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 Type ( tyVarsOfType )
23 import CoreUtils ( exprIsTrivial, isDefaultAlt, mkCoerceI, isExpandableApp )
24 import Coercion ( CoercionI(..), mkSymCoI )
26 import Name ( localiseName )
32 import Maybes ( orElse )
33 import Digraph ( SCC(..), stronglyConnCompFromEdgedVerticesR )
34 import PrelNames ( buildIdKey, foldrIdKey, runSTRepIdKey, augmentIdKey )
35 import Unique ( Unique )
36 import UniqFM ( keysUFM, intersectUFM_C, foldUFM_Directly )
37 import Util ( mapAndUnzip, filterOut )
45 %************************************************************************
47 \subsection[OccurAnal-main]{Counting occurrences: main function}
49 %************************************************************************
51 Here's the externally-callable interface:
54 occurAnalysePgm :: [CoreBind] -> [CoreRule] -> [CoreBind]
55 occurAnalysePgm binds rules
56 = snd (go initOccEnv binds)
58 initial_details = addIdOccs emptyDetails (rulesFreeVars rules)
59 -- The RULES keep things alive!
61 go :: OccEnv -> [CoreBind] -> (UsageDetails, [CoreBind])
63 = (initial_details, [])
65 = (final_usage, bind' ++ binds')
67 (bs_usage, binds') = go env binds
68 (final_usage, bind') = occAnalBind env env bind bs_usage
70 occurAnalyseExpr :: CoreExpr -> CoreExpr
71 -- Do occurrence analysis, and discard occurence info returned
72 occurAnalyseExpr expr = snd (occAnal initOccEnv expr)
76 %************************************************************************
78 \subsection[OccurAnal-main]{Counting occurrences: main function}
80 %************************************************************************
86 occAnalBind :: OccEnv -- The incoming OccEnv
87 -> OccEnv -- Same, but trimmed by (binderOf bind)
89 -> UsageDetails -- Usage details of scope
90 -> (UsageDetails, -- Of the whole let(rec)
93 occAnalBind env _ (NonRec binder rhs) body_usage
94 | isTyVar binder -- A type let; we don't gather usage info
95 = (body_usage, [NonRec binder rhs])
97 | not (binder `usedIn` body_usage) -- It's not mentioned
100 | otherwise -- It's mentioned in the body
101 = (body_usage' +++ addRuleUsage rhs_usage binder, -- Note [Rules are extra RHSs]
102 [NonRec tagged_binder rhs'])
104 (body_usage', tagged_binder) = tagBinder body_usage binder
105 (rhs_usage, rhs') = occAnalRhs env tagged_binder rhs
110 Dropping dead code for recursive bindings is done in a very simple way:
112 the entire set of bindings is dropped if none of its binders are
113 mentioned in its body; otherwise none are.
115 This seems to miss an obvious improvement.
127 Now 'f' is unused! But it's OK! Dependency analysis will sort this
128 out into a letrec for 'g' and a 'let' for 'f', and then 'f' will get
129 dropped. It isn't easy to do a perfect job in one blow. Consider
140 Note [Loop breaking and RULES]
141 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
142 Loop breaking is surprisingly subtle. First read the section 4 of
143 "Secrets of the GHC inliner". This describes our basic plan.
145 However things are made quite a bit more complicated by RULES. Remember
147 * Note [Rules are extra RHSs]
148 ~~~~~~~~~~~~~~~~~~~~~~~~~~~
149 A RULE for 'f' is like an extra RHS for 'f'. That way the "parent"
150 keeps the specialised "children" alive. If the parent dies
151 (because it isn't referenced any more), then the children will die
152 too (unless they are already referenced directly).
154 To that end, we build a Rec group for each cyclic strongly
156 *treating f's rules as extra RHSs for 'f'*.
158 When we make the Rec groups we include variables free in *either*
159 LHS *or* RHS of the rule. The former might seems silly, but see
160 Note [Rule dependency info].
162 So in Example [eftInt], eftInt and eftIntFB will be put in the
163 same Rec, even though their 'main' RHSs are both non-recursive.
165 * Note [Rules are visible in their own rec group]
166 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
167 We want the rules for 'f' to be visible in f's right-hand side.
168 And we'd like them to be visible in other functions in f's Rec
169 group. E.g. in Example [Specialisation rules] we want f' rule
170 to be visible in both f's RHS, and fs's RHS.
172 This means that we must simplify the RULEs first, before looking
173 at any of the definitions. This is done by Simplify.simplRecBind,
174 when it calls addLetIdInfo.
176 * Note [Choosing loop breakers]
177 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
178 We avoid infinite inlinings by choosing loop breakers, and
179 ensuring that a loop breaker cuts each loop. But what is a
180 "loop"? In particular, a RULE is like an equation for 'f' that
181 is *always* inlined if it is applicable. We do *not* disable
182 rules for loop-breakers. It's up to whoever makes the rules to
183 make sure that the rules themselves always terminate. See Note
184 [Rules for recursive functions] in Simplify.lhs
187 f's RHS mentions g, and
188 g has a RULE that mentions h, and
189 h has a RULE that mentions f
191 then we *must* choose f to be a loop breaker. In general, take the
192 free variables of f's RHS, and augment it with all the variables
193 reachable by RULES from those starting points. That is the whole
194 reason for computing rule_fv_env in occAnalBind. (Of course we
195 only consider free vars that are also binders in this Rec group.)
197 Note that when we compute this rule_fv_env, we only consider variables
198 free in the *RHS* of the rule, in contrast to the way we build the
199 Rec group in the first place (Note [Rule dependency info])
201 Note that in Example [eftInt], *neither* eftInt *nor* eftIntFB is
202 chosen as a loop breaker, because their RHSs don't mention each other.
203 And indeed both can be inlined safely.
205 Note that the edges of the graph we use for computing loop breakers
206 are not the same as the edges we use for computing the Rec blocks.
207 That's why we compute
208 rec_edges for the Rec block analysis
209 loop_breaker_edges for the loop breaker analysis
212 * Note [Weak loop breakers]
213 ~~~~~~~~~~~~~~~~~~~~~~~~~
214 There is a last nasty wrinkle. Suppose we have
224 Remmber that we simplify the RULES before any RHS (see Note
225 [Rules are visible in their own rec group] above).
227 So we must *not* postInlineUnconditionally 'g', even though
228 its RHS turns out to be trivial. (I'm assuming that 'g' is
229 not choosen as a loop breaker.) Why not? Because then we
230 drop the binding for 'g', which leaves it out of scope in the
233 We "solve" this by making g a "weak" or "rules-only" loop breaker,
234 with OccInfo = IAmLoopBreaker True. A normal "strong" loop breaker
235 has IAmLoopBreaker False. So
237 Inline postInlineUnconditionally
238 IAmLoopBreaker False no no
239 IAmLoopBreaker True yes no
242 The **sole** reason for this kind of loop breaker is so that
243 postInlineUnconditionally does not fire. Ugh.
245 * Note [Rule dependency info]
246 ~~~~~~~~~~~~~~~~~~~~~~~~~~~
247 The VarSet in a SpecInfo is used for dependency analysis in the
248 occurrence analyser. We must track free vars in *both* lhs and rhs.
249 Hence use of idRuleVars, rather than idRuleRhsVars in addRuleUsage.
253 Then if we substitute y for x, we'd better do so in the
254 rule's LHS too, so we'd better ensure the dependency is respected
257 * Note [Inline rules]
259 None of the above stuff about RULES applies to Inline Rules,
260 stored in a CoreUnfolding. The unfolding, if any, is simplified
261 at the same time as the regular RHS of the function, so it should
262 be treated *exactly* like an extra RHS.
