2 % (c) The AQUA Project, Glasgow University, 1993-1998
4 \section[SimplUtils]{The simplifier utilities}
9 mkLam, mkCase, prepareAlts,
12 preInlineUnconditionally, postInlineUnconditionally,
13 activeUnfolding, activeUnfInRule, activeRule,
14 simplEnvForGHCi, simplEnvForRules, updModeForInlineRules,
16 -- The continuation type
17 SimplCont(..), DupFlag(..), ArgInfo(..),
18 contIsDupable, contResultType, contIsTrivial, contArgs, dropArgs,
19 pushArgs, countValArgs, countArgs, addArgTo,
20 mkBoringStop, mkRhsStop, mkLazyArgStop, contIsRhsOrArg,
21 interestingCallContext,
23 interestingArg, mkArgInfo,
28 #include "HsVersions.h"
31 import CoreMonad ( SimplifierMode(..), Tick(..) )
35 import qualified CoreSubst
39 import CoreArity ( etaExpand, exprEtaExpandArity )
43 import Var ( isCoVar )
46 import Type hiding( substTy )
47 import Coercion ( coercionKind )
49 import Unify ( dataConCannotMatch )
61 %************************************************************************
65 %************************************************************************
67 A SimplCont allows the simplifier to traverse the expression in a
68 zipper-like fashion. The SimplCont represents the rest of the expression,
69 "above" the point of interest.
71 You can also think of a SimplCont as an "evaluation context", using
72 that term in the way it is used for operational semantics. This is the
73 way I usually think of it, For example you'll often see a syntax for
74 evaluation context looking like
75 C ::= [] | C e | case C of alts | C `cast` co
76 That's the kind of thing we are doing here, and I use that syntax in
81 * A SimplCont describes a *strict* context (just like
82 evaluation contexts do). E.g. Just [] is not a SimplCont
84 * A SimplCont describes a context that *does not* bind
85 any variables. E.g. \x. [] is not a SimplCont
89 = Stop -- An empty context, or hole, []
90 CallCtxt -- True <=> There is something interesting about
91 -- the context, and hence the inliner
92 -- should be a bit keener (see interestingCallContext)
94 -- This is an argument of a function that has RULES
95 -- Inlining the call might allow the rule to fire
97 | CoerceIt -- C `cast` co
98 OutCoercion -- The coercion simplified
103 InExpr StaticEnv -- The argument and its static env
106 | Select -- case C of alts
108 InId [InAlt] StaticEnv -- The case binder, alts, and subst-env
111 -- The two strict forms have no DupFlag, because we never duplicate them
112 | StrictBind -- (\x* \xs. e) C
113 InId [InBndr] -- let x* = [] in e
114 InExpr StaticEnv -- is a special case
117 | StrictArg -- f e1 ..en C
118 ArgInfo -- Specifies f, e1..en, Whether f has rules, etc
119 -- plus strictness flags for *further* args
120 CallCtxt -- Whether *this* argument position is interesting
125 ai_fun :: Id, -- The function
126 ai_args :: [OutExpr], -- ...applied to these args (which are in *reverse* order)
127 ai_rules :: [CoreRule], -- Rules for this function
129 ai_encl :: Bool, -- Flag saying whether this function
130 -- or an enclosing one has rules (recursively)
131 -- True => be keener to inline in all args
133 ai_strs :: [Bool], -- Strictness of remaining arguments
134 -- Usually infinite, but if it is finite it guarantees
135 -- that the function diverges after being given
136 -- that number of args
137 ai_discs :: [Int] -- Discounts for remaining arguments; non-zero => be keener to inline
141 addArgTo :: ArgInfo -> OutExpr -> ArgInfo
142 addArgTo ai arg = ai { ai_args = arg : ai_args ai }
144 instance Outputable SimplCont where
145 ppr (Stop interesting) = ptext (sLit "Stop") <> brackets (ppr interesting)
146 ppr (ApplyTo dup arg _ cont) = ((ptext (sLit "ApplyTo") <+> ppr dup <+> pprParendExpr arg)
147 {- $$ nest 2 (pprSimplEnv se) -}) $$ ppr cont
148 ppr (StrictBind b _ _ _ cont) = (ptext (sLit "StrictBind") <+> ppr b) $$ ppr cont
149 ppr (StrictArg ai _ cont) = (ptext (sLit "StrictArg") <+> ppr (ai_fun ai)) $$ ppr cont
150 ppr (Select dup bndr alts se cont) = (ptext (sLit "Select") <+> ppr dup <+> ppr bndr) $$
151 (nest 2 $ vcat [ppr (seTvSubst se), ppr alts]) $$ ppr cont
152 ppr (CoerceIt co cont) = (ptext (sLit "CoerceIt") <+> ppr co) $$ ppr cont
154 data DupFlag = OkToDup | NoDup
156 instance Outputable DupFlag where
157 ppr OkToDup = ptext (sLit "ok")
158 ppr NoDup = ptext (sLit "nodup")
163 mkBoringStop :: SimplCont
164 mkBoringStop = Stop BoringCtxt
166 mkRhsStop :: SimplCont -- See Note [RHS of lets] in CoreUnfold
167 mkRhsStop = Stop (ArgCtxt False)
169 mkLazyArgStop :: CallCtxt -> SimplCont
170 mkLazyArgStop cci = Stop cci
173 contIsRhsOrArg :: SimplCont -> Bool
174 contIsRhsOrArg (Stop {}) = True
175 contIsRhsOrArg (StrictBind {}) = True
176 contIsRhsOrArg (StrictArg {}) = True
177 contIsRhsOrArg _ = False
180 contIsDupable :: SimplCont -> Bool
181 contIsDupable (Stop {}) = True
182 contIsDupable (ApplyTo OkToDup _ _ _) = True
183 contIsDupable (Select OkToDup _ _ _ _) = True
184 contIsDupable (CoerceIt _ cont) = contIsDupable cont
185 contIsDupable _ = False
188 contIsTrivial :: SimplCont -> Bool
189 contIsTrivial (Stop {}) = True
190 contIsTrivial (ApplyTo _ (Type _) _ cont) = contIsTrivial cont
191 contIsTrivial (CoerceIt _ cont) = contIsTrivial cont
192 contIsTrivial _ = False
195 contResultType :: SimplEnv -> OutType -> SimplCont -> OutType
196 contResultType env ty cont
199 subst_ty se ty = substTy (se `setInScope` env) ty
202 go (CoerceIt co cont) _ = go cont (snd (coercionKind co))
203 go (StrictBind _ bs body se cont) _ = go cont (subst_ty se (exprType (mkLams bs body)))
204 go (StrictArg ai _ cont) _ = go cont (funResultTy (argInfoResultTy ai))
205 go (Select _ _ alts se cont) _ = go cont (subst_ty se (coreAltsType alts))
206 go (ApplyTo _ arg se cont) ty = go cont (apply_to_arg ty arg se)
208 apply_to_arg ty (Type ty_arg) se = applyTy ty (subst_ty se ty_arg)
209 apply_to_arg ty _ _ = funResultTy ty
211 argInfoResultTy :: ArgInfo -> OutType
212 argInfoResultTy (ArgInfo { ai_fun = fun, ai_args = args })
213 = foldr (\arg fn_ty -> applyTypeToArg fn_ty arg) (idType fun) args
216 countValArgs :: SimplCont -> Int
217 countValArgs (ApplyTo _ (Type _) _ cont) = countValArgs cont
218 countValArgs (ApplyTo _ _ _ cont) = 1 + countValArgs cont
221 countArgs :: SimplCont -> Int
222 countArgs (ApplyTo _ _ _ cont) = 1 + countArgs cont
225 contArgs :: SimplCont -> (Bool, [ArgSummary], SimplCont)
226 -- Uses substitution to turn each arg into an OutExpr
227 contArgs cont@(ApplyTo {})
228 = case go [] cont of { (args, cont') -> (False, args, cont') }
230 go args (ApplyTo _ arg se cont)
231 | isTypeArg arg = go args cont
232 | otherwise = go (is_interesting arg se : args) cont
233 go args cont = (reverse args, cont)
235 is_interesting arg se = interestingArg (substExpr (text "contArgs") se arg)
236 -- Do *not* use short-cutting substitution here
237 -- because we want to get as much IdInfo as possible
239 contArgs cont = (True, [], cont)
241 pushArgs :: SimplEnv -> [CoreExpr] -> SimplCont -> SimplCont
242 pushArgs _env [] cont = cont
243 pushArgs env (arg:args) cont = ApplyTo NoDup arg env (pushArgs env args cont)
245 dropArgs :: Int -> SimplCont -> SimplCont
246 dropArgs 0 cont = cont
247 dropArgs n (ApplyTo _ _ _ cont) = dropArgs (n-1) cont
248 dropArgs n other = pprPanic "dropArgs" (ppr n <+> ppr other)
252 Note [Interesting call context]
253 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
254 We want to avoid inlining an expression where there can't possibly be
255 any gain, such as in an argument position. Hence, if the continuation
256 is interesting (eg. a case scrutinee, application etc.) then we
257 inline, otherwise we don't.
