2 % (c) The GRASP/AQUA Project, Glasgow University, 1992-1998
4 \section[CoreUtils]{Utility functions on @Core@ syntax}
9 mkInlineMe, mkSCC, mkCoerce,
10 bindNonRec, needsCaseBinding,
11 mkIfThenElse, mkAltExpr, mkPiType, mkPiTypes,
13 -- Taking expressions apart
14 findDefault, findAlt, isDefaultAlt, mergeAlts,
16 -- Properties of expressions
17 exprType, coreAltType,
18 exprIsDupable, exprIsTrivial, exprIsCheap,
19 exprIsHNF,exprOkForSpeculation, exprIsBig,
20 exprIsConApp_maybe, exprIsBottom,
23 -- Arity and eta expansion
24 manifestArity, exprArity,
25 exprEtaExpandArity, etaExpand,
34 cheapEqExpr, tcEqExpr, tcEqExprX, applyTypeToArgs, applyTypeToArg,
39 #include "HsVersions.h"
42 import GLAEXTS -- For `xori`
45 import CoreFVs ( exprFreeVars )
46 import PprCore ( pprCoreExpr )
47 import Var ( Var, TyVar, CoVar, isCoVar, tyVarKind, setVarUnique,
48 mkCoVar, mkTyVar, mkCoVar )
49 import VarSet ( unionVarSet )
51 import Name ( hashName, mkSysTvName )
53 import Packages ( isDllName )
55 import Literal ( hashLiteral, literalType, litIsDupable,
56 litIsTrivial, isZeroLit, Literal( MachLabel ) )
57 import DataCon ( DataCon, dataConRepArity,
58 isVanillaDataCon, dataConTyCon, dataConRepArgTys,
59 dataConUnivTyVars, dataConExTyVars, dataConEqSpec )
60 import PrimOp ( PrimOp(..), primOpOkForSpeculation, primOpIsCheap )
61 import Id ( Id, idType, globalIdDetails, idNewStrictness,
62 mkWildId, idArity, idName, idUnfolding, idInfo,
63 isOneShotBndr, isStateHackType, isDataConWorkId_maybe, mkSysLocal,
64 isDataConWorkId, isBottomingId, isDictId
66 import IdInfo ( GlobalIdDetails(..), megaSeqIdInfo )
67 import NewDemand ( appIsBottom )
68 import Type ( Type, mkFunTy, mkForAllTy, splitFunTy_maybe,
69 splitFunTy, tcEqTypeX,
70 applyTys, isUnLiftedType, seqType, mkTyVarTy,
71 splitForAllTy_maybe, isForAllTy, splitRecNewType_maybe,
72 splitTyConApp_maybe, coreEqType, funResultTy, applyTy,
75 import Coercion ( Coercion, mkTransCoercion, coercionKind,
76 splitNewTypeRepCo_maybe, mkSymCoercion, mkLeftCoercion,
77 mkRightCoercion, decomposeCo, coercionKindPredTy,
78 splitCoercionKind, mkEqPred )
79 import TyCon ( tyConArity )
80 import TysWiredIn ( boolTy, trueDataCon, falseDataCon )
81 import CostCentre ( CostCentre )
82 import BasicTypes ( Arity )
83 import PackageConfig ( PackageId )
84 import Unique ( Unique )
86 import DynFlags ( DynFlags, DynFlag(Opt_DictsCheap), dopt )
87 import TysPrim ( alphaTy ) -- Debugging only
88 import Util ( equalLength, lengthAtLeast, foldl2 )
92 %************************************************************************
94 \subsection{Find the type of a Core atom/expression}
96 %************************************************************************
99 exprType :: CoreExpr -> Type
101 exprType (Var var) = idType var
102 exprType (Lit lit) = literalType lit
103 exprType (Let _ body) = exprType body
104 exprType (Case _ _ ty alts) = ty
106 = let (_, ty) = coercionKind co in ty
107 exprType (Note other_note e) = exprType e
108 exprType (Lam binder expr) = mkPiType binder (exprType expr)
110 = case collectArgs e of
111 (fun, args) -> applyTypeToArgs e (exprType fun) args
113 exprType other = pprTrace "exprType" (pprCoreExpr other) alphaTy
115 coreAltType :: CoreAlt -> Type
116 coreAltType (_,_,rhs) = exprType rhs
119 @mkPiType@ makes a (->) type or a forall type, depending on whether
120 it is given a type variable or a term variable. We cleverly use the
121 lbvarinfo field to figure out the right annotation for the arrove in
122 case of a term variable.
