2 % (c) The AQUA Project, Glasgow University, 1993-1998
4 \section[Simplify]{The main module of the simplifier}
7 module Simplify ( simplTopBinds, simplExpr ) where
9 #include "HsVersions.h"
13 import Type hiding ( substTy, extendTvSubst )
16 import FamInstEnv ( FamInstEnv )
18 import MkId ( mkImpossibleExpr, seqId )
21 import Name ( mkSystemVarName )
23 import FamInstEnv ( topNormaliseType )
24 import DataCon ( DataCon, dataConWorkId, dataConRepStrictness )
26 import NewDemand ( isStrictDmd, splitStrictSig )
27 import PprCore ( pprParendExpr, pprCoreExpr )
28 import CoreUnfold ( mkUnfolding, mkCoreUnfolding, mkInlineRule,
29 exprIsConApp_maybe, callSiteInline, CallCtxt(..) )
31 import qualified CoreSubst
32 import CoreArity ( exprArity )
33 import Rules ( lookupRule, getRules )
34 import BasicTypes ( isMarkedStrict, Arity )
35 import CostCentre ( currentCCS, pushCCisNop )
36 import TysPrim ( realWorldStatePrimTy )
37 import PrelInfo ( realWorldPrimId )
38 import BasicTypes ( TopLevelFlag(..), isTopLevel,
39 RecFlag(..), isNonRuleLoopBreaker )
40 import MonadUtils ( foldlM, mapAccumLM )
41 import Maybes ( orElse )
42 import Data.List ( mapAccumL )
48 The guts of the simplifier is in this module, but the driver loop for
49 the simplifier is in SimplCore.lhs.
52 -----------------------------------------
53 *** IMPORTANT NOTE ***
54 -----------------------------------------
55 The simplifier used to guarantee that the output had no shadowing, but
56 it does not do so any more. (Actually, it never did!) The reason is
57 documented with simplifyArgs.
60 -----------------------------------------
61 *** IMPORTANT NOTE ***
62 -----------------------------------------
63 Many parts of the simplifier return a bunch of "floats" as well as an
64 expression. This is wrapped as a datatype SimplUtils.FloatsWith.
66 All "floats" are let-binds, not case-binds, but some non-rec lets may
67 be unlifted (with RHS ok-for-speculation).
71 -----------------------------------------
72 ORGANISATION OF FUNCTIONS
73 -----------------------------------------
75 - simplify all top-level binders
76 - for NonRec, call simplRecOrTopPair
77 - for Rec, call simplRecBind
80 ------------------------------
81 simplExpr (applied lambda) ==> simplNonRecBind
82 simplExpr (Let (NonRec ...) ..) ==> simplNonRecBind
83 simplExpr (Let (Rec ...) ..) ==> simplify binders; simplRecBind
85 ------------------------------
86 simplRecBind [binders already simplfied]
87 - use simplRecOrTopPair on each pair in turn
89 simplRecOrTopPair [binder already simplified]
90 Used for: recursive bindings (top level and nested)
91 top-level non-recursive bindings
93 - check for PreInlineUnconditionally
97 Used for: non-top-level non-recursive bindings
98 beta reductions (which amount to the same thing)
99 Because it can deal with strict arts, it takes a
100 "thing-inside" and returns an expression
102 - check for PreInlineUnconditionally
103 - simplify binder, including its IdInfo
112 simplNonRecX: [given a *simplified* RHS, but an *unsimplified* binder]
113 Used for: binding case-binder and constr args in a known-constructor case
114 - check for PreInLineUnconditionally
118 ------------------------------
119 simplLazyBind: [binder already simplified, RHS not]
120 Used for: recursive bindings (top level and nested)
121 top-level non-recursive bindings
122 non-top-level, but *lazy* non-recursive bindings
123 [must not be strict or unboxed]
124 Returns floats + an augmented environment, not an expression
125 - substituteIdInfo and add result to in-scope
126 [so that rules are available in rec rhs]
129 - float if exposes constructor or PAP
133 completeNonRecX: [binder and rhs both simplified]
134 - if the the thing needs case binding (unlifted and not ok-for-spec)
140 completeBind: [given a simplified RHS]
141 [used for both rec and non-rec bindings, top level and not]
142 - try PostInlineUnconditionally
143 - add unfolding [this is the only place we add an unfolding]
148 Right hand sides and arguments
149 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
150 In many ways we want to treat
151 (a) the right hand side of a let(rec), and
152 (b) a function argument
153 in the same way. But not always! In particular, we would
154 like to leave these arguments exactly as they are, so they
155 will match a RULE more easily.
160 It's harder to make the rule match if we ANF-ise the constructor,
161 or eta-expand the PAP:
163 f (let { a = g x; b = h x } in (a,b))
166 On the other hand if we see the let-defns
171 then we *do* want to ANF-ise and eta-expand, so that p and q
172 can be safely inlined.
174 Even floating lets out is a bit dubious. For let RHS's we float lets
175 out if that exposes a value, so that the value can be inlined more vigorously.
178 r = let x = e in (x,x)
180 Here, if we float the let out we'll expose a nice constructor. We did experiments
181 that showed this to be a generally good thing. But it was a bad thing to float
182 lets out unconditionally, because that meant they got allocated more often.
184 For function arguments, there's less reason to expose a constructor (it won't
185 get inlined). Just possibly it might make a rule match, but I'm pretty skeptical.
186 So for the moment we don't float lets out of function arguments either.
191 For eta expansion, we want to catch things like
193 case e of (a,b) -> \x -> case a of (p,q) -> \y -> r
195 If the \x was on the RHS of a let, we'd eta expand to bring the two
196 lambdas together. And in general that's a good thing to do. Perhaps
197 we should eta expand wherever we find a (value) lambda? Then the eta
198 expansion at a let RHS can concentrate solely on the PAP case.
201 %************************************************************************
203 \subsection{Bindings}
205 %************************************************************************
208 simplTopBinds :: SimplEnv -> [InBind] -> SimplM SimplEnv
210 simplTopBinds env0 binds0
211 = do { -- Put all the top-level binders into scope at the start
212 -- so that if a transformation rule has unexpectedly brought
213 -- anything into scope, then we don't get a complaint about that.
214 -- It's rather as if the top-level binders were imported.
215 ; env1 <- simplRecBndrs env0 (bindersOfBinds binds0)
216 ; dflags <- getDOptsSmpl
217 ; let dump_flag = dopt Opt_D_dump_inlinings dflags ||
218 dopt Opt_D_dump_rule_firings dflags
219 ; env2 <- simpl_binds dump_flag env1 binds0
220 ; freeTick SimplifierDone
223 -- We need to track the zapped top-level binders, because
224 -- they should have their fragile IdInfo zapped (notably occurrence info)
225 -- That's why we run down binds and bndrs' simultaneously.
227 -- The dump-flag emits a trace for each top-level binding, which
228 -- helps to locate the tracing for inlining and rule firing
229 simpl_binds :: Bool -> SimplEnv -> [InBind] -> SimplM SimplEnv
230 simpl_binds _ env [] = return env
231 simpl_binds dump env (bind:binds) = do { env' <- trace_bind dump bind $
233 ; simpl_binds dump env' binds }
235 trace_bind True bind = pprTrace "SimplBind" (ppr (bindersOf bind))
236 trace_bind False _ = \x -> x
238 simpl_bind env (Rec pairs) = simplRecBind env TopLevel pairs
239 simpl_bind env (NonRec b r) = simplRecOrTopPair env' TopLevel b b' r
241 (env', b') = addBndrRules env b (lookupRecBndr env b)
245 %************************************************************************
247 \subsection{Lazy bindings}
249 %************************************************************************
251 simplRecBind is used for
252 * recursive bindings only
255 simplRecBind :: SimplEnv -> TopLevelFlag
258 simplRecBind env0 top_lvl pairs0
259 = do { let (env_with_info, triples) = mapAccumL add_rules env0 pairs0
260 ; env1 <- go (zapFloats env_with_info) triples
261 ; return (env0 `addRecFloats` env1) }
262 -- addFloats adds the floats from env1,
263 -- _and_ updates env0 with the in-scope set from env1
265 add_rules :: SimplEnv -> (InBndr,InExpr) -> (SimplEnv, (InBndr, OutBndr, InExpr))
266 -- Add the (substituted) rules to the binder
267 add_rules env (bndr, rhs) = (env', (bndr, bndr', rhs))
269 (env', bndr') = addBndrRules env bndr (lookupRecBndr env bndr)
271 go env [] = return env
273 go env ((old_bndr, new_bndr, rhs) : pairs)
274 = do { env' <- simplRecOrTopPair env top_lvl old_bndr new_bndr rhs
278 simplOrTopPair is used for
279 * recursive bindings (whether top level or not)
280 * top-level non-recursive bindings
282 It assumes the binder has already been simplified, but not its IdInfo.
285 simplRecOrTopPair :: SimplEnv
287 -> InId -> OutBndr -> InExpr -- Binder and rhs
288 -> SimplM SimplEnv -- Returns an env that includes the binding
290 simplRecOrTopPair env top_lvl old_bndr new_bndr rhs
291 | preInlineUnconditionally env top_lvl old_bndr rhs -- Check for unconditional inline
292 = do { tick (PreInlineUnconditionally old_bndr)
293 ; return (extendIdSubst env old_bndr (mkContEx env rhs)) }
296 = simplLazyBind env top_lvl Recursive old_bndr new_bndr rhs env
297 -- May not actually be recursive, but it doesn't matter
301 simplLazyBind is used for
302 * [simplRecOrTopPair] recursive bindings (whether top level or not)
303 * [simplRecOrTopPair] top-level non-recursive bindings
304 * [simplNonRecE] non-top-level *lazy* non-recursive bindings
307 1. It assumes that the binder is *already* simplified,
308 and is in scope, and its IdInfo too, except unfolding
310 2. It assumes that the binder type is lifted.
312 3. It does not check for pre-inline-unconditionallly;
313 that should have been done already.
316 simplLazyBind :: SimplEnv
317 -> TopLevelFlag -> RecFlag
318 -> InId -> OutId -- Binder, both pre-and post simpl
319 -- The OutId has IdInfo, except arity, unfolding
320 -> InExpr -> SimplEnv -- The RHS and its environment
323 simplLazyBind env top_lvl is_rec bndr bndr1 rhs rhs_se
324 = do { let rhs_env = rhs_se `setInScope` env
325 (tvs, body) = case collectTyBinders rhs of
326 (tvs, body) | not_lam body -> (tvs,body)
327 | otherwise -> ([], rhs)
328 not_lam (Lam _ _) = False
330 -- Do not do the "abstract tyyvar" thing if there's
331 -- a lambda inside, becuase it defeats eta-reduction
332 -- f = /\a. \x. g a x
335 ; (body_env, tvs') <- simplBinders rhs_env tvs
336 -- See Note [Floating and type abstraction] in SimplUtils
339 ; (body_env1, body1) <- simplExprF body_env body mkRhsStop
340 -- ANF-ise a constructor or PAP rhs
341 ; (body_env2, body2) <- prepareRhs body_env1 bndr1 body1
344 <- if not (doFloatFromRhs top_lvl is_rec False body2 body_env2)
345 then -- No floating, just wrap up!
