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 CmdLineOpts ( dopt, DynFlag(Opt_D_dump_inlinings),
15 import SimplUtils ( mkCase, mkLam, newId, prepareAlts,
16 simplBinder, simplBinders, simplLamBndrs, simplRecBndrs, simplLetBndr,
17 SimplCont(..), DupFlag(..), LetRhsFlag(..),
18 mkStop, mkBoringStop, pushContArgs,
19 contResultType, countArgs, contIsDupable, contIsRhsOrArg,
20 getContArgs, interestingCallContext, interestingArg, isStrictType
22 import Var ( mustHaveLocalBinding )
24 import Id ( Id, idType, idInfo, idArity, isDataConWorkId,
25 setIdUnfolding, isDeadBinder,
26 idNewDemandInfo, setIdInfo,
27 setIdOccInfo, zapLamIdInfo, setOneShotLambda,
29 import OccName ( encodeFS )
30 import IdInfo ( OccInfo(..), isLoopBreaker,
35 import NewDemand ( isStrictDmd )
36 import DataCon ( dataConNumInstArgs, dataConRepStrictness )
38 import PprCore ( pprParendExpr, pprCoreExpr )
39 import CoreUnfold ( mkOtherCon, mkUnfolding, callSiteInline )
40 import CoreUtils ( exprIsDupable, exprIsTrivial, needsCaseBinding,
41 exprIsConApp_maybe, mkPiTypes, findAlt,
42 exprType, exprIsValue,
43 exprOkForSpeculation, exprArity,
44 mkCoerce, mkCoerce2, mkSCC, mkInlineMe, mkAltExpr, applyTypeToArg
46 import Rules ( lookupRule )
47 import BasicTypes ( isMarkedStrict )
48 import CostCentre ( currentCCS )
49 import Type ( isUnLiftedType, seqType, tyConAppArgs, funArgTy,
50 splitFunTy_maybe, splitFunTy, eqType
52 import Subst ( mkSubst, substTy, substExpr,
53 isInScope, lookupIdSubst, simplIdInfo
55 import TysPrim ( realWorldStatePrimTy )
56 import PrelInfo ( realWorldPrimId )
57 import BasicTypes ( TopLevelFlag(..), isTopLevel,
61 import Maybe ( Maybe )
63 import Util ( notNull )
67 The guts of the simplifier is in this module, but the driver loop for
68 the simplifier is in SimplCore.lhs.
71 -----------------------------------------
72 *** IMPORTANT NOTE ***
73 -----------------------------------------
74 The simplifier used to guarantee that the output had no shadowing, but
75 it does not do so any more. (Actually, it never did!) The reason is
76 documented with simplifyArgs.
79 -----------------------------------------
80 *** IMPORTANT NOTE ***
81 -----------------------------------------
82 Many parts of the simplifier return a bunch of "floats" as well as an
83 expression. This is wrapped as a datatype SimplUtils.FloatsWith.
85 All "floats" are let-binds, not case-binds, but some non-rec lets may
86 be unlifted (with RHS ok-for-speculation).
90 -----------------------------------------
91 ORGANISATION OF FUNCTIONS
92 -----------------------------------------
94 - simplify all top-level binders
95 - for NonRec, call simplRecOrTopPair
96 - for Rec, call simplRecBind
99 ------------------------------
100 simplExpr (applied lambda) ==> simplNonRecBind
101 simplExpr (Let (NonRec ...) ..) ==> simplNonRecBind
102 simplExpr (Let (Rec ...) ..) ==> simplify binders; simplRecBind
104 ------------------------------
105 simplRecBind [binders already simplfied]
106 - use simplRecOrTopPair on each pair in turn
108 simplRecOrTopPair [binder already simplified]
109 Used for: recursive bindings (top level and nested)
110 top-level non-recursive bindings
112 - check for PreInlineUnconditionally
116 Used for: non-top-level non-recursive bindings
117 beta reductions (which amount to the same thing)
118 Because it can deal with strict arts, it takes a
119 "thing-inside" and returns an expression
121 - check for PreInlineUnconditionally
122 - simplify binder, including its IdInfo
131 simplNonRecX: [given a *simplified* RHS, but an *unsimplified* binder]
132 Used for: binding case-binder and constr args in a known-constructor case
133 - check for PreInLineUnconditionally
137 ------------------------------
138 simplLazyBind: [binder already simplified, RHS not]
139 Used for: recursive bindings (top level and nested)
140 top-level non-recursive bindings
141 non-top-level, but *lazy* non-recursive bindings
142 [must not be strict or unboxed]
143 Returns floats + an augmented environment, not an expression
144 - substituteIdInfo and add result to in-scope
145 [so that rules are available in rec rhs]
148 - float if exposes constructor or PAP
152 completeNonRecX: [binder and rhs both simplified]
153 - if the the thing needs case binding (unlifted and not ok-for-spec)
159 completeLazyBind: [given a simplified RHS]
160 [used for both rec and non-rec bindings, top level and not]
161 - try PostInlineUnconditionally
162 - add unfolding [this is the only place we add an unfolding]
167 Right hand sides and arguments
168 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
169 In many ways we want to treat
170 (a) the right hand side of a let(rec), and
171 (b) a function argument
172 in the same way. But not always! In particular, we would
173 like to leave these arguments exactly as they are, so they
174 will match a RULE more easily.
179 It's harder to make the rule match if we ANF-ise the constructor,
180 or eta-expand the PAP:
182 f (let { a = g x; b = h x } in (a,b))
185 On the other hand if we see the let-defns
190 then we *do* want to ANF-ise and eta-expand, so that p and q
191 can be safely inlined.
193 Even floating lets out is a bit dubious. For let RHS's we float lets
194 out if that exposes a value, so that the value can be inlined more vigorously.
197 r = let x = e in (x,x)
199 Here, if we float the let out we'll expose a nice constructor. We did experiments
200 that showed this to be a generally good thing. But it was a bad thing to float
201 lets out unconditionally, because that meant they got allocated more often.
203 For function arguments, there's less reason to expose a constructor (it won't
204 get inlined). Just possibly it might make a rule match, but I'm pretty skeptical.
205 So for the moment we don't float lets out of function arguments either.
210 For eta expansion, we want to catch things like
212 case e of (a,b) -> \x -> case a of (p,q) -> \y -> r
214 If the \x was on the RHS of a let, we'd eta expand to bring the two
215 lambdas together. And in general that's a good thing to do. Perhaps
216 we should eta expand wherever we find a (value) lambda? Then the eta
217 expansion at a let RHS can concentrate solely on the PAP case.
220 %************************************************************************
222 \subsection{Bindings}
224 %************************************************************************
227 simplTopBinds :: SimplEnv -> [InBind] -> SimplM [OutBind]
229 simplTopBinds env binds
230 = -- Put all the top-level binders into scope at the start
231 -- so that if a transformation rule has unexpectedly brought
232 -- anything into scope, then we don't get a complaint about that.
233 -- It's rather as if the top-level binders were imported.
234 simplRecBndrs env (bindersOfBinds binds) `thenSmpl` \ (env, bndrs') ->
235 simpl_binds env binds bndrs' `thenSmpl` \ (floats, _) ->
236 freeTick SimplifierDone `thenSmpl_`
237 returnSmpl (floatBinds floats)
239 -- We need to track the zapped top-level binders, because
240 -- they should have their fragile IdInfo zapped (notably occurrence info)
241 -- That's why we run down binds and bndrs' simultaneously.
242 simpl_binds :: SimplEnv -> [InBind] -> [OutId] -> SimplM (FloatsWith ())
243 simpl_binds env [] bs = ASSERT( null bs ) returnSmpl (emptyFloats env, ())
244 simpl_binds env (bind:binds) bs = simpl_bind env bind bs `thenSmpl` \ (floats,env) ->
245 addFloats env floats $ \env ->
246 simpl_binds env binds (drop_bs bind bs)
248 drop_bs (NonRec _ _) (_ : bs) = bs
249 drop_bs (Rec prs) bs = drop (length prs) bs
251 simpl_bind env bind bs
252 = getDOptsSmpl `thenSmpl` \ dflags ->
253 if dopt Opt_D_dump_inlinings dflags then
254 pprTrace "SimplBind" (ppr (bindersOf bind)) $ simpl_bind1 env bind bs
256 simpl_bind1 env bind bs
258 simpl_bind1 env (NonRec b r) (b':_) = simplRecOrTopPair env TopLevel b b' r
259 simpl_bind1 env (Rec pairs) bs' = simplRecBind env TopLevel pairs bs'
263 %************************************************************************
265 \subsection{simplNonRec}
267 %************************************************************************
269 simplNonRecBind is used for
270 * non-top-level non-recursive lets in expressions
274 * An unsimplified (binder, rhs) pair
275 * The env for the RHS. It may not be the same as the
276 current env because the bind might occur via (\x.E) arg
278 It uses the CPS form because the binding might be strict, in which
279 case we might discard the continuation:
280 let x* = error "foo" in (...x...)
