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 occurAnalysePgm, occurAnalyseExpr
18 #include "HsVersions.h"
22 import CoreUtils ( exprIsTrivial, isDefaultAlt )
30 import Maybes ( orElse )
31 import Digraph ( stronglyConnCompR, SCC(..) )
32 import PrelNames ( buildIdKey, foldrIdKey, runSTRepIdKey, augmentIdKey )
33 import Unique ( Unique )
34 import UniqFM ( keysUFM, intersectUFM_C, foldUFM_Directly )
35 import Util ( mapAndUnzip )
42 %************************************************************************
44 \subsection[OccurAnal-main]{Counting occurrences: main function}
46 %************************************************************************
48 Here's the externally-callable interface:
51 occurAnalysePgm :: [CoreBind] -> [CoreBind]
53 = snd (go initOccEnv binds)
55 go :: OccEnv -> [CoreBind] -> (UsageDetails, [CoreBind])
59 = (final_usage, bind' ++ binds')
61 (bs_usage, binds') = go env binds
62 (final_usage, bind') = occAnalBind env bind bs_usage
64 occurAnalyseExpr :: CoreExpr -> CoreExpr
65 -- Do occurrence analysis, and discard occurence info returned
66 occurAnalyseExpr expr = snd (occAnal initOccEnv expr)
70 %************************************************************************
72 \subsection[OccurAnal-main]{Counting occurrences: main function}
74 %************************************************************************
82 -> UsageDetails -- Usage details of scope
83 -> (UsageDetails, -- Of the whole let(rec)
86 occAnalBind env (NonRec binder rhs) body_usage
87 | not (binder `usedIn` body_usage) -- It's not mentioned
90 | otherwise -- It's mentioned in the body
91 = (body_usage' +++ addRuleUsage rhs_usage binder, -- Note [Rules are extra RHSs]
92 [NonRec tagged_binder rhs'])
94 (body_usage', tagged_binder) = tagBinder body_usage binder
95 (rhs_usage, rhs') = occAnalRhs env tagged_binder rhs
100 Dropping dead code for recursive bindings is done in a very simple way:
102 the entire set of bindings is dropped if none of its binders are
103 mentioned in its body; otherwise none are.
105 This seems to miss an obvious improvement.
117 Now 'f' is unused! But it's OK! Dependency analysis will sort this
118 out into a letrec for 'g' and a 'let' for 'f', and then 'f' will get
119 dropped. It isn't easy to do a perfect job in one blow. Consider
130 Note [Loop breaking and RULES]
131 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
132 Loop breaking is surprisingly subtle. First read the section 4 of
133 "Secrets of the GHC inliner". This describes our basic plan.
135 However things are made quite a bit more complicated by RULES. Remember
137 * Note [Rules are extra RHSs]
138 ~~~~~~~~~~~~~~~~~~~~~~~~~~~
139 A RULE for 'f' is like an extra RHS for 'f'. That way the "parent"
140 keeps the specialised "children" alive. If the parent dies
141 (because it isn't referenced any more), then the children will die
142 too (unless they are already referenced directly).
144 To that end, we build a Rec group for each cyclic strongly
146 *treating f's rules as extra RHSs for 'f'*.
148 When we make the Rec groups we include variables free in *either*
149 LHS *or* RHS of the rule. The former might seems silly, but see
150 Note [Rule dependency info].
152 So in Example [eftInt], eftInt and eftIntFB will be put in the
153 same Rec, even though their 'main' RHSs are both non-recursive.
155 * Note [Rules are visible in their own rec group]
156 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
157 We want the rules for 'f' to be visible in f's right-hand side.
158 And we'd like them to be visible in other functions in f's Rec
159 group. E.g. in Example [Specialisation rules] we want f' rule
160 to be visible in both f's RHS, and fs's RHS.
162 This means that we must simplify the RULEs first, before looking
163 at any of the definitions. This is done by Simplify.simplRecBind,
164 when it calls addLetIdInfo.
166 * Note [Choosing loop breakers]
167 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
168 We avoid infinite inlinings by choosing loop breakers, and
169 ensuring that a loop breaker cuts each loop. But what is a
170 "loop"? In particular, a RULES is like an equation for 'f' that
171 is *always* inlined if it are applicable. We do *not* disable
172 rules for loop-breakers. It's up to whoever makes the rules to
173 make sure that the rules themselves alwasys terminate. See Note
174 [Rules for recursive functions] in Simplify.lhs
177 f's RHS mentions g, and
178 g has a RULE that mentions h, and
179 h has a RULE that mentions f
181 then we *must* choose f to be a loop breaker. In general, take the
182 free variables of f's RHS, and augment it with all the variables
183 reachable by RULES from those starting points. That is the whole
184 reason for computing rule_fv_env in occAnalBind. (Of course we
185 only consider free vars that are also binders in this Rec group.)
187 Note that when we compute this rule_fv_env, we only consider variables
188 free in the *RHS* of the rule, in contrast to the way we build the
189 Rec group in the first place (Note [Rule dependency info])
191 Note that in Example [eftInt], *neither* eftInt *nor* eftIntFB is
192 chosen as a loop breaker, because their RHSs don't mention each other.
