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 )
20 import FamInstEnv ( topNormaliseType )
21 import DataCon ( dataConRepStrictness, dataConUnivTyVars )
23 import NewDemand ( isStrictDmd )
24 import PprCore ( pprParendExpr, pprCoreExpr )
25 import CoreUnfold ( mkUnfolding, callSiteInline, CallCtxt(..) )
27 import Rules ( lookupRule )
28 import BasicTypes ( isMarkedStrict )
29 import CostCentre ( currentCCS )
30 import TysPrim ( realWorldStatePrimTy )
31 import PrelInfo ( realWorldPrimId )
32 import BasicTypes ( TopLevelFlag(..), isTopLevel,
33 RecFlag(..), isNonRuleLoopBreaker )
34 import Maybes ( orElse )
35 import Data.List ( mapAccumL )
40 The guts of the simplifier is in this module, but the driver loop for
41 the simplifier is in SimplCore.lhs.
44 -----------------------------------------
45 *** IMPORTANT NOTE ***
46 -----------------------------------------
47 The simplifier used to guarantee that the output had no shadowing, but
48 it does not do so any more. (Actually, it never did!) The reason is
49 documented with simplifyArgs.
52 -----------------------------------------
53 *** IMPORTANT NOTE ***
54 -----------------------------------------
55 Many parts of the simplifier return a bunch of "floats" as well as an
56 expression. This is wrapped as a datatype SimplUtils.FloatsWith.
58 All "floats" are let-binds, not case-binds, but some non-rec lets may
59 be unlifted (with RHS ok-for-speculation).
63 -----------------------------------------
64 ORGANISATION OF FUNCTIONS
65 -----------------------------------------
67 - simplify all top-level binders
68 - for NonRec, call simplRecOrTopPair
69 - for Rec, call simplRecBind
72 ------------------------------
73 simplExpr (applied lambda) ==> simplNonRecBind
74 simplExpr (Let (NonRec ...) ..) ==> simplNonRecBind
75 simplExpr (Let (Rec ...) ..) ==> simplify binders; simplRecBind
77 ------------------------------
78 simplRecBind [binders already simplfied]
79 - use simplRecOrTopPair on each pair in turn
81 simplRecOrTopPair [binder already simplified]
82 Used for: recursive bindings (top level and nested)
83 top-level non-recursive bindings
85 - check for PreInlineUnconditionally
89 Used for: non-top-level non-recursive bindings
90 beta reductions (which amount to the same thing)
91 Because it can deal with strict arts, it takes a
92 "thing-inside" and returns an expression
94 - check for PreInlineUnconditionally
95 - simplify binder, including its IdInfo
104 simplNonRecX: [given a *simplified* RHS, but an *unsimplified* binder]
105 Used for: binding case-binder and constr args in a known-constructor case
106 - check for PreInLineUnconditionally
110 ------------------------------
111 simplLazyBind: [binder already simplified, RHS not]
112 Used for: recursive bindings (top level and nested)
113 top-level non-recursive bindings
114 non-top-level, but *lazy* non-recursive bindings
115 [must not be strict or unboxed]
116 Returns floats + an augmented environment, not an expression
117 - substituteIdInfo and add result to in-scope
118 [so that rules are available in rec rhs]
121 - float if exposes constructor or PAP
125 completeNonRecX: [binder and rhs both simplified]
126 - if the the thing needs case binding (unlifted and not ok-for-spec)
132 completeBind: [given a simplified RHS]
133 [used for both rec and non-rec bindings, top level and not]
134 - try PostInlineUnconditionally
135 - add unfolding [this is the only place we add an unfolding]
140 Right hand sides and arguments
141 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
142 In many ways we want to treat
143 (a) the right hand side of a let(rec), and
144 (b) a function argument
145 in the same way. But not always! In particular, we would
146 like to leave these arguments exactly as they are, so they
147 will match a RULE more easily.
152 It's harder to make the rule match if we ANF-ise the constructor,
153 or eta-expand the PAP:
155 f (let { a = g x; b = h x } in (a,b))
158 On the other hand if we see the let-defns
163 then we *do* want to ANF-ise and eta-expand, so that p and q
164 can be safely inlined.
166 Even floating lets out is a bit dubious. For let RHS's we float lets
167 out if that exposes a value, so that the value can be inlined more vigorously.
170 r = let x = e in (x,x)
172 Here, if we float the let out we'll expose a nice constructor. We did experiments
173 that showed this to be a generally good thing. But it was a bad thing to float
174 lets out unconditionally, because that meant they got allocated more often.
176 For function arguments, there's less reason to expose a constructor (it won't
177 get inlined). Just possibly it might make a rule match, but I'm pretty skeptical.
178 So for the moment we don't float lets out of function arguments either.
183 For eta expansion, we want to catch things like
185 case e of (a,b) -> \x -> case a of (p,q) -> \y -> r
187 If the \x was on the RHS of a let, we'd eta expand to bring the two
188 lambdas together. And in general that's a good thing to do. Perhaps
189 we should eta expand wherever we find a (value) lambda? Then the eta
190 expansion at a let RHS can concentrate solely on the PAP case.
193 %************************************************************************
195 \subsection{Bindings}
197 %************************************************************************
200 simplTopBinds :: SimplEnv -> [InBind] -> SimplM [OutBind]
202 simplTopBinds env0 binds0
203 = do { -- Put all the top-level binders into scope at the start
204 -- so that if a transformation rule has unexpectedly brought
205 -- anything into scope, then we don't get a complaint about that.
206 -- It's rather as if the top-level binders were imported.
207 ; env1 <- simplRecBndrs env0 (bindersOfBinds binds0)
208 ; dflags <- getDOptsSmpl
209 ; let dump_flag = dopt Opt_D_dump_inlinings dflags ||
210 dopt Opt_D_dump_rule_firings dflags
211 ; env2 <- simpl_binds dump_flag env1 binds0
212 ; freeTick SimplifierDone
213 ; return (getFloats env2) }
215 -- We need to track the zapped top-level binders, because
216 -- they should have their fragile IdInfo zapped (notably occurrence info)
217 -- That's why we run down binds and bndrs' simultaneously.
219 -- The dump-flag emits a trace for each top-level binding, which
220 -- helps to locate the tracing for inlining and rule firing
221 simpl_binds :: Bool -> SimplEnv -> [InBind] -> SimplM SimplEnv
222 simpl_binds _ env [] = return env
223 simpl_binds dump env (bind:binds) = do { env' <- trace_bind dump bind $
225 ; simpl_binds dump env' binds }
227 trace_bind True bind = pprTrace "SimplBind" (ppr (bindersOf bind))
228 trace_bind False _ = \x -> x
230 simpl_bind env (Rec pairs) = simplRecBind env TopLevel pairs
231 simpl_bind env (NonRec b r) = simplRecOrTopPair env' TopLevel b b' r
233 (env', b') = addBndrRules env b (lookupRecBndr env b)
237 %************************************************************************
239 \subsection{Lazy bindings}
241 %************************************************************************
243 simplRecBind is used for
244 * recursive bindings only
247 simplRecBind :: SimplEnv -> TopLevelFlag
250 simplRecBind env0 top_lvl pairs0
251 = do { let (env_with_info, triples) = mapAccumL add_rules env0 pairs0
252 ; env1 <- go (zapFloats env_with_info) triples
253 ; return (env0 `addRecFloats` env1) }
254 -- addFloats adds the floats from env1,
255 -- *and* updates env0 with the in-scope set from env1
257 add_rules :: SimplEnv -> (InBndr,InExpr) -> (SimplEnv, (InBndr, OutBndr, InExpr))
258 -- Add the (substituted) rules to the binder
259 add_rules env (bndr, rhs) = (env', (bndr, bndr', rhs))
261 (env', bndr') = addBndrRules env bndr (lookupRecBndr env bndr)
263 go env [] = return env
265 go env ((old_bndr, new_bndr, rhs) : pairs)
266 = do { env' <- simplRecOrTopPair env top_lvl old_bndr new_bndr rhs
270 simplOrTopPair is used for
271 * recursive bindings (whether top level or not)
272 * top-level non-recursive bindings
274 It assumes the binder has already been simplified, but not its IdInfo.
277 simplRecOrTopPair :: SimplEnv
279 -> InId -> OutBndr -> InExpr -- Binder and rhs
280 -> SimplM SimplEnv -- Returns an env that includes the binding
282 simplRecOrTopPair env top_lvl old_bndr new_bndr rhs
283 | preInlineUnconditionally env top_lvl old_bndr rhs -- Check for unconditional inline
284 = do { tick (PreInlineUnconditionally old_bndr)
285 ; return (extendIdSubst env old_bndr (mkContEx env rhs)) }
288 = simplLazyBind env top_lvl Recursive old_bndr new_bndr rhs env
289 -- May not actually be recursive, but it doesn't matter
293 simplLazyBind is used for
294 * [simplRecOrTopPair] recursive bindings (whether top level or not)
295 * [simplRecOrTopPair] top-level non-recursive bindings
296 * [simplNonRecE] non-top-level *lazy* non-recursive bindings
299 1. It assumes that the binder is *already* simplified,
300 and is in scope, and its IdInfo too, except unfolding
302 2. It assumes that the binder type is lifted.
304 3. It does not check for pre-inline-unconditionallly;
305 that should have been done already.
308 simplLazyBind :: SimplEnv
309 -> TopLevelFlag -> RecFlag
310 -> InId -> OutId -- Binder, both pre-and post simpl
311 -- The OutId has IdInfo, except arity, unfolding
312 -> InExpr -> SimplEnv -- The RHS and its environment
315 simplLazyBind env top_lvl is_rec bndr bndr1 rhs rhs_se
316 = do { let rhs_env = rhs_se `setInScope` env
317 (tvs, body) = collectTyBinders rhs
318 ; (body_env, tvs') <- simplBinders rhs_env tvs
319 -- See Note [Floating and type abstraction]
322 -- Simplify the RHS; note the mkRhsStop, which tells
323 -- the simplifier that this is the RHS of a let.
