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 )
22 import FamInstEnv ( topNormaliseType )
23 import DataCon ( DataCon, dataConWorkId, dataConRepStrictness )
25 import NewDemand ( isStrictDmd, splitStrictSig )
26 import PprCore ( pprParendExpr, pprCoreExpr )
27 import CoreUnfold ( mkUnfolding, mkCoreUnfolding, mkInlineRule,
28 exprIsConApp_maybe, callSiteInline, CallCtxt(..) )
30 import qualified CoreSubst
31 import CoreArity ( exprArity )
32 import Rules ( lookupRule, getRules )
33 import BasicTypes ( isMarkedStrict, Arity )
34 import CostCentre ( currentCCS, pushCCisNop )
35 import TysPrim ( realWorldStatePrimTy )
36 import PrelInfo ( realWorldPrimId )
37 import BasicTypes ( TopLevelFlag(..), isTopLevel,
38 RecFlag(..), isNonRuleLoopBreaker )
39 import MonadUtils ( foldlM )
40 import Maybes ( orElse )
41 import Data.List ( mapAccumL )
47 The guts of the simplifier is in this module, but the driver loop for
48 the simplifier is in SimplCore.lhs.
51 -----------------------------------------
52 *** IMPORTANT NOTE ***
53 -----------------------------------------
54 The simplifier used to guarantee that the output had no shadowing, but
55 it does not do so any more. (Actually, it never did!) The reason is
56 documented with simplifyArgs.
59 -----------------------------------------
60 *** IMPORTANT NOTE ***
61 -----------------------------------------
62 Many parts of the simplifier return a bunch of "floats" as well as an
63 expression. This is wrapped as a datatype SimplUtils.FloatsWith.
65 All "floats" are let-binds, not case-binds, but some non-rec lets may
66 be unlifted (with RHS ok-for-speculation).
70 -----------------------------------------
71 ORGANISATION OF FUNCTIONS
72 -----------------------------------------
74 - simplify all top-level binders
75 - for NonRec, call simplRecOrTopPair
76 - for Rec, call simplRecBind
79 ------------------------------
80 simplExpr (applied lambda) ==> simplNonRecBind
81 simplExpr (Let (NonRec ...) ..) ==> simplNonRecBind
82 simplExpr (Let (Rec ...) ..) ==> simplify binders; simplRecBind
84 ------------------------------
85 simplRecBind [binders already simplfied]
86 - use simplRecOrTopPair on each pair in turn
88 simplRecOrTopPair [binder already simplified]
89 Used for: recursive bindings (top level and nested)
90 top-level non-recursive bindings
92 - check for PreInlineUnconditionally
96 Used for: non-top-level non-recursive bindings
97 beta reductions (which amount to the same thing)
98 Because it can deal with strict arts, it takes a
99 "thing-inside" and returns an expression
101 - check for PreInlineUnconditionally
102 - simplify binder, including its IdInfo
111 simplNonRecX: [given a *simplified* RHS, but an *unsimplified* binder]
112 Used for: binding case-binder and constr args in a known-constructor case
113 - check for PreInLineUnconditionally
117 ------------------------------
118 simplLazyBind: [binder already simplified, RHS not]
119 Used for: recursive bindings (top level and nested)
120 top-level non-recursive bindings
121 non-top-level, but *lazy* non-recursive bindings
122 [must not be strict or unboxed]
123 Returns floats + an augmented environment, not an expression
124 - substituteIdInfo and add result to in-scope
125 [so that rules are available in rec rhs]
128 - float if exposes constructor or PAP
132 completeNonRecX: [binder and rhs both simplified]
133 - if the the thing needs case binding (unlifted and not ok-for-spec)
139 completeBind: [given a simplified RHS]
140 [used for both rec and non-rec bindings, top level and not]
141 - try PostInlineUnconditionally
142 - add unfolding [this is the only place we add an unfolding]
147 Right hand sides and arguments
148 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
149 In many ways we want to treat
150 (a) the right hand side of a let(rec), and
151 (b) a function argument
152 in the same way. But not always! In particular, we would
153 like to leave these arguments exactly as they are, so they
154 will match a RULE more easily.
159 It's harder to make the rule match if we ANF-ise the constructor,
160 or eta-expand the PAP:
162 f (let { a = g x; b = h x } in (a,b))
165 On the other hand if we see the let-defns
170 then we *do* want to ANF-ise and eta-expand, so that p and q
171 can be safely inlined.
173 Even floating lets out is a bit dubious. For let RHS's we float lets
174 out if that exposes a value, so that the value can be inlined more vigorously.
177 r = let x = e in (x,x)
179 Here, if we float the let out we'll expose a nice constructor. We did experiments
180 that showed this to be a generally good thing. But it was a bad thing to float
181 lets out unconditionally, because that meant they got allocated more often.
183 For function arguments, there's less reason to expose a constructor (it won't
184 get inlined). Just possibly it might make a rule match, but I'm pretty skeptical.
185 So for the moment we don't float lets out of function arguments either.
190 For eta expansion, we want to catch things like
192 case e of (a,b) -> \x -> case a of (p,q) -> \y -> r
194 If the \x was on the RHS of a let, we'd eta expand to bring the two
195 lambdas together. And in general that's a good thing to do. Perhaps
196 we should eta expand wherever we find a (value) lambda? Then the eta
197 expansion at a let RHS can concentrate solely on the PAP case.
200 %************************************************************************
202 \subsection{Bindings}
204 %************************************************************************
207 simplTopBinds :: SimplEnv -> [InBind] -> SimplM SimplEnv
209 simplTopBinds env0 binds0
210 = do { -- Put all the top-level binders into scope at the start
211 -- so that if a transformation rule has unexpectedly brought
212 -- anything into scope, then we don't get a complaint about that.
213 -- It's rather as if the top-level binders were imported.
214 ; env1 <- simplRecBndrs env0 (bindersOfBinds binds0)
215 ; dflags <- getDOptsSmpl
216 ; let dump_flag = dopt Opt_D_dump_inlinings dflags ||
217 dopt Opt_D_dump_rule_firings dflags
218 ; env2 <- simpl_binds dump_flag env1 binds0
219 ; freeTick SimplifierDone
222 -- We need to track the zapped top-level binders, because
223 -- they should have their fragile IdInfo zapped (notably occurrence info)
224 -- That's why we run down binds and bndrs' simultaneously.
226 -- The dump-flag emits a trace for each top-level binding, which
227 -- helps to locate the tracing for inlining and rule firing
228 simpl_binds :: Bool -> SimplEnv -> [InBind] -> SimplM SimplEnv
229 simpl_binds _ env [] = return env
230 simpl_binds dump env (bind:binds) = do { env' <- trace_bind dump bind $
232 ; simpl_binds dump env' binds }
234 trace_bind True bind = pprTrace "SimplBind" (ppr (bindersOf bind))
235 trace_bind False _ = \x -> x
237 simpl_bind env (Rec pairs) = simplRecBind env TopLevel pairs
238 simpl_bind env (NonRec b r) = simplRecOrTopPair env' TopLevel b b' r
240 (env', b') = addBndrRules env b (lookupRecBndr env b)
244 %************************************************************************
246 \subsection{Lazy bindings}
248 %************************************************************************
250 simplRecBind is used for
251 * recursive bindings only
254 simplRecBind :: SimplEnv -> TopLevelFlag
257 simplRecBind env0 top_lvl pairs0
258 = do { let (env_with_info, triples) = mapAccumL add_rules env0 pairs0
259 ; env1 <- go (zapFloats env_with_info) triples
260 ; return (env0 `addRecFloats` env1) }
261 -- addFloats adds the floats from env1,
262 -- _and_ updates env0 with the in-scope set from env1
264 add_rules :: SimplEnv -> (InBndr,InExpr) -> (SimplEnv, (InBndr, OutBndr, InExpr))
265 -- Add the (substituted) rules to the binder
266 add_rules env (bndr, rhs) = (env', (bndr, bndr', rhs))
268 (env', bndr') = addBndrRules env bndr (lookupRecBndr env bndr)
270 go env [] = return env
272 go env ((old_bndr, new_bndr, rhs) : pairs)
273 = do { env' <- simplRecOrTopPair env top_lvl old_bndr new_bndr rhs
277 simplOrTopPair is used for
278 * recursive bindings (whether top level or not)
279 * top-level non-recursive bindings
281 It assumes the binder has already been simplified, but not its IdInfo.
284 simplRecOrTopPair :: SimplEnv
286 -> InId -> OutBndr -> InExpr -- Binder and rhs
287 -> SimplM SimplEnv -- Returns an env that includes the binding
289 simplRecOrTopPair env top_lvl old_bndr new_bndr rhs
290 | preInlineUnconditionally env top_lvl old_bndr rhs -- Check for unconditional inline
291 = do { tick (PreInlineUnconditionally old_bndr)
292 ; return (extendIdSubst env old_bndr (mkContEx env rhs)) }
295 = simplLazyBind env top_lvl Recursive old_bndr new_bndr rhs env
296 -- May not actually be recursive, but it doesn't matter
300 simplLazyBind is used for
301 * [simplRecOrTopPair] recursive bindings (whether top level or not)
302 * [simplRecOrTopPair] top-level non-recursive bindings
303 * [simplNonRecE] non-top-level *lazy* non-recursive bindings
306 1. It assumes that the binder is *already* simplified,
307 and is in scope, and its IdInfo too, except unfolding
309 2. It assumes that the binder type is lifted.
311 3. It does not check for pre-inline-unconditionallly;
312 that should have been done already.
315 simplLazyBind :: SimplEnv
316 -> TopLevelFlag -> RecFlag
317 -> InId -> OutId -- Binder, both pre-and post simpl
318 -- The OutId has IdInfo, except arity, unfolding
319 -> InExpr -> SimplEnv -- The RHS and its environment
322 simplLazyBind env top_lvl is_rec bndr bndr1 rhs rhs_se
323 = do { let rhs_env = rhs_se `setInScope` env
324 (tvs, body) = case collectTyBinders rhs of
325 (tvs, body) | not_lam body -> (tvs,body)
326 | otherwise -> ([], rhs)
327 not_lam (Lam _ _) = False
329 -- Do not do the "abstract tyyvar" thing if there's
330 -- a lambda inside, becuase it defeats eta-reduction
331 -- f = /\a. \x. g a x
334 ; (body_env, tvs') <- simplBinders rhs_env tvs
335 -- See Note [Floating and type abstraction] in SimplUtils
338 ; (body_env1, body1) <- simplExprF body_env body mkRhsStop
339 -- ANF-ise a constructor or PAP rhs
340 ; (body_env2, body2) <- prepareRhs body_env1 body1
343 <- if not (doFloatFromRhs top_lvl is_rec False body2 body_env2)
344 then -- No floating, just wrap up!
