info for b.
969 -- - We can't postInlineUnconditionally because the RHS is big, and
970 -- that risks exponential behaviour
971 -- - We can't call-site inline, because the rhs is big
975 active = isActive (sm_phase (getMode env)) (idInlineActivation bndr)
976 -- See Note [pre/postInlineUnconditionally in gentle mode]
979 Note [Top level and postInlineUnconditionally]
980 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
981 We don't do postInlineUnconditionally for top-level things (even for
982 ones that are trivial):
984 * Doing so will inline top-level error expressions that have been
985 carefully floated out by FloatOut. More generally, it might
986 replace static allocation with dynamic.
988 * Even for trivial expressions there's a problem. Consider
989 {-# RULE "foo" forall (xs::[T]). reverse xs = ruggle xs #-}
992 In one simplifier pass we might fire the rule, getting
994 but in *that* simplifier pass we must not do postInlineUnconditionally
995 on 'ruggle' because then we'll have an unbound occurrence of 'ruggle'
997 If the rhs is trivial it'll be inlined by callSiteInline, and then
998 the binding will be dead and discarded by the next use of OccurAnal
1000 * There is less point, because the main goal is to get rid of local
1001 bindings used in multiple case branches.
1004 Note [InlineRule and postInlineUnconditionally]
1005 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
1006 Do not do postInlineUnconditionally if the Id has an InlineRule, otherwise
1007 we lose the unfolding. Example
1009 -- f has InlineRule with rhs (e |> co)
1013 Then there's a danger we'll optimise to
1018 and now postInlineUnconditionally, losing the InlineRule on f. Now f'
1019 won't inline because 'e' is too big.
1021 c.f. Note [InlineRule and preInlineUnconditionally]
1024 %************************************************************************
1028 %************************************************************************
1031 mkLam :: SimplEnv -> [OutBndr] -> OutExpr -> SimplM OutExpr
1032 -- mkLam tries three things
1033 -- a) eta reduction, if that gives a trivial expression
1034 -- b) eta expansion [only if there are some value lambdas]
1038 mkLam _env bndrs body
1039 = do { dflags <- getDOptsSmpl
1040 ; mkLam' dflags bndrs body }
1042 mkLam' :: DynFlags -> [OutBndr] -> OutExpr -> SimplM OutExpr
1043 mkLam' dflags bndrs (Cast body co)
1044 | not (any bad bndrs)
1045 -- Note [Casts and lambdas]
1046 = do { lam <- mkLam' dflags bndrs body
1047 ; return (mkCoerce (mkPiCos bndrs co) lam) }
1049 co_vars = tyCoVarsOfCo co
1050 bad bndr = isCoVar bndr && bndr `elemVarSet` co_vars
1052 mkLam' dflags bndrs body@(Lam {})
1053 = mkLam' dflags (bndrs ++ bndrs1) body1
1055 (bndrs1, body1) = collectBinders body
1057 mkLam' dflags bndrs body
1058 | dopt Opt_DoEtaReduction dflags
1059 , Just etad_lam <- tryEtaReduce bndrs body
1060 = do { tick (EtaReduction (head bndrs))
1064 = return (mkLams bndrs body)
1068 Note [Casts and lambdas]
1069 ~~~~~~~~~~~~~~~~~~~~~~~~
1071 (\x. (\y. e) `cast` g1) `cast` g2
1072 There is a danger here that the two lambdas look separated, and the
1073 full laziness pass might float an expression to between the two.