125 mkPiType :: Var -> Type -> Type -- The more polymorphic version
126 mkPiTypes :: [Var] -> Type -> Type -- doesn't work...
128 mkPiTypes vs ty = foldr mkPiType ty vs
131 | isId v = mkFunTy (idType v) ty
132 | otherwise = mkForAllTy v ty
136 applyTypeToArg :: Type -> CoreExpr -> Type
137 applyTypeToArg fun_ty (Type arg_ty) = applyTy fun_ty arg_ty
138 applyTypeToArg fun_ty other_arg = funResultTy fun_ty
140 applyTypeToArgs :: CoreExpr -> Type -> [CoreExpr] -> Type
141 -- A more efficient version of applyTypeToArg
142 -- when we have several args
143 -- The first argument is just for debugging
144 applyTypeToArgs e op_ty [] = op_ty
146 applyTypeToArgs e op_ty (Type ty : args)
147 = -- Accumulate type arguments so we can instantiate all at once
150 go rev_tys (Type ty : args) = go (ty:rev_tys) args
151 go rev_tys rest_args = applyTypeToArgs e op_ty' rest_args
153 op_ty' = applyTys op_ty (reverse rev_tys)
155 applyTypeToArgs e op_ty (other_arg : args)
156 = case (splitFunTy_maybe op_ty) of
157 Just (_, res_ty) -> applyTypeToArgs e res_ty args
158 Nothing -> pprPanic "applyTypeToArgs" (pprCoreExpr e $$ ppr op_ty)
163 %************************************************************************
165 \subsection{Attaching notes}
167 %************************************************************************
169 mkNote removes redundant coercions, and SCCs where possible
173 mkNote :: Note -> CoreExpr -> CoreExpr
174 mkNote (SCC cc) expr = mkSCC cc expr
175 mkNote InlineMe expr = mkInlineMe expr
176 mkNote note expr = Note note expr
180 Drop trivial InlineMe's. This is somewhat important, because if we have an unfolding
181 that looks like (Note InlineMe (Var v)), the InlineMe doesn't go away because it may
182 not be *applied* to anything.
184 We don't use exprIsTrivial here, though, because we sometimes generate worker/wrapper
187 f = inline_me (coerce t fw)
188 As usual, the inline_me prevents the worker from getting inlined back into the wrapper.
189 We want the split, so that the coerces can cancel at the call site.
191 However, we can get left with tiresome type applications. Notably, consider
192 f = /\ a -> let t = e in (t, w)
193 Then lifting the let out of the big lambda gives
195 f = /\ a -> let t = inline_me (t' a) in (t, w)
196 The inline_me is to stop the simplifier inlining t' right back
197 into t's RHS. In the next phase we'll substitute for t (since
198 its rhs is trivial) and *then* we could get rid of the inline_me.
