2 % (c) The GRASP/AQUA Project, Glasgow University, 1992-1998
4 %************************************************************************
6 \section[OccurAnal]{Occurrence analysis pass}
8 %************************************************************************
10 The occurrence analyser re-typechecks a core expression, returning a new
11 core expression with (hopefully) improved usage information.
15 -- The above warning supression flag is a temporary kludge.
16 -- While working on this module you are encouraged to remove it and fix
17 -- any warnings in the module. See
18 -- http://hackage.haskell.org/trac/ghc/wiki/Commentary/CodingStyle#Warnings
22 occurAnalysePgm, occurAnalyseExpr
25 #include "HsVersions.h"
29 import CoreUtils ( exprIsTrivial, isDefaultAlt )
32 import BasicTypes ( OccInfo(..), isOneOcc, InterestingCxt )
37 import Maybes ( orElse )
38 import Digraph ( stronglyConnCompR, SCC(..) )
39 import PrelNames ( buildIdKey, foldrIdKey, runSTRepIdKey, augmentIdKey )
40 import Unique ( Unique )
41 import UniqFM ( keysUFM, intersectsUFM, intersectUFM_C, foldUFM_Directly )
42 import Util ( mapAndUnzip )
49 %************************************************************************
51 \subsection[OccurAnal-main]{Counting occurrences: main function}
53 %************************************************************************
55 Here's the externally-callable interface:
58 occurAnalysePgm :: [CoreBind] -> [CoreBind]
60 = snd (go initOccEnv binds)
62 go :: OccEnv -> [CoreBind] -> (UsageDetails, [CoreBind])
66 = (final_usage, bind' ++ binds')
68 (bs_usage, binds') = go env binds
69 (final_usage, bind') = occAnalBind env bind bs_usage
71 occurAnalyseExpr :: CoreExpr -> CoreExpr
72 -- Do occurrence analysis, and discard occurence info returned
73 occurAnalyseExpr expr = snd (occAnal initOccEnv expr)
77 %************************************************************************
79 \subsection[OccurAnal-main]{Counting occurrences: main function}
81 %************************************************************************
89 -> UsageDetails -- Usage details of scope
90 -> (UsageDetails, -- Of the whole let(rec)
93 occAnalBind env (NonRec binder rhs) body_usage
94 | not (binder `usedIn` body_usage) -- It's not mentioned
97 | otherwise -- It's mentioned in the body
98 = (body_usage' +++ addRuleUsage rhs_usage binder, -- Note [Rules are extra RHSs]
99 [NonRec tagged_binder rhs'])
101 (body_usage', tagged_binder) = tagBinder body_usage binder
102 (rhs_usage, rhs') = occAnalRhs env tagged_binder rhs
107 Dropping dead code for recursive bindings is done in a very simple way:
109 the entire set of bindings is dropped if none of its binders are
110 mentioned in its body; otherwise none are.
112 This seems to miss an obvious improvement.
124 Now 'f' is unused! But it's OK! Dependency analysis will sort this
125 out into a letrec for 'g' and a 'let' for 'f', and then 'f' will get
126 dropped. It isn't easy to do a perfect job in one blow. Consider
137 Note [Loop breaking and RULES]
138 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
139 Loop breaking is surprisingly subtle. First read the section 4 of
140 "Secrets of the GHC inliner". This describes our basic plan.
142 However things are made quite a bit more complicated by RULES. Remember
144 * Note [Rules are extra RHSs]
145 ~~~~~~~~~~~~~~~~~~~~~~~~~~~
146 A RULE for 'f' is like an extra RHS for 'f'. That way the "parent"
147 keeps the specialised "children" alive. If the parent dies
148 (because it isn't referenced any more), then the children will die
149 too (unless they are already referenced directly).
151 To that end, we build a Rec group for each cyclic strongly
153 *treating f's rules as extra RHSs for 'f'*.
155 When we make the Rec groups we include variables free in *either*
156 LHS *or* RHS of the rule. The former might seems silly, but see
157 Note [Rule dependency info].
159 So in Example [eftInt], eftInt and eftIntFB will be put in the
160 same Rec, even though their 'main' RHSs are both non-recursive.
162 * Note [Rules are visible in their own rec group]
163 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
164 We want the rules for 'f' to be visible in f's right-hand side.
165 And we'd like them to be visible in other functions in f's Rec
166 group. E.g. in Example [Specialisation rules] we want f' rule
167 to be visible in both f's RHS, and fs's RHS.
169 This means that we must simplify the RULEs first, before looking
170 at any of the definitions. This is done by Simplify.simplRecBind,
171 when it calls addLetIdInfo.
