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, coreEqType, substTy, mkTyVarTys
54 import VarEnv ( elemVarEnv )
55 import Subst ( SubstResult(..), emptySubst, substExpr,
56 substId, simplIdInfo )
57 import TysPrim ( realWorldStatePrimTy )
58 import PrelInfo ( realWorldPrimId )
59 import BasicTypes ( TopLevelFlag(..), isTopLevel,
63 import Maybe ( Maybe )
64 import Maybes ( orElse )
66 import Util ( notNull )
70 The guts of the simplifier is in this module, but the driver loop for
71 the simplifier is in SimplCore.lhs.
74 -----------------------------------------
75 *** IMPORTANT NOTE ***
76 -----------------------------------------
77 The simplifier used to guarantee that the output had no shadowing, but
78 it does not do so any more. (Actually, it never did!) The reason is
79 documented with simplifyArgs.
82 -----------------------------------------
83 *** IMPORTANT NOTE ***
84 -----------------------------------------
85 Many parts of the simplifier return a bunch of "floats" as well as an
86 expression. This is wrapped as a datatype SimplUtils.FloatsWith.
88 All "floats" are let-binds, not case-binds, but some non-rec lets may
89 be unlifted (with RHS ok-for-speculation).
93 -----------------------------------------
94 ORGANISATION OF FUNCTIONS
95 -----------------------------------------
97 - simplify all top-level binders
98 - for NonRec, call simplRecOrTopPair
99 - for Rec, call simplRecBind
102 ------------------------------
103 simplExpr (applied lambda) ==> simplNonRecBind
104 simplExpr (Let (NonRec ...) ..) ==> simplNonRecBind
105 simplExpr (Let (Rec ...) ..) ==> simplify binders; simplRecBind
107 ------------------------------
108 simplRecBind [binders already simplfied]
109 - use simplRecOrTopPair on each pair in turn
111 simplRecOrTopPair [binder already simplified]
112 Used for: recursive bindings (top level and nested)
113 top-level non-recursive bindings
115 - check for PreInlineUnconditionally
119 Used for: non-top-level non-recursive bindings
120 beta reductions (which amount to the same thing)
121 Because it can deal with strict arts, it takes a
122 "thing-inside" and returns an expression
124 - check for PreInlineUnconditionally
125 - simplify binder, including its IdInfo
134 simplNonRecX: [given a *simplified* RHS, but an *unsimplified* binder]
135 Used for: binding case-binder and constr args in a known-constructor case
136 - check for PreInLineUnconditionally
140 ------------------------------
141 simplLazyBind: [binder already simplified, RHS not]
142 Used for: recursive bindings (top level and nested)
143 top-level non-recursive bindings
144 non-top-level, but *lazy* non-recursive bindings
145 [must not be strict or unboxed]
146 Returns floats + an augmented environment, not an expression
147 - substituteIdInfo and add result to in-scope
148 [so that rules are available in rec rhs]
151 - float if exposes constructor or PAP
155 completeNonRecX: [binder and rhs both simplified]
156 - if the the thing needs case binding (unlifted and not ok-for-spec)
162 completeLazyBind: [given a simplified RHS]
163 [used for both rec and non-rec bindings, top level and not]
164 - try PostInlineUnconditionally
165 - add unfolding [this is the only place we add an unfolding]
170 Right hand sides and arguments
171 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
172 In many ways we want to treat
173 (a) the right hand side of a let(rec), and
174 (b) a function argument
175 in the same way. But not always! In particular, we would
176 like to leave these arguments exactly as they are, so they
177 will match a RULE more easily.
182 It's harder to make the rule match if we ANF-ise the constructor,
183 or eta-expand the PAP:
185 f (let { a = g x; b = h x } in (a,b))
188 On the other hand if we see the let-defns
193 then we *do* want to ANF-ise and eta-expand, so that p and q
194 can be safely inlined.
196 Even floating lets out is a bit dubious. For let RHS's we float lets
197 out if that exposes a value, so that the value can be inlined more vigorously.
200 r = let x = e in (x,x)
202 Here, if we float the let out we'll expose a nice constructor. We did experiments
203 that showed this to be a generally good thing. But it was a bad thing to float
204 lets out unconditionally, because that meant they got allocated more often.
206 For function arguments, there's less reason to expose a constructor (it won't
207 get inlined). Just possibly it might make a rule match, but I'm pretty skeptical.
208 So for the moment we don't float lets out of function arguments either.
213 For eta expansion, we want to catch things like
215 case e of (a,b) -> \x -> case a of (p,q) -> \y -> r
217 If the \x was on the RHS of a let, we'd eta expand to bring the two
218 lambdas together. And in general that's a good thing to do. Perhaps
219 we should eta expand wherever we find a (value) lambda? Then the eta
220 expansion at a let RHS can concentrate solely on the PAP case.
223 %************************************************************************
225 \subsection{Bindings}
227 %************************************************************************
230 simplTopBinds :: SimplEnv -> [InBind] -> SimplM [OutBind]
232 simplTopBinds env binds
233 = -- Put all the top-level binders into scope at the start
234 -- so that if a transformation rule has unexpectedly brought
235 -- anything into scope, then we don't get a complaint about that.
236 -- It's rather as if the top-level binders were imported.
237 simplRecBndrs env (bindersOfBinds binds) `thenSmpl` \ (env, bndrs') ->
238 simpl_binds env binds bndrs' `thenSmpl` \ (floats, _) ->
239 freeTick SimplifierDone `thenSmpl_`
240 returnSmpl (floatBinds floats)
242 -- We need to track the zapped top-level binders, because
243 -- they should have their fragile IdInfo zapped (notably occurrence info)
244 -- That's why we run down binds and bndrs' simultaneously.
245 simpl_binds :: SimplEnv -> [InBind] -> [OutId] -> SimplM (FloatsWith ())
246 simpl_binds env [] bs = ASSERT( null bs ) returnSmpl (emptyFloats env, ())
247 simpl_binds env (bind:binds) bs = simpl_bind env bind bs `thenSmpl` \ (floats,env) ->
248 addFloats env floats $ \env ->
249 simpl_binds env binds (drop_bs bind bs)
251 drop_bs (NonRec _ _) (_ : bs) = bs
252 drop_bs (Rec prs) bs = drop (length prs) bs
254 simpl_bind env bind bs
255 = getDOptsSmpl `thenSmpl` \ dflags ->
256 if dopt Opt_D_dump_inlinings dflags then
257 pprTrace "SimplBind" (ppr (bindersOf bind)) $ simpl_bind1 env bind bs
259 simpl_bind1 env bind bs
261 simpl_bind1 env (NonRec b r) (b':_) = simplRecOrTopPair env TopLevel b b' r
262 simpl_bind1 env (Rec pairs) bs' = simplRecBind env TopLevel pairs bs'
266 %************************************************************************
268 \subsection{simplNonRec}
270 %************************************************************************
272 simplNonRecBind is used for
273 * non-top-level non-recursive lets in expressions
277 * An unsimplified (binder, rhs) pair
278 * The env for the RHS. It may not be the same as the
279 current env because the bind might occur via (\x.E) arg
281 It uses the CPS form because the binding might be strict, in which
282 case we might discard the continuation:
283 let x* = error "foo" in (...x...)
285 It needs to turn unlifted bindings into a @case@. They can arise
286 from, say: (\x -> e) (4# + 3#)
289 simplNonRecBind :: SimplEnv
291 -> InExpr -> SimplEnv -- Arg, with its subst-env
292 -> OutType -- Type of thing computed by the context
293 -> (SimplEnv -> SimplM FloatsWithExpr) -- The body
294 -> SimplM FloatsWithExpr
296 simplNonRecBind env bndr rhs rhs_se cont_ty thing_inside
298 = pprPanic "simplNonRecBind" (ppr bndr <+> ppr rhs)
301 simplNonRecBind env bndr rhs rhs_se cont_ty thing_inside
302 | preInlineUnconditionally env NotTopLevel bndr
303 = tick (PreInlineUnconditionally bndr) `thenSmpl_`
304 thing_inside (extendIdSubst env bndr (ContEx (getSubst rhs_se) rhs))
307 | isStrictDmd (idNewDemandInfo bndr) || isStrictType (idType bndr) -- A strict let
308 = -- Don't use simplBinder because that doesn't keep
309 -- fragile occurrence info in the substitution
310 simplLetBndr env bndr `thenSmpl` \ (env, bndr1) ->
311 simplStrictArg AnRhs env rhs rhs_se (idType bndr1) cont_ty $ \ env1 rhs1 ->
313 -- Now complete the binding and simplify the body
315 -- simplLetBndr doesn't deal with the IdInfo, so we must
316 -- do so here (c.f. simplLazyBind)
317 bndr2 = bndr1 `setIdInfo` simplIdInfo (getSubst env) (idInfo bndr)
318 env2 = modifyInScope env1 bndr2 bndr2
320 completeNonRecX env2 True {- strict -} bndr bndr2 rhs1 thing_inside
322 | otherwise -- Normal, lazy case
323 = -- Don't use simplBinder because that doesn't keep
324 -- fragile occurrence info in the substitution
325 simplLetBndr env bndr `thenSmpl` \ (env, bndr') ->
326 simplLazyBind env NotTopLevel NonRecursive
327 bndr bndr' rhs rhs_se `thenSmpl` \ (floats, env) ->
328 addFloats env floats thing_inside
331 A specialised variant of simplNonRec used when the RHS is already simplified, notably
332 in knownCon. It uses case-binding where necessary.