282 It needs to turn unlifted bindings into a @case@. They can arise
283 from, say: (\x -> e) (4# + 3#)
286 simplNonRecBind :: SimplEnv
288 -> InExpr -> SimplEnv -- Arg, with its subst-env
289 -> OutType -- Type of thing computed by the context
290 -> (SimplEnv -> SimplM FloatsWithExpr) -- The body
291 -> SimplM FloatsWithExpr
293 simplNonRecBind env bndr rhs rhs_se cont_ty thing_inside
295 = pprPanic "simplNonRecBind" (ppr bndr <+> ppr rhs)
298 simplNonRecBind env bndr rhs rhs_se cont_ty thing_inside
299 | preInlineUnconditionally env NotTopLevel bndr
300 = tick (PreInlineUnconditionally bndr) `thenSmpl_`
301 thing_inside (extendSubst env bndr (ContEx (getSubstEnv rhs_se) rhs))
304 | isStrictDmd (idNewDemandInfo bndr) || isStrictType (idType bndr) -- A strict let
305 = -- Don't use simplBinder because that doesn't keep
306 -- fragile occurrence info in the substitution
307 simplLetBndr env bndr `thenSmpl` \ (env, bndr1) ->
308 simplStrictArg AnRhs env rhs rhs_se (idType bndr1) cont_ty $ \ env1 rhs1 ->
310 -- Now complete the binding and simplify the body
312 -- simplLetBndr doesn't deal with the IdInfo, so we must
313 -- do so here (c.f. simplLazyBind)
314 bndr2 = bndr1 `setIdInfo` simplIdInfo (getSubst env) (idInfo bndr)
315 env2 = modifyInScope env1 bndr2 bndr2
317 completeNonRecX env2 True {- strict -} bndr bndr2 rhs1 thing_inside
319 | otherwise -- Normal, lazy case
320 = -- Don't use simplBinder because that doesn't keep
321 -- fragile occurrence info in the substitution
322 simplLetBndr env bndr `thenSmpl` \ (env, bndr') ->
323 simplLazyBind env NotTopLevel NonRecursive
324 bndr bndr' rhs rhs_se `thenSmpl` \ (floats, env) ->
325 addFloats env floats thing_inside
328 A specialised variant of simplNonRec used when the RHS is already simplified, notably
329 in knownCon. It uses case-binding where necessary.
332 simplNonRecX :: SimplEnv
333 -> InId -- Old binder
334 -> OutExpr -- Simplified RHS
335 -> (SimplEnv -> SimplM FloatsWithExpr)
336 -> SimplM FloatsWithExpr
338 simplNonRecX env bndr new_rhs thing_inside
339 | needsCaseBinding (idType bndr) new_rhs
340 -- Make this test *before* the preInlineUnconditionally
341 -- Consider case I# (quotInt# x y) of
342 -- I# v -> let w = J# v in ...
343 -- If we gaily inline (quotInt# x y) for v, we end up building an
345 -- let w = J# (quotInt# x y) in ...
346 -- because quotInt# can fail.
347 = simplBinder env bndr `thenSmpl` \ (env, bndr') ->
348 thing_inside env `thenSmpl` \ (floats, body) ->
349 returnSmpl (emptyFloats env, Case new_rhs bndr' [(DEFAULT, [], wrapFloats floats body)])
351 | preInlineUnconditionally env NotTopLevel bndr
352 -- This happens; for example, the case_bndr during case of
353 -- known constructor: case (a,b) of x { (p,q) -> ... }
354 -- Here x isn't mentioned in the RHS, so we don't want to
355 -- create the (dead) let-binding let x = (a,b) in ...
357 -- Similarly, single occurrences can be inlined vigourously
358 -- e.g. case (f x, g y) of (a,b) -> ....
359 -- If a,b occur once we can avoid constructing the let binding for them.
360 = thing_inside (extendSubst env bndr (ContEx emptySubstEnv new_rhs))
363 = simplBinder env bndr `thenSmpl` \ (env, bndr') ->
364 completeNonRecX env False {- Non-strict; pessimistic -}
365 bndr bndr' new_rhs thing_inside
367 completeNonRecX env is_strict old_bndr new_bndr new_rhs thing_inside
368 = mkAtomicArgs is_strict
369 True {- OK to float unlifted -}
370 new_rhs `thenSmpl` \ (aux_binds, rhs2) ->
372 -- Make the arguments atomic if necessary,
373 -- adding suitable bindings
374 addAtomicBindsE env (fromOL aux_binds) $ \ env ->
375 completeLazyBind env NotTopLevel
376 old_bndr new_bndr rhs2 `thenSmpl` \ (floats, env) ->
377 addFloats env floats thing_inside
381 %************************************************************************
383 \subsection{Lazy bindings}
385 %************************************************************************
387 simplRecBind is used for
388 * recursive bindings only
391 simplRecBind :: SimplEnv -> TopLevelFlag
392 -> [(InId, InExpr)] -> [OutId]
393 -> SimplM (FloatsWith SimplEnv)
394 simplRecBind env top_lvl pairs bndrs'
395 = go env pairs bndrs' `thenSmpl` \ (floats, env) ->
396 returnSmpl (flattenFloats floats, env)
398 go env [] _ = returnSmpl (emptyFloats env, env)
400 go env ((bndr, rhs) : pairs) (bndr' : bndrs')
401 = simplRecOrTopPair env top_lvl bndr bndr' rhs `thenSmpl` \ (floats, env) ->
402 addFloats env floats (\env -> go env pairs bndrs')
406 simplRecOrTopPair is used for
407 * recursive bindings (whether top level or not)
408 * top-level non-recursive bindings
410 It assumes the binder has already been simplified, but not its IdInfo.
413 simplRecOrTopPair :: SimplEnv
415 -> InId -> OutId -- Binder, both pre-and post simpl
416 -> InExpr -- The RHS and its environment
417 -> SimplM (FloatsWith SimplEnv)
419 simplRecOrTopPair env top_lvl bndr bndr' rhs
420 | preInlineUnconditionally env top_lvl bndr -- Check for unconditional inline
421 = tick (PreInlineUnconditionally bndr) `thenSmpl_`
422 returnSmpl (emptyFloats env, extendSubst env bndr (ContEx (getSubstEnv env) rhs))
425 = simplLazyBind env top_lvl Recursive bndr bndr' rhs env
426 -- May not actually be recursive, but it doesn't matter
430 simplLazyBind is used for
431 * recursive bindings (whether top level or not)
432 * top-level non-recursive bindings
433 * non-top-level *lazy* non-recursive bindings
435 [Thus it deals with the lazy cases from simplNonRecBind, and all cases
436 from SimplRecOrTopBind]
439 1. It assumes that the binder is *already* simplified,
440 and is in scope, but not its IdInfo
442 2. It assumes that the binder type is lifted.
444 3. It does not check for pre-inline-unconditionallly;
445 that should have been done already.
448 simplLazyBind :: SimplEnv
449 -> TopLevelFlag -> RecFlag
450 -> InId -> OutId -- Binder, both pre-and post simpl
451 -> InExpr -> SimplEnv -- The RHS and its environment
452 -> SimplM (FloatsWith SimplEnv)
454 simplLazyBind env top_lvl is_rec bndr bndr1 rhs rhs_se
455 = let -- Transfer the IdInfo of the original binder to the new binder
456 -- This is crucial: we must preserve
460 -- etc. To do this we must apply the current substitution,
461 -- which incorporates earlier substitutions in this very letrec group.
463 -- NB 1. We do this *before* processing the RHS of the binder, so that
464 -- its substituted rules are visible in its own RHS.
465 -- This is important. Manuel found cases where he really, really
466 -- wanted a RULE for a recursive function to apply in that function's
467 -- own right-hand side.
469 -- NB 2: We do not transfer the arity (see Subst.substIdInfo)
470 -- The arity of an Id should not be visible
471 -- in its own RHS, else we eta-reduce
475 -- which isn't sound. And it makes the arity in f's IdInfo greater than
476 -- the manifest arity, which isn't good.
477 -- The arity will get added later.
