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, substTyVar )
16 import FamInstEnv ( FamInstEnv )
18 import MkId ( seqId, realWorldPrimId )
19 import MkCore ( mkImpossibleExpr )
22 import Name ( mkSystemVarName, isExternalName )
24 import OptCoercion ( optCoercion )
25 import FamInstEnv ( topNormaliseType )
26 import DataCon ( DataCon, dataConWorkId, dataConRepStrictness )
27 import CoreMonad ( Tick(..), SimplifierMode(..) )
29 import Demand ( isStrictDmd )
30 import PprCore ( pprParendExpr, pprCoreExpr )
31 import CoreUnfold ( mkUnfolding, mkCoreUnfolding
32 , mkInlineUnfolding, mkSimpleUnfolding
33 , exprIsConApp_maybe, callSiteInline, CallCtxt(..) )
35 import qualified CoreSubst
37 import Rules ( lookupRule, getRules )
38 import BasicTypes ( isMarkedStrict, Arity )
39 import CostCentre ( currentCCS, pushCCisNop )
40 import TysPrim ( realWorldStatePrimTy )
41 import BasicTypes ( TopLevelFlag(..), isTopLevel, RecFlag(..) )
42 import MonadUtils ( foldlM, mapAccumLM )
43 import Maybes ( orElse )
44 import Data.List ( mapAccumL )
50 The guts of the simplifier is in this module, but the driver loop for
51 the simplifier is in SimplCore.lhs.
54 -----------------------------------------
55 *** IMPORTANT NOTE ***
56 -----------------------------------------
57 The simplifier used to guarantee that the output had no shadowing, but
58 it does not do so any more. (Actually, it never did!) The reason is
59 documented with simplifyArgs.
62 -----------------------------------------
63 *** IMPORTANT NOTE ***
64 -----------------------------------------
65 Many parts of the simplifier return a bunch of "floats" as well as an
66 expression. This is wrapped as a datatype SimplUtils.FloatsWith.
68 All "floats" are let-binds, not case-binds, but some non-rec lets may
69 be unlifted (with RHS ok-for-speculation).
73 -----------------------------------------
74 ORGANISATION OF FUNCTIONS
75 -----------------------------------------
77 - simplify all top-level binders
78 - for NonRec, call simplRecOrTopPair
79 - for Rec, call simplRecBind
82 ------------------------------
83 simplExpr (applied lambda) ==> simplNonRecBind
84 simplExpr (Let (NonRec ...) ..) ==> simplNonRecBind
85 simplExpr (Let (Rec ...) ..) ==> simplify binders; simplRecBind
87 ------------------------------
88 simplRecBind [binders already simplfied]
89 - use simplRecOrTopPair on each pair in turn
91 simplRecOrTopPair [binder already simplified]
92 Used for: recursive bindings (top level and nested)
93 top-level non-recursive bindings
95 - check for PreInlineUnconditionally
99 Used for: non-top-level non-recursive bindings
100 beta reductions (which amount to the same thing)
101 Because it can deal with strict arts, it takes a
102 "thing-inside" and returns an expression
104 - check for PreInlineUnconditionally
105 - simplify binder, including its IdInfo
114 simplNonRecX: [given a *simplified* RHS, but an *unsimplified* binder]
115 Used for: binding case-binder and constr args in a known-constructor case
116 - check for PreInLineUnconditionally
120 ------------------------------
121 simplLazyBind: [binder already simplified, RHS not]
122 Used for: recursive bindings (top level and nested)
123 top-level non-recursive bindings
124 non-top-level, but *lazy* non-recursive bindings
125 [must not be strict or unboxed]
126 Returns floats + an augmented environment, not an expression
127 - substituteIdInfo and add result to in-scope
128 [so that rules are available in rec rhs]
131 - float if exposes constructor or PAP
135 completeNonRecX: [binder and rhs both simplified]
136 - if the the thing needs case binding (unlifted and not ok-for-spec)
142 completeBind: [given a simplified RHS]
143 [used for both rec and non-rec bindings, top level and not]
144 - try PostInlineUnconditionally
145 - add unfolding [this is the only place we add an unfolding]
150 Right hand sides and arguments
151 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
152 In many ways we want to treat
153 (a) the right hand side of a let(rec), and
154 (b) a function argument
155 in the same way. But not always! In particular, we would
156 like to leave these arguments exactly as they are, so they
157 will match a RULE more easily.
162 It's harder to make the rule match if we ANF-ise the constructor,
163 or eta-expand the PAP:
165 f (let { a = g x; b = h x } in (a,b))
168 On the other hand if we see the let-defns
173 then we *do* want to ANF-ise and eta-expand, so that p and q
174 can be safely inlined.
176 Even floating lets out is a bit dubious. For let RHS's we float lets
177 out if that exposes a value, so that the value can be inlined more vigorously.
180 r = let x = e in (x,x)
182 Here, if we float the let out we'll expose a nice constructor. We did experiments
183 that showed this to be a generally good thing. But it was a bad thing to float
184 lets out unconditionally, because that meant they got allocated more often.
186 For function arguments, there's less reason to expose a constructor (it won't
187 get inlined). Just possibly it might make a rule match, but I'm pretty skeptical.
188 So for the moment we don't float lets out of function arguments either.
193 For eta expansion, we want to catch things like
195 case e of (a,b) -> \x -> case a of (p,q) -> \y -> r
197 If the \x was on the RHS of a let, we'd eta expand to bring the two
198 lambdas together. And in general that's a good thing to do. Perhaps
199 we should eta expand wherever we find a (value) lambda? Then the eta
200 expansion at a let RHS can concentrate solely on the PAP case.
203 %************************************************************************
205 \subsection{Bindings}
207 %************************************************************************
210 simplTopBinds :: SimplEnv -> [InBind] -> SimplM SimplEnv
212 simplTopBinds env0 binds0
213 = do { -- Put all the top-level binders into scope at the start
214 -- so that if a transformation rule has unexpectedly brought
215 -- anything into scope, then we don't get a complaint about that.
216 -- It's rather as if the top-level binders were imported.
217 ; env1 <- simplRecBndrs env0 (bindersOfBinds binds0)
218 ; dflags <- getDOptsSmpl
219 ; let dump_flag = dopt Opt_D_verbose_core2core dflags
220 ; env2 <- simpl_binds dump_flag env1 binds0
221 ; freeTick SimplifierDone
224 -- We need to track the zapped top-level binders, because
225 -- they should have their fragile IdInfo zapped (notably occurrence info)
226 -- That's why we run down binds and bndrs' simultaneously.
228 -- The dump-flag emits a trace for each top-level binding, which
229 -- helps to locate the tracing for inlining and rule firing
230 simpl_binds :: Bool -> SimplEnv -> [InBind] -> SimplM SimplEnv
231 simpl_binds _ env [] = return env
232 simpl_binds dump env (bind:binds) = do { env' <- trace_bind dump bind $
234 ; simpl_binds dump env' binds }
236 trace_bind True bind = pprTrace "SimplBind" (ppr (bindersOf bind))
237 trace_bind False _ = \x -> x
239 simpl_bind env (Rec pairs) = simplRecBind env TopLevel pairs
240 simpl_bind env (NonRec b r) = simplRecOrTopPair env' TopLevel NonRecursive b b' r
242 (env', b') = addBndrRules env b (lookupRecBndr env b)
246 %************************************************************************
248 \subsection{Lazy bindings}
250 %************************************************************************
252 simplRecBind is used for
253 * recursive bindings only
256 simplRecBind :: SimplEnv -> TopLevelFlag
259 simplRecBind env0 top_lvl pairs0
260 = do { let (env_with_info, triples) = mapAccumL add_rules env0 pairs0
261 ; env1 <- go (zapFloats env_with_info) triples
262 ; return (env0 `addRecFloats` env1) }
263 -- addFloats adds the floats from env1,
264 -- _and_ updates env0 with the in-scope set from env1
266 add_rules :: SimplEnv -> (InBndr,InExpr) -> (SimplEnv, (InBndr, OutBndr, InExpr))
267 -- Add the (substituted) rules to the binder
268 add_rules env (bndr, rhs) = (env', (bndr, bndr', rhs))
270 (env', bndr') = addBndrRules env bndr (lookupRecBndr env bndr)
272 go env [] = return env
274 go env ((old_bndr, new_bndr, rhs) : pairs)
275 = do { env' <- simplRecOrTopPair env top_lvl Recursive old_bndr new_bndr rhs
279 simplOrTopPair is used for
280 * recursive bindings (whether top level or not)
281 * top-level non-recursive bindings
283 It assumes the binder has already been simplified, but not its IdInfo.
286 simplRecOrTopPair :: SimplEnv
287 -> TopLevelFlag -> RecFlag
288 -> InId -> OutBndr -> InExpr -- Binder and rhs
289 -> SimplM SimplEnv -- Returns an env that includes the binding
291 simplRecOrTopPair env top_lvl is_rec old_bndr new_bndr rhs
292 | preInlineUnconditionally env top_lvl old_bndr rhs -- Check for unconditional inline
293 = do { tick (PreInlineUnconditionally old_bndr)
294 ; return (extendIdSubst env old_bndr (mkContEx env rhs)) }
297 = simplLazyBind env top_lvl is_rec old_bndr new_bndr rhs env
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 = -- pprTrace "simplLazyBind" ((ppr bndr <+> ppr bndr1) $$ ppr rhs $$ ppr (seIdSubst rhs_se)) $
325 do { let rhs_env = rhs_se `setInScope` env
326 (tvs, body) = case collectTyBinders rhs of
327 (tvs, body) | not_lam body -> (tvs,body)
328 | otherwise -> ([], rhs)
329 not_lam (Lam _ _) = False
331 -- Do not do the "abstract tyyvar" thing if there's
332 -- a lambda inside, becuase it defeats eta-reduction
333 -- f = /\a. \x. g a x
336 ; (body_env, tvs') <- simplBinders rhs_env tvs
337 -- See Note [Floating and type abstraction] in SimplUtils
340 ; (body_env1, body1) <- simplExprF body_env body mkRhsStop
341 -- ANF-ise a constructor or PAP rhs
342 ; (body_env2, body2) <- prepareRhs top_lvl body_env1 bndr1 body1
345 <- if not (doFloatFromRhs top_lvl is_rec False body2 body_env2)
346 then -- No floating, revert to body1
347 do { rhs' <- mkLam env tvs' (wrapFloats body_env1 body1)
348 ; return (env, rhs') }
350 else if null tvs then -- Simple floating
351 do { tick LetFloatFromLet
352 ; return (addFloats env body_env2, body2) }
354 else -- Do type-abstraction first
355 do { tick LetFloatFromLet
356 ; (poly_binds, body3) <- abstractFloats tvs' body_env2 body2
357 ; rhs' <- mkLam env tvs' body3
358 ; env' <- foldlM (addPolyBind top_lvl) env poly_binds
359 ; return (env', rhs') }
361 ; completeBind env' top_lvl bndr bndr1 rhs' }
364 A specialised variant of simplNonRec used when the RHS is already simplified,
365 notably in knownCon. It uses case-binding where necessary.
