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 mkRhsStop, mkBoringStop, pushContArgs,
19 contResultType, countArgs, contIsDupable, contIsRhsOrArg,
20 getContArgs, interestingCallContext, interestingArg, isStrictType
22 import Id ( Id, idType, idInfo, idArity, isDataConWorkId,
23 setIdUnfolding, isDeadBinder,
24 idNewDemandInfo, setIdInfo,
25 setIdOccInfo, zapLamIdInfo, setOneShotLambda,
27 import MkId ( eRROR_ID )
28 import Literal ( mkStringLit )
29 import OccName ( encodeFS )
30 import IdInfo ( OccInfo(..), isLoopBreaker,
31 setArityInfo, zapDemandInfo,
35 import NewDemand ( isStrictDmd )
36 import Unify ( coreRefineTys )
37 import DataCon ( dataConTyCon, dataConRepStrictness, isVanillaDataCon, dataConResTy )
38 import TyCon ( tyConArity )
40 import PprCore ( pprParendExpr, pprCoreExpr )
41 import CoreUnfold ( mkOtherCon, mkUnfolding, callSiteInline )
42 import CoreUtils ( exprIsDupable, exprIsTrivial, needsCaseBinding,
43 exprIsConApp_maybe, mkPiTypes, findAlt,
44 exprType, exprIsValue,
45 exprOkForSpeculation, exprArity,
46 mkCoerce, mkCoerce2, mkSCC, mkInlineMe, applyTypeToArg
48 import Rules ( lookupRule )
49 import BasicTypes ( isMarkedStrict )
50 import CostCentre ( currentCCS )
51 import Type ( TvSubstEnv, isUnLiftedType, seqType, tyConAppArgs, funArgTy,
52 splitFunTy_maybe, splitFunTy, eqType, substTy,
53 mkTyVarTys, mkTyConApp
55 import VarEnv ( elemVarEnv )
56 import Subst ( SubstResult(..), emptySubst, substExpr,
57 substId, simplIdInfo )
58 import TysPrim ( realWorldStatePrimTy )
59 import PrelInfo ( realWorldPrimId )
60 import BasicTypes ( TopLevelFlag(..), isTopLevel,
64 import Maybe ( Maybe )
65 import Maybes ( orElse )
67 import Util ( notNull, equalLength )
71 The guts of the simplifier is in this module, but the driver loop for
72 the simplifier is in SimplCore.lhs.
75 -----------------------------------------
76 *** IMPORTANT NOTE ***
77 -----------------------------------------
78 The simplifier used to guarantee that the output had no shadowing, but
79 it does not do so any more. (Actually, it never did!) The reason is
80 documented with simplifyArgs.
83 -----------------------------------------
84 *** IMPORTANT NOTE ***
85 -----------------------------------------
86 Many parts of the simplifier return a bunch of "floats" as well as an
87 expression. This is wrapped as a datatype SimplUtils.FloatsWith.
89 All "floats" are let-binds, not case-binds, but some non-rec lets may
90 be unlifted (with RHS ok-for-speculation).
94 -----------------------------------------
95 ORGANISATION OF FUNCTIONS
96 -----------------------------------------
98 - simplify all top-level binders
99 - for NonRec, call simplRecOrTopPair
100 - for Rec, call simplRecBind
103 ------------------------------
104 simplExpr (applied lambda) ==> simplNonRecBind
105 simplExpr (Let (NonRec ...) ..) ==> simplNonRecBind
106 simplExpr (Let (Rec ...) ..) ==> simplify binders; simplRecBind
108 ------------------------------
109 simplRecBind [binders already simplfied]
110 - use simplRecOrTopPair on each pair in turn
112 simplRecOrTopPair [binder already simplified]
113 Used for: recursive bindings (top level and nested)
114 top-level non-recursive bindings
116 - check for PreInlineUnconditionally
120 Used for: non-top-level non-recursive bindings
121 beta reductions (which amount to the same thing)
122 Because it can deal with strict arts, it takes a
123 "thing-inside" and returns an expression
125 - check for PreInlineUnconditionally
126 - simplify binder, including its IdInfo
135 simplNonRecX: [given a *simplified* RHS, but an *unsimplified* binder]
136 Used for: binding case-binder and constr args in a known-constructor case
137 - check for PreInLineUnconditionally
141 ------------------------------
142 simplLazyBind: [binder already simplified, RHS not]
143 Used for: recursive bindings (top level and nested)
144 top-level non-recursive bindings
145 non-top-level, but *lazy* non-recursive bindings
146 [must not be strict or unboxed]
147 Returns floats + an augmented environment, not an expression
148 - substituteIdInfo and add result to in-scope
149 [so that rules are available in rec rhs]
152 - float if exposes constructor or PAP
156 completeNonRecX: [binder and rhs both simplified]
157 - if the the thing needs case binding (unlifted and not ok-for-spec)
163 completeLazyBind: [given a simplified RHS]
164 [used for both rec and non-rec bindings, top level and not]
165 - try PostInlineUnconditionally
166 - add unfolding [this is the only place we add an unfolding]
171 Right hand sides and arguments
172 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
173 In many ways we want to treat
174 (a) the right hand side of a let(rec), and
175 (b) a function argument
176 in the same way. But not always! In particular, we would
177 like to leave these arguments exactly as they are, so they
178 will match a RULE more easily.
183 It's harder to make the rule match if we ANF-ise the constructor,
184 or eta-expand the PAP:
186 f (let { a = g x; b = h x } in (a,b))
189 On the other hand if we see the let-defns
194 then we *do* want to ANF-ise and eta-expand, so that p and q
195 can be safely inlined.
197 Even floating lets out is a bit dubious. For let RHS's we float lets
198 out if that exposes a value, so that the value can be inlined more vigorously.
201 r = let x = e in (x,x)
203 Here, if we float the let out we'll expose a nice constructor. We did experiments
204 that showed this to be a generally good thing. But it was a bad thing to float
205 lets out unconditionally, because that meant they got allocated more often.
207 For function arguments, there's less reason to expose a constructor (it won't
208 get inlined). Just possibly it might make a rule match, but I'm pretty skeptical.
209 So for the moment we don't float lets out of function arguments either.
214 For eta expansion, we want to catch things like
216 case e of (a,b) -> \x -> case a of (p,q) -> \y -> r
218 If the \x was on the RHS of a let, we'd eta expand to bring the two
219 lambdas together. And in general that's a good thing to do. Perhaps
220 we should eta expand wherever we find a (value) lambda? Then the eta
221 expansion at a let RHS can concentrate solely on the PAP case.
224 %************************************************************************
226 \subsection{Bindings}
228 %************************************************************************
231 simplTopBinds :: SimplEnv -> [InBind] -> SimplM [OutBind]
233 simplTopBinds env binds
234 = -- Put all the top-level binders into scope at the start
235 -- so that if a transformation rule has unexpectedly brought
236 -- anything into scope, then we don't get a complaint about that.
237 -- It's rather as if the top-level binders were imported.
238 simplRecBndrs env (bindersOfBinds binds) `thenSmpl` \ (env, bndrs') ->
239 simpl_binds env binds bndrs' `thenSmpl` \ (floats, _) ->
240 freeTick SimplifierDone `thenSmpl_`
241 returnSmpl (floatBinds floats)
243 -- We need to track the zapped top-level binders, because
244 -- they should have their fragile IdInfo zapped (notably occurrence info)
245 -- That's why we run down binds and bndrs' simultaneously.
246 simpl_binds :: SimplEnv -> [InBind] -> [OutId] -> SimplM (FloatsWith ())
247 simpl_binds env [] bs = ASSERT( null bs ) returnSmpl (emptyFloats env, ())
248 simpl_binds env (bind:binds) bs = simpl_bind env bind bs `thenSmpl` \ (floats,env) ->
249 addFloats env floats $ \env ->
250 simpl_binds env binds (drop_bs bind bs)
252 drop_bs (NonRec _ _) (_ : bs) = bs
253 drop_bs (Rec prs) bs = drop (length prs) bs
255 simpl_bind env bind bs
256 = getDOptsSmpl `thenSmpl` \ dflags ->
257 if dopt Opt_D_dump_inlinings dflags then
258 pprTrace "SimplBind" (ppr (bindersOf bind)) $ simpl_bind1 env bind bs
260 simpl_bind1 env bind bs
262 simpl_bind1 env (NonRec b r) (b':_) = simplRecOrTopPair env TopLevel b b' r
263 simpl_bind1 env (Rec pairs) bs' = simplRecBind env TopLevel pairs bs'
267 %************************************************************************
269 \subsection{simplNonRec}
271 %************************************************************************
273 simplNonRecBind is used for
274 * non-top-level non-recursive lets in expressions
278 * An unsimplified (binder, rhs) pair
279 * The env for the RHS. It may not be the same as the
280 current env because the bind might occur via (\x.E) arg
282 It uses the CPS form because the binding might be strict, in which
283 case we might discard the continuation:
284 let x* = error "foo" in (...x...)