199 But it hardly seems worth it, so I don't bother.
202 mkInlineMe (Var v) = Var v
203 mkInlineMe e = Note InlineMe e
209 mkCoerce :: Coercion -> CoreExpr -> CoreExpr
210 mkCoerce co (Cast expr co2)
211 = ASSERT(let { (from_ty, to_ty) = coercionKind co;
212 (from_ty2, to_ty2) = coercionKind co2} in
213 from_ty `coreEqType` to_ty2 )
214 mkCoerce (mkTransCoercion co2 co) expr
217 = let (from_ty, to_ty) = coercionKind co in
218 -- if to_ty `coreEqType` from_ty
221 ASSERT2(from_ty `coreEqType` (exprType expr), text "Trying to coerce" <+> text "(" <> ppr expr $$ text "::" <+> ppr (exprType expr) <> text ")" $$ ppr co $$ ppr (coercionKindPredTy co))
226 mkSCC :: CostCentre -> Expr b -> Expr b
227 -- Note: Nested SCC's *are* preserved for the benefit of
228 -- cost centre stack profiling
229 mkSCC cc (Lit lit) = Lit lit
230 mkSCC cc (Lam x e) = Lam x (mkSCC cc e) -- Move _scc_ inside lambda
231 mkSCC cc (Note (SCC cc') e) = Note (SCC cc) (Note (SCC cc') e)
232 mkSCC cc (Note n e) = Note n (mkSCC cc e) -- Move _scc_ inside notes
233 mkSCC cc (Cast e co) = Cast (mkSCC cc e) co -- Move _scc_ inside cast
234 mkSCC cc expr = Note (SCC cc) expr
238 %************************************************************************
240 \subsection{Other expression construction}
242 %************************************************************************
245 bindNonRec :: Id -> CoreExpr -> CoreExpr -> CoreExpr
246 -- (bindNonRec x r b) produces either
249 -- case r of x { _DEFAULT_ -> b }
251 -- depending on whether x is unlifted or not
252 -- It's used by the desugarer to avoid building bindings
253 -- that give Core Lint a heart attack. Actually the simplifier
254 -- deals with them perfectly well.
256 bindNonRec bndr rhs body
257 | needsCaseBinding (idType bndr) rhs = Case rhs bndr (exprType body) [(DEFAULT,[],body)]
258 | otherwise = Let (NonRec bndr rhs) body
260 needsCaseBinding ty rhs = isUnLiftedType ty && not (exprOkForSpeculation rhs)
261 -- Make a case expression instead of a let
262 -- These can arise either from the desugarer,
263 -- or from beta reductions: (\x.e) (x +# y)
267 mkAltExpr :: AltCon -> [CoreBndr] -> [Type] -> CoreExpr
268 -- This guy constructs the value that the scrutinee must have
269 -- when you are in one particular branch of a case
270 mkAltExpr (DataAlt con) args inst_tys
271 = mkConApp con (map Type inst_tys ++ varsToCoreExprs args)
272 mkAltExpr (LitAlt lit) [] []
275 mkIfThenElse :: CoreExpr -> CoreExpr -> CoreExpr -> CoreExpr
276 mkIfThenElse guard then_expr else_expr
277 -- Not going to be refining, so okay to take the type of the "then" clause
278 = Case guard (mkWildId boolTy) (exprType then_expr)
279 [ (DataAlt falseDataCon, [], else_expr), -- Increasing order of tag!
280 (DataAlt trueDataCon, [], then_expr) ]
284 %************************************************************************
286 \subsection{Taking expressions apart}
288 %************************************************************************
290 The default alternative must be first, if it exists at all.
291 This makes it easy to find, though it makes matching marginally harder.
294 findDefault :: [CoreAlt] -> ([CoreAlt], Maybe CoreExpr)
295 findDefault ((DEFAULT,args,rhs) : alts) = ASSERT( null args ) (alts, Just rhs)
296 findDefault alts = (alts, Nothing)
298 findAlt :: AltCon -> [CoreAlt] -> CoreAlt
301 (deflt@(DEFAULT,_,_):alts) -> go alts deflt
302 other -> go alts panic_deflt
304 panic_deflt = pprPanic "Missing alternative" (ppr con $$ vcat (map ppr alts))
307 go (alt@(con1,_,_) : alts) deflt
308 = case con `cmpAltCon` con1 of
309 LT -> deflt -- Missed it already; the alts are in increasing order
311 GT -> ASSERT( not (con1 == DEFAULT) ) go alts deflt
313 isDefaultAlt :: CoreAlt -> Bool
314 isDefaultAlt (DEFAULT, _, _) = True
315 isDefaultAlt other = False
317 ---------------------------------
318 mergeAlts :: [CoreAlt] -> [CoreAlt] -> [CoreAlt]
319 -- Merge preserving order; alternatives in the first arg
320 -- shadow ones in the second
321 mergeAlts [] as2 = as2
322 mergeAlts as1 [] = as1
323 mergeAlts (a1:as1) (a2:as2)
324 = case a1 `cmpAlt` a2 of
325 LT -> a1 : mergeAlts as1 (a2:as2)
326 EQ -> a1 : mergeAlts as1 as2 -- Discard a2
327 GT -> a2 : mergeAlts (a1:as1) as2
331 %************************************************************************
333 \subsection{Figuring out things about expressions}
335 %************************************************************************
337 @exprIsTrivial@ is true of expressions we are unconditionally happy to
338 duplicate; simple variables and constants, and type
339 applications. Note that primop Ids aren't considered
342 @exprIsBottom@ is true of expressions that are guaranteed to diverge
345 There used to be a gruesome test for (hasNoBinding v) in the
347 exprIsTrivial (Var v) | hasNoBinding v = idArity v == 0
348 The idea here is that a constructor worker, like $wJust, is
349 really short for (\x -> $wJust x), becuase $wJust has no binding.