479 -- NB 3: It's important that we *do* transer the loop-breaker OccInfo,
480 -- because that's what stops the Id getting inlined infinitely, in the body
483 -- NB 4: does no harm for non-recursive bindings
485 bndr2 = bndr1 `setIdInfo` simplIdInfo (getSubst env) (idInfo bndr)
486 env1 = modifyInScope env bndr2 bndr2
487 rhs_env = setInScope rhs_se env1
488 is_top_level = isTopLevel top_lvl
489 ok_float_unlifted = not is_top_level && isNonRec is_rec
490 rhs_cont = mkStop (idType bndr1) AnRhs
492 -- Simplify the RHS; note the mkStop, which tells
493 -- the simplifier that this is the RHS of a let.
494 simplExprF rhs_env rhs rhs_cont `thenSmpl` \ (floats, rhs1) ->
496 -- If any of the floats can't be floated, give up now
497 -- (The allLifted predicate says True for empty floats.)
498 if (not ok_float_unlifted && not (allLifted floats)) then
499 completeLazyBind env1 top_lvl bndr bndr2
500 (wrapFloats floats rhs1)
503 -- ANF-ise a constructor or PAP rhs
504 mkAtomicArgs False {- Not strict -}
505 ok_float_unlifted rhs1 `thenSmpl` \ (aux_binds, rhs2) ->
507 -- If the result is a PAP, float the floats out, else wrap them
508 -- By this time it's already been ANF-ised (if necessary)
509 if isEmptyFloats floats && isNilOL aux_binds then -- Shortcut a common case
510 completeLazyBind env1 top_lvl bndr bndr2 rhs2
512 else if is_top_level || exprIsTrivial rhs2 || exprIsValue rhs2 then
513 -- WARNING: long dodgy argument coming up
514 -- WANTED: a better way to do this
516 -- We can't use "exprIsCheap" instead of exprIsValue,
517 -- because that causes a strictness bug.
518 -- x = let y* = E in case (scc y) of { T -> F; F -> T}
519 -- The case expression is 'cheap', but it's wrong to transform to
520 -- y* = E; x = case (scc y) of {...}
521 -- Either we must be careful not to float demanded non-values, or
522 -- we must use exprIsValue for the test, which ensures that the
523 -- thing is non-strict. So exprIsValue => bindings are non-strict
524 -- I think. The WARN below tests for this.
526 -- We use exprIsTrivial here because we want to reveal lone variables.
527 -- E.g. let { x = letrec { y = E } in y } in ...
528 -- Here we definitely want to float the y=E defn.
529 -- exprIsValue definitely isn't right for that.
531 -- Again, the floated binding can't be strict; if it's recursive it'll
532 -- be non-strict; if it's non-recursive it'd be inlined.
534 -- Note [SCC-and-exprIsTrivial]
536 -- y = let { x* = E } in scc "foo" x
537 -- then we do *not* want to float out the x binding, because
538 -- it's strict! Fortunately, exprIsTrivial replies False to
541 -- There's a subtlety here. There may be a binding (x* = e) in the
542 -- floats, where the '*' means 'will be demanded'. So is it safe
543 -- to float it out? Answer no, but it won't matter because
544 -- we only float if (a) arg' is a WHNF, or (b) it's going to top level
545 -- and so there can't be any 'will be demanded' bindings in the floats.
547 ASSERT2( is_top_level || not (any demanded_float (floatBinds floats)),
548 ppr (filter demanded_float (floatBinds floats)) )
550 tick LetFloatFromLet `thenSmpl_` (
551 addFloats env1 floats $ \ env2 ->
552 addAtomicBinds env2 (fromOL aux_binds) $ \ env3 ->
553 completeLazyBind env3 top_lvl bndr bndr2 rhs2)
556 completeLazyBind env1 top_lvl bndr bndr2 (wrapFloats floats rhs1)
559 demanded_float (NonRec b r) = isStrictDmd (idNewDemandInfo b) && not (isUnLiftedType (idType b))
560 -- Unlifted-type (cheap-eagerness) lets may well have a demanded flag on them
561 demanded_float (Rec _) = False
566 %************************************************************************
568 \subsection{Completing a lazy binding}
570 %************************************************************************
573 * deals only with Ids, not TyVars
574 * takes an already-simplified binder and RHS
575 * is used for both recursive and non-recursive bindings
576 * is used for both top-level and non-top-level bindings
578 It does the following:
579 - tries discarding a dead binding
580 - tries PostInlineUnconditionally
581 - add unfolding [this is the only place we add an unfolding]
584 It does *not* attempt to do let-to-case. Why? Because it is used for
585 - top-level bindings (when let-to-case is impossible)
586 - many situations where the "rhs" is known to be a WHNF
587 (so let-to-case is inappropriate).
590 completeLazyBind :: SimplEnv
591 -> TopLevelFlag -- Flag stuck into unfolding
592 -> InId -- Old binder
593 -> OutId -- New binder
594 -> OutExpr -- Simplified RHS
595 -> SimplM (FloatsWith SimplEnv)
596 -- We return a new SimplEnv, because completeLazyBind may choose to do its work
597 -- by extending the substitution (e.g. let x = y in ...)
598 -- The new binding (if any) is returned as part of the floats.
599 -- NB: the returned SimplEnv has the right SubstEnv, but you should
600 -- (as usual) use the in-scope-env from the floats
602 completeLazyBind env top_lvl old_bndr new_bndr new_rhs
603 | postInlineUnconditionally env new_bndr occ_info new_rhs
604 = -- Drop the binding
605 tick (PostInlineUnconditionally old_bndr) `thenSmpl_`
606 returnSmpl (emptyFloats env, extendSubst env old_bndr (DoneEx new_rhs))
607 -- Use the substitution to make quite, quite sure that the substitution
608 -- will happen, since we are going to discard the binding
613 new_bndr_info = idInfo new_bndr `setArityInfo` exprArity new_rhs
615 -- Add the unfolding *only* for non-loop-breakers
616 -- Making loop breakers not have an unfolding at all
617 -- means that we can avoid tests in exprIsConApp, for example.
618 -- This is important: if exprIsConApp says 'yes' for a recursive
619 -- thing, then we can get into an infinite loop
620 info_w_unf | loop_breaker = new_bndr_info
621 | otherwise = new_bndr_info `setUnfoldingInfo` unfolding
622 unfolding = mkUnfolding (isTopLevel top_lvl) new_rhs
624 final_id = new_bndr `setIdInfo` info_w_unf
626 -- These seqs forces the Id, and hence its IdInfo,
627 -- and hence any inner substitutions
629 returnSmpl (unitFloat env final_id new_rhs, env)
632 loop_breaker = isLoopBreaker occ_info
633 old_info = idInfo old_bndr
634 occ_info = occInfo old_info
639 %************************************************************************
641 \subsection[Simplify-simplExpr]{The main function: simplExpr}
643 %************************************************************************
645 The reason for this OutExprStuff stuff is that we want to float *after*
646 simplifying a RHS, not before. If we do so naively we get quadratic
647 behaviour as things float out.
649 To see why it's important to do it after, consider this (real) example:
663 a -- Can't inline a this round, cos it appears twice
667 Each of the ==> steps is a round of simplification. We'd save a
668 whole round if we float first. This can cascade. Consider
673 let f = let d1 = ..d.. in \y -> e
677 in \x -> ...(\y ->e)...
679 Only in this second round can the \y be applied, and it
680 might do the same again.
684 simplExpr :: SimplEnv -> CoreExpr -> SimplM CoreExpr
685 simplExpr env expr = simplExprC env expr (mkStop expr_ty' AnArg)
687 expr_ty' = substTy (getSubst env) (exprType expr)
688 -- The type in the Stop continuation, expr_ty', is usually not used
689 -- It's only needed when discarding continuations after finding
690 -- a function that returns bottom.
