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, isExpandableApp, mkCoerce )
26 import Name ( Name, localiseName )
34 import Maybes ( orElse )
35 import Digraph ( SCC(..), stronglyConnCompFromEdgedVerticesR )
36 import PrelNames ( buildIdKey, foldrIdKey, runSTRepIdKey, augmentIdKey )
39 import Util ( mapAndUnzip, filterOut )
47 %************************************************************************
49 \subsection[OccurAnal-main]{Counting occurrences: main function}
51 %************************************************************************
53 Here's the externally-callable interface:
56 occurAnalysePgm :: Maybe (Activation -> Bool) -> [CoreRule] -> [CoreVect]
57 -> [CoreBind] -> [CoreBind]
58 occurAnalysePgm active_rule imp_rules vects binds
59 = snd (go (initOccEnv active_rule imp_rules) binds)
61 initial_uds = addIdOccs emptyDetails
62 (rulesFreeVars imp_rules `unionVarSet` vectsFreeVars vects)
63 -- The RULES and VECTORISE declarations keep things alive!
65 go :: OccEnv -> [CoreBind] -> (UsageDetails, [CoreBind])
69 = (final_usage, bind' ++ binds')
71 (bs_usage, binds') = go env binds
72 (final_usage, bind') = occAnalBind env env bind bs_usage
74 occurAnalyseExpr :: CoreExpr -> CoreExpr
75 -- Do occurrence analysis, and discard occurence info returned
77 = snd (occAnal (initOccEnv all_active_rules []) expr)
79 -- To be conservative, we say that all inlines and rules are active
80 all_active_rules = Just (\_ -> True)
84 %************************************************************************
86 \subsection[OccurAnal-main]{Counting occurrences: main function}
88 %************************************************************************
94 occAnalBind :: OccEnv -- The incoming OccEnv
95 -> OccEnv -- Same, but trimmed by (binderOf bind)
97 -> UsageDetails -- Usage details of scope
98 -> (UsageDetails, -- Of the whole let(rec)
101 occAnalBind env _ (NonRec binder rhs) body_usage
102 | isTyVar binder -- A type let; we don't gather usage info
103 = (body_usage, [NonRec binder rhs])
105 | not (binder `usedIn` body_usage) -- It's not mentioned
108 | otherwise -- It's mentioned in the body
109 = (body_usage' +++ rhs_usage3, [NonRec tagged_binder rhs'])
111 (body_usage', tagged_binder) = tagBinder body_usage binder
112 (rhs_usage1, rhs') = occAnalRhs env (Just tagged_binder) rhs
113 rhs_usage2 = addIdOccs rhs_usage1 (idUnfoldingVars binder)
114 rhs_usage3 = addIdOccs rhs_usage2 (idRuleVars binder)
115 -- See Note [Rules are extra RHSs] and Note [Rule dependency info]
120 Dropping dead code for recursive bindings is done in a very simple way:
122 the entire set of bindings is dropped if none of its binders are
123 mentioned in its body; otherwise none are.
125 This seems to miss an obvious improvement.
137 Now 'f' is unused! But it's OK! Dependency analysis will sort this
138 out into a letrec for 'g' and a 'let' for 'f', and then 'f' will get
139 dropped. It isn't easy to do a perfect job in one blow. Consider
150 Note [Loop breaking and RULES]
151 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
152 Loop breaking is surprisingly subtle. First read the section 4 of
153 "Secrets of the GHC inliner". This describes our basic plan.
155 However things are made quite a bit more complicated by RULES. Remember
157 * Note [Rules are extra RHSs]
158 ~~~~~~~~~~~~~~~~~~~~~~~~~~~
159 A RULE for 'f' is like an extra RHS for 'f'. That way the "parent"
160 keeps the specialised "children" alive. If the parent dies
161 (because it isn't referenced any more), then the children will die
162 too (unless they are already referenced directly).
164 To that end, we build a Rec group for each cyclic strongly
166 *treating f's rules as extra RHSs for 'f'*.
167 More concretely, the SCC analysis runs on a graph with an edge
168 from f -> g iff g is mentioned in
173 Under (b) we include variables free in *either* LHS *or* RHS of
174 the rule. The former might seems silly, but see Note [Rule
175 dependency info]. So in Example [eftInt], eftInt and eftIntFB
176 will be put in the same Rec, even though their 'main' RHSs are
179 * Note [Rules are visible in their own rec group]
180 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
181 We want the rules for 'f' to be visible in f's right-hand side.
182 And we'd like them to be visible in other functions in f's Rec
183 group. E.g. in Example [Specialisation rules] we want f' rule
184 to be visible in both f's RHS, and fs's RHS.
186 This means that we must simplify the RULEs first, before looking
187 at any of the definitions. This is done by Simplify.simplRecBind,
188 when it calls addLetIdInfo.
190 * Note [Choosing loop breakers]
191 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
192 We avoid infinite inlinings by choosing loop breakers, and
193 ensuring that a loop breaker cuts each loop. But what is a
194 "loop"? In particular, a RULE is like an equation for 'f' that
195 is *always* inlined if it is applicable. We do *not* disable
196 rules for loop-breakers. It's up to whoever makes the rules to
197 make sure that the rules themselves always terminate. See Note
198 [Rules for recursive functions] in Simplify.lhs
201 f's RHS (or its INLINE template if it has one) mentions g, and
202 g has a RULE that mentions h, and
203 h has a RULE that mentions f
205 then we *must* choose f to be a loop breaker. In general, take the
206 free variables of f's RHS, and augment it with all the variables
207 reachable by RULES from those starting points. That is the whole
208 reason for computing rule_fv_env in occAnalBind. (Of course we
209 only consider free vars that are also binders in this Rec group.)
210 See also Note [Finding rule RHS free vars]
212 Note that when we compute this rule_fv_env, we only consider variables
213 free in the *RHS* of the rule, in contrast to the way we build the
214 Rec group in the first place (Note [Rule dependency info])
216 Note that if 'g' has RHS that mentions 'w', we should add w to
217 g's loop-breaker edges. More concretely there is an edge from f -> g
219 (a) g is mentioned in f's RHS
220 (b) h is mentioned in f's RHS, and
221 g appears in the RHS of a RULE of h
222 or a transitive sequence of rules starting with h
224 Note that in Example [eftInt], *neither* eftInt *nor* eftIntFB is
225 chosen as a loop breaker, because their RHSs don't mention each other.
226 And indeed both can be inlined safely.
228 Note that the edges of the graph we use for computing loop breakers
229 are not the same as the edges we use for computing the Rec blocks.
230 That's why we compute
231 rec_edges for the Rec block analysis
232 loop_breaker_edges for the loop breaker analysis
234 * Note [Finding rule RHS free vars]
235 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
236 Consider this real example from Data Parallel Haskell
237 tagZero :: Array Int -> Array Tag
238 {-# INLINE [1] tagZeroes #-}
239 tagZero xs = pmap (\x -> fromBool (x==0)) xs
241 {-# RULES "tagZero" [~1] forall xs n.
242 pmap fromBool <blah blah> = tagZero xs #-}
243 So tagZero's RHS mentions pmap, and pmap's RULE mentions tagZero.
244 However, tagZero can only be inlined in phase 1 and later, while
245 the RULE is only active *before* phase 1. So there's no problem.
247 To make this work, we look for the RHS free vars only for
248 *active* rules. That's the reason for the is_active argument
249 to idRhsRuleVars, and the occ_rule_act field of the OccEnv.
251 * Note [Weak loop breakers]
252 ~~~~~~~~~~~~~~~~~~~~~~~~~
253 There is a last nasty wrinkle. Suppose we have
263 Remember that we simplify the RULES before any RHS (see Note
264 [Rules are visible in their own rec group] above).
266 So we must *not* postInlineUnconditionally 'g', even though
267 its RHS turns out to be trivial. (I'm assuming that 'g' is
268 not choosen as a loop breaker.) Why not? Because then we
269 drop the binding for 'g', which leaves it out of scope in the
272 We "solve" this by making g a "weak" or "rules-only" loop breaker,
273 with OccInfo = IAmLoopBreaker True. A normal "strong" loop breaker
274 has IAmLoopBreaker False. So
276 Inline postInlineUnconditionally
277 IAmLoopBreaker False no no
278 IAmLoopBreaker True yes no
281 The **sole** reason for this kind of loop breaker is so that
282 postInlineUnconditionally does not fire. Ugh.
284 * Note [Rule dependency info]
285 ~~~~~~~~~~~~~~~~~~~~~~~~~~~
286 The VarSet in a SpecInfo is used for dependency analysis in the
287 occurrence analyser. We must track free vars in *both* lhs and rhs.