368 simplNonRecX :: SimplEnv
369 -> InId -- Old binder
370 -> OutExpr -- Simplified RHS
373 simplNonRecX env bndr new_rhs
374 | isDeadBinder bndr -- Not uncommon; e.g. case (a,b) of b { (p,q) -> p }
375 = return env -- Here b is dead, and we avoid creating
376 | otherwise -- the binding b = (a,b)
377 = do { (env', bndr') <- simplBinder env bndr
378 ; completeNonRecX NotTopLevel env' (isStrictId bndr) bndr bndr' new_rhs }
379 -- simplNonRecX is only used for NotTopLevel things
381 completeNonRecX :: TopLevelFlag -> SimplEnv
383 -> InId -- Old binder
384 -> OutId -- New binder
385 -> OutExpr -- Simplified RHS
388 completeNonRecX top_lvl env is_strict old_bndr new_bndr new_rhs
389 = do { (env1, rhs1) <- prepareRhs top_lvl (zapFloats env) new_bndr new_rhs
391 if doFloatFromRhs NotTopLevel NonRecursive is_strict rhs1 env1
392 then do { tick LetFloatFromLet
393 ; return (addFloats env env1, rhs1) } -- Add the floats to the main env
394 else return (env, wrapFloats env1 rhs1) -- Wrap the floats around the RHS
395 ; completeBind env2 NotTopLevel old_bndr new_bndr rhs2 }
398 {- No, no, no! Do not try preInlineUnconditionally in completeNonRecX
399 Doing so risks exponential behaviour, because new_rhs has been simplified once already
400 In the cases described by the folowing commment, postInlineUnconditionally will
401 catch many of the relevant cases.
402 -- This happens; for example, the case_bndr during case of
403 -- known constructor: case (a,b) of x { (p,q) -> ... }
404 -- Here x isn't mentioned in the RHS, so we don't want to
405 -- create the (dead) let-binding let x = (a,b) in ...
407 -- Similarly, single occurrences can be inlined vigourously
408 -- e.g. case (f x, g y) of (a,b) -> ....
409 -- If a,b occur once we can avoid constructing the let binding for them.
411 Furthermore in the case-binding case preInlineUnconditionally risks extra thunks
412 -- Consider case I# (quotInt# x y) of
413 -- I# v -> let w = J# v in ...
414 -- If we gaily inline (quotInt# x y) for v, we end up building an
416 -- let w = J# (quotInt# x y) in ...
417 -- because quotInt# can fail.
419 | preInlineUnconditionally env NotTopLevel bndr new_rhs
420 = thing_inside (extendIdSubst env bndr (DoneEx new_rhs))
423 ----------------------------------
424 prepareRhs takes a putative RHS, checks whether it's a PAP or
425 constructor application and, if so, converts it to ANF, so that the
426 resulting thing can be inlined more easily. Thus
433 We also want to deal well cases like this
434 v = (f e1 `cast` co) e2
435 Here we want to make e1,e2 trivial and get
436 x1 = e1; x2 = e2; v = (f x1 `cast` co) v2
437 That's what the 'go' loop in prepareRhs does
440 prepareRhs :: TopLevelFlag -> SimplEnv -> OutId -> OutExpr -> SimplM (SimplEnv, OutExpr)
441 -- Adds new floats to the env iff that allows us to return a good RHS
442 prepareRhs top_lvl env id (Cast rhs co) -- Note [Float coercions]
443 | (ty1, _ty2) <- coercionKind co -- Do *not* do this if rhs has an unlifted type
444 , not (isUnLiftedType ty1) -- see Note [Float coercions (unlifted)]
445 = do { (env', rhs') <- makeTrivialWithInfo top_lvl env sanitised_info rhs
446 ; return (env', Cast rhs' co) }
448 sanitised_info = vanillaIdInfo `setStrictnessInfo` strictnessInfo info
449 `setDemandInfo` demandInfo info
452 prepareRhs top_lvl env0 _ rhs0
453 = do { (_is_exp, env1, rhs1) <- go 0 env0 rhs0
454 ; return (env1, rhs1) }
456 go n_val_args env (Cast rhs co)
457 = do { (is_exp, env', rhs') <- go n_val_args env rhs
458 ; return (is_exp, env', Cast rhs' co) }
459 go n_val_args env (App fun (Type ty))
460 = do { (is_exp, env', rhs') <- go n_val_args env fun
461 ; return (is_exp, env', App rhs' (Type ty)) }
462 go n_val_args env (App fun arg)
463 = do { (is_exp, env', fun') <- go (n_val_args+1) env fun
465 True -> do { (env'', arg') <- makeTrivial top_lvl env' arg
466 ; return (True, env'', App fun' arg') }
467 False -> return (False, env, App fun arg) }
468 go n_val_args env (Var fun)
469 = return (is_exp, env, Var fun)
471 is_exp = isExpandableApp fun n_val_args -- The fun a constructor or PAP
472 -- See Note [CONLIKE pragma] in BasicTypes
473 -- The definition of is_exp should match that in
474 -- OccurAnal.occAnalApp
477 = return (False, env, other)
481 Note [Float coercions]
482 ~~~~~~~~~~~~~~~~~~~~~~
483 When we find the binding
485 we'd like to transform it to
487 x = x `cast` co -- A trivial binding
488 There's a chance that e will be a constructor application or function, or something
489 like that, so moving the coerion to the usage site may well cancel the coersions
490 and lead to further optimisation. Example:
493 data instance T Int = T Int
495 foo :: Int -> Int -> Int
500 go n = case x of { T m -> go (n-m) }
501 -- This case should optimise
503 Note [Preserve strictness when floating coercions]
504 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
505 In the Note [Float coercions] transformation, keep the strictness info.
507 f = e `cast` co -- f has strictness SSL
509 f' = e -- f' also has strictness SSL
510 f = f' `cast` co -- f still has strictness SSL
512 Its not wrong to drop it on the floor, but better to keep it.
514 Note [Float coercions (unlifted)]
515 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
516 BUT don't do [Float coercions] if 'e' has an unlifted type.
519 foo :: Int = (error (# Int,Int #) "urk")
520 `cast` CoUnsafe (# Int,Int #) Int
522 If do the makeTrivial thing to the error call, we'll get
523 foo = case error (# Int,Int #) "urk" of v -> v `cast` ...
524 But 'v' isn't in scope!
526 These strange casts can happen as a result of case-of-case
527 bar = case (case x of { T -> (# 2,3 #); F -> error "urk" }) of
532 makeTrivial :: TopLevelFlag -> SimplEnv -> OutExpr -> SimplM (SimplEnv, OutExpr)
533 -- Binds the expression to a variable, if it's not trivial, returning the variable
534 makeTrivial top_lvl env expr = makeTrivialWithInfo top_lvl env vanillaIdInfo expr
536 makeTrivialWithInfo :: TopLevelFlag -> SimplEnv -> IdInfo
537 -> OutExpr -> SimplM (SimplEnv, OutExpr)
538 -- Propagate strictness and demand info to the new binder
539 -- Note [Preserve strictness when floating coercions]
540 -- Returned SimplEnv has same substitution as incoming one
541 makeTrivialWithInfo top_lvl env info expr
542 | exprIsTrivial expr -- Already trivial
543 || not (bindingOk top_lvl expr expr_ty) -- Cannot trivialise
544 -- See Note [Cannot trivialise]
546 | otherwise -- See Note [Take care] below
547 = do { uniq <- getUniqueM
548 ; let name = mkSystemVarName uniq (fsLit "a")
549 var = mkLocalIdWithInfo name expr_ty info
550 ; env' <- completeNonRecX top_lvl env False var var expr
551 ; expr' <- simplVar env' var
552 ; return (env', expr') }
553 -- The simplVar is needed becase we're constructing a new binding
555 -- And if rhs is of form (rhs1 |> co), then we might get
558 -- and now a's RHS is trivial and can be substituted out, and that
559 -- is what completeNonRecX will do
560 -- To put it another way, it's as if we'd simplified
561 -- let var = e in var
563 expr_ty = exprType expr
565 bindingOk :: TopLevelFlag -> CoreExpr -> Type -> Bool
566 -- True iff we can have a binding of this expression at this level
567 -- Precondition: the type is the type of the expression
568 bindingOk top_lvl _ expr_ty
569 | isTopLevel top_lvl = not (isUnLiftedType expr_ty)
573 Note [Cannot trivialise]
574 ~~~~~~~~~~~~~~~~~~~~~~~~
581 Then we can't ANF-ise foo, even though we'd like to, because
582 we can't make a top-level binding for the Addr# (f 3). And if
583 so we don't want to turn it into
584 foo = let x = f 3 in Bar x
585 because we'll just end up inlining x back, and that makes the
586 simplifier loop. Better not to ANF-ise it at all.
588 A case in point is literal strings (a MachStr is not regarded as
593 We don't want to ANF-ise this.
595 %************************************************************************
597 \subsection{Completing a lazy binding}
599 %************************************************************************
602 * deals only with Ids, not TyVars
603 * takes an already-simplified binder and RHS
604 * is used for both recursive and non-recursive bindings
605 * is used for both top-level and non-top-level bindings
607 It does the following:
608 - tries discarding a dead binding
609 - tries PostInlineUnconditionally
610 - add unfolding [this is the only place we add an unfolding]
613 It does *not* attempt to do let-to-case. Why? Because it is used for
614 - top-level bindings (when let-to-case is impossible)
615 - many situations where the "rhs" is known to be a WHNF
616 (so let-to-case is inappropriate).
618 Nor does it do the atomic-argument thing
621 completeBind :: SimplEnv
622 -> TopLevelFlag -- Flag stuck into unfolding
623 -> InId -- Old binder
624 -> OutId -> OutExpr -- New binder and RHS
626 -- completeBind may choose to do its work
627 -- * by extending the substitution (e.g. let x = y in ...)
628 -- * or by adding to the floats in the envt
630 completeBind env top_lvl old_bndr new_bndr new_rhs
631 = ASSERT( isId new_bndr )
632 do { let old_info = idInfo old_bndr
633 old_unf = unfoldingInfo old_info
634 occ_info = occInfo old_info
636 -- Do eta-expansion on the RHS of the binding
637 -- See Note [Eta-expanding at let bindings] in SimplUtils
638 ; (new_arity, final_rhs) <- tryEtaExpand env new_bndr new_rhs
640 -- Simplify the unfolding
641 ; new_unfolding <- simplUnfolding env top_lvl old_bndr occ_info final_rhs old_unf
643 ; if postInlineUnconditionally env top_lvl new_bndr occ_info final_rhs new_unfolding
644 -- Inline and discard the binding
645 then do { tick (PostInlineUnconditionally old_bndr)
646 ; -- pprTrace "postInlineUnconditionally"
647 -- (ppr old_bndr <+> equals <+> ppr final_rhs $$ ppr occ_info) $
648 return (extendIdSubst env old_bndr (DoneEx final_rhs)) }
649 -- Use the substitution to make quite, quite sure that the
650 -- substitution will happen, since we are going to discard the binding
652 do { let info1 = idInfo new_bndr `setArityInfo` new_arity
654 -- Unfolding info: Note [Setting the new unfolding]
655 info2 = info1 `setUnfoldingInfo` new_unfolding
657 -- Demand info: Note [Setting the demand info]
658 info3 | isEvaldUnfolding new_unfolding = zapDemandInfo info2 `orElse` info2
661 final_id = new_bndr `setIdInfo` info3
663 ; -- pprTrace "Binding" (ppr final_id <+> ppr unfolding) $
664 return (addNonRec env final_id final_rhs) } }
665 -- The addNonRec adds it to the in-scope set too
667 ------------------------------
668 addPolyBind :: TopLevelFlag -> SimplEnv -> OutBind -> SimplM SimplEnv
669 -- Add a new binding to the environment, complete with its unfolding
670 -- but *do not* do postInlineUnconditionally, because we have already
671 -- processed some of the scope of the binding
672 -- We still want the unfolding though. Consider
674 -- x = /\a. let y = ... in Just y
676 -- Then we float the y-binding out (via abstractFloats and addPolyBind)
677 -- but 'x' may well then be inlined in 'body' in which case we'd like the
678 -- opportunity to inline 'y' too.
