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 )
28 import Name ( Name, localiseName )
32 import Var ( varUnique )
33 import Maybes ( orElse )
34 import Digraph ( SCC(..), stronglyConnCompFromEdgedVerticesR )
35 import PrelNames ( buildIdKey, foldrIdKey, runSTRepIdKey, augmentIdKey )
38 import Util ( mapAndUnzip, filterOut )
46 %************************************************************************
48 \subsection[OccurAnal-main]{Counting occurrences: main function}
50 %************************************************************************
52 Here's the externally-callable interface:
55 occurAnalysePgm :: Maybe (Activation -> Bool) -> [CoreRule]
56 -> [CoreBind] -> [CoreBind]
57 occurAnalysePgm active_rule imp_rules binds
58 = snd (go (initOccEnv active_rule imp_rules) binds)
60 initial_uds = addIdOccs emptyDetails (rulesFreeVars imp_rules)
61 -- The RULES keep things alive!
63 go :: OccEnv -> [CoreBind] -> (UsageDetails, [CoreBind])
67 = (final_usage, bind' ++ binds')
69 (bs_usage, binds') = go env binds
70 (final_usage, bind') = occAnalBind env env bind bs_usage
72 occurAnalyseExpr :: CoreExpr -> CoreExpr
73 -- Do occurrence analysis, and discard occurence info returned
75 = snd (occAnal (initOccEnv all_active_rules []) expr)
77 -- To be conservative, we say that all inlines and rules are active
78 all_active_rules = Just (\_ -> True)
82 %************************************************************************
84 \subsection[OccurAnal-main]{Counting occurrences: main function}
86 %************************************************************************
92 occAnalBind :: OccEnv -- The incoming OccEnv
93 -> OccEnv -- Same, but trimmed by (binderOf bind)
95 -> UsageDetails -- Usage details of scope
96 -> (UsageDetails, -- Of the whole let(rec)
99 occAnalBind env _ (NonRec binder rhs) body_usage
100 | isTyCoVar binder -- A type let; we don't gather usage info
101 = (body_usage, [NonRec binder rhs])
103 | not (binder `usedIn` body_usage) -- It's not mentioned
106 | otherwise -- It's mentioned in the body
107 = (body_usage' +++ rhs_usage3, [NonRec tagged_binder rhs'])
109 (body_usage', tagged_binder) = tagBinder body_usage binder
110 (rhs_usage1, rhs') = occAnalRhs env (idOccInfo tagged_binder) rhs
111 rhs_usage2 = addIdOccs rhs_usage1 (idUnfoldingVars binder)
112 rhs_usage3 = addIdOccs rhs_usage2 (idRuleVars binder)
113 -- See Note [Rules are extra RHSs] and Note [Rule dependency info]
118 Dropping dead code for recursive bindings is done in a very simple way:
120 the entire set of bindings is dropped if none of its binders are
121 mentioned in its body; otherwise none are.
123 This seems to miss an obvious improvement.
135 Now 'f' is unused! But it's OK! Dependency analysis will sort this
136 out into a letrec for 'g' and a 'let' for 'f', and then 'f' will get
137 dropped. It isn't easy to do a perfect job in one blow. Consider
148 Note [Loop breaking and RULES]
149 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
150 Loop breaking is surprisingly subtle. First read the section 4 of
151 "Secrets of the GHC inliner". This describes our basic plan.
153 However things are made quite a bit more complicated by RULES. Remember
155 * Note [Rules are extra RHSs]
156 ~~~~~~~~~~~~~~~~~~~~~~~~~~~
157 A RULE for 'f' is like an extra RHS for 'f'. That way the "parent"
158 keeps the specialised "children" alive. If the parent dies
159 (because it isn't referenced any more), then the children will die
160 too (unless they are already referenced directly).
162 To that end, we build a Rec group for each cyclic strongly
164 *treating f's rules as extra RHSs for 'f'*.
165 More concretely, the SCC analysis runs on a graph with an edge
166 from f -> g iff g is mentioned in
171 Under (b) we include variables free in *either* LHS *or* RHS of
172 the rule. The former might seems silly, but see Note [Rule
173 dependency info]. So in Example [eftInt], eftInt and eftIntFB
174 will be put in the same Rec, even though their 'main' RHSs are
177 * Note [Rules are visible in their own rec group]
178 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
179 We want the rules for 'f' to be visible in f's right-hand side.
180 And we'd like them to be visible in other functions in f's Rec
181 group. E.g. in Example [Specialisation rules] we want f' rule
182 to be visible in both f's RHS, and fs's RHS.
184 This means that we must simplify the RULEs first, before looking
185 at any of the definitions. This is done by Simplify.simplRecBind,
186 when it calls addLetIdInfo.
188 * Note [Choosing loop breakers]
189 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
190 We avoid infinite inlinings by choosing loop breakers, and
191 ensuring that a loop breaker cuts each loop. But what is a
192 "loop"? In particular, a RULE is like an equation for 'f' that
193 is *always* inlined if it is applicable. We do *not* disable
194 rules for loop-breakers. It's up to whoever makes the rules to
195 make sure that the rules themselves always terminate. See Note
196 [Rules for recursive functions] in Simplify.lhs
199 f's RHS (or its INLINE template if it has one) mentions g, and
200 g has a RULE that mentions h, and
201 h has a RULE that mentions f
203 then we *must* choose f to be a loop breaker. In general, take the
204 free variables of f's RHS, and augment it with all the variables
205 reachable by RULES from those starting points. That is the whole
206 reason for computing rule_fv_env in occAnalBind. (Of course we
207 only consider free vars that are also binders in this Rec group.)
208 See also Note [Finding rule RHS free vars]
210 Note that when we compute this rule_fv_env, we only consider variables
211 free in the *RHS* of the rule, in contrast to the way we build the
212 Rec group in the first place (Note [Rule dependency info])
214 Note that if 'g' has RHS that mentions 'w', we should add w to
215 g's loop-breaker edges. More concretely there is an edge from f -> g
217 (a) g is mentioned in f's RHS
218 (b) h is mentioned in f's RHS, and
219 g appears in the RHS of a RULE of h
220 or a transitive sequence of rules starting with h
222 Note that in Example [eftInt], *neither* eftInt *nor* eftIntFB is
223 chosen as a loop breaker, because their RHSs don't mention each other.
224 And indeed both can be inlined safely.
226 Note that the edges of the graph we use for computing loop breakers
227 are not the same as the edges we use for computing the Rec blocks.
228 That's why we compute
229 rec_edges for the Rec block analysis
230 loop_breaker_edges for the loop breaker analysis
232 * Note [Finding rule RHS free vars]
233 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
234 Consider this real example from Data Parallel Haskell
235 tagZero :: Array Int -> Array Tag
236 {-# INLINE [1] tagZeroes #-}
237 tagZero xs = pmap (\x -> fromBool (x==0)) xs
239 {-# RULES "tagZero" [~1] forall xs n.
240 pmap fromBool <blah blah> = tagZero xs #-}
241 So tagZero's RHS mentions pmap, and pmap's RULE mentions tagZero.
242 However, tagZero can only be inlined in phase 1 and later, while
243 the RULE is only active *before* phase 1. So there's no problem.
245 To make this work, we look for the RHS free vars only for
246 *active* rules. That's the reason for the is_active argument
247 to idRhsRuleVars, and the occ_rule_act field of the OccEnv.
249 * Note [Weak loop breakers]
250 ~~~~~~~~~~~~~~~~~~~~~~~~~
251 There is a last nasty wrinkle. Suppose we have
261 Remember that we simplify the RULES before any RHS (see Note
262 [Rules are visible in their own rec group] above).
264 So we must *not* postInlineUnconditionally 'g', even though
265 its RHS turns out to be trivial. (I'm assuming that 'g' is
266 not choosen as a loop breaker.) Why not? Because then we
267 drop the binding for 'g', which leaves it out of scope in the
270 We "solve" this by making g a "weak" or "rules-only" loop breaker,
271 with OccInfo = IAmLoopBreaker True. A normal "strong" loop breaker
272 has IAmLoopBreaker False. So
274 Inline postInlineUnconditionally
275 IAmLoopBreaker False no no
276 IAmLoopBreaker True yes no
279 The **sole** reason for this kind of loop breaker is so that
280 postInlineUnconditionally does not fire. Ugh.
282 * Note [Rule dependency info]
283 ~~~~~~~~~~~~~~~~~~~~~~~~~~~
284 The VarSet in a SpecInfo is used for dependency analysis in the
285 occurrence analyser. We must track free vars in *both* lhs and rhs.
