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"
11 import DynFlags ( dopt, DynFlag(Opt_D_dump_inlinings),
16 import SimplUtils ( mkCase, mkLam,
17 SimplCont(..), DupFlag(..), LetRhsFlag(..),
18 mkRhsStop, mkBoringStop, mkLazyArgStop, pushContArgs,
19 contResultType, countArgs, contIsDupable, contIsRhsOrArg,
20 getContArgs, interestingCallContext, interestingArg, isStrictType,
21 preInlineUnconditionally, postInlineUnconditionally,
22 interestingArgContext, inlineMode, activeInline, activeRule
24 import Id ( Id, idType, idInfo, idArity, isDataConWorkId,
25 idUnfolding, setIdUnfolding, isDeadBinder,
26 idNewDemandInfo, setIdInfo,
27 setIdOccInfo, zapLamIdInfo, setOneShotLambda
29 import IdInfo ( OccInfo(..), setArityInfo, zapDemandInfo,
30 setUnfoldingInfo, occInfo
32 import NewDemand ( isStrictDmd )
33 import TcGadt ( dataConCanMatch )
34 import DataCon ( dataConTyCon, dataConRepStrictness )
35 import TyCon ( tyConArity, isAlgTyCon, isNewTyCon, tyConDataCons_maybe )
37 import PprCore ( pprParendExpr, pprCoreExpr )
38 import CoreUnfold ( mkUnfolding, callSiteInline )
39 import CoreUtils ( exprIsDupable, exprIsTrivial, needsCaseBinding,
40 exprIsConApp_maybe, mkPiTypes, findAlt,
41 exprType, exprIsHNF, findDefault, mergeAlts,
42 exprOkForSpeculation, exprArity,
43 mkCoerce, mkSCC, mkInlineMe, applyTypeToArg,
46 import Rules ( lookupRule )
47 import BasicTypes ( isMarkedStrict )
48 import CostCentre ( currentCCS )
49 import Type ( TvSubstEnv, isUnLiftedType, seqType, tyConAppArgs, funArgTy,
50 coreEqType, splitTyConApp_maybe,
51 isTyVarTy, isFunTy, tcEqType
53 import Coercion ( Coercion, coercionKind,
54 mkTransCoercion, mkSymCoercion, splitCoercionKind_maybe, decomposeCo )
55 import VarEnv ( elemVarEnv, emptyVarEnv )
56 import TysPrim ( realWorldStatePrimTy )
57 import PrelInfo ( realWorldPrimId )
58 import BasicTypes ( TopLevelFlag(..), isTopLevel,
59 RecFlag(..), isNonRec, isNonRuleLoopBreaker
63 import Maybes ( orElse )
65 import Util ( notNull, filterOut )
69 The guts of the simplifier is in this module, but the driver loop for
70 the simplifier is in SimplCore.lhs.
73 -----------------------------------------
74 *** IMPORTANT NOTE ***
75 -----------------------------------------
76 The simplifier used to guarantee that the output had no shadowing, but
77 it does not do so any more. (Actually, it never did!) The reason is
78 documented with simplifyArgs.
81 -----------------------------------------
82 *** IMPORTANT NOTE ***
83 -----------------------------------------
84 Many parts of the simplifier return a bunch of "floats" as well as an
85 expression. This is wrapped as a datatype SimplUtils.FloatsWith.
87 All "floats" are let-binds, not case-binds, but some non-rec lets may
88 be unlifted (with RHS ok-for-speculation).
92 -----------------------------------------
93 ORGANISATION OF FUNCTIONS
94 -----------------------------------------
96 - simplify all top-level binders
97 - for NonRec, call simplRecOrTopPair
98 - for Rec, call simplRecBind
101 ------------------------------
102 simplExpr (applied lambda) ==> simplNonRecBind
103 simplExpr (Let (NonRec ...) ..) ==> simplNonRecBind
104 simplExpr (Let (Rec ...) ..) ==> simplify binders; simplRecBind
106 ------------------------------
107 simplRecBind [binders already simplfied]
108 - use simplRecOrTopPair on each pair in turn
110 simplRecOrTopPair [binder already simplified]
111 Used for: recursive bindings (top level and nested)
112 top-level non-recursive bindings
114 - check for PreInlineUnconditionally
118 Used for: non-top-level non-recursive bindings
119 beta reductions (which amount to the same thing)
120 Because it can deal with strict arts, it takes a
121 "thing-inside" and returns an expression
123 - check for PreInlineUnconditionally
124 - simplify binder, including its IdInfo
133 simplNonRecX: [given a *simplified* RHS, but an *unsimplified* binder]
134 Used for: binding case-binder and constr args in a known-constructor case
135 - check for PreInLineUnconditionally
139 ------------------------------
140 simplLazyBind: [binder already simplified, RHS not]
141 Used for: recursive bindings (top level and nested)
142 top-level non-recursive bindings
143 non-top-level, but *lazy* non-recursive bindings
144 [must not be strict or unboxed]
145 Returns floats + an augmented environment, not an expression
146 - substituteIdInfo and add result to in-scope
147 [so that rules are available in rec rhs]
150 - float if exposes constructor or PAP
154 completeNonRecX: [binder and rhs both simplified]
155 - if the the thing needs case binding (unlifted and not ok-for-spec)
161 completeLazyBind: [given a simplified RHS]
162 [used for both rec and non-rec bindings, top level and not]
163 - try PostInlineUnconditionally
164 - add unfolding [this is the only place we add an unfolding]
169 Right hand sides and arguments
170 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
171 In many ways we want to treat
172 (a) the right hand side of a let(rec), and
173 (b) a function argument
174 in the same way. But not always! In particular, we would
175 like to leave these arguments exactly as they are, so they
176 will match a RULE more easily.
181 It's harder to make the rule match if we ANF-ise the constructor,
182 or eta-expand the PAP:
184 f (let { a = g x; b = h x } in (a,b))
187 On the other hand if we see the let-defns
192 then we *do* want to ANF-ise and eta-expand, so that p and q
193 can be safely inlined.
195 Even floating lets out is a bit dubious. For let RHS's we float lets
196 out if that exposes a value, so that the value can be inlined more vigorously.
199 r = let x = e in (x,x)
201 Here, if we float the let out we'll expose a nice constructor. We did experiments
202 that showed this to be a generally good thing. But it was a bad thing to float
203 lets out unconditionally, because that meant they got allocated more often.
205 For function arguments, there's less reason to expose a constructor (it won't
206 get inlined). Just possibly it might make a rule match, but I'm pretty skeptical.
207 So for the moment we don't float lets out of function arguments either.
212 For eta expansion, we want to catch things like
214 case e of (a,b) -> \x -> case a of (p,q) -> \y -> r
216 If the \x was on the RHS of a let, we'd eta expand to bring the two
217 lambdas together. And in general that's a good thing to do. Perhaps
218 we should eta expand wherever we find a (value) lambda? Then the eta
219 expansion at a let RHS can concentrate solely on the PAP case.
222 %************************************************************************
224 \subsection{Bindings}
226 %************************************************************************
229 simplTopBinds :: SimplEnv -> [InBind] -> SimplM [OutBind]
231 simplTopBinds env binds
232 = -- Put all the top-level binders into scope at the start
233 -- so that if a transformation rule has unexpectedly brought
234 -- anything into scope, then we don't get a complaint about that.
235 -- It's rather as if the top-level binders were imported.
236 simplRecBndrs env (bindersOfBinds binds) `thenSmpl` \ (env, bndrs') ->
237 simpl_binds env binds bndrs' `thenSmpl` \ (floats, _) ->
238 freeTick SimplifierDone `thenSmpl_`
239 returnSmpl (floatBinds floats)
241 -- We need to track the zapped top-level binders, because
242 -- they should have their fragile IdInfo zapped (notably occurrence info)
243 -- That's why we run down binds and bndrs' simultaneously.
244 simpl_binds :: SimplEnv -> [InBind] -> [OutId] -> SimplM (FloatsWith ())
245 simpl_binds env [] bs = ASSERT( null bs ) returnSmpl (emptyFloats env, ())
246 simpl_binds env (bind:binds) bs = simpl_bind env bind bs `thenSmpl` \ (floats,env) ->
247 addFloats env floats $ \env ->
248 simpl_binds env binds (drop_bs bind bs)
250 drop_bs (NonRec _ _) (_ : bs) = bs
251 drop_bs (Rec prs) bs = drop (length prs) bs
253 simpl_bind env bind bs
254 = getDOptsSmpl `thenSmpl` \ dflags ->
255 if dopt Opt_D_dump_inlinings dflags then
256 pprTrace "SimplBind" (ppr (bindersOf bind)) $ simpl_bind1 env bind bs
258 simpl_bind1 env bind bs
260 simpl_bind1 env (NonRec b r) (b':_) = simplRecOrTopPair env TopLevel b b' r
261 simpl_bind1 env (Rec pairs) bs' = simplRecBind env TopLevel pairs bs'
265 %************************************************************************
267 \subsection{simplNonRec}
269 %************************************************************************
271 simplNonRecBind is used for
272 * non-top-level non-recursive lets in expressions
276 * An unsimplified (binder, rhs) pair
277 * The env for the RHS. It may not be the same as the
278 current env because the bind might occur via (\x.E) arg
280 It uses the CPS form because the binding might be strict, in which
281 case we might discard the continuation:
282 let x* = error "foo" in (...x...)
284 It needs to turn unlifted bindings into a @case@. They can arise
285 from, say: (\x -> e) (4# + 3#)
288 simplNonRecBind :: SimplEnv
290 -> InExpr -> SimplEnv -- Arg, with its subst-env
291 -> OutType -- Type of thing computed by the context
292 -> (SimplEnv -> SimplM FloatsWithExpr) -- The body
293 -> SimplM FloatsWithExpr
295 simplNonRecBind env bndr rhs rhs_se cont_ty thing_inside
297 = pprPanic "simplNonRecBind" (ppr bndr <+> ppr rhs)
300 simplNonRecBind env bndr rhs rhs_se cont_ty thing_inside
301 = simplNonRecBind' env bndr rhs rhs_se cont_ty thing_inside
303 simplNonRecBind' env bndr rhs rhs_se cont_ty thing_inside
304 | preInlineUnconditionally env NotTopLevel bndr rhs
305 = tick (PreInlineUnconditionally bndr) `thenSmpl_`
306 thing_inside (extendIdSubst env bndr (mkContEx rhs_se rhs))
308 | isStrictDmd (idNewDemandInfo bndr) || isStrictType bndr_ty -- A strict let
309 = -- Don't use simplBinder because that doesn't keep
310 -- fragile occurrence info in the substitution
311 simplNonRecBndr env bndr `thenSmpl` \ (env, bndr1) ->
312 simplStrictArg AnRhs env rhs rhs_se (idType bndr1) cont_ty $ \ env1 rhs1 ->
314 -- Now complete the binding and simplify the body
316 (env2,bndr2) = addLetIdInfo env1 bndr bndr1
318 completeNonRecX env2 True {- strict -} bndr bndr2 rhs1 thing_inside
320 | otherwise -- Normal, lazy case
321 = -- Don't use simplBinder because that doesn't keep
322 -- fragile occurrence info in the substitution
323 simplNonRecBndr env bndr `thenSmpl` \ (env, bndr') ->
324 simplLazyBind env NotTopLevel NonRecursive
325 bndr bndr' rhs rhs_se `thenSmpl` \ (floats, env) ->
326 addFloats env floats thing_inside
329 bndr_ty = idType bndr
332 A specialised variant of simplNonRec used when the RHS is already simplified, notably
333 in knownCon. It uses case-binding where necessary.