346 do { rhs' <- mkLam env tvs' (wrapFloats body_env2 body2)
347 ; return (env, rhs') }
349 else if null tvs then -- Simple floating
350 do { tick LetFloatFromLet
351 ; return (addFloats env body_env2, body2) }
353 else -- Do type-abstraction first
354 do { tick LetFloatFromLet
355 ; (poly_binds, body3) <- abstractFloats tvs' body_env2 body2
356 ; rhs' <- mkLam env tvs' body3
357 ; env' <- foldlM (addPolyBind top_lvl) env poly_binds
358 ; return (env', rhs') }
360 ; completeBind env' top_lvl bndr bndr1 rhs' }
363 A specialised variant of simplNonRec used when the RHS is already simplified,
364 notably in knownCon. It uses case-binding where necessary.
367 simplNonRecX :: SimplEnv
368 -> InId -- Old binder
369 -> OutExpr -- Simplified RHS
372 simplNonRecX env bndr new_rhs
373 | isDeadBinder bndr -- Not uncommon; e.g. case (a,b) of b { (p,q) -> p }
374 = return env -- Here b is dead, and we avoid creating
375 | otherwise -- the binding b = (a,b)
376 = do { (env', bndr') <- simplBinder env bndr
377 ; completeNonRecX env' (isStrictId bndr) bndr bndr' new_rhs }
379 completeNonRecX :: SimplEnv
381 -> InId -- Old binder
382 -> OutId -- New binder
383 -> OutExpr -- Simplified RHS
386 completeNonRecX env is_strict old_bndr new_bndr new_rhs
387 = do { (env1, rhs1) <- prepareRhs (zapFloats env) new_bndr new_rhs
389 if doFloatFromRhs NotTopLevel NonRecursive is_strict rhs1 env1
390 then do { tick LetFloatFromLet
391 ; return (addFloats env env1, rhs1) } -- Add the floats to the main env
392 else return (env, wrapFloats env1 rhs1) -- Wrap the floats around the RHS
393 ; completeBind env2 NotTopLevel old_bndr new_bndr rhs2 }
396 {- No, no, no! Do not try preInlineUnconditionally in completeNonRecX
397 Doing so risks exponential behaviour, because new_rhs has been simplified once already
398 In the cases described by the folowing commment, postInlineUnconditionally will
399 catch many of the relevant cases.
400 -- This happens; for example, the case_bndr during case of
401 -- known constructor: case (a,b) of x { (p,q) -> ... }
402 -- Here x isn't mentioned in the RHS, so we don't want to
403 -- create the (dead) let-binding let x = (a,b) in ...
405 -- Similarly, single occurrences can be inlined vigourously
406 -- e.g. case (f x, g y) of (a,b) -> ....
407 -- If a,b occur once we can avoid constructing the let binding for them.
409 Furthermore in the case-binding case preInlineUnconditionally risks extra thunks
410 -- Consider case I# (quotInt# x y) of
411 -- I# v -> let w = J# v in ...
412 -- If we gaily inline (quotInt# x y) for v, we end up building an
414 -- let w = J# (quotInt# x y) in ...
415 -- because quotInt# can fail.
417 | preInlineUnconditionally env NotTopLevel bndr new_rhs
418 = thing_inside (extendIdSubst env bndr (DoneEx new_rhs))
421 ----------------------------------
422 prepareRhs takes a putative RHS, checks whether it's a PAP or
423 constructor application and, if so, converts it to ANF, so that the
424 resulting thing can be inlined more easily. Thus
431 We also want to deal well cases like this
432 v = (f e1 `cast` co) e2
433 Here we want to make e1,e2 trivial and get
434 x1 = e1; x2 = e2; v = (f x1 `cast` co) v2
435 That's what the 'go' loop in prepareRhs does
438 prepareRhs :: SimplEnv -> OutId -> OutExpr -> SimplM (SimplEnv, OutExpr)
439 -- Adds new floats to the env iff that allows us to return a good RHS
440 prepareRhs env id (Cast rhs co) -- Note [Float coercions]
441 | (ty1, _ty2) <- coercionKind co -- Do *not* do this if rhs has an unlifted type
442 , not (isUnLiftedType ty1) -- see Note [Float coercions (unlifted)]
443 = do { (env', rhs') <- makeTrivialWithInfo env sanitised_info rhs
444 ; return (env', Cast rhs' co) }
446 sanitised_info = vanillaIdInfo `setNewStrictnessInfo` newStrictnessInfo info
447 `setNewDemandInfo` newDemandInfo info
450 prepareRhs env0 _ rhs0
451 = do { (_is_val, env1, rhs1) <- go 0 env0 rhs0
452 ; return (env1, rhs1) }
454 go n_val_args env (Cast rhs co)
455 = do { (is_val, env', rhs') <- go n_val_args env rhs
456 ; return (is_val, env', Cast rhs' co) }
457 go n_val_args env (App fun (Type ty))
458 = do { (is_val, env', rhs') <- go n_val_args env fun
459 ; return (is_val, env', App rhs' (Type ty)) }
460 go n_val_args env (App fun arg)
461 = do { (is_val, env', fun') <- go (n_val_args+1) env fun
463 True -> do { (env'', arg') <- makeTrivial env' arg
464 ; return (True, env'', App fun' arg') }
465 False -> return (False, env, App fun arg) }
466 go n_val_args env (Var fun)
467 = return (is_val, env, Var fun)
469 is_val = n_val_args > 0 -- There is at least one arg
470 -- ...and the fun a constructor or PAP
471 && (isConLikeId fun || n_val_args < idArity fun)
472 -- See Note [CONLIKE pragma] in BasicTypes
474 = return (False, env, other)
478 Note [Float coercions]
479 ~~~~~~~~~~~~~~~~~~~~~~
480 When we find the binding
482 we'd like to transform it to
484 x = x `cast` co -- A trivial binding
485 There's a chance that e will be a constructor application or function, or something
486 like that, so moving the coerion to the usage site may well cancel the coersions
487 and lead to further optimisation. Example:
490 data instance T Int = T Int
492 foo :: Int -> Int -> Int
497 go n = case x of { T m -> go (n-m) }
498 -- This case should optimise
500 Note [Preserve strictness when floating coercions]
501 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
502 In the Note [Float coercions] transformation, keep the strictness info.
504 f = e `cast` co -- f has strictness SSL
506 f' = e -- f' also has strictness SSL
507 f = f' `cast` co -- f still has strictness SSL
509 Its not wrong to drop it on the floor, but better to keep it.
511 Note [Float coercions (unlifted)]
512 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
513 BUT don't do [Float coercions] if 'e' has an unlifted type.
516 foo :: Int = (error (# Int,Int #) "urk")
517 `cast` CoUnsafe (# Int,Int #) Int
519 If do the makeTrivial thing to the error call, we'll get
520 foo = case error (# Int,Int #) "urk" of v -> v `cast` ...
521 But 'v' isn't in scope!
523 These strange casts can happen as a result of case-of-case
524 bar = case (case x of { T -> (# 2,3 #); F -> error "urk" }) of
529 makeTrivial :: SimplEnv -> OutExpr -> SimplM (SimplEnv, OutExpr)
530 -- Binds the expression to a variable, if it's not trivial, returning the variable
531 makeTrivial env expr = makeTrivialWithInfo env vanillaIdInfo expr
533 makeTrivialWithInfo :: SimplEnv -> IdInfo -> OutExpr -> SimplM (SimplEnv, OutExpr)
534 -- Propagate strictness and demand info to the new binder
535 -- Note [Preserve strictness when floating coercions]
536 makeTrivialWithInfo env info expr
539 | otherwise -- See Note [Take care] below
540 = do { uniq <- getUniqueM
541 ; let name = mkSystemVarName uniq (fsLit "a")
542 var = mkLocalIdWithInfo name (exprType expr) info
543 ; env' <- completeNonRecX env False var var expr
544 ; return (env', substExpr env' (Var var)) }
545 -- The substitution is needed becase we're constructing a new binding
547 -- And if rhs is of form (rhs1 |> co), then we might get
550 -- and now a's RHS is trivial and can be substituted out, and that
551 -- is what completeNonRecX will do
555 %************************************************************************
557 \subsection{Completing a lazy binding}
559 %************************************************************************
562 * deals only with Ids, not TyVars
563 * takes an already-simplified binder and RHS
564 * is used for both recursive and non-recursive bindings
565 * is used for both top-level and non-top-level bindings
567 It does the following:
568 - tries discarding a dead binding
569 - tries PostInlineUnconditionally
570 - add unfolding [this is the only place we add an unfolding]
573 It does *not* attempt to do let-to-case. Why? Because it is used for
574 - top-level bindings (when let-to-case is impossible)
575 - many situations where the "rhs" is known to be a WHNF
576 (so let-to-case is inappropriate).
578 Nor does it do the atomic-argument thing
581 completeBind :: SimplEnv
582 -> TopLevelFlag -- Flag stuck into unfolding
583 -> InId -- Old binder
584 -> OutId -> OutExpr -- New binder and RHS
586 -- completeBind may choose to do its work
587 -- * by extending the substitution (e.g. let x = y in ...)
588 -- * or by adding to the floats in the envt
590 completeBind env top_lvl old_bndr new_bndr new_rhs
591 = do { let old_info = idInfo old_bndr
592 old_unf = unfoldingInfo old_info
593 occ_info = occInfo old_info
595 ; new_unfolding <- simplUnfolding env top_lvl old_bndr occ_info new_rhs old_unf
597 ; if postInlineUnconditionally env top_lvl new_bndr occ_info new_rhs new_unfolding
598 -- Inline and discard the binding
599 then do { tick (PostInlineUnconditionally old_bndr)
600 ; return (extendIdSubst env old_bndr (DoneEx new_rhs)) }
601 -- Use the substitution to make quite, quite sure that the
602 -- substitution will happen, since we are going to discard the binding
604 else return (addNonRecWithUnf env new_bndr new_rhs new_unfolding) }
606 ------------------------------
607 addPolyBind :: TopLevelFlag -> SimplEnv -> OutBind -> SimplM SimplEnv
608 -- Add a new binding to the environment, complete with its unfolding
609 -- but *do not* do postInlineUnconditionally, because we have already
610 -- processed some of the scope of the binding
611 -- We still want the unfolding though. Consider
613 -- x = /\a. let y = ... in Just y
615 -- Then we float the y-binding out (via abstractFloats and addPolyBind)
616 -- but 'x' may well then be inlined in 'body' in which case we'd like the
617 -- opportunity to inline 'y' too.