335 simplNonRecX :: SimplEnv
336 -> InId -- Old binder
337 -> OutExpr -- Simplified RHS
338 -> (SimplEnv -> SimplM FloatsWithExpr)
339 -> SimplM FloatsWithExpr
341 simplNonRecX env bndr new_rhs thing_inside
342 | needsCaseBinding (idType bndr) new_rhs
343 -- Make this test *before* the preInlineUnconditionally
344 -- Consider case I# (quotInt# x y) of
345 -- I# v -> let w = J# v in ...
346 -- If we gaily inline (quotInt# x y) for v, we end up building an
348 -- let w = J# (quotInt# x y) in ...
349 -- because quotInt# can fail.
350 = simplBinder env bndr `thenSmpl` \ (env, bndr') ->
351 thing_inside env `thenSmpl` \ (floats, body) ->
352 let body' = wrapFloats floats body in
353 returnSmpl (emptyFloats env, Case new_rhs bndr' (exprType body') [(DEFAULT, [], body')])
355 | preInlineUnconditionally env NotTopLevel bndr
356 -- This happens; for example, the case_bndr during case of
357 -- known constructor: case (a,b) of x { (p,q) -> ... }
358 -- Here x isn't mentioned in the RHS, so we don't want to
359 -- create the (dead) let-binding let x = (a,b) in ...
361 -- Similarly, single occurrences can be inlined vigourously
362 -- e.g. case (f x, g y) of (a,b) -> ....
363 -- If a,b occur once we can avoid constructing the let binding for them.
364 = thing_inside (extendIdSubst env bndr (ContEx emptySubst new_rhs))
367 = simplBinder env bndr `thenSmpl` \ (env, bndr') ->
368 completeNonRecX env False {- Non-strict; pessimistic -}
369 bndr bndr' new_rhs thing_inside
371 completeNonRecX env is_strict old_bndr new_bndr new_rhs thing_inside
372 = mkAtomicArgs is_strict
373 True {- OK to float unlifted -}
374 new_rhs `thenSmpl` \ (aux_binds, rhs2) ->
376 -- Make the arguments atomic if necessary,
377 -- adding suitable bindings
378 addAtomicBindsE env (fromOL aux_binds) $ \ env ->
379 completeLazyBind env NotTopLevel
380 old_bndr new_bndr rhs2 `thenSmpl` \ (floats, env) ->
381 addFloats env floats thing_inside
385 %************************************************************************
387 \subsection{Lazy bindings}
389 %************************************************************************
391 simplRecBind is used for
392 * recursive bindings only
395 simplRecBind :: SimplEnv -> TopLevelFlag
396 -> [(InId, InExpr)] -> [OutId]
397 -> SimplM (FloatsWith SimplEnv)
398 simplRecBind env top_lvl pairs bndrs'
399 = go env pairs bndrs' `thenSmpl` \ (floats, env) ->
400 returnSmpl (flattenFloats floats, env)
402 go env [] _ = returnSmpl (emptyFloats env, env)
404 go env ((bndr, rhs) : pairs) (bndr' : bndrs')
405 = simplRecOrTopPair env top_lvl bndr bndr' rhs `thenSmpl` \ (floats, env) ->
406 addFloats env floats (\env -> go env pairs bndrs')
410 simplRecOrTopPair is used for
411 * recursive bindings (whether top level or not)
412 * top-level non-recursive bindings
414 It assumes the binder has already been simplified, but not its IdInfo.
417 simplRecOrTopPair :: SimplEnv
419 -> InId -> OutId -- Binder, both pre-and post simpl
420 -> InExpr -- The RHS and its environment
421 -> SimplM (FloatsWith SimplEnv)
423 simplRecOrTopPair env top_lvl bndr bndr' rhs
424 | preInlineUnconditionally env top_lvl bndr -- Check for unconditional inline
425 = tick (PreInlineUnconditionally bndr) `thenSmpl_`
426 returnSmpl (emptyFloats env, extendIdSubst env bndr (ContEx (getSubst env) rhs))
429 = simplLazyBind env top_lvl Recursive bndr bndr' rhs env
430 -- May not actually be recursive, but it doesn't matter
434 simplLazyBind is used for
435 * recursive bindings (whether top level or not)
436 * top-level non-recursive bindings
437 * non-top-level *lazy* non-recursive bindings
439 [Thus it deals with the lazy cases from simplNonRecBind, and all cases
440 from SimplRecOrTopBind]
443 1. It assumes that the binder is *already* simplified,
444 and is in scope, but not its IdInfo
446 2. It assumes that the binder type is lifted.
448 3. It does not check for pre-inline-unconditionallly;
449 that should have been done already.
452 simplLazyBind :: SimplEnv
453 -> TopLevelFlag -> RecFlag
454 -> InId -> OutId -- Binder, both pre-and post simpl
455 -> InExpr -> SimplEnv -- The RHS and its environment
456 -> SimplM (FloatsWith SimplEnv)
458 simplLazyBind env top_lvl is_rec bndr bndr1 rhs rhs_se
459 = let -- Transfer the IdInfo of the original binder to the new binder
460 -- This is crucial: we must preserve
464 -- etc. To do this we must apply the current substitution,
465 -- which incorporates earlier substitutions in this very letrec group.
467 -- NB 1. We do this *before* processing the RHS of the binder, so that
468 -- its substituted rules are visible in its own RHS.
469 -- This is important. Manuel found cases where he really, really
470 -- wanted a RULE for a recursive function to apply in that function's
471 -- own right-hand side.
473 -- NB 2: We do not transfer the arity (see Subst.substIdInfo)
474 -- The arity of an Id should not be visible
475 -- in its own RHS, else we eta-reduce
479 -- which isn't sound. And it makes the arity in f's IdInfo greater than
480 -- the manifest arity, which isn't good.
481 -- The arity will get added later.
483 -- NB 3: It's important that we *do* transer the loop-breaker OccInfo,
484 -- because that's what stops the Id getting inlined infinitely, in the body
487 -- NB 4: does no harm for non-recursive bindings
489 bndr2 = bndr1 `setIdInfo` simplIdInfo (getSubst env) (idInfo bndr)
490 env1 = modifyInScope env bndr2 bndr2
491 rhs_env = setInScope rhs_se env1
492 is_top_level = isTopLevel top_lvl
493 ok_float_unlifted = not is_top_level && isNonRec is_rec
494 rhs_cont = mkRhsStop (idType bndr1)
496 -- Simplify the RHS; note the mkRhsStop, which tells
497 -- the simplifier that this is the RHS of a let.
498 simplExprF rhs_env rhs rhs_cont `thenSmpl` \ (floats, rhs1) ->
500 -- If any of the floats can't be floated, give up now
501 -- (The allLifted predicate says True for empty floats.)
502 if (not ok_float_unlifted && not (allLifted floats)) then
503 completeLazyBind env1 top_lvl bndr bndr2
504 (wrapFloats floats rhs1)
507 -- ANF-ise a constructor or PAP rhs
508 mkAtomicArgs False {- Not strict -}
509 ok_float_unlifted rhs1 `thenSmpl` \ (aux_binds, rhs2) ->
511 -- If the result is a PAP, float the floats out, else wrap them
512 -- By this time it's already been ANF-ised (if necessary)
513 if isEmptyFloats floats && isNilOL aux_binds then -- Shortcut a common case
514 completeLazyBind env1 top_lvl bndr bndr2 rhs2
516 else if is_top_level || exprIsTrivial rhs2 || exprIsValue rhs2 then
517 -- WARNING: long dodgy argument coming up
518 -- WANTED: a better way to do this
520 -- We can't use "exprIsCheap" instead of exprIsValue,
521 -- because that causes a strictness bug.
522 -- x = let y* = E in case (scc y) of { T -> F; F -> T}
523 -- The case expression is 'cheap', but it's wrong to transform to
524 -- y* = E; x = case (scc y) of {...}
525 -- Either we must be careful not to float demanded non-values, or
526 -- we must use exprIsValue for the test, which ensures that the
527 -- thing is non-strict. So exprIsValue => bindings are non-strict
528 -- I think. The WARN below tests for this.
530 -- We use exprIsTrivial here because we want to reveal lone variables.
531 -- E.g. let { x = letrec { y = E } in y } in ...
532 -- Here we definitely want to float the y=E defn.
533 -- exprIsValue definitely isn't right for that.
535 -- Again, the floated binding can't be strict; if it's recursive it'll
536 -- be non-strict; if it's non-recursive it'd be inlined.