336 simplNonRecX :: SimplEnv
337 -> InId -- Old binder
338 -> OutExpr -- Simplified RHS
339 -> (SimplEnv -> SimplM FloatsWithExpr)
340 -> SimplM FloatsWithExpr
342 simplNonRecX env bndr new_rhs thing_inside
343 = do { (env, bndr') <- simplBinder env bndr
344 ; completeNonRecX env False {- Non-strict; pessimistic -}
345 bndr bndr' new_rhs thing_inside }
348 completeNonRecX :: SimplEnv
349 -> Bool -- Strict binding
350 -> InId -- Old binder
351 -> OutId -- New binder
352 -> OutExpr -- Simplified RHS
353 -> (SimplEnv -> SimplM FloatsWithExpr)
354 -> SimplM FloatsWithExpr
356 completeNonRecX env is_strict old_bndr new_bndr new_rhs thing_inside
357 | needsCaseBinding (idType new_bndr) new_rhs
358 -- Make this test *before* the preInlineUnconditionally
359 -- Consider case I# (quotInt# x y) of
360 -- I# v -> let w = J# v in ...
361 -- If we gaily inline (quotInt# x y) for v, we end up building an
363 -- let w = J# (quotInt# x y) in ...
364 -- because quotInt# can fail.
365 = do { (floats, body) <- thing_inside env
366 ; let body' = wrapFloats floats body
367 ; return (emptyFloats env, Case new_rhs new_bndr (exprType body)
368 [(DEFAULT, [], body')]) }
371 = -- Make the arguments atomic if necessary,
372 -- adding suitable bindings
373 mkAtomicArgsE env is_strict new_rhs $ \ env new_rhs ->
374 completeLazyBind env NotTopLevel
375 old_bndr new_bndr new_rhs `thenSmpl` \ (floats, env) ->
376 addFloats env floats thing_inside
378 {- No, no, no! Do not try preInlineUnconditionally in completeNonRecX
379 Doing so risks exponential behaviour, because new_rhs has been simplified once already
380 In the cases described by the folowing commment, postInlineUnconditionally will
381 catch many of the relevant cases.
382 -- This happens; for example, the case_bndr during case of
383 -- known constructor: case (a,b) of x { (p,q) -> ... }
384 -- Here x isn't mentioned in the RHS, so we don't want to
385 -- create the (dead) let-binding let x = (a,b) in ...
387 -- Similarly, single occurrences can be inlined vigourously
388 -- e.g. case (f x, g y) of (a,b) -> ....
389 -- If a,b occur once we can avoid constructing the let binding for them.
390 | preInlineUnconditionally env NotTopLevel bndr new_rhs
391 = thing_inside (extendIdSubst env bndr (DoneEx new_rhs))
393 -- NB: completeLazyBind uses postInlineUnconditionally; no need to do that here
398 %************************************************************************
400 \subsection{Lazy bindings}
402 %************************************************************************
404 simplRecBind is used for
405 * recursive bindings only
408 simplRecBind :: SimplEnv -> TopLevelFlag
409 -> [(InId, InExpr)] -> [OutId]
410 -> SimplM (FloatsWith SimplEnv)
411 simplRecBind env top_lvl pairs bndrs'
412 = go env pairs bndrs' `thenSmpl` \ (floats, env) ->
413 returnSmpl (flattenFloats floats, env)
415 go env [] _ = returnSmpl (emptyFloats env, env)
417 go env ((bndr, rhs) : pairs) (bndr' : bndrs')
418 = simplRecOrTopPair env top_lvl bndr bndr' rhs `thenSmpl` \ (floats, env) ->
419 addFloats env floats (\env -> go env pairs bndrs')
423 simplRecOrTopPair is used for
424 * recursive bindings (whether top level or not)
425 * top-level non-recursive bindings
427 It assumes the binder has already been simplified, but not its IdInfo.
430 simplRecOrTopPair :: SimplEnv
432 -> InId -> OutId -- Binder, both pre-and post simpl
433 -> InExpr -- The RHS and its environment
434 -> SimplM (FloatsWith SimplEnv)
436 simplRecOrTopPair env top_lvl bndr bndr' rhs
437 | preInlineUnconditionally env top_lvl bndr rhs -- Check for unconditional inline
438 = tick (PreInlineUnconditionally bndr) `thenSmpl_`
439 returnSmpl (emptyFloats env, extendIdSubst env bndr (mkContEx env rhs))
442 = simplLazyBind env top_lvl Recursive bndr bndr' rhs env
443 -- May not actually be recursive, but it doesn't matter
447 simplLazyBind is used for
448 * recursive bindings (whether top level or not)
449 * top-level non-recursive bindings
450 * non-top-level *lazy* non-recursive bindings
452 [Thus it deals with the lazy cases from simplNonRecBind, and all cases
453 from SimplRecOrTopBind]
456 1. It assumes that the binder is *already* simplified,
457 and is in scope, but not its IdInfo
459 2. It assumes that the binder type is lifted.
461 3. It does not check for pre-inline-unconditionallly;
462 that should have been done already.
465 simplLazyBind :: SimplEnv
466 -> TopLevelFlag -> RecFlag
467 -> InId -> OutId -- Binder, both pre-and post simpl
468 -> InExpr -> SimplEnv -- The RHS and its environment
469 -> SimplM (FloatsWith SimplEnv)
471 simplLazyBind env top_lvl is_rec bndr bndr1 rhs rhs_se
473 (env1,bndr2) = addLetIdInfo env bndr bndr1
474 rhs_env = setInScope rhs_se env1
475 is_top_level = isTopLevel top_lvl
476 ok_float_unlifted = not is_top_level && isNonRec is_rec
477 rhs_cont = mkRhsStop (idType bndr2)
479 -- Simplify the RHS; note the mkRhsStop, which tells
480 -- the simplifier that this is the RHS of a let.
481 simplExprF rhs_env rhs rhs_cont `thenSmpl` \ (floats, rhs1) ->
483 -- If any of the floats can't be floated, give up now
484 -- (The allLifted predicate says True for empty floats.)
485 if (not ok_float_unlifted && not (allLifted floats)) then
486 completeLazyBind env1 top_lvl bndr bndr2
487 (wrapFloats floats rhs1)
490 -- ANF-ise a constructor or PAP rhs
491 mkAtomicArgs ok_float_unlifted rhs1 `thenSmpl` \ (aux_binds, rhs2) ->
493 -- If the result is a PAP, float the floats out, else wrap them
494 -- By this time it's already been ANF-ised (if necessary)
495 if isEmptyFloats floats && isNilOL aux_binds then -- Shortcut a common case
496 completeLazyBind env1 top_lvl bndr bndr2 rhs2
498 else if is_top_level || exprIsTrivial rhs2 || exprIsHNF rhs2 then
499 -- WARNING: long dodgy argument coming up
500 -- WANTED: a better way to do this
502 -- We can't use "exprIsCheap" instead of exprIsHNF,
503 -- because that causes a strictness bug.
504 -- x = let y* = E in case (scc y) of { T -> F; F -> T}
505 -- The case expression is 'cheap', but it's wrong to transform to
506 -- y* = E; x = case (scc y) of {...}
507 -- Either we must be careful not to float demanded non-values, or
508 -- we must use exprIsHNF for the test, which ensures that the
509 -- thing is non-strict. So exprIsHNF => bindings are non-strict
510 -- I think. The WARN below tests for this.
512 -- We use exprIsTrivial here because we want to reveal lone variables.
513 -- E.g. let { x = letrec { y = E } in y } in ...
514 -- Here we definitely want to float the y=E defn.
515 -- exprIsHNF definitely isn't right for that.
517 -- Again, the floated binding can't be strict; if it's recursive it'll
518 -- be non-strict; if it's non-recursive it'd be inlined.
520 -- Note [SCC-and-exprIsTrivial]
522 -- y = let { x* = E } in scc "foo" x
523 -- then we do *not* want to float out the x binding, because
524 -- it's strict! Fortunately, exprIsTrivial replies False to
527 -- There's a subtlety here. There may be a binding (x* = e) in the
528 -- floats, where the '*' means 'will be demanded'. So is it safe
529 -- to float it out? Answer no, but it won't matter because
530 -- we only float if (a) arg' is a WHNF, or (b) it's going to top level
531 -- and so there can't be any 'will be demanded' bindings in the floats.
533 WARN( not (is_top_level || not (any demanded_float (floatBinds floats))),
534 ppr (filter demanded_float (floatBinds floats)) )
536 tick LetFloatFromLet `thenSmpl_` (
537 addFloats env1 floats $ \ env2 ->
538 addAtomicBinds env2 (fromOL aux_binds) $ \ env3 ->
539 completeLazyBind env3 top_lvl bndr bndr2 rhs2)
542 completeLazyBind env1 top_lvl bndr bndr2 (wrapFloats floats rhs1)
545 demanded_float (NonRec b r) = isStrictDmd (idNewDemandInfo b) && not (isUnLiftedType (idType b))
546 -- Unlifted-type (cheap-eagerness) lets may well have a demanded flag on them
547 demanded_float (Rec _) = False
552 %************************************************************************
554 \subsection{Completing a lazy binding}
556 %************************************************************************
559 * deals only with Ids, not TyVars
560 * takes an already-simplified binder and RHS
561 * is used for both recursive and non-recursive bindings
562 * is used for both top-level and non-top-level bindings
564 It does the following:
565 - tries discarding a dead binding
566 - tries PostInlineUnconditionally
567 - add unfolding [this is the only place we add an unfolding]
570 It does *not* attempt to do let-to-case. Why? Because it is used for
571 - top-level bindings (when let-to-case is impossible)
572 - many situations where the "rhs" is known to be a WHNF
573 (so let-to-case is inappropriate).