288 Hence use of idRuleVars, rather than idRuleRhsVars in occAnalBind.
292 Then if we substitute y for x, we'd better do so in the
293 rule's LHS too, so we'd better ensure the dependency is respected
296 * Note [Inline rules]
298 None of the above stuff about RULES applies to Inline Rules,
299 stored in a CoreUnfolding. The unfolding, if any, is simplified
300 at the same time as the regular RHS of the function, so it should
301 be treated *exactly* like an extra RHS.
303 There is a danger that we'll be sub-optimal if we see this
305 [INLINE f = ..no f...]
306 where f is recursive, but the INLINE is not. This can just about
307 happen with a sufficiently odd set of rules; eg
310 {-# INLINE [1] foo #-}
314 {-# INLINE [1] bar #-}
317 {-# RULES "foo" [~1] forall x. foo x = bar x #-}
319 Here the RULE makes bar recursive; but it's INLINE pragma remains
320 non-recursive. It's tempting to then say that 'bar' should not be
321 a loop breaker, but an attempt to do so goes wrong in two ways:
325 [INLINE $cfoo = ...no-$df...]
326 But we want $cfoo to depend on $df explicitly so that we
327 put the bindings in the right order to inline $df in $cfoo
328 and perhaps break the loop altogether. (Maybe this
335 Example (from GHC.Enum):
337 eftInt :: Int# -> Int# -> [Int]
338 eftInt x y = ...(non-recursive)...
340 {-# INLINE [0] eftIntFB #-}
341 eftIntFB :: (Int -> r -> r) -> r -> Int# -> Int# -> r
342 eftIntFB c n x y = ...(non-recursive)...
345 "eftInt" [~1] forall x y. eftInt x y = build (\ c n -> eftIntFB c n x y)
346 "eftIntList" [1] eftIntFB (:) [] = eftInt
349 Example [Specialisation rules]
350 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
351 Consider this group, which is typical of what SpecConstr builds:
353 fs a = ....f (C a)....
354 f x = ....f (C a)....
355 {-# RULE f (C a) = fs a #-}
357 So 'f' and 'fs' are in the same Rec group (since f refers to fs via its RULE).
359 But watch out! If 'fs' is not chosen as a loop breaker, we may get an infinite loop:
360 - the RULE is applied in f's RHS (see Note [Self-recursive rules] in Simplify
361 - fs is inlined (say it's small)
362 - now there's another opportunity to apply the RULE
364 This showed up when compiling Control.Concurrent.Chan.getChanContents.
368 occAnalBind _ env (Rec pairs) body_usage
369 = foldr (occAnalRec env) (body_usage, []) sccs
370 -- For a recursive group, we
371 -- * occ-analyse all the RHSs
372 -- * compute strongly-connected components
373 -- * feed those components to occAnalRec
375 -------------Dependency analysis ------------------------------
376 bndr_set = mkVarSet (map fst pairs)
378 sccs :: [SCC (Node Details)]
379 sccs = {-# SCC "occAnalBind.scc" #-} stronglyConnCompFromEdgedVerticesR rec_edges
381 rec_edges :: [Node Details]
382 rec_edges = {-# SCC "occAnalBind.assoc" #-} map make_node pairs
384 make_node (bndr, rhs)
385 = (details, varUnique bndr, keysUFM out_edges)
387 details = ND { nd_bndr = bndr, nd_rhs = rhs'
388 , nd_uds = rhs_usage3, nd_inl = inl_fvs}
390 (rhs_usage1, rhs') = occAnalRhs env Nothing rhs
391 rhs_usage2 = addIdOccs rhs_usage1 rule_fvs -- Note [Rules are extra RHSs]
392 rhs_usage3 = addIdOccs rhs_usage2 unf_fvs
393 unf = realIdUnfolding bndr -- Ignore any current loop-breaker flag
394 unf_fvs = stableUnfoldingVars unf
395 rule_fvs = idRuleVars bndr -- See Note [Rule dependency info]
397 inl_fvs = rhs_fvs `unionVarSet` unf_fvs
398 rhs_fvs = intersectUFM_C (\b _ -> b) bndr_set rhs_usage1
399 out_edges = intersectUFM_C (\b _ -> b) bndr_set rhs_usage3
400 -- (a -> b) means a mentions b
401 -- Given the usage details (a UFM that gives occ info for each free var of
402 -- the RHS) we can get the list of free vars -- or rather their Int keys --
403 -- by just extracting the keys from the finite map. Grimy, but fast.
404 -- Previously we had this:
405 -- [ bndr | bndr <- bndrs,
406 -- maybeToBool (lookupVarEnv rhs_usage bndr)]
407 -- which has n**2 cost, and this meant that edges_from alone
408 -- consumed 10% of total runtime!
410 -----------------------------
411 occAnalRec :: OccEnv -> SCC (Node Details)
412 -> (UsageDetails, [CoreBind])
413 -> (UsageDetails, [CoreBind])
415 -- The NonRec case is just like a Let (NonRec ...) above
416 occAnalRec _ (AcyclicSCC (ND { nd_bndr = bndr, nd_rhs = rhs, nd_uds = rhs_usage}, _, _))
418 | not (bndr `usedIn` body_usage)
419 = (body_usage, binds)
421 | otherwise -- It's mentioned in the body
422 = (body_usage' +++ rhs_usage,
423 NonRec tagged_bndr rhs : binds)
425 (body_usage', tagged_bndr) = tagBinder body_usage bndr
428 -- The Rec case is the interesting one
429 -- See Note [Loop breaking]
430 occAnalRec env (CyclicSCC nodes) (body_usage, binds)
431 | not (any (`usedIn` body_usage) bndrs) -- NB: look at body_usage, not total_usage
432 = (body_usage, binds) -- Dead code
434 | otherwise -- At this point we always build a single Rec
435 = (final_usage, Rec pairs : binds)
438 bndrs = [b | (ND { nd_bndr = b }, _, _) <- nodes]
439 bndr_set = mkVarSet bndrs
440 non_boring bndr = isId bndr &&
441 (isStableUnfolding (realIdUnfolding bndr) || idHasRules bndr)
443 ----------------------------
444 -- Tag the binders with their occurrence info
445 total_usage = foldl add_usage body_usage nodes
446 add_usage usage_so_far (ND { nd_uds = rhs_usage }, _, _) = usage_so_far +++ rhs_usage
447 (final_usage, tagged_nodes) = mapAccumL tag_node total_usage nodes
449 tag_node :: UsageDetails -> Node Details -> (UsageDetails, Node Details)
450 -- (a) Tag the binders in the details with occ info
451 -- (b) Mark the binder with "weak loop-breaker" OccInfo
452 -- saying "no preInlineUnconditionally" if it is used
453 -- in any rule (lhs or rhs) of the recursive group
454 -- See Note [Weak loop breakers]
455 tag_node usage (details@ND { nd_bndr = bndr }, k, ks)
456 = (usage `delVarEnv` bndr, (details { nd_bndr = bndr2 }, k, ks))
458 bndr2 | bndr `elemVarSet` all_rule_fvs = makeLoopBreaker True bndr1
460 bndr1 = setBinderOcc usage bndr
461 all_rule_fvs = bndr_set `intersectVarSet` foldr (unionVarSet . idRuleVars)
464 ----------------------------
465 -- Now reconstruct the cycle
466 pairs | any non_boring bndrs
467 = foldr (reOrderRec 0) [] $
468 stronglyConnCompFromEdgedVerticesR loop_breaker_edges
470 = reOrderCycle 0 tagged_nodes []
472 -- See Note [Choosing loop breakers] for loop_breaker_edges
473 loop_breaker_edges = map mk_node tagged_nodes
474 mk_node (details@(ND { nd_inl = inl_fvs }), k, _) = (details, k, new_ks)
476 new_ks = keysUFM (fst (extendFvs rule_fv_env inl_fvs))
478 ------------------------------------
479 rule_fv_env :: IdEnv IdSet -- Variables from this group mentioned in RHS of rules
480 -- Domain is *subset* of bound vars (others have no rule fvs)
481 rule_fv_env = transClosureFV init_rule_fvs
483 | Just is_active <- occ_rule_act env -- See Note [Finding rule RHS free vars]
487 , let rule_fvs = idRuleRhsVars is_active b
488 `intersectVarSet` bndr_set
489 , not (isEmptyVarSet rule_fvs)]
494 @reOrderRec@ is applied to the list of (binder,rhs) pairs for a cyclic
495 strongly connected component (there's guaranteed to be a cycle). It returns the
497 a) in a better order,
498 b) with some of the Ids having a IAmALoopBreaker pragma
500 The "loop-breaker" Ids are sufficient to break all cycles in the SCC. This means
501 that the simplifier can guarantee not to loop provided it never records an inlining
502 for these no-inline guys.