259 Previously some_benefit used to return True only if the variable was
260 applied to some value arguments. This didn't work:
262 let x = _coerce_ (T Int) Int (I# 3) in
263 case _coerce_ Int (T Int) x of
266 we want to inline x, but can't see that it's a constructor in a case
267 scrutinee position, and some_benefit is False.
271 dMonadST = _/\_ t -> :Monad (g1 _@_ t, g2 _@_ t, g3 _@_ t)
273 .... case dMonadST _@_ x0 of (a,b,c) -> ....
275 we'd really like to inline dMonadST here, but we *don't* want to
276 inline if the case expression is just
278 case x of y { DEFAULT -> ... }
280 since we can just eliminate this case instead (x is in WHNF). Similar
281 applies when x is bound to a lambda expression. Hence
282 contIsInteresting looks for case expressions with just a single
287 interestingCallContext :: SimplCont -> CallCtxt
288 -- See Note [Interesting call context]
289 interestingCallContext cont
292 interesting (Select _ bndr _ _ _)
293 | isDeadBinder bndr = CaseCtxt
294 | otherwise = ArgCtxt False -- If the binder is used, this
295 -- is like a strict let
296 -- See Note [RHS of lets] in CoreUnfold
298 interesting (ApplyTo _ arg _ cont)
299 | isTypeArg arg = interesting cont
300 | otherwise = ValAppCtxt -- Can happen if we have (f Int |> co) y
301 -- If f has an INLINE prag we need to give it some
302 -- motivation to inline. See Note [Cast then apply]
305 interesting (StrictArg _ cci _) = cci
306 interesting (StrictBind {}) = BoringCtxt
307 interesting (Stop cci) = cci
308 interesting (CoerceIt _ cont) = interesting cont
309 -- If this call is the arg of a strict function, the context
310 -- is a bit interesting. If we inline here, we may get useful
311 -- evaluation information to avoid repeated evals: e.g.
313 -- Here the contIsInteresting makes the '*' keener to inline,
314 -- which in turn exposes a constructor which makes the '+' inline.
315 -- Assuming that +,* aren't small enough to inline regardless.
317 -- It's also very important to inline in a strict context for things
320 -- Here, the context of (f x) is strict, and if f's unfolding is
321 -- a build it's *great* to inline it here. So we must ensure that
322 -- the context for (f x) is not totally uninteresting.
327 -> [CoreRule] -- Rules for function
328 -> Int -- Number of value args
329 -> SimplCont -- Context of the call
332 mkArgInfo fun rules n_val_args call_cont
333 | n_val_args < idArity fun -- Note [Unsaturated functions]
334 = ArgInfo { ai_fun = fun, ai_args = [], ai_rules = rules
336 , ai_strs = vanilla_stricts
337 , ai_discs = vanilla_discounts }
339 = ArgInfo { ai_fun = fun, ai_args = [], ai_rules = rules
340 , ai_encl = interestingArgContext rules call_cont
341 , ai_strs = add_type_str (idType fun) arg_stricts
342 , ai_discs = arg_discounts }
344 vanilla_discounts, arg_discounts :: [Int]
345 vanilla_discounts = repeat 0
346 arg_discounts = case idUnfolding fun of
347 CoreUnfolding {uf_guidance = UnfIfGoodArgs {ug_args = discounts}}
348 -> discounts ++ vanilla_discounts
349 _ -> vanilla_discounts
351 vanilla_stricts, arg_stricts :: [Bool]
352 vanilla_stricts = repeat False
355 = case splitStrictSig (idStrictness fun) of
356 (demands, result_info)
357 | not (demands `lengthExceeds` n_val_args)
358 -> -- Enough args, use the strictness given.
359 -- For bottoming functions we used to pretend that the arg
360 -- is lazy, so that we don't treat the arg as an
361 -- interesting context. This avoids substituting
362 -- top-level bindings for (say) strings into
363 -- calls to error. But now we are more careful about
364 -- inlining lone variables, so its ok (see SimplUtils.analyseCont)
365 if isBotRes result_info then
366 map isStrictDmd demands -- Finite => result is bottom
368 map isStrictDmd demands ++ vanilla_stricts
370 -> WARN( True, text "More demands than arity" <+> ppr fun <+> ppr (idArity fun)
371 <+> ppr n_val_args <+> ppr demands )
372 vanilla_stricts -- Not enough args, or no strictness
374 add_type_str :: Type -> [Bool] -> [Bool]
375 -- If the function arg types are strict, record that in the 'strictness bits'
376 -- No need to instantiate because unboxed types (which dominate the strict
377 -- types) can't instantiate type variables.
378 -- add_type_str is done repeatedly (for each call); might be better
379 -- once-for-all in the function
380 -- But beware primops/datacons with no strictness
381 add_type_str _ [] = []
382 add_type_str fun_ty strs -- Look through foralls
383 | Just (_, fun_ty') <- splitForAllTy_maybe fun_ty -- Includes coercions
384 = add_type_str fun_ty' strs
385 add_type_str fun_ty (str:strs) -- Add strict-type info
386 | Just (arg_ty, fun_ty') <- splitFunTy_maybe fun_ty
387 = (str || isStrictType arg_ty) : add_type_str fun_ty' strs
391 {- Note [Unsaturated functions]
392 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
393 Consider (test eyeball/inline4)
396 where f has arity 2. Then we do not want to inline 'x', because
397 it'll just be floated out again. Even if f has lots of discounts
398 on its first argument -- it must be saturated for these to kick in
401 interestingArgContext :: [CoreRule] -> SimplCont -> Bool
402 -- If the argument has form (f x y), where x,y are boring,
403 -- and f is marked INLINE, then we don't want to inline f.
404 -- But if the context of the argument is
406 -- where g has rules, then we *do* want to inline f, in case it
407 -- exposes a rule that might fire. Similarly, if the context is
409 -- where h has rules, then we do want to inline f; hence the
410 -- call_cont argument to interestingArgContext
412 -- The ai-rules flag makes this happen; if it's
413 -- set, the inliner gets just enough keener to inline f
414 -- regardless of how boring f's arguments are, if it's marked INLINE
416 -- The alternative would be to *always* inline an INLINE function,
417 -- regardless of how boring its context is; but that seems overkill
418 -- For example, it'd mean that wrapper functions were always inlined
419 interestingArgContext rules call_cont
420 = notNull rules || enclosing_fn_has_rules
422 enclosing_fn_has_rules = go call_cont
424 go (Select {}) = False
425 go (ApplyTo {}) = False
426 go (StrictArg _ cci _) = interesting cci
427 go (StrictBind {}) = False -- ??
428 go (CoerceIt _ c) = go c
429 go (Stop cci) = interesting cci
431 interesting (ArgCtxt rules) = rules
432 interesting _ = False
437 %************************************************************************
439 \subsection{Decisions about inlining}
441 %************************************************************************
443 Inlining is controlled partly by the SimplifierMode switch. This has two
446 SimplGently (a) Simplifying before specialiser/full laziness
447 (b) Simplifiying inside InlineRules
448 (c) Simplifying the LHS of a rule
449 (d) Simplifying a GHCi expression or Template
452 SimplPhase n _ Used at all other times
456 Gentle mode has a separate boolean flag to control
457 a) inlining (sm_inline flag)
458 b) rules (sm_rules flag)
459 A key invariant about Gentle mode is that it is treated as the EARLIEST
460 phase. Something is inlined if the sm_inline flag is on AND the thing
461 is inlinable in the earliest phase. This is important. Example
463 {-# INLINE [~1] g #-}
469 If we were to inline g into f's inlining, then an importing module would
471 f e --> g (g e) ---> RULE fires
472 because the InlineRule for f has had g inlined into it.