173 * Note [Choosing loop breakers]
174 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
175 We avoid infinite inlinings by choosing loop breakers, and
176 ensuring that a loop breaker cuts each loop. But what is a
177 "loop"? In particular, a RULES is like an equation for 'f' that
178 is *always* inlined if it are applicable. We do *not* disable
179 rules for loop-breakers. It's up to whoever makes the rules to
180 make sure that the rules themselves alwasys terminate. See Note
181 [Rules for recursive functions] in Simplify.lhs
184 f's RHS mentions g, and
185 g has a RULE that mentions h, and
186 h has a RULE that mentions f
188 then we *must* choose f to be a loop breaker. In general, take the
189 free variables of f's RHS, and augment it with all the variables
190 reachable by RULES from those starting points. That is the whole
191 reason for computing rule_fv_env in occAnalBind. (Of course we
192 only consider free vars that are also binders in this Rec group.)
194 Note that when we compute this rule_fv_env, we only consider variables
195 free in the *RHS* of the rule, in contrast to the way we build the
196 Rec group in the first place (Note [Rule dependency info])
198 Note that in Example [eftInt], *neither* eftInt *nor* eftIntFB is
199 chosen as a loop breaker, because their RHSs don't mention each other.
200 And indeed both can be inlined safely.
202 Note that the edges of the graph we use for computing loop breakers
203 are not the same as the edges we use for computing the Rec blocks.
204 That's why we compute
205 rec_edges for the Rec block analysis
206 loop_breaker_edges for the loop breaker analysis
209 * Note [Weak loop breakers]
210 ~~~~~~~~~~~~~~~~~~~~~~~~~
211 There is a last nasty wrinkle. Suppose we have
221 Remmber that we simplify the RULES before any RHS (see Note
222 [Rules are visible in their own rec group] above).
224 So we must *not* postInlineUnconditionally 'g', even though
225 its RHS turns out to be trivial. (I'm assuming that 'g' is
226 not choosen as a loop breaker.)
228 We "solve" this by making g a "weak" or "rules-only" loop breaker,
229 with OccInfo = IAmLoopBreaker True. A normal "strong" loop breaker
230 has IAmLoopBreaker False. So
232 Inline postInlineUnconditinoally
233 IAmLoopBreaker False no no
234 IAmLoopBreaker True yes no
237 The **sole** reason for this kind of loop breaker is so that
238 postInlineUnconditionally does not fire. Ugh.
240 * Note [Rule dependency info]
241 ~~~~~~~~~~~~~~~~~~~~~~~~~~~
242 The VarSet in a SpecInfo is used for dependency analysis in the
243 occurrence analyser. We must track free vars in *both* lhs and rhs. Why both?
247 Then if we substitute y for x, we'd better do so in the
248 rule's LHS too, so we'd better ensure the dependency is respected
253 Example (from GHC.Enum):
255 eftInt :: Int# -> Int# -> [Int]
256 eftInt x y = ...(non-recursive)...
258 {-# INLINE [0] eftIntFB #-}
259 eftIntFB :: (Int -> r -> r) -> r -> Int# -> Int# -> r
260 eftIntFB c n x y = ...(non-recursive)...
263 "eftInt" [~1] forall x y. eftInt x y = build (\ c n -> eftIntFB c n x y)
264 "eftIntList" [1] eftIntFB (:) [] = eftInt
267 Example [Specialisation rules]
268 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
269 Consider this group, which is typical of what SpecConstr builds:
271 fs a = ....f (C a)....
272 f x = ....f (C a)....
273 {-# RULE f (C a) = fs a #-}
275 So 'f' and 'fs' are in the same Rec group (since f refers to fs via its RULE).
277 But watch out! If 'fs' is not chosen as a loop breaker, we may get an infinite loop:
278 - the RULE is applied in f's RHS (see Note [Self-recursive rules] in Simplify
279 - fs is inlined (say it's small)