286 It needs to turn unlifted bindings into a @case@. They can arise
287 from, say: (\x -> e) (4# + 3#)
290 simplNonRecBind :: SimplEnv
292 -> InExpr -> SimplEnv -- Arg, with its subst-env
293 -> OutType -- Type of thing computed by the context
294 -> (SimplEnv -> SimplM FloatsWithExpr) -- The body
295 -> SimplM FloatsWithExpr
297 simplNonRecBind env bndr rhs rhs_se cont_ty thing_inside
299 = pprPanic "simplNonRecBind" (ppr bndr <+> ppr rhs)
302 simplNonRecBind env bndr rhs rhs_se cont_ty thing_inside
303 | preInlineUnconditionally env NotTopLevel bndr
304 = tick (PreInlineUnconditionally bndr) `thenSmpl_`
305 thing_inside (extendIdSubst env bndr (ContEx (getSubst rhs_se) rhs))
308 | isStrictDmd (idNewDemandInfo bndr) || isStrictType (idType bndr) -- A strict let
309 = -- Don't use simplBinder because that doesn't keep
310 -- fragile occurrence info in the substitution
311 simplLetBndr 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 -- simplLetBndr doesn't deal with the IdInfo, so we must
317 -- do so here (c.f. simplLazyBind)
318 bndr2 = bndr1 `setIdInfo` simplIdInfo (getSubst env) (idInfo bndr)
319 env2 = modifyInScope env1 bndr2 bndr2
321 completeNonRecX env2 True {- strict -} bndr bndr2 rhs1 thing_inside
323 | otherwise -- Normal, lazy case
324 = -- Don't use simplBinder because that doesn't keep
325 -- fragile occurrence info in the substitution
326 simplLetBndr env bndr `thenSmpl` \ (env, bndr') ->
327 simplLazyBind env NotTopLevel NonRecursive
328 bndr bndr' rhs rhs_se `thenSmpl` \ (floats, env) ->
329 addFloats env floats thing_inside
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 | needsCaseBinding (idType bndr) new_rhs
344 -- Make this test *before* the preInlineUnconditionally
345 -- Consider case I# (quotInt# x y) of
346 -- I# v -> let w = J# v in ...
347 -- If we gaily inline (quotInt# x y) for v, we end up building an
349 -- let w = J# (quotInt# x y) in ...
350 -- because quotInt# can fail.
351 = simplBinder env bndr `thenSmpl` \ (env, bndr') ->
352 thing_inside env `thenSmpl` \ (floats, body) ->
354 let body' = wrapFloats floats body in
355 returnSmpl (emptyFloats env, Case new_rhs bndr' (exprType body') [(DEFAULT, [], body')])
357 | preInlineUnconditionally env NotTopLevel bndr
358 -- This happens; for example, the case_bndr during case of
359 -- known constructor: case (a,b) of x { (p,q) -> ... }
360 -- Here x isn't mentioned in the RHS, so we don't want to
361 -- create the (dead) let-binding let x = (a,b) in ...
363 -- Similarly, single occurrences can be inlined vigourously
364 -- e.g. case (f x, g y) of (a,b) -> ....
365 -- If a,b occur once we can avoid constructing the let binding for them.
366 = thing_inside (extendIdSubst env bndr (ContEx emptySubst new_rhs))
369 = simplBinder env bndr `thenSmpl` \ (env, bndr') ->
370 completeNonRecX env False {- Non-strict; pessimistic -}
371 bndr bndr' new_rhs thing_inside
373 completeNonRecX env is_strict old_bndr new_bndr new_rhs thing_inside
374 = mkAtomicArgs is_strict
375 True {- OK to float unlifted -}
376 new_rhs `thenSmpl` \ (aux_binds, rhs2) ->
378 -- Make the arguments atomic if necessary,
379 -- adding suitable bindings
380 addAtomicBindsE env (fromOL aux_binds) $ \ env ->
381 completeLazyBind env NotTopLevel
382 old_bndr new_bndr rhs2 `thenSmpl` \ (floats, env) ->
383 addFloats env floats thing_inside
387 %************************************************************************
389 \subsection{Lazy bindings}
391 %************************************************************************
393 simplRecBind is used for
394 * recursive bindings only
397 simplRecBind :: SimplEnv -> TopLevelFlag
398 -> [(InId, InExpr)] -> [OutId]
399 -> SimplM (FloatsWith SimplEnv)
400 simplRecBind env top_lvl pairs bndrs'
401 = go env pairs bndrs' `thenSmpl` \ (floats, env) ->
402 returnSmpl (flattenFloats floats, env)
404 go env [] _ = returnSmpl (emptyFloats env, env)
406 go env ((bndr, rhs) : pairs) (bndr' : bndrs')
407 = simplRecOrTopPair env top_lvl bndr bndr' rhs `thenSmpl` \ (floats, env) ->
408 addFloats env floats (\env -> go env pairs bndrs')
412 simplRecOrTopPair is used for
413 * recursive bindings (whether top level or not)
414 * top-level non-recursive bindings
416 It assumes the binder has already been simplified, but not its IdInfo.
419 simplRecOrTopPair :: SimplEnv
421 -> InId -> OutId -- Binder, both pre-and post simpl
422 -> InExpr -- The RHS and its environment
423 -> SimplM (FloatsWith SimplEnv)
425 simplRecOrTopPair env top_lvl bndr bndr' rhs
426 | preInlineUnconditionally env top_lvl bndr -- Check for unconditional inline
427 = tick (PreInlineUnconditionally bndr) `thenSmpl_`
428 returnSmpl (emptyFloats env, extendIdSubst env bndr (ContEx (getSubst env) rhs))
431 = simplLazyBind env top_lvl Recursive bndr bndr' rhs env
432 -- May not actually be recursive, but it doesn't matter
436 simplLazyBind is used for
437 * recursive bindings (whether top level or not)
438 * top-level non-recursive bindings
439 * non-top-level *lazy* non-recursive bindings
441 [Thus it deals with the lazy cases from simplNonRecBind, and all cases
442 from SimplRecOrTopBind]
445 1. It assumes that the binder is *already* simplified,
446 and is in scope, but not its IdInfo
448 2. It assumes that the binder type is lifted.
450 3. It does not check for pre-inline-unconditionallly;
451 that should have been done already.
454 simplLazyBind :: SimplEnv
455 -> TopLevelFlag -> RecFlag
456 -> InId -> OutId -- Binder, both pre-and post simpl
457 -> InExpr -> SimplEnv -- The RHS and its environment
458 -> SimplM (FloatsWith SimplEnv)
460 simplLazyBind env top_lvl is_rec bndr bndr1 rhs rhs_se
461 = let -- Transfer the IdInfo of the original binder to the new binder
462 -- This is crucial: we must preserve
466 -- etc. To do this we must apply the current substitution,
467 -- which incorporates earlier substitutions in this very letrec group.
469 -- NB 1. We do this *before* processing the RHS of the binder, so that
470 -- its substituted rules are visible in its own RHS.
471 -- This is important. Manuel found cases where he really, really
472 -- wanted a RULE for a recursive function to apply in that function's
473 -- own right-hand side.
475 -- NB 2: We do not transfer the arity (see Subst.substIdInfo)
476 -- The arity of an Id should not be visible
477 -- in its own RHS, else we eta-reduce
481 -- which isn't sound. And it makes the arity in f's IdInfo greater than
482 -- the manifest arity, which isn't good.
483 -- The arity will get added later.
485 -- NB 3: It's important that we *do* transer the loop-breaker OccInfo,
486 -- because that's what stops the Id getting inlined infinitely, in the body
489 -- NB 4: does no harm for non-recursive bindings
491 bndr2 = bndr1 `setIdInfo` simplIdInfo (getSubst env) (idInfo bndr)
492 env1 = modifyInScope env bndr2 bndr2
493 rhs_env = setInScope rhs_se env1
494 is_top_level = isTopLevel top_lvl
495 ok_float_unlifted = not is_top_level && isNonRec is_rec
496 rhs_cont = mkRhsStop (idType bndr1)
498 -- Simplify the RHS; note the mkRhsStop, which tells
499 -- the simplifier that this is the RHS of a let.
500 simplExprF rhs_env rhs rhs_cont `thenSmpl` \ (floats, rhs1) ->
502 -- If any of the floats can't be floated, give up now
503 -- (The allLifted predicate says True for empty floats.)
504 if (not ok_float_unlifted && not (allLifted floats)) then
505 completeLazyBind env1 top_lvl bndr bndr2
506 (wrapFloats floats rhs1)
509 -- ANF-ise a constructor or PAP rhs
510 mkAtomicArgs False {- Not strict -}
511 ok_float_unlifted rhs1 `thenSmpl` \ (aux_binds, rhs2) ->
513 -- If the result is a PAP, float the floats out, else wrap them
514 -- By this time it's already been ANF-ised (if necessary)
515 if isEmptyFloats floats && isNilOL aux_binds then -- Shortcut a common case
516 completeLazyBind env1 top_lvl bndr bndr2 rhs2
518 else if is_top_level || exprIsTrivial rhs2 || exprIsValue rhs2 then
519 -- WARNING: long dodgy argument coming up
520 -- WANTED: a better way to do this
522 -- We can't use "exprIsCheap" instead of exprIsValue,
523 -- because that causes a strictness bug.