324 ; let rhs_cont = mkRhsStop (applyTys (idType bndr1) (mkTyVarTys tvs'))
325 ; (body_env1, body1) <- simplExprF body_env body rhs_cont
327 -- ANF-ise a constructor or PAP rhs
328 ; (body_env2, body2) <- prepareRhs body_env1 body1
331 <- if not (doFloatFromRhs top_lvl is_rec False body2 body_env2)
332 then -- No floating, just wrap up!
333 do { rhs' <- mkLam tvs' (wrapFloats body_env2 body2)
334 ; return (env, rhs') }
336 else if null tvs then -- Simple floating
337 do { tick LetFloatFromLet
338 ; return (addFloats env body_env2, body2) }
340 else -- Do type-abstraction first
341 do { tick LetFloatFromLet
342 ; (poly_binds, body3) <- abstractFloats tvs' body_env2 body2
343 ; rhs' <- mkLam tvs' body3
344 ; return (extendFloats env poly_binds, rhs') }
346 ; completeBind env' top_lvl bndr bndr1 rhs' }
349 A specialised variant of simplNonRec used when the RHS is already simplified,
350 notably in knownCon. It uses case-binding where necessary.
353 simplNonRecX :: SimplEnv
354 -> InId -- Old binder
355 -> OutExpr -- Simplified RHS
358 simplNonRecX env bndr new_rhs
359 = do { (env', bndr') <- simplBinder env bndr
360 ; completeNonRecX env' NotTopLevel NonRecursive
361 (isStrictId bndr) bndr bndr' new_rhs }
363 completeNonRecX :: SimplEnv
364 -> TopLevelFlag -> RecFlag -> Bool
365 -> InId -- Old binder
366 -> OutId -- New binder
367 -> OutExpr -- Simplified RHS
370 completeNonRecX env top_lvl is_rec is_strict old_bndr new_bndr new_rhs
371 = do { (env1, rhs1) <- prepareRhs (zapFloats env) new_rhs
373 if doFloatFromRhs top_lvl is_rec is_strict rhs1 env1
374 then do { tick LetFloatFromLet
375 ; return (addFloats env env1, rhs1) } -- Add the floats to the main env
376 else return (env, wrapFloats env1 rhs1) -- Wrap the floats around the RHS
377 ; completeBind env2 NotTopLevel old_bndr new_bndr rhs2 }
380 {- No, no, no! Do not try preInlineUnconditionally in completeNonRecX
381 Doing so risks exponential behaviour, because new_rhs has been simplified once already
382 In the cases described by the folowing commment, postInlineUnconditionally will
383 catch many of the relevant cases.
384 -- This happens; for example, the case_bndr during case of
385 -- known constructor: case (a,b) of x { (p,q) -> ... }
386 -- Here x isn't mentioned in the RHS, so we don't want to
387 -- create the (dead) let-binding let x = (a,b) in ...
389 -- Similarly, single occurrences can be inlined vigourously
390 -- e.g. case (f x, g y) of (a,b) -> ....
391 -- If a,b occur once we can avoid constructing the let binding for them.
393 Furthermore in the case-binding case preInlineUnconditionally risks extra thunks
394 -- Consider case I# (quotInt# x y) of
395 -- I# v -> let w = J# v in ...
396 -- If we gaily inline (quotInt# x y) for v, we end up building an
398 -- let w = J# (quotInt# x y) in ...
399 -- because quotInt# can fail.
401 | preInlineUnconditionally env NotTopLevel bndr new_rhs
402 = thing_inside (extendIdSubst env bndr (DoneEx new_rhs))
405 ----------------------------------
406 prepareRhs takes a putative RHS, checks whether it's a PAP or
407 constructor application and, if so, converts it to ANF, so that the
408 resulting thing can be inlined more easily. Thus
415 We also want to deal well cases like this
416 v = (f e1 `cast` co) e2
417 Here we want to make e1,e2 trivial and get
418 x1 = e1; x2 = e2; v = (f x1 `cast` co) v2
419 That's what the 'go' loop in prepareRhs does
422 prepareRhs :: SimplEnv -> OutExpr -> SimplM (SimplEnv, OutExpr)
423 -- Adds new floats to the env iff that allows us to return a good RHS
424 prepareRhs env (Cast rhs co) -- Note [Float coercions]
425 | (ty1, _ty2) <- coercionKind co -- Do *not* do this if rhs has an unlifted type
426 , not (isUnLiftedType ty1) -- see Note [Float coercions (unlifted)]
427 = do { (env', rhs') <- makeTrivial env rhs
428 ; return (env', Cast rhs' co) }
431 = do { (_is_val, env1, rhs1) <- go 0 env0 rhs0
432 ; return (env1, rhs1) }
434 go n_val_args env (Cast rhs co)
435 = do { (is_val, env', rhs') <- go n_val_args env rhs
436 ; return (is_val, env', Cast rhs' co) }
437 go n_val_args env (App fun (Type ty))
438 = do { (is_val, env', rhs') <- go n_val_args env fun
439 ; return (is_val, env', App rhs' (Type ty)) }
440 go n_val_args env (App fun arg)
441 = do { (is_val, env', fun') <- go (n_val_args+1) env fun
443 True -> do { (env'', arg') <- makeTrivial env' arg
444 ; return (True, env'', App fun' arg') }
445 False -> return (False, env, App fun arg) }
446 go n_val_args env (Var fun)
447 = return (is_val, env, Var fun)
449 is_val = n_val_args > 0 -- There is at least one arg
450 -- ...and the fun a constructor or PAP
451 && (isDataConWorkId fun || n_val_args < idArity fun)
453 = return (False, env, other)
457 Note [Float coercions]
458 ~~~~~~~~~~~~~~~~~~~~~~
459 When we find the binding
461 we'd like to transform it to
463 x = x `cast` co -- A trivial binding
464 There's a chance that e will be a constructor application or function, or something
465 like that, so moving the coerion to the usage site may well cancel the coersions
466 and lead to further optimisation. Example:
469 data instance T Int = T Int
471 foo :: Int -> Int -> Int
476 go n = case x of { T m -> go (n-m) }
477 -- This case should optimise
479 Note [Float coercions (unlifted)]
480 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
481 BUT don't do [Float coercions] if 'e' has an unlifted type.
484 foo :: Int = (error (# Int,Int #) "urk")
485 `cast` CoUnsafe (# Int,Int #) Int
487 If do the makeTrivial thing to the error call, we'll get
488 foo = case error (# Int,Int #) "urk" of v -> v `cast` ...
489 But 'v' isn't in scope!
491 These strange casts can happen as a result of case-of-case
492 bar = case (case x of { T -> (# 2,3 #); F -> error "urk" }) of
497 makeTrivial :: SimplEnv -> OutExpr -> SimplM (SimplEnv, OutExpr)
498 -- Binds the expression to a variable, if it's not trivial, returning the variable
502 | otherwise -- See Note [Take care] below
503 = do { var <- newId FSLIT("a") (exprType expr)
504 ; env' <- completeNonRecX env NotTopLevel NonRecursive
506 ; return (env', substExpr env' (Var var)) }
510 %************************************************************************
512 \subsection{Completing a lazy binding}
514 %************************************************************************
517 * deals only with Ids, not TyVars
518 * takes an already-simplified binder and RHS
519 * is used for both recursive and non-recursive bindings
520 * is used for both top-level and non-top-level bindings
522 It does the following:
523 - tries discarding a dead binding
524 - tries PostInlineUnconditionally
525 - add unfolding [this is the only place we add an unfolding]
528 It does *not* attempt to do let-to-case. Why? Because it is used for
529 - top-level bindings (when let-to-case is impossible)
530 - many situations where the "rhs" is known to be a WHNF
531 (so let-to-case is inappropriate).
533 Nor does it do the atomic-argument thing
536 completeBind :: SimplEnv
537 -> TopLevelFlag -- Flag stuck into unfolding
538 -> InId -- Old binder
539 -> OutId -> OutExpr -- New binder and RHS
541 -- completeBind may choose to do its work
542 -- * by extending the substitution (e.g. let x = y in ...)
543 -- * or by adding to the floats in the envt
545 completeBind env top_lvl old_bndr new_bndr new_rhs
546 | postInlineUnconditionally env top_lvl new_bndr occ_info new_rhs unfolding
547 -- Inline and discard the binding
548 = do { tick (PostInlineUnconditionally old_bndr)
549 ; -- pprTrace "postInlineUnconditionally" (ppr old_bndr <+> ppr new_bndr <+> ppr new_rhs) $
550 return (extendIdSubst env old_bndr (DoneEx new_rhs)) }
551 -- Use the substitution to make quite, quite sure that the
552 -- substitution will happen, since we are going to discard the binding
557 new_bndr_info = idInfo new_bndr `setArityInfo` exprArity new_rhs
560 -- Add the unfolding *only* for non-loop-breakers
561 -- Making loop breakers not have an unfolding at all
562 -- means that we can avoid tests in exprIsConApp, for example.
563 -- This is important: if exprIsConApp says 'yes' for a recursive
564 -- thing, then we can get into an infinite loop
567 -- If the unfolding is a value, the demand info may
568 -- go pear-shaped, so we nuke it. Example:
570 -- case x of (p,q) -> h p q x
571 -- Here x is certainly demanded. But after we've nuked
572 -- the case, we'll get just
573 -- let x = (a,b) in h a b x
574 -- and now x is not demanded (I'm assuming h is lazy)
575 -- This really happens. Similarly
576 -- let f = \x -> e in ...f..f...
577 -- After inlining f at some of its call sites the original binding may
578 -- (for example) be no longer strictly demanded.
