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"
77 import GHC.Exts -- For `xori`
81 %************************************************************************
83 \subsection{Find the type of a Core atom/expression}
85 %************************************************************************
88 exprType :: CoreExpr -> Type
90 exprType (Var var) = idType var
91 exprType (Lit lit) = literalType lit
92 exprType (Let _ body) = exprType body
93 exprType (Case _ _ ty alts) = ty
95 = let (_, ty) = coercionKind co in ty
96 exprType (Note other_note e) = exprType e
97 exprType (Lam binder expr) = mkPiType binder (exprType expr)
99 = case collectArgs e of
100 (fun, args) -> applyTypeToArgs e (exprType fun) args
102 exprType other = pprTrace "exprType" (pprCoreExpr other) alphaTy
104 coreAltType :: CoreAlt -> Type
105 coreAltType (_,_,rhs) = exprType rhs
108 @mkPiType@ makes a (->) type or a forall type, depending on whether
109 it is given a type variable or a term variable. We cleverly use the
110 lbvarinfo field to figure out the right annotation for the arrove in
111 case of a term variable.
114 mkPiType :: Var -> Type -> Type -- The more polymorphic version
115 mkPiTypes :: [Var] -> Type -> Type -- doesn't work...
117 mkPiTypes vs ty = foldr mkPiType ty vs
120 | isId v = mkFunTy (idType v) ty
121 | otherwise = mkForAllTy v ty
125 applyTypeToArg :: Type -> CoreExpr -> Type
126 applyTypeToArg fun_ty (Type arg_ty) = applyTy fun_ty arg_ty
127 applyTypeToArg fun_ty other_arg = funResultTy fun_ty
129 applyTypeToArgs :: CoreExpr -> Type -> [CoreExpr] -> Type
130 -- A more efficient version of applyTypeToArg
131 -- when we have several args
132 -- The first argument is just for debugging
133 applyTypeToArgs e op_ty [] = op_ty
135 applyTypeToArgs e op_ty (Type ty : args)
136 = -- Accumulate type arguments so we can instantiate all at once
139 go rev_tys (Type ty : args) = go (ty:rev_tys) args
140 go rev_tys rest_args = applyTypeToArgs e op_ty' rest_args
142 op_ty' = applyTys op_ty (reverse rev_tys)
144 applyTypeToArgs e op_ty (other_arg : args)
145 = case (splitFunTy_maybe op_ty) of
146 Just (_, res_ty) -> applyTypeToArgs e res_ty args
147 Nothing -> pprPanic "applyTypeToArgs" (pprCoreExpr e $$ ppr op_ty)
152 %************************************************************************
154 \subsection{Attaching notes}
156 %************************************************************************
158 mkNote removes redundant coercions, and SCCs where possible
162 mkNote :: Note -> CoreExpr -> CoreExpr
163 mkNote (SCC cc) expr = mkSCC cc expr
164 mkNote InlineMe expr = mkInlineMe expr
165 mkNote note expr = Note note expr
169 Drop trivial InlineMe's. This is somewhat important, because if we have an unfolding
170 that looks like (Note InlineMe (Var v)), the InlineMe doesn't go away because it may
171 not be *applied* to anything.
173 We don't use exprIsTrivial here, though, because we sometimes generate worker/wrapper
176 f = inline_me (coerce t fw)
177 As usual, the inline_me prevents the worker from getting inlined back into the wrapper.
178 We want the split, so that the coerces can cancel at the call site.
180 However, we can get left with tiresome type applications. Notably, consider
181 f = /\ a -> let t = e in (t, w)
182 Then lifting the let out of the big lambda gives
184 f = /\ a -> let t = inline_me (t' a) in (t, w)
185 The inline_me is to stop the simplifier inlining t' right back
186 into t's RHS. In the next phase we'll substitute for t (since
187 its rhs is trivial) and *then* we could get rid of the inline_me.
188 But it hardly seems worth it, so I don't bother.
191 mkInlineMe (Var v) = Var v
192 mkInlineMe e = Note InlineMe e
198 mkCoerce :: Coercion -> CoreExpr -> CoreExpr
199 mkCoerce co (Cast expr co2)
200 = ASSERT(let { (from_ty, _to_ty) = coercionKind co;
201 (_from_ty2, to_ty2) = coercionKind co2} in
202 from_ty `coreEqType` to_ty2 )
203 mkCoerce (mkTransCoercion co2 co) expr
206 = let (from_ty, to_ty) = coercionKind co in
207 -- if to_ty `coreEqType` from_ty
210 ASSERT2(from_ty `coreEqType` (exprType expr), text "Trying to coerce" <+> text "(" <> ppr expr $$ text "::" <+> ppr (exprType expr) <> text ")" $$ ppr co $$ ppr (coercionKindPredTy co))
215 mkSCC :: CostCentre -> Expr b -> Expr b
216 -- Note: Nested SCC's *are* preserved for the benefit of
217 -- cost centre stack profiling
218 mkSCC cc (Lit lit) = Lit lit
219 mkSCC cc (Lam x e) = Lam x (mkSCC cc e) -- Move _scc_ inside lambda
220 mkSCC cc (Note (SCC cc') e) = Note (SCC cc) (Note (SCC cc') e)
221 mkSCC cc (Note n e) = Note n (mkSCC cc e) -- Move _scc_ inside notes
222 mkSCC cc (Cast e co) = Cast (mkSCC cc e) co -- Move _scc_ inside cast
223 mkSCC cc expr = Note (SCC cc) expr
227 %************************************************************************
229 \subsection{Other expression construction}
231 %************************************************************************
234 bindNonRec :: Id -> CoreExpr -> CoreExpr -> CoreExpr
235 -- (bindNonRec x r b) produces either
238 -- case r of x { _DEFAULT_ -> b }
240 -- depending on whether x is unlifted or not
241 -- It's used by the desugarer to avoid building bindings
242 -- that give Core Lint a heart attack. Actually the simplifier
243 -- deals with them perfectly well.
245 bindNonRec bndr rhs body
246 | needsCaseBinding (idType bndr) rhs = Case rhs bndr (exprType body) [(DEFAULT,[],body)]
247 | otherwise = Let (NonRec bndr rhs) body
249 needsCaseBinding ty rhs = isUnLiftedType ty && not (exprOkForSpeculation rhs)
250 -- Make a case expression instead of a let
251 -- These can arise either from the desugarer,
252 -- or from beta reductions: (\x.e) (x +# y)
256 mkAltExpr :: AltCon -> [CoreBndr] -> [Type] -> CoreExpr
257 -- This guy constructs the value that the scrutinee must have
258 -- when you are in one particular branch of a case
259 mkAltExpr (DataAlt con) args inst_tys
260 = mkConApp con (map Type inst_tys ++ varsToCoreExprs args)
261 mkAltExpr (LitAlt lit) [] []
264 mkIfThenElse :: CoreExpr -> CoreExpr -> CoreExpr -> CoreExpr
265 mkIfThenElse guard then_expr else_expr
266 -- Not going to be refining, so okay to take the type of the "then" clause
267 = Case guard (mkWildId boolTy) (exprType then_expr)
268 [ (DataAlt falseDataCon, [], else_expr), -- Increasing order of tag!