1075 So this equation in mkLam' floats the g1 out, thus:
1076 (\x. e `cast` g1) --> (\x.e) `cast` (tx -> g1)
1079 In general, this floats casts outside lambdas, where (I hope) they
1080 might meet and cancel with some other cast:
1081 \x. e `cast` co ===> (\x. e) `cast` (tx -> co)
1082 /\a. e `cast` co ===> (/\a. e) `cast` (/\a. co)
1083 /\g. e `cast` co ===> (/\g. e) `cast` (/\g. co)
1084 (if not (g `in` co))
1086 Notice that it works regardless of 'e'. Originally it worked only
1087 if 'e' was itself a lambda, but in some cases that resulted in
1088 fruitless iteration in the simplifier. A good example was when
1089 compiling Text.ParserCombinators.ReadPrec, where we had a definition
1090 like (\x. Get `cast` g)
1091 where Get is a constructor with nonzero arity. Then mkLam eta-expanded
1092 the Get, and the next iteration eta-reduced it, and then eta-expanded
1095 Note also the side condition for the case of coercion binders.
1096 It does not make sense to transform
1097 /\g. e `cast` g ==> (/\g.e) `cast` (/\g.g)
1098 because the latter is not well-kinded.
1100 %************************************************************************
1104 %************************************************************************
1107 tryEtaExpand :: SimplEnv -> OutId -> OutExpr -> SimplM (Arity, OutExpr)
1108 -- See Note [Eta-expanding at let bindings]
1109 tryEtaExpand env bndr rhs
1110 = do { dflags <- getDOptsSmpl
1111 ; (new_arity, new_rhs) <- try_expand dflags
1113 ; WARN( new_arity < old_arity || new_arity < _dmd_arity,
1114 (ptext (sLit "Arity decrease:") <+> (ppr bndr <+> ppr old_arity
1115 <+> ppr new_arity <+> ppr _dmd_arity) $$ ppr new_rhs) )
1116 -- Note [Arity decrease]
1117 return (new_arity, new_rhs) }
1120 | sm_eta_expand (getMode env) -- Provided eta-expansion is on
1121 , not (exprIsTrivial rhs)
1122 , let dicts_cheap = dopt Opt_DictsCheap dflags
1123 new_arity = findArity dicts_cheap bndr rhs old_arity
1124 , new_arity > rhs_arity
1125 = do { tick (EtaExpansion bndr)
1126 ; return (new_arity, etaExpand new_arity rhs) }
1128 = return (rhs_arity, rhs)
1130 rhs_arity = exprArity rhs
1131 old_arity = idArity bndr
1132 _dmd_arity = length $ fst $ splitStrictSig $ idStrictness bndr
1134 findArity :: Bool -> Id -> CoreExpr -> Arity -> Arity
1135 -- This implements the fixpoint loop for arity analysis
1136 -- See Note [Arity analysis]
1137 findArity dicts_cheap bndr rhs old_arity
1138 = go (exprEtaExpandArity (mk_cheap_fn dicts_cheap init_cheap_app) rhs)
1139 -- We always call exprEtaExpandArity once, but usually
1140 -- that produces a result equal to old_arity, and then
1141 -- we stop right away (since arities should not decrease)
1142 -- Result: the common case is that there is just one iteration
1144 go :: Arity -> Arity
1146 | cur_arity <= old_arity = cur_arity
1147 | new_arity == cur_arity = cur_arity
1148 | otherwise = ASSERT( new_arity < cur_arity )
1149 pprTrace "Exciting arity"
1150 (vcat [ ppr bndr <+> ppr cur_arity <+> ppr new_arity
1154 new_arity = exprEtaExpandArity (mk_cheap_fn dicts_cheap cheap_app) rhs
1156 cheap_app :: CheapAppFun
1157 cheap_app fn n_val_args
1158 | fn == bndr = n_val_args < cur_arity
1159 | otherwise = isCheapApp fn n_val_args
1161 init_cheap_app :: CheapAppFun
1162 init_cheap_app fn n_val_args
1164 | otherwise = isCheapApp fn n_val_args
1166 mk_cheap_fn :: Bool -> CheapAppFun -> CheapFun
1167 mk_cheap_fn dicts_cheap cheap_app
1169 = \e _ -> exprIsCheap' cheap_app e
1171 = \e mb_ty -> exprIsCheap' cheap_app e
1174 Just ty -> isDictLikeTy ty
1175 -- If the experimental -fdicts-cheap flag is on, we eta-expand through
1176 -- dictionary bindings. This improves arities. Thereby, it also
1177 -- means that full laziness is less prone to floating out the
1178 -- application of a function to its dictionary arguments, which
1179 -- can thereby lose opportunities for fusion. Example:
1180 -- foo :: Ord a => a -> ...