524 -- x = let y* = E in case (scc y) of { T -> F; F -> T}
525 -- The case expression is 'cheap', but it's wrong to transform to
526 -- y* = E; x = case (scc y) of {...}
527 -- Either we must be careful not to float demanded non-values, or
528 -- we must use exprIsValue for the test, which ensures that the
529 -- thing is non-strict. So exprIsValue => bindings are non-strict
530 -- I think. The WARN below tests for this.
532 -- We use exprIsTrivial here because we want to reveal lone variables.
533 -- E.g. let { x = letrec { y = E } in y } in ...
534 -- Here we definitely want to float the y=E defn.
535 -- exprIsValue definitely isn't right for that.
537 -- Again, the floated binding can't be strict; if it's recursive it'll
538 -- be non-strict; if it's non-recursive it'd be inlined.
540 -- Note [SCC-and-exprIsTrivial]
542 -- y = let { x* = E } in scc "foo" x
543 -- then we do *not* want to float out the x binding, because
544 -- it's strict! Fortunately, exprIsTrivial replies False to
547 -- There's a subtlety here. There may be a binding (x* = e) in the
548 -- floats, where the '*' means 'will be demanded'. So is it safe
549 -- to float it out? Answer no, but it won't matter because
550 -- we only float if (a) arg' is a WHNF, or (b) it's going to top level
551 -- and so there can't be any 'will be demanded' bindings in the floats.
553 ASSERT2( is_top_level || not (any demanded_float (floatBinds floats)),
554 ppr (filter demanded_float (floatBinds floats)) )
556 tick LetFloatFromLet `thenSmpl_` (
557 addFloats env1 floats $ \ env2 ->
558 addAtomicBinds env2 (fromOL aux_binds) $ \ env3 ->
559 completeLazyBind env3 top_lvl bndr bndr2 rhs2)
562 completeLazyBind env1 top_lvl bndr bndr2 (wrapFloats floats rhs1)
565 demanded_float (NonRec b r) = isStrictDmd (idNewDemandInfo b) && not (isUnLiftedType (idType b))
566 -- Unlifted-type (cheap-eagerness) lets may well have a demanded flag on them
567 demanded_float (Rec _) = False
572 %************************************************************************
574 \subsection{Completing a lazy binding}
576 %************************************************************************
579 * deals only with Ids, not TyVars
580 * takes an already-simplified binder and RHS
581 * is used for both recursive and non-recursive bindings
582 * is used for both top-level and non-top-level bindings
584 It does the following:
585 - tries discarding a dead binding
586 - tries PostInlineUnconditionally
587 - add unfolding [this is the only place we add an unfolding]
590 It does *not* attempt to do let-to-case. Why? Because it is used for
591 - top-level bindings (when let-to-case is impossible)
592 - many situations where the "rhs" is known to be a WHNF
593 (so let-to-case is inappropriate).
596 completeLazyBind :: SimplEnv
597 -> TopLevelFlag -- Flag stuck into unfolding
598 -> InId -- Old binder
599 -> OutId -- New binder
600 -> OutExpr -- Simplified RHS
601 -> SimplM (FloatsWith SimplEnv)
602 -- We return a new SimplEnv, because completeLazyBind may choose to do its work
603 -- by extending the substitution (e.g. let x = y in ...)
604 -- The new binding (if any) is returned as part of the floats.
605 -- NB: the returned SimplEnv has the right SubstEnv, but you should
606 -- (as usual) use the in-scope-env from the floats
608 completeLazyBind env top_lvl old_bndr new_bndr new_rhs
609 | postInlineUnconditionally env new_bndr occ_info new_rhs
610 = -- Drop the binding
611 tick (PostInlineUnconditionally old_bndr) `thenSmpl_`
612 returnSmpl (emptyFloats env, extendIdSubst env old_bndr (DoneEx new_rhs))
613 -- Use the substitution to make quite, quite sure that the substitution
614 -- will happen, since we are going to discard the binding
619 new_bndr_info = idInfo new_bndr `setArityInfo` exprArity new_rhs
621 -- Add the unfolding *only* for non-loop-breakers
622 -- Making loop breakers not have an unfolding at all
623 -- means that we can avoid tests in exprIsConApp, for example.
624 -- This is important: if exprIsConApp says 'yes' for a recursive
625 -- thing, then we can get into an infinite loop
627 -- If the unfolding is a value, the demand info may
628 -- go pear-shaped, so we nuke it. Example:
630 -- case x of (p,q) -> h p q x
631 -- Here x is certainly demanded. But after we've nuked
632 -- the case, we'll get just
633 -- let x = (a,b) in h a b x
634 -- and now x is not demanded (I'm assuming h is lazy)
635 -- This really happens. Similarly
636 -- let f = \x -> e in ...f..f...
637 -- After inling f at some of its call sites the original binding may
638 -- (for example) be no longer strictly demanded.
639 -- The solution here is a bit ad hoc...
640 unfolding = mkUnfolding (isTopLevel top_lvl) new_rhs
641 info_w_unf = new_bndr_info `setUnfoldingInfo` unfolding
642 final_info | loop_breaker = new_bndr_info
643 | isEvaldUnfolding unfolding = zapDemandInfo info_w_unf `orElse` info_w_unf
644 | otherwise = info_w_unf
646 final_id = new_bndr `setIdInfo` final_info
648 -- These seqs forces the Id, and hence its IdInfo,
649 -- and hence any inner substitutions
651 returnSmpl (unitFloat env final_id new_rhs, env)
654 loop_breaker = isLoopBreaker occ_info
655 old_info = idInfo old_bndr
656 occ_info = occInfo old_info
661 %************************************************************************
663 \subsection[Simplify-simplExpr]{The main function: simplExpr}
665 %************************************************************************
667 The reason for this OutExprStuff stuff is that we want to float *after*
668 simplifying a RHS, not before. If we do so naively we get quadratic
669 behaviour as things float out.
671 To see why it's important to do it after, consider this (real) example:
685 a -- Can't inline a this round, cos it appears twice
689 Each of the ==> steps is a round of simplification. We'd save a
690 whole round if we float first. This can cascade. Consider
695 let f = let d1 = ..d.. in \y -> e
699 in \x -> ...(\y ->e)...
701 Only in this second round can the \y be applied, and it
702 might do the same again.
706 simplExpr :: SimplEnv -> CoreExpr -> SimplM CoreExpr
707 simplExpr env expr = simplExprC env expr (mkBoringStop expr_ty')
709 expr_ty' = substTy (getTvSubst env) (exprType expr)
710 -- The type in the Stop continuation, expr_ty', is usually not used
711 -- It's only needed when discarding continuations after finding
712 -- a function that returns bottom.