286 Hence use of idRuleVars, rather than idRuleRhsVars in occAnalBind.
290 Then if we substitute y for x, we'd better do so in the
291 rule's LHS too, so we'd better ensure the dependency is respected
294 * Note [Inline rules]
296 None of the above stuff about RULES applies to Inline Rules,
297 stored in a CoreUnfolding. The unfolding, if any, is simplified
298 at the same time as the regular RHS of the function, so it should
299 be treated *exactly* like an extra RHS.
301 There is a danger that we'll be sub-optimal if we see this
303 [INLINE f = ..no f...]
304 where f is recursive, but the INLINE is not. This can just about
305 happen with a sufficiently odd set of rules; eg
308 {-# INLINE [1] foo #-}
312 {-# INLINE [1] bar #-}
315 {-# RULES "foo" [~1] forall x. foo x = bar x #-}
317 Here the RULE makes bar recursive; but it's INLINE pragma remains
318 non-recursive. It's tempting to then say that 'bar' should not be
319 a loop breaker, but an attempt to do so goes wrong in two ways:
323 [INLINE $cfoo = ...no-$df...]
324 But we want $cfoo to depend on $df explicitly so that we
325 put the bindings in the right order to inline $df in $cfoo
326 and perhaps break the loop altogether. (Maybe this
333 Example (from GHC.Enum):
335 eftInt :: Int# -> Int# -> [Int]
336 eftInt x y = ...(non-recursive)...
338 {-# INLINE [0] eftIntFB #-}
339 eftIntFB :: (Int -> r -> r) -> r -> Int# -> Int# -> r
340 eftIntFB c n x y = ...(non-recursive)...
343 "eftInt" [~1] forall x y. eftInt x y = build (\ c n -> eftIntFB c n x y)
344 "eftIntList" [1] eftIntFB (:) [] = eftInt
347 Example [Specialisation rules]
348 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
349 Consider this group, which is typical of what SpecConstr builds:
351 fs a = ....f (C a)....
352 f x = ....f (C a)....
353 {-# RULE f (C a) = fs a #-}
355 So 'f' and 'fs' are in the same Rec group (since f refers to fs via its RULE).
357 But watch out! If 'fs' is not chosen as a loop breaker, we may get an infinite loop:
358 - the RULE is applied in f's RHS (see Note [Self-recursive rules] in Simplify
359 - fs is inlined (say it's small)
360 - now there's another opportunity to apply the RULE
362 This showed up when compiling Control.Concurrent.Chan.getChanContents.
366 occAnalBind _ env (Rec pairs) body_usage
367 = foldr (occAnalRec env) (body_usage, []) sccs
368 -- For a recursive group, we
369 -- * occ-analyse all the RHSs
370 -- * compute strongly-connected components
371 -- * feed those components to occAnalRec
373 -------------Dependency analysis ------------------------------
374 bndr_set = mkVarSet (map fst pairs)
376 sccs :: [SCC (Node Details)]
377 sccs = {-# SCC "occAnalBind.scc" #-} stronglyConnCompFromEdgedVerticesR rec_edges
379 rec_edges :: [Node Details]
380 rec_edges = {-# SCC "occAnalBind.assoc" #-} map make_node pairs
382 make_node (bndr, rhs)
383 = (details, varUnique bndr, keysUFM out_edges)
385 details = ND { nd_bndr = bndr, nd_rhs = rhs'
386 , nd_uds = rhs_usage3, nd_inl = inl_fvs}
388 (rhs_usage1, rhs') = occAnalRhs env NoOccInfo rhs
389 rhs_usage2 = addIdOccs rhs_usage1 rule_fvs -- Note [Rules are extra RHSs]
390 rhs_usage3 = addIdOccs rhs_usage2 unf_fvs
391 unf = realIdUnfolding bndr -- Ignore any current loop-breaker flag
392 unf_fvs = stableUnfoldingVars unf
393 rule_fvs = idRuleVars bndr -- See Note [Rule dependency info]
395 inl_fvs = rhs_fvs `unionVarSet` unf_fvs
396 rhs_fvs = intersectUFM_C (\b _ -> b) bndr_set rhs_usage1
397 out_edges = intersectUFM_C (\b _ -> b) bndr_set rhs_usage3
398 -- (a -> b) means a mentions b
399 -- Given the usage details (a UFM that gives occ info for each free var of
400 -- the RHS) we can get the list of free vars -- or rather their Int keys --
401 -- by just extracting the keys from the finite map. Grimy, but fast.
402 -- Previously we had this:
403 -- [ bndr | bndr <- bndrs,
404 -- maybeToBool (lookupVarEnv rhs_usage bndr)]
405 -- which has n**2 cost, and this meant that edges_from alone
406 -- consumed 10% of total runtime!
408 -----------------------------
409 occAnalRec :: OccEnv -> SCC (Node Details)
410 -> (UsageDetails, [CoreBind])
411 -> (UsageDetails, [CoreBind])
413 -- The NonRec case is just like a Let (NonRec ...) above
414 occAnalRec _ (AcyclicSCC (ND { nd_bndr = bndr, nd_rhs = rhs, nd_uds = rhs_usage}, _, _))
416 | not (bndr `usedIn` body_usage)
417 = (body_usage, binds)
419 | otherwise -- It's mentioned in the body
420 = (body_usage' +++ rhs_usage,
421 NonRec tagged_bndr rhs : binds)
423 (body_usage', tagged_bndr) = tagBinder body_usage bndr
426 -- The Rec case is the interesting one
427 -- See Note [Loop breaking]
428 occAnalRec env (CyclicSCC nodes) (body_usage, binds)
429 | not (any (`usedIn` body_usage) bndrs) -- NB: look at body_usage, not total_usage
430 = (body_usage, binds) -- Dead code
432 | otherwise -- At this point we always build a single Rec
433 = (final_usage, Rec pairs : binds)
436 bndrs = [b | (ND { nd_bndr = b }, _, _) <- nodes]
437 bndr_set = mkVarSet bndrs
438 non_boring bndr = isId bndr &&
439 (isStableUnfolding (realIdUnfolding bndr) || idHasRules bndr)
441 ----------------------------
442 -- Tag the binders with their occurrence info
443 total_usage = foldl add_usage body_usage nodes
444 add_usage usage_so_far (ND { nd_uds = rhs_usage }, _, _) = usage_so_far +++ rhs_usage
445 (final_usage, tagged_nodes) = mapAccumL tag_node total_usage nodes
447 tag_node :: UsageDetails -> Node Details -> (UsageDetails, Node Details)
448 -- (a) Tag the binders in the details with occ info
449 -- (b) Mark the binder with "weak loop-breaker" OccInfo
450 -- saying "no preInlineUnconditionally" if it is used
451 -- in any rule (lhs or rhs) of the recursive group
452 -- See Note [Weak loop breakers]
453 tag_node usage (details@ND { nd_bndr = bndr }, k, ks)
454 = (usage `delVarEnv` bndr, (details { nd_bndr = bndr2 }, k, ks))
456 bndr2 | bndr `elemVarSet` all_rule_fvs = makeLoopBreaker True bndr1
458 bndr1 = setBinderOcc usage bndr
459 all_rule_fvs = bndr_set `intersectVarSet` foldr (unionVarSet . idRuleVars)
462 ----------------------------
463 -- Now reconstruct the cycle
464 pairs | any non_boring bndrs
465 = foldr (reOrderRec 0) [] $
466 stronglyConnCompFromEdgedVerticesR loop_breaker_edges
468 = reOrderCycle 0 tagged_nodes []
470 -- See Note [Choosing loop breakers] for loop_breaker_edges
471 loop_breaker_edges = map mk_node tagged_nodes
472 mk_node (details@(ND { nd_inl = inl_fvs }), k, _) = (details, k, new_ks)
474 new_ks = keysUFM (fst (extendFvs rule_fv_env inl_fvs))
476 ------------------------------------
477 rule_fv_env :: IdEnv IdSet -- Variables from this group mentioned in RHS of rules
478 -- Domain is *subset* of bound vars (others have no rule fvs)
479 rule_fv_env = transClosureFV init_rule_fvs
481 | Just is_active <- occ_rule_act env -- See Note [Finding rule RHS free vars]
485 , let rule_fvs = idRuleRhsVars is_active b
486 `intersectVarSet` bndr_set
487 , not (isEmptyVarSet rule_fvs)]
492 @reOrderRec@ is applied to the list of (binder,rhs) pairs for a cyclic
493 strongly connected component (there's guaranteed to be a cycle). It returns the
495 a) in a better order,
496 b) with some of the Ids having a IAmALoopBreaker pragma
498 The "loop-breaker" Ids are sufficient to break all cycles in the SCC. This means
499 that the simplifier can guarantee not to loop provided it never records an inlining
500 for these no-inline guys.