269 (DataAlt trueDataCon, [], then_expr) ]
273 %************************************************************************
275 \subsection{Taking expressions apart}
277 %************************************************************************
279 The default alternative must be first, if it exists at all.
280 This makes it easy to find, though it makes matching marginally harder.
283 findDefault :: [CoreAlt] -> ([CoreAlt], Maybe CoreExpr)
284 findDefault ((DEFAULT,args,rhs) : alts) = ASSERT( null args ) (alts, Just rhs)
285 findDefault alts = (alts, Nothing)
287 findAlt :: AltCon -> [CoreAlt] -> CoreAlt
290 (deflt@(DEFAULT,_,_):alts) -> go alts deflt
291 other -> go alts panic_deflt
293 panic_deflt = pprPanic "Missing alternative" (ppr con $$ vcat (map ppr alts))
296 go (alt@(con1,_,_) : alts) deflt
297 = case con `cmpAltCon` con1 of
298 LT -> deflt -- Missed it already; the alts are in increasing order
300 GT -> ASSERT( not (con1 == DEFAULT) ) go alts deflt
302 isDefaultAlt :: CoreAlt -> Bool
303 isDefaultAlt (DEFAULT, _, _) = True
304 isDefaultAlt other = False
306 ---------------------------------
307 mergeAlts :: [CoreAlt] -> [CoreAlt] -> [CoreAlt]
308 -- Merge preserving order; alternatives in the first arg
309 -- shadow ones in the second
310 mergeAlts [] as2 = as2
311 mergeAlts as1 [] = as1
312 mergeAlts (a1:as1) (a2:as2)
313 = case a1 `cmpAlt` a2 of
314 LT -> a1 : mergeAlts as1 (a2:as2)
315 EQ -> a1 : mergeAlts as1 as2 -- Discard a2
316 GT -> a2 : mergeAlts (a1:as1) as2
320 %************************************************************************
322 \subsection{Figuring out things about expressions}
324 %************************************************************************
326 @exprIsTrivial@ is true of expressions we are unconditionally happy to
327 duplicate; simple variables and constants, and type
328 applications. Note that primop Ids aren't considered
331 @exprIsBottom@ is true of expressions that are guaranteed to diverge
334 There used to be a gruesome test for (hasNoBinding v) in the
336 exprIsTrivial (Var v) | hasNoBinding v = idArity v == 0
337 The idea here is that a constructor worker, like $wJust, is
338 really short for (\x -> $wJust x), becuase $wJust has no binding.
339 So it should be treated like a lambda. Ditto unsaturated primops.
340 But now constructor workers are not "have-no-binding" Ids. And
341 completely un-applied primops and foreign-call Ids are sufficiently
342 rare that I plan to allow them to be duplicated and put up with
345 SCC notes. We do not treat (_scc_ "foo" x) as trivial, because
346 a) it really generates code, (and a heap object when it's
347 a function arg) to capture the cost centre
348 b) see the note [SCC-and-exprIsTrivial] in Simplify.simplLazyBind
351 exprIsTrivial (Var v) = True -- See notes above
352 exprIsTrivial (Type _) = True
353 exprIsTrivial (Lit lit) = litIsTrivial lit
354 exprIsTrivial (App e arg) = not (isRuntimeArg arg) && exprIsTrivial e
355 exprIsTrivial (Note (SCC _) e) = False -- See notes above
356 exprIsTrivial (Note _ e) = exprIsTrivial e
357 exprIsTrivial (Cast e co) = exprIsTrivial e
358 exprIsTrivial (Lam b body) = not (isRuntimeVar b) && exprIsTrivial body
359 exprIsTrivial other = False
363 @exprIsDupable@ is true of expressions that can be duplicated at a modest
364 cost in code size. This will only happen in different case
365 branches, so there's no issue about duplicating work.
367 That is, exprIsDupable returns True of (f x) even if
368 f is very very expensive to call.
370 Its only purpose is to avoid fruitless let-binding
371 and then inlining of case join points
375 exprIsDupable (Type _) = True
376 exprIsDupable (Var v) = True
377 exprIsDupable (Lit lit) = litIsDupable lit
378 exprIsDupable (Note InlineMe e) = True
379 exprIsDupable (Note _ e) = exprIsDupable e
380 exprIsDupable (Cast e co) = exprIsDupable e
384 go (Var v) n_args = True
385 go (App f a) n_args = n_args < dupAppSize
388 go other n_args = False
391 dupAppSize = 4 -- Size of application we are prepared to duplicate
394 @exprIsCheap@ looks at a Core expression and returns \tr{True} if
395 it is obviously in weak head normal form, or is cheap to get to WHNF.
396 [Note that that's not the same as exprIsDupable; an expression might be
397 big, and hence not dupable, but still cheap.]
399 By ``cheap'' we mean a computation we're willing to:
400 push inside a lambda, or
401 inline at more than one place
402 That might mean it gets evaluated more than once, instead of being
403 shared. The main examples of things which aren't WHNF but are
408 (where e, and all the ei are cheap)
411 (where e and b are cheap)
414 (where op is a cheap primitive operator)
417 (because we are happy to substitute it inside a lambda)
419 Notice that a variable is considered 'cheap': we can push it inside a lambda,
420 because sharing will make sure it is only evaluated once.
423 exprIsCheap :: CoreExpr -> Bool
424 exprIsCheap (Lit lit) = True
425 exprIsCheap (Type _) = True
426 exprIsCheap (Var _) = True
427 exprIsCheap (Note InlineMe e) = True
428 exprIsCheap (Note _ e) = exprIsCheap e
429 exprIsCheap (Cast e co) = exprIsCheap e
430 exprIsCheap (Lam x e) = isRuntimeVar x || exprIsCheap e
431 exprIsCheap (Case e _ _ alts) = exprIsCheap e &&
432 and [exprIsCheap rhs | (_,_,rhs) <- alts]
433 -- Experimentally, treat (case x of ...) as cheap
434 -- (and case __coerce x etc.)
435 -- This improves arities of overloaded functions where
436 -- there is only dictionary selection (no construction) involved
437 exprIsCheap (Let (NonRec x _) e)
438 | isUnLiftedType (idType x) = exprIsCheap e
440 -- strict lets always have cheap right hand sides,
441 -- and do no allocation.