504 Furthermore, the order of the binds is such that if we neglect dependencies
505 on the no-inline Ids then the binds are topologically sorted. This means
506 that the simplifier will generally do a good job if it works from top bottom,
507 recording inlinings for any Ids which aren't marked as "no-inline" as it goes.
510 [June 98: I don't understand the following paragraphs, and I've
511 changed the a=b case again so that it isn't a special case any more.]
513 Here's a case that bit me:
521 Re-ordering doesn't change the order of bindings, but there was no loop-breaker.
523 My solution was to make a=b bindings record b as Many, rather like INLINE bindings.
524 Perhaps something cleverer would suffice.
529 type Node details = (details, Unique, [Unique]) -- The Ints are gotten from the Unique,
530 -- which is gotten from the Id.
532 = ND { nd_bndr :: Id -- Binder
533 , nd_rhs :: CoreExpr -- RHS
535 , nd_uds :: UsageDetails -- Usage from RHS,
536 -- including RULES and InlineRule unfolding
538 , nd_inl :: IdSet -- Other binders *from this Rec group* mentioned in
539 } -- its InlineRule unfolding (if present)
541 -- but *excluding* any RULES
542 -- This is the IdSet that may be used if the Id is inlined
544 reOrderRec :: Int -> SCC (Node Details)
545 -> [(Id,CoreExpr)] -> [(Id,CoreExpr)]
546 -- Sorted into a plausible order. Enough of the Ids have
547 -- IAmALoopBreaker pragmas that there are no loops left.
548 reOrderRec _ (AcyclicSCC (ND { nd_bndr = bndr, nd_rhs = rhs }, _, _))
549 pairs = (bndr, rhs) : pairs
550 reOrderRec depth (CyclicSCC cycle) pairs = reOrderCycle depth cycle pairs
552 reOrderCycle :: Int -> [Node Details] -> [(Id,CoreExpr)] -> [(Id,CoreExpr)]
554 = panic "reOrderCycle"
555 reOrderCycle _ [(ND { nd_bndr = bndr, nd_rhs = rhs }, _, _)] pairs
556 = -- Common case of simple self-recursion
557 (makeLoopBreaker False bndr, rhs) : pairs
559 reOrderCycle depth (bind : binds) pairs
560 = -- Choose a loop breaker, mark it no-inline,
561 -- do SCC analysis on the rest, and recursively sort them out
562 -- pprTrace "reOrderCycle" (ppr [b | (ND { nd_bndr = b }, _, _) <- bind:binds]) $
563 foldr (reOrderRec new_depth)
564 ([ (makeLoopBreaker False bndr, rhs)
565 | (ND { nd_bndr = bndr, nd_rhs = rhs }, _, _) <- chosen_binds] ++ pairs)
566 (stronglyConnCompFromEdgedVerticesR unchosen)
568 (chosen_binds, unchosen) = choose_loop_breaker [bind] (score bind) [] binds
570 approximate_loop_breaker = depth >= 2
571 new_depth | approximate_loop_breaker = 0
572 | otherwise = depth+1
573 -- After two iterations (d=0, d=1) give up
574 -- and approximate, returning to d=0
576 -- This loop looks for the bind with the lowest score
577 -- to pick as the loop breaker. The rest accumulate in
578 choose_loop_breaker loop_binds _loop_sc acc []
579 = (loop_binds, acc) -- Done
581 -- If approximate_loop_breaker is True, we pick *all*
582 -- nodes with lowest score, else just one
583 -- See Note [Complexity of loop breaking]
584 choose_loop_breaker loop_binds loop_sc acc (bind : binds)
585 | sc < loop_sc -- Lower score so pick this new one
586 = choose_loop_breaker [bind] sc (loop_binds ++ acc) binds
588 | approximate_loop_breaker && sc == loop_sc
589 = choose_loop_breaker (bind : loop_binds) loop_sc acc binds
591 | otherwise -- Higher score so don't pick it
592 = choose_loop_breaker loop_binds loop_sc (bind : acc) binds
596 score :: Node Details -> Int -- Higher score => less likely to be picked as loop breaker
597 score (ND { nd_bndr = bndr, nd_rhs = rhs }, _, _)
598 | not (isId bndr) = 100 -- A type or cercion variable is never a loop breaker
600 | isDFunId bndr = 9 -- Never choose a DFun as a loop breaker
601 -- Note [DFuns should not be loop breakers]
603 | Just inl_source <- isStableCoreUnfolding_maybe (idUnfolding bndr)
605 InlineWrapper {} -> 10 -- Note [INLINE pragmas]
606 _other -> 3 -- Data structures are more important than this
607 -- so that dictionary/method recursion unravels
608 -- Note that this case hits all InlineRule things, so we
609 -- never look at 'rhs for InlineRule stuff. That's right, because
610 -- 'rhs' is irrelevant for inlining things with an InlineRule
612 | is_con_app rhs = 5 -- Data types help with cases: Note [Constructor applications]
614 | exprIsTrivial rhs = 10 -- Practically certain to be inlined
615 -- Used to have also: && not (isExportedId bndr)
616 -- But I found this sometimes cost an extra iteration when we have
617 -- rec { d = (a,b); a = ...df...; b = ...df...; df = d }
618 -- where df is the exported dictionary. Then df makes a really
619 -- bad choice for loop breaker
622 -- If an Id is marked "never inline" then it makes a great loop breaker
623 -- The only reason for not checking that here is that it is rare
624 -- and I've never seen a situation where it makes a difference,
625 -- so it probably isn't worth the time to test on every binder
626 -- | isNeverActive (idInlinePragma bndr) = -10
628 | isOneOcc (idOccInfo bndr) = 2 -- Likely to be inlined
630 | canUnfold (realIdUnfolding bndr) = 1
631 -- The Id has some kind of unfolding
632 -- Ignore loop-breaker-ness here because that is what we are setting!
636 -- Checking for a constructor application
637 -- Cheap and cheerful; the simplifer moves casts out of the way
638 -- The lambda case is important to spot x = /\a. C (f a)
639 -- which comes up when C is a dictionary constructor and
640 -- f is a default method.
641 -- Example: the instance for Show (ST s a) in GHC.ST
643 -- However we *also* treat (\x. C p q) as a con-app-like thing,
644 -- Note [Closure conversion]
645 is_con_app (Var v) = isConLikeId v
646 is_con_app (App f _) = is_con_app f
647 is_con_app (Lam _ e) = is_con_app e
648 is_con_app (Note _ e) = is_con_app e
651 makeLoopBreaker :: Bool -> Id -> Id
652 -- Set the loop-breaker flag: see Note [Weak loop breakers]
653 makeLoopBreaker weak bndr
654 = ASSERT2( isId bndr, ppr bndr ) setIdOccInfo bndr (IAmALoopBreaker weak)
657 Note [Complexity of loop breaking]
658 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
659 The loop-breaking algorithm knocks out one binder at a time, and
660 performs a new SCC analysis on the remaining binders. That can
661 behave very badly in tightly-coupled groups of bindings; in the
662 worst case it can be (N**2)*log N, because it does a full SCC
663 on N, then N-1, then N-2 and so on.
665 To avoid this, we switch plans after 2 (or whatever) attempts:
666 Plan A: pick one binder with the lowest score, make it
667 a loop breaker, and try again
668 Plan B: pick *all* binders with the lowest score, make them
669 all loop breakers, and try again
670 Since there are only a small finite number of scores, this will
671 terminate in a constant number of iterations, rather than O(N)
674 You might thing that it's very unlikely, but RULES make it much
675 more likely. Here's a real example from Trac #1969:
676 Rec { $dm = \d.\x. op d
677 {-# RULES forall d. $dm Int d = $s$dm1
678 forall d. $dm Bool d = $s$dm2 #-}
680 dInt = MkD .... opInt ...
681 dInt = MkD .... opBool ...
686 $s$dm2 = \x. op dBool }
687 The RULES stuff means that we can't choose $dm as a loop breaker
688 (Note [Choosing loop breakers]), so we must choose at least (say)
689 opInt *and* opBool, and so on. The number of loop breakders is
690 linear in the number of instance declarations.
692 Note [INLINE pragmas]
693 ~~~~~~~~~~~~~~~~~~~~~
694 Avoid choosing a function with an INLINE pramga as the loop breaker!
695 If such a function is mutually-recursive with a non-INLINE thing,
696 then the latter should be the loop-breaker.
698 Usually this is just a question of optimisation. But a particularly
699 bad case is wrappers generated by the demand analyser: if you make
700 then into a loop breaker you may get an infinite inlining loop. For
703 $wfoo x = ....foo x....
705 {-loop brk-} foo x = ...$wfoo x...