579 -- The solution here is a bit ad hoc...
580 info_w_unf = new_bndr_info `setUnfoldingInfo` unfolding
581 `setWorkerInfo` worker_info
583 final_info | loop_breaker = new_bndr_info
584 | isEvaldUnfolding unfolding = zapDemandInfo info_w_unf `orElse` info_w_unf
585 | otherwise = info_w_unf
587 final_id = new_bndr `setIdInfo` final_info
589 -- These seqs forces the Id, and hence its IdInfo,
590 -- and hence any inner substitutions
592 -- pprTrace "Binding" (ppr final_id <+> ppr unfolding) $
593 return (addNonRec env final_id new_rhs)
595 unfolding = mkUnfolding (isTopLevel top_lvl) new_rhs
596 worker_info = substWorker env (workerInfo old_info)
597 loop_breaker = isNonRuleLoopBreaker occ_info
598 old_info = idInfo old_bndr
599 occ_info = occInfo old_info
604 %************************************************************************
606 \subsection[Simplify-simplExpr]{The main function: simplExpr}
608 %************************************************************************
610 The reason for this OutExprStuff stuff is that we want to float *after*
611 simplifying a RHS, not before. If we do so naively we get quadratic
612 behaviour as things float out.
614 To see why it's important to do it after, consider this (real) example:
628 a -- Can't inline a this round, cos it appears twice
632 Each of the ==> steps is a round of simplification. We'd save a
633 whole round if we float first. This can cascade. Consider
638 let f = let d1 = ..d.. in \y -> e
642 in \x -> ...(\y ->e)...
644 Only in this second round can the \y be applied, and it
645 might do the same again.
649 simplExpr :: SimplEnv -> CoreExpr -> SimplM CoreExpr
650 simplExpr env expr = simplExprC env expr (mkBoringStop expr_ty')
652 expr_ty' = substTy env (exprType expr)
653 -- The type in the Stop continuation, expr_ty', is usually not used
654 -- It's only needed when discarding continuations after finding
655 -- a function that returns bottom.
656 -- Hence the lazy substitution
659 simplExprC :: SimplEnv -> CoreExpr -> SimplCont -> SimplM CoreExpr
660 -- Simplify an expression, given a continuation
661 simplExprC env expr cont
662 = -- pprTrace "simplExprC" (ppr expr $$ ppr cont {- $$ ppr (seIdSubst env) -} $$ ppr (seFloats env) ) $
663 do { (env', expr') <- simplExprF (zapFloats env) expr cont
664 ; -- pprTrace "simplExprC ret" (ppr expr $$ ppr expr') $
665 -- pprTrace "simplExprC ret3" (ppr (seInScope env')) $
666 -- pprTrace "simplExprC ret4" (ppr (seFloats env')) $
667 return (wrapFloats env' expr') }
669 --------------------------------------------------
670 simplExprF :: SimplEnv -> InExpr -> SimplCont
671 -> SimplM (SimplEnv, OutExpr)
673 simplExprF env e cont
674 = -- pprTrace "simplExprF" (ppr e $$ ppr cont $$ ppr (seTvSubst env) $$ ppr (seIdSubst env) {- $$ ppr (seFloats env) -} ) $
675 simplExprF' env e cont
677 simplExprF' :: SimplEnv -> InExpr -> SimplCont
678 -> SimplM (SimplEnv, OutExpr)
679 simplExprF' env (Var v) cont = simplVar env v cont
680 simplExprF' env (Lit lit) cont = rebuild env (Lit lit) cont
681 simplExprF' env (Note n expr) cont = simplNote env n expr cont
682 simplExprF' env (Cast body co) cont = simplCast env body co cont
683 simplExprF' env (App fun arg) cont = simplExprF env fun $
684 ApplyTo NoDup arg env cont
686 simplExprF' env expr@(Lam _ _) cont
687 = simplLam env (map zap bndrs) body cont
688 -- The main issue here is under-saturated lambdas
689 -- (\x1. \x2. e) arg1
690 -- Here x1 might have "occurs-once" occ-info, because occ-info
691 -- is computed assuming that a group of lambdas is applied
692 -- all at once. If there are too few args, we must zap the
695 n_args = countArgs cont
696 n_params = length bndrs
697 (bndrs, body) = collectBinders expr
698 zap | n_args >= n_params = \b -> b
699 | otherwise = \b -> if isTyVar b then b
701 -- NB: we count all the args incl type args
702 -- so we must count all the binders (incl type lambdas)
704 simplExprF' env (Type ty) cont
705 = ASSERT( contIsRhsOrArg cont )
706 do { ty' <- simplType env ty
707 ; rebuild env (Type ty') cont }
709 simplExprF' env (Case scrut bndr case_ty alts) cont
710 | not (switchIsOn (getSwitchChecker env) NoCaseOfCase)
711 = -- Simplify the scrutinee with a Select continuation
712 simplExprF env scrut (Select NoDup bndr alts env cont)
715 = -- If case-of-case is off, simply simplify the case expression
716 -- in a vanilla Stop context, and rebuild the result around it
717 do { case_expr' <- simplExprC env scrut case_cont
718 ; rebuild env case_expr' cont }
720 case_cont = Select NoDup bndr alts env (mkBoringStop case_ty')
721 case_ty' = substTy env case_ty -- c.f. defn of simplExpr
723 simplExprF' env (Let (Rec pairs) body) cont
724 = do { env' <- simplRecBndrs env (map fst pairs)
725 -- NB: bndrs' don't have unfoldings or rules
726 -- We add them as we go down
728 ; env'' <- simplRecBind env' NotTopLevel pairs
729 ; simplExprF env'' body cont }
731 simplExprF' env (Let (NonRec bndr rhs) body) cont
732 = simplNonRecE env bndr (rhs, env) ([], body) cont
734 ---------------------------------
735 simplType :: SimplEnv -> InType -> SimplM OutType
736 -- Kept monadic just so we can do the seqType
738 = -- pprTrace "simplType" (ppr ty $$ ppr (seTvSubst env)) $
739 seqType new_ty `seq` return new_ty
741 new_ty = substTy env ty
745 %************************************************************************
747 \subsection{The main rebuilder}
749 %************************************************************************
752 rebuild :: SimplEnv -> OutExpr -> SimplCont -> SimplM (SimplEnv, OutExpr)
753 -- At this point the substitution in the SimplEnv should be irrelevant
754 -- only the in-scope set and floats should matter
755 rebuild env expr cont0
756 = -- pprTrace "rebuild" (ppr expr $$ ppr cont0 $$ ppr (seFloats env)) $
758 Stop {} -> return (env, expr)
759 CoerceIt co cont -> rebuild env (mkCoerce co expr) cont
760 Select _ bndr alts se cont -> rebuildCase (se `setFloats` env) expr bndr alts cont
761 StrictArg fun ty _ info cont -> rebuildCall env (fun `App` expr) (funResultTy ty) info cont
762 StrictBind b bs body se cont -> do { env' <- simplNonRecX (se `setFloats` env) b expr
763 ; simplLam env' bs body cont }
764 ApplyTo _ arg se cont -> do { arg' <- simplExpr (se `setInScope` env) arg
765 ; rebuild env (App expr arg') cont }
769 %************************************************************************
773 %************************************************************************
776 simplCast :: SimplEnv -> InExpr -> Coercion -> SimplCont
777 -> SimplM (SimplEnv, OutExpr)
778 simplCast env body co0 cont0
779 = do { co1 <- simplType env co0
780 ; simplExprF env body (addCoerce co1 cont0) }
782 addCoerce co cont = add_coerce co (coercionKind co) cont
784 add_coerce _co (s1, k1) cont -- co :: ty~ty
785 | s1 `coreEqType` k1 = cont -- is a no-op
787 add_coerce co1 (s1, _k2) (CoerceIt co2 cont)
788 | (_l1, t1) <- coercionKind co2
789 -- coerce T1 S1 (coerce S1 K1 e)
792 -- coerce T1 K1 e, otherwise
794 -- For example, in the initial form of a worker
795 -- we may find (coerce T (coerce S (\x.e))) y
796 -- and we'd like it to simplify to e[y/x] in one round
798 , s1 `coreEqType` t1 = cont -- The coerces cancel out
799 | otherwise = CoerceIt (mkTransCoercion co1 co2) cont
801 add_coerce co (s1s2, _t1t2) (ApplyTo dup (Type arg_ty) arg_se cont)
802 -- (f `cast` g) ty ---> (f ty) `cast` (g @ ty)
803 -- This implements the PushT rule from the paper
804 | Just (tyvar,_) <- splitForAllTy_maybe s1s2
805 , not (isCoVar tyvar)
806 = ApplyTo dup (Type ty') (zapSubstEnv env) (addCoerce (mkInstCoercion co ty') cont)
808 ty' = substTy arg_se arg_ty
810 -- ToDo: the PushC rule is not implemented at all
812 add_coerce co (s1s2, _t1t2) (ApplyTo dup arg arg_se cont)
813 | not (isTypeArg arg) -- This implements the Push rule from the paper
814 , isFunTy s1s2 -- t1t2 must be a function type, becuase it's applied
815 -- co : s1s2 :=: t1t2
816 -- (coerce (T1->T2) (S1->S2) F) E
818 -- coerce T2 S2 (F (coerce S1 T1 E))
820 -- t1t2 must be a function type, T1->T2, because it's applied
821 -- to something but s1s2 might conceivably not be
823 -- When we build the ApplyTo we can't mix the out-types
824 -- with the InExpr in the argument, so we simply substitute
825 -- to make it all consistent. It's a bit messy.
826 -- But it isn't a common case.