502 Furthermore, the order of the binds is such that if we neglect dependencies
503 on the no-inline Ids then the binds are topologically sorted. This means
504 that the simplifier will generally do a good job if it works from top bottom,
505 recording inlinings for any Ids which aren't marked as "no-inline" as it goes.
508 [June 98: I don't understand the following paragraphs, and I've
509 changed the a=b case again so that it isn't a special case any more.]
511 Here's a case that bit me:
519 Re-ordering doesn't change the order of bindings, but there was no loop-breaker.
521 My solution was to make a=b bindings record b as Many, rather like INLINE bindings.
522 Perhaps something cleverer would suffice.
527 type Node details = (details, Unique, [Unique]) -- The Ints are gotten from the Unique,
528 -- which is gotten from the Id.
530 = ND { nd_bndr :: Id -- Binder
531 , nd_rhs :: CoreExpr -- RHS
533 , nd_uds :: UsageDetails -- Usage from RHS,
534 -- including RULES and InlineRule unfolding
536 , nd_inl :: IdSet -- Other binders *from this Rec group* mentioned in
537 } -- its InlineRule unfolding (if present)
539 -- but *excluding* any RULES
540 -- This is the IdSet that may be used if the Id is inlined
542 reOrderRec :: Int -> SCC (Node Details)
543 -> [(Id,CoreExpr)] -> [(Id,CoreExpr)]
544 -- Sorted into a plausible order. Enough of the Ids have
545 -- IAmALoopBreaker pragmas that there are no loops left.
546 reOrderRec _ (AcyclicSCC (ND { nd_bndr = bndr, nd_rhs = rhs }, _, _))
547 pairs = (bndr, rhs) : pairs
548 reOrderRec depth (CyclicSCC cycle) pairs = reOrderCycle depth cycle pairs
550 reOrderCycle :: Int -> [Node Details] -> [(Id,CoreExpr)] -> [(Id,CoreExpr)]
552 = panic "reOrderCycle"
553 reOrderCycle _ [(ND { nd_bndr = bndr, nd_rhs = rhs }, _, _)] pairs
554 = -- Common case of simple self-recursion
555 (makeLoopBreaker False bndr, rhs) : pairs
557 reOrderCycle depth (bind : binds) pairs
558 = -- Choose a loop breaker, mark it no-inline,
559 -- do SCC analysis on the rest, and recursively sort them out
560 -- pprTrace "reOrderCycle" (ppr [b | (ND { nd_bndr = b }, _, _) <- bind:binds]) $
561 foldr (reOrderRec new_depth)
562 ([ (makeLoopBreaker False bndr, rhs)
563 | (ND { nd_bndr = bndr, nd_rhs = rhs }, _, _) <- chosen_binds] ++ pairs)
564 (stronglyConnCompFromEdgedVerticesR unchosen)
566 (chosen_binds, unchosen) = choose_loop_breaker [bind] (score bind) [] binds
568 approximate_loop_breaker = depth >= 2
569 new_depth | approximate_loop_breaker = 0
570 | otherwise = depth+1
571 -- After two iterations (d=0, d=1) give up
572 -- and approximate, returning to d=0
574 -- This loop looks for the bind with the lowest score
575 -- to pick as the loop breaker. The rest accumulate in
576 choose_loop_breaker loop_binds _loop_sc acc []
577 = (loop_binds, acc) -- Done
579 -- If approximate_loop_breaker is True, we pick *all*
580 -- nodes with lowest score, else just one
581 -- See Note [Complexity of loop breaking]
582 choose_loop_breaker loop_binds loop_sc acc (bind : binds)
583 | sc < loop_sc -- Lower score so pick this new one
584 = choose_loop_breaker [bind] sc (loop_binds ++ acc) binds
586 | approximate_loop_breaker && sc == loop_sc
587 = choose_loop_breaker (bind : loop_binds) loop_sc acc binds
589 | otherwise -- Higher score so don't pick it
590 = choose_loop_breaker loop_binds loop_sc (bind : acc) binds
594 score :: Node Details -> Int -- Higher score => less likely to be picked as loop breaker
595 score (ND { nd_bndr = bndr, nd_rhs = rhs }, _, _)
596 | not (isId bndr) = 100 -- A type or cercion variable is never a loop breaker
598 | isDFunId bndr = 9 -- Never choose a DFun as a loop breaker
599 -- Note [DFuns should not be loop breakers]
601 | Just inl_source <- isStableCoreUnfolding_maybe (idUnfolding bndr)
603 InlineWrapper {} -> 10 -- Note [INLINE pragmas]
604 _other -> 3 -- Data structures are more important than this
605 -- so that dictionary/method recursion unravels
606 -- Note that this case hits all InlineRule things, so we
607 -- never look at 'rhs for InlineRule stuff. That's right, because
608 -- 'rhs' is irrelevant for inlining things with an InlineRule
610 | is_con_app rhs = 5 -- Data types help with cases: Note [Constructor applications]
612 | exprIsTrivial rhs = 10 -- Practically certain to be inlined
613 -- Used to have also: && not (isExportedId bndr)
614 -- But I found this sometimes cost an extra iteration when we have
615 -- rec { d = (a,b); a = ...df...; b = ...df...; df = d }
616 -- where df is the exported dictionary. Then df makes a really
617 -- bad choice for loop breaker
620 -- If an Id is marked "never inline" then it makes a great loop breaker
621 -- The only reason for not checking that here is that it is rare
622 -- and I've never seen a situation where it makes a difference,
623 -- so it probably isn't worth the time to test on every binder
624 -- | isNeverActive (idInlinePragma bndr) = -10
626 | isOneOcc (idOccInfo bndr) = 2 -- Likely to be inlined
628 | canUnfold (realIdUnfolding bndr) = 1
629 -- The Id has some kind of unfolding
630 -- Ignore loop-breaker-ness here because that is what we are setting!
634 -- Checking for a constructor application
635 -- Cheap and cheerful; the simplifer moves casts out of the way
636 -- The lambda case is important to spot x = /\a. C (f a)
637 -- which comes up when C is a dictionary constructor and
638 -- f is a default method.
639 -- Example: the instance for Show (ST s a) in GHC.ST
641 -- However we *also* treat (\x. C p q) as a con-app-like thing,
642 -- Note [Closure conversion]
643 is_con_app (Var v) = isConLikeId v
644 is_con_app (App f _) = is_con_app f
645 is_con_app (Lam _ e) = is_con_app e
646 is_con_app (Note _ e) = is_con_app e
649 makeLoopBreaker :: Bool -> Id -> Id
650 -- Set the loop-breaker flag: see Note [Weak loop breakers]
651 makeLoopBreaker weak bndr
652 = ASSERT2( isId bndr, ppr bndr ) setIdOccInfo bndr (IAmALoopBreaker weak)
655 Note [Complexity of loop breaking]
656 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
657 The loop-breaking algorithm knocks out one binder at a time, and
658 performs a new SCC analysis on the remaining binders. That can
659 behave very badly in tightly-coupled groups of bindings; in the
660 worst case it can be (N**2)*log N, because it does a full SCC
661 on N, then N-1, then N-2 and so on.
663 To avoid this, we switch plans after 2 (or whatever) attempts:
664 Plan A: pick one binder with the lowest score, make it
665 a loop breaker, and try again
666 Plan B: pick *all* binders with the lowest score, make them
667 all loop breakers, and try again
668 Since there are only a small finite number of scores, this will
669 terminate in a constant number of iterations, rather than O(N)
672 You might thing that it's very unlikely, but RULES make it much
673 more likely. Here's a real example from Trac #1969:
674 Rec { $dm = \d.\x. op d
675 {-# RULES forall d. $dm Int d = $s$dm1
676 forall d. $dm Bool d = $s$dm2 #-}
678 dInt = MkD .... opInt ...
679 dInt = MkD .... opBool ...
684 $s$dm2 = \x. op dBool }
685 The RULES stuff means that we can't choose $dm as a loop breaker
686 (Note [Choosing loop breakers]), so we must choose at least (say)
687 opInt *and* opBool, and so on. The number of loop breakders is
688 linear in the number of instance declarations.