680 addPolyBind top_lvl env (NonRec poly_id rhs)
681 = do { unfolding <- simplUnfolding env top_lvl poly_id NoOccInfo rhs noUnfolding
682 -- Assumes that poly_id did not have an INLINE prag
683 -- which is perhaps wrong. ToDo: think about this
684 ; let final_id = setIdInfo poly_id $
685 idInfo poly_id `setUnfoldingInfo` unfolding
686 `setArityInfo` exprArity rhs
688 ; return (addNonRec env final_id rhs) }
690 addPolyBind _ env bind@(Rec _)
691 = return (extendFloats env bind)
692 -- Hack: letrecs are more awkward, so we extend "by steam"
693 -- without adding unfoldings etc. At worst this leads to
694 -- more simplifier iterations
696 ------------------------------
697 simplUnfolding :: SimplEnv-> TopLevelFlag
699 -> OccInfo -> OutExpr
700 -> Unfolding -> SimplM Unfolding
701 -- Note [Setting the new unfolding]
702 simplUnfolding env _ _ _ _ (DFunUnfolding ar con ops)
703 = return (DFunUnfolding ar con ops')
705 ops' = map (fmap (substExpr (text "simplUnfolding") env)) ops
707 simplUnfolding env top_lvl id _ _
708 (CoreUnfolding { uf_tmpl = expr, uf_arity = arity
709 , uf_src = src, uf_guidance = guide })
711 = do { expr' <- simplExpr rule_env expr
712 ; let src' = CoreSubst.substUnfoldingSource (mkCoreSubst (text "inline-unf") env) src
713 is_top_lvl = isTopLevel top_lvl
716 -- We need to force bottoming, or the new unfolding holds
717 -- on to the old unfolding (which is part of the id).
718 let bottoming = isBottomingId id
719 in bottoming `seq` return (mkUnfolding src' is_top_lvl bottoming expr')
720 -- If the guidance is UnfIfGoodArgs, this is an INLINABLE
721 -- unfolding, and we need to make sure the guidance is kept up
722 -- to date with respect to any changes in the unfolding.
724 return (mkCoreUnfolding src' is_top_lvl expr' arity guide)
725 -- See Note [Top-level flag on inline rules] in CoreUnfold
728 act = idInlineActivation id
729 rule_env = updMode (updModeForInlineRules act) env
730 -- See Note [Simplifying inside InlineRules] in SimplUtils
732 simplUnfolding _ top_lvl id _occ_info new_rhs _
733 = -- We need to force bottoming, or the new unfolding holds
734 -- on to the old unfolding (which is part of the id).
735 let bottoming = isBottomingId id
736 in bottoming `seq` return (mkUnfolding InlineRhs (isTopLevel top_lvl) bottoming new_rhs)
737 -- We make an unfolding *even for loop-breakers*.
738 -- Reason: (a) It might be useful to know that they are WHNF
739 -- (b) In TidyPgm we currently assume that, if we want to
740 -- expose the unfolding then indeed we *have* an unfolding
741 -- to expose. (We could instead use the RHS, but currently
742 -- we don't.) The simple thing is always to have one.
745 Note [Arity decrease]
746 ~~~~~~~~~~~~~~~~~~~~~
747 Generally speaking the arity of a binding should not decrease. But it *can*
748 legitimately happen becuase of RULES. Eg
750 where g has arity 2, will have arity 2. But if there's a rewrite rule
752 where h has arity 1, then f's arity will decrease. Here's a real-life example,
753 which is in the output of Specialise:
756 $dm {Arity 2} = \d.\x. op d
757 {-# RULES forall d. $dm Int d = $s$dm #-}
759 dInt = MkD .... opInt ...
760 opInt {Arity 1} = $dm dInt
762 $s$dm {Arity 0} = \x. op dInt }
764 Here opInt has arity 1; but when we apply the rule its arity drops to 0.
765 That's why Specialise goes to a little trouble to pin the right arity
766 on specialised functions too.
768 Note [Setting the new unfolding]
769 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
770 * If there's an INLINE pragma, we simplify the RHS gently. Maybe we
771 should do nothing at all, but simplifying gently might get rid of
774 * If not, we make an unfolding from the new RHS. But *only* for
775 non-loop-breakers. Making loop breakers not have an unfolding at all
776 means that we can avoid tests in exprIsConApp, for example. This is
777 important: if exprIsConApp says 'yes' for a recursive thing, then we
778 can get into an infinite loop
780 If there's an InlineRule on a loop breaker, we hang on to the inlining.
781 It's pretty dodgy, but the user did say 'INLINE'. May need to revisit
784 Note [Setting the demand info]
785 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
786 If the unfolding is a value, the demand info may
787 go pear-shaped, so we nuke it. Example:
789 case x of (p,q) -> h p q x
790 Here x is certainly demanded. But after we've nuked
791 the case, we'll get just
792 let x = (a,b) in h a b x
793 and now x is not demanded (I'm assuming h is lazy)
794 This really happens. Similarly
795 let f = \x -> e in ...f..f...
796 After inlining f at some of its call sites the original binding may
797 (for example) be no longer strictly demanded.
798 The solution here is a bit ad hoc...
801 %************************************************************************
803 \subsection[Simplify-simplExpr]{The main function: simplExpr}
805 %************************************************************************
807 The reason for this OutExprStuff stuff is that we want to float *after*
808 simplifying a RHS, not before. If we do so naively we get quadratic
809 behaviour as things float out.
811 To see why it's important to do it after, consider this (real) example:
825 a -- Can't inline a this round, cos it appears twice
829 Each of the ==> steps is a round of simplification. We'd save a
830 whole round if we float first. This can cascade. Consider
835 let f = let d1 = ..d.. in \y -> e
839 in \x -> ...(\y ->e)...
841 Only in this second round can the \y be applied, and it
842 might do the same again.
846 simplExpr :: SimplEnv -> CoreExpr -> SimplM CoreExpr
847 simplExpr env expr = simplExprC env expr mkBoringStop
849 simplExprC :: SimplEnv -> CoreExpr -> SimplCont -> SimplM CoreExpr
850 -- Simplify an expression, given a continuation
851 simplExprC env expr cont
852 = -- pprTrace "simplExprC" (ppr expr $$ ppr cont {- $$ ppr (seIdSubst env) -} $$ ppr (seFloats env) ) $
853 do { (env', expr') <- simplExprF (zapFloats env) expr cont
854 ; -- pprTrace "simplExprC ret" (ppr expr $$ ppr expr') $
855 -- pprTrace "simplExprC ret3" (ppr (seInScope env')) $
856 -- pprTrace "simplExprC ret4" (ppr (seFloats env')) $
857 return (wrapFloats env' expr') }
859 --------------------------------------------------
860 simplExprF :: SimplEnv -> InExpr -> SimplCont
861 -> SimplM (SimplEnv, OutExpr)
863 simplExprF env e cont
864 = -- pprTrace "simplExprF" (ppr e $$ ppr cont $$ ppr (seTvSubst env) $$ ppr (seIdSubst env) {- $$ ppr (seFloats env) -} ) $
865 simplExprF' env e cont
867 simplExprF' :: SimplEnv -> InExpr -> SimplCont
868 -> SimplM (SimplEnv, OutExpr)
869 simplExprF' env (Var v) cont = simplVarF env v cont
870 simplExprF' env (Lit lit) cont = rebuild env (Lit lit) cont
871 simplExprF' env (Note n expr) cont = simplNote env n expr cont
872 simplExprF' env (Cast body co) cont = simplCast env body co cont
873 simplExprF' env (App fun arg) cont = simplExprF env fun $
874 ApplyTo NoDup arg env cont
876 simplExprF' env expr@(Lam _ _) cont
877 = simplLam env zapped_bndrs body cont
878 -- The main issue here is under-saturated lambdas
879 -- (\x1. \x2. e) arg1
880 -- Here x1 might have "occurs-once" occ-info, because occ-info
881 -- is computed assuming that a group of lambdas is applied
882 -- all at once. If there are too few args, we must zap the
883 -- occ-info, UNLESS the remaining binders are one-shot
885 (bndrs, body) = collectBinders expr
886 zapped_bndrs | need_to_zap = map zap bndrs
889 need_to_zap = any zappable_bndr (drop n_args bndrs)
890 n_args = countArgs cont
891 -- NB: countArgs counts all the args (incl type args)
892 -- and likewise drop counts all binders (incl type lambdas)
894 zappable_bndr b = isId b && not (isOneShotBndr b)
895 zap b | isTyCoVar b = b
896 | otherwise = zapLamIdInfo b
898 simplExprF' env (Type ty) cont
899 = ASSERT( contIsRhsOrArg cont )
900 do { ty' <- simplCoercion env ty
901 ; rebuild env (Type ty') cont }
903 simplExprF' env (Case scrut bndr _ alts) cont
904 | sm_case_case (getMode env)
905 = -- Simplify the scrutinee with a Select continuation
906 simplExprF env scrut (Select NoDup bndr alts env cont)
909 = -- If case-of-case is off, simply simplify the case expression
910 -- in a vanilla Stop context, and rebuild the result around it
911 do { case_expr' <- simplExprC env scrut
912 (Select NoDup bndr alts env mkBoringStop)
913 ; rebuild env case_expr' cont }
915 simplExprF' env (Let (Rec pairs) body) cont
916 = do { env' <- simplRecBndrs env (map fst pairs)
917 -- NB: bndrs' don't have unfoldings or rules
918 -- We add them as we go down
920 ; env'' <- simplRecBind env' NotTopLevel pairs
921 ; simplExprF env'' body cont }
923 simplExprF' env (Let (NonRec bndr rhs) body) cont
924 = simplNonRecE env bndr (rhs, env) ([], body) cont
926 ---------------------------------
927 simplType :: SimplEnv -> InType -> SimplM OutType
928 -- Kept monadic just so we can do the seqType
930 = -- pprTrace "simplType" (ppr ty $$ ppr (seTvSubst env)) $
931 seqType new_ty `seq` return new_ty
933 new_ty = substTy env ty
935 ---------------------------------
936 simplCoercion :: SimplEnv -> InType -> SimplM OutType
937 -- The InType isn't *necessarily* a coercion, but it might be
938 -- (in a type application, say) and optCoercion is a no-op on types
940 = seqType new_co `seq` return new_co
942 new_co = optCoercion (getTvSubst env) co
946 %************************************************************************
948 \subsection{The main rebuilder}
950 %************************************************************************
953 rebuild :: SimplEnv -> OutExpr -> SimplCont -> SimplM (SimplEnv, OutExpr)
954 -- At this point the substitution in the SimplEnv should be irrelevant
955 -- only the in-scope set and floats should matter
956 rebuild env expr cont
958 Stop {} -> return (env, expr)
959 CoerceIt co cont -> rebuild env (mkCoerce co expr) cont
960 Select _ bndr alts se cont -> rebuildCase (se `setFloats` env) expr bndr alts cont
961 StrictArg info _ cont -> rebuildCall env (info `addArgTo` expr) cont
962 StrictBind b bs body se cont -> do { env' <- simplNonRecX (se `setFloats` env) b expr
963 ; simplLam env' bs body cont }
964 ApplyTo dup_flag arg se cont -- See Note [Avoid redundant simplification]
965 | isSimplified dup_flag -> rebuild env (App expr arg) cont
966 | otherwise -> do { arg' <- simplExpr (se `setInScope` env) arg
967 ; rebuild env (App expr arg') cont }
971 %************************************************************************
975 %************************************************************************
978 simplCast :: SimplEnv -> InExpr -> Coercion -> SimplCont
979 -> SimplM (SimplEnv, OutExpr)
980 simplCast env body co0 cont0
981 = do { co1 <- simplCoercion env co0
982 ; simplExprF env body (addCoerce co1 cont0) }
984 addCoerce co cont = add_coerce co (coercionKind co) cont
986 add_coerce _co (s1, k1) cont -- co :: ty~ty
987 | s1 `coreEqType` k1 = cont -- is a no-op
989 add_coerce co1 (s1, _k2) (CoerceIt co2 cont)
990 | (_l1, t1) <- coercionKind co2
991 -- e |> (g1 :: S1~L) |> (g2 :: L~T1)
994 -- e |> (g1 . g2 :: S1~T1) otherwise
996 -- For example, in the initial form of a worker
997 -- we may find (coerce T (coerce S (\x.e))) y
998 -- and we'd like it to simplify to e[y/x] in one round
1000 , s1 `coreEqType` t1 = cont -- The coerces cancel out
1001 | otherwise = CoerceIt (mkTransCoercion co1 co2) cont
1003 add_coerce co (s1s2, _t1t2) (ApplyTo dup (Type arg_ty) arg_se cont)
1004 -- (f |> g) ty ---> (f ty) |> (g @ ty)
1005 -- This implements the PushT and PushC rules from the paper
1006 | Just (tyvar,_) <- splitForAllTy_maybe s1s2
1008 (new_arg_ty, new_cast)
1009 | isCoVar tyvar = (new_arg_co, mkCselRCoercion co) -- PushC rule
1010 | otherwise = (ty', mkInstCoercion co ty') -- PushT rule
1012 ApplyTo dup (Type new_arg_ty) (zapSubstEnv arg_se) (addCoerce new_cast cont)
1014 ty' = substTy (arg_se `setInScope` env) arg_ty
1015 new_arg_co = mkCsel1Coercion co `mkTransCoercion`
1016 ty' `mkTransCoercion`
1017 mkSymCoercion (mkCsel2Coercion co)
1019 add_coerce co (s1s2, _t1t2) (ApplyTo dup arg arg_se cont)
1020 | not (isTypeArg arg) -- This implements the Push rule from the paper
1021 , isFunTy s1s2 -- t1t2 must be a function type, becuase it's applied
1022 -- (e |> (g :: s1s2 ~ t1->t2)) f
1024 -- (e (f |> (arg g :: t1~s1))
1025 -- |> (res g :: s2->t2)
1027 -- t1t2 must be a function type, t1->t2, because it's applied
1028 -- to something but s1s2 might conceivably not be
1030 -- When we build the ApplyTo we can't mix the out-types
1031 -- with the InExpr in the argument, so we simply substitute
1032 -- to make it all consistent. It's a bit messy.