619 addPolyBind top_lvl env (NonRec poly_id rhs)
620 = do { unfolding <- simplUnfolding env top_lvl poly_id NoOccInfo rhs noUnfolding
621 -- Assumes that poly_id did not have an INLINE prag
622 -- which is perhaps wrong. ToDo: think about this
623 ; return (addNonRecWithUnf env poly_id rhs unfolding) }
625 addPolyBind _ env bind@(Rec _) = return (extendFloats env bind)
626 -- Hack: letrecs are more awkward, so we extend "by steam"
627 -- without adding unfoldings etc. At worst this leads to
628 -- more simplifier iterations
630 ------------------------------
631 addNonRecWithUnf :: SimplEnv
632 -> OutId -> OutExpr -- New binder and RHS
633 -> Unfolding -- New unfolding
635 addNonRecWithUnf env new_bndr new_rhs new_unfolding
636 = let new_arity = exprArity new_rhs
637 old_arity = idArity new_bndr
638 info1 = idInfo new_bndr `setArityInfo` new_arity
640 -- Unfolding info: Note [Setting the new unfolding]
641 info2 = info1 `setUnfoldingInfo` new_unfolding
643 -- Demand info: Note [Setting the demand info]
644 info3 | isEvaldUnfolding new_unfolding = zapDemandInfo info2 `orElse` info2
647 final_id = new_bndr `setIdInfo` info3
648 dmd_arity = length $ fst $ splitStrictSig $ idNewStrictness new_bndr
650 ASSERT( isId new_bndr )
651 WARN( new_arity < old_arity || new_arity < dmd_arity,
652 (ptext (sLit "Arity decrease:") <+> ppr final_id <+> ppr old_arity
653 <+> ppr new_arity <+> ppr dmd_arity) )
654 -- Note [Arity decrease]
656 final_id `seq` -- This seq forces the Id, and hence its IdInfo,
657 -- and hence any inner substitutions
658 -- pprTrace "Binding" (ppr final_id <+> ppr unfolding) $
659 addNonRec env final_id new_rhs
660 -- The addNonRec adds it to the in-scope set too
662 ------------------------------
663 simplUnfolding :: SimplEnv-> TopLevelFlag
664 -> Id -- Debug output only
665 -> OccInfo -> OutExpr
666 -> Unfolding -> SimplM Unfolding
667 -- Note [Setting the new unfolding]
668 simplUnfolding env _ _ _ _ (DFunUnfolding con ops)
669 = return (DFunUnfolding con ops')
671 ops' = map (CoreSubst.substExpr (mkCoreSubst env)) ops
673 simplUnfolding env top_lvl _ _ _
674 (CoreUnfolding { uf_tmpl = expr, uf_arity = arity
675 , uf_guidance = guide@(InlineRule {}) })
676 = do { expr' <- simplExpr (updMode updModeForInlineRules env) expr
677 -- See Note [Simplifying gently inside InlineRules] in SimplUtils
678 ; let mb_wkr' = CoreSubst.substInlineRuleInfo (mkCoreSubst env) (ir_info guide)
679 ; return (mkCoreUnfolding (isTopLevel top_lvl) expr' arity
680 (guide { ir_info = mb_wkr' })) }
681 -- See Note [Top-level flag on inline rules] in CoreUnfold
683 simplUnfolding _ top_lvl _ occ_info new_rhs _
684 | omit_unfolding = return NoUnfolding
685 | otherwise = return (mkUnfolding (isTopLevel top_lvl) new_rhs)
687 omit_unfolding = isNonRuleLoopBreaker occ_info
690 Note [Arity decrease]
691 ~~~~~~~~~~~~~~~~~~~~~
692 Generally speaking the arity of a binding should not decrease. But it *can*
693 legitimately happen becuase of RULES. Eg
695 where g has arity 2, will have arity 2. But if there's a rewrite rule
697 where h has arity 1, then f's arity will decrease. Here's a real-life example,
698 which is in the output of Specialise:
701 $dm {Arity 2} = \d.\x. op d
702 {-# RULES forall d. $dm Int d = $s$dm #-}
704 dInt = MkD .... opInt ...
705 opInt {Arity 1} = $dm dInt
707 $s$dm {Arity 0} = \x. op dInt }
709 Here opInt has arity 1; but when we apply the rule its arity drops to 0.
710 That's why Specialise goes to a little trouble to pin the right arity
711 on specialised functions too.
713 Note [Setting the new unfolding]
714 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
715 * If there's an INLINE pragma, we simplify the RHS gently. Maybe we
716 should do nothing at all, but simplifying gently might get rid of
719 * If not, we make an unfolding from the new RHS. But *only* for
720 non-loop-breakers. Making loop breakers not have an unfolding at all
721 means that we can avoid tests in exprIsConApp, for example. This is
722 important: if exprIsConApp says 'yes' for a recursive thing, then we
723 can get into an infinite loop
725 If there's an InlineRule on a loop breaker, we hang on to the inlining.
726 It's pretty dodgy, but the user did say 'INLINE'. May need to revisit
729 Note [Setting the demand info]
730 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
731 If the unfolding is a value, the demand info may
732 go pear-shaped, so we nuke it. Example:
734 case x of (p,q) -> h p q x
735 Here x is certainly demanded. But after we've nuked
736 the case, we'll get just
737 let x = (a,b) in h a b x
738 and now x is not demanded (I'm assuming h is lazy)
739 This really happens. Similarly
740 let f = \x -> e in ...f..f...
741 After inlining f at some of its call sites the original binding may
742 (for example) be no longer strictly demanded.
743 The solution here is a bit ad hoc...
746 %************************************************************************
748 \subsection[Simplify-simplExpr]{The main function: simplExpr}
750 %************************************************************************
752 The reason for this OutExprStuff stuff is that we want to float *after*
753 simplifying a RHS, not before. If we do so naively we get quadratic
754 behaviour as things float out.
756 To see why it's important to do it after, consider this (real) example:
770 a -- Can't inline a this round, cos it appears twice
774 Each of the ==> steps is a round of simplification. We'd save a
775 whole round if we float first. This can cascade. Consider
780 let f = let d1 = ..d.. in \y -> e
784 in \x -> ...(\y ->e)...
786 Only in this second round can the \y be applied, and it
787 might do the same again.
791 simplExpr :: SimplEnv -> CoreExpr -> SimplM CoreExpr
792 simplExpr env expr = simplExprC env expr mkBoringStop
794 simplExprC :: SimplEnv -> CoreExpr -> SimplCont -> SimplM CoreExpr
795 -- Simplify an expression, given a continuation
796 simplExprC env expr cont
797 = -- pprTrace "simplExprC" (ppr expr $$ ppr cont {- $$ ppr (seIdSubst env) -} $$ ppr (seFloats env) ) $
798 do { (env', expr') <- simplExprF (zapFloats env) expr cont
799 ; -- pprTrace "simplExprC ret" (ppr expr $$ ppr expr') $
800 -- pprTrace "simplExprC ret3" (ppr (seInScope env')) $
801 -- pprTrace "simplExprC ret4" (ppr (seFloats env')) $
802 return (wrapFloats env' expr') }
804 --------------------------------------------------
805 simplExprF :: SimplEnv -> InExpr -> SimplCont
806 -> SimplM (SimplEnv, OutExpr)
808 simplExprF env e cont
809 = -- pprTrace "simplExprF" (ppr e $$ ppr cont $$ ppr (seTvSubst env) $$ ppr (seIdSubst env) {- $$ ppr (seFloats env) -} ) $
810 simplExprF' env e cont
812 simplExprF' :: SimplEnv -> InExpr -> SimplCont
813 -> SimplM (SimplEnv, OutExpr)
814 simplExprF' env (Var v) cont = simplVar env v cont
815 simplExprF' env (Lit lit) cont = rebuild env (Lit lit) cont
816 simplExprF' env (Note n expr) cont = simplNote env n expr cont
817 simplExprF' env (Cast body co) cont = simplCast env body co cont
818 simplExprF' env (App fun arg) cont = simplExprF env fun $
819 ApplyTo NoDup arg env cont
821 simplExprF' env expr@(Lam _ _) cont
822 = simplLam env (map zap bndrs) body cont
823 -- The main issue here is under-saturated lambdas
824 -- (\x1. \x2. e) arg1
825 -- Here x1 might have "occurs-once" occ-info, because occ-info
826 -- is computed assuming that a group of lambdas is applied
827 -- all at once. If there are too few args, we must zap the
830 n_args = countArgs cont
831 n_params = length bndrs
832 (bndrs, body) = collectBinders expr
833 zap | n_args >= n_params = \b -> b
834 | otherwise = \b -> if isTyVar b then b
836 -- NB: we count all the args incl type args
837 -- so we must count all the binders (incl type lambdas)
839 simplExprF' env (Type ty) cont
840 = ASSERT( contIsRhsOrArg cont )
841 do { ty' <- simplCoercion env ty
842 ; rebuild env (Type ty') cont }
844 simplExprF' env (Case scrut bndr _ alts) cont
845 | not (switchIsOn (getSwitchChecker env) NoCaseOfCase)
846 = -- Simplify the scrutinee with a Select continuation
847 simplExprF env scrut (Select NoDup bndr alts env cont)
850 = -- If case-of-case is off, simply simplify the case expression
851 -- in a vanilla Stop context, and rebuild the result around it
852 do { case_expr' <- simplExprC env scrut case_cont
853 ; rebuild env case_expr' cont }
855 case_cont = Select NoDup bndr alts env mkBoringStop
857 simplExprF' env (Let (Rec pairs) body) cont
858 = do { env' <- simplRecBndrs env (map fst pairs)
859 -- NB: bndrs' don't have unfoldings or rules
860 -- We add them as we go down
862 ; env'' <- simplRecBind env' NotTopLevel pairs
863 ; simplExprF env'' body cont }
865 simplExprF' env (Let (NonRec bndr rhs) body) cont
866 = simplNonRecE env bndr (rhs, env) ([], body) cont
868 ---------------------------------
869 simplType :: SimplEnv -> InType -> SimplM OutType
870 -- Kept monadic just so we can do the seqType
872 = -- pprTrace "simplType" (ppr ty $$ ppr (seTvSubst env)) $
873 seqType new_ty `seq` return new_ty
875 new_ty = substTy env ty
877 ---------------------------------
878 simplCoercion :: SimplEnv -> InType -> SimplM OutType
879 -- The InType isn't *necessarily* a coercion, but it might be
880 -- (in a type application, say) and optCoercion is a no-op on types
882 = do { co' <- simplType env co
883 ; return (optCoercion co') }
887 %************************************************************************
889 \subsection{The main rebuilder}
891 %************************************************************************
894 rebuild :: SimplEnv -> OutExpr -> SimplCont -> SimplM (SimplEnv, OutExpr)
895 -- At this point the substitution in the SimplEnv should be irrelevant
896 -- only the in-scope set and floats should matter
897 rebuild env expr cont0
898 = -- pprTrace "rebuild" (ppr expr $$ ppr cont0 $$ ppr (seFloats env)) $
900 Stop {} -> return (env, expr)
901 CoerceIt co cont -> rebuild env (mkCoerce co expr) cont
902 Select _ bndr alts se cont -> rebuildCase (se `setFloats` env) expr bndr alts cont
903 StrictArg info _ cont -> rebuildCall env (info `addArgTo` expr) cont
904 StrictBind b bs body se cont -> do { env' <- simplNonRecX (se `setFloats` env) b expr
905 ; simplLam env' bs body cont }
906 ApplyTo _ arg se cont -> do { arg' <- simplExpr (se `setInScope` env) arg
907 ; rebuild env (App expr arg') cont }
911 %************************************************************************
915 %************************************************************************
918 simplCast :: SimplEnv -> InExpr -> Coercion -> SimplCont
919 -> SimplM (SimplEnv, OutExpr)
920 simplCast env body co0 cont0
921 = do { co1 <- simplCoercion env co0
922 ; simplExprF env body (addCoerce co1 cont0) }
924 addCoerce co cont = add_coerce co (coercionKind co) cont
926 add_coerce _co (s1, k1) cont -- co :: ty~ty
927 | s1 `coreEqType` k1 = cont -- is a no-op
929 add_coerce co1 (s1, _k2) (CoerceIt co2 cont)
930 | (_l1, t1) <- coercionKind co2
931 -- e |> (g1 :: S1~L) |> (g2 :: L~T1)
934 -- e |> (g1 . g2 :: S1~T1) otherwise
936 -- For example, in the initial form of a worker
937 -- we may find (coerce T (coerce S (\x.e))) y
938 -- and we'd like it to simplify to e[y/x] in one round
940 , s1 `coreEqType` t1 = cont -- The coerces cancel out
941 | otherwise = CoerceIt (mkTransCoercion co1 co2) cont
943 add_coerce co (s1s2, _t1t2) (ApplyTo dup (Type arg_ty) arg_se cont)
944 -- (f |> g) ty ---> (f ty) |> (g @ ty)
945 -- This implements the PushT and PushC rules from the paper
946 | Just (tyvar,_) <- splitForAllTy_maybe s1s2
948 (new_arg_ty, new_cast)
949 | isCoVar tyvar = (new_arg_co, mkCselRCoercion co) -- PushC rule
950 | otherwise = (ty', mkInstCoercion co ty') -- PushT rule
952 ApplyTo dup (Type new_arg_ty) (zapSubstEnv arg_se) (addCoerce new_cast cont)
954 ty' = substTy (arg_se `setInScope` env) arg_ty
955 new_arg_co = mkCsel1Coercion co `mkTransCoercion`
956 ty' `mkTransCoercion`
957 mkSymCoercion (mkCsel2Coercion co)
959 add_coerce co (s1s2, _t1t2) (ApplyTo dup arg arg_se cont)
960 | not (isTypeArg arg) -- This implements the Push rule from the paper
961 , isFunTy s1s2 -- t1t2 must be a function type, becuase it's applied
962 -- (e |> (g :: s1s2 ~ t1->t2)) f
964 -- (e (f |> (arg g :: t1~s1))
965 -- |> (res g :: s2->t2)
967 -- t1t2 must be a function type, t1->t2, because it's applied
968 -- to something but s1s2 might conceivably not be
970 -- When we build the ApplyTo we can't mix the out-types
971 -- with the InExpr in the argument, so we simply substitute
972 -- to make it all consistent. It's a bit messy.