690 Note [INLINE pragmas]
691 ~~~~~~~~~~~~~~~~~~~~~
692 Avoid choosing a function with an INLINE pramga as the loop breaker!
693 If such a function is mutually-recursive with a non-INLINE thing,
694 then the latter should be the loop-breaker.
696 Usually this is just a question of optimisation. But a particularly
697 bad case is wrappers generated by the demand analyser: if you make
698 then into a loop breaker you may get an infinite inlining loop. For
701 $wfoo x = ....foo x....
703 {-loop brk-} foo x = ...$wfoo x...
705 The interface file sees the unfolding for $wfoo, and sees that foo is
706 strict (and hence it gets an auto-generated wrapper). Result: an
707 infinite inlining in the importing scope. So be a bit careful if you
708 change this. A good example is Tree.repTree in
709 nofib/spectral/minimax. If the repTree wrapper is chosen as the loop
710 breaker then compiling Game.hs goes into an infinite loop. This
711 happened when we gave is_con_app a lower score than inline candidates:
714 = __inline_me (/\a. \w w1 w2 ->
715 case Tree.$wrepTree @ a w w1 w2 of
716 { (# ww1, ww2 #) -> Branch @ a ww1 ww2 })
719 (# w2_smP, map a (Tree a) (Tree.repTree a w1 w) (w w2) #)
721 Here we do *not* want to choose 'repTree' as the loop breaker.
723 Note [DFuns should not be loop breakers]
724 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
725 It's particularly bad to make a DFun into a loop breaker. See
726 Note [How instance declarations are translated] in TcInstDcls
728 We give DFuns a higher score than ordinary CONLIKE things because
729 if there's a choice we want the DFun to be the non-looop breker. Eg
731 rec { sc = /\ a \$dC. $fBWrap (T a) ($fCT @ a $dC)
733 $fCT :: forall a_afE. (Roman.C a_afE) => Roman.C (Roman.T a_afE)
735 $fCT = /\a \$dC. MkD (T a) ((sc @ a $dC) |> blah) ($ctoF @ a $dC)
738 Here 'sc' (the superclass) looks CONLIKE, but we'll never get to it
739 if we can't unravel the DFun first.
741 Note [Constructor applications]
742 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
743 It's really really important to inline dictionaries. Real
744 example (the Enum Ordering instance from GHC.Base):
746 rec f = \ x -> case d of (p,q,r) -> p x
747 g = \ x -> case d of (p,q,r) -> q x
750 Here, f and g occur just once; but we can't inline them into d.
751 On the other hand we *could* simplify those case expressions if
752 we didn't stupidly choose d as the loop breaker.
753 But we won't because constructor args are marked "Many".
754 Inlining dictionaries is really essential to unravelling
755 the loops in static numeric dictionaries, see GHC.Float.
757 Note [Closure conversion]
758 ~~~~~~~~~~~~~~~~~~~~~~~~~
759 We treat (\x. C p q) as a high-score candidate in the letrec scoring algorithm.
760 The immediate motivation came from the result of a closure-conversion transformation
761 which generated code like this:
763 data Clo a b = forall c. Clo (c -> a -> b) c
765 ($:) :: Clo a b -> a -> b
766 Clo f env $: x = f env x
768 rec { plus = Clo plus1 ()
770 ; plus1 _ n = Clo plus2 n
773 ; plus2 (Succ m) n = Succ (plus $: m $: n) }
775 If we inline 'plus' and 'plus1', everything unravels nicely. But if
776 we choose 'plus1' as the loop breaker (which is entirely possible
777 otherwise), the loop does not unravel nicely.
780 @occAnalRhs@ deals with the question of bindings where the Id is marked
781 by an INLINE pragma. For these we record that anything which occurs
782 in its RHS occurs many times. This pessimistically assumes that ths
783 inlined binder also occurs many times in its scope, but if it doesn't
784 we'll catch it next time round. At worst this costs an extra simplifier pass.
785 ToDo: try using the occurrence info for the inline'd binder.
787 [March 97] We do the same for atomic RHSs. Reason: see notes with reOrderRec.
788 [June 98, SLPJ] I've undone this change; I don't understand it. See notes with reOrderRec.
793 -> OccInfo -> CoreExpr -- Binder and rhs
794 -- For non-recs the binder is alrady tagged
795 -- with occurrence info
796 -> (UsageDetails, CoreExpr)
797 -- Returned usage details covers only the RHS,
798 -- and *not* the RULE or INLINE template for the Id
799 occAnalRhs env occ rhs
802 ctxt | certainly_inline = env
803 | otherwise = rhsCtxt env
804 -- Note that we generally use an rhsCtxt. This tells the occ anal n
805 -- that it's looking at an RHS, which has an effect in occAnalApp
807 -- But there's a problem. Consider
812 -- First time round, it looks as if x1 and x2 occur as an arg of a
813 -- let-bound constructor ==> give them a many-occurrence.
814 -- But then x3 is inlined (unconditionally as it happens) and
815 -- next time round, x2 will be, and the next time round x1 will be
816 -- Result: multiple simplifier iterations. Sigh.
817 -- Crude solution: use rhsCtxt for things that occur just once...
819 certainly_inline = case occ of
820 OneOcc in_lam one_br _ -> not in_lam && one_br
824 addIdOccs :: UsageDetails -> VarSet -> UsageDetails
825 addIdOccs usage id_set = foldVarSet add usage id_set
827 add v u | isId v = addOneOcc u v NoOccInfo
829 -- Give a non-committal binder info (i.e NoOccInfo) because
830 -- a) Many copies of the specialised thing can appear
831 -- b) We don't want to substitute a BIG expression inside a RULE
832 -- even if that's the only occurrence of the thing
833 -- (Same goes for INLINE.)
841 -> (UsageDetails, -- Gives info only about the "interesting" Ids
844 occAnal _ (Type t) = (emptyDetails, Type t)
845 occAnal env (Var v) = (mkOneOcc env v False, Var v)
846 -- At one stage, I gathered the idRuleVars for v here too,
847 -- which in a way is the right thing to do.
848 -- But that went wrong right after specialisation, when
849 -- the *occurrences* of the overloaded function didn't have any
850 -- rules in them, so the *specialised* versions looked as if they
851 -- weren't used at all.
854 We regard variables that occur as constructor arguments as "dangerousToDup":
858 f x = let y = expensive x in
860 (case z of {(p,q)->q}, case z of {(p,q)->q})
863 We feel free to duplicate the WHNF (True,y), but that means
864 that y may be duplicated thereby.
866 If we aren't careful we duplicate the (expensive x) call!
867 Constructors are rather like lambdas in this way.
870 occAnal _ expr@(Lit _) = (emptyDetails, expr)
874 occAnal env (Note note@(SCC _) body)
875 = case occAnal env body of { (usage, body') ->
876 (mapVarEnv markInsideSCC usage, Note note body')
879 occAnal env (Note note body)
880 = case occAnal env body of { (usage, body') ->
881 (usage, Note note body')
884 occAnal env (Cast expr co)
885 = case occAnal env expr of { (usage, expr') ->
886 (markManyIf (isRhsEnv env) usage, Cast expr' co)
887 -- If we see let x = y `cast` co
888 -- then mark y as 'Many' so that we don't
889 -- immediately inline y again.
894 occAnal env app@(App _ _)
895 = occAnalApp env (collectArgs app)
897 -- Ignore type variables altogether
898 -- (a) occurrences inside type lambdas only not marked as InsideLam
899 -- (b) type variables not in environment
901 occAnal env (Lam x body) | isTyCoVar x
902 = case occAnal env body of { (body_usage, body') ->
903 (body_usage, Lam x body')
906 -- For value lambdas we do a special hack. Consider
908 -- If we did nothing, x is used inside the \y, so would be marked
909 -- as dangerous to dup. But in the common case where the abstraction
910 -- is applied to two arguments this is over-pessimistic.
911 -- So instead, we just mark each binder with its occurrence
912 -- info in the *body* of the multiple lambda.
913 -- Then, the simplifier is careful when partially applying lambdas.
915 occAnal env expr@(Lam _ _)
916 = case occAnal env_body body of { (body_usage, body') ->
918 (final_usage, tagged_binders) = tagLamBinders body_usage binders'
919 -- Use binders' to put one-shot info on the lambdas
921 -- URGH! Sept 99: we don't seem to be able to use binders' here, because
922 -- we get linear-typed things in the resulting program that we can't handle yet.
