2 % (c) The AQUA Project, Glasgow University, 1993-1998
4 \section[Simplify]{The main module of the simplifier}
7 module Simplify ( simplTopBinds, simplExpr ) where
9 #include "HsVersions.h"
13 import Type hiding ( substTy, extendTvSubst )
16 import Literal ( mkStringLit )
17 import MkId ( rUNTIME_ERROR_ID )
22 import FamInstEnv ( topNormaliseType )
23 import DataCon ( dataConRepStrictness, dataConUnivTyVars )
25 import NewDemand ( isStrictDmd )
26 import PprCore ( pprParendExpr, pprCoreExpr )
27 import CoreUnfold ( mkUnfolding, callSiteInline, CallCtxt(..) )
29 import Rules ( lookupRule )
30 import BasicTypes ( isMarkedStrict )
31 import CostCentre ( currentCCS )
32 import TysPrim ( realWorldStatePrimTy )
33 import PrelInfo ( realWorldPrimId )
34 import BasicTypes ( TopLevelFlag(..), isTopLevel,
35 RecFlag(..), isNonRuleLoopBreaker )
36 import Maybes ( orElse )
37 import Data.List ( mapAccumL )
44 The guts of the simplifier is in this module, but the driver loop for
45 the simplifier is in SimplCore.lhs.
48 -----------------------------------------
49 *** IMPORTANT NOTE ***
50 -----------------------------------------
51 The simplifier used to guarantee that the output had no shadowing, but
52 it does not do so any more. (Actually, it never did!) The reason is
53 documented with simplifyArgs.
56 -----------------------------------------
57 *** IMPORTANT NOTE ***
58 -----------------------------------------
59 Many parts of the simplifier return a bunch of "floats" as well as an
60 expression. This is wrapped as a datatype SimplUtils.FloatsWith.
62 All "floats" are let-binds, not case-binds, but some non-rec lets may
63 be unlifted (with RHS ok-for-speculation).
67 -----------------------------------------
68 ORGANISATION OF FUNCTIONS
69 -----------------------------------------
71 - simplify all top-level binders
72 - for NonRec, call simplRecOrTopPair
73 - for Rec, call simplRecBind
76 ------------------------------
77 simplExpr (applied lambda) ==> simplNonRecBind
78 simplExpr (Let (NonRec ...) ..) ==> simplNonRecBind
79 simplExpr (Let (Rec ...) ..) ==> simplify binders; simplRecBind
81 ------------------------------
82 simplRecBind [binders already simplfied]
83 - use simplRecOrTopPair on each pair in turn
85 simplRecOrTopPair [binder already simplified]
86 Used for: recursive bindings (top level and nested)
87 top-level non-recursive bindings
89 - check for PreInlineUnconditionally
93 Used for: non-top-level non-recursive bindings
94 beta reductions (which amount to the same thing)
95 Because it can deal with strict arts, it takes a
96 "thing-inside" and returns an expression
98 - check for PreInlineUnconditionally
99 - simplify binder, including its IdInfo
108 simplNonRecX: [given a *simplified* RHS, but an *unsimplified* binder]
109 Used for: binding case-binder and constr args in a known-constructor case
110 - check for PreInLineUnconditionally
114 ------------------------------
115 simplLazyBind: [binder already simplified, RHS not]
116 Used for: recursive bindings (top level and nested)
117 top-level non-recursive bindings
118 non-top-level, but *lazy* non-recursive bindings
119 [must not be strict or unboxed]
120 Returns floats + an augmented environment, not an expression
121 - substituteIdInfo and add result to in-scope
122 [so that rules are available in rec rhs]
125 - float if exposes constructor or PAP
129 completeNonRecX: [binder and rhs both simplified]
130 - if the the thing needs case binding (unlifted and not ok-for-spec)
136 completeBind: [given a simplified RHS]
137 [used for both rec and non-rec bindings, top level and not]
138 - try PostInlineUnconditionally
139 - add unfolding [this is the only place we add an unfolding]
144 Right hand sides and arguments
145 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
146 In many ways we want to treat
147 (a) the right hand side of a let(rec), and
148 (b) a function argument
149 in the same way. But not always! In particular, we would
150 like to leave these arguments exactly as they are, so they
151 will match a RULE more easily.
156 It's harder to make the rule match if we ANF-ise the constructor,
157 or eta-expand the PAP:
159 f (let { a = g x; b = h x } in (a,b))
162 On the other hand if we see the let-defns
167 then we *do* want to ANF-ise and eta-expand, so that p and q
168 can be safely inlined.
170 Even floating lets out is a bit dubious. For let RHS's we float lets
171 out if that exposes a value, so that the value can be inlined more vigorously.
174 r = let x = e in (x,x)
176 Here, if we float the let out we'll expose a nice constructor. We did experiments
177 that showed this to be a generally good thing. But it was a bad thing to float
178 lets out unconditionally, because that meant they got allocated more often.
180 For function arguments, there's less reason to expose a constructor (it won't
181 get inlined). Just possibly it might make a rule match, but I'm pretty skeptical.
182 So for the moment we don't float lets out of function arguments either.
187 For eta expansion, we want to catch things like
189 case e of (a,b) -> \x -> case a of (p,q) -> \y -> r
191 If the \x was on the RHS of a let, we'd eta expand to bring the two
192 lambdas together. And in general that's a good thing to do. Perhaps
193 we should eta expand wherever we find a (value) lambda? Then the eta
194 expansion at a let RHS can concentrate solely on the PAP case.
197 %************************************************************************
199 \subsection{Bindings}
201 %************************************************************************
204 simplTopBinds :: SimplEnv -> [InBind] -> SimplM [OutBind]
206 simplTopBinds env0 binds0
207 = do { -- Put all the top-level binders into scope at the start
208 -- so that if a transformation rule has unexpectedly brought
209 -- anything into scope, then we don't get a complaint about that.
210 -- It's rather as if the top-level binders were imported.
211 ; env1 <- simplRecBndrs env0 (bindersOfBinds binds0)
212 ; dflags <- getDOptsSmpl
213 ; let dump_flag = dopt Opt_D_dump_inlinings dflags ||
214 dopt Opt_D_dump_rule_firings dflags
215 ; env2 <- simpl_binds dump_flag env1 binds0
216 ; freeTick SimplifierDone
217 ; return (getFloats env2) }
219 -- We need to track the zapped top-level binders, because
220 -- they should have their fragile IdInfo zapped (notably occurrence info)
221 -- That's why we run down binds and bndrs' simultaneously.
223 -- The dump-flag emits a trace for each top-level binding, which
224 -- helps to locate the tracing for inlining and rule firing
225 simpl_binds :: Bool -> SimplEnv -> [InBind] -> SimplM SimplEnv
226 simpl_binds _ env [] = return env
227 simpl_binds dump env (bind:binds) = do { env' <- trace_bind dump bind $
229 ; simpl_binds dump env' binds }
231 trace_bind True bind = pprTrace "SimplBind" (ppr (bindersOf bind))
232 trace_bind False _ = \x -> x
234 simpl_bind env (Rec pairs) = simplRecBind env TopLevel pairs
235 simpl_bind env (NonRec b r) = simplRecOrTopPair env' TopLevel b b' r
237 (env', b') = addBndrRules env b (lookupRecBndr env b)
241 %************************************************************************
243 \subsection{Lazy bindings}
245 %************************************************************************
247 simplRecBind is used for
248 * recursive bindings only
251 simplRecBind :: SimplEnv -> TopLevelFlag
254 simplRecBind env0 top_lvl pairs0
255 = do { let (env_with_info, triples) = mapAccumL add_rules env0 pairs0
256 ; env1 <- go (zapFloats env_with_info) triples
257 ; return (env0 `addRecFloats` env1) }
258 -- addFloats adds the floats from env1,
259 -- *and* updates env0 with the in-scope set from env1
261 add_rules :: SimplEnv -> (InBndr,InExpr) -> (SimplEnv, (InBndr, OutBndr, InExpr))
262 -- Add the (substituted) rules to the binder
263 add_rules env (bndr, rhs) = (env', (bndr, bndr', rhs))
265 (env', bndr') = addBndrRules env bndr (lookupRecBndr env bndr)
267 go env [] = return env
269 go env ((old_bndr, new_bndr, rhs) : pairs)
270 = do { env' <- simplRecOrTopPair env top_lvl old_bndr new_bndr rhs
274 simplOrTopPair is used for
275 * recursive bindings (whether top level or not)
276 * top-level non-recursive bindings
278 It assumes the binder has already been simplified, but not its IdInfo.
281 simplRecOrTopPair :: SimplEnv
283 -> InId -> OutBndr -> InExpr -- Binder and rhs
284 -> SimplM SimplEnv -- Returns an env that includes the binding
286 simplRecOrTopPair env top_lvl old_bndr new_bndr rhs
287 | preInlineUnconditionally env top_lvl old_bndr rhs -- Check for unconditional inline
288 = do { tick (PreInlineUnconditionally old_bndr)
289 ; return (extendIdSubst env old_bndr (mkContEx env rhs)) }
292 = simplLazyBind env top_lvl Recursive old_bndr new_bndr rhs env
293 -- May not actually be recursive, but it doesn't matter
297 simplLazyBind is used for
298 * [simplRecOrTopPair] recursive bindings (whether top level or not)
299 * [simplRecOrTopPair] top-level non-recursive bindings
300 * [simplNonRecE] non-top-level *lazy* non-recursive bindings
303 1. It assumes that the binder is *already* simplified,
304 and is in scope, and its IdInfo too, except unfolding
306 2. It assumes that the binder type is lifted.
308 3. It does not check for pre-inline-unconditionallly;
309 that should have been done already.