1181 -- foo = /\a \(d:Ord a). let d' = ...d... in \(x:a). ....
1182 -- -- So foo has arity 1
1184 -- f = \x. foo dInt $ bar x
1186 -- The (foo DInt) is floated out, and makes ineffective a RULE
1187 -- foo (bar x) = ...
1189 -- One could go further and make exprIsCheap reply True to any
1190 -- dictionary-typed expression, but that's more work.
1192 -- See Note [Dictionary-like types] in TcType.lhs for why we use
1193 -- isDictLikeTy here rather than isDictTy
1196 Note [Eta-expanding at let bindings]
1197 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
1198 We now eta expand at let-bindings, which is where the payoff
1201 One useful consequence is this example:
1202 genMap :: C a => ...
1203 {-# INLINE genMap #-}
1207 {-# INLINE myMap #-}
1210 Notice that 'genMap' should only inline if applied to two arguments.
1211 In the InlineRule for myMap we'll have the unfolding
1212 (\d -> genMap Int (..d..))
1213 We do not want to eta-expand to
1214 (\d f xs -> genMap Int (..d..) f xs)
1215 because then 'genMap' will inline, and it really shouldn't: at least
1216 as far as the programmer is concerned, it's not applied to two
1219 Note [Arity analysis]
1220 ~~~~~~~~~~~~~~~~~~~~~
1221 The motivating example for arity analysis is this:
1223 f = \x. let g = f (x+1)
1226 What arity does f have? Really it should have arity 2, but a naive
1227 look at the RHS won't see that. You need a fixpoint analysis which
1228 says it has arity "infinity" the first time round.
1230 This example happens a lot; it first showed up in Andy Gill's thesis,
1231 fifteen years ago! It also shows up in the code for 'rnf' on lists
1234 The analysis is easy to achieve because exprEtaExpandArity takes an
1236 type CheapFun = CoreExpr -> Maybe Type -> Bool
1237 used to decide if an expression is cheap enough to push inside a
1238 lambda. And exprIsCheap' in turn takes an argument
1239 type CheapAppFun = Id -> Int -> Bool
1240 which tells when an application is cheap. This makes it easy to
1241 write the analysis loop.
1243 The analysis is cheap-and-cheerful because it doesn't deal with
1244 mutual recursion. But the self-recursive case is the important one.
1247 %************************************************************************
1249 \subsection{Floating lets out of big lambdas}
1251 %************************************************************************
1253 Note [Floating and type abstraction]
1254 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
1257 We'd like to float this to
1260 x = /\a. C (y1 a) (y2 a)
1261 for the usual reasons: we want to inline x rather vigorously.
1263 You may think that this kind of thing is rare. But in some programs it is
1264 common. For example, if you do closure conversion you might get:
1266 data a :-> b = forall e. (e -> a -> b) :$ e
1268 f_cc :: forall a. a :-> a
1269 f_cc = /\a. (\e. id a) :$ ()
1271 Now we really want to inline that f_cc thing so that the
1272 construction of the closure goes away.
1274 So I have elaborated simplLazyBind to understand right-hand sides that look
1278 and treat them specially. The real work is done in SimplUtils.abstractFloats,
1279 but there is quite a bit of plumbing in simplLazyBind as well.
1281 The same transformation is good when there are lets in the body:
1283 /\abc -> let(rec) x = e in b
1285 let(rec) x' = /\abc -> let x = x' a b c in e
1287 /\abc -> let x = x' a b c in b
1289 This is good because it can turn things like:
1291 let f = /\a -> letrec g = ... g ... in g
1293 letrec g' = /\a -> ... g' a ...
1295 let f = /\ a -> g' a
1297 which is better. In effect, it means that big lambdas don't impede
1300 This optimisation is CRUCIAL in eliminating the junk introduced by
1301 desugaring mutually recursive definitions. Don't eliminate it lightly!