193 And indeed both can be inlined safely.
195 Note that the edges of the graph we use for computing loop breakers
196 are not the same as the edges we use for computing the Rec blocks.
197 That's why we compute
198 rec_edges for the Rec block analysis
199 loop_breaker_edges for the loop breaker analysis
202 * Note [Weak loop breakers]
203 ~~~~~~~~~~~~~~~~~~~~~~~~~
204 There is a last nasty wrinkle. Suppose we have
214 Remmber that we simplify the RULES before any RHS (see Note
215 [Rules are visible in their own rec group] above).
217 So we must *not* postInlineUnconditionally 'g', even though
218 its RHS turns out to be trivial. (I'm assuming that 'g' is
219 not choosen as a loop breaker.)
221 We "solve" this by making g a "weak" or "rules-only" loop breaker,
222 with OccInfo = IAmLoopBreaker True. A normal "strong" loop breaker
223 has IAmLoopBreaker False. So
225 Inline postInlineUnconditinoally
226 IAmLoopBreaker False no no
227 IAmLoopBreaker True yes no
230 The **sole** reason for this kind of loop breaker is so that
231 postInlineUnconditionally does not fire. Ugh.
233 * Note [Rule dependency info]
234 ~~~~~~~~~~~~~~~~~~~~~~~~~~~
235 The VarSet in a SpecInfo is used for dependency analysis in the
236 occurrence analyser. We must track free vars in *both* lhs and rhs. Why both?
240 Then if we substitute y for x, we'd better do so in the
241 rule's LHS too, so we'd better ensure the dependency is respected
246 Example (from GHC.Enum):
248 eftInt :: Int# -> Int# -> [Int]
249 eftInt x y = ...(non-recursive)...
251 {-# INLINE [0] eftIntFB #-}
252 eftIntFB :: (Int -> r -> r) -> r -> Int# -> Int# -> r
253 eftIntFB c n x y = ...(non-recursive)...
256 "eftInt" [~1] forall x y. eftInt x y = build (\ c n -> eftIntFB c n x y)
257 "eftIntList" [1] eftIntFB (:) [] = eftInt
260 Example [Specialisation rules]
261 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
262 Consider this group, which is typical of what SpecConstr builds:
264 fs a = ....f (C a)....
265 f x = ....f (C a)....
266 {-# RULE f (C a) = fs a #-}
268 So 'f' and 'fs' are in the same Rec group (since f refers to fs via its RULE).
270 But watch out! If 'fs' is not chosen as a loop breaker, we may get an infinite loop:
271 - the RULE is applied in f's RHS (see Note [Self-recursive rules] in Simplify
272 - fs is inlined (say it's small)
273 - now there's another opportunity to apply the RULE
275 This showed up when compiling Control.Concurrent.Chan.getChanContents.
279 occAnalBind env (Rec pairs) body_usage
280 = foldr occAnalRec (body_usage, []) sccs
281 -- For a recursive group, we
282 -- * occ-analyse all the RHSs
283 -- * compute strongly-connected components
284 -- * feed those components to occAnalRec
286 -------------Dependency analysis ------------------------------
287 bndr_set = mkVarSet (map fst pairs)
289 sccs :: [SCC (Node Details)]
290 sccs = {-# SCC "occAnalBind.scc" #-} stronglyConnCompR rec_edges
292 rec_edges :: [Node Details]
293 rec_edges = {-# SCC "occAnalBind.assoc" #-} map make_node pairs
295 make_node (bndr, rhs)
296 = (ND bndr rhs' rhs_usage rhs_fvs, idUnique bndr, out_edges)
298 (rhs_usage, rhs') = occAnalRhs env bndr rhs
299 rhs_fvs = intersectUFM_C (\b _ -> b) bndr_set rhs_usage
300 out_edges = keysUFM (rhs_fvs `unionVarSet` idRuleVars bndr)
301 -- (a -> b) means a mentions b
302 -- Given the usage details (a UFM that gives occ info for each free var of
303 -- the RHS) we can get the list of free vars -- or rather their Int keys --
304 -- by just extracting the keys from the finite map. Grimy, but fast.
305 -- Previously we had this:
306 -- [ bndr | bndr <- bndrs,
307 -- maybeToBool (lookupVarEnv rhs_usage bndr)]
308 -- which has n**2 cost, and this meant that edges_from alone
309 -- consumed 10% of total runtime!