312 simplLazyBind :: SimplEnv
313 -> TopLevelFlag -> RecFlag
314 -> InId -> OutId -- Binder, both pre-and post simpl
315 -- The OutId has IdInfo, except arity, unfolding
316 -> InExpr -> SimplEnv -- The RHS and its environment
319 simplLazyBind env top_lvl is_rec bndr bndr1 rhs rhs_se
320 = do { let rhs_env = rhs_se `setInScope` env
321 (tvs, body) = case collectTyBinders rhs of
322 (tvs, body) | not_lam body -> (tvs,body)
323 | otherwise -> ([], rhs)
324 not_lam (Lam _ _) = False
326 -- Do not do the "abstract tyyvar" thing if there's
327 -- a lambda inside, becuase it defeats eta-reduction
328 -- f = /\a. \x. g a x
331 ; (body_env, tvs') <- simplBinders rhs_env tvs
332 -- See Note [Floating and type abstraction] in SimplUtils
335 ; (body_env1, body1) <- simplExprF body_env body mkBoringStop
337 -- ANF-ise a constructor or PAP rhs
338 ; (body_env2, body2) <- prepareRhs body_env1 body1
341 <- if not (doFloatFromRhs top_lvl is_rec False body2 body_env2)
342 then -- No floating, just wrap up!
343 do { rhs' <- mkLam tvs' (wrapFloats body_env2 body2)
344 ; return (env, rhs') }
346 else if null tvs then -- Simple floating
347 do { tick LetFloatFromLet
348 ; return (addFloats env body_env2, body2) }
350 else -- Do type-abstraction first
351 do { tick LetFloatFromLet
352 ; (poly_binds, body3) <- abstractFloats tvs' body_env2 body2
353 ; rhs' <- mkLam tvs' body3
354 ; env' <- foldlM add_poly_bind env poly_binds
355 ; return (env', rhs') }
357 ; completeBind env' top_lvl bndr bndr1 rhs' }
359 add_poly_bind env (NonRec poly_id rhs)
360 = completeBind env top_lvl poly_id poly_id rhs
361 -- completeBind adds the new binding in the
362 -- proper way (ie complete with unfolding etc),
363 -- and extends the in-scope set
364 add_poly_bind env bind@(Rec _)
365 = return (extendFloats env bind)
366 -- Hack: letrecs are more awkward, so we extend "by steam"
367 -- without adding unfoldings etc. At worst this leads to
368 -- more simplifier iterations
371 A specialised variant of simplNonRec used when the RHS is already simplified,
372 notably in knownCon. It uses case-binding where necessary.
375 simplNonRecX :: SimplEnv
376 -> InId -- Old binder
377 -> OutExpr -- Simplified RHS
380 simplNonRecX env bndr new_rhs
381 = do { (env', bndr') <- simplBinder env bndr
382 ; completeNonRecX env' (isStrictId bndr) bndr bndr' new_rhs }
384 completeNonRecX :: SimplEnv
386 -> InId -- Old binder
387 -> OutId -- New binder
388 -> OutExpr -- Simplified RHS
391 completeNonRecX env is_strict old_bndr new_bndr new_rhs
392 = do { (env1, rhs1) <- prepareRhs (zapFloats env) new_rhs
394 if doFloatFromRhs NotTopLevel NonRecursive is_strict rhs1 env1
395 then do { tick LetFloatFromLet
396 ; return (addFloats env env1, rhs1) } -- Add the floats to the main env
397 else return (env, wrapFloats env1 rhs1) -- Wrap the floats around the RHS
398 ; completeBind env2 NotTopLevel old_bndr new_bndr rhs2 }
401 {- No, no, no! Do not try preInlineUnconditionally in completeNonRecX
402 Doing so risks exponential behaviour, because new_rhs has been simplified once already
403 In the cases described by the folowing commment, postInlineUnconditionally will
404 catch many of the relevant cases.
405 -- This happens; for example, the case_bndr during case of
406 -- known constructor: case (a,b) of x { (p,q) -> ... }
407 -- Here x isn't mentioned in the RHS, so we don't want to
408 -- create the (dead) let-binding let x = (a,b) in ...
410 -- Similarly, single occurrences can be inlined vigourously
411 -- e.g. case (f x, g y) of (a,b) -> ....
412 -- If a,b occur once we can avoid constructing the let binding for them.
414 Furthermore in the case-binding case preInlineUnconditionally risks extra thunks
415 -- Consider case I# (quotInt# x y) of
416 -- I# v -> let w = J# v in ...
417 -- If we gaily inline (quotInt# x y) for v, we end up building an
419 -- let w = J# (quotInt# x y) in ...
420 -- because quotInt# can fail.
422 | preInlineUnconditionally env NotTopLevel bndr new_rhs
423 = thing_inside (extendIdSubst env bndr (DoneEx new_rhs))
426 ----------------------------------
427 prepareRhs takes a putative RHS, checks whether it's a PAP or
428 constructor application and, if so, converts it to ANF, so that the
429 resulting thing can be inlined more easily. Thus
436 We also want to deal well cases like this
437 v = (f e1 `cast` co) e2
438 Here we want to make e1,e2 trivial and get
439 x1 = e1; x2 = e2; v = (f x1 `cast` co) v2
440 That's what the 'go' loop in prepareRhs does
443 prepareRhs :: SimplEnv -> OutExpr -> SimplM (SimplEnv, OutExpr)
444 -- Adds new floats to the env iff that allows us to return a good RHS
445 prepareRhs env (Cast rhs co) -- Note [Float coercions]
446 | (ty1, _ty2) <- coercionKind co -- Do *not* do this if rhs has an unlifted type
447 , not (isUnLiftedType ty1) -- see Note [Float coercions (unlifted)]
448 = do { (env', rhs') <- makeTrivial env rhs
449 ; return (env', Cast rhs' co) }
452 = do { (_is_val, env1, rhs1) <- go 0 env0 rhs0
453 ; return (env1, rhs1) }
455 go n_val_args env (Cast rhs co)
456 = do { (is_val, env', rhs') <- go n_val_args env rhs
457 ; return (is_val, env', Cast rhs' co) }
458 go n_val_args env (App fun (Type ty))
459 = do { (is_val, env', rhs') <- go n_val_args env fun
460 ; return (is_val, env', App rhs' (Type ty)) }
461 go n_val_args env (App fun arg)
462 = do { (is_val, env', fun') <- go (n_val_args+1) env fun
464 True -> do { (env'', arg') <- makeTrivial env' arg
465 ; return (True, env'', App fun' arg') }
466 False -> return (False, env, App fun arg) }
467 go n_val_args env (Var fun)
468 = return (is_val, env, Var fun)
470 is_val = n_val_args > 0 -- There is at least one arg
471 -- ...and the fun a constructor or PAP
472 && (isDataConWorkId fun || n_val_args < idArity fun)
474 = return (False, env, other)
478 Note [Float coercions]
479 ~~~~~~~~~~~~~~~~~~~~~~
480 When we find the binding
482 we'd like to transform it to
484 x = x `cast` co -- A trivial binding
485 There's a chance that e will be a constructor application or function, or something
486 like that, so moving the coerion to the usage site may well cancel the coersions
487 and lead to further optimisation. Example:
490 data instance T Int = T Int
492 foo :: Int -> Int -> Int
497 go n = case x of { T m -> go (n-m) }
498 -- This case should optimise
500 Note [Float coercions (unlifted)]
501 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
502 BUT don't do [Float coercions] if 'e' has an unlifted type.
505 foo :: Int = (error (# Int,Int #) "urk")
506 `cast` CoUnsafe (# Int,Int #) Int
508 If do the makeTrivial thing to the error call, we'll get
509 foo = case error (# Int,Int #) "urk" of v -> v `cast` ...
510 But 'v' isn't in scope!
512 These strange casts can happen as a result of case-of-case
513 bar = case (case x of { T -> (# 2,3 #); F -> error "urk" }) of
518 makeTrivial :: SimplEnv -> OutExpr -> SimplM (SimplEnv, OutExpr)
519 -- Binds the expression to a variable, if it's not trivial, returning the variable
523 | otherwise -- See Note [Take care] below
524 = do { var <- newId (fsLit "a") (exprType expr)
525 ; env' <- completeNonRecX env False var var expr
526 ; return (env', substExpr env' (Var var)) }
530 %************************************************************************
532 \subsection{Completing a lazy binding}
534 %************************************************************************
537 * deals only with Ids, not TyVars
538 * takes an already-simplified binder and RHS
539 * is used for both recursive and non-recursive bindings
540 * is used for both top-level and non-top-level bindings
542 It does the following:
543 - tries discarding a dead binding
544 - tries PostInlineUnconditionally
545 - add unfolding [this is the only place we add an unfolding]
548 It does *not* attempt to do let-to-case. Why? Because it is used for
549 - top-level bindings (when let-to-case is impossible)
550 - many situations where the "rhs" is known to be a WHNF
551 (so let-to-case is inappropriate).
553 Nor does it do the atomic-argument thing
556 completeBind :: SimplEnv
557 -> TopLevelFlag -- Flag stuck into unfolding
558 -> InId -- Old binder
559 -> OutId -> OutExpr -- New binder and RHS
561 -- completeBind may choose to do its work
562 -- * by extending the substitution (e.g. let x = y in ...)
563 -- * or by adding to the floats in the envt
565 completeBind env top_lvl old_bndr new_bndr new_rhs
566 | postInlineUnconditionally env top_lvl new_bndr occ_info new_rhs unfolding
567 -- Inline and discard the binding
568 = do { tick (PostInlineUnconditionally old_bndr)
569 ; -- pprTrace "postInlineUnconditionally" (ppr old_bndr <+> ppr new_bndr <+> ppr new_rhs) $
570 return (extendIdSubst env old_bndr (DoneEx new_rhs)) }
571 -- Use the substitution to make quite, quite sure that the
572 -- substitution will happen, since we are going to discard the binding
577 new_bndr_info = idInfo new_bndr `setArityInfo` exprArity new_rhs
580 -- Add the unfolding *only* for non-loop-breakers
581 -- Making loop breakers not have an unfolding at all
582 -- means that we can avoid tests in exprIsConApp, for example.
