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 ( intSwitchSet,
12 opt_SccProfilingOn, opt_PprStyle_Debug, opt_SimplDoEtaReduction,
13 opt_SimplNoPreInlining, opt_DictsStrict, opt_SimplPedanticBottoms,
17 import SimplUtils ( mkCase, transformRhs, findAlt,
18 simplBinder, simplBinders, simplIds, findDefault, mkCoerce
20 import Var ( TyVar, mkSysTyVar, tyVarKind, maybeModifyIdInfo )
23 import Id ( Id, idType, idInfo, idUnique,
24 getIdUnfolding, setIdUnfolding, isExportedId,
25 getIdSpecialisation, setIdSpecialisation,
26 getIdDemandInfo, setIdDemandInfo,
27 getIdArity, setIdArity, setIdInfo,
29 setInlinePragma, getInlinePragma, idMustBeINLINEd,
32 import IdInfo ( InlinePragInfo(..), OccInfo(..), StrictnessInfo(..),
33 ArityInfo(..), atLeastArity, arityLowerBound, unknownArity,
34 specInfo, inlinePragInfo, zapLamIdInfo, setArityInfo, setInlinePragInfo, setUnfoldingInfo
36 import Demand ( Demand, isStrict, wwLazy )
37 import Const ( isWHNFCon, conOkForAlt )
38 import ConFold ( tryPrimOp )
39 import PrimOp ( PrimOp, primOpStrictness, primOpType )
40 import DataCon ( DataCon, dataConNumInstArgs, dataConRepStrictness, dataConSig, dataConArgTys )
41 import Const ( Con(..) )
42 import Name ( isLocallyDefined )
44 import CoreFVs ( exprFreeVars )
45 import CoreUnfold ( Unfolding, mkOtherCon, mkUnfolding, otherCons,
46 callSiteInline, blackListed, hasSomeUnfolding
48 import CoreUtils ( cheapEqExpr, exprIsDupable, exprIsCheap, exprIsTrivial,
49 coreExprType, coreAltsType, exprArity, exprIsValue,
52 import Rules ( lookupRule )
53 import CostCentre ( isSubsumedCCS, currentCCS, isEmptyCC )
54 import Type ( Type, mkTyVarTy, mkTyVarTys, isUnLiftedType,
55 mkFunTy, splitFunTys, splitTyConApp_maybe, splitFunTy_maybe,
56 funResultTy, isDictTy, isDataType, applyTy, applyTys, mkFunTys
58 import Subst ( Subst, mkSubst, emptySubst, substExpr, substTy,
59 substEnv, lookupInScope, lookupSubst, substIdInfo
61 import TyCon ( isDataTyCon, tyConDataCons, tyConClass_maybe, tyConArity, isDataTyCon )
62 import TysPrim ( realWorldStatePrimTy )
63 import PrelInfo ( realWorldPrimId )
64 import BasicTypes ( TopLevelFlag(..), isTopLevel )
65 import Maybes ( maybeToBool )
66 import Util ( zipWithEqual, stretchZipEqual, lengthExceeds )
72 The guts of the simplifier is in this module, but the driver
73 loop for the simplifier is in SimplCore.lhs.
76 %************************************************************************
80 %************************************************************************
83 simplTopBinds :: [InBind] -> SimplM [OutBind]
86 = -- Put all the top-level binders into scope at the start
87 -- so that if a transformation rule has unexpectedly brought
88 -- anything into scope, then we don't get a complaint about that.
89 -- It's rather as if the top-level binders were imported.
90 extendInScopes top_binders $
91 simpl_binds binds `thenSmpl` \ (binds', _) ->
92 freeTick SimplifierDone `thenSmpl_`
95 top_binders = bindersOfBinds binds
97 simpl_binds [] = returnSmpl ([], panic "simplTopBinds corner")
98 simpl_binds (NonRec bndr rhs : binds) = simplLazyBind TopLevel bndr bndr rhs (simpl_binds binds)
99 simpl_binds (Rec pairs : binds) = simplRecBind TopLevel pairs (map fst pairs) (simpl_binds binds)
102 simplRecBind :: TopLevelFlag -> [(InId, InExpr)] -> [OutId]
103 -> SimplM (OutStuff a) -> SimplM (OutStuff a)
104 simplRecBind top_lvl pairs bndrs' thing_inside
105 = go pairs bndrs' `thenSmpl` \ (binds', stuff) ->
106 returnSmpl (addBind (Rec (flattenBinds binds')) stuff)
108 go [] _ = thing_inside `thenSmpl` \ stuff ->
109 returnSmpl ([], stuff)
111 go ((bndr, rhs) : pairs) (bndr' : bndrs')
112 = simplLazyBind top_lvl bndr bndr' rhs (go pairs bndrs')
113 -- Don't float unboxed bindings out,
114 -- because we can't "rec" them
118 %************************************************************************
120 \subsection[Simplify-simplExpr]{The main function: simplExpr}
122 %************************************************************************
125 addBind :: CoreBind -> OutStuff a -> OutStuff a
126 addBind bind (binds, res) = (bind:binds, res)
128 addBinds :: [CoreBind] -> OutStuff a -> OutStuff a
129 addBinds [] stuff = stuff
130 addBinds binds1 (binds2, res) = (binds1++binds2, res)
133 The reason for this OutExprStuff stuff is that we want to float *after*
134 simplifying a RHS, not before. If we do so naively we get quadratic
135 behaviour as things float out.
137 To see why it's important to do it after, consider this (real) example:
151 a -- Can't inline a this round, cos it appears twice
155 Each of the ==> steps is a round of simplification. We'd save a
156 whole round if we float first. This can cascade. Consider
161 let f = let d1 = ..d.. in \y -> e
165 in \x -> ...(\y ->e)...
167 Only in this second round can the \y be applied, and it
168 might do the same again.
172 simplExpr :: CoreExpr -> SimplM CoreExpr
173 simplExpr expr = getSubst `thenSmpl` \ subst ->
174 simplExprC expr (Stop (substTy subst (coreExprType expr)))
175 -- The type in the Stop continuation is usually not used
176 -- It's only needed when discarding continuations after finding
177 -- a function that returns bottom
179 simplExprC :: CoreExpr -> SimplCont -> SimplM CoreExpr
180 -- Simplify an expression, given a continuation
182 simplExprC expr cont = simplExprF expr cont `thenSmpl` \ (floats, (_, body)) ->
183 returnSmpl (mkLets floats body)
185 simplExprF :: InExpr -> SimplCont -> SimplM OutExprStuff
186 -- Simplify an expression, returning floated binds
188 simplExprF (Var v) cont
191 simplExprF expr@(Con (PrimOp op) args) cont
192 = getSubstEnv `thenSmpl` \ se ->
195 (primOpStrictness op)
196 (pushArgs se args cont) $ \ args1 cont1 ->
199 -- Boring... we may have too many arguments now, so we push them back
201 args2 = ASSERT( length args1 >= n_args )
203 cont2 = pushArgs emptySubstEnv (drop n_args args1) cont1
205 -- Try the prim op simplification
206 -- It's really worth trying simplExpr again if it succeeds,
207 -- because you can find
208 -- case (eqChar# x 'a') of ...