350 So it should be treated like a lambda. Ditto unsaturated primops.
351 But now constructor workers are not "have-no-binding" Ids. And
352 completely un-applied primops and foreign-call Ids are sufficiently
353 rare that I plan to allow them to be duplicated and put up with
356 SCC notes. We do not treat (_scc_ "foo" x) as trivial, because
357 a) it really generates code, (and a heap object when it's
358 a function arg) to capture the cost centre
359 b) see the note [SCC-and-exprIsTrivial] in Simplify.simplLazyBind
362 exprIsTrivial (Var v) = True -- See notes above
363 exprIsTrivial (Type _) = True
364 exprIsTrivial (Lit lit) = litIsTrivial lit
365 exprIsTrivial (App e arg) = not (isRuntimeArg arg) && exprIsTrivial e
366 exprIsTrivial (Note (SCC _) e) = False -- See notes above
367 exprIsTrivial (Note _ e) = exprIsTrivial e
368 exprIsTrivial (Cast e co) = exprIsTrivial e
369 exprIsTrivial (Lam b body) = not (isRuntimeVar b) && exprIsTrivial body
370 exprIsTrivial other = False
374 @exprIsDupable@ is true of expressions that can be duplicated at a modest
375 cost in code size. This will only happen in different case
376 branches, so there's no issue about duplicating work.
378 That is, exprIsDupable returns True of (f x) even if
379 f is very very expensive to call.
381 Its only purpose is to avoid fruitless let-binding
382 and then inlining of case join points
386 exprIsDupable (Type _) = True
387 exprIsDupable (Var v) = True
388 exprIsDupable (Lit lit) = litIsDupable lit
389 exprIsDupable (Note InlineMe e) = True
390 exprIsDupable (Note _ e) = exprIsDupable e
391 exprIsDupable (Cast e co) = exprIsDupable e
395 go (Var v) n_args = True
396 go (App f a) n_args = n_args < dupAppSize
399 go other n_args = False
402 dupAppSize = 4 -- Size of application we are prepared to duplicate
405 @exprIsCheap@ looks at a Core expression and returns \tr{True} if
406 it is obviously in weak head normal form, or is cheap to get to WHNF.
407 [Note that that's not the same as exprIsDupable; an expression might be
408 big, and hence not dupable, but still cheap.]
410 By ``cheap'' we mean a computation we're willing to:
411 push inside a lambda, or
412 inline at more than one place
413 That might mean it gets evaluated more than once, instead of being
414 shared. The main examples of things which aren't WHNF but are
419 (where e, and all the ei are cheap)
422 (where e and b are cheap)
425 (where op is a cheap primitive operator)
428 (because we are happy to substitute it inside a lambda)
430 Notice that a variable is considered 'cheap': we can push it inside a lambda,
431 because sharing will make sure it is only evaluated once.
434 exprIsCheap :: CoreExpr -> Bool
435 exprIsCheap (Lit lit) = True
436 exprIsCheap (Type _) = True
437 exprIsCheap (Var _) = True
438 exprIsCheap (Note InlineMe e) = True
439 exprIsCheap (Note _ e) = exprIsCheap e
440 exprIsCheap (Cast e co) = exprIsCheap e
441 exprIsCheap (Lam x e) = isRuntimeVar x || exprIsCheap e
442 exprIsCheap (Case e _ _ alts) = exprIsCheap e &&
443 and [exprIsCheap rhs | (_,_,rhs) <- alts]
444 -- Experimentally, treat (case x of ...) as cheap
445 -- (and case __coerce x etc.)
