2 % (c) The University of Glasgow 2006
3 % (c) The GRASP/AQUA Project, Glasgow University, 1992-1998
6 Utility functions on @Core@ syntax
11 mkInlineMe, mkSCC, mkCoerce,
12 bindNonRec, needsCaseBinding,
13 mkIfThenElse, mkAltExpr, mkPiType, mkPiTypes,
15 -- Taking expressions apart
16 findDefault, findAlt, isDefaultAlt, mergeAlts,
18 -- Properties of expressions
19 exprType, coreAltType,
20 exprIsDupable, exprIsTrivial, exprIsCheap,
21 exprIsHNF,exprOkForSpeculation, exprIsBig,
22 exprIsConApp_maybe, exprIsBottom,
25 -- Arity and eta expansion
26 manifestArity, exprArity,
27 exprEtaExpandArity, etaExpand,
36 cheapEqExpr, tcEqExpr, tcEqExprX, applyTypeToArgs, applyTypeToArg,
38 dataConOrigInstPat, dataConRepInstPat, dataConRepFSInstPat
41 #include "HsVersions.h"
74 import GHC.Exts -- For `xori`
78 %************************************************************************
80 \subsection{Find the type of a Core atom/expression}
82 %************************************************************************
85 exprType :: CoreExpr -> Type
87 exprType (Var var) = idType var
88 exprType (Lit lit) = literalType lit
89 exprType (Let _ body) = exprType body
90 exprType (Case _ _ ty alts) = ty
92 = let (_, ty) = coercionKind co in ty
93 exprType (Note other_note e) = exprType e
94 exprType (Lam binder expr) = mkPiType binder (exprType expr)
96 = case collectArgs e of
97 (fun, args) -> applyTypeToArgs e (exprType fun) args
99 exprType other = pprTrace "exprType" (pprCoreExpr other) alphaTy
101 coreAltType :: CoreAlt -> Type
102 coreAltType (_,_,rhs) = exprType rhs
105 @mkPiType@ makes a (->) type or a forall type, depending on whether
106 it is given a type variable or a term variable. We cleverly use the
107 lbvarinfo field to figure out the right annotation for the arrove in
108 case of a term variable.
111 mkPiType :: Var -> Type -> Type -- The more polymorphic version
112 mkPiTypes :: [Var] -> Type -> Type -- doesn't work...
114 mkPiTypes vs ty = foldr mkPiType ty vs
117 | isId v = mkFunTy (idType v) ty
118 | otherwise = mkForAllTy v ty
122 applyTypeToArg :: Type -> CoreExpr -> Type
123 applyTypeToArg fun_ty (Type arg_ty) = applyTy fun_ty arg_ty
124 applyTypeToArg fun_ty other_arg = funResultTy fun_ty
126 applyTypeToArgs :: CoreExpr -> Type -> [CoreExpr] -> Type
127 -- A more efficient version of applyTypeToArg
128 -- when we have several args
129 -- The first argument is just for debugging
130 applyTypeToArgs e op_ty [] = op_ty
132 applyTypeToArgs e op_ty (Type ty : args)
133 = -- Accumulate type arguments so we can instantiate all at once
136 go rev_tys (Type ty : args) = go (ty:rev_tys) args
137 go rev_tys rest_args = applyTypeToArgs e op_ty' rest_args
139 op_ty' = applyTys op_ty (reverse rev_tys)
141 applyTypeToArgs e op_ty (other_arg : args)
142 = case (splitFunTy_maybe op_ty) of
143 Just (_, res_ty) -> applyTypeToArgs e res_ty args
144 Nothing -> pprPanic "applyTypeToArgs" (pprCoreExpr e $$ ppr op_ty)
149 %************************************************************************
151 \subsection{Attaching notes}
153 %************************************************************************
155 mkNote removes redundant coercions, and SCCs where possible
159 mkNote :: Note -> CoreExpr -> CoreExpr
160 mkNote (SCC cc) expr = mkSCC cc expr
161 mkNote InlineMe expr = mkInlineMe expr
162 mkNote note expr = Note note expr
166 Drop trivial InlineMe's. This is somewhat important, because if we have an unfolding
167 that looks like (Note InlineMe (Var v)), the InlineMe doesn't go away because it may
168 not be *applied* to anything.
170 We don't use exprIsTrivial here, though, because we sometimes generate worker/wrapper
173 f = inline_me (coerce t fw)
174 As usual, the inline_me prevents the worker from getting inlined back into the wrapper.
175 We want the split, so that the coerces can cancel at the call site.
177 However, we can get left with tiresome type applications. Notably, consider
178 f = /\ a -> let t = e in (t, w)
179 Then lifting the let out of the big lambda gives
181 f = /\ a -> let t = inline_me (t' a) in (t, w)
182 The inline_me is to stop the simplifier inlining t' right back
183 into t's RHS. In the next phase we'll substitute for t (since
184 its rhs is trivial) and *then* we could get rid of the inline_me.
185 But it hardly seems worth it, so I don't bother.
188 mkInlineMe (Var v) = Var v
189 mkInlineMe e = Note InlineMe e
195 mkCoerce :: Coercion -> CoreExpr -> CoreExpr
196 mkCoerce co (Cast expr co2)
197 = ASSERT(let { (from_ty, _to_ty) = coercionKind co;
198 (_from_ty2, to_ty2) = coercionKind co2} in
199 from_ty `coreEqType` to_ty2 )
200 mkCoerce (mkTransCoercion co2 co) expr
203 = let (from_ty, to_ty) = coercionKind co in
204 -- if to_ty `coreEqType` from_ty
207 ASSERT2(from_ty `coreEqType` (exprType expr), text "Trying to coerce" <+> text "(" <> ppr expr $$ text "::" <+> ppr (exprType expr) <> text ")" $$ ppr co $$ ppr (coercionKindPredTy co))
212 mkSCC :: CostCentre -> Expr b -> Expr b
213 -- Note: Nested SCC's *are* preserved for the benefit of
214 -- cost centre stack profiling
215 mkSCC cc (Lit lit) = Lit lit
216 mkSCC cc (Lam x e) = Lam x (mkSCC cc e) -- Move _scc_ inside lambda
217 mkSCC cc (Note (SCC cc') e) = Note (SCC cc) (Note (SCC cc') e)
218 mkSCC cc (Note n e) = Note n (mkSCC cc e) -- Move _scc_ inside notes
219 mkSCC cc (Cast e co) = Cast (mkSCC cc e) co -- Move _scc_ inside cast
220 mkSCC cc expr = Note (SCC cc) expr
224 %************************************************************************
226 \subsection{Other expression construction}
228 %************************************************************************
231 bindNonRec :: Id -> CoreExpr -> CoreExpr -> CoreExpr
232 -- (bindNonRec x r b) produces either
235 -- case r of x { _DEFAULT_ -> b }
237 -- depending on whether x is unlifted or not
238 -- It's used by the desugarer to avoid building bindings
239 -- that give Core Lint a heart attack. Actually the simplifier
240 -- deals with them perfectly well.