311 -----------------------------
312 occAnalRec :: SCC (Node Details) -> (UsageDetails, [CoreBind])
313 -> (UsageDetails, [CoreBind])
315 -- The NonRec case is just like a Let (NonRec ...) above
316 occAnalRec (AcyclicSCC (ND bndr rhs rhs_usage _, _, _)) (body_usage, binds)
317 | not (bndr `usedIn` body_usage)
318 = (body_usage, binds)
320 | otherwise -- It's mentioned in the body
321 = (body_usage' +++ addRuleUsage rhs_usage bndr, -- Note [Rules are extra RHSs]
322 NonRec tagged_bndr rhs : binds)
324 (body_usage', tagged_bndr) = tagBinder body_usage bndr
327 -- The Rec case is the interesting one
328 -- See Note [Loop breaking]
329 occAnalRec (CyclicSCC nodes) (body_usage, binds)
330 | not (any (`usedIn` body_usage) bndrs) -- NB: look at body_usage, not total_usage
331 = (body_usage, binds) -- Dead code
333 | otherwise -- At this point we always build a single Rec
334 = (final_usage, Rec pairs : binds)
337 bndrs = [b | (ND b _ _ _, _, _) <- nodes]
338 bndr_set = mkVarSet bndrs
340 ----------------------------
341 -- Tag the binders with their occurrence info
342 total_usage = foldl add_usage body_usage nodes
343 add_usage body_usage (ND bndr _ rhs_usage _, _, _)
344 = body_usage +++ addRuleUsage rhs_usage bndr
345 (final_usage, tagged_nodes) = mapAccumL tag_node total_usage nodes
347 tag_node :: UsageDetails -> Node Details -> (UsageDetails, Node Details)
348 -- (a) Tag the binders in the details with occ info
349 -- (b) Mark the binder with OccInfo saying "no preInlineUnconditionally" if
350 -- it is used in any rule (lhs or rhs) of the recursive group
351 -- See Note [Weak loop breakers]
352 tag_node usage (ND bndr rhs rhs_usage rhs_fvs, k, ks)
353 = (usage `delVarEnv` bndr, (ND bndr2 rhs rhs_usage rhs_fvs, k, ks))
355 bndr2 | bndr `elemVarSet` all_rule_fvs = makeLoopBreaker True bndr1
357 bndr1 = setBinderOcc usage bndr
358 all_rule_fvs = bndr_set `intersectVarSet` foldr (unionVarSet . idRuleVars)
361 ----------------------------
362 -- Now reconstruct the cycle
363 pairs | no_rules = reOrderCycle tagged_nodes
364 | otherwise = concatMap reOrderRec (stronglyConnCompR loop_breaker_edges)
366 -- See Note [Choosing loop breakers] for looop_breaker_edges
367 loop_breaker_edges = map mk_node tagged_nodes
368 mk_node (details@(ND _ _ _ rhs_fvs), k, _) = (details, k, new_ks)
370 new_ks = keysUFM (extendFvs rule_fv_env rhs_fvs rhs_fvs)
372 ------------------------------------
373 rule_fv_env :: IdEnv IdSet -- Variables from this group mentioned in RHS of rules
374 -- Domain is *subset* of bound vars (others have no rule fvs)
375 rule_fv_env = rule_loop init_rule_fvs
377 no_rules = null init_rule_fvs
378 init_rule_fvs = [(b, rule_fvs)
380 , let rule_fvs = idRuleRhsVars b `intersectVarSet` bndr_set
381 , not (isEmptyVarSet rule_fvs)]
383 rule_loop :: [(Id,IdSet)] -> IdEnv IdSet -- Finds fixpoint
386 | otherwise = rule_loop new_fv_list
388 env = mkVarEnv init_rule_fvs
389 (no_change, new_fv_list) = mapAccumL bump True fv_list
390 bump no_change (b,fvs)
391 | new_fvs `subVarSet` fvs = (no_change, (b,fvs))
392 | otherwise = (False, (b,new_fvs `unionVarSet` fvs))
394 new_fvs = extendFvs env emptyVarSet fvs
396 idRuleRhsVars :: Id -> VarSet
397 -- Just the variables free on the *rhs* of a rule
398 -- See Note [Choosing loop breakers]
399 idRuleRhsVars id = foldr (unionVarSet . ruleRhsFreeVars) emptyVarSet (idCoreRules id)
401 extendFvs :: IdEnv IdSet -> IdSet -> IdSet -> IdSet
402 -- (extendFVs env fvs s) returns (fvs `union` env(s))
403 extendFvs env fvs id_set
404 = foldUFM_Directly add fvs id_set
407 = case lookupVarEnv_Directly env uniq of
408 Just fvs' -> fvs' `unionVarSet` fvs
412 @reOrderRec@ is applied to the list of (binder,rhs) pairs for a cyclic
413 strongly connected component (there's guaranteed to be a cycle). It returns the
415 a) in a better order,
416 b) with some of the Ids having a IAmALoopBreaker pragma
418 The "loop-breaker" Ids are sufficient to break all cycles in the SCC. This means
419 that the simplifier can guarantee not to loop provided it never records an inlining
420 for these no-inline guys.
422 Furthermore, the order of the binds is such that if we neglect dependencies
423 on the no-inline Ids then the binds are topologically sorted. This means
424 that the simplifier will generally do a good job if it works from top bottom,
425 recording inlinings for any Ids which aren't marked as "no-inline" as it goes.
428 [June 98: I don't understand the following paragraphs, and I've
429 changed the a=b case again so that it isn't a special case any more.]