443 exprIsCheap other_expr -- Applications and variables
446 -- Accumulate value arguments, then decide
447 go (App f a) val_args | isRuntimeArg a = go f (a:val_args)
448 | otherwise = go f val_args
450 go (Var f) [] = True -- Just a type application of a variable
451 -- (f t1 t2 t3) counts as WHNF
453 = case globalIdDetails f of
454 RecordSelId {} -> go_sel args
455 ClassOpId _ -> go_sel args
456 PrimOpId op -> go_primop op args
458 DataConWorkId _ -> go_pap args
459 other | length args < idArity f -> go_pap args
461 other -> isBottomingId f
462 -- Application of a function which
463 -- always gives bottom; we treat this as cheap
464 -- because it certainly doesn't need to be shared!
466 go other args = False
469 go_pap args = all exprIsTrivial args
470 -- For constructor applications and primops, check that all
471 -- the args are trivial. We don't want to treat as cheap, say,
473 -- We'll put up with one constructor application, but not dozens
476 go_primop op args = primOpIsCheap op && all exprIsCheap args
477 -- In principle we should worry about primops
478 -- that return a type variable, since the result
479 -- might be applied to something, but I'm not going
480 -- to bother to check the number of args
483 go_sel [arg] = exprIsCheap arg -- I'm experimenting with making record selection
484 go_sel other = False -- look cheap, so we will substitute it inside a
485 -- lambda. Particularly for dictionary field selection.
486 -- BUT: Take care with (sel d x)! The (sel d) might be cheap, but
487 -- there's no guarantee that (sel d x) will be too. Hence (n_val_args == 1)
490 exprOkForSpeculation returns True of an expression that it is
492 * safe to evaluate even if normal order eval might not
493 evaluate the expression at all, or
495 * safe *not* to evaluate even if normal order would do so
499 the expression guarantees to terminate,
501 without raising an exception,
502 without causing a side effect (e.g. writing a mutable variable)
504 NB: if exprIsHNF e, then exprOkForSpecuation e
507 let x = case y# +# 1# of { r# -> I# r# }
510 case y# +# 1# of { r# ->
515 We can only do this if the (y+1) is ok for speculation: it has no
516 side effects, and can't diverge or raise an exception.
519 exprOkForSpeculation :: CoreExpr -> Bool
520 exprOkForSpeculation (Lit _) = True
521 exprOkForSpeculation (Type _) = True
522 exprOkForSpeculation (Var v) = isUnLiftedType (idType v)
523 exprOkForSpeculation (Note _ e) = exprOkForSpeculation e
524 exprOkForSpeculation (Cast e co) = exprOkForSpeculation e
525 exprOkForSpeculation other_expr
526 = case collectArgs other_expr of
527 (Var f, args) -> spec_ok (globalIdDetails f) args
531 spec_ok (DataConWorkId _) args
532 = True -- The strictness of the constructor has already
533 -- been expressed by its "wrapper", so we don't need
534 -- to take the arguments into account
536 spec_ok (PrimOpId op) args
537 | isDivOp op, -- Special case for dividing operations that fail
538 [arg1, Lit lit] <- args -- only if the divisor is zero
539 = not (isZeroLit lit) && exprOkForSpeculation arg1
540 -- Often there is a literal divisor, and this
541 -- can get rid of a thunk in an inner looop
544 = primOpOkForSpeculation op &&
545 all exprOkForSpeculation args
546 -- A bit conservative: we don't really need
547 -- to care about lazy arguments, but this is easy
549 spec_ok other args = False
551 isDivOp :: PrimOp -> Bool
552 -- True of dyadic operators that can fail
553 -- only if the second arg is zero
554 -- This function probably belongs in PrimOp, or even in
555 -- an automagically generated file.. but it's such a
556 -- special case I thought I'd leave it here for now.
557 isDivOp IntQuotOp = True
558 isDivOp IntRemOp = True
559 isDivOp WordQuotOp = True
560 isDivOp WordRemOp = True
561 isDivOp IntegerQuotRemOp = True
562 isDivOp IntegerDivModOp = True
563 isDivOp FloatDivOp = True
564 isDivOp DoubleDivOp = True
565 isDivOp other = False
570 exprIsBottom :: CoreExpr -> Bool -- True => definitely bottom
571 exprIsBottom e = go 0 e
573 -- n is the number of args
574 go n (Note _ e) = go n e
575 go n (Cast e co) = go n e
576 go n (Let _ e) = go n e
577 go n (Case e _ _ _) = go 0 e -- Just check the scrut
578 go n (App e _) = go (n+1) e
579 go n (Var v) = idAppIsBottom v n
581 go n (Lam _ _) = False
582 go n (Type _) = False
584 idAppIsBottom :: Id -> Int -> Bool
585 idAppIsBottom id n_val_args = appIsBottom (idNewStrictness id) n_val_args
588 @exprIsHNF@ returns true for expressions that are certainly *already*
589 evaluated to *head* normal form. This is used to decide whether it's ok
592 case x of _ -> e ===> e
594 and to decide whether it's safe to discard a `seq`
596 So, it does *not* treat variables as evaluated, unless they say they are.
598 But it *does* treat partial applications and constructor applications
599 as values, even if their arguments are non-trivial, provided the argument
601 e.g. (:) (f x) (map f xs) is a value
602 map (...redex...) is a value
603 Because `seq` on such things completes immediately
605 For unlifted argument types, we have to be careful:
607 Suppose (f x) diverges; then C (f x) is not a value. True, but
608 this form is illegal (see the invariants in CoreSyn). Args of unboxed
609 type must be ok-for-speculation (or trivial).
612 exprIsHNF :: CoreExpr -> Bool -- True => Value-lambda, constructor, PAP
613 exprIsHNF (Var v) -- NB: There are no value args at this point
614 = isDataConWorkId v -- Catches nullary constructors,
615 -- so that [] and () are values, for example
616 || idArity v > 0 -- Catches (e.g.) primops that don't have unfoldings
617 || isEvaldUnfolding (idUnfolding v)
618 -- Check the thing's unfolding; it might be bound to a value
619 -- A worry: what if an Id's unfolding is just itself:
620 -- then we could get an infinite loop...