713 -- Hence the lazy substitution
716 simplExprC :: SimplEnv -> CoreExpr -> SimplCont -> SimplM CoreExpr
717 -- Simplify an expression, given a continuation
718 simplExprC env expr cont
719 = simplExprF env expr cont `thenSmpl` \ (floats, expr) ->
720 returnSmpl (wrapFloats floats expr)
722 simplExprF :: SimplEnv -> InExpr -> SimplCont -> SimplM FloatsWithExpr
723 -- Simplify an expression, returning floated binds
725 simplExprF env (Var v) cont = simplVar env v cont
726 simplExprF env (Lit lit) cont = rebuild env (Lit lit) cont
727 simplExprF env expr@(Lam _ _) cont = simplLam env expr cont
728 simplExprF env (Note note expr) cont = simplNote env note expr cont
729 simplExprF env (App fun arg) cont = simplExprF env fun (ApplyTo NoDup arg env cont)
731 simplExprF env (Type ty) cont
732 = ASSERT( contIsRhsOrArg cont )
733 simplType env ty `thenSmpl` \ ty' ->
734 rebuild env (Type ty') cont
737 simplExprF env (Case scrut bndr case_ty alts) cont
738 | not (switchIsOn (getSwitchChecker env) NoCaseOfCase)
739 = -- Simplify the scrutinee with a Select continuation
740 simplExprF env scrut (Select NoDup bndr alts env cont)
743 = -- If case-of-case is off, simply simplify the case expression
744 -- in a vanilla Stop context, and rebuild the result around it
745 simplExprC env scrut case_cont `thenSmpl` \ case_expr' ->
746 rebuild env case_expr' cont
748 case_cont = Select NoDup bndr alts env (mkBoringStop case_ty')
749 case_ty' = substTy (getTvSubst env) case_ty -- c.f. defn of simplExpr
751 simplExprF env (Let (Rec pairs) body) cont
752 = simplRecBndrs env (map fst pairs) `thenSmpl` \ (env, bndrs') ->
753 -- NB: bndrs' don't have unfoldings or rules
754 -- We add them as we go down
756 simplRecBind env NotTopLevel pairs bndrs' `thenSmpl` \ (floats, env) ->
757 addFloats env floats $ \ env ->
758 simplExprF env body cont
760 -- A non-recursive let is dealt with by simplNonRecBind
761 simplExprF env (Let (NonRec bndr rhs) body) cont
762 = simplNonRecBind env bndr rhs env (contResultType cont) $ \ env ->
763 simplExprF env body cont
766 ---------------------------------
767 simplType :: SimplEnv -> InType -> SimplM OutType
768 -- Kept monadic just so we can do the seqType
770 = seqType new_ty `seq` returnSmpl new_ty
772 new_ty = substTy (getTvSubst env) ty
776 %************************************************************************
780 %************************************************************************
783 simplLam env fun cont
786 zap_it = mkLamBndrZapper fun (countArgs cont)
787 cont_ty = contResultType cont
789 -- Type-beta reduction
790 go env (Lam bndr body) (ApplyTo _ (Type ty_arg) arg_se body_cont)
791 = ASSERT( isTyVar bndr )
792 tick (BetaReduction bndr) `thenSmpl_`
793 simplType (setInScope arg_se env) ty_arg `thenSmpl` \ ty_arg' ->
794 go (extendTvSubst env bndr ty_arg') body body_cont
796 -- Ordinary beta reduction
797 go env (Lam bndr body) cont@(ApplyTo _ arg arg_se body_cont)
798 = tick (BetaReduction bndr) `thenSmpl_`
799 simplNonRecBind env (zap_it bndr) arg arg_se cont_ty $ \ env ->
800 go env body body_cont
802 -- Not enough args, so there are real lambdas left to put in the result
803 go env lam@(Lam _ _) cont
804 = simplLamBndrs env bndrs `thenSmpl` \ (env, bndrs') ->
805 simplExpr env body `thenSmpl` \ body' ->
806 mkLam env bndrs' body' cont `thenSmpl` \ (floats, new_lam) ->
807 addFloats env floats $ \ env ->
808 rebuild env new_lam cont
810 (bndrs,body) = collectBinders lam
812 -- Exactly enough args
813 go env expr cont = simplExprF env expr cont
815 mkLamBndrZapper :: CoreExpr -- Function
816 -> Int -- Number of args supplied, *including* type args
817 -> Id -> Id -- Use this to zap the binders
818 mkLamBndrZapper fun n_args
819 | n_args >= n_params fun = \b -> b -- Enough args
820 | otherwise = \b -> zapLamIdInfo b
822 -- NB: we count all the args incl type args
823 -- so we must count all the binders (incl type lambdas)
824 n_params (Note _ e) = n_params e
825 n_params (Lam b e) = 1 + n_params e
826 n_params other = 0::Int
830 %************************************************************************
834 %************************************************************************
837 simplNote env (Coerce to from) body cont
839 addCoerce s1 k1 (CoerceIt t1 cont)
840 -- coerce T1 S1 (coerce S1 K1 e)
843 -- coerce T1 K1 e, otherwise
845 -- For example, in the initial form of a worker
846 -- we may find (coerce T (coerce S (\x.e))) y
847 -- and we'd like it to simplify to e[y/x] in one round
849 | t1 `eqType` k1 = cont -- The coerces cancel out
850 | otherwise = CoerceIt t1 cont -- They don't cancel, but
851 -- the inner one is redundant
853 addCoerce t1t2 s1s2 (ApplyTo dup arg arg_se cont)
854 | not (isTypeArg arg), -- This whole case only works for value args
855 -- Could upgrade to have equiv thing for type apps too
856 Just (s1, s2) <- splitFunTy_maybe s1s2
857 -- (coerce (T1->T2) (S1->S2) F) E
859 -- coerce T2 S2 (F (coerce S1 T1 E))
861 -- t1t2 must be a function type, T1->T2, because it's applied to something
862 -- but s1s2 might conceivably not be
864 -- When we build the ApplyTo we can't mix the out-types
865 -- with the InExpr in the argument, so we simply substitute
866 -- to make it all consistent. It's a bit messy.
867 -- But it isn't a common case.
869 (t1,t2) = splitFunTy t1t2
870 new_arg = mkCoerce2 s1 t1 (substExpr subst arg)
871 subst = getSubst (setInScope arg_se env)
873 ApplyTo dup new_arg (zapSubstEnv env) (addCoerce t2 s2 cont)
875 addCoerce to' _ cont = CoerceIt to' cont
877 simplType env to `thenSmpl` \ to' ->
878 simplType env from `thenSmpl` \ from' ->
879 simplExprF env body (addCoerce to' from' cont)
882 -- Hack: we only distinguish subsumed cost centre stacks for the purposes of
883 -- inlining. All other CCCSs are mapped to currentCCS.
884 simplNote env (SCC cc) e cont
885 = simplExpr (setEnclosingCC env currentCCS) e `thenSmpl` \ e' ->
886 rebuild env (mkSCC cc e') cont
888 simplNote env InlineCall e cont
889 = simplExprF env e (InlinePlease cont)
891 -- See notes with SimplMonad.inlineMode
892 simplNote env InlineMe e cont
893 | contIsRhsOrArg cont -- Totally boring continuation; see notes above
894 = -- Don't inline inside an INLINE expression
895 simplExpr (setMode inlineMode env ) e `thenSmpl` \ e' ->
896 rebuild env (mkInlineMe e') cont
898 | otherwise -- Dissolve the InlineMe note if there's
899 -- an interesting context of any kind to combine with
900 -- (even a type application -- anything except Stop)
901 = simplExprF env e cont
903 simplNote env (CoreNote s) e cont
904 = simplExpr env e `thenSmpl` \ e' ->
905 rebuild env (Note (CoreNote s) e') cont
909 %************************************************************************
911 \subsection{Dealing with calls}
913 %************************************************************************
916 simplVar env var cont
917 = case substId (getSubst env) var of
918 DoneEx e -> simplExprF (zapSubstEnv env) e cont
919 ContEx se e -> simplExprF (setSubstEnv env se) e cont
920 DoneId var1 occ -> completeCall (zapSubstEnv env) var1 occ cont
921 -- Note [zapSubstEnv]
922 -- The template is already simplified, so don't re-substitute.
923 -- This is VITAL. Consider
925 -- let y = \z -> ...x... in
927 -- We'll clone the inner \x, adding x->x' in the id_subst
928 -- Then when we inline y, we must *not* replace x by x' in
929 -- the inlined copy!!
931 ---------------------------------------------------------
932 -- Dealing with a call site
934 completeCall env var occ_info cont
935 = -- Simplify the arguments
936 getDOptsSmpl `thenSmpl` \ dflags ->
938 chkr = getSwitchChecker env
939 (args, call_cont, inline_call) = getContArgs chkr var cont
942 simplifyArgs env fn_ty args (contResultType call_cont) $ \ env args ->
944 -- Next, look for rules or specialisations that match
946 -- It's important to simplify the args first, because the rule-matcher
947 -- doesn't do substitution as it goes. We don't want to use subst_args
948 -- (defined in the 'where') because that throws away useful occurrence info,
949 -- and perhaps-very-important specialisations.
951 -- Some functions have specialisations *and* are strict; in this case,
952 -- we don't want to inline the wrapper of the non-specialised thing; better
953 -- to call the specialised thing instead.
954 -- We used to use the black-listing mechanism to ensure that inlining of
955 -- the wrapper didn't occur for things that have specialisations till a
956 -- later phase, so but now we just try RULES first
958 -- You might think that we shouldn't apply rules for a loop breaker:
959 -- doing so might give rise to an infinite loop, because a RULE is
960 -- rather like an extra equation for the function:
961 -- RULE: f (g x) y = x+y
964 -- But it's too drastic to disable rules for loop breakers.
965 -- Even the foldr/build rule would be disabled, because foldr
966 -- is recursive, and hence a loop breaker:
967 -- foldr k z (build g) = g k z
968 -- So it's up to the programmer: rules can cause divergence
971 in_scope = getInScope env
972 maybe_rule = case activeRule env of
973 Nothing -> Nothing -- No rules apply
974 Just act_fn -> lookupRule act_fn in_scope var args
977 Just (rule_name, rule_rhs) ->
978 tick (RuleFired rule_name) `thenSmpl_`
979 (if dopt Opt_D_dump_inlinings dflags then
980 pprTrace "Rule fired" (vcat [
981 text "Rule:" <+> ftext rule_name,
982 text "Before:" <+> ppr var <+> sep (map pprParendExpr args),
983 text "After: " <+> pprCoreExpr rule_rhs,
984 text "Cont: " <+> ppr call_cont])
987 simplExprF env rule_rhs call_cont ;
989 Nothing -> -- No rules
991 -- Next, look for an inlining
993 arg_infos = [ interestingArg arg | arg <- args, isValArg arg]
995 interesting_cont = interestingCallContext (notNull args)
999 active_inline = activeInline env var occ_info
1000 maybe_inline = callSiteInline dflags active_inline inline_call occ_info
1001 var arg_infos interesting_cont
1003 case maybe_inline of {
1004 Just unfolding -- There is an inlining!
1005 -> tick (UnfoldingDone var) `thenSmpl_`
1006 makeThatCall env var unfolding args call_cont
1009 Nothing -> -- No inlining!
1012 rebuild env (mkApps (Var var) args) call_cont
1015 makeThatCall :: SimplEnv
1017 -> InExpr -- Inlined function rhs
1018 -> [OutExpr] -- Arguments, already simplified
1019 -> SimplCont -- After the call
1020 -> SimplM FloatsWithExpr
1021 -- Similar to simplLam, but this time
1022 -- the arguments are already simplified
1023 makeThatCall orig_env var fun@(Lam _ _) args cont
1024 = go orig_env fun args
1026 zap_it = mkLamBndrZapper fun (length args)
1028 -- Type-beta reduction
1029 go env (Lam bndr body) (Type ty_arg : args)
1030 = ASSERT( isTyVar bndr )
1031 tick (BetaReduction bndr) `thenSmpl_`
1032 go (extendTvSubst env bndr ty_arg) body args
1034 -- Ordinary beta reduction
1035 go env (Lam bndr body) (arg : args)
1036 = tick (BetaReduction bndr) `thenSmpl_`
1037 simplNonRecX env (zap_it bndr) arg $ \ env ->
1040 -- Not enough args, so there are real lambdas left to put in the result
1042 = simplExprF env fun (pushContArgs orig_env args cont)
1043 -- NB: orig_env; the correct environment to capture with
1044 -- the arguments.... env has been augmented with substitutions
1045 -- from the beta reductions.