267 Example (from GHC.Enum):
269 eftInt :: Int# -> Int# -> [Int]
270 eftInt x y = ...(non-recursive)...
272 {-# INLINE [0] eftIntFB #-}
273 eftIntFB :: (Int -> r -> r) -> r -> Int# -> Int# -> r
274 eftIntFB c n x y = ...(non-recursive)...
277 "eftInt" [~1] forall x y. eftInt x y = build (\ c n -> eftIntFB c n x y)
278 "eftIntList" [1] eftIntFB (:) [] = eftInt
281 Example [Specialisation rules]
282 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
283 Consider this group, which is typical of what SpecConstr builds:
285 fs a = ....f (C a)....
286 f x = ....f (C a)....
287 {-# RULE f (C a) = fs a #-}
289 So 'f' and 'fs' are in the same Rec group (since f refers to fs via its RULE).
291 But watch out! If 'fs' is not chosen as a loop breaker, we may get an infinite loop:
292 - the RULE is applied in f's RHS (see Note [Self-recursive rules] in Simplify
293 - fs is inlined (say it's small)
294 - now there's another opportunity to apply the RULE
296 This showed up when compiling Control.Concurrent.Chan.getChanContents.
300 occAnalBind _ env (Rec pairs) body_usage
301 = foldr occAnalRec (body_usage, []) sccs
302 -- For a recursive group, we
303 -- * occ-analyse all the RHSs
304 -- * compute strongly-connected components
305 -- * feed those components to occAnalRec
307 -------------Dependency analysis ------------------------------
308 bndr_set = mkVarSet (map fst pairs)
310 sccs :: [SCC (Node Details)]
311 sccs = {-# SCC "occAnalBind.scc" #-} stronglyConnCompFromEdgedVerticesR rec_edges
313 rec_edges :: [Node Details]
314 rec_edges = {-# SCC "occAnalBind.assoc" #-} map make_node pairs
316 make_node (bndr, rhs)
317 = (ND bndr rhs' all_rhs_usage rhs_fvs, idUnique bndr, out_edges)
319 (rhs_usage, rhs') = occAnalRhs env bndr rhs
320 all_rhs_usage = addRuleUsage rhs_usage bndr -- Note [Rules are extra RHSs]
321 rhs_fvs = intersectUFM_C (\b _ -> b) bndr_set rhs_usage
322 out_edges = keysUFM (rhs_fvs `unionVarSet` idRuleVars bndr)
323 -- (a -> b) means a mentions b
324 -- Given the usage details (a UFM that gives occ info for each free var of
325 -- the RHS) we can get the list of free vars -- or rather their Int keys --
326 -- by just extracting the keys from the finite map. Grimy, but fast.
327 -- Previously we had this:
328 -- [ bndr | bndr <- bndrs,
329 -- maybeToBool (lookupVarEnv rhs_usage bndr)]
330 -- which has n**2 cost, and this meant that edges_from alone
331 -- consumed 10% of total runtime!
333 -----------------------------
334 occAnalRec :: SCC (Node Details) -> (UsageDetails, [CoreBind])
335 -> (UsageDetails, [CoreBind])
337 -- The NonRec case is just like a Let (NonRec ...) above
338 occAnalRec (AcyclicSCC (ND bndr rhs rhs_usage _, _, _)) (body_usage, binds)
339 | not (bndr `usedIn` body_usage)
340 = (body_usage, binds)
342 | otherwise -- It's mentioned in the body
343 = (body_usage' +++ rhs_usage,
344 NonRec tagged_bndr rhs : binds)
346 (body_usage', tagged_bndr) = tagBinder body_usage bndr
349 -- The Rec case is the interesting one
350 -- See Note [Loop breaking]
351 occAnalRec (CyclicSCC nodes) (body_usage, binds)
352 | not (any (`usedIn` body_usage) bndrs) -- NB: look at body_usage, not total_usage
353 = (body_usage, binds) -- Dead code
355 | otherwise -- At this point we always build a single Rec
356 = (final_usage, Rec pairs : binds)
359 bndrs = [b | (ND b _ _ _, _, _) <- nodes]
360 bndr_set = mkVarSet bndrs
362 ----------------------------
363 -- Tag the binders with their occurrence info
364 total_usage = foldl add_usage body_usage nodes
365 add_usage usage_so_far (ND _ _ rhs_usage _, _, _) = usage_so_far +++ rhs_usage
366 (final_usage, tagged_nodes) = mapAccumL tag_node total_usage nodes
368 tag_node :: UsageDetails -> Node Details -> (UsageDetails, Node Details)
369 -- (a) Tag the binders in the details with occ info
370 -- (b) Mark the binder with "weak loop-breaker" OccInfo
371 -- saying "no preInlineUnconditionally" if it is used
372 -- in any rule (lhs or rhs) of the recursive group
373 -- See Note [Weak loop breakers]
374 tag_node usage (ND bndr rhs rhs_usage rhs_fvs, k, ks)
375 = (usage `delVarEnv` bndr, (ND bndr2 rhs rhs_usage rhs_fvs, k, ks))
377 bndr2 | bndr `elemVarSet` all_rule_fvs = makeLoopBreaker True bndr1
379 bndr1 = setBinderOcc usage bndr
380 all_rule_fvs = bndr_set `intersectVarSet` foldr (unionVarSet . idRuleVars)
383 ----------------------------
384 -- Now reconstruct the cycle
385 pairs | no_rules = reOrderCycle 0 tagged_nodes []
386 | otherwise = foldr (reOrderRec 0) [] $
387 stronglyConnCompFromEdgedVerticesR loop_breaker_edges
389 -- See Note [Choosing loop breakers] for loop_breaker_edges
390 loop_breaker_edges = map mk_node tagged_nodes
391 mk_node (details@(ND _ _ _ rhs_fvs), k, _) = (details, k, new_ks)
393 new_ks = keysUFM (extendFvs rule_fv_env rhs_fvs rhs_fvs)
395 ------------------------------------
396 rule_fv_env :: IdEnv IdSet -- Variables from this group mentioned in RHS of rules
397 -- Domain is *subset* of bound vars (others have no rule fvs)
398 rule_fv_env = rule_loop init_rule_fvs
400 no_rules = null init_rule_fvs
401 init_rule_fvs = [(b, rule_fvs)
403 , let rule_fvs = idRuleRhsVars b `intersectVarSet` bndr_set
404 , not (isEmptyVarSet rule_fvs)]
406 rule_loop :: [(Id,IdSet)] -> IdEnv IdSet -- Finds fixpoint
409 | otherwise = rule_loop new_fv_list
411 env = mkVarEnv init_rule_fvs
412 (no_change, new_fv_list) = mapAccumL bump True fv_list
413 bump no_change (b,fvs)
414 | new_fvs `subVarSet` fvs = (no_change, (b,fvs))
415 | otherwise = (False, (b,new_fvs `unionVarSet` fvs))
417 new_fvs = extendFvs env emptyVarSet fvs
419 extendFvs :: IdEnv IdSet -> IdSet -> IdSet -> IdSet
420 -- (extendFVs env fvs s) returns (fvs `union` env(s))
421 extendFvs env fvs id_set
422 = foldUFM_Directly add fvs id_set
425 = case lookupVarEnv_Directly env uniq of
426 Just fvs' -> fvs' `unionVarSet` fvs
430 @reOrderRec@ is applied to the list of (binder,rhs) pairs for a cyclic
431 strongly connected component (there's guaranteed to be a cycle). It returns the
433 a) in a better order,
434 b) with some of the Ids having a IAmALoopBreaker pragma
436 The "loop-breaker" Ids are sufficient to break all cycles in the SCC. This means
437 that the simplifier can guarantee not to loop provided it never records an inlining
438 for these no-inline guys.