691 -- Hence the lazy substitution
694 simplExprC :: SimplEnv -> CoreExpr -> SimplCont -> SimplM CoreExpr
695 -- Simplify an expression, given a continuation
696 simplExprC env expr cont
697 = simplExprF env expr cont `thenSmpl` \ (floats, expr) ->
698 returnSmpl (wrapFloats floats expr)
700 simplExprF :: SimplEnv -> InExpr -> SimplCont -> SimplM FloatsWithExpr
701 -- Simplify an expression, returning floated binds
703 simplExprF env (Var v) cont = simplVar env v cont
704 simplExprF env (Lit lit) cont = rebuild env (Lit lit) cont
705 simplExprF env expr@(Lam _ _) cont = simplLam env expr cont
706 simplExprF env (Note note expr) cont = simplNote env note expr cont
707 simplExprF env (App fun arg) cont = simplExprF env fun (ApplyTo NoDup arg env cont)
709 simplExprF env (Type ty) cont
710 = ASSERT( contIsRhsOrArg cont )
711 simplType env ty `thenSmpl` \ ty' ->
712 rebuild env (Type ty') cont
714 simplExprF env (Case scrut bndr alts) cont
715 | not (switchIsOn (getSwitchChecker env) NoCaseOfCase)
716 = -- Simplify the scrutinee with a Select continuation
717 simplExprF env scrut (Select NoDup bndr alts env cont)
720 = -- If case-of-case is off, simply simplify the case expression
721 -- in a vanilla Stop context, and rebuild the result around it
722 simplExprC env scrut case_cont `thenSmpl` \ case_expr' ->
723 rebuild env case_expr' cont
725 case_cont = Select NoDup bndr alts env (mkBoringStop (contResultType cont))
727 simplExprF env (Let (Rec pairs) body) cont
728 = simplRecBndrs env (map fst pairs) `thenSmpl` \ (env, bndrs') ->
729 -- NB: bndrs' don't have unfoldings or rules
730 -- We add them as we go down
732 simplRecBind env NotTopLevel pairs bndrs' `thenSmpl` \ (floats, env) ->
733 addFloats env floats $ \ env ->
734 simplExprF env body cont
736 -- A non-recursive let is dealt with by simplNonRecBind
737 simplExprF env (Let (NonRec bndr rhs) body) cont
738 = simplNonRecBind env bndr rhs env (contResultType cont) $ \ env ->
739 simplExprF env body cont
742 ---------------------------------
743 simplType :: SimplEnv -> InType -> SimplM OutType
744 -- Kept monadic just so we can do the seqType
746 = seqType new_ty `seq` returnSmpl new_ty
748 new_ty = substTy (getSubst env) ty
752 %************************************************************************
756 %************************************************************************
759 simplLam env fun cont
762 zap_it = mkLamBndrZapper fun (countArgs cont)
763 cont_ty = contResultType cont
765 -- Type-beta reduction
766 go env (Lam bndr body) (ApplyTo _ (Type ty_arg) arg_se body_cont)
767 = ASSERT( isTyVar bndr )
768 tick (BetaReduction bndr) `thenSmpl_`
769 simplType (setInScope arg_se env) ty_arg `thenSmpl` \ ty_arg' ->
770 go (extendSubst env bndr (DoneTy ty_arg')) body body_cont
772 -- Ordinary beta reduction
773 go env (Lam bndr body) cont@(ApplyTo _ arg arg_se body_cont)
774 = tick (BetaReduction bndr) `thenSmpl_`
775 simplNonRecBind env (zap_it bndr) arg arg_se cont_ty $ \ env ->
776 go env body body_cont
778 -- Not enough args, so there are real lambdas left to put in the result
779 go env lam@(Lam _ _) cont
780 = simplLamBndrs env bndrs `thenSmpl` \ (env, bndrs') ->
781 simplExpr env body `thenSmpl` \ body' ->
782 mkLam env bndrs' body' cont `thenSmpl` \ (floats, new_lam) ->
783 addFloats env floats $ \ env ->
784 rebuild env new_lam cont
786 (bndrs,body) = collectBinders lam
788 -- Exactly enough args
789 go env expr cont = simplExprF env expr cont
791 mkLamBndrZapper :: CoreExpr -- Function
792 -> Int -- Number of args supplied, *including* type args
793 -> Id -> Id -- Use this to zap the binders
794 mkLamBndrZapper fun n_args
795 | n_args >= n_params fun = \b -> b -- Enough args
796 | otherwise = \b -> zapLamIdInfo b
798 -- NB: we count all the args incl type args
799 -- so we must count all the binders (incl type lambdas)
800 n_params (Note _ e) = n_params e
801 n_params (Lam b e) = 1 + n_params e
802 n_params other = 0::Int
806 %************************************************************************
810 %************************************************************************
813 simplNote env (Coerce to from) body cont
815 in_scope = getInScope env
817 addCoerce s1 k1 (CoerceIt t1 cont)
818 -- coerce T1 S1 (coerce S1 K1 e)
821 -- coerce T1 K1 e, otherwise
823 -- For example, in the initial form of a worker
824 -- we may find (coerce T (coerce S (\x.e))) y
825 -- and we'd like it to simplify to e[y/x] in one round
827 | t1 `eqType` k1 = cont -- The coerces cancel out
828 | otherwise = CoerceIt t1 cont -- They don't cancel, but
829 -- the inner one is redundant
831 addCoerce t1t2 s1s2 (ApplyTo dup arg arg_se cont)
832 | not (isTypeArg arg), -- This whole case only works for value args
833 -- Could upgrade to have equiv thing for type apps too
834 Just (s1, s2) <- splitFunTy_maybe s1s2
835 -- (coerce (T1->T2) (S1->S2) F) E
837 -- coerce T2 S2 (F (coerce S1 T1 E))
839 -- t1t2 must be a function type, T1->T2, because it's applied to something
840 -- but s1s2 might conceivably not be
842 -- When we build the ApplyTo we can't mix the out-types
843 -- with the InExpr in the argument, so we simply substitute
844 -- to make it all consistent. It's a bit messy.
845 -- But it isn't a common case.
847 (t1,t2) = splitFunTy t1t2
848 new_arg = mkCoerce2 s1 t1 (substExpr (mkSubst in_scope (getSubstEnv arg_se)) arg)
850 ApplyTo dup new_arg (zapSubstEnv env) (addCoerce t2 s2 cont)
852 addCoerce to' _ cont = CoerceIt to' cont
854 simplType env to `thenSmpl` \ to' ->
855 simplType env from `thenSmpl` \ from' ->
856 simplExprF env body (addCoerce to' from' cont)
859 -- Hack: we only distinguish subsumed cost centre stacks for the purposes of
860 -- inlining. All other CCCSs are mapped to currentCCS.
861 simplNote env (SCC cc) e cont
862 = simplExpr (setEnclosingCC env currentCCS) e `thenSmpl` \ e' ->
863 rebuild env (mkSCC cc e') cont
865 simplNote env InlineCall e cont
866 = simplExprF env e (InlinePlease cont)
868 -- See notes with SimplMonad.inlineMode
869 simplNote env InlineMe e cont
870 | contIsRhsOrArg cont -- Totally boring continuation; see notes above
871 = -- Don't inline inside an INLINE expression
872 simplExpr (setMode inlineMode env ) e `thenSmpl` \ e' ->
873 rebuild env (mkInlineMe e') cont
875 | otherwise -- Dissolve the InlineMe note if there's
876 -- an interesting context of any kind to combine with
877 -- (even a type application -- anything except Stop)
878 = simplExprF env e cont
880 simplNote env (CoreNote s) e cont
881 = simplExpr env e `thenSmpl` \ e' ->
882 rebuild env (Note (CoreNote s) e') cont
886 %************************************************************************
888 \subsection{Dealing with calls}
890 %************************************************************************
893 simplVar env var cont
894 = case lookupIdSubst (getSubst env) var of
895 DoneEx e -> simplExprF (zapSubstEnv env) e cont
896 ContEx se e -> simplExprF (setSubstEnv env se) e cont
897 DoneId var1 occ -> WARN( not (isInScope var1 (getSubst env)) && mustHaveLocalBinding var1,
898 text "simplVar:" <+> ppr var )
899 completeCall (zapSubstEnv env) var1 occ cont
900 -- The template is already simplified, so don't re-substitute.
901 -- This is VITAL. Consider
903 -- let y = \z -> ...x... in
905 -- We'll clone the inner \x, adding x->x' in the id_subst
906 -- Then when we inline y, we must *not* replace x by x' in
907 -- the inlined copy!!
909 ---------------------------------------------------------
910 -- Dealing with a call site
912 completeCall env var occ_info cont
913 = -- Simplify the arguments
914 getDOptsSmpl `thenSmpl` \ dflags ->
916 chkr = getSwitchChecker env
917 (args, call_cont, inline_call) = getContArgs chkr var cont
920 simplifyArgs env fn_ty args (contResultType call_cont) $ \ env args ->
922 -- Next, look for rules or specialisations that match
924 -- It's important to simplify the args first, because the rule-matcher
925 -- doesn't do substitution as it goes. We don't want to use subst_args
926 -- (defined in the 'where') because that throws away useful occurrence info,
927 -- and perhaps-very-important specialisations.
929 -- Some functions have specialisations *and* are strict; in this case,
930 -- we don't want to inline the wrapper of the non-specialised thing; better
931 -- to call the specialised thing instead.