1047 makeThatCall env var fun args cont
1048 = simplExprF env fun (pushContArgs env args cont)
1052 %************************************************************************
1054 \subsection{Arguments}
1056 %************************************************************************
1059 ---------------------------------------------------------
1060 -- Simplifying the arguments of a call
1062 simplifyArgs :: SimplEnv
1063 -> OutType -- Type of the function
1064 -> [(InExpr, SimplEnv, Bool)] -- Details of the arguments
1065 -> OutType -- Type of the continuation
1066 -> (SimplEnv -> [OutExpr] -> SimplM FloatsWithExpr)
1067 -> SimplM FloatsWithExpr
1069 -- [CPS-like because of strict arguments]
1071 -- Simplify the arguments to a call.
1072 -- This part of the simplifier may break the no-shadowing invariant
1074 -- f (...(\a -> e)...) (case y of (a,b) -> e')
1075 -- where f is strict in its second arg
1076 -- If we simplify the innermost one first we get (...(\a -> e)...)
1077 -- Simplifying the second arg makes us float the case out, so we end up with
1078 -- case y of (a,b) -> f (...(\a -> e)...) e'
1079 -- So the output does not have the no-shadowing invariant. However, there is
1080 -- no danger of getting name-capture, because when the first arg was simplified
1081 -- we used an in-scope set that at least mentioned all the variables free in its
1082 -- static environment, and that is enough.
1084 -- We can't just do innermost first, or we'd end up with a dual problem:
1085 -- case x of (a,b) -> f e (...(\a -> e')...)
1087 -- I spent hours trying to recover the no-shadowing invariant, but I just could
1088 -- not think of an elegant way to do it. The simplifier is already knee-deep in
1089 -- continuations. We have to keep the right in-scope set around; AND we have
1090 -- to get the effect that finding (error "foo") in a strict arg position will
1091 -- discard the entire application and replace it with (error "foo"). Getting
1092 -- all this at once is TOO HARD!
1094 simplifyArgs env fn_ty args cont_ty thing_inside
1095 = go env fn_ty args thing_inside
1097 go env fn_ty [] thing_inside = thing_inside env []
1098 go env fn_ty (arg:args) thing_inside = simplifyArg env fn_ty arg cont_ty $ \ env arg' ->
1099 go env (applyTypeToArg fn_ty arg') args $ \ env args' ->
1100 thing_inside env (arg':args')
1102 simplifyArg env fn_ty (Type ty_arg, se, _) cont_ty thing_inside
1103 = simplType (setInScope se env) ty_arg `thenSmpl` \ new_ty_arg ->
1104 thing_inside env (Type new_ty_arg)
1106 simplifyArg env fn_ty (val_arg, arg_se, is_strict) cont_ty thing_inside
1108 = simplStrictArg AnArg env val_arg arg_se arg_ty cont_ty thing_inside
1110 | otherwise -- Lazy argument
1111 -- DO NOT float anything outside, hence simplExprC
1112 -- There is no benefit (unlike in a let-binding), and we'd
1113 -- have to be very careful about bogus strictness through
1114 -- floating a demanded let.
1115 = simplExprC (setInScope arg_se env) val_arg
1116 (mkBoringStop arg_ty) `thenSmpl` \ arg1 ->
1117 thing_inside env arg1
1119 arg_ty = funArgTy fn_ty
1122 simplStrictArg :: LetRhsFlag
1123 -> SimplEnv -- The env of the call
1124 -> InExpr -> SimplEnv -- The arg plus its env
1125 -> OutType -- arg_ty: type of the argument
1126 -> OutType -- cont_ty: Type of thing computed by the context
1127 -> (SimplEnv -> OutExpr -> SimplM FloatsWithExpr)
1128 -- Takes an expression of type rhs_ty,
1129 -- returns an expression of type cont_ty
1130 -- The env passed to this continuation is the
1131 -- env of the call, plus any new in-scope variables
1132 -> SimplM FloatsWithExpr -- An expression of type cont_ty
1134 simplStrictArg is_rhs call_env arg arg_env arg_ty cont_ty thing_inside
1135 = simplExprF (setInScope arg_env call_env) arg
1136 (ArgOf is_rhs arg_ty cont_ty (\ new_env -> thing_inside (setInScope call_env new_env)))
1137 -- Notice the way we use arg_env (augmented with in-scope vars from call_env)
1138 -- to simplify the argument
1139 -- and call-env (augmented with in-scope vars from the arg) to pass to the continuation
1143 %************************************************************************
1145 \subsection{mkAtomicArgs}
1147 %************************************************************************
1149 mkAtomicArgs takes a putative RHS, checks whether it's a PAP or
1150 constructor application and, if so, converts it to ANF, so that the
1151 resulting thing can be inlined more easily. Thus
1158 There are three sorts of binding context, specified by the two
1164 N N Top-level or recursive Only bind args of lifted type
1166 N Y Non-top-level and non-recursive, Bind args of lifted type, or
1167 but lazy unlifted-and-ok-for-speculation
1169 Y Y Non-top-level, non-recursive, Bind all args
1170 and strict (demanded)
1177 there is no point in transforming to
1179 x = case (y div# z) of r -> MkC r
1181 because the (y div# z) can't float out of the let. But if it was
1182 a *strict* let, then it would be a good thing to do. Hence the
1183 context information.
1186 mkAtomicArgs :: Bool -- A strict binding
1187 -> Bool -- OK to float unlifted args
1189 -> SimplM (OrdList (OutId,OutExpr), -- The floats (unusually) may include
1190 OutExpr) -- things that need case-binding,
1191 -- if the strict-binding flag is on
1193 mkAtomicArgs is_strict ok_float_unlifted rhs
1194 | (Var fun, args) <- collectArgs rhs, -- It's an application
1195 isDataConWorkId fun || valArgCount args < idArity fun -- And it's a constructor or PAP
1196 = go fun nilOL [] args -- Have a go
1198 | otherwise = bale_out -- Give up
1201 bale_out = returnSmpl (nilOL, rhs)
1203 go fun binds rev_args []
1204 = returnSmpl (binds, mkApps (Var fun) (reverse rev_args))
1206 go fun binds rev_args (arg : args)
1207 | exprIsTrivial arg -- Easy case
1208 = go fun binds (arg:rev_args) args
1210 | not can_float_arg -- Can't make this arg atomic
1211 = bale_out -- ... so give up
1213 | otherwise -- Don't forget to do it recursively
1214 -- E.g. x = a:b:c:[]
1215 = mkAtomicArgs is_strict ok_float_unlifted arg `thenSmpl` \ (arg_binds, arg') ->
1216 newId FSLIT("a") arg_ty `thenSmpl` \ arg_id ->
1217 go fun ((arg_binds `snocOL` (arg_id,arg')) `appOL` binds)
1218 (Var arg_id : rev_args) args
1220 arg_ty = exprType arg
1221 can_float_arg = is_strict
1222 || not (isUnLiftedType arg_ty)
1223 || (ok_float_unlifted && exprOkForSpeculation arg)
1226 addAtomicBinds :: SimplEnv -> [(OutId,OutExpr)]
1227 -> (SimplEnv -> SimplM (FloatsWith a))
1228 -> SimplM (FloatsWith a)
1229 addAtomicBinds env [] thing_inside = thing_inside env
1230 addAtomicBinds env ((v,r):bs) thing_inside = addAuxiliaryBind env (NonRec v r) $ \ env ->
1231 addAtomicBinds env bs thing_inside
1233 addAtomicBindsE :: SimplEnv -> [(OutId,OutExpr)]
1234 -> (SimplEnv -> SimplM FloatsWithExpr)
1235 -> SimplM FloatsWithExpr
1236 -- Same again, but this time we're in an expression context,
1237 -- and may need to do some case bindings
1239 addAtomicBindsE env [] thing_inside
1241 addAtomicBindsE env ((v,r):bs) thing_inside
1242 | needsCaseBinding (idType v) r
1243 = addAtomicBindsE (addNewInScopeIds env [v]) bs thing_inside `thenSmpl` \ (floats, expr) ->
1244 WARN( exprIsTrivial expr, ppr v <+> pprCoreExpr expr )
1245 (let body = wrapFloats floats expr in
1246 returnSmpl (emptyFloats env, Case r v (exprType body) [(DEFAULT,[],body)]))
1249 = addAuxiliaryBind env (NonRec v r) $ \ env ->
1250 addAtomicBindsE env bs thing_inside
1254 %************************************************************************
1256 \subsection{The main rebuilder}
1258 %************************************************************************
1261 rebuild :: SimplEnv -> OutExpr -> SimplCont -> SimplM FloatsWithExpr
1263 rebuild env expr (Stop _ _ _) = rebuildDone env expr
1264 rebuild env expr (ArgOf _ _ _ cont_fn) = cont_fn env expr
1265 rebuild env expr (CoerceIt to_ty cont) = rebuild env (mkCoerce to_ty expr) cont
1266 rebuild env expr (InlinePlease cont) = rebuild env (Note InlineCall expr) cont
1267 rebuild env expr (Select _ bndr alts se cont) = rebuildCase (setInScope se env) expr bndr alts cont
1268 rebuild env expr (ApplyTo _ arg se cont) = rebuildApp (setInScope se env) expr arg cont
1270 rebuildApp env fun arg cont
1271 = simplExpr env arg `thenSmpl` \ arg' ->
1272 rebuild env (App fun arg') cont
1274 rebuildDone env expr = returnSmpl (emptyFloats env, expr)