210 -- case (case x of 'a' -> True; other -> False) of ...
211 case tryPrimOp op args2 of
212 Just e' -> zapSubstEnv (simplExprF e' cont2)
213 Nothing -> rebuild (Con (PrimOp op) args2) cont2
215 simplExprF (Con con@(DataCon _) args) cont
216 = freeTick LeafVisit `thenSmpl_`
217 simplConArgs args ( \ args' ->
218 rebuild (Con con args') cont)
220 simplExprF expr@(Con con@(Literal _) args) cont
221 = ASSERT( null args )
222 freeTick LeafVisit `thenSmpl_`
225 simplExprF (App fun arg) cont
226 = getSubstEnv `thenSmpl` \ se ->
227 simplExprF fun (ApplyTo NoDup arg se cont)
229 simplExprF (Case scrut bndr alts) cont
230 = getSubstEnv `thenSmpl` \ se ->
231 simplExprF scrut (Select NoDup bndr alts se cont)
234 simplExprF (Let (Rec pairs) body) cont
235 = simplIds (map fst pairs) $ \ bndrs' ->
236 -- NB: bndrs' don't have unfoldings or spec-envs
237 -- We add them as we go down, using simplPrags
239 simplRecBind NotTopLevel pairs bndrs' (simplExprF body cont)
241 simplExprF expr@(Lam _ _) cont = simplLam expr cont
243 simplExprF (Type ty) cont
244 = ASSERT( case cont of { Stop _ -> True; ArgOf _ _ _ -> True; other -> False } )
245 simplType ty `thenSmpl` \ ty' ->
246 rebuild (Type ty') cont
248 simplExprF (Note (Coerce to from) e) cont
249 | to == from = simplExprF e cont
250 | otherwise = getSubst `thenSmpl` \ subst ->
251 simplExprF e (CoerceIt (substTy subst to) cont)
253 -- hack: we only distinguish subsumed cost centre stacks for the purposes of
254 -- inlining. All other CCCSs are mapped to currentCCS.
255 simplExprF (Note (SCC cc) e) cont
256 = setEnclosingCC currentCCS $
257 simplExpr e `thenSmpl` \ e ->
258 rebuild (mkNote (SCC cc) e) cont
260 simplExprF (Note InlineCall e) cont
261 = simplExprF e (InlinePlease cont)
263 -- Comments about the InlineMe case
264 -- ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
265 -- Don't inline in the RHS of something that has an
266 -- inline pragma. But be careful that the InScopeEnv that
267 -- we return does still have inlinings on!
269 -- It really is important to switch off inlinings. This function
270 -- may be inlinined in other modules, so we don't want to remove
271 -- (by inlining) calls to functions that have specialisations, or
272 -- that may have transformation rules in an importing scope.
273 -- E.g. {-# INLINE f #-}
275 -- and suppose that g is strict *and* has specialisations.
276 -- If we inline g's wrapper, we deny f the chance of getting
277 -- the specialised version of g when f is inlined at some call site
278 -- (perhaps in some other module).
280 simplExprF (Note InlineMe e) cont
282 Stop _ -> -- Totally boring continuation
283 -- Don't inline inside an INLINE expression
284 switchOffInlining (simplExpr e) `thenSmpl` \ e' ->
285 rebuild (mkNote InlineMe e') cont
287 other -> -- Dissolve the InlineMe note if there's
288 -- an interesting context of any kind to combine with
289 -- (even a type application -- anything except Stop)
292 -- A non-recursive let is dealt with by simplBeta
293 simplExprF (Let (NonRec bndr rhs) body) cont
294 = getSubstEnv `thenSmpl` \ se ->
295 simplBeta bndr rhs se (contResultType cont) $
300 ---------------------------------
306 zap_it = mkLamBndrZapper fun (countArgs cont)
307 cont_ty = contResultType cont
309 -- Type-beta reduction
310 go (Lam bndr body) (ApplyTo _ (Type ty_arg) arg_se body_cont)
311 = ASSERT( isTyVar bndr )
312 tick (BetaReduction bndr) `thenSmpl_`
313 getInScope `thenSmpl` \ in_scope ->
315 ty' = substTy (mkSubst in_scope arg_se) ty_arg
317 extendSubst bndr (DoneTy ty')
320 -- Ordinary beta reduction
321 go (Lam bndr body) cont@(ApplyTo _ arg arg_se body_cont)
322 = tick (BetaReduction bndr) `thenSmpl_`
323 simplBeta zapped_bndr arg arg_se cont_ty
326 zapped_bndr = zap_it bndr
329 go lam@(Lam _ _) cont = completeLam [] lam cont
331 -- Exactly enough args
332 go expr cont = simplExprF expr cont
335 -- completeLam deals with the case where a lambda doesn't have an ApplyTo
336 -- continuation. Try for eta reduction, but *only* if we get all
337 -- the way to an exprIsTrivial expression.
338 -- 'acc' holds the simplified binders, in reverse order
340 completeLam acc (Lam bndr body) cont
341 = simplBinder bndr $ \ bndr' ->
342 completeLam (bndr':acc) body cont
344 completeLam acc body cont
345 = simplExpr body `thenSmpl` \ body' ->
347 case (opt_SimplDoEtaReduction, check_eta acc body') of
348 (True, Just body'') -- Eta reduce!
349 -> tick (EtaReduction (head acc)) `thenSmpl_`
352 other -> -- No eta reduction
353 rebuild (foldl (flip Lam) body' acc) cont
354 -- Remember, acc is the reversed binders
356 -- NB: the binders are reversed
357 check_eta (b : bs) (App fun arg)
358 | (varToCoreExpr b `cheapEqExpr` arg)
362 | exprIsTrivial body && -- ONLY if the body is trivial
363 not (any (`elemVarSet` body_fvs) acc)
364 = Just body -- Success!
366 body_fvs = exprFreeVars body
368 check_eta _ _ = Nothing -- Bale out
370 mkLamBndrZapper :: CoreExpr -- Function
371 -> Int -- Number of args
372 -> Id -> Id -- Use this to zap the binders
373 mkLamBndrZapper fun n_args
374 | n_args >= n_params fun = \b -> b -- Enough args
375 | otherwise = \b -> maybeModifyIdInfo zapLamIdInfo b
377 n_params (Lam b e) | isId b = 1 + n_params e
378 | otherwise = n_params e
379 n_params other = 0::Int
383 ---------------------------------
384 simplConArgs makes sure that the arguments all end up being atomic.
385 That means it may generate some Lets, hence the strange type
388 simplConArgs :: [InArg] -> ([OutArg] -> SimplM OutExprStuff) -> SimplM OutExprStuff
389 simplConArgs [] thing_inside
392 simplConArgs (arg:args) thing_inside
393 = switchOffInlining (simplExpr arg) `thenSmpl` \ arg' ->
394 -- Simplify the RHS with inlining switched off, so that
395 -- only absolutely essential things will happen.