474 On the other hand, it is bad not to do ANY inlining into an
475 InlineRule, because then recursive knots in instance declarations
476 don't get unravelled.
478 However, *sometimes* SimplGently must do no call-site inlining at all.
479 Before full laziness we must be careful not to inline wrappers,
480 because doing so inhibits floating
481 e.g. ...(case f x of ...)...
482 ==> ...(case (case x of I# x# -> fw x#) of ...)...
483 ==> ...(case x of I# x# -> case fw x# of ...)...
484 and now the redex (f x) isn't floatable any more.
486 The no-inlining thing is also important for Template Haskell. You might be
487 compiling in one-shot mode with -O2; but when TH compiles a splice before
488 running it, we don't want to use -O2. Indeed, we don't want to inline
489 anything, because the byte-code interpreter might get confused about
490 unboxed tuples and suchlike.
492 Note [RULEs enabled in SimplGently]
493 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
494 RULES are enabled when doing "gentle" simplification. Two reasons:
496 * We really want the class-op cancellation to happen:
497 op (df d1 d2) --> $cop3 d1 d2
498 because this breaks the mutual recursion between 'op' and 'df'
502 to work in Template Haskell when simplifying
503 splices, so we get simpler code for literal strings
505 Note [Simplifying inside InlineRules]
506 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
507 We must take care with simplification inside InlineRules (which come from
510 First, consider the following example
515 in ...g...g...g...g...g...
516 Now, if that's the ONLY occurrence of f, it might be inlined inside g,
517 and thence copied multiple times when g is inlined. HENCE we treat
518 any occurrence in an InlineRule as a multiple occurrence, not a single
519 one; see OccurAnal.addRuleUsage.
521 Second, we do want *do* to some modest rules/inlining stuff in InlineRules,
522 partly to eliminate senseless crap, and partly to break the recursive knots
523 generated by instance declarations. To keep things simple, we always set
524 the phase to 'gentle' when processing InlineRules. OK, so suppose we have
525 {-# INLINE <act> f #-}
527 meaning "inline f in phases p where activation <act>(p) holds".
528 Then what inlinings/rules can we apply to the copy of <rhs> captured in
529 f's InlineRule? Our model is that literally <rhs> is substituted for
530 f when it is inlined. So our conservative plan (implemented by
531 updModeForInlineRules) is this:
533 -------------------------------------------------------------
534 When simplifying the RHS of an InlineRule,
535 If the InlineRule becomes active in phase p, then
536 if the current phase is *earlier than* p,
537 make no inlinings or rules active when simplifying the RHS
539 set the phase to p when simplifying the RHS
540 -------------------------------------------------------------
544 a) Rules/inlinings that *cease* being active before p will
545 not apply to the InlineRule rhs, consistent with it being
546 inlined in its *original* form in phase p.
548 b) Rules/inlinings that only become active *after* p will
549 not apply to the InlineRule rhs, again to be consistent with
550 inlining the *original* rhs in phase p.
556 {-# NOINLINE [1] g #-}
559 {-# RULE h g = ... #-}
560 Here we must not inline g into f's RHS, even when we get to phase 0,
561 because when f is later inlined into some other module we want the
569 and suppose that there are auto-generated specialisations and a strictness
570 wrapper for g. The specialisations get activation AlwaysActive, and the
571 strictness wrapper get activation (ActiveAfter 0). So the strictness
572 wrepper fails the test and won't be inlined into f's InlineRule. That
573 means f can inline, expose the specialised call to g, so the specialisation
576 A note about wrappers
577 ~~~~~~~~~~~~~~~~~~~~~
578 It's also important not to inline a worker back into a wrapper.
580 wraper = inline_me (\x -> ...worker... )
581 Normally, the inline_me prevents the worker getting inlined into
582 the wrapper (initially, the worker's only call site!). But,
583 if the wrapper is sure to be called, the strictness analyser will
584 mark it 'demanded', so when the RHS is simplified, it'll get an ArgOf
588 simplEnvForGHCi :: SimplEnv
589 simplEnvForGHCi = mkSimplEnv allOffSwitchChecker $
590 SimplGently { sm_rules = False, sm_inline = False }
591 -- Do not do any inlining, in case we expose some unboxed
592 -- tuple stuff that confuses the bytecode interpreter
594 simplEnvForRules :: SimplEnv
595 simplEnvForRules = mkSimplEnv allOffSwitchChecker $
596 SimplGently { sm_rules = True, sm_inline = False }
598 updModeForInlineRules :: Activation -> SimplifierMode -> SimplifierMode
599 -- See Note [Simplifying inside InlineRules]
600 -- Treat Gentle as phase "infinity"
601 -- If current_phase `earlier than` inline_rule_start_phase
604 -- if current_phase `same phase` inline_rule_start_phase
605 -- then current_phase (keep gentle flags)
606 -- else inline_rule_start_phase
607 updModeForInlineRules inline_rule_act current_mode
608 = case inline_rule_act of
610 AlwaysActive -> mk_gentle current_mode
611 ActiveBefore {} -> mk_gentle current_mode
612 ActiveAfter n -> mk_phase n current_mode
614 no_op = SimplGently { sm_rules = False, sm_inline = False }
616 mk_gentle (SimplGently {}) = current_mode
617 mk_gentle _ = SimplGently { sm_rules = True, sm_inline = True }
619 mk_phase n (SimplPhase _ ss) = SimplPhase n ss
620 mk_phase n (SimplGently {}) = SimplPhase n ["gentle-rules"]
624 preInlineUnconditionally
625 ~~~~~~~~~~~~~~~~~~~~~~~~
626 @preInlineUnconditionally@ examines a bndr to see if it is used just
627 once in a completely safe way, so that it is safe to discard the
628 binding inline its RHS at the (unique) usage site, REGARDLESS of how
629 big the RHS might be. If this is the case we don't simplify the RHS
630 first, but just inline it un-simplified.
632 This is much better than first simplifying a perhaps-huge RHS and then
633 inlining and re-simplifying it. Indeed, it can be at least quadratically
642 We may end up simplifying e1 N times, e2 N-1 times, e3 N-3 times etc.
643 This can happen with cascades of functions too:
650 THE MAIN INVARIANT is this:
652 ---- preInlineUnconditionally invariant -----
653 IF preInlineUnconditionally chooses to inline x = <rhs>
654 THEN doing the inlining should not change the occurrence
655 info for the free vars of <rhs>
656 ----------------------------------------------
658 For example, it's tempting to look at trivial binding like
660 and inline it unconditionally. But suppose x is used many times,
661 but this is the unique occurrence of y. Then inlining x would change
662 y's occurrence info, which breaks the invariant. It matters: y
663 might have a BIG rhs, which will now be dup'd at every occurrenc of x.
666 Even RHSs labelled InlineMe aren't caught here, because there might be
667 no benefit from inlining at the call site.
669 [Sept 01] Don't unconditionally inline a top-level thing, because that
670 can simply make a static thing into something built dynamically. E.g.
674 [Remember that we treat \s as a one-shot lambda.] No point in
675 inlining x unless there is something interesting about the call site.
677 But watch out: if you aren't careful, some useful foldr/build fusion
678 can be lost (most notably in spectral/hartel/parstof) because the
679 foldr didn't see the build. Doing the dynamic allocation isn't a big
680 deal, in fact, but losing the fusion can be. But the right thing here
681 seems to be to do a callSiteInline based on the fact that there is
682 something interesting about the call site (it's strict). Hmm. That
685 Conclusion: inline top level things gaily until Phase 0 (the last
686 phase), at which point don't.
688 Note [pre/postInlineUnconditionally in gentle mode]
689 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
690 Even in gentle mode we want to do preInlineUnconditionally. The
691 reason is that too little clean-up happens if you don't inline
692 use-once things. Also a bit of inlining is *good* for full laziness;
693 it can expose constant sub-expressions. Example in
694 spectral/mandel/Mandel.hs, where the mandelset function gets a useful
695 let-float if you inline windowToViewport
697 However, as usual for Gentle mode, do not inline things that are
698 inactive in the intial stages. See Note [Gentle mode].