242 bindNonRec bndr rhs body
243 | needsCaseBinding (idType bndr) rhs = Case rhs bndr (exprType body) [(DEFAULT,[],body)]
244 | otherwise = Let (NonRec bndr rhs) body
246 needsCaseBinding ty rhs = isUnLiftedType ty && not (exprOkForSpeculation rhs)
247 -- Make a case expression instead of a let
248 -- These can arise either from the desugarer,
249 -- or from beta reductions: (\x.e) (x +# y)
253 mkAltExpr :: AltCon -> [CoreBndr] -> [Type] -> CoreExpr
254 -- This guy constructs the value that the scrutinee must have
255 -- when you are in one particular branch of a case
256 mkAltExpr (DataAlt con) args inst_tys
257 = mkConApp con (map Type inst_tys ++ varsToCoreExprs args)
258 mkAltExpr (LitAlt lit) [] []
261 mkIfThenElse :: CoreExpr -> CoreExpr -> CoreExpr -> CoreExpr
262 mkIfThenElse guard then_expr else_expr
263 -- Not going to be refining, so okay to take the type of the "then" clause
264 = Case guard (mkWildId boolTy) (exprType then_expr)
265 [ (DataAlt falseDataCon, [], else_expr), -- Increasing order of tag!
266 (DataAlt trueDataCon, [], then_expr) ]
270 %************************************************************************
272 \subsection{Taking expressions apart}
274 %************************************************************************
276 The default alternative must be first, if it exists at all.
277 This makes it easy to find, though it makes matching marginally harder.
280 findDefault :: [CoreAlt] -> ([CoreAlt], Maybe CoreExpr)
281 findDefault ((DEFAULT,args,rhs) : alts) = ASSERT( null args ) (alts, Just rhs)
282 findDefault alts = (alts, Nothing)
284 findAlt :: AltCon -> [CoreAlt] -> CoreAlt
287 (deflt@(DEFAULT,_,_):alts) -> go alts deflt
288 other -> go alts panic_deflt
290 panic_deflt = pprPanic "Missing alternative" (ppr con $$ vcat (map ppr alts))
293 go (alt@(con1,_,_) : alts) deflt
294 = case con `cmpAltCon` con1 of
295 LT -> deflt -- Missed it already; the alts are in increasing order
297 GT -> ASSERT( not (con1 == DEFAULT) ) go alts deflt
299 isDefaultAlt :: CoreAlt -> Bool
300 isDefaultAlt (DEFAULT, _, _) = True
301 isDefaultAlt other = False
303 ---------------------------------
304 mergeAlts :: [CoreAlt] -> [CoreAlt] -> [CoreAlt]
305 -- Merge preserving order; alternatives in the first arg
306 -- shadow ones in the second
307 mergeAlts [] as2 = as2
308 mergeAlts as1 [] = as1
309 mergeAlts (a1:as1) (a2:as2)
310 = case a1 `cmpAlt` a2 of
311 LT -> a1 : mergeAlts as1 (a2:as2)
312 EQ -> a1 : mergeAlts as1 as2 -- Discard a2
313 GT -> a2 : mergeAlts (a1:as1) as2
317 %************************************************************************
319 \subsection{Figuring out things about expressions}
321 %************************************************************************
323 @exprIsTrivial@ is true of expressions we are unconditionally happy to
324 duplicate; simple variables and constants, and type
325 applications. Note that primop Ids aren't considered
328 @exprIsBottom@ is true of expressions that are guaranteed to diverge
331 There used to be a gruesome test for (hasNoBinding v) in the
333 exprIsTrivial (Var v) | hasNoBinding v = idArity v == 0
334 The idea here is that a constructor worker, like $wJust, is
335 really short for (\x -> $wJust x), becuase $wJust has no binding.
336 So it should be treated like a lambda. Ditto unsaturated primops.
337 But now constructor workers are not "have-no-binding" Ids. And
338 completely un-applied primops and foreign-call Ids are sufficiently
339 rare that I plan to allow them to be duplicated and put up with
342 SCC notes. We do not treat (_scc_ "foo" x) as trivial, because
343 a) it really generates code, (and a heap object when it's
344 a function arg) to capture the cost centre
345 b) see the note [SCC-and-exprIsTrivial] in Simplify.simplLazyBind
348 exprIsTrivial (Var v) = True -- See notes above
349 exprIsTrivial (Type _) = True
350 exprIsTrivial (Lit lit) = litIsTrivial lit
351 exprIsTrivial (App e arg) = not (isRuntimeArg arg) && exprIsTrivial e
352 exprIsTrivial (Note (SCC _) e) = False -- See notes above
353 exprIsTrivial (Note _ e) = exprIsTrivial e
354 exprIsTrivial (Cast e co) = exprIsTrivial e
355 exprIsTrivial (Lam b body) = not (isRuntimeVar b) && exprIsTrivial body
356 exprIsTrivial other = False
360 @exprIsDupable@ is true of expressions that can be duplicated at a modest
361 cost in code size. This will only happen in different case
362 branches, so there's no issue about duplicating work.
364 That is, exprIsDupable returns True of (f x) even if
365 f is very very expensive to call.
367 Its only purpose is to avoid fruitless let-binding
368 and then inlining of case join points
372 exprIsDupable (Type _) = True
373 exprIsDupable (Var v) = True
374 exprIsDupable (Lit lit) = litIsDupable lit
375 exprIsDupable (Note InlineMe e) = True
376 exprIsDupable (Note _ e) = exprIsDupable e
377 exprIsDupable (Cast e co) = exprIsDupable e
381 go (Var v) n_args = True
382 go (App f a) n_args = n_args < dupAppSize
385 go other n_args = False
388 dupAppSize = 4 -- Size of application we are prepared to duplicate
391 @exprIsCheap@ looks at a Core expression and returns \tr{True} if
392 it is obviously in weak head normal form, or is cheap to get to WHNF.
393 [Note that that's not the same as exprIsDupable; an expression might be
394 big, and hence not dupable, but still cheap.]
396 By ``cheap'' we mean a computation we're willing to:
397 push inside a lambda, or
398 inline at more than one place
399 That might mean it gets evaluated more than once, instead of being
400 shared. The main examples of things which aren't WHNF but are
405 (where e, and all the ei are cheap)
408 (where e and b are cheap)
411 (where op is a cheap primitive operator)
414 (because we are happy to substitute it inside a lambda)
416 Notice that a variable is considered 'cheap': we can push it inside a lambda,
417 because sharing will make sure it is only evaluated once.
420 exprIsCheap :: CoreExpr -> Bool
421 exprIsCheap (Lit lit) = True
422 exprIsCheap (Type _) = True
423 exprIsCheap (Var _) = True
424 exprIsCheap (Note InlineMe e) = True
425 exprIsCheap (Note _ e) = exprIsCheap e
426 exprIsCheap (Cast e co) = exprIsCheap e
427 exprIsCheap (Lam x e) = isRuntimeVar x || exprIsCheap e
428 exprIsCheap (Case e _ _ alts) = exprIsCheap e &&
429 and [exprIsCheap rhs | (_,_,rhs) <- alts]
430 -- Experimentally, treat (case x of ...) as cheap
431 -- (and case __coerce x etc.)