446 -- This improves arities of overloaded functions where
447 -- there is only dictionary selection (no construction) involved
448 exprIsCheap (Let (NonRec x _) e)
449 | isUnLiftedType (idType x) = exprIsCheap e
451 -- strict lets always have cheap right hand sides,
452 -- and do no allocation.
454 exprIsCheap other_expr -- Applications and variables
457 -- Accumulate value arguments, then decide
458 go (App f a) val_args | isRuntimeArg a = go f (a:val_args)
459 | otherwise = go f val_args
461 go (Var f) [] = True -- Just a type application of a variable
462 -- (f t1 t2 t3) counts as WHNF
464 = case globalIdDetails f of
465 RecordSelId {} -> go_sel args
466 ClassOpId _ -> go_sel args
467 PrimOpId op -> go_primop op args
469 DataConWorkId _ -> go_pap args
470 other | length args < idArity f -> go_pap args
472 other -> isBottomingId f
473 -- Application of a function which
474 -- always gives bottom; we treat this as cheap
475 -- because it certainly doesn't need to be shared!
477 go other args = False
480 go_pap args = all exprIsTrivial args
481 -- For constructor applications and primops, check that all
482 -- the args are trivial. We don't want to treat as cheap, say,
484 -- We'll put up with one constructor application, but not dozens
487 go_primop op args = primOpIsCheap op && all exprIsCheap args
488 -- In principle we should worry about primops
489 -- that return a type variable, since the result
490 -- might be applied to something, but I'm not going
491 -- to bother to check the number of args
494 go_sel [arg] = exprIsCheap arg -- I'm experimenting with making record selection
495 go_sel other = False -- look cheap, so we will substitute it inside a
496 -- lambda. Particularly for dictionary field selection.
497 -- BUT: Take care with (sel d x)! The (sel d) might be cheap, but
498 -- there's no guarantee that (sel d x) will be too. Hence (n_val_args == 1)
501 exprOkForSpeculation returns True of an expression that it is
503 * safe to evaluate even if normal order eval might not
504 evaluate the expression at all, or
506 * safe *not* to evaluate even if normal order would do so
510 the expression guarantees to terminate,
512 without raising an exception,
513 without causing a side effect (e.g. writing a mutable variable)
515 NB: if exprIsHNF e, then exprOkForSpecuation e
518 let x = case y# +# 1# of { r# -> I# r# }
521 case y# +# 1# of { r# ->
526 We can only do this if the (y+1) is ok for speculation: it has no
527 side effects, and can't diverge or raise an exception.
530 exprOkForSpeculation :: CoreExpr -> Bool
531 exprOkForSpeculation (Lit _) = True
532 exprOkForSpeculation (Type _) = True
533 exprOkForSpeculation (Var v) = isUnLiftedType (idType v)
534 exprOkForSpeculation (Note _ e) = exprOkForSpeculation e
535 exprOkForSpeculation (Cast e co) = exprOkForSpeculation e
536 exprOkForSpeculation other_expr
537 = case collectArgs other_expr of
538 (Var f, args) -> spec_ok (globalIdDetails f) args
542 spec_ok (DataConWorkId _) args
543 = True -- The strictness of the constructor has already
544 -- been expressed by its "wrapper", so we don't need
545 -- to take the arguments into account
547 spec_ok (PrimOpId op) args
548 | isDivOp op, -- Special case for dividing operations that fail
549 [arg1, Lit lit] <- args -- only if the divisor is zero
550 = not (isZeroLit lit) && exprOkForSpeculation arg1
551 -- Often there is a literal divisor, and this
552 -- can get rid of a thunk in an inner looop
555 = primOpOkForSpeculation op &&
556 all exprOkForSpeculation args
557 -- A bit conservative: we don't really need
558 -- to care about lazy arguments, but this is easy
560 spec_ok other args = False
562 isDivOp :: PrimOp -> Bool
563 -- True of dyadic operators that can fail
564 -- only if the second arg is zero
565 -- This function probably belongs in PrimOp, or even in
566 -- an automagically generated file.. but it's such a
567 -- special case I thought I'd leave it here for now.