700 Note [Top-level botomming Ids]
701 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
702 Don't inline top-level Ids that are bottoming, even if they are used just
703 once, because FloatOut has gone to some trouble to extract them out.
704 Inlining them won't make the program run faster!
707 preInlineUnconditionally :: SimplEnv -> TopLevelFlag -> InId -> InExpr -> Bool
708 preInlineUnconditionally env top_lvl bndr rhs
710 | isTopLevel top_lvl && isBottomingId bndr = False -- Note [Top-level bottoming Ids]
711 | opt_SimplNoPreInlining = False
712 | otherwise = case idOccInfo bndr of
713 IAmDead -> True -- Happens in ((\x.1) v)
714 OneOcc in_lam True int_cxt -> try_once in_lam int_cxt
718 active = case phase of
719 SimplGently {} -> isEarlyActive act
720 -- See Note [pre/postInlineUnconditionally in gentle mode]
721 SimplPhase n _ -> isActive n act
722 act = idInlineActivation bndr
723 try_once in_lam int_cxt -- There's one textual occurrence
724 | not in_lam = isNotTopLevel top_lvl || early_phase
725 | otherwise = int_cxt && canInlineInLam rhs
727 -- Be very careful before inlining inside a lambda, because (a) we must not
728 -- invalidate occurrence information, and (b) we want to avoid pushing a
729 -- single allocation (here) into multiple allocations (inside lambda).
730 -- Inlining a *function* with a single *saturated* call would be ok, mind you.
731 -- || (if is_cheap && not (canInlineInLam rhs) then pprTrace "preinline" (ppr bndr <+> ppr rhs) ok else ok)
733 -- is_cheap = exprIsCheap rhs
734 -- ok = is_cheap && int_cxt
736 -- int_cxt The context isn't totally boring
737 -- E.g. let f = \ab.BIG in \y. map f xs
738 -- Don't want to substitute for f, because then we allocate
739 -- its closure every time the \y is called
740 -- But: let f = \ab.BIG in \y. map (f y) xs
741 -- Now we do want to substitute for f, even though it's not
742 -- saturated, because we're going to allocate a closure for
743 -- (f y) every time round the loop anyhow.
745 -- canInlineInLam => free vars of rhs are (Once in_lam) or Many,
746 -- so substituting rhs inside a lambda doesn't change the occ info.
747 -- Sadly, not quite the same as exprIsHNF.
748 canInlineInLam (Lit _) = True
749 canInlineInLam (Lam b e) = isRuntimeVar b || canInlineInLam e
750 canInlineInLam (Note _ e) = canInlineInLam e
751 canInlineInLam _ = False
753 early_phase = case phase of
754 SimplPhase 0 _ -> False
756 -- If we don't have this early_phase test, consider
757 -- x = length [1,2,3]
758 -- The full laziness pass carefully floats all the cons cells to
759 -- top level, and preInlineUnconditionally floats them all back in.
760 -- Result is (a) static allocation replaced by dynamic allocation
761 -- (b) many simplifier iterations because this tickles
762 -- a related problem; only one inlining per pass
764 -- On the other hand, I have seen cases where top-level fusion is
765 -- lost if we don't inline top level thing (e.g. string constants)
766 -- Hence the test for phase zero (which is the phase for all the final
767 -- simplifications). Until phase zero we take no special notice of
768 -- top level things, but then we become more leery about inlining
773 postInlineUnconditionally
774 ~~~~~~~~~~~~~~~~~~~~~~~~~
775 @postInlineUnconditionally@ decides whether to unconditionally inline
776 a thing based on the form of its RHS; in particular if it has a
777 trivial RHS. If so, we can inline and discard the binding altogether.
779 NB: a loop breaker has must_keep_binding = True and non-loop-breakers
780 only have *forward* references Hence, it's safe to discard the binding
782 NOTE: This isn't our last opportunity to inline. We're at the binding
783 site right now, and we'll get another opportunity when we get to the
786 Note that we do this unconditional inlining only for trival RHSs.
787 Don't inline even WHNFs inside lambdas; doing so may simply increase
788 allocation when the function is called. This isn't the last chance; see
791 NB: Even inline pragmas (e.g. IMustBeINLINEd) are ignored here Why?
792 Because we don't even want to inline them into the RHS of constructor
793 arguments. See NOTE above
795 NB: At one time even NOINLINE was ignored here: if the rhs is trivial
796 it's best to inline it anyway. We often get a=E; b=a from desugaring,
797 with both a and b marked NOINLINE. But that seems incompatible with
798 our new view that inlining is like a RULE, so I'm sticking to the 'active'
802 postInlineUnconditionally
803 :: SimplEnv -> TopLevelFlag
804 -> OutId -- The binder (an InId would be fine too)
805 -> OccInfo -- From the InId
809 postInlineUnconditionally env top_lvl bndr occ_info rhs unfolding
811 | isLoopBreaker occ_info = False -- If it's a loop-breaker of any kind, don't inline
812 -- because it might be referred to "earlier"
813 | isExportedId bndr = False
814 | isStableUnfolding unfolding = False -- Note [InlineRule and postInlineUnconditionally]
815 | exprIsTrivial rhs = True
816 | isTopLevel top_lvl = False -- Note [Top level and postInlineUnconditionally]
819 -- The point of examining occ_info here is that for *non-values*
820 -- that occur outside a lambda, the call-site inliner won't have
821 -- a chance (becuase it doesn't know that the thing
822 -- only occurs once). The pre-inliner won't have gotten
823 -- it either, if the thing occurs in more than one branch
824 -- So the main target is things like
827 -- True -> case x of ...
828 -- False -> case x of ...
829 -- This is very important in practice; e.g. wheel-seive1 doubles
830 -- in allocation if you miss this out
831 OneOcc in_lam _one_br int_cxt -- OneOcc => no code-duplication issue
832 -> smallEnoughToInline unfolding -- Small enough to dup
833 -- ToDo: consider discount on smallEnoughToInline if int_cxt is true
835 -- NB: Do NOT inline arbitrarily big things, even if one_br is True
836 -- Reason: doing so risks exponential behaviour. We simplify a big
837 -- expression, inline it, and simplify it again. But if the
838 -- very same thing happens in the big expression, we get
840 -- PRINCIPLE: when we've already simplified an expression once,
841 -- make sure that we only inline it if it's reasonably small.
844 -- Outside a lambda, we want to be reasonably aggressive
845 -- about inlining into multiple branches of case
846 -- e.g. let x = <non-value>
847 -- in case y of { C1 -> ..x..; C2 -> ..x..; C3 -> ... }
848 -- Inlining can be a big win if C3 is the hot-spot, even if
849 -- the uses in C1, C2 are not 'interesting'
850 -- An example that gets worse if you add int_cxt here is 'clausify'
852 (isCheapUnfolding unfolding && int_cxt))
853 -- isCheap => acceptable work duplication; in_lam may be true
854 -- int_cxt to prevent us inlining inside a lambda without some
855 -- good reason. See the notes on int_cxt in preInlineUnconditionally
857 IAmDead -> True -- This happens; for example, the case_bndr during case of
858 -- known constructor: case (a,b) of x { (p,q) -> ... }
859 -- Here x isn't mentioned in the RHS, so we don't want to
860 -- create the (dead) let-binding let x = (a,b) in ...
864 -- Here's an example that we don't handle well:
865 -- let f = if b then Left (\x.BIG) else Right (\y.BIG)
866 -- in \y. ....case f of {...} ....
867 -- Here f is used just once, and duplicating the case work is fine (exprIsCheap).
869 -- - We can't preInlineUnconditionally because that woud invalidate
870 -- the occ info for b.