397 simplConArgs args $ \ args' ->
399 -- If the argument ain't trivial, then let-bind it
400 if exprIsTrivial arg' then
401 thing_inside (arg' : args')
403 newId (coreExprType arg') $ \ arg_id ->
404 thing_inside (Var arg_id : args') `thenSmpl` \ res ->
405 returnSmpl (addBind (NonRec arg_id arg') res)
409 ---------------------------------
411 simplType :: InType -> SimplM OutType
413 = getSubst `thenSmpl` \ subst ->
414 returnSmpl (substTy subst ty)
418 %************************************************************************
422 %************************************************************************
424 @simplBeta@ is used for non-recursive lets in expressions,
425 as well as true beta reduction.
427 Very similar to @simplLazyBind@, but not quite the same.
430 simplBeta :: InId -- Binder
431 -> InExpr -> SubstEnv -- Arg, with its subst-env
432 -> OutType -- Type of thing computed by the context
433 -> SimplM OutExprStuff -- The body
434 -> SimplM OutExprStuff
436 simplBeta bndr rhs rhs_se cont_ty thing_inside
438 = pprPanic "simplBeta" (ppr bndr <+> ppr rhs)
441 simplBeta bndr rhs rhs_se cont_ty thing_inside
442 | preInlineUnconditionally bndr && not opt_SimplNoPreInlining
443 = tick (PreInlineUnconditionally bndr) `thenSmpl_`
444 extendSubst bndr (ContEx rhs_se rhs) thing_inside
447 = -- Simplify the RHS
448 simplBinder bndr $ \ bndr' ->
449 simplArg (idType bndr') (getIdDemandInfo bndr)
450 rhs rhs_se cont_ty $ \ rhs' ->
452 -- Now complete the binding and simplify the body
453 completeBeta bndr bndr' rhs' thing_inside
455 completeBeta bndr bndr' rhs' thing_inside
456 | isUnLiftedType (idType bndr') && not (exprOkForSpeculation rhs')
457 -- Make a case expression instead of a let
458 -- These can arise either from the desugarer,
459 -- or from beta reductions: (\x.e) (x +# y)
460 = getInScope `thenSmpl` \ in_scope ->
461 thing_inside `thenSmpl` \ (floats, (_, body)) ->
462 returnSmpl ([], (in_scope, Case rhs' bndr' [(DEFAULT, [], mkLets floats body)]))
465 = completeBinding bndr bndr' rhs' thing_inside
470 simplArg :: OutType -> Demand
471 -> InExpr -> SubstEnv
472 -> OutType -- Type of thing computed by the context
473 -> (OutExpr -> SimplM OutExprStuff)
474 -> SimplM OutExprStuff
475 simplArg arg_ty demand arg arg_se cont_ty thing_inside
477 isUnLiftedType arg_ty ||
478 (opt_DictsStrict && isDictTy arg_ty && isDataType arg_ty)
479 -- Return true only for dictionary types where the dictionary
480 -- has more than one component (else we risk poking on the component
481 -- of a newtype dictionary)
482 = getSubstEnv `thenSmpl` \ body_se ->
483 transformRhs arg `thenSmpl` \ t_arg ->
484 setSubstEnv arg_se (simplExprF t_arg (ArgOf NoDup cont_ty $ \ arg' ->
485 setSubstEnv body_se (thing_inside arg')
486 )) -- NB: we must restore body_se before carrying on with thing_inside!!
489 = simplRhs NotTopLevel True arg_ty arg arg_se thing_inside
494 - deals only with Ids, not TyVars
495 - take an already-simplified RHS
497 It does *not* attempt to do let-to-case. Why? Because they are used for
500 (when let-to-case is impossible)
502 - many situations where the "rhs" is known to be a WHNF
503 (so let-to-case is inappropriate).
506 completeBinding :: InId -- Binder
507 -> OutId -- New binder
508 -> OutExpr -- Simplified RHS
509 -> SimplM (OutStuff a) -- Thing inside
510 -> SimplM (OutStuff a)
512 completeBinding old_bndr new_bndr new_rhs thing_inside
513 | isDeadBinder old_bndr -- This happens; for example, the case_bndr during case of
514 -- known constructor: case (a,b) of x { (p,q) -> ... }
515 -- Here x isn't mentioned in the RHS, so we don't want to
516 -- create the (dead) let-binding let x = (a,b) in ...
519 | postInlineUnconditionally old_bndr new_rhs
520 -- Maybe we don't need a let-binding! Maybe we can just
521 -- inline it right away. Unlike the preInlineUnconditionally case
522 -- we are allowed to look at the RHS.
524 -- NB: a loop breaker never has postInlineUnconditionally True
525 -- and non-loop-breakers only have *forward* references
526 -- Hence, it's safe to discard the binding
527 = tick (PostInlineUnconditionally old_bndr) `thenSmpl_`
528 extendSubst old_bndr (DoneEx new_rhs)
532 = getSubst `thenSmpl` \ subst ->
534 -- We make new IdInfo for the new binder by starting from the old binder,
535 -- doing appropriate substitutions,
536 old_bndr_info = idInfo old_bndr
537 new_bndr_info = substIdInfo subst old_bndr_info
538 `setArityInfo` ArityAtLeast (exprArity new_rhs)
540 -- At the *binding* site we want to zap the now-out-of-date inline
541 -- pragma, in case the expression is simplified a second time.
542 -- This has already been done in new_bndr, so we get it from there
543 binding_site_id = new_bndr `setIdInfo`
544 (new_bndr_info `setInlinePragInfo` getInlinePragma new_bndr)
546 -- At the occurrence sites we want to know the unfolding,
547 -- We want the occurrence info of the *original*, which is already
549 occ_site_id = new_bndr `setIdInfo`
550 (new_bndr_info `setUnfoldingInfo` mkUnfolding new_rhs)
552 modifyInScope occ_site_id thing_inside `thenSmpl` \ stuff ->
553 returnSmpl (addBind (NonRec binding_site_id new_rhs) stuff)
557 %************************************************************************
559 \subsection{simplLazyBind}
561 %************************************************************************
563 simplLazyBind basically just simplifies the RHS of a let(rec).
564 It does two important optimisations though:
566 * It floats let(rec)s out of the RHS, even if they
567 are hidden by big lambdas
569 * It does eta expansion
572 simplLazyBind :: TopLevelFlag
575 -> SimplM (OutStuff a) -- The body of the binding
576 -> SimplM (OutStuff a)
577 -- When called, the subst env is correct for the entire let-binding
578 -- and hence right for the RHS.