568 isDivOp IntQuotOp = True
569 isDivOp IntRemOp = True
570 isDivOp WordQuotOp = True
571 isDivOp WordRemOp = True
572 isDivOp IntegerQuotRemOp = True
573 isDivOp IntegerDivModOp = True
574 isDivOp FloatDivOp = True
575 isDivOp DoubleDivOp = True
576 isDivOp other = False
581 exprIsBottom :: CoreExpr -> Bool -- True => definitely bottom
582 exprIsBottom e = go 0 e
584 -- n is the number of args
585 go n (Note _ e) = go n e
586 go n (Cast e co) = go n e
587 go n (Let _ e) = go n e
588 go n (Case e _ _ _) = go 0 e -- Just check the scrut
589 go n (App e _) = go (n+1) e
590 go n (Var v) = idAppIsBottom v n
592 go n (Lam _ _) = False
593 go n (Type _) = False
595 idAppIsBottom :: Id -> Int -> Bool
596 idAppIsBottom id n_val_args = appIsBottom (idNewStrictness id) n_val_args
599 @exprIsHNF@ returns true for expressions that are certainly *already*
600 evaluated to *head* normal form. This is used to decide whether it's ok
603 case x of _ -> e ===> e
605 and to decide whether it's safe to discard a `seq`
607 So, it does *not* treat variables as evaluated, unless they say they are.
609 But it *does* treat partial applications and constructor applications
610 as values, even if their arguments are non-trivial, provided the argument
612 e.g. (:) (f x) (map f xs) is a value
613 map (...redex...) is a value
614 Because `seq` on such things completes immediately
616 For unlifted argument types, we have to be careful:
618 Suppose (f x) diverges; then C (f x) is not a value. True, but
619 this form is illegal (see the invariants in CoreSyn). Args of unboxed
620 type must be ok-for-speculation (or trivial).
623 exprIsHNF :: CoreExpr -> Bool -- True => Value-lambda, constructor, PAP
624 exprIsHNF (Var v) -- NB: There are no value args at this point
625 = isDataConWorkId v -- Catches nullary constructors,
626 -- so that [] and () are values, for example
627 || idArity v > 0 -- Catches (e.g.) primops that don't have unfoldings
628 || isEvaldUnfolding (idUnfolding v)
629 -- Check the thing's unfolding; it might be bound to a value
630 -- A worry: what if an Id's unfolding is just itself:
631 -- then we could get an infinite loop...
633 exprIsHNF (Lit l) = True
634 exprIsHNF (Type ty) = True -- Types are honorary Values;
635 -- we don't mind copying them
636 exprIsHNF (Lam b e) = isRuntimeVar b || exprIsHNF e
637 exprIsHNF (Note _ e) = exprIsHNF e
638 exprIsHNF (Cast e co) = exprIsHNF e
639 exprIsHNF (App e (Type _)) = exprIsHNF e
640 exprIsHNF (App e a) = app_is_value e [a]
641 exprIsHNF other = False
643 -- There is at least one value argument
644 app_is_value (Var fun) args
645 | isDataConWorkId fun -- Constructor apps are values
646 || idArity fun > valArgCount args -- Under-applied function
647 = check_args (idType fun) args
648 app_is_value (App f a) as = app_is_value f (a:as)
649 app_is_value other as = False
651 -- 'check_args' checks that unlifted-type args
652 -- are in fact guaranteed non-divergent
653 check_args fun_ty [] = True
654 check_args fun_ty (Type _ : args) = case splitForAllTy_maybe fun_ty of
655 Just (_, ty) -> check_args ty args
656 check_args fun_ty (arg : args)
657 | isUnLiftedType arg_ty = exprOkForSpeculation arg
658 | otherwise = check_args res_ty args
660 (arg_ty, res_ty) = splitFunTy fun_ty
664 -- deep applies a TyConApp coercion as a substitution to a reflexive coercion
665 -- deepCast t [a1,...,an] co corresponds to deep(t, [a1,...,an], co) from
667 deepCast :: Type -> [TyVar] -> Coercion -> Coercion