576 completeLazyBind :: SimplEnv
577 -> TopLevelFlag -- Flag stuck into unfolding
578 -> InId -- Old binder
579 -> OutId -- New binder
580 -> OutExpr -- Simplified RHS
581 -> SimplM (FloatsWith SimplEnv)
582 -- We return a new SimplEnv, because completeLazyBind may choose to do its work
583 -- by extending the substitution (e.g. let x = y in ...)
584 -- The new binding (if any) is returned as part of the floats.
585 -- NB: the returned SimplEnv has the right SubstEnv, but you should
586 -- (as usual) use the in-scope-env from the floats
588 completeLazyBind env top_lvl old_bndr new_bndr new_rhs
589 | postInlineUnconditionally env top_lvl new_bndr occ_info new_rhs unfolding
590 = -- Drop the binding
591 tick (PostInlineUnconditionally old_bndr) `thenSmpl_`
592 -- pprTrace "Inline unconditionally" (ppr old_bndr <+> ppr new_bndr <+> ppr new_rhs) $
593 returnSmpl (emptyFloats env, extendIdSubst env old_bndr (DoneEx new_rhs))
594 -- Use the substitution to make quite, quite sure that the substitution
595 -- will happen, since we are going to discard the binding
600 new_bndr_info = idInfo new_bndr `setArityInfo` exprArity new_rhs
603 -- Add the unfolding *only* for non-loop-breakers
604 -- Making loop breakers not have an unfolding at all
605 -- means that we can avoid tests in exprIsConApp, for example.
606 -- This is important: if exprIsConApp says 'yes' for a recursive
607 -- thing, then we can get into an infinite loop
610 -- If the unfolding is a value, the demand info may
611 -- go pear-shaped, so we nuke it. Example:
613 -- case x of (p,q) -> h p q x
614 -- Here x is certainly demanded. But after we've nuked
615 -- the case, we'll get just
616 -- let x = (a,b) in h a b x
617 -- and now x is not demanded (I'm assuming h is lazy)
618 -- This really happens. Similarly
619 -- let f = \x -> e in ...f..f...
620 -- After inlining f at some of its call sites the original binding may
621 -- (for example) be no longer strictly demanded.
622 -- The solution here is a bit ad hoc...
623 info_w_unf = new_bndr_info `setUnfoldingInfo` unfolding
624 final_info | loop_breaker = new_bndr_info
625 | isEvaldUnfolding unfolding = zapDemandInfo info_w_unf `orElse` info_w_unf
626 | otherwise = info_w_unf
628 final_id = new_bndr `setIdInfo` final_info
630 -- These seqs forces the Id, and hence its IdInfo,
631 -- and hence any inner substitutions
633 -- pprTrace "Binding" (ppr final_id <+> ppr unfolding) $
634 returnSmpl (unitFloat env final_id new_rhs, env)
636 unfolding = mkUnfolding (isTopLevel top_lvl) new_rhs
637 loop_breaker = isNonRuleLoopBreaker occ_info
638 old_info = idInfo old_bndr
639 occ_info = occInfo old_info
644 %************************************************************************
646 \subsection[Simplify-simplExpr]{The main function: simplExpr}
648 %************************************************************************
650 The reason for this OutExprStuff stuff is that we want to float *after*
651 simplifying a RHS, not before. If we do so naively we get quadratic
652 behaviour as things float out.
654 To see why it's important to do it after, consider this (real) example:
668 a -- Can't inline a this round, cos it appears twice
672 Each of the ==> steps is a round of simplification. We'd save a
673 whole round if we float first. This can cascade. Consider
678 let f = let d1 = ..d.. in \y -> e
682 in \x -> ...(\y ->e)...
684 Only in this second round can the \y be applied, and it
685 might do the same again.
689 simplExpr :: SimplEnv -> CoreExpr -> SimplM CoreExpr
690 simplExpr env expr = simplExprC env expr (mkBoringStop expr_ty')
692 expr_ty' = substTy env (exprType expr)
693 -- The type in the Stop continuation, expr_ty', is usually not used
694 -- It's only needed when discarding continuations after finding
695 -- a function that returns bottom.
696 -- Hence the lazy substitution
699 simplExprC :: SimplEnv -> CoreExpr -> SimplCont -> SimplM CoreExpr
700 -- Simplify an expression, given a continuation
701 simplExprC env expr cont
702 = simplExprF env expr cont `thenSmpl` \ (floats, expr) ->
703 returnSmpl (wrapFloats floats expr)
705 simplExprF :: SimplEnv -> InExpr -> SimplCont -> SimplM FloatsWithExpr
706 -- Simplify an expression, returning floated binds
708 simplExprF env (Var v) cont = simplVar env v cont
709 simplExprF env (Lit lit) cont = rebuild env (Lit lit) cont
710 simplExprF env expr@(Lam _ _) cont = simplLam env expr cont
711 simplExprF env (Note note expr) cont = simplNote env note expr cont
712 simplExprF env (Cast body co) cont = simplCast env body co cont
713 simplExprF env (App fun arg) cont = simplExprF env fun
714 (ApplyTo NoDup arg (Just env) cont)
716 simplExprF env (Type ty) cont
717 = ASSERT( contIsRhsOrArg cont )
718 simplType env ty `thenSmpl` \ ty' ->
719 rebuild env (Type ty') cont
721 simplExprF env (Case scrut bndr case_ty alts) cont
722 | not (switchIsOn (getSwitchChecker env) NoCaseOfCase)
723 = -- Simplify the scrutinee with a Select continuation
724 simplExprF env scrut (Select NoDup bndr alts env cont)
727 = -- If case-of-case is off, simply simplify the case expression
728 -- in a vanilla Stop context, and rebuild the result around it
729 simplExprC env scrut case_cont `thenSmpl` \ case_expr' ->
730 rebuild env case_expr' cont
732 case_cont = Select NoDup bndr alts env (mkBoringStop case_ty')
733 case_ty' = substTy env case_ty -- c.f. defn of simplExpr
735 simplExprF env (Let (Rec pairs) body) cont
736 = simplRecBndrs env (map fst pairs) `thenSmpl` \ (env, bndrs') ->
737 -- NB: bndrs' don't have unfoldings or rules
738 -- We add them as we go down
740 simplRecBind env NotTopLevel pairs bndrs' `thenSmpl` \ (floats, env) ->
741 addFloats env floats $ \ env ->
742 simplExprF env body cont
744 -- A non-recursive let is dealt with by simplNonRecBind
745 simplExprF env (Let (NonRec bndr rhs) body) cont
746 = simplNonRecBind env bndr rhs env (contResultType cont) $ \ env ->
747 simplExprF env body cont
750 ---------------------------------
751 simplType :: SimplEnv -> InType -> SimplM OutType
752 -- Kept monadic just so we can do the seqType
754 = seqType new_ty `seq` returnSmpl new_ty
756 new_ty = substTy env ty
760 %************************************************************************
764 %************************************************************************
767 simplCast :: SimplEnv -> InExpr -> Coercion -> SimplCont -> SimplM FloatsWithExpr
768 simplCast env body co cont
771 | (s1, k1) <- coercionKind co
772 , s1 `tcEqType` k1 = cont
773 addCoerce co1 (CoerceIt co2 cont)
774 | (s1, k1) <- coercionKind co1
775 , (l1, t1) <- coercionKind co2
776 -- coerce T1 S1 (coerce S1 K1 e)
779 -- coerce T1 K1 e, otherwise
781 -- For example, in the initial form of a worker
782 -- we may find (coerce T (coerce S (\x.e))) y
783 -- and we'd like it to simplify to e[y/x] in one round
785 , s1 `coreEqType` t1 = cont -- The coerces cancel out
786 | otherwise = CoerceIt (mkTransCoercion co1 co2) cont
788 addCoerce co (ApplyTo dup arg arg_se cont)
789 | not (isTypeArg arg) -- This whole case only works for value args
790 -- Could upgrade to have equiv thing for type apps too
791 , Just (s1s2, t1t2) <- splitCoercionKind_maybe co
793 -- co : s1s2 :=: t1t2
794 -- (coerce (T1->T2) (S1->S2) F) E
796 -- coerce T2 S2 (F (coerce S1 T1 E))
798 -- t1t2 must be a function type, T1->T2, because it's applied
799 -- to something but s1s2 might conceivably not be
801 -- When we build the ApplyTo we can't mix the out-types
802 -- with the InExpr in the argument, so we simply substitute
803 -- to make it all consistent. It's a bit messy.
804 -- But it isn't a common case.
807 -- we split coercion t1->t2 :=: s1->s2 into t1 :=: s1 and
808 -- t2 :=: s2 with left and right on the curried form:
809 -- (->) t1 t2 :=: (->) s1 s2
810 [co1, co2] = decomposeCo 2 co
811 new_arg = mkCoerce (mkSymCoercion co1) arg'
812 arg' = case arg_se of
814 Just arg_se -> substExpr (setInScope arg_se env) arg
815 result = ApplyTo dup new_arg (Just $ zapSubstEnv env)
817 addCoerce co cont = CoerceIt co cont
819 simplType env co `thenSmpl` \ co' ->
820 simplExprF env body (addCoerce co' cont)
823 %************************************************************************
827 %************************************************************************
830 simplLam env fun cont
833 zap_it = mkLamBndrZapper fun (countArgs cont)
834 cont_ty = contResultType cont
836 -- Type-beta reduction
837 go env (Lam bndr body) (ApplyTo _ (Type ty_arg) mb_arg_se body_cont)
838 = ASSERT( isTyVar bndr )
839 do { tick (BetaReduction bndr)
840 ; ty_arg' <- case mb_arg_se of
841 Just arg_se -> simplType (setInScope arg_se env) ty_arg
842 Nothing -> return ty_arg
843 ; go (extendTvSubst env bndr ty_arg') body body_cont }
845 -- Ordinary beta reduction
846 go env (Lam bndr body) cont@(ApplyTo _ arg (Just arg_se) body_cont)
847 = do { tick (BetaReduction bndr)
848 ; simplNonRecBind env (zap_it bndr) arg arg_se cont_ty $ \ env ->
849 go env body body_cont }
851 go env (Lam bndr body) cont@(ApplyTo _ arg Nothing body_cont)
852 = do { tick (BetaReduction bndr)
853 ; simplNonRecX env (zap_it bndr) arg $ \ env ->
854 go env body body_cont }
856 -- Not enough args, so there are real lambdas left to put in the result
857 go env lam@(Lam _ _) cont
858 = do { (env, bndrs') <- simplLamBndrs env bndrs
859 ; body' <- simplExpr env body
860 ; (floats, new_lam) <- mkLam env bndrs' body' cont
861 ; addFloats env floats $ \ env ->
862 rebuild env new_lam cont }
864 (bndrs,body) = collectBinders lam
866 -- Exactly enough args
867 go env expr cont = simplExprF env expr cont
869 mkLamBndrZapper :: CoreExpr -- Function
870 -> Int -- Number of args supplied, *including* type args
871 -> Id -> Id -- Use this to zap the binders
872 mkLamBndrZapper fun n_args
873 | n_args >= n_params fun = \b -> b -- Enough args
874 | otherwise = \b -> zapLamIdInfo b
876 -- NB: we count all the args incl type args
877 -- so we must count all the binders (incl type lambdas)
878 n_params (Note _ e) = n_params e
879 n_params (Lam b e) = 1 + n_params e
880 n_params other = 0::Int
884 %************************************************************************
888 %************************************************************************
893 -- Hack: we only distinguish subsumed cost centre stacks for the purposes of
894 -- inlining. All other CCCSs are mapped to currentCCS.