932 -- We used to use the black-listing mechanism to ensure that inlining of
933 -- the wrapper didn't occur for things that have specialisations till a
934 -- later phase, so but now we just try RULES first
936 -- You might think that we shouldn't apply rules for a loop breaker:
937 -- doing so might give rise to an infinite loop, because a RULE is
938 -- rather like an extra equation for the function:
939 -- RULE: f (g x) y = x+y
942 -- But it's too drastic to disable rules for loop breakers.
943 -- Even the foldr/build rule would be disabled, because foldr
944 -- is recursive, and hence a loop breaker:
945 -- foldr k z (build g) = g k z
946 -- So it's up to the programmer: rules can cause divergence
949 in_scope = getInScope env
950 maybe_rule = case activeRule env of
951 Nothing -> Nothing -- No rules apply
952 Just act_fn -> lookupRule act_fn in_scope var args
955 Just (rule_name, rule_rhs) ->
956 tick (RuleFired rule_name) `thenSmpl_`
957 (if dopt Opt_D_dump_inlinings dflags then
958 pprTrace "Rule fired" (vcat [
959 text "Rule:" <+> ftext rule_name,
960 text "Before:" <+> ppr var <+> sep (map pprParendExpr args),
961 text "After: " <+> pprCoreExpr rule_rhs,
962 text "Cont: " <+> ppr call_cont])
965 simplExprF env rule_rhs call_cont ;
967 Nothing -> -- No rules
969 -- Next, look for an inlining
971 arg_infos = [ interestingArg arg | arg <- args, isValArg arg]
973 interesting_cont = interestingCallContext (notNull args)
977 active_inline = activeInline env var occ_info
978 maybe_inline = callSiteInline dflags active_inline inline_call occ_info
979 var arg_infos interesting_cont
981 case maybe_inline of {
982 Just unfolding -- There is an inlining!
983 -> tick (UnfoldingDone var) `thenSmpl_`
984 makeThatCall env var unfolding args call_cont
987 Nothing -> -- No inlining!
990 rebuild env (mkApps (Var var) args) call_cont
993 makeThatCall :: SimplEnv
995 -> InExpr -- Inlined function rhs
996 -> [OutExpr] -- Arguments, already simplified
997 -> SimplCont -- After the call
998 -> SimplM FloatsWithExpr
999 -- Similar to simplLam, but this time
1000 -- the arguments are already simplified
1001 makeThatCall orig_env var fun@(Lam _ _) args cont
1002 = go orig_env fun args
1004 zap_it = mkLamBndrZapper fun (length args)
1006 -- Type-beta reduction
1007 go env (Lam bndr body) (Type ty_arg : args)
1008 = ASSERT( isTyVar bndr )
1009 tick (BetaReduction bndr) `thenSmpl_`
1010 go (extendSubst env bndr (DoneTy ty_arg)) body args
1012 -- Ordinary beta reduction
1013 go env (Lam bndr body) (arg : args)
1014 = tick (BetaReduction bndr) `thenSmpl_`
1015 simplNonRecX env (zap_it bndr) arg $ \ env ->
1018 -- Not enough args, so there are real lambdas left to put in the result
1020 = simplExprF env fun (pushContArgs orig_env args cont)
1021 -- NB: orig_env; the correct environment to capture with
1022 -- the arguments.... env has been augmented with substitutions
1023 -- from the beta reductions.
1025 makeThatCall env var fun args cont
1026 = simplExprF env fun (pushContArgs env args cont)
1030 %************************************************************************
1032 \subsection{Arguments}
1034 %************************************************************************
1037 ---------------------------------------------------------
1038 -- Simplifying the arguments of a call
1040 simplifyArgs :: SimplEnv
1041 -> OutType -- Type of the function
1042 -> [(InExpr, SimplEnv, Bool)] -- Details of the arguments
1043 -> OutType -- Type of the continuation
1044 -> (SimplEnv -> [OutExpr] -> SimplM FloatsWithExpr)
1045 -> SimplM FloatsWithExpr
1047 -- [CPS-like because of strict arguments]
1049 -- Simplify the arguments to a call.
1050 -- This part of the simplifier may break the no-shadowing invariant
1052 -- f (...(\a -> e)...) (case y of (a,b) -> e')
1053 -- where f is strict in its second arg
1054 -- If we simplify the innermost one first we get (...(\a -> e)...)
1055 -- Simplifying the second arg makes us float the case out, so we end up with
1056 -- case y of (a,b) -> f (...(\a -> e)...) e'
1057 -- So the output does not have the no-shadowing invariant. However, there is
1058 -- no danger of getting name-capture, because when the first arg was simplified
1059 -- we used an in-scope set that at least mentioned all the variables free in its
1060 -- static environment, and that is enough.
1062 -- We can't just do innermost first, or we'd end up with a dual problem:
1063 -- case x of (a,b) -> f e (...(\a -> e')...)
1065 -- I spent hours trying to recover the no-shadowing invariant, but I just could
1066 -- not think of an elegant way to do it. The simplifier is already knee-deep in
1067 -- continuations. We have to keep the right in-scope set around; AND we have
1068 -- to get the effect that finding (error "foo") in a strict arg position will
1069 -- discard the entire application and replace it with (error "foo"). Getting
1070 -- all this at once is TOO HARD!
1072 simplifyArgs env fn_ty args cont_ty thing_inside
1073 = go env fn_ty args thing_inside
1075 go env fn_ty [] thing_inside = thing_inside env []
1076 go env fn_ty (arg:args) thing_inside = simplifyArg env fn_ty arg cont_ty $ \ env arg' ->
1077 go env (applyTypeToArg fn_ty arg') args $ \ env args' ->
1078 thing_inside env (arg':args')
1080 simplifyArg env fn_ty (Type ty_arg, se, _) cont_ty thing_inside
1081 = simplType (setInScope se env) ty_arg `thenSmpl` \ new_ty_arg ->
1082 thing_inside env (Type new_ty_arg)
1084 simplifyArg env fn_ty (val_arg, arg_se, is_strict) cont_ty thing_inside
1086 = simplStrictArg AnArg env val_arg arg_se arg_ty cont_ty thing_inside
1088 | otherwise -- Lazy argument
1089 -- DO NOT float anything outside, hence simplExprC
1090 -- There is no benefit (unlike in a let-binding), and we'd
1091 -- have to be very careful about bogus strictness through
1092 -- floating a demanded let.
1093 = simplExprC (setInScope arg_se env) val_arg
1094 (mkStop arg_ty AnArg) `thenSmpl` \ arg1 ->
1095 thing_inside env arg1
1097 arg_ty = funArgTy fn_ty
1100 simplStrictArg :: LetRhsFlag
1101 -> SimplEnv -- The env of the call
1102 -> InExpr -> SimplEnv -- The arg plus its env
1103 -> OutType -- arg_ty: type of the argument
1104 -> OutType -- cont_ty: Type of thing computed by the context
1105 -> (SimplEnv -> OutExpr -> SimplM FloatsWithExpr)
1106 -- Takes an expression of type rhs_ty,
1107 -- returns an expression of type cont_ty
1108 -- The env passed to this continuation is the
1109 -- env of the call, plus any new in-scope variables
1110 -> SimplM FloatsWithExpr -- An expression of type cont_ty
1112 simplStrictArg is_rhs call_env arg arg_env arg_ty cont_ty thing_inside
1113 = simplExprF (setInScope arg_env call_env) arg
1114 (ArgOf is_rhs arg_ty cont_ty (\ new_env -> thing_inside (setInScope call_env new_env)))
1115 -- Notice the way we use arg_env (augmented with in-scope vars from call_env)
1116 -- to simplify the argument
1117 -- and call-env (augmented with in-scope vars from the arg) to pass to the continuation
1121 %************************************************************************
1123 \subsection{mkAtomicArgs}
1125 %************************************************************************
1127 mkAtomicArgs takes a putative RHS, checks whether it's a PAP or
1128 constructor application and, if so, converts it to ANF, so that the
1129 resulting thing can be inlined more easily. Thus
1136 There are three sorts of binding context, specified by the two
1142 N N Top-level or recursive Only bind args of lifted type
1144 N Y Non-top-level and non-recursive, Bind args of lifted type, or
1145 but lazy unlifted-and-ok-for-speculation
1147 Y Y Non-top-level, non-recursive, Bind all args
1148 and strict (demanded)
1155 there is no point in transforming to
1157 x = case (y div# z) of r -> MkC r
1159 because the (y div# z) can't float out of the let. But if it was
1160 a *strict* let, then it would be a good thing to do. Hence the
1161 context information.