538 -- Note [SCC-and-exprIsTrivial]
540 -- y = let { x* = E } in scc "foo" x
541 -- then we do *not* want to float out the x binding, because
542 -- it's strict! Fortunately, exprIsTrivial replies False to
545 -- There's a subtlety here. There may be a binding (x* = e) in the
546 -- floats, where the '*' means 'will be demanded'. So is it safe
547 -- to float it out? Answer no, but it won't matter because
548 -- we only float if (a) arg' is a WHNF, or (b) it's going to top level
549 -- and so there can't be any 'will be demanded' bindings in the floats.
551 ASSERT2( is_top_level || not (any demanded_float (floatBinds floats)),
552 ppr (filter demanded_float (floatBinds floats)) )
554 tick LetFloatFromLet `thenSmpl_` (
555 addFloats env1 floats $ \ env2 ->
556 addAtomicBinds env2 (fromOL aux_binds) $ \ env3 ->
557 completeLazyBind env3 top_lvl bndr bndr2 rhs2)
560 completeLazyBind env1 top_lvl bndr bndr2 (wrapFloats floats rhs1)
563 demanded_float (NonRec b r) = isStrictDmd (idNewDemandInfo b) && not (isUnLiftedType (idType b))
564 -- Unlifted-type (cheap-eagerness) lets may well have a demanded flag on them
565 demanded_float (Rec _) = False
570 %************************************************************************
572 \subsection{Completing a lazy binding}
574 %************************************************************************
577 * deals only with Ids, not TyVars
578 * takes an already-simplified binder and RHS
579 * is used for both recursive and non-recursive bindings
580 * is used for both top-level and non-top-level bindings
582 It does the following:
583 - tries discarding a dead binding
584 - tries PostInlineUnconditionally
585 - add unfolding [this is the only place we add an unfolding]
588 It does *not* attempt to do let-to-case. Why? Because it is used for
589 - top-level bindings (when let-to-case is impossible)
590 - many situations where the "rhs" is known to be a WHNF
591 (so let-to-case is inappropriate).
594 completeLazyBind :: SimplEnv
595 -> TopLevelFlag -- Flag stuck into unfolding
596 -> InId -- Old binder
597 -> OutId -- New binder
598 -> OutExpr -- Simplified RHS
599 -> SimplM (FloatsWith SimplEnv)
600 -- We return a new SimplEnv, because completeLazyBind may choose to do its work
601 -- by extending the substitution (e.g. let x = y in ...)
602 -- The new binding (if any) is returned as part of the floats.
603 -- NB: the returned SimplEnv has the right SubstEnv, but you should
604 -- (as usual) use the in-scope-env from the floats
606 completeLazyBind env top_lvl old_bndr new_bndr new_rhs
607 | postInlineUnconditionally env new_bndr occ_info new_rhs
608 = -- Drop the binding
609 tick (PostInlineUnconditionally old_bndr) `thenSmpl_`
610 returnSmpl (emptyFloats env, extendIdSubst env old_bndr (DoneEx new_rhs))
611 -- Use the substitution to make quite, quite sure that the substitution
612 -- will happen, since we are going to discard the binding
617 new_bndr_info = idInfo new_bndr `setArityInfo` exprArity new_rhs
619 -- Add the unfolding *only* for non-loop-breakers
620 -- Making loop breakers not have an unfolding at all
621 -- means that we can avoid tests in exprIsConApp, for example.
622 -- This is important: if exprIsConApp says 'yes' for a recursive
623 -- thing, then we can get into an infinite loop
625 -- If the unfolding is a value, the demand info may
626 -- go pear-shaped, so we nuke it. Example:
628 -- case x of (p,q) -> h p q x
629 -- Here x is certainly demanded. But after we've nuked
630 -- the case, we'll get just
631 -- let x = (a,b) in h a b x
632 -- and now x is not demanded (I'm assuming h is lazy)
633 -- This really happens. Similarly
634 -- let f = \x -> e in ...f..f...
635 -- After inling f at some of its call sites the original binding may
636 -- (for example) be no longer strictly demanded.
637 -- The solution here is a bit ad hoc...
638 unfolding = mkUnfolding (isTopLevel top_lvl) new_rhs
639 info_w_unf = new_bndr_info `setUnfoldingInfo` unfolding
640 final_info | loop_breaker = new_bndr_info
641 | isEvaldUnfolding unfolding = zapDemandInfo info_w_unf `orElse` info_w_unf
642 | otherwise = info_w_unf
644 final_id = new_bndr `setIdInfo` final_info
646 -- These seqs forces the Id, and hence its IdInfo,
647 -- and hence any inner substitutions
649 returnSmpl (unitFloat env final_id new_rhs, env)
652 loop_breaker = isLoopBreaker occ_info
653 old_info = idInfo old_bndr
654 occ_info = occInfo old_info
659 %************************************************************************
661 \subsection[Simplify-simplExpr]{The main function: simplExpr}
663 %************************************************************************
665 The reason for this OutExprStuff stuff is that we want to float *after*
666 simplifying a RHS, not before. If we do so naively we get quadratic
667 behaviour as things float out.
669 To see why it's important to do it after, consider this (real) example:
683 a -- Can't inline a this round, cos it appears twice
687 Each of the ==> steps is a round of simplification. We'd save a
688 whole round if we float first. This can cascade. Consider
693 let f = let d1 = ..d.. in \y -> e
697 in \x -> ...(\y ->e)...
699 Only in this second round can the \y be applied, and it
700 might do the same again.
704 simplExpr :: SimplEnv -> CoreExpr -> SimplM CoreExpr
705 simplExpr env expr = simplExprC env expr (mkBoringStop expr_ty')
707 expr_ty' = substTy (getTvSubst env) (exprType expr)
708 -- The type in the Stop continuation, expr_ty', is usually not used
709 -- It's only needed when discarding continuations after finding
710 -- a function that returns bottom.
711 -- Hence the lazy substitution
714 simplExprC :: SimplEnv -> CoreExpr -> SimplCont -> SimplM CoreExpr
715 -- Simplify an expression, given a continuation
716 simplExprC env expr cont
717 = simplExprF env expr cont `thenSmpl` \ (floats, expr) ->
718 returnSmpl (wrapFloats floats expr)
720 simplExprF :: SimplEnv -> InExpr -> SimplCont -> SimplM FloatsWithExpr
721 -- Simplify an expression, returning floated binds
723 simplExprF env (Var v) cont = simplVar env v cont
724 simplExprF env (Lit lit) cont = rebuild env (Lit lit) cont
725 simplExprF env expr@(Lam _ _) cont = simplLam env expr cont
726 simplExprF env (Note note expr) cont = simplNote env note expr cont
727 simplExprF env (App fun arg) cont = simplExprF env fun (ApplyTo NoDup arg env cont)
729 simplExprF env (Type ty) cont
730 = ASSERT( contIsRhsOrArg cont )
731 simplType env ty `thenSmpl` \ ty' ->
732 rebuild env (Type ty') cont
734 simplExprF env (Case scrut bndr case_ty alts) cont
735 | not (switchIsOn (getSwitchChecker env) NoCaseOfCase)
736 = -- Simplify the scrutinee with a Select continuation
737 simplExprF env scrut (Select NoDup bndr alts env cont)
740 = -- If case-of-case is off, simply simplify the case expression
741 -- in a vanilla Stop context, and rebuild the result around it
742 simplExprC env scrut case_cont `thenSmpl` \ case_expr' ->
743 rebuild env case_expr' cont
745 case_cont = Select NoDup bndr alts env (mkBoringStop case_ty')
746 case_ty' = substTy (getTvSubst env) case_ty -- c.f. defn of simplExpr
748 simplExprF env (Let (Rec pairs) body) cont
749 = simplRecBndrs env (map fst pairs) `thenSmpl` \ (env, bndrs') ->
750 -- NB: bndrs' don't have unfoldings or rules
751 -- We add them as we go down
753 simplRecBind env NotTopLevel pairs bndrs' `thenSmpl` \ (floats, env) ->
754 addFloats env floats $ \ env ->
755 simplExprF env body cont
757 -- A non-recursive let is dealt with by simplNonRecBind
758 simplExprF env (Let (NonRec bndr rhs) body) cont
759 = simplNonRecBind env bndr rhs env (contResultType cont) $ \ env ->
760 simplExprF env body cont
763 ---------------------------------
764 simplType :: SimplEnv -> InType -> SimplM OutType
765 -- Kept monadic just so we can do the seqType
767 = seqType new_ty `seq` returnSmpl new_ty
769 new_ty = substTy (getTvSubst env) ty
773 %************************************************************************
777 %************************************************************************
780 simplLam env fun cont
783 zap_it = mkLamBndrZapper fun (countArgs cont)
784 cont_ty = contResultType cont
786 -- Type-beta reduction
787 go env (Lam bndr body) (ApplyTo _ (Type ty_arg) arg_se body_cont)
788 = ASSERT( isTyVar bndr )
789 tick (BetaReduction bndr) `thenSmpl_`
790 simplType (setInScope arg_se env) ty_arg `thenSmpl` \ ty_arg' ->