345 do { rhs' <- mkLam env tvs' (wrapFloats body_env2 body2)
346 ; return (env, rhs') }
348 else if null tvs then -- Simple floating
349 do { tick LetFloatFromLet
350 ; return (addFloats env body_env2, body2) }
352 else -- Do type-abstraction first
353 do { tick LetFloatFromLet
354 ; (poly_binds, body3) <- abstractFloats tvs' body_env2 body2
355 ; rhs' <- mkLam env tvs' body3
356 ; env' <- foldlM (addPolyBind top_lvl) env poly_binds
357 ; return (env', rhs') }
359 ; completeBind env' top_lvl bndr bndr1 rhs' }
362 A specialised variant of simplNonRec used when the RHS is already simplified,
363 notably in knownCon. It uses case-binding where necessary.
366 simplNonRecX :: SimplEnv
367 -> InId -- Old binder
368 -> OutExpr -- Simplified RHS
371 simplNonRecX env bndr new_rhs
372 | isDeadBinder bndr -- Not uncommon; e.g. case (a,b) of b { (p,q) -> p }
373 = return env -- Here b is dead, and we avoid creating
374 | otherwise -- the binding b = (a,b)
375 = do { (env', bndr') <- simplBinder env bndr
376 ; completeNonRecX env' (isStrictId bndr) bndr bndr' new_rhs }
378 completeNonRecX :: SimplEnv
380 -> InId -- Old binder
381 -> OutId -- New binder
382 -> OutExpr -- Simplified RHS
385 completeNonRecX env is_strict old_bndr new_bndr new_rhs
386 = do { (env1, rhs1) <- prepareRhs (zapFloats env) new_rhs
388 if doFloatFromRhs NotTopLevel NonRecursive is_strict rhs1 env1
389 then do { tick LetFloatFromLet
390 ; return (addFloats env env1, rhs1) } -- Add the floats to the main env
391 else return (env, wrapFloats env1 rhs1) -- Wrap the floats around the RHS
392 ; completeBind env2 NotTopLevel old_bndr new_bndr rhs2 }
395 {- No, no, no! Do not try preInlineUnconditionally in completeNonRecX
396 Doing so risks exponential behaviour, because new_rhs has been simplified once already
397 In the cases described by the folowing commment, postInlineUnconditionally will
398 catch many of the relevant cases.
399 -- This happens; for example, the case_bndr during case of
400 -- known constructor: case (a,b) of x { (p,q) -> ... }
401 -- Here x isn't mentioned in the RHS, so we don't want to
402 -- create the (dead) let-binding let x = (a,b) in ...
404 -- Similarly, single occurrences can be inlined vigourously
405 -- e.g. case (f x, g y) of (a,b) -> ....
406 -- If a,b occur once we can avoid constructing the let binding for them.
408 Furthermore in the case-binding case preInlineUnconditionally risks extra thunks
409 -- Consider case I# (quotInt# x y) of
410 -- I# v -> let w = J# v in ...
411 -- If we gaily inline (quotInt# x y) for v, we end up building an
413 -- let w = J# (quotInt# x y) in ...
414 -- because quotInt# can fail.
416 | preInlineUnconditionally env NotTopLevel bndr new_rhs
417 = thing_inside (extendIdSubst env bndr (DoneEx new_rhs))
420 ----------------------------------
421 prepareRhs takes a putative RHS, checks whether it's a PAP or
422 constructor application and, if so, converts it to ANF, so that the
423 resulting thing can be inlined more easily. Thus
430 We also want to deal well cases like this
431 v = (f e1 `cast` co) e2
432 Here we want to make e1,e2 trivial and get
433 x1 = e1; x2 = e2; v = (f x1 `cast` co) v2
434 That's what the 'go' loop in prepareRhs does
437 prepareRhs :: SimplEnv -> OutExpr -> SimplM (SimplEnv, OutExpr)
438 -- Adds new floats to the env iff that allows us to return a good RHS
439 prepareRhs env (Cast rhs co) -- Note [Float coercions]
440 | (ty1, _ty2) <- coercionKind co -- Do *not* do this if rhs has an unlifted type
441 , not (isUnLiftedType ty1) -- see Note [Float coercions (unlifted)]
442 = do { (env', rhs') <- makeTrivial env rhs
443 ; return (env', Cast rhs' co) }
446 = do { (_is_val, env1, rhs1) <- go 0 env0 rhs0
447 ; return (env1, rhs1) }
449 go n_val_args env (Cast rhs co)
450 = do { (is_val, env', rhs') <- go n_val_args env rhs
451 ; return (is_val, env', Cast rhs' co) }
452 go n_val_args env (App fun (Type ty))
453 = do { (is_val, env', rhs') <- go n_val_args env fun
454 ; return (is_val, env', App rhs' (Type ty)) }
455 go n_val_args env (App fun arg)
456 = do { (is_val, env', fun') <- go (n_val_args+1) env fun
458 True -> do { (env'', arg') <- makeTrivial env' arg
459 ; return (True, env'', App fun' arg') }
460 False -> return (False, env, App fun arg) }
461 go n_val_args env (Var fun)
462 = return (is_val, env, Var fun)
464 is_val = n_val_args > 0 -- There is at least one arg
465 -- ...and the fun a constructor or PAP
466 && (isConLikeId fun || n_val_args < idArity fun)
467 -- See Note [CONLIKE pragma] in BasicTypes
469 = return (False, env, other)
473 Note [Float coercions]
474 ~~~~~~~~~~~~~~~~~~~~~~
475 When we find the binding
477 we'd like to transform it to
479 x = x `cast` co -- A trivial binding
480 There's a chance that e will be a constructor application or function, or something
481 like that, so moving the coerion to the usage site may well cancel the coersions
482 and lead to further optimisation. Example:
485 data instance T Int = T Int
487 foo :: Int -> Int -> Int
492 go n = case x of { T m -> go (n-m) }
493 -- This case should optimise
495 Note [Float coercions (unlifted)]
496 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
497 BUT don't do [Float coercions] if 'e' has an unlifted type.
500 foo :: Int = (error (# Int,Int #) "urk")
501 `cast` CoUnsafe (# Int,Int #) Int
503 If do the makeTrivial thing to the error call, we'll get
504 foo = case error (# Int,Int #) "urk" of v -> v `cast` ...
505 But 'v' isn't in scope!
507 These strange casts can happen as a result of case-of-case
508 bar = case (case x of { T -> (# 2,3 #); F -> error "urk" }) of
513 makeTrivial :: SimplEnv -> OutExpr -> SimplM (SimplEnv, OutExpr)
514 -- Binds the expression to a variable, if it's not trivial, returning the variable
518 | otherwise -- See Note [Take care] below
519 = do { var <- newId (fsLit "a") (exprType expr)
520 ; env' <- completeNonRecX env False var var expr
521 -- pprTrace "makeTrivial" (vcat [ppr var <+> ppr (exprArity (substExpr env' (Var var)))
523 -- , ppr (substExpr env' (Var var))
524 -- , ppr (idArity (fromJust (lookupInScope (seInScope env') var))) ]) $
525 ; return (env', substExpr env' (Var var)) }
526 -- The substitution is needed becase we're constructing a new binding
528 -- And if rhs is of form (rhs1 |> co), then we might get
531 -- and now a's RHS is trivial and can be substituted out, and that
532 -- is what completeNonRecX will do
536 %************************************************************************
538 \subsection{Completing a lazy binding}
540 %************************************************************************
543 * deals only with Ids, not TyVars
544 * takes an already-simplified binder and RHS
545 * is used for both recursive and non-recursive bindings
546 * is used for both top-level and non-top-level bindings
548 It does the following:
549 - tries discarding a dead binding
550 - tries PostInlineUnconditionally
551 - add unfolding [this is the only place we add an unfolding]
554 It does *not* attempt to do let-to-case. Why? Because it is used for
555 - top-level bindings (when let-to-case is impossible)
556 - many situations where the "rhs" is known to be a WHNF
557 (so let-to-case is inappropriate).
559 Nor does it do the atomic-argument thing
562 completeBind :: SimplEnv
563 -> TopLevelFlag -- Flag stuck into unfolding
564 -> InId -- Old binder
565 -> OutId -> OutExpr -- New binder and RHS
567 -- completeBind may choose to do its work
568 -- * by extending the substitution (e.g. let x = y in ...)
569 -- * or by adding to the floats in the envt
571 completeBind env top_lvl old_bndr new_bndr new_rhs
572 = do { let old_info = idInfo old_bndr
573 old_unf = unfoldingInfo old_info
574 occ_info = occInfo old_info
576 ; new_unfolding <- simplUnfolding env top_lvl old_bndr occ_info new_rhs old_unf
578 ; if postInlineUnconditionally env top_lvl new_bndr occ_info new_rhs new_unfolding
579 -- Inline and discard the binding
580 then do { tick (PostInlineUnconditionally old_bndr)
581 ; return (extendIdSubst env old_bndr (DoneEx new_rhs)) }
582 -- Use the substitution to make quite, quite sure that the
583 -- substitution will happen, since we are going to discard the binding
585 else return (addNonRecWithUnf env new_bndr new_rhs new_unfolding) }
587 ------------------------------
588 addPolyBind :: TopLevelFlag -> SimplEnv -> OutBind -> SimplM SimplEnv
589 -- Add a new binding to the environment, complete with its unfolding
590 -- but *do not* do postInlineUnconditionally, because we have already
591 -- processed some of the scope of the binding
592 -- We still want the unfolding though. Consider
594 -- x = /\a. let y = ... in Just y
596 -- Then we float the y-binding out (via abstractFloats and addPolyBind)
597 -- but 'x' may well then be inlined in 'body' in which case we'd like the
598 -- opportunity to inline 'y' too.