579 -- Also the binder has already been simplified, and hence is in scope
581 simplLazyBind top_lvl bndr bndr' rhs thing_inside
582 | preInlineUnconditionally bndr && not opt_SimplNoPreInlining
583 = tick (PreInlineUnconditionally bndr) `thenSmpl_`
584 getSubstEnv `thenSmpl` \ rhs_se ->
585 (extendSubst bndr (ContEx rhs_se rhs) thing_inside)
588 = -- Simplify the RHS
589 getSubstEnv `thenSmpl` \ rhs_se ->
591 simplRhs top_lvl False {- Not ok to float unboxed -}
593 rhs rhs_se $ \ rhs' ->
595 -- Now compete the binding and simplify the body
596 completeBinding bndr bndr' rhs' thing_inside
602 simplRhs :: TopLevelFlag
603 -> Bool -- True <=> OK to float unboxed (speculative) bindings
604 -> OutType -> InExpr -> SubstEnv
605 -> (OutExpr -> SimplM (OutStuff a))
606 -> SimplM (OutStuff a)
607 simplRhs top_lvl float_ubx rhs_ty rhs rhs_se thing_inside
608 = -- Swizzle the inner lets past the big lambda (if any)
609 -- and try eta expansion
610 transformRhs rhs `thenSmpl` \ t_rhs ->
613 setSubstEnv rhs_se (simplExprF t_rhs (Stop rhs_ty)) `thenSmpl` \ (floats, (in_scope', rhs')) ->
615 -- Float lets out of RHS
617 (floats_out, rhs'') | float_ubx = (floats, rhs')
618 | otherwise = splitFloats floats rhs'
620 if (isTopLevel top_lvl || exprIsCheap rhs') && -- Float lets if (a) we're at the top level
621 not (null floats_out) -- or (b) it exposes a cheap (i.e. duplicatable) expression
623 tickLetFloat floats_out `thenSmpl_`
626 -- There's a subtlety here. There may be a binding (x* = e) in the
627 -- floats, where the '*' means 'will be demanded'. So is it safe
628 -- to float it out? Answer no, but it won't matter because
629 -- we only float if arg' is a WHNF,
630 -- and so there can't be any 'will be demanded' bindings in the floats.
632 WARN( any demanded_float floats_out, ppr floats_out )
633 setInScope in_scope' (thing_inside rhs'') `thenSmpl` \ stuff ->
634 -- in_scope' may be excessive, but that's OK;
635 -- it's a superset of what's in scope
636 returnSmpl (addBinds floats_out stuff)
638 -- Don't do the float
639 thing_inside (mkLets floats rhs')
641 -- In a let-from-let float, we just tick once, arbitrarily
642 -- choosing the first floated binder to identify it
643 tickLetFloat (NonRec b r : fs) = tick (LetFloatFromLet b)
644 tickLetFloat (Rec ((b,r):prs) : fs) = tick (LetFloatFromLet b)
646 demanded_float (NonRec b r) = isStrict (getIdDemandInfo b) && not (isUnLiftedType (idType b))
647 -- Unlifted-type (cheap-eagerness) lets may well have a demanded flag on them
648 demanded_float (Rec _) = False
650 -- Don't float any unlifted bindings out, because the context
651 -- is either a Rec group, or the top level, neither of which
652 -- can tolerate them.
653 splitFloats floats rhs
657 go (f:fs) | must_stay f = ([], mkLets (f:fs) rhs)
658 | otherwise = case go fs of
659 (out, rhs') -> (f:out, rhs')
661 must_stay (Rec prs) = False -- No unlifted bindings in here
662 must_stay (NonRec b r) = isUnLiftedType (idType b)
667 %************************************************************************
669 \subsection{Variables}
671 %************************************************************************
675 = freeTick LeafVisit `thenSmpl_`
676 getSubst `thenSmpl` \ subst ->
677 case lookupSubst subst var of
678 Just (DoneEx (Var v)) -> zapSubstEnv (simplVar v cont)
679 Just (DoneEx e) -> zapSubstEnv (simplExprF e cont)
680 Just (ContEx env' e) -> setSubstEnv env' (simplExprF e cont)
683 var' = case lookupInScope subst var of
687 if isLocallyDefined var && not (idMustBeINLINEd var)
688 -- The idMustBeINLINEd test accouunts for the fact
689 -- that class dictionary constructors don't have top level
690 -- bindings and hence aren't in scope.
693 pprTrace "simplVar:" (ppr var) var
698 getBlackList `thenSmpl` \ black_list ->
699 getInScope `thenSmpl` \ in_scope ->
700 completeCall black_list in_scope var' cont
702 ---------------------------------------------------------
703 -- Dealing with a call
705 completeCall black_list_fn in_scope var cont
706 -- Look for rules or specialisations that match
707 -- Do this *before* trying inlining because some functions
708 -- have specialisations *and* are strict; we don't want to
709 -- inline the wrapper of the non-specialised thing... better
710 -- to call the specialised thing instead.
711 | maybeToBool maybe_rule_match
712 = tick (RuleFired rule_name) `thenSmpl_`
713 zapSubstEnv (simplExprF rule_rhs (pushArgs emptySubstEnv rule_args result_cont))
714 -- See note below about zapping the substitution here
716 -- Look for an unfolding. There's a binding for the
717 -- thing, but perhaps we want to inline it anyway
718 | maybeToBool maybe_inline
719 = tick (UnfoldingDone var) `thenSmpl_`
720 zapSubstEnv (completeInlining var unf_template discard_inline_cont)
721 -- The template is already simplified, so don't re-substitute.
722 -- This is VITAL. Consider
724 -- let y = \z -> ...x... in
726 -- We'll clone the inner \x, adding x->x' in the id_subst
727 -- Then when we inline y, we must *not* replace x by x' in
728 -- the inlined copy!!
730 | otherwise -- Neither rule nor inlining
731 -- Use prepareArgs to use function strictness
732 = prepareArgs (ppr var) (idType var) (get_str var) cont $ \ args' cont' ->
733 rebuild (mkApps (Var var) args') cont'
736 get_str var = case getIdStrictness var of
737 NoStrictnessInfo -> (repeat wwLazy, False)
738 StrictnessInfo demands result_bot -> (demands, result_bot)
741 (args', result_cont) = contArgs in_scope cont
742 val_args = filter isValArg args'
743 arg_infos = map (interestingArg in_scope) val_args
744 inline_call = contIsInline result_cont
745 interesting_cont = contIsInteresting result_cont
746 discard_inline_cont | inline_call = discardInline cont
749 ---------- Unfolding stuff
750 maybe_inline = callSiteInline black_listed inline_call
751 var arg_infos interesting_cont
752 Just unf_template = maybe_inline
753 black_listed = black_list_fn var
755 ---------- Specialisation stuff
756 maybe_rule_match = lookupRule in_scope var args'
757 Just (rule_name, rule_rhs, rule_args) = maybe_rule_match
761 -- An argument is interesting if it has *some* structure
762 -- We are here trying to avoid unfolding a function that
763 -- is applied only to variables that have no unfolding
764 -- (i.e. they are probably lambda bound): f x y z
765 -- There is little point in inlining f here.
766 interestingArg in_scope (Type _) = False
767 interestingArg in_scope (App fn (Type _)) = interestingArg in_scope fn
768 interestingArg in_scope (Var v) = hasSomeUnfolding (getIdUnfolding v')
770 v' = case lookupVarSet in_scope v of
773 interestingArg in_scope other = True
776 -- First a special case
777 -- Don't actually inline the scrutinee when we see
778 -- case x of y { .... }
779 -- and x has unfolding (C a b). Why not? Because
780 -- we get a silly binding y = C a b. If we don't
781 -- inline knownCon can directly substitute x for y instead.