871 -- - We can't postInlineUnconditionally because the RHS is big, and
872 -- that risks exponential behaviour
873 -- - We can't call-site inline, because the rhs is big
877 active = case getMode env of
878 SimplGently {} -> isEarlyActive act
879 -- See Note [pre/postInlineUnconditionally in gentle mode]
880 SimplPhase n _ -> isActive n act
881 act = idInlineActivation bndr
883 activeUnfolding :: SimplEnv -> IdUnfoldingFun
885 = case getMode env of
886 SimplGently { sm_inline = False } -> active_unfolding_minimal
887 SimplGently { sm_inline = True } -> active_unfolding_gentle
888 SimplPhase n _ -> active_unfolding n
890 activeUnfInRule :: SimplEnv -> IdUnfoldingFun
891 -- When matching in RULE, we want to "look through" an unfolding
892 -- if *rules* are on, even if *inlinings* are not. A notable example
893 -- is DFuns, which really we want to match in rules like (op dfun)
896 = case getMode env of
897 SimplGently { sm_rules = False } -> active_unfolding_minimal
898 SimplGently { sm_rules = True } -> active_unfolding_gentle
899 SimplPhase n _ -> active_unfolding n
901 active_unfolding_minimal :: IdUnfoldingFun
902 -- Compuslory unfoldings only
903 -- Ignore SimplGently, because we want to inline regardless;
904 -- the Id has no top-level binding at all
906 -- NB: we used to have a second exception, for data con wrappers.
907 -- On the grounds that we use gentle mode for rule LHSs, and
908 -- they match better when data con wrappers are inlined.
909 -- But that only really applies to the trivial wrappers (like (:)),
910 -- and they are now constructed as Compulsory unfoldings (in MkId)
911 -- so they'll happen anyway.
912 active_unfolding_minimal id
913 | isCompulsoryUnfolding unf = unf
914 | otherwise = NoUnfolding
916 unf = realIdUnfolding id -- Never a loop breaker
918 active_unfolding_gentle :: IdUnfoldingFun
919 -- Anything that is early-active
920 -- See Note [Gentle mode]
921 active_unfolding_gentle id
922 | isEarlyActive (idInlineActivation id) = idUnfolding id
923 | otherwise = NoUnfolding
924 -- idUnfolding checks for loop-breakers
925 -- Things with an INLINE pragma may have
926 -- an unfolding *and* be a loop breaker
927 -- (maybe the knot is not yet untied)
929 active_unfolding :: CompilerPhase -> IdUnfoldingFun
930 active_unfolding n id
931 | isActive n (idInlineActivation id) = idUnfolding id
932 | otherwise = NoUnfolding
934 activeRule :: DynFlags -> SimplEnv -> Maybe (Activation -> Bool)
935 -- Nothing => No rules at all
936 activeRule dflags env
937 | not (dopt Opt_EnableRewriteRules dflags)
938 = Nothing -- Rewriting is off
940 = case getMode env of
941 SimplGently { sm_rules = rules_on }
942 | rules_on -> Just isEarlyActive -- Note [RULEs enabled in SimplGently]
943 | otherwise -> Nothing
944 SimplPhase n _ -> Just (isActive n)
947 Note [Top level and postInlineUnconditionally]
948 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
949 We don't do postInlineUnconditionally for top-level things (exept ones that
951 * There is no point, because the main goal is to get rid of local
952 bindings used in multiple case branches.
953 * Doing so will inline top-level error expressions that have been
954 carefully floated out by FloatOut. More generally, it might
955 replace static allocation with dynamic.
957 Note [InlineRule and postInlineUnconditionally]
958 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
959 Do not do postInlineUnconditionally if the Id has an InlineRule, otherwise
960 we lose the unfolding. Example
962 -- f has InlineRule with rhs (e |> co)
966 Then there's a danger we'll optimise to
971 and now postInlineUnconditionally, losing the InlineRule on f. Now f'
972 won't inline because 'e' is too big.
975 %************************************************************************
979 %************************************************************************
982 mkLam :: SimplEnv -> [OutBndr] -> OutExpr -> SimplM OutExpr
983 -- mkLam tries three things
984 -- a) eta reduction, if that gives a trivial expression
985 -- b) eta expansion [only if there are some value lambdas]
990 = do { dflags <- getDOptsSmpl
991 ; mkLam' dflags bndrs body }
993 mkLam' :: DynFlags -> [OutBndr] -> OutExpr -> SimplM OutExpr
994 mkLam' dflags bndrs (Cast body co)
995 | not (any bad bndrs)
996 -- Note [Casts and lambdas]
997 = do { lam <- mkLam' dflags bndrs body
998 ; return (mkCoerce (mkPiTypes bndrs co) lam) }
1000 co_vars = tyVarsOfType co
1001 bad bndr = isCoVar bndr && bndr `elemVarSet` co_vars
1003 mkLam' dflags bndrs body
1004 | dopt Opt_DoEtaReduction dflags,
1005 Just etad_lam <- tryEtaReduce bndrs body
1006 = do { tick (EtaReduction (head bndrs))
1009 | dopt Opt_DoLambdaEtaExpansion dflags,
1010 not (inGentleMode env), -- In gentle mode don't eta-expansion
1011 -- because it can clutter up the code
1012 -- with casts etc that may not be removed
1013 not (all isTyVar bndrs) -- Don't eta expand type abstractions
1014 = do { let body' = tryEtaExpansion dflags body
1015 ; return (mkLams bndrs body') }
1018 = return (mkLams bndrs body)
1021 Note [Casts and lambdas]
1022 ~~~~~~~~~~~~~~~~~~~~~~~~
1024 (\x. (\y. e) `cast` g1) `cast` g2
1025 There is a danger here that the two lambdas look separated, and the
1026 full laziness pass might float an expression to between the two.
1028 So this equation in mkLam' floats the g1 out, thus:
1029 (\x. e `cast` g1) --> (\x.e) `cast` (tx -> g1)
1032 In general, this floats casts outside lambdas, where (I hope) they
1033 might meet and cancel with some other cast:
1034 \x. e `cast` co ===> (\x. e) `cast` (tx -> co)
1035 /\a. e `cast` co ===> (/\a. e) `cast` (/\a. co)
1036 /\g. e `cast` co ===> (/\g. e) `cast` (/\g. co)
1037 (if not (g `in` co))
1039 Notice that it works regardless of 'e'. Originally it worked only
1040 if 'e' was itself a lambda, but in some cases that resulted in
1041 fruitless iteration in the simplifier. A good example was when
1042 compiling Text.ParserCombinators.ReadPrec, where we had a definition
1043 like (\x. Get `cast` g)
1044 where Get is a constructor with nonzero arity. Then mkLam eta-expanded
1045 the Get, and the next iteration eta-reduced it, and then eta-expanded
1048 Note also the side condition for the case of coercion binders.
1049 It does not make sense to transform
1050 /\g. e `cast` g ==> (/\g.e) `cast` (/\g.g)
1051 because the latter is not well-kinded.
1053 -- c) floating lets out through big lambdas
1054 -- [only if all tyvar lambdas, and only if this lambda
1055 -- is the RHS of a let]
1057 {- Sept 01: I'm experimenting with getting the
1058 full laziness pass to float out past big lambdsa
1059 | all isTyVar bndrs, -- Only for big lambdas
1060 contIsRhs cont -- Only try the rhs type-lambda floating
1061 -- if this is indeed a right-hand side; otherwise
1062 -- we end up floating the thing out, only for float-in
1063 -- to float it right back in again!
1064 = do (floats, body') <- tryRhsTyLam env bndrs body
1065 return (floats, mkLams bndrs body')
1069 %************************************************************************
1073 %************************************************************************
1075 Note [Eta reduction conditions]
1076 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
1077 We try for eta reduction here, but *only* if we get all the way to an
1078 trivial expression. We don't want to remove extra lambdas unless we
1079 are going to avoid allocating this thing altogether.
1081 There are some particularly delicate points here:
1083 * Eta reduction is not valid in general:
1085 This matters, partly for old-fashioned correctness reasons but,
1086 worse, getting it wrong can yield a seg fault. Consider
1088 h y = case (case y of { True -> f `seq` True; False -> False }) of
1089 True -> ...; False -> ...
1091 If we (unsoundly) eta-reduce f to get f=f, the strictness analyser
1092 says f=bottom, and replaces the (f `seq` True) with just
1093 (f `cast` unsafe-co). BUT, as thing stand, 'f' got arity 1, and it
1094 *keeps* arity 1 (perhaps also wrongly). So CorePrep eta-expands
1095 the definition again, so that it does not termninate after all.