600 addPolyBind top_lvl env (NonRec poly_id rhs)
601 = do { unfolding <- simplUnfolding env top_lvl poly_id NoOccInfo rhs noUnfolding
602 -- Assumes that poly_id did not have an INLINE prag
603 -- which is perhaps wrong. ToDo: think about this
604 ; return (addNonRecWithUnf env poly_id rhs unfolding) }
606 addPolyBind _ env bind@(Rec _) = return (extendFloats env bind)
607 -- Hack: letrecs are more awkward, so we extend "by steam"
608 -- without adding unfoldings etc. At worst this leads to
609 -- more simplifier iterations
611 ------------------------------
612 addNonRecWithUnf :: SimplEnv
613 -> OutId -> OutExpr -- New binder and RHS
614 -> Unfolding -- New unfolding
616 addNonRecWithUnf env new_bndr new_rhs new_unfolding
617 = let new_arity = exprArity new_rhs
618 old_arity = idArity new_bndr
619 info1 = idInfo new_bndr `setArityInfo` new_arity
621 -- Unfolding info: Note [Setting the new unfolding]
622 info2 = info1 `setUnfoldingInfo` new_unfolding
624 -- Demand info: Note [Setting the demand info]
625 info3 | isEvaldUnfolding new_unfolding = zapDemandInfo info2 `orElse` info2
628 final_id = new_bndr `setIdInfo` info3
629 dmd_arity = length $ fst $ splitStrictSig $ idNewStrictness new_bndr
631 ASSERT( isId new_bndr )
632 WARN( new_arity < old_arity || new_arity < dmd_arity,
633 (ptext (sLit "Arity decrease:") <+> ppr final_id <+> ppr old_arity
634 <+> ppr new_arity <+> ppr dmd_arity) )
635 -- Note [Arity decrease]
637 final_id `seq` -- This seq forces the Id, and hence its IdInfo,
638 -- and hence any inner substitutions
639 -- pprTrace "Binding" (ppr final_id <+> ppr unfolding) $
640 addNonRec env final_id new_rhs
641 -- The addNonRec adds it to the in-scope set too
643 ------------------------------
644 simplUnfolding :: SimplEnv-> TopLevelFlag
645 -> Id -- Debug output only
646 -> OccInfo -> OutExpr
647 -> Unfolding -> SimplM Unfolding
648 -- Note [Setting the new unfolding]
649 simplUnfolding env _ _ _ _ (DFunUnfolding con ops)
650 = return (DFunUnfolding con ops')
652 ops' = map (CoreSubst.substExpr (mkCoreSubst env)) ops
654 simplUnfolding env top_lvl _ _ _
655 (CoreUnfolding { uf_tmpl = expr, uf_arity = arity
656 , uf_guidance = guide@(InlineRule {}) })
657 = do { expr' <- simplExpr (setMode simplGentlyForInlineRules env) expr
658 -- See Note [Simplifying gently inside InlineRules] in SimplUtils
659 ; let mb_wkr' = CoreSubst.substInlineRuleInfo (mkCoreSubst env) (ir_info guide)
660 ; return (mkCoreUnfolding (isTopLevel top_lvl) expr' arity
661 (guide { ir_info = mb_wkr' })) }
662 -- See Note [Top-level flag on inline rules] in CoreUnfold
664 simplUnfolding _ top_lvl _ occ_info new_rhs _
665 | omit_unfolding = return NoUnfolding
666 | otherwise = return (mkUnfolding (isTopLevel top_lvl) new_rhs)
668 omit_unfolding = isNonRuleLoopBreaker occ_info
671 Note [Arity decrease]
672 ~~~~~~~~~~~~~~~~~~~~~
673 Generally speaking the arity of a binding should not decrease. But it *can*
674 legitimately happen becuase of RULES. Eg
676 where g has arity 2, will have arity 2. But if there's a rewrite rule
678 where h has arity 1, then f's arity will decrease. Here's a real-life example,
679 which is in the output of Specialise:
682 $dm {Arity 2} = \d.\x. op d
683 {-# RULES forall d. $dm Int d = $s$dm #-}
685 dInt = MkD .... opInt ...
686 opInt {Arity 1} = $dm dInt
688 $s$dm {Arity 0} = \x. op dInt }
690 Here opInt has arity 1; but when we apply the rule its arity drops to 0.
691 That's why Specialise goes to a little trouble to pin the right arity
692 on specialised functions too.
694 Note [Setting the new unfolding]
695 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
696 * If there's an INLINE pragma, we simplify the RHS gently. Maybe we
697 should do nothing at all, but simplifying gently might get rid of
700 * If not, we make an unfolding from the new RHS. But *only* for
701 non-loop-breakers. Making loop breakers not have an unfolding at all
702 means that we can avoid tests in exprIsConApp, for example. This is
703 important: if exprIsConApp says 'yes' for a recursive thing, then we
704 can get into an infinite loop
706 If there's an InlineRule on a loop breaker, we hang on to the inlining.
707 It's pretty dodgy, but the user did say 'INLINE'. May need to revisit
710 Note [Setting the demand info]
711 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
712 If the unfolding is a value, the demand info may
713 go pear-shaped, so we nuke it. Example:
715 case x of (p,q) -> h p q x
716 Here x is certainly demanded. But after we've nuked
717 the case, we'll get just
718 let x = (a,b) in h a b x
719 and now x is not demanded (I'm assuming h is lazy)
720 This really happens. Similarly
721 let f = \x -> e in ...f..f...
722 After inlining f at some of its call sites the original binding may
723 (for example) be no longer strictly demanded.
724 The solution here is a bit ad hoc...
727 %************************************************************************
729 \subsection[Simplify-simplExpr]{The main function: simplExpr}
731 %************************************************************************
733 The reason for this OutExprStuff stuff is that we want to float *after*
734 simplifying a RHS, not before. If we do so naively we get quadratic
735 behaviour as things float out.
737 To see why it's important to do it after, consider this (real) example:
751 a -- Can't inline a this round, cos it appears twice
755 Each of the ==> steps is a round of simplification. We'd save a
756 whole round if we float first. This can cascade. Consider
761 let f = let d1 = ..d.. in \y -> e
765 in \x -> ...(\y ->e)...
767 Only in this second round can the \y be applied, and it
768 might do the same again.
772 simplExpr :: SimplEnv -> CoreExpr -> SimplM CoreExpr
773 simplExpr env expr = simplExprC env expr mkBoringStop
775 simplExprC :: SimplEnv -> CoreExpr -> SimplCont -> SimplM CoreExpr
776 -- Simplify an expression, given a continuation
777 simplExprC env expr cont
778 = -- pprTrace "simplExprC" (ppr expr $$ ppr cont {- $$ ppr (seIdSubst env) -} $$ ppr (seFloats env) ) $
779 do { (env', expr') <- simplExprF (zapFloats env) expr cont
780 ; -- pprTrace "simplExprC ret" (ppr expr $$ ppr expr') $
781 -- pprTrace "simplExprC ret3" (ppr (seInScope env')) $
782 -- pprTrace "simplExprC ret4" (ppr (seFloats env')) $
783 return (wrapFloats env' expr') }
785 --------------------------------------------------
786 simplExprF :: SimplEnv -> InExpr -> SimplCont
787 -> SimplM (SimplEnv, OutExpr)
789 simplExprF env e cont
790 = -- pprTrace "simplExprF" (ppr e $$ ppr cont $$ ppr (seTvSubst env) $$ ppr (seIdSubst env) {- $$ ppr (seFloats env) -} ) $
791 simplExprF' env e cont
793 simplExprF' :: SimplEnv -> InExpr -> SimplCont
794 -> SimplM (SimplEnv, OutExpr)
795 simplExprF' env (Var v) cont = simplVar env v cont
796 simplExprF' env (Lit lit) cont = rebuild env (Lit lit) cont
797 simplExprF' env (Note n expr) cont = simplNote env n expr cont
798 simplExprF' env (Cast body co) cont = simplCast env body co cont
799 simplExprF' env (App fun arg) cont = simplExprF env fun $
800 ApplyTo NoDup arg env cont
802 simplExprF' env expr@(Lam _ _) cont
803 = simplLam env (map zap bndrs) body cont
804 -- The main issue here is under-saturated lambdas
805 -- (\x1. \x2. e) arg1
806 -- Here x1 might have "occurs-once" occ-info, because occ-info
807 -- is computed assuming that a group of lambdas is applied
808 -- all at once. If there are too few args, we must zap the
811 n_args = countArgs cont
812 n_params = length bndrs
813 (bndrs, body) = collectBinders expr
814 zap | n_args >= n_params = \b -> b
815 | otherwise = \b -> if isTyVar b then b
817 -- NB: we count all the args incl type args
818 -- so we must count all the binders (incl type lambdas)
820 simplExprF' env (Type ty) cont
821 = ASSERT( contIsRhsOrArg cont )
822 do { ty' <- simplCoercion env ty
823 ; rebuild env (Type ty') cont }
825 simplExprF' env (Case scrut bndr _ alts) cont
826 | not (switchIsOn (getSwitchChecker env) NoCaseOfCase)
827 = -- Simplify the scrutinee with a Select continuation
828 simplExprF env scrut (Select NoDup bndr alts env cont)
831 = -- If case-of-case is off, simply simplify the case expression
832 -- in a vanilla Stop context, and rebuild the result around it
833 do { case_expr' <- simplExprC env scrut case_cont
834 ; rebuild env case_expr' cont }
836 case_cont = Select NoDup bndr alts env mkBoringStop
838 simplExprF' env (Let (Rec pairs) body) cont
839 = do { env' <- simplRecBndrs env (map fst pairs)
840 -- NB: bndrs' don't have unfoldings or rules
841 -- We add them as we go down
843 ; env'' <- simplRecBind env' NotTopLevel pairs
844 ; simplExprF env'' body cont }
846 simplExprF' env (Let (NonRec bndr rhs) body) cont
847 = simplNonRecE env bndr (rhs, env) ([], body) cont
849 ---------------------------------
850 simplType :: SimplEnv -> InType -> SimplM OutType
851 -- Kept monadic just so we can do the seqType
853 = -- pprTrace "simplType" (ppr ty $$ ppr (seTvSubst env)) $
854 seqType new_ty `seq` return new_ty
856 new_ty = substTy env ty
858 ---------------------------------
859 simplCoercion :: SimplEnv -> InType -> SimplM OutType
860 -- The InType isn't *necessarily* a coercion, but it might be
861 -- (in a type application, say) and optCoercion is a no-op on types
863 = do { co' <- simplType env co
864 ; return (optCoercion co') }
868 %************************************************************************
870 \subsection{The main rebuilder}
872 %************************************************************************
875 rebuild :: SimplEnv -> OutExpr -> SimplCont -> SimplM (SimplEnv, OutExpr)
876 -- At this point the substitution in the SimplEnv should be irrelevant
877 -- only the in-scope set and floats should matter
878 rebuild env expr cont0
879 = -- pprTrace "rebuild" (ppr expr $$ ppr cont0 $$ ppr (seFloats env)) $
881 Stop {} -> return (env, expr)
882 CoerceIt co cont -> rebuild env (mkCoerce co expr) cont
883 Select _ bndr alts se cont -> rebuildCase (se `setFloats` env) expr bndr alts cont
884 StrictArg fun _ info cont -> rebuildCall env (fun `App` expr) info cont
885 StrictBind b bs body se cont -> do { env' <- simplNonRecX (se `setFloats` env) b expr
886 ; simplLam env' bs body cont }
887 ApplyTo _ arg se cont -> do { arg' <- simplExpr (se `setInScope` env) arg
888 ; rebuild env (App expr arg') cont }
892 %************************************************************************
896 %************************************************************************
899 simplCast :: SimplEnv -> InExpr -> Coercion -> SimplCont
900 -> SimplM (SimplEnv, OutExpr)
901 simplCast env body co0 cont0
902 = do { co1 <- simplCoercion env co0
903 ; simplExprF env body (addCoerce co1 cont0) }
905 addCoerce co cont = add_coerce co (coercionKind co) cont
907 add_coerce _co (s1, k1) cont -- co :: ty~ty
908 | s1 `coreEqType` k1 = cont -- is a no-op
910 add_coerce co1 (s1, _k2) (CoerceIt co2 cont)
911 | (_l1, t1) <- coercionKind co2
912 -- e |> (g1 :: S1~L) |> (g2 :: L~T1)
915 -- e |> (g1 . g2 :: S1~T1) otherwise
917 -- For example, in the initial form of a worker
918 -- we may find (coerce T (coerce S (\x.e))) y
919 -- and we'd like it to simplify to e[y/x] in one round
921 , s1 `coreEqType` t1 = cont -- The coerces cancel out
922 | otherwise = CoerceIt (mkTransCoercion co1 co2) cont
924 add_coerce co (s1s2, _t1t2) (ApplyTo dup (Type arg_ty) arg_se cont)
925 -- (f |> g) ty ---> (f ty) |> (g @ ty)
926 -- This implements the PushT rule from the paper
927 | Just (tyvar,_) <- splitForAllTy_maybe s1s2
928 , not (isCoVar tyvar)
929 = ApplyTo dup (Type ty') (zapSubstEnv env) (addCoerce (mkInstCoercion co ty') cont)
931 ty' = substTy (arg_se `setInScope` env) arg_ty
933 -- ToDo: the PushC rule is not implemented at all
935 add_coerce co (s1s2, _t1t2) (ApplyTo dup arg arg_se cont)
936 | not (isTypeArg arg) -- This implements the Push rule from the paper
937 , isFunTy s1s2 -- t1t2 must be a function type, becuase it's applied
938 -- (e |> (g :: s1s2 ~ t1->t2)) f
940 -- (e (f |> (arg g :: t1~s1))
941 -- |> (res g :: s2->t2)
943 -- t1t2 must be a function type, t1->t2, because it's applied
944 -- to something but s1s2 might conceivably not be
946 -- When we build the ApplyTo we can't mix the out-types
947 -- with the InExpr in the argument, so we simply substitute
948 -- to make it all consistent. It's a bit messy.