782 completeInlining var (Con con con_args) (Select _ bndr alts se cont)
784 = knownCon (Var var) con con_args bndr alts se cont
786 -- Now the normal case
787 completeInlining var unfolding cont
788 = simplExprF unfolding cont
790 ----------- costCentreOk
791 -- costCentreOk checks that it's ok to inline this thing
792 -- The time it *isn't* is this:
794 -- f x = let y = E in
795 -- scc "foo" (...y...)
797 -- Here y has a "current cost centre", and we can't inline it inside "foo",
798 -- regardless of whether E is a WHNF or not.
800 costCentreOk ccs_encl cc_rhs
801 = not opt_SccProfilingOn
802 || isSubsumedCCS ccs_encl -- can unfold anything into a subsumed scope
803 || not (isEmptyCC cc_rhs) -- otherwise need a cc on the unfolding
808 ---------------------------------------------------------
809 -- Preparing arguments for a call
811 prepareArgs :: SDoc -- Error message info
812 -> OutType -> ([Demand],Bool) -> SimplCont
813 -> ([OutExpr] -> SimplCont -> SimplM OutExprStuff)
814 -> SimplM OutExprStuff
816 prepareArgs pp_fun orig_fun_ty (fun_demands, result_bot) orig_cont thing_inside
817 = go [] demands orig_fun_ty orig_cont
819 not_enough_args = fun_demands `lengthExceeds` countValArgs orig_cont
820 -- "No strictness info" is signalled by an infinite list of wwLazy
822 demands | not_enough_args = repeat wwLazy -- Not enough args, or no strictness
823 | result_bot = fun_demands -- Enough args, and function returns bottom
824 | otherwise = fun_demands ++ repeat wwLazy -- Enough args and function does not return bottom
825 -- NB: demands is finite iff enough args and result_bot is True
827 -- Main game plan: loop through the arguments, simplifying
828 -- each of them in turn. We carry with us a list of demands,
829 -- and the type of the function-applied-to-earlier-args
832 go acc ds fun_ty (ApplyTo _ arg@(Type ty_arg) se cont)
833 = getInScope `thenSmpl` \ in_scope ->
835 ty_arg' = substTy (mkSubst in_scope se) ty_arg
836 res_ty = applyTy fun_ty ty_arg'
838 go (Type ty_arg' : acc) ds res_ty cont
841 go acc (d:ds) fun_ty (ApplyTo _ val_arg se cont)
842 = case splitFunTy_maybe fun_ty of {
843 Nothing -> pprTrace "prepareArgs" (pp_fun $$ ppr orig_fun_ty $$ ppr orig_cont)
844 (thing_inside (reverse acc) cont) ;
845 Just (arg_ty, res_ty) ->
846 simplArg arg_ty d val_arg se (contResultType cont) $ \ arg' ->
847 go (arg':acc) ds res_ty cont }
849 -- We've run out of demands, which only happens for functions
850 -- we *know* now return bottom
852 -- * case (error "hello") of { ... }
853 -- * (error "Hello") arg
854 -- * f (error "Hello") where f is strict
856 go acc [] fun_ty cont = tick_case_of_error cont `thenSmpl_`
857 thing_inside (reverse acc) (discardCont cont)
859 -- We're run out of arguments
860 go acc ds fun_ty cont = thing_inside (reverse acc) cont
862 -- Boring: we must only record a tick if there was an interesting
863 -- continuation to discard. If not, we tick forever.
864 tick_case_of_error (Stop _) = returnSmpl ()
865 tick_case_of_error (CoerceIt _ (Stop _)) = returnSmpl ()
866 tick_case_of_error other = tick BottomFound
869 %************************************************************************
871 \subsection{Decisions about inlining}
873 %************************************************************************
876 preInlineUnconditionally :: InId -> Bool
877 -- Examines a bndr to see if it is used just once in a
878 -- completely safe way, so that it is safe to discard the binding
879 -- inline its RHS at the (unique) usage site, REGARDLESS of how
880 -- big the RHS might be. If this is the case we don't simplify
881 -- the RHS first, but just inline it un-simplified.
883 -- This is much better than first simplifying a perhaps-huge RHS
884 -- and then inlining and re-simplifying it.
886 -- NB: we don't even look at the RHS to see if it's trivial
889 -- where x is used many times, but this is the unique occurrence
890 -- of y. We should NOT inline x at all its uses, because then
891 -- we'd do the same for y -- aargh! So we must base this
892 -- pre-rhs-simplification decision solely on x's occurrences, not
895 -- Evne RHSs labelled InlineMe aren't caught here, because
896 -- there might be no benefit from inlining at the call site.
897 -- But things labelled 'IMustBeINLINEd' *are* caught. We use this
898 -- for the trivial bindings introduced by SimplUtils.mkRhsTyLam
899 preInlineUnconditionally bndr
900 = case getInlinePragma bndr of
901 IMustBeINLINEd -> True
902 ICanSafelyBeINLINEd NotInsideLam True -> True -- Not inside a lambda,
903 -- one occurrence ==> safe!
907 postInlineUnconditionally :: InId -> OutExpr -> Bool
908 -- Examines a (bndr = rhs) binding, AFTER the rhs has been simplified
909 -- It returns True if it's ok to discard the binding and inline the
910 -- RHS at every use site.
912 -- NOTE: This isn't our last opportunity to inline.
913 -- We're at the binding site right now, and
914 -- we'll get another opportunity when we get to the ocurrence(s)
916 postInlineUnconditionally bndr rhs
920 = case getInlinePragma bndr of
921 IAmALoopBreaker -> False
923 ICanSafelyBeINLINEd InsideLam one_branch -> exprIsTrivial rhs
924 -- Don't inline even WHNFs inside lambdas; doing so may
925 -- simply increase allocation when the function is called
926 -- This isn't the last chance; see NOTE above.
928 ICanSafelyBeINLINEd not_in_lam one_branch -> one_branch || exprIsTrivial rhs
929 -- Was 'exprIsDupable' instead of 'exprIsTrivial' but the
930 -- decision about duplicating code is best left to callSiteInline
932 other -> exprIsTrivial rhs -- Duplicating is *free*
933 -- NB: Even InlineMe and IMustBeINLINEd are ignored here
934 -- Why? Because we don't even want to inline them into the
935 -- RHS of constructor arguments. See NOTE above
936 -- NB: Even IMustBeINLINEd is ignored here: if the rhs is trivial
937 -- it's best to inline it anyway. We often get a=E; b=a
938 -- from desugaring, with both a and b marked NOINLINE.