440 Furthermore, the order of the binds is such that if we neglect dependencies
441 on the no-inline Ids then the binds are topologically sorted. This means
442 that the simplifier will generally do a good job if it works from top bottom,
443 recording inlinings for any Ids which aren't marked as "no-inline" as it goes.
446 [June 98: I don't understand the following paragraphs, and I've
447 changed the a=b case again so that it isn't a special case any more.]
449 Here's a case that bit me:
457 Re-ordering doesn't change the order of bindings, but there was no loop-breaker.
459 My solution was to make a=b bindings record b as Many, rather like INLINE bindings.
460 Perhaps something cleverer would suffice.
465 type Node details = (details, Unique, [Unique]) -- The Ints are gotten from the Unique,
466 -- which is gotten from the Id.
467 data Details = ND Id -- Binder
470 UsageDetails -- Full usage from RHS,
471 -- including *both* RULES *and* InlineRule unfolding
473 IdSet -- Other binders *from this Rec group* mentioned in
475 -- * any InlineRule unfolding
476 -- but *excluding* any RULES
478 reOrderRec :: Int -> SCC (Node Details)
479 -> [(Id,CoreExpr)] -> [(Id,CoreExpr)]
480 -- Sorted into a plausible order. Enough of the Ids have
481 -- IAmALoopBreaker pragmas that there are no loops left.
482 reOrderRec _ (AcyclicSCC (ND bndr rhs _ _, _, _)) pairs = (bndr, rhs) : pairs
483 reOrderRec depth (CyclicSCC cycle) pairs = reOrderCycle depth cycle pairs
485 reOrderCycle :: Int -> [Node Details] -> [(Id,CoreExpr)] -> [(Id,CoreExpr)]
487 = panic "reOrderCycle"
488 reOrderCycle _ [bind] pairs -- Common case of simple self-recursion
489 = (makeLoopBreaker False bndr, rhs) : pairs
491 (ND bndr rhs _ _, _, _) = bind
493 reOrderCycle depth (bind : binds) pairs
494 = -- Choose a loop breaker, mark it no-inline,
495 -- do SCC analysis on the rest, and recursively sort them out
496 -- pprTrace "reOrderCycle" (ppr [b | (ND b _ _ _, _, _) <- bind:binds]) $
497 foldr (reOrderRec new_depth)
498 ([ (makeLoopBreaker False bndr, rhs)
499 | (ND bndr rhs _ _, _, _) <- chosen_binds] ++ pairs)
500 (stronglyConnCompFromEdgedVerticesR unchosen)
502 (chosen_binds, unchosen) = choose_loop_breaker [bind] (score bind) [] binds
504 approximate_loop_breaker = depth >= 2
505 new_depth | approximate_loop_breaker = 0
506 | otherwise = depth+1
507 -- After two iterations (d=0, d=1) give up
508 -- and approximate, returning to d=0
510 -- This loop looks for the bind with the lowest score
511 -- to pick as the loop breaker. The rest accumulate in
512 choose_loop_breaker loop_binds _loop_sc acc []
513 = (loop_binds, acc) -- Done
515 -- If approximate_loop_breaker is True, we pick *all*
516 -- nodes with lowest score, else just one
517 -- See Note [Complexity of loop breaking]
518 choose_loop_breaker loop_binds loop_sc acc (bind : binds)
519 | sc < loop_sc -- Lower score so pick this new one
520 = choose_loop_breaker [bind] sc (loop_binds ++ acc) binds
522 | approximate_loop_breaker && sc == loop_sc
523 = choose_loop_breaker (bind : loop_binds) loop_sc acc binds
525 | otherwise -- Higher score so don't pick it
526 = choose_loop_breaker loop_binds loop_sc (bind : acc) binds
530 score :: Node Details -> Int -- Higher score => less likely to be picked as loop breaker
531 score (ND bndr rhs _ _, _, _)
532 | isDFunId bndr = 9 -- Never choose a DFun as a loop breaker
533 -- Note [DFuns should not be loop breakers]
535 | Just (inl_source, _) <- isInlineRule_maybe (idUnfolding bndr)
537 InlineWrapper {} -> 10 -- Note [INLINE pragmas]
538 _other -> 3 -- Data structures are more important than this
539 -- so that dictionary/method recursion unravels
540 -- Note that this case hits all InlineRule things, so we
541 -- never look at 'rhs for InlineRule stuff. That's right, because
542 -- 'rhs' is irrelevant for inlining things with an InlineRule
544 | is_con_app rhs = 5 -- Data types help with cases: Note [Constructor applications]
546 | exprIsTrivial rhs = 10 -- Practically certain to be inlined
547 -- Used to have also: && not (isExportedId bndr)
548 -- But I found this sometimes cost an extra iteration when we have
549 -- rec { d = (a,b); a = ...df...; b = ...df...; df = d }
550 -- where df is the exported dictionary. Then df makes a really
551 -- bad choice for loop breaker
554 -- If an Id is marked "never inline" then it makes a great loop breaker
555 -- The only reason for not checking that here is that it is rare
556 -- and I've never seen a situation where it makes a difference,
557 -- so it probably isn't worth the time to test on every binder
558 -- | isNeverActive (idInlinePragma bndr) = -10
560 | isOneOcc (idOccInfo bndr) = 2 -- Likely to be inlined
562 | canUnfold (realIdUnfolding bndr) = 1
563 -- The Id has some kind of unfolding
564 -- Ignore loop-breaker-ness here because that is what we are setting!
568 -- Checking for a constructor application
569 -- Cheap and cheerful; the simplifer moves casts out of the way
570 -- The lambda case is important to spot x = /\a. C (f a)
571 -- which comes up when C is a dictionary constructor and
572 -- f is a default method.
573 -- Example: the instance for Show (ST s a) in GHC.ST
575 -- However we *also* treat (\x. C p q) as a con-app-like thing,
576 -- Note [Closure conversion]
577 is_con_app (Var v) = isConLikeId v
578 is_con_app (App f _) = is_con_app f
579 is_con_app (Lam _ e) = is_con_app e
580 is_con_app (Note _ e) = is_con_app e
583 makeLoopBreaker :: Bool -> Id -> Id
584 -- Set the loop-breaker flag: see Note [Weak loop breakers]
585 makeLoopBreaker weak bndr = setIdOccInfo bndr (IAmALoopBreaker weak)
588 Note [Complexity of loop breaking]
589 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
590 The loop-breaking algorithm knocks out one binder at a time, and
591 performs a new SCC analysis on the remaining binders. That can
592 behave very badly in tightly-coupled groups of bindings; in the
593 worst case it can be (N**2)*log N, because it does a full SCC
594 on N, then N-1, then N-2 and so on.
596 To avoid this, we switch plans after 2 (or whatever) attempts:
597 Plan A: pick one binder with the lowest score, make it
598 a loop breaker, and try again
599 Plan B: pick *all* binders with the lowest score, make them
600 all loop breakers, and try again
601 Since there are only a small finite number of scores, this will
602 terminate in a constant number of iterations, rather than O(N)
605 You might thing that it's very unlikely, but RULES make it much
606 more likely. Here's a real example from Trac #1969:
607 Rec { $dm = \d.\x. op d
608 {-# RULES forall d. $dm Int d = $s$dm1
609 forall d. $dm Bool d = $s$dm2 #-}
611 dInt = MkD .... opInt ...
612 dInt = MkD .... opBool ...
617 $s$dm2 = \x. op dBool }
618 The RULES stuff means that we can't choose $dm as a loop breaker
619 (Note [Choosing loop breakers]), so we must choose at least (say)
620 opInt *and* opBool, and so on. The number of loop breakders is
621 linear in the number of instance declarations.