1033 -- But it isn't a common case.
1035 -- Example of use: Trac #995
1036 = ApplyTo dup new_arg (zapSubstEnv arg_se) (addCoerce co2 cont)
1038 -- we split coercion t1->t2 ~ s1->s2 into t1 ~ s1 and
1039 -- t2 ~ s2 with left and right on the curried form:
1040 -- (->) t1 t2 ~ (->) s1 s2
1041 [co1, co2] = decomposeCo 2 co
1042 new_arg = mkCoerce (mkSymCoercion co1) arg'
1043 arg' = substExpr (text "move-cast") (arg_se `setInScope` env) arg
1045 add_coerce co _ cont = CoerceIt co cont
1049 %************************************************************************
1051 \subsection{Lambdas}
1053 %************************************************************************
1056 simplLam :: SimplEnv -> [InId] -> InExpr -> SimplCont
1057 -> SimplM (SimplEnv, OutExpr)
1059 simplLam env [] body cont = simplExprF env body cont
1062 simplLam env (bndr:bndrs) body (ApplyTo _ arg arg_se cont)
1063 = do { tick (BetaReduction bndr)
1064 ; simplNonRecE env bndr (arg, arg_se) (bndrs, body) cont }
1066 -- Not enough args, so there are real lambdas left to put in the result
1067 simplLam env bndrs body cont
1068 = do { (env', bndrs') <- simplLamBndrs env bndrs
1069 ; body' <- simplExpr env' body
1070 ; new_lam <- mkLam env' bndrs' body'
1071 ; rebuild env' new_lam cont }
1074 simplNonRecE :: SimplEnv
1075 -> InBndr -- The binder
1076 -> (InExpr, SimplEnv) -- Rhs of binding (or arg of lambda)
1077 -> ([InBndr], InExpr) -- Body of the let/lambda
1080 -> SimplM (SimplEnv, OutExpr)
1082 -- simplNonRecE is used for
1083 -- * non-top-level non-recursive lets in expressions
1086 -- It deals with strict bindings, via the StrictBind continuation,
1087 -- which may abort the whole process
1089 -- The "body" of the binding comes as a pair of ([InId],InExpr)
1090 -- representing a lambda; so we recurse back to simplLam
1091 -- Why? Because of the binder-occ-info-zapping done before
1092 -- the call to simplLam in simplExprF (Lam ...)
1094 -- First deal with type applications and type lets
1095 -- (/\a. e) (Type ty) and (let a = Type ty in e)
1096 simplNonRecE env bndr (Type ty_arg, rhs_se) (bndrs, body) cont
1097 = ASSERT( isTyCoVar bndr )
1098 do { ty_arg' <- simplType (rhs_se `setInScope` env) ty_arg
1099 ; simplLam (extendTvSubst env bndr ty_arg') bndrs body cont }
1101 simplNonRecE env bndr (rhs, rhs_se) (bndrs, body) cont
1102 | preInlineUnconditionally env NotTopLevel bndr rhs
1103 = do { tick (PreInlineUnconditionally bndr)
1104 ; -- pprTrace "preInlineUncond" (ppr bndr <+> ppr rhs) $
1105 simplLam (extendIdSubst env bndr (mkContEx rhs_se rhs)) bndrs body cont }
1108 = do { simplExprF (rhs_se `setFloats` env) rhs
1109 (StrictBind bndr bndrs body env cont) }
1112 = ASSERT( not (isTyCoVar bndr) )
1113 do { (env1, bndr1) <- simplNonRecBndr env bndr
1114 ; let (env2, bndr2) = addBndrRules env1 bndr bndr1
1115 ; env3 <- simplLazyBind env2 NotTopLevel NonRecursive bndr bndr2 rhs rhs_se
1116 ; simplLam env3 bndrs body cont }
1120 %************************************************************************
1124 %************************************************************************
1127 -- Hack alert: we only distinguish subsumed cost centre stacks for the
1128 -- purposes of inlining. All other CCCSs are mapped to currentCCS.
1129 simplNote :: SimplEnv -> Note -> CoreExpr -> SimplCont
1130 -> SimplM (SimplEnv, OutExpr)
1131 simplNote env (SCC cc) e cont
1132 | pushCCisNop cc (getEnclosingCC env) -- scc "f" (...(scc "f" e)...)
1133 = simplExprF env e cont -- ==> scc "f" (...e...)
1135 = do { e' <- simplExpr (setEnclosingCC env currentCCS) e
1136 ; rebuild env (mkSCC cc e') cont }
1138 simplNote env (CoreNote s) e cont
1139 = do { e' <- simplExpr env e
1140 ; rebuild env (Note (CoreNote s) e') cont }
1144 %************************************************************************
1148 %************************************************************************
1151 simplVar :: SimplEnv -> InVar -> SimplM OutExpr
1152 -- Look up an InVar in the environment
1155 = return (Type (substTyVar env var))
1157 = case substId env var of
1158 DoneId var1 -> return (Var var1)
1159 DoneEx e -> return e
1160 ContEx tvs ids e -> simplExpr (setSubstEnv env tvs ids) e
1162 simplVarF :: SimplEnv -> InId -> SimplCont -> SimplM (SimplEnv, OutExpr)
1163 simplVarF env var cont
1164 = case substId env var of
1165 DoneEx e -> simplExprF (zapSubstEnv env) e cont
1166 ContEx tvs ids e -> simplExprF (setSubstEnv env tvs ids) e cont
1167 DoneId var1 -> completeCall env var1 cont
1168 -- Note [zapSubstEnv]
1169 -- The template is already simplified, so don't re-substitute.
1170 -- This is VITAL. Consider
1172 -- let y = \z -> ...x... in
1174 -- We'll clone the inner \x, adding x->x' in the id_subst
1175 -- Then when we inline y, we must *not* replace x by x' in
1176 -- the inlined copy!!
1178 ---------------------------------------------------------
1179 -- Dealing with a call site
1181 completeCall :: SimplEnv -> Id -> SimplCont -> SimplM (SimplEnv, OutExpr)
1182 completeCall env var cont
1183 = do { ------------- Try inlining ----------------
1184 dflags <- getDOptsSmpl
1185 ; let (lone_variable, arg_infos, call_cont) = contArgs cont
1186 -- The args are OutExprs, obtained by *lazily* substituting
1187 -- in the args found in cont. These args are only examined
1188 -- to limited depth (unless a rule fires). But we must do
1189 -- the substitution; rule matching on un-simplified args would
1192 n_val_args = length arg_infos
1193 interesting_cont = interestingCallContext call_cont
1194 unfolding = activeUnfolding env var
1195 maybe_inline = callSiteInline dflags var unfolding
1196 lone_variable arg_infos interesting_cont
1197 ; case maybe_inline of {
1198 Just expr -- There is an inlining!
1199 -> do { tick (UnfoldingDone var)
1200 ; trace_inline dflags expr cont $
1201 simplExprF (zapSubstEnv env) expr cont }
1203 ; Nothing -> do -- No inlining!
1205 { rule_base <- getSimplRules
1206 ; let info = mkArgInfo var (getRules rule_base var) n_val_args call_cont
1207 ; rebuildCall env info cont
1210 trace_inline dflags unfolding cont stuff
1211 | not (dopt Opt_D_dump_inlinings dflags) = stuff
1212 | not (dopt Opt_D_verbose_core2core dflags)
1213 = if isExternalName (idName var) then
1214 pprTrace "Inlining done:" (ppr var) stuff
1217 = pprTrace ("Inlining done: " ++ showSDoc (ppr var))
1218 (vcat [text "Inlined fn: " <+> nest 2 (ppr unfolding),
1219 text "Cont: " <+> ppr cont])
1222 rebuildCall :: SimplEnv
1225 -> SimplM (SimplEnv, OutExpr)
1226 rebuildCall env (ArgInfo { ai_fun = fun, ai_args = rev_args, ai_strs = [] }) cont
1227 -- When we run out of strictness args, it means
1228 -- that the call is definitely bottom; see SimplUtils.mkArgInfo
1229 -- Then we want to discard the entire strict continuation. E.g.
1230 -- * case (error "hello") of { ... }
1231 -- * (error "Hello") arg
1232 -- * f (error "Hello") where f is strict
1234 -- Then, especially in the first of these cases, we'd like to discard
1235 -- the continuation, leaving just the bottoming expression. But the
1236 -- type might not be right, so we may have to add a coerce.