1278 %************************************************************************
1280 \subsection{Functions dealing with a case}
1282 %************************************************************************
1284 Blob of helper functions for the "case-of-something-else" situation.
1287 ---------------------------------------------------------
1288 -- Eliminate the case if possible
1290 rebuildCase :: SimplEnv
1291 -> OutExpr -- Scrutinee
1292 -> InId -- Case binder
1293 -> [InAlt] -- Alternatives
1295 -> SimplM FloatsWithExpr
1297 rebuildCase env scrut case_bndr alts cont
1298 | Just (con,args) <- exprIsConApp_maybe scrut
1299 -- Works when the scrutinee is a variable with a known unfolding
1300 -- as well as when it's an explicit constructor application
1301 = knownCon env (DataAlt con) args case_bndr alts cont
1303 | Lit lit <- scrut -- No need for same treatment as constructors
1304 -- because literals are inlined more vigorously
1305 = knownCon env (LitAlt lit) [] case_bndr alts cont
1308 = prepareAlts scrut case_bndr alts `thenSmpl` \ (better_alts, handled_cons) ->
1310 -- Deal with the case binder, and prepare the continuation;
1311 -- The new subst_env is in place
1312 prepareCaseCont env better_alts cont `thenSmpl` \ (floats, (dup_cont, nondup_cont)) ->
1313 addFloats env floats $ \ env ->
1316 -- The case expression is annotated with the result type of the continuation
1317 -- This may differ from the type originally on the case. For example
1318 -- case(T) (case(Int#) a of { True -> 1#; False -> 0# }) of
1321 -- let j a# = <blob>
1322 -- in case(T) a of { True -> j 1#; False -> j 0# }
1323 -- Note that the case that scrutinises a now returns a T not an Int#
1324 res_ty' = contResultType dup_cont
1327 -- Deal with variable scrutinee
1328 simplCaseBinder env scrut case_bndr `thenSmpl` \ (alt_env, case_bndr') ->
1330 -- Deal with the case alternatives
1331 simplAlts alt_env handled_cons
1332 case_bndr' better_alts dup_cont `thenSmpl` \ alts' ->
1334 -- Put the case back together
1335 mkCase scrut case_bndr' res_ty' alts' `thenSmpl` \ case_expr ->
1337 -- Notice that rebuildDone returns the in-scope set from env, not alt_env
1338 -- The case binder *not* scope over the whole returned case-expression
1339 rebuild env case_expr nondup_cont
1342 simplCaseBinder checks whether the scrutinee is a variable, v. If so,
1343 try to eliminate uses of v in the RHSs in favour of case_bndr; that
1344 way, there's a chance that v will now only be used once, and hence
1349 There is a time we *don't* want to do that, namely when
1350 -fno-case-of-case is on. This happens in the first simplifier pass,
1351 and enhances full laziness. Here's the bad case:
1352 f = \ y -> ...(case x of I# v -> ...(case x of ...) ... )
1353 If we eliminate the inner case, we trap it inside the I# v -> arm,
1354 which might prevent some full laziness happening. I've seen this
1355 in action in spectral/cichelli/Prog.hs:
1356 [(m,n) | m <- [1..max], n <- [1..max]]
1357 Hence the check for NoCaseOfCase.
1361 There is another situation when we don't want to do it. If we have
1363 case x of w1 { DEFAULT -> case x of w2 { A -> e1; B -> e2 }
1364 ...other cases .... }
1366 We'll perform the binder-swap for the outer case, giving
1368 case x of w1 { DEFAULT -> case w1 of w2 { A -> e1; B -> e2 }
1369 ...other cases .... }
1371 But there is no point in doing it for the inner case, because w1 can't
1372 be inlined anyway. Furthermore, doing the case-swapping involves
1373 zapping w2's occurrence info (see paragraphs that follow), and that
1374 forces us to bind w2 when doing case merging. So we get
1376 case x of w1 { A -> let w2 = w1 in e1
1377 B -> let w2 = w1 in e2
1378 ...other cases .... }
1380 This is plain silly in the common case where w2 is dead.
1382 Even so, I can't see a good way to implement this idea. I tried
1383 not doing the binder-swap if the scrutinee was already evaluated
1384 but that failed big-time:
1388 case v of w { MkT x ->
1389 case x of x1 { I# y1 ->
1390 case x of x2 { I# y2 -> ...
1392 Notice that because MkT is strict, x is marked "evaluated". But to
1393 eliminate the last case, we must either make sure that x (as well as
1394 x1) has unfolding MkT y1. THe straightforward thing to do is to do
1395 the binder-swap. So this whole note is a no-op.
1399 If we replace the scrutinee, v, by tbe case binder, then we have to nuke
1400 any occurrence info (eg IAmDead) in the case binder, because the
1401 case-binder now effectively occurs whenever v does. AND we have to do
1402 the same for the pattern-bound variables! Example:
1404 (case x of { (a,b) -> a }) (case x of { (p,q) -> q })
1406 Here, b and p are dead. But when we move the argment inside the first
1407 case RHS, and eliminate the second case, we get
1409 case x of { (a,b) -> a b }
1411 Urk! b is alive! Reason: the scrutinee was a variable, and case elimination
1414 Indeed, this can happen anytime the case binder isn't dead:
1415 case <any> of x { (a,b) ->
1416 case x of { (p,q) -> p } }
1417 Here (a,b) both look dead, but come alive after the inner case is eliminated.
1418 The point is that we bring into the envt a binding
1420 after the outer case, and that makes (a,b) alive. At least we do unless
1421 the case binder is guaranteed dead.
1424 simplCaseBinder env (Var v) case_bndr
1425 | not (switchIsOn (getSwitchChecker env) NoCaseOfCase)
1427 -- Failed try [see Note 2 above]
1428 -- not (isEvaldUnfolding (idUnfolding v))
1430 = simplBinder env (zap case_bndr) `thenSmpl` \ (env, case_bndr') ->
1431 returnSmpl (modifyInScope env v case_bndr', case_bndr')
1432 -- We could extend the substitution instead, but it would be
1433 -- a hack because then the substitution wouldn't be idempotent
1434 -- any more (v is an OutId). And this does just as well.
1436 zap b = b `setIdOccInfo` NoOccInfo
1438 simplCaseBinder env other_scrut case_bndr
1439 = simplBinder env case_bndr `thenSmpl` \ (env, case_bndr') ->
1440 returnSmpl (env, case_bndr')
1446 simplAlts :: SimplEnv
1447 -> [AltCon] -- Alternatives the scrutinee can't be
1448 -- in the default case
1449 -> OutId -- Case binder
1450 -> [InAlt] -> SimplCont
1451 -> SimplM [OutAlt] -- Includes the continuation
1453 simplAlts env handled_cons case_bndr' alts cont'
1454 = mapSmpl simpl_alt alts
1456 simpl_alt alt = simplAlt env handled_cons case_bndr' alt cont' `thenSmpl` \ (_, alt') ->
1459 simplAlt :: SimplEnv -> [AltCon] -> OutId -> InAlt -> SimplCont
1460 -> SimplM (Maybe TvSubstEnv, OutAlt)
1461 -- Simplify an alternative, returning the type refinement for the
1462 -- alternative, if the alternative does any refinement at all
1464 simplAlt env handled_cons case_bndr' (DEFAULT, bndrs, rhs) cont'
1465 = ASSERT( null bndrs )
1466 simplExprC env' rhs cont' `thenSmpl` \ rhs' ->
1467 returnSmpl (Nothing, (DEFAULT, [], rhs'))
1469 env' = mk_rhs_env env case_bndr' (mkOtherCon handled_cons)
1470 -- Record the constructors that the case-binder *can't* be.