623 Note [INLINE pragmas]
624 ~~~~~~~~~~~~~~~~~~~~~
625 Avoid choosing a function with an INLINE pramga as the loop breaker!
626 If such a function is mutually-recursive with a non-INLINE thing,
627 then the latter should be the loop-breaker.
629 Usually this is just a question of optimisation. But a particularly
630 bad case is wrappers generated by the demand analyser: if you make
631 then into a loop breaker you may get an infinite inlining loop. For
634 $wfoo x = ....foo x....
636 {-loop brk-} foo x = ...$wfoo x...
638 The interface file sees the unfolding for $wfoo, and sees that foo is
639 strict (and hence it gets an auto-generated wrapper). Result: an
640 infinite inlining in the importing scope. So be a bit careful if you
641 change this. A good example is Tree.repTree in
642 nofib/spectral/minimax. If the repTree wrapper is chosen as the loop
643 breaker then compiling Game.hs goes into an infinite loop. This
644 happened when we gave is_con_app a lower score than inline candidates:
647 = __inline_me (/\a. \w w1 w2 ->
648 case Tree.$wrepTree @ a w w1 w2 of
649 { (# ww1, ww2 #) -> Branch @ a ww1 ww2 })
652 (# w2_smP, map a (Tree a) (Tree.repTree a w1 w) (w w2) #)
654 Here we do *not* want to choose 'repTree' as the loop breaker.
656 Note [DFuns should not be loop breakers]
657 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
658 It's particularly bad to make a DFun into a loop breaker. See
659 Note [How instance declarations are translated] in TcInstDcls
661 We give DFuns a higher score than ordinary CONLIKE things because
662 if there's a choice we want the DFun to be the non-looop breker. Eg
664 rec { sc = /\ a \$dC. $fBWrap (T a) ($fCT @ a $dC)
666 $fCT :: forall a_afE. (Roman.C a_afE) => Roman.C (Roman.T a_afE)
668 $fCT = /\a \$dC. MkD (T a) ((sc @ a $dC) |> blah) ($ctoF @ a $dC)
671 Here 'sc' (the superclass) looks CONLIKE, but we'll never get to it
672 if we can't unravel the DFun first.
674 Note [Constructor applications]
675 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
676 It's really really important to inline dictionaries. Real
677 example (the Enum Ordering instance from GHC.Base):
679 rec f = \ x -> case d of (p,q,r) -> p x
680 g = \ x -> case d of (p,q,r) -> q x
683 Here, f and g occur just once; but we can't inline them into d.
684 On the other hand we *could* simplify those case expressions if
685 we didn't stupidly choose d as the loop breaker.
686 But we won't because constructor args are marked "Many".
687 Inlining dictionaries is really essential to unravelling
688 the loops in static numeric dictionaries, see GHC.Float.
690 Note [Closure conversion]
691 ~~~~~~~~~~~~~~~~~~~~~~~~~
692 We treat (\x. C p q) as a high-score candidate in the letrec scoring algorithm.
693 The immediate motivation came from the result of a closure-conversion transformation
694 which generated code like this:
696 data Clo a b = forall c. Clo (c -> a -> b) c
698 ($:) :: Clo a b -> a -> b
699 Clo f env $: x = f env x
701 rec { plus = Clo plus1 ()
703 ; plus1 _ n = Clo plus2 n
706 ; plus2 (Succ m) n = Succ (plus $: m $: n) }
708 If we inline 'plus' and 'plus1', everything unravels nicely. But if
709 we choose 'plus1' as the loop breaker (which is entirely possible
710 otherwise), the loop does not unravel nicely.
713 @occAnalRhs@ deals with the question of bindings where the Id is marked
714 by an INLINE pragma. For these we record that anything which occurs
715 in its RHS occurs many times. This pessimistically assumes that ths
716 inlined binder also occurs many times in its scope, but if it doesn't
717 we'll catch it next time round. At worst this costs an extra simplifier pass.
718 ToDo: try using the occurrence info for the inline'd binder.
720 [March 97] We do the same for atomic RHSs. Reason: see notes with reOrderRec.
721 [June 98, SLPJ] I've undone this change; I don't understand it. See notes with reOrderRec.
726 -> Id -> CoreExpr -- Binder and rhs
727 -- For non-recs the binder is alrady tagged
728 -- with occurrence info
729 -> (UsageDetails, CoreExpr)
730 -- Returned usage details includes any INLINE rhs
732 occAnalRhs env id rhs
733 = (addIdOccs rhs_usage (idUnfoldingVars id), rhs')
734 -- Include occurrences for the "extra RHS" from a CoreUnfolding
736 (rhs_usage, rhs') = occAnal ctxt rhs
737 ctxt | certainly_inline id = env
738 | otherwise = rhsCtxt env
739 -- Note that we generally use an rhsCtxt. This tells the occ anal n
740 -- that it's looking at an RHS, which has an effect in occAnalApp
742 -- But there's a problem. Consider
747 -- First time round, it looks as if x1 and x2 occur as an arg of a
748 -- let-bound constructor ==> give them a many-occurrence.
749 -- But then x3 is inlined (unconditionally as it happens) and
750 -- next time round, x2 will be, and the next time round x1 will be
751 -- Result: multiple simplifier iterations. Sigh.
752 -- Crude solution: use rhsCtxt for things that occur just once...
754 certainly_inline id = case idOccInfo id of
755 OneOcc in_lam one_br _ -> not in_lam && one_br
762 addRuleUsage :: UsageDetails -> Id -> UsageDetails
763 -- Add the usage from RULES in Id to the usage
764 addRuleUsage usage id = addIdOccs usage (idRuleVars id)
765 -- idRuleVars here: see Note [Rule dependency info]
767 addIdOccs :: UsageDetails -> VarSet -> UsageDetails
768 addIdOccs usage id_set = foldVarSet add usage id_set
770 add v u | isId v = addOneOcc u v NoOccInfo
772 -- Give a non-committal binder info (i.e NoOccInfo) because
773 -- a) Many copies of the specialised thing can appear
774 -- b) We don't want to substitute a BIG expression inside a RULE
775 -- even if that's the only occurrence of the thing
776 -- (Same goes for INLINE.)
784 -> (UsageDetails, -- Gives info only about the "interesting" Ids
787 occAnal _ (Type t) = (emptyDetails, Type t)
788 occAnal env (Var v) = (mkOneOcc env v False, Var v)
789 -- At one stage, I gathered the idRuleVars for v here too,
790 -- which in a way is the right thing to do.
791 -- But that went wrong right after specialisation, when
792 -- the *occurrences* of the overloaded function didn't have any
793 -- rules in them, so the *specialised* versions looked as if they
794 -- weren't used at all.
797 We regard variables that occur as constructor arguments as "dangerousToDup":
801 f x = let y = expensive x in
803 (case z of {(p,q)->q}, case z of {(p,q)->q})
806 We feel free to duplicate the WHNF (True,y), but that means
807 that y may be duplicated thereby.
809 If we aren't careful we duplicate the (expensive x) call!
810 Constructors are rather like lambdas in this way.
813 occAnal _ expr@(Lit _) = (emptyDetails, expr)
817 occAnal env (Note note@(SCC _) body)
818 = case occAnal env body of { (usage, body') ->
819 (mapVarEnv markInsideSCC usage, Note note body')
822 occAnal env (Note note body)
823 = case occAnal env body of { (usage, body') ->
824 (usage, Note note body')
827 occAnal env (Cast expr co)
828 = case occAnal env expr of { (usage, expr') ->
829 (markRhsUds env True usage, Cast expr' co)
830 -- If we see let x = y `cast` co
831 -- then mark y as 'Many' so that we don't
832 -- immediately inline y again.