431 Here's a case that bit me:
439 Re-ordering doesn't change the order of bindings, but there was no loop-breaker.
441 My solution was to make a=b bindings record b as Many, rather like INLINE bindings.
442 Perhaps something cleverer would suffice.
447 type Node details = (details, Unique, [Unique]) -- The Ints are gotten from the Unique,
448 -- which is gotten from the Id.
449 data Details = ND Id -- Binder
451 UsageDetails -- Full usage from RHS (*not* including rules)
452 IdSet -- Other binders from this Rec group mentioned on RHS
453 -- (derivable from UsageDetails but cached here)
455 reOrderRec :: SCC (Node Details)
457 -- Sorted into a plausible order. Enough of the Ids have
458 -- IAmALoopBreaker pragmas that there are no loops left.
459 reOrderRec (AcyclicSCC (ND bndr rhs _ _, _, _)) = [(bndr, rhs)]
460 reOrderRec (CyclicSCC cycle) = reOrderCycle cycle
462 reOrderCycle :: [Node Details] -> [(Id,CoreExpr)]
464 = panic "reOrderCycle"
465 reOrderCycle [bind] -- Common case of simple self-recursion
466 = [(makeLoopBreaker False bndr, rhs)]
468 (ND bndr rhs _ _, _, _) = bind
470 reOrderCycle (bind : binds)
471 = -- Choose a loop breaker, mark it no-inline,
472 -- do SCC analysis on the rest, and recursively sort them out
473 concatMap reOrderRec (stronglyConnCompR unchosen) ++
474 [(makeLoopBreaker False bndr, rhs)]
477 (chosen_bind, unchosen) = choose_loop_breaker bind (score bind) [] binds
478 ND bndr rhs _ _ = chosen_bind
480 -- This loop looks for the bind with the lowest score
481 -- to pick as the loop breaker. The rest accumulate in
482 choose_loop_breaker (details,_,_) _loop_sc acc []
483 = (details, acc) -- Done
485 choose_loop_breaker loop_bind loop_sc acc (bind : binds)
486 | sc < loop_sc -- Lower score so pick this new one
487 = choose_loop_breaker bind sc (loop_bind : acc) binds
489 | otherwise -- No lower so don't pick it
490 = choose_loop_breaker loop_bind loop_sc (bind : acc) binds
494 score :: Node Details -> Int -- Higher score => less likely to be picked as loop breaker
495 score (ND bndr rhs _ _, _, _)
496 | workerExists (idWorkerInfo bndr) = 10
497 -- Note [Worker inline loop]
499 | exprIsTrivial rhs = 4 -- Practically certain to be inlined
500 -- Used to have also: && not (isExportedId bndr)
501 -- But I found this sometimes cost an extra iteration when we have
502 -- rec { d = (a,b); a = ...df...; b = ...df...; df = d }
503 -- where df is the exported dictionary. Then df makes a really
504 -- bad choice for loop breaker
506 | is_con_app rhs = 2 -- Data types help with cases
509 | inlineCandidate bndr rhs = 1 -- Likely to be inlined
510 -- Note [Inline candidates]
514 inlineCandidate :: Id -> CoreExpr -> Bool
515 inlineCandidate _ (Note InlineMe _) = True
516 inlineCandidate id _ = isOneOcc (idOccInfo id)
520 -- It's really really important to inline dictionaries. Real
521 -- example (the Enum Ordering instance from GHC.Base):
523 -- rec f = \ x -> case d of (p,q,r) -> p x
524 -- g = \ x -> case d of (p,q,r) -> q x
527 -- Here, f and g occur just once; but we can't inline them into d.
528 -- On the other hand we *could* simplify those case expressions if
529 -- we didn't stupidly choose d as the loop breaker.
530 -- But we won't because constructor args are marked "Many".
531 -- Inlining dictionaries is really essential to unravelling
532 -- the loops in static numeric dictionaries, see GHC.Float.
534 -- Cheap and cheerful; the simplifer moves casts out of the way
535 -- The lambda case is important to spot x = /\a. C (f a)
536 -- which comes up when C is a dictionary constructor and
537 -- f is a default method.
538 -- Example: the instance for Show (ST s a) in GHC.ST
540 -- However we *also* treat (\x. C p q) as a con-app-like thing,
541 -- Note [Closure conversion]
542 is_con_app (Var v) = isDataConWorkId v
543 is_con_app (App f _) = is_con_app f
544 is_con_app (Lam _ e) = is_con_app e
545 is_con_app (Note _ e) = is_con_app e
548 makeLoopBreaker :: Bool -> Id -> Id
549 -- Set the loop-breaker flag
550 -- See Note [Weak loop breakers]
551 makeLoopBreaker weak bndr = setIdOccInfo bndr (IAmALoopBreaker weak)
554 Note [Worker inline loop]
555 ~~~~~~~~~~~~~~~~~~~~~~~~
556 Never choose a wrapper as the loop breaker! Because
557 wrappers get auto-generated inlinings when importing, and
558 that can lead to an infinite inlining loop. For example:
560 $wfoo x = ....foo x....