1164 mkAtomicArgs :: Bool -- A strict binding
1165 -> Bool -- OK to float unlifted args
1167 -> SimplM (OrdList (OutId,OutExpr), -- The floats (unusually) may include
1168 OutExpr) -- things that need case-binding,
1169 -- if the strict-binding flag is on
1171 mkAtomicArgs is_strict ok_float_unlifted rhs
1172 | (Var fun, args) <- collectArgs rhs, -- It's an application
1173 isDataConWorkId fun || valArgCount args < idArity fun -- And it's a constructor or PAP
1174 = go fun nilOL [] args -- Have a go
1176 | otherwise = bale_out -- Give up
1179 bale_out = returnSmpl (nilOL, rhs)
1181 go fun binds rev_args []
1182 = returnSmpl (binds, mkApps (Var fun) (reverse rev_args))
1184 go fun binds rev_args (arg : args)
1185 | exprIsTrivial arg -- Easy case
1186 = go fun binds (arg:rev_args) args
1188 | not can_float_arg -- Can't make this arg atomic
1189 = bale_out -- ... so give up
1191 | otherwise -- Don't forget to do it recursively
1192 -- E.g. x = a:b:c:[]
1193 = mkAtomicArgs is_strict ok_float_unlifted arg `thenSmpl` \ (arg_binds, arg') ->
1194 newId FSLIT("a") arg_ty `thenSmpl` \ arg_id ->
1195 go fun ((arg_binds `snocOL` (arg_id,arg')) `appOL` binds)
1196 (Var arg_id : rev_args) args
1198 arg_ty = exprType arg
1199 can_float_arg = is_strict
1200 || not (isUnLiftedType arg_ty)
1201 || (ok_float_unlifted && exprOkForSpeculation arg)
1204 addAtomicBinds :: SimplEnv -> [(OutId,OutExpr)]
1205 -> (SimplEnv -> SimplM (FloatsWith a))
1206 -> SimplM (FloatsWith a)
1207 addAtomicBinds env [] thing_inside = thing_inside env
1208 addAtomicBinds env ((v,r):bs) thing_inside = addAuxiliaryBind env (NonRec v r) $ \ env ->
1209 addAtomicBinds env bs thing_inside
1211 addAtomicBindsE :: SimplEnv -> [(OutId,OutExpr)]
1212 -> (SimplEnv -> SimplM FloatsWithExpr)
1213 -> SimplM FloatsWithExpr
1214 -- Same again, but this time we're in an expression context,
1215 -- and may need to do some case bindings
1217 addAtomicBindsE env [] thing_inside
1219 addAtomicBindsE env ((v,r):bs) thing_inside
1220 | needsCaseBinding (idType v) r
1221 = addAtomicBindsE (addNewInScopeIds env [v]) bs thing_inside `thenSmpl` \ (floats, expr) ->
1222 WARN( exprIsTrivial expr, ppr v <+> pprCoreExpr expr )
1223 returnSmpl (emptyFloats env, Case r v [(DEFAULT,[], wrapFloats floats expr)])
1226 = addAuxiliaryBind env (NonRec v r) $ \ env ->
1227 addAtomicBindsE env bs thing_inside
1231 %************************************************************************
1233 \subsection{The main rebuilder}
1235 %************************************************************************
1238 rebuild :: SimplEnv -> OutExpr -> SimplCont -> SimplM FloatsWithExpr
1240 rebuild env expr (Stop _ _ _) = rebuildDone env expr
1241 rebuild env expr (ArgOf _ _ _ cont_fn) = cont_fn env expr
1242 rebuild env expr (CoerceIt to_ty cont) = rebuild env (mkCoerce to_ty expr) cont
1243 rebuild env expr (InlinePlease cont) = rebuild env (Note InlineCall expr) cont
1244 rebuild env expr (Select _ bndr alts se cont) = rebuildCase (setInScope se env) expr bndr alts cont
1245 rebuild env expr (ApplyTo _ arg se cont) = rebuildApp (setInScope se env) expr arg cont
1247 rebuildApp env fun arg cont
1248 = simplExpr env arg `thenSmpl` \ arg' ->
1249 rebuild env (App fun arg') cont
1251 rebuildDone env expr = returnSmpl (emptyFloats env, expr)
1255 %************************************************************************
1257 \subsection{Functions dealing with a case}
1259 %************************************************************************
1261 Blob of helper functions for the "case-of-something-else" situation.
1264 ---------------------------------------------------------
1265 -- Eliminate the case if possible
1267 rebuildCase :: SimplEnv
1268 -> OutExpr -- Scrutinee
1269 -> InId -- Case binder
1270 -> [InAlt] -- Alternatives
1272 -> SimplM FloatsWithExpr
1274 rebuildCase env scrut case_bndr alts cont
1275 | Just (con,args) <- exprIsConApp_maybe scrut
1276 -- Works when the scrutinee is a variable with a known unfolding
1277 -- as well as when it's an explicit constructor application
1278 = knownCon env (DataAlt con) args case_bndr alts cont
1280 | Lit lit <- scrut -- No need for same treatment as constructors
1281 -- because literals are inlined more vigorously
1282 = knownCon env (LitAlt lit) [] case_bndr alts cont
1285 = prepareAlts scrut case_bndr alts `thenSmpl` \ (better_alts, handled_cons) ->
1287 -- Deal with the case binder, and prepare the continuation;
1288 -- The new subst_env is in place
1289 prepareCaseCont env better_alts cont `thenSmpl` \ (floats, (dup_cont, nondup_cont)) ->
1290 addFloats env floats $ \ env ->
1292 -- Deal with variable scrutinee
1293 simplCaseBinder env scrut case_bndr `thenSmpl` \ (alt_env, case_bndr', zap_occ_info) ->
1295 -- Deal with the case alternatives
1296 simplAlts alt_env zap_occ_info handled_cons
1297 case_bndr' better_alts dup_cont `thenSmpl` \ alts' ->
1299 -- Put the case back together
1300 mkCase scrut case_bndr' alts' `thenSmpl` \ case_expr ->
1302 -- Notice that rebuildDone returns the in-scope set from env, not alt_env
1303 -- The case binder *not* scope over the whole returned case-expression
1304 rebuild env case_expr nondup_cont
1307 simplCaseBinder checks whether the scrutinee is a variable, v. If so,
1308 try to eliminate uses of v in the RHSs in favour of case_bndr; that
1309 way, there's a chance that v will now only be used once, and hence
1314 There is a time we *don't* want to do that, namely when
1315 -fno-case-of-case is on. This happens in the first simplifier pass,
1316 and enhances full laziness. Here's the bad case:
1317 f = \ y -> ...(case x of I# v -> ...(case x of ...) ... )
1318 If we eliminate the inner case, we trap it inside the I# v -> arm,
1319 which might prevent some full laziness happening. I've seen this
1320 in action in spectral/cichelli/Prog.hs:
1321 [(m,n) | m <- [1..max], n <- [1..max]]
1322 Hence the check for NoCaseOfCase.
1326 There is another situation when we don't want to do it. If we have
1328 case x of w1 { DEFAULT -> case x of w2 { A -> e1; B -> e2 }
1329 ...other cases .... }
1331 We'll perform the binder-swap for the outer case, giving
1333 case x of w1 { DEFAULT -> case w1 of w2 { A -> e1; B -> e2 }
1334 ...other cases .... }
1336 But there is no point in doing it for the inner case, because w1 can't
1337 be inlined anyway. Furthermore, doing the case-swapping involves
1338 zapping w2's occurrence info (see paragraphs that follow), and that
1339 forces us to bind w2 when doing case merging. So we get
1341 case x of w1 { A -> let w2 = w1 in e1
1342 B -> let w2 = w1 in e2
1343 ...other cases .... }
1345 This is plain silly in the common case where w2 is dead.
1347 Even so, I can't see a good way to implement this idea. I tried
1348 not doing the binder-swap if the scrutinee was already evaluated
1349 but that failed big-time:
1353 case v of w { MkT x ->
1354 case x of x1 { I# y1 ->
1355 case x of x2 { I# y2 -> ...
1357 Notice that because MkT is strict, x is marked "evaluated". But to
1358 eliminate the last case, we must either make sure that x (as well as
1359 x1) has unfolding MkT y1. THe straightforward thing to do is to do
1360 the binder-swap. So this whole note is a no-op.