1237 | not (contIsTrivial cont) -- Only do this if there is a non-trivial
1238 = return (env, mk_coerce res) -- contination to discard, else we do it
1239 where -- again and again!
1240 res = mkApps (Var fun) (reverse rev_args)
1241 res_ty = exprType res
1242 cont_ty = contResultType env res_ty cont
1243 co = mkUnsafeCoercion res_ty cont_ty
1244 mk_coerce expr | cont_ty `coreEqType` res_ty = expr
1245 | otherwise = mkCoerce co expr
1247 rebuildCall env info (ApplyTo _ (Type arg_ty) se cont)
1248 = do { ty' <- simplCoercion (se `setInScope` env) arg_ty
1249 ; rebuildCall env (info `addArgTo` Type ty') cont }
1251 rebuildCall env info@(ArgInfo { ai_encl = encl_rules
1252 , ai_strs = str:strs, ai_discs = disc:discs })
1253 (ApplyTo dup_flag arg arg_se cont)
1254 | isSimplified dup_flag -- See Note [Avoid redundant simplification]
1255 = rebuildCall env (addArgTo info' arg) cont
1257 | str -- Strict argument
1258 = -- pprTrace "Strict Arg" (ppr arg $$ ppr (seIdSubst env) $$ ppr (seInScope env)) $
1259 simplExprF (arg_se `setFloats` env) arg
1260 (StrictArg info' cci cont)
1263 | otherwise -- Lazy argument
1264 -- DO NOT float anything outside, hence simplExprC
1265 -- There is no benefit (unlike in a let-binding), and we'd
1266 -- have to be very careful about bogus strictness through
1267 -- floating a demanded let.
1268 = do { arg' <- simplExprC (arg_se `setInScope` env) arg
1270 ; rebuildCall env (addArgTo info' arg') cont }
1272 info' = info { ai_strs = strs, ai_discs = discs }
1273 cci | encl_rules || disc > 0 = ArgCtxt encl_rules -- Be keener here
1274 | otherwise = BoringCtxt -- Nothing interesting
1276 rebuildCall env (ArgInfo { ai_fun = fun, ai_args = rev_args, ai_rules = rules }) cont
1277 = do { -- We've accumulated a simplified call in <fun,rev_args>
1278 -- so try rewrite rules; see Note [RULEs apply to simplified arguments]
1279 -- See also Note [Rules for recursive functions]
1280 ; let args = reverse rev_args
1281 env' = zapSubstEnv env
1282 ; mb_rule <- tryRules env rules fun args cont
1284 Just (n_args, rule_rhs) -> simplExprF env' rule_rhs $
1285 pushSimplifiedArgs env' (drop n_args args) cont ;
1286 -- n_args says how many args the rule consumed
1287 ; Nothing -> rebuild env (mkApps (Var fun) args) cont -- No rules
1291 Note [RULES apply to simplified arguments]
1292 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
1293 It's very desirable to try RULES once the arguments have been simplified, because
1294 doing so ensures that rule cascades work in one pass. Consider
1295 {-# RULES g (h x) = k x
1298 Then we want to rewrite (g (h x)) to (k x) and only then try f's rules. If
1299 we match f's rules against the un-simplified RHS, it won't match. This
1300 makes a particularly big difference when superclass selectors are involved:
1301 op ($p1 ($p2 (df d)))
1302 We want all this to unravel in one sweeep.
1304 Note [Avoid redundant simplification]
1305 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
1306 Because RULES apply to simplified arguments, there's a danger of repeatedly
1307 simplifying already-simplified arguments. An important example is that of
1309 Here e1, e2 are simplified before the rule is applied, but don't really
1310 participate in the rule firing. So we mark them as Simplified to avoid
1311 re-simplifying them.
1315 This part of the simplifier may break the no-shadowing invariant
1317 f (...(\a -> e)...) (case y of (a,b) -> e')
1318 where f is strict in its second arg
1319 If we simplify the innermost one first we get (...(\a -> e)...)
1320 Simplifying the second arg makes us float the case out, so we end up with
1321 case y of (a,b) -> f (...(\a -> e)...) e'
1322 So the output does not have the no-shadowing invariant. However, there is
1323 no danger of getting name-capture, because when the first arg was simplified
1324 we used an in-scope set that at least mentioned all the variables free in its
1325 static environment, and that is enough.
1327 We can't just do innermost first, or we'd end up with a dual problem:
1328 case x of (a,b) -> f e (...(\a -> e')...)
1330 I spent hours trying to recover the no-shadowing invariant, but I just could
1331 not think of an elegant way to do it. The simplifier is already knee-deep in
1332 continuations. We have to keep the right in-scope set around; AND we have
1333 to get the effect that finding (error "foo") in a strict arg position will
1334 discard the entire application and replace it with (error "foo"). Getting
1335 all this at once is TOO HARD!
1338 %************************************************************************
1342 %************************************************************************
1345 tryRules :: SimplEnv -> [CoreRule]
1346 -> Id -> [OutExpr] -> SimplCont
1347 -> SimplM (Maybe (Arity, CoreExpr)) -- The arity is the number of
1348 -- args consumed by the rule
1349 tryRules env rules fn args call_cont
1353 = do { dflags <- getDOptsSmpl
1354 ; case activeRule dflags env of {
1355 Nothing -> return Nothing ; -- No rules apply
1357 case lookupRule act_fn (getUnfoldingInRuleMatch env) (getInScope env) fn args rules of {
1358 Nothing -> return Nothing ; -- No rule matches
1359 Just (rule, rule_rhs) ->
1361 do { tick (RuleFired (ru_name rule))
1362 ; trace_dump dflags rule rule_rhs $
1363 return (Just (ruleArity rule, rule_rhs)) }}}}
1365 trace_dump dflags rule rule_rhs stuff
1366 | not (dopt Opt_D_dump_rule_firings dflags)
1367 , not (dopt Opt_D_dump_rule_rewrites dflags) = stuff
1368 | not (dopt Opt_D_dump_rule_rewrites dflags)
1370 = pprTrace "Rule fired:" (ftext (ru_name rule)) stuff
1372 = pprTrace "Rule fired"
1373 (vcat [text "Rule:" <+> ftext (ru_name rule),
1374 text "Before:" <+> hang (ppr fn) 2 (sep (map pprParendExpr args)),
1375 text "After: " <+> pprCoreExpr rule_rhs,
1376 text "Cont: " <+> ppr call_cont])
1380 Note [Rules for recursive functions]
1381 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
1382 You might think that we shouldn't apply rules for a loop breaker:
1383 doing so might give rise to an infinite loop, because a RULE is
1384 rather like an extra equation for the function:
1385 RULE: f (g x) y = x+y
1388 But it's too drastic to disable rules for loop breakers.
1389 Even the foldr/build rule would be disabled, because foldr
1390 is recursive, and hence a loop breaker:
1391 foldr k z (build g) = g k z
1392 So it's up to the programmer: rules can cause divergence
1395 %************************************************************************
1397 Rebuilding a cse expression
1399 %************************************************************************
1401 Note [Case elimination]
1402 ~~~~~~~~~~~~~~~~~~~~~~~
1403 The case-elimination transformation discards redundant case expressions.
1404 Start with a simple situation:
1406 case x# of ===> e[x#/y#]
1409 (when x#, y# are of primitive type, of course). We can't (in general)
1410 do this for algebraic cases, because we might turn bottom into
1413 The code in SimplUtils.prepareAlts has the effect of generalise this
1414 idea to look for a case where we're scrutinising a variable, and we
1415 know that only the default case can match. For example:
1419 DEFAULT -> ...(case x of
1423 Here the inner case is first trimmed to have only one alternative, the
1424 DEFAULT, after which it's an instance of the previous case. This
1425 really only shows up in eliminating error-checking code.
1427 We also make sure that we deal with this very common case:
1432 Here we are using the case as a strict let; if x is used only once
1433 then we want to inline it. We have to be careful that this doesn't
1434 make the program terminate when it would have diverged before, so we
1436 - e is already evaluated (it may so if e is a variable)
1437 - x is used strictly, or
1439 Lastly, the code in SimplUtils.mkCase combines identical RHSs. So
1441 case e of ===> case e of DEFAULT -> r
1445 Now again the case may be elminated by the CaseElim transformation.
1447 Note [CaseElimination: lifted case]
1448 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
1449 We do not use exprOkForSpeculation in the lifted case. Consider
1450 case (case a ># b of { True -> (p,q); False -> (q,p) }) of
1452 The scrutinee is ok-for-speculation (it looks inside cases), but we do
1453 not want to transform to
1454 let r = case a ># b of { True -> (p,q); False -> (q,p) }
1456 because that builds an unnecessary thunk.
1459 Further notes about case elimination
1460 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
1461 Consider: test :: Integer -> IO ()
1464 Turns out that this compiles to:
1467 eta1 :: State# RealWorld ->
1468 case PrelNum.< eta PrelNum.zeroInteger of wild { __DEFAULT ->
1470 (PrelNum.jtos eta ($w[] @ Char))
1472 of wild1 { (# new_s, a4 #) -> PrelIO.lvl23 new_s }}
1474 Notice the strange '<' which has no effect at all. This is a funny one.
1475 It started like this:
1477 f x y = if x < 0 then jtos x
1478 else if y==0 then "" else jtos x
1480 At a particular call site we have (f v 1). So we inline to get
1482 if v < 0 then jtos x
1483 else if 1==0 then "" else jtos x
1485 Now simplify the 1==0 conditional:
1487 if v<0 then jtos v else jtos v
1489 Now common-up the two branches of the case:
1491 case (v<0) of DEFAULT -> jtos v
1493 Why don't we drop the case? Because it's strict in v. It's technically
1494 wrong to drop even unnecessary evaluations, and in practice they
1495 may be a result of 'seq' so we *definitely* don't want to drop those.