973 -- But it isn't a common case.
975 -- Example of use: Trac #995
976 = ApplyTo dup new_arg (zapSubstEnv arg_se) (addCoerce co2 cont)
978 -- we split coercion t1->t2 ~ s1->s2 into t1 ~ s1 and
979 -- t2 ~ s2 with left and right on the curried form:
980 -- (->) t1 t2 ~ (->) s1 s2
981 [co1, co2] = decomposeCo 2 co
982 new_arg = mkCoerce (mkSymCoercion co1) arg'
983 arg' = substExpr (arg_se `setInScope` env) arg
985 add_coerce co _ cont = CoerceIt co cont
989 %************************************************************************
993 %************************************************************************
996 simplLam :: SimplEnv -> [InId] -> InExpr -> SimplCont
997 -> SimplM (SimplEnv, OutExpr)
999 simplLam env [] body cont = simplExprF env body cont
1002 simplLam env (bndr:bndrs) body (ApplyTo _ arg arg_se cont)
1003 = do { tick (BetaReduction bndr)
1004 ; simplNonRecE env bndr (arg, arg_se) (bndrs, body) cont }
1006 -- Not enough args, so there are real lambdas left to put in the result
1007 simplLam env bndrs body cont
1008 = do { (env', bndrs') <- simplLamBndrs env bndrs
1009 ; body' <- simplExpr env' body
1010 ; new_lam <- mkLam env' bndrs' body'
1011 ; rebuild env' new_lam cont }
1014 simplNonRecE :: SimplEnv
1015 -> InBndr -- The binder
1016 -> (InExpr, SimplEnv) -- Rhs of binding (or arg of lambda)
1017 -> ([InBndr], InExpr) -- Body of the let/lambda
1020 -> SimplM (SimplEnv, OutExpr)
1022 -- simplNonRecE is used for
1023 -- * non-top-level non-recursive lets in expressions
1026 -- It deals with strict bindings, via the StrictBind continuation,
1027 -- which may abort the whole process
1029 -- The "body" of the binding comes as a pair of ([InId],InExpr)
1030 -- representing a lambda; so we recurse back to simplLam
1031 -- Why? Because of the binder-occ-info-zapping done before
1032 -- the call to simplLam in simplExprF (Lam ...)
1034 -- First deal with type applications and type lets
1035 -- (/\a. e) (Type ty) and (let a = Type ty in e)
1036 simplNonRecE env bndr (Type ty_arg, rhs_se) (bndrs, body) cont
1037 = ASSERT( isTyVar bndr )
1038 do { ty_arg' <- simplType (rhs_se `setInScope` env) ty_arg
1039 ; simplLam (extendTvSubst env bndr ty_arg') bndrs body cont }
1041 simplNonRecE env bndr (rhs, rhs_se) (bndrs, body) cont
1042 | preInlineUnconditionally env NotTopLevel bndr rhs
1043 = do { tick (PreInlineUnconditionally bndr)
1044 ; simplLam (extendIdSubst env bndr (mkContEx rhs_se rhs)) bndrs body cont }
1047 = do { simplExprF (rhs_se `setFloats` env) rhs
1048 (StrictBind bndr bndrs body env cont) }
1051 = ASSERT( not (isTyVar bndr) )
1052 do { (env1, bndr1) <- simplNonRecBndr env bndr
1053 ; let (env2, bndr2) = addBndrRules env1 bndr bndr1
1054 ; env3 <- simplLazyBind env2 NotTopLevel NonRecursive bndr bndr2 rhs rhs_se
1055 ; simplLam env3 bndrs body cont }
1059 %************************************************************************
1063 %************************************************************************
1066 -- Hack alert: we only distinguish subsumed cost centre stacks for the
1067 -- purposes of inlining. All other CCCSs are mapped to currentCCS.
1068 simplNote :: SimplEnv -> Note -> CoreExpr -> SimplCont
1069 -> SimplM (SimplEnv, OutExpr)
1070 simplNote env (SCC cc) e cont
1071 | pushCCisNop cc (getEnclosingCC env) -- scc "f" (...(scc "f" e)...)
1072 = simplExprF env e cont -- ==> scc "f" (...e...)
1074 = do { e' <- simplExpr (setEnclosingCC env currentCCS) e
1075 ; rebuild env (mkSCC cc e') cont }
1077 simplNote env (CoreNote s) e cont
1078 = do { e' <- simplExpr env e
1079 ; rebuild env (Note (CoreNote s) e') cont }
1083 %************************************************************************
1085 \subsection{Dealing with calls}
1087 %************************************************************************
1090 simplVar :: SimplEnv -> Id -> SimplCont -> SimplM (SimplEnv, OutExpr)
1091 simplVar env var cont
1092 = case substId env var of
1093 DoneEx e -> simplExprF (zapSubstEnv env) e cont
1094 ContEx tvs ids e -> simplExprF (setSubstEnv env tvs ids) e cont
1095 DoneId var1 -> completeCall env var1 cont
1096 -- Note [zapSubstEnv]
1097 -- The template is already simplified, so don't re-substitute.
1098 -- This is VITAL. Consider
1100 -- let y = \z -> ...x... in
1102 -- We'll clone the inner \x, adding x->x' in the id_subst
1103 -- Then when we inline y, we must *not* replace x by x' in
1104 -- the inlined copy!!
1106 ---------------------------------------------------------
1107 -- Dealing with a call site
1109 completeCall :: SimplEnv -> Id -> SimplCont -> SimplM (SimplEnv, OutExpr)
1110 completeCall env var cont
1111 = do { ------------- Try inlining ----------------
1112 dflags <- getDOptsSmpl
1113 ; let (args,call_cont) = contArgs cont
1114 -- The args are OutExprs, obtained by *lazily* substituting
1115 -- in the args found in cont. These args are only examined
1116 -- to limited depth (unless a rule fires). But we must do
1117 -- the substitution; rule matching on un-simplified args would
1120 arg_infos = [interestingArg arg | arg <- args, isValArg arg]
1121 n_val_args = length arg_infos
1122 interesting_cont = interestingCallContext call_cont
1123 active_inline = activeInline env var
1124 maybe_inline = callSiteInline dflags active_inline var
1125 (null args) arg_infos interesting_cont
1126 ; case maybe_inline of {
1127 Just unfolding -- There is an inlining!
1128 -> do { tick (UnfoldingDone var)
1129 ; (if dopt Opt_D_dump_inlinings dflags then
1130 pprTrace ("Inlining done: " ++ showSDoc (ppr var)) (vcat [
1131 text "Before:" <+> ppr var <+> sep (map pprParendExpr args),
1132 text "Inlined fn: " <+> nest 2 (ppr unfolding),
1133 text "Cont: " <+> ppr call_cont])
1136 simplExprF (zapSubstEnv env) unfolding cont }
1138 ; Nothing -> do -- No inlining!
1140 { rule_base <- getSimplRules
1141 ; let info = mkArgInfo var (getRules rule_base var) n_val_args call_cont
1142 ; rebuildCall env info cont
1145 rebuildCall :: SimplEnv
1148 -> SimplM (SimplEnv, OutExpr)
1149 rebuildCall env (ArgInfo { ai_fun = fun, ai_args = rev_args, ai_strs = [] }) cont
1150 -- When we run out of strictness args, it means
1151 -- that the call is definitely bottom; see SimplUtils.mkArgInfo
1152 -- Then we want to discard the entire strict continuation. E.g.
1153 -- * case (error "hello") of { ... }
1154 -- * (error "Hello") arg
1155 -- * f (error "Hello") where f is strict
1157 -- Then, especially in the first of these cases, we'd like to discard
1158 -- the continuation, leaving just the bottoming expression. But the
1159 -- type might not be right, so we may have to add a coerce.