432 -- This improves arities of overloaded functions where
433 -- there is only dictionary selection (no construction) involved
434 exprIsCheap (Let (NonRec x _) e)
435 | isUnLiftedType (idType x) = exprIsCheap e
437 -- strict lets always have cheap right hand sides,
438 -- and do no allocation.
440 exprIsCheap other_expr -- Applications and variables
443 -- Accumulate value arguments, then decide
444 go (App f a) val_args | isRuntimeArg a = go f (a:val_args)
445 | otherwise = go f val_args
447 go (Var f) [] = True -- Just a type application of a variable
448 -- (f t1 t2 t3) counts as WHNF
450 = case globalIdDetails f of
451 RecordSelId {} -> go_sel args
452 ClassOpId _ -> go_sel args
453 PrimOpId op -> go_primop op args
455 DataConWorkId _ -> go_pap args
456 other | length args < idArity f -> go_pap args
458 other -> isBottomingId f
459 -- Application of a function which
460 -- always gives bottom; we treat this as cheap
461 -- because it certainly doesn't need to be shared!
463 go other args = False
466 go_pap args = all exprIsTrivial args
467 -- For constructor applications and primops, check that all
468 -- the args are trivial. We don't want to treat as cheap, say,
470 -- We'll put up with one constructor application, but not dozens
473 go_primop op args = primOpIsCheap op && all exprIsCheap args
474 -- In principle we should worry about primops
475 -- that return a type variable, since the result
476 -- might be applied to something, but I'm not going
477 -- to bother to check the number of args
480 go_sel [arg] = exprIsCheap arg -- I'm experimenting with making record selection
481 go_sel other = False -- look cheap, so we will substitute it inside a
482 -- lambda. Particularly for dictionary field selection.
483 -- BUT: Take care with (sel d x)! The (sel d) might be cheap, but
484 -- there's no guarantee that (sel d x) will be too. Hence (n_val_args == 1)
487 exprOkForSpeculation returns True of an expression that it is
489 * safe to evaluate even if normal order eval might not
490 evaluate the expression at all, or
492 * safe *not* to evaluate even if normal order would do so
496 the expression guarantees to terminate,
498 without raising an exception,
499 without causing a side effect (e.g. writing a mutable variable)
501 NB: if exprIsHNF e, then exprOkForSpecuation e
504 let x = case y# +# 1# of { r# -> I# r# }
507 case y# +# 1# of { r# ->
512 We can only do this if the (y+1) is ok for speculation: it has no
513 side effects, and can't diverge or raise an exception.
516 exprOkForSpeculation :: CoreExpr -> Bool
517 exprOkForSpeculation (Lit _) = True
518 exprOkForSpeculation (Type _) = True
519 exprOkForSpeculation (Var v) = isUnLiftedType (idType v)
520 exprOkForSpeculation (Note _ e) = exprOkForSpeculation e
521 exprOkForSpeculation (Cast e co) = exprOkForSpeculation e
522 exprOkForSpeculation other_expr
523 = case collectArgs other_expr of
524 (Var f, args) -> spec_ok (globalIdDetails f) args
528 spec_ok (DataConWorkId _) args
529 = True -- The strictness of the constructor has already
530 -- been expressed by its "wrapper", so we don't need
531 -- to take the arguments into account
533 spec_ok (PrimOpId op) args
534 | isDivOp op, -- Special case for dividing operations that fail
535 [arg1, Lit lit] <- args -- only if the divisor is zero
536 = not (isZeroLit lit) && exprOkForSpeculation arg1
537 -- Often there is a literal divisor, and this
538 -- can get rid of a thunk in an inner looop
541 = primOpOkForSpeculation op &&
542 all exprOkForSpeculation args
543 -- A bit conservative: we don't really need
544 -- to care about lazy arguments, but this is easy
546 spec_ok other args = False
548 isDivOp :: PrimOp -> Bool
549 -- True of dyadic operators that can fail
550 -- only if the second arg is zero
551 -- This function probably belongs in PrimOp, or even in
552 -- an automagically generated file.. but it's such a
553 -- special case I thought I'd leave it here for now.
554 isDivOp IntQuotOp = True
555 isDivOp IntRemOp = True
556 isDivOp WordQuotOp = True
557 isDivOp WordRemOp = True
558 isDivOp IntegerQuotRemOp = True
559 isDivOp IntegerDivModOp = True
560 isDivOp FloatDivOp = True
561 isDivOp DoubleDivOp = True
562 isDivOp other = False
567 exprIsBottom :: CoreExpr -> Bool -- True => definitely bottom
568 exprIsBottom e = go 0 e
570 -- n is the number of args
571 go n (Note _ e) = go n e
572 go n (Cast e co) = go n e
573 go n (Let _ e) = go n e
574 go n (Case e _ _ _) = go 0 e -- Just check the scrut
575 go n (App e _) = go (n+1) e
576 go n (Var v) = idAppIsBottom v n
578 go n (Lam _ _) = False
579 go n (Type _) = False
581 idAppIsBottom :: Id -> Int -> Bool
582 idAppIsBottom id n_val_args = appIsBottom (idNewStrictness id) n_val_args
585 @exprIsHNF@ returns true for expressions that are certainly *already*
586 evaluated to *head* normal form. This is used to decide whether it's ok
589 case x of _ -> e ===> e
591 and to decide whether it's safe to discard a `seq`
593 So, it does *not* treat variables as evaluated, unless they say they are.
595 But it *does* treat partial applications and constructor applications
596 as values, even if their arguments are non-trivial, provided the argument
598 e.g. (:) (f x) (map f xs) is a value
599 map (...redex...) is a value
600 Because `seq` on such things completes immediately
602 For unlifted argument types, we have to be careful:
604 Suppose (f x) diverges; then C (f x) is not a value. True, but
605 this form is illegal (see the invariants in CoreSyn). Args of unboxed
606 type must be ok-for-speculation (or trivial).
609 exprIsHNF :: CoreExpr -> Bool -- True => Value-lambda, constructor, PAP
610 exprIsHNF (Var v) -- NB: There are no value args at this point
611 = isDataConWorkId v -- Catches nullary constructors,
612 -- so that [] and () are values, for example
613 || idArity v > 0 -- Catches (e.g.) primops that don't have unfoldings
614 || isEvaldUnfolding (idUnfolding v)
615 -- Check the thing's unfolding; it might be bound to a value
616 -- A worry: what if an Id's unfolding is just itself:
617 -- then we could get an infinite loop...