949 -- But it isn't a common case.
951 -- Example of use: Trac #995
952 = ApplyTo dup new_arg (zapSubstEnv env) (addCoerce co2 cont)
954 -- we split coercion t1->t2 ~ s1->s2 into t1 ~ s1 and
955 -- t2 ~ s2 with left and right on the curried form:
956 -- (->) t1 t2 ~ (->) s1 s2
957 [co1, co2] = decomposeCo 2 co
958 new_arg = mkCoerce (mkSymCoercion co1) arg'
959 arg' = substExpr (arg_se `setInScope` env) arg
961 add_coerce co _ cont = CoerceIt co cont
965 %************************************************************************
969 %************************************************************************
972 simplLam :: SimplEnv -> [InId] -> InExpr -> SimplCont
973 -> SimplM (SimplEnv, OutExpr)
975 simplLam env [] body cont = simplExprF env body cont
978 simplLam env (bndr:bndrs) body (ApplyTo _ arg arg_se cont)
979 = do { tick (BetaReduction bndr)
980 ; simplNonRecE env bndr (arg, arg_se) (bndrs, body) cont }
982 -- Not enough args, so there are real lambdas left to put in the result
983 simplLam env bndrs body cont
984 = do { (env', bndrs') <- simplLamBndrs env bndrs
985 ; body' <- simplExpr env' body
986 ; new_lam <- mkLam env' bndrs' body'
987 ; rebuild env' new_lam cont }
990 simplNonRecE :: SimplEnv
991 -> InBndr -- The binder
992 -> (InExpr, SimplEnv) -- Rhs of binding (or arg of lambda)
993 -> ([InBndr], InExpr) -- Body of the let/lambda
996 -> SimplM (SimplEnv, OutExpr)
998 -- simplNonRecE is used for
999 -- * non-top-level non-recursive lets in expressions
1002 -- It deals with strict bindings, via the StrictBind continuation,
1003 -- which may abort the whole process
1005 -- The "body" of the binding comes as a pair of ([InId],InExpr)
1006 -- representing a lambda; so we recurse back to simplLam
1007 -- Why? Because of the binder-occ-info-zapping done before
1008 -- the call to simplLam in simplExprF (Lam ...)
1010 -- First deal with type applications and type lets
1011 -- (/\a. e) (Type ty) and (let a = Type ty in e)
1012 simplNonRecE env bndr (Type ty_arg, rhs_se) (bndrs, body) cont
1013 = ASSERT( isTyVar bndr )
1014 do { ty_arg' <- simplType (rhs_se `setInScope` env) ty_arg
1015 ; simplLam (extendTvSubst env bndr ty_arg') bndrs body cont }
1017 simplNonRecE env bndr (rhs, rhs_se) (bndrs, body) cont
1018 | preInlineUnconditionally env NotTopLevel bndr rhs
1019 = do { tick (PreInlineUnconditionally bndr)
1020 ; simplLam (extendIdSubst env bndr (mkContEx rhs_se rhs)) bndrs body cont }
1023 = do { simplExprF (rhs_se `setFloats` env) rhs
1024 (StrictBind bndr bndrs body env cont) }
1027 = ASSERT( not (isTyVar bndr) )
1028 do { (env1, bndr1) <- simplNonRecBndr env bndr
1029 ; let (env2, bndr2) = addBndrRules env1 bndr bndr1
1030 ; env3 <- simplLazyBind env2 NotTopLevel NonRecursive bndr bndr2 rhs rhs_se
1031 ; simplLam env3 bndrs body cont }
1035 %************************************************************************
1039 %************************************************************************
1042 -- Hack alert: we only distinguish subsumed cost centre stacks for the
1043 -- purposes of inlining. All other CCCSs are mapped to currentCCS.
1044 simplNote :: SimplEnv -> Note -> CoreExpr -> SimplCont
1045 -> SimplM (SimplEnv, OutExpr)
1046 simplNote env (SCC cc) e cont
1047 | pushCCisNop cc (getEnclosingCC env) -- scc "f" (...(scc "f" e)...)
1048 = simplExprF env e cont -- ==> scc "f" (...e...)
1050 = do { e' <- simplExpr (setEnclosingCC env currentCCS) e
1051 ; rebuild env (mkSCC cc e') cont }
1053 simplNote env (CoreNote s) e cont
1054 = do { e' <- simplExpr env e
1055 ; rebuild env (Note (CoreNote s) e') cont }
1059 %************************************************************************
1061 \subsection{Dealing with calls}
1063 %************************************************************************
1066 simplVar :: SimplEnv -> Id -> SimplCont -> SimplM (SimplEnv, OutExpr)
1067 simplVar env var cont
1068 = case substId env var of
1069 DoneEx e -> simplExprF (zapSubstEnv env) e cont
1070 ContEx tvs ids e -> simplExprF (setSubstEnv env tvs ids) e cont
1071 DoneId var1 -> completeCall (zapSubstEnv env) var1 cont
1072 -- Note [zapSubstEnv]
1073 -- The template is already simplified, so don't re-substitute.
1074 -- This is VITAL. Consider
1076 -- let y = \z -> ...x... in
1078 -- We'll clone the inner \x, adding x->x' in the id_subst
1079 -- Then when we inline y, we must *not* replace x by x' in
1080 -- the inlined copy!!
1082 ---------------------------------------------------------
1083 -- Dealing with a call site
1085 completeCall :: SimplEnv -> Id -> SimplCont -> SimplM (SimplEnv, OutExpr)
1086 completeCall env var cont
1087 = do { let (args,call_cont) = contArgs cont
1088 -- The args are OutExprs, obtained by *lazily* substituting
1089 -- in the args found in cont. These args are only examined
1090 -- to limited depth (unless a rule fires). But we must do
1091 -- the substitution; rule matching on un-simplified args would
1094 ------------- First try rules ----------------
1095 -- Do this before trying inlining. Some functions have
1096 -- rules *and* are strict; in this case, we don't want to
1097 -- inline the wrapper of the non-specialised thing; better
1098 -- to call the specialised thing instead.
1100 -- We used to use the black-listing mechanism to ensure that inlining of
1101 -- the wrapper didn't occur for things that have specialisations till a
1102 -- later phase, so but now we just try RULES first
1104 -- See also Note [Rules for recursive functions]
1105 ; rule_base <- getSimplRules
1106 ; let rules = getRules rule_base var
1107 ; mb_rule <- tryRules env var rules args call_cont
1109 Just (n_args, rule_rhs) -> simplExprF env rule_rhs (dropArgs n_args cont) ;
1110 -- The ruleArity says how many args the rule consumed
1111 ; Nothing -> do -- No rules
1114 ------------- Next try inlining ----------------
1115 { dflags <- getDOptsSmpl
1116 ; let arg_infos = [interestingArg arg | arg <- args, isValArg arg]
1117 n_val_args = length arg_infos
1118 interesting_cont = interestingCallContext call_cont
1119 active_inline = activeInline env var
1120 maybe_inline = callSiteInline dflags active_inline var
1121 (null args) arg_infos interesting_cont
1122 ; case maybe_inline of {
1123 Just unfolding -- There is an inlining!
1124 -> do { tick (UnfoldingDone var)
1125 ; (if dopt Opt_D_dump_inlinings dflags then
1126 pprTrace ("Inlining done: " ++ showSDoc (ppr var)) (vcat [
1127 text "Before:" <+> ppr var <+> sep (map pprParendExpr args),
1128 text "Inlined fn: " <+> nest 2 (ppr unfolding),
1129 text "Cont: " <+> ppr call_cont])
1132 simplExprF env unfolding cont }
1134 ; Nothing -> -- No inlining!
1136 ------------- No inlining! ----------------
1137 -- Next, look for rules or specialisations that match
1139 rebuildCall env (Var var)
1140 (mkArgInfo var rules n_val_args call_cont)
1144 rebuildCall :: SimplEnv
1145 -> OutExpr -- Function
1148 -> SimplM (SimplEnv, OutExpr)
1149 rebuildCall env fun (ArgInfo { 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 fun) -- contination to discard, else we do it
1162 where -- again and again!