1160 | not (contIsTrivial cont) -- Only do this if there is a non-trivial
1161 = return (env, mk_coerce res) -- contination to discard, else we do it
1162 where -- again and again!
1163 res = mkApps (Var fun) (reverse rev_args)
1164 res_ty = exprType res
1165 cont_ty = contResultType env res_ty cont
1166 co = mkUnsafeCoercion res_ty cont_ty
1167 mk_coerce expr | cont_ty `coreEqType` res_ty = expr
1168 | otherwise = mkCoerce co expr
1170 rebuildCall env info (ApplyTo _ (Type arg_ty) se cont)
1171 = do { ty' <- simplCoercion (se `setInScope` env) arg_ty
1172 ; rebuildCall env (info `addArgTo` Type ty') cont }
1174 rebuildCall env info@(ArgInfo { ai_encl = encl_rules
1175 , ai_strs = str:strs, ai_discs = disc:discs })
1176 (ApplyTo _ arg arg_se cont)
1177 | str -- Strict argument
1178 = -- pprTrace "Strict Arg" (ppr arg $$ ppr (seIdSubst env) $$ ppr (seInScope env)) $
1179 simplExprF (arg_se `setFloats` env) arg
1180 (StrictArg info' cci cont)
1183 | otherwise -- Lazy argument
1184 -- DO NOT float anything outside, hence simplExprC
1185 -- There is no benefit (unlike in a let-binding), and we'd
1186 -- have to be very careful about bogus strictness through
1187 -- floating a demanded let.
1188 = do { arg' <- simplExprC (arg_se `setInScope` env) arg
1190 ; rebuildCall env (addArgTo info' arg') cont }
1192 info' = info { ai_strs = strs, ai_discs = discs }
1193 cci | encl_rules || disc > 0 = ArgCtxt encl_rules -- Be keener here
1194 | otherwise = BoringCtxt -- Nothing interesting
1196 rebuildCall env (ArgInfo { ai_fun = fun, ai_args = rev_args, ai_rules = rules }) cont
1197 = do { -- We've accumulated a simplified call in <fun,rev_args>
1198 -- so try rewrite rules; see Note [RULEs apply to simplified arguments]
1199 -- See also Note [Rules for recursive functions]
1200 ; let args = reverse rev_args
1201 env' = zapSubstEnv env
1202 ; mb_rule <- tryRules env rules fun args cont
1204 Just (n_args, rule_rhs) -> simplExprF env' rule_rhs $
1205 pushArgs env' (drop n_args args) cont ;
1206 -- n_args says how many args the rule consumed
1207 ; Nothing -> rebuild env (mkApps (Var fun) args) cont -- No rules
1211 Note [RULES apply to simplified arguments]
1212 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
1213 It's very desirable to try RULES once the arguments have been simplified, because
1214 doing so ensures that rule cascades work in one pass. Consider
1215 {-# RULES g (h x) = k x
1218 Then we want to rewrite (g (h x)) to (k x) and only then try f's rules. If
1219 we match f's rules against the un-simplified RHS, it won't match. This
1220 makes a particularly big difference when superclass selectors are involved:
1221 op ($p1 ($p2 (df d)))
1222 We want all this to unravel in one sweeep.
1226 This part of the simplifier may break the no-shadowing invariant
1228 f (...(\a -> e)...) (case y of (a,b) -> e')
1229 where f is strict in its second arg
1230 If we simplify the innermost one first we get (...(\a -> e)...)
1231 Simplifying the second arg makes us float the case out, so we end up with
1232 case y of (a,b) -> f (...(\a -> e)...) e'
1233 So the output does not have the no-shadowing invariant. However, there is
1234 no danger of getting name-capture, because when the first arg was simplified
1235 we used an in-scope set that at least mentioned all the variables free in its
1236 static environment, and that is enough.
1238 We can't just do innermost first, or we'd end up with a dual problem:
1239 case x of (a,b) -> f e (...(\a -> e')...)
1241 I spent hours trying to recover the no-shadowing invariant, but I just could
1242 not think of an elegant way to do it. The simplifier is already knee-deep in
1243 continuations. We have to keep the right in-scope set around; AND we have
1244 to get the effect that finding (error "foo") in a strict arg position will
1245 discard the entire application and replace it with (error "foo"). Getting
1246 all this at once is TOO HARD!
1249 %************************************************************************
1253 %************************************************************************
1256 tryRules :: SimplEnv -> [CoreRule]
1257 -> Id -> [OutExpr] -> SimplCont
1258 -> SimplM (Maybe (Arity, CoreExpr)) -- The arity is the number of
1259 -- args consumed by the rule
1260 tryRules env rules fn args call_cont
1264 = do { dflags <- getDOptsSmpl
1265 ; case activeRule dflags env of {
1266 Nothing -> return Nothing ; -- No rules apply
1268 case lookupRule act_fn (getInScope env) fn args rules of {
1269 Nothing -> return Nothing ; -- No rule matches
1270 Just (rule, rule_rhs) ->
1272 do { tick (RuleFired (ru_name rule))
1273 ; (if dopt Opt_D_dump_rule_firings dflags then
1274 pprTrace "Rule fired" (vcat [
1275 text "Rule:" <+> ftext (ru_name rule),
1276 text "Before:" <+> ppr fn <+> sep (map pprParendExpr args),
1277 text "After: " <+> pprCoreExpr rule_rhs,
1278 text "Cont: " <+> ppr call_cont])
1281 return (Just (ruleArity rule, rule_rhs)) }}}}
1284 Note [Rules for recursive functions]
1285 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
1286 You might think that we shouldn't apply rules for a loop breaker:
1287 doing so might give rise to an infinite loop, because a RULE is
1288 rather like an extra equation for the function:
1289 RULE: f (g x) y = x+y
1292 But it's too drastic to disable rules for loop breakers.
1293 Even the foldr/build rule would be disabled, because foldr
1294 is recursive, and hence a loop breaker:
1295 foldr k z (build g) = g k z
1296 So it's up to the programmer: rules can cause divergence
1299 %************************************************************************
1301 Rebuilding a cse expression
1303 %************************************************************************
1305 Note [Case elimination]
1306 ~~~~~~~~~~~~~~~~~~~~~~~
1307 The case-elimination transformation discards redundant case expressions.
1308 Start with a simple situation:
1310 case x# of ===> e[x#/y#]
1313 (when x#, y# are of primitive type, of course). We can't (in general)
1314 do this for algebraic cases, because we might turn bottom into
1317 The code in SimplUtils.prepareAlts has the effect of generalise this
1318 idea to look for a case where we're scrutinising a variable, and we
1319 know that only the default case can match. For example:
1323 DEFAULT -> ...(case x of
1327 Here the inner case is first trimmed to have only one alternative, the
1328 DEFAULT, after which it's an instance of the previous case. This
1329 really only shows up in eliminating error-checking code.
1331 We also make sure that we deal with this very common case:
1336 Here we are using the case as a strict let; if x is used only once
1337 then we want to inline it. We have to be careful that this doesn't
1338 make the program terminate when it would have diverged before, so we
1340 - e is already evaluated (it may so if e is a variable)
1341 - x is used strictly, or
1343 Lastly, the code in SimplUtils.mkCase combines identical RHSs. So
1345 case e of ===> case e of DEFAULT -> r
1349 Now again the case may be elminated by the CaseElim transformation.
1352 Further notes about case elimination
1353 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
1354 Consider: test :: Integer -> IO ()
1357 Turns out that this compiles to:
1360 eta1 :: State# RealWorld ->
1361 case PrelNum.< eta PrelNum.zeroInteger of wild { __DEFAULT ->
1363 (PrelNum.jtos eta ($w[] @ Char))
1365 of wild1 { (# new_s, a4 #) -> PrelIO.lvl23 new_s }}
1367 Notice the strange '<' which has no effect at all. This is a funny one.
1368 It started like this:
1370 f x y = if x < 0 then jtos x
1371 else if y==0 then "" else jtos x
1373 At a particular call site we have (f v 1). So we inline to get
1375 if v < 0 then jtos x
1376 else if 1==0 then "" else jtos x
1378 Now simplify the 1==0 conditional:
1380 if v<0 then jtos v else jtos v
1382 Now common-up the two branches of the case:
1384 case (v<0) of DEFAULT -> jtos v
1386 Why don't we drop the case? Because it's strict in v. It's technically
1387 wrong to drop even unnecessary evaluations, and in practice they
1388 may be a result of 'seq' so we *definitely* don't want to drop those.
1389 I don't really know how to improve this situation.
1392 ---------------------------------------------------------
1393 -- Eliminate the case if possible
1395 rebuildCase, reallyRebuildCase
1397 -> OutExpr -- Scrutinee
1398 -> InId -- Case binder
1399 -> [InAlt] -- Alternatives (inceasing order)
1401 -> SimplM (SimplEnv, OutExpr)
1403 --------------------------------------------------
1404 -- 1. Eliminate the case if there's a known constructor
1405 --------------------------------------------------
1407 rebuildCase env scrut case_bndr alts cont
1408 | Lit lit <- scrut -- No need for same treatment as constructors
1409 -- because literals are inlined more vigorously
1410 = do { tick (KnownBranch case_bndr)
1411 ; case findAlt (LitAlt lit) alts of
1412 Nothing -> missingAlt env case_bndr alts cont
1413 Just (_, bs, rhs) -> simple_rhs bs rhs }
1415 | Just (con, ty_args, other_args) <- exprIsConApp_maybe scrut
1416 -- Works when the scrutinee is a variable with a known unfolding
1417 -- as well as when it's an explicit constructor application
1418 = do { tick (KnownBranch case_bndr)
1419 ; case findAlt (DataAlt con) alts of
1420 Nothing -> missingAlt env case_bndr alts cont
1421 Just (DEFAULT, bs, rhs) -> simple_rhs bs rhs
1422 Just (_, bs, rhs) -> knownCon env scrut con ty_args other_args
1423 case_bndr bs rhs cont
1426 simple_rhs bs rhs = ASSERT( null bs )
1427 do { env' <- simplNonRecX env case_bndr scrut
1428 ; simplExprF env' rhs cont }
1431 --------------------------------------------------
1432 -- 2. Eliminate the case if scrutinee is evaluated
1433 --------------------------------------------------
1435 rebuildCase env scrut case_bndr [(_, bndrs, rhs)] cont
1436 -- See if we can get rid of the case altogether
1437 -- See Note [Case eliminiation]
1438 -- mkCase made sure that if all the alternatives are equal,
1439 -- then there is now only one (DEFAULT) rhs
1440 | all isDeadBinder bndrs -- bndrs are [InId]
1442 -- Check that the scrutinee can be let-bound instead of case-bound
1443 , exprOkForSpeculation scrut
1444 -- OK not to evaluate it
1445 -- This includes things like (==# a# b#)::Bool
1446 -- so that we simplify
1447 -- case ==# a# b# of { True -> x; False -> x }
1450 -- This particular example shows up in default methods for
1451 -- comparision operations (e.g. in (>=) for Int.Int32)
1452 || exprIsHNF scrut -- It's already evaluated
1453 || var_demanded_later scrut -- It'll be demanded later
1455 -- || not opt_SimplPedanticBottoms) -- Or we don't care!