619 exprIsHNF (Lit l) = True
620 exprIsHNF (Type ty) = True -- Types are honorary Values;
621 -- we don't mind copying them
622 exprIsHNF (Lam b e) = isRuntimeVar b || exprIsHNF e
623 exprIsHNF (Note _ e) = exprIsHNF e
624 exprIsHNF (Cast e co) = exprIsHNF e
625 exprIsHNF (App e (Type _)) = exprIsHNF e
626 exprIsHNF (App e a) = app_is_value e [a]
627 exprIsHNF other = False
629 -- There is at least one value argument
630 app_is_value (Var fun) args
631 | isDataConWorkId fun -- Constructor apps are values
632 || idArity fun > valArgCount args -- Under-applied function
633 = check_args (idType fun) args
634 app_is_value (App f a) as = app_is_value f (a:as)
635 app_is_value other as = False
637 -- 'check_args' checks that unlifted-type args
638 -- are in fact guaranteed non-divergent
639 check_args fun_ty [] = True
640 check_args fun_ty (Type _ : args) = case splitForAllTy_maybe fun_ty of
641 Just (_, ty) -> check_args ty args
642 check_args fun_ty (arg : args)
643 | isUnLiftedType arg_ty = exprOkForSpeculation arg
644 | otherwise = check_args res_ty args
646 (arg_ty, res_ty) = splitFunTy fun_ty
650 -- These InstPat functions go here to avoid circularity between DataCon and Id
651 dataConRepInstPat = dataConInstPat dataConRepArgTys (repeat (FSLIT("ipv")))
652 dataConRepFSInstPat = dataConInstPat dataConRepArgTys
653 dataConOrigInstPat = dataConInstPat dc_arg_tys (repeat (FSLIT("ipv")))
655 dc_arg_tys dc = map mkPredTy (dataConTheta dc) ++ dataConOrigArgTys dc
656 -- Remember to include the existential dictionaries
658 dataConInstPat :: (DataCon -> [Type]) -- function used to find arg tys
659 -> [FastString] -- A long enough list of FSs to use for names
660 -> [Unique] -- An equally long list of uniques, at least one for each binder
662 -> [Type] -- Types to instantiate the universally quantified tyvars
663 -> ([TyVar], [CoVar], [Id]) -- Return instantiated variables
664 -- dataConInstPat arg_fun fss us con inst_tys returns a triple
665 -- (ex_tvs, co_tvs, arg_ids),
667 -- ex_tvs are intended to be used as binders for existential type args
669 -- co_tvs are intended to be used as binders for coercion args and the kinds
670 -- of these vars have been instantiated by the inst_tys and the ex_tys
672 -- arg_ids are indended to be used as binders for value arguments, including
673 -- dicts, and their types have been instantiated with inst_tys and ex_tys
676 -- The following constructor T1
679 -- T1 :: forall b. Int -> b -> T(a,b)
682 -- has representation type
683 -- forall a. forall a1. forall b. (a :=: (a1,b)) =>
686 -- dataConInstPat fss us T1 (a1',b') will return
688 -- ([a1'', b''], [c :: (a1', b'):=:(a1'', b'')], [x :: Int, y :: b''])
690 -- where the double-primed variables are created with the FastStrings and
691 -- Uniques given as fss and us
692 dataConInstPat arg_fun fss uniqs con inst_tys
693 = (ex_bndrs, co_bndrs, id_bndrs)
695 univ_tvs = dataConUnivTyVars con
696 ex_tvs = dataConExTyVars con
697 arg_tys = arg_fun con
698 eq_spec = dataConEqSpec con
699 eq_preds = eqSpecPreds eq_spec
702 n_co = length eq_spec
704 -- split the Uniques and FastStrings
705 (ex_uniqs, uniqs') = splitAt n_ex uniqs
706 (co_uniqs, id_uniqs) = splitAt n_co uniqs'
708 (ex_fss, fss') = splitAt n_ex fss
709 (co_fss, id_fss) = splitAt n_co fss'
711 -- Make existential type variables
712 ex_bndrs = zipWith3 mk_ex_var ex_uniqs ex_fss ex_tvs
713 mk_ex_var uniq fs var = mkTyVar new_name kind
715 new_name = mkSysTvName uniq fs
718 -- Make the instantiating substitution
719 subst = zipOpenTvSubst (univ_tvs ++ ex_tvs) (inst_tys ++ map mkTyVarTy ex_bndrs)
721 -- Make new coercion vars, instantiating kind
722 co_bndrs = zipWith3 mk_co_var co_uniqs co_fss eq_preds
723 mk_co_var uniq fs eq_pred = mkCoVar new_name co_kind
725 new_name = mkSysTvName uniq fs
726 co_kind = substTy subst (mkPredTy eq_pred)
728 -- make value vars, instantiating types
729 mk_id_var uniq fs ty = mkUserLocal (mkVarOccFS fs) uniq (substTy subst ty) noSrcLoc
730 id_bndrs = zipWith3 mk_id_var id_uniqs id_fss arg_tys
732 exprIsConApp_maybe :: CoreExpr -> Maybe (DataCon, [CoreExpr])
733 -- Returns (Just (dc, [x1..xn])) if the argument expression is
734 -- a constructor application of the form (dc x1 .. xn)
735 exprIsConApp_maybe (Cast expr co)
736 = -- Here we do the PushC reduction rule as described in the FC paper
737 case exprIsConApp_maybe expr of {
739 Just (dc, dc_args) ->
741 -- The transformation applies iff we have
742 -- (C e1 ... en) `cast` co
743 -- where co :: (T t1 .. tn) :=: (T s1 ..sn)
744 -- That is, with a T at the top of both sides
745 -- The left-hand one must be a T, because exprIsConApp returned True
746 -- but the right-hand one might not be. (Though it usually will.)
748 let (from_ty, to_ty) = coercionKind co
749 (from_tc, from_tc_arg_tys) = splitTyConApp from_ty
750 -- The inner one must be a TyConApp
752 case splitTyConApp_maybe to_ty of {
754 Just (to_tc, to_tc_arg_tys)
755 | from_tc /= to_tc -> Nothing
756 -- These two Nothing cases are possible; we might see
757 -- (C x y) `cast` (g :: T a ~ S [a]),
758 -- where S is a type function. In fact, exprIsConApp
759 -- will probably not be called in such circumstances,
760 -- but there't nothing wrong with it
764 tc_arity = tyConArity from_tc
766 (univ_args, rest1) = splitAt tc_arity dc_args
767 (ex_args, rest2) = splitAt n_ex_tvs rest1
768 (co_args, val_args) = splitAt n_cos rest2
770 arg_tys = dataConRepArgTys dc
771 dc_univ_tyvars = dataConUnivTyVars dc
772 dc_ex_tyvars = dataConExTyVars dc
773 dc_eq_spec = dataConEqSpec dc
774 dc_tyvars = dc_univ_tyvars ++ dc_ex_tyvars
775 n_ex_tvs = length dc_ex_tyvars
776 n_cos = length dc_eq_spec
778 -- Make the "theta" from Fig 3 of the paper
779 gammas = decomposeCo tc_arity co
780 new_tys = gammas ++ map (\ (Type t) -> t) ex_args
781 theta = zipOpenTvSubst dc_tyvars new_tys
783 -- First we cast the existential coercion arguments
784 cast_co (tv,ty) (Type co) = Type $ mkSymCoercion (substTyVar theta tv)
786 `mkTransCoercion` (substTy theta ty)
787 new_co_args = zipWith cast_co dc_eq_spec co_args
789 -- ...and now value arguments
790 new_val_args = zipWith cast_arg arg_tys val_args
791 cast_arg arg_ty arg = mkCoerce (substTy theta arg_ty) arg
794 ASSERT( length univ_args == tc_arity )
795 ASSERT( from_tc == dataConTyCon dc )
796 ASSERT( and (zipWith coreEqType [t | Type t <- univ_args] from_tc_arg_tys) )
797 ASSERT( all isTypeArg (univ_args ++ ex_args) )
798 ASSERT2( equalLength val_args arg_tys, ppr dc $$ ppr dc_tyvars $$ ppr dc_ex_tyvars $$ ppr arg_tys $$ ppr dc_args $$ ppr univ_args $$ ppr ex_args $$ ppr val_args $$ ppr arg_tys )
800 Just (dc, map Type to_tc_arg_tys ++ ex_args ++ new_co_args ++ new_val_args)
803 exprIsConApp_maybe (Note _ expr)
804 = exprIsConApp_maybe expr
805 -- We ignore InlineMe notes in case we have
806 -- x = __inline_me__ (a,b)
807 -- All part of making sure that INLINE pragmas never hurt
808 -- Marcin tripped on this one when making dictionaries more inlinable
810 -- In fact, we ignore all notes. For example,
811 -- case _scc_ "foo" (C a b) of
813 -- should be optimised away, but it will be only if we look
814 -- through the SCC note.