622 exprIsHNF (Lit l) = True
623 exprIsHNF (Type ty) = True -- Types are honorary Values;
624 -- we don't mind copying them
625 exprIsHNF (Lam b e) = isRuntimeVar b || exprIsHNF e
626 exprIsHNF (Note _ e) = exprIsHNF e
627 exprIsHNF (Cast e co) = exprIsHNF e
628 exprIsHNF (App e (Type _)) = exprIsHNF e
629 exprIsHNF (App e a) = app_is_value e [a]
630 exprIsHNF other = False
632 -- There is at least one value argument
633 app_is_value (Var fun) args
634 | isDataConWorkId fun -- Constructor apps are values
635 || idArity fun > valArgCount args -- Under-applied function
636 = check_args (idType fun) args
637 app_is_value (App f a) as = app_is_value f (a:as)
638 app_is_value other as = False
640 -- 'check_args' checks that unlifted-type args
641 -- are in fact guaranteed non-divergent
642 check_args fun_ty [] = True
643 check_args fun_ty (Type _ : args) = case splitForAllTy_maybe fun_ty of
644 Just (_, ty) -> check_args ty args
645 check_args fun_ty (arg : args)
646 | isUnLiftedType arg_ty = exprOkForSpeculation arg
647 | otherwise = check_args res_ty args
649 (arg_ty, res_ty) = splitFunTy fun_ty
653 -- These InstPat functions go here to avoid circularity between DataCon and Id
654 dataConRepInstPat = dataConInstPat dataConRepArgTys (repeat (FSLIT("ipv")))
655 dataConRepFSInstPat = dataConInstPat dataConRepArgTys
656 dataConOrigInstPat = dataConInstPat dc_arg_tys (repeat (FSLIT("ipv")))
658 dc_arg_tys dc = map mkPredTy (dataConTheta dc) ++ dataConOrigArgTys dc
659 -- Remember to include the existential dictionaries
661 dataConInstPat :: (DataCon -> [Type]) -- function used to find arg tys
662 -> [FastString] -- A long enough list of FSs to use for names
663 -> [Unique] -- An equally long list of uniques, at least one for each binder
665 -> [Type] -- Types to instantiate the universally quantified tyvars
666 -> ([TyVar], [CoVar], [Id]) -- Return instantiated variables
667 -- dataConInstPat arg_fun fss us con inst_tys returns a triple
668 -- (ex_tvs, co_tvs, arg_ids),
670 -- ex_tvs are intended to be used as binders for existential type args
672 -- co_tvs are intended to be used as binders for coercion args and the kinds
673 -- of these vars have been instantiated by the inst_tys and the ex_tys
675 -- arg_ids are indended to be used as binders for value arguments, including
676 -- dicts, and their types have been instantiated with inst_tys and ex_tys
679 -- The following constructor T1
682 -- T1 :: forall b. Int -> b -> T(a,b)
685 -- has representation type
686 -- forall a. forall a1. forall b. (a :=: (a1,b)) =>
689 -- dataConInstPat fss us T1 (a1',b') will return
691 -- ([a1'', b''], [c :: (a1', b'):=:(a1'', b'')], [x :: Int, y :: b''])
693 -- where the double-primed variables are created with the FastStrings and
694 -- Uniques given as fss and us
695 dataConInstPat arg_fun fss uniqs con inst_tys
696 = (ex_bndrs, co_bndrs, id_bndrs)
698 univ_tvs = dataConUnivTyVars con
699 ex_tvs = dataConExTyVars con
700 arg_tys = arg_fun con
701 eq_spec = dataConEqSpec con
702 eq_preds = eqSpecPreds eq_spec
705 n_co = length eq_spec
707 -- split the Uniques and FastStrings
708 (ex_uniqs, uniqs') = splitAt n_ex uniqs
709 (co_uniqs, id_uniqs) = splitAt n_co uniqs'
711 (ex_fss, fss') = splitAt n_ex fss
712 (co_fss, id_fss) = splitAt n_co fss'
714 -- Make existential type variables
715 ex_bndrs = zipWith3 mk_ex_var ex_uniqs ex_fss ex_tvs
716 mk_ex_var uniq fs var = mkTyVar new_name kind
718 new_name = mkSysTvName uniq fs
721 -- Make the instantiating substitution
722 subst = zipOpenTvSubst (univ_tvs ++ ex_tvs) (inst_tys ++ map mkTyVarTy ex_bndrs)
724 -- Make new coercion vars, instantiating kind
725 co_bndrs = zipWith3 mk_co_var co_uniqs co_fss eq_preds
726 mk_co_var uniq fs eq_pred = mkCoVar new_name co_kind
728 new_name = mkSysTvName uniq fs
729 co_kind = substTy subst (mkPredTy eq_pred)
731 -- make value vars, instantiating types
732 mk_id_var uniq fs ty = mkUserLocal (mkVarOccFS fs) uniq (substTy subst ty) noSrcLoc
733 id_bndrs = zipWith3 mk_id_var id_uniqs id_fss arg_tys
735 exprIsConApp_maybe :: CoreExpr -> Maybe (DataCon, [CoreExpr])
736 -- Returns (Just (dc, [x1..xn])) if the argument expression is
737 -- a constructor application of the form (dc x1 .. xn)
738 exprIsConApp_maybe (Cast expr co)
739 = -- Here we do the PushC reduction rule as described in the FC paper
740 case exprIsConApp_maybe expr of {
742 Just (dc, dc_args) ->
744 -- The transformation applies iff we have
745 -- (C e1 ... en) `cast` co
746 -- where co :: (T t1 .. tn) :=: (T s1 ..sn)
747 -- That is, with a T at the top of both sides
748 -- The left-hand one must be a T, because exprIsConApp returned True
749 -- but the right-hand one might not be. (Though it usually will.)
751 let (from_ty, to_ty) = coercionKind co
752 (from_tc, from_tc_arg_tys) = splitTyConApp from_ty
753 -- The inner one must be a TyConApp
755 case splitTyConApp_maybe to_ty of {
757 Just (to_tc, to_tc_arg_tys)
758 | from_tc /= to_tc -> Nothing
759 -- These two Nothing cases are possible; we might see
760 -- (C x y) `cast` (g :: T a ~ S [a]),
761 -- where S is a type function. In fact, exprIsConApp
762 -- will probably not be called in such circumstances,
763 -- but there't nothing wrong with it
767 tc_arity = tyConArity from_tc
769 (univ_args, rest1) = splitAt tc_arity dc_args
770 (ex_args, rest2) = splitAt n_ex_tvs rest1
771 (co_args, val_args) = splitAt n_cos rest2
773 arg_tys = dataConRepArgTys dc
774 dc_univ_tyvars = dataConUnivTyVars dc
775 dc_ex_tyvars = dataConExTyVars dc
776 dc_eq_spec = dataConEqSpec dc
777 dc_tyvars = dc_univ_tyvars ++ dc_ex_tyvars
778 n_ex_tvs = length dc_ex_tyvars
779 n_cos = length dc_eq_spec
781 -- Make the "theta" from Fig 3 of the paper
782 gammas = decomposeCo tc_arity co
783 new_tys = gammas ++ map (\ (Type t) -> t) ex_args
784 theta = zipOpenTvSubst dc_tyvars new_tys
786 -- First we cast the existential coercion arguments
787 cast_co (tv,ty) (Type co) = Type $ mkSymCoercion (substTyVar theta tv)
789 `mkTransCoercion` (substTy theta ty)
790 new_co_args = zipWith cast_co dc_eq_spec co_args
792 -- ...and now value arguments
793 new_val_args = zipWith cast_arg arg_tys val_args
794 cast_arg arg_ty arg = mkCoerce (substTy theta arg_ty) arg
797 ASSERT( length univ_args == tc_arity )
798 ASSERT( from_tc == dataConTyCon dc )
799 ASSERT( and (zipWith coreEqType [t | Type t <- univ_args] from_tc_arg_tys) )
800 ASSERT( all isTypeArg (univ_args ++ ex_args) )
801 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 )
803 Just (dc, map Type to_tc_arg_tys ++ ex_args ++ new_co_args ++ new_val_args)
806 exprIsConApp_maybe (Note _ expr)
807 = exprIsConApp_maybe expr
808 -- We ignore InlineMe notes in case we have
809 -- x = __inline_me__ (a,b)
810 -- All part of making sure that INLINE pragmas never hurt
811 -- Marcin tripped on this one when making dictionaries more inlinable
813 -- In fact, we ignore all notes. For example,
814 -- case _scc_ "foo" (C a b) of
816 -- should be optimised away, but it will be only if we look
817 -- through the SCC note.