1163 fun_ty = exprType fun
1164 cont_ty = contResultType env fun_ty cont
1165 co = mkUnsafeCoercion fun_ty cont_ty
1166 mk_coerce expr | cont_ty `coreEqType` fun_ty = expr
1167 | otherwise = mkCoerce co expr
1169 rebuildCall env fun info (ApplyTo _ (Type arg_ty) se cont)
1170 = do { ty' <- simplCoercion (se `setInScope` env) arg_ty
1171 ; rebuildCall env (fun `App` Type ty') info cont }
1174 (ArgInfo { ai_rules = has_rules, ai_strs = str:strs, ai_discs = disc:discs })
1175 (ApplyTo _ arg arg_se cont)
1176 | str -- Strict argument
1177 = -- pprTrace "Strict Arg" (ppr arg $$ ppr (seIdSubst env) $$ ppr (seInScope env)) $
1178 simplExprF (arg_se `setFloats` env) arg
1179 (StrictArg fun cci arg_info' cont)
1182 | otherwise -- Lazy argument
1183 -- DO NOT float anything outside, hence simplExprC
1184 -- There is no benefit (unlike in a let-binding), and we'd
1185 -- have to be very careful about bogus strictness through
1186 -- floating a demanded let.
1187 = do { arg' <- simplExprC (arg_se `setInScope` env) arg
1189 ; rebuildCall env (fun `App` arg') arg_info' cont }
1191 arg_info' = ArgInfo { ai_rules = has_rules, ai_strs = strs, ai_discs = discs }
1192 cci | has_rules || disc > 0 = ArgCtxt has_rules -- Be keener here
1193 | otherwise = BoringCtxt -- Nothing interesting
1195 rebuildCall env fun _ cont
1196 = rebuild env fun cont
1201 This part of the simplifier may break the no-shadowing invariant
1203 f (...(\a -> e)...) (case y of (a,b) -> e')
1204 where f is strict in its second arg
1205 If we simplify the innermost one first we get (...(\a -> e)...)
1206 Simplifying the second arg makes us float the case out, so we end up with
1207 case y of (a,b) -> f (...(\a -> e)...) e'
1208 So the output does not have the no-shadowing invariant. However, there is
1209 no danger of getting name-capture, because when the first arg was simplified
1210 we used an in-scope set that at least mentioned all the variables free in its
1211 static environment, and that is enough.
1213 We can't just do innermost first, or we'd end up with a dual problem:
1214 case x of (a,b) -> f e (...(\a -> e')...)
1216 I spent hours trying to recover the no-shadowing invariant, but I just could
1217 not think of an elegant way to do it. The simplifier is already knee-deep in
1218 continuations. We have to keep the right in-scope set around; AND we have
1219 to get the effect that finding (error "foo") in a strict arg position will
1220 discard the entire application and replace it with (error "foo"). Getting
1221 all this at once is TOO HARD!
1224 %************************************************************************
1228 %************************************************************************
1231 tryRules :: SimplEnv
1232 -> Id -> [CoreRule] -> [OutExpr] -> SimplCont
1233 -> SimplM (Maybe (Arity, CoreExpr)) -- The arity is the number of
1234 -- args consumed by the rule
1235 tryRules env fn rules args call_cont
1239 = do { dflags <- getDOptsSmpl
1240 ; case activeRule dflags env of {
1241 Nothing -> return Nothing ; -- No rules apply
1244 case lookupRule act_fn (getInScope env) fn args rules of {
1245 Nothing -> return Nothing ; -- No rule matches
1246 Just (rule, rule_rhs) ->
1248 do { tick (RuleFired (ru_name rule))
1249 ; (if dopt Opt_D_dump_rule_firings dflags then
1250 pprTrace "Rule fired" (vcat [
1251 text "Rule:" <+> ftext (ru_name rule),
1252 text "Before:" <+> ppr fn <+> sep (map pprParendExpr args),
1253 text "After: " <+> pprCoreExpr rule_rhs,
1254 text "Cont: " <+> ppr call_cont])
1257 return (Just (ruleArity rule, rule_rhs)) }}}}
1260 Note [Rules for recursive functions]
1261 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
1262 You might think that we shouldn't apply rules for a loop breaker:
1263 doing so might give rise to an infinite loop, because a RULE is
1264 rather like an extra equation for the function:
1265 RULE: f (g x) y = x+y
1268 But it's too drastic to disable rules for loop breakers.
1269 Even the foldr/build rule would be disabled, because foldr
1270 is recursive, and hence a loop breaker:
1271 foldr k z (build g) = g k z
1272 So it's up to the programmer: rules can cause divergence
1275 %************************************************************************
1277 Rebuilding a cse expression
1279 %************************************************************************
1281 Note [Case elimination]
1282 ~~~~~~~~~~~~~~~~~~~~~~~
1283 The case-elimination transformation discards redundant case expressions.
1284 Start with a simple situation:
1286 case x# of ===> e[x#/y#]
1289 (when x#, y# are of primitive type, of course). We can't (in general)
1290 do this for algebraic cases, because we might turn bottom into
1293 The code in SimplUtils.prepareAlts has the effect of generalise this
1294 idea to look for a case where we're scrutinising a variable, and we
1295 know that only the default case can match. For example:
1299 DEFAULT -> ...(case x of
1303 Here the inner case is first trimmed to have only one alternative, the
1304 DEFAULT, after which it's an instance of the previous case. This
1305 really only shows up in eliminating error-checking code.
1307 We also make sure that we deal with this very common case:
1312 Here we are using the case as a strict let; if x is used only once
1313 then we want to inline it. We have to be careful that this doesn't
1314 make the program terminate when it would have diverged before, so we
1316 - e is already evaluated (it may so if e is a variable)
1317 - x is used strictly, or
1319 Lastly, the code in SimplUtils.mkCase combines identical RHSs. So
1321 case e of ===> case e of DEFAULT -> r
1325 Now again the case may be elminated by the CaseElim transformation.
1328 Further notes about case elimination
1329 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
1330 Consider: test :: Integer -> IO ()
1333 Turns out that this compiles to:
1336 eta1 :: State# RealWorld ->
1337 case PrelNum.< eta PrelNum.zeroInteger of wild { __DEFAULT ->
1339 (PrelNum.jtos eta ($w[] @ Char))
1341 of wild1 { (# new_s, a4 #) -> PrelIO.lvl23 new_s }}
1343 Notice the strange '<' which has no effect at all. This is a funny one.
1344 It started like this:
1346 f x y = if x < 0 then jtos x
1347 else if y==0 then "" else jtos x
1349 At a particular call site we have (f v 1). So we inline to get
1351 if v < 0 then jtos x
1352 else if 1==0 then "" else jtos x
1354 Now simplify the 1==0 conditional:
1356 if v<0 then jtos v else jtos v
1358 Now common-up the two branches of the case:
1360 case (v<0) of DEFAULT -> jtos v
1362 Why don't we drop the case? Because it's strict in v. It's technically
1363 wrong to drop even unnecessary evaluations, and in practice they
1364 may be a result of 'seq' so we *definitely* don't want to drop those.
1365 I don't really know how to improve this situation.
1368 ---------------------------------------------------------
1369 -- Eliminate the case if possible
1371 rebuildCase, reallyRebuildCase
1373 -> OutExpr -- Scrutinee
1374 -> InId -- Case binder
1375 -> [InAlt] -- Alternatives (inceasing order)
1377 -> SimplM (SimplEnv, OutExpr)
1379 --------------------------------------------------
1380 -- 1. Eliminate the case if there's a known constructor
1381 --------------------------------------------------
1383 rebuildCase env scrut case_bndr alts cont
1384 | Lit lit <- scrut -- No need for same treatment as constructors
1385 -- because literals are inlined more vigorously
1386 = do { tick (KnownBranch case_bndr)
1387 ; case findAlt (LitAlt lit) alts of
1388 Nothing -> missingAlt env case_bndr alts cont
1389 Just (_, bs, rhs) -> simple_rhs bs rhs }
1391 | Just (con, ty_args, other_args) <- exprIsConApp_maybe scrut
1392 -- Works when the scrutinee is a variable with a known unfolding
1393 -- as well as when it's an explicit constructor application
1394 = do { tick (KnownBranch case_bndr)
1395 ; case findAlt (DataAlt con) alts of
1396 Nothing -> missingAlt env case_bndr alts cont
1397 Just (DEFAULT, bs, rhs) -> simple_rhs bs rhs
1398 Just (_, bs, rhs) -> knownCon env scrut con ty_args other_args
1399 case_bndr bs rhs cont
1402 simple_rhs bs rhs = ASSERT( null bs )
1403 do { env' <- simplNonRecX env case_bndr scrut
1404 ; simplExprF env' rhs cont }
1407 --------------------------------------------------
1408 -- 2. Eliminate the case if scrutinee is evaluated
1409 --------------------------------------------------
1411 rebuildCase env scrut case_bndr [(_, bndrs, rhs)] cont
1412 -- See if we can get rid of the case altogether
1413 -- See Note [Case eliminiation]
1414 -- mkCase made sure that if all the alternatives are equal,
1415 -- then there is now only one (DEFAULT) rhs
1416 | all isDeadBinder bndrs -- bndrs are [InId]
1418 -- Check that the scrutinee can be let-bound instead of case-bound
1419 , exprOkForSpeculation scrut
1420 -- OK not to evaluate it
1421 -- This includes things like (==# a# b#)::Bool
1422 -- so that we simplify
1423 -- case ==# a# b# of { True -> x; False -> x }
1426 -- This particular example shows up in default methods for
1427 -- comparision operations (e.g. in (>=) for Int.Int32)
1428 || exprIsHNF scrut -- It's already evaluated
1429 || var_demanded_later scrut -- It'll be demanded later
1431 -- || not opt_SimplPedanticBottoms) -- Or we don't care!
1432 -- We used to allow improving termination by discarding cases, unless -fpedantic-bottoms was on,
1433 -- but that breaks badly for the dataToTag# primop, which relies on a case to evaluate
1434 -- its argument: case x of { y -> dataToTag# y }
1435 -- Here we must *not* discard the case, because dataToTag# just fetches the tag from
1436 -- the info pointer. So we'll be pedantic all the time, and see if that gives any
1438 -- Also we don't want to discard 'seq's
1439 = do { tick (CaseElim case_bndr)
1440 ; env' <- simplNonRecX env case_bndr scrut
1441 ; simplExprF env' rhs cont }
1443 -- The case binder is going to be evaluated later,
1444 -- and the scrutinee is a simple variable
1445 var_demanded_later (Var v) = isStrictDmd (idNewDemandInfo case_bndr)
1446 && not (isTickBoxOp v)
1447 -- ugly hack; covering this case is what
1448 -- exprOkForSpeculation was intended for.