1096 Result: seg-fault because the boolean case actually gets a function value.
1099 So it's important to to the right thing.
1101 * Note [Arity care]: we need to be careful if we just look at f's
1102 arity. Currently (Dec07), f's arity is visible in its own RHS (see
1103 Note [Arity robustness] in SimplEnv) so we must *not* trust the
1104 arity when checking that 'f' is a value. Otherwise we will
1109 Which might change a terminiating program (think (f `seq` e)) to a
1110 non-terminating one. So we check for being a loop breaker first.
1112 However for GlobalIds we can look at the arity; and for primops we
1113 must, since they have no unfolding.
1115 * Regardless of whether 'f' is a value, we always want to
1116 reduce (/\a -> f a) to f
1117 This came up in a RULE: foldr (build (/\a -> g a))
1118 did not match foldr (build (/\b -> ...something complex...))
1119 The type checker can insert these eta-expanded versions,
1120 with both type and dictionary lambdas; hence the slightly
1123 * Never *reduce* arity. For example
1125 Then if h has arity 1 we don't want to eta-reduce because then
1126 f's arity would decrease, and that is bad
1128 These delicacies are why we don't use exprIsTrivial and exprIsHNF here.
1132 tryEtaReduce :: [OutBndr] -> OutExpr -> Maybe OutExpr
1133 tryEtaReduce bndrs body
1134 = go (reverse bndrs) body
1136 incoming_arity = count isId bndrs
1138 go (b : bs) (App fun arg) | ok_arg b arg = go bs fun -- Loop round
1139 go [] fun | ok_fun fun = Just fun -- Success!
1140 go _ _ = Nothing -- Failure!
1142 -- Note [Eta reduction conditions]
1143 ok_fun (App fun (Type ty))
1144 | not (any (`elemVarSet` tyVarsOfType ty) bndrs)
1147 = not (fun_id `elem` bndrs)
1148 && (ok_fun_id fun_id || all ok_lam bndrs)
1151 ok_fun_id fun = fun_arity fun >= incoming_arity
1153 fun_arity fun -- See Note [Arity care]
1154 | isLocalId fun && isLoopBreaker (idOccInfo fun) = 0
1155 | otherwise = idArity fun
1157 ok_lam v = isTyVar v || isDictId v
1159 ok_arg b arg = varToCoreExpr b `cheapEqExpr` arg
1163 %************************************************************************
1167 %************************************************************************
1171 f = \x1..xn -> N ==> f = \x1..xn y1..ym -> N y1..ym
1174 where (in both cases)
1176 * The xi can include type variables
1178 * The yi are all value variables
1180 * N is a NORMAL FORM (i.e. no redexes anywhere)
1181 wanting a suitable number of extra args.
1183 The biggest reason for doing this is for cases like
1189 Here we want to get the lambdas together. A good exmaple is the nofib
1190 program fibheaps, which gets 25% more allocation if you don't do this
1193 We may have to sandwich some coerces between the lambdas
1194 to make the types work. exprEtaExpandArity looks through coerces
1195 when computing arity; and etaExpand adds the coerces as necessary when
1196 actually computing the expansion.
1199 tryEtaExpansion :: DynFlags -> OutExpr -> OutExpr
1200 -- There is at least one runtime binder in the binders
1201 tryEtaExpansion dflags body
1202 = etaExpand fun_arity body
1204 fun_arity = exprEtaExpandArity dflags body
1208 %************************************************************************
1210 \subsection{Floating lets out of big lambdas}
1212 %************************************************************************
1214 Note [Floating and type abstraction]
1215 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
1218 We'd like to float this to
1221 x = /\a. C (y1 a) (y2 a)
1222 for the usual reasons: we want to inline x rather vigorously.
1224 You may think that this kind of thing is rare. But in some programs it is
1225 common. For example, if you do closure conversion you might get:
1227 data a :-> b = forall e. (e -> a -> b) :$ e
1229 f_cc :: forall a. a :-> a
1230 f_cc = /\a. (\e. id a) :$ ()
1232 Now we really want to inline that f_cc thing so that the
1233 construction of the closure goes away.
1235 So I have elaborated simplLazyBind to understand right-hand sides that look
1239 and treat them specially. The real work is done in SimplUtils.abstractFloats,
1240 but there is quite a bit of plumbing in simplLazyBind as well.
1242 The same transformation is good when there are lets in the body:
1244 /\abc -> let(rec) x = e in b
1246 let(rec) x' = /\abc -> let x = x' a b c in e
1248 /\abc -> let x = x' a b c in b
1250 This is good because it can turn things like:
1252 let f = /\a -> letrec g = ... g ... in g
1254 letrec g' = /\a -> ... g' a ...
1256 let f = /\ a -> g' a
1258 which is better. In effect, it means that big lambdas don't impede
1261 This optimisation is CRUCIAL in eliminating the junk introduced by
1262 desugaring mutually recursive definitions. Don't eliminate it lightly!
1264 [May 1999] If we do this transformation *regardless* then we can
1265 end up with some pretty silly stuff. For example,
1268 st = /\ s -> let { x1=r1 ; x2=r2 } in ...
1273 st = /\s -> ...[y1 s/x1, y2 s/x2]
1276 Unless the "..." is a WHNF there is really no point in doing this.
1277 Indeed it can make things worse. Suppose x1 is used strictly,
1280 x1* = case f y of { (a,b) -> e }
1282 If we abstract this wrt the tyvar we then can't do the case inline
1283 as we would normally do.
1285 That's why the whole transformation is part of the same process that
1286 floats let-bindings and constructor arguments out of RHSs. In particular,
1287 it is guarded by the doFloatFromRhs call in simplLazyBind.
1291 abstractFloats :: [OutTyVar] -> SimplEnv -> OutExpr -> SimplM ([OutBind], OutExpr)
1292 abstractFloats main_tvs body_env body
1293 = ASSERT( notNull body_floats )
1294 do { (subst, float_binds) <- mapAccumLM abstract empty_subst body_floats
1295 ; return (float_binds, CoreSubst.substExpr (text "abstract_floats1") subst body) }
1297 main_tv_set = mkVarSet main_tvs
1298 body_floats = getFloats body_env
1299 empty_subst = CoreSubst.mkEmptySubst (seInScope body_env)
1301 abstract :: CoreSubst.Subst -> OutBind -> SimplM (CoreSubst.Subst, OutBind)
1302 abstract subst (NonRec id rhs)
1303 = do { (poly_id, poly_app) <- mk_poly tvs_here id
1304 ; let poly_rhs = mkLams tvs_here rhs'
1305 subst' = CoreSubst.extendIdSubst subst id poly_app
1306 ; return (subst', (NonRec poly_id poly_rhs)) }
1308 rhs' = CoreSubst.substExpr (text "abstract_floats2") subst rhs
1309 tvs_here | any isCoVar main_tvs = main_tvs -- Note [Abstract over coercions]
1311 = varSetElems (main_tv_set `intersectVarSet` exprSomeFreeVars isTyVar rhs')
1313 -- Abstract only over the type variables free in the rhs
1314 -- wrt which the new binding is abstracted. But the naive
1315 -- approach of abstract wrt the tyvars free in the Id's type
1317 -- /\ a b -> let t :: (a,b) = (e1, e2)
1320 -- Here, b isn't free in x's type, but we must nevertheless
1321 -- abstract wrt b as well, because t's type mentions b.
1322 -- Since t is floated too, we'd end up with the bogus:
1323 -- poly_t = /\ a b -> (e1, e2)
1324 -- poly_x = /\ a -> fst (poly_t a *b*)
1325 -- So for now we adopt the even more naive approach of
1326 -- abstracting wrt *all* the tyvars. We'll see if that
1327 -- gives rise to problems. SLPJ June 98
1329 abstract subst (Rec prs)
1330 = do { (poly_ids, poly_apps) <- mapAndUnzipM (mk_poly tvs_here) ids
1331 ; let subst' = CoreSubst.extendSubstList subst (ids `zip` poly_apps)
1332 poly_rhss = [mkLams tvs_here (CoreSubst.substExpr (text "abstract_floats3") subst' rhs)
1334 ; return (subst', Rec (poly_ids `zip` poly_rhss)) }
1336 (ids,rhss) = unzip prs
1337 -- For a recursive group, it's a bit of a pain to work out the minimal
1338 -- set of tyvars over which to abstract:
1339 -- /\ a b c. let x = ...a... in
1340 -- letrec { p = ...x...q...