828 -- Example of use: Trac #995
829 = ApplyTo dup new_arg (zapSubstEnv env) (addCoerce co2 cont)
831 -- we split coercion t1->t2 :=: s1->s2 into t1 :=: s1 and
832 -- t2 :=: s2 with left and right on the curried form:
833 -- (->) t1 t2 :=: (->) s1 s2
834 [co1, co2] = decomposeCo 2 co
835 new_arg = mkCoerce (mkSymCoercion co1) arg'
836 arg' = substExpr arg_se arg
838 add_coerce co _ cont = CoerceIt co cont
842 %************************************************************************
846 %************************************************************************
849 simplLam :: SimplEnv -> [InId] -> InExpr -> SimplCont
850 -> SimplM (SimplEnv, OutExpr)
852 simplLam env [] body cont = simplExprF env body cont
854 -- Type-beta reduction
855 simplLam env (bndr:bndrs) body (ApplyTo _ (Type ty_arg) arg_se cont)
856 = ASSERT( isTyVar bndr )
857 do { tick (BetaReduction bndr)
858 ; ty_arg' <- simplType (arg_se `setInScope` env) ty_arg
859 ; simplLam (extendTvSubst env bndr ty_arg') bndrs body cont }
861 -- Ordinary beta reduction
862 simplLam env (bndr:bndrs) body (ApplyTo _ arg arg_se cont)
863 = do { tick (BetaReduction bndr)
864 ; simplNonRecE env bndr (arg, arg_se) (bndrs, body) cont }
866 -- Not enough args, so there are real lambdas left to put in the result
867 simplLam env bndrs body cont
868 = do { (env', bndrs') <- simplLamBndrs env bndrs
869 ; body' <- simplExpr env' body
870 ; new_lam <- mkLam bndrs' body'
871 ; rebuild env' new_lam cont }
874 simplNonRecE :: SimplEnv
875 -> InId -- The binder
876 -> (InExpr, SimplEnv) -- Rhs of binding (or arg of lambda)
877 -> ([InId], InExpr) -- Body of the let/lambda
880 -> SimplM (SimplEnv, OutExpr)
882 -- simplNonRecE is used for
883 -- * non-top-level non-recursive lets in expressions
886 -- It deals with strict bindings, via the StrictBind continuation,
887 -- which may abort the whole process
889 -- The "body" of the binding comes as a pair of ([InId],InExpr)
890 -- representing a lambda; so we recurse back to simplLam
891 -- Why? Because of the binder-occ-info-zapping done before
892 -- the call to simplLam in simplExprF (Lam ...)
894 simplNonRecE env bndr (rhs, rhs_se) (bndrs, body) cont
895 | preInlineUnconditionally env NotTopLevel bndr rhs
896 = do { tick (PreInlineUnconditionally bndr)
897 ; simplLam (extendIdSubst env bndr (mkContEx rhs_se rhs)) bndrs body cont }
900 = do { simplExprF (rhs_se `setFloats` env) rhs
901 (StrictBind bndr bndrs body env cont) }
904 = do { (env1, bndr1) <- simplNonRecBndr env bndr
905 ; let (env2, bndr2) = addBndrRules env1 bndr bndr1
906 ; env3 <- simplLazyBind env2 NotTopLevel NonRecursive bndr bndr2 rhs rhs_se
907 ; simplLam env3 bndrs body cont }
911 %************************************************************************
915 %************************************************************************
918 -- Hack alert: we only distinguish subsumed cost centre stacks for the
919 -- purposes of inlining. All other CCCSs are mapped to currentCCS.
920 simplNote :: SimplEnv -> Note -> CoreExpr -> SimplCont
921 -> SimplM (SimplEnv, OutExpr)
922 simplNote env (SCC cc) e cont
923 = do { e' <- simplExpr (setEnclosingCC env currentCCS) e
924 ; rebuild env (mkSCC cc e') cont }
926 -- See notes with SimplMonad.inlineMode
927 simplNote env InlineMe e cont
928 | Just (inside, outside) <- splitInlineCont cont -- Boring boring continuation; see notes above
929 = do { -- Don't inline inside an INLINE expression
930 e' <- simplExprC (setMode inlineMode env) e inside
931 ; rebuild env (mkInlineMe e') outside }
933 | otherwise -- Dissolve the InlineMe note if there's
934 -- an interesting context of any kind to combine with
935 -- (even a type application -- anything except Stop)
936 = simplExprF env e cont
938 simplNote env (CoreNote s) e cont = do
939 e' <- simplExpr env e
940 rebuild env (Note (CoreNote s) e') cont
944 %************************************************************************
946 \subsection{Dealing with calls}
948 %************************************************************************
951 simplVar :: SimplEnv -> Id -> SimplCont -> SimplM (SimplEnv, OutExpr)
952 simplVar env var cont
953 = case substId env var of
954 DoneEx e -> simplExprF (zapSubstEnv env) e cont
955 ContEx tvs ids e -> simplExprF (setSubstEnv env tvs ids) e cont
956 DoneId var1 -> completeCall (zapSubstEnv env) var1 cont
957 -- Note [zapSubstEnv]
958 -- The template is already simplified, so don't re-substitute.
959 -- This is VITAL. Consider
961 -- let y = \z -> ...x... in
963 -- We'll clone the inner \x, adding x->x' in the id_subst
964 -- Then when we inline y, we must *not* replace x by x' in
965 -- the inlined copy!!
967 ---------------------------------------------------------
968 -- Dealing with a call site
970 completeCall :: SimplEnv -> Id -> SimplCont -> SimplM (SimplEnv, OutExpr)
971 completeCall env var cont
972 = do { dflags <- getDOptsSmpl
973 ; let (args,call_cont) = contArgs cont
974 -- The args are OutExprs, obtained by *lazily* substituting
975 -- in the args found in cont. These args are only examined
976 -- to limited depth (unless a rule fires). But we must do
977 -- the substitution; rule matching on un-simplified args would
980 ------------- First try rules ----------------
981 -- Do this before trying inlining. Some functions have
982 -- rules *and* are strict; in this case, we don't want to
983 -- inline the wrapper of the non-specialised thing; better
984 -- to call the specialised thing instead.
986 -- We used to use the black-listing mechanism to ensure that inlining of
987 -- the wrapper didn't occur for things that have specialisations till a
988 -- later phase, so but now we just try RULES first
990 -- Note [Rules for recursive functions]
991 -- ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
992 -- You might think that we shouldn't apply rules for a loop breaker:
993 -- doing so might give rise to an infinite loop, because a RULE is
994 -- rather like an extra equation for the function:
995 -- RULE: f (g x) y = x+y
998 -- But it's too drastic to disable rules for loop breakers.
999 -- Even the foldr/build rule would be disabled, because foldr
1000 -- is recursive, and hence a loop breaker:
1001 -- foldr k z (build g) = g k z
1002 -- So it's up to the programmer: rules can cause divergence
1004 ; let in_scope = getInScope env
1005 maybe_rule = case activeRule dflags env of
1006 Nothing -> Nothing -- No rules apply
1007 Just act_fn -> lookupRule act_fn in_scope
1009 ; case maybe_rule of {
1010 Just (rule, rule_rhs) -> do
1011 tick (RuleFired (ru_name rule))
1012 (if dopt Opt_D_dump_rule_firings dflags then
1013 pprTrace "Rule fired" (vcat [
1014 text "Rule:" <+> ftext (ru_name rule),
1015 text "Before:" <+> ppr var <+> sep (map pprParendExpr args),
1016 text "After: " <+> pprCoreExpr rule_rhs,
1017 text "Cont: " <+> ppr call_cont])
1020 simplExprF env rule_rhs (dropArgs (ruleArity rule) cont)
1021 -- The ruleArity says how many args the rule consumed
1023 ; Nothing -> do -- No rules
1025 ------------- Next try inlining ----------------
1026 { let arg_infos = [interestingArg arg | arg <- args, isValArg arg]
1027 n_val_args = length arg_infos
1028 interesting_cont = interestingCallContext call_cont
1029 active_inline = activeInline env var
1030 maybe_inline = callSiteInline dflags active_inline var
1031 (null args) arg_infos interesting_cont
1032 ; case maybe_inline of {
1033 Just unfolding -- There is an inlining!
1034 -> do { tick (UnfoldingDone var)
1035 ; (if dopt Opt_D_dump_inlinings dflags then
1036 pprTrace ("Inlining done" ++ showSDoc (ppr var)) (vcat [
1037 text "Before:" <+> ppr var <+> sep (map pprParendExpr args),
1038 text "Inlined fn: " <+> nest 2 (ppr unfolding),
1039 text "Cont: " <+> ppr call_cont])
1042 simplExprF env unfolding cont }
1044 ; Nothing -> -- No inlining!
1046 ------------- No inlining! ----------------
1047 -- Next, look for rules or specialisations that match
1049 rebuildCall env (Var var) (idType var)
1050 (mkArgInfo var n_val_args call_cont) cont
1053 rebuildCall :: SimplEnv
1054 -> OutExpr -> OutType -- Function and its type
1057 -> SimplM (SimplEnv, OutExpr)
1058 rebuildCall env fun fun_ty (ArgInfo { ai_strs = [] }) cont
1059 -- When we run out of strictness args, it means
1060 -- that the call is definitely bottom; see SimplUtils.mkArgInfo
1061 -- Then we want to discard the entire strict continuation. E.g.
1062 -- * case (error "hello") of { ... }
1063 -- * (error "Hello") arg
1064 -- * f (error "Hello") where f is strict
1066 -- Then, especially in the first of these cases, we'd like to discard
1067 -- the continuation, leaving just the bottoming expression. But the
1068 -- type might not be right, so we may have to add a coerce.
1069 | not (contIsTrivial cont) -- Only do this if there is a non-trivial
1070 = return (env, mk_coerce fun) -- contination to discard, else we do it
1071 where -- again and again!