707 The interface file sees the unfolding for $wfoo, and sees that foo is
708 strict (and hence it gets an auto-generated wrapper). Result: an
709 infinite inlining in the importing scope. So be a bit careful if you
710 change this. A good example is Tree.repTree in
711 nofib/spectral/minimax. If the repTree wrapper is chosen as the loop
712 breaker then compiling Game.hs goes into an infinite loop. This
713 happened when we gave is_con_app a lower score than inline candidates:
716 = __inline_me (/\a. \w w1 w2 ->
717 case Tree.$wrepTree @ a w w1 w2 of
718 { (# ww1, ww2 #) -> Branch @ a ww1 ww2 })
721 (# w2_smP, map a (Tree a) (Tree.repTree a w1 w) (w w2) #)
723 Here we do *not* want to choose 'repTree' as the loop breaker.
725 Note [DFuns should not be loop breakers]
726 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
727 It's particularly bad to make a DFun into a loop breaker. See
728 Note [How instance declarations are translated] in TcInstDcls
730 We give DFuns a higher score than ordinary CONLIKE things because
731 if there's a choice we want the DFun to be the non-looop breker. Eg
733 rec { sc = /\ a \$dC. $fBWrap (T a) ($fCT @ a $dC)
735 $fCT :: forall a_afE. (Roman.C a_afE) => Roman.C (Roman.T a_afE)
737 $fCT = /\a \$dC. MkD (T a) ((sc @ a $dC) |> blah) ($ctoF @ a $dC)
740 Here 'sc' (the superclass) looks CONLIKE, but we'll never get to it
741 if we can't unravel the DFun first.
743 Note [Constructor applications]
744 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
745 It's really really important to inline dictionaries. Real
746 example (the Enum Ordering instance from GHC.Base):
748 rec f = \ x -> case d of (p,q,r) -> p x
749 g = \ x -> case d of (p,q,r) -> q x
752 Here, f and g occur just once; but we can't inline them into d.
753 On the other hand we *could* simplify those case expressions if
754 we didn't stupidly choose d as the loop breaker.
755 But we won't because constructor args are marked "Many".
756 Inlining dictionaries is really essential to unravelling
757 the loops in static numeric dictionaries, see GHC.Float.
759 Note [Closure conversion]
760 ~~~~~~~~~~~~~~~~~~~~~~~~~
761 We treat (\x. C p q) as a high-score candidate in the letrec scoring algorithm.
762 The immediate motivation came from the result of a closure-conversion transformation
763 which generated code like this:
765 data Clo a b = forall c. Clo (c -> a -> b) c
767 ($:) :: Clo a b -> a -> b
768 Clo f env $: x = f env x
770 rec { plus = Clo plus1 ()
772 ; plus1 _ n = Clo plus2 n
775 ; plus2 (Succ m) n = Succ (plus $: m $: n) }
777 If we inline 'plus' and 'plus1', everything unravels nicely. But if
778 we choose 'plus1' as the loop breaker (which is entirely possible
779 otherwise), the loop does not unravel nicely.
782 @occAnalRhs@ deals with the question of bindings where the Id is marked
783 by an INLINE pragma. For these we record that anything which occurs
784 in its RHS occurs many times. This pessimistically assumes that ths
785 inlined binder also occurs many times in its scope, but if it doesn't
786 we'll catch it next time round. At worst this costs an extra simplifier pass.
787 ToDo: try using the occurrence info for the inline'd binder.
789 [March 97] We do the same for atomic RHSs. Reason: see notes with reOrderRec.
790 [June 98, SLPJ] I've undone this change; I don't understand it. See notes with reOrderRec.
795 -> Maybe Id -> CoreExpr -- Binder and rhs
796 -- Just b => non-rec, and alrady tagged with occurrence info
797 -- Nothing => Rec, no occ info
798 -> (UsageDetails, CoreExpr)
799 -- Returned usage details covers only the RHS,
800 -- and *not* the RULE or INLINE template for the Id
801 occAnalRhs env mb_bndr rhs
804 -- See Note [Cascading inlines]
805 ctxt = case mb_bndr of
806 Just b | certainly_inline b -> env
807 _other -> rhsCtxt env
809 certainly_inline bndr -- See Note [Cascading inlines]
810 = case idOccInfo bndr of
811 OneOcc in_lam one_br _ -> not in_lam && one_br && active && not_stable
814 active = isAlwaysActive (idInlineActivation bndr)
815 not_stable = not (isStableUnfolding (idUnfolding bndr))
817 addIdOccs :: UsageDetails -> VarSet -> UsageDetails
818 addIdOccs usage id_set = foldVarSet add usage id_set
820 add v u | isId v = addOneOcc u v NoOccInfo
822 -- Give a non-committal binder info (i.e NoOccInfo) because
823 -- a) Many copies of the specialised thing can appear
824 -- b) We don't want to substitute a BIG expression inside a RULE
825 -- even if that's the only occurrence of the thing
826 -- (Same goes for INLINE.)
829 Note [Cascading inlines]
830 ~~~~~~~~~~~~~~~~~~~~~~~~
831 By default we use an rhsCtxt for the RHS of a binding. This tells the
832 occ anal n that it's looking at an RHS, which has an effect in
833 occAnalApp. In particular, for constructor applications, it makes
834 the arguments appear to have NoOccInfo, so that we don't inline into
837 we do not want to inline x.
839 But there's a problem. Consider
844 First time round, it looks as if x1 and x2 occur as an arg of a
845 let-bound constructor ==> give them a many-occurrence.
846 But then x3 is inlined (unconditionally as it happens) and
847 next time round, x2 will be, and the next time round x1 will be
848 Result: multiple simplifier iterations. Sigh.
850 So, when analysing the RHS of x3 we notice that x3 will itself
851 definitely inline the next time round, and so we analyse x3's rhs in
852 an ordinary context, not rhsCtxt. Hence the "certainly_inline" stuff.
854 Annoyingly, we have to approximiate SimplUtils.preInlineUnconditionally.
855 If we say "yes" when preInlineUnconditionally says "no" the simplifier iterates
865 This is worse than the slow cascade, so we only want to say "certainly_inline"
866 if it really is certain. Look at the note with preInlineUnconditionally
867 for the various clauses.
874 -> (UsageDetails, -- Gives info only about the "interesting" Ids
877 occAnal _ expr@(Type _) = (emptyDetails, expr)
878 occAnal _ expr@(Lit _) = (emptyDetails, expr)
879 occAnal env expr@(Var v) = (mkOneOcc env v False, expr)
880 -- At one stage, I gathered the idRuleVars for v here too,
881 -- which in a way is the right thing to do.
882 -- But that went wrong right after specialisation, when
883 -- the *occurrences* of the overloaded function didn't have any
884 -- rules in them, so the *specialised* versions looked as if they
885 -- weren't used at all.
887 occAnal _ (Coercion co)
888 = (addIdOccs emptyDetails (coVarsOfCo co), Coercion co)
889 -- See Note [Gather occurrences of coercion veriables]
892 Note [Gather occurrences of coercion veriables]
893 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
894 We need to gather info about what coercion variables appear, so that
895 we can sort them into the right place when doing dependency analysis.
901 occAnal env (Note note@(SCC _) body)
902 = case occAnal env body of { (usage, body') ->
903 (mapVarEnv markInsideSCC usage, Note note body')
906 occAnal env (Note note body)
907 = case occAnal env body of { (usage, body') ->
908 (usage, Note note body')
911 occAnal env (Cast expr co)
912 = case occAnal env expr of { (usage, expr') ->
913 let usage1 = markManyIf (isRhsEnv env) usage
914 usage2 = addIdOccs usage1 (coVarsOfCo co)
915 -- See Note [Gather occurrences of coercion veriables]
916 in (usage2, Cast expr' co)
917 -- If we see let x = y `cast` co
918 -- then mark y as 'Many' so that we don't
919 -- immediately inline y again.
924 occAnal env app@(App _ _)
925 = occAnalApp env (collectArgs app)
927 -- Ignore type variables altogether
928 -- (a) occurrences inside type lambdas only not marked as InsideLam
929 -- (b) type variables not in environment
931 occAnal env (Lam x body) | isTyVar x
932 = case occAnal env body of { (body_usage, body') ->
933 (body_usage, Lam x body')
936 -- For value lambdas we do a special hack. Consider
938 -- If we did nothing, x is used inside the \y, so would be marked
939 -- as dangerous to dup. But in the common case where the abstraction
940 -- is applied to two arguments this is over-pessimistic.
941 -- So instead, we just mark each binder with its occurrence
942 -- info in the *body* of the multiple lambda.
943 -- Then, the simplifier is careful when partially applying lambdas.
945 occAnal env expr@(Lam _ _)
946 = case occAnal env_body body of { (body_usage, body') ->
948 (final_usage, tagged_binders) = tagLamBinders body_usage binders'
949 -- Use binders' to put one-shot info on the lambdas
951 -- URGH! Sept 99: we don't seem to be able to use binders' here, because
952 -- we get linear-typed things in the resulting program that we can't handle yet.