1496 I don't really know how to improve this situation.
1499 ---------------------------------------------------------
1500 -- Eliminate the case if possible
1502 rebuildCase, reallyRebuildCase
1504 -> OutExpr -- Scrutinee
1505 -> InId -- Case binder
1506 -> [InAlt] -- Alternatives (inceasing order)
1508 -> SimplM (SimplEnv, OutExpr)
1510 --------------------------------------------------
1511 -- 1. Eliminate the case if there's a known constructor
1512 --------------------------------------------------
1514 rebuildCase env scrut case_bndr alts cont
1515 | Lit lit <- scrut -- No need for same treatment as constructors
1516 -- because literals are inlined more vigorously
1517 = do { tick (KnownBranch case_bndr)
1518 ; case findAlt (LitAlt lit) alts of
1519 Nothing -> missingAlt env case_bndr alts cont
1520 Just (_, bs, rhs) -> simple_rhs bs rhs }
1522 | Just (con, ty_args, other_args) <- exprIsConApp_maybe (getUnfoldingInRuleMatch env) scrut
1523 -- Works when the scrutinee is a variable with a known unfolding
1524 -- as well as when it's an explicit constructor application
1525 = do { tick (KnownBranch case_bndr)
1526 ; case findAlt (DataAlt con) alts of
1527 Nothing -> missingAlt env case_bndr alts cont
1528 Just (DEFAULT, bs, rhs) -> simple_rhs bs rhs
1529 Just (_, bs, rhs) -> knownCon env scrut con ty_args other_args
1530 case_bndr bs rhs cont
1533 simple_rhs bs rhs = ASSERT( null bs )
1534 do { env' <- simplNonRecX env case_bndr scrut
1535 ; simplExprF env' rhs cont }
1538 --------------------------------------------------
1539 -- 2. Eliminate the case if scrutinee is evaluated
1540 --------------------------------------------------
1542 rebuildCase env scrut case_bndr [(_, bndrs, rhs)] cont
1543 -- See if we can get rid of the case altogether
1544 -- See Note [Case elimination]
1545 -- mkCase made sure that if all the alternatives are equal,
1546 -- then there is now only one (DEFAULT) rhs
1547 | all isDeadBinder bndrs -- bndrs are [InId]
1549 -- Check that the scrutinee can be let-bound instead of case-bound
1550 , if isUnLiftedType (idType case_bndr)
1551 then exprOkForSpeculation scrut
1552 -- Satisfy the let-binding invariant
1553 -- This includes things like (==# a# b#)::Bool
1554 -- so that we simplify
1555 -- case ==# a# b# of { True -> x; False -> x }
1558 -- This particular example shows up in default methods for
1559 -- comparision operations (e.g. in (>=) for Int.Int32)
1561 else exprIsHNF scrut || var_demanded_later scrut
1562 -- It's already evaluated, or will be demanded later
1563 -- See Note [Case elimination: lifted case]
1564 = do { tick (CaseElim case_bndr)
1565 ; env' <- simplNonRecX env case_bndr scrut
1566 -- If case_bndr is deads, simplNonRecX will discard
1567 ; simplExprF env' rhs cont }
1569 -- The case binder is going to be evaluated later,
1570 -- and the scrutinee is a simple variable
1571 var_demanded_later (Var v) = isStrictDmd (idDemandInfo case_bndr)
1572 && not (isTickBoxOp v)
1573 -- ugly hack; covering this case is what
1574 -- exprOkForSpeculation was intended for.
1575 var_demanded_later _ = False
1577 --------------------------------------------------
1578 -- 3. Try seq rules; see Note [User-defined RULES for seq] in MkId
1579 --------------------------------------------------
1581 rebuildCase env scrut case_bndr alts@[(_, bndrs, rhs)] cont
1582 | all isDeadBinder (case_bndr : bndrs) -- So this is just 'seq'
1583 = do { let rhs' = substExpr (text "rebuild-case") env rhs
1584 out_args = [Type (substTy env (idType case_bndr)),
1585 Type (exprType rhs'), scrut, rhs']
1586 -- Lazily evaluated, so we don't do most of this
1588 ; rule_base <- getSimplRules
1589 ; mb_rule <- tryRules env (getRules rule_base seqId) seqId out_args cont
1591 Just (n_args, res) -> simplExprF (zapSubstEnv env)
1592 (mkApps res (drop n_args out_args))
1594 Nothing -> reallyRebuildCase env scrut case_bndr alts cont }
1596 rebuildCase env scrut case_bndr alts cont
1597 = reallyRebuildCase env scrut case_bndr alts cont
1599 --------------------------------------------------
1600 -- 3. Catch-all case
1601 --------------------------------------------------
1603 reallyRebuildCase env scrut case_bndr alts cont
1604 = do { -- Prepare the continuation;
1605 -- The new subst_env is in place
1606 (env', dup_cont, nodup_cont) <- prepareCaseCont env alts cont
1608 -- Simplify the alternatives
1609 ; (scrut', case_bndr', alts') <- simplAlts env' scrut case_bndr alts dup_cont
1611 -- Check for empty alternatives
1612 ; if null alts' then missingAlt env case_bndr alts cont
1614 { dflags <- getDOptsSmpl
1615 ; case_expr <- mkCase dflags scrut' case_bndr' alts'
1617 -- Notice that rebuild gets the in-scope set from env', not alt_env
1618 -- (which in any case is only build in simplAlts)
1619 -- The case binder *not* scope over the whole returned case-expression
1620 ; rebuild env' case_expr nodup_cont } }
1623 simplCaseBinder checks whether the scrutinee is a variable, v. If so,
1624 try to eliminate uses of v in the RHSs in favour of case_bndr; that
1625 way, there's a chance that v will now only be used once, and hence
1628 Historical note: we use to do the "case binder swap" in the Simplifier
1629 so there were additional complications if the scrutinee was a variable.
1630 Now the binder-swap stuff is done in the occurrence analyer; see
1631 OccurAnal Note [Binder swap].
1635 If the case binder is not dead, then neither are the pattern bound
1637 case <any> of x { (a,b) ->
1638 case x of { (p,q) -> p } }
1639 Here (a,b) both look dead, but come alive after the inner case is eliminated.
1640 The point is that we bring into the envt a binding
1642 after the outer case, and that makes (a,b) alive. At least we do unless
1643 the case binder is guaranteed dead.
1645 In practice, the scrutinee is almost always a variable, so we pretty
1646 much always zap the OccInfo of the binders. It doesn't matter much though.
1651 Consider case (v `cast` co) of x { I# y ->
1652 ... (case (v `cast` co) of {...}) ...
1653 We'd like to eliminate the inner case. We can get this neatly by
1654 arranging that inside the outer case we add the unfolding
1655 v |-> x `cast` (sym co)
1656 to v. Then we should inline v at the inner case, cancel the casts, and away we go
1658 Note [Improving seq]
1661 type family F :: * -> *
1662 type instance F Int = Int
1664 ... case e of x { DEFAULT -> rhs } ...
1666 where x::F Int. Then we'd like to rewrite (F Int) to Int, getting
1668 case e `cast` co of x'::Int
1669 I# x# -> let x = x' `cast` sym co
1672 so that 'rhs' can take advantage of the form of x'.
1674 Notice that Note [Case of cast] may then apply to the result.
1676 Nota Bene: We only do the [Improving seq] transformation if the
1677 case binder 'x' is actually used in the rhs; that is, if the case
1678 is *not* a *pure* seq.
1679 a) There is no point in adding the cast to a pure seq.
1680 b) There is a good reason not to: doing so would interfere
1681 with seq rules (Note [Built-in RULES for seq] in MkId).
1682 In particular, this [Improving seq] thing *adds* a cast
1683 while [Built-in RULES for seq] *removes* one, so they
1686 You might worry about
1687 case v of x { __DEFAULT ->
1688 ... case (v `cast` co) of y { I# -> ... }}
1689 This is a pure seq (since x is unused), so [Improving seq] won't happen.
1690 But it's ok: the simplifier will replace 'v' by 'x' in the rhs to get
1691 case v of x { __DEFAULT ->
1692 ... case (x `cast` co) of y { I# -> ... }}
1693 Now the outer case is not a pure seq, so [Improving seq] will happen,
1694 and then the inner case will disappear.
1696 The need for [Improving seq] showed up in Roman's experiments. Example:
1697 foo :: F Int -> Int -> Int
1698 foo t n = t `seq` bar n
1701 bar n = bar (n - case t of TI i -> i)
1702 Here we'd like to avoid repeated evaluating t inside the loop, by
1703 taking advantage of the `seq`.
1705 At one point I did transformation in LiberateCase, but it's more
1706 robust here. (Otherwise, there's a danger that we'll simply drop the
1707 'seq' altogether, before LiberateCase gets to see it.)
1710 simplAlts :: SimplEnv
1712 -> InId -- Case binder
1713 -> [InAlt] -- Non-empty
1715 -> SimplM (OutExpr, OutId, [OutAlt]) -- Includes the continuation
1716 -- Like simplExpr, this just returns the simplified alternatives;
1717 -- it does not return an environment
1719 simplAlts env scrut case_bndr alts cont'
1720 = -- pprTrace "simplAlts" (ppr alts $$ ppr (seTvSubst env)) $
1721 do { let env0 = zapFloats env
1723 ; (env1, case_bndr1) <- simplBinder env0 case_bndr
1725 ; fam_envs <- getFamEnvs
1726 ; (alt_env', scrut', case_bndr') <- improveSeq fam_envs env1 scrut
1727 case_bndr case_bndr1 alts
1729 ; (imposs_deflt_cons, in_alts) <- prepareAlts scrut' case_bndr' alts
1731 ; alts' <- mapM (simplAlt alt_env' imposs_deflt_cons case_bndr' cont') in_alts
1732 ; return (scrut', case_bndr', alts') }
1735 ------------------------------------
1736 improveSeq :: (FamInstEnv, FamInstEnv) -> SimplEnv
1737 -> OutExpr -> InId -> OutId -> [InAlt]
1738 -> SimplM (SimplEnv, OutExpr, OutId)
1739 -- Note [Improving seq]
1740 improveSeq fam_envs env scrut case_bndr case_bndr1 [(DEFAULT,_,_)]
1741 | not (isDeadBinder case_bndr) -- Not a pure seq! See the Note!
1742 , Just (co, ty2) <- topNormaliseType fam_envs (idType case_bndr1)
1743 = do { case_bndr2 <- newId (fsLit "nt") ty2
1744 ; let rhs = DoneEx (Var case_bndr2 `Cast` mkSymCoercion co)
1745 env2 = extendIdSubst env case_bndr rhs
1746 ; return (env2, scrut `Cast` co, case_bndr2) }
1748 improveSeq _ env scrut _ case_bndr1 _
1749 = return (env, scrut, case_bndr1)
1752 ------------------------------------
1753 simplAlt :: SimplEnv
1754 -> [AltCon] -- These constructors can't be present when
1755 -- matching the DEFAULT alternative
1756 -> OutId -- The case binder
1761 simplAlt env imposs_deflt_cons case_bndr' cont' (DEFAULT, bndrs, rhs)
1762 = ASSERT( null bndrs )
1763 do { let env' = addBinderOtherCon env case_bndr' imposs_deflt_cons
1764 -- Record the constructors that the case-binder *can't* be.
1765 ; rhs' <- simplExprC env' rhs cont'
1766 ; return (DEFAULT, [], rhs') }
1768 simplAlt env _ case_bndr' cont' (LitAlt lit, bndrs, rhs)
1769 = ASSERT( null bndrs )
1770 do { let env' = addBinderUnfolding env case_bndr' (Lit lit)
1771 ; rhs' <- simplExprC env' rhs cont'
1772 ; return (LitAlt lit, [], rhs') }
1774 simplAlt env _ case_bndr' cont' (DataAlt con, vs, rhs)
1775 = do { -- Deal with the pattern-bound variables
1776 -- Mark the ones that are in ! positions in the
1777 -- data constructor as certainly-evaluated.
1778 -- NB: simplLamBinders preserves this eval info
1779 let vs_with_evals = add_evals (dataConRepStrictness con)
1780 ; (env', vs') <- simplLamBndrs env vs_with_evals
1782 -- Bind the case-binder to (con args)
1783 ; let inst_tys' = tyConAppArgs (idType case_bndr')
1784 con_args = map Type inst_tys' ++ varsToCoreExprs vs'
1785 env'' = addBinderUnfolding env' case_bndr'
1786 (mkConApp con con_args)
1788 ; rhs' <- simplExprC env'' rhs cont'
1789 ; return (DataAlt con, vs', rhs') }
1791 -- add_evals records the evaluated-ness of the bound variables of
1792 -- a case pattern. This is *important*. Consider
1793 -- data T = T !Int !Int
1795 -- case x of { T a b -> T (a+1) b }
1797 -- We really must record that b is already evaluated so that we don't
1798 -- go and re-evaluate it when constructing the result.