1456 -- We used to allow improving termination by discarding cases, unless -fpedantic-bottoms was on,
1457 -- but that breaks badly for the dataToTag# primop, which relies on a case to evaluate
1458 -- its argument: case x of { y -> dataToTag# y }
1459 -- Here we must *not* discard the case, because dataToTag# just fetches the tag from
1460 -- the info pointer. So we'll be pedantic all the time, and see if that gives any
1462 -- Also we don't want to discard 'seq's
1463 = do { tick (CaseElim case_bndr)
1464 ; env' <- simplNonRecX env case_bndr scrut
1465 ; simplExprF env' rhs cont }
1467 -- The case binder is going to be evaluated later,
1468 -- and the scrutinee is a simple variable
1469 var_demanded_later (Var v) = isStrictDmd (idNewDemandInfo case_bndr)
1470 && not (isTickBoxOp v)
1471 -- ugly hack; covering this case is what
1472 -- exprOkForSpeculation was intended for.
1473 var_demanded_later _ = False
1475 --------------------------------------------------
1476 -- 3. Try seq rules; see Note [User-defined RULES for seq] in MkId
1477 --------------------------------------------------
1479 rebuildCase env scrut case_bndr alts@[(_, bndrs, rhs)] cont
1480 | all isDeadBinder (case_bndr : bndrs) -- So this is just 'seq'
1481 = do { let rhs' = substExpr env rhs
1482 out_args = [Type (substTy env (idType case_bndr)),
1483 Type (exprType rhs'), scrut, rhs']
1484 -- Lazily evaluated, so we don't do most of this
1486 ; rule_base <- getSimplRules
1487 ; mb_rule <- tryRules env (getRules rule_base seqId) seqId out_args cont
1489 Just (n_args, res) -> simplExprF (zapSubstEnv env)
1490 (mkApps res (drop n_args out_args))
1492 Nothing -> reallyRebuildCase env scrut case_bndr alts cont }
1494 rebuildCase env scrut case_bndr alts cont
1495 = reallyRebuildCase env scrut case_bndr alts cont
1497 --------------------------------------------------
1498 -- 3. Catch-all case
1499 --------------------------------------------------
1501 reallyRebuildCase env scrut case_bndr alts cont
1502 = do { -- Prepare the continuation;
1503 -- The new subst_env is in place
1504 (env', dup_cont, nodup_cont) <- prepareCaseCont env alts cont
1506 -- Simplify the alternatives
1507 ; (scrut', case_bndr', alts') <- simplAlts env' scrut case_bndr alts dup_cont
1509 -- Check for empty alternatives
1510 ; if null alts' then missingAlt env case_bndr alts cont
1512 { dflags <- getDOptsSmpl
1513 ; case_expr <- mkCase dflags scrut' case_bndr' alts'
1515 -- Notice that rebuild gets the in-scope set from env', not alt_env
1516 -- (which in any case is only build in simplAlts)
1517 -- The case binder *not* scope over the whole returned case-expression
1518 ; rebuild env' case_expr nodup_cont } }
1521 simplCaseBinder checks whether the scrutinee is a variable, v. If so,
1522 try to eliminate uses of v in the RHSs in favour of case_bndr; that
1523 way, there's a chance that v will now only be used once, and hence
1526 Historical note: we use to do the "case binder swap" in the Simplifier
1527 so there were additional complications if the scrutinee was a variable.
1528 Now the binder-swap stuff is done in the occurrence analyer; see
1529 OccurAnal Note [Binder swap].
1533 If the case binder is not dead, then neither are the pattern bound
1535 case <any> of x { (a,b) ->
1536 case x of { (p,q) -> p } }
1537 Here (a,b) both look dead, but come alive after the inner case is eliminated.
1538 The point is that we bring into the envt a binding
1540 after the outer case, and that makes (a,b) alive. At least we do unless
1541 the case binder is guaranteed dead.
1543 In practice, the scrutinee is almost always a variable, so we pretty
1544 much always zap the OccInfo of the binders. It doesn't matter much though.
1549 Consider case (v `cast` co) of x { I# y ->
1550 ... (case (v `cast` co) of {...}) ...
1551 We'd like to eliminate the inner case. We can get this neatly by
1552 arranging that inside the outer case we add the unfolding
1553 v |-> x `cast` (sym co)
1554 to v. Then we should inline v at the inner case, cancel the casts, and away we go
1556 Note [Improving seq]
1559 type family F :: * -> *
1560 type instance F Int = Int
1562 ... case e of x { DEFAULT -> rhs } ...
1564 where x::F Int. Then we'd like to rewrite (F Int) to Int, getting
1566 case e `cast` co of x'::Int
1567 I# x# -> let x = x' `cast` sym co
1570 so that 'rhs' can take advantage of the form of x'.
1572 Notice that Note [Case of cast] may then apply to the result.
1574 Nota Bene: We only do the [Improving seq] transformation if the
1575 case binder 'x' is actually used in the rhs; that is, if the case
1576 is *not* a *pure* seq.
1577 a) There is no point in adding the cast to a pure seq.
1578 b) There is a good reason not to: doing so would interfere
1579 with seq rules (Note [Built-in RULES for seq] in MkId).
1580 In particular, this [Improving seq] thing *adds* a cast
1581 while [Built-in RULES for seq] *removes* one, so they
1584 You might worry about
1585 case v of x { __DEFAULT ->
1586 ... case (v `cast` co) of y { I# -> ... }}
1587 This is a pure seq (since x is unused), so [Improving seq] won't happen.
1588 But it's ok: the simplifier will replace 'v' by 'x' in the rhs to get
1589 case v of x { __DEFAULT ->
1590 ... case (x `cast` co) of y { I# -> ... }}
1591 Now the outer case is not a pure seq, so [Improving seq] will happen,
1592 and then the inner case will disappear.
1594 The need for [Improving seq] showed up in Roman's experiments. Example:
1595 foo :: F Int -> Int -> Int
1596 foo t n = t `seq` bar n
1599 bar n = bar (n - case t of TI i -> i)
1600 Here we'd like to avoid repeated evaluating t inside the loop, by
1601 taking advantage of the `seq`.
1603 At one point I did transformation in LiberateCase, but it's more
1604 robust here. (Otherwise, there's a danger that we'll simply drop the
1605 'seq' altogether, before LiberateCase gets to see it.)
1608 simplAlts :: SimplEnv
1610 -> InId -- Case binder
1611 -> [InAlt] -- Non-empty
1613 -> SimplM (OutExpr, OutId, [OutAlt]) -- Includes the continuation
1614 -- Like simplExpr, this just returns the simplified alternatives;
1615 -- it does not return an environment
1617 simplAlts env scrut case_bndr alts cont'
1618 = -- pprTrace "simplAlts" (ppr alts $$ ppr (seIdSubst env)) $
1619 do { let env0 = zapFloats env
1621 ; (env1, case_bndr1) <- simplBinder env0 case_bndr
1623 ; fam_envs <- getFamEnvs
1624 ; (alt_env', scrut', case_bndr') <- improveSeq fam_envs env1 scrut
1625 case_bndr case_bndr1 alts
1627 ; (imposs_deflt_cons, in_alts) <- prepareAlts scrut' case_bndr' alts
1629 ; alts' <- mapM (simplAlt alt_env' imposs_deflt_cons case_bndr' cont') in_alts
1630 ; return (scrut', case_bndr', alts') }
1633 ------------------------------------
1634 improveSeq :: (FamInstEnv, FamInstEnv) -> SimplEnv
1635 -> OutExpr -> InId -> OutId -> [InAlt]
1636 -> SimplM (SimplEnv, OutExpr, OutId)
1637 -- Note [Improving seq]
1638 improveSeq fam_envs env scrut case_bndr case_bndr1 [(DEFAULT,_,_)]
1639 | not (isDeadBinder case_bndr) -- Not a pure seq! See the Note!
1640 , Just (co, ty2) <- topNormaliseType fam_envs (idType case_bndr1)
1641 = do { case_bndr2 <- newId (fsLit "nt") ty2
1642 ; let rhs = DoneEx (Var case_bndr2 `Cast` mkSymCoercion co)
1643 env2 = extendIdSubst env case_bndr rhs
1644 ; return (env2, scrut `Cast` co, case_bndr2) }
1646 improveSeq _ env scrut _ case_bndr1 _
1647 = return (env, scrut, case_bndr1)
1650 ------------------------------------
1651 simplAlt :: SimplEnv
1652 -> [AltCon] -- These constructors can't be present when
1653 -- matching the DEFAULT alternative
1654 -> OutId -- The case binder
1659 simplAlt env imposs_deflt_cons case_bndr' cont' (DEFAULT, bndrs, rhs)
1660 = ASSERT( null bndrs )
1661 do { let env' = addBinderOtherCon env case_bndr' imposs_deflt_cons
1662 -- Record the constructors that the case-binder *can't* be.
1663 ; rhs' <- simplExprC env' rhs cont'
1664 ; return (DEFAULT, [], rhs') }
1666 simplAlt env _ case_bndr' cont' (LitAlt lit, bndrs, rhs)
1667 = ASSERT( null bndrs )
1668 do { let env' = addBinderUnfolding env case_bndr' (Lit lit)
1669 ; rhs' <- simplExprC env' rhs cont'
1670 ; return (LitAlt lit, [], rhs') }
1672 simplAlt env _ case_bndr' cont' (DataAlt con, vs, rhs)
1673 = do { -- Deal with the pattern-bound variables
1674 -- Mark the ones that are in ! positions in the
1675 -- data constructor as certainly-evaluated.
1676 -- NB: simplLamBinders preserves this eval info
1677 let vs_with_evals = add_evals (dataConRepStrictness con)
1678 ; (env', vs') <- simplLamBndrs env vs_with_evals
1680 -- Bind the case-binder to (con args)
1681 ; let inst_tys' = tyConAppArgs (idType case_bndr')
1682 con_args = map Type inst_tys' ++ varsToCoreExprs vs'
1683 env'' = addBinderUnfolding env' case_bndr'
1684 (mkConApp con con_args)
1686 ; rhs' <- simplExprC env'' rhs cont'
1687 ; return (DataAlt con, vs', rhs') }
1689 -- add_evals records the evaluated-ness of the bound variables of
1690 -- a case pattern. This is *important*. Consider
1691 -- data T = T !Int !Int
1693 -- case x of { T a b -> T (a+1) b }
1695 -- We really must record that b is already evaluated so that we don't
1696 -- go and re-evaluate it when constructing the result.