791 go (extendTvSubst env bndr ty_arg') body body_cont
793 -- Ordinary beta reduction
794 go env (Lam bndr body) cont@(ApplyTo _ arg arg_se body_cont)
795 = tick (BetaReduction bndr) `thenSmpl_`
796 simplNonRecBind env (zap_it bndr) arg arg_se cont_ty $ \ env ->
797 go env body body_cont
799 -- Not enough args, so there are real lambdas left to put in the result
800 go env lam@(Lam _ _) cont
801 = simplLamBndrs env bndrs `thenSmpl` \ (env, bndrs') ->
802 simplExpr env body `thenSmpl` \ body' ->
803 mkLam env bndrs' body' cont `thenSmpl` \ (floats, new_lam) ->
804 addFloats env floats $ \ env ->
805 rebuild env new_lam cont
807 (bndrs,body) = collectBinders lam
809 -- Exactly enough args
810 go env expr cont = simplExprF env expr cont
812 mkLamBndrZapper :: CoreExpr -- Function
813 -> Int -- Number of args supplied, *including* type args
814 -> Id -> Id -- Use this to zap the binders
815 mkLamBndrZapper fun n_args
816 | n_args >= n_params fun = \b -> b -- Enough args
817 | otherwise = \b -> zapLamIdInfo b
819 -- NB: we count all the args incl type args
820 -- so we must count all the binders (incl type lambdas)
821 n_params (Note _ e) = n_params e
822 n_params (Lam b e) = 1 + n_params e
823 n_params other = 0::Int
827 %************************************************************************
831 %************************************************************************
834 simplNote env (Coerce to from) body cont
836 addCoerce s1 k1 (CoerceIt t1 cont)
837 -- coerce T1 S1 (coerce S1 K1 e)
840 -- coerce T1 K1 e, otherwise
842 -- For example, in the initial form of a worker
843 -- we may find (coerce T (coerce S (\x.e))) y
844 -- and we'd like it to simplify to e[y/x] in one round
846 | t1 `coreEqType` k1 = cont -- The coerces cancel out
847 | otherwise = CoerceIt t1 cont -- They don't cancel, but
848 -- the inner one is redundant
850 addCoerce t1t2 s1s2 (ApplyTo dup arg arg_se cont)
851 | not (isTypeArg arg), -- This whole case only works for value args
852 -- Could upgrade to have equiv thing for type apps too
853 Just (s1, s2) <- splitFunTy_maybe s1s2
854 -- (coerce (T1->T2) (S1->S2) F) E
856 -- coerce T2 S2 (F (coerce S1 T1 E))
858 -- t1t2 must be a function type, T1->T2, because it's applied to something
859 -- but s1s2 might conceivably not be
861 -- When we build the ApplyTo we can't mix the out-types
862 -- with the InExpr in the argument, so we simply substitute
863 -- to make it all consistent. It's a bit messy.
864 -- But it isn't a common case.
866 (t1,t2) = splitFunTy t1t2
867 new_arg = mkCoerce2 s1 t1 (substExpr subst arg)
868 subst = getSubst (setInScope arg_se env)
870 ApplyTo dup new_arg (zapSubstEnv env) (addCoerce t2 s2 cont)
872 addCoerce to' _ cont = CoerceIt to' cont
874 simplType env to `thenSmpl` \ to' ->
875 simplType env from `thenSmpl` \ from' ->
876 simplExprF env body (addCoerce to' from' cont)
879 -- Hack: we only distinguish subsumed cost centre stacks for the purposes of
880 -- inlining. All other CCCSs are mapped to currentCCS.
881 simplNote env (SCC cc) e cont
882 = simplExpr (setEnclosingCC env currentCCS) e `thenSmpl` \ e' ->
883 rebuild env (mkSCC cc e') cont
885 simplNote env InlineCall e cont
886 = simplExprF env e (InlinePlease cont)
888 -- See notes with SimplMonad.inlineMode
889 simplNote env InlineMe e cont
890 | contIsRhsOrArg cont -- Totally boring continuation; see notes above
891 = -- Don't inline inside an INLINE expression
892 simplExpr (setMode inlineMode env ) e `thenSmpl` \ e' ->
893 rebuild env (mkInlineMe e') cont
895 | otherwise -- Dissolve the InlineMe note if there's
896 -- an interesting context of any kind to combine with
897 -- (even a type application -- anything except Stop)
898 = simplExprF env e cont
900 simplNote env (CoreNote s) e cont
901 = simplExpr env e `thenSmpl` \ e' ->
902 rebuild env (Note (CoreNote s) e') cont
906 %************************************************************************
908 \subsection{Dealing with calls}
910 %************************************************************************
913 simplVar env var cont
914 = case substId (getSubst env) var of
915 DoneEx e -> simplExprF (zapSubstEnv env) e cont
916 ContEx se e -> simplExprF (setSubstEnv env se) e cont
917 DoneId var1 occ -> completeCall (zapSubstEnv env) var1 occ cont
918 -- Note [zapSubstEnv]
919 -- The template is already simplified, so don't re-substitute.
920 -- This is VITAL. Consider
922 -- let y = \z -> ...x... in
924 -- We'll clone the inner \x, adding x->x' in the id_subst
925 -- Then when we inline y, we must *not* replace x by x' in
926 -- the inlined copy!!
928 ---------------------------------------------------------
929 -- Dealing with a call site
931 completeCall env var occ_info cont
932 = -- Simplify the arguments
933 getDOptsSmpl `thenSmpl` \ dflags ->
935 chkr = getSwitchChecker env
936 (args, call_cont, inline_call) = getContArgs chkr var cont
939 simplifyArgs env fn_ty args (contResultType call_cont) $ \ env args ->
941 -- Next, look for rules or specialisations that match
943 -- It's important to simplify the args first, because the rule-matcher
944 -- doesn't do substitution as it goes. We don't want to use subst_args
945 -- (defined in the 'where') because that throws away useful occurrence info,
946 -- and perhaps-very-important specialisations.
948 -- Some functions have specialisations *and* are strict; in this case,
949 -- we don't want to inline the wrapper of the non-specialised thing; better
950 -- to call the specialised thing instead.
951 -- We used to use the black-listing mechanism to ensure that inlining of
952 -- the wrapper didn't occur for things that have specialisations till a
953 -- later phase, so but now we just try RULES first
955 -- You might think that we shouldn't apply rules for a loop breaker:
956 -- doing so might give rise to an infinite loop, because a RULE is
957 -- rather like an extra equation for the function:
958 -- RULE: f (g x) y = x+y
961 -- But it's too drastic to disable rules for loop breakers.
962 -- Even the foldr/build rule would be disabled, because foldr
963 -- is recursive, and hence a loop breaker:
964 -- foldr k z (build g) = g k z
965 -- So it's up to the programmer: rules can cause divergence
968 in_scope = getInScope env
969 maybe_rule = case activeRule env of
970 Nothing -> Nothing -- No rules apply
971 Just act_fn -> lookupRule act_fn in_scope var args
974 Just (rule_name, rule_rhs) ->
975 tick (RuleFired rule_name) `thenSmpl_`
976 (if dopt Opt_D_dump_inlinings dflags then
977 pprTrace "Rule fired" (vcat [
978 text "Rule:" <+> ftext rule_name,
979 text "Before:" <+> ppr var <+> sep (map pprParendExpr args),
980 text "After: " <+> pprCoreExpr rule_rhs,
981 text "Cont: " <+> ppr call_cont])
984 simplExprF env rule_rhs call_cont ;
986 Nothing -> -- No rules
988 -- Next, look for an inlining
990 arg_infos = [ interestingArg arg | arg <- args, isValArg arg]
992 interesting_cont = interestingCallContext (notNull args)
996 active_inline = activeInline env var occ_info
997 maybe_inline = callSiteInline dflags active_inline inline_call occ_info
998 var arg_infos interesting_cont
1000 case maybe_inline of {
1001 Just unfolding -- There is an inlining!
1002 -> tick (UnfoldingDone var) `thenSmpl_`
1003 makeThatCall env var unfolding args call_cont
1006 Nothing -> -- No inlining!
1009 rebuild env (mkApps (Var var) args) call_cont
1012 makeThatCall :: SimplEnv
1014 -> InExpr -- Inlined function rhs
1015 -> [OutExpr] -- Arguments, already simplified
1016 -> SimplCont -- After the call
1017 -> SimplM FloatsWithExpr
1018 -- Similar to simplLam, but this time
1019 -- the arguments are already simplified
1020 makeThatCall orig_env var fun@(Lam _ _) args cont
1021 = go orig_env fun args
1023 zap_it = mkLamBndrZapper fun (length args)
1025 -- Type-beta reduction
1026 go env (Lam bndr body) (Type ty_arg : args)
1027 = ASSERT( isTyVar bndr )
1028 tick (BetaReduction bndr) `thenSmpl_`
1029 go (extendTvSubst env bndr ty_arg) body args
1031 -- Ordinary beta reduction
1032 go env (Lam bndr body) (arg : args)
1033 = tick (BetaReduction bndr) `thenSmpl_`
1034 simplNonRecX env (zap_it bndr) arg $ \ env ->
1037 -- Not enough args, so there are real lambdas left to put in the result
1039 = simplExprF env fun (pushContArgs orig_env args cont)
1040 -- NB: orig_env; the correct environment to capture with
1041 -- the arguments.... env has been augmented with substitutions
1042 -- from the beta reductions.