819 exprIsConApp_maybe expr = analyse (collectArgs expr)
821 analyse (Var fun, args)
822 | Just con <- isDataConWorkId_maybe fun,
823 args `lengthAtLeast` dataConRepArity con
824 -- Might be > because the arity excludes type args
827 -- Look through unfoldings, but only cheap ones, because
828 -- we are effectively duplicating the unfolding
829 analyse (Var fun, [])
830 | let unf = idUnfolding fun,
832 = exprIsConApp_maybe (unfoldingTemplate unf)
834 analyse other = Nothing
839 %************************************************************************
841 \subsection{Eta reduction and expansion}
843 %************************************************************************
846 exprEtaExpandArity :: DynFlags -> CoreExpr -> Arity
847 {- The Arity returned is the number of value args the
848 thing can be applied to without doing much work
850 exprEtaExpandArity is used when eta expanding
853 It returns 1 (or more) to:
854 case x of p -> \s -> ...
855 because for I/O ish things we really want to get that \s to the top.
856 We are prepared to evaluate x each time round the loop in order to get that
858 It's all a bit more subtle than it looks:
862 Consider one-shot lambdas
863 let x = expensive in \y z -> E
864 We want this to have arity 2 if the \y-abstraction is a 1-shot lambda
865 Hence the ArityType returned by arityType
867 2. The state-transformer hack
869 The one-shot lambda special cause is particularly important/useful for
870 IO state transformers, where we often get
871 let x = E in \ s -> ...
873 and the \s is a real-world state token abstraction. Such abstractions
874 are almost invariably 1-shot, so we want to pull the \s out, past the
875 let x=E, even if E is expensive. So we treat state-token lambdas as
876 one-shot even if they aren't really. The hack is in Id.isOneShotBndr.
878 3. Dealing with bottom
881 f = \x -> error "foo"
882 Here, arity 1 is fine. But if it is
886 then we want to get arity 2. Tecnically, this isn't quite right, because
888 should diverge, but it'll converge if we eta-expand f. Nevertheless, we
889 do so; it improves some programs significantly, and increasing convergence
890 isn't a bad thing. Hence the ABot/ATop in ArityType.
892 Actually, the situation is worse. Consider
896 Can we eta-expand here? At first the answer looks like "yes of course", but
899 This should diverge! But if we eta-expand, it won't. Again, we ignore this
900 "problem", because being scrupulous would lose an important transformation for
906 Non-recursive newtypes are transparent, and should not get in the way.
907 We do (currently) eta-expand recursive newtypes too. So if we have, say
909 newtype T = MkT ([T] -> Int)
913 where f has arity 1. Then: etaExpandArity e = 1;
914 that is, etaExpandArity looks through the coerce.
916 When we eta-expand e to arity 1: eta_expand 1 e T
917 we want to get: coerce T (\x::[T] -> (coerce ([T]->Int) e) x)
919 HOWEVER, note that if you use coerce bogusly you can ge
921 And since negate has arity 2, you might try to eta expand. But you can't
922 decopose Int to a function type. Hence the final case in eta_expand.
926 exprEtaExpandArity dflags e = arityDepth (arityType dflags e)
928 -- A limited sort of function type
929 data ArityType = AFun Bool ArityType -- True <=> one-shot
930 | ATop -- Know nothing
933 arityDepth :: ArityType -> Arity
934 arityDepth (AFun _ ty) = 1 + arityDepth ty
937 andArityType ABot at2 = at2
938 andArityType ATop at2 = ATop
939 andArityType (AFun t1 at1) (AFun t2 at2) = AFun (t1 && t2) (andArityType at1 at2)
940 andArityType at1 at2 = andArityType at2 at1
942 arityType :: DynFlags -> CoreExpr -> ArityType
943 -- (go1 e) = [b1,..,bn]
944 -- means expression can be rewritten \x_b1 -> ... \x_bn -> body
945 -- where bi is True <=> the lambda is one-shot
947 arityType dflags (Note n e) = arityType dflags e
948 -- Not needed any more: etaExpand is cleverer
949 -- | ok_note n = arityType dflags e
950 -- | otherwise = ATop
952 arityType dflags (Cast e co) = arityType dflags e
954 arityType dflags (Var v)
955 = mk (idArity v) (arg_tys (idType v))
957 mk :: Arity -> [Type] -> ArityType
958 -- The argument types are only to steer the "state hack"
959 -- Consider case x of
961 -- False -> \(s:RealWorld) -> e
962 -- where foo has arity 1. Then we want the state hack to
963 -- apply to foo too, so we can eta expand the case.
964 mk 0 tys | isBottomingId v = ABot
965 | (ty:tys) <- tys, isStateHackType ty = AFun True ATop
967 mk n (ty:tys) = AFun (isStateHackType ty) (mk (n-1) tys)
968 mk n [] = AFun False (mk (n-1) [])
970 arg_tys :: Type -> [Type] -- Ignore for-alls
972 | Just (_, ty') <- splitForAllTy_maybe ty = arg_tys ty'
973 | Just (arg,res) <- splitFunTy_maybe ty = arg : arg_tys res
976 -- Lambdas; increase arity
977 arityType dflags (Lam x e)
978 | isId x = AFun (isOneShotBndr x) (arityType dflags e)
979 | otherwise = arityType dflags e
981 -- Applications; decrease arity
982 arityType dflags (App f (Type _)) = arityType dflags f
983 arityType dflags (App f a) = case arityType dflags f of
984 AFun one_shot xs | exprIsCheap a -> xs
987 -- Case/Let; keep arity if either the expression is cheap
988 -- or it's a 1-shot lambda
989 -- The former is not really right for Haskell
990 -- f x = case x of { (a,b) -> \y. e }
992 -- f x y = case x of { (a,b) -> e }
993 -- The difference is observable using 'seq'
994 arityType dflags (Case scrut _ _ alts)
995 = case foldr1 andArityType [arityType dflags rhs | (_,_,rhs) <- alts] of
996 xs | exprIsCheap scrut -> xs
997 xs@(AFun one_shot _) | one_shot -> AFun True ATop
1000 arityType dflags (Let b e)
1001 = case arityType dflags e of
1002 xs | cheap_bind b -> xs
1003 xs@(AFun one_shot _) | one_shot -> AFun True ATop
1006 cheap_bind (NonRec b e) = is_cheap (b,e)
1007 cheap_bind (Rec prs) = all is_cheap prs
1008 is_cheap (b,e) = (dopt Opt_DictsCheap dflags && isDictId b)
1010 -- If the experimental -fdicts-cheap flag is on, we eta-expand through
1011 -- dictionary bindings. This improves arities. Thereby, it also
1012 -- means that full laziness is less prone to floating out the
1013 -- application of a function to its dictionary arguments, which
1014 -- can thereby lose opportunities for fusion. Example:
1015 -- foo :: Ord a => a -> ...