1799 -- See Note [Data-con worker strictness] in MkId.lhs
1804 go (v:vs') strs | isTyCoVar v = v : go vs' strs
1805 go (v:vs') (str:strs)
1806 | isMarkedStrict str = evald_v : go vs' strs
1807 | otherwise = zapped_v : go vs' strs
1809 zapped_v = zap_occ_info v
1810 evald_v = zapped_v `setIdUnfolding` evaldUnfolding
1811 go _ _ = pprPanic "cat_evals" (ppr con $$ ppr vs $$ ppr the_strs)
1813 -- See Note [zapOccInfo]
1814 -- zap_occ_info: if the case binder is alive, then we add the unfolding
1816 -- to the envt; so vs are now very much alive
1817 -- Note [Aug06] I can't see why this actually matters, but it's neater
1818 -- case e of t { (a,b) -> ...(case t of (p,q) -> p)... }
1819 -- ==> case e of t { (a,b) -> ...(a)... }
1820 -- Look, Ma, a is alive now.
1821 zap_occ_info = zapCasePatIdOcc case_bndr'
1823 addBinderUnfolding :: SimplEnv -> Id -> CoreExpr -> SimplEnv
1824 addBinderUnfolding env bndr rhs
1825 = modifyInScope env (bndr `setIdUnfolding` mkSimpleUnfolding rhs)
1827 addBinderOtherCon :: SimplEnv -> Id -> [AltCon] -> SimplEnv
1828 addBinderOtherCon env bndr cons
1829 = modifyInScope env (bndr `setIdUnfolding` mkOtherCon cons)
1831 zapCasePatIdOcc :: Id -> Id -> Id
1832 -- Consider case e of b { (a,b) -> ... }
1833 -- Then if we bind b to (a,b) in "...", and b is not dead,
1834 -- then we must zap the deadness info on a,b
1835 zapCasePatIdOcc case_bndr
1836 | isDeadBinder case_bndr = \ pat_id -> pat_id
1837 | otherwise = \ pat_id -> zapIdOccInfo pat_id
1841 %************************************************************************
1843 \subsection{Known constructor}
1845 %************************************************************************
1847 We are a bit careful with occurrence info. Here's an example
1849 (\x* -> case x of (a*, b) -> f a) (h v, e)
1851 where the * means "occurs once". This effectively becomes
1852 case (h v, e) of (a*, b) -> f a)
1854 let a* = h v; b = e in f a
1858 All this should happen in one sweep.
1861 knownCon :: SimplEnv
1862 -> OutExpr -- The scrutinee
1863 -> DataCon -> [OutType] -> [OutExpr] -- The scrutinee (in pieces)
1864 -> InId -> [InBndr] -> InExpr -- The alternative
1866 -> SimplM (SimplEnv, OutExpr)
1868 knownCon env scrut dc dc_ty_args dc_args bndr bs rhs cont
1869 = do { env' <- bind_args env bs dc_args
1870 ; env'' <- bind_case_bndr env'
1871 ; simplExprF env'' rhs cont }
1873 zap_occ = zapCasePatIdOcc bndr -- bndr is an InId
1876 bind_args env' [] _ = return env'
1878 bind_args env' (b:bs') (Type ty : args)
1879 = ASSERT( isTyCoVar b )
1880 bind_args (extendTvSubst env' b ty) bs' args
1882 bind_args env' (b:bs') (arg : args)
1884 do { let b' = zap_occ b
1885 -- Note that the binder might be "dead", because it doesn't
1886 -- occur in the RHS; and simplNonRecX may therefore discard
1887 -- it via postInlineUnconditionally.
1888 -- Nevertheless we must keep it if the case-binder is alive,
1889 -- because it may be used in the con_app. See Note [zapOccInfo]
1890 ; env'' <- simplNonRecX env' b' arg
1891 ; bind_args env'' bs' args }
1894 pprPanic "bind_args" $ ppr dc $$ ppr bs $$ ppr dc_args $$
1895 text "scrut:" <+> ppr scrut
1897 -- It's useful to bind bndr to scrut, rather than to a fresh
1898 -- binding x = Con arg1 .. argn
1899 -- because very often the scrut is a variable, so we avoid
1900 -- creating, and then subsequently eliminating, a let-binding
1901 -- BUT, if scrut is a not a variable, we must be careful
1902 -- about duplicating the arg redexes; in that case, make
1903 -- a new con-app from the args
1905 | isDeadBinder bndr = return env
1906 | exprIsTrivial scrut = return (extendIdSubst env bndr (DoneEx scrut))
1907 | otherwise = do { dc_args <- mapM (simplVar env) bs
1908 -- dc_ty_args are aready OutTypes,
1909 -- but bs are InBndrs
1910 ; let con_app = Var (dataConWorkId dc)
1911 `mkTyApps` dc_ty_args
1913 ; simplNonRecX env bndr con_app }
1916 missingAlt :: SimplEnv -> Id -> [InAlt] -> SimplCont -> SimplM (SimplEnv, OutExpr)
1917 -- This isn't strictly an error, although it is unusual.
1918 -- It's possible that the simplifer might "see" that
1919 -- an inner case has no accessible alternatives before
1920 -- it "sees" that the entire branch of an outer case is
1921 -- inaccessible. So we simply put an error case here instead.
1922 missingAlt env case_bndr alts cont
1923 = WARN( True, ptext (sLit "missingAlt") <+> ppr case_bndr )
1924 return (env, mkImpossibleExpr res_ty)
1926 res_ty = contResultType env (substTy env (coreAltsType alts)) cont
1930 %************************************************************************
1932 \subsection{Duplicating continuations}
1934 %************************************************************************
1937 prepareCaseCont :: SimplEnv
1938 -> [InAlt] -> SimplCont
1939 -> SimplM (SimplEnv, SimplCont,SimplCont)
1940 -- Return a duplicatable continuation, a non-duplicable part
1941 -- plus some extra bindings (that scope over the entire
1944 -- No need to make it duplicatable if there's only one alternative
1945 prepareCaseCont env [_] cont = return (env, cont, mkBoringStop)
1946 prepareCaseCont env _ cont = mkDupableCont env cont
1950 mkDupableCont :: SimplEnv -> SimplCont
1951 -> SimplM (SimplEnv, SimplCont, SimplCont)
1953 mkDupableCont env cont
1954 | contIsDupable cont
1955 = return (env, cont, mkBoringStop)
1957 mkDupableCont _ (Stop {}) = panic "mkDupableCont" -- Handled by previous eqn
1959 mkDupableCont env (CoerceIt ty cont)
1960 = do { (env', dup, nodup) <- mkDupableCont env cont
1961 ; return (env', CoerceIt ty dup, nodup) }
1963 mkDupableCont env cont@(StrictBind {})
1964 = return (env, mkBoringStop, cont)
1965 -- See Note [Duplicating StrictBind]
1967 mkDupableCont env (StrictArg info cci cont)
1968 -- See Note [Duplicating StrictArg]
1969 = do { (env', dup, nodup) <- mkDupableCont env cont
1970 ; (env'', args') <- mapAccumLM (makeTrivial NotTopLevel) env' (ai_args info)
1971 ; return (env'', StrictArg (info { ai_args = args' }) cci dup, nodup) }
1973 mkDupableCont env (ApplyTo _ arg se cont)
1974 = -- e.g. [...hole...] (...arg...)
1976 -- let a = ...arg...
1977 -- in [...hole...] a
1978 do { (env', dup_cont, nodup_cont) <- mkDupableCont env cont
1979 ; arg' <- simplExpr (se `setInScope` env') arg
1980 ; (env'', arg'') <- makeTrivial NotTopLevel env' arg'
1981 ; let app_cont = ApplyTo OkToDup arg'' (zapSubstEnv env'') dup_cont
1982 ; return (env'', app_cont, nodup_cont) }
1984 mkDupableCont env cont@(Select _ case_bndr [(_, bs, _rhs)] _ _)
1985 -- See Note [Single-alternative case]
1986 -- | not (exprIsDupable rhs && contIsDupable case_cont)
1987 -- | not (isDeadBinder case_bndr)
1988 | all isDeadBinder bs -- InIds
1989 && not (isUnLiftedType (idType case_bndr))
1990 -- Note [Single-alternative-unlifted]
1991 = return (env, mkBoringStop, cont)
1993 mkDupableCont env (Select _ case_bndr alts se cont)
1994 = -- e.g. (case [...hole...] of { pi -> ei })
1996 -- let ji = \xij -> ei
1997 -- in case [...hole...] of { pi -> ji xij }
1998 do { tick (CaseOfCase case_bndr)
1999 ; (env', dup_cont, nodup_cont) <- mkDupableCont env cont
2000 -- NB: call mkDupableCont here, *not* prepareCaseCont
2001 -- We must make a duplicable continuation, whereas prepareCaseCont
2002 -- doesn't when there is a single case branch
2004 ; let alt_env = se `setInScope` env'
2005 ; (alt_env', case_bndr') <- simplBinder alt_env case_bndr
2006 ; alts' <- mapM (simplAlt alt_env' [] case_bndr' dup_cont) alts
2007 -- Safe to say that there are no handled-cons for the DEFAULT case
2008 -- NB: simplBinder does not zap deadness occ-info, so
2009 -- a dead case_bndr' will still advertise its deadness
2010 -- This is really important because in
2011 -- case e of b { (# p,q #) -> ... }
2012 -- b is always dead, and indeed we are not allowed to bind b to (# p,q #),
2013 -- which might happen if e was an explicit unboxed pair and b wasn't marked dead.
2014 -- In the new alts we build, we have the new case binder, so it must retain
2016 -- NB: we don't use alt_env further; it has the substEnv for
2017 -- the alternatives, and we don't want that
2019 ; (env'', alts'') <- mkDupableAlts env' case_bndr' alts'
2020 ; return (env'', -- Note [Duplicated env]
2021 Select OkToDup case_bndr' alts'' (zapSubstEnv env'') mkBoringStop,
2025 mkDupableAlts :: SimplEnv -> OutId -> [InAlt]
2026 -> SimplM (SimplEnv, [InAlt])
2027 -- Absorbs the continuation into the new alternatives
2029 mkDupableAlts env case_bndr' the_alts
2032 go env0 [] = return (env0, [])
2034 = do { (env1, alt') <- mkDupableAlt env0 case_bndr' alt
2035 ; (env2, alts') <- go env1 alts
2036 ; return (env2, alt' : alts' ) }
2038 mkDupableAlt :: SimplEnv -> OutId -> (AltCon, [CoreBndr], CoreExpr)
2039 -> SimplM (SimplEnv, (AltCon, [CoreBndr], CoreExpr))
2040 mkDupableAlt env case_bndr (con, bndrs', rhs')
2041 | exprIsDupable rhs' -- Note [Small alternative rhs]
2042 = return (env, (con, bndrs', rhs'))
2044 = do { let rhs_ty' = exprType rhs'
2045 scrut_ty = idType case_bndr
2048 DEFAULT -> case_bndr
2049 DataAlt dc -> setIdUnfolding case_bndr unf
2051 -- See Note [Case binders and join points]
2052 unf = mkInlineUnfolding Nothing rhs
2053 rhs = mkConApp dc (map Type (tyConAppArgs scrut_ty)
2054 ++ varsToCoreExprs bndrs')
2056 LitAlt {} -> WARN( True, ptext (sLit "mkDupableAlt")
2057 <+> ppr case_bndr <+> ppr con )
2059 -- The case binder is alive but trivial, so why has
2060 -- it not been substituted away?