923 -- (e.g. PrelShow) TODO
925 really_final_usage = if linear then
928 mapVarEnv markInsideLam final_usage
931 mkLams tagged_binders body') }
933 env_body = vanillaCtxt (trimOccEnv env binders)
934 -- Body is (no longer) an RhsContext
935 (binders, body) = collectBinders expr
936 binders' = oneShotGroup env binders
937 linear = all is_one_shot binders'
938 is_one_shot b = isId b && isOneShotBndr b
940 occAnal env (Case scrut bndr ty alts)
941 = case occ_anal_scrut scrut alts of { (scrut_usage, scrut') ->
942 case mapAndUnzip occ_anal_alt alts of { (alts_usage_s, alts') ->
944 alts_usage = foldr1 combineAltsUsageDetails alts_usage_s
945 (alts_usage1, tagged_bndr) = tag_case_bndr alts_usage bndr
946 total_usage = scrut_usage +++ alts_usage1
948 total_usage `seq` (total_usage, Case scrut' tagged_bndr ty alts') }}
950 -- Note [Case binder usage]
951 -- ~~~~~~~~~~~~~~~~~~~~~~~~
952 -- The case binder gets a usage of either "many" or "dead", never "one".
953 -- Reason: we like to inline single occurrences, to eliminate a binding,
954 -- but inlining a case binder *doesn't* eliminate a binding.
955 -- We *don't* want to transform
956 -- case x of w { (p,q) -> f w }
958 -- case x of w { (p,q) -> f (p,q) }
959 tag_case_bndr usage bndr
960 = case lookupVarEnv usage bndr of
961 Nothing -> (usage, setIdOccInfo bndr IAmDead)
962 Just _ -> (usage `delVarEnv` bndr, setIdOccInfo bndr NoOccInfo)
964 alt_env = mkAltEnv env scrut bndr
965 occ_anal_alt = occAnalAlt alt_env bndr
967 occ_anal_scrut (Var v) (alt1 : other_alts)
968 | not (null other_alts) || not (isDefaultAlt alt1)
969 = (mkOneOcc env v True, Var v) -- The 'True' says that the variable occurs
970 -- in an interesting context; the case has
971 -- at least one non-default alternative
972 occ_anal_scrut scrut _alts
973 = occAnal (vanillaCtxt env) scrut -- No need for rhsCtxt
975 occAnal env (Let bind body)
976 = case occAnal env_body body of { (body_usage, body') ->
977 case occAnalBind env env_body bind body_usage of { (final_usage, new_binds) ->
978 (final_usage, mkLets new_binds body') }}
980 env_body = trimOccEnv env (bindersOf bind)
982 occAnalArgs :: OccEnv -> [CoreExpr] -> (UsageDetails, [CoreExpr])
984 = case mapAndUnzip (occAnal arg_env) args of { (arg_uds_s, args') ->
985 (foldr (+++) emptyDetails arg_uds_s, args')}
987 arg_env = vanillaCtxt env
990 Applications are dealt with specially because we want
991 the "build hack" to work.
995 -> (Expr CoreBndr, [Arg CoreBndr])
996 -> (UsageDetails, Expr CoreBndr)
997 occAnalApp env (Var fun, args)
998 = case args_stuff of { (args_uds, args') ->
1000 final_args_uds = markManyIf (isRhsEnv env && is_exp) args_uds
1001 -- We mark the free vars of the argument of a constructor or PAP
1002 -- as "many", if it is the RHS of a let(rec).
1003 -- This means that nothing gets inlined into a constructor argument
1004 -- position, which is what we want. Typically those constructor
1005 -- arguments are just variables, or trivial expressions.
1007 -- This is the *whole point* of the isRhsEnv predicate
1009 (fun_uds +++ final_args_uds, mkApps (Var fun) args') }
1011 fun_uniq = idUnique fun
1012 fun_uds = mkOneOcc env fun (valArgCount args > 0)
1013 is_exp = isExpandableApp fun (valArgCount args)
1014 -- See Note [CONLIKE pragma] in BasicTypes
1015 -- The definition of is_exp should match that in
1016 -- Simplify.prepareRhs
1018 -- Hack for build, fold, runST
1019 args_stuff | fun_uniq == buildIdKey = appSpecial env 2 [True,True] args
1020 | fun_uniq == augmentIdKey = appSpecial env 2 [True,True] args
1021 | fun_uniq == foldrIdKey = appSpecial env 3 [False,True] args
1022 | fun_uniq == runSTRepIdKey = appSpecial env 2 [True] args
1023 -- (foldr k z xs) may call k many times, but it never
1024 -- shares a partial application of k; hence [False,True]
1025 -- This means we can optimise
1026 -- foldr (\x -> let v = ...x... in \y -> ...v...) z xs
1027 -- by floating in the v
1029 | otherwise = occAnalArgs env args
1032 occAnalApp env (fun, args)
1033 = case occAnal (addAppCtxt env args) fun of { (fun_uds, fun') ->
1034 -- The addAppCtxt is a bit cunning. One iteration of the simplifier
1035 -- often leaves behind beta redexs like
1036 -- (\x y -> e) a1 a2
1037 -- Here we would like to mark x,y as one-shot, and treat the whole
1038 -- thing much like a let. We do this by pushing some True items
1039 -- onto the context stack.
1041 case occAnalArgs env args of { (args_uds, args') ->
1043 final_uds = fun_uds +++ args_uds
1045 (final_uds, mkApps fun' args') }}
1048 markManyIf :: Bool -- If this is true
1049 -> UsageDetails -- Then do markMany on this
1051 markManyIf True uds = mapVarEnv markMany uds
1052 markManyIf False uds = uds
1054 appSpecial :: OccEnv
1055 -> Int -> CtxtTy -- Argument number, and context to use for it
1057 -> (UsageDetails, [CoreExpr])
1058 appSpecial env n ctxt args
1061 arg_env = vanillaCtxt env
1063 go _ [] = (emptyDetails, []) -- Too few args
1065 go 1 (arg:args) -- The magic arg
1066 = case occAnal (setCtxtTy arg_env ctxt) arg of { (arg_uds, arg') ->
1067 case occAnalArgs env args of { (args_uds, args') ->
1068 (arg_uds +++ args_uds, arg':args') }}
1071 = case occAnal arg_env arg of { (arg_uds, arg') ->
1072 case go (n-1) args of { (args_uds, args') ->
1073 (arg_uds +++ args_uds, arg':args') }}
1077 Note [Binders in case alternatives]
1078 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
1080 case x of y { (a,b) -> f y }
1081 We treat 'a', 'b' as dead, because they don't physically occur in the
1082 case alternative. (Indeed, a variable is dead iff it doesn't occur in
1083 its scope in the output of OccAnal.) It really helps to know when
1084 binders are unused. See esp the call to isDeadBinder in
1085 Simplify.mkDupableAlt
1087 In this example, though, the Simplifier will bring 'a' and 'b' back to
1088 life, beause it binds 'y' to (a,b) (imagine got inlined and
1092 occAnalAlt :: OccEnv
1095 -> (UsageDetails, Alt IdWithOccInfo)
1096 occAnalAlt env case_bndr (con, bndrs, rhs)
1098 env' = trimOccEnv env bndrs
1100 case occAnal env' rhs of { (rhs_usage1, rhs1) ->
1102 proxies = getProxies env' case_bndr
1103 (rhs_usage2, rhs2) = foldrBag wrapProxy (rhs_usage1, rhs1) proxies
1104 (alt_usg, tagged_bndrs) = tagLamBinders rhs_usage2 bndrs
1105 bndrs' = tagged_bndrs -- See Note [Binders in case alternatives]
1107 (alt_usg, (con, bndrs', rhs2)) }
1109 wrapProxy :: ProxyBind -> (UsageDetails, CoreExpr) -> (UsageDetails, CoreExpr)
1110 wrapProxy (bndr, rhs_var, co) (body_usg, body)
1111 | not (bndr `usedIn` body_usg)
1114 = (body_usg' +++ rhs_usg, Let (NonRec tagged_bndr rhs) body)
1116 (body_usg', tagged_bndr) = tagBinder body_usg bndr
1117 rhs_usg = unitVarEnv rhs_var NoOccInfo -- We don't need exact info
1118 rhs = mkCoerceI co (Var rhs_var)
1122 %************************************************************************
1126 %************************************************************************
1130 = OccEnv { occ_encl :: !OccEncl -- Enclosing context information
1131 , occ_ctxt :: !CtxtTy -- Tells about linearity
1132 , occ_proxy :: ProxyEnv
1133 , occ_rule_fvs :: ImpRuleUsage
1134 , occ_rule_act :: Maybe (Activation -> Bool) -- Nothing => Rules are inactive
1135 -- See Note [Finding rule RHS free vars]
1139 -----------------------------
1140 -- OccEncl is used to control whether to inline into constructor arguments
1142 -- x = (p,q) -- Don't inline p or q
1143 -- y = /\a -> (p a, q a) -- Still don't inline p or q
1144 -- z = f (p,q) -- Do inline p,q; it may make a rule fire
1145 -- So OccEncl tells enought about the context to know what to do when
1146 -- we encounter a contructor application or PAP.