280 - now there's another opportunity to apply the RULE
282 This showed up when compiling Control.Concurrent.Chan.getChanContents.
286 occAnalBind env (Rec pairs) body_usage
287 | not (any (`usedIn` body_usage) bndrs) -- NB: look at body_usage, not total_usage
288 = (body_usage, []) -- Dead code
290 = (final_usage, map ({-# SCC "occAnalBind.dofinal" #-} do_final_bind) sccs)
292 bndrs = map fst pairs
293 bndr_set = mkVarSet bndrs
295 ---------------------------------------
296 -- See Note [Loop breaking]
297 ---------------------------------------
299 -------------Dependency analysis ------------------------------
300 occ_anald :: [(Id, (UsageDetails, CoreExpr))]
301 -- The UsageDetails here are strictly those arising from the RHS
302 -- *not* from any rules in the Id
303 occ_anald = [(bndr, occAnalRhs env bndr rhs) | (bndr,rhs) <- pairs]
305 total_usage = foldl add_usage body_usage occ_anald
306 add_usage body_usage (bndr, (rhs_usage, _))
307 = body_usage +++ addRuleUsage rhs_usage bndr
309 (final_usage, tagged_bndrs) = tagBinders total_usage bndrs
310 final_bndrs | isEmptyVarSet all_rule_fvs = tagged_bndrs
311 | otherwise = map tag_rule_var tagged_bndrs
313 tag_rule_var bndr | bndr `elemVarSet` all_rule_fvs = makeLoopBreaker True bndr
315 all_rule_fvs = bndr_set `intersectVarSet` foldr (unionVarSet . idRuleVars) emptyVarSet bndrs
316 -- Mark the binder with OccInfo saying "no preInlineUnconditionally" if
317 -- it is used in any rule (lhs or rhs) of the recursive group
319 ---- stuff for dependency analysis of binds -------------------------------
320 sccs :: [SCC (Node Details)]
321 sccs = {-# SCC "occAnalBind.scc" #-} stronglyConnCompR rec_edges
323 rec_edges :: [Node Details] -- The binders are tagged with correct occ-info
324 rec_edges = {-# SCC "occAnalBind.assoc" #-} zipWith make_node final_bndrs occ_anald
325 make_node tagged_bndr (_bndr, (rhs_usage, rhs))
326 = ((tagged_bndr, rhs, rhs_fvs), idUnique tagged_bndr, out_edges)
328 rhs_fvs = intersectUFM_C (\b _ -> b) bndr_set rhs_usage
329 out_edges = keysUFM (rhs_fvs `unionVarSet` idRuleVars tagged_bndr)
332 -- (a -> b) means a mentions b
333 -- Given the usage details (a UFM that gives occ info for each free var of
334 -- the RHS) we can get the list of free vars -- or rather their Int keys --
335 -- by just extracting the keys from the finite map. Grimy, but fast.
336 -- Previously we had this:
337 -- [ bndr | bndr <- bndrs,
338 -- maybeToBool (lookupVarEnv rhs_usage bndr)]
339 -- which has n**2 cost, and this meant that edges_from alone
340 -- consumed 10% of total runtime!
342 ---- Stuff to "re-constitute" bindings from dependency-analysis info ------
343 do_final_bind (AcyclicSCC ((bndr, rhs, _), _, _)) = NonRec bndr rhs
344 do_final_bind (CyclicSCC cycle)
345 | no_rules = Rec (reOrderCycle cycle)
346 | otherwise = Rec (concatMap reOrderRec (stronglyConnCompR loop_breaker_edges))
347 where -- See Note [Choosing loop breakers] for looop_breker_edges
348 loop_breaker_edges = map mk_node cycle
349 mk_node (details@(bndr, rhs, rhs_fvs), k, _) = (details, k, new_ks)
351 new_ks = keysUFM (extendFvs rule_fv_env rhs_fvs rhs_fvs)
354 ------------------------------------
355 rule_fv_env :: IdEnv IdSet -- Variables from this group mentioned in RHS of rules
356 -- Domain is *subset* of bound vars (others have no rule fvs)
357 rule_fv_env = rule_loop init_rule_fvs
359 no_rules = null init_rule_fvs
360 init_rule_fvs = [(b, rule_fvs)
362 , let rule_fvs = idRuleRhsVars b `intersectVarSet` bndr_set
363 , not (isEmptyVarSet rule_fvs)]
365 rule_loop :: [(Id,IdSet)] -> IdEnv IdSet -- Finds fixpoint
368 | otherwise = rule_loop new_fv_list
370 env = mkVarEnv init_rule_fvs
371 (no_change, new_fv_list) = mapAccumL bump True fv_list
372 bump no_change (b,fvs)
373 | new_fvs `subVarSet` fvs = (no_change, (b,fvs))
374 | otherwise = (False, (b,new_fvs `unionVarSet` fvs))
376 new_fvs = extendFvs env emptyVarSet fvs
378 idRuleRhsVars :: Id -> VarSet
379 -- Just the variables free on the *rhs* of a rule
380 -- See Note [Choosing loop breakers]
381 idRuleRhsVars id = foldr (unionVarSet . ruleRhsFreeVars) emptyVarSet (idCoreRules id)
383 extendFvs :: IdEnv IdSet -> IdSet -> IdSet -> IdSet
384 -- (extendFVs env fvs s) returns (fvs `union` env(s))
385 extendFvs env fvs id_set
386 = foldUFM_Directly add fvs id_set
389 = case lookupVarEnv_Directly env uniq of
390 Just fvs' -> fvs' `unionVarSet` fvs
394 @reOrderRec@ is applied to the list of (binder,rhs) pairs for a cyclic
395 strongly connected component (there's guaranteed to be a cycle). It returns the
397 a) in a better order,
398 b) with some of the Ids having a IAmALoopBreaker pragma
400 The "loop-breaker" Ids are sufficient to break all cycles in the SCC. This means
401 that the simplifier can guarantee not to loop provided it never records an inlining
402 for these no-inline guys.
404 Furthermore, the order of the binds is such that if we neglect dependencies
405 on the no-inline Ids then the binds are topologically sorted. This means
406 that the simplifier will generally do a good job if it works from top bottom,
407 recording inlinings for any Ids which aren't marked as "no-inline" as it goes.