895 simplNote env (SCC cc) e cont
896 = simplExpr (setEnclosingCC env currentCCS) e `thenSmpl` \ e' ->
897 rebuild env (mkSCC cc e') cont
899 -- See notes with SimplMonad.inlineMode
900 simplNote env InlineMe e cont
901 | contIsRhsOrArg cont -- Totally boring continuation; see notes above
902 = -- Don't inline inside an INLINE expression
903 simplExpr (setMode inlineMode env ) e `thenSmpl` \ e' ->
904 rebuild env (mkInlineMe e') cont
906 | otherwise -- Dissolve the InlineMe note if there's
907 -- an interesting context of any kind to combine with
908 -- (even a type application -- anything except Stop)
909 = simplExprF env e cont
911 simplNote env (CoreNote s) e cont
912 = simplExpr env e `thenSmpl` \ e' ->
913 rebuild env (Note (CoreNote s) e') cont
915 simplNote env note@(TickBox {}) e cont
916 = simplExpr env e `thenSmpl` \ e' ->
917 rebuild env (Note note e') cont
919 simplNote env note@(BinaryTickBox {}) e cont
920 = simplExpr env e `thenSmpl` \ e' ->
921 rebuild env (Note note e') cont
925 %************************************************************************
927 \subsection{Dealing with calls}
929 %************************************************************************
932 simplVar env var cont
933 = case substId env var of
934 DoneEx e -> simplExprF (zapSubstEnv env) e cont
935 ContEx tvs ids e -> simplExprF (setSubstEnv env tvs ids) e cont
936 DoneId var1 -> completeCall (zapSubstEnv env) var1 cont
937 -- Note [zapSubstEnv]
938 -- The template is already simplified, so don't re-substitute.
939 -- This is VITAL. Consider
941 -- let y = \z -> ...x... in
943 -- We'll clone the inner \x, adding x->x' in the id_subst
944 -- Then when we inline y, we must *not* replace x by x' in
945 -- the inlined copy!!
947 ---------------------------------------------------------
948 -- Dealing with a call site
950 completeCall env var cont
951 = -- Simplify the arguments
952 getDOptsSmpl `thenSmpl` \ dflags ->
954 chkr = getSwitchChecker env
955 (args, call_cont) = getContArgs chkr var cont
958 simplifyArgs env fn_ty (interestingArgContext var call_cont) args
959 (contResultType call_cont) $ \ env args ->
961 -- Next, look for rules or specialisations that match
963 -- It's important to simplify the args first, because the rule-matcher
964 -- doesn't do substitution as it goes. We don't want to use subst_args
965 -- (defined in the 'where') because that throws away useful occurrence info,
966 -- and perhaps-very-important specialisations.
968 -- Some functions have specialisations *and* are strict; in this case,
969 -- we don't want to inline the wrapper of the non-specialised thing; better
970 -- to call the specialised thing instead.
971 -- We used to use the black-listing mechanism to ensure that inlining of
972 -- the wrapper didn't occur for things that have specialisations till a
973 -- later phase, so but now we just try RULES first
975 -- You might think that we shouldn't apply rules for a loop breaker:
976 -- doing so might give rise to an infinite loop, because a RULE is
977 -- rather like an extra equation for the function:
978 -- RULE: f (g x) y = x+y
981 -- But it's too drastic to disable rules for loop breakers.
982 -- Even the foldr/build rule would be disabled, because foldr
983 -- is recursive, and hence a loop breaker:
984 -- foldr k z (build g) = g k z
985 -- So it's up to the programmer: rules can cause divergence
988 in_scope = getInScope env
990 maybe_rule = case activeRule env of
991 Nothing -> Nothing -- No rules apply
992 Just act_fn -> lookupRule act_fn in_scope rules var args
995 Just (rule_name, rule_rhs) ->
996 tick (RuleFired rule_name) `thenSmpl_`
997 (if dopt Opt_D_dump_inlinings dflags then
998 pprTrace "Rule fired" (vcat [
999 text "Rule:" <+> ftext rule_name,
1000 text "Before:" <+> ppr var <+> sep (map pprParendExpr args),
1001 text "After: " <+> pprCoreExpr rule_rhs,
1002 text "Cont: " <+> ppr call_cont])
1005 simplExprF env rule_rhs call_cont ;
1007 Nothing -> -- No rules
1009 -- Next, look for an inlining
1011 arg_infos = [ interestingArg arg | arg <- args, isValArg arg]
1012 interesting_cont = interestingCallContext (notNull args)
1015 active_inline = activeInline env var
1016 maybe_inline = callSiteInline dflags active_inline
1017 var arg_infos interesting_cont
1019 case maybe_inline of {
1020 Just unfolding -- There is an inlining!
1021 -> tick (UnfoldingDone var) `thenSmpl_`
1022 (if dopt Opt_D_dump_inlinings dflags then
1023 pprTrace "Inlining done" (vcat [
1024 text "Before:" <+> ppr var <+> sep (map pprParendExpr args),
1025 text "Inlined fn: " $$ nest 2 (ppr unfolding),
1026 text "Cont: " <+> ppr call_cont])
1029 simplExprF env unfolding (pushContArgs args call_cont)
1032 Nothing -> -- No inlining!
1035 rebuild env (mkApps (Var var) args) call_cont
1039 %************************************************************************
1041 \subsection{Arguments}
1043 %************************************************************************
1046 ---------------------------------------------------------
1047 -- Simplifying the arguments of a call
1049 simplifyArgs :: SimplEnv
1050 -> OutType -- Type of the function
1051 -> Bool -- True if the fn has RULES
1052 -> [(InExpr, Maybe SimplEnv, Bool)] -- Details of the arguments
1053 -> OutType -- Type of the continuation
1054 -> (SimplEnv -> [OutExpr] -> SimplM FloatsWithExpr)
1055 -> SimplM FloatsWithExpr
1057 -- [CPS-like because of strict arguments]
1059 -- Simplify the arguments to a call.
1060 -- This part of the simplifier may break the no-shadowing invariant
1062 -- f (...(\a -> e)...) (case y of (a,b) -> e')
1063 -- where f is strict in its second arg
1064 -- If we simplify the innermost one first we get (...(\a -> e)...)
1065 -- Simplifying the second arg makes us float the case out, so we end up with
1066 -- case y of (a,b) -> f (...(\a -> e)...) e'
1067 -- So the output does not have the no-shadowing invariant. However, there is
1068 -- no danger of getting name-capture, because when the first arg was simplified
1069 -- we used an in-scope set that at least mentioned all the variables free in its
1070 -- static environment, and that is enough.
1072 -- We can't just do innermost first, or we'd end up with a dual problem:
1073 -- case x of (a,b) -> f e (...(\a -> e')...)
1075 -- I spent hours trying to recover the no-shadowing invariant, but I just could
1076 -- not think of an elegant way to do it. The simplifier is already knee-deep in
1077 -- continuations. We have to keep the right in-scope set around; AND we have
1078 -- to get the effect that finding (error "foo") in a strict arg position will
1079 -- discard the entire application and replace it with (error "foo"). Getting
1080 -- all this at once is TOO HARD!
1082 simplifyArgs env fn_ty has_rules args cont_ty thing_inside
1083 = go env fn_ty args thing_inside
1085 go env fn_ty [] thing_inside = thing_inside env []
1086 go env fn_ty (arg:args) thing_inside = simplifyArg env fn_ty has_rules arg cont_ty $ \ env arg' ->
1087 go env (applyTypeToArg fn_ty arg') args $ \ env args' ->
1088 thing_inside env (arg':args')
1090 simplifyArg env fn_ty has_rules (arg, Nothing, _) cont_ty thing_inside
1091 = thing_inside env arg -- Already simplified
1093 simplifyArg env fn_ty has_rules (Type ty_arg, Just se, _) cont_ty thing_inside
1094 = simplType (setInScope se env) ty_arg `thenSmpl` \ new_ty_arg ->
1095 thing_inside env (Type new_ty_arg)
1097 simplifyArg env fn_ty has_rules (val_arg, Just arg_se, is_strict) cont_ty thing_inside
1099 = simplStrictArg AnArg env val_arg arg_se arg_ty cont_ty thing_inside
1101 | otherwise -- Lazy argument
1102 -- DO NOT float anything outside, hence simplExprC
1103 -- There is no benefit (unlike in a let-binding), and we'd
1104 -- have to be very careful about bogus strictness through
1105 -- floating a demanded let.