1364 If we replace the scrutinee, v, by tbe case binder, then we have to nuke
1365 any occurrence info (eg IAmDead) in the case binder, because the
1366 case-binder now effectively occurs whenever v does. AND we have to do
1367 the same for the pattern-bound variables! Example:
1369 (case x of { (a,b) -> a }) (case x of { (p,q) -> q })
1371 Here, b and p are dead. But when we move the argment inside the first
1372 case RHS, and eliminate the second case, we get
1374 case x or { (a,b) -> a b }
1376 Urk! b is alive! Reason: the scrutinee was a variable, and case elimination
1377 happened. Hence the zap_occ_info function returned by simplCaseBinder
1380 simplCaseBinder env (Var v) case_bndr
1381 | not (switchIsOn (getSwitchChecker env) NoCaseOfCase)
1383 -- Failed try [see Note 2 above]
1384 -- not (isEvaldUnfolding (idUnfolding v))
1386 = simplBinder env (zap case_bndr) `thenSmpl` \ (env, case_bndr') ->
1387 returnSmpl (modifyInScope env v case_bndr', case_bndr', zap)
1388 -- We could extend the substitution instead, but it would be
1389 -- a hack because then the substitution wouldn't be idempotent
1390 -- any more (v is an OutId). And this just just as well.
1392 zap b = b `setIdOccInfo` NoOccInfo
1394 simplCaseBinder env other_scrut case_bndr
1395 = simplBinder env case_bndr `thenSmpl` \ (env, case_bndr') ->
1396 returnSmpl (env, case_bndr', \ bndr -> bndr) -- NoOp on bndr
1402 simplAlts :: SimplEnv
1403 -> (InId -> InId) -- Occ-info zapper
1404 -> [AltCon] -- Alternatives the scrutinee can't be
1405 -- in the default case
1406 -> OutId -- Case binder
1407 -> [InAlt] -> SimplCont
1408 -> SimplM [OutAlt] -- Includes the continuation
1410 simplAlts env zap_occ_info handled_cons case_bndr' alts cont'
1411 = mapSmpl simpl_alt alts
1413 inst_tys' = tyConAppArgs (idType case_bndr')
1415 simpl_alt (DEFAULT, _, rhs)
1417 -- In the default case we record the constructors that the
1418 -- case-binder *can't* be.
1419 -- We take advantage of any OtherCon info in the case scrutinee
1420 case_bndr_w_unf = case_bndr' `setIdUnfolding` mkOtherCon handled_cons
1421 env_with_unf = modifyInScope env case_bndr' case_bndr_w_unf
1423 simplExprC env_with_unf rhs cont' `thenSmpl` \ rhs' ->
1424 returnSmpl (DEFAULT, [], rhs')
1426 simpl_alt (con, vs, rhs)
1427 = -- Deal with the pattern-bound variables
1428 -- Mark the ones that are in ! positions in the data constructor
1429 -- as certainly-evaluated.
1430 -- NB: it happens that simplBinders does *not* erase the OtherCon
1431 -- form of unfolding, so it's ok to add this info before
1432 -- doing simplBinders
1433 simplBinders env (add_evals con vs) `thenSmpl` \ (env, vs') ->
1435 -- Bind the case-binder to (con args)
1437 unfolding = mkUnfolding False (mkAltExpr con vs' inst_tys')
1438 env_with_unf = modifyInScope env case_bndr' (case_bndr' `setIdUnfolding` unfolding)
1440 simplExprC env_with_unf rhs cont' `thenSmpl` \ rhs' ->
1441 returnSmpl (con, vs', rhs')
1444 -- add_evals records the evaluated-ness of the bound variables of
1445 -- a case pattern. This is *important*. Consider
1446 -- data T = T !Int !Int
1448 -- case x of { T a b -> T (a+1) b }
1450 -- We really must record that b is already evaluated so that we don't
1451 -- go and re-evaluate it when constructing the result.
1453 add_evals (DataAlt dc) vs = cat_evals vs (dataConRepStrictness dc)
1454 add_evals other_con vs = vs
1456 cat_evals [] [] = []
1457 cat_evals (v:vs) (str:strs)
1458 | isTyVar v = v : cat_evals vs (str:strs)
1459 | isMarkedStrict str = evald_v : cat_evals vs strs
1460 | otherwise = zapped_v : cat_evals vs strs
1462 zapped_v = zap_occ_info v
1463 evald_v = zapped_v `setIdUnfolding` mkOtherCon []
1467 %************************************************************************
1469 \subsection{Known constructor}
1471 %************************************************************************
1473 We are a bit careful with occurrence info. Here's an example
1475 (\x* -> case x of (a*, b) -> f a) (h v, e)
1477 where the * means "occurs once". This effectively becomes
1478 case (h v, e) of (a*, b) -> f a)
1480 let a* = h v; b = e in f a
1484 All this should happen in one sweep.
1487 knownCon :: SimplEnv -> AltCon -> [OutExpr]
1488 -> InId -> [InAlt] -> SimplCont
1489 -> SimplM FloatsWithExpr
1491 knownCon env con args bndr alts cont
1492 = tick (KnownBranch bndr) `thenSmpl_`
1493 case findAlt con alts of
1494 (DEFAULT, bs, rhs) -> ASSERT( null bs )
1495 simplNonRecX env bndr scrut $ \ env ->
1496 -- This might give rise to a binding with non-atomic args
1497 -- like x = Node (f x) (g x)
1498 -- but no harm will be done
1499 simplExprF env rhs cont
1502 LitAlt lit -> Lit lit
1503 DataAlt dc -> mkConApp dc args
1505 (LitAlt lit, bs, rhs) -> ASSERT( null bs )
1506 simplNonRecX env bndr (Lit lit) $ \ env ->
1507 simplExprF env rhs cont
1509 (DataAlt dc, bs, rhs) -> ASSERT( length bs + n_tys == length args )
1510 bind_args env bs (drop n_tys args) $ \ env ->
1512 con_app = mkConApp dc (take n_tys args ++ con_args)
1513 con_args = [substExpr (getSubst env) (varToCoreExpr b) | b <- bs]
1514 -- args are aready OutExprs, but bs are InIds
1516 simplNonRecX env bndr con_app $ \ env ->
1517 simplExprF env rhs cont
1519 n_tys = dataConNumInstArgs dc -- Non-existential type args
1521 bind_args env [] _ thing_inside = thing_inside env
1523 bind_args env (b:bs) (Type ty : args) thing_inside
1524 = bind_args (extendSubst env b (DoneTy ty)) bs args thing_inside
1526 bind_args env (b:bs) (arg : args) thing_inside
1527 = simplNonRecX env b arg $ \ env ->
1528 bind_args env bs args thing_inside
1532 %************************************************************************
1534 \subsection{Duplicating continuations}
1536 %************************************************************************
1539 prepareCaseCont :: SimplEnv
1540 -> [InAlt] -> SimplCont
1541 -> SimplM (FloatsWith (SimplCont,SimplCont))
1542 -- Return a duplicatable continuation, a non-duplicable part
1543 -- plus some extra bindings
1545 -- No need to make it duplicatable if there's only one alternative
1546 prepareCaseCont env [alt] cont = returnSmpl (emptyFloats env, (cont, mkBoringStop (contResultType cont)))
1547 prepareCaseCont env alts cont = mkDupableCont env cont
1551 mkDupableCont :: SimplEnv -> SimplCont
1552 -> SimplM (FloatsWith (SimplCont, SimplCont))
1554 mkDupableCont env cont
1555 | contIsDupable cont
1556 = returnSmpl (emptyFloats env, (cont, mkBoringStop (contResultType cont)))
1558 mkDupableCont env (CoerceIt ty cont)
1559 = mkDupableCont env cont `thenSmpl` \ (floats, (dup_cont, nondup_cont)) ->
1560 returnSmpl (floats, (CoerceIt ty dup_cont, nondup_cont))
1562 mkDupableCont env (InlinePlease cont)
1563 = mkDupableCont env cont `thenSmpl` \ (floats, (dup_cont, nondup_cont)) ->
1564 returnSmpl (floats, (InlinePlease dup_cont, nondup_cont))
1566 mkDupableCont env cont@(ArgOf _ arg_ty _ _)
1567 = returnSmpl (emptyFloats env, (mkBoringStop arg_ty, cont))
1568 -- Do *not* duplicate an ArgOf continuation
1569 -- Because ArgOf continuations are opaque, we gain nothing by
1570 -- propagating them into the expressions, and we do lose a lot.