562 {-loop brk-} foo x = ...$wfoo x...
565 The interface file sees the unfolding for $wfoo, and sees that foo is
566 strict (and hence it gets an auto-generated wrapper). Result: an
567 infinite inlining in the importing scope. So be a bit careful if you
568 change this. A good example is Tree.repTree in
569 nofib/spectral/minimax. If the repTree wrapper is chosen as the loop
570 breaker then compiling Game.hs goes into an infinite loop (this
571 happened when we gave is_con_app a lower score than inline candidates).
573 Note [Closure conversion]
574 ~~~~~~~~~~~~~~~~~~~~~~~~~
575 We treat (\x. C p q) as a high-score candidate in the letrec scoring algorithm.
576 The immediate motivation came from the result of a closure-conversion transformation
577 which generated code like this:
579 data Clo a b = forall c. Clo (c -> a -> b) c
581 ($:) :: Clo a b -> a -> b
582 Clo f env $: x = f env x
584 rec { plus = Clo plus1 ()
586 ; plus1 _ n = Clo plus2 n
589 ; plus2 (Succ m) n = Succ (plus $: m $: n) }
591 If we inline 'plus' and 'plus1', everything unravels nicely. But if
592 we choose 'plus1' as the loop breaker (which is entirely possible
593 otherwise), the loop does not unravel nicely.
596 @occAnalRhs@ deals with the question of bindings where the Id is marked
597 by an INLINE pragma. For these we record that anything which occurs
598 in its RHS occurs many times. This pessimistically assumes that ths
599 inlined binder also occurs many times in its scope, but if it doesn't
600 we'll catch it next time round. At worst this costs an extra simplifier pass.
601 ToDo: try using the occurrence info for the inline'd binder.
603 [March 97] We do the same for atomic RHSs. Reason: see notes with reOrderRec.
604 [June 98, SLPJ] I've undone this change; I don't understand it. See notes with reOrderRec.
609 -> Id -> CoreExpr -- Binder and rhs
610 -- For non-recs the binder is alrady tagged
611 -- with occurrence info
612 -> (UsageDetails, CoreExpr)
614 occAnalRhs env id rhs
617 ctxt | certainly_inline id = env
618 | otherwise = rhsCtxt
619 -- Note that we generally use an rhsCtxt. This tells the occ anal n
620 -- that it's looking at an RHS, which has an effect in occAnalApp
622 -- But there's a problem. Consider
627 -- First time round, it looks as if x1 and x2 occur as an arg of a
628 -- let-bound constructor ==> give them a many-occurrence.
629 -- But then x3 is inlined (unconditionally as it happens) and
630 -- next time round, x2 will be, and the next time round x1 will be
631 -- Result: multiple simplifier iterations. Sigh.
632 -- Crude solution: use rhsCtxt for things that occur just once...
634 certainly_inline id = case idOccInfo id of
635 OneOcc in_lam one_br _ -> not in_lam && one_br
642 addRuleUsage :: UsageDetails -> Id -> UsageDetails
643 -- Add the usage from RULES in Id to the usage
644 addRuleUsage usage id
645 = foldVarSet add usage (idRuleVars id)
647 add v u = addOneOcc u v NoOccInfo -- Give a non-committal binder info
648 -- (i.e manyOcc) because many copies
649 -- of the specialised thing can appear
657 -> (UsageDetails, -- Gives info only about the "interesting" Ids
660 occAnal _ (Type t) = (emptyDetails, Type t)
661 occAnal env (Var v) = (mkOneOcc env v False, Var v)
662 -- At one stage, I gathered the idRuleVars for v here too,
663 -- which in a way is the right thing to do.
664 -- But that went wrong right after specialisation, when
665 -- the *occurrences* of the overloaded function didn't have any
666 -- rules in them, so the *specialised* versions looked as if they
667 -- weren't used at all.
670 We regard variables that occur as constructor arguments as "dangerousToDup":
674 f x = let y = expensive x in
676 (case z of {(p,q)->q}, case z of {(p,q)->q})
679 We feel free to duplicate the WHNF (True,y), but that means
680 that y may be duplicated thereby.
682 If we aren't careful we duplicate the (expensive x) call!
683 Constructors are rather like lambdas in this way.
686 occAnal _ expr@(Lit _) = (emptyDetails, expr)
690 occAnal env (Note InlineMe body)
691 = case occAnal env body of { (usage, body') ->
692 (mapVarEnv markMany usage, Note InlineMe body')
695 occAnal env (Note note@(SCC _) body)
696 = case occAnal env body of { (usage, body') ->
697 (mapVarEnv markInsideSCC usage, Note note body')
700 occAnal env (Note note body)
701 = case occAnal env body of { (usage, body') ->
702 (usage, Note note body')
705 occAnal env (Cast expr co)
706 = case occAnal env expr of { (usage, expr') ->
707 (markRhsUds env True usage, Cast expr' co)
708 -- If we see let x = y `cast` co
709 -- then mark y as 'Many' so that we don't
710 -- immediately inline y again.