410 [June 98: I don't understand the following paragraphs, and I've
411 changed the a=b case again so that it isn't a special case any more.]
413 Here's a case that bit me:
421 Re-ordering doesn't change the order of bindings, but there was no loop-breaker.
423 My solution was to make a=b bindings record b as Many, rather like INLINE bindings.
424 Perhaps something cleverer would suffice.
429 type Node details = (details, Unique, [Unique]) -- The Ints are gotten from the Unique,
430 -- which is gotten from the Id.
431 type Details = (Id, -- Binder
433 IdSet) -- RHS free vars (*not* include rules)
435 reOrderRec :: SCC (Node Details)
437 -- Sorted into a plausible order. Enough of the Ids have
438 -- IAmALoopBreaker pragmas that there are no loops left.
439 reOrderRec (AcyclicSCC ((bndr, rhs, _), _, _)) = [(bndr, rhs)]
440 reOrderRec (CyclicSCC cycle) = reOrderCycle cycle
442 reOrderCycle :: [Node Details] -> [(Id,CoreExpr)]
444 = panic "reOrderCycle"
445 reOrderCycle [bind] -- Common case of simple self-recursion
446 = [(makeLoopBreaker False bndr, rhs)]
448 ((bndr, rhs, _), _, _) = bind
450 reOrderCycle (bind : binds)
451 = -- Choose a loop breaker, mark it no-inline,
452 -- do SCC analysis on the rest, and recursively sort them out
453 concatMap reOrderRec (stronglyConnCompR unchosen) ++
454 [(makeLoopBreaker False bndr, rhs)]
457 (chosen_bind, unchosen) = choose_loop_breaker bind (score bind) [] binds
458 (bndr, rhs, _) = chosen_bind
460 -- This loop looks for the bind with the lowest score
461 -- to pick as the loop breaker. The rest accumulate in
462 choose_loop_breaker (details,_,_) loop_sc acc []
463 = (details, acc) -- Done
465 choose_loop_breaker loop_bind loop_sc acc (bind : binds)
466 | sc < loop_sc -- Lower score so pick this new one
467 = choose_loop_breaker bind sc (loop_bind : acc) binds
469 | otherwise -- No lower so don't pick it
470 = choose_loop_breaker loop_bind loop_sc (bind : acc) binds
474 score :: Node Details -> Int -- Higher score => less likely to be picked as loop breaker
475 score ((bndr, rhs, _), _, _)
476 | workerExists (idWorkerInfo bndr) = 10
477 -- Note [Worker inline loop]
479 | exprIsTrivial rhs = 4 -- Practically certain to be inlined
480 -- Used to have also: && not (isExportedId bndr)
481 -- But I found this sometimes cost an extra iteration when we have
482 -- rec { d = (a,b); a = ...df...; b = ...df...; df = d }
483 -- where df is the exported dictionary. Then df makes a really
484 -- bad choice for loop breaker
486 | is_con_app rhs = 2 -- Data types help with cases
489 | inlineCandidate bndr rhs = 1 -- Likely to be inlined
490 -- Note [Inline candidates]
494 inlineCandidate :: Id -> CoreExpr -> Bool
495 inlineCandidate id (Note InlineMe _) = True
496 inlineCandidate id rhs = isOneOcc (idOccInfo id)
500 -- It's really really important to inline dictionaries. Real
501 -- example (the Enum Ordering instance from GHC.Base):
503 -- rec f = \ x -> case d of (p,q,r) -> p x
504 -- g = \ x -> case d of (p,q,r) -> q x
507 -- Here, f and g occur just once; but we can't inline them into d.
508 -- On the other hand we *could* simplify those case expressions if
509 -- we didn't stupidly choose d as the loop breaker.
510 -- But we won't because constructor args are marked "Many".
511 -- Inlining dictionaries is really essential to unravelling
512 -- the loops in static numeric dictionaries, see GHC.Float.
514 -- Cheap and cheerful; the simplifer moves casts out of the way
515 -- The lambda case is important to spot x = /\a. C (f a)
516 -- which comes up when C is a dictionary constructor and
517 -- f is a default method.
518 -- Example: the instance for Show (ST s a) in GHC.ST
520 -- However we *also* treat (\x. C p q) as a con-app-like thing,
521 -- Note [Closure conversion]
522 is_con_app (Var v) = isDataConWorkId v
523 is_con_app (App f _) = is_con_app f
524 is_con_app (Lam b e) = is_con_app e
525 is_con_app (Note _ e) = is_con_app e
526 is_con_app other = False
528 makeLoopBreaker :: Bool -> Id -> Id
529 -- Set the loop-breaker flag
530 -- See Note [Weak loop breakers]
531 makeLoopBreaker weak bndr = setIdOccInfo bndr (IAmALoopBreaker weak)
534 Note [Worker inline loop]
535 ~~~~~~~~~~~~~~~~~~~~~~~~
536 Never choose a wrapper as the loop breaker! Because
537 wrappers get auto-generated inlinings when importing, and
538 that can lead to an infinite inlining loop. For example:
540 $wfoo x = ....foo x....