583 -- This is important: if exprIsConApp says 'yes' for a recursive
584 -- thing, then we can get into an infinite loop
587 -- If the unfolding is a value, the demand info may
588 -- go pear-shaped, so we nuke it. Example:
590 -- case x of (p,q) -> h p q x
591 -- Here x is certainly demanded. But after we've nuked
592 -- the case, we'll get just
593 -- let x = (a,b) in h a b x
594 -- and now x is not demanded (I'm assuming h is lazy)
595 -- This really happens. Similarly
596 -- let f = \x -> e in ...f..f...
597 -- After inlining f at some of its call sites the original binding may
598 -- (for example) be no longer strictly demanded.
599 -- The solution here is a bit ad hoc...
600 info_w_unf = new_bndr_info `setUnfoldingInfo` unfolding
601 `setWorkerInfo` worker_info
603 final_info | omit_unfolding = new_bndr_info
604 | isEvaldUnfolding unfolding = zapDemandInfo info_w_unf `orElse` info_w_unf
605 | otherwise = info_w_unf
607 final_id = new_bndr `setIdInfo` final_info
609 -- These seqs forces the Id, and hence its IdInfo,
610 -- and hence any inner substitutions
612 -- pprTrace "Binding" (ppr final_id <+> ppr unfolding) $
613 return (addNonRec env final_id new_rhs)
614 -- The addNonRec adds it to the in-scope set too
616 unfolding = mkUnfolding (isTopLevel top_lvl) new_rhs
617 worker_info = substWorker env (workerInfo old_info)
618 omit_unfolding = isNonRuleLoopBreaker occ_info || not (activeInline env old_bndr)
619 old_info = idInfo old_bndr
620 occ_info = occInfo old_info
625 %************************************************************************
627 \subsection[Simplify-simplExpr]{The main function: simplExpr}
629 %************************************************************************
631 The reason for this OutExprStuff stuff is that we want to float *after*
632 simplifying a RHS, not before. If we do so naively we get quadratic
633 behaviour as things float out.
635 To see why it's important to do it after, consider this (real) example:
649 a -- Can't inline a this round, cos it appears twice
653 Each of the ==> steps is a round of simplification. We'd save a
654 whole round if we float first. This can cascade. Consider
659 let f = let d1 = ..d.. in \y -> e
663 in \x -> ...(\y ->e)...
665 Only in this second round can the \y be applied, and it
666 might do the same again.
670 simplExpr :: SimplEnv -> CoreExpr -> SimplM CoreExpr
671 simplExpr env expr = simplExprC env expr mkBoringStop
673 simplExprC :: SimplEnv -> CoreExpr -> SimplCont -> SimplM CoreExpr
674 -- Simplify an expression, given a continuation
675 simplExprC env expr cont
676 = -- pprTrace "simplExprC" (ppr expr $$ ppr cont {- $$ ppr (seIdSubst env) -} $$ ppr (seFloats env) ) $
677 do { (env', expr') <- simplExprF (zapFloats env) expr cont
678 ; -- pprTrace "simplExprC ret" (ppr expr $$ ppr expr') $
679 -- pprTrace "simplExprC ret3" (ppr (seInScope env')) $
680 -- pprTrace "simplExprC ret4" (ppr (seFloats env')) $
681 return (wrapFloats env' expr') }
683 --------------------------------------------------
684 simplExprF :: SimplEnv -> InExpr -> SimplCont
685 -> SimplM (SimplEnv, OutExpr)
687 simplExprF env e cont
688 = -- pprTrace "simplExprF" (ppr e $$ ppr cont $$ ppr (seTvSubst env) $$ ppr (seIdSubst env) {- $$ ppr (seFloats env) -} ) $
689 simplExprF' env e cont
691 simplExprF' :: SimplEnv -> InExpr -> SimplCont
692 -> SimplM (SimplEnv, OutExpr)
693 simplExprF' env (Var v) cont = simplVar env v cont
694 simplExprF' env (Lit lit) cont = rebuild env (Lit lit) cont
695 simplExprF' env (Note n expr) cont = simplNote env n expr cont
696 simplExprF' env (Cast body co) cont = simplCast env body co cont
697 simplExprF' env (App fun arg) cont = simplExprF env fun $
698 ApplyTo NoDup arg env cont
700 simplExprF' env expr@(Lam _ _) cont
701 = simplLam env (map zap bndrs) body cont
702 -- The main issue here is under-saturated lambdas
703 -- (\x1. \x2. e) arg1
704 -- Here x1 might have "occurs-once" occ-info, because occ-info
705 -- is computed assuming that a group of lambdas is applied
706 -- all at once. If there are too few args, we must zap the
709 n_args = countArgs cont
710 n_params = length bndrs
711 (bndrs, body) = collectBinders expr
712 zap | n_args >= n_params = \b -> b
713 | otherwise = \b -> if isTyVar b then b
715 -- NB: we count all the args incl type args
716 -- so we must count all the binders (incl type lambdas)
718 simplExprF' env (Type ty) cont
719 = ASSERT( contIsRhsOrArg cont )
720 do { ty' <- simplType env ty
721 ; rebuild env (Type ty') cont }
723 simplExprF' env (Case scrut bndr _ alts) cont
724 | not (switchIsOn (getSwitchChecker env) NoCaseOfCase)
725 = -- Simplify the scrutinee with a Select continuation
726 simplExprF env scrut (Select NoDup bndr alts env cont)
729 = -- If case-of-case is off, simply simplify the case expression
730 -- in a vanilla Stop context, and rebuild the result around it
731 do { case_expr' <- simplExprC env scrut case_cont
732 ; rebuild env case_expr' cont }
734 case_cont = Select NoDup bndr alts env mkBoringStop
736 simplExprF' env (Let (Rec pairs) body) cont
737 = do { env' <- simplRecBndrs env (map fst pairs)
738 -- NB: bndrs' don't have unfoldings or rules
739 -- We add them as we go down
741 ; env'' <- simplRecBind env' NotTopLevel pairs
742 ; simplExprF env'' body cont }
744 simplExprF' env (Let (NonRec bndr rhs) body) cont
745 = simplNonRecE env bndr (rhs, env) ([], body) cont
747 ---------------------------------
748 simplType :: SimplEnv -> InType -> SimplM OutType
749 -- Kept monadic just so we can do the seqType
751 = -- pprTrace "simplType" (ppr ty $$ ppr (seTvSubst env)) $
752 seqType new_ty `seq` return new_ty
754 new_ty = substTy env ty
758 %************************************************************************
760 \subsection{The main rebuilder}
762 %************************************************************************
765 rebuild :: SimplEnv -> OutExpr -> SimplCont -> SimplM (SimplEnv, OutExpr)
766 -- At this point the substitution in the SimplEnv should be irrelevant
767 -- only the in-scope set and floats should matter
768 rebuild env expr cont0
769 = -- pprTrace "rebuild" (ppr expr $$ ppr cont0 $$ ppr (seFloats env)) $
771 Stop {} -> return (env, expr)
772 CoerceIt co cont -> rebuild env (mkCoerce co expr) cont
773 Select _ bndr alts se cont -> rebuildCase (se `setFloats` env) expr bndr alts cont
774 StrictArg fun _ info cont -> rebuildCall env (fun `App` expr) info cont
775 StrictBind b bs body se cont -> do { env' <- simplNonRecX (se `setFloats` env) b expr
776 ; simplLam env' bs body cont }
777 ApplyTo _ arg se cont -> do { arg' <- simplExpr (se `setInScope` env) arg
778 ; rebuild env (App expr arg') cont }
782 %************************************************************************
786 %************************************************************************
789 simplCast :: SimplEnv -> InExpr -> Coercion -> SimplCont
790 -> SimplM (SimplEnv, OutExpr)
791 simplCast env body co0 cont0
792 = do { co1 <- simplType env co0
793 ; simplExprF env body (addCoerce co1 cont0) }
795 addCoerce co cont = add_coerce co (coercionKind co) cont
797 add_coerce _co (s1, k1) cont -- co :: ty~ty
798 | s1 `coreEqType` k1 = cont -- is a no-op
800 add_coerce co1 (s1, _k2) (CoerceIt co2 cont)
801 | (_l1, t1) <- coercionKind co2
802 -- coerce T1 S1 (coerce S1 K1 e)
805 -- coerce T1 K1 e, otherwise
807 -- For example, in the initial form of a worker
808 -- we may find (coerce T (coerce S (\x.e))) y
809 -- and we'd like it to simplify to e[y/x] in one round
811 , s1 `coreEqType` t1 = cont -- The coerces cancel out
812 | otherwise = CoerceIt (mkTransCoercion co1 co2) cont
814 add_coerce co (s1s2, _t1t2) (ApplyTo dup (Type arg_ty) arg_se cont)
815 -- (f `cast` g) ty ---> (f ty) `cast` (g @ ty)
816 -- This implements the PushT rule from the paper
817 | Just (tyvar,_) <- splitForAllTy_maybe s1s2
818 , not (isCoVar tyvar)
819 = ApplyTo dup (Type ty') (zapSubstEnv env) (addCoerce (mkInstCoercion co ty') cont)
821 ty' = substTy (arg_se `setInScope` env) arg_ty
823 -- ToDo: the PushC rule is not implemented at all
825 add_coerce co (s1s2, _t1t2) (ApplyTo dup arg arg_se cont)
826 | not (isTypeArg arg) -- This implements the Push rule from the paper
827 , isFunTy s1s2 -- t1t2 must be a function type, becuase it's applied
828 -- co : s1s2 :=: t1t2
829 -- (coerce (T1->T2) (S1->S2) F) E
831 -- coerce T2 S2 (F (coerce S1 T1 E))
833 -- t1t2 must be a function type, T1->T2, because it's applied
834 -- to something but s1s2 might conceivably not be
836 -- When we build the ApplyTo we can't mix the out-types
837 -- with the InExpr in the argument, so we simply substitute
838 -- to make it all consistent. It's a bit messy.