943 %************************************************************************
945 \subsection{The main rebuilder}
947 %************************************************************************
950 -------------------------------------------------------------------
953 = getInScope `thenSmpl` \ in_scope ->
954 returnSmpl ([], (in_scope, expr))
956 ---------------------------------------------------------
957 rebuild :: OutExpr -> SimplCont -> SimplM OutExprStuff
960 rebuild expr (Stop _) = rebuild_done expr
962 -- ArgOf continuation
963 rebuild expr (ArgOf _ _ cont_fn) = cont_fn expr
965 -- ApplyTo continuation
966 rebuild expr cont@(ApplyTo _ arg se cont')
967 = setSubstEnv se (simplExpr arg) `thenSmpl` \ arg' ->
968 rebuild (App expr arg') cont'
970 -- Coerce continuation
971 rebuild expr (CoerceIt to_ty cont)
972 = rebuild (mkCoerce to_ty expr) cont
974 -- Inline continuation
975 rebuild expr (InlinePlease cont)
976 = rebuild (Note InlineCall expr) cont
978 -- Case of known constructor or literal
979 rebuild expr@(Con con args) (Select _ bndr alts se cont)
980 | conOkForAlt con -- Knocks out PrimOps and NoRepLits
981 = knownCon expr con args bndr alts se cont
984 ---------------------------------------------------------
985 -- The other Select cases
987 rebuild scrut (Select _ bndr alts se cont)
988 | -- Check that the RHSs are all the same, and
989 -- don't use the binders in the alternatives
990 -- This test succeeds rapidly in the common case of
991 -- a single DEFAULT alternative
992 all (cheapEqExpr rhs1) other_rhss && all binders_unused alts
994 -- Check that the scrutinee can be let-bound instead of case-bound
995 && ( exprOkForSpeculation scrut
996 -- OK not to evaluate it
997 -- This includes things like (==# a# b#)::Bool
998 -- so that we simplify
999 -- case ==# a# b# of { True -> x; False -> x }
1002 -- This particular example shows up in default methods for
1003 -- comparision operations (e.g. in (>=) for Int.Int32)
1004 || exprIsValue scrut -- It's already evaluated
1005 || var_demanded_later scrut -- It'll be demanded later
1007 -- || not opt_SimplPedanticBottoms) -- Or we don't care!
1008 -- We used to allow improving termination by discarding cases, unless -fpedantic-bottoms was on,
1009 -- but that breaks badly for the dataToTag# primop, which relies on a case to evaluate
1010 -- its argument: case x of { y -> dataToTag# y }
1011 -- Here we must *not* discard the case, because dataToTag# just fetches the tag from
1012 -- the info pointer. So we'll be pedantic all the time, and see if that gives any
1016 -- && opt_SimplDoCaseElim
1017 -- [June 99; don't test this flag. The code generator dies if it sees
1018 -- case (\x.e) of f -> ...
1019 -- so better to always do it
1021 -- Get rid of the case altogether
1022 -- See the extensive notes on case-elimination below
1023 -- Remember to bind the binder though!
1024 = tick (CaseElim bndr) `thenSmpl_` (
1026 simplBinder bndr $ \ bndr' ->
1027 completeBinding bndr bndr' scrut $
1028 simplExprF rhs1 cont)
1031 = rebuild_case scrut bndr alts se cont
1033 (rhs1:other_rhss) = [rhs | (_,_,rhs) <- alts]
1034 binders_unused (_, bndrs, _) = all isDeadBinder bndrs
1036 var_demanded_later (Var v) = isStrict (getIdDemandInfo bndr) -- It's going to be evaluated later
1037 var_demanded_later other = False
1040 Case elimination [see the code above]
1042 Start with a simple situation:
1044 case x# of ===> e[x#/y#]
1047 (when x#, y# are of primitive type, of course). We can't (in general)
1048 do this for algebraic cases, because we might turn bottom into
1051 Actually, we generalise this idea to look for a case where we're
1052 scrutinising a variable, and we know that only the default case can
1057 other -> ...(case x of
1061 Here the inner case can be eliminated. This really only shows up in
1062 eliminating error-checking code.
1064 We also make sure that we deal with this very common case:
1069 Here we are using the case as a strict let; if x is used only once
1070 then we want to inline it. We have to be careful that this doesn't
1071 make the program terminate when it would have diverged before, so we
1073 - x is used strictly, or
1074 - e is already evaluated (it may so if e is a variable)
1076 Lastly, we generalise the transformation to handle this:
1082 We only do this for very cheaply compared r's (constructors, literals
1083 and variables). If pedantic bottoms is on, we only do it when the
1084 scrutinee is a PrimOp which can't fail.
1086 We do it *here*, looking at un-simplified alternatives, because we
1087 have to check that r doesn't mention the variables bound by the
1088 pattern in each alternative, so the binder-info is rather useful.
1090 So the case-elimination algorithm is:
1092 1. Eliminate alternatives which can't match
1094 2. Check whether all the remaining alternatives
1095 (a) do not mention in their rhs any of the variables bound in their pattern
1096 and (b) have equal rhss
1098 3. Check we can safely ditch the case:
1099 * PedanticBottoms is off,
1100 or * the scrutinee is an already-evaluated variable
1101 or * the scrutinee is a primop which is ok for speculation
1102 -- ie we want to preserve divide-by-zero errors, and
1103 -- calls to error itself!
1105 or * [Prim cases] the scrutinee is a primitive variable
1107 or * [Alg cases] the scrutinee is a variable and
1108 either * the rhs is the same variable
1109 (eg case x of C a b -> x ===> x)
1110 or * there is only one alternative, the default alternative,
1111 and the binder is used strictly in its scope.
1112 [NB this is helped by the "use default binder where
1113 possible" transformation; see below.]
1116 If so, then we can replace the case with one of the rhss.
1119 Blob of helper functions for the "case-of-something-else" situation.
1122 ---------------------------------------------------------
1123 -- Case of something else
1125 rebuild_case scrut case_bndr alts se cont
1126 = -- Prepare case alternatives
1127 prepareCaseAlts case_bndr (splitTyConApp_maybe (idType case_bndr))
1128 scrut_cons alts `thenSmpl` \ better_alts ->
1130 -- Set the new subst-env in place (before dealing with the case binder)
1133 -- Deal with the case binder, and prepare the continuation;
1134 -- The new subst_env is in place
1135 prepareCaseCont better_alts cont $ \ cont' ->
1138 -- Deal with variable scrutinee
1139 ( simplBinder case_bndr $ \ case_bndr' ->
1140 substForVarScrut scrut case_bndr' $ \ zap_occ_info ->
1142 case_bndr'' = zap_occ_info case_bndr'
1145 -- Deal with the case alternaatives
1146 simplAlts zap_occ_info scrut_cons
1147 case_bndr'' better_alts cont' `thenSmpl` \ alts' ->
1149 mkCase scrut case_bndr'' alts'
1150 ) `thenSmpl` \ case_expr ->
1152 -- Notice that the simplBinder, prepareCaseCont, etc, do *not* scope
1153 -- over the rebuild_done; rebuild_done returns the in-scope set, and
1154 -- that should not include these chaps!