1072 cont_ty = contResultType cont
1073 co = mkUnsafeCoercion fun_ty cont_ty
1074 mk_coerce expr | cont_ty `coreEqType` fun_ty = expr
1075 | otherwise = mkCoerce co expr
1077 rebuildCall env fun fun_ty info (ApplyTo _ (Type arg_ty) se cont)
1078 = do { ty' <- simplType (se `setInScope` env) arg_ty
1079 ; rebuildCall env (fun `App` Type ty') (applyTy fun_ty ty') info cont }
1081 rebuildCall env fun fun_ty
1082 (ArgInfo { ai_rules = has_rules, ai_strs = str:strs, ai_discs = disc:discs })
1083 (ApplyTo _ arg arg_se cont)
1084 | str || isStrictType arg_ty -- Strict argument
1085 = -- pprTrace "Strict Arg" (ppr arg $$ ppr (seIdSubst env) $$ ppr (seInScope env)) $
1086 simplExprF (arg_se `setFloats` env) arg
1087 (StrictArg fun fun_ty cci arg_info' cont)
1090 | otherwise -- Lazy argument
1091 -- DO NOT float anything outside, hence simplExprC
1092 -- There is no benefit (unlike in a let-binding), and we'd
1093 -- have to be very careful about bogus strictness through
1094 -- floating a demanded let.
1095 = do { arg' <- simplExprC (arg_se `setInScope` env) arg
1096 (mkLazyArgStop arg_ty cci)
1097 ; rebuildCall env (fun `App` arg') res_ty arg_info' cont }
1099 (arg_ty, res_ty) = splitFunTy fun_ty
1100 arg_info' = ArgInfo { ai_rules = has_rules, ai_strs = strs, ai_discs = discs }
1101 cci | has_rules || disc > 0 = ArgCtxt has_rules disc -- Be keener here
1102 | otherwise = BoringCtxt -- Nothing interesting
1104 rebuildCall env fun _ _ cont
1105 = rebuild env fun cont
1110 This part of the simplifier may break the no-shadowing invariant
1112 f (...(\a -> e)...) (case y of (a,b) -> e')
1113 where f is strict in its second arg
1114 If we simplify the innermost one first we get (...(\a -> e)...)
1115 Simplifying the second arg makes us float the case out, so we end up with
1116 case y of (a,b) -> f (...(\a -> e)...) e'
1117 So the output does not have the no-shadowing invariant. However, there is
1118 no danger of getting name-capture, because when the first arg was simplified
1119 we used an in-scope set that at least mentioned all the variables free in its
1120 static environment, and that is enough.
1122 We can't just do innermost first, or we'd end up with a dual problem:
1123 case x of (a,b) -> f e (...(\a -> e')...)
1125 I spent hours trying to recover the no-shadowing invariant, but I just could
1126 not think of an elegant way to do it. The simplifier is already knee-deep in
1127 continuations. We have to keep the right in-scope set around; AND we have
1128 to get the effect that finding (error "foo") in a strict arg position will
1129 discard the entire application and replace it with (error "foo"). Getting
1130 all this at once is TOO HARD!
1132 %************************************************************************
1134 Rebuilding a cse expression
1136 %************************************************************************
1138 Blob of helper functions for the "case-of-something-else" situation.
1141 ---------------------------------------------------------
1142 -- Eliminate the case if possible
1144 rebuildCase :: SimplEnv
1145 -> OutExpr -- Scrutinee
1146 -> InId -- Case binder
1147 -> [InAlt] -- Alternatives (inceasing order)
1149 -> SimplM (SimplEnv, OutExpr)
1151 --------------------------------------------------
1152 -- 1. Eliminate the case if there's a known constructor
1153 --------------------------------------------------
1155 rebuildCase env scrut case_bndr alts cont
1156 | Just (con,args) <- exprIsConApp_maybe scrut
1157 -- Works when the scrutinee is a variable with a known unfolding
1158 -- as well as when it's an explicit constructor application
1159 = knownCon env scrut (DataAlt con) args case_bndr alts cont
1161 | Lit lit <- scrut -- No need for same treatment as constructors
1162 -- because literals are inlined more vigorously
1163 = knownCon env scrut (LitAlt lit) [] case_bndr alts cont
1166 --------------------------------------------------
1167 -- 2. Eliminate the case if scrutinee is evaluated
1168 --------------------------------------------------
1170 rebuildCase env scrut case_bndr [(_, bndrs, rhs)] cont
1171 -- See if we can get rid of the case altogether
1172 -- See the extensive notes on case-elimination above
1173 -- mkCase made sure that if all the alternatives are equal,
1174 -- then there is now only one (DEFAULT) rhs
1175 | all isDeadBinder bndrs -- bndrs are [InId]
1177 -- Check that the scrutinee can be let-bound instead of case-bound
1178 , exprOkForSpeculation scrut
1179 -- OK not to evaluate it
1180 -- This includes things like (==# a# b#)::Bool
1181 -- so that we simplify
1182 -- case ==# a# b# of { True -> x; False -> x }
1185 -- This particular example shows up in default methods for
1186 -- comparision operations (e.g. in (>=) for Int.Int32)
1187 || exprIsHNF scrut -- It's already evaluated
1188 || var_demanded_later scrut -- It'll be demanded later
1190 -- || not opt_SimplPedanticBottoms) -- Or we don't care!
1191 -- We used to allow improving termination by discarding cases, unless -fpedantic-bottoms was on,
1192 -- but that breaks badly for the dataToTag# primop, which relies on a case to evaluate
1193 -- its argument: case x of { y -> dataToTag# y }
1194 -- Here we must *not* discard the case, because dataToTag# just fetches the tag from
1195 -- the info pointer. So we'll be pedantic all the time, and see if that gives any
1197 -- Also we don't want to discard 'seq's
1198 = do { tick (CaseElim case_bndr)
1199 ; env' <- simplNonRecX env case_bndr scrut
1200 ; simplExprF env' rhs cont }
1202 -- The case binder is going to be evaluated later,
1203 -- and the scrutinee is a simple variable
1204 var_demanded_later (Var v) = isStrictDmd (idNewDemandInfo case_bndr)
1205 && not (isTickBoxOp v)
1206 -- ugly hack; covering this case is what
1207 -- exprOkForSpeculation was intended for.
1208 var_demanded_later _ = False
1211 --------------------------------------------------
1212 -- 3. Catch-all case
1213 --------------------------------------------------
1215 rebuildCase env scrut case_bndr alts cont
1216 = do { -- Prepare the continuation;
1217 -- The new subst_env is in place
1218 (env', dup_cont, nodup_cont) <- prepareCaseCont env alts cont
1220 -- Simplify the alternatives
1221 ; (scrut', case_bndr', alts') <- simplAlts env' scrut case_bndr alts dup_cont
1222 ; let res_ty' = contResultType dup_cont
1223 ; case_expr <- mkCase scrut' case_bndr' res_ty' alts'
1225 -- Notice that rebuildDone returns the in-scope set from env', not alt_env
1226 -- The case binder *not* scope over the whole returned case-expression
1227 ; rebuild env' case_expr nodup_cont }
1230 simplCaseBinder checks whether the scrutinee is a variable, v. If so,
1231 try to eliminate uses of v in the RHSs in favour of case_bndr; that
1232 way, there's a chance that v will now only be used once, and hence
1235 Note [no-case-of-case]
1236 ~~~~~~~~~~~~~~~~~~~~~~
1237 There is a time we *don't* want to do that, namely when
1238 -fno-case-of-case is on. This happens in the first simplifier pass,
1239 and enhances full laziness. Here's the bad case:
1240 f = \ y -> ...(case x of I# v -> ...(case x of ...) ... )
1241 If we eliminate the inner case, we trap it inside the I# v -> arm,
1242 which might prevent some full laziness happening. I've seen this
1243 in action in spectral/cichelli/Prog.hs:
1244 [(m,n) | m <- [1..max], n <- [1..max]]
1245 Hence the check for NoCaseOfCase.
1247 Note [Suppressing the case binder-swap]
1248 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
1249 There is another situation when it might make sense to suppress the
1250 case-expression binde-swap. If we have
1252 case x of w1 { DEFAULT -> case x of w2 { A -> e1; B -> e2 }
1253 ...other cases .... }
1255 We'll perform the binder-swap for the outer case, giving
1257 case x of w1 { DEFAULT -> case w1 of w2 { A -> e1; B -> e2 }
1258 ...other cases .... }
1260 But there is no point in doing it for the inner case, because w1 can't
1261 be inlined anyway. Furthermore, doing the case-swapping involves
1262 zapping w2's occurrence info (see paragraphs that follow), and that
1263 forces us to bind w2 when doing case merging. So we get
1265 case x of w1 { A -> let w2 = w1 in e1
1266 B -> let w2 = w1 in e2
1267 ...other cases .... }
1269 This is plain silly in the common case where w2 is dead.
1271 Even so, I can't see a good way to implement this idea. I tried
1272 not doing the binder-swap if the scrutinee was already evaluated
1273 but that failed big-time:
1277 case v of w { MkT x ->
1278 case x of x1 { I# y1 ->
1279 case x of x2 { I# y2 -> ...
1281 Notice that because MkT is strict, x is marked "evaluated". But to
1282 eliminate the last case, we must either make sure that x (as well as
1283 x1) has unfolding MkT y1. THe straightforward thing to do is to do
1284 the binder-swap. So this whole note is a no-op.