953 -- (e.g. PrelShow) TODO
955 really_final_usage = if linear then
958 mapVarEnv markInsideLam final_usage
961 mkLams tagged_binders body') }
963 env_body = vanillaCtxt (trimOccEnv env binders)
964 -- Body is (no longer) an RhsContext
965 (binders, body) = collectBinders expr
966 binders' = oneShotGroup env binders
967 linear = all is_one_shot binders'
968 is_one_shot b = isId b && isOneShotBndr b
970 occAnal env (Case scrut bndr ty alts)
971 = case occ_anal_scrut scrut alts of { (scrut_usage, scrut') ->
972 case mapAndUnzip occ_anal_alt alts of { (alts_usage_s, alts') ->
974 alts_usage = foldr1 combineAltsUsageDetails alts_usage_s
975 (alts_usage1, tagged_bndr) = tag_case_bndr alts_usage bndr
976 total_usage = scrut_usage +++ alts_usage1
978 total_usage `seq` (total_usage, Case scrut' tagged_bndr ty alts') }}
980 -- Note [Case binder usage]
981 -- ~~~~~~~~~~~~~~~~~~~~~~~~
982 -- The case binder gets a usage of either "many" or "dead", never "one".
983 -- Reason: we like to inline single occurrences, to eliminate a binding,
984 -- but inlining a case binder *doesn't* eliminate a binding.
985 -- We *don't* want to transform
986 -- case x of w { (p,q) -> f w }
988 -- case x of w { (p,q) -> f (p,q) }
989 tag_case_bndr usage bndr
990 = case lookupVarEnv usage bndr of
991 Nothing -> (usage, setIdOccInfo bndr IAmDead)
992 Just _ -> (usage `delVarEnv` bndr, setIdOccInfo bndr NoOccInfo)
994 alt_env = mkAltEnv env scrut bndr
995 occ_anal_alt = occAnalAlt alt_env bndr
997 occ_anal_scrut (Var v) (alt1 : other_alts)
998 | not (null other_alts) || not (isDefaultAlt alt1)
999 = (mkOneOcc env v True, Var v) -- The 'True' says that the variable occurs
1000 -- in an interesting context; the case has
1001 -- at least one non-default alternative
1002 occ_anal_scrut scrut _alts
1003 = occAnal (vanillaCtxt env) scrut -- No need for rhsCtxt
1005 occAnal env (Let bind body)
1006 = case occAnal env_body body of { (body_usage, body') ->
1007 case occAnalBind env env_body bind body_usage of { (final_usage, new_binds) ->
1008 (final_usage, mkLets new_binds body') }}
1010 env_body = trimOccEnv env (bindersOf bind)
1012 occAnalArgs :: OccEnv -> [CoreExpr] -> (UsageDetails, [CoreExpr])
1013 occAnalArgs env args
1014 = case mapAndUnzip (occAnal arg_env) args of { (arg_uds_s, args') ->
1015 (foldr (+++) emptyDetails arg_uds_s, args')}
1017 arg_env = vanillaCtxt env
1020 Applications are dealt with specially because we want
1021 the "build hack" to work.
1023 Note [Arguments of let-bound constructors]
1024 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
1026 f x = let y = expensive x in
1028 (case z of {(p,q)->q}, case z of {(p,q)->q})
1029 We feel free to duplicate the WHNF (True,y), but that means
1030 that y may be duplicated thereby.
1032 If we aren't careful we duplicate the (expensive x) call!
1033 Constructors are rather like lambdas in this way.
1036 occAnalApp :: OccEnv
1037 -> (Expr CoreBndr, [Arg CoreBndr])
1038 -> (UsageDetails, Expr CoreBndr)
1039 occAnalApp env (Var fun, args)
1040 = case args_stuff of { (args_uds, args') ->
1042 final_args_uds = markManyIf (isRhsEnv env && is_exp) args_uds
1043 -- We mark the free vars of the argument of a constructor or PAP
1044 -- as "many", if it is the RHS of a let(rec).
1045 -- This means that nothing gets inlined into a constructor argument
1046 -- position, which is what we want. Typically those constructor
1047 -- arguments are just variables, or trivial expressions.
1049 -- This is the *whole point* of the isRhsEnv predicate
1050 -- See Note [Arguments of let-bound constructors]
1052 (fun_uds +++ final_args_uds, mkApps (Var fun) args') }
1054 fun_uniq = idUnique fun
1055 fun_uds = mkOneOcc env fun (valArgCount args > 0)
1056 is_exp = isExpandableApp fun (valArgCount args)
1057 -- See Note [CONLIKE pragma] in BasicTypes
1058 -- The definition of is_exp should match that in
1059 -- Simplify.prepareRhs
1061 -- Hack for build, fold, runST
1062 args_stuff | fun_uniq == buildIdKey = appSpecial env 2 [True,True] args
1063 | fun_uniq == augmentIdKey = appSpecial env 2 [True,True] args
1064 | fun_uniq == foldrIdKey = appSpecial env 3 [False,True] args
1065 | fun_uniq == runSTRepIdKey = appSpecial env 2 [True] args
1066 -- (foldr k z xs) may call k many times, but it never
1067 -- shares a partial application of k; hence [False,True]
1068 -- This means we can optimise
1069 -- foldr (\x -> let v = ...x... in \y -> ...v...) z xs
1070 -- by floating in the v
1072 | otherwise = occAnalArgs env args
1075 occAnalApp env (fun, args)
1076 = case occAnal (addAppCtxt env args) fun of { (fun_uds, fun') ->
1077 -- The addAppCtxt is a bit cunning. One iteration of the simplifier
1078 -- often leaves behind beta redexs like
1079 -- (\x y -> e) a1 a2
1080 -- Here we would like to mark x,y as one-shot, and treat the whole
1081 -- thing much like a let. We do this by pushing some True items
1082 -- onto the context stack.
1084 case occAnalArgs env args of { (args_uds, args') ->
1086 final_uds = fun_uds +++ args_uds
1088 (final_uds, mkApps fun' args') }}
1091 markManyIf :: Bool -- If this is true
1092 -> UsageDetails -- Then do markMany on this
1094 markManyIf True uds = mapVarEnv markMany uds
1095 markManyIf False uds = uds
1097 appSpecial :: OccEnv
1098 -> Int -> CtxtTy -- Argument number, and context to use for it
1100 -> (UsageDetails, [CoreExpr])
1101 appSpecial env n ctxt args
1104 arg_env = vanillaCtxt env
1106 go _ [] = (emptyDetails, []) -- Too few args
1108 go 1 (arg:args) -- The magic arg
1109 = case occAnal (setCtxtTy arg_env ctxt) arg of { (arg_uds, arg') ->
1110 case occAnalArgs env args of { (args_uds, args') ->
1111 (arg_uds +++ args_uds, arg':args') }}
1114 = case occAnal arg_env arg of { (arg_uds, arg') ->
1115 case go (n-1) args of { (args_uds, args') ->
1116 (arg_uds +++ args_uds, arg':args') }}
1120 Note [Binders in case alternatives]
1121 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
1123 case x of y { (a,b) -> f y }
1124 We treat 'a', 'b' as dead, because they don't physically occur in the
1125 case alternative. (Indeed, a variable is dead iff it doesn't occur in
1126 its scope in the output of OccAnal.) It really helps to know when
1127 binders are unused. See esp the call to isDeadBinder in
1128 Simplify.mkDupableAlt
1130 In this example, though, the Simplifier will bring 'a' and 'b' back to
1131 life, beause it binds 'y' to (a,b) (imagine got inlined and
1135 occAnalAlt :: OccEnv
1138 -> (UsageDetails, Alt IdWithOccInfo)
1139 occAnalAlt env case_bndr (con, bndrs, rhs)
1141 env' = trimOccEnv env bndrs
1143 case occAnal env' rhs of { (rhs_usage1, rhs1) ->
1145 proxies = getProxies env' case_bndr
1146 (rhs_usage2, rhs2) = foldrBag wrapProxy (rhs_usage1, rhs1) proxies
1147 (alt_usg, tagged_bndrs) = tagLamBinders rhs_usage2 bndrs
1148 bndrs' = tagged_bndrs -- See Note [Binders in case alternatives]
1150 (alt_usg, (con, bndrs', rhs2)) }
1152 wrapProxy :: ProxyBind -> (UsageDetails, CoreExpr) -> (UsageDetails, CoreExpr)
1153 wrapProxy (bndr, rhs_var, co) (body_usg, body)
1154 | not (bndr `usedIn` body_usg)
1157 = (body_usg' +++ rhs_usg, Let (NonRec tagged_bndr rhs) body)
1159 (body_usg', tagged_bndr) = tagBinder body_usg bndr
1160 rhs_usg = unitVarEnv rhs_var NoOccInfo -- We don't need exact info
1161 rhs = mkCoerce co (Var (zapIdOccInfo rhs_var)) -- See Note [Zap case binders in proxy bindings]
1165 %************************************************************************
1169 %************************************************************************
1173 = OccEnv { occ_encl :: !OccEncl -- Enclosing context information
1174 , occ_ctxt :: !CtxtTy -- Tells about linearity
1175 , occ_proxy :: ProxyEnv
1176 , occ_rule_fvs :: ImpRuleUsage
1177 , occ_rule_act :: Maybe (Activation -> Bool) -- Nothing => Rules are inactive
1178 -- See Note [Finding rule RHS free vars]
1182 -----------------------------
1183 -- OccEncl is used to control whether to inline into constructor arguments
1185 -- x = (p,q) -- Don't inline p or q
1186 -- y = /\a -> (p a, q a) -- Still don't inline p or q
1187 -- z = f (p,q) -- Do inline p,q; it may make a rule fire
1188 -- So OccEncl tells enought about the context to know what to do when
1189 -- we encounter a contructor application or PAP.