837 occAnal env app@(App _ _)
838 = occAnalApp env (collectArgs app)
840 -- Ignore type variables altogether
841 -- (a) occurrences inside type lambdas only not marked as InsideLam
842 -- (b) type variables not in environment
844 occAnal env (Lam x body) | isTyVar x
845 = case occAnal env body of { (body_usage, body') ->
846 (body_usage, Lam x body')
849 -- For value lambdas we do a special hack. Consider
851 -- If we did nothing, x is used inside the \y, so would be marked
852 -- as dangerous to dup. But in the common case where the abstraction
853 -- is applied to two arguments this is over-pessimistic.
854 -- So instead, we just mark each binder with its occurrence
855 -- info in the *body* of the multiple lambda.
856 -- Then, the simplifier is careful when partially applying lambdas.
858 occAnal env expr@(Lam _ _)
859 = case occAnal env_body body of { (body_usage, body') ->
861 (final_usage, tagged_binders) = tagLamBinders body_usage binders'
862 -- Use binders' to put one-shot info on the lambdas
864 -- URGH! Sept 99: we don't seem to be able to use binders' here, because
865 -- we get linear-typed things in the resulting program that we can't handle yet.
866 -- (e.g. PrelShow) TODO
868 really_final_usage = if linear then
871 mapVarEnv markInsideLam final_usage
874 mkLams tagged_binders body') }
876 env_body = vanillaCtxt (trimOccEnv env binders)
877 -- Body is (no longer) an RhsContext
878 (binders, body) = collectBinders expr
879 binders' = oneShotGroup env binders
880 linear = all is_one_shot binders'
881 is_one_shot b = isId b && isOneShotBndr b
883 occAnal env (Case scrut bndr ty alts)
884 = case occ_anal_scrut scrut alts of { (scrut_usage, scrut') ->
885 case mapAndUnzip occ_anal_alt alts of { (alts_usage_s, alts') ->
887 alts_usage = foldr1 combineAltsUsageDetails alts_usage_s
888 (alts_usage1, tagged_bndr) = tag_case_bndr alts_usage bndr
889 total_usage = scrut_usage +++ alts_usage1
891 total_usage `seq` (total_usage, Case scrut' tagged_bndr ty alts') }}
893 -- Note [Case binder usage]
894 -- ~~~~~~~~~~~~~~~~~~~~~~~~
895 -- The case binder gets a usage of either "many" or "dead", never "one".
896 -- Reason: we like to inline single occurrences, to eliminate a binding,
897 -- but inlining a case binder *doesn't* eliminate a binding.
898 -- We *don't* want to transform
899 -- case x of w { (p,q) -> f w }
901 -- case x of w { (p,q) -> f (p,q) }
902 tag_case_bndr usage bndr
903 = case lookupVarEnv usage bndr of
904 Nothing -> (usage, setIdOccInfo bndr IAmDead)
905 Just _ -> (usage `delVarEnv` bndr, setIdOccInfo bndr NoOccInfo)
907 alt_env = mkAltEnv env scrut bndr
908 occ_anal_alt = occAnalAlt alt_env bndr
910 occ_anal_scrut (Var v) (alt1 : other_alts)
911 | not (null other_alts) || not (isDefaultAlt alt1)
912 = (mkOneOcc env v True, Var v) -- The 'True' says that the variable occurs
913 -- in an interesting context; the case has
914 -- at least one non-default alternative
915 occ_anal_scrut scrut _alts
916 = occAnal (vanillaCtxt env) scrut -- No need for rhsCtxt
918 occAnal env (Let bind body)
919 = case occAnal env_body body of { (body_usage, body') ->
920 case occAnalBind env env_body bind body_usage of { (final_usage, new_binds) ->
921 (final_usage, mkLets new_binds body') }}
923 env_body = trimOccEnv env (bindersOf bind)
925 occAnalArgs :: OccEnv -> [CoreExpr] -> (UsageDetails, [CoreExpr])
927 = case mapAndUnzip (occAnal arg_env) args of { (arg_uds_s, args') ->
928 (foldr (+++) emptyDetails arg_uds_s, args')}
930 arg_env = vanillaCtxt env
933 Applications are dealt with specially because we want
934 the "build hack" to work.
938 -> (Expr CoreBndr, [Arg CoreBndr])
939 -> (UsageDetails, Expr CoreBndr)
940 occAnalApp env (Var fun, args)
941 = case args_stuff of { (args_uds, args') ->
943 final_args_uds = markRhsUds env is_exp args_uds
945 (fun_uds +++ final_args_uds, mkApps (Var fun) args') }
947 fun_uniq = idUnique fun
948 fun_uds = mkOneOcc env fun (valArgCount args > 0)
949 is_exp = isExpandableApp fun (valArgCount args)
950 -- See Note [CONLIKE pragma] in BasicTypes
951 -- The definition of is_exp should match that in
952 -- Simplify.prepareRhs
954 -- Hack for build, fold, runST
955 args_stuff | fun_uniq == buildIdKey = appSpecial env 2 [True,True] args
956 | fun_uniq == augmentIdKey = appSpecial env 2 [True,True] args
957 | fun_uniq == foldrIdKey = appSpecial env 3 [False,True] args
958 | fun_uniq == runSTRepIdKey = appSpecial env 2 [True] args
959 -- (foldr k z xs) may call k many times, but it never
960 -- shares a partial application of k; hence [False,True]
961 -- This means we can optimise
962 -- foldr (\x -> let v = ...x... in \y -> ...v...) z xs
963 -- by floating in the v
965 | otherwise = occAnalArgs env args
968 occAnalApp env (fun, args)
969 = case occAnal (addAppCtxt env args) fun of { (fun_uds, fun') ->
970 -- The addAppCtxt is a bit cunning. One iteration of the simplifier
971 -- often leaves behind beta redexs like
973 -- Here we would like to mark x,y as one-shot, and treat the whole
974 -- thing much like a let. We do this by pushing some True items
975 -- onto the context stack.
977 case occAnalArgs env args of { (args_uds, args') ->
979 final_uds = fun_uds +++ args_uds
981 (final_uds, mkApps fun' args') }}
984 markRhsUds :: OccEnv -- Check if this is a RhsEnv
985 -> Bool -- and this is true
986 -> UsageDetails -- The do markMany on this
988 -- We mark the free vars of the argument of a constructor or PAP
989 -- as "many", if it is the RHS of a let(rec).
990 -- This means that nothing gets inlined into a constructor argument
991 -- position, which is what we want. Typically those constructor
992 -- arguments are just variables, or trivial expressions.