1106 = simplExprC (setInScope arg_se env) val_arg
1107 (mkLazyArgStop arg_ty has_rules) `thenSmpl` \ arg1 ->
1108 thing_inside env arg1
1110 arg_ty = funArgTy fn_ty
1113 simplStrictArg :: LetRhsFlag
1114 -> SimplEnv -- The env of the call
1115 -> InExpr -> SimplEnv -- The arg plus its env
1116 -> OutType -- arg_ty: type of the argument
1117 -> OutType -- cont_ty: Type of thing computed by the context
1118 -> (SimplEnv -> OutExpr -> SimplM FloatsWithExpr)
1119 -- Takes an expression of type rhs_ty,
1120 -- returns an expression of type cont_ty
1121 -- The env passed to this continuation is the
1122 -- env of the call, plus any new in-scope variables
1123 -> SimplM FloatsWithExpr -- An expression of type cont_ty
1125 simplStrictArg is_rhs call_env arg arg_env arg_ty cont_ty thing_inside
1126 = simplExprF (setInScope arg_env call_env) arg
1127 (ArgOf is_rhs arg_ty cont_ty (\ new_env -> thing_inside (setInScope call_env new_env)))
1128 -- Notice the way we use arg_env (augmented with in-scope vars from call_env)
1129 -- to simplify the argument
1130 -- and call-env (augmented with in-scope vars from the arg) to pass to the continuation
1134 %************************************************************************
1136 \subsection{mkAtomicArgs}
1138 %************************************************************************
1140 mkAtomicArgs takes a putative RHS, checks whether it's a PAP or
1141 constructor application and, if so, converts it to ANF, so that the
1142 resulting thing can be inlined more easily. Thus
1149 There are three sorts of binding context, specified by the two
1155 N N Top-level or recursive Only bind args of lifted type
1157 N Y Non-top-level and non-recursive, Bind args of lifted type, or
1158 but lazy unlifted-and-ok-for-speculation
1160 Y Y Non-top-level, non-recursive, Bind all args
1161 and strict (demanded)
1167 there is no point in transforming to
1169 x = case (y div# z) of r -> MkC r
1171 because the (y div# z) can't float out of the let. But if it was
1172 a *strict* let, then it would be a good thing to do. Hence the
1173 context information.
1175 Note [Float coercions]
1176 ~~~~~~~~~~~~~~~~~~~~~~
1177 When we find the binding
1179 we'd like to transform it to
1181 x = x `cast` co -- A trivial binding
1182 There's a chance that e will be a constructor application or function, or something
1183 like that, so moving the coerion to the usage site may well cancel the coersions
1184 and lead to further optimisation. Example:
1186 data family T a :: *
1187 data instance T Int = T Int
1189 foo :: Int -> Int -> Int
1194 go n = case x of { T m -> go (n-m) }
1195 -- This case should optimise
1198 mkAtomicArgsE :: SimplEnv
1199 -> Bool -- A strict binding
1200 -> OutExpr -- The rhs
1201 -> (SimplEnv -> OutExpr -> SimplM FloatsWithExpr)
1202 -- Consumer for the simpler rhs
1203 -> SimplM FloatsWithExpr
1205 mkAtomicArgsE env is_strict (Cast rhs co) thing_inside
1206 | not (exprIsTrivial rhs)
1207 -- Note [Float coersions]
1208 -- See also Note [Take care] below
1209 = do { id <- newId FSLIT("a") (exprType rhs)
1210 ; completeNonRecX env False id id rhs $ \ env ->
1211 thing_inside env (Cast (substExpr env (Var id)) co) }
1213 mkAtomicArgsE env is_strict rhs thing_inside
1214 | (Var fun, args) <- collectArgs rhs, -- It's an application
1215 isDataConWorkId fun || valArgCount args < idArity fun -- And it's a constructor or PAP
1216 = go env (Var fun) args
1218 | otherwise = thing_inside env rhs
1221 go env fun [] = thing_inside env fun
1223 go env fun (arg : args)
1224 | exprIsTrivial arg -- Easy case
1225 || no_float_arg -- Can't make it atomic
1226 = go env (App fun arg) args
1229 = do { arg_id <- newId FSLIT("a") arg_ty
1230 ; completeNonRecX env False {- pessimistic -} arg_id arg_id arg $ \env ->
1231 go env (App fun (substExpr env (Var arg_id))) args }
1232 -- Note [Take care]:
1233 -- If completeNonRecX was to do a postInlineUnconditionally
1234 -- (undoing the effect of introducing the let-binding), we'd find arg_id had
1235 -- no binding; hence the substExpr. This happens if we see
1237 -- Then we start by making a variable a1, thus
1238 -- let a1 = D x `cast` g in C a1
1239 -- But then we deal with the rhs of a1, getting
1240 -- let a2 = D x, a1 = a1 `cast` g in C a1
1241 -- And now the preInlineUnconditionally kicks in, and we substitute for a1
1244 arg_ty = exprType arg
1245 no_float_arg = not is_strict && (isUnLiftedType arg_ty) && not (exprOkForSpeculation arg)
1248 -- Old code: consider rewriting to be more like mkAtomicArgsE
1250 mkAtomicArgs :: Bool -- OK to float unlifted args
1252 -> SimplM (OrdList (OutId,OutExpr), -- The floats (unusually) may include
1253 OutExpr) -- things that need case-binding,
1254 -- if the strict-binding flag is on
1256 mkAtomicArgs ok_float_unlifted (Cast rhs co)
1257 | not (exprIsTrivial rhs)
1258 -- Note [Float coersions]
1259 = do { id <- newId FSLIT("a") (exprType rhs)
1260 ; (binds, rhs') <- mkAtomicArgs ok_float_unlifted rhs
1261 ; return (binds `snocOL` (id, rhs'), Cast (Var id) co) }
1263 mkAtomicArgs ok_float_unlifted rhs
1264 | (Var fun, args) <- collectArgs rhs, -- It's an application
1265 isDataConWorkId fun || valArgCount args < idArity fun -- And it's a constructor or PAP
1266 = go fun nilOL [] args -- Have a go
1268 | otherwise = bale_out -- Give up
1271 bale_out = returnSmpl (nilOL, rhs)
1273 go fun binds rev_args []
1274 = returnSmpl (binds, mkApps (Var fun) (reverse rev_args))
1276 go fun binds rev_args (arg : args)
1277 | exprIsTrivial arg -- Easy case
1278 = go fun binds (arg:rev_args) args
1280 | not can_float_arg -- Can't make this arg atomic
1281 = bale_out -- ... so give up
1283 | otherwise -- Don't forget to do it recursively
1284 -- E.g. x = a:b:c:[]
1285 = mkAtomicArgs ok_float_unlifted arg `thenSmpl` \ (arg_binds, arg') ->
1286 newId FSLIT("a") arg_ty `thenSmpl` \ arg_id ->
1287 go fun ((arg_binds `snocOL` (arg_id,arg')) `appOL` binds)
1288 (Var arg_id : rev_args) args
1290 arg_ty = exprType arg
1291 can_float_arg = not (isUnLiftedType arg_ty)
1292 || (ok_float_unlifted && exprOkForSpeculation arg)
1295 addAtomicBinds :: SimplEnv -> [(OutId,OutExpr)]
1296 -> (SimplEnv -> SimplM (FloatsWith a))
1297 -> SimplM (FloatsWith a)
1298 addAtomicBinds env [] thing_inside = thing_inside env
1299 addAtomicBinds env ((v,r):bs) thing_inside = addAuxiliaryBind env (NonRec v r) $ \ env ->
1300 addAtomicBinds env bs thing_inside
1304 %************************************************************************
1306 \subsection{The main rebuilder}
1308 %************************************************************************
1311 rebuild :: SimplEnv -> OutExpr -> SimplCont -> SimplM FloatsWithExpr
1313 rebuild env expr (Stop _ _ _) = rebuildDone env expr
1314 rebuild env expr (ArgOf _ _ _ cont_fn) = cont_fn env expr
1315 rebuild env expr (CoerceIt co cont) = rebuild env (mkCoerce co expr) cont
1316 rebuild env expr (Select _ bndr alts se cont) = rebuildCase (setInScope se env) expr bndr alts cont
1317 rebuild env expr (ApplyTo _ arg mb_se cont) = rebuildApp env expr arg mb_se cont
1319 rebuildApp env fun arg mb_se cont
1320 = do { arg' <- simplArg env arg mb_se
1321 ; rebuild env (App fun arg') cont }
1323 simplArg :: SimplEnv -> CoreExpr -> Maybe SimplEnv -> SimplM CoreExpr
1324 simplArg env arg Nothing = return arg -- The arg is already simplified
1325 simplArg env arg (Just arg_env) = simplExpr (setInScope arg_env env) arg
1327 rebuildDone env expr = returnSmpl (emptyFloats env, expr)
1331 %************************************************************************
1333 \subsection{Functions dealing with a case}
1335 %************************************************************************
1337 Blob of helper functions for the "case-of-something-else" situation.
1340 ---------------------------------------------------------
1341 -- Eliminate the case if possible
1343 rebuildCase :: SimplEnv
1344 -> OutExpr -- Scrutinee
1345 -> InId -- Case binder
1346 -> [InAlt] -- Alternatives (inceasing order)
1348 -> SimplM FloatsWithExpr
1350 rebuildCase env scrut case_bndr alts cont
1351 | Just (con,args) <- exprIsConApp_maybe scrut
1352 -- Works when the scrutinee is a variable with a known unfolding
1353 -- as well as when it's an explicit constructor application
1354 = knownCon env scrut (DataAlt con) args case_bndr alts cont
1356 | Lit lit <- scrut -- No need for same treatment as constructors
1357 -- because literals are inlined more vigorously
1358 = knownCon env scrut (LitAlt lit) [] case_bndr alts cont
1361 = -- Prepare the continuation;
1362 -- The new subst_env is in place
1363 prepareCaseCont env alts cont `thenSmpl` \ (floats, (dup_cont, nondup_cont)) ->
1364 addFloats env floats $ \ env ->
1367 -- The case expression is annotated with the result type of the continuation
1368 -- This may differ from the type originally on the case. For example
1369 -- case(T) (case(Int#) a of { True -> 1#; False -> 0# }) of
1372 -- let j a# = <blob>
1373 -- in case(T) a of { True -> j 1#; False -> j 0# }
1374 -- Note that the case that scrutinises a now returns a T not an Int#
1375 res_ty' = contResultType dup_cont
1378 -- Deal with case binder
1379 simplCaseBinder env scrut case_bndr `thenSmpl` \ (alt_env, case_bndr') ->
1381 -- Deal with the case alternatives
1382 simplAlts alt_env scrut case_bndr' alts dup_cont `thenSmpl` \ alts' ->
1384 -- Put the case back together
1385 mkCase scrut case_bndr' res_ty' alts' `thenSmpl` \ case_expr ->
1387 -- Notice that rebuildDone returns the in-scope set from env, not alt_env
1388 -- The case binder *not* scope over the whole returned case-expression
1389 rebuild env case_expr nondup_cont
1392 simplCaseBinder checks whether the scrutinee is a variable, v. If so,
1393 try to eliminate uses of v in the RHSs in favour of case_bndr; that
1394 way, there's a chance that v will now only be used once, and hence
1397 Note [no-case-of-case]
1398 ~~~~~~~~~~~~~~~~~~~~~~
1399 There is a time we *don't* want to do that, namely when
1400 -fno-case-of-case is on. This happens in the first simplifier pass,
1401 and enhances full laziness. Here's the bad case:
1402 f = \ y -> ...(case x of I# v -> ...(case x of ...) ... )
1403 If we eliminate the inner case, we trap it inside the I# v -> arm,
1404 which might prevent some full laziness happening. I've seen this
1405 in action in spectral/cichelli/Prog.hs:
1406 [(m,n) | m <- [1..max], n <- [1..max]]
1407 Hence the check for NoCaseOfCase.
1411 Consider case (v `cast` co) of x { I# ->
1412 ... (case (v `cast` co) of {...}) ...