1192 = OccRhs -- RHS of let(rec), albeit perhaps inside a type lambda
1193 -- Don't inline into constructor args here
1194 | OccVanilla -- Argument of function, body of lambda, scruintee of case etc.
1195 -- Do inline into constructor args here
1197 instance Outputable OccEncl where
1198 ppr OccRhs = ptext (sLit "occRhs")
1199 ppr OccVanilla = ptext (sLit "occVanilla")
1201 type CtxtTy = [Bool]
1204 -- True:ctxt Analysing a function-valued expression that will be
1205 -- applied just once
1207 -- False:ctxt Analysing a function-valued expression that may
1208 -- be applied many times; but when it is,
1209 -- the CtxtTy inside applies
1211 initOccEnv :: Maybe (Activation -> Bool) -> [CoreRule]
1213 initOccEnv active_rule imp_rules
1214 = OccEnv { occ_encl = OccVanilla
1216 , occ_proxy = PE emptyVarEnv emptyVarSet
1217 , occ_rule_fvs = findImpRuleUsage active_rule imp_rules
1218 , occ_rule_act = active_rule }
1220 vanillaCtxt :: OccEnv -> OccEnv
1221 vanillaCtxt env = env { occ_encl = OccVanilla, occ_ctxt = [] }
1223 rhsCtxt :: OccEnv -> OccEnv
1224 rhsCtxt env = env { occ_encl = OccRhs, occ_ctxt = [] }
1226 setCtxtTy :: OccEnv -> CtxtTy -> OccEnv
1227 setCtxtTy env ctxt = env { occ_ctxt = ctxt }
1229 isRhsEnv :: OccEnv -> Bool
1230 isRhsEnv (OccEnv { occ_encl = OccRhs }) = True
1231 isRhsEnv (OccEnv { occ_encl = OccVanilla }) = False
1233 oneShotGroup :: OccEnv -> [CoreBndr] -> [CoreBndr]
1234 -- The result binders have one-shot-ness set that they might not have had originally.
1235 -- This happens in (build (\cn -> e)). Here the occurrence analyser
1236 -- linearity context knows that c,n are one-shot, and it records that fact in
1237 -- the binder. This is useful to guide subsequent float-in/float-out tranformations
1239 oneShotGroup (OccEnv { occ_ctxt = ctxt }) bndrs
1242 go _ [] rev_bndrs = reverse rev_bndrs
1244 go (lin_ctxt:ctxt) (bndr:bndrs) rev_bndrs
1245 | isId bndr = go ctxt bndrs (bndr':rev_bndrs)
1247 bndr' | lin_ctxt = setOneShotLambda bndr
1250 go ctxt (bndr:bndrs) rev_bndrs = go ctxt bndrs (bndr:rev_bndrs)
1252 addAppCtxt :: OccEnv -> [Arg CoreBndr] -> OccEnv
1253 addAppCtxt env@(OccEnv { occ_ctxt = ctxt }) args
1254 = env { occ_ctxt = replicate (valArgCount args) True ++ ctxt }
1257 %************************************************************************
1261 %************************************************************************
1264 type ImpRuleUsage = NameEnv UsageDetails
1265 -- Maps an *imported* Id f to the UsageDetails for *local* Ids
1266 -- used on the RHS for a *local* rule for f.
1271 Consider this, where A.g is an imported Id
1274 {-# RULE "foo" forall x. A.g x = f x #-}
1276 Obviously there's a loop, but the danger is that the occurrence analyser
1277 will say that 'f' is not a loop breaker. Then the simplifier will
1280 and then gaily inline 'f'. Result infinite loop. More realistically,
1281 these kind of rules are generated when specialising imported INLINABLE Ids.
1283 Solution: treat an occurrence of A.g as an occurrence of all the local Ids
1284 that occur on the RULE's RHS. This mapping from imported Id to local Ids
1285 is held in occ_rule_fvs.
1288 findImpRuleUsage :: Maybe (Activation -> Bool) -> [CoreRule] -> ImpRuleUsage
1289 -- Find the *local* Ids that can be reached transitively,
1290 -- via local rules, from each *imported* Id.
1291 -- Sigh: this function seems more complicated than it is really worth
1292 findImpRuleUsage Nothing _ = emptyNameEnv
1293 findImpRuleUsage (Just is_active) rules
1294 = mkNameEnv [ (f, mapUFM (\_ -> NoOccInfo) ls)
1296 , let ls = find_lcl_deps f
1297 , not (isEmptyVarSet ls) ]
1299 rule_names = map ru_fn rules
1300 rule_name_set = mkNameSet rule_names
1302 imp_deps :: NameEnv VarSet
1303 -- (f,g) means imported Id 'g' appears in RHS of
1304 -- rule for imported Id 'f', *or* does so transitively
1305 imp_deps = foldr add_imp emptyNameEnv rules
1307 | is_active (ruleActivation rule)
1308 = extendNameEnv_C unionVarSet acc (ru_fn rule)
1309 (exprSomeFreeVars keep_imp (ru_rhs rule))
1311 keep_imp v = isId v && (idName v `elemNameSet` rule_name_set)
1312 full_imp_deps = transClosureFV (ufmToList imp_deps)
1314 lcl_deps :: NameEnv VarSet
1315 -- (f, l) means localId 'l' appears immediately
1316 -- in the RHS of a rule for imported Id 'f'
1317 -- Remember, many rules might have the same ru_fn
1318 -- so we do need to fold
1319 lcl_deps = foldr add_lcl emptyNameEnv rules
1320 add_lcl rule acc = extendNameEnv_C unionVarSet acc (ru_fn rule)
1321 (exprFreeIds (ru_rhs rule))
1323 find_lcl_deps :: Name -> VarSet
1325 = foldVarSet (unionVarSet . lookup_lcl . idName) (lookup_lcl f)
1326 (lookupNameEnv full_imp_deps f `orElse` emptyVarSet)
1327 lookup_lcl :: Name -> VarSet
1328 lookup_lcl g = lookupNameEnv lcl_deps g `orElse` emptyVarSet
1331 transClosureFV :: Uniquable a => [(a, VarSet)] -> UniqFM VarSet
1332 -- If (f,g), (g,h) are in the input, then (f,h) is in the output
1333 transClosureFV fv_list
1335 | otherwise = transClosureFV new_fv_list
1337 env = listToUFM fv_list
1338 (no_change, new_fv_list) = mapAccumL bump True fv_list
1339 bump no_change (b,fvs)
1340 | no_change_here = (no_change, (b,fvs))
1341 | otherwise = (False, (b,new_fvs))
1343 (new_fvs, no_change_here) = extendFvs env fvs
1346 extendFvs :: UniqFM VarSet -> VarSet -> (VarSet, Bool)
1347 -- (extendFVs env s) returns
1348 -- (s `union` env(s), env(s) `subset` s)
1350 = foldVarSet add (s, True) s
1352 add v (vs, no_change_so_far)
1353 = case lookupUFM env v of
1354 Just fvs | not (fvs `subVarSet` s)
1355 -> (vs `unionVarSet` fvs, False)
1356 _ -> (vs, no_change_so_far)
1360 %************************************************************************
1364 %************************************************************************
1367 data ProxyEnv -- See Note [ProxyEnv]
1368 = PE (IdEnv -- Domain = scrutinee variables
1369 (Id, -- The scrutinee variable again
1370 [(Id,Coercion)])) -- The case binders that it maps to
1371 VarSet -- Free variables of both range and domain
1376 The ProxyEnv keeps track of the connection between case binders and
1377 scrutinee. Specifically, if
1378 sc |-> (sc, [...(cb, co)...])