1016 -- foo = /\a \(d:Ord a). let d' = ...d... in \(x:a). ....
1017 -- -- So foo has arity 1
1019 -- f = \x. foo dInt $ bar x
1021 -- The (foo DInt) is floated out, and makes ineffective a RULE
1022 -- foo (bar x) = ...
1024 -- One could go further and make exprIsCheap reply True to any
1025 -- dictionary-typed expression, but that's more work.
1027 arityType dflags other = ATop
1029 {- NOT NEEDED ANY MORE: etaExpand is cleverer
1030 ok_note InlineMe = False
1031 ok_note other = True
1032 -- Notice that we do not look through __inline_me__
1033 -- This may seem surprising, but consider
1034 -- f = _inline_me (\x -> e)
1035 -- We DO NOT want to eta expand this to
1036 -- f = \x -> (_inline_me (\x -> e)) x
1037 -- because the _inline_me gets dropped now it is applied,
1046 etaExpand :: Arity -- Result should have this number of value args
1048 -> CoreExpr -> Type -- Expression and its type
1050 -- (etaExpand n us e ty) returns an expression with
1051 -- the same meaning as 'e', but with arity 'n'.
1053 -- Given e' = etaExpand n us e ty
1055 -- ty = exprType e = exprType e'
1057 -- Note that SCCs are not treated specially. If we have
1058 -- etaExpand 2 (\x -> scc "foo" e)
1059 -- = (\xy -> (scc "foo" e) y)
1060 -- So the costs of evaluating 'e' (not 'e y') are attributed to "foo"
1062 etaExpand n us expr ty
1063 | manifestArity expr >= n = expr -- The no-op case
1065 = eta_expand n us expr ty
1068 -- manifestArity sees how many leading value lambdas there are
1069 manifestArity :: CoreExpr -> Arity
1070 manifestArity (Lam v e) | isId v = 1 + manifestArity e
1071 | otherwise = manifestArity e
1072 manifestArity (Note _ e) = manifestArity e
1073 manifestArity (Cast e _) = manifestArity e
1076 -- etaExpand deals with for-alls. For example:
1078 -- where E :: forall a. a -> a
1080 -- (/\b. \y::a -> E b y)
1082 -- It deals with coerces too, though they are now rare
1083 -- so perhaps the extra code isn't worth it
1085 eta_expand n us expr ty
1087 -- The ILX code generator requires eta expansion for type arguments
1088 -- too, but alas the 'n' doesn't tell us how many of them there
1089 -- may be. So we eagerly eta expand any big lambdas, and just
1090 -- cross our fingers about possible loss of sharing in the ILX case.
1091 -- The Right Thing is probably to make 'arity' include
1092 -- type variables throughout the compiler. (ToDo.)
1094 -- Saturated, so nothing to do
1097 -- Short cut for the case where there already
1098 -- is a lambda; no point in gratuitously adding more
1099 eta_expand n us (Lam v body) ty
1101 = Lam v (eta_expand n us body (applyTy ty (mkTyVarTy v)))
1104 = Lam v (eta_expand (n-1) us body (funResultTy ty))
1106 -- We used to have a special case that stepped inside Coerces here,
1107 -- thus: eta_expand n us (Note note@(Coerce _ ty) e) _
1108 -- = Note note (eta_expand n us e ty)
1109 -- BUT this led to an infinite loop
1110 -- Example: newtype T = MkT (Int -> Int)
1111 -- eta_expand 1 (coerce (Int->Int) e)
1112 -- --> coerce (Int->Int) (eta_expand 1 T e)
1114 -- --> coerce (Int->Int) (coerce T
1115 -- (\x::Int -> eta_expand 1 (coerce (Int->Int) e)))
1116 -- by the splitNewType_maybe case below
1119 eta_expand n us expr ty
1120 = ASSERT2 (exprType expr `coreEqType` ty, ppr (exprType expr) $$ ppr ty)
1121 case splitForAllTy_maybe ty of {
1124 Lam lam_tv (eta_expand n us2 (App expr (Type (mkTyVarTy lam_tv))) (substTyWith [tv] [mkTyVarTy lam_tv] ty'))
1126 lam_tv = mkTyVar (mkSysTvName uniq FSLIT("etaT")) (tyVarKind tv)
1130 case splitFunTy_maybe ty of {
1131 Just (arg_ty, res_ty) -> Lam arg1 (eta_expand (n-1) us2 (App expr (Var arg1)) res_ty)
1133 arg1 = mkSysLocal FSLIT("eta") uniq arg_ty
1139 -- newtype T = MkT ([T] -> Int)
1140 -- Consider eta-expanding this
1143 -- coerce T (\x::[T] -> (coerce ([T]->Int) e) x)
1145 case splitNewTypeRepCo_maybe ty of {
1147 mkCoerce (mkSymCoercion co) (eta_expand n us (mkCoerce co expr) ty1) ;
1150 -- We have an expression of arity > 0, but its type isn't a function
1151 -- This *can* legitmately happen: e.g. coerce Int (\x. x)
1152 -- Essentially the programmer is playing fast and loose with types
1153 -- (Happy does this a lot). So we simply decline to eta-expand.
1158 exprArity is a cheap-and-cheerful version of exprEtaExpandArity.
1159 It tells how many things the expression can be applied to before doing
1160 any work. It doesn't look inside cases, lets, etc. The idea is that
1161 exprEtaExpandArity will do the hard work, leaving something that's easy
1162 for exprArity to grapple with. In particular, Simplify uses exprArity to
1163 compute the ArityInfo for the Id.
1165 Originally I thought that it was enough just to look for top-level lambdas, but
1166 it isn't. I've seen this
1168 foo = PrelBase.timesInt
1170 We want foo to get arity 2 even though the eta-expander will leave it
1171 unchanged, in the expectation that it'll be inlined. But occasionally it
1172 isn't, because foo is blacklisted (used in a rule).
1174 Similarly, see the ok_note check in exprEtaExpandArity. So
1175 f = __inline_me (\x -> e)
1176 won't be eta-expanded.