994 -- This is the *whole point* of the isRhsEnv predicate
995 markRhsUds env is_pap arg_uds
996 | isRhsEnv env && is_pap = mapVarEnv markMany arg_uds
997 | otherwise = arg_uds
1000 appSpecial :: OccEnv
1001 -> Int -> CtxtTy -- Argument number, and context to use for it
1003 -> (UsageDetails, [CoreExpr])
1004 appSpecial env n ctxt args
1007 arg_env = vanillaCtxt env
1009 go _ [] = (emptyDetails, []) -- Too few args
1011 go 1 (arg:args) -- The magic arg
1012 = case occAnal (setCtxtTy arg_env ctxt) arg of { (arg_uds, arg') ->
1013 case occAnalArgs env args of { (args_uds, args') ->
1014 (arg_uds +++ args_uds, arg':args') }}
1017 = case occAnal arg_env arg of { (arg_uds, arg') ->
1018 case go (n-1) args of { (args_uds, args') ->
1019 (arg_uds +++ args_uds, arg':args') }}
1023 Note [Binders in case alternatives]
1024 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
1026 case x of y { (a,b) -> f y }
1027 We treat 'a', 'b' as dead, because they don't physically occur in the
1028 case alternative. (Indeed, a variable is dead iff it doesn't occur in
1029 its scope in the output of OccAnal.) It really helps to know when
1030 binders are unused. See esp the call to isDeadBinder in
1031 Simplify.mkDupableAlt
1033 In this example, though, the Simplifier will bring 'a' and 'b' back to
1034 life, beause it binds 'y' to (a,b) (imagine got inlined and
1038 occAnalAlt :: OccEnv
1041 -> (UsageDetails, Alt IdWithOccInfo)
1042 occAnalAlt env case_bndr (con, bndrs, rhs)
1044 env' = trimOccEnv env bndrs
1046 case occAnal env' rhs of { (rhs_usage1, rhs1) ->
1048 proxies = getProxies env' case_bndr
1049 (rhs_usage2, rhs2) = foldrBag wrapProxy (rhs_usage1, rhs1) proxies
1050 (alt_usg, tagged_bndrs) = tagLamBinders rhs_usage2 bndrs
1051 bndrs' = tagged_bndrs -- See Note [Binders in case alternatives]
1053 (alt_usg, (con, bndrs', rhs2)) }
1055 wrapProxy :: ProxyBind -> (UsageDetails, CoreExpr) -> (UsageDetails, CoreExpr)
1056 wrapProxy (bndr, rhs_var, co) (body_usg, body)
1057 | not (bndr `usedIn` body_usg)
1060 = (body_usg' +++ rhs_usg, Let (NonRec tagged_bndr rhs) body)
1062 (body_usg', tagged_bndr) = tagBinder body_usg bndr
1063 rhs_usg = unitVarEnv rhs_var NoOccInfo -- We don't need exact info
1064 rhs = mkCoerceI co (Var rhs_var)
1068 %************************************************************************
1072 %************************************************************************
1076 = OccEnv { occ_encl :: !OccEncl -- Enclosing context information
1077 , occ_ctxt :: !CtxtTy -- Tells about linearity
1078 , occ_proxy :: ProxyEnv }
1081 -----------------------------
1082 -- OccEncl is used to control whether to inline into constructor arguments
1084 -- x = (p,q) -- Don't inline p or q
1085 -- y = /\a -> (p a, q a) -- Still don't inline p or q
1086 -- z = f (p,q) -- Do inline p,q; it may make a rule fire
1087 -- So OccEncl tells enought about the context to know what to do when
1088 -- we encounter a contructor application or PAP.
1091 = OccRhs -- RHS of let(rec), albeit perhaps inside a type lambda
1092 -- Don't inline into constructor args here
1093 | OccVanilla -- Argument of function, body of lambda, scruintee of case etc.
1094 -- Do inline into constructor args here
1096 type CtxtTy = [Bool]
1099 -- True:ctxt Analysing a function-valued expression that will be
1100 -- applied just once
1102 -- False:ctxt Analysing a function-valued expression that may
1103 -- be applied many times; but when it is,
1104 -- the CtxtTy inside applies
1106 initOccEnv :: OccEnv
1107 initOccEnv = OccEnv { occ_encl = OccVanilla
1109 , occ_proxy = PE emptyVarEnv emptyVarSet }
1111 vanillaCtxt :: OccEnv -> OccEnv
1112 vanillaCtxt env = OccEnv { occ_encl = OccVanilla
1114 , occ_proxy = occ_proxy env }
1116 rhsCtxt :: OccEnv -> OccEnv
1117 rhsCtxt env = OccEnv { occ_encl = OccRhs, occ_ctxt = []
1118 , occ_proxy = occ_proxy env }
1120 setCtxtTy :: OccEnv -> CtxtTy -> OccEnv
1121 setCtxtTy env ctxt = env { occ_ctxt = ctxt }
1123 isRhsEnv :: OccEnv -> Bool
1124 isRhsEnv (OccEnv { occ_encl = OccRhs }) = True
1125 isRhsEnv (OccEnv { occ_encl = OccVanilla }) = False
1127 oneShotGroup :: OccEnv -> [CoreBndr] -> [CoreBndr]
1128 -- The result binders have one-shot-ness set that they might not have had originally.
1129 -- This happens in (build (\cn -> e)). Here the occurrence analyser
1130 -- linearity context knows that c,n are one-shot, and it records that fact in
1131 -- the binder. This is useful to guide subsequent float-in/float-out tranformations
1133 oneShotGroup (OccEnv { occ_ctxt = ctxt }) bndrs
1136 go _ [] rev_bndrs = reverse rev_bndrs
1138 go (lin_ctxt:ctxt) (bndr:bndrs) rev_bndrs
1139 | isId bndr = go ctxt bndrs (bndr':rev_bndrs)
1141 bndr' | lin_ctxt = setOneShotLambda bndr
1144 go ctxt (bndr:bndrs) rev_bndrs = go ctxt bndrs (bndr:rev_bndrs)
1146 addAppCtxt :: OccEnv -> [Arg CoreBndr] -> OccEnv
1147 addAppCtxt env@(OccEnv { occ_ctxt = ctxt }) args
1148 = env { occ_ctxt = replicate (valArgCount args) True ++ ctxt }
1151 %************************************************************************
1155 %************************************************************************
1159 = PE (IdEnv (Id, [(Id,CoercionI)])) VarSet
1160 -- Main env, and its free variables (of both range and domain)
1165 The ProxyEnv keeps track of the connection between case binders and
1166 scrutinee. Specifically, if
1167 sc |-> (sc, [...(cb, co)...])
1168 is a binding in the ProxyEnv, then
1170 Typically we add such a binding when encountering the case expression
1171 case (sc |> coi) of cb { ... }
1174 * The domain of the ProxyEnv is the variable (or casted variable)
1175 scrutinees of enclosing cases. This is additionally used
1176 to ensure we gather occurrence info even for GlobalId scrutinees;
1177 see Note [Binder swap for GlobalId scrutinee]
1179 * The ProxyEnv is just an optimisation; you can throw away any
1180 element without losing correctness. And we do so when pushing
1181 it inside a binding (see trimProxyEnv).
1183 * Once scrutinee might map to many case binders: Eg
1184 case sc of cb1 { DEFAULT -> ....case sc of cb2 { ... } .. }
1187 * If sc1 |-> (sc2, [...(cb, co)...]), then sc1==sc2
1188 It's a UniqFM and we sometimes need the domain Id
1190 * Any particular case binder 'cb' occurs only once in entire range
1194 The Main Reason for having a ProxyEnv is so that when we encounter
1195 case e of cb { pi -> ri }
1196 we can find all the in-scope variables derivable from 'cb',
1197 and effectively add let-bindings for them thus:
1198 case e of cb { pi -> let { x = ..cb..; y = ...cb.. }
1200 The function getProxies finds these bindings; then we
1201 add just the necessary ones, using wrapProxy.
1203 More info under Note [Binder swap]
1207 We do these two transformations right here:
1209 (1) case x of b { pi -> ri }
1211 case x of b { pi -> let x=b in ri }
1213 (2) case (x |> co) of b { pi -> ri }
1215 case (x |> co) of b { pi -> let x = b |> sym co in ri }
1217 Why (2)? See Note [Case of cast]
1219 In both cases, in a particular alternative (pi -> ri), we only
1221 (a) x occurs free in (pi -> ri)
1222 (ie it occurs in ri, but is not bound in pi)
1223 (b) the pi does not bind b (or the free vars of co)
1224 We need (a) and (b) for the inserted binding to be correct.
1226 For the alternatives where we inject the binding, we can transfer
1227 all x's OccInfo to b. And that is the point.