1149 = OccRhs -- RHS of let(rec), albeit perhaps inside a type lambda
1150 -- Don't inline into constructor args here
1151 | OccVanilla -- Argument of function, body of lambda, scruintee of case etc.
1152 -- Do inline into constructor args here
1154 instance Outputable OccEncl where
1155 ppr OccRhs = ptext (sLit "occRhs")
1156 ppr OccVanilla = ptext (sLit "occVanilla")
1158 type CtxtTy = [Bool]
1161 -- True:ctxt Analysing a function-valued expression that will be
1162 -- applied just once
1164 -- False:ctxt Analysing a function-valued expression that may
1165 -- be applied many times; but when it is,
1166 -- the CtxtTy inside applies
1168 initOccEnv :: Maybe (Activation -> Bool) -> [CoreRule]
1170 initOccEnv active_rule imp_rules
1171 = OccEnv { occ_encl = OccVanilla
1173 , occ_proxy = PE emptyVarEnv emptyVarSet
1174 , occ_rule_fvs = findImpRuleUsage active_rule imp_rules
1175 , occ_rule_act = active_rule }
1177 vanillaCtxt :: OccEnv -> OccEnv
1178 vanillaCtxt env = env { occ_encl = OccVanilla, occ_ctxt = [] }
1180 rhsCtxt :: OccEnv -> OccEnv
1181 rhsCtxt env = env { occ_encl = OccRhs, occ_ctxt = [] }
1183 setCtxtTy :: OccEnv -> CtxtTy -> OccEnv
1184 setCtxtTy env ctxt = env { occ_ctxt = ctxt }
1186 isRhsEnv :: OccEnv -> Bool
1187 isRhsEnv (OccEnv { occ_encl = OccRhs }) = True
1188 isRhsEnv (OccEnv { occ_encl = OccVanilla }) = False
1190 oneShotGroup :: OccEnv -> [CoreBndr] -> [CoreBndr]
1191 -- The result binders have one-shot-ness set that they might not have had originally.
1192 -- This happens in (build (\cn -> e)). Here the occurrence analyser
1193 -- linearity context knows that c,n are one-shot, and it records that fact in
1194 -- the binder. This is useful to guide subsequent float-in/float-out tranformations
1196 oneShotGroup (OccEnv { occ_ctxt = ctxt }) bndrs
1199 go _ [] rev_bndrs = reverse rev_bndrs
1201 go (lin_ctxt:ctxt) (bndr:bndrs) rev_bndrs
1202 | isId bndr = go ctxt bndrs (bndr':rev_bndrs)
1204 bndr' | lin_ctxt = setOneShotLambda bndr
1207 go ctxt (bndr:bndrs) rev_bndrs = go ctxt bndrs (bndr:rev_bndrs)
1209 addAppCtxt :: OccEnv -> [Arg CoreBndr] -> OccEnv
1210 addAppCtxt env@(OccEnv { occ_ctxt = ctxt }) args
1211 = env { occ_ctxt = replicate (valArgCount args) True ++ ctxt }
1214 %************************************************************************
1218 %************************************************************************
1221 type ImpRuleUsage = NameEnv UsageDetails
1222 -- Maps an *imported* Id f to the UsageDetails for *local* Ids
1223 -- used on the RHS for a *local* rule for f.
1228 Consider this, where A.g is an imported Id
1231 {-# RULE "foo" forall x. A.g x = f x #-}
1233 Obviously there's a loop, but the danger is that the occurrence analyser
1234 will say that 'f' is not a loop breaker. Then the simplifier will
1237 and then gaily inline 'f'. Result infinite loop. More realistically,
1238 these kind of rules are generated when specialising imported INLINABLE Ids.
1240 Solution: treat an occurrence of A.g as an occurrence of all the local Ids
1241 that occur on the RULE's RHS. This mapping from imported Id to local Ids
1242 is held in occ_rule_fvs.
1245 findImpRuleUsage :: Maybe (Activation -> Bool) -> [CoreRule] -> ImpRuleUsage
1246 -- Find the *local* Ids that can be reached transitively,
1247 -- via local rules, from each *imported* Id.
1248 -- Sigh: this function seems more complicated than it is really worth
1249 findImpRuleUsage Nothing _ = emptyNameEnv
1250 findImpRuleUsage (Just is_active) rules
1251 = mkNameEnv [ (f, mapUFM (\_ -> NoOccInfo) ls)
1253 , let ls = find_lcl_deps f
1254 , not (isEmptyVarSet ls) ]
1256 rule_names = map ru_fn rules
1257 rule_name_set = mkNameSet rule_names
1259 imp_deps :: NameEnv VarSet
1260 -- (f,g) means imported Id 'g' appears in RHS of
1261 -- rule for imported Id 'f', *or* does so transitively
1262 imp_deps = foldr add_imp emptyNameEnv rules
1264 | is_active (ruleActivation rule)
1265 = extendNameEnv_C unionVarSet acc (ru_fn rule)
1266 (exprSomeFreeVars keep_imp (ru_rhs rule))
1268 keep_imp v = isId v && (idName v `elemNameSet` rule_name_set)
1269 full_imp_deps = transClosureFV (ufmToList imp_deps)
1271 lcl_deps :: NameEnv VarSet
1272 -- (f, l) means localId 'l' appears immediately
1273 -- in the RHS of a rule for imported Id 'f'
1274 -- Remember, many rules might have the same ru_fn
1275 -- so we do need to fold
1276 lcl_deps = foldr add_lcl emptyNameEnv rules
1277 add_lcl rule acc = extendNameEnv_C unionVarSet acc (ru_fn rule)
1278 (exprFreeIds (ru_rhs rule))
1280 find_lcl_deps :: Name -> VarSet
1282 = foldVarSet (unionVarSet . lookup_lcl . idName) (lookup_lcl f)
1283 (lookupNameEnv full_imp_deps f `orElse` emptyVarSet)
1284 lookup_lcl :: Name -> VarSet
1285 lookup_lcl g = lookupNameEnv lcl_deps g `orElse` emptyVarSet
1288 transClosureFV :: Uniquable a => [(a, VarSet)] -> UniqFM VarSet
1289 -- If (f,g), (g,h) are in the input, then (f,h) is in the output
1290 transClosureFV fv_list
1292 | otherwise = transClosureFV new_fv_list
1294 env = listToUFM fv_list
1295 (no_change, new_fv_list) = mapAccumL bump True fv_list
1296 bump no_change (b,fvs)
1297 | no_change_here = (no_change, (b,fvs))
1298 | otherwise = (False, (b,new_fvs))
1300 (new_fvs, no_change_here) = extendFvs env fvs
1303 extendFvs :: UniqFM VarSet -> VarSet -> (VarSet, Bool)
1304 -- (extendFVs env s) returns
1305 -- (s `union` env(s), env(s) `subset` s)
1307 = foldVarSet add (s, True) s
1309 add v (vs, no_change_so_far)
1310 = case lookupUFM env v of
1311 Just fvs | not (fvs `subVarSet` s)
1312 -> (vs `unionVarSet` fvs, False)
1313 _ -> (vs, no_change_so_far)
1317 %************************************************************************
1321 %************************************************************************
1325 = PE (IdEnv (Id, [(Id,CoercionI)])) VarSet
1326 -- Main env, and its free variables (of both range and domain)
1331 The ProxyEnv keeps track of the connection between case binders and
1332 scrutinee. Specifically, if
1333 sc |-> (sc, [...(cb, co)...])
1334 is a binding in the ProxyEnv, then
1336 Typically we add such a binding when encountering the case expression
1337 case (sc |> coi) of cb { ... }
1340 * The domain of the ProxyEnv is the variable (or casted variable)
1341 scrutinees of enclosing cases. This is additionally used
1342 to ensure we gather occurrence info even for GlobalId scrutinees;
1343 see Note [Binder swap for GlobalId scrutinee]
1345 * The ProxyEnv is just an optimisation; you can throw away any
1346 element without losing correctness. And we do so when pushing
1347 it inside a binding (see trimProxyEnv).