542 {-loop brk-} foo x = ...$wfoo x...
545 The interface file sees the unfolding for $wfoo, and sees that foo is
546 strict (and hence it gets an auto-generated wrapper). Result: an
547 infinite inlining in the importing scope. So be a bit careful if you
548 change this. A good example is Tree.repTree in
549 nofib/spectral/minimax. If the repTree wrapper is chosen as the loop
550 breaker then compiling Game.hs goes into an infinite loop (this
551 happened when we gave is_con_app a lower score than inline candidates).
553 Note [Closure conversion]
554 ~~~~~~~~~~~~~~~~~~~~~~~~~
555 We treat (\x. C p q) as a high-score candidate in the letrec scoring algorithm.
556 The immediate motivation came from the result of a closure-conversion transformation
557 which generated code like this:
559 data Clo a b = forall c. Clo (c -> a -> b) c
561 ($:) :: Clo a b -> a -> b
562 Clo f env $: x = f env x
564 rec { plus = Clo plus1 ()
566 ; plus1 _ n = Clo plus2 n
569 ; plus2 (Succ m) n = Succ (plus $: m $: n) }
571 If we inline 'plus' and 'plus1', everything unravels nicely. But if
572 we choose 'plus1' as the loop breaker (which is entirely possible
573 otherwise), the loop does not unravel nicely.
576 @occAnalRhs@ deals with the question of bindings where the Id is marked
577 by an INLINE pragma. For these we record that anything which occurs
578 in its RHS occurs many times. This pessimistically assumes that ths
579 inlined binder also occurs many times in its scope, but if it doesn't
580 we'll catch it next time round. At worst this costs an extra simplifier pass.
581 ToDo: try using the occurrence info for the inline'd binder.
583 [March 97] We do the same for atomic RHSs. Reason: see notes with reOrderRec.
584 [June 98, SLPJ] I've undone this change; I don't understand it. See notes with reOrderRec.
589 -> Id -> CoreExpr -- Binder and rhs
590 -- For non-recs the binder is alrady tagged
591 -- with occurrence info
592 -> (UsageDetails, CoreExpr)
594 occAnalRhs env id rhs
597 ctxt | certainly_inline id = env
598 | otherwise = rhsCtxt
599 -- Note that we generally use an rhsCtxt. This tells the occ anal n
600 -- that it's looking at an RHS, which has an effect in occAnalApp
602 -- But there's a problem. Consider
607 -- First time round, it looks as if x1 and x2 occur as an arg of a
608 -- let-bound constructor ==> give them a many-occurrence.
609 -- But then x3 is inlined (unconditionally as it happens) and
610 -- next time round, x2 will be, and the next time round x1 will be
611 -- Result: multiple simplifier iterations. Sigh.
612 -- Crude solution: use rhsCtxt for things that occur just once...
614 certainly_inline id = case idOccInfo id of
615 OneOcc in_lam one_br _ -> not in_lam && one_br
622 addRuleUsage :: UsageDetails -> Id -> UsageDetails
623 -- Add the usage from RULES in Id to the usage
624 addRuleUsage usage id
625 = foldVarSet add usage (idRuleVars id)
627 add v u = addOneOcc u v NoOccInfo -- Give a non-committal binder info
628 -- (i.e manyOcc) because many copies
629 -- of the specialised thing can appear
637 -> (UsageDetails, -- Gives info only about the "interesting" Ids
640 occAnal env (Type t) = (emptyDetails, Type t)
641 occAnal env (Var v) = (mkOneOcc env v False, Var v)
642 -- At one stage, I gathered the idRuleVars for v here too,
643 -- which in a way is the right thing to do.
644 -- Btu that went wrong right after specialisation, when
645 -- the *occurrences* of the overloaded function didn't have any
646 -- rules in them, so the *specialised* versions looked as if they
647 -- weren't used at all.
650 We regard variables that occur as constructor arguments as "dangerousToDup":
654 f x = let y = expensive x in
656 (case z of {(p,q)->q}, case z of {(p,q)->q})
659 We feel free to duplicate the WHNF (True,y), but that means
660 that y may be duplicated thereby.
662 If we aren't careful we duplicate the (expensive x) call!
663 Constructors are rather like lambdas in this way.
666 occAnal env expr@(Lit lit) = (emptyDetails, expr)
670 occAnal env (Note InlineMe body)
671 = case occAnal env body of { (usage, body') ->
672 (mapVarEnv markMany usage, Note InlineMe body')
675 occAnal env (Note note@(SCC cc) body)
676 = case occAnal env body of { (usage, body') ->
677 (mapVarEnv markInsideSCC usage, Note note body')
680 occAnal env (Note note body)
681 = case occAnal env body of { (usage, body') ->
682 (usage, Note note body')
685 occAnal env (Cast expr co)
686 = case occAnal env expr of { (usage, expr') ->
687 (markRhsUds env True usage, Cast expr' co)
688 -- If we see let x = y `cast` co
689 -- then mark y as 'Many' so that we don't
690 -- immediately inline y again.