1449 var_demanded_later _ = False
1451 rebuildCase env scrut case_bndr alts@[(_, bndrs, rhs)] cont
1452 | all isDeadBinder (case_bndr : bndrs) -- So this is just 'seq'
1453 = -- For this case, see Note [User-defined RULES for seq] in MkId
1454 do { let rhs' = substExpr env rhs
1455 out_args = [Type (substTy env (idType case_bndr)),
1456 Type (exprType rhs'), scrut, rhs']
1457 -- Lazily evaluated, so we don't do most of this
1459 ; rule_base <- getSimplRules
1460 ; let rules = getRules rule_base seqId
1461 ; mb_rule <- tryRules env seqId rules out_args cont
1463 Just (n_args, res) -> simplExprF (zapSubstEnv env)
1464 (mkApps res (drop n_args out_args))
1466 Nothing -> reallyRebuildCase env scrut case_bndr alts cont }
1468 rebuildCase env scrut case_bndr alts cont
1469 = reallyRebuildCase env scrut case_bndr alts cont
1471 --------------------------------------------------
1472 -- 3. Catch-all case
1473 --------------------------------------------------
1475 reallyRebuildCase env scrut case_bndr alts cont
1476 = do { -- Prepare the continuation;
1477 -- The new subst_env is in place
1478 (env', dup_cont, nodup_cont) <- prepareCaseCont env alts cont
1480 -- Simplify the alternatives
1481 ; (scrut', case_bndr', alts') <- simplAlts env' scrut case_bndr alts dup_cont
1483 -- Check for empty alternatives
1484 ; if null alts' then missingAlt env case_bndr alts cont
1486 { case_expr <- mkCase scrut' case_bndr' alts'
1488 -- Notice that rebuild gets the in-scope set from env, not alt_env
1489 -- The case binder *not* scope over the whole returned case-expression
1490 ; rebuild env' case_expr nodup_cont } }
1493 simplCaseBinder checks whether the scrutinee is a variable, v. If so,
1494 try to eliminate uses of v in the RHSs in favour of case_bndr; that
1495 way, there's a chance that v will now only be used once, and hence
1498 Historical note: we use to do the "case binder swap" in the Simplifier
1499 so there were additional complications if the scrutinee was a variable.
1500 Now the binder-swap stuff is done in the occurrence analyer; see
1501 OccurAnal Note [Binder swap].
1505 If the case binder is not dead, then neither are the pattern bound
1507 case <any> of x { (a,b) ->
1508 case x of { (p,q) -> p } }
1509 Here (a,b) both look dead, but come alive after the inner case is eliminated.
1510 The point is that we bring into the envt a binding
1512 after the outer case, and that makes (a,b) alive. At least we do unless
1513 the case binder is guaranteed dead.
1515 In practice, the scrutinee is almost always a variable, so we pretty
1516 much always zap the OccInfo of the binders. It doesn't matter much though.
1521 Consider case (v `cast` co) of x { I# y ->
1522 ... (case (v `cast` co) of {...}) ...
1523 We'd like to eliminate the inner case. We can get this neatly by
1524 arranging that inside the outer case we add the unfolding
1525 v |-> x `cast` (sym co)
1526 to v. Then we should inline v at the inner case, cancel the casts, and away we go
1528 Note [Improving seq]
1531 type family F :: * -> *
1532 type instance F Int = Int
1534 ... case e of x { DEFAULT -> rhs } ...
1536 where x::F Int. Then we'd like to rewrite (F Int) to Int, getting
1538 case e `cast` co of x'::Int
1539 I# x# -> let x = x' `cast` sym co
1542 so that 'rhs' can take advantage of the form of x'.
1544 Notice that Note [Case of cast] may then apply to the result.
1546 Nota Bene: We only do the [Improving seq] transformation if the
1547 case binder 'x' is actually used in the rhs; that is, if the case
1548 is *not* a *pure* seq.
1549 a) There is no point in adding the cast to a pure seq.
1550 b) There is a good reason not to: doing so would interfere
1551 with seq rules (Note [Built-in RULES for seq] in MkId).
1552 In particular, this [Improving seq] thing *adds* a cast
1553 while [Built-in RULES for seq] *removes* one, so they
1556 You might worry about
1557 case v of x { __DEFAULT ->
1558 ... case (v `cast` co) of y { I# -> ... }}
1559 This is a pure seq (since x is unused), so [Improving seq] won't happen.
1560 But it's ok: the simplifier will replace 'v' by 'x' in the rhs to get
1561 case v of x { __DEFAULT ->
1562 ... case (x `cast` co) of y { I# -> ... }}
1563 Now the outer case is not a pure seq, so [Improving seq] will happen,
1564 and then the inner case will disappear.
1566 The need for [Improving seq] showed up in Roman's experiments. Example:
1567 foo :: F Int -> Int -> Int
1568 foo t n = t `seq` bar n
1571 bar n = bar (n - case t of TI i -> i)
1572 Here we'd like to avoid repeated evaluating t inside the loop, by
1573 taking advantage of the `seq`.
1575 At one point I did transformation in LiberateCase, but it's more
1576 robust here. (Otherwise, there's a danger that we'll simply drop the
1577 'seq' altogether, before LiberateCase gets to see it.)
1581 improveSeq :: (FamInstEnv, FamInstEnv) -> SimplEnv
1582 -> OutExpr -> InId -> OutId -> [InAlt]
1583 -> SimplM (SimplEnv, OutExpr, OutId)
1584 -- Note [Improving seq]
1585 improveSeq fam_envs env scrut case_bndr case_bndr1 [(DEFAULT,_,_)]
1586 | not (isDeadBinder case_bndr) -- Not a pure seq! See the Note!
1587 , Just (co, ty2) <- topNormaliseType fam_envs (idType case_bndr1)
1588 = do { case_bndr2 <- newId (fsLit "nt") ty2
1589 ; let rhs = DoneEx (Var case_bndr2 `Cast` mkSymCoercion co)
1590 env2 = extendIdSubst env case_bndr rhs
1591 ; return (env2, scrut `Cast` co, case_bndr2) }
1593 improveSeq _ env scrut _ case_bndr1 _
1594 = return (env, scrut, case_bndr1)
1598 simplAlts does two things:
1600 1. Eliminate alternatives that cannot match, including the
1601 DEFAULT alternative.
1603 2. If the DEFAULT alternative can match only one possible constructor,
1604 then make that constructor explicit.
1606 case e of x { DEFAULT -> rhs }
1608 case e of x { (a,b) -> rhs }
1609 where the type is a single constructor type. This gives better code
1610 when rhs also scrutinises x or e.
1612 Here "cannot match" includes knowledge from GADTs
1614 It's a good idea do do this stuff before simplifying the alternatives, to
1615 avoid simplifying alternatives we know can't happen, and to come up with
1616 the list of constructors that are handled, to put into the IdInfo of the
1617 case binder, for use when simplifying the alternatives.
1619 Eliminating the default alternative in (1) isn't so obvious, but it can
1622 data Colour = Red | Green | Blue
1631 DEFAULT -> [ case y of ... ]
1633 If we inline h into f, the default case of the inlined h can't happen.
1634 If we don't notice this, we may end up filtering out *all* the cases
1635 of the inner case y, which give us nowhere to go!
1639 simplAlts :: SimplEnv
1641 -> InId -- Case binder
1642 -> [InAlt] -- Non-empty
1644 -> SimplM (OutExpr, OutId, [OutAlt]) -- Includes the continuation
1645 -- Like simplExpr, this just returns the simplified alternatives;
1646 -- it not return an environment
1648 simplAlts env scrut case_bndr alts cont'
1649 = -- pprTrace "simplAlts" (ppr alts $$ ppr (seIdSubst env)) $
1650 do { let env0 = zapFloats env
1652 ; (env1, case_bndr1) <- simplBinder env0 case_bndr
1654 ; fam_envs <- getFamEnvs
1655 ; (alt_env', scrut', case_bndr') <- improveSeq fam_envs env1 scrut
1656 case_bndr case_bndr1 alts
1658 ; (imposs_deflt_cons, in_alts) <- prepareAlts alt_env' scrut' case_bndr' alts
1660 ; alts' <- mapM (simplAlt alt_env' imposs_deflt_cons case_bndr' cont') in_alts
1661 ; return (scrut', case_bndr', alts') }
1663 ------------------------------------
1664 simplAlt :: SimplEnv
1665 -> [AltCon] -- These constructors can't be present when
1666 -- matching the DEFAULT alternative
1667 -> OutId -- The case binder
1672 simplAlt env imposs_deflt_cons case_bndr' cont' (DEFAULT, bndrs, rhs)
1673 = ASSERT( null bndrs )
1674 do { let env' = addBinderOtherCon env case_bndr' imposs_deflt_cons
1675 -- Record the constructors that the case-binder *can't* be.
1676 ; rhs' <- simplExprC env' rhs cont'
1677 ; return (DEFAULT, [], rhs') }
1679 simplAlt env _ case_bndr' cont' (LitAlt lit, bndrs, rhs)
1680 = ASSERT( null bndrs )
1681 do { let env' = addBinderUnfolding env case_bndr' (Lit lit)
1682 ; rhs' <- simplExprC env' rhs cont'
1683 ; return (LitAlt lit, [], rhs') }
1685 simplAlt env _ case_bndr' cont' (DataAlt con, vs, rhs)
1686 = do { -- Deal with the pattern-bound variables
1687 -- Mark the ones that are in ! positions in the
1688 -- data constructor as certainly-evaluated.
1689 -- NB: simplLamBinders preserves this eval info
1690 let vs_with_evals = add_evals (dataConRepStrictness con)
1691 ; (env', vs') <- simplLamBndrs env vs_with_evals
1693 -- Bind the case-binder to (con args)
1694 ; let inst_tys' = tyConAppArgs (idType case_bndr')
1695 con_args = map Type inst_tys' ++ varsToCoreExprs vs'
1696 env'' = addBinderUnfolding env' case_bndr'
1697 (mkConApp con con_args)
1699 ; rhs' <- simplExprC env'' rhs cont'
1700 ; return (DataAlt con, vs', rhs') }
1702 -- add_evals records the evaluated-ness of the bound variables of
1703 -- a case pattern. This is *important*. Consider
1704 -- data T = T !Int !Int
1706 -- case x of { T a b -> T (a+1) b }
1708 -- We really must record that b is already evaluated so that we don't
1709 -- go and re-evaluate it when constructing the result.
1710 -- See Note [Data-con worker strictness] in MkId.lhs
1715 go (v:vs') strs | isTyVar v = v : go vs' strs
1716 go (v:vs') (str:strs)
1717 | isMarkedStrict str = evald_v : go vs' strs
1718 | otherwise = zapped_v : go vs' strs
1720 zapped_v = zap_occ_info v
1721 evald_v = zapped_v `setIdUnfolding` evaldUnfolding
1722 go _ _ = pprPanic "cat_evals" (ppr con $$ ppr vs $$ ppr the_strs)
1724 -- See Note [zapOccInfo]
1725 -- zap_occ_info: if the case binder is alive, then we add the unfolding
1727 -- to the envt; so vs are now very much alive
1728 -- Note [Aug06] I can't see why this actually matters, but it's neater
1729 -- case e of t { (a,b) -> ...(case t of (p,q) -> p)... }
1730 -- ==> case e of t { (a,b) -> ...(a)... }
1731 -- Look, Ma, a is alive now.