1288 If we replace the scrutinee, v, by tbe case binder, then we have to nuke
1289 any occurrence info (eg IAmDead) in the case binder, because the
1290 case-binder now effectively occurs whenever v does. AND we have to do
1291 the same for the pattern-bound variables! Example:
1293 (case x of { (a,b) -> a }) (case x of { (p,q) -> q })
1295 Here, b and p are dead. But when we move the argment inside the first
1296 case RHS, and eliminate the second case, we get
1298 case x of { (a,b) -> a b }
1300 Urk! b is alive! Reason: the scrutinee was a variable, and case elimination
1303 Indeed, this can happen anytime the case binder isn't dead:
1304 case <any> of x { (a,b) ->
1305 case x of { (p,q) -> p } }
1306 Here (a,b) both look dead, but come alive after the inner case is eliminated.
1307 The point is that we bring into the envt a binding
1309 after the outer case, and that makes (a,b) alive. At least we do unless
1310 the case binder is guaranteed dead.
1314 Consider case (v `cast` co) of x { I# ->
1315 ... (case (v `cast` co) of {...}) ...
1316 We'd like to eliminate the inner case. We can get this neatly by
1317 arranging that inside the outer case we add the unfolding
1318 v |-> x `cast` (sym co)
1319 to v. Then we should inline v at the inner case, cancel the casts, and away we go
1321 Note [Improving seq]
1324 type family F :: * -> *
1325 type instance F Int = Int
1327 ... case e of x { DEFAULT -> rhs } ...
1329 where x::F Int. Then we'd like to rewrite (F Int) to Int, getting
1331 case e `cast` co of x'::Int
1332 I# x# -> let x = x' `cast` sym co
1335 so that 'rhs' can take advantage of the form of x'. Notice that Note
1336 [Case of cast] may then apply to the result.
1338 This showed up in Roman's experiments. Example:
1339 foo :: F Int -> Int -> Int
1340 foo t n = t `seq` bar n
1343 bar n = bar (n - case t of TI i -> i)
1344 Here we'd like to avoid repeated evaluating t inside the loop, by
1345 taking advantage of the `seq`.
1347 At one point I did transformation in LiberateCase, but it's more robust here.
1348 (Otherwise, there's a danger that we'll simply drop the 'seq' altogether, before
1349 LiberateCase gets to see it.)
1351 Note [Case elimination]
1352 ~~~~~~~~~~~~~~~~~~~~~~~
1353 The case-elimination transformation discards redundant case expressions.
1354 Start with a simple situation:
1356 case x# of ===> e[x#/y#]
1359 (when x#, y# are of primitive type, of course). We can't (in general)
1360 do this for algebraic cases, because we might turn bottom into
1363 The code in SimplUtils.prepareAlts has the effect of generalise this
1364 idea to look for a case where we're scrutinising a variable, and we
1365 know that only the default case can match. For example:
1369 DEFAULT -> ...(case x of
1373 Here the inner case is first trimmed to have only one alternative, the
1374 DEFAULT, after which it's an instance of the previous case. This
1375 really only shows up in eliminating error-checking code.
1377 We also make sure that we deal with this very common case:
1382 Here we are using the case as a strict let; if x is used only once
1383 then we want to inline it. We have to be careful that this doesn't
1384 make the program terminate when it would have diverged before, so we
1386 - e is already evaluated (it may so if e is a variable)
1387 - x is used strictly, or
1389 Lastly, the code in SimplUtils.mkCase combines identical RHSs. So
1391 case e of ===> case e of DEFAULT -> r
1395 Now again the case may be elminated by the CaseElim transformation.
1398 Further notes about case elimination
1399 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
1400 Consider: test :: Integer -> IO ()
1403 Turns out that this compiles to:
1406 eta1 :: State# RealWorld ->
1407 case PrelNum.< eta PrelNum.zeroInteger of wild { __DEFAULT ->
1409 (PrelNum.jtos eta ($w[] @ Char))
1411 of wild1 { (# new_s, a4 #) -> PrelIO.lvl23 new_s }}
1413 Notice the strange '<' which has no effect at all. This is a funny one.
1414 It started like this:
1416 f x y = if x < 0 then jtos x
1417 else if y==0 then "" else jtos x
1419 At a particular call site we have (f v 1). So we inline to get
1421 if v < 0 then jtos x
1422 else if 1==0 then "" else jtos x
1424 Now simplify the 1==0 conditional:
1426 if v<0 then jtos v else jtos v
1428 Now common-up the two branches of the case:
1430 case (v<0) of DEFAULT -> jtos v
1432 Why don't we drop the case? Because it's strict in v. It's technically
1433 wrong to drop even unnecessary evaluations, and in practice they
1434 may be a result of 'seq' so we *definitely* don't want to drop those.
1435 I don't really know how to improve this situation.
1439 simplCaseBinder :: SimplEnv -> OutExpr -> OutId -> [InAlt]
1440 -> SimplM (SimplEnv, OutExpr, OutId)
1441 simplCaseBinder env0 scrut0 case_bndr0 alts
1442 = do { (env1, case_bndr1) <- simplBinder env0 case_bndr0
1444 ; fam_envs <- getFamEnvs
1445 ; (env2, scrut2, case_bndr2) <- improve_seq fam_envs env1 scrut0
1446 case_bndr0 case_bndr1 alts
1447 -- Note [Improving seq]
1449 ; let (env3, case_bndr3) = improve_case_bndr env2 scrut2 case_bndr2
1450 -- Note [Case of cast]
1452 ; return (env3, scrut2, case_bndr3) }
1455 improve_seq fam_envs env scrut case_bndr case_bndr1 [(DEFAULT,_,_)]
1456 | Just (co, ty2) <- topNormaliseType fam_envs (idType case_bndr1)
1457 = do { case_bndr2 <- newId FSLIT("nt") ty2
1458 ; let rhs = DoneEx (Var case_bndr2 `Cast` mkSymCoercion co)
1459 env2 = extendIdSubst env case_bndr rhs
1460 ; return (env2, scrut `Cast` co, case_bndr2) }
1462 improve_seq _ env scrut _ case_bndr1 _
1463 = return (env, scrut, case_bndr1)
1466 improve_case_bndr env scrut case_bndr
1467 | switchIsOn (getSwitchChecker env) NoCaseOfCase
1468 -- See Note [no-case-of-case]
1471 | otherwise -- Failed try; see Note [Suppressing the case binder-swap]
1472 -- not (isEvaldUnfolding (idUnfolding v))
1474 Var v -> (modifyInScope env1 v case_bndr', case_bndr')
1475 -- Note about using modifyInScope for v here
1476 -- We could extend the substitution instead, but it would be
1477 -- a hack because then the substitution wouldn't be idempotent
1478 -- any more (v is an OutId). And this does just as well.
1480 Cast (Var v) co -> (addBinderUnfolding env1 v rhs, case_bndr')
1482 rhs = Cast (Var case_bndr') (mkSymCoercion co)
1484 _ -> (env, case_bndr)
1486 case_bndr' = zapOccInfo case_bndr
1487 env1 = modifyInScope env case_bndr case_bndr'
1490 zapOccInfo :: InId -> InId -- See Note [zapOccInfo]
1491 zapOccInfo b = b `setIdOccInfo` NoOccInfo
1495 simplAlts does two things:
1497 1. Eliminate alternatives that cannot match, including the
1498 DEFAULT alternative.
1500 2. If the DEFAULT alternative can match only one possible constructor,
1501 then make that constructor explicit.
1503 case e of x { DEFAULT -> rhs }
1505 case e of x { (a,b) -> rhs }
1506 where the type is a single constructor type. This gives better code
1507 when rhs also scrutinises x or e.
1509 Here "cannot match" includes knowledge from GADTs
1511 It's a good idea do do this stuff before simplifying the alternatives, to
1512 avoid simplifying alternatives we know can't happen, and to come up with
1513 the list of constructors that are handled, to put into the IdInfo of the
1514 case binder, for use when simplifying the alternatives.
1516 Eliminating the default alternative in (1) isn't so obvious, but it can
1519 data Colour = Red | Green | Blue
1528 DEFAULT -> [ case y of ... ]
1530 If we inline h into f, the default case of the inlined h can't happen.
1531 If we don't notice this, we may end up filtering out *all* the cases
1532 of the inner case y, which give us nowhere to go!
1536 simplAlts :: SimplEnv
1538 -> InId -- Case binder
1539 -> [InAlt] -> SimplCont
1540 -> SimplM (OutExpr, OutId, [OutAlt]) -- Includes the continuation
1541 -- Like simplExpr, this just returns the simplified alternatives;
1542 -- it not return an environment
1544 simplAlts env scrut case_bndr alts cont'
1545 = -- pprTrace "simplAlts" (ppr alts $$ ppr (seIdSubst env)) $
1546 do { let alt_env = zapFloats env
1547 ; (alt_env', scrut', case_bndr') <- simplCaseBinder alt_env scrut case_bndr alts
1549 ; (imposs_deflt_cons, in_alts) <- prepareAlts alt_env' scrut case_bndr' alts
1551 ; alts' <- mapM (simplAlt alt_env' imposs_deflt_cons case_bndr' cont') in_alts
1552 ; return (scrut', case_bndr', alts') }
1554 ------------------------------------
1555 simplAlt :: SimplEnv
1556 -> [AltCon] -- These constructors can't be present when
1557 -- matching the DEFAULT alternative
1558 -> OutId -- The case binder
1563 simplAlt env imposs_deflt_cons case_bndr' cont' (DEFAULT, bndrs, rhs)
1564 = ASSERT( null bndrs )
1565 do { let env' = addBinderOtherCon env case_bndr' imposs_deflt_cons
1566 -- Record the constructors that the case-binder *can't* be.
1567 ; rhs' <- simplExprC env' rhs cont'
1568 ; return (DEFAULT, [], rhs') }
1570 simplAlt env _ case_bndr' cont' (LitAlt lit, bndrs, rhs)
1571 = ASSERT( null bndrs )
1572 do { let env' = addBinderUnfolding env case_bndr' (Lit lit)
1573 ; rhs' <- simplExprC env' rhs cont'
1574 ; return (LitAlt lit, [], rhs') }
1576 simplAlt env _ case_bndr' cont' (DataAlt con, vs, rhs)
1577 = do { -- Deal with the pattern-bound variables
1578 -- Mark the ones that are in ! positions in the
1579 -- data constructor as certainly-evaluated.