1178 And in any case it seems more robust to have exprArity be a bit more intelligent.
1179 But note that (\x y z -> f x y z)
1180 should have arity 3, regardless of f's arity.
1183 exprArity :: CoreExpr -> Arity
1186 go (Var v) = idArity v
1187 go (Lam x e) | isId x = go e + 1
1189 go (Note n e) = go e
1190 go (Cast e _) = go e
1191 go (App e (Type t)) = go e
1192 go (App f a) | exprIsCheap a = (go f - 1) `max` 0
1193 -- NB: exprIsCheap a!
1194 -- f (fac x) does not have arity 2,
1195 -- even if f has arity 3!
1196 -- NB: `max 0`! (\x y -> f x) has arity 2, even if f is
1197 -- unknown, hence arity 0
1201 %************************************************************************
1203 \subsection{Equality}
1205 %************************************************************************
1207 @cheapEqExpr@ is a cheap equality test which bales out fast!
1208 True => definitely equal
1209 False => may or may not be equal
1212 cheapEqExpr :: Expr b -> Expr b -> Bool
1214 cheapEqExpr (Var v1) (Var v2) = v1==v2
1215 cheapEqExpr (Lit lit1) (Lit lit2) = lit1 == lit2
1216 cheapEqExpr (Type t1) (Type t2) = t1 `coreEqType` t2
1218 cheapEqExpr (App f1 a1) (App f2 a2)
1219 = f1 `cheapEqExpr` f2 && a1 `cheapEqExpr` a2
1221 cheapEqExpr _ _ = False
1223 exprIsBig :: Expr b -> Bool
1224 -- Returns True of expressions that are too big to be compared by cheapEqExpr
1225 exprIsBig (Lit _) = False
1226 exprIsBig (Var v) = False
1227 exprIsBig (Type t) = False
1228 exprIsBig (App f a) = exprIsBig f || exprIsBig a
1229 exprIsBig (Cast e _) = exprIsBig e -- Hopefully coercions are not too big!
1230 exprIsBig other = True
1235 tcEqExpr :: CoreExpr -> CoreExpr -> Bool
1236 -- Used in rule matching, so does *not* look through
1237 -- newtypes, predicate types; hence tcEqExpr
1239 tcEqExpr e1 e2 = tcEqExprX rn_env e1 e2
1241 rn_env = mkRnEnv2 (mkInScopeSet (exprFreeVars e1 `unionVarSet` exprFreeVars e2))
1243 tcEqExprX :: RnEnv2 -> CoreExpr -> CoreExpr -> Bool
1244 tcEqExprX env (Var v1) (Var v2) = rnOccL env v1 == rnOccR env v2
1245 tcEqExprX env (Lit lit1) (Lit lit2) = lit1 == lit2
1246 tcEqExprX env (App f1 a1) (App f2 a2) = tcEqExprX env f1 f2 && tcEqExprX env a1 a2
1247 tcEqExprX env (Lam v1 e1) (Lam v2 e2) = tcEqExprX (rnBndr2 env v1 v2) e1 e2
1248 tcEqExprX env (Let (NonRec v1 r1) e1)
1249 (Let (NonRec v2 r2) e2) = tcEqExprX env r1 r2
1250 && tcEqExprX (rnBndr2 env v1 v2) e1 e2
1251 tcEqExprX env (Let (Rec ps1) e1)
1252 (Let (Rec ps2) e2) = equalLength ps1 ps2
1253 && and (zipWith eq_rhs ps1 ps2)
1254 && tcEqExprX env' e1 e2
1256 env' = foldl2 rn_bndr2 env ps2 ps2
1257 rn_bndr2 env (b1,_) (b2,_) = rnBndr2 env b1 b2
1258 eq_rhs (_,r1) (_,r2) = tcEqExprX env' r1 r2
1259 tcEqExprX env (Case e1 v1 t1 a1)
1260 (Case e2 v2 t2 a2) = tcEqExprX env e1 e2
1261 && tcEqTypeX env t1 t2
1262 && equalLength a1 a2
1263 && and (zipWith (eq_alt env') a1 a2)
1265 env' = rnBndr2 env v1 v2
1267 tcEqExprX env (Note n1 e1) (Note n2 e2) = eq_note env n1 n2 && tcEqExprX env e1 e2
1268 tcEqExprX env (Cast e1 co1) (Cast e2 co2) = tcEqTypeX env co1 co2 && tcEqExprX env e1 e2
1269 tcEqExprX env (Type t1) (Type t2) = tcEqTypeX env t1 t2
1270 tcEqExprX env e1 e2 = False
1272 eq_alt env (c1,vs1,r1) (c2,vs2,r2) = c1==c2 && tcEqExprX (rnBndrs2 env vs1 vs2) r1 r2
1274 eq_note env (SCC cc1) (SCC cc2) = cc1 == cc2
1275 eq_note env (CoreNote s1) (CoreNote s2) = s1 == s2
1276 eq_note env other1 other2 = False
1280 %************************************************************************
1282 \subsection{The size of an expression}
1284 %************************************************************************
1287 coreBindsSize :: [CoreBind] -> Int
1288 coreBindsSize bs = foldr ((+) . bindSize) 0 bs
1290 exprSize :: CoreExpr -> Int
1291 -- A measure of the size of the expressions
1292 -- It also forces the expression pretty drastically as a side effect
1293 exprSize (Var v) = v `seq` 1
1294 exprSize (Lit lit) = lit `seq` 1
1295 exprSize (App f a) = exprSize f + exprSize a
1296 exprSize (Lam b e) = varSize b + exprSize e
1297 exprSize (Let b e) = bindSize b + exprSize e
1298 exprSize (Case e b t as) = seqType t `seq` exprSize e + varSize b + 1 + foldr ((+) . altSize) 0 as
1299 exprSize (Cast e co) = (seqType co `seq` 1) + exprSize e
1300 exprSize (Note n e) = noteSize n + exprSize e
1301 exprSize (Type t) = seqType t `seq` 1
1303 noteSize (SCC cc) = cc `seq` 1
1304 noteSize InlineMe = 1
1305 noteSize (CoreNote s) = s `seq` 1 -- hdaume: core annotations
1307 varSize :: Var -> Int
1308 varSize b | isTyVar b = 1
1309 | otherwise = seqType (idType b) `seq`
1310 megaSeqIdInfo (idInfo b) `seq`
1313 varsSize = foldr ((+) . varSize) 0
1315 bindSize (NonRec b e) = varSize b + exprSize e
1316 bindSize (Rec prs) = foldr ((+) . pairSize) 0 prs
1318 pairSize (b,e) = varSize b + exprSize e
1320 altSize (c,bs,e) = c `seq` varsSize bs + exprSize e
1324 %************************************************************************
1326 \subsection{Hashing}
1328 %************************************************************************
1331 hashExpr :: CoreExpr -> Int
1332 -- Two expressions that hash to the same Int may be equal (but may not be)
1333 -- Two expressions that hash to the different Ints are definitely unequal
1335 -- But "unequal" here means "not identical"; two alpha-equivalent
1336 -- expressions may hash to the different Ints
1338 -- The emphasis is on a crude, fast hash, rather than on high precision
1340 hashExpr e | hash < 0 = 77 -- Just in case we hit -maxInt
1343 hash = abs (hash_expr e) -- Negative numbers kill UniqFM
1345 hash_expr (Note _ e) = hash_expr e
1346 hash_expr (Cast e co) = hash_expr e
1347 hash_expr (Let (NonRec b r) e) = hashId b
1348 hash_expr (Let (Rec ((b,r):_)) e) = hashId b
1349 hash_expr (Case _ b _ _) = hashId b
1350 hash_expr (App f e) = hash_expr f * fast_hash_expr e
1351 hash_expr (Var v) = hashId v
1352 hash_expr (Lit lit) = hashLiteral lit
1353 hash_expr (Lam b _) = hashId b
1354 hash_expr (Type t) = trace "hash_expr: type" 1 -- Shouldn't happen
1356 fast_hash_expr (Var v) = hashId v
1357 fast_hash_expr (Lit lit) = hashLiteral lit
1358 fast_hash_expr (App f (Type _)) = fast_hash_expr f
1359 fast_hash_expr (App f a) = fast_hash_expr a
1360 fast_hash_expr (Lam b _) = hashId b
1361 fast_hash_expr other = 1
1364 hashId id = hashName (idName id)
1367 %************************************************************************
1369 \subsection{Determining non-updatable right-hand-sides}
1371 %************************************************************************
1373 Top-level constructor applications can usually be allocated
1374 statically, but they can't if the constructor, or any of the
1375 arguments, come from another DLL (because we can't refer to static
1376 labels in other DLLs).