1413 We'd like to eliminate the inner case. We can get this neatly by
1414 arranging that inside the outer case we add the unfolding
1415 v |-> x `cast` (sym co)
1416 to v. Then we should inline v at the inner case, cancel the casts, and away we go
1420 There is another situation when we don't want to do it. If we have
1422 case x of w1 { DEFAULT -> case x of w2 { A -> e1; B -> e2 }
1423 ...other cases .... }
1425 We'll perform the binder-swap for the outer case, giving
1427 case x of w1 { DEFAULT -> case w1 of w2 { A -> e1; B -> e2 }
1428 ...other cases .... }
1430 But there is no point in doing it for the inner case, because w1 can't
1431 be inlined anyway. Furthermore, doing the case-swapping involves
1432 zapping w2's occurrence info (see paragraphs that follow), and that
1433 forces us to bind w2 when doing case merging. So we get
1435 case x of w1 { A -> let w2 = w1 in e1
1436 B -> let w2 = w1 in e2
1437 ...other cases .... }
1439 This is plain silly in the common case where w2 is dead.
1441 Even so, I can't see a good way to implement this idea. I tried
1442 not doing the binder-swap if the scrutinee was already evaluated
1443 but that failed big-time:
1447 case v of w { MkT x ->
1448 case x of x1 { I# y1 ->
1449 case x of x2 { I# y2 -> ...
1451 Notice that because MkT is strict, x is marked "evaluated". But to
1452 eliminate the last case, we must either make sure that x (as well as
1453 x1) has unfolding MkT y1. THe straightforward thing to do is to do
1454 the binder-swap. So this whole note is a no-op.
1458 If we replace the scrutinee, v, by tbe case binder, then we have to nuke
1459 any occurrence info (eg IAmDead) in the case binder, because the
1460 case-binder now effectively occurs whenever v does. AND we have to do
1461 the same for the pattern-bound variables! Example:
1463 (case x of { (a,b) -> a }) (case x of { (p,q) -> q })
1465 Here, b and p are dead. But when we move the argment inside the first
1466 case RHS, and eliminate the second case, we get
1468 case x of { (a,b) -> a b }
1470 Urk! b is alive! Reason: the scrutinee was a variable, and case elimination
1473 Indeed, this can happen anytime the case binder isn't dead:
1474 case <any> of x { (a,b) ->
1475 case x of { (p,q) -> p } }
1476 Here (a,b) both look dead, but come alive after the inner case is eliminated.
1477 The point is that we bring into the envt a binding
1479 after the outer case, and that makes (a,b) alive. At least we do unless
1480 the case binder is guaranteed dead.
1483 simplCaseBinder env scrut case_bndr
1484 | switchIsOn (getSwitchChecker env) NoCaseOfCase
1485 -- See Note [no-case-of-case]
1486 = do { (env, case_bndr') <- simplBinder env case_bndr
1487 ; return (env, case_bndr') }
1489 simplCaseBinder env (Var v) case_bndr
1490 -- Failed try [see Note 2 above]
1491 -- not (isEvaldUnfolding (idUnfolding v))
1492 = do { (env, case_bndr') <- simplBinder env (zapOccInfo case_bndr)
1493 ; return (modifyInScope env v case_bndr', case_bndr') }
1494 -- We could extend the substitution instead, but it would be
1495 -- a hack because then the substitution wouldn't be idempotent
1496 -- any more (v is an OutId). And this does just as well.
1498 simplCaseBinder env (Cast (Var v) co) case_bndr -- Note [Case of cast]
1499 = do { (env, case_bndr') <- simplBinder env (zapOccInfo case_bndr)
1500 ; let rhs = Cast (Var case_bndr') (mkSymCoercion co)
1501 ; return (addBinderUnfolding env v rhs, case_bndr') }
1503 simplCaseBinder env other_scrut case_bndr
1504 = do { (env, case_bndr') <- simplBinder env case_bndr
1505 ; return (env, case_bndr') }
1507 zapOccInfo :: InId -> InId -- See Note [zapOccInfo]
1508 zapOccInfo b = b `setIdOccInfo` NoOccInfo
1512 simplAlts does two things:
1514 1. Eliminate alternatives that cannot match, including the
1515 DEFAULT alternative.
1517 2. If the DEFAULT alternative can match only one possible constructor,
1518 then make that constructor explicit.
1520 case e of x { DEFAULT -> rhs }
1522 case e of x { (a,b) -> rhs }
1523 where the type is a single constructor type. This gives better code
1524 when rhs also scrutinises x or e.
1526 Here "cannot match" includes knowledge from GADTs
1528 It's a good idea do do this stuff before simplifying the alternatives, to
1529 avoid simplifying alternatives we know can't happen, and to come up with
1530 the list of constructors that are handled, to put into the IdInfo of the
1531 case binder, for use when simplifying the alternatives.
1533 Eliminating the default alternative in (1) isn't so obvious, but it can
1536 data Colour = Red | Green | Blue
1545 DEFAULT -> [ case y of ... ]
1547 If we inline h into f, the default case of the inlined h can't happen.
1548 If we don't notice this, we may end up filtering out *all* the cases
1549 of the inner case y, which give us nowhere to go!
1553 simplAlts :: SimplEnv
1555 -> OutId -- Case binder
1556 -> [InAlt] -> SimplCont
1557 -> SimplM [OutAlt] -- Includes the continuation
1559 simplAlts env scrut case_bndr' alts cont'
1560 = do { mb_alts <- mapSmpl (simplAlt env imposs_cons case_bndr' cont') alts_wo_default
1561 ; default_alts <- simplDefault env case_bndr' imposs_deflt_cons cont' maybe_deflt
1562 ; return (mergeAlts default_alts [alt' | Just (_, alt') <- mb_alts]) }
1563 -- We need the mergeAlts in case the new default_alt
1564 -- has turned into a constructor alternative.
1566 (alts_wo_default, maybe_deflt) = findDefault alts
1567 imposs_cons = case scrut of
1568 Var v -> otherCons (idUnfolding v)
1571 -- "imposs_deflt_cons" are handled either by the context,
1572 -- OR by a branch in this case expression. (Don't include DEFAULT!!)
1573 imposs_deflt_cons = nub (imposs_cons ++ [con | (con,_,_) <- alts_wo_default])
1575 simplDefault :: SimplEnv
1576 -> OutId -- Case binder; need just for its type. Note that as an
1577 -- OutId, it has maximum information; this is important.
1578 -- Test simpl013 is an example
1579 -> [AltCon] -- These cons can't happen when matching the default
1582 -> SimplM [OutAlt] -- One branch or none; we use a list because it's what
1583 -- mergeAlts expects
1586 simplDefault env case_bndr' imposs_cons cont Nothing
1587 = return [] -- No default branch
1589 simplDefault env case_bndr' imposs_cons cont (Just rhs)
1590 | -- This branch handles the case where we are
1591 -- scrutinisng an algebraic data type
1592 Just (tycon, inst_tys) <- splitTyConApp_maybe (idType case_bndr'),
1593 isAlgTyCon tycon, -- It's a data type, tuple, or unboxed tuples.
1594 not (isNewTyCon tycon), -- We can have a newtype, if we are just doing an eval:
1595 -- case x of { DEFAULT -> e }
1596 -- and we don't want to fill in a default for them!
1597 Just all_cons <- tyConDataCons_maybe tycon,
1598 not (null all_cons), -- This is a tricky corner case. If the data type has no constructors,
1599 -- which GHC allows, then the case expression will have at most a default
1600 -- alternative. We don't want to eliminate that alternative, because the
1601 -- invariant is that there's always one alternative. It's more convenient
1603 -- case x of { DEFAULT -> e }
1604 -- as it is, rather than transform it to
1605 -- error "case cant match"
1606 -- which would be quite legitmate. But it's a really obscure corner, and
1607 -- not worth wasting code on.
1609 let imposs_data_cons = [con | DataAlt con <- imposs_cons] -- We now know it's a data type
1610 poss_data_cons = filterOut (`elem` imposs_data_cons) all_cons
1611 gadt_imposs | all isTyVarTy inst_tys = []
1612 | otherwise = filter (cant_match inst_tys) poss_data_cons
1613 final_poss = filterOut (`elem` gadt_imposs) poss_data_cons
1615 = case final_poss of
1616 [] -> returnSmpl [] -- Eliminate the default alternative
1617 -- altogether if it can't match
1619 [con] -> -- It matches exactly one constructor, so fill it in
1620 do { tick (FillInCaseDefault case_bndr')
1621 ; us <- getUniquesSmpl
1622 ; let (ex_tvs, co_tvs, arg_ids) =
1623 dataConRepInstPat us con inst_tys
1624 ; let con_alt = (DataAlt con, ex_tvs ++ co_tvs ++ arg_ids, rhs)
1625 ; Just (_, alt') <- simplAlt env [] case_bndr' cont con_alt
1626 -- The simplAlt must succeed with Just because we have
1627 -- already filtered out construtors that can't match
1630 two_or_more -> simplify_default (map DataAlt gadt_imposs ++ imposs_cons)
1633 = simplify_default imposs_cons
1635 cant_match tys data_con = not (dataConCanMatch data_con tys)
1637 simplify_default imposs_cons
1638 = do { let env' = addBinderOtherCon env case_bndr' imposs_cons
1639 -- Record the constructors that the case-binder *can't* be.