1580 -- NB: simplLamBinders preserves this eval info
1581 let vs_with_evals = add_evals (dataConRepStrictness con)
1582 ; (env', vs') <- simplLamBndrs env vs_with_evals
1584 -- Bind the case-binder to (con args)
1585 ; let inst_tys' = tyConAppArgs (idType case_bndr')
1586 con_args = map Type inst_tys' ++ varsToCoreExprs vs'
1587 env'' = addBinderUnfolding env' case_bndr'
1588 (mkConApp con con_args)
1590 ; rhs' <- simplExprC env'' rhs cont'
1591 ; return (DataAlt con, vs', rhs') }
1593 -- add_evals records the evaluated-ness of the bound variables of
1594 -- a case pattern. This is *important*. Consider
1595 -- data T = T !Int !Int
1597 -- case x of { T a b -> T (a+1) b }
1599 -- We really must record that b is already evaluated so that we don't
1600 -- go and re-evaluate it when constructing the result.
1601 -- See Note [Data-con worker strictness] in MkId.lhs
1606 go (v:vs') strs | isTyVar v = v : go vs' strs
1607 go (v:vs') (str:strs)
1608 | isMarkedStrict str = evald_v : go vs' strs
1609 | otherwise = zapped_v : go vs' strs
1611 zapped_v = zap_occ_info v
1612 evald_v = zapped_v `setIdUnfolding` evaldUnfolding
1613 go _ _ = pprPanic "cat_evals" (ppr con $$ ppr vs $$ ppr the_strs)
1615 -- zap_occ_info: if the case binder is alive, then we add the unfolding
1617 -- to the envt; so vs are now very much alive
1618 -- Note [Aug06] I can't see why this actually matters, but it's neater
1619 -- case e of t { (a,b) -> ...(case t of (p,q) -> p)... }
1620 -- ==> case e of t { (a,b) -> ...(a)... }
1621 -- Look, Ma, a is alive now.
1622 zap_occ_info | isDeadBinder case_bndr' = \ident -> ident
1623 | otherwise = zapOccInfo
1625 addBinderUnfolding :: SimplEnv -> Id -> CoreExpr -> SimplEnv
1626 addBinderUnfolding env bndr rhs
1627 = modifyInScope env bndr (bndr `setIdUnfolding` mkUnfolding False rhs)
1629 addBinderOtherCon :: SimplEnv -> Id -> [AltCon] -> SimplEnv
1630 addBinderOtherCon env bndr cons
1631 = modifyInScope env bndr (bndr `setIdUnfolding` mkOtherCon cons)
1635 %************************************************************************
1637 \subsection{Known constructor}
1639 %************************************************************************
1641 We are a bit careful with occurrence info. Here's an example
1643 (\x* -> case x of (a*, b) -> f a) (h v, e)
1645 where the * means "occurs once". This effectively becomes
1646 case (h v, e) of (a*, b) -> f a)
1648 let a* = h v; b = e in f a
1652 All this should happen in one sweep.
1655 knownCon :: SimplEnv -> OutExpr -> AltCon -> [OutExpr]
1656 -> InId -> [InAlt] -> SimplCont
1657 -> SimplM (SimplEnv, OutExpr)
1659 knownCon env scrut con args bndr alts cont
1660 = do { tick (KnownBranch bndr)
1661 ; knownAlt env scrut args bndr (findAlt con alts) cont }
1663 knownAlt :: SimplEnv -> OutExpr -> [OutExpr]
1664 -> InId -> (AltCon, [CoreBndr], InExpr) -> SimplCont
1665 -> SimplM (SimplEnv, OutExpr)
1666 knownAlt env scrut _ bndr (DEFAULT, bs, rhs) cont
1668 do { env' <- simplNonRecX env bndr scrut
1669 -- This might give rise to a binding with non-atomic args
1670 -- like x = Node (f x) (g x)
1671 -- but simplNonRecX will atomic-ify it
1672 ; simplExprF env' rhs cont }
1674 knownAlt env scrut _ bndr (LitAlt _, bs, rhs) cont
1676 do { env' <- simplNonRecX env bndr scrut
1677 ; simplExprF env' rhs cont }
1679 knownAlt env scrut the_args bndr (DataAlt dc, bs, rhs) cont
1680 = do { let dead_bndr = isDeadBinder bndr -- bndr is an InId
1681 n_drop_tys = length (dataConUnivTyVars dc)
1682 ; env' <- bind_args env dead_bndr bs (drop n_drop_tys the_args)
1684 -- It's useful to bind bndr to scrut, rather than to a fresh
1685 -- binding x = Con arg1 .. argn
1686 -- because very often the scrut is a variable, so we avoid
1687 -- creating, and then subsequently eliminating, a let-binding
1688 -- BUT, if scrut is a not a variable, we must be careful
1689 -- about duplicating the arg redexes; in that case, make
1690 -- a new con-app from the args
1691 bndr_rhs = case scrut of
1694 con_app = mkConApp dc (take n_drop_tys the_args ++ con_args)
1695 con_args = [substExpr env' (varToCoreExpr b) | b <- bs]
1696 -- args are aready OutExprs, but bs are InIds
1698 ; env'' <- simplNonRecX env' bndr bndr_rhs
1699 ; -- pprTrace "knownCon2" (ppr bs $$ ppr rhs $$ ppr (seIdSubst env'')) $
1700 simplExprF env'' rhs cont }
1703 bind_args env' _ [] _ = return env'
1705 bind_args env' dead_bndr (b:bs') (Type ty : args)
1706 = ASSERT( isTyVar b )
1707 bind_args (extendTvSubst env' b ty) dead_bndr bs' args
1709 bind_args env' dead_bndr (b:bs') (arg : args)
1711 do { let b' = if dead_bndr then b else zapOccInfo b
1712 -- Note that the binder might be "dead", because it doesn't
1713 -- occur in the RHS; and simplNonRecX may therefore discard
1714 -- it via postInlineUnconditionally.
1715 -- Nevertheless we must keep it if the case-binder is alive,
1716 -- because it may be used in the con_app. See Note [zapOccInfo]
1717 ; env'' <- simplNonRecX env' b' arg
1718 ; bind_args env'' dead_bndr bs' args }
1721 pprPanic "bind_args" $ ppr dc $$ ppr bs $$ ppr the_args $$
1722 text "scrut:" <+> ppr scrut
1726 %************************************************************************
1728 \subsection{Duplicating continuations}
1730 %************************************************************************
1733 prepareCaseCont :: SimplEnv
1734 -> [InAlt] -> SimplCont
1735 -> SimplM (SimplEnv, SimplCont,SimplCont)
1736 -- Return a duplicatable continuation, a non-duplicable part
1737 -- plus some extra bindings (that scope over the entire
1740 -- No need to make it duplicatable if there's only one alternative
1741 prepareCaseCont env [_] cont = return (env, cont, mkBoringStop (contResultType cont))
1742 prepareCaseCont env _ cont = mkDupableCont env cont
1746 mkDupableCont :: SimplEnv -> SimplCont
1747 -> SimplM (SimplEnv, SimplCont, SimplCont)
1749 mkDupableCont env cont
1750 | contIsDupable cont
1751 = return (env, cont, mkBoringStop (contResultType cont))
1753 mkDupableCont _ (Stop {}) = panic "mkDupableCont" -- Handled by previous eqn
1755 mkDupableCont env (CoerceIt ty cont)
1756 = do { (env', dup, nodup) <- mkDupableCont env cont
1757 ; return (env', CoerceIt ty dup, nodup) }
1759 mkDupableCont env cont@(StrictBind bndr _ _ se _)
1760 = return (env, mkBoringStop (substTy se (idType bndr)), cont)
1761 -- See Note [Duplicating strict continuations]
1763 mkDupableCont env cont@(StrictArg _ fun_ty _ _ _)
1764 = return (env, mkBoringStop (funArgTy fun_ty), cont)
1765 -- See Note [Duplicating strict continuations]
1767 mkDupableCont env (ApplyTo _ arg se cont)
1768 = -- e.g. [...hole...] (...arg...)
1770 -- let a = ...arg...
1771 -- in [...hole...] a
1772 do { (env', dup_cont, nodup_cont) <- mkDupableCont env cont
1773 ; arg' <- simplExpr (se `setInScope` env') arg
1774 ; (env'', arg'') <- makeTrivial env' arg'
1775 ; let app_cont = ApplyTo OkToDup arg'' (zapSubstEnv env') dup_cont
1776 ; return (env'', app_cont, nodup_cont) }
1778 mkDupableCont env cont@(Select _ case_bndr [(_, bs, _rhs)] se _case_cont)
1779 -- See Note [Single-alternative case]
1780 -- | not (exprIsDupable rhs && contIsDupable case_cont)
1781 -- | not (isDeadBinder case_bndr)
1782 | all isDeadBinder bs -- InIds
1783 = return (env, mkBoringStop scrut_ty, cont)
1785 scrut_ty = substTy se (idType case_bndr)
1787 mkDupableCont env (Select _ case_bndr alts se cont)
1788 = -- e.g. (case [...hole...] of { pi -> ei })
1790 -- let ji = \xij -> ei
1791 -- in case [...hole...] of { pi -> ji xij }
1792 do { tick (CaseOfCase case_bndr)
1793 ; (env', dup_cont, nodup_cont) <- mkDupableCont env cont
1794 -- NB: call mkDupableCont here, *not* prepareCaseCont
1795 -- We must make a duplicable continuation, whereas prepareCaseCont
1796 -- doesn't when there is a single case branch
1798 ; let alt_env = se `setInScope` env'
1799 ; (alt_env', case_bndr') <- simplBinder alt_env case_bndr
1800 ; alts' <- mapM (simplAlt alt_env' [] case_bndr' dup_cont) alts
1801 -- Safe to say that there are no handled-cons for the DEFAULT case
1802 -- NB: simplBinder does not zap deadness occ-info, so
1803 -- a dead case_bndr' will still advertise its deadness
1804 -- This is really important because in
1805 -- case e of b { (# p,q #) -> ... }
1806 -- b is always dead, and indeed we are not allowed to bind b to (# p,q #),
1807 -- which might happen if e was an explicit unboxed pair and b wasn't marked dead.