715 occAnal env app@(App _ _)
716 = occAnalApp env (collectArgs app)
718 -- Ignore type variables altogether
719 -- (a) occurrences inside type lambdas only not marked as InsideLam
720 -- (b) type variables not in environment
722 occAnal env (Lam x body) | isTyVar x
723 = case occAnal env body of { (body_usage, body') ->
724 (body_usage, Lam x body')
727 -- For value lambdas we do a special hack. Consider
729 -- If we did nothing, x is used inside the \y, so would be marked
730 -- as dangerous to dup. But in the common case where the abstraction
731 -- is applied to two arguments this is over-pessimistic.
732 -- So instead, we just mark each binder with its occurrence
733 -- info in the *body* of the multiple lambda.
734 -- Then, the simplifier is careful when partially applying lambdas.
736 occAnal env expr@(Lam _ _)
737 = case occAnal env_body body of { (body_usage, body') ->
739 (final_usage, tagged_binders) = tagBinders body_usage binders
740 -- URGH! Sept 99: we don't seem to be able to use binders' here, because
741 -- we get linear-typed things in the resulting program that we can't handle yet.
742 -- (e.g. PrelShow) TODO
744 really_final_usage = if linear then
747 mapVarEnv markInsideLam final_usage
750 mkLams tagged_binders body') }
752 env_body = vanillaCtxt -- Body is (no longer) an RhsContext
753 (binders, body) = collectBinders expr
754 binders' = oneShotGroup env binders
755 linear = all is_one_shot binders'
756 is_one_shot b = isId b && isOneShotBndr b
758 occAnal env (Case scrut bndr ty alts)
759 = case occ_anal_scrut scrut alts of { (scrut_usage, scrut') ->
760 case mapAndUnzip (occAnalAlt alt_env bndr) alts of { (alts_usage_s, alts') ->
762 alts_usage = foldr1 combineAltsUsageDetails alts_usage_s
763 alts_usage' = addCaseBndrUsage alts_usage
764 (alts_usage1, tagged_bndr) = tagBinder alts_usage' bndr
765 total_usage = scrut_usage +++ alts_usage1
767 total_usage `seq` (total_usage, Case scrut' tagged_bndr ty alts') }}
769 -- The case binder gets a usage of either "many" or "dead", never "one".
770 -- Reason: we like to inline single occurrences, to eliminate a binding,
771 -- but inlining a case binder *doesn't* eliminate a binding.
772 -- We *don't* want to transform
773 -- case x of w { (p,q) -> f w }
775 -- case x of w { (p,q) -> f (p,q) }
776 addCaseBndrUsage usage = case lookupVarEnv usage bndr of
778 Just occ -> extendVarEnv usage bndr (markMany occ)
780 alt_env = setVanillaCtxt env
781 -- Consider x = case v of { True -> (p,q); ... }
782 -- Then it's fine to inline p and q
784 occ_anal_scrut (Var v) (alt1 : other_alts)
785 | not (null other_alts) || not (isDefaultAlt alt1)
786 = (mkOneOcc env v True, Var v)
787 occ_anal_scrut scrut _alts = occAnal vanillaCtxt scrut
788 -- No need for rhsCtxt
790 occAnal env (Let bind body)
791 = case occAnal env body of { (body_usage, body') ->
792 case occAnalBind env bind body_usage of { (final_usage, new_binds) ->
793 (final_usage, mkLets new_binds body') }}
795 occAnalArgs :: OccEnv -> [CoreExpr] -> (UsageDetails, [CoreExpr])
796 occAnalArgs _env args
797 = case mapAndUnzip (occAnal arg_env) args of { (arg_uds_s, args') ->
798 (foldr (+++) emptyDetails arg_uds_s, args')}
800 arg_env = vanillaCtxt
803 Applications are dealt with specially because we want
804 the "build hack" to work.
808 -> (Expr CoreBndr, [Arg CoreBndr])
809 -> (UsageDetails, Expr CoreBndr)
810 occAnalApp env (Var fun, args)
811 = case args_stuff of { (args_uds, args') ->
813 final_args_uds = markRhsUds env is_pap args_uds
815 (fun_uds +++ final_args_uds, mkApps (Var fun) args') }
817 fun_uniq = idUnique fun
818 fun_uds = mkOneOcc env fun (valArgCount args > 0)
819 is_pap = isDataConWorkId fun || valArgCount args < idArity fun
821 -- Hack for build, fold, runST
822 args_stuff | fun_uniq == buildIdKey = appSpecial env 2 [True,True] args
823 | fun_uniq == augmentIdKey = appSpecial env 2 [True,True] args
824 | fun_uniq == foldrIdKey = appSpecial env 3 [False,True] args
825 | fun_uniq == runSTRepIdKey = appSpecial env 2 [True] args
826 -- (foldr k z xs) may call k many times, but it never
827 -- shares a partial application of k; hence [False,True]
828 -- This means we can optimise
829 -- foldr (\x -> let v = ...x... in \y -> ...v...) z xs
830 -- by floating in the v
832 | otherwise = occAnalArgs env args
835 occAnalApp env (fun, args)
836 = case occAnal (addAppCtxt env args) fun of { (fun_uds, fun') ->
837 -- The addAppCtxt is a bit cunning. One iteration of the simplifier
838 -- often leaves behind beta redexs like
840 -- Here we would like to mark x,y as one-shot, and treat the whole
841 -- thing much like a let. We do this by pushing some True items
842 -- onto the context stack.