1303 [May 1999] If we do this transformation *regardless* then we can
1304 end up with some pretty silly stuff. For example,
1307 st = /\ s -> let { x1=r1 ; x2=r2 } in ...
1312 st = /\s -> ...[y1 s/x1, y2 s/x2]
1315 Unless the "..." is a WHNF there is really no point in doing this.
1316 Indeed it can make things worse. Suppose x1 is used strictly,
1319 x1* = case f y of { (a,b) -> e }
1321 If we abstract this wrt the tyvar we then can't do the case inline
1322 as we would normally do.
1324 That's why the whole transformation is part of the same process that
1325 floats let-bindings and constructor arguments out of RHSs. In particular,
1326 it is guarded by the doFloatFromRhs call in simplLazyBind.
1330 abstractFloats :: [OutTyVar] -> SimplEnv -> OutExpr -> SimplM ([OutBind], OutExpr)
1331 abstractFloats main_tvs body_env body
1332 = ASSERT( notNull body_floats )
1333 do { (subst, float_binds) <- mapAccumLM abstract empty_subst body_floats
1334 ; return (float_binds, CoreSubst.substExpr (text "abstract_floats1") subst body) }
1336 main_tv_set = mkVarSet main_tvs
1337 body_floats = getFloats body_env
1338 empty_subst = CoreSubst.mkEmptySubst (seInScope body_env)
1340 abstract :: CoreSubst.Subst -> OutBind -> SimplM (CoreSubst.Subst, OutBind)
1341 abstract subst (NonRec id rhs)
1342 = do { (poly_id, poly_app) <- mk_poly tvs_here id
1343 ; let poly_rhs = mkLams tvs_here rhs'
1344 subst' = CoreSubst.extendIdSubst subst id poly_app
1345 ; return (subst', (NonRec poly_id poly_rhs)) }
1347 rhs' = CoreSubst.substExpr (text "abstract_floats2") subst rhs
1348 tvs_here = varSetElems (main_tv_set `intersectVarSet` exprSomeFreeVars isTyVar rhs')
1350 -- Abstract only over the type variables free in the rhs
1351 -- wrt which the new binding is abstracted. But the naive
1352 -- approach of abstract wrt the tyvars free in the Id's type
1354 -- /\ a b -> let t :: (a,b) = (e1, e2)
1357 -- Here, b isn't free in x's type, but we must nevertheless
1358 -- abstract wrt b as well, because t's type mentions b.
1359 -- Since t is floated too, we'd end up with the bogus:
1360 -- poly_t = /\ a b -> (e1, e2)
1361 -- poly_x = /\ a -> fst (poly_t a *b*)
1362 -- So for now we adopt the even more naive approach of
1363 -- abstracting wrt *all* the tyvars. We'll see if that
1364 -- gives rise to problems. SLPJ June 98
1366 abstract subst (Rec prs)
1367 = do { (poly_ids, poly_apps) <- mapAndUnzipM (mk_poly tvs_here) ids
1368 ; let subst' = CoreSubst.extendSubstList subst (ids `zip` poly_apps)
1369 poly_rhss = [mkLams tvs_here (CoreSubst.substExpr (text "abstract_floats3") subst' rhs)
1371 ; return (subst', Rec (poly_ids `zip` poly_rhss)) }
1373 (ids,rhss) = unzip prs
1374 -- For a recursive group, it's a bit of a pain to work out the minimal
1375 -- set of tyvars over which to abstract:
1376 -- /\ a b c. let x = ...a... in
1377 -- letrec { p = ...x...q...
1378 -- q = .....p...b... } in
1380 -- Since 'x' is abstracted over 'a', the {p,q} group must be abstracted
1381 -- over 'a' (because x is replaced by (poly_x a)) as well as 'b'.
1382 -- Since it's a pain, we just use the whole set, which is always safe
1384 -- If you ever want to be more selective, remember this bizarre case too:
1386 -- Here, we must abstract 'x' over 'a'.