1155 rebuild_done case_expr
1157 -- scrut_cons tells what constructors the scrutinee can't possibly match
1158 scrut_cons = case scrut of
1159 Var v -> otherCons (getIdUnfolding v)
1163 knownCon expr con args bndr alts se cont
1164 = tick (KnownBranch bndr) `thenSmpl_`
1166 simplBinder bndr $ \ bndr' ->
1167 case findAlt con alts of
1168 (DEFAULT, bs, rhs) -> ASSERT( null bs )
1169 completeBinding bndr bndr' expr $
1170 -- Don't use completeBeta here. The expr might be
1171 -- an unboxed literal, like 3, or a variable
1172 -- whose unfolding is an unboxed literal... and
1173 -- completeBeta will just construct another case
1177 (Literal lit, bs, rhs) -> ASSERT( null bs )
1178 extendSubst bndr (DoneEx expr) $
1179 -- Unconditionally substitute, because expr must
1180 -- be a variable or a literal. It can't be a
1181 -- NoRep literal because they don't occur in
1185 (DataCon dc, bs, rhs) -> ASSERT( length bs == length real_args )
1186 completeBinding bndr bndr' expr $
1188 extendSubstList bs (map mk real_args) $
1191 real_args = drop (dataConNumInstArgs dc) args
1192 mk (Type ty) = DoneTy ty
1193 mk other = DoneEx other
1198 prepareCaseCont :: [InAlt] -> SimplCont
1199 -> (SimplCont -> SimplM (OutStuff a))
1200 -> SimplM (OutStuff a)
1201 -- Polymorphic recursion here!
1203 prepareCaseCont [alt] cont thing_inside = thing_inside cont
1204 prepareCaseCont alts cont thing_inside = mkDupableCont (coreAltsType alts) cont thing_inside
1207 substForVarScrut checks whether the scrutinee is a variable, v.
1208 If so, try to eliminate uses of v in the RHSs in favour of case_bndr;
1209 that way, there's a chance that v will now only be used once, and hence inlined.
1211 If we do this, then we have to nuke any occurrence info (eg IAmDead)
1212 in the case binder, because the case-binder now effectively occurs
1213 whenever v does. AND we have to do the same for the pattern-bound
1216 (case x of { (a,b) -> a }) (case x of { (p,q) -> q })
1218 Here, b and p are dead. But when we move the argment inside the first
1219 case RHS, and eliminate the second case, we get
1221 case x or { (a,b) -> a b }
1223 Urk! b is alive! Reason: the scrutinee was a variable, and case elimination
1224 happened. Hence the zap_occ_info function returned by substForVarScrut
1227 substForVarScrut (Var v) case_bndr' thing_inside
1228 | isLocallyDefined v -- No point for imported things
1229 = modifyInScope (v `setIdUnfolding` mkUnfolding (Var case_bndr')
1230 `setInlinePragma` IMustBeINLINEd) $
1231 -- We could extend the substitution instead, but it would be
1232 -- a hack because then the substitution wouldn't be idempotent
1234 thing_inside (\ bndr -> bndr `setInlinePragma` NoInlinePragInfo)
1236 substForVarScrut other_scrut case_bndr' thing_inside
1237 = thing_inside (\ bndr -> bndr) -- NoOp on bndr
1240 prepareCaseAlts does two things:
1242 1. Remove impossible alternatives
1244 2. If the DEFAULT alternative can match only one possible constructor,
1245 then make that constructor explicit.
1247 case e of x { DEFAULT -> rhs }
1249 case e of x { (a,b) -> rhs }
1250 where the type is a single constructor type. This gives better code
1251 when rhs also scrutinises x or e.
1254 prepareCaseAlts bndr (Just (tycon, inst_tys)) scrut_cons alts
1256 = case (findDefault filtered_alts, missing_cons) of
1258 ((alts_no_deflt, Just rhs), [data_con]) -- Just one missing constructor!
1259 -> tick (FillInCaseDefault bndr) `thenSmpl_`
1261 (_,_,ex_tyvars,_,_,_) = dataConSig data_con
1263 getUniquesSmpl (length ex_tyvars) `thenSmpl` \ tv_uniqs ->
1265 ex_tyvars' = zipWithEqual "simpl_alt" mk tv_uniqs ex_tyvars
1266 mk uniq tv = mkSysTyVar uniq (tyVarKind tv)
1268 newIds (dataConArgTys
1270 (inst_tys ++ mkTyVarTys ex_tyvars')) $ \ bndrs ->
1271 returnSmpl ((DataCon data_con, ex_tyvars' ++ bndrs, rhs) : alts_no_deflt)
1273 other -> returnSmpl filtered_alts
1275 -- Filter out alternatives that can't possibly match
1276 filtered_alts = case scrut_cons of
1278 other -> [alt | alt@(con,_,_) <- alts, not (con `elem` scrut_cons)]
1280 missing_cons = [data_con | data_con <- tyConDataCons tycon,
1281 not (data_con `elem` handled_data_cons)]
1282 handled_data_cons = [data_con | DataCon data_con <- scrut_cons] ++
1283 [data_con | (DataCon data_con, _, _) <- filtered_alts]
1286 prepareCaseAlts _ _ scrut_cons alts
1287 = returnSmpl alts -- Functions
1290 ----------------------
1291 simplAlts zap_occ_info scrut_cons case_bndr'' alts cont'
1292 = mapSmpl simpl_alt alts
1294 inst_tys' = case splitTyConApp_maybe (idType case_bndr'') of
1295 Just (tycon, inst_tys) -> inst_tys
1297 -- handled_cons is all the constructors that are dealt
1298 -- with, either by being impossible, or by there being an alternative
1299 handled_cons = scrut_cons ++ [con | (con,_,_) <- alts, con /= DEFAULT]
1301 simpl_alt (DEFAULT, _, rhs)
1302 = -- In the default case we record the constructors that the
1303 -- case-binder *can't* be.
1304 -- We take advantage of any OtherCon info in the case scrutinee
1305 modifyInScope (case_bndr'' `setIdUnfolding` mkOtherCon handled_cons) $
1306 simplExprC rhs cont' `thenSmpl` \ rhs' ->
1307 returnSmpl (DEFAULT, [], rhs')
1309 simpl_alt (con, vs, rhs)
1310 = -- Deal with the pattern-bound variables
1311 -- Mark the ones that are in ! positions in the data constructor
1312 -- as certainly-evaluated
1313 simplBinders (add_evals con vs) $ \ vs' ->
1315 -- Bind the case-binder to (Con args)
1317 con_app = Con con (map Type inst_tys' ++ map varToCoreExpr vs')
1319 modifyInScope (case_bndr'' `setIdUnfolding` mkUnfolding con_app) $
1320 simplExprC rhs cont' `thenSmpl` \ rhs' ->
1321 returnSmpl (con, vs', rhs')
1324 -- add_evals records the evaluated-ness of the bound variables of
1325 -- a case pattern. This is *important*. Consider
1326 -- data T = T !Int !Int
1328 -- case x of { T a b -> T (a+1) b }
1330 -- We really must record that b is already evaluated so that we don't
1331 -- go and re-evaluate it when constructing the result.