839 -- But it isn't a common case.
841 -- Example of use: Trac #995
842 = ApplyTo dup new_arg (zapSubstEnv env) (addCoerce co2 cont)
844 -- we split coercion t1->t2 :=: s1->s2 into t1 :=: s1 and
845 -- t2 :=: s2 with left and right on the curried form:
846 -- (->) t1 t2 :=: (->) s1 s2
847 [co1, co2] = decomposeCo 2 co
848 new_arg = mkCoerce (mkSymCoercion co1) arg'
849 arg' = substExpr (arg_se `setInScope` env) arg
851 add_coerce co _ cont = CoerceIt co cont
855 %************************************************************************
859 %************************************************************************
862 simplLam :: SimplEnv -> [InId] -> InExpr -> SimplCont
863 -> SimplM (SimplEnv, OutExpr)
865 simplLam env [] body cont = simplExprF env body cont
868 simplLam env (bndr:bndrs) body (ApplyTo _ arg arg_se cont)
869 = do { tick (BetaReduction bndr)
870 ; simplNonRecE env bndr (arg, arg_se) (bndrs, body) cont }
872 -- Not enough args, so there are real lambdas left to put in the result
873 simplLam env bndrs body cont
874 = do { (env', bndrs') <- simplLamBndrs env bndrs
875 ; body' <- simplExpr env' body
876 ; new_lam <- mkLam bndrs' body'
877 ; rebuild env' new_lam cont }
880 simplNonRecE :: SimplEnv
881 -> InId -- The binder
882 -> (InExpr, SimplEnv) -- Rhs of binding (or arg of lambda)
883 -> ([InBndr], InExpr) -- Body of the let/lambda
886 -> SimplM (SimplEnv, OutExpr)
888 -- simplNonRecE is used for
889 -- * non-top-level non-recursive lets in expressions
892 -- It deals with strict bindings, via the StrictBind continuation,
893 -- which may abort the whole process
895 -- The "body" of the binding comes as a pair of ([InId],InExpr)
896 -- representing a lambda; so we recurse back to simplLam
897 -- Why? Because of the binder-occ-info-zapping done before
898 -- the call to simplLam in simplExprF (Lam ...)
900 -- First deal with type applications and type lets
901 -- (/\a. e) (Type ty) and (let a = Type ty in e)
902 simplNonRecE env bndr (Type ty_arg, rhs_se) (bndrs, body) cont
903 = ASSERT( isTyVar bndr )
904 do { ty_arg' <- simplType (rhs_se `setInScope` env) ty_arg
905 ; simplLam (extendTvSubst env bndr ty_arg') bndrs body cont }
907 simplNonRecE env bndr (rhs, rhs_se) (bndrs, body) cont
908 | preInlineUnconditionally env NotTopLevel bndr rhs
909 = do { tick (PreInlineUnconditionally bndr)
910 ; simplLam (extendIdSubst env bndr (mkContEx rhs_se rhs)) bndrs body cont }
913 = do { simplExprF (rhs_se `setFloats` env) rhs
914 (StrictBind bndr bndrs body env cont) }
917 = do { (env1, bndr1) <- simplNonRecBndr env bndr
918 ; let (env2, bndr2) = addBndrRules env1 bndr bndr1
919 ; env3 <- simplLazyBind env2 NotTopLevel NonRecursive bndr bndr2 rhs rhs_se
920 ; simplLam env3 bndrs body cont }
924 %************************************************************************
928 %************************************************************************
931 -- Hack alert: we only distinguish subsumed cost centre stacks for the
932 -- purposes of inlining. All other CCCSs are mapped to currentCCS.
933 simplNote :: SimplEnv -> Note -> CoreExpr -> SimplCont
934 -> SimplM (SimplEnv, OutExpr)
935 simplNote env (SCC cc) e cont
936 = do { e' <- simplExpr (setEnclosingCC env currentCCS) e
937 ; rebuild env (mkSCC cc e') cont }
939 -- See notes with SimplMonad.inlineMode
940 simplNote env InlineMe e cont
941 | Just (inside, outside) <- splitInlineCont cont -- Boring boring continuation; see notes above
942 = do { -- Don't inline inside an INLINE expression
943 e' <- simplExprC (setMode inlineMode env) e inside
944 ; rebuild env (mkInlineMe e') outside }
946 | otherwise -- Dissolve the InlineMe note if there's
947 -- an interesting context of any kind to combine with
948 -- (even a type application -- anything except Stop)
949 = simplExprF env e cont
951 simplNote env (CoreNote s) e cont = do
952 e' <- simplExpr env e
953 rebuild env (Note (CoreNote s) e') cont
957 %************************************************************************
959 \subsection{Dealing with calls}
961 %************************************************************************
964 simplVar :: SimplEnv -> Id -> SimplCont -> SimplM (SimplEnv, OutExpr)
965 simplVar env var cont
966 = case substId env var of
967 DoneEx e -> simplExprF (zapSubstEnv env) e cont
968 ContEx tvs ids e -> simplExprF (setSubstEnv env tvs ids) e cont
969 DoneId var1 -> completeCall (zapSubstEnv env) var1 cont
970 -- Note [zapSubstEnv]
971 -- The template is already simplified, so don't re-substitute.
972 -- This is VITAL. Consider
974 -- let y = \z -> ...x... in
976 -- We'll clone the inner \x, adding x->x' in the id_subst
977 -- Then when we inline y, we must *not* replace x by x' in
978 -- the inlined copy!!
980 ---------------------------------------------------------
981 -- Dealing with a call site
983 completeCall :: SimplEnv -> Id -> SimplCont -> SimplM (SimplEnv, OutExpr)
984 completeCall env var cont
985 = do { dflags <- getDOptsSmpl
986 ; let (args,call_cont) = contArgs cont
987 -- The args are OutExprs, obtained by *lazily* substituting
988 -- in the args found in cont. These args are only examined
989 -- to limited depth (unless a rule fires). But we must do
990 -- the substitution; rule matching on un-simplified args would
993 ------------- First try rules ----------------
994 -- Do this before trying inlining. Some functions have
995 -- rules *and* are strict; in this case, we don't want to
996 -- inline the wrapper of the non-specialised thing; better
997 -- to call the specialised thing instead.
999 -- We used to use the black-listing mechanism to ensure that inlining of
1000 -- the wrapper didn't occur for things that have specialisations till a
1001 -- later phase, so but now we just try RULES first
1003 -- Note [Rules for recursive functions]
1004 -- ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
1005 -- You might think that we shouldn't apply rules for a loop breaker:
1006 -- doing so might give rise to an infinite loop, because a RULE is
1007 -- rather like an extra equation for the function:
1008 -- RULE: f (g x) y = x+y
1011 -- But it's too drastic to disable rules for loop breakers.
1012 -- Even the foldr/build rule would be disabled, because foldr
1013 -- is recursive, and hence a loop breaker:
1014 -- foldr k z (build g) = g k z
1015 -- So it's up to the programmer: rules can cause divergence
1017 ; let in_scope = getInScope env
1018 maybe_rule = case activeRule dflags env of
1019 Nothing -> Nothing -- No rules apply
1020 Just act_fn -> lookupRule act_fn in_scope
1022 ; case maybe_rule of {
1023 Just (rule, rule_rhs) -> do
1024 tick (RuleFired (ru_name rule))
1025 (if dopt Opt_D_dump_rule_firings dflags then
1026 pprTrace "Rule fired" (vcat [
1027 text "Rule:" <+> ftext (ru_name rule),
1028 text "Before:" <+> ppr var <+> sep (map pprParendExpr args),
1029 text "After: " <+> pprCoreExpr rule_rhs,
1030 text "Cont: " <+> ppr call_cont])
1033 simplExprF env rule_rhs (dropArgs (ruleArity rule) cont)
1034 -- The ruleArity says how many args the rule consumed
1036 ; Nothing -> do -- No rules
1038 ------------- Next try inlining ----------------
1039 { let arg_infos = [interestingArg arg | arg <- args, isValArg arg]
1040 n_val_args = length arg_infos
1041 interesting_cont = interestingCallContext call_cont
1042 active_inline = activeInline env var
1043 maybe_inline = callSiteInline dflags active_inline var
1044 (null args) arg_infos interesting_cont
1045 ; case maybe_inline of {
1046 Just unfolding -- There is an inlining!
1047 -> do { tick (UnfoldingDone var)
1048 ; (if dopt Opt_D_dump_inlinings dflags then
1049 pprTrace ("Inlining done" ++ showSDoc (ppr var)) (vcat [
1050 text "Before:" <+> ppr var <+> sep (map pprParendExpr args),
1051 text "Inlined fn: " <+> nest 2 (ppr unfolding),
1052 text "Cont: " <+> ppr call_cont])
1055 simplExprF env unfolding cont }
1057 ; Nothing -> -- No inlining!
1059 ------------- No inlining! ----------------
1060 -- Next, look for rules or specialisations that match
1062 rebuildCall env (Var var)
1063 (mkArgInfo var n_val_args call_cont) cont
1066 rebuildCall :: SimplEnv
1067 -> OutExpr -- Function
1070 -> SimplM (SimplEnv, OutExpr)
1071 rebuildCall env fun (ArgInfo { ai_strs = [] }) cont
1072 -- When we run out of strictness args, it means
1073 -- that the call is definitely bottom; see SimplUtils.mkArgInfo
1074 -- Then we want to discard the entire strict continuation. E.g.
1075 -- * case (error "hello") of { ... }
1076 -- * (error "Hello") arg
1077 -- * f (error "Hello") where f is strict
1079 -- Then, especially in the first of these cases, we'd like to discard
1080 -- the continuation, leaving just the bottoming expression. But the
1081 -- type might not be right, so we may have to add a coerce.
1082 | not (contIsTrivial cont) -- Only do this if there is a non-trivial
1083 = return (env, mk_coerce fun) -- contination to discard, else we do it
1084 where -- again and again!