1044 makeThatCall env var fun args cont
1045 = simplExprF env fun (pushContArgs env args cont)
1049 %************************************************************************
1051 \subsection{Arguments}
1053 %************************************************************************
1056 ---------------------------------------------------------
1057 -- Simplifying the arguments of a call
1059 simplifyArgs :: SimplEnv
1060 -> OutType -- Type of the function
1061 -> [(InExpr, SimplEnv, Bool)] -- Details of the arguments
1062 -> OutType -- Type of the continuation
1063 -> (SimplEnv -> [OutExpr] -> SimplM FloatsWithExpr)
1064 -> SimplM FloatsWithExpr
1066 -- [CPS-like because of strict arguments]
1068 -- Simplify the arguments to a call.
1069 -- This part of the simplifier may break the no-shadowing invariant
1071 -- f (...(\a -> e)...) (case y of (a,b) -> e')
1072 -- where f is strict in its second arg
1073 -- If we simplify the innermost one first we get (...(\a -> e)...)
1074 -- Simplifying the second arg makes us float the case out, so we end up with
1075 -- case y of (a,b) -> f (...(\a -> e)...) e'
1076 -- So the output does not have the no-shadowing invariant. However, there is
1077 -- no danger of getting name-capture, because when the first arg was simplified
1078 -- we used an in-scope set that at least mentioned all the variables free in its
1079 -- static environment, and that is enough.
1081 -- We can't just do innermost first, or we'd end up with a dual problem:
1082 -- case x of (a,b) -> f e (...(\a -> e')...)
1084 -- I spent hours trying to recover the no-shadowing invariant, but I just could
1085 -- not think of an elegant way to do it. The simplifier is already knee-deep in
1086 -- continuations. We have to keep the right in-scope set around; AND we have
1087 -- to get the effect that finding (error "foo") in a strict arg position will
1088 -- discard the entire application and replace it with (error "foo"). Getting
1089 -- all this at once is TOO HARD!
1091 simplifyArgs env fn_ty args cont_ty thing_inside
1092 = go env fn_ty args thing_inside
1094 go env fn_ty [] thing_inside = thing_inside env []
1095 go env fn_ty (arg:args) thing_inside = simplifyArg env fn_ty arg cont_ty $ \ env arg' ->
1096 go env (applyTypeToArg fn_ty arg') args $ \ env args' ->
1097 thing_inside env (arg':args')
1099 simplifyArg env fn_ty (Type ty_arg, se, _) cont_ty thing_inside
1100 = simplType (setInScope se env) ty_arg `thenSmpl` \ new_ty_arg ->
1101 thing_inside env (Type new_ty_arg)
1103 simplifyArg env fn_ty (val_arg, arg_se, is_strict) cont_ty thing_inside
1105 = simplStrictArg AnArg env val_arg arg_se arg_ty cont_ty thing_inside
1107 | otherwise -- Lazy argument
1108 -- DO NOT float anything outside, hence simplExprC
1109 -- There is no benefit (unlike in a let-binding), and we'd
1110 -- have to be very careful about bogus strictness through
1111 -- floating a demanded let.
1112 = simplExprC (setInScope arg_se env) val_arg
1113 (mkBoringStop arg_ty) `thenSmpl` \ arg1 ->
1114 thing_inside env arg1
1116 arg_ty = funArgTy fn_ty
1119 simplStrictArg :: LetRhsFlag
1120 -> SimplEnv -- The env of the call
1121 -> InExpr -> SimplEnv -- The arg plus its env
1122 -> OutType -- arg_ty: type of the argument
1123 -> OutType -- cont_ty: Type of thing computed by the context
1124 -> (SimplEnv -> OutExpr -> SimplM FloatsWithExpr)
1125 -- Takes an expression of type rhs_ty,
1126 -- returns an expression of type cont_ty
1127 -- The env passed to this continuation is the
1128 -- env of the call, plus any new in-scope variables
1129 -> SimplM FloatsWithExpr -- An expression of type cont_ty
1131 simplStrictArg is_rhs call_env arg arg_env arg_ty cont_ty thing_inside
1132 = simplExprF (setInScope arg_env call_env) arg
1133 (ArgOf is_rhs arg_ty cont_ty (\ new_env -> thing_inside (setInScope call_env new_env)))
1134 -- Notice the way we use arg_env (augmented with in-scope vars from call_env)
1135 -- to simplify the argument
1136 -- and call-env (augmented with in-scope vars from the arg) to pass to the continuation
1140 %************************************************************************
1142 \subsection{mkAtomicArgs}
1144 %************************************************************************
1146 mkAtomicArgs takes a putative RHS, checks whether it's a PAP or
1147 constructor application and, if so, converts it to ANF, so that the
1148 resulting thing can be inlined more easily. Thus
1155 There are three sorts of binding context, specified by the two
1161 N N Top-level or recursive Only bind args of lifted type
1163 N Y Non-top-level and non-recursive, Bind args of lifted type, or
1164 but lazy unlifted-and-ok-for-speculation
1166 Y Y Non-top-level, non-recursive, Bind all args
1167 and strict (demanded)
1174 there is no point in transforming to
1176 x = case (y div# z) of r -> MkC r
1178 because the (y div# z) can't float out of the let. But if it was
1179 a *strict* let, then it would be a good thing to do. Hence the
1180 context information.
1183 mkAtomicArgs :: Bool -- A strict binding
1184 -> Bool -- OK to float unlifted args
1186 -> SimplM (OrdList (OutId,OutExpr), -- The floats (unusually) may include
1187 OutExpr) -- things that need case-binding,
1188 -- if the strict-binding flag is on
1190 mkAtomicArgs is_strict ok_float_unlifted rhs
1191 | (Var fun, args) <- collectArgs rhs, -- It's an application
1192 isDataConWorkId fun || valArgCount args < idArity fun -- And it's a constructor or PAP
1193 = go fun nilOL [] args -- Have a go
1195 | otherwise = bale_out -- Give up
1198 bale_out = returnSmpl (nilOL, rhs)
1200 go fun binds rev_args []
1201 = returnSmpl (binds, mkApps (Var fun) (reverse rev_args))
1203 go fun binds rev_args (arg : args)
1204 | exprIsTrivial arg -- Easy case
1205 = go fun binds (arg:rev_args) args
1207 | not can_float_arg -- Can't make this arg atomic
1208 = bale_out -- ... so give up
1210 | otherwise -- Don't forget to do it recursively
1211 -- E.g. x = a:b:c:[]
1212 = mkAtomicArgs is_strict ok_float_unlifted arg `thenSmpl` \ (arg_binds, arg') ->
1213 newId FSLIT("a") arg_ty `thenSmpl` \ arg_id ->
1214 go fun ((arg_binds `snocOL` (arg_id,arg')) `appOL` binds)
1215 (Var arg_id : rev_args) args
1217 arg_ty = exprType arg
1218 can_float_arg = is_strict
1219 || not (isUnLiftedType arg_ty)
1220 || (ok_float_unlifted && exprOkForSpeculation arg)
1223 addAtomicBinds :: SimplEnv -> [(OutId,OutExpr)]
1224 -> (SimplEnv -> SimplM (FloatsWith a))
1225 -> SimplM (FloatsWith a)
1226 addAtomicBinds env [] thing_inside = thing_inside env
1227 addAtomicBinds env ((v,r):bs) thing_inside = addAuxiliaryBind env (NonRec v r) $ \ env ->
1228 addAtomicBinds env bs thing_inside
1230 addAtomicBindsE :: SimplEnv -> [(OutId,OutExpr)]
1231 -> (SimplEnv -> SimplM FloatsWithExpr)
1232 -> SimplM FloatsWithExpr
1233 -- Same again, but this time we're in an expression context,
1234 -- and may need to do some case bindings
1236 addAtomicBindsE env [] thing_inside
1238 addAtomicBindsE env ((v,r):bs) thing_inside
1239 | needsCaseBinding (idType v) r
1240 = addAtomicBindsE (addNewInScopeIds env [v]) bs thing_inside `thenSmpl` \ (floats, expr) ->
1241 WARN( exprIsTrivial expr, ppr v <+> pprCoreExpr expr )
1242 (let body = wrapFloats floats expr in
1243 returnSmpl (emptyFloats env, Case r v (exprType body) [(DEFAULT,[],body)]))
1246 = addAuxiliaryBind env (NonRec v r) $ \ env ->
1247 addAtomicBindsE env bs thing_inside
1251 %************************************************************************
1253 \subsection{The main rebuilder}
1255 %************************************************************************
1258 rebuild :: SimplEnv -> OutExpr -> SimplCont -> SimplM FloatsWithExpr
1260 rebuild env expr (Stop _ _ _) = rebuildDone env expr
1261 rebuild env expr (ArgOf _ _ _ cont_fn) = cont_fn env expr
1262 rebuild env expr (CoerceIt to_ty cont) = rebuild env (mkCoerce to_ty expr) cont
1263 rebuild env expr (InlinePlease cont) = rebuild env (Note InlineCall expr) cont
1264 rebuild env expr (Select _ bndr alts se cont) = rebuildCase (setInScope se env) expr bndr alts cont
1265 rebuild env expr (ApplyTo _ arg se cont) = rebuildApp (setInScope se env) expr arg cont
1267 rebuildApp env fun arg cont
1268 = simplExpr env arg `thenSmpl` \ arg' ->
1269 rebuild env (App fun arg') cont
1271 rebuildDone env expr = returnSmpl (emptyFloats env, expr)