1732 zap_occ_info = zapCasePatIdOcc case_bndr'
1734 addBinderUnfolding :: SimplEnv -> Id -> CoreExpr -> SimplEnv
1735 addBinderUnfolding env bndr rhs
1736 = modifyInScope env (bndr `setIdUnfolding` mkUnfolding False rhs)
1738 addBinderOtherCon :: SimplEnv -> Id -> [AltCon] -> SimplEnv
1739 addBinderOtherCon env bndr cons
1740 = modifyInScope env (bndr `setIdUnfolding` mkOtherCon cons)
1742 zapCasePatIdOcc :: Id -> Id -> Id
1743 -- Consider case e of b { (a,b) -> ... }
1744 -- Then if we bind b to (a,b) in "...", and b is not dead,
1745 -- then we must zap the deadness info on a,b
1746 zapCasePatIdOcc case_bndr
1747 | isDeadBinder case_bndr = \ pat_id -> pat_id
1748 | otherwise = \ pat_id -> zapIdOccInfo pat_id
1752 %************************************************************************
1754 \subsection{Known constructor}
1756 %************************************************************************
1758 We are a bit careful with occurrence info. Here's an example
1760 (\x* -> case x of (a*, b) -> f a) (h v, e)
1762 where the * means "occurs once". This effectively becomes
1763 case (h v, e) of (a*, b) -> f a)
1765 let a* = h v; b = e in f a
1769 All this should happen in one sweep.
1772 knownCon :: SimplEnv
1773 -> OutExpr -- The scrutinee
1774 -> DataCon -> [OutType] -> [OutExpr] -- The scrutinee (in pieces)
1775 -> InId -> [InBndr] -> InExpr -- The alternative
1777 -> SimplM (SimplEnv, OutExpr)
1779 knownCon env scrut dc dc_ty_args dc_args bndr bs rhs cont
1780 = do { env' <- bind_args env bs dc_args
1782 -- It's useful to bind bndr to scrut, rather than to a fresh
1783 -- binding x = Con arg1 .. argn
1784 -- because very often the scrut is a variable, so we avoid
1785 -- creating, and then subsequently eliminating, a let-binding
1786 -- BUT, if scrut is a not a variable, we must be careful
1787 -- about duplicating the arg redexes; in that case, make
1788 -- a new con-app from the args
1789 bndr_rhs | exprIsTrivial scrut = scrut
1790 | otherwise = con_app
1791 con_app = Var (dataConWorkId dc)
1792 `mkTyApps` dc_ty_args
1793 `mkApps` [substExpr env' (varToCoreExpr b) | b <- bs]
1794 -- dc_ty_args are aready OutTypes, but bs are InBndrs
1796 ; env'' <- simplNonRecX env' bndr bndr_rhs
1797 ; simplExprF env'' rhs cont }
1799 zap_occ = zapCasePatIdOcc bndr -- bndr is an InId
1802 bind_args env' [] _ = return env'
1804 bind_args env' (b:bs') (Type ty : args)
1805 = ASSERT( isTyVar b )
1806 bind_args (extendTvSubst env' b ty) bs' args
1808 bind_args env' (b:bs') (arg : args)
1810 do { let b' = zap_occ b
1811 -- Note that the binder might be "dead", because it doesn't
1812 -- occur in the RHS; and simplNonRecX may therefore discard
1813 -- it via postInlineUnconditionally.
1814 -- Nevertheless we must keep it if the case-binder is alive,
1815 -- because it may be used in the con_app. See Note [zapOccInfo]
1816 ; env'' <- simplNonRecX env' b' arg
1817 ; bind_args env'' bs' args }
1820 pprPanic "bind_args" $ ppr dc $$ ppr bs $$ ppr dc_args $$
1821 text "scrut:" <+> ppr scrut
1824 missingAlt :: SimplEnv -> Id -> [InAlt] -> SimplCont -> SimplM (SimplEnv, OutExpr)
1825 -- This isn't strictly an error, although it is unusual.
1826 -- It's possible that the simplifer might "see" that
1827 -- an inner case has no accessible alternatives before
1828 -- it "sees" that the entire branch of an outer case is
1829 -- inaccessible. So we simply put an error case here instead.
1830 missingAlt env case_bndr alts cont
1831 = WARN( True, ptext (sLit "missingAlt") <+> ppr case_bndr )
1832 return (env, mkImpossibleExpr res_ty)
1834 res_ty = contResultType env (substTy env (coreAltsType alts)) cont
1838 %************************************************************************
1840 \subsection{Duplicating continuations}
1842 %************************************************************************
1845 prepareCaseCont :: SimplEnv
1846 -> [InAlt] -> SimplCont
1847 -> SimplM (SimplEnv, SimplCont,SimplCont)
1848 -- Return a duplicatable continuation, a non-duplicable part
1849 -- plus some extra bindings (that scope over the entire
1852 -- No need to make it duplicatable if there's only one alternative
1853 prepareCaseCont env [_] cont = return (env, cont, mkBoringStop)
1854 prepareCaseCont env _ cont = mkDupableCont env cont
1858 mkDupableCont :: SimplEnv -> SimplCont
1859 -> SimplM (SimplEnv, SimplCont, SimplCont)
1861 mkDupableCont env cont
1862 | contIsDupable cont
1863 = return (env, cont, mkBoringStop)
1865 mkDupableCont _ (Stop {}) = panic "mkDupableCont" -- Handled by previous eqn
1867 mkDupableCont env (CoerceIt ty cont)
1868 = do { (env', dup, nodup) <- mkDupableCont env cont
1869 ; return (env', CoerceIt ty dup, nodup) }
1871 mkDupableCont env cont@(StrictBind {})
1872 = return (env, mkBoringStop, cont)
1873 -- See Note [Duplicating StrictBind]
1875 mkDupableCont env (StrictArg fun cci ai cont)
1876 -- See Note [Duplicating StrictArg]
1877 = do { (env', dup, nodup) <- mkDupableCont env cont
1878 ; (env'', fun') <- mk_dupable_call env' fun
1879 ; return (env'', StrictArg fun' cci ai dup, nodup) }
1881 mk_dupable_call env (Var v) = return (env, Var v)
1882 mk_dupable_call env (App fun arg) = do { (env', fun') <- mk_dupable_call env fun
1883 ; (env'', arg') <- makeTrivial env' arg
1884 ; return (env'', fun' `App` arg') }
1885 mk_dupable_call _ other = pprPanic "mk_dupable_call" (ppr other)
1886 -- The invariant of StrictArg is that the first arg is always an App chain
1888 mkDupableCont env (ApplyTo _ arg se cont)
1889 = -- e.g. [...hole...] (...arg...)
1891 -- let a = ...arg...
1892 -- in [...hole...] a
1893 do { (env', dup_cont, nodup_cont) <- mkDupableCont env cont
1894 ; arg' <- simplExpr (se `setInScope` env') arg
1895 ; (env'', arg'') <- makeTrivial env' arg'
1896 ; let app_cont = ApplyTo OkToDup arg'' (zapSubstEnv env'') dup_cont
1897 ; return (env'', app_cont, nodup_cont) }
1899 mkDupableCont env cont@(Select _ case_bndr [(_, bs, _rhs)] _ _)
1900 -- See Note [Single-alternative case]
1901 -- | not (exprIsDupable rhs && contIsDupable case_cont)
1902 -- | not (isDeadBinder case_bndr)
1903 | all isDeadBinder bs -- InIds
1904 && not (isUnLiftedType (idType case_bndr))
1905 -- Note [Single-alternative-unlifted]
1906 = return (env, mkBoringStop, cont)
1908 mkDupableCont env (Select _ case_bndr alts se cont)
1909 = -- e.g. (case [...hole...] of { pi -> ei })
1911 -- let ji = \xij -> ei
1912 -- in case [...hole...] of { pi -> ji xij }
1913 do { tick (CaseOfCase case_bndr)
1914 ; (env', dup_cont, nodup_cont) <- mkDupableCont env cont
1915 -- NB: call mkDupableCont here, *not* prepareCaseCont
1916 -- We must make a duplicable continuation, whereas prepareCaseCont
1917 -- doesn't when there is a single case branch
1919 ; let alt_env = se `setInScope` env'
1920 ; (alt_env', case_bndr') <- simplBinder alt_env case_bndr
1921 ; alts' <- mapM (simplAlt alt_env' [] case_bndr' dup_cont) alts
1922 -- Safe to say that there are no handled-cons for the DEFAULT case
1923 -- NB: simplBinder does not zap deadness occ-info, so
1924 -- a dead case_bndr' will still advertise its deadness
1925 -- This is really important because in
1926 -- case e of b { (# p,q #) -> ... }
1927 -- b is always dead, and indeed we are not allowed to bind b to (# p,q #),
1928 -- which might happen if e was an explicit unboxed pair and b wasn't marked dead.
1929 -- In the new alts we build, we have the new case binder, so it must retain
1931 -- NB: we don't use alt_env further; it has the substEnv for
1932 -- the alternatives, and we don't want that
1934 ; (env'', alts'') <- mkDupableAlts env' case_bndr' alts'
1935 ; return (env'', -- Note [Duplicated env]
1936 Select OkToDup case_bndr' alts'' (zapSubstEnv env'') mkBoringStop,
1940 mkDupableAlts :: SimplEnv -> OutId -> [InAlt]
1941 -> SimplM (SimplEnv, [InAlt])
1942 -- Absorbs the continuation into the new alternatives
1944 mkDupableAlts env case_bndr' the_alts
1947 go env0 [] = return (env0, [])
1949 = do { (env1, alt') <- mkDupableAlt env0 case_bndr' alt
1950 ; (env2, alts') <- go env1 alts
1951 ; return (env2, alt' : alts' ) }
1953 mkDupableAlt :: SimplEnv -> OutId -> (AltCon, [CoreBndr], CoreExpr)
1954 -> SimplM (SimplEnv, (AltCon, [CoreBndr], CoreExpr))
1955 mkDupableAlt env case_bndr (con, bndrs', rhs')
1956 | exprIsDupable rhs' -- Note [Small alternative rhs]
1957 = return (env, (con, bndrs', rhs'))
1959 = do { let rhs_ty' = exprType rhs'
1960 scrut_ty = idType case_bndr
1963 DEFAULT -> case_bndr
1964 DataAlt dc -> setIdUnfolding case_bndr unf
1966 -- See Note [Case binders and join points]
1967 unf = mkInlineRule InlSat rhs 0
1968 rhs = mkConApp dc (map Type (tyConAppArgs scrut_ty)
1969 ++ varsToCoreExprs bndrs')
1971 LitAlt {} -> WARN( True, ptext (sLit "mkDupableAlt")
1972 <+> ppr case_bndr <+> ppr con )
1974 -- The case binder is alive but trivial, so why has
1975 -- it not been substituted away?