1085 fun_ty = exprType fun
1086 cont_ty = contResultType env fun_ty cont
1087 co = mkUnsafeCoercion fun_ty cont_ty
1088 mk_coerce expr | cont_ty `coreEqType` fun_ty = expr
1089 | otherwise = mkCoerce co expr
1091 rebuildCall env fun info (ApplyTo _ (Type arg_ty) se cont)
1092 = do { ty' <- simplType (se `setInScope` env) arg_ty
1093 ; rebuildCall env (fun `App` Type ty') info cont }
1096 (ArgInfo { ai_rules = has_rules, ai_strs = str:strs, ai_discs = disc:discs })
1097 (ApplyTo _ arg arg_se cont)
1098 | str -- Strict argument
1099 = -- pprTrace "Strict Arg" (ppr arg $$ ppr (seIdSubst env) $$ ppr (seInScope env)) $
1100 simplExprF (arg_se `setFloats` env) arg
1101 (StrictArg fun cci arg_info' cont)
1104 | otherwise -- Lazy argument
1105 -- DO NOT float anything outside, hence simplExprC
1106 -- There is no benefit (unlike in a let-binding), and we'd
1107 -- have to be very careful about bogus strictness through
1108 -- floating a demanded let.
1109 = do { arg' <- simplExprC (arg_se `setInScope` env) arg
1111 ; rebuildCall env (fun `App` arg') arg_info' cont }
1113 arg_info' = ArgInfo { ai_rules = has_rules, ai_strs = strs, ai_discs = discs }
1114 cci | has_rules || disc > 0 = ArgCtxt has_rules disc -- Be keener here
1115 | otherwise = BoringCtxt -- Nothing interesting
1117 rebuildCall env fun _ cont
1118 = rebuild env fun cont
1123 This part of the simplifier may break the no-shadowing invariant
1125 f (...(\a -> e)...) (case y of (a,b) -> e')
1126 where f is strict in its second arg
1127 If we simplify the innermost one first we get (...(\a -> e)...)
1128 Simplifying the second arg makes us float the case out, so we end up with
1129 case y of (a,b) -> f (...(\a -> e)...) e'
1130 So the output does not have the no-shadowing invariant. However, there is
1131 no danger of getting name-capture, because when the first arg was simplified
1132 we used an in-scope set that at least mentioned all the variables free in its
1133 static environment, and that is enough.
1135 We can't just do innermost first, or we'd end up with a dual problem:
1136 case x of (a,b) -> f e (...(\a -> e')...)
1138 I spent hours trying to recover the no-shadowing invariant, but I just could
1139 not think of an elegant way to do it. The simplifier is already knee-deep in
1140 continuations. We have to keep the right in-scope set around; AND we have
1141 to get the effect that finding (error "foo") in a strict arg position will
1142 discard the entire application and replace it with (error "foo"). Getting
1143 all this at once is TOO HARD!
1145 %************************************************************************
1147 Rebuilding a cse expression
1149 %************************************************************************
1151 Blob of helper functions for the "case-of-something-else" situation.
1154 ---------------------------------------------------------
1155 -- Eliminate the case if possible
1157 rebuildCase :: SimplEnv
1158 -> OutExpr -- Scrutinee
1159 -> InId -- Case binder
1160 -> [InAlt] -- Alternatives (inceasing order)
1162 -> SimplM (SimplEnv, OutExpr)
1164 --------------------------------------------------
1165 -- 1. Eliminate the case if there's a known constructor
1166 --------------------------------------------------
1168 rebuildCase env scrut case_bndr alts cont
1169 | Just (con,args) <- exprIsConApp_maybe scrut
1170 -- Works when the scrutinee is a variable with a known unfolding
1171 -- as well as when it's an explicit constructor application
1172 = knownCon env scrut (DataAlt con) args case_bndr alts cont
1174 | Lit lit <- scrut -- No need for same treatment as constructors
1175 -- because literals are inlined more vigorously
1176 = knownCon env scrut (LitAlt lit) [] case_bndr alts cont
1179 --------------------------------------------------
1180 -- 2. Eliminate the case if scrutinee is evaluated
1181 --------------------------------------------------
1183 rebuildCase env scrut case_bndr [(_, bndrs, rhs)] cont
1184 -- See if we can get rid of the case altogether
1185 -- See the extensive notes on case-elimination above
1186 -- mkCase made sure that if all the alternatives are equal,
1187 -- then there is now only one (DEFAULT) rhs
1188 | all isDeadBinder bndrs -- bndrs are [InId]
1190 -- Check that the scrutinee can be let-bound instead of case-bound
1191 , exprOkForSpeculation scrut
1192 -- OK not to evaluate it
1193 -- This includes things like (==# a# b#)::Bool
1194 -- so that we simplify
1195 -- case ==# a# b# of { True -> x; False -> x }
1198 -- This particular example shows up in default methods for
1199 -- comparision operations (e.g. in (>=) for Int.Int32)
1200 || exprIsHNF scrut -- It's already evaluated
1201 || var_demanded_later scrut -- It'll be demanded later
1203 -- || not opt_SimplPedanticBottoms) -- Or we don't care!
1204 -- We used to allow improving termination by discarding cases, unless -fpedantic-bottoms was on,
1205 -- but that breaks badly for the dataToTag# primop, which relies on a case to evaluate
1206 -- its argument: case x of { y -> dataToTag# y }
1207 -- Here we must *not* discard the case, because dataToTag# just fetches the tag from
1208 -- the info pointer. So we'll be pedantic all the time, and see if that gives any
1210 -- Also we don't want to discard 'seq's
1211 = do { tick (CaseElim case_bndr)
1212 ; env' <- simplNonRecX env case_bndr scrut
1213 ; simplExprF env' rhs cont }
1215 -- The case binder is going to be evaluated later,
1216 -- and the scrutinee is a simple variable
1217 var_demanded_later (Var v) = isStrictDmd (idNewDemandInfo case_bndr)
1218 && not (isTickBoxOp v)
1219 -- ugly hack; covering this case is what
1220 -- exprOkForSpeculation was intended for.
1221 var_demanded_later _ = False
1224 --------------------------------------------------
1225 -- 3. Catch-all case
1226 --------------------------------------------------
1228 rebuildCase env scrut case_bndr alts cont
1229 = do { -- Prepare the continuation;
1230 -- The new subst_env is in place
1231 (env', dup_cont, nodup_cont) <- prepareCaseCont env alts cont
1233 -- Simplify the alternatives
1234 ; (scrut', case_bndr', alts') <- simplAlts env' scrut case_bndr alts dup_cont
1236 -- Check for empty alternatives
1237 ; if null alts' then
1238 -- This isn't strictly an error, although it is unusual.
1239 -- It's possible that the simplifer might "see" that
1240 -- an inner case has no accessible alternatives before
1241 -- it "sees" that the entire branch of an outer case is
1242 -- inaccessible. So we simply put an error case here instead.
1243 pprTrace "mkCase: null alts" (ppr case_bndr <+> ppr scrut) $
1244 let res_ty' = contResultType env' (substTy env' (coreAltsType alts)) dup_cont
1245 lit = Lit (mkStringLit "Impossible alternative")
1246 in return (env', mkApps (Var rUNTIME_ERROR_ID) [Type res_ty', lit])
1249 { case_expr <- mkCase scrut' case_bndr' alts'
1251 -- Notice that rebuild gets the in-scope set from env, not alt_env
1252 -- The case binder *not* scope over the whole returned case-expression
1253 ; rebuild env' case_expr nodup_cont } }
1256 simplCaseBinder checks whether the scrutinee is a variable, v. If so,
1257 try to eliminate uses of v in the RHSs in favour of case_bndr; that
1258 way, there's a chance that v will now only be used once, and hence
1261 Note [no-case-of-case]
1262 ~~~~~~~~~~~~~~~~~~~~~~
1263 We *used* to suppress the binder-swap in case expressoins when
1264 -fno-case-of-case is on. Old remarks:
1265 "This happens in the first simplifier pass,
1266 and enhances full laziness. Here's the bad case:
1267 f = \ y -> ...(case x of I# v -> ...(case x of ...) ... )
1268 If we eliminate the inner case, we trap it inside the I# v -> arm,
1269 which might prevent some full laziness happening. I've seen this
1270 in action in spectral/cichelli/Prog.hs:
1271 [(m,n) | m <- [1..max], n <- [1..max]]
1272 Hence the check for NoCaseOfCase."
1273 However, now the full-laziness pass itself reverses the binder-swap, so this
1274 check is no longer necessary.
1276 Note [Suppressing the case binder-swap]
1277 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
1278 There is another situation when it might make sense to suppress the
1279 case-expression binde-swap. If we have
1281 case x of w1 { DEFAULT -> case x of w2 { A -> e1; B -> e2 }
1282 ...other cases .... }
1284 We'll perform the binder-swap for the outer case, giving
1286 case x of w1 { DEFAULT -> case w1 of w2 { A -> e1; B -> e2 }
1287 ...other cases .... }
1289 But there is no point in doing it for the inner case, because w1 can't
1290 be inlined anyway. Furthermore, doing the case-swapping involves
1291 zapping w2's occurrence info (see paragraphs that follow), and that
1292 forces us to bind w2 when doing case merging. So we get
1294 case x of w1 { A -> let w2 = w1 in e1
1295 B -> let w2 = w1 in e2
1296 ...other cases .... }
1298 This is plain silly in the common case where w2 is dead.
1300 Even so, I can't see a good way to implement this idea. I tried
1301 not doing the binder-swap if the scrutinee was already evaluated
1302 but that failed big-time:
1306 case v of w { MkT x ->
1307 case x of x1 { I# y1 ->
1308 case x of x2 { I# y2 -> ...
1310 Notice that because MkT is strict, x is marked "evaluated". But to
1311 eliminate the last case, we must either make sure that x (as well as
1312 x1) has unfolding MkT y1. THe straightforward thing to do is to do
1313 the binder-swap. So this whole note is a no-op.