695 occAnal env app@(App fun arg)
696 = occAnalApp env (collectArgs app) False
698 -- Ignore type variables altogether
699 -- (a) occurrences inside type lambdas only not marked as InsideLam
700 -- (b) type variables not in environment
702 occAnal env expr@(Lam x body) | isTyVar x
703 = case occAnal env body of { (body_usage, body') ->
704 (body_usage, Lam x body')
707 -- For value lambdas we do a special hack. Consider
709 -- If we did nothing, x is used inside the \y, so would be marked
710 -- as dangerous to dup. But in the common case where the abstraction
711 -- is applied to two arguments this is over-pessimistic.
712 -- So instead, we just mark each binder with its occurrence
713 -- info in the *body* of the multiple lambda.
714 -- Then, the simplifier is careful when partially applying lambdas.
716 occAnal env expr@(Lam _ _)
717 = case occAnal env_body body of { (body_usage, body') ->
719 (final_usage, tagged_binders) = tagBinders body_usage binders
720 -- URGH! Sept 99: we don't seem to be able to use binders' here, because
721 -- we get linear-typed things in the resulting program that we can't handle yet.
722 -- (e.g. PrelShow) TODO
724 really_final_usage = if linear then
727 mapVarEnv markInsideLam final_usage
730 mkLams tagged_binders body') }
732 env_body = vanillaCtxt -- Body is (no longer) an RhsContext
733 (binders, body) = collectBinders expr
734 binders' = oneShotGroup env binders
735 linear = all is_one_shot binders'
736 is_one_shot b = isId b && isOneShotBndr b
738 occAnal env (Case scrut bndr ty alts)
739 = case occ_anal_scrut scrut alts of { (scrut_usage, scrut') ->
740 case mapAndUnzip (occAnalAlt alt_env bndr) alts of { (alts_usage_s, alts') ->
742 alts_usage = foldr1 combineAltsUsageDetails alts_usage_s
743 alts_usage' = addCaseBndrUsage alts_usage
744 (alts_usage1, tagged_bndr) = tagBinder alts_usage' bndr
745 total_usage = scrut_usage +++ alts_usage1
747 total_usage `seq` (total_usage, Case scrut' tagged_bndr ty alts') }}
749 -- The case binder gets a usage of either "many" or "dead", never "one".
750 -- Reason: we like to inline single occurrences, to eliminate a binding,
751 -- but inlining a case binder *doesn't* eliminate a binding.
752 -- We *don't* want to transform
753 -- case x of w { (p,q) -> f w }
755 -- case x of w { (p,q) -> f (p,q) }
756 addCaseBndrUsage usage = case lookupVarEnv usage bndr of
758 Just occ -> extendVarEnv usage bndr (markMany occ)
760 alt_env = setVanillaCtxt env
761 -- Consider x = case v of { True -> (p,q); ... }
762 -- Then it's fine to inline p and q
764 occ_anal_scrut (Var v) (alt1 : other_alts)
765 | not (null other_alts) || not (isDefaultAlt alt1)
766 = (mkOneOcc env v True, Var v)
767 occ_anal_scrut scrut alts = occAnal vanillaCtxt scrut
768 -- No need for rhsCtxt
770 occAnal env (Let bind body)
771 = case occAnal env body of { (body_usage, body') ->
772 case occAnalBind env bind body_usage of { (final_usage, new_binds) ->
773 (final_usage, mkLets new_binds body') }}
776 = case mapAndUnzip (occAnal arg_env) args of { (arg_uds_s, args') ->
777 (foldr (+++) emptyDetails arg_uds_s, args')}
779 arg_env = vanillaCtxt
782 Applications are dealt with specially because we want
783 the "build hack" to work.
786 occAnalApp env (Var fun, args) is_rhs
787 = case args_stuff of { (args_uds, args') ->
789 final_args_uds = markRhsUds env is_pap args_uds
791 (fun_uds +++ final_args_uds, mkApps (Var fun) args') }
793 fun_uniq = idUnique fun
794 fun_uds = mkOneOcc env fun (valArgCount args > 0)
795 is_pap = isDataConWorkId fun || valArgCount args < idArity fun
797 -- Hack for build, fold, runST
798 args_stuff | fun_uniq == buildIdKey = appSpecial env 2 [True,True] args
799 | fun_uniq == augmentIdKey = appSpecial env 2 [True,True] args
800 | fun_uniq == foldrIdKey = appSpecial env 3 [False,True] args
801 | fun_uniq == runSTRepIdKey = appSpecial env 2 [True] args
802 -- (foldr k z xs) may call k many times, but it never
803 -- shares a partial application of k; hence [False,True]
804 -- This means we can optimise
805 -- foldr (\x -> let v = ...x... in \y -> ...v...) z xs
806 -- by floating in the v
808 | otherwise = occAnalArgs env args
811 occAnalApp env (fun, args) is_rhs
812 = case occAnal (addAppCtxt env args) fun of { (fun_uds, fun') ->
813 -- The addAppCtxt is a bit cunning. One iteration of the simplifier
814 -- often leaves behind beta redexs like
816 -- Here we would like to mark x,y as one-shot, and treat the whole
817 -- thing much like a let. We do this by pushing some True items
818 -- onto the context stack.