816 exprIsConApp_maybe expr = analyse (collectArgs expr)
818 analyse (Var fun, args)
819 | Just con <- isDataConWorkId_maybe fun,
820 args `lengthAtLeast` dataConRepArity con
821 -- Might be > because the arity excludes type args
824 -- Look through unfoldings, but only cheap ones, because
825 -- we are effectively duplicating the unfolding
826 analyse (Var fun, [])
827 | let unf = idUnfolding fun,
829 = exprIsConApp_maybe (unfoldingTemplate unf)
831 analyse other = Nothing
836 %************************************************************************
838 \subsection{Eta reduction and expansion}
840 %************************************************************************
843 exprEtaExpandArity :: DynFlags -> CoreExpr -> Arity
844 {- The Arity returned is the number of value args the
845 thing can be applied to without doing much work
847 exprEtaExpandArity is used when eta expanding
850 It returns 1 (or more) to:
851 case x of p -> \s -> ...
852 because for I/O ish things we really want to get that \s to the top.
853 We are prepared to evaluate x each time round the loop in order to get that
855 It's all a bit more subtle than it looks:
859 Consider one-shot lambdas
860 let x = expensive in \y z -> E
861 We want this to have arity 2 if the \y-abstraction is a 1-shot lambda
862 Hence the ArityType returned by arityType
864 2. The state-transformer hack
866 The one-shot lambda special cause is particularly important/useful for
867 IO state transformers, where we often get
868 let x = E in \ s -> ...
870 and the \s is a real-world state token abstraction. Such abstractions
871 are almost invariably 1-shot, so we want to pull the \s out, past the
872 let x=E, even if E is expensive. So we treat state-token lambdas as
873 one-shot even if they aren't really. The hack is in Id.isOneShotBndr.
875 3. Dealing with bottom
878 f = \x -> error "foo"
879 Here, arity 1 is fine. But if it is
883 then we want to get arity 2. Tecnically, this isn't quite right, because
885 should diverge, but it'll converge if we eta-expand f. Nevertheless, we
886 do so; it improves some programs significantly, and increasing convergence
887 isn't a bad thing. Hence the ABot/ATop in ArityType.
889 Actually, the situation is worse. Consider
893 Can we eta-expand here? At first the answer looks like "yes of course", but
896 This should diverge! But if we eta-expand, it won't. Again, we ignore this
897 "problem", because being scrupulous would lose an important transformation for
903 Non-recursive newtypes are transparent, and should not get in the way.
904 We do (currently) eta-expand recursive newtypes too. So if we have, say
906 newtype T = MkT ([T] -> Int)
910 where f has arity 1. Then: etaExpandArity e = 1;
911 that is, etaExpandArity looks through the coerce.
913 When we eta-expand e to arity 1: eta_expand 1 e T
914 we want to get: coerce T (\x::[T] -> (coerce ([T]->Int) e) x)
916 HOWEVER, note that if you use coerce bogusly you can ge
918 And since negate has arity 2, you might try to eta expand. But you can't
919 decopose Int to a function type. Hence the final case in eta_expand.
923 exprEtaExpandArity dflags e = arityDepth (arityType dflags e)
925 -- A limited sort of function type
926 data ArityType = AFun Bool ArityType -- True <=> one-shot
927 | ATop -- Know nothing
930 arityDepth :: ArityType -> Arity
931 arityDepth (AFun _ ty) = 1 + arityDepth ty
934 andArityType ABot at2 = at2
935 andArityType ATop at2 = ATop
936 andArityType (AFun t1 at1) (AFun t2 at2) = AFun (t1 && t2) (andArityType at1 at2)
937 andArityType at1 at2 = andArityType at2 at1
939 arityType :: DynFlags -> CoreExpr -> ArityType
940 -- (go1 e) = [b1,..,bn]
941 -- means expression can be rewritten \x_b1 -> ... \x_bn -> body
942 -- where bi is True <=> the lambda is one-shot
944 arityType dflags (Note n e) = arityType dflags e
945 -- Not needed any more: etaExpand is cleverer
946 -- | ok_note n = arityType dflags e
947 -- | otherwise = ATop
949 arityType dflags (Cast e co) = arityType dflags e
951 arityType dflags (Var v)
952 = mk (idArity v) (arg_tys (idType v))
954 mk :: Arity -> [Type] -> ArityType
955 -- The argument types are only to steer the "state hack"
956 -- Consider case x of
958 -- False -> \(s:RealWorld) -> e
959 -- where foo has arity 1. Then we want the state hack to
960 -- apply to foo too, so we can eta expand the case.