1230 * The deliberate shadowing of 'x'.
1231 * That (a) rapidly becomes false, so no bindings are injected.
1233 The reason for doing these transformations here is because it allows
1234 us to adjust the OccInfo for 'x' and 'b' as we go.
1236 * Suppose the only occurrences of 'x' are the scrutinee and in the
1237 ri; then this transformation makes it occur just once, and hence
1238 get inlined right away.
1240 * If we do this in the Simplifier, we don't know whether 'x' is used
1241 in ri, so we are forced to pessimistically zap b's OccInfo even
1242 though it is typically dead (ie neither it nor x appear in the
1243 ri). There's nothing actually wrong with zapping it, except that
1244 it's kind of nice to know which variables are dead. My nose
1245 tells me to keep this information as robustly as possible.
1247 The Maybe (Id,CoreExpr) passed to occAnalAlt is the extra let-binding
1248 {x=b}; it's Nothing if the binder-swap doesn't happen.
1250 There is a danger though. Consider
1252 in case (f v) of w -> ...v...v...
1253 And suppose that (f v) expands to just v. Then we'd like to
1254 use 'w' instead of 'v' in the alternative. But it may be too
1255 late; we may have substituted the (cheap) x+#y for v in the
1256 same simplifier pass that reduced (f v) to v.
1258 I think this is just too bad. CSE will recover some of it.
1262 Consider case (x `cast` co) of b { I# ->
1263 ... (case (x `cast` co) of {...}) ...
1264 We'd like to eliminate the inner case. That is the motivation for
1265 equation (2) in Note [Binder swap]. When we get to the inner case, we
1266 inline x, cancel the casts, and away we go.
1268 Note [Binder swap on GlobalId scrutinees]
1269 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
1270 When the scrutinee is a GlobalId we must take care in two ways
1272 i) In order to *know* whether 'x' occurs free in the RHS, we need its
1273 occurrence info. BUT, we don't gather occurrence info for
1274 GlobalIds. That's one use for the (small) occ_proxy env in OccEnv is
1275 for: it says "gather occurrence info for these.
1277 ii) We must call localiseId on 'x' first, in case it's a GlobalId, or
1278 has an External Name. See, for example, SimplEnv Note [Global Ids in
1281 Note [getProxies is subtle]
1282 ~~~~~~~~~~~~~~~~~~~~~~~~~~~
1283 The code for getProxies isn't all that obvious. Consider
1285 case v |> cov of x { DEFAULT ->
1286 case x |> cox1 of y { DEFAULT ->
1287 case x |> cox2 of z { DEFAULT -> r
1289 These will give us a ProxyEnv looking like:
1290 x |-> (x, [(y, cox1), (z, cox2)])
1291 v |-> (v, [(x, cov)])
1293 From this we want to extract the bindings
1298 Notice that later bindings may mention earlier ones, and that
1299 we need to go "both ways".
1301 Historical note [no-case-of-case]
1302 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
1303 We *used* to suppress the binder-swap in case expressions when
1304 -fno-case-of-case is on. Old remarks:
1305 "This happens in the first simplifier pass,
1306 and enhances full laziness. Here's the bad case:
1307 f = \ y -> ...(case x of I# v -> ...(case x of ...) ... )
1308 If we eliminate the inner case, we trap it inside the I# v -> arm,
1309 which might prevent some full laziness happening. I've seen this
1310 in action in spectral/cichelli/Prog.hs:
1311 [(m,n) | m <- [1..max], n <- [1..max]]
1312 Hence the check for NoCaseOfCase."
1313 However, now the full-laziness pass itself reverses the binder-swap, so this
1314 check is no longer necessary.
1316 Historical note [Suppressing the case binder-swap]
1317 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
1318 This old note describes a problem that is also fixed by doing the
1319 binder-swap in OccAnal:
1321 There is another situation when it might make sense to suppress the
1322 case-expression binde-swap. If we have
1324 case x of w1 { DEFAULT -> case x of w2 { A -> e1; B -> e2 }
1325 ...other cases .... }
1327 We'll perform the binder-swap for the outer case, giving
1329 case x of w1 { DEFAULT -> case w1 of w2 { A -> e1; B -> e2 }
1330 ...other cases .... }
1332 But there is no point in doing it for the inner case, because w1 can't
1333 be inlined anyway. Furthermore, doing the case-swapping involves
1334 zapping w2's occurrence info (see paragraphs that follow), and that
1335 forces us to bind w2 when doing case merging. So we get
1337 case x of w1 { A -> let w2 = w1 in e1
1338 B -> let w2 = w1 in e2
1339 ...other cases .... }
1341 This is plain silly in the common case where w2 is dead.
1343 Even so, I can't see a good way to implement this idea. I tried
1344 not doing the binder-swap if the scrutinee was already evaluated
1345 but that failed big-time:
1349 case v of w { MkT x ->
1350 case x of x1 { I# y1 ->
1351 case x of x2 { I# y2 -> ...
1353 Notice that because MkT is strict, x is marked "evaluated". But to
1354 eliminate the last case, we must either make sure that x (as well as
1355 x1) has unfolding MkT y1. THe straightforward thing to do is to do
1356 the binder-swap. So this whole note is a no-op.
1358 It's fixed by doing the binder-swap in OccAnal because we can do the
1359 binder-swap unconditionally and still get occurrence analysis
1363 extendProxyEnv :: ProxyEnv -> Id -> CoercionI -> Id -> ProxyEnv
1364 -- (extendPE x co y) typically arises from
1365 -- case (x |> co) of y { ... }
1366 -- It extends the proxy env with the binding
1368 extendProxyEnv pe scrut co case_bndr
1369 | scrut == case_bndr = PE env1 fvs1 -- If case_bndr shadows scrut,
1370 | otherwise = PE env2 fvs2 -- don't extend
1372 PE env1 fvs1 = trimProxyEnv pe [case_bndr]
1373 env2 = extendVarEnv_Acc add single env1 scrut1 (case_bndr,co)
1374 single cb_co = (scrut1, [cb_co])
1375 add cb_co (x, cb_cos) = (x, cb_co:cb_cos)
1376 fvs2 = fvs1 `unionVarSet` freeVarsCoI co
1377 `extendVarSet` case_bndr
1378 `extendVarSet` scrut1
1380 scrut1 = mkLocalId (localiseName (idName scrut)) (idType scrut)
1381 -- Localise the scrut_var before shadowing it; we're making a
1382 -- new binding for it, and it might have an External Name, or
1383 -- even be a GlobalId; Note [Binder swap on GlobalId scrutinees]
1384 -- Also we don't want any INLILNE or NOINLINE pragmas!
1387 type ProxyBind = (Id, Id, CoercionI)
1389 getProxies :: OccEnv -> Id -> Bag ProxyBind
1390 -- Return a bunch of bindings [...(xi,ei)...]