668 deepCast ty tyVars co
669 = ASSERT( let {(lty, rty) = coercionKind co;
670 Just (tc1, lArgs) = splitTyConApp_maybe lty;
671 Just (tc2, rArgs) = splitTyConApp_maybe rty}
673 tc1 == tc2 && length lArgs == length rArgs &&
674 length lArgs == length tyVars )
675 substTyWith tyVars coArgs ty
677 -- coArgs = [right (left (left co)), right (left co), right co]
678 coArgs = decomposeCo (length tyVars) co
680 -- This goes here to avoid circularity between DataCon and Id
681 dataConInstPat :: [Unique] -- An infinite list of uniques
683 -> [Type] -- Types to instantiate the universally quantified tyvars
684 -> ([TyVar], [CoVar], [Id]) -- Return instantiated variables
685 dataConInstPat uniqs con inst_tys
686 = (ex_bndrs, co_bndrs, id_bndrs)
688 univ_tvs = dataConUnivTyVars con
689 ex_tvs = dataConExTyVars con
690 arg_tys = dataConRepArgTys con
691 eq_spec = dataConEqSpec con
692 eq_preds = [ mkEqPred (mkTyVarTy tv, ty) | (tv,ty) <- eq_spec ]
695 n_co = length eq_spec
696 n_id = length arg_tys
699 (ex_uniqs, uniqs') = splitAt n_ex uniqs
700 (co_uniqs, id_uniqs) = splitAt n_co uniqs'
702 -- make existential type variables
703 mk_ex_var uniq var = setVarUnique var uniq
704 ex_bndrs = zipWith mk_ex_var ex_uniqs ex_tvs
706 -- make the instantiation substitution
707 inst_subst = substTyWith (univ_tvs ++ ex_tvs) (inst_tys ++ map mkTyVarTy ex_bndrs)
709 -- make a new coercion vars, instantiating kind
710 mk_co_var uniq eq_pred = mkCoVar new_name (inst_subst (mkPredTy eq_pred))
712 new_name = mkSysTvName uniq FSLIT("co")
714 co_bndrs = zipWith mk_co_var co_uniqs eq_preds
716 -- make value vars, instantiating types
717 mk_id_var uniq ty = mkSysLocal FSLIT("ca") uniq (inst_subst ty)
719 id_bndrs = zipWith mk_id_var id_uniqs arg_tys
722 exprIsConApp_maybe :: CoreExpr -> Maybe (DataCon, [CoreExpr])
723 -- Returns (Just (dc, [x1..xn])) if the argument expression is
724 -- a constructor application of the form (dc x1 .. xn)
726 exprIsConApp_maybe (Cast expr co)
727 = -- Maybe this is over the top, but here we try to turn
728 -- coerce (S,T) ( x, y )
730 -- ( coerce S x, coerce T y )
731 -- This happens in anger in PrelArrExts which has a coerce
732 -- case coerce memcpy a b of
734 -- where the memcpy is in the IO monad, but the call is in
736 case exprIsConApp_maybe expr of {
740 let (from_ty, to_ty) = coercionKind co in
742 case splitTyConApp_maybe to_ty of {
744 Just (tc, tc_arg_tys) | tc /= dataConTyCon dc -> Nothing
745 -- | not (isVanillaDataCon dc) -> Nothing
747 -- Type constructor must match datacon
749 case splitTyConApp_maybe from_ty of {
751 Just (tc', tc_arg_tys') | tc /= tc' -> Nothing
752 -- Both sides of coercion must have the same type constructor
756 -- here we do the PushC reduction rule as described in the FC paper
757 arity = tyConArity tc
758 n_ex_tvs = length dc_ex_tyvars
760 (univ_args, rest) = splitAt arity args
761 (ex_args, val_args) = splitAt n_ex_tvs rest
763 arg_tys = dataConRepArgTys dc
764 dc_tyvars = dataConUnivTyVars dc
765 dc_ex_tyvars = dataConExTyVars dc
767 deep arg_ty = deepCast arg_ty dc_tyvars co
769 -- first we appropriately cast the value arguments
770 arg_cos = map deep arg_tys
771 new_val_args = zipWith mkCoerce (map deep arg_tys) val_args
773 -- then we cast the existential coercion arguments
774 orig_tvs = dc_tyvars ++ dc_ex_tyvars
775 gammas = decomposeCo arity co
776 new_tys = gammas ++ (map (\ (Type t) -> t) ex_args)
777 theta = substTyWith orig_tvs new_tys