1275 %************************************************************************
1277 \subsection{Functions dealing with a case}
1279 %************************************************************************
1281 Blob of helper functions for the "case-of-something-else" situation.
1284 ---------------------------------------------------------
1285 -- Eliminate the case if possible
1287 rebuildCase :: SimplEnv
1288 -> OutExpr -- Scrutinee
1289 -> InId -- Case binder
1290 -> [InAlt] -- Alternatives (inceasing order)
1292 -> SimplM FloatsWithExpr
1294 rebuildCase env scrut case_bndr alts cont
1295 | Just (con,args) <- exprIsConApp_maybe scrut
1296 -- Works when the scrutinee is a variable with a known unfolding
1297 -- as well as when it's an explicit constructor application
1298 = knownCon env (DataAlt con) args case_bndr alts cont
1300 | Lit lit <- scrut -- No need for same treatment as constructors
1301 -- because literals are inlined more vigorously
1302 = knownCon env (LitAlt lit) [] case_bndr alts cont
1305 = prepareAlts scrut case_bndr alts `thenSmpl` \ (better_alts, handled_cons) ->
1307 -- Deal with the case binder, and prepare the continuation;
1308 -- The new subst_env is in place
1309 prepareCaseCont env better_alts cont `thenSmpl` \ (floats, (dup_cont, nondup_cont)) ->
1310 addFloats env floats $ \ env ->
1313 -- The case expression is annotated with the result type of the continuation
1314 -- This may differ from the type originally on the case. For example
1315 -- case(T) (case(Int#) a of { True -> 1#; False -> 0# }) of
1318 -- let j a# = <blob>
1319 -- in case(T) a of { True -> j 1#; False -> j 0# }
1320 -- Note that the case that scrutinises a now returns a T not an Int#
1321 res_ty' = contResultType dup_cont
1324 -- Deal with variable scrutinee
1325 simplCaseBinder env scrut case_bndr `thenSmpl` \ (alt_env, case_bndr') ->
1327 -- Deal with the case alternatives
1328 simplAlts alt_env handled_cons
1329 case_bndr' better_alts dup_cont `thenSmpl` \ alts' ->
1331 -- Put the case back together
1332 mkCase scrut case_bndr' res_ty' alts' `thenSmpl` \ case_expr ->
1334 -- Notice that rebuildDone returns the in-scope set from env, not alt_env
1335 -- The case binder *not* scope over the whole returned case-expression
1336 rebuild env case_expr nondup_cont
1339 simplCaseBinder checks whether the scrutinee is a variable, v. If so,
1340 try to eliminate uses of v in the RHSs in favour of case_bndr; that
1341 way, there's a chance that v will now only be used once, and hence
1346 There is a time we *don't* want to do that, namely when
1347 -fno-case-of-case is on. This happens in the first simplifier pass,
1348 and enhances full laziness. Here's the bad case:
1349 f = \ y -> ...(case x of I# v -> ...(case x of ...) ... )
1350 If we eliminate the inner case, we trap it inside the I# v -> arm,
1351 which might prevent some full laziness happening. I've seen this
1352 in action in spectral/cichelli/Prog.hs:
1353 [(m,n) | m <- [1..max], n <- [1..max]]
1354 Hence the check for NoCaseOfCase.
1358 There is another situation when we don't want to do it. If we have
1360 case x of w1 { DEFAULT -> case x of w2 { A -> e1; B -> e2 }
1361 ...other cases .... }
1363 We'll perform the binder-swap for the outer case, giving
1365 case x of w1 { DEFAULT -> case w1 of w2 { A -> e1; B -> e2 }
1366 ...other cases .... }
1368 But there is no point in doing it for the inner case, because w1 can't
1369 be inlined anyway. Furthermore, doing the case-swapping involves
1370 zapping w2's occurrence info (see paragraphs that follow), and that
1371 forces us to bind w2 when doing case merging. So we get
1373 case x of w1 { A -> let w2 = w1 in e1
1374 B -> let w2 = w1 in e2
1375 ...other cases .... }
1377 This is plain silly in the common case where w2 is dead.
1379 Even so, I can't see a good way to implement this idea. I tried
1380 not doing the binder-swap if the scrutinee was already evaluated
1381 but that failed big-time:
1385 case v of w { MkT x ->
1386 case x of x1 { I# y1 ->
1387 case x of x2 { I# y2 -> ...
1389 Notice that because MkT is strict, x is marked "evaluated". But to
1390 eliminate the last case, we must either make sure that x (as well as
1391 x1) has unfolding MkT y1. THe straightforward thing to do is to do
1392 the binder-swap. So this whole note is a no-op.
1396 If we replace the scrutinee, v, by tbe case binder, then we have to nuke
1397 any occurrence info (eg IAmDead) in the case binder, because the
1398 case-binder now effectively occurs whenever v does. AND we have to do
1399 the same for the pattern-bound variables! Example:
1401 (case x of { (a,b) -> a }) (case x of { (p,q) -> q })
1403 Here, b and p are dead. But when we move the argment inside the first
1404 case RHS, and eliminate the second case, we get
1406 case x of { (a,b) -> a b }
1408 Urk! b is alive! Reason: the scrutinee was a variable, and case elimination
1411 Indeed, this can happen anytime the case binder isn't dead:
1412 case <any> of x { (a,b) ->
1413 case x of { (p,q) -> p } }
1414 Here (a,b) both look dead, but come alive after the inner case is eliminated.
1415 The point is that we bring into the envt a binding
1417 after the outer case, and that makes (a,b) alive. At least we do unless
1418 the case binder is guaranteed dead.
1421 simplCaseBinder env (Var v) case_bndr
1422 | not (switchIsOn (getSwitchChecker env) NoCaseOfCase)
1424 -- Failed try [see Note 2 above]
1425 -- not (isEvaldUnfolding (idUnfolding v))
1427 = simplBinder env (zap case_bndr) `thenSmpl` \ (env, case_bndr') ->
1428 returnSmpl (modifyInScope env v case_bndr', case_bndr')
1429 -- We could extend the substitution instead, but it would be
1430 -- a hack because then the substitution wouldn't be idempotent
1431 -- any more (v is an OutId). And this does just as well.
1433 zap b = b `setIdOccInfo` NoOccInfo
1435 simplCaseBinder env other_scrut case_bndr
1436 = simplBinder env case_bndr `thenSmpl` \ (env, case_bndr') ->
1437 returnSmpl (env, case_bndr')
1443 simplAlts :: SimplEnv
1444 -> [AltCon] -- Alternatives the scrutinee can't be
1445 -- in the default case
1446 -> OutId -- Case binder
1447 -> [InAlt] -> SimplCont
1448 -> SimplM [OutAlt] -- Includes the continuation
1450 simplAlts env handled_cons case_bndr' alts cont'
1451 = mapSmpl simpl_alt alts
1453 simpl_alt alt = simplAlt env handled_cons case_bndr' alt cont' `thenSmpl` \ (_, alt') ->
1456 simplAlt :: SimplEnv -> [AltCon] -> OutId -> InAlt -> SimplCont
1457 -> SimplM (Maybe TvSubstEnv, OutAlt)
1458 -- Simplify an alternative, returning the type refinement for the
1459 -- alternative, if the alternative does any refinement at all
1461 simplAlt env handled_cons case_bndr' (DEFAULT, bndrs, rhs) cont'
1462 = ASSERT( null bndrs )
1463 simplExprC env' rhs cont' `thenSmpl` \ rhs' ->
1464 returnSmpl (Nothing, (DEFAULT, [], rhs'))
1466 env' = mk_rhs_env env case_bndr' (mkOtherCon handled_cons)
1467 -- Record the constructors that the case-binder *can't* be.
1469 simplAlt env handled_cons case_bndr' (LitAlt lit, bndrs, rhs) cont'
1470 = ASSERT( null bndrs )
1471 simplExprC env' rhs cont' `thenSmpl` \ rhs' ->
1472 returnSmpl (Nothing, (LitAlt lit, [], rhs'))
1474 env' = mk_rhs_env env case_bndr' (mkUnfolding False (Lit lit))
1476 simplAlt env handled_cons case_bndr' (DataAlt con, vs, rhs) cont'
1477 | isVanillaDataCon con
1478 = -- Deal with the pattern-bound variables
1479 -- Mark the ones that are in ! positions in the data constructor
1480 -- as certainly-evaluated.