1317 If we replace the scrutinee, v, by tbe case binder, then we have to nuke
1318 any occurrence info (eg IAmDead) in the case binder, because the
1319 case-binder now effectively occurs whenever v does. AND we have to do
1320 the same for the pattern-bound variables! Example:
1322 (case x of { (a,b) -> a }) (case x of { (p,q) -> q })
1324 Here, b and p are dead. But when we move the argment inside the first
1325 case RHS, and eliminate the second case, we get
1327 case x of { (a,b) -> a b }
1329 Urk! b is alive! Reason: the scrutinee was a variable, and case elimination
1332 Indeed, this can happen anytime the case binder isn't dead:
1333 case <any> of x { (a,b) ->
1334 case x of { (p,q) -> p } }
1335 Here (a,b) both look dead, but come alive after the inner case is eliminated.
1336 The point is that we bring into the envt a binding
1338 after the outer case, and that makes (a,b) alive. At least we do unless
1339 the case binder is guaranteed dead.
1343 Consider case (v `cast` co) of x { I# ->
1344 ... (case (v `cast` co) of {...}) ...
1345 We'd like to eliminate the inner case. We can get this neatly by
1346 arranging that inside the outer case we add the unfolding
1347 v |-> x `cast` (sym co)
1348 to v. Then we should inline v at the inner case, cancel the casts, and away we go
1350 Note [Improving seq]
1353 type family F :: * -> *
1354 type instance F Int = Int
1356 ... case e of x { DEFAULT -> rhs } ...
1358 where x::F Int. Then we'd like to rewrite (F Int) to Int, getting
1360 case e `cast` co of x'::Int
1361 I# x# -> let x = x' `cast` sym co
1364 so that 'rhs' can take advantage of the form of x'. Notice that Note
1365 [Case of cast] may then apply to the result.
1367 This showed up in Roman's experiments. Example:
1368 foo :: F Int -> Int -> Int
1369 foo t n = t `seq` bar n
1372 bar n = bar (n - case t of TI i -> i)
1373 Here we'd like to avoid repeated evaluating t inside the loop, by
1374 taking advantage of the `seq`.
1376 At one point I did transformation in LiberateCase, but it's more robust here.
1377 (Otherwise, there's a danger that we'll simply drop the 'seq' altogether, before
1378 LiberateCase gets to see it.)
1380 Note [Case elimination]
1381 ~~~~~~~~~~~~~~~~~~~~~~~
1382 The case-elimination transformation discards redundant case expressions.
1383 Start with a simple situation:
1385 case x# of ===> e[x#/y#]
1388 (when x#, y# are of primitive type, of course). We can't (in general)
1389 do this for algebraic cases, because we might turn bottom into
1392 The code in SimplUtils.prepareAlts has the effect of generalise this
1393 idea to look for a case where we're scrutinising a variable, and we
1394 know that only the default case can match. For example:
1398 DEFAULT -> ...(case x of
1402 Here the inner case is first trimmed to have only one alternative, the
1403 DEFAULT, after which it's an instance of the previous case. This
1404 really only shows up in eliminating error-checking code.
1406 We also make sure that we deal with this very common case:
1411 Here we are using the case as a strict let; if x is used only once
1412 then we want to inline it. We have to be careful that this doesn't
1413 make the program terminate when it would have diverged before, so we
1415 - e is already evaluated (it may so if e is a variable)
1416 - x is used strictly, or
1418 Lastly, the code in SimplUtils.mkCase combines identical RHSs. So
1420 case e of ===> case e of DEFAULT -> r
1424 Now again the case may be elminated by the CaseElim transformation.
1427 Further notes about case elimination
1428 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
1429 Consider: test :: Integer -> IO ()
1432 Turns out that this compiles to:
1435 eta1 :: State# RealWorld ->
1436 case PrelNum.< eta PrelNum.zeroInteger of wild { __DEFAULT ->
1438 (PrelNum.jtos eta ($w[] @ Char))
1440 of wild1 { (# new_s, a4 #) -> PrelIO.lvl23 new_s }}
1442 Notice the strange '<' which has no effect at all. This is a funny one.
1443 It started like this:
1445 f x y = if x < 0 then jtos x
1446 else if y==0 then "" else jtos x
1448 At a particular call site we have (f v 1). So we inline to get
1450 if v < 0 then jtos x
1451 else if 1==0 then "" else jtos x
1453 Now simplify the 1==0 conditional:
1455 if v<0 then jtos v else jtos v
1457 Now common-up the two branches of the case:
1459 case (v<0) of DEFAULT -> jtos v
1461 Why don't we drop the case? Because it's strict in v. It's technically
1462 wrong to drop even unnecessary evaluations, and in practice they
1463 may be a result of 'seq' so we *definitely* don't want to drop those.
1464 I don't really know how to improve this situation.
1468 simplCaseBinder :: SimplEnv -> OutExpr -> OutId -> [InAlt]
1469 -> SimplM (SimplEnv, OutExpr, OutId)
1470 simplCaseBinder env0 scrut0 case_bndr0 alts
1471 = do { (env1, case_bndr1) <- simplBinder env0 case_bndr0
1473 ; fam_envs <- getFamEnvs
1474 ; (env2, scrut2, case_bndr2) <- improve_seq fam_envs env1 scrut0
1475 case_bndr0 case_bndr1 alts
1476 -- Note [Improving seq]
1478 ; let (env3, case_bndr3) = improve_case_bndr env2 scrut2 case_bndr2
1479 -- Note [Case of cast]
1481 ; return (env3, scrut2, case_bndr3) }
1484 improve_seq fam_envs env scrut case_bndr case_bndr1 [(DEFAULT,_,_)]
1485 | Just (co, ty2) <- topNormaliseType fam_envs (idType case_bndr1)
1486 = do { case_bndr2 <- newId (fsLit "nt") ty2
1487 ; let rhs = DoneEx (Var case_bndr2 `Cast` mkSymCoercion co)
1488 env2 = extendIdSubst env case_bndr rhs
1489 ; return (env2, scrut `Cast` co, case_bndr2) }
1491 improve_seq _ env scrut _ case_bndr1 _
1492 = return (env, scrut, case_bndr1)
1495 improve_case_bndr env scrut case_bndr
1496 -- See Note [no-case-of-case]
1497 -- | switchIsOn (getSwitchChecker env) NoCaseOfCase
1498 -- = (env, case_bndr)
1500 | otherwise -- Failed try; see Note [Suppressing the case binder-swap]
1501 -- not (isEvaldUnfolding (idUnfolding v))
1503 Var v -> (modifyInScope env1 v case_bndr', case_bndr')
1504 -- Note about using modifyInScope for v here
1505 -- We could extend the substitution instead, but it would be
1506 -- a hack because then the substitution wouldn't be idempotent
1507 -- any more (v is an OutId). And this does just as well.
1509 Cast (Var v) co -> (addBinderUnfolding env1 v rhs, case_bndr')
1511 rhs = Cast (Var case_bndr') (mkSymCoercion co)
1513 _ -> (env, case_bndr)
1515 case_bndr' = zapOccInfo case_bndr
1516 env1 = modifyInScope env case_bndr case_bndr'
1519 zapOccInfo :: InId -> InId -- See Note [zapOccInfo]
1520 zapOccInfo b = b `setIdOccInfo` NoOccInfo
1524 simplAlts does two things:
1526 1. Eliminate alternatives that cannot match, including the
1527 DEFAULT alternative.
1529 2. If the DEFAULT alternative can match only one possible constructor,
1530 then make that constructor explicit.
1532 case e of x { DEFAULT -> rhs }
1534 case e of x { (a,b) -> rhs }
1535 where the type is a single constructor type. This gives better code
1536 when rhs also scrutinises x or e.
1538 Here "cannot match" includes knowledge from GADTs
1540 It's a good idea do do this stuff before simplifying the alternatives, to
1541 avoid simplifying alternatives we know can't happen, and to come up with
1542 the list of constructors that are handled, to put into the IdInfo of the
1543 case binder, for use when simplifying the alternatives.
1545 Eliminating the default alternative in (1) isn't so obvious, but it can
1548 data Colour = Red | Green | Blue
1557 DEFAULT -> [ case y of ... ]
1559 If we inline h into f, the default case of the inlined h can't happen.
1560 If we don't notice this, we may end up filtering out *all* the cases
1561 of the inner case y, which give us nowhere to go!
1565 simplAlts :: SimplEnv
1567 -> InId -- Case binder
1568 -> [InAlt] -- Non-empty
1570 -> SimplM (OutExpr, OutId, [OutAlt]) -- Includes the continuation
1571 -- Like simplExpr, this just returns the simplified alternatives;
1572 -- it not return an environment
1574 simplAlts env scrut case_bndr alts cont'
1575 = -- pprTrace "simplAlts" (ppr alts $$ ppr (seIdSubst env)) $
1576 do { let alt_env = zapFloats env
1577 ; (alt_env', scrut', case_bndr') <- simplCaseBinder alt_env scrut case_bndr alts
1579 ; (imposs_deflt_cons, in_alts) <- prepareAlts alt_env' scrut case_bndr' alts
1581 ; alts' <- mapM (simplAlt alt_env' imposs_deflt_cons case_bndr' cont') in_alts
1582 ; return (scrut', case_bndr', alts') }
1584 ------------------------------------
1585 simplAlt :: SimplEnv
1586 -> [AltCon] -- These constructors can't be present when
1587 -- matching the DEFAULT alternative
1588 -> OutId -- The case binder
1593 simplAlt env imposs_deflt_cons case_bndr' cont' (DEFAULT, bndrs, rhs)
1594 = ASSERT( null bndrs )
1595 do { let env' = addBinderOtherCon env case_bndr' imposs_deflt_cons
1596 -- Record the constructors that the case-binder *can't* be.