1341 -- q = .....p...b... } in
1343 -- Since 'x' is abstracted over 'a', the {p,q} group must be abstracted
1344 -- over 'a' (because x is replaced by (poly_x a)) as well as 'b'.
1345 -- Since it's a pain, we just use the whole set, which is always safe
1347 -- If you ever want to be more selective, remember this bizarre case too:
1349 -- Here, we must abstract 'x' over 'a'.
1352 mk_poly tvs_here var
1353 = do { uniq <- getUniqueM
1354 ; let poly_name = setNameUnique (idName var) uniq -- Keep same name
1355 poly_ty = mkForAllTys tvs_here (idType var) -- But new type of course
1356 poly_id = transferPolyIdInfo var tvs_here $ -- Note [transferPolyIdInfo] in Id.lhs
1357 mkLocalId poly_name poly_ty
1358 ; return (poly_id, mkTyApps (Var poly_id) (mkTyVarTys tvs_here)) }
1359 -- In the olden days, it was crucial to copy the occInfo of the original var,
1360 -- because we were looking at occurrence-analysed but as yet unsimplified code!
1361 -- In particular, we mustn't lose the loop breakers. BUT NOW we are looking
1362 -- at already simplified code, so it doesn't matter
1364 -- It's even right to retain single-occurrence or dead-var info:
1365 -- Suppose we started with /\a -> let x = E in B
1366 -- where x occurs once in B. Then we transform to:
1367 -- let x' = /\a -> E in /\a -> let x* = x' a in B
1368 -- where x* has an INLINE prag on it. Now, once x* is inlined,
1369 -- the occurrences of x' will be just the occurrences originally
1373 Note [Abstract over coercions]
1374 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
1375 If a coercion variable (g :: a ~ Int) is free in the RHS, then so is the
1376 type variable a. Rather than sort this mess out, we simply bale out and abstract
1377 wrt all the type variables if any of them are coercion variables.
1380 Historical note: if you use let-bindings instead of a substitution, beware of this:
1382 -- Suppose we start with:
1384 -- x = /\ a -> let g = G in E
1386 -- Then we'll float to get
1388 -- x = let poly_g = /\ a -> G
1389 -- in /\ a -> let g = poly_g a in E
1391 -- But now the occurrence analyser will see just one occurrence
1392 -- of poly_g, not inside a lambda, so the simplifier will
1393 -- PreInlineUnconditionally poly_g back into g! Badk to square 1!
1394 -- (I used to think that the "don't inline lone occurrences" stuff
1395 -- would stop this happening, but since it's the *only* occurrence,
1396 -- PreInlineUnconditionally kicks in first!)
1398 -- Solution: put an INLINE note on g's RHS, so that poly_g seems
1399 -- to appear many times. (NB: mkInlineMe eliminates
1400 -- such notes on trivial RHSs, so do it manually.)
1402 %************************************************************************
1406 %************************************************************************
1408 prepareAlts tries these things:
1410 1. Eliminate alternatives that cannot match, including the
1411 DEFAULT alternative.
1413 2. If the DEFAULT alternative can match only one possible constructor,
1414 then make that constructor explicit.
1416 case e of x { DEFAULT -> rhs }
1418 case e of x { (a,b) -> rhs }
1419 where the type is a single constructor type. This gives better code
1420 when rhs also scrutinises x or e.
1422 3. Returns a list of the constructors that cannot holds in the
1423 DEFAULT alternative (if there is one)
1425 Here "cannot match" includes knowledge from GADTs
1427 It's a good idea do do this stuff before simplifying the alternatives, to
1428 avoid simplifying alternatives we know can't happen, and to come up with
1429 the list of constructors that are handled, to put into the IdInfo of the
1430 case binder, for use when simplifying the alternatives.
1432 Eliminating the default alternative in (1) isn't so obvious, but it can
1435 data Colour = Red | Green | Blue
1444 DEFAULT -> [ case y of ... ]
1446 If we inline h into f, the default case of the inlined h can't happen.
1447 If we don't notice this, we may end up filtering out *all* the cases
1448 of the inner case y, which give us nowhere to go!
1451 prepareAlts :: OutExpr -> OutId -> [InAlt] -> SimplM ([AltCon], [InAlt])
1452 prepareAlts scrut case_bndr' alts
1453 = do { let (alts_wo_default, maybe_deflt) = findDefault alts
1454 alt_cons = [con | (con,_,_) <- alts_wo_default]
1455 imposs_deflt_cons = nub (imposs_cons ++ alt_cons)
1456 -- "imposs_deflt_cons" are handled
1457 -- EITHER by the context,
1458 -- OR by a non-DEFAULT branch in this case expression.
1460 ; default_alts <- prepareDefault case_bndr' mb_tc_app
1461 imposs_deflt_cons maybe_deflt
1463 ; let trimmed_alts = filterOut impossible_alt alts_wo_default
1464 merged_alts = mergeAlts trimmed_alts default_alts
1465 -- We need the mergeAlts in case the new default_alt
1466 -- has turned into a constructor alternative.
1467 -- The merge keeps the inner DEFAULT at the front, if there is one
1468 -- and interleaves the alternatives in the right order
1470 ; return (imposs_deflt_cons, merged_alts) }
1472 mb_tc_app = splitTyConApp_maybe (idType case_bndr')
1473 Just (_, inst_tys) = mb_tc_app
1475 imposs_cons = case scrut of
1476 Var v -> otherCons (idUnfolding v)
1479 impossible_alt :: CoreAlt -> Bool
1480 impossible_alt (con, _, _) | con `elem` imposs_cons = True
1481 impossible_alt (DataAlt con, _, _) = dataConCannotMatch inst_tys con
1482 impossible_alt _ = False
1485 prepareDefault :: OutId -- Case binder; need just for its type. Note that as an
1486 -- OutId, it has maximum information; this is important.
1487 -- Test simpl013 is an example
1488 -> Maybe (TyCon, [Type]) -- Type of scrutinee, decomposed
1489 -> [AltCon] -- These cons can't happen when matching the default
1490 -> Maybe InExpr -- Rhs
1491 -> SimplM [InAlt] -- Still unsimplified
1492 -- We use a list because it's what mergeAlts expects,
1494 --------- Fill in known constructor -----------
1495 prepareDefault case_bndr (Just (tycon, inst_tys)) imposs_cons (Just deflt_rhs)
1496 | -- This branch handles the case where we are
1497 -- scrutinisng an algebraic data type
1498 isAlgTyCon tycon -- It's a data type, tuple, or unboxed tuples.
1499 , not (isNewTyCon tycon) -- We can have a newtype, if we are just doing an eval:
1500 -- case x of { DEFAULT -> e }
1501 -- and we don't want to fill in a default for them!
1502 , Just all_cons <- tyConDataCons_maybe tycon
1503 , not (null all_cons) -- This is a tricky corner case. If the data type has no constructors,
1504 -- which GHC allows, then the case expression will have at most a default
1505 -- alternative. We don't want to eliminate that alternative, because the
1506 -- invariant is that there's always one alternative. It's more convenient
1508 -- case x of { DEFAULT -> e }
1509 -- as it is, rather than transform it to
1510 -- error "case cant match"
1511 -- which would be quite legitmate. But it's a really obscure corner, and
1512 -- not worth wasting code on.