1697 -- See Note [Data-con worker strictness] in MkId.lhs
1702 go (v:vs') strs | isTyVar v = v : go vs' strs
1703 go (v:vs') (str:strs)
1704 | isMarkedStrict str = evald_v : go vs' strs
1705 | otherwise = zapped_v : go vs' strs
1707 zapped_v = zap_occ_info v
1708 evald_v = zapped_v `setIdUnfolding` evaldUnfolding
1709 go _ _ = pprPanic "cat_evals" (ppr con $$ ppr vs $$ ppr the_strs)
1711 -- See Note [zapOccInfo]
1712 -- zap_occ_info: if the case binder is alive, then we add the unfolding
1714 -- to the envt; so vs are now very much alive
1715 -- Note [Aug06] I can't see why this actually matters, but it's neater
1716 -- case e of t { (a,b) -> ...(case t of (p,q) -> p)... }
1717 -- ==> case e of t { (a,b) -> ...(a)... }
1718 -- Look, Ma, a is alive now.
1719 zap_occ_info = zapCasePatIdOcc case_bndr'
1721 addBinderUnfolding :: SimplEnv -> Id -> CoreExpr -> SimplEnv
1722 addBinderUnfolding env bndr rhs
1723 = modifyInScope env (bndr `setIdUnfolding` mkUnfolding False rhs)
1725 addBinderOtherCon :: SimplEnv -> Id -> [AltCon] -> SimplEnv
1726 addBinderOtherCon env bndr cons
1727 = modifyInScope env (bndr `setIdUnfolding` mkOtherCon cons)
1729 zapCasePatIdOcc :: Id -> Id -> Id
1730 -- Consider case e of b { (a,b) -> ... }
1731 -- Then if we bind b to (a,b) in "...", and b is not dead,
1732 -- then we must zap the deadness info on a,b
1733 zapCasePatIdOcc case_bndr
1734 | isDeadBinder case_bndr = \ pat_id -> pat_id
1735 | otherwise = \ pat_id -> zapIdOccInfo pat_id
1739 %************************************************************************
1741 \subsection{Known constructor}
1743 %************************************************************************
1745 We are a bit careful with occurrence info. Here's an example
1747 (\x* -> case x of (a*, b) -> f a) (h v, e)
1749 where the * means "occurs once". This effectively becomes
1750 case (h v, e) of (a*, b) -> f a)
1752 let a* = h v; b = e in f a
1756 All this should happen in one sweep.
1759 knownCon :: SimplEnv
1760 -> OutExpr -- The scrutinee
1761 -> DataCon -> [OutType] -> [OutExpr] -- The scrutinee (in pieces)
1762 -> InId -> [InBndr] -> InExpr -- The alternative
1764 -> SimplM (SimplEnv, OutExpr)
1766 knownCon env scrut dc dc_ty_args dc_args bndr bs rhs cont
1767 = do { env' <- bind_args env bs dc_args
1769 -- It's useful to bind bndr to scrut, rather than to a fresh
1770 -- binding x = Con arg1 .. argn
1771 -- because very often the scrut is a variable, so we avoid
1772 -- creating, and then subsequently eliminating, a let-binding
1773 -- BUT, if scrut is a not a variable, we must be careful
1774 -- about duplicating the arg redexes; in that case, make
1775 -- a new con-app from the args
1776 bndr_rhs | exprIsTrivial scrut = scrut
1777 | otherwise = con_app
1778 con_app = Var (dataConWorkId dc)
1779 `mkTyApps` dc_ty_args
1780 `mkApps` [substExpr env' (varToCoreExpr b) | b <- bs]
1781 -- dc_ty_args are aready OutTypes, but bs are InBndrs
1783 ; env'' <- simplNonRecX env' bndr bndr_rhs
1784 ; simplExprF env'' rhs cont }
1786 zap_occ = zapCasePatIdOcc bndr -- bndr is an InId
1789 bind_args env' [] _ = return env'
1791 bind_args env' (b:bs') (Type ty : args)
1792 = ASSERT( isTyVar b )
1793 bind_args (extendTvSubst env' b ty) bs' args
1795 bind_args env' (b:bs') (arg : args)
1797 do { let b' = zap_occ b
1798 -- Note that the binder might be "dead", because it doesn't
1799 -- occur in the RHS; and simplNonRecX may therefore discard
1800 -- it via postInlineUnconditionally.
1801 -- Nevertheless we must keep it if the case-binder is alive,
1802 -- because it may be used in the con_app. See Note [zapOccInfo]
1803 ; env'' <- simplNonRecX env' b' arg
1804 ; bind_args env'' bs' args }
1807 pprPanic "bind_args" $ ppr dc $$ ppr bs $$ ppr dc_args $$
1808 text "scrut:" <+> ppr scrut
1811 missingAlt :: SimplEnv -> Id -> [InAlt] -> SimplCont -> SimplM (SimplEnv, OutExpr)
1812 -- This isn't strictly an error, although it is unusual.
1813 -- It's possible that the simplifer might "see" that
1814 -- an inner case has no accessible alternatives before
1815 -- it "sees" that the entire branch of an outer case is
1816 -- inaccessible. So we simply put an error case here instead.
1817 missingAlt env case_bndr alts cont
1818 = WARN( True, ptext (sLit "missingAlt") <+> ppr case_bndr )
1819 return (env, mkImpossibleExpr res_ty)
1821 res_ty = contResultType env (substTy env (coreAltsType alts)) cont
1825 %************************************************************************
1827 \subsection{Duplicating continuations}
1829 %************************************************************************
1832 prepareCaseCont :: SimplEnv
1833 -> [InAlt] -> SimplCont
1834 -> SimplM (SimplEnv, SimplCont,SimplCont)
1835 -- Return a duplicatable continuation, a non-duplicable part
1836 -- plus some extra bindings (that scope over the entire
1839 -- No need to make it duplicatable if there's only one alternative
1840 prepareCaseCont env [_] cont = return (env, cont, mkBoringStop)
1841 prepareCaseCont env _ cont = mkDupableCont env cont
1845 mkDupableCont :: SimplEnv -> SimplCont
1846 -> SimplM (SimplEnv, SimplCont, SimplCont)
1848 mkDupableCont env cont
1849 | contIsDupable cont
1850 = return (env, cont, mkBoringStop)
1852 mkDupableCont _ (Stop {}) = panic "mkDupableCont" -- Handled by previous eqn
1854 mkDupableCont env (CoerceIt ty cont)
1855 = do { (env', dup, nodup) <- mkDupableCont env cont
1856 ; return (env', CoerceIt ty dup, nodup) }
1858 mkDupableCont env cont@(StrictBind {})
1859 = return (env, mkBoringStop, cont)
1860 -- See Note [Duplicating StrictBind]
1862 mkDupableCont env (StrictArg info cci cont)
1863 -- See Note [Duplicating StrictArg]
1864 = do { (env', dup, nodup) <- mkDupableCont env cont
1865 ; (env'', args') <- mapAccumLM makeTrivial env' (ai_args info)
1866 ; return (env'', StrictArg (info { ai_args = args' }) cci dup, nodup) }
1868 mkDupableCont env (ApplyTo _ arg se cont)
1869 = -- e.g. [...hole...] (...arg...)
1871 -- let a = ...arg...
1872 -- in [...hole...] a
1873 do { (env', dup_cont, nodup_cont) <- mkDupableCont env cont
1874 ; arg' <- simplExpr (se `setInScope` env') arg
1875 ; (env'', arg'') <- makeTrivial env' arg'
1876 ; let app_cont = ApplyTo OkToDup arg'' (zapSubstEnv env'') dup_cont
1877 ; return (env'', app_cont, nodup_cont) }
1879 mkDupableCont env cont@(Select _ case_bndr [(_, bs, _rhs)] _ _)
1880 -- See Note [Single-alternative case]
1881 -- | not (exprIsDupable rhs && contIsDupable case_cont)
1882 -- | not (isDeadBinder case_bndr)
1883 | all isDeadBinder bs -- InIds
1884 && not (isUnLiftedType (idType case_bndr))
1885 -- Note [Single-alternative-unlifted]
1886 = return (env, mkBoringStop, cont)
1888 mkDupableCont env (Select _ case_bndr alts se cont)
1889 = -- e.g. (case [...hole...] of { pi -> ei })
1891 -- let ji = \xij -> ei
1892 -- in case [...hole...] of { pi -> ji xij }
1893 do { tick (CaseOfCase case_bndr)
1894 ; (env', dup_cont, nodup_cont) <- mkDupableCont env cont
1895 -- NB: call mkDupableCont here, *not* prepareCaseCont
1896 -- We must make a duplicable continuation, whereas prepareCaseCont
1897 -- doesn't when there is a single case branch
1899 ; let alt_env = se `setInScope` env'
1900 ; (alt_env', case_bndr') <- simplBinder alt_env case_bndr
1901 ; alts' <- mapM (simplAlt alt_env' [] case_bndr' dup_cont) alts
1902 -- Safe to say that there are no handled-cons for the DEFAULT case
1903 -- NB: simplBinder does not zap deadness occ-info, so
1904 -- a dead case_bndr' will still advertise its deadness
1905 -- This is really important because in
1906 -- case e of b { (# p,q #) -> ... }
1907 -- b is always dead, and indeed we are not allowed to bind b to (# p,q #),
1908 -- which might happen if e was an explicit unboxed pair and b wasn't marked dead.
1909 -- In the new alts we build, we have the new case binder, so it must retain
1911 -- NB: we don't use alt_env further; it has the substEnv for
1912 -- the alternatives, and we don't want that
1914 ; (env'', alts'') <- mkDupableAlts env' case_bndr' alts'
1915 ; return (env'', -- Note [Duplicated env]
1916 Select OkToDup case_bndr' alts'' (zapSubstEnv env'') mkBoringStop,
1920 mkDupableAlts :: SimplEnv -> OutId -> [InAlt]
1921 -> SimplM (SimplEnv, [InAlt])
1922 -- Absorbs the continuation into the new alternatives
1924 mkDupableAlts env case_bndr' the_alts
1927 go env0 [] = return (env0, [])
1929 = do { (env1, alt') <- mkDupableAlt env0 case_bndr' alt
1930 ; (env2, alts') <- go env1 alts
1931 ; return (env2, alt' : alts' ) }
1933 mkDupableAlt :: SimplEnv -> OutId -> (AltCon, [CoreBndr], CoreExpr)
1934 -> SimplM (SimplEnv, (AltCon, [CoreBndr], CoreExpr))
1935 mkDupableAlt env case_bndr (con, bndrs', rhs')
1936 | exprIsDupable rhs' -- Note [Small alternative rhs]
1937 = return (env, (con, bndrs', rhs'))
1939 = do { let rhs_ty' = exprType rhs'
1940 scrut_ty = idType case_bndr
1943 DEFAULT -> case_bndr
1944 DataAlt dc -> setIdUnfolding case_bndr unf
1946 -- See Note [Case binders and join points]
1947 unf = mkInlineRule InlSat rhs 0
1948 rhs = mkConApp dc (map Type (tyConAppArgs scrut_ty)
1949 ++ varsToCoreExprs bndrs')
1951 LitAlt {} -> WARN( True, ptext (sLit "mkDupableAlt")
1952 <+> ppr case_bndr <+> ppr con )
1954 -- The case binder is alive but trivial, so why has
1955 -- it not been substituted away?