1379 is a binding in the ProxyEnv, then
1381 Typically we add such a binding when encountering the case expression
1382 case (sc |> coi) of cb { ... }
1385 * The domain of the ProxyEnv is the variable (or casted variable)
1386 scrutinees of enclosing cases. This is additionally used
1387 to ensure we gather occurrence info even for GlobalId scrutinees;
1388 see Note [Binder swap for GlobalId scrutinee]
1390 * The ProxyEnv is just an optimisation; you can throw away any
1391 element without losing correctness. And we do so when pushing
1392 it inside a binding (see trimProxyEnv).
1394 * One scrutinee might map to many case binders: Eg
1395 case sc of cb1 { DEFAULT -> ....case sc of cb2 { ... } .. }
1398 * If sc1 |-> (sc2, [...(cb, co)...]), then sc1==sc2
1399 It's a UniqFM and we sometimes need the domain Id
1401 * Any particular case binder 'cb' occurs only once in entire range
1405 The Main Reason for having a ProxyEnv is so that when we encounter
1406 case e of cb { pi -> ri }
1407 we can find all the in-scope variables derivable from 'cb',
1408 and effectively add let-bindings for them (or at least for the
1409 ones *mentioned* in ri) thus:
1410 case e of cb { pi -> let { x = ..cb..; y = ...cb.. }
1412 In this way we'll replace occurrences of 'x', 'y' with 'cb',
1413 which implements the Binder-swap idea (see Note [Binder swap])
1415 The function getProxies finds these bindings; then we
1416 add just the necessary ones, using wrapProxy.
1420 We do these two transformations right here:
1422 (1) case x of b { pi -> ri }
1424 case x of b { pi -> let x=b in ri }
1426 (2) case (x |> co) of b { pi -> ri }
1428 case (x |> co) of b { pi -> let x = b |> sym co in ri }
1430 Why (2)? See Note [Case of cast]
1432 In both cases, in a particular alternative (pi -> ri), we only
1434 (a) x occurs free in (pi -> ri)
1435 (ie it occurs in ri, but is not bound in pi)
1436 (b) the pi does not bind b (or the free vars of co)
1437 We need (a) and (b) for the inserted binding to be correct.
1439 For the alternatives where we inject the binding, we can transfer
1440 all x's OccInfo to b. And that is the point.
1443 * The deliberate shadowing of 'x'.
1444 * That (a) rapidly becomes false, so no bindings are injected.
1446 The reason for doing these transformations here is because it allows
1447 us to adjust the OccInfo for 'x' and 'b' as we go.
1449 * Suppose the only occurrences of 'x' are the scrutinee and in the
1450 ri; then this transformation makes it occur just once, and hence
1451 get inlined right away.
1453 * If we do this in the Simplifier, we don't know whether 'x' is used
1454 in ri, so we are forced to pessimistically zap b's OccInfo even
1455 though it is typically dead (ie neither it nor x appear in the
1456 ri). There's nothing actually wrong with zapping it, except that
1457 it's kind of nice to know which variables are dead. My nose
1458 tells me to keep this information as robustly as possible.
1460 The Maybe (Id,CoreExpr) passed to occAnalAlt is the extra let-binding
1461 {x=b}; it's Nothing if the binder-swap doesn't happen.
1463 There is a danger though. Consider
1465 in case (f v) of w -> ...v...v...
1466 And suppose that (f v) expands to just v. Then we'd like to
1467 use 'w' instead of 'v' in the alternative. But it may be too
1468 late; we may have substituted the (cheap) x+#y for v in the
1469 same simplifier pass that reduced (f v) to v.
1471 I think this is just too bad. CSE will recover some of it.
1475 Consider case (x `cast` co) of b { I# ->
1476 ... (case (x `cast` co) of {...}) ...
1477 We'd like to eliminate the inner case. That is the motivation for
1478 equation (2) in Note [Binder swap]. When we get to the inner case, we
1479 inline x, cancel the casts, and away we go.
1481 Note [Binder swap on GlobalId scrutinees]
1482 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
1483 When the scrutinee is a GlobalId we must take care in two ways
1485 i) In order to *know* whether 'x' occurs free in the RHS, we need its
1486 occurrence info. BUT, we don't gather occurrence info for
1487 GlobalIds. That's one use for the (small) occ_proxy env in OccEnv is
1488 for: it says "gather occurrence info for these.
1490 ii) We must call localiseId on 'x' first, in case it's a GlobalId, or
1491 has an External Name. See, for example, SimplEnv Note [Global Ids in
1494 Note [getProxies is subtle]
1495 ~~~~~~~~~~~~~~~~~~~~~~~~~~~
1496 The code for getProxies isn't all that obvious. Consider
1498 case v |> cov of x { DEFAULT ->
1499 case x |> cox1 of y { DEFAULT ->
1500 case x |> cox2 of z { DEFAULT -> r
1502 These will give us a ProxyEnv looking like:
1503 x |-> (x, [(y, cox1), (z, cox2)])
1504 v |-> (v, [(x, cov)])
1506 From this we want to extract the bindings
1511 Notice that later bindings may mention earlier ones, and that
1512 we need to go "both ways".
1514 Note [Zap case binders in proxy bindings]
1515 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
1517 case x of cb(dead) { p -> ...x... }
1519 case x of cb(live) { p -> let x = cb in ...x... }
1521 Core Lint never expects to find an *occurence* of an Id marked
1522 as Dead, so we must zap the OccInfo on cb before making the
1523 binding x = cb. See Trac #5028.
1525 Historical note [no-case-of-case]
1526 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
1527 We *used* to suppress the binder-swap in case expressions when
1528 -fno-case-of-case is on. Old remarks:
1529 "This happens in the first simplifier pass,
1530 and enhances full laziness. Here's the bad case:
1531 f = \ y -> ...(case x of I# v -> ...(case x of ...) ... )
1532 If we eliminate the inner case, we trap it inside the I# v -> arm,
1533 which might prevent some full laziness happening. I've seen this
1534 in action in spectral/cichelli/Prog.hs:
1535 [(m,n) | m <- [1..max], n <- [1..max]]
1536 Hence the check for NoCaseOfCase."
1537 However, now the full-laziness pass itself reverses the binder-swap, so this
1538 check is no longer necessary.
1540 Historical note [Suppressing the case binder-swap]
1541 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
1542 This old note describes a problem that is also fixed by doing the
1543 binder-swap in OccAnal:
1545 There is another situation when it might make sense to suppress the
1546 case-expression binde-swap. If we have
1548 case x of w1 { DEFAULT -> case x of w2 { A -> e1; B -> e2 }
1549 ...other cases .... }
1551 We'll perform the binder-swap for the outer case, giving
1553 case x of w1 { DEFAULT -> case w1 of w2 { A -> e1; B -> e2 }
1554 ...other cases .... }
1556 But there is no point in doing it for the inner case, because w1 can't
1557 be inlined anyway. Furthermore, doing the case-swapping involves
1558 zapping w2's occurrence info (see paragraphs that follow), and that
1559 forces us to bind w2 when doing case merging. So we get
1561 case x of w1 { A -> let w2 = w1 in e1
1562 B -> let w2 = w1 in e2
1563 ...other cases .... }
1565 This is plain silly in the common case where w2 is dead.
1567 Even so, I can't see a good way to implement this idea. I tried
1568 not doing the binder-swap if the scrutinee was already evaluated
1569 but that failed big-time:
1573 case v of w { MkT x ->
1574 case x of x1 { I# y1 ->
1575 case x of x2 { I# y2 -> ...
1577 Notice that because MkT is strict, x is marked "evaluated". But to
1578 eliminate the last case, we must either make sure that x (as well as
1579 x1) has unfolding MkT y1. THe straightforward thing to do is to do
1580 the binder-swap. So this whole note is a no-op.
1582 It's fixed by doing the binder-swap in OccAnal because we can do the
1583 binder-swap unconditionally and still get occurrence analysis
1587 extendProxyEnv :: ProxyEnv -> Id -> Coercion -> Id -> ProxyEnv
1588 -- (extendPE x co y) typically arises from
1589 -- case (x |> co) of y { ... }
1590 -- It extends the proxy env with the binding
1592 extendProxyEnv pe scrut co case_bndr
1593 | scrut == case_bndr = PE env1 fvs1 -- If case_bndr shadows scrut,
1594 | otherwise = PE env2 fvs2 -- don't extend
1596 PE env1 fvs1 = trimProxyEnv pe [case_bndr]
1597 env2 = extendVarEnv_Acc add single env1 scrut1 (case_bndr,co)
1598 single cb_co = (scrut1, [cb_co])
1599 add cb_co (x, cb_cos) = (x, cb_co:cb_cos)
1600 fvs2 = fvs1 `unionVarSet` tyCoVarsOfCo co
1601 `extendVarSet` case_bndr
1602 `extendVarSet` scrut1
1604 scrut1 = mkLocalId (localiseName (idName scrut)) (idType scrut)
1605 -- Localise the scrut_var before shadowing it; we're making a
1606 -- new binding for it, and it might have an External Name, or
1607 -- even be a GlobalId; Note [Binder swap on GlobalId scrutinees]
1608 -- Also we don't want any INLINE or NOINLINE pragmas!