961 mk 0 tys | isBottomingId v = ABot
962 | (ty:tys) <- tys, isStateHackType ty = AFun True ATop
964 mk n (ty:tys) = AFun (isStateHackType ty) (mk (n-1) tys)
965 mk n [] = AFun False (mk (n-1) [])
967 arg_tys :: Type -> [Type] -- Ignore for-alls
969 | Just (_, ty') <- splitForAllTy_maybe ty = arg_tys ty'
970 | Just (arg,res) <- splitFunTy_maybe ty = arg : arg_tys res
973 -- Lambdas; increase arity
974 arityType dflags (Lam x e)
975 | isId x = AFun (isOneShotBndr x) (arityType dflags e)
976 | otherwise = arityType dflags e
978 -- Applications; decrease arity
979 arityType dflags (App f (Type _)) = arityType dflags f
980 arityType dflags (App f a) = case arityType dflags f of
981 AFun one_shot xs | exprIsCheap a -> xs
984 -- Case/Let; keep arity if either the expression is cheap
985 -- or it's a 1-shot lambda
986 -- The former is not really right for Haskell
987 -- f x = case x of { (a,b) -> \y. e }
989 -- f x y = case x of { (a,b) -> e }
990 -- The difference is observable using 'seq'
991 arityType dflags (Case scrut _ _ alts)
992 = case foldr1 andArityType [arityType dflags rhs | (_,_,rhs) <- alts] of
993 xs | exprIsCheap scrut -> xs
994 xs@(AFun one_shot _) | one_shot -> AFun True ATop
997 arityType dflags (Let b e)
998 = case arityType dflags e of
999 xs | cheap_bind b -> xs
1000 xs@(AFun one_shot _) | one_shot -> AFun True ATop
1003 cheap_bind (NonRec b e) = is_cheap (b,e)
1004 cheap_bind (Rec prs) = all is_cheap prs
1005 is_cheap (b,e) = (dopt Opt_DictsCheap dflags && isDictId b)
1007 -- If the experimental -fdicts-cheap flag is on, we eta-expand through
1008 -- dictionary bindings. This improves arities. Thereby, it also
1009 -- means that full laziness is less prone to floating out the
1010 -- application of a function to its dictionary arguments, which
1011 -- can thereby lose opportunities for fusion. Example:
1012 -- foo :: Ord a => a -> ...
1013 -- foo = /\a \(d:Ord a). let d' = ...d... in \(x:a). ....
1014 -- -- So foo has arity 1
1016 -- f = \x. foo dInt $ bar x
1018 -- The (foo DInt) is floated out, and makes ineffective a RULE
1019 -- foo (bar x) = ...
1021 -- One could go further and make exprIsCheap reply True to any
1022 -- dictionary-typed expression, but that's more work.
1024 arityType dflags other = ATop
1026 {- NOT NEEDED ANY MORE: etaExpand is cleverer
1027 ok_note InlineMe = False
1028 ok_note other = True
1029 -- Notice that we do not look through __inline_me__
1030 -- This may seem surprising, but consider
1031 -- f = _inline_me (\x -> e)
1032 -- We DO NOT want to eta expand this to
1033 -- f = \x -> (_inline_me (\x -> e)) x
1034 -- because the _inline_me gets dropped now it is applied,
1043 etaExpand :: Arity -- Result should have this number of value args
1045 -> CoreExpr -> Type -- Expression and its type
1047 -- (etaExpand n us e ty) returns an expression with
1048 -- the same meaning as 'e', but with arity 'n'.
1050 -- Given e' = etaExpand n us e ty
1052 -- ty = exprType e = exprType e'
1054 -- Note that SCCs are not treated specially. If we have
1055 -- etaExpand 2 (\x -> scc "foo" e)
1056 -- = (\xy -> (scc "foo" e) y)
1057 -- So the costs of evaluating 'e' (not 'e y') are attributed to "foo"
1059 etaExpand n us expr ty
1060 | manifestArity expr >= n = expr -- The no-op case
1062 = eta_expand n us expr ty
1065 -- manifestArity sees how many leading value lambdas there are
1066 manifestArity :: CoreExpr -> Arity
1067 manifestArity (Lam v e) | isId v = 1 + manifestArity e
1068 | otherwise = manifestArity e
1069 manifestArity (Note _ e) = manifestArity e
1070 manifestArity (Cast e _) = manifestArity e
1073 -- etaExpand deals with for-alls. For example:
1075 -- where E :: forall a. a -> a
1077 -- (/\b. \y::a -> E b y)
1079 -- It deals with coerces too, though they are now rare
1080 -- so perhaps the extra code isn't worth it
1082 eta_expand n us expr ty
1084 -- The ILX code generator requires eta expansion for type arguments
1085 -- too, but alas the 'n' doesn't tell us how many of them there
1086 -- may be. So we eagerly eta expand any big lambdas, and just
1087 -- cross our fingers about possible loss of sharing in the ILX case.
1088 -- The Right Thing is probably to make 'arity' include
1089 -- type variables throughout the compiler. (ToDo.)
1091 -- Saturated, so nothing to do
1094 -- Short cut for the case where there already
1095 -- is a lambda; no point in gratuitously adding more
1096 eta_expand n us (Lam v body) ty
1098 = Lam v (eta_expand n us body (applyTy ty (mkTyVarTy v)))
1101 = Lam v (eta_expand (n-1) us body (funResultTy ty))
1103 -- We used to have a special case that stepped inside Coerces here,
1104 -- thus: eta_expand n us (Note note@(Coerce _ ty) e) _
1105 -- = Note note (eta_expand n us e ty)
1106 -- BUT this led to an infinite loop
1107 -- Example: newtype T = MkT (Int -> Int)
1108 -- eta_expand 1 (coerce (Int->Int) e)
1109 -- --> coerce (Int->Int) (eta_expand 1 T e)
1111 -- --> coerce (Int->Int) (coerce T
1112 -- (\x::Int -> eta_expand 1 (coerce (Int->Int) e)))
1113 -- by the splitNewType_maybe case below
1116 eta_expand n us expr ty
1117 = ASSERT2 (exprType expr `coreEqType` ty, ppr (exprType expr) $$ ppr ty)
1118 case splitForAllTy_maybe ty of {
1121 Lam lam_tv (eta_expand n us2 (App expr (Type (mkTyVarTy lam_tv))) (substTyWith [tv] [mkTyVarTy lam_tv] ty'))
1123 lam_tv = setVarName tv (mkSysTvName uniq FSLIT("etaT"))
1124 -- Using tv as a base retains its tyvar/covar-ness
1128 case splitFunTy_maybe ty of {
1129 Just (arg_ty, res_ty) -> Lam arg1 (eta_expand (n-1) us2 (App expr (Var arg1)) res_ty)
1131 arg1 = mkSysLocal FSLIT("eta") uniq arg_ty
1137 -- newtype T = MkT ([T] -> Int)
1138 -- Consider eta-expanding this
1141 -- coerce T (\x::[T] -> (coerce ([T]->Int) e) x)
1143 case splitNewTypeRepCo_maybe ty of {
1145 mkCoerce (mkSymCoercion co) (eta_expand n us (mkCoerce co expr) ty1) ;
1148 -- We have an expression of arity > 0, but its type isn't a function
1149 -- This *can* legitmately happen: e.g. coerce Int (\x. x)
1150 -- Essentially the programmer is playing fast and loose with types
1151 -- (Happy does this a lot). So we simply decline to eta-expand.
1156 exprArity is a cheap-and-cheerful version of exprEtaExpandArity.
1157 It tells how many things the expression can be applied to before doing
1158 any work. It doesn't look inside cases, lets, etc. The idea is that
1159 exprEtaExpandArity will do the hard work, leaving something that's easy
1160 for exprArity to grapple with. In particular, Simplify uses exprArity to
1161 compute the ArityInfo for the Id.
1163 Originally I thought that it was enough just to look for top-level lambdas, but
1164 it isn't. I've seen this
1166 foo = PrelBase.timesInt
1168 We want foo to get arity 2 even though the eta-expander will leave it
1169 unchanged, in the expectation that it'll be inlined. But occasionally it
1170 isn't, because foo is blacklisted (used in a rule).