1349 * One scrutinee might map to many case binders: Eg
1350 case sc of cb1 { DEFAULT -> ....case sc of cb2 { ... } .. }
1353 * If sc1 |-> (sc2, [...(cb, co)...]), then sc1==sc2
1354 It's a UniqFM and we sometimes need the domain Id
1356 * Any particular case binder 'cb' occurs only once in entire range
1360 The Main Reason for having a ProxyEnv is so that when we encounter
1361 case e of cb { pi -> ri }
1362 we can find all the in-scope variables derivable from 'cb',
1363 and effectively add let-bindings for them (or at least for the
1364 ones *mentioned* in ri) thus:
1365 case e of cb { pi -> let { x = ..cb..; y = ...cb.. }
1367 In this way we'll replace occurrences of 'x', 'y' with 'cb',
1368 which implements the Binder-swap idea (see Note [Binder swap])
1370 The function getProxies finds these bindings; then we
1371 add just the necessary ones, using wrapProxy.
1375 We do these two transformations right here:
1377 (1) case x of b { pi -> ri }
1379 case x of b { pi -> let x=b in ri }
1381 (2) case (x |> co) of b { pi -> ri }
1383 case (x |> co) of b { pi -> let x = b |> sym co in ri }
1385 Why (2)? See Note [Case of cast]
1387 In both cases, in a particular alternative (pi -> ri), we only
1389 (a) x occurs free in (pi -> ri)
1390 (ie it occurs in ri, but is not bound in pi)
1391 (b) the pi does not bind b (or the free vars of co)
1392 We need (a) and (b) for the inserted binding to be correct.
1394 For the alternatives where we inject the binding, we can transfer
1395 all x's OccInfo to b. And that is the point.
1398 * The deliberate shadowing of 'x'.
1399 * That (a) rapidly becomes false, so no bindings are injected.
1401 The reason for doing these transformations here is because it allows
1402 us to adjust the OccInfo for 'x' and 'b' as we go.
1404 * Suppose the only occurrences of 'x' are the scrutinee and in the
1405 ri; then this transformation makes it occur just once, and hence
1406 get inlined right away.
1408 * If we do this in the Simplifier, we don't know whether 'x' is used
1409 in ri, so we are forced to pessimistically zap b's OccInfo even
1410 though it is typically dead (ie neither it nor x appear in the
1411 ri). There's nothing actually wrong with zapping it, except that
1412 it's kind of nice to know which variables are dead. My nose
1413 tells me to keep this information as robustly as possible.
1415 The Maybe (Id,CoreExpr) passed to occAnalAlt is the extra let-binding
1416 {x=b}; it's Nothing if the binder-swap doesn't happen.
1418 There is a danger though. Consider
1420 in case (f v) of w -> ...v...v...
1421 And suppose that (f v) expands to just v. Then we'd like to
1422 use 'w' instead of 'v' in the alternative. But it may be too
1423 late; we may have substituted the (cheap) x+#y for v in the
1424 same simplifier pass that reduced (f v) to v.
1426 I think this is just too bad. CSE will recover some of it.
1430 Consider case (x `cast` co) of b { I# ->
1431 ... (case (x `cast` co) of {...}) ...
1432 We'd like to eliminate the inner case. That is the motivation for
1433 equation (2) in Note [Binder swap]. When we get to the inner case, we
1434 inline x, cancel the casts, and away we go.
1436 Note [Binder swap on GlobalId scrutinees]
1437 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
1438 When the scrutinee is a GlobalId we must take care in two ways
1440 i) In order to *know* whether 'x' occurs free in the RHS, we need its
1441 occurrence info. BUT, we don't gather occurrence info for
1442 GlobalIds. That's one use for the (small) occ_proxy env in OccEnv is
1443 for: it says "gather occurrence info for these.
1445 ii) We must call localiseId on 'x' first, in case it's a GlobalId, or
1446 has an External Name. See, for example, SimplEnv Note [Global Ids in
1449 Note [getProxies is subtle]
1450 ~~~~~~~~~~~~~~~~~~~~~~~~~~~
1451 The code for getProxies isn't all that obvious. Consider
1453 case v |> cov of x { DEFAULT ->
1454 case x |> cox1 of y { DEFAULT ->
1455 case x |> cox2 of z { DEFAULT -> r
1457 These will give us a ProxyEnv looking like:
1458 x |-> (x, [(y, cox1), (z, cox2)])
1459 v |-> (v, [(x, cov)])
1461 From this we want to extract the bindings
1466 Notice that later bindings may mention earlier ones, and that
1467 we need to go "both ways".
1469 Historical note [no-case-of-case]
1470 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
1471 We *used* to suppress the binder-swap in case expressions when
1472 -fno-case-of-case is on. Old remarks:
1473 "This happens in the first simplifier pass,
1474 and enhances full laziness. Here's the bad case:
1475 f = \ y -> ...(case x of I# v -> ...(case x of ...) ... )
1476 If we eliminate the inner case, we trap it inside the I# v -> arm,
1477 which might prevent some full laziness happening. I've seen this
1478 in action in spectral/cichelli/Prog.hs:
1479 [(m,n) | m <- [1..max], n <- [1..max]]
1480 Hence the check for NoCaseOfCase."
1481 However, now the full-laziness pass itself reverses the binder-swap, so this
1482 check is no longer necessary.
1484 Historical note [Suppressing the case binder-swap]
1485 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
1486 This old note describes a problem that is also fixed by doing the
1487 binder-swap in OccAnal:
1489 There is another situation when it might make sense to suppress the
1490 case-expression binde-swap. If we have
1492 case x of w1 { DEFAULT -> case x of w2 { A -> e1; B -> e2 }
1493 ...other cases .... }
1495 We'll perform the binder-swap for the outer case, giving
1497 case x of w1 { DEFAULT -> case w1 of w2 { A -> e1; B -> e2 }
1498 ...other cases .... }
1500 But there is no point in doing it for the inner case, because w1 can't
1501 be inlined anyway. Furthermore, doing the case-swapping involves
1502 zapping w2's occurrence info (see paragraphs that follow), and that
1503 forces us to bind w2 when doing case merging. So we get
1505 case x of w1 { A -> let w2 = w1 in e1
1506 B -> let w2 = w1 in e2
1507 ...other cases .... }
1509 This is plain silly in the common case where w2 is dead.
1511 Even so, I can't see a good way to implement this idea. I tried
1512 not doing the binder-swap if the scrutinee was already evaluated
1513 but that failed big-time:
1517 case v of w { MkT x ->
1518 case x of x1 { I# y1 ->
1519 case x of x2 { I# y2 -> ...
1521 Notice that because MkT is strict, x is marked "evaluated". But to
1522 eliminate the last case, we must either make sure that x (as well as
1523 x1) has unfolding MkT y1. THe straightforward thing to do is to do
1524 the binder-swap. So this whole note is a no-op.
1526 It's fixed by doing the binder-swap in OccAnal because we can do the
1527 binder-swap unconditionally and still get occurrence analysis
1531 extendProxyEnv :: ProxyEnv -> Id -> CoercionI -> Id -> ProxyEnv
1532 -- (extendPE x co y) typically arises from
1533 -- case (x |> co) of y { ... }
1534 -- It extends the proxy env with the binding
1536 extendProxyEnv pe scrut co case_bndr
1537 | scrut == case_bndr = PE env1 fvs1 -- If case_bndr shadows scrut,
1538 | otherwise = PE env2 fvs2 -- don't extend
1540 PE env1 fvs1 = trimProxyEnv pe [case_bndr]
1541 env2 = extendVarEnv_Acc add single env1 scrut1 (case_bndr,co)
1542 single cb_co = (scrut1, [cb_co])
1543 add cb_co (x, cb_cos) = (x, cb_co:cb_cos)
1544 fvs2 = fvs1 `unionVarSet` freeVarsCoI co
1545 `extendVarSet` case_bndr
1546 `extendVarSet` scrut1
1548 scrut1 = mkLocalId (localiseName (idName scrut)) (idType scrut)
1549 -- Localise the scrut_var before shadowing it; we're making a
1550 -- new binding for it, and it might have an External Name, or
1551 -- even be a GlobalId; Note [Binder swap on GlobalId scrutinees]
1552 -- Also we don't want any INLILNE or NOINLINE pragmas!
1555 type ProxyBind = (Id, Id, CoercionI)
1557 getProxies :: OccEnv -> Id -> Bag ProxyBind
1558 -- Return a bunch of bindings [...(xi,ei)...]