1481 -- NB: it happens that simplBinders does *not* erase the OtherCon
1482 -- form of unfolding, so it's ok to add this info before
1483 -- doing simplBinders
1484 simplBinders env (add_evals con vs) `thenSmpl` \ (env, vs') ->
1486 -- Bind the case-binder to (con args)
1487 let unf = mkUnfolding False (mkConApp con con_args)
1488 inst_tys' = tyConAppArgs (idType case_bndr')
1489 con_args = map Type inst_tys' ++ map varToCoreExpr vs'
1490 env' = mk_rhs_env env case_bndr' unf
1492 simplExprC env' rhs cont' `thenSmpl` \ rhs' ->
1493 returnSmpl (Nothing, (DataAlt con, vs', rhs'))
1495 | otherwise -- GADT case
1497 (tvs,ids) = span isTyVar vs
1499 simplBinders env tvs `thenSmpl` \ (env1, tvs') ->
1501 pat_res_ty = dataConResTy con (mkTyVarTys tvs')
1502 tv_subst = getTvSubst env1
1504 case coreRefineTys tvs' tv_subst pat_res_ty (idType case_bndr') of {
1505 Nothing -- Dead code; for now, I'm just going to put in an
1506 -- error case so I can see them
1507 -> let rhs' = mkApps (Var eRROR_ID)
1508 [Type (substTy tv_subst (exprType rhs)),
1509 Lit (mkStringLit "Impossible alternative (GADT)")]
1511 simplBinders env1 ids `thenSmpl` \ (env2, ids') ->
1512 returnSmpl (Nothing, (DataAlt con, tvs' ++ ids', rhs')) ;
1514 Just tv_subst_env -> -- The normal case
1517 env2 = setTvSubstEnv env1 tv_subst_env
1518 -- Simplify the Ids in the refined environment, so their types
1519 -- reflect the refinement. Usually this doesn't matter, but it helps
1520 -- in mkDupableAlt, when we want to float a lambda that uses these binders
1522 simplBinders env2 (add_evals con ids) `thenSmpl` \ (env3, ids') ->
1523 let unf = mkUnfolding False con_app
1524 con_app = mkConApp con con_args
1525 con_args = map varToCoreExpr vs' -- NB: no inst_tys'
1526 env_w_unf = mk_rhs_env env3 case_bndr' unf
1529 simplExprC env_w_unf rhs cont' `thenSmpl` \ rhs' ->
1530 returnSmpl (Just tv_subst_env, (DataAlt con, vs', rhs')) }
1533 -- add_evals records the evaluated-ness of the bound variables of
1534 -- a case pattern. This is *important*. Consider
1535 -- data T = T !Int !Int
1537 -- case x of { T a b -> T (a+1) b }
1539 -- We really must record that b is already evaluated so that we don't
1540 -- go and re-evaluate it when constructing the result.
1541 add_evals dc vs = cat_evals dc vs (dataConRepStrictness dc)
1543 cat_evals dc vs strs
1547 go (v:vs) strs | isTyVar v = v : go vs strs
1548 go (v:vs) (str:strs)
1549 | isMarkedStrict str = evald_v : go vs strs
1550 | otherwise = zapped_v : go vs strs
1552 zapped_v = zap_occ_info v
1553 evald_v = zapped_v `setIdUnfolding` mkOtherCon []
1554 go _ _ = pprPanic "cat_evals" (ppr dc $$ ppr vs $$ ppr strs)
1556 -- If the case binder is alive, then we add the unfolding
1558 -- to the envt; so vs are now very much alive
1559 zap_occ_info | isDeadBinder case_bndr' = \id -> id
1560 | otherwise = \id -> id `setIdOccInfo` NoOccInfo
1562 mk_rhs_env env case_bndr' case_bndr_unf
1563 = modifyInScope env case_bndr' (case_bndr' `setIdUnfolding` case_bndr_unf)
1567 %************************************************************************
1569 \subsection{Known constructor}
1571 %************************************************************************
1573 We are a bit careful with occurrence info. Here's an example
1575 (\x* -> case x of (a*, b) -> f a) (h v, e)
1577 where the * means "occurs once". This effectively becomes
1578 case (h v, e) of (a*, b) -> f a)
1580 let a* = h v; b = e in f a
1584 All this should happen in one sweep.
1587 knownCon :: SimplEnv -> AltCon -> [OutExpr]
1588 -> InId -> [InAlt] -> SimplCont
1589 -> SimplM FloatsWithExpr
1591 knownCon env con args bndr alts cont
1592 = tick (KnownBranch bndr) `thenSmpl_`
1593 case findAlt con alts of
1594 (DEFAULT, bs, rhs) -> ASSERT( null bs )
1595 simplNonRecX env bndr scrut $ \ env ->
1596 -- This might give rise to a binding with non-atomic args
1597 -- like x = Node (f x) (g x)
1598 -- but no harm will be done
1599 simplExprF env rhs cont
1602 LitAlt lit -> Lit lit
1603 DataAlt dc -> mkConApp dc args
1605 (LitAlt lit, bs, rhs) -> ASSERT( null bs )
1606 simplNonRecX env bndr (Lit lit) $ \ env ->
1607 simplExprF env rhs cont
1609 (DataAlt dc, bs, rhs)
1610 -> ASSERT( n_drop_tys + length bs == length args )
1611 bind_args env bs (drop n_drop_tys args) $ \ env ->
1613 con_app = mkConApp dc (take n_drop_tys args ++ con_args)
1614 con_args = [substExpr (getSubst env) (varToCoreExpr b) | b <- bs]
1615 -- args are aready OutExprs, but bs are InIds
1617 simplNonRecX env bndr con_app $ \ env ->
1618 simplExprF env rhs cont
1620 n_drop_tys | isVanillaDataCon dc = tyConArity (dataConTyCon dc)
1622 -- Vanilla data constructors lack type arguments in the pattern
1625 bind_args env [] _ thing_inside = thing_inside env
1627 bind_args env (b:bs) (Type ty : args) thing_inside
1628 = ASSERT( isTyVar b )
1629 bind_args (extendTvSubst env b ty) bs args thing_inside
1631 bind_args env (b:bs) (arg : args) thing_inside
1633 simplNonRecX env b arg $ \ env ->
1634 bind_args env bs args thing_inside
1638 %************************************************************************
1640 \subsection{Duplicating continuations}
1642 %************************************************************************
1645 prepareCaseCont :: SimplEnv
1646 -> [InAlt] -> SimplCont
1647 -> SimplM (FloatsWith (SimplCont,SimplCont))
1648 -- Return a duplicatable continuation, a non-duplicable part
1649 -- plus some extra bindings
1651 -- No need to make it duplicatable if there's only one alternative
1652 prepareCaseCont env [alt] cont = returnSmpl (emptyFloats env, (cont, mkBoringStop (contResultType cont)))
1653 prepareCaseCont env alts cont = mkDupableCont env cont
1657 mkDupableCont :: SimplEnv -> SimplCont
1658 -> SimplM (FloatsWith (SimplCont, SimplCont))
1660 mkDupableCont env cont
1661 | contIsDupable cont
1662 = returnSmpl (emptyFloats env, (cont, mkBoringStop (contResultType cont)))
1664 mkDupableCont env (CoerceIt ty cont)
1665 = mkDupableCont env cont `thenSmpl` \ (floats, (dup_cont, nondup_cont)) ->
1666 returnSmpl (floats, (CoerceIt ty dup_cont, nondup_cont))
1668 mkDupableCont env (InlinePlease cont)
1669 = mkDupableCont env cont `thenSmpl` \ (floats, (dup_cont, nondup_cont)) ->
1670 returnSmpl (floats, (InlinePlease dup_cont, nondup_cont))
1672 mkDupableCont env cont@(ArgOf _ arg_ty _ _)
1673 = returnSmpl (emptyFloats env, (mkBoringStop arg_ty, cont))
1674 -- Do *not* duplicate an ArgOf continuation
1675 -- Because ArgOf continuations are opaque, we gain nothing by
1676 -- propagating them into the expressions, and we do lose a lot.
1677 -- Here's an example:
1678 -- && (case x of { T -> F; F -> T }) E
1679 -- Now, && is strict so we end up simplifying the case with
1680 -- an ArgOf continuation. If we let-bind it, we get
1682 -- let $j = \v -> && v E
1683 -- in simplExpr (case x of { T -> F; F -> T })
1684 -- (ArgOf (\r -> $j r)
1685 -- And after simplifying more we get
1687 -- let $j = \v -> && v E
1688 -- in case of { T -> $j F; F -> $j T }
1689 -- Which is a Very Bad Thing
1691 -- The desire not to duplicate is the entire reason that
1692 -- mkDupableCont returns a pair of continuations.