1172 Similarly, see the ok_note check in exprEtaExpandArity. So
1173 f = __inline_me (\x -> e)
1174 won't be eta-expanded.
1176 And in any case it seems more robust to have exprArity be a bit more intelligent.
1177 But note that (\x y z -> f x y z)
1178 should have arity 3, regardless of f's arity.
1181 exprArity :: CoreExpr -> Arity
1184 go (Var v) = idArity v
1185 go (Lam x e) | isId x = go e + 1
1187 go (Note n e) = go e
1188 go (Cast e _) = go e
1189 go (App e (Type t)) = go e
1190 go (App f a) | exprIsCheap a = (go f - 1) `max` 0
1191 -- NB: exprIsCheap a!
1192 -- f (fac x) does not have arity 2,
1193 -- even if f has arity 3!
1194 -- NB: `max 0`! (\x y -> f x) has arity 2, even if f is
1195 -- unknown, hence arity 0
1199 %************************************************************************
1201 \subsection{Equality}
1203 %************************************************************************
1205 @cheapEqExpr@ is a cheap equality test which bales out fast!
1206 True => definitely equal
1207 False => may or may not be equal
1210 cheapEqExpr :: Expr b -> Expr b -> Bool
1212 cheapEqExpr (Var v1) (Var v2) = v1==v2
1213 cheapEqExpr (Lit lit1) (Lit lit2) = lit1 == lit2
1214 cheapEqExpr (Type t1) (Type t2) = t1 `coreEqType` t2
1216 cheapEqExpr (App f1 a1) (App f2 a2)
1217 = f1 `cheapEqExpr` f2 && a1 `cheapEqExpr` a2
1219 cheapEqExpr _ _ = False
1221 exprIsBig :: Expr b -> Bool
1222 -- Returns True of expressions that are too big to be compared by cheapEqExpr
1223 exprIsBig (Lit _) = False
1224 exprIsBig (Var v) = False
1225 exprIsBig (Type t) = False
1226 exprIsBig (App f a) = exprIsBig f || exprIsBig a
1227 exprIsBig (Cast e _) = exprIsBig e -- Hopefully coercions are not too big!
1228 exprIsBig other = True
1233 tcEqExpr :: CoreExpr -> CoreExpr -> Bool
1234 -- Used in rule matching, so does *not* look through
1235 -- newtypes, predicate types; hence tcEqExpr
1237 tcEqExpr e1 e2 = tcEqExprX rn_env e1 e2
1239 rn_env = mkRnEnv2 (mkInScopeSet (exprFreeVars e1 `unionVarSet` exprFreeVars e2))
1241 tcEqExprX :: RnEnv2 -> CoreExpr -> CoreExpr -> Bool
1242 tcEqExprX env (Var v1) (Var v2) = rnOccL env v1 == rnOccR env v2
1243 tcEqExprX env (Lit lit1) (Lit lit2) = lit1 == lit2
1244 tcEqExprX env (App f1 a1) (App f2 a2) = tcEqExprX env f1 f2 && tcEqExprX env a1 a2
1245 tcEqExprX env (Lam v1 e1) (Lam v2 e2) = tcEqExprX (rnBndr2 env v1 v2) e1 e2
1246 tcEqExprX env (Let (NonRec v1 r1) e1)
1247 (Let (NonRec v2 r2) e2) = tcEqExprX env r1 r2
1248 && tcEqExprX (rnBndr2 env v1 v2) e1 e2
1249 tcEqExprX env (Let (Rec ps1) e1)
1250 (Let (Rec ps2) e2) = equalLength ps1 ps2
1251 && and (zipWith eq_rhs ps1 ps2)
1252 && tcEqExprX env' e1 e2
1254 env' = foldl2 rn_bndr2 env ps2 ps2
1255 rn_bndr2 env (b1,_) (b2,_) = rnBndr2 env b1 b2
1256 eq_rhs (_,r1) (_,r2) = tcEqExprX env' r1 r2
1257 tcEqExprX env (Case e1 v1 t1 a1)
1258 (Case e2 v2 t2 a2) = tcEqExprX env e1 e2
1259 && tcEqTypeX env t1 t2
1260 && equalLength a1 a2
1261 && and (zipWith (eq_alt env') a1 a2)
1263 env' = rnBndr2 env v1 v2
1265 tcEqExprX env (Note n1 e1) (Note n2 e2) = eq_note env n1 n2 && tcEqExprX env e1 e2
1266 tcEqExprX env (Cast e1 co1) (Cast e2 co2) = tcEqTypeX env co1 co2 && tcEqExprX env e1 e2
1267 tcEqExprX env (Type t1) (Type t2) = tcEqTypeX env t1 t2
1268 tcEqExprX env e1 e2 = False
1270 eq_alt env (c1,vs1,r1) (c2,vs2,r2) = c1==c2 && tcEqExprX (rnBndrs2 env vs1 vs2) r1 r2
1272 eq_note env (SCC cc1) (SCC cc2) = cc1 == cc2
1273 eq_note env (CoreNote s1) (CoreNote s2) = s1 == s2
1274 eq_note env other1 other2 = False
1278 %************************************************************************
1280 \subsection{The size of an expression}
1282 %************************************************************************
1285 coreBindsSize :: [CoreBind] -> Int
1286 coreBindsSize bs = foldr ((+) . bindSize) 0 bs
1288 exprSize :: CoreExpr -> Int
1289 -- A measure of the size of the expressions
1290 -- It also forces the expression pretty drastically as a side effect
1291 exprSize (Var v) = v `seq` 1
1292 exprSize (Lit lit) = lit `seq` 1
1293 exprSize (App f a) = exprSize f + exprSize a
1294 exprSize (Lam b e) = varSize b + exprSize e
1295 exprSize (Let b e) = bindSize b + exprSize e
1296 exprSize (Case e b t as) = seqType t `seq` exprSize e + varSize b + 1 + foldr ((+) . altSize) 0 as
1297 exprSize (Cast e co) = (seqType co `seq` 1) + exprSize e
1298 exprSize (Note n e) = noteSize n + exprSize e
1299 exprSize (Type t) = seqType t `seq` 1
1301 noteSize (SCC cc) = cc `seq` 1
1302 noteSize InlineMe = 1
1303 noteSize (CoreNote s) = s `seq` 1 -- hdaume: core annotations
1305 varSize :: Var -> Int
1306 varSize b | isTyVar b = 1
1307 | otherwise = seqType (idType b) `seq`
1308 megaSeqIdInfo (idInfo b) `seq`
1311 varsSize = foldr ((+) . varSize) 0
1313 bindSize (NonRec b e) = varSize b + exprSize e
1314 bindSize (Rec prs) = foldr ((+) . pairSize) 0 prs
1316 pairSize (b,e) = varSize b + exprSize e
1318 altSize (c,bs,e) = c `seq` varsSize bs + exprSize e
1322 %************************************************************************
1324 \subsection{Hashing}
1326 %************************************************************************
1329 hashExpr :: CoreExpr -> Int
1330 -- Two expressions that hash to the same Int may be equal (but may not be)
1331 -- Two expressions that hash to the different Ints are definitely unequal
1333 -- But "unequal" here means "not identical"; two alpha-equivalent
1334 -- expressions may hash to the different Ints
1336 -- The emphasis is on a crude, fast hash, rather than on high precision
1338 hashExpr e | hash < 0 = 77 -- Just in case we hit -maxInt
1341 hash = abs (hash_expr e) -- Negative numbers kill UniqFM
1343 hash_expr (Note _ e) = hash_expr e
1344 hash_expr (Cast e co) = hash_expr e
1345 hash_expr (Let (NonRec b r) e) = hashId b
1346 hash_expr (Let (Rec ((b,r):_)) e) = hashId b
1347 hash_expr (Case _ b _ _) = hashId b
1348 hash_expr (App f e) = hash_expr f * fast_hash_expr e
1349 hash_expr (Var v) = hashId v
1350 hash_expr (Lit lit) = hashLiteral lit
1351 hash_expr (Lam b _) = hashId b
1352 hash_expr (Type t) = trace "hash_expr: type" 1 -- Shouldn't happen
1354 fast_hash_expr (Var v) = hashId v
1355 fast_hash_expr (Lit lit) = hashLiteral lit
1356 fast_hash_expr (App f (Type _)) = fast_hash_expr f
1357 fast_hash_expr (App f a) = fast_hash_expr a
1358 fast_hash_expr (Lam b _) = hashId b
1359 fast_hash_expr other = 1
1362 hashId id = hashName (idName id)
1365 %************************************************************************
1367 \subsection{Determining non-updatable right-hand-sides}
1369 %************************************************************************
1371 Top-level constructor applications can usually be allocated
1372 statically, but they can't if the constructor, or any of the
1373 arguments, come from another DLL (because we can't refer to static
1374 labels in other DLLs).