1571 -- Here's an example:
1572 -- && (case x of { T -> F; F -> T }) E
1573 -- Now, && is strict so we end up simplifying the case with
1574 -- an ArgOf continuation. If we let-bind it, we get
1576 -- let $j = \v -> && v E
1577 -- in simplExpr (case x of { T -> F; F -> T })
1578 -- (ArgOf (\r -> $j r)
1579 -- And after simplifying more we get
1581 -- let $j = \v -> && v E
1582 -- in case of { T -> $j F; F -> $j T }
1583 -- Which is a Very Bad Thing
1585 -- The desire not to duplicate is the entire reason that
1586 -- mkDupableCont returns a pair of continuations.
1588 -- The original plan had:
1589 -- e.g. (...strict-fn...) [...hole...]
1591 -- let $j = \a -> ...strict-fn...
1592 -- in $j [...hole...]
1594 mkDupableCont env (ApplyTo _ arg se cont)
1595 = -- e.g. [...hole...] (...arg...)
1597 -- let a = ...arg...
1598 -- in [...hole...] a
1599 simplExpr (setInScope se env) arg `thenSmpl` \ arg' ->
1601 mkDupableCont env cont `thenSmpl` \ (floats, (dup_cont, nondup_cont)) ->
1602 addFloats env floats $ \ env ->
1604 if exprIsDupable arg' then
1605 returnSmpl (emptyFloats env, (ApplyTo OkToDup arg' (zapSubstEnv se) dup_cont, nondup_cont))
1607 newId FSLIT("a") (exprType arg') `thenSmpl` \ arg_id ->
1609 tick (CaseOfCase arg_id) `thenSmpl_`
1610 -- Want to tick here so that we go round again,
1611 -- and maybe copy or inline the code.
1612 -- Not strictly CaseOfCase, but never mind
1614 returnSmpl (unitFloat env arg_id arg',
1615 (ApplyTo OkToDup (Var arg_id) (zapSubstEnv se) dup_cont,
1617 -- But what if the arg should be case-bound?
1618 -- This has been this way for a long time, so I'll leave it,
1619 -- but I can't convince myself that it's right.
1622 mkDupableCont env (Select _ case_bndr alts se cont)
1623 = -- e.g. (case [...hole...] of { pi -> ei })
1625 -- let ji = \xij -> ei
1626 -- in case [...hole...] of { pi -> ji xij }
1627 tick (CaseOfCase case_bndr) `thenSmpl_`
1629 alt_env = setInScope se env
1631 prepareCaseCont alt_env alts cont `thenSmpl` \ (floats1, (dup_cont, nondup_cont)) ->
1632 addFloats alt_env floats1 $ \ alt_env ->
1634 simplBinder alt_env case_bndr `thenSmpl` \ (alt_env, case_bndr') ->
1635 -- NB: simplBinder does not zap deadness occ-info, so
1636 -- a dead case_bndr' will still advertise its deadness
1637 -- This is really important because in
1638 -- case e of b { (# a,b #) -> ... }
1639 -- b is always dead, and indeed we are not allowed to bind b to (# a,b #),
1640 -- which might happen if e was an explicit unboxed pair and b wasn't marked dead.
1641 -- In the new alts we build, we have the new case binder, so it must retain
1644 mkDupableAlts alt_env case_bndr' alts dup_cont `thenSmpl` \ (floats2, alts') ->
1645 addFloats alt_env floats2 $ \ alt_env ->
1646 returnSmpl (emptyFloats alt_env,
1647 (Select OkToDup case_bndr' alts' (zapSubstEnv se)
1648 (mkBoringStop (contResultType dup_cont)),
1651 mkDupableAlts :: SimplEnv -> OutId -> [InAlt] -> SimplCont
1652 -> SimplM (FloatsWith [InAlt])
1653 -- Absorbs the continuation into the new alternatives
1655 mkDupableAlts env case_bndr' alts dupable_cont
1658 go env [] = returnSmpl (emptyFloats env, [])
1660 = mkDupableAlt env case_bndr' dupable_cont alt `thenSmpl` \ (floats1, alt') ->
1661 addFloats env floats1 $ \ env ->
1662 go env alts `thenSmpl` \ (floats2, alts') ->
1663 returnSmpl (floats2, alt' : alts')
1665 mkDupableAlt env case_bndr' cont alt@(con, bndrs, rhs)
1666 = simplBinders env bndrs `thenSmpl` \ (env, bndrs') ->
1667 simplExprC env rhs cont `thenSmpl` \ rhs' ->
1669 if exprIsDupable rhs' then
1670 returnSmpl (emptyFloats env, (con, bndrs', rhs'))
1671 -- It is worth checking for a small RHS because otherwise we
1672 -- get extra let bindings that may cause an extra iteration of the simplifier to
1673 -- inline back in place. Quite often the rhs is just a variable or constructor.
1674 -- The Ord instance of Maybe in PrelMaybe.lhs, for example, took several extra
1675 -- iterations because the version with the let bindings looked big, and so wasn't
1676 -- inlined, but after the join points had been inlined it looked smaller, and so
1679 -- NB: we have to check the size of rhs', not rhs.
1680 -- Duplicating a small InAlt might invalidate occurrence information
1681 -- However, if it *is* dupable, we return the *un* simplified alternative,
1682 -- because otherwise we'd need to pair it up with an empty subst-env....
1683 -- but we only have one env shared between all the alts.
1684 -- (Remember we must zap the subst-env before re-simplifying something).
1685 -- Rather than do this we simply agree to re-simplify the original (small) thing later.
1689 rhs_ty' = exprType rhs'
1690 used_bndrs' = filter (not . isDeadBinder) (case_bndr' : bndrs')
1691 -- The deadness info on the new binders is unscathed
1693 -- If we try to lift a primitive-typed something out
1694 -- for let-binding-purposes, we will *caseify* it (!),
1695 -- with potentially-disastrous strictness results. So
1696 -- instead we turn it into a function: \v -> e
1697 -- where v::State# RealWorld#. The value passed to this function
1698 -- is realworld#, which generates (almost) no code.
1700 -- There's a slight infelicity here: we pass the overall
1701 -- case_bndr to all the join points if it's used in *any* RHS,
1702 -- because we don't know its usage in each RHS separately
1704 -- We used to say "&& isUnLiftedType rhs_ty'" here, but now
1705 -- we make the join point into a function whenever used_bndrs'
1706 -- is empty. This makes the join-point more CPR friendly.
1707 -- Consider: let j = if .. then I# 3 else I# 4
1708 -- in case .. of { A -> j; B -> j; C -> ... }
1710 -- Now CPR doesn't w/w j because it's a thunk, so
1711 -- that means that the enclosing function can't w/w either,
1712 -- which is a lose. Here's the example that happened in practice:
1713 -- kgmod :: Int -> Int -> Int
1714 -- kgmod x y = if x > 0 && y < 0 || x < 0 && y > 0
1718 -- I have seen a case alternative like this:
1719 -- True -> \v -> ...
1720 -- It's a bit silly to add the realWorld dummy arg in this case, making
1723 -- (the \v alone is enough to make CPR happy) but I think it's rare
1725 ( if null used_bndrs'
1726 then newId FSLIT("w") realWorldStatePrimTy `thenSmpl` \ rw_id ->
1727 returnSmpl ([rw_id], [Var realWorldPrimId])
1729 returnSmpl (used_bndrs', map varToCoreExpr used_bndrs')
1730 ) `thenSmpl` \ (final_bndrs', final_args) ->
1732 -- See comment about "$j" name above
1733 newId (encodeFS FSLIT("$j")) (mkPiTypes final_bndrs' rhs_ty') `thenSmpl` \ join_bndr ->
1734 -- Notice the funky mkPiTypes. If the contructor has existentials
1735 -- it's possible that the join point will be abstracted over
1736 -- type varaibles as well as term variables.
1737 -- Example: Suppose we have
1738 -- data T = forall t. C [t]
1740 -- case (case e of ...) of
1741 -- C t xs::[t] -> rhs
1742 -- We get the join point
1743 -- let j :: forall t. [t] -> ...
1744 -- j = /\t \xs::[t] -> rhs
1746 -- case (case e of ...) of
1747 -- C t xs::[t] -> j t xs
1749 -- We make the lambdas into one-shot-lambdas. The
1750 -- join point is sure to be applied at most once, and doing so
1751 -- prevents the body of the join point being floated out by
1752 -- the full laziness pass
1753 really_final_bndrs = map one_shot final_bndrs'
1754 one_shot v | isId v = setOneShotLambda v
1756 join_rhs = mkLams really_final_bndrs rhs'
1757 join_call = mkApps (Var join_bndr) final_args
1759 returnSmpl (unitFloat env join_bndr join_rhs, (con, bndrs', join_call))