1808 -- In the new alts we build, we have the new case binder, so it must retain
1810 -- NB: we don't use alt_env further; it has the substEnv for
1811 -- the alternatives, and we don't want that
1813 ; (env'', alts'') <- mkDupableAlts env' case_bndr' alts'
1814 ; return (env'', -- Note [Duplicated env]
1815 Select OkToDup case_bndr' alts'' (zapSubstEnv env'')
1816 (mkBoringStop (contResultType dup_cont)),
1820 mkDupableAlts :: SimplEnv -> OutId -> [InAlt]
1821 -> SimplM (SimplEnv, [InAlt])
1822 -- Absorbs the continuation into the new alternatives
1824 mkDupableAlts env case_bndr' the_alts
1827 go env0 [] = return (env0, [])
1829 = do { (env1, alt') <- mkDupableAlt env0 case_bndr' alt
1830 ; (env2, alts') <- go env1 alts
1831 ; return (env2, alt' : alts' ) }
1833 mkDupableAlt :: SimplEnv -> OutId -> (AltCon, [CoreBndr], CoreExpr)
1834 -> SimplM (SimplEnv, (AltCon, [CoreBndr], CoreExpr))
1835 mkDupableAlt env case_bndr' (con, bndrs', rhs')
1836 | exprIsDupable rhs' -- Note [Small alternative rhs]
1837 = return (env, (con, bndrs', rhs'))
1839 = do { let rhs_ty' = exprType rhs'
1840 used_bndrs' = filter abstract_over (case_bndr' : bndrs')
1842 | isTyVar bndr = True -- Abstract over all type variables just in case
1843 | otherwise = not (isDeadBinder bndr)
1844 -- The deadness info on the new Ids is preserved by simplBinders
1846 ; (final_bndrs', final_args) -- Note [Join point abstraction]
1847 <- if (any isId used_bndrs')
1848 then return (used_bndrs', varsToCoreExprs used_bndrs')
1849 else do { rw_id <- newId FSLIT("w") realWorldStatePrimTy
1850 ; return ([rw_id], [Var realWorldPrimId]) }
1852 ; join_bndr <- newId FSLIT("$j") (mkPiTypes final_bndrs' rhs_ty')
1853 -- Note [Funky mkPiTypes]
1855 ; let -- We make the lambdas into one-shot-lambdas. The
1856 -- join point is sure to be applied at most once, and doing so
1857 -- prevents the body of the join point being floated out by
1858 -- the full laziness pass
1859 really_final_bndrs = map one_shot final_bndrs'
1860 one_shot v | isId v = setOneShotLambda v
1862 join_rhs = mkLams really_final_bndrs rhs'
1863 join_call = mkApps (Var join_bndr) final_args
1865 ; return (addNonRec env join_bndr join_rhs, (con, bndrs', join_call)) }
1866 -- See Note [Duplicated env]
1869 Note [Duplicated env]
1870 ~~~~~~~~~~~~~~~~~~~~~
1871 Some of the alternatives are simplified, but have not been turned into a join point
1872 So they *must* have an zapped subst-env. So we can't use completeNonRecX to
1873 bind the join point, because it might to do PostInlineUnconditionally, and
1874 we'd lose that when zapping the subst-env. We could have a per-alt subst-env,
1875 but zapping it (as we do in mkDupableCont, the Select case) is safe, and
1876 at worst delays the join-point inlining.
1878 Note [Small alterantive rhs]
1879 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~
1880 It is worth checking for a small RHS because otherwise we
1881 get extra let bindings that may cause an extra iteration of the simplifier to
1882 inline back in place. Quite often the rhs is just a variable or constructor.
1883 The Ord instance of Maybe in PrelMaybe.lhs, for example, took several extra
1884 iterations because the version with the let bindings looked big, and so wasn't
1885 inlined, but after the join points had been inlined it looked smaller, and so
1888 NB: we have to check the size of rhs', not rhs.
1889 Duplicating a small InAlt might invalidate occurrence information
1890 However, if it *is* dupable, we return the *un* simplified alternative,
1891 because otherwise we'd need to pair it up with an empty subst-env....
1892 but we only have one env shared between all the alts.
1893 (Remember we must zap the subst-env before re-simplifying something).
1894 Rather than do this we simply agree to re-simplify the original (small) thing later.
1896 Note [Funky mkPiTypes]
1897 ~~~~~~~~~~~~~~~~~~~~~~
1898 Notice the funky mkPiTypes. If the contructor has existentials
1899 it's possible that the join point will be abstracted over
1900 type varaibles as well as term variables.
1901 Example: Suppose we have
1902 data T = forall t. C [t]
1904 case (case e of ...) of
1906 We get the join point
1907 let j :: forall t. [t] -> ...
1908 j = /\t \xs::[t] -> rhs
1910 case (case e of ...) of
1911 C t xs::[t] -> j t xs
1913 Note [Join point abstaction]
1914 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~
1915 If we try to lift a primitive-typed something out
1916 for let-binding-purposes, we will *caseify* it (!),
1917 with potentially-disastrous strictness results. So
1918 instead we turn it into a function: \v -> e
1919 where v::State# RealWorld#. The value passed to this function
1920 is realworld#, which generates (almost) no code.
1922 There's a slight infelicity here: we pass the overall
1923 case_bndr to all the join points if it's used in *any* RHS,
1924 because we don't know its usage in each RHS separately
1926 We used to say "&& isUnLiftedType rhs_ty'" here, but now
1927 we make the join point into a function whenever used_bndrs'
1928 is empty. This makes the join-point more CPR friendly.
1929 Consider: let j = if .. then I# 3 else I# 4
1930 in case .. of { A -> j; B -> j; C -> ... }
1932 Now CPR doesn't w/w j because it's a thunk, so
1933 that means that the enclosing function can't w/w either,
1934 which is a lose. Here's the example that happened in practice:
1935 kgmod :: Int -> Int -> Int
1936 kgmod x y = if x > 0 && y < 0 || x < 0 && y > 0
1940 I have seen a case alternative like this:
1942 It's a bit silly to add the realWorld dummy arg in this case, making
1945 (the \v alone is enough to make CPR happy) but I think it's rare
1947 Note [Duplicating strict continuations]
1948 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
1949 Do *not* duplicate StrictBind and StritArg continuations. We gain
1950 nothing by propagating them into the expressions, and we do lose a
1951 lot. Here's an example:
1952 && (case x of { T -> F; F -> T }) E
1953 Now, && is strict so we end up simplifying the case with
1954 an ArgOf continuation. If we let-bind it, we get
1956 let $j = \v -> && v E
1957 in simplExpr (case x of { T -> F; F -> T })
1959 And after simplifying more we get
1961 let $j = \v -> && v E
1962 in case x of { T -> $j F; F -> $j T }
1963 Which is a Very Bad Thing
1965 The desire not to duplicate is the entire reason that
1966 mkDupableCont returns a pair of continuations.
1968 The original plan had:
1969 e.g. (...strict-fn...) [...hole...]
1971 let $j = \a -> ...strict-fn...
1974 Note [Single-alternative cases]
1975 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
1976 This case is just like the ArgOf case. Here's an example:
1980 case (case x of I# x' ->
1982 True -> I# (negate# x')
1983 False -> I# x') of y {
1985 Because the (case x) has only one alternative, we'll transform to
1987 case (case x' <# 0# of
1988 True -> I# (negate# x')
1989 False -> I# x') of y {
1991 But now we do *NOT* want to make a join point etc, giving
1993 let $j = \y -> MkT y
1995 True -> $j (I# (negate# x'))
1997 In this case the $j will inline again, but suppose there was a big
1998 strict computation enclosing the orginal call to MkT. Then, it won't
1999 "see" the MkT any more, because it's big and won't get duplicated.
2000 And, what is worse, nothing was gained by the case-of-case transform.
2002 When should use this case of mkDupableCont?
2003 However, matching on *any* single-alternative case is a *disaster*;
2004 e.g. case (case ....) of (a,b) -> (# a,b #)
2005 We must push the outer case into the inner one!
2008 * Match [(DEFAULT,_,_)], but in the common case of Int,
2009 the alternative-filling-in code turned the outer case into
2010 case (...) of y { I# _ -> MkT y }
2012 * Match on single alternative plus (not (isDeadBinder case_bndr))
2013 Rationale: pushing the case inwards won't eliminate the construction.
2014 But there's a risk of
2015 case (...) of y { (a,b) -> let z=(a,b) in ... }
2016 Now y looks dead, but it'll come alive again. Still, this
2017 seems like the best option at the moment.
2019 * Match on single alternative plus (all (isDeadBinder bndrs))
2020 Rationale: this is essentially seq.
2022 * Match when the rhs is *not* duplicable, and hence would lead to a
2023 join point. This catches the disaster-case above. We can test
2024 the *un-simplified* rhs, which is fine. It might get bigger or
2025 smaller after simplification; if it gets smaller, this case might
2026 fire next time round. NB also that we must test contIsDupable
2027 case_cont *btoo, because case_cont might be big!
2029 HOWEVER: I found that this version doesn't work well, because
2030 we can get let x = case (...) of { small } in ...case x...
2031 When x is inlined into its full context, we find that it was a bad
2032 idea to have pushed the outer case inside the (...) case.