1472 simplAlt env handled_cons case_bndr' (LitAlt lit, bndrs, rhs) cont'
1473 = ASSERT( null bndrs )
1474 simplExprC env' rhs cont' `thenSmpl` \ rhs' ->
1475 returnSmpl (Nothing, (LitAlt lit, [], rhs'))
1477 env' = mk_rhs_env env case_bndr' (mkUnfolding False (Lit lit))
1479 simplAlt env handled_cons case_bndr' (DataAlt con, vs, rhs) cont'
1480 | isVanillaDataCon con
1481 = -- Deal with the pattern-bound variables
1482 -- Mark the ones that are in ! positions in the data constructor
1483 -- as certainly-evaluated.
1484 -- NB: it happens that simplBinders does *not* erase the OtherCon
1485 -- form of unfolding, so it's ok to add this info before
1486 -- doing simplBinders
1487 simplBinders env (add_evals con vs) `thenSmpl` \ (env, vs') ->
1489 -- Bind the case-binder to (con args)
1490 let unf = mkUnfolding False (mkConApp con con_args)
1491 inst_tys' = tyConAppArgs (idType case_bndr')
1492 con_args = map Type inst_tys' ++ map varToCoreExpr vs'
1493 env' = mk_rhs_env env case_bndr' unf
1495 simplExprC env' rhs cont' `thenSmpl` \ rhs' ->
1496 returnSmpl (Nothing, (DataAlt con, vs', rhs'))
1498 | otherwise -- GADT case
1500 (tvs,ids) = span isTyVar vs
1502 simplBinders env tvs `thenSmpl` \ (env1, tvs') ->
1504 pat_res_ty = dataConResTy con (mkTyVarTys tvs')
1505 tv_subst = getTvSubst env1
1507 case coreRefineTys tvs' tv_subst pat_res_ty (idType case_bndr') of {
1508 Nothing -- Dead code; for now, I'm just going to put in an
1509 -- error case so I can see them
1510 -> let rhs' = mkApps (Var eRROR_ID)
1511 [Type (substTy tv_subst (exprType rhs)),
1512 Lit (mkStringLit "Impossible alternative (GADT)")]
1514 simplBinders env1 ids `thenSmpl` \ (env2, ids') ->
1515 returnSmpl (Nothing, (DataAlt con, tvs' ++ ids', rhs')) ;
1517 Just tv_subst_env -> -- The normal case
1520 env2 = setTvSubstEnv env1 tv_subst_env
1521 -- Simplify the Ids in the refined environment, so their types
1522 -- reflect the refinement. Usually this doesn't matter, but it helps
1523 -- in mkDupableAlt, when we want to float a lambda that uses these binders
1525 simplBinders env2 (add_evals con ids) `thenSmpl` \ (env3, ids') ->
1526 let unf = mkUnfolding False con_app
1527 con_app = mkConApp con con_args
1528 con_args = map varToCoreExpr vs' -- NB: no inst_tys'
1529 env_w_unf = mk_rhs_env env3 case_bndr' unf
1532 simplExprC env_w_unf rhs cont' `thenSmpl` \ rhs' ->
1533 returnSmpl (Just tv_subst_env, (DataAlt con, vs', rhs')) }
1536 -- add_evals records the evaluated-ness of the bound variables of
1537 -- a case pattern. This is *important*. Consider
1538 -- data T = T !Int !Int
1540 -- case x of { T a b -> T (a+1) b }
1542 -- We really must record that b is already evaluated so that we don't
1543 -- go and re-evaluate it when constructing the result.
1544 add_evals dc vs = cat_evals dc vs (dataConRepStrictness dc)
1546 cat_evals dc vs strs
1550 go (v:vs) strs | isTyVar v = v : go vs strs
1551 go (v:vs) (str:strs)
1552 | isMarkedStrict str = evald_v : go vs strs
1553 | otherwise = zapped_v : go vs strs
1555 zapped_v = zap_occ_info v
1556 evald_v = zapped_v `setIdUnfolding` mkOtherCon []
1557 go _ _ = pprPanic "cat_evals" (ppr dc $$ ppr vs $$ ppr strs)
1559 -- If the case binder is alive, then we add the unfolding
1561 -- to the envt; so vs are now very much alive
1562 zap_occ_info | isDeadBinder case_bndr' = \id -> id
1563 | otherwise = \id -> id `setIdOccInfo` NoOccInfo
1565 mk_rhs_env env case_bndr' case_bndr_unf
1566 = modifyInScope env case_bndr' (case_bndr' `setIdUnfolding` case_bndr_unf)
1570 %************************************************************************
1572 \subsection{Known constructor}
1574 %************************************************************************
1576 We are a bit careful with occurrence info. Here's an example
1578 (\x* -> case x of (a*, b) -> f a) (h v, e)
1580 where the * means "occurs once". This effectively becomes
1581 case (h v, e) of (a*, b) -> f a)
1583 let a* = h v; b = e in f a
1587 All this should happen in one sweep.
1590 knownCon :: SimplEnv -> AltCon -> [OutExpr]
1591 -> InId -> [InAlt] -> SimplCont
1592 -> SimplM FloatsWithExpr
1594 knownCon env con args bndr alts cont
1595 = tick (KnownBranch bndr) `thenSmpl_`
1596 case findAlt con alts of
1597 (DEFAULT, bs, rhs) -> ASSERT( null bs )
1598 simplNonRecX env bndr scrut $ \ env ->
1599 -- This might give rise to a binding with non-atomic args
1600 -- like x = Node (f x) (g x)
1601 -- but no harm will be done
1602 simplExprF env rhs cont
1605 LitAlt lit -> Lit lit
1606 DataAlt dc -> mkConApp dc args
1608 (LitAlt lit, bs, rhs) -> ASSERT( null bs )
1609 simplNonRecX env bndr (Lit lit) $ \ env ->
1610 simplExprF env rhs cont
1612 (DataAlt dc, bs, rhs)
1613 -> ASSERT( n_drop_tys + length bs == length args )
1614 bind_args env bs (drop n_drop_tys args) $ \ env ->
1616 con_app = mkConApp dc (take n_drop_tys args ++ con_args)
1617 con_args = [substExpr (getSubst env) (varToCoreExpr b) | b <- bs]
1618 -- args are aready OutExprs, but bs are InIds
1620 simplNonRecX env bndr con_app $ \ env ->
1621 simplExprF env rhs cont
1623 n_drop_tys | isVanillaDataCon dc = tyConArity (dataConTyCon dc)
1625 -- Vanilla data constructors lack type arguments in the pattern
1628 bind_args env [] _ thing_inside = thing_inside env
1630 bind_args env (b:bs) (Type ty : args) thing_inside
1631 = ASSERT( isTyVar b )
1632 bind_args (extendTvSubst env b ty) bs args thing_inside
1634 bind_args env (b:bs) (arg : args) thing_inside
1636 simplNonRecX env b arg $ \ env ->
1637 bind_args env bs args thing_inside
1641 %************************************************************************
1643 \subsection{Duplicating continuations}
1645 %************************************************************************
1648 prepareCaseCont :: SimplEnv
1649 -> [InAlt] -> SimplCont
1650 -> SimplM (FloatsWith (SimplCont,SimplCont))
1651 -- Return a duplicatable continuation, a non-duplicable part
1652 -- plus some extra bindings
1654 -- No need to make it duplicatable if there's only one alternative
1655 prepareCaseCont env [alt] cont = returnSmpl (emptyFloats env, (cont, mkBoringStop (contResultType cont)))
1656 prepareCaseCont env alts cont = mkDupableCont env cont
1660 mkDupableCont :: SimplEnv -> SimplCont
1661 -> SimplM (FloatsWith (SimplCont, SimplCont))
1663 mkDupableCont env cont
1664 | contIsDupable cont
1665 = returnSmpl (emptyFloats env, (cont, mkBoringStop (contResultType cont)))
1667 mkDupableCont env (CoerceIt ty cont)
1668 = mkDupableCont env cont `thenSmpl` \ (floats, (dup_cont, nondup_cont)) ->
1669 returnSmpl (floats, (CoerceIt ty dup_cont, nondup_cont))
1671 mkDupableCont env (InlinePlease cont)
1672 = mkDupableCont env cont `thenSmpl` \ (floats, (dup_cont, nondup_cont)) ->
1673 returnSmpl (floats, (InlinePlease dup_cont, nondup_cont))
1675 mkDupableCont env cont@(ArgOf _ arg_ty _ _)
1676 = returnSmpl (emptyFloats env, (mkBoringStop arg_ty, cont))
1677 -- Do *not* duplicate an ArgOf continuation
1678 -- Because ArgOf continuations are opaque, we gain nothing by
1679 -- propagating them into the expressions, and we do lose a lot.
1680 -- Here's an example:
1681 -- && (case x of { T -> F; F -> T }) E
1682 -- Now, && is strict so we end up simplifying the case with
1683 -- an ArgOf continuation. If we let-bind it, we get
1685 -- let $j = \v -> && v E
1686 -- in simplExpr (case x of { T -> F; F -> T })
1687 -- (ArgOf (\r -> $j r)
1688 -- And after simplifying more we get
1690 -- let $j = \v -> && v E
1691 -- in case of { T -> $j F; F -> $j T }
1692 -- Which is a Very Bad Thing
1694 -- The desire not to duplicate is the entire reason that
1695 -- mkDupableCont returns a pair of continuations.