1391 -- such that let { ...; xi=ei; ... } binds the xi using y alone
1392 -- See Note [getProxies is subtle]
1393 getProxies (OccEnv { occ_proxy = PE pe _ }) case_bndr
1394 = -- pprTrace "wrapProxies" (ppr case_bndr) $
1397 fwd_pe :: IdEnv (Id, CoercionI)
1398 fwd_pe = foldVarEnv add1 emptyVarEnv pe
1400 add1 (x,ycos) env = foldr (add2 x) env ycos
1401 add2 x (y,co) env = extendVarEnv env y (x,co)
1403 go_fwd :: Id -> Bag ProxyBind
1404 -- Return bindings derivable from case_bndr
1405 go_fwd case_bndr = -- pprTrace "go_fwd" (vcat [ppr case_bndr, text "fwd_pe =" <+> ppr fwd_pe,
1406 -- text "pe =" <+> ppr pe]) $
1410 | Just (scrut, co) <- lookupVarEnv fwd_pe case_bndr
1411 = unitBag (scrut, case_bndr, mkSymCoI co)
1412 `unionBags` go_fwd scrut
1413 `unionBags` go_bwd scrut [pr | pr@(cb,_) <- lookup_bwd scrut
1418 lookup_bwd :: Id -> [(Id, CoercionI)]
1419 -- Return case_bndrs that are connected to scrut
1420 lookup_bwd scrut = case lookupVarEnv pe scrut of
1422 Just (_, cb_cos) -> cb_cos
1424 go_bwd :: Id -> [(Id, CoercionI)] -> Bag ProxyBind
1425 go_bwd scrut cb_cos = foldr (unionBags . go_bwd1 scrut) emptyBag cb_cos
1427 go_bwd1 :: Id -> (Id, CoercionI) -> Bag ProxyBind
1428 go_bwd1 scrut (case_bndr, co)
1429 = -- pprTrace "go_bwd1" (ppr case_bndr) $
1430 unitBag (case_bndr, scrut, co)
1431 `unionBags` go_bwd case_bndr (lookup_bwd case_bndr)
1434 mkAltEnv :: OccEnv -> CoreExpr -> Id -> OccEnv
1435 -- Does two things: a) makes the occ_ctxt = OccVanilla
1436 -- b) extends the ProxyEnv if possible
1437 mkAltEnv env scrut cb
1438 = env { occ_encl = OccVanilla, occ_proxy = pe' }
1442 Var v -> extendProxyEnv pe v IdCo cb
1443 Cast (Var v) co -> extendProxyEnv pe v (ACo co) cb
1444 _other -> trimProxyEnv pe [cb]
1447 trimOccEnv :: OccEnv -> [CoreBndr] -> OccEnv
1448 trimOccEnv env bndrs = env { occ_proxy = trimProxyEnv (occ_proxy env) bndrs }
1451 trimProxyEnv :: ProxyEnv -> [CoreBndr] -> ProxyEnv
1452 -- We are about to push this ProxyEnv inside a binding for 'bndrs'
1453 -- So dump any ProxyEnv bindings which mention any of the bndrs
1454 trimProxyEnv (PE pe fvs) bndrs
1455 | not (bndr_set `intersectsVarSet` fvs)
1458 = PE pe' (fvs `minusVarSet` bndr_set)
1460 pe' = mapVarEnv trim pe
1461 bndr_set = mkVarSet bndrs
1462 trim (scrut, cb_cos) | scrut `elemVarSet` bndr_set = (scrut, [])
1463 | otherwise = (scrut, filterOut discard cb_cos)
1464 discard (cb,co) = bndr_set `intersectsVarSet`
1465 extendVarSet (freeVarsCoI co) cb
1468 freeVarsCoI :: CoercionI -> VarSet
1469 freeVarsCoI IdCo = emptyVarSet
1470 freeVarsCoI (ACo co) = tyVarsOfType co
1474 %************************************************************************
1476 \subsection[OccurAnal-types]{OccEnv}
1478 %************************************************************************
1481 type UsageDetails = IdEnv OccInfo -- A finite map from ids to their usage
1482 -- INVARIANT: never IAmDead
1483 -- (Deadness is signalled by not being in the map at all)
1485 (+++), combineAltsUsageDetails
1486 :: UsageDetails -> UsageDetails -> UsageDetails
1489 = plusVarEnv_C addOccInfo usage1 usage2
1491 combineAltsUsageDetails usage1 usage2
1492 = plusVarEnv_C orOccInfo usage1 usage2
1494 addOneOcc :: UsageDetails -> Id -> OccInfo -> UsageDetails
1495 addOneOcc usage id info
1496 = plusVarEnv_C addOccInfo usage (unitVarEnv id info)
1497 -- ToDo: make this more efficient
1499 emptyDetails :: UsageDetails
1500 emptyDetails = (emptyVarEnv :: UsageDetails)
1502 localUsedIn, usedIn :: Id -> UsageDetails -> Bool
1503 v `localUsedIn` details = v `elemVarEnv` details
1504 v `usedIn` details = isExportedId v || v `localUsedIn` details
1506 type IdWithOccInfo = Id
1508 tagLamBinders :: UsageDetails -- Of scope
1510 -> (UsageDetails, -- Details with binders removed
1511 [IdWithOccInfo]) -- Tagged binders
1512 -- Used for lambda and case binders
1513 -- It copes with the fact that lambda bindings can have InlineRule
1514 -- unfoldings, used for join points
1515 tagLamBinders usage binders = usage' `seq` (usage', bndrs')
1517 (usage', bndrs') = mapAccumR tag_lam usage binders
1518 tag_lam usage bndr = (usage2, setBinderOcc usage bndr)
1520 usage1 = usage `delVarEnv` bndr
1521 usage2 | isId bndr = addIdOccs usage1 (idUnfoldingVars bndr)
1522 | otherwise = usage1
1524 tagBinder :: UsageDetails -- Of scope
1526 -> (UsageDetails, -- Details with binders removed
1527 IdWithOccInfo) -- Tagged binders
1529 tagBinder usage binder
1531 usage' = usage `delVarEnv` binder
1532 binder' = setBinderOcc usage binder
1534 usage' `seq` (usage', binder')
1536 setBinderOcc :: UsageDetails -> CoreBndr -> CoreBndr
1537 setBinderOcc usage bndr
1538 | isTyVar bndr = bndr
1539 | isExportedId bndr = case idOccInfo bndr of
1541 _ -> setIdOccInfo bndr NoOccInfo
1542 -- Don't use local usage info for visible-elsewhere things
1543 -- BUT *do* erase any IAmALoopBreaker annotation, because we're
1544 -- about to re-generate it and it shouldn't be "sticky"
1546 | otherwise = setIdOccInfo bndr occ_info
1548 occ_info = lookupVarEnv usage bndr `orElse` IAmDead
1552 %************************************************************************
1554 \subsection{Operations over OccInfo}
1556 %************************************************************************
1559 mkOneOcc :: OccEnv -> Id -> InterestingCxt -> UsageDetails
1560 mkOneOcc env id int_cxt
1561 | isLocalId id = unitVarEnv id (OneOcc False True int_cxt)
1562 | PE env _ <- occ_proxy env
1563 , id `elemVarEnv` env = unitVarEnv id NoOccInfo
1564 | otherwise = emptyDetails
1566 markMany, markInsideLam, markInsideSCC :: OccInfo -> OccInfo
1568 markMany _ = NoOccInfo
1570 markInsideSCC occ = markMany occ
1572 markInsideLam (OneOcc _ one_br int_cxt) = OneOcc True one_br int_cxt
1573 markInsideLam occ = occ
1575 addOccInfo, orOccInfo :: OccInfo -> OccInfo -> OccInfo
1577 addOccInfo a1 a2 = ASSERT( not (isDeadOcc a1 || isDeadOcc a2) )
1578 NoOccInfo -- Both branches are at least One
1579 -- (Argument is never IAmDead)
1581 -- (orOccInfo orig new) is used
1582 -- when combining occurrence info from branches of a case
1584 orOccInfo (OneOcc in_lam1 _ int_cxt1)
1585 (OneOcc in_lam2 _ int_cxt2)
1586 = OneOcc (in_lam1 || in_lam2)
1587 False -- False, because it occurs in both branches
1588 (int_cxt1 && int_cxt2)
1589 orOccInfo a1 a2 = ASSERT( not (isDeadOcc a1 || isDeadOcc a2) )