1957 used_bndrs' | isDeadBinder case_bndr = filter abstract_over bndrs'
1958 | otherwise = bndrs' ++ [case_bndr_w_unf]
1961 | isTyVar bndr = True -- Abstract over all type variables just in case
1962 | otherwise = not (isDeadBinder bndr)
1963 -- The deadness info on the new Ids is preserved by simplBinders
1965 ; (final_bndrs', final_args) -- Note [Join point abstraction]
1966 <- if (any isId used_bndrs')
1967 then return (used_bndrs', varsToCoreExprs used_bndrs')
1968 else do { rw_id <- newId (fsLit "w") realWorldStatePrimTy
1969 ; return ([rw_id], [Var realWorldPrimId]) }
1971 ; join_bndr <- newId (fsLit "$j") (mkPiTypes final_bndrs' rhs_ty')
1972 -- Note [Funky mkPiTypes]
1974 ; let -- We make the lambdas into one-shot-lambdas. The
1975 -- join point is sure to be applied at most once, and doing so
1976 -- prevents the body of the join point being floated out by
1977 -- the full laziness pass
1978 really_final_bndrs = map one_shot final_bndrs'
1979 one_shot v | isId v = setOneShotLambda v
1981 join_rhs = mkLams really_final_bndrs rhs'
1982 join_call = mkApps (Var join_bndr) final_args
1984 ; env' <- addPolyBind NotTopLevel env (NonRec join_bndr join_rhs)
1985 ; return (env', (con, bndrs', join_call)) }
1986 -- See Note [Duplicated env]
1989 Note [Case binders and join points]
1990 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
1992 case (case .. ) of c {
1995 If we make a join point with c but not c# we get
1996 $j = \c -> ....c....
1998 But if later inlining scrutines the c, thus
2000 $j = \c -> ... case c of { I# y -> ... } ...
2002 we won't see that 'c' has already been scrutinised. This actually
2003 happens in the 'tabulate' function in wave4main, and makes a significant
2004 difference to allocation.
2006 An alternative plan is this:
2008 $j = \c# -> let c = I# c# in ...c....
2010 but that is bad if 'c' is *not* later scrutinised.
2012 So instead we do both: we pass 'c' and 'c#' , and record in c's inlining
2013 that it's really I# c#, thus
2015 $j = \c# -> \c[=I# c#] -> ...c....
2017 Absence analysis may later discard 'c'.
2020 Note [Duplicated env]
2021 ~~~~~~~~~~~~~~~~~~~~~
2022 Some of the alternatives are simplified, but have not been turned into a join point
2023 So they *must* have an zapped subst-env. So we can't use completeNonRecX to
2024 bind the join point, because it might to do PostInlineUnconditionally, and
2025 we'd lose that when zapping the subst-env. We could have a per-alt subst-env,
2026 but zapping it (as we do in mkDupableCont, the Select case) is safe, and
2027 at worst delays the join-point inlining.
2029 Note [Small alternative rhs]
2030 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~
2031 It is worth checking for a small RHS because otherwise we
2032 get extra let bindings that may cause an extra iteration of the simplifier to
2033 inline back in place. Quite often the rhs is just a variable or constructor.
2034 The Ord instance of Maybe in PrelMaybe.lhs, for example, took several extra
2035 iterations because the version with the let bindings looked big, and so wasn't
2036 inlined, but after the join points had been inlined it looked smaller, and so
2039 NB: we have to check the size of rhs', not rhs.
2040 Duplicating a small InAlt might invalidate occurrence information
2041 However, if it *is* dupable, we return the *un* simplified alternative,
2042 because otherwise we'd need to pair it up with an empty subst-env....
2043 but we only have one env shared between all the alts.
2044 (Remember we must zap the subst-env before re-simplifying something).
2045 Rather than do this we simply agree to re-simplify the original (small) thing later.
2047 Note [Funky mkPiTypes]
2048 ~~~~~~~~~~~~~~~~~~~~~~
2049 Notice the funky mkPiTypes. If the contructor has existentials
2050 it's possible that the join point will be abstracted over
2051 type varaibles as well as term variables.
2052 Example: Suppose we have
2053 data T = forall t. C [t]
2055 case (case e of ...) of
2057 We get the join point
2058 let j :: forall t. [t] -> ...
2059 j = /\t \xs::[t] -> rhs
2061 case (case e of ...) of
2062 C t xs::[t] -> j t xs
2064 Note [Join point abstaction]
2065 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~
2066 If we try to lift a primitive-typed something out
2067 for let-binding-purposes, we will *caseify* it (!),
2068 with potentially-disastrous strictness results. So
2069 instead we turn it into a function: \v -> e
2070 where v::State# RealWorld#. The value passed to this function
2071 is realworld#, which generates (almost) no code.
2073 There's a slight infelicity here: we pass the overall
2074 case_bndr to all the join points if it's used in *any* RHS,
2075 because we don't know its usage in each RHS separately
2077 We used to say "&& isUnLiftedType rhs_ty'" here, but now
2078 we make the join point into a function whenever used_bndrs'
2079 is empty. This makes the join-point more CPR friendly.
2080 Consider: let j = if .. then I# 3 else I# 4
2081 in case .. of { A -> j; B -> j; C -> ... }
2083 Now CPR doesn't w/w j because it's a thunk, so
2084 that means that the enclosing function can't w/w either,
2085 which is a lose. Here's the example that happened in practice:
2086 kgmod :: Int -> Int -> Int
2087 kgmod x y = if x > 0 && y < 0 || x < 0 && y > 0
2091 I have seen a case alternative like this:
2093 It's a bit silly to add the realWorld dummy arg in this case, making
2096 (the \v alone is enough to make CPR happy) but I think it's rare
2098 Note [Duplicating StrictArg]
2099 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~
2100 The original plan had (where E is a big argument)
2102 ==> let $j = \a -> f E a
2105 But this is terrible! Here's an example:
2106 && E (case x of { T -> F; F -> T })
2107 Now, && is strict so we end up simplifying the case with
2108 an ArgOf continuation. If we let-bind it, we get
2109 let $j = \v -> && E v
2110 in simplExpr (case x of { T -> F; F -> T })
2112 And after simplifying more we get
2113 let $j = \v -> && E v
2114 in case x of { T -> $j F; F -> $j T }
2115 Which is a Very Bad Thing
2117 What we do now is this
2121 Now if the thing in the hole is a case expression (which is when
2122 we'll call mkDupableCont), we'll push the function call into the
2123 branches, which is what we want. Now RULES for f may fire, and
2124 call-pattern specialisation. Here's an example from Trac #3116
2127 _ -> Chunk p fpc (o+1) (l-1) bs')
2128 If we can push the call for 'go' inside the case, we get
2129 call-pattern specialisation for 'go', which is *crucial* for
2132 Here is the (&&) example:
2133 && E (case x of { T -> F; F -> T })
2135 case x of { T -> && a F; F -> && a T }
2139 * Arguments to f *after* the strict one are handled by
2140 the ApplyTo case of mkDupableCont. Eg
2143 * We can only do the let-binding of E because the function
2144 part of a StrictArg continuation is an explicit syntax
2145 tree. In earlier versions we represented it as a function
2146 (CoreExpr -> CoreEpxr) which we couldn't take apart.
2148 Do *not* duplicate StrictBind and StritArg continuations. We gain
2149 nothing by propagating them into the expressions, and we do lose a
2152 The desire not to duplicate is the entire reason that
2153 mkDupableCont returns a pair of continuations.
2155 Note [Duplicating StrictBind]
2156 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
2157 Unlike StrictArg, there doesn't seem anything to gain from
2158 duplicating a StrictBind continuation, so we don't.
2160 The desire not to duplicate is the entire reason that
2161 mkDupableCont returns a pair of continuations.
2164 Note [Single-alternative cases]
2165 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
2166 This case is just like the ArgOf case. Here's an example:
2170 case (case x of I# x' ->
2172 True -> I# (negate# x')
2173 False -> I# x') of y {
2175 Because the (case x) has only one alternative, we'll transform to
2177 case (case x' <# 0# of
2178 True -> I# (negate# x')
2179 False -> I# x') of y {
2181 But now we do *NOT* want to make a join point etc, giving
2183 let $j = \y -> MkT y
2185 True -> $j (I# (negate# x'))
2187 In this case the $j will inline again, but suppose there was a big
2188 strict computation enclosing the orginal call to MkT. Then, it won't
2189 "see" the MkT any more, because it's big and won't get duplicated.
2190 And, what is worse, nothing was gained by the case-of-case transform.
2192 When should use this case of mkDupableCont?
2193 However, matching on *any* single-alternative case is a *disaster*;
2194 e.g. case (case ....) of (a,b) -> (# a,b #)
2195 We must push the outer case into the inner one!
2198 * Match [(DEFAULT,_,_)], but in the common case of Int,
2199 the alternative-filling-in code turned the outer case into
2200 case (...) of y { I# _ -> MkT y }
2202 * Match on single alternative plus (not (isDeadBinder case_bndr))
2203 Rationale: pushing the case inwards won't eliminate the construction.
2204 But there's a risk of
2205 case (...) of y { (a,b) -> let z=(a,b) in ... }
2206 Now y looks dead, but it'll come alive again. Still, this
2207 seems like the best option at the moment.
2209 * Match on single alternative plus (all (isDeadBinder bndrs))
2210 Rationale: this is essentially seq.
2212 * Match when the rhs is *not* duplicable, and hence would lead to a
2213 join point. This catches the disaster-case above. We can test
2214 the *un-simplified* rhs, which is fine. It might get bigger or
2215 smaller after simplification; if it gets smaller, this case might
2216 fire next time round. NB also that we must test contIsDupable
2217 case_cont *btoo, because case_cont might be big!
2219 HOWEVER: I found that this version doesn't work well, because
2220 we can get let x = case (...) of { small } in ...case x...
2221 When x is inlined into its full context, we find that it was a bad
2222 idea to have pushed the outer case inside the (...) case.
2224 Note [Single-alternative-unlifted]
2225 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
2226 Here's another single-alternative where we really want to do case-of-case:
2234 case y_s6X of tpl_s7m {
2235 M1.Mk1 ipv_s70 -> ipv_s70;
2236 M1.Mk2 ipv_s72 -> ipv_s72;
2242 case x_s74 of tpl_s7n {
2243 M1.Mk1 ipv_s77 -> ipv_s77;
2244 M1.Mk2 ipv_s79 -> ipv_s79;
2248 { __DEFAULT -> ==# [wild1_s7b wild_s7c];
2252 So the outer case is doing *nothing at all*, other than serving as a
2253 join-point. In this case we really want to do case-of-case and decide
2254 whether to use a real join point or just duplicate the continuation.
2256 Hence: check whether the case binder's type is unlifted, because then
2257 the outer case is *not* a seq.