1977 used_bndrs' | isDeadBinder case_bndr = filter abstract_over bndrs'
1978 | otherwise = bndrs' ++ [case_bndr_w_unf]
1981 | isTyVar bndr = True -- Abstract over all type variables just in case
1982 | otherwise = not (isDeadBinder bndr)
1983 -- The deadness info on the new Ids is preserved by simplBinders
1985 ; (final_bndrs', final_args) -- Note [Join point abstraction]
1986 <- if (any isId used_bndrs')
1987 then return (used_bndrs', varsToCoreExprs used_bndrs')
1988 else do { rw_id <- newId (fsLit "w") realWorldStatePrimTy
1989 ; return ([rw_id], [Var realWorldPrimId]) }
1991 ; join_bndr <- newId (fsLit "$j") (mkPiTypes final_bndrs' rhs_ty')
1992 -- Note [Funky mkPiTypes]
1994 ; let -- We make the lambdas into one-shot-lambdas. The
1995 -- join point is sure to be applied at most once, and doing so
1996 -- prevents the body of the join point being floated out by
1997 -- the full laziness pass
1998 really_final_bndrs = map one_shot final_bndrs'
1999 one_shot v | isId v = setOneShotLambda v
2001 join_rhs = mkLams really_final_bndrs rhs'
2002 join_call = mkApps (Var join_bndr) final_args
2004 ; env' <- addPolyBind NotTopLevel env (NonRec join_bndr join_rhs)
2005 ; return (env', (con, bndrs', join_call)) }
2006 -- See Note [Duplicated env]
2009 Note [Case binders and join points]
2010 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
2012 case (case .. ) of c {
2015 If we make a join point with c but not c# we get
2016 $j = \c -> ....c....
2018 But if later inlining scrutines the c, thus
2020 $j = \c -> ... case c of { I# y -> ... } ...
2022 we won't see that 'c' has already been scrutinised. This actually
2023 happens in the 'tabulate' function in wave4main, and makes a significant
2024 difference to allocation.
2026 An alternative plan is this:
2028 $j = \c# -> let c = I# c# in ...c....
2030 but that is bad if 'c' is *not* later scrutinised.
2032 So instead we do both: we pass 'c' and 'c#' , and record in c's inlining
2033 that it's really I# c#, thus
2035 $j = \c# -> \c[=I# c#] -> ...c....
2037 Absence analysis may later discard 'c'.
2040 Note [Duplicated env]
2041 ~~~~~~~~~~~~~~~~~~~~~
2042 Some of the alternatives are simplified, but have not been turned into a join point
2043 So they *must* have an zapped subst-env. So we can't use completeNonRecX to
2044 bind the join point, because it might to do PostInlineUnconditionally, and
2045 we'd lose that when zapping the subst-env. We could have a per-alt subst-env,
2046 but zapping it (as we do in mkDupableCont, the Select case) is safe, and
2047 at worst delays the join-point inlining.
2049 Note [Small alternative rhs]
2050 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~
2051 It is worth checking for a small RHS because otherwise we
2052 get extra let bindings that may cause an extra iteration of the simplifier to
2053 inline back in place. Quite often the rhs is just a variable or constructor.
2054 The Ord instance of Maybe in PrelMaybe.lhs, for example, took several extra
2055 iterations because the version with the let bindings looked big, and so wasn't
2056 inlined, but after the join points had been inlined it looked smaller, and so
2059 NB: we have to check the size of rhs', not rhs.
2060 Duplicating a small InAlt might invalidate occurrence information
2061 However, if it *is* dupable, we return the *un* simplified alternative,
2062 because otherwise we'd need to pair it up with an empty subst-env....
2063 but we only have one env shared between all the alts.
2064 (Remember we must zap the subst-env before re-simplifying something).
2065 Rather than do this we simply agree to re-simplify the original (small) thing later.
2067 Note [Funky mkPiTypes]
2068 ~~~~~~~~~~~~~~~~~~~~~~
2069 Notice the funky mkPiTypes. If the contructor has existentials
2070 it's possible that the join point will be abstracted over
2071 type varaibles as well as term variables.
2072 Example: Suppose we have
2073 data T = forall t. C [t]
2075 case (case e of ...) of
2077 We get the join point
2078 let j :: forall t. [t] -> ...
2079 j = /\t \xs::[t] -> rhs
2081 case (case e of ...) of
2082 C t xs::[t] -> j t xs
2084 Note [Join point abstaction]
2085 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~
2086 If we try to lift a primitive-typed something out
2087 for let-binding-purposes, we will *caseify* it (!),
2088 with potentially-disastrous strictness results. So
2089 instead we turn it into a function: \v -> e
2090 where v::State# RealWorld#. The value passed to this function
2091 is realworld#, which generates (almost) no code.
2093 There's a slight infelicity here: we pass the overall
2094 case_bndr to all the join points if it's used in *any* RHS,
2095 because we don't know its usage in each RHS separately
2097 We used to say "&& isUnLiftedType rhs_ty'" here, but now
2098 we make the join point into a function whenever used_bndrs'
2099 is empty. This makes the join-point more CPR friendly.
2100 Consider: let j = if .. then I# 3 else I# 4
2101 in case .. of { A -> j; B -> j; C -> ... }
2103 Now CPR doesn't w/w j because it's a thunk, so
2104 that means that the enclosing function can't w/w either,
2105 which is a lose. Here's the example that happened in practice:
2106 kgmod :: Int -> Int -> Int
2107 kgmod x y = if x > 0 && y < 0 || x < 0 && y > 0
2111 I have seen a case alternative like this:
2113 It's a bit silly to add the realWorld dummy arg in this case, making
2116 (the \v alone is enough to make CPR happy) but I think it's rare
2118 Note [Duplicating StrictArg]
2119 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~
2120 The original plan had (where E is a big argument)
2122 ==> let $j = \a -> f E a
2125 But this is terrible! Here's an example:
2126 && E (case x of { T -> F; F -> T })
2127 Now, && is strict so we end up simplifying the case with
2128 an ArgOf continuation. If we let-bind it, we get
2129 let $j = \v -> && E v
2130 in simplExpr (case x of { T -> F; F -> T })
2132 And after simplifying more we get
2133 let $j = \v -> && E v
2134 in case x of { T -> $j F; F -> $j T }
2135 Which is a Very Bad Thing
2137 What we do now is this
2141 Now if the thing in the hole is a case expression (which is when
2142 we'll call mkDupableCont), we'll push the function call into the
2143 branches, which is what we want. Now RULES for f may fire, and
2144 call-pattern specialisation. Here's an example from Trac #3116
2147 _ -> Chunk p fpc (o+1) (l-1) bs')
2148 If we can push the call for 'go' inside the case, we get
2149 call-pattern specialisation for 'go', which is *crucial* for
2152 Here is the (&&) example:
2153 && E (case x of { T -> F; F -> T })
2155 case x of { T -> && a F; F -> && a T }
2159 * Arguments to f *after* the strict one are handled by
2160 the ApplyTo case of mkDupableCont. Eg
2163 * We can only do the let-binding of E because the function
2164 part of a StrictArg continuation is an explicit syntax
2165 tree. In earlier versions we represented it as a function
2166 (CoreExpr -> CoreEpxr) which we couldn't take apart.
2168 Do *not* duplicate StrictBind and StritArg continuations. We gain
2169 nothing by propagating them into the expressions, and we do lose a
2172 The desire not to duplicate is the entire reason that
2173 mkDupableCont returns a pair of continuations.
2175 Note [Duplicating StrictBind]
2176 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
2177 Unlike StrictArg, there doesn't seem anything to gain from
2178 duplicating a StrictBind continuation, so we don't.
2180 The desire not to duplicate is the entire reason that
2181 mkDupableCont returns a pair of continuations.
2184 Note [Single-alternative cases]
2185 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
2186 This case is just like the ArgOf case. Here's an example:
2190 case (case x of I# x' ->
2192 True -> I# (negate# x')
2193 False -> I# x') of y {
2195 Because the (case x) has only one alternative, we'll transform to
2197 case (case x' <# 0# of
2198 True -> I# (negate# x')
2199 False -> I# x') of y {
2201 But now we do *NOT* want to make a join point etc, giving
2203 let $j = \y -> MkT y
2205 True -> $j (I# (negate# x'))
2207 In this case the $j will inline again, but suppose there was a big
2208 strict computation enclosing the orginal call to MkT. Then, it won't
2209 "see" the MkT any more, because it's big and won't get duplicated.
2210 And, what is worse, nothing was gained by the case-of-case transform.
2212 When should use this case of mkDupableCont?
2213 However, matching on *any* single-alternative case is a *disaster*;
2214 e.g. case (case ....) of (a,b) -> (# a,b #)
2215 We must push the outer case into the inner one!
2218 * Match [(DEFAULT,_,_)], but in the common case of Int,
2219 the alternative-filling-in code turned the outer case into
2220 case (...) of y { I# _ -> MkT y }
2222 * Match on single alternative plus (not (isDeadBinder case_bndr))
2223 Rationale: pushing the case inwards won't eliminate the construction.
2224 But there's a risk of
2225 case (...) of y { (a,b) -> let z=(a,b) in ... }
2226 Now y looks dead, but it'll come alive again. Still, this
2227 seems like the best option at the moment.
2229 * Match on single alternative plus (all (isDeadBinder bndrs))
2230 Rationale: this is essentially seq.
2232 * Match when the rhs is *not* duplicable, and hence would lead to a
2233 join point. This catches the disaster-case above. We can test
2234 the *un-simplified* rhs, which is fine. It might get bigger or
2235 smaller after simplification; if it gets smaller, this case might
2236 fire next time round. NB also that we must test contIsDupable
2237 case_cont *btoo, because case_cont might be big!
2239 HOWEVER: I found that this version doesn't work well, because
2240 we can get let x = case (...) of { small } in ...case x...
2241 When x is inlined into its full context, we find that it was a bad
2242 idea to have pushed the outer case inside the (...) case.
2244 Note [Single-alternative-unlifted]
2245 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
2246 Here's another single-alternative where we really want to do case-of-case:
2254 case y_s6X of tpl_s7m {
2255 M1.Mk1 ipv_s70 -> ipv_s70;
2256 M1.Mk2 ipv_s72 -> ipv_s72;
2262 case x_s74 of tpl_s7n {
2263 M1.Mk1 ipv_s77 -> ipv_s77;
2264 M1.Mk2 ipv_s79 -> ipv_s79;
2268 { __DEFAULT -> ==# [wild1_s7b wild_s7c];
2272 So the outer case is doing *nothing at all*, other than serving as a
2273 join-point. In this case we really want to do case-of-case and decide
2274 whether to use a real join point or just duplicate the continuation.
2276 Hence: check whether the case binder's type is unlifted, because then
2277 the outer case is *not* a seq.