1611 type ProxyBind = (Id, Id, Coercion)
1612 -- (scrut variable, case-binder variable, coercion)
1614 getProxies :: OccEnv -> Id -> Bag ProxyBind
1615 -- Return a bunch of bindings [...(xi,ei)...]
1616 -- such that let { ...; xi=ei; ... } binds the xi using y alone
1617 -- See Note [getProxies is subtle]
1618 getProxies (OccEnv { occ_proxy = PE pe _ }) case_bndr
1619 = -- pprTrace "wrapProxies" (ppr case_bndr) $
1622 fwd_pe :: IdEnv (Id, Coercion)
1623 fwd_pe = foldVarEnv add1 emptyVarEnv pe
1625 add1 (x,ycos) env = foldr (add2 x) env ycos
1626 add2 x (y,co) env = extendVarEnv env y (x,co)
1628 go_fwd :: Id -> Bag ProxyBind
1629 -- Return bindings derivable from case_bndr
1630 go_fwd case_bndr = -- pprTrace "go_fwd" (vcat [ppr case_bndr, text "fwd_pe =" <+> ppr fwd_pe,
1631 -- text "pe =" <+> ppr pe]) $
1635 | Just (scrut, co) <- lookupVarEnv fwd_pe case_bndr
1636 = unitBag (scrut, case_bndr, mkSymCo co)
1637 `unionBags` go_fwd scrut
1638 `unionBags` go_bwd scrut [pr | pr@(cb,_) <- lookup_bwd scrut
1643 lookup_bwd :: Id -> [(Id, Coercion)]
1644 -- Return case_bndrs that are connected to scrut
1645 lookup_bwd scrut = case lookupVarEnv pe scrut of
1647 Just (_, cb_cos) -> cb_cos
1649 go_bwd :: Id -> [(Id, Coercion)] -> Bag ProxyBind
1650 go_bwd scrut cb_cos = foldr (unionBags . go_bwd1 scrut) emptyBag cb_cos
1652 go_bwd1 :: Id -> (Id, Coercion) -> Bag ProxyBind
1653 go_bwd1 scrut (case_bndr, co)
1654 = -- pprTrace "go_bwd1" (ppr case_bndr) $
1655 unitBag (case_bndr, scrut, co)
1656 `unionBags` go_bwd case_bndr (lookup_bwd case_bndr)
1659 mkAltEnv :: OccEnv -> CoreExpr -> Id -> OccEnv
1660 -- Does two things: a) makes the occ_ctxt = OccVanilla
1661 -- b) extends the ProxyEnv if possible
1662 mkAltEnv env scrut cb
1663 = env { occ_encl = OccVanilla, occ_proxy = pe' }
1667 Var v -> extendProxyEnv pe v (mkReflCo (idType v)) cb
1668 Cast (Var v) co -> extendProxyEnv pe v co cb
1669 _other -> trimProxyEnv pe [cb]
1672 trimOccEnv :: OccEnv -> [CoreBndr] -> OccEnv
1673 trimOccEnv env bndrs = env { occ_proxy = trimProxyEnv (occ_proxy env) bndrs }
1676 trimProxyEnv :: ProxyEnv -> [CoreBndr] -> ProxyEnv
1677 -- We are about to push this ProxyEnv inside a binding for 'bndrs'
1678 -- So dump any ProxyEnv bindings which mention any of the bndrs
1679 trimProxyEnv (PE pe fvs) bndrs
1680 | not (bndr_set `intersectsVarSet` fvs)
1683 = PE pe' (fvs `minusVarSet` bndr_set)
1685 pe' = mapVarEnv trim pe
1686 bndr_set = mkVarSet bndrs
1687 trim (scrut, cb_cos) | scrut `elemVarSet` bndr_set = (scrut, [])
1688 | otherwise = (scrut, filterOut discard cb_cos)
1689 discard (cb,co) = bndr_set `intersectsVarSet`
1690 extendVarSet (tyCoVarsOfCo co) cb
1694 %************************************************************************
1696 \subsection[OccurAnal-types]{OccEnv}
1698 %************************************************************************
1701 type UsageDetails = IdEnv OccInfo -- A finite map from ids to their usage
1702 -- INVARIANT: never IAmDead
1703 -- (Deadness is signalled by not being in the map at all)
1705 (+++), combineAltsUsageDetails
1706 :: UsageDetails -> UsageDetails -> UsageDetails
1709 = plusVarEnv_C addOccInfo usage1 usage2
1711 combineAltsUsageDetails usage1 usage2
1712 = plusVarEnv_C orOccInfo usage1 usage2
1714 addOneOcc :: UsageDetails -> Id -> OccInfo -> UsageDetails
1715 addOneOcc usage id info
1716 = plusVarEnv_C addOccInfo usage (unitVarEnv id info)
1717 -- ToDo: make this more efficient
1719 emptyDetails :: UsageDetails
1720 emptyDetails = (emptyVarEnv :: UsageDetails)
1722 usedIn :: Id -> UsageDetails -> Bool
1723 v `usedIn` details = isExportedId v || v `elemVarEnv` details
1725 type IdWithOccInfo = Id
1727 tagLamBinders :: UsageDetails -- Of scope
1729 -> (UsageDetails, -- Details with binders removed
1730 [IdWithOccInfo]) -- Tagged binders
1731 -- Used for lambda and case binders
1732 -- It copes with the fact that lambda bindings can have InlineRule
1733 -- unfoldings, used for join points
1734 tagLamBinders usage binders = usage' `seq` (usage', bndrs')
1736 (usage', bndrs') = mapAccumR tag_lam usage binders
1737 tag_lam usage bndr = (usage2, setBinderOcc usage bndr)
1739 usage1 = usage `delVarEnv` bndr
1740 usage2 | isId bndr = addIdOccs usage1 (idUnfoldingVars bndr)
1741 | otherwise = usage1
1743 tagBinder :: UsageDetails -- Of scope
1745 -> (UsageDetails, -- Details with binders removed
1746 IdWithOccInfo) -- Tagged binders
1748 tagBinder usage binder
1750 usage' = usage `delVarEnv` binder
1751 binder' = setBinderOcc usage binder
1753 usage' `seq` (usage', binder')
1755 setBinderOcc :: UsageDetails -> CoreBndr -> CoreBndr
1756 setBinderOcc usage bndr
1757 | isTyVar bndr = bndr
1758 | isExportedId bndr = case idOccInfo bndr of
1760 _ -> setIdOccInfo bndr NoOccInfo
1761 -- Don't use local usage info for visible-elsewhere things
1762 -- BUT *do* erase any IAmALoopBreaker annotation, because we're
1763 -- about to re-generate it and it shouldn't be "sticky"
1765 | otherwise = setIdOccInfo bndr occ_info
1767 occ_info = lookupVarEnv usage bndr `orElse` IAmDead
1771 %************************************************************************
1773 \subsection{Operations over OccInfo}
1775 %************************************************************************
1778 mkOneOcc :: OccEnv -> Id -> InterestingCxt -> UsageDetails
1779 mkOneOcc env id int_cxt
1780 | isLocalId id = unitVarEnv id (OneOcc False True int_cxt)
1781 | PE env _ <- occ_proxy env
1782 , id `elemVarEnv` env = unitVarEnv id NoOccInfo
1783 | Just uds <- lookupNameEnv (occ_rule_fvs env) (idName id)
1785 | otherwise = emptyDetails
1787 markMany, markInsideLam, markInsideSCC :: OccInfo -> OccInfo
1789 markMany _ = NoOccInfo
1791 markInsideSCC occ = markMany occ
1793 markInsideLam (OneOcc _ one_br int_cxt) = OneOcc True one_br int_cxt
1794 markInsideLam occ = occ
1796 addOccInfo, orOccInfo :: OccInfo -> OccInfo -> OccInfo
1798 addOccInfo a1 a2 = ASSERT( not (isDeadOcc a1 || isDeadOcc a2) )
1799 NoOccInfo -- Both branches are at least One
1800 -- (Argument is never IAmDead)
1802 -- (orOccInfo orig new) is used
1803 -- when combining occurrence info from branches of a case
1805 orOccInfo (OneOcc in_lam1 _ int_cxt1)
1806 (OneOcc in_lam2 _ int_cxt2)
1807 = OneOcc (in_lam1 || in_lam2)
1808 False -- False, because it occurs in both branches
1809 (int_cxt1 && int_cxt2)
1810 orOccInfo a1 a2 = ASSERT( not (isDeadOcc a1 || isDeadOcc a2) )