1389 mk_poly tvs_here var
1390 = do { uniq <- getUniqueM
1391 ; let poly_name = setNameUnique (idName var) uniq -- Keep same name
1392 poly_ty = mkForAllTys tvs_here (idType var) -- But new type of course
1393 poly_id = transferPolyIdInfo var tvs_here $ -- Note [transferPolyIdInfo] in Id.lhs
1394 mkLocalId poly_name poly_ty
1395 ; return (poly_id, mkTyApps (Var poly_id) (mkTyVarTys tvs_here)) }
1396 -- In the olden days, it was crucial to copy the occInfo of the original var,
1397 -- because we were looking at occurrence-analysed but as yet unsimplified code!
1398 -- In particular, we mustn't lose the loop breakers. BUT NOW we are looking
1399 -- at already simplified code, so it doesn't matter
1401 -- It's even right to retain single-occurrence or dead-var info:
1402 -- Suppose we started with /\a -> let x = E in B
1403 -- where x occurs once in B. Then we transform to:
1404 -- let x' = /\a -> E in /\a -> let x* = x' a in B
1405 -- where x* has an INLINE prag on it. Now, once x* is inlined,
1406 -- the occurrences of x' will be just the occurrences originally
1410 Note [Abstract over coercions]
1411 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
1412 If a coercion variable (g :: a ~ Int) is free in the RHS, then so is the
1413 type variable a. Rather than sort this mess out, we simply bale out and abstract
1414 wrt all the type variables if any of them are coercion variables.
1417 Historical note: if you use let-bindings instead of a substitution, beware of this:
1419 -- Suppose we start with:
1421 -- x = /\ a -> let g = G in E
1423 -- Then we'll float to get
1425 -- x = let poly_g = /\ a -> G
1426 -- in /\ a -> let g = poly_g a in E
1428 -- But now the occurrence analyser will see just one occurrence
1429 -- of poly_g, not inside a lambda, so the simplifier will
1430 -- PreInlineUnconditionally poly_g back into g! Badk to square 1!
1431 -- (I used to think that the "don't inline lone occurrences" stuff
1432 -- would stop this happening, but since it's the *only* occurrence,
1433 -- PreInlineUnconditionally kicks in first!)
1435 -- Solution: put an INLINE note on g's RHS, so that poly_g seems
1436 -- to appear many times. (NB: mkInlineMe eliminates
1437 -- such notes on trivial RHSs, so do it manually.)
1439 %************************************************************************
1443 %************************************************************************
1445 prepareAlts tries these things:
1447 1. Eliminate alternatives that cannot match, including the
1448 DEFAULT alternative.
1450 2. If the DEFAULT alternative can match only one possible constructor,
1451 then make that constructor explicit.
1453 case e of x { DEFAULT -> rhs }
1455 case e of x { (a,b) -> rhs }
1456 where the type is a single constructor type. This gives better code
1457 when rhs also scrutinises x or e.
1459 3. Returns a list of the constructors that cannot holds in the
1460 DEFAULT alternative (if there is one)
1462 Here "cannot match" includes knowledge from GADTs
1464 It's a good idea do do this stuff before simplifying the alternatives, to
1465 avoid simplifying alternatives we know can't happen, and to come up with
1466 the list of constructors that are handled, to put into the IdInfo of the
1467 case binder, for use when simplifying the alternatives.
1469 Eliminating the default alternative in (1) isn't so obvious, but it can
1472 data Colour = Red | Green | Blue
1481 DEFAULT -> [ case y of ... ]
1483 If we inline h into f, the default case of the inlined h can't happen.
1484 If we don't notice this, we may end up filtering out *all* the cases
1485 of the inner case y, which give us nowhere to go!
1488 prepareAlts :: OutExpr -> OutId -> [InAlt] -> SimplM ([AltCon], [InAlt])
1489 prepareAlts scrut case_bndr' alts
1490 = do { let (alts_wo_default, maybe_deflt) = findDefault alts
1491 alt_cons = [con | (con,_,_) <- alts_wo_default]
1492 imposs_deflt_cons = nub (imposs_cons ++ alt_cons)
1493 -- "imposs_deflt_cons" are handled
1494 -- EITHER by the context,
1495 -- OR by a non-DEFAULT branch in this case expression.