1597 ; rhs' <- simplExprC env' rhs cont'
1598 ; return (DEFAULT, [], rhs') }
1600 simplAlt env _ case_bndr' cont' (LitAlt lit, bndrs, rhs)
1601 = ASSERT( null bndrs )
1602 do { let env' = addBinderUnfolding env case_bndr' (Lit lit)
1603 ; rhs' <- simplExprC env' rhs cont'
1604 ; return (LitAlt lit, [], rhs') }
1606 simplAlt env _ case_bndr' cont' (DataAlt con, vs, rhs)
1607 = do { -- Deal with the pattern-bound variables
1608 -- Mark the ones that are in ! positions in the
1609 -- data constructor as certainly-evaluated.
1610 -- NB: simplLamBinders preserves this eval info
1611 let vs_with_evals = add_evals (dataConRepStrictness con)
1612 ; (env', vs') <- simplLamBndrs env vs_with_evals
1614 -- Bind the case-binder to (con args)
1615 ; let inst_tys' = tyConAppArgs (idType case_bndr')
1616 con_args = map Type inst_tys' ++ varsToCoreExprs vs'
1617 env'' = addBinderUnfolding env' case_bndr'
1618 (mkConApp con con_args)
1620 ; rhs' <- simplExprC env'' rhs cont'
1621 ; return (DataAlt con, vs', rhs') }
1623 -- add_evals records the evaluated-ness of the bound variables of
1624 -- a case pattern. This is *important*. Consider
1625 -- data T = T !Int !Int
1627 -- case x of { T a b -> T (a+1) b }
1629 -- We really must record that b is already evaluated so that we don't
1630 -- go and re-evaluate it when constructing the result.
1631 -- See Note [Data-con worker strictness] in MkId.lhs
1636 go (v:vs') strs | isTyVar v = v : go vs' strs
1637 go (v:vs') (str:strs)
1638 | isMarkedStrict str = evald_v : go vs' strs
1639 | otherwise = zapped_v : go vs' strs
1641 zapped_v = zap_occ_info v
1642 evald_v = zapped_v `setIdUnfolding` evaldUnfolding
1643 go _ _ = pprPanic "cat_evals" (ppr con $$ ppr vs $$ ppr the_strs)
1645 -- zap_occ_info: if the case binder is alive, then we add the unfolding
1647 -- to the envt; so vs are now very much alive
1648 -- Note [Aug06] I can't see why this actually matters, but it's neater
1649 -- case e of t { (a,b) -> ...(case t of (p,q) -> p)... }
1650 -- ==> case e of t { (a,b) -> ...(a)... }
1651 -- Look, Ma, a is alive now.
1652 zap_occ_info | isDeadBinder case_bndr' = \ident -> ident
1653 | otherwise = zapOccInfo
1655 addBinderUnfolding :: SimplEnv -> Id -> CoreExpr -> SimplEnv
1656 addBinderUnfolding env bndr rhs
1657 = modifyInScope env bndr (bndr `setIdUnfolding` mkUnfolding False rhs)
1659 addBinderOtherCon :: SimplEnv -> Id -> [AltCon] -> SimplEnv
1660 addBinderOtherCon env bndr cons
1661 = modifyInScope env bndr (bndr `setIdUnfolding` mkOtherCon cons)
1665 %************************************************************************
1667 \subsection{Known constructor}
1669 %************************************************************************
1671 We are a bit careful with occurrence info. Here's an example
1673 (\x* -> case x of (a*, b) -> f a) (h v, e)
1675 where the * means "occurs once". This effectively becomes
1676 case (h v, e) of (a*, b) -> f a)
1678 let a* = h v; b = e in f a
1682 All this should happen in one sweep.
1685 knownCon :: SimplEnv -> OutExpr -> AltCon
1686 -> [OutExpr] -- Args *including* the universal args
1687 -> InId -> [InAlt] -> SimplCont
1688 -> SimplM (SimplEnv, OutExpr)
1690 knownCon env scrut con args bndr alts cont
1691 = do { tick (KnownBranch bndr)
1692 ; knownAlt env scrut args bndr (findAlt con alts) cont }
1694 knownAlt :: SimplEnv -> OutExpr -> [OutExpr]
1695 -> InId -> (AltCon, [CoreBndr], InExpr) -> SimplCont
1696 -> SimplM (SimplEnv, OutExpr)
1697 knownAlt env scrut _ bndr (DEFAULT, bs, rhs) cont
1699 do { env' <- simplNonRecX env bndr scrut
1700 -- This might give rise to a binding with non-atomic args
1701 -- like x = Node (f x) (g x)
1702 -- but simplNonRecX will atomic-ify it
1703 ; simplExprF env' rhs cont }
1705 knownAlt env scrut _ bndr (LitAlt _, bs, rhs) cont
1707 do { env' <- simplNonRecX env bndr scrut
1708 ; simplExprF env' rhs cont }
1710 knownAlt env scrut the_args bndr (DataAlt dc, bs, rhs) cont
1711 = do { let dead_bndr = isDeadBinder bndr -- bndr is an InId
1712 n_drop_tys = length (dataConUnivTyVars dc)
1713 ; env' <- bind_args env dead_bndr bs (drop n_drop_tys the_args)
1715 -- It's useful to bind bndr to scrut, rather than to a fresh
1716 -- binding x = Con arg1 .. argn
1717 -- because very often the scrut is a variable, so we avoid
1718 -- creating, and then subsequently eliminating, a let-binding
1719 -- BUT, if scrut is a not a variable, we must be careful
1720 -- about duplicating the arg redexes; in that case, make
1721 -- a new con-app from the args
1722 bndr_rhs = case scrut of
1725 con_app = mkConApp dc (take n_drop_tys the_args ++ con_args)
1726 con_args = [substExpr env' (varToCoreExpr b) | b <- bs]
1727 -- args are aready OutExprs, but bs are InIds
1729 ; env'' <- simplNonRecX env' bndr bndr_rhs
1730 ; -- pprTrace "knownCon2" (ppr bs $$ ppr rhs $$ ppr (seIdSubst env'')) $
1731 simplExprF env'' rhs cont }
1734 bind_args env' _ [] _ = return env'
1736 bind_args env' dead_bndr (b:bs') (Type ty : args)
1737 = ASSERT( isTyVar b )
1738 bind_args (extendTvSubst env' b ty) dead_bndr bs' args
1740 bind_args env' dead_bndr (b:bs') (arg : args)
1742 do { let b' = if dead_bndr then b else zapOccInfo b
1743 -- Note that the binder might be "dead", because it doesn't
1744 -- occur in the RHS; and simplNonRecX may therefore discard
1745 -- it via postInlineUnconditionally.
1746 -- Nevertheless we must keep it if the case-binder is alive,
1747 -- because it may be used in the con_app. See Note [zapOccInfo]
1748 ; env'' <- simplNonRecX env' b' arg
1749 ; bind_args env'' dead_bndr bs' args }
1752 pprPanic "bind_args" $ ppr dc $$ ppr bs $$ ppr the_args $$
1753 text "scrut:" <+> ppr scrut
1757 %************************************************************************
1759 \subsection{Duplicating continuations}
1761 %************************************************************************
1764 prepareCaseCont :: SimplEnv
1765 -> [InAlt] -> SimplCont
1766 -> SimplM (SimplEnv, SimplCont,SimplCont)
1767 -- Return a duplicatable continuation, a non-duplicable part
1768 -- plus some extra bindings (that scope over the entire
1771 -- No need to make it duplicatable if there's only one alternative
1772 prepareCaseCont env [_] cont = return (env, cont, mkBoringStop)
1773 prepareCaseCont env _ cont = mkDupableCont env cont
1777 mkDupableCont :: SimplEnv -> SimplCont
1778 -> SimplM (SimplEnv, SimplCont, SimplCont)
1780 mkDupableCont env cont
1781 | contIsDupable cont
1782 = return (env, cont, mkBoringStop)
1784 mkDupableCont _ (Stop {}) = panic "mkDupableCont" -- Handled by previous eqn
1786 mkDupableCont env (CoerceIt ty cont)
1787 = do { (env', dup, nodup) <- mkDupableCont env cont
1788 ; return (env', CoerceIt ty dup, nodup) }
1790 mkDupableCont env cont@(StrictBind {})
1791 = return (env, mkBoringStop, cont)
1792 -- See Note [Duplicating strict continuations]
1794 mkDupableCont env cont@(StrictArg {})
1795 = return (env, mkBoringStop, cont)
1796 -- See Note [Duplicating strict continuations]
1798 mkDupableCont env (ApplyTo _ arg se cont)
1799 = -- e.g. [...hole...] (...arg...)
1801 -- let a = ...arg...
1802 -- in [...hole...] a
1803 do { (env', dup_cont, nodup_cont) <- mkDupableCont env cont
1804 ; arg' <- simplExpr (se `setInScope` env') arg
1805 ; (env'', arg'') <- makeTrivial env' arg'
1806 ; let app_cont = ApplyTo OkToDup arg'' (zapSubstEnv env') dup_cont
1807 ; return (env'', app_cont, nodup_cont) }
1809 mkDupableCont env cont@(Select _ _ [(_, bs, _rhs)] _ _)
1810 -- See Note [Single-alternative case]
1811 -- | not (exprIsDupable rhs && contIsDupable case_cont)
1812 -- | not (isDeadBinder case_bndr)
1813 | all isDeadBinder bs -- InIds
1814 = return (env, mkBoringStop, cont)
1816 mkDupableCont env (Select _ case_bndr alts se cont)
1817 = -- e.g. (case [...hole...] of { pi -> ei })
1819 -- let ji = \xij -> ei
1820 -- in case [...hole...] of { pi -> ji xij }
1821 do { tick (CaseOfCase case_bndr)
1822 ; (env', dup_cont, nodup_cont) <- mkDupableCont env cont
1823 -- NB: call mkDupableCont here, *not* prepareCaseCont
1824 -- We must make a duplicable continuation, whereas prepareCaseCont
1825 -- doesn't when there is a single case branch
1827 ; let alt_env = se `setInScope` env'
1828 ; (alt_env', case_bndr') <- simplBinder alt_env case_bndr
1829 ; alts' <- mapM (simplAlt alt_env' [] case_bndr' dup_cont) alts
1830 -- Safe to say that there are no handled-cons for the DEFAULT case
1831 -- NB: simplBinder does not zap deadness occ-info, so
1832 -- a dead case_bndr' will still advertise its deadness
1833 -- This is really important because in
1834 -- case e of b { (# p,q #) -> ... }
1835 -- b is always dead, and indeed we are not allowed to bind b to (# p,q #),
1836 -- which might happen if e was an explicit unboxed pair and b wasn't marked dead.