1694 -- The original plan had:
1695 -- e.g. (...strict-fn...) [...hole...]
1697 -- let $j = \a -> ...strict-fn...
1698 -- in $j [...hole...]
1700 mkDupableCont env (ApplyTo _ arg se cont)
1701 = -- e.g. [...hole...] (...arg...)
1703 -- let a = ...arg...
1704 -- in [...hole...] a
1705 simplExpr (setInScope se env) arg `thenSmpl` \ arg' ->
1707 mkDupableCont env cont `thenSmpl` \ (floats, (dup_cont, nondup_cont)) ->
1708 addFloats env floats $ \ env ->
1710 if exprIsDupable arg' then
1711 returnSmpl (emptyFloats env, (ApplyTo OkToDup arg' (zapSubstEnv se) dup_cont, nondup_cont))
1713 newId FSLIT("a") (exprType arg') `thenSmpl` \ arg_id ->
1715 tick (CaseOfCase arg_id) `thenSmpl_`
1716 -- Want to tick here so that we go round again,
1717 -- and maybe copy or inline the code.
1718 -- Not strictly CaseOfCase, but never mind
1720 returnSmpl (unitFloat env arg_id arg',
1721 (ApplyTo OkToDup (Var arg_id) (zapSubstEnv se) dup_cont,
1723 -- But what if the arg should be case-bound?
1724 -- This has been this way for a long time, so I'll leave it,
1725 -- but I can't convince myself that it's right.
1728 mkDupableCont env (Select _ case_bndr alts se cont)
1729 = -- e.g. (case [...hole...] of { pi -> ei })
1731 -- let ji = \xij -> ei
1732 -- in case [...hole...] of { pi -> ji xij }
1733 tick (CaseOfCase case_bndr) `thenSmpl_`
1735 alt_env = setInScope se env
1737 prepareCaseCont alt_env alts cont `thenSmpl` \ (floats1, (dup_cont, nondup_cont)) ->
1738 addFloats alt_env floats1 $ \ alt_env ->
1740 simplBinder alt_env case_bndr `thenSmpl` \ (alt_env, case_bndr') ->
1741 -- NB: simplBinder does not zap deadness occ-info, so
1742 -- a dead case_bndr' will still advertise its deadness
1743 -- This is really important because in
1744 -- case e of b { (# a,b #) -> ... }
1745 -- b is always dead, and indeed we are not allowed to bind b to (# a,b #),
1746 -- which might happen if e was an explicit unboxed pair and b wasn't marked dead.
1747 -- In the new alts we build, we have the new case binder, so it must retain
1750 mkDupableAlts alt_env case_bndr' alts dup_cont `thenSmpl` \ (floats2, alts') ->
1751 addFloats alt_env floats2 $ \ alt_env ->
1752 returnSmpl (emptyFloats alt_env,
1753 (Select OkToDup case_bndr' alts' (zapSubstEnv se)
1754 (mkBoringStop (contResultType dup_cont)),
1757 mkDupableAlts :: SimplEnv -> OutId -> [InAlt] -> SimplCont
1758 -> SimplM (FloatsWith [InAlt])
1759 -- Absorbs the continuation into the new alternatives
1761 mkDupableAlts env case_bndr' alts dupable_cont
1764 go env [] = returnSmpl (emptyFloats env, [])
1766 = mkDupableAlt env case_bndr' dupable_cont alt `thenSmpl` \ (floats1, alt') ->
1767 addFloats env floats1 $ \ env ->
1768 go env alts `thenSmpl` \ (floats2, alts') ->
1769 returnSmpl (floats2, alt' : alts')
1771 mkDupableAlt env case_bndr' cont alt
1772 = simplAlt env [] case_bndr' alt cont `thenSmpl` \ (mb_reft, (con, bndrs', rhs')) ->
1773 -- Safe to say that there are no handled-cons for the DEFAULT case
1775 if exprIsDupable rhs' then
1776 returnSmpl (emptyFloats env, (con, bndrs', rhs'))
1777 -- It is worth checking for a small RHS because otherwise we
1778 -- get extra let bindings that may cause an extra iteration of the simplifier to
1779 -- inline back in place. Quite often the rhs is just a variable or constructor.
1780 -- The Ord instance of Maybe in PrelMaybe.lhs, for example, took several extra
1781 -- iterations because the version with the let bindings looked big, and so wasn't
1782 -- inlined, but after the join points had been inlined it looked smaller, and so
1785 -- NB: we have to check the size of rhs', not rhs.
1786 -- Duplicating a small InAlt might invalidate occurrence information
1787 -- However, if it *is* dupable, we return the *un* simplified alternative,
1788 -- because otherwise we'd need to pair it up with an empty subst-env....
1789 -- but we only have one env shared between all the alts.
1790 -- (Remember we must zap the subst-env before re-simplifying something).
1791 -- Rather than do this we simply agree to re-simplify the original (small) thing later.
1795 rhs_ty' = exprType rhs'
1796 used_bndrs' = filter abstract_over (case_bndr' : bndrs')
1798 | isTyVar bndr = not (mb_reft `refines` bndr)
1799 -- Don't abstract over tyvar binders which are refined away
1800 | otherwise = not (isDeadBinder bndr)
1801 -- The deadness info on the new Ids is preserved by simplBinders
1802 refines Nothing bndr = False
1803 refines (Just tv_subst) bndr = bndr `elemVarEnv` tv_subst
1804 -- See Note [Refinement] below
1806 -- If we try to lift a primitive-typed something out
1807 -- for let-binding-purposes, we will *caseify* it (!),
1808 -- with potentially-disastrous strictness results. So
1809 -- instead we turn it into a function: \v -> e
1810 -- where v::State# RealWorld#. The value passed to this function
1811 -- is realworld#, which generates (almost) no code.
1813 -- There's a slight infelicity here: we pass the overall
1814 -- case_bndr to all the join points if it's used in *any* RHS,
1815 -- because we don't know its usage in each RHS separately
1817 -- We used to say "&& isUnLiftedType rhs_ty'" here, but now
1818 -- we make the join point into a function whenever used_bndrs'
1819 -- is empty. This makes the join-point more CPR friendly.
1820 -- Consider: let j = if .. then I# 3 else I# 4
1821 -- in case .. of { A -> j; B -> j; C -> ... }
1823 -- Now CPR doesn't w/w j because it's a thunk, so
1824 -- that means that the enclosing function can't w/w either,
1825 -- which is a lose. Here's the example that happened in practice:
1826 -- kgmod :: Int -> Int -> Int
1827 -- kgmod x y = if x > 0 && y < 0 || x < 0 && y > 0
1831 -- I have seen a case alternative like this:
1832 -- True -> \v -> ...
1833 -- It's a bit silly to add the realWorld dummy arg in this case, making
1836 -- (the \v alone is enough to make CPR happy) but I think it's rare
1838 ( if not (any isId used_bndrs')
1839 then newId FSLIT("w") realWorldStatePrimTy `thenSmpl` \ rw_id ->
1840 returnSmpl ([rw_id], [Var realWorldPrimId])
1842 returnSmpl (used_bndrs', map varToCoreExpr used_bndrs')
1843 ) `thenSmpl` \ (final_bndrs', final_args) ->
1845 -- See comment about "$j" name above
1846 newId (encodeFS FSLIT("$j")) (mkPiTypes final_bndrs' rhs_ty') `thenSmpl` \ join_bndr ->
1847 -- Notice the funky mkPiTypes. If the contructor has existentials
1848 -- it's possible that the join point will be abstracted over
1849 -- type varaibles as well as term variables.
1850 -- Example: Suppose we have
1851 -- data T = forall t. C [t]
1853 -- case (case e of ...) of
1854 -- C t xs::[t] -> rhs
1855 -- We get the join point
1856 -- let j :: forall t. [t] -> ...
1857 -- j = /\t \xs::[t] -> rhs
1859 -- case (case e of ...) of
1860 -- C t xs::[t] -> j t xs
1862 -- We make the lambdas into one-shot-lambdas. The
1863 -- join point is sure to be applied at most once, and doing so
1864 -- prevents the body of the join point being floated out by
1865 -- the full laziness pass
1866 really_final_bndrs = map one_shot final_bndrs'
1867 one_shot v | isId v = setOneShotLambda v
1869 join_rhs = mkLams really_final_bndrs rhs'
1870 join_call = mkApps (Var join_bndr) final_args
1872 returnSmpl (unitFloat env join_bndr join_rhs, (con, bndrs', join_call))
1879 MkT :: a -> b -> T a
1883 MkT a' b (p::a') (q::b) -> [p,w]
1885 The danger is that we'll make a join point
1889 and that's ill-typed, because (p::a') but (w::a).
1891 Solution so far: don't abstract over a', because the type refinement
1892 maps [a' -> a] . Ultimately that won't work when real refinement goes on.
1894 Then we must abstract over any refined free variables. Hmm. Maybe we
1895 could just abstract over *all* free variables, thereby lambda-lifting
1896 the join point? We should try this.