1497 ; default_alts <- prepareDefault case_bndr' mb_tc_app
1498 imposs_deflt_cons maybe_deflt
1500 ; let trimmed_alts = filterOut impossible_alt alts_wo_default
1501 merged_alts = mergeAlts trimmed_alts default_alts
1502 -- We need the mergeAlts in case the new default_alt
1503 -- has turned into a constructor alternative.
1504 -- The merge keeps the inner DEFAULT at the front, if there is one
1505 -- and interleaves the alternatives in the right order
1507 ; return (imposs_deflt_cons, merged_alts) }
1509 mb_tc_app = splitTyConApp_maybe (idType case_bndr')
1510 Just (_, inst_tys) = mb_tc_app
1512 imposs_cons = case scrut of
1513 Var v -> otherCons (idUnfolding v)
1516 impossible_alt :: CoreAlt -> Bool
1517 impossible_alt (con, _, _) | con `elem` imposs_cons = True
1518 impossible_alt (DataAlt con, _, _) = dataConCannotMatch inst_tys con
1519 impossible_alt _ = False
1522 prepareDefault :: OutId -- Case binder; need just for its type. Note that as an
1523 -- OutId, it has maximum information; this is important.
1524 -- Test simpl013 is an example
1525 -> Maybe (TyCon, [Type]) -- Type of scrutinee, decomposed
1526 -> [AltCon] -- These cons can't happen when matching the default
1527 -> Maybe InExpr -- Rhs
1528 -> SimplM [InAlt] -- Still unsimplified
1529 -- We use a list because it's what mergeAlts expects,
1531 --------- Fill in known constructor -----------
1532 prepareDefault case_bndr (Just (tycon, inst_tys)) imposs_cons (Just deflt_rhs)
1533 | -- This branch handles the case where we are
1534 -- scrutinisng an algebraic data type
1535 isAlgTyCon tycon -- It's a data type, tuple, or unboxed tuples.
1536 , not (isNewTyCon tycon) -- We can have a newtype, if we are just doing an eval:
1537 -- case x of { DEFAULT -> e }
1538 -- and we don't want to fill in a default for them!
1539 , Just all_cons <- tyConDataCons_maybe tycon
1540 , not (null all_cons)
1541 -- This is a tricky corner case. If the data type has no constructors,
1542 -- which GHC allows, then the case expression will have at most a default
1543 -- alternative. We don't want to eliminate that alternative, because the
1544 -- invariant is that there's always one alternative. It's more convenient
1546 -- case x of { DEFAULT -> e }
1547 -- as it is, rather than transform it to
1548 -- error "case cant match"
1549 -- which would be quite legitmate. But it's a really obscure corner, and
1550 -- not worth wasting code on.
1551 , let imposs_data_cons = [con | DataAlt con <- imposs_cons] -- We now know it's a data type
1552 impossible con = con `elem` imposs_data_cons || dataConCannotMatch inst_tys con
1553 = case filterOut impossible all_cons of
1554 [] -> return [] -- Eliminate the default alternative
1555 -- altogether if it can't match
1557 [con] -> -- It matches exactly one constructor, so fill it in
1558 do { tick (FillInCaseDefault case_bndr)
1560 ; let (ex_tvs, arg_ids) = dataConRepInstPat us con inst_tys
1561 ; return [(DataAlt con, ex_tvs ++ arg_ids, deflt_rhs)] }
1563 _ -> return [(DEFAULT, [], deflt_rhs)]
1565 | debugIsOn, isAlgTyCon tycon
1566 , null (tyConDataCons tycon)
1567 , not (isFamilyTyCon tycon || isAbstractTyCon tycon)
1568 -- Check for no data constructors
1569 -- This can legitimately happen for abstract types and type families,
1570 -- so don't report that
1571 = pprTrace "prepareDefault" (ppr case_bndr <+> ppr tycon)
1572 $ return [(DEFAULT, [], deflt_rhs)]
1574 --------- Catch-all cases -----------
1575 prepareDefault _case_bndr _bndr_ty _imposs_cons (Just deflt_rhs)
1576 = return [(DEFAULT, [], deflt_rhs)]
1578 prepareDefault _case_bndr _bndr_ty _imposs_cons Nothing
1579 = return [] -- No default branch
1584 %************************************************************************
1588 %************************************************************************
1590 mkCase tries these things
1592 1. Merge Nested Cases
1594 case e of b { ==> case e of b {
1595 p1 -> rhs1 p1 -> rhs1
1597 pm -> rhsm pm -> rhsm
1598 _ -> case b of b' { pn -> let b'=b in rhsn
1600 ... po -> let b'=b in rhso
1601 po -> rhso _ -> let b'=b in rhsd
1605 which merges two cases in one case when -- the default alternative of
1606 the outer case scrutises the same variable as the outer case. This
1607 transformation is called Case Merging. It avoids that the same
1608 variable is scrutinised multiple times.