1837 -- In the new alts we build, we have the new case binder, so it must retain
1839 -- NB: we don't use alt_env further; it has the substEnv for
1840 -- the alternatives, and we don't want that
1842 ; (env'', alts'') <- mkDupableAlts env' case_bndr' alts'
1843 ; return (env'', -- Note [Duplicated env]
1844 Select OkToDup case_bndr' alts'' (zapSubstEnv env'') mkBoringStop,
1848 mkDupableAlts :: SimplEnv -> OutId -> [InAlt]
1849 -> SimplM (SimplEnv, [InAlt])
1850 -- Absorbs the continuation into the new alternatives
1852 mkDupableAlts env case_bndr' the_alts
1855 go env0 [] = return (env0, [])
1857 = do { (env1, alt') <- mkDupableAlt env0 case_bndr' alt
1858 ; (env2, alts') <- go env1 alts
1859 ; return (env2, alt' : alts' ) }
1861 mkDupableAlt :: SimplEnv -> OutId -> (AltCon, [CoreBndr], CoreExpr)
1862 -> SimplM (SimplEnv, (AltCon, [CoreBndr], CoreExpr))
1863 mkDupableAlt env case_bndr' (con, bndrs', rhs')
1864 | exprIsDupable rhs' -- Note [Small alternative rhs]
1865 = return (env, (con, bndrs', rhs'))
1867 = do { let rhs_ty' = exprType rhs'
1868 used_bndrs' = filter abstract_over (case_bndr' : bndrs')
1870 | isTyVar bndr = True -- Abstract over all type variables just in case
1871 | otherwise = not (isDeadBinder bndr)
1872 -- The deadness info on the new Ids is preserved by simplBinders
1874 ; (final_bndrs', final_args) -- Note [Join point abstraction]
1875 <- if (any isId used_bndrs')
1876 then return (used_bndrs', varsToCoreExprs used_bndrs')
1877 else do { rw_id <- newId (fsLit "w") realWorldStatePrimTy
1878 ; return ([rw_id], [Var realWorldPrimId]) }
1880 ; join_bndr <- newId (fsLit "$j") (mkPiTypes final_bndrs' rhs_ty')
1881 -- Note [Funky mkPiTypes]
1883 ; let -- We make the lambdas into one-shot-lambdas. The
1884 -- join point is sure to be applied at most once, and doing so
1885 -- prevents the body of the join point being floated out by
1886 -- the full laziness pass
1887 really_final_bndrs = map one_shot final_bndrs'
1888 one_shot v | isId v = setOneShotLambda v
1890 join_rhs = mkLams really_final_bndrs rhs'
1891 join_call = mkApps (Var join_bndr) final_args
1893 ; return (addNonRec env join_bndr join_rhs, (con, bndrs', join_call)) }
1894 -- See Note [Duplicated env]
1897 Note [Duplicated env]
1898 ~~~~~~~~~~~~~~~~~~~~~
1899 Some of the alternatives are simplified, but have not been turned into a join point
1900 So they *must* have an zapped subst-env. So we can't use completeNonRecX to
1901 bind the join point, because it might to do PostInlineUnconditionally, and
1902 we'd lose that when zapping the subst-env. We could have a per-alt subst-env,
1903 but zapping it (as we do in mkDupableCont, the Select case) is safe, and
1904 at worst delays the join-point inlining.
1906 Note [Small alterantive rhs]
1907 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~
1908 It is worth checking for a small RHS because otherwise we
1909 get extra let bindings that may cause an extra iteration of the simplifier to
1910 inline back in place. Quite often the rhs is just a variable or constructor.
1911 The Ord instance of Maybe in PrelMaybe.lhs, for example, took several extra
1912 iterations because the version with the let bindings looked big, and so wasn't
1913 inlined, but after the join points had been inlined it looked smaller, and so
1916 NB: we have to check the size of rhs', not rhs.
1917 Duplicating a small InAlt might invalidate occurrence information
1918 However, if it *is* dupable, we return the *un* simplified alternative,
1919 because otherwise we'd need to pair it up with an empty subst-env....
1920 but we only have one env shared between all the alts.
1921 (Remember we must zap the subst-env before re-simplifying something).
1922 Rather than do this we simply agree to re-simplify the original (small) thing later.
1924 Note [Funky mkPiTypes]
1925 ~~~~~~~~~~~~~~~~~~~~~~
1926 Notice the funky mkPiTypes. If the contructor has existentials
1927 it's possible that the join point will be abstracted over
1928 type varaibles as well as term variables.
1929 Example: Suppose we have
1930 data T = forall t. C [t]
1932 case (case e of ...) of
1934 We get the join point
1935 let j :: forall t. [t] -> ...
1936 j = /\t \xs::[t] -> rhs
1938 case (case e of ...) of
1939 C t xs::[t] -> j t xs
1941 Note [Join point abstaction]
1942 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~
1943 If we try to lift a primitive-typed something out
1944 for let-binding-purposes, we will *caseify* it (!),
1945 with potentially-disastrous strictness results. So
1946 instead we turn it into a function: \v -> e
1947 where v::State# RealWorld#. The value passed to this function
1948 is realworld#, which generates (almost) no code.
1950 There's a slight infelicity here: we pass the overall
1951 case_bndr to all the join points if it's used in *any* RHS,
1952 because we don't know its usage in each RHS separately
1954 We used to say "&& isUnLiftedType rhs_ty'" here, but now
1955 we make the join point into a function whenever used_bndrs'
1956 is empty. This makes the join-point more CPR friendly.
1957 Consider: let j = if .. then I# 3 else I# 4
1958 in case .. of { A -> j; B -> j; C -> ... }
1960 Now CPR doesn't w/w j because it's a thunk, so
1961 that means that the enclosing function can't w/w either,
1962 which is a lose. Here's the example that happened in practice:
1963 kgmod :: Int -> Int -> Int
1964 kgmod x y = if x > 0 && y < 0 || x < 0 && y > 0
1968 I have seen a case alternative like this:
1970 It's a bit silly to add the realWorld dummy arg in this case, making
1973 (the \v alone is enough to make CPR happy) but I think it's rare
1975 Note [Duplicating strict continuations]
1976 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
1977 Do *not* duplicate StrictBind and StritArg continuations. We gain
1978 nothing by propagating them into the expressions, and we do lose a
1979 lot. Here's an example:
1980 && (case x of { T -> F; F -> T }) E
1981 Now, && is strict so we end up simplifying the case with
1982 an ArgOf continuation. If we let-bind it, we get
1984 let $j = \v -> && v E
1985 in simplExpr (case x of { T -> F; F -> T })
1987 And after simplifying more we get
1989 let $j = \v -> && v E
1990 in case x of { T -> $j F; F -> $j T }
1991 Which is a Very Bad Thing
1993 The desire not to duplicate is the entire reason that
1994 mkDupableCont returns a pair of continuations.
1996 The original plan had:
1997 e.g. (...strict-fn...) [...hole...]
1999 let $j = \a -> ...strict-fn...
2002 Note [Single-alternative cases]
2003 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
2004 This case is just like the ArgOf case. Here's an example:
2008 case (case x of I# x' ->
2010 True -> I# (negate# x')
2011 False -> I# x') of y {
2013 Because the (case x) has only one alternative, we'll transform to
2015 case (case x' <# 0# of
2016 True -> I# (negate# x')
2017 False -> I# x') of y {
2019 But now we do *NOT* want to make a join point etc, giving
2021 let $j = \y -> MkT y
2023 True -> $j (I# (negate# x'))
2025 In this case the $j will inline again, but suppose there was a big
2026 strict computation enclosing the orginal call to MkT. Then, it won't
2027 "see" the MkT any more, because it's big and won't get duplicated.
2028 And, what is worse, nothing was gained by the case-of-case transform.
2030 When should use this case of mkDupableCont?
2031 However, matching on *any* single-alternative case is a *disaster*;
2032 e.g. case (case ....) of (a,b) -> (# a,b #)
2033 We must push the outer case into the inner one!
2036 * Match [(DEFAULT,_,_)], but in the common case of Int,
2037 the alternative-filling-in code turned the outer case into
2038 case (...) of y { I# _ -> MkT y }
2040 * Match on single alternative plus (not (isDeadBinder case_bndr))
2041 Rationale: pushing the case inwards won't eliminate the construction.
2042 But there's a risk of
2043 case (...) of y { (a,b) -> let z=(a,b) in ... }
2044 Now y looks dead, but it'll come alive again. Still, this
2045 seems like the best option at the moment.
2047 * Match on single alternative plus (all (isDeadBinder bndrs))
2048 Rationale: this is essentially seq.
2050 * Match when the rhs is *not* duplicable, and hence would lead to a
2051 join point. This catches the disaster-case above. We can test
2052 the *un-simplified* rhs, which is fine. It might get bigger or
2053 smaller after simplification; if it gets smaller, this case might
2054 fire next time round. NB also that we must test contIsDupable
2055 case_cont *btoo, because case_cont might be big!
2057 HOWEVER: I found that this version doesn't work well, because
2058 we can get let x = case (...) of { small } in ...case x...
2059 When x is inlined into its full context, we find that it was a bad
2060 idea to have pushed the outer case inside the (...) case.