1378 If this happens we simply make the RHS into an updatable thunk,
1379 and 'exectute' it rather than allocating it statically.
1382 rhsIsStatic :: PackageId -> CoreExpr -> Bool
1383 -- This function is called only on *top-level* right-hand sides
1384 -- Returns True if the RHS can be allocated statically, with
1385 -- no thunks involved at all.
1387 -- It's called (i) in TidyPgm.hasCafRefs to decide if the rhs is, or
1388 -- refers to, CAFs; and (ii) in CoreToStg to decide whether to put an
1389 -- update flag on it.
1391 -- The basic idea is that rhsIsStatic returns True only if the RHS is
1392 -- (a) a value lambda
1393 -- (b) a saturated constructor application with static args
1395 -- BUT watch out for
1396 -- (i) Any cross-DLL references kill static-ness completely
1397 -- because they must be 'executed' not statically allocated
1398 -- ("DLL" here really only refers to Windows DLLs, on other platforms,
1399 -- this is not necessary)
1401 -- (ii) We treat partial applications as redexes, because in fact we
1402 -- make a thunk for them that runs and builds a PAP
1403 -- at run-time. The only appliations that are treated as
1404 -- static are *saturated* applications of constructors.
1406 -- We used to try to be clever with nested structures like this:
1407 -- ys = (:) w ((:) w [])
1408 -- on the grounds that CorePrep will flatten ANF-ise it later.
1409 -- But supporting this special case made the function much more
1410 -- complicated, because the special case only applies if there are no
1411 -- enclosing type lambdas:
1412 -- ys = /\ a -> Foo (Baz ([] a))
1413 -- Here the nested (Baz []) won't float out to top level in CorePrep.
1415 -- But in fact, even without -O, nested structures at top level are
1416 -- flattened by the simplifier, so we don't need to be super-clever here.
1420 -- f = \x::Int. x+7 TRUE
1421 -- p = (True,False) TRUE
1423 -- d = (fst p, False) FALSE because there's a redex inside
1424 -- (this particular one doesn't happen but...)
1426 -- h = D# (1.0## /## 2.0##) FALSE (redex again)
1427 -- n = /\a. Nil a TRUE
1429 -- t = /\a. (:) (case w a of ...) (Nil a) FALSE (redex)
1432 -- This is a bit like CoreUtils.exprIsHNF, with the following differences:
1433 -- a) scc "foo" (\x -> ...) is updatable (so we catch the right SCC)
1435 -- b) (C x xs), where C is a contructors is updatable if the application is
1438 -- c) don't look through unfolding of f in (f x).
1440 -- When opt_RuntimeTypes is on, we keep type lambdas and treat
1441 -- them as making the RHS re-entrant (non-updatable).
1443 rhsIsStatic this_pkg rhs = is_static False rhs
1445 is_static :: Bool -- True <=> in a constructor argument; must be atomic
1448 is_static False (Lam b e) = isRuntimeVar b || is_static False e
1450 is_static in_arg (Note (SCC _) e) = False
1451 is_static in_arg (Note _ e) = is_static in_arg e
1452 is_static in_arg (Cast e co) = is_static in_arg e
1454 is_static in_arg (Lit lit)
1456 MachLabel _ _ -> False
1458 -- A MachLabel (foreign import "&foo") in an argument
1459 -- prevents a constructor application from being static. The
1460 -- reason is that it might give rise to unresolvable symbols
1461 -- in the object file: under Linux, references to "weak"
1462 -- symbols from the data segment give rise to "unresolvable
1463 -- relocation" errors at link time This might be due to a bug
1464 -- in the linker, but we'll work around it here anyway.
1467 is_static in_arg other_expr = go other_expr 0
1469 go (Var f) n_val_args
1470 #if mingw32_TARGET_OS
1471 | not (isDllName this_pkg (idName f))
1473 = saturated_data_con f n_val_args
1474 || (in_arg && n_val_args == 0)
1475 -- A naked un-applied variable is *not* deemed a static RHS
1477 -- Reason: better to update so that the indirection gets shorted
1478 -- out, and the true value will be seen
1479 -- NB: if you change this, you'll break the invariant that THUNK_STATICs
1480 -- are always updatable. If you do so, make sure that non-updatable
1481 -- ones have enough space for their static link field!
1483 go (App f a) n_val_args
1484 | isTypeArg a = go f n_val_args
1485 | not in_arg && is_static True a = go f (n_val_args + 1)
1486 -- The (not in_arg) checks that we aren't in a constructor argument;
1487 -- if we are, we don't allow (value) applications of any sort
1489 -- NB. In case you wonder, args are sometimes not atomic. eg.
1490 -- x = D# (1.0## /## 2.0##)
1491 -- can't float because /## can fail.
1493 go (Note (SCC _) f) n_val_args = False
1494 go (Note _ f) n_val_args = go f n_val_args
1495 go (Cast e co) n_val_args = go e n_val_args
1497 go other n_val_args = False
1499 saturated_data_con f n_val_args
1500 = case isDataConWorkId_maybe f of
1501 Just dc -> n_val_args == dataConRepArity dc