1610 2. Eliminate Identity Case
1616 and similar friends.
1618 3. Merge identical alternatives.
1619 If several alternatives are identical, merge them into
1620 a single DEFAULT alternative. I've occasionally seen this
1621 making a big difference:
1623 case e of =====> case e of
1624 C _ -> f x D v -> ....v....
1625 D v -> ....v.... DEFAULT -> f x
1628 The point is that we merge common RHSs, at least for the DEFAULT case.
1629 [One could do something more elaborate but I've never seen it needed.]
1630 To avoid an expensive test, we just merge branches equal to the *first*
1631 alternative; this picks up the common cases
1632 a) all branches equal
1633 b) some branches equal to the DEFAULT (which occurs first)
1635 The case where Merge Identical Alternatives transformation showed up
1636 was like this (base/Foreign/C/Err/Error.lhs):
1642 where @is@ was something like
1644 p `is` n = p /= (-1) && p == n
1646 This gave rise to a horrible sequence of cases
1653 and similarly in cascade for all the join points!
1657 mkCase, mkCase1, mkCase2
1660 -> [OutAlt] -- Alternatives in standard (increasing) order
1663 --------------------------------------------------
1664 -- 1. Merge Nested Cases
1665 --------------------------------------------------
1667 mkCase dflags scrut outer_bndr ((DEFAULT, _, deflt_rhs) : outer_alts)
1668 | dopt Opt_CaseMerge dflags
1669 , Case (Var inner_scrut_var) inner_bndr _ inner_alts <- deflt_rhs
1670 , inner_scrut_var == outer_bndr
1671 = do { tick (CaseMerge outer_bndr)
1673 ; let wrap_alt (con, args, rhs) = ASSERT( outer_bndr `notElem` args )
1674 (con, args, wrap_rhs rhs)
1675 -- Simplifier's no-shadowing invariant should ensure
1676 -- that outer_bndr is not shadowed by the inner patterns
1677 wrap_rhs rhs = Let (NonRec inner_bndr (Var outer_bndr)) rhs
1678 -- The let is OK even for unboxed binders,
1680 wrapped_alts | isDeadBinder inner_bndr = inner_alts
1681 | otherwise = map wrap_alt inner_alts
1683 merged_alts = mergeAlts outer_alts wrapped_alts
1684 -- NB: mergeAlts gives priority to the left
1687 -- DEFAULT -> case x of
1690 -- When we merge, we must ensure that e1 takes
1691 -- precedence over e2 as the value for A!
1693 ; mkCase1 dflags scrut outer_bndr merged_alts
1695 -- Warning: don't call mkCase recursively!
1696 -- Firstly, there's no point, because inner alts have already had
1697 -- mkCase applied to them, so they won't have a case in their default
1698 -- Secondly, if you do, you get an infinite loop, because the bindCaseBndr
1699 -- in munge_rhs may put a case into the DEFAULT branch!