820 case occAnalArgs env args of { (args_uds, args') ->
822 final_uds = fun_uds +++ args_uds
824 (final_uds, mkApps fun' args') }}
827 markRhsUds :: OccEnv -- Check if this is a RhsEnv
828 -> Bool -- and this is true
829 -> UsageDetails -- The do markMany on this
831 -- We mark the free vars of the argument of a constructor or PAP
832 -- as "many", if it is the RHS of a let(rec).
833 -- This means that nothing gets inlined into a constructor argument
834 -- position, which is what we want. Typically those constructor
835 -- arguments are just variables, or trivial expressions.
837 -- This is the *whole point* of the isRhsEnv predicate
838 markRhsUds env is_pap arg_uds
839 | isRhsEnv env && is_pap = mapVarEnv markMany arg_uds
840 | otherwise = arg_uds
844 -> Int -> CtxtTy -- Argument number, and context to use for it
846 -> (UsageDetails, [CoreExpr])
847 appSpecial env n ctxt args
850 arg_env = vanillaCtxt
852 go n [] = (emptyDetails, []) -- Too few args
854 go 1 (arg:args) -- The magic arg
855 = case occAnal (setCtxt arg_env ctxt) arg of { (arg_uds, arg') ->
856 case occAnalArgs env args of { (args_uds, args') ->
857 (arg_uds +++ args_uds, arg':args') }}
860 = case occAnal arg_env arg of { (arg_uds, arg') ->
861 case go (n-1) args of { (args_uds, args') ->
862 (arg_uds +++ args_uds, arg':args') }}
868 If the case binder occurs at all, the other binders effectively do too.
870 case e of x { (a,b) -> rhs }
873 If e turns out to be (e1,e2) we indeed get something like
874 let a = e1; b = e2; x = (a,b) in rhs
876 Note [Aug 06]: I don't think this is necessary any more, and it helpe
877 to know when binders are unused. See esp the call to
878 isDeadBinder in Simplify.mkDupableAlt
881 occAnalAlt env case_bndr (con, bndrs, rhs)
882 = case occAnal env rhs of { (rhs_usage, rhs') ->
884 (final_usage, tagged_bndrs) = tagBinders rhs_usage bndrs
885 final_bndrs = tagged_bndrs -- See Note [Aug06] above
887 final_bndrs | case_bndr `elemVarEnv` final_usage = bndrs
888 | otherwise = tagged_bndrs
889 -- Leave the binders untagged if the case
890 -- binder occurs at all; see note above
893 (final_usage, (con, final_bndrs, rhs')) }
897 %************************************************************************
899 \subsection[OccurAnal-types]{OccEnv}
901 %************************************************************************
905 = OccEnv OccEncl -- Enclosing context information
906 CtxtTy -- Tells about linearity
908 -- OccEncl is used to control whether to inline into constructor arguments
910 -- x = (p,q) -- Don't inline p or q
911 -- y = /\a -> (p a, q a) -- Still don't inline p or q
912 -- z = f (p,q) -- Do inline p,q; it may make a rule fire
913 -- So OccEncl tells enought about the context to know what to do when
914 -- we encounter a contructor application or PAP.
917 = OccRhs -- RHS of let(rec), albeit perhaps inside a type lambda
918 -- Don't inline into constructor args here
919 | OccVanilla -- Argument of function, body of lambda, scruintee of case etc.
920 -- Do inline into constructor args here
925 -- True:ctxt Analysing a function-valued expression that will be
928 -- False:ctxt Analysing a function-valued expression that may
929 -- be applied many times; but when it is,
930 -- the CtxtTy inside applies
933 initOccEnv = OccEnv OccRhs []
935 vanillaCtxt = OccEnv OccVanilla []
936 rhsCtxt = OccEnv OccRhs []
938 isRhsEnv (OccEnv OccRhs _) = True
939 isRhsEnv (OccEnv OccVanilla _) = False
941 setVanillaCtxt :: OccEnv -> OccEnv
942 setVanillaCtxt (OccEnv OccRhs ctxt_ty) = OccEnv OccVanilla ctxt_ty
943 setVanillaCtxt other_env = other_env
945 setCtxt :: OccEnv -> CtxtTy -> OccEnv
946 setCtxt (OccEnv encl _) ctxt = OccEnv encl ctxt
948 oneShotGroup :: OccEnv -> [CoreBndr] -> [CoreBndr]
949 -- The result binders have one-shot-ness set that they might not have had originally.