1640 ; rhs' <- simplExprC env' rhs cont
1641 ; return [(DEFAULT, [], rhs')] }
1643 simplAlt :: SimplEnv
1644 -> [AltCon] -- These constructors can't be present when
1645 -- matching this alternative
1646 -> OutId -- The case binder
1649 -> SimplM (Maybe (TvSubstEnv, OutAlt))
1651 -- Simplify an alternative, returning the type refinement for the
1652 -- alternative, if the alternative does any refinement at all
1653 -- Nothing => the alternative is inaccessible
1655 simplAlt env imposs_cons case_bndr' cont' (con, bndrs, rhs)
1656 | con `elem` imposs_cons -- This case can't match
1659 simplAlt env handled_cons case_bndr' cont' (DEFAULT, bndrs, rhs)
1660 -- TURGID DUPLICATION, needed only for the simplAlt call
1661 -- in mkDupableAlt. Clean this up when moving to FC
1662 = ASSERT( null bndrs )
1663 simplExprC env' rhs cont' `thenSmpl` \ rhs' ->
1664 returnSmpl (Just (emptyVarEnv, (DEFAULT, [], rhs')))
1666 env' = addBinderOtherCon env case_bndr' handled_cons
1667 -- Record the constructors that the case-binder *can't* be.
1669 simplAlt env handled_cons case_bndr' cont' (LitAlt lit, bndrs, rhs)
1670 = ASSERT( null bndrs )
1671 simplExprC env' rhs cont' `thenSmpl` \ rhs' ->
1672 returnSmpl (Just (emptyVarEnv, (LitAlt lit, [], rhs')))
1674 env' = addBinderUnfolding env case_bndr' (Lit lit)
1676 simplAlt env handled_cons case_bndr' cont' (DataAlt con, vs, rhs)
1677 = -- Deal with the pattern-bound variables
1678 -- Mark the ones that are in ! positions in the data constructor
1679 -- as certainly-evaluated.
1680 -- NB: it happens that simplBinders does *not* erase the OtherCon
1681 -- form of unfolding, so it's ok to add this info before
1682 -- doing simplBinders
1683 simplBinders env (add_evals con vs) `thenSmpl` \ (env, vs') ->
1685 -- Bind the case-binder to (con args)
1686 let inst_tys' = tyConAppArgs (idType case_bndr')
1687 con_args = map Type inst_tys' ++ varsToCoreExprs vs'
1688 env' = addBinderUnfolding env case_bndr' (mkConApp con con_args)
1690 simplExprC env' rhs cont' `thenSmpl` \ rhs' ->
1691 returnSmpl (Just (emptyVarEnv, (DataAlt con, vs', rhs')))
1693 -- add_evals records the evaluated-ness of the bound variables of
1694 -- a case pattern. This is *important*. Consider
1695 -- data T = T !Int !Int
1697 -- case x of { T a b -> T (a+1) b }
1699 -- We really must record that b is already evaluated so that we don't
1700 -- go and re-evaluate it when constructing the result.
1701 add_evals dc vs = cat_evals dc vs (dataConRepStrictness dc)
1703 cat_evals dc vs strs
1707 go (v:vs) strs | isTyVar v = v : go vs strs
1708 go (v:vs) (str:strs)
1709 | isMarkedStrict str = evald_v : go vs strs
1710 | otherwise = zapped_v : go vs strs
1712 zapped_v = zap_occ_info v
1713 evald_v = zapped_v `setIdUnfolding` evaldUnfolding
1714 go _ _ = pprPanic "cat_evals" (ppr dc $$ ppr vs $$ ppr strs)
1716 -- If the case binder is alive, then we add the unfolding
1718 -- to the envt; so vs are now very much alive
1719 -- Note [Aug06] I can't see why this actually matters
1720 zap_occ_info | isDeadBinder case_bndr' = \id -> id
1721 | otherwise = zapOccInfo
1723 addBinderUnfolding :: SimplEnv -> Id -> CoreExpr -> SimplEnv
1724 addBinderUnfolding env bndr rhs
1725 = modifyInScope env bndr (bndr `setIdUnfolding` mkUnfolding False rhs)
1727 addBinderOtherCon :: SimplEnv -> Id -> [AltCon] -> SimplEnv
1728 addBinderOtherCon env bndr cons
1729 = modifyInScope env bndr (bndr `setIdUnfolding` mkOtherCon cons)
1733 %************************************************************************
1735 \subsection{Known constructor}
1737 %************************************************************************
1739 We are a bit careful with occurrence info. Here's an example
1741 (\x* -> case x of (a*, b) -> f a) (h v, e)
1743 where the * means "occurs once". This effectively becomes
1744 case (h v, e) of (a*, b) -> f a)
1746 let a* = h v; b = e in f a
1750 All this should happen in one sweep.
1753 knownCon :: SimplEnv -> OutExpr -> AltCon -> [OutExpr]
1754 -> InId -> [InAlt] -> SimplCont
1755 -> SimplM FloatsWithExpr
1757 knownCon env scrut con args bndr alts cont
1758 = tick (KnownBranch bndr) `thenSmpl_`
1759 case findAlt con alts of
1760 (DEFAULT, bs, rhs) -> ASSERT( null bs )
1761 simplNonRecX env bndr scrut $ \ env ->
1762 -- This might give rise to a binding with non-atomic args
1763 -- like x = Node (f x) (g x)
1764 -- but simplNonRecX will atomic-ify it
1765 simplExprF env rhs cont
1767 (LitAlt lit, bs, rhs) -> ASSERT( null bs )
1768 simplNonRecX env bndr scrut $ \ env ->
1769 simplExprF env rhs cont
1771 (DataAlt dc, bs, rhs)
1772 -> -- ASSERT( n_drop_tys + length bs == length args )
1773 bind_args env dead_bndr bs (drop n_drop_tys args) $ \ env ->
1775 -- It's useful to bind bndr to scrut, rather than to a fresh
1776 -- binding x = Con arg1 .. argn
1777 -- because very often the scrut is a variable, so we avoid
1778 -- creating, and then subsequently eliminating, a let-binding
1779 -- BUT, if scrut is a not a variable, we must be careful
1780 -- about duplicating the arg redexes; in that case, make
1781 -- a new con-app from the args
1782 bndr_rhs = case scrut of
1785 con_app = mkConApp dc (take n_drop_tys args ++ con_args)
1786 con_args = [substExpr env (varToCoreExpr b) | b <- bs]
1787 -- args are aready OutExprs, but bs are InIds
1789 simplNonRecX env bndr bndr_rhs $ \ env ->
1790 simplExprF env rhs cont
1792 dead_bndr = isDeadBinder bndr
1793 n_drop_tys = tyConArity (dataConTyCon dc)
1796 bind_args env dead_bndr [] _ thing_inside = thing_inside env
1798 bind_args env dead_bndr (b:bs) (Type ty : args) thing_inside
1799 = ASSERT( isTyVar b )
1800 bind_args (extendTvSubst env b ty) dead_bndr bs args thing_inside
1802 bind_args env dead_bndr (b:bs) (arg : args) thing_inside
1805 b' = if dead_bndr then b else zapOccInfo b
1806 -- Note that the binder might be "dead", because it doesn't occur
1807 -- in the RHS; and simplNonRecX may therefore discard it via postInlineUnconditionally
1808 -- Nevertheless we must keep it if the case-binder is alive, because it may
1809 -- be used in the con_app. See Note [zapOccInfo]
1811 simplNonRecX env b' arg $ \ env ->
1812 bind_args env dead_bndr bs args thing_inside
1816 %************************************************************************
1818 \subsection{Duplicating continuations}
1820 %************************************************************************
1823 prepareCaseCont :: SimplEnv
1824 -> [InAlt] -> SimplCont
1825 -> SimplM (FloatsWith (SimplCont,SimplCont))
1826 -- Return a duplicatable continuation, a non-duplicable part
1827 -- plus some extra bindings (that scope over the entire
1830 -- No need to make it duplicatable if there's only one alternative
1831 prepareCaseCont env [alt] cont = returnSmpl (emptyFloats env, (cont, mkBoringStop (contResultType cont)))
1832 prepareCaseCont env alts cont = mkDupableCont env cont
1836 mkDupableCont :: SimplEnv -> SimplCont
1837 -> SimplM (FloatsWith (SimplCont, SimplCont))
1839 mkDupableCont env cont
1840 | contIsDupable cont
1841 = returnSmpl (emptyFloats env, (cont, mkBoringStop (contResultType cont)))
1843 mkDupableCont env (CoerceIt ty cont)
1844 = mkDupableCont env cont `thenSmpl` \ (floats, (dup_cont, nondup_cont)) ->
1845 returnSmpl (floats, (CoerceIt ty dup_cont, nondup_cont))
1847 mkDupableCont env cont@(ArgOf _ arg_ty _ _)
1848 = returnSmpl (emptyFloats env, (mkBoringStop arg_ty, cont))
1849 -- Do *not* duplicate an ArgOf continuation
1850 -- Because ArgOf continuations are opaque, we gain nothing by
1851 -- propagating them into the expressions, and we do lose a lot.
1852 -- Here's an example:
1853 -- && (case x of { T -> F; F -> T }) E
1854 -- Now, && is strict so we end up simplifying the case with
1855 -- an ArgOf continuation. If we let-bind it, we get
1857 -- let $j = \v -> && v E
1858 -- in simplExpr (case x of { T -> F; F -> T })
1859 -- (ArgOf (\r -> $j r)
1860 -- And after simplifying more we get
1862 -- let $j = \v -> && v E
1863 -- in case of { T -> $j F; F -> $j T }
1864 -- Which is a Very Bad Thing
1866 -- The desire not to duplicate is the entire reason that
1867 -- mkDupableCont returns a pair of continuations.
1869 -- The original plan had:
1870 -- e.g. (...strict-fn...) [...hole...]
1872 -- let $j = \a -> ...strict-fn...
1873 -- in $j [...hole...]
1875 mkDupableCont env (ApplyTo _ arg mb_se cont)
1876 = -- e.g. [...hole...] (...arg...)
1878 -- let a = ...arg...
1879 -- in [...hole...] a
1880 do { (floats, (dup_cont, nondup_cont)) <- mkDupableCont env cont
1881 ; addFloats env floats $ \ env -> do
1882 { arg1 <- simplArg env arg mb_se
1883 ; (floats2, arg2) <- mkDupableArg env arg1
1884 ; return (floats2, (ApplyTo OkToDup arg2 Nothing dup_cont, nondup_cont)) }}
1886 mkDupableCont env cont@(Select _ case_bndr [(_,bs,rhs)] se case_cont)
1887 -- | not (exprIsDupable rhs && contIsDupable case_cont) -- See notes below
1888 -- | not (isDeadBinder case_bndr)
1889 | all isDeadBinder bs
1890 = returnSmpl (emptyFloats env, (mkBoringStop scrut_ty, cont))
1892 scrut_ty = substTy se (idType case_bndr)
1894 {- Note [Single-alternative cases]
1895 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
1896 This case is just like the ArgOf case. Here's an example:
1900 case (case x of I# x' ->
1902 True -> I# (negate# x')
1903 False -> I# x') of y {
1905 Because the (case x) has only one alternative, we'll transform to
1907 case (case x' <# 0# of
1908 True -> I# (negate# x')
1909 False -> I# x') of y {
1911 But now we do *NOT* want to make a join point etc, giving
1913 let $j = \y -> MkT y
1915 True -> $j (I# (negate# x'))
1917 In this case the $j will inline again, but suppose there was a big
1918 strict computation enclosing the orginal call to MkT. Then, it won't
1919 "see" the MkT any more, because it's big and won't get duplicated.