1701 mkCase dflags scrut bndr alts = mkCase1 dflags scrut bndr alts
1703 --------------------------------------------------
1704 -- 2. Eliminate Identity Case
1705 --------------------------------------------------
1707 mkCase1 _dflags scrut case_bndr alts -- Identity case
1708 | all identity_alt alts
1709 = do { tick (CaseIdentity case_bndr)
1710 ; return (re_cast scrut) }
1712 identity_alt (con, args, rhs) = check_eq con args (de_cast rhs)
1714 check_eq DEFAULT _ (Var v) = v == case_bndr
1715 check_eq (LitAlt lit') _ (Lit lit) = lit == lit'
1716 check_eq (DataAlt con) args rhs = rhs `cheapEqExpr` mkConApp con (arg_tys ++ varsToCoreExprs args)
1717 || rhs `cheapEqExpr` Var case_bndr
1718 check_eq _ _ _ = False
1720 arg_tys = map Type (tyConAppArgs (idType case_bndr))
1723 -- case e of x { _ -> x `cast` c }
1724 -- And we definitely want to eliminate this case, to give
1726 -- So we throw away the cast from the RHS, and reconstruct
1727 -- it at the other end. All the RHS casts must be the same
1728 -- if (all identity_alt alts) holds.
1730 -- Don't worry about nested casts, because the simplifier combines them
1731 de_cast (Cast e _) = e
1734 re_cast scrut = case head alts of
1735 (_,_,Cast _ co) -> Cast scrut co
1738 --------------------------------------------------
1739 -- 3. Merge Identical Alternatives
1740 --------------------------------------------------
1741 mkCase1 dflags scrut case_bndr ((_con1,bndrs1,rhs1) : con_alts)
1742 | all isDeadBinder bndrs1 -- Remember the default
1743 , length filtered_alts < length con_alts -- alternative comes first
1744 -- Also Note [Dead binders]
1745 = do { tick (AltMerge case_bndr)
1746 ; mkCase2 dflags scrut case_bndr alts' }
1748 alts' = (DEFAULT, [], rhs1) : filtered_alts
1749 filtered_alts = filter keep con_alts
1750 keep (_con,bndrs,rhs) = not (all isDeadBinder bndrs && rhs `cheapEqExpr` rhs1)
1752 mkCase1 dflags scrut bndr alts = mkCase2 dflags scrut bndr alts
1754 --------------------------------------------------
1756 --------------------------------------------------
1757 mkCase2 _dflags scrut bndr alts
1758 = return (Case scrut bndr (coreAltsType alts) alts)
1762 ~~~~~~~~~~~~~~~~~~~~
1763 Note that dead-ness is maintained by the simplifier, so that it is
1764 accurate after simplification as well as before.
1767 Note [Cascading case merge]
1768 ~~~~~~~~~~~~~~~~~~~~~~~~~~~
1769 Case merging should cascade in one sweep, because it
1773 DEFAULT -> case a of b
1774 DEFAULT -> case b of c {
1781 DEFAULT -> case a of b
1782 DEFAULT -> let c = b in e
1783 A -> let c = b in ea
1788 DEFAULT -> let b = a in let c = b in e
1789 A -> let b = a in let c = b in ea
1790 B -> let b = a in eb
1794 However here's a tricky case that we still don't catch, and I don't
1795 see how to catch it in one pass:
1797 case x of c1 { I# a1 ->
1800 DEFAULT -> case x of c3 { I# a2 ->
1803 After occurrence analysis (and its binder-swap) we get this
1805 case x of c1 { I# a1 ->
1806 let x = c1 in -- Binder-swap addition
1809 DEFAULT -> case x of c3 { I# a2 ->
1812 When we simplify the inner case x, we'll see that
1813 x=c1=I# a1. So we'll bind a2 to a1, and get
1815 case x of c1 { I# a1 ->
1818 DEFAULT -> case a1 of ...
1820 This is corect, but we can't do a case merge in this sweep
1821 because c2 /= a1. Reason: the binding c1=I# a1 went inwards
1822 without getting changed to c1=I# c2.
1824 I don't think this is worth fixing, even if I knew how. It'll
1825 all come out in the next pass anyway.