1559 -- such that let { ...; xi=ei; ... } binds the xi using y alone
1560 -- See Note [getProxies is subtle]
1561 getProxies (OccEnv { occ_proxy = PE pe _ }) case_bndr
1562 = -- pprTrace "wrapProxies" (ppr case_bndr) $
1565 fwd_pe :: IdEnv (Id, CoercionI)
1566 fwd_pe = foldVarEnv add1 emptyVarEnv pe
1568 add1 (x,ycos) env = foldr (add2 x) env ycos
1569 add2 x (y,co) env = extendVarEnv env y (x,co)
1571 go_fwd :: Id -> Bag ProxyBind
1572 -- Return bindings derivable from case_bndr
1573 go_fwd case_bndr = -- pprTrace "go_fwd" (vcat [ppr case_bndr, text "fwd_pe =" <+> ppr fwd_pe,
1574 -- text "pe =" <+> ppr pe]) $
1578 | Just (scrut, co) <- lookupVarEnv fwd_pe case_bndr
1579 = unitBag (scrut, case_bndr, mkSymCoI co)
1580 `unionBags` go_fwd scrut
1581 `unionBags` go_bwd scrut [pr | pr@(cb,_) <- lookup_bwd scrut
1586 lookup_bwd :: Id -> [(Id, CoercionI)]
1587 -- Return case_bndrs that are connected to scrut
1588 lookup_bwd scrut = case lookupVarEnv pe scrut of
1590 Just (_, cb_cos) -> cb_cos
1592 go_bwd :: Id -> [(Id, CoercionI)] -> Bag ProxyBind
1593 go_bwd scrut cb_cos = foldr (unionBags . go_bwd1 scrut) emptyBag cb_cos
1595 go_bwd1 :: Id -> (Id, CoercionI) -> Bag ProxyBind
1596 go_bwd1 scrut (case_bndr, co)
1597 = -- pprTrace "go_bwd1" (ppr case_bndr) $
1598 unitBag (case_bndr, scrut, co)
1599 `unionBags` go_bwd case_bndr (lookup_bwd case_bndr)
1602 mkAltEnv :: OccEnv -> CoreExpr -> Id -> OccEnv
1603 -- Does two things: a) makes the occ_ctxt = OccVanilla
1604 -- b) extends the ProxyEnv if possible
1605 mkAltEnv env scrut cb
1606 = env { occ_encl = OccVanilla, occ_proxy = pe' }
1610 Var v -> extendProxyEnv pe v (IdCo (idType v)) cb
1611 Cast (Var v) co -> extendProxyEnv pe v (ACo co) cb
1612 _other -> trimProxyEnv pe [cb]
1615 trimOccEnv :: OccEnv -> [CoreBndr] -> OccEnv
1616 trimOccEnv env bndrs = env { occ_proxy = trimProxyEnv (occ_proxy env) bndrs }
1619 trimProxyEnv :: ProxyEnv -> [CoreBndr] -> ProxyEnv
1620 -- We are about to push this ProxyEnv inside a binding for 'bndrs'
1621 -- So dump any ProxyEnv bindings which mention any of the bndrs
1622 trimProxyEnv (PE pe fvs) bndrs
1623 | not (bndr_set `intersectsVarSet` fvs)
1626 = PE pe' (fvs `minusVarSet` bndr_set)
1628 pe' = mapVarEnv trim pe
1629 bndr_set = mkVarSet bndrs
1630 trim (scrut, cb_cos) | scrut `elemVarSet` bndr_set = (scrut, [])
1631 | otherwise = (scrut, filterOut discard cb_cos)
1632 discard (cb,co) = bndr_set `intersectsVarSet`
1633 extendVarSet (freeVarsCoI co) cb
1636 freeVarsCoI :: CoercionI -> VarSet
1637 freeVarsCoI (IdCo t) = tyVarsOfType t
1638 freeVarsCoI (ACo co) = tyVarsOfType co
1642 %************************************************************************
1644 \subsection[OccurAnal-types]{OccEnv}
1646 %************************************************************************
1649 type UsageDetails = IdEnv OccInfo -- A finite map from ids to their usage
1650 -- INVARIANT: never IAmDead
1651 -- (Deadness is signalled by not being in the map at all)
1653 (+++), combineAltsUsageDetails
1654 :: UsageDetails -> UsageDetails -> UsageDetails
1657 = plusVarEnv_C addOccInfo usage1 usage2
1659 combineAltsUsageDetails usage1 usage2
1660 = plusVarEnv_C orOccInfo usage1 usage2
1662 addOneOcc :: UsageDetails -> Id -> OccInfo -> UsageDetails
1663 addOneOcc usage id info
1664 = plusVarEnv_C addOccInfo usage (unitVarEnv id info)
1665 -- ToDo: make this more efficient
1667 emptyDetails :: UsageDetails
1668 emptyDetails = (emptyVarEnv :: UsageDetails)
1670 usedIn :: Id -> UsageDetails -> Bool
1671 v `usedIn` details = isExportedId v || v `elemVarEnv` details
1673 type IdWithOccInfo = Id
1675 tagLamBinders :: UsageDetails -- Of scope
1677 -> (UsageDetails, -- Details with binders removed
1678 [IdWithOccInfo]) -- Tagged binders
1679 -- Used for lambda and case binders
1680 -- It copes with the fact that lambda bindings can have InlineRule
1681 -- unfoldings, used for join points
1682 tagLamBinders usage binders = usage' `seq` (usage', bndrs')
1684 (usage', bndrs') = mapAccumR tag_lam usage binders
1685 tag_lam usage bndr = (usage2, setBinderOcc usage bndr)
1687 usage1 = usage `delVarEnv` bndr
1688 usage2 | isId bndr = addIdOccs usage1 (idUnfoldingVars bndr)
1689 | otherwise = usage1
1691 tagBinder :: UsageDetails -- Of scope
1693 -> (UsageDetails, -- Details with binders removed
1694 IdWithOccInfo) -- Tagged binders
1696 tagBinder usage binder
1698 usage' = usage `delVarEnv` binder
1699 binder' = setBinderOcc usage binder
1701 usage' `seq` (usage', binder')
1703 setBinderOcc :: UsageDetails -> CoreBndr -> CoreBndr
1704 setBinderOcc usage bndr
1705 | isTyCoVar bndr = bndr
1706 | isExportedId bndr = case idOccInfo bndr of
1708 _ -> setIdOccInfo bndr NoOccInfo
1709 -- Don't use local usage info for visible-elsewhere things
1710 -- BUT *do* erase any IAmALoopBreaker annotation, because we're
1711 -- about to re-generate it and it shouldn't be "sticky"
1713 | otherwise = setIdOccInfo bndr occ_info
1715 occ_info = lookupVarEnv usage bndr `orElse` IAmDead
1719 %************************************************************************
1721 \subsection{Operations over OccInfo}
1723 %************************************************************************
1726 mkOneOcc :: OccEnv -> Id -> InterestingCxt -> UsageDetails
1727 mkOneOcc env id int_cxt
1728 | isLocalId id = unitVarEnv id (OneOcc False True int_cxt)
1729 | PE env _ <- occ_proxy env
1730 , id `elemVarEnv` env = unitVarEnv id NoOccInfo
1731 | Just uds <- lookupNameEnv (occ_rule_fvs env) (idName id)
1733 | otherwise = emptyDetails
1735 markMany, markInsideLam, markInsideSCC :: OccInfo -> OccInfo
1737 markMany _ = NoOccInfo
1739 markInsideSCC occ = markMany occ
1741 markInsideLam (OneOcc _ one_br int_cxt) = OneOcc True one_br int_cxt
1742 markInsideLam occ = occ
1744 addOccInfo, orOccInfo :: OccInfo -> OccInfo -> OccInfo
1746 addOccInfo a1 a2 = ASSERT( not (isDeadOcc a1 || isDeadOcc a2) )
1747 NoOccInfo -- Both branches are at least One
1748 -- (Argument is never IAmDead)
1750 -- (orOccInfo orig new) is used
1751 -- when combining occurrence info from branches of a case
1753 orOccInfo (OneOcc in_lam1 _ int_cxt1)
1754 (OneOcc in_lam2 _ int_cxt2)
1755 = OneOcc (in_lam1 || in_lam2)
1756 False -- False, because it occurs in both branches
1757 (int_cxt1 && int_cxt2)
1758 orOccInfo a1 a2 = ASSERT( not (isDeadOcc a1 || isDeadOcc a2) )