844 case occAnalArgs env args of { (args_uds, args') ->
846 final_uds = fun_uds +++ args_uds
848 (final_uds, mkApps fun' args') }}
851 markRhsUds :: OccEnv -- Check if this is a RhsEnv
852 -> Bool -- and this is true
853 -> UsageDetails -- The do markMany on this
855 -- We mark the free vars of the argument of a constructor or PAP
856 -- as "many", if it is the RHS of a let(rec).
857 -- This means that nothing gets inlined into a constructor argument
858 -- position, which is what we want. Typically those constructor
859 -- arguments are just variables, or trivial expressions.
861 -- This is the *whole point* of the isRhsEnv predicate
862 markRhsUds env is_pap arg_uds
863 | isRhsEnv env && is_pap = mapVarEnv markMany arg_uds
864 | otherwise = arg_uds
868 -> Int -> CtxtTy -- Argument number, and context to use for it
870 -> (UsageDetails, [CoreExpr])
871 appSpecial env n ctxt args
874 arg_env = vanillaCtxt
876 go _ [] = (emptyDetails, []) -- Too few args
878 go 1 (arg:args) -- The magic arg
879 = case occAnal (setCtxt arg_env ctxt) arg of { (arg_uds, arg') ->
880 case occAnalArgs env args of { (args_uds, args') ->
881 (arg_uds +++ args_uds, arg':args') }}
884 = case occAnal arg_env arg of { (arg_uds, arg') ->
885 case go (n-1) args of { (args_uds, args') ->
886 (arg_uds +++ args_uds, arg':args') }}
892 If the case binder occurs at all, the other binders effectively do too.
894 case e of x { (a,b) -> rhs }
897 If e turns out to be (e1,e2) we indeed get something like
898 let a = e1; b = e2; x = (a,b) in rhs
900 Note [Aug 06]: I don't think this is necessary any more, and it helpe
901 to know when binders are unused. See esp the call to
902 isDeadBinder in Simplify.mkDupableAlt
908 -> (UsageDetails, Alt IdWithOccInfo)
909 occAnalAlt env _case_bndr (con, bndrs, rhs)
910 = case occAnal env rhs of { (rhs_usage, rhs') ->
912 (final_usage, tagged_bndrs) = tagBinders rhs_usage bndrs
913 final_bndrs = tagged_bndrs -- See Note [Aug06] above
915 final_bndrs | case_bndr `elemVarEnv` final_usage = bndrs
916 | otherwise = tagged_bndrs
917 -- Leave the binders untagged if the case
918 -- binder occurs at all; see note above
921 (final_usage, (con, final_bndrs, rhs')) }
925 %************************************************************************
927 \subsection[OccurAnal-types]{OccEnv}
929 %************************************************************************
933 = OccEnv OccEncl -- Enclosing context information
934 CtxtTy -- Tells about linearity
936 -- OccEncl is used to control whether to inline into constructor arguments
938 -- x = (p,q) -- Don't inline p or q
939 -- y = /\a -> (p a, q a) -- Still don't inline p or q
940 -- z = f (p,q) -- Do inline p,q; it may make a rule fire
941 -- So OccEncl tells enought about the context to know what to do when
942 -- we encounter a contructor application or PAP.
945 = OccRhs -- RHS of let(rec), albeit perhaps inside a type lambda
946 -- Don't inline into constructor args here
947 | OccVanilla -- Argument of function, body of lambda, scruintee of case etc.
948 -- Do inline into constructor args here
953 -- True:ctxt Analysing a function-valued expression that will be
956 -- False:ctxt Analysing a function-valued expression that may
957 -- be applied many times; but when it is,
958 -- the CtxtTy inside applies
961 initOccEnv = OccEnv OccRhs []
963 vanillaCtxt :: OccEnv
964 vanillaCtxt = OccEnv OccVanilla []
967 rhsCtxt = OccEnv OccRhs []
969 isRhsEnv :: OccEnv -> Bool
970 isRhsEnv (OccEnv OccRhs _) = True
971 isRhsEnv (OccEnv OccVanilla _) = False
973 setVanillaCtxt :: OccEnv -> OccEnv
974 setVanillaCtxt (OccEnv OccRhs ctxt_ty) = OccEnv OccVanilla ctxt_ty
975 setVanillaCtxt other_env = other_env
977 setCtxt :: OccEnv -> CtxtTy -> OccEnv
978 setCtxt (OccEnv encl _) ctxt = OccEnv encl ctxt
980 oneShotGroup :: OccEnv -> [CoreBndr] -> [CoreBndr]
981 -- The result binders have one-shot-ness set that they might not have had originally.