1333 add_evals (DataCon dc) vs = cat_evals vs (dataConRepStrictness dc)
1334 add_evals other_con vs = vs
1336 cat_evals [] [] = []
1337 cat_evals (v:vs) (str:strs)
1338 | isTyVar v = v : cat_evals vs (str:strs)
1339 | isStrict str = (v' `setIdUnfolding` mkOtherCon []) : cat_evals vs strs
1340 | otherwise = v' : cat_evals vs strs
1346 %************************************************************************
1348 \subsection{Duplicating continuations}
1350 %************************************************************************
1353 mkDupableCont :: InType -- Type of the thing to be given to the continuation
1355 -> (SimplCont -> SimplM (OutStuff a))
1356 -> SimplM (OutStuff a)
1357 mkDupableCont ty cont thing_inside
1358 | contIsDupable cont
1361 mkDupableCont _ (CoerceIt ty cont) thing_inside
1362 = mkDupableCont ty cont $ \ cont' ->
1363 thing_inside (CoerceIt ty cont')
1365 mkDupableCont ty (InlinePlease cont) thing_inside
1366 = mkDupableCont ty cont $ \ cont' ->
1367 thing_inside (InlinePlease cont')
1369 mkDupableCont join_arg_ty (ArgOf _ cont_ty cont_fn) thing_inside
1370 = -- Build the RHS of the join point
1371 simplType join_arg_ty `thenSmpl` \ join_arg_ty' ->
1372 newId join_arg_ty' ( \ arg_id ->
1373 getSwitchChecker `thenSmpl` \ chkr ->
1374 cont_fn (Var arg_id) `thenSmpl` \ (binds, (_, rhs)) ->
1375 returnSmpl (Lam (setOneShotLambda arg_id) (mkLets binds rhs))
1376 ) `thenSmpl` \ join_rhs ->
1378 -- Build the join Id and continuation
1379 newId (coreExprType join_rhs) $ \ join_id ->
1381 new_cont = ArgOf OkToDup cont_ty
1382 (\arg' -> rebuild_done (App (Var join_id) arg'))
1385 -- Do the thing inside
1386 thing_inside new_cont `thenSmpl` \ res ->
1387 returnSmpl (addBind (NonRec join_id join_rhs) res)
1389 mkDupableCont ty (ApplyTo _ arg se cont) thing_inside
1390 = mkDupableCont (funResultTy ty) cont $ \ cont' ->
1391 setSubstEnv se (simplExpr arg) `thenSmpl` \ arg' ->
1392 if exprIsDupable arg' then
1393 thing_inside (ApplyTo OkToDup arg' emptySubstEnv cont')
1395 newId (coreExprType arg') $ \ bndr ->
1396 thing_inside (ApplyTo OkToDup (Var bndr) emptySubstEnv cont') `thenSmpl` \ res ->
1397 returnSmpl (addBind (NonRec bndr arg') res)
1399 mkDupableCont ty (Select _ case_bndr alts se cont) thing_inside
1400 = tick (CaseOfCase case_bndr) `thenSmpl_`
1402 simplBinder case_bndr $ \ case_bndr' ->
1403 prepareCaseCont alts cont $ \ cont' ->
1404 mapAndUnzipSmpl (mkDupableAlt case_bndr case_bndr' cont') alts `thenSmpl` \ (alt_binds_s, alts') ->
1405 returnSmpl (concat alt_binds_s, alts')
1406 ) `thenSmpl` \ (alt_binds, alts') ->
1408 extendInScopes [b | NonRec b _ <- alt_binds] $
1410 -- NB that the new alternatives, alts', are still InAlts, using the original
1411 -- binders. That means we can keep the case_bndr intact. This is important
1412 -- because another case-of-case might strike, and so we want to keep the
1413 -- info that the case_bndr is dead (if it is, which is often the case).
1414 -- This is VITAL when the type of case_bndr is an unboxed pair (often the
1415 -- case in I/O rich code. We aren't allowed a lambda bound
1416 -- arg of unboxed tuple type, and indeed such a case_bndr is always dead
1417 thing_inside (Select OkToDup case_bndr alts' se (Stop (contResultType cont))) `thenSmpl` \ res ->
1419 returnSmpl (addBinds alt_binds res)
1422 mkDupableAlt :: InId -> OutId -> SimplCont -> InAlt -> SimplM (OutStuff InAlt)
1423 mkDupableAlt case_bndr case_bndr' (Stop _) alt@(con, bndrs, rhs)
1425 = -- It is worth checking for a small RHS because otherwise we
1426 -- get extra let bindings that may cause an extra iteration of the simplifier to
1427 -- inline back in place. Quite often the rhs is just a variable or constructor.
1428 -- The Ord instance of Maybe in PrelMaybe.lhs, for example, took several extra
1429 -- iterations because the version with the let bindings looked big, and so wasn't
1430 -- inlined, but after the join points had been inlined it looked smaller, and so
1433 -- But since the continuation is absorbed into the rhs, we only do this
1434 -- for a Stop continuation.
1435 returnSmpl ([], alt)
1437 mkDupableAlt case_bndr case_bndr' cont alt@(con, bndrs, rhs)
1439 = -- Not worth checking whether the rhs is small; the
1440 -- inliner will inline it if so.
1441 simplBinders bndrs $ \ bndrs' ->
1442 simplExprC rhs cont `thenSmpl` \ rhs' ->
1444 rhs_ty' = coreExprType rhs'
1445 (used_bndrs, used_bndrs')
1446 = unzip [pr | pr@(bndr,bndr') <- zip (case_bndr : bndrs)
1447 (case_bndr' : bndrs'),
1448 not (isDeadBinder bndr)]
1449 -- The new binders have lost their occurrence info,
1450 -- so we have to extract it from the old ones
1452 ( if null used_bndrs'
1453 -- If we try to lift a primitive-typed something out
1454 -- for let-binding-purposes, we will *caseify* it (!),
1455 -- with potentially-disastrous strictness results. So
1456 -- instead we turn it into a function: \v -> e
1457 -- where v::State# RealWorld#. The value passed to this function
1458 -- is realworld#, which generates (almost) no code.
1460 -- There's a slight infelicity here: we pass the overall
1461 -- case_bndr to all the join points if it's used in *any* RHS,
1462 -- because we don't know its usage in each RHS separately
1464 -- We used to say "&& isUnLiftedType rhs_ty'" here, but now
1465 -- we make the join point into a function whenever used_bndrs'
1466 -- is empty. This makes the join-point more CPR friendly.
1467 -- Consider: let j = if .. then I# 3 else I# 4
1468 -- in case .. of { A -> j; B -> j; C -> ... }
1470 -- Now CPR should not w/w j because it's a thunk, so
1471 -- that means that the enclosing function can't w/w either,
1472 -- which is a lose. Here's the example that happened in practice:
1473 -- kgmod :: Int -> Int -> Int
1474 -- kgmod x y = if x > 0 && y < 0 || x < 0 && y > 0
1478 then newId realWorldStatePrimTy $ \ rw_id ->
1479 returnSmpl ([rw_id], [Var realWorldPrimId])
1481 returnSmpl (used_bndrs', map varToCoreExpr used_bndrs)
1483 `thenSmpl` \ (final_bndrs', final_args) ->
1485 newId (foldr (mkFunTy . idType) rhs_ty' final_bndrs') $ \ join_bndr ->
1487 -- Notice that we make the lambdas into one-shot-lambdas. The
1488 -- join point is sure to be applied at most once, and doing so
1489 -- prevents the body of the join point being floated out by
1490 -- the full laziness pass
1491 returnSmpl ([NonRec join_bndr (mkLams (map setOneShotLambda final_bndrs') rhs')],
1492 (con, bndrs, mkApps (Var join_bndr) final_args))