1697 -- The original plan had:
1698 -- e.g. (...strict-fn...) [...hole...]
1700 -- let $j = \a -> ...strict-fn...
1701 -- in $j [...hole...]
1703 mkDupableCont env (ApplyTo _ arg se cont)
1704 = -- e.g. [...hole...] (...arg...)
1706 -- let a = ...arg...
1707 -- in [...hole...] a
1708 simplExpr (setInScope se env) arg `thenSmpl` \ arg' ->
1710 mkDupableCont env cont `thenSmpl` \ (floats, (dup_cont, nondup_cont)) ->
1711 addFloats env floats $ \ env ->
1713 if exprIsDupable arg' then
1714 returnSmpl (emptyFloats env, (ApplyTo OkToDup arg' (zapSubstEnv se) dup_cont, nondup_cont))
1716 newId FSLIT("a") (exprType arg') `thenSmpl` \ arg_id ->
1718 tick (CaseOfCase arg_id) `thenSmpl_`
1719 -- Want to tick here so that we go round again,
1720 -- and maybe copy or inline the code.
1721 -- Not strictly CaseOfCase, but never mind
1723 returnSmpl (unitFloat env arg_id arg',
1724 (ApplyTo OkToDup (Var arg_id) (zapSubstEnv se) dup_cont,
1726 -- But what if the arg should be case-bound?
1727 -- This has been this way for a long time, so I'll leave it,
1728 -- but I can't convince myself that it's right.
1731 mkDupableCont env (Select _ case_bndr alts se cont)
1732 = -- e.g. (case [...hole...] of { pi -> ei })
1734 -- let ji = \xij -> ei
1735 -- in case [...hole...] of { pi -> ji xij }
1736 tick (CaseOfCase case_bndr) `thenSmpl_`
1738 alt_env = setInScope se env
1740 prepareCaseCont alt_env alts cont `thenSmpl` \ (floats1, (dup_cont, nondup_cont)) ->
1741 addFloats alt_env floats1 $ \ alt_env ->
1743 simplBinder alt_env case_bndr `thenSmpl` \ (alt_env, case_bndr') ->
1744 -- NB: simplBinder does not zap deadness occ-info, so
1745 -- a dead case_bndr' will still advertise its deadness
1746 -- This is really important because in
1747 -- case e of b { (# a,b #) -> ... }
1748 -- b is always dead, and indeed we are not allowed to bind b to (# a,b #),
1749 -- which might happen if e was an explicit unboxed pair and b wasn't marked dead.
1750 -- In the new alts we build, we have the new case binder, so it must retain
1753 mkDupableAlts alt_env case_bndr' alts dup_cont `thenSmpl` \ (floats2, alts') ->
1754 addFloats alt_env floats2 $ \ alt_env ->
1755 returnSmpl (emptyFloats alt_env,
1756 (Select OkToDup case_bndr' alts' (zapSubstEnv se)
1757 (mkBoringStop (contResultType dup_cont)),
1760 mkDupableAlts :: SimplEnv -> OutId -> [InAlt] -> SimplCont
1761 -> SimplM (FloatsWith [InAlt])
1762 -- Absorbs the continuation into the new alternatives
1764 mkDupableAlts env case_bndr' alts dupable_cont
1767 go env [] = returnSmpl (emptyFloats env, [])
1769 = mkDupableAlt env case_bndr' dupable_cont alt `thenSmpl` \ (floats1, alt') ->
1770 addFloats env floats1 $ \ env ->
1771 go env alts `thenSmpl` \ (floats2, alts') ->
1772 returnSmpl (floats2, alt' : alts')
1774 mkDupableAlt env case_bndr' cont alt
1775 = simplAlt env [] case_bndr' alt cont `thenSmpl` \ (mb_reft, (con, bndrs', rhs')) ->
1776 -- Safe to say that there are no handled-cons for the DEFAULT case
1778 if exprIsDupable rhs' then
1779 returnSmpl (emptyFloats env, (con, bndrs', rhs'))
1780 -- It is worth checking for a small RHS because otherwise we
1781 -- get extra let bindings that may cause an extra iteration of the simplifier to
1782 -- inline back in place. Quite often the rhs is just a variable or constructor.
1783 -- The Ord instance of Maybe in PrelMaybe.lhs, for example, took several extra
1784 -- iterations because the version with the let bindings looked big, and so wasn't
1785 -- inlined, but after the join points had been inlined it looked smaller, and so
1788 -- NB: we have to check the size of rhs', not rhs.
1789 -- Duplicating a small InAlt might invalidate occurrence information
1790 -- However, if it *is* dupable, we return the *un* simplified alternative,
1791 -- because otherwise we'd need to pair it up with an empty subst-env....
1792 -- but we only have one env shared between all the alts.
1793 -- (Remember we must zap the subst-env before re-simplifying something).
1794 -- Rather than do this we simply agree to re-simplify the original (small) thing later.
1798 rhs_ty' = exprType rhs'
1799 used_bndrs' = filter abstract_over (case_bndr' : bndrs')
1801 | isTyVar bndr = not (mb_reft `refines` bndr)
1802 -- Don't abstract over tyvar binders which are refined away
1803 | otherwise = not (isDeadBinder bndr)
1804 -- The deadness info on the new Ids is preserved by simplBinders
1805 refines Nothing bndr = False
1806 refines (Just tv_subst) bndr = bndr `elemVarEnv` tv_subst
1807 -- See Note [Refinement] below
1809 -- If we try to lift a primitive-typed something out
1810 -- for let-binding-purposes, we will *caseify* it (!),
1811 -- with potentially-disastrous strictness results. So
1812 -- instead we turn it into a function: \v -> e
1813 -- where v::State# RealWorld#. The value passed to this function
1814 -- is realworld#, which generates (almost) no code.
1816 -- There's a slight infelicity here: we pass the overall
1817 -- case_bndr to all the join points if it's used in *any* RHS,
1818 -- because we don't know its usage in each RHS separately
1820 -- We used to say "&& isUnLiftedType rhs_ty'" here, but now
1821 -- we make the join point into a function whenever used_bndrs'
1822 -- is empty. This makes the join-point more CPR friendly.
1823 -- Consider: let j = if .. then I# 3 else I# 4
1824 -- in case .. of { A -> j; B -> j; C -> ... }
1826 -- Now CPR doesn't w/w j because it's a thunk, so
1827 -- that means that the enclosing function can't w/w either,
1828 -- which is a lose. Here's the example that happened in practice:
1829 -- kgmod :: Int -> Int -> Int
1830 -- kgmod x y = if x > 0 && y < 0 || x < 0 && y > 0
1834 -- I have seen a case alternative like this:
1835 -- True -> \v -> ...
1836 -- It's a bit silly to add the realWorld dummy arg in this case, making
1839 -- (the \v alone is enough to make CPR happy) but I think it's rare
1841 ( if not (any isId used_bndrs')
1842 then newId FSLIT("w") realWorldStatePrimTy `thenSmpl` \ rw_id ->
1843 returnSmpl ([rw_id], [Var realWorldPrimId])
1845 returnSmpl (used_bndrs', map varToCoreExpr used_bndrs')
1846 ) `thenSmpl` \ (final_bndrs', final_args) ->
1848 -- See comment about "$j" name above
1849 newId (encodeFS FSLIT("$j")) (mkPiTypes final_bndrs' rhs_ty') `thenSmpl` \ join_bndr ->
1850 -- Notice the funky mkPiTypes. If the contructor has existentials
1851 -- it's possible that the join point will be abstracted over
1852 -- type varaibles as well as term variables.
1853 -- Example: Suppose we have
1854 -- data T = forall t. C [t]
1856 -- case (case e of ...) of
1857 -- C t xs::[t] -> rhs
1858 -- We get the join point
1859 -- let j :: forall t. [t] -> ...
1860 -- j = /\t \xs::[t] -> rhs
1862 -- case (case e of ...) of
1863 -- C t xs::[t] -> j t xs
1865 -- We make the lambdas into one-shot-lambdas. The
1866 -- join point is sure to be applied at most once, and doing so
1867 -- prevents the body of the join point being floated out by
1868 -- the full laziness pass
1869 really_final_bndrs = map one_shot final_bndrs'
1870 one_shot v | isId v = setOneShotLambda v
1872 join_rhs = mkLams really_final_bndrs rhs'
1873 join_call = mkApps (Var join_bndr) final_args
1875 returnSmpl (unitFloat env join_bndr join_rhs, (con, bndrs', join_call))
1882 MkT :: a -> b -> T a
1886 MkT a' b (p::a') (q::b) -> [p,w]
1888 The danger is that we'll make a join point
1892 and that's ill-typed, because (p::a') but (w::a).
1894 Solution so far: don't abstract over a', because the type refinement
1895 maps [a' -> a] . Ultimately that won't work when real refinement goes on.
1897 Then we must abstract over any refined free variables. Hmm. Maybe we
1898 could just abstract over *all* free variables, thereby lambda-lifting
1899 the join point? We should try this.