982 -- This happens in (build (\cn -> e)). Here the occurrence analyser
983 -- linearity context knows that c,n are one-shot, and it records that fact in
984 -- the binder. This is useful to guide subsequent float-in/float-out tranformations
986 oneShotGroup (OccEnv _encl ctxt) bndrs
989 go _ [] rev_bndrs = reverse rev_bndrs
991 go (lin_ctxt:ctxt) (bndr:bndrs) rev_bndrs
992 | isId bndr = go ctxt bndrs (bndr':rev_bndrs)
994 bndr' | lin_ctxt = setOneShotLambda bndr
997 go ctxt (bndr:bndrs) rev_bndrs = go ctxt bndrs (bndr:rev_bndrs)
999 addAppCtxt :: OccEnv -> [Arg CoreBndr] -> OccEnv
1000 addAppCtxt (OccEnv encl ctxt) args
1001 = OccEnv encl (replicate (valArgCount args) True ++ ctxt)
1004 %************************************************************************
1006 \subsection[OccurAnal-types]{OccEnv}
1008 %************************************************************************
1011 type UsageDetails = IdEnv OccInfo -- A finite map from ids to their usage
1013 (+++), combineAltsUsageDetails
1014 :: UsageDetails -> UsageDetails -> UsageDetails
1017 = plusVarEnv_C addOccInfo usage1 usage2
1019 combineAltsUsageDetails usage1 usage2
1020 = plusVarEnv_C orOccInfo usage1 usage2
1022 addOneOcc :: UsageDetails -> Id -> OccInfo -> UsageDetails
1023 addOneOcc usage id info
1024 = plusVarEnv_C addOccInfo usage (unitVarEnv id info)
1025 -- ToDo: make this more efficient
1027 emptyDetails :: UsageDetails
1028 emptyDetails = (emptyVarEnv :: UsageDetails)
1030 usedIn :: Id -> UsageDetails -> Bool
1031 v `usedIn` details = isExportedId v || v `elemVarEnv` details
1033 type IdWithOccInfo = Id
1035 tagBinders :: UsageDetails -- Of scope
1037 -> (UsageDetails, -- Details with binders removed
1038 [IdWithOccInfo]) -- Tagged binders
1040 tagBinders usage binders
1042 usage' = usage `delVarEnvList` binders
1043 uss = map (setBinderOcc usage) binders
1045 usage' `seq` (usage', uss)
1047 tagBinder :: UsageDetails -- Of scope
1049 -> (UsageDetails, -- Details with binders removed
1050 IdWithOccInfo) -- Tagged binders
1052 tagBinder usage binder
1054 usage' = usage `delVarEnv` binder
1055 binder' = setBinderOcc usage binder
1057 usage' `seq` (usage', binder')
1059 setBinderOcc :: UsageDetails -> CoreBndr -> CoreBndr
1060 setBinderOcc usage bndr
1061 | isTyVar bndr = bndr
1062 | isExportedId bndr = case idOccInfo bndr of
1064 _ -> setIdOccInfo bndr NoOccInfo
1065 -- Don't use local usage info for visible-elsewhere things
1066 -- BUT *do* erase any IAmALoopBreaker annotation, because we're
1067 -- about to re-generate it and it shouldn't be "sticky"
1069 | otherwise = setIdOccInfo bndr occ_info
1071 occ_info = lookupVarEnv usage bndr `orElse` IAmDead
1075 %************************************************************************
1077 \subsection{Operations over OccInfo}
1079 %************************************************************************
1082 mkOneOcc :: OccEnv -> Id -> InterestingCxt -> UsageDetails
1083 mkOneOcc _env id int_cxt
1084 | isLocalId id = unitVarEnv id (OneOcc False True int_cxt)
1085 | otherwise = emptyDetails
1087 markMany, markInsideLam, markInsideSCC :: OccInfo -> OccInfo
1089 markMany IAmDead = IAmDead
1090 markMany _ = NoOccInfo
1092 markInsideSCC occ = markMany occ
1094 markInsideLam (OneOcc _ one_br int_cxt) = OneOcc True one_br int_cxt
1095 markInsideLam occ = occ
1097 addOccInfo, orOccInfo :: OccInfo -> OccInfo -> OccInfo
1099 addOccInfo IAmDead info2 = info2
1100 addOccInfo info1 IAmDead = info1
1101 addOccInfo _ _ = NoOccInfo
1103 -- (orOccInfo orig new) is used
1104 -- when combining occurrence info from branches of a case
1106 orOccInfo IAmDead info2 = info2
1107 orOccInfo info1 IAmDead = info1
1108 orOccInfo (OneOcc in_lam1 _ int_cxt1)
1109 (OneOcc in_lam2 _ int_cxt2)
1110 = OneOcc (in_lam1 || in_lam2)
1111 False -- False, because it occurs in both branches
1112 (int_cxt1 && int_cxt2)
1113 orOccInfo _ _ = NoOccInfo