1376 If this happens we simply make the RHS into an updatable thunk,
1377 and 'exectute' it rather than allocating it statically.
1380 rhsIsStatic :: PackageId -> CoreExpr -> Bool
1381 -- This function is called only on *top-level* right-hand sides
1382 -- Returns True if the RHS can be allocated statically, with
1383 -- no thunks involved at all.
1385 -- It's called (i) in TidyPgm.hasCafRefs to decide if the rhs is, or
1386 -- refers to, CAFs; and (ii) in CoreToStg to decide whether to put an
1387 -- update flag on it.
1389 -- The basic idea is that rhsIsStatic returns True only if the RHS is
1390 -- (a) a value lambda
1391 -- (b) a saturated constructor application with static args
1393 -- BUT watch out for
1394 -- (i) Any cross-DLL references kill static-ness completely
1395 -- because they must be 'executed' not statically allocated
1396 -- ("DLL" here really only refers to Windows DLLs, on other platforms,
1397 -- this is not necessary)
1399 -- (ii) We treat partial applications as redexes, because in fact we
1400 -- make a thunk for them that runs and builds a PAP
1401 -- at run-time. The only appliations that are treated as
1402 -- static are *saturated* applications of constructors.
1404 -- We used to try to be clever with nested structures like this:
1405 -- ys = (:) w ((:) w [])
1406 -- on the grounds that CorePrep will flatten ANF-ise it later.
1407 -- But supporting this special case made the function much more
1408 -- complicated, because the special case only applies if there are no
1409 -- enclosing type lambdas:
1410 -- ys = /\ a -> Foo (Baz ([] a))
1411 -- Here the nested (Baz []) won't float out to top level in CorePrep.
1413 -- But in fact, even without -O, nested structures at top level are
1414 -- flattened by the simplifier, so we don't need to be super-clever here.
1418 -- f = \x::Int. x+7 TRUE
1419 -- p = (True,False) TRUE
1421 -- d = (fst p, False) FALSE because there's a redex inside
1422 -- (this particular one doesn't happen but...)
1424 -- h = D# (1.0## /## 2.0##) FALSE (redex again)
1425 -- n = /\a. Nil a TRUE
1427 -- t = /\a. (:) (case w a of ...) (Nil a) FALSE (redex)
1430 -- This is a bit like CoreUtils.exprIsHNF, with the following differences:
1431 -- a) scc "foo" (\x -> ...) is updatable (so we catch the right SCC)
1433 -- b) (C x xs), where C is a contructors is updatable if the application is
1436 -- c) don't look through unfolding of f in (f x).
1438 -- When opt_RuntimeTypes is on, we keep type lambdas and treat
1439 -- them as making the RHS re-entrant (non-updatable).
1441 rhsIsStatic this_pkg rhs = is_static False rhs
1443 is_static :: Bool -- True <=> in a constructor argument; must be atomic
1446 is_static False (Lam b e) = isRuntimeVar b || is_static False e
1448 is_static in_arg (Note (SCC _) e) = False
1449 is_static in_arg (Note _ e) = is_static in_arg e
1450 is_static in_arg (Cast e co) = is_static in_arg e
1452 is_static in_arg (Lit lit)
1454 MachLabel _ _ -> False
1456 -- A MachLabel (foreign import "&foo") in an argument
1457 -- prevents a constructor application from being static. The
1458 -- reason is that it might give rise to unresolvable symbols
1459 -- in the object file: under Linux, references to "weak"
1460 -- symbols from the data segment give rise to "unresolvable
1461 -- relocation" errors at link time This might be due to a bug
1462 -- in the linker, but we'll work around it here anyway.
1465 is_static in_arg other_expr = go other_expr 0
1467 go (Var f) n_val_args
1468 #if mingw32_TARGET_OS
1469 | not (isDllName this_pkg (idName f))
1471 = saturated_data_con f n_val_args
1472 || (in_arg && n_val_args == 0)
1473 -- A naked un-applied variable is *not* deemed a static RHS
1475 -- Reason: better to update so that the indirection gets shorted
1476 -- out, and the true value will be seen
1477 -- NB: if you change this, you'll break the invariant that THUNK_STATICs
1478 -- are always updatable. If you do so, make sure that non-updatable
1479 -- ones have enough space for their static link field!
1481 go (App f a) n_val_args
1482 | isTypeArg a = go f n_val_args
1483 | not in_arg && is_static True a = go f (n_val_args + 1)
1484 -- The (not in_arg) checks that we aren't in a constructor argument;
1485 -- if we are, we don't allow (value) applications of any sort
1487 -- NB. In case you wonder, args are sometimes not atomic. eg.
1488 -- x = D# (1.0## /## 2.0##)
1489 -- can't float because /## can fail.
1491 go (Note (SCC _) f) n_val_args = False
1492 go (Note _ f) n_val_args = go f n_val_args
1493 go (Cast e co) n_val_args = go e n_val_args
1495 go other n_val_args = False
1497 saturated_data_con f n_val_args
1498 = case isDataConWorkId_maybe f of
1499 Just dc -> n_val_args == dataConRepArity dc