1920 And, what is worse, nothing was gained by the case-of-case transform.
1922 When should use this case of mkDupableCont?
1923 However, matching on *any* single-alternative case is a *disaster*;
1924 e.g. case (case ....) of (a,b) -> (# a,b #)
1925 We must push the outer case into the inner one!
1928 * Match [(DEFAULT,_,_)], but in the common case of Int,
1929 the alternative-filling-in code turned the outer case into
1930 case (...) of y { I# _ -> MkT y }
1932 * Match on single alternative plus (not (isDeadBinder case_bndr))
1933 Rationale: pushing the case inwards won't eliminate the construction.
1934 But there's a risk of
1935 case (...) of y { (a,b) -> let z=(a,b) in ... }
1936 Now y looks dead, but it'll come alive again. Still, this
1937 seems like the best option at the moment.
1939 * Match on single alternative plus (all (isDeadBinder bndrs))
1940 Rationale: this is essentially seq.
1942 * Match when the rhs is *not* duplicable, and hence would lead to a
1943 join point. This catches the disaster-case above. We can test
1944 the *un-simplified* rhs, which is fine. It might get bigger or
1945 smaller after simplification; if it gets smaller, this case might
1946 fire next time round. NB also that we must test contIsDupable
1947 case_cont *btoo, because case_cont might be big!
1949 HOWEVER: I found that this version doesn't work well, because
1950 we can get let x = case (...) of { small } in ...case x...
1951 When x is inlined into its full context, we find that it was a bad
1952 idea to have pushed the outer case inside the (...) case.
1955 mkDupableCont env (Select _ case_bndr alts se cont)
1956 = -- e.g. (case [...hole...] of { pi -> ei })
1958 -- let ji = \xij -> ei
1959 -- in case [...hole...] of { pi -> ji xij }
1960 do { tick (CaseOfCase case_bndr)
1961 ; let alt_env = setInScope se env
1962 ; (floats1, (dup_cont, nondup_cont)) <- mkDupableCont alt_env cont
1963 -- NB: call mkDupableCont here, *not* prepareCaseCont
1964 -- We must make a duplicable continuation, whereas prepareCaseCont
1965 -- doesn't when there is a single case branch
1966 ; addFloats alt_env floats1 $ \ alt_env -> do
1968 { (alt_env, case_bndr') <- simplBinder alt_env case_bndr
1969 -- NB: simplBinder does not zap deadness occ-info, so
1970 -- a dead case_bndr' will still advertise its deadness
1971 -- This is really important because in
1972 -- case e of b { (# a,b #) -> ... }
1973 -- b is always dead, and indeed we are not allowed to bind b to (# a,b #),
1974 -- which might happen if e was an explicit unboxed pair and b wasn't marked dead.
1975 -- In the new alts we build, we have the new case binder, so it must retain
1978 ; (floats2, alts') <- mkDupableAlts alt_env case_bndr' alts dup_cont
1979 ; return (floats2, (Select OkToDup case_bndr' alts' (zapSubstEnv se)
1980 (mkBoringStop (contResultType dup_cont)),
1984 mkDupableArg :: SimplEnv -> OutExpr -> SimplM (FloatsWith OutExpr)
1985 -- Let-bind the thing if necessary
1986 mkDupableArg env arg
1988 = return (emptyFloats env, arg)
1990 = do { arg_id <- newId FSLIT("a") (exprType arg)
1991 ; tick (CaseOfCase arg_id)
1992 -- Want to tick here so that we go round again,
1993 -- and maybe copy or inline the code.
1994 -- Not strictly CaseOfCase, but never mind
1995 ; return (unitFloat env arg_id arg, Var arg_id) }
1996 -- What if the arg should be case-bound?
1997 -- This has been this way for a long time, so I'll leave it,
1998 -- but I can't convince myself that it's right.
2000 mkDupableAlts :: SimplEnv -> OutId -> [InAlt] -> SimplCont
2001 -> SimplM (FloatsWith [InAlt])
2002 -- Absorbs the continuation into the new alternatives
2004 mkDupableAlts env case_bndr' alts dupable_cont
2007 go env [] = returnSmpl (emptyFloats env, [])
2009 = do { (floats1, mb_alt') <- mkDupableAlt env case_bndr' dupable_cont alt
2010 ; addFloats env floats1 $ \ env -> do
2011 { (floats2, alts') <- go env alts
2012 ; returnSmpl (floats2, case mb_alt' of
2013 Just alt' -> alt' : alts'
2017 mkDupableAlt env case_bndr' cont alt
2018 = simplAlt env [] case_bndr' cont alt `thenSmpl` \ mb_stuff ->
2020 Nothing -> returnSmpl (emptyFloats env, Nothing) ;
2022 Just (reft, (con, bndrs', rhs')) ->
2023 -- Safe to say that there are no handled-cons for the DEFAULT case
2025 if exprIsDupable rhs' then
2026 returnSmpl (emptyFloats env, Just (con, bndrs', rhs'))
2027 -- It is worth checking for a small RHS because otherwise we
2028 -- get extra let bindings that may cause an extra iteration of the simplifier to
2029 -- inline back in place. Quite often the rhs is just a variable or constructor.
2030 -- The Ord instance of Maybe in PrelMaybe.lhs, for example, took several extra
2031 -- iterations because the version with the let bindings looked big, and so wasn't
2032 -- inlined, but after the join points had been inlined it looked smaller, and so
2035 -- NB: we have to check the size of rhs', not rhs.
2036 -- Duplicating a small InAlt might invalidate occurrence information
2037 -- However, if it *is* dupable, we return the *un* simplified alternative,
2038 -- because otherwise we'd need to pair it up with an empty subst-env....
2039 -- but we only have one env shared between all the alts.
2040 -- (Remember we must zap the subst-env before re-simplifying something).
2041 -- Rather than do this we simply agree to re-simplify the original (small) thing later.
2045 rhs_ty' = exprType rhs'
2046 used_bndrs' = filter abstract_over (case_bndr' : bndrs')
2048 | isTyVar bndr = not (bndr `elemVarEnv` reft)
2049 -- Don't abstract over tyvar binders which are refined away
2050 -- See Note [Refinement] below
2051 | otherwise = not (isDeadBinder bndr)
2052 -- The deadness info on the new Ids is preserved by simplBinders
2054 -- If we try to lift a primitive-typed something out
2055 -- for let-binding-purposes, we will *caseify* it (!),
2056 -- with potentially-disastrous strictness results. So
2057 -- instead we turn it into a function: \v -> e
2058 -- where v::State# RealWorld#. The value passed to this function
2059 -- is realworld#, which generates (almost) no code.
2061 -- There's a slight infelicity here: we pass the overall
2062 -- case_bndr to all the join points if it's used in *any* RHS,
2063 -- because we don't know its usage in each RHS separately
2065 -- We used to say "&& isUnLiftedType rhs_ty'" here, but now
2066 -- we make the join point into a function whenever used_bndrs'
2067 -- is empty. This makes the join-point more CPR friendly.
2068 -- Consider: let j = if .. then I# 3 else I# 4
2069 -- in case .. of { A -> j; B -> j; C -> ... }
2071 -- Now CPR doesn't w/w j because it's a thunk, so
2072 -- that means that the enclosing function can't w/w either,
2073 -- which is a lose. Here's the example that happened in practice:
2074 -- kgmod :: Int -> Int -> Int
2075 -- kgmod x y = if x > 0 && y < 0 || x < 0 && y > 0
2079 -- I have seen a case alternative like this:
2080 -- True -> \v -> ...
2081 -- It's a bit silly to add the realWorld dummy arg in this case, making
2084 -- (the \v alone is enough to make CPR happy) but I think it's rare
2086 ( if not (any isId used_bndrs')
2087 then newId FSLIT("w") realWorldStatePrimTy `thenSmpl` \ rw_id ->
2088 returnSmpl ([rw_id], [Var realWorldPrimId])
2090 returnSmpl (used_bndrs', varsToCoreExprs used_bndrs')
2091 ) `thenSmpl` \ (final_bndrs', final_args) ->
2093 -- See comment about "$j" name above
2094 newId FSLIT("$j") (mkPiTypes final_bndrs' rhs_ty') `thenSmpl` \ join_bndr ->
2095 -- Notice the funky mkPiTypes. If the contructor has existentials
2096 -- it's possible that the join point will be abstracted over
2097 -- type varaibles as well as term variables.
2098 -- Example: Suppose we have
2099 -- data T = forall t. C [t]
2101 -- case (case e of ...) of
2102 -- C t xs::[t] -> rhs
2103 -- We get the join point
2104 -- let j :: forall t. [t] -> ...
2105 -- j = /\t \xs::[t] -> rhs
2107 -- case (case e of ...) of
2108 -- C t xs::[t] -> j t xs
2110 -- We make the lambdas into one-shot-lambdas. The
2111 -- join point is sure to be applied at most once, and doing so
2112 -- prevents the body of the join point being floated out by
2113 -- the full laziness pass
2114 really_final_bndrs = map one_shot final_bndrs'
2115 one_shot v | isId v = setOneShotLambda v
2117 join_rhs = mkLams really_final_bndrs rhs'
2118 join_call = mkApps (Var join_bndr) final_args
2120 returnSmpl (unitFloat env join_bndr join_rhs, Just (con, bndrs', join_call)) }
2127 MkT :: a -> b -> T a
2131 MkT a' b (p::a') (q::b) -> [p,w]
2133 The danger is that we'll make a join point
2137 and that's ill-typed, because (p::a') but (w::a).
2139 Solution so far: don't abstract over a', because the type refinement
2140 maps [a' -> a] . Ultimately that won't work when real refinement goes on.
2142 Then we must abstract over any refined free variables. Hmm. Maybe we
2143 could just abstract over *all* free variables, thereby lambda-lifting
2144 the join point? We should try this.