1513 , let imposs_data_cons = [con | DataAlt con <- imposs_cons] -- We now know it's a data type
1514 impossible con = con `elem` imposs_data_cons || dataConCannotMatch inst_tys con
1515 = case filterOut impossible all_cons of
1516 [] -> return [] -- Eliminate the default alternative
1517 -- altogether if it can't match
1519 [con] -> -- It matches exactly one constructor, so fill it in
1520 do { tick (FillInCaseDefault case_bndr)
1522 ; let (ex_tvs, co_tvs, arg_ids) =
1523 dataConRepInstPat us con inst_tys
1524 ; return [(DataAlt con, ex_tvs ++ co_tvs ++ arg_ids, deflt_rhs)] }
1526 _ -> return [(DEFAULT, [], deflt_rhs)]
1528 | debugIsOn, isAlgTyCon tycon, not (isOpenTyCon tycon), null (tyConDataCons tycon)
1529 -- Check for no data constructors
1530 -- This can legitimately happen for type families, so don't report that
1531 = pprTrace "prepareDefault" (ppr case_bndr <+> ppr tycon)
1532 $ return [(DEFAULT, [], deflt_rhs)]
1534 --------- Catch-all cases -----------
1535 prepareDefault _case_bndr _bndr_ty _imposs_cons (Just deflt_rhs)
1536 = return [(DEFAULT, [], deflt_rhs)]
1538 prepareDefault _case_bndr _bndr_ty _imposs_cons Nothing
1539 = return [] -- No default branch
1544 %************************************************************************
1548 %************************************************************************
1550 mkCase tries these things
1552 1. Merge Nested Cases
1554 case e of b { ==> case e of b {
1555 p1 -> rhs1 p1 -> rhs1
1557 pm -> rhsm pm -> rhsm
1558 _ -> case b of b' { pn -> let b'=b in rhsn
1560 ... po -> let b'=b in rhso
1561 po -> rhso _ -> let b'=b in rhsd
1565 which merges two cases in one case when -- the default alternative of
1566 the outer case scrutises the same variable as the outer case. This
1567 transformation is called Case Merging. It avoids that the same
1568 variable is scrutinised multiple times.
1570 2. Eliminate Identity Case
1576 and similar friends.
1578 3. Merge identical alternatives.
1579 If several alternatives are identical, merge them into
1580 a single DEFAULT alternative. I've occasionally seen this
1581 making a big difference:
1583 case e of =====> case e of
1584 C _ -> f x D v -> ....v....
1585 D v -> ....v.... DEFAULT -> f x
1588 The point is that we merge common RHSs, at least for the DEFAULT case.
1589 [One could do something more elaborate but I've never seen it needed.]
1590 To avoid an expensive test, we just merge branches equal to the *first*
1591 alternative; this picks up the common cases
1592 a) all branches equal
1593 b) some branches equal to the DEFAULT (which occurs first)
1595 The case where Merge Identical Alternatives transformation showed up
1596 was like this (base/Foreign/C/Err/Error.lhs):
1602 where @is@ was something like
1604 p `is` n = p /= (-1) && p == n
1606 This gave rise to a horrible sequence of cases
1613 and similarly in cascade for all the join points!
1617 mkCase, mkCase1, mkCase2
1620 -> [OutAlt] -- Alternatives in standard (increasing) order
1623 --------------------------------------------------
1624 -- 1. Merge Nested Cases
1625 --------------------------------------------------
1627 mkCase dflags scrut outer_bndr ((DEFAULT, _, deflt_rhs) : outer_alts)
1628 | dopt Opt_CaseMerge dflags
1629 , Case (Var inner_scrut_var) inner_bndr _ inner_alts <- deflt_rhs
1630 , inner_scrut_var == outer_bndr
1631 = do { tick (CaseMerge outer_bndr)
1633 ; let wrap_alt (con, args, rhs) = ASSERT( outer_bndr `notElem` args )
1634 (con, args, wrap_rhs rhs)
1635 -- Simplifier's no-shadowing invariant should ensure
1636 -- that outer_bndr is not shadowed by the inner patterns
1637 wrap_rhs rhs = Let (NonRec inner_bndr (Var outer_bndr)) rhs
1638 -- The let is OK even for unboxed binders,
1640 wrapped_alts | isDeadBinder inner_bndr = inner_alts
1641 | otherwise = map wrap_alt inner_alts
1643 merged_alts = mergeAlts outer_alts wrapped_alts
1644 -- NB: mergeAlts gives priority to the left
1647 -- DEFAULT -> case x of
1650 -- When we merge, we must ensure that e1 takes
1651 -- precedence over e2 as the value for A!
1653 ; mkCase1 dflags scrut outer_bndr merged_alts
1655 -- Warning: don't call mkCase recursively!
1656 -- Firstly, there's no point, because inner alts have already had
1657 -- mkCase applied to them, so they won't have a case in their default
1658 -- Secondly, if you do, you get an infinite loop, because the bindCaseBndr
1659 -- in munge_rhs may put a case into the DEFAULT branch!
1661 mkCase dflags scrut bndr alts = mkCase1 dflags scrut bndr alts
1663 --------------------------------------------------
1664 -- 2. Eliminate Identity Case
1665 --------------------------------------------------
1667 mkCase1 _dflags scrut case_bndr alts -- Identity case
1668 | all identity_alt alts
1669 = do { tick (CaseIdentity case_bndr)
1670 ; return (re_cast scrut) }
1672 identity_alt (con, args, rhs) = check_eq con args (de_cast rhs)
1674 check_eq DEFAULT _ (Var v) = v == case_bndr
1675 check_eq (LitAlt lit') _ (Lit lit) = lit == lit'
1676 check_eq (DataAlt con) args rhs = rhs `cheapEqExpr` mkConApp con (arg_tys ++ varsToCoreExprs args)
1677 || rhs `cheapEqExpr` Var case_bndr
1678 check_eq _ _ _ = False
1680 arg_tys = map Type (tyConAppArgs (idType case_bndr))
1683 -- case e of x { _ -> x `cast` c }
1684 -- And we definitely want to eliminate this case, to give
1686 -- So we throw away the cast from the RHS, and reconstruct
1687 -- it at the other end. All the RHS casts must be the same
1688 -- if (all identity_alt alts) holds.
1690 -- Don't worry about nested casts, because the simplifier combines them
1691 de_cast (Cast e _) = e
1694 re_cast scrut = case head alts of
1695 (_,_,Cast _ co) -> Cast scrut co
1698 --------------------------------------------------
1699 -- 3. Merge Identical Alternatives
1700 --------------------------------------------------
1701 mkCase1 dflags scrut case_bndr ((_con1,bndrs1,rhs1) : con_alts)
1702 | all isDeadBinder bndrs1 -- Remember the default
1703 , length filtered_alts < length con_alts -- alternative comes first
1704 -- Also Note [Dead binders]
1705 = do { tick (AltMerge case_bndr)
1706 ; mkCase2 dflags scrut case_bndr alts' }
1708 alts' = (DEFAULT, [], rhs1) : filtered_alts
1709 filtered_alts = filter keep con_alts
1710 keep (_con,bndrs,rhs) = not (all isDeadBinder bndrs && rhs `cheapEqExpr` rhs1)
1712 mkCase1 dflags scrut bndr alts = mkCase2 dflags scrut bndr alts
1714 --------------------------------------------------
1716 --------------------------------------------------
1717 mkCase2 _dflags scrut bndr alts
1718 = return (Case scrut bndr (coreAltsType alts) alts)
1722 ~~~~~~~~~~~~~~~~~~~~
1723 Note that dead-ness is maintained by the simplifier, so that it is
1724 accurate after simplification as well as before.
1727 Note [Cascading case merge]
1728 ~~~~~~~~~~~~~~~~~~~~~~~~~~~
1729 Case merging should cascade in one sweep, because it
1733 DEFAULT -> case a of b
1734 DEFAULT -> case b of c {
1741 DEFAULT -> case a of b
1742 DEFAULT -> let c = b in e
1743 A -> let c = b in ea
1748 DEFAULT -> let b = a in let c = b in e
1749 A -> let b = a in let c = b in ea
1750 B -> let b = a in eb
1754 However here's a tricky case that we still don't catch, and I don't
1755 see how to catch it in one pass:
1757 case x of c1 { I# a1 ->
1760 DEFAULT -> case x of c3 { I# a2 ->
1763 After occurrence analysis (and its binder-swap) we get this
1765 case x of c1 { I# a1 ->
1766 let x = c1 in -- Binder-swap addition
1769 DEFAULT -> case x of c3 { I# a2 ->
1772 When we simplify the inner case x, we'll see that
1773 x=c1=I# a1. So we'll bind a2 to a1, and get
1775 case x of c1 { I# a1 ->
1778 DEFAULT -> case a1 of ...
1780 This is corect, but we can't do a case merge in this sweep
1781 because c2 /= a1. Reason: the binding c1=I# a1 went inwards
1782 without getting changed to c1=I# c2.
1784 I don't think this is worth fixing, even if I knew how. It'll
1785 all come out in the next pass anyway.