950 -- This happens in (build (\cn -> e)). Here the occurrence analyser
951 -- linearity context knows that c,n are one-shot, and it records that fact in
952 -- the binder. This is useful to guide subsequent float-in/float-out tranformations
954 oneShotGroup (OccEnv encl ctxt) bndrs
957 go ctxt [] rev_bndrs = reverse rev_bndrs
959 go (lin_ctxt:ctxt) (bndr:bndrs) rev_bndrs
960 | isId bndr = go ctxt bndrs (bndr':rev_bndrs)
962 bndr' | lin_ctxt = setOneShotLambda bndr
965 go ctxt (bndr:bndrs) rev_bndrs = go ctxt bndrs (bndr:rev_bndrs)
967 addAppCtxt (OccEnv encl ctxt) args
968 = OccEnv encl (replicate (valArgCount args) True ++ ctxt)
971 %************************************************************************
973 \subsection[OccurAnal-types]{OccEnv}
975 %************************************************************************
978 type UsageDetails = IdEnv OccInfo -- A finite map from ids to their usage
980 (+++), combineAltsUsageDetails
981 :: UsageDetails -> UsageDetails -> UsageDetails
984 = plusVarEnv_C addOccInfo usage1 usage2
986 combineAltsUsageDetails usage1 usage2
987 = plusVarEnv_C orOccInfo usage1 usage2
989 addOneOcc :: UsageDetails -> Id -> OccInfo -> UsageDetails
990 addOneOcc usage id info
991 = plusVarEnv_C addOccInfo usage (unitVarEnv id info)
992 -- ToDo: make this more efficient
994 emptyDetails = (emptyVarEnv :: UsageDetails)
996 usedIn :: Id -> UsageDetails -> Bool
997 v `usedIn` details = isExportedId v || v `elemVarEnv` details
999 type IdWithOccInfo = Id
1001 tagBinders :: UsageDetails -- Of scope
1003 -> (UsageDetails, -- Details with binders removed
1004 [IdWithOccInfo]) -- Tagged binders
1006 tagBinders usage binders
1008 usage' = usage `delVarEnvList` binders
1009 uss = map (setBinderOcc usage) binders
1011 usage' `seq` (usage', uss)
1013 tagBinder :: UsageDetails -- Of scope
1015 -> (UsageDetails, -- Details with binders removed
1016 IdWithOccInfo) -- Tagged binders
1018 tagBinder usage binder
1020 usage' = usage `delVarEnv` binder
1021 binder' = setBinderOcc usage binder
1023 usage' `seq` (usage', binder')
1025 setBinderOcc :: UsageDetails -> CoreBndr -> CoreBndr
1026 setBinderOcc usage bndr
1027 | isTyVar bndr = bndr
1028 | isExportedId bndr = case idOccInfo bndr of
1030 other -> setIdOccInfo bndr NoOccInfo
1031 -- Don't use local usage info for visible-elsewhere things
1032 -- BUT *do* erase any IAmALoopBreaker annotation, because we're
1033 -- about to re-generate it and it shouldn't be "sticky"
1035 | otherwise = setIdOccInfo bndr occ_info
1037 occ_info = lookupVarEnv usage bndr `orElse` IAmDead
1041 %************************************************************************
1043 \subsection{Operations over OccInfo}
1045 %************************************************************************
1048 mkOneOcc :: OccEnv -> Id -> InterestingCxt -> UsageDetails
1049 mkOneOcc env id int_cxt
1050 | isLocalId id = unitVarEnv id (OneOcc False True int_cxt)
1051 | otherwise = emptyDetails
1053 markMany, markInsideLam, markInsideSCC :: OccInfo -> OccInfo
1055 markMany IAmDead = IAmDead
1056 markMany other = NoOccInfo
1058 markInsideSCC occ = markMany occ
1060 markInsideLam (OneOcc _ one_br int_cxt) = OneOcc True one_br int_cxt
1061 markInsideLam occ = occ
1063 addOccInfo, orOccInfo :: OccInfo -> OccInfo -> OccInfo
1065 addOccInfo IAmDead info2 = info2
1066 addOccInfo info1 IAmDead = info1
1067 addOccInfo info1 info2 = NoOccInfo
1069 -- (orOccInfo orig new) is used
1070 -- when combining occurrence info from branches of a case
1072 orOccInfo IAmDead info2 = info2
1073 orOccInfo info1 IAmDead = info1
1074 orOccInfo (OneOcc in_lam1 one_branch1 int_cxt1)
1075 (OneOcc in_lam2 one_branch2 int_cxt2)
1076 = OneOcc (in_lam1 || in_lam2)
1077 False -- False, because it occurs in both branches
1078 (int_cxt1 && int_cxt2)
1079 orOccInfo info1 info2 = NoOccInfo