2062 used_bndrs' | isDeadBinder case_bndr = filter abstract_over bndrs'
2063 | otherwise = bndrs' ++ [case_bndr_w_unf]
2066 | isTyCoVar bndr = True -- Abstract over all type variables just in case
2067 | otherwise = not (isDeadBinder bndr)
2068 -- The deadness info on the new Ids is preserved by simplBinders
2070 ; (final_bndrs', final_args) -- Note [Join point abstraction]
2071 <- if (any isId used_bndrs')
2072 then return (used_bndrs', varsToCoreExprs used_bndrs')
2073 else do { rw_id <- newId (fsLit "w") realWorldStatePrimTy
2074 ; return ([rw_id], [Var realWorldPrimId]) }
2076 ; join_bndr <- newId (fsLit "$j") (mkPiTypes final_bndrs' rhs_ty')
2077 -- Note [Funky mkPiTypes]
2079 ; let -- We make the lambdas into one-shot-lambdas. The
2080 -- join point is sure to be applied at most once, and doing so
2081 -- prevents the body of the join point being floated out by
2082 -- the full laziness pass
2083 really_final_bndrs = map one_shot final_bndrs'
2084 one_shot v | isId v = setOneShotLambda v
2086 join_rhs = mkLams really_final_bndrs rhs'
2087 join_call = mkApps (Var join_bndr) final_args
2089 ; env' <- addPolyBind NotTopLevel env (NonRec join_bndr join_rhs)
2090 ; return (env', (con, bndrs', join_call)) }
2091 -- See Note [Duplicated env]
2094 Note [Case binders and join points]
2095 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
2097 case (case .. ) of c {
2100 If we make a join point with c but not c# we get
2101 $j = \c -> ....c....
2103 But if later inlining scrutines the c, thus
2105 $j = \c -> ... case c of { I# y -> ... } ...
2107 we won't see that 'c' has already been scrutinised. This actually
2108 happens in the 'tabulate' function in wave4main, and makes a significant
2109 difference to allocation.
2111 An alternative plan is this:
2113 $j = \c# -> let c = I# c# in ...c....
2115 but that is bad if 'c' is *not* later scrutinised.
2117 So instead we do both: we pass 'c' and 'c#' , and record in c's inlining
2118 (an InlineRule) that it's really I# c#, thus
2120 $j = \c# -> \c[=I# c#] -> ...c....
2122 Absence analysis may later discard 'c'.
2124 NB: take great care when doing strictness analysis;
2125 see Note [Lamba-bound unfoldings] in DmdAnal.
2127 Also note that we can still end up passing stuff that isn't used. Before
2128 strictness analysis we have
2129 let $j x y c{=(x,y)} = (h c, ...)
2131 After strictness analysis we see that h is strict, we end up with
2132 let $j x y c{=(x,y)} = ($wh x y, ...)
2135 Note [Duplicated env]
2136 ~~~~~~~~~~~~~~~~~~~~~
2137 Some of the alternatives are simplified, but have not been turned into a join point
2138 So they *must* have an zapped subst-env. So we can't use completeNonRecX to
2139 bind the join point, because it might to do PostInlineUnconditionally, and
2140 we'd lose that when zapping the subst-env. We could have a per-alt subst-env,
2141 but zapping it (as we do in mkDupableCont, the Select case) is safe, and
2142 at worst delays the join-point inlining.
2144 Note [Small alternative rhs]
2145 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~
2146 It is worth checking for a small RHS because otherwise we
2147 get extra let bindings that may cause an extra iteration of the simplifier to
2148 inline back in place. Quite often the rhs is just a variable or constructor.
2149 The Ord instance of Maybe in PrelMaybe.lhs, for example, took several extra
2150 iterations because the version with the let bindings looked big, and so wasn't
2151 inlined, but after the join points had been inlined it looked smaller, and so
2154 NB: we have to check the size of rhs', not rhs.
2155 Duplicating a small InAlt might invalidate occurrence information
2156 However, if it *is* dupable, we return the *un* simplified alternative,
2157 because otherwise we'd need to pair it up with an empty subst-env....
2158 but we only have one env shared between all the alts.
2159 (Remember we must zap the subst-env before re-simplifying something).
2160 Rather than do this we simply agree to re-simplify the original (small) thing later.
2162 Note [Funky mkPiTypes]
2163 ~~~~~~~~~~~~~~~~~~~~~~
2164 Notice the funky mkPiTypes. If the contructor has existentials
2165 it's possible that the join point will be abstracted over
2166 type varaibles as well as term variables.
2167 Example: Suppose we have
2168 data T = forall t. C [t]
2170 case (case e of ...) of
2172 We get the join point
2173 let j :: forall t. [t] -> ...
2174 j = /\t \xs::[t] -> rhs
2176 case (case e of ...) of
2177 C t xs::[t] -> j t xs
2179 Note [Join point abstaction]
2180 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~
2181 If we try to lift a primitive-typed something out
2182 for let-binding-purposes, we will *caseify* it (!),
2183 with potentially-disastrous strictness results. So
2184 instead we turn it into a function: \v -> e
2185 where v::State# RealWorld#. The value passed to this function
2186 is realworld#, which generates (almost) no code.
2188 There's a slight infelicity here: we pass the overall
2189 case_bndr to all the join points if it's used in *any* RHS,
2190 because we don't know its usage in each RHS separately
2192 We used to say "&& isUnLiftedType rhs_ty'" here, but now
2193 we make the join point into a function whenever used_bndrs'
2194 is empty. This makes the join-point more CPR friendly.
2195 Consider: let j = if .. then I# 3 else I# 4
2196 in case .. of { A -> j; B -> j; C -> ... }
2198 Now CPR doesn't w/w j because it's a thunk, so
2199 that means that the enclosing function can't w/w either,
2200 which is a lose. Here's the example that happened in practice:
2201 kgmod :: Int -> Int -> Int
2202 kgmod x y = if x > 0 && y < 0 || x < 0 && y > 0
2206 I have seen a case alternative like this:
2208 It's a bit silly to add the realWorld dummy arg in this case, making
2211 (the \v alone is enough to make CPR happy) but I think it's rare
2213 Note [Duplicating StrictArg]
2214 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~
2215 The original plan had (where E is a big argument)
2217 ==> let $j = \a -> f E a
2220 But this is terrible! Here's an example:
2221 && E (case x of { T -> F; F -> T })
2222 Now, && is strict so we end up simplifying the case with
2223 an ArgOf continuation. If we let-bind it, we get
2224 let $j = \v -> && E v
2225 in simplExpr (case x of { T -> F; F -> T })
2227 And after simplifying more we get
2228 let $j = \v -> && E v
2229 in case x of { T -> $j F; F -> $j T }
2230 Which is a Very Bad Thing
2232 What we do now is this
2236 Now if the thing in the hole is a case expression (which is when
2237 we'll call mkDupableCont), we'll push the function call into the
2238 branches, which is what we want. Now RULES for f may fire, and
2239 call-pattern specialisation. Here's an example from Trac #3116
2242 _ -> Chunk p fpc (o+1) (l-1) bs')
2243 If we can push the call for 'go' inside the case, we get
2244 call-pattern specialisation for 'go', which is *crucial* for
2247 Here is the (&&) example:
2248 && E (case x of { T -> F; F -> T })
2250 case x of { T -> && a F; F -> && a T }
2254 * Arguments to f *after* the strict one are handled by
2255 the ApplyTo case of mkDupableCont. Eg
2258 * We can only do the let-binding of E because the function
2259 part of a StrictArg continuation is an explicit syntax
2260 tree. In earlier versions we represented it as a function
2261 (CoreExpr -> CoreEpxr) which we couldn't take apart.
2263 Do *not* duplicate StrictBind and StritArg continuations. We gain
2264 nothing by propagating them into the expressions, and we do lose a
2267 The desire not to duplicate is the entire reason that
2268 mkDupableCont returns a pair of continuations.
2270 Note [Duplicating StrictBind]
2271 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
2272 Unlike StrictArg, there doesn't seem anything to gain from
2273 duplicating a StrictBind continuation, so we don't.
2275 The desire not to duplicate is the entire reason that
2276 mkDupableCont returns a pair of continuations.
2279 Note [Single-alternative cases]
2280 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
2281 This case is just like the ArgOf case. Here's an example:
2285 case (case x of I# x' ->
2287 True -> I# (negate# x')
2288 False -> I# x') of y {
2290 Because the (case x) has only one alternative, we'll transform to
2292 case (case x' <# 0# of
2293 True -> I# (negate# x')
2294 False -> I# x') of y {
2296 But now we do *NOT* want to make a join point etc, giving
2298 let $j = \y -> MkT y
2300 True -> $j (I# (negate# x'))
2302 In this case the $j will inline again, but suppose there was a big
2303 strict computation enclosing the orginal call to MkT. Then, it won't
2304 "see" the MkT any more, because it's big and won't get duplicated.
2305 And, what is worse, nothing was gained by the case-of-case transform.
2307 So, in circumstances like these, we don't want to build join points
2308 and push the outer case into the branches of the inner one. Instead,
2309 don't duplicate the continuation.
2311 When should we use this strategy? We should not use it on *every*
2312 single-alternative case:
2313 e.g. case (case ....) of (a,b) -> (# a,b #)
2314 Here we must push the outer case into the inner one!
2317 * Match [(DEFAULT,_,_)], but in the common case of Int,
2318 the alternative-filling-in code turned the outer case into
2319 case (...) of y { I# _ -> MkT y }
2321 * Match on single alternative plus (not (isDeadBinder case_bndr))
2322 Rationale: pushing the case inwards won't eliminate the construction.
2323 But there's a risk of
2324 case (...) of y { (a,b) -> let z=(a,b) in ... }
2325 Now y looks dead, but it'll come alive again. Still, this
2326 seems like the best option at the moment.
2328 * Match on single alternative plus (all (isDeadBinder bndrs))
2329 Rationale: this is essentially seq.
2331 * Match when the rhs is *not* duplicable, and hence would lead to a
2332 join point. This catches the disaster-case above. We can test
2333 the *un-simplified* rhs, which is fine. It might get bigger or
2334 smaller after simplification; if it gets smaller, this case might
2335 fire next time round. NB also that we must test contIsDupable
2336 case_cont *too, because case_cont might be big!
2338 HOWEVER: I found that this version doesn't work well, because
2339 we can get let x = case (...) of { small } in ...case x...
2340 When x is inlined into its full context, we find that it was a bad
2341 idea to have pushed the outer case inside the (...) case.
2343 Note [Single-alternative-unlifted]
2344 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
2345 Here's another single-alternative where we really want to do case-of-case:
2353 case y_s6X of tpl_s7m {
2354 M1.Mk1 ipv_s70 -> ipv_s70;
2355 M1.Mk2 ipv_s72 -> ipv_s72;
2361 case x_s74 of tpl_s7n {
2362 M1.Mk1 ipv_s77 -> ipv_s77;
2363 M1.Mk2 ipv_s79 -> ipv_s79;
2367 { __DEFAULT -> ==# [wild1_s7b wild_s7c];
2371 So the outer case is doing *nothing at all*, other than serving as a
2372 join-point. In this case we really want to do case-of-case and decide
2373 whether to use a real join point or just duplicate the continuation.
2375 Hence: check whether the case binder's type is unlifted, because then
2376 the outer case is *not* a seq.