780 , (ty1, ty2) <- splitCoercionKind (tyVarKind tv)
781 = Type $ mkTransCoercion (mkSymCoercion (theta ty1))
782 (mkTransCoercion ty (theta ty2))
785 new_ex_args = zipWith cast_ty dc_ex_tyvars ex_args
788 ASSERT( all isTypeArg (take arity args) )
789 ASSERT( equalLength val_args arg_tys )
790 Just (dc, map Type tc_arg_tys ++ new_ex_args ++ new_val_args)
793 exprIsConApp_maybe (Note _ expr)
794 = exprIsConApp_maybe expr
795 -- We ignore InlineMe notes in case we have
796 -- x = __inline_me__ (a,b)
797 -- All part of making sure that INLINE pragmas never hurt
798 -- Marcin tripped on this one when making dictionaries more inlinable
800 -- In fact, we ignore all notes. For example,
801 -- case _scc_ "foo" (C a b) of
803 -- should be optimised away, but it will be only if we look
804 -- through the SCC note.
806 exprIsConApp_maybe expr = analyse (collectArgs expr)
808 analyse (Var fun, args)
809 | Just con <- isDataConWorkId_maybe fun,
810 args `lengthAtLeast` dataConRepArity con
811 -- Might be > because the arity excludes type args
814 -- Look through unfoldings, but only cheap ones, because
815 -- we are effectively duplicating the unfolding
816 analyse (Var fun, [])
817 | let unf = idUnfolding fun,
819 = exprIsConApp_maybe (unfoldingTemplate unf)
821 analyse other = Nothing
826 %************************************************************************
828 \subsection{Eta reduction and expansion}
830 %************************************************************************
833 exprEtaExpandArity :: DynFlags -> CoreExpr -> Arity
834 {- The Arity returned is the number of value args the
835 thing can be applied to without doing much work
837 exprEtaExpandArity is used when eta expanding
840 It returns 1 (or more) to:
841 case x of p -> \s -> ...
842 because for I/O ish things we really want to get that \s to the top.
843 We are prepared to evaluate x each time round the loop in order to get that
845 It's all a bit more subtle than it looks:
849 Consider one-shot lambdas
850 let x = expensive in \y z -> E
851 We want this to have arity 2 if the \y-abstraction is a 1-shot lambda
852 Hence the ArityType returned by arityType
854 2. The state-transformer hack
856 The one-shot lambda special cause is particularly important/useful for
857 IO state transformers, where we often get
858 let x = E in \ s -> ...
860 and the \s is a real-world state token abstraction. Such abstractions
861 are almost invariably 1-shot, so we want to pull the \s out, past the
862 let x=E, even if E is expensive. So we treat state-token lambdas as
863 one-shot even if they aren't really. The hack is in Id.isOneShotBndr.
865 3. Dealing with bottom
868 f = \x -> error "foo"
869 Here, arity 1 is fine. But if it is
873 then we want to get arity 2. Tecnically, this isn't quite right, because
875 should diverge, but it'll converge if we eta-expand f. Nevertheless, we
876 do so; it improves some programs significantly, and increasing convergence
877 isn't a bad thing. Hence the ABot/ATop in ArityType.
879 Actually, the situation is worse. Consider
883 Can we eta-expand here? At first the answer looks like "yes of course", but
886 This should diverge! But if we eta-expand, it won't. Again, we ignore this
887 "problem", because being scrupulous would lose an important transformation for
893 Non-recursive newtypes are transparent, and should not get in the way.
894 We do (currently) eta-expand recursive newtypes too. So if we have, say
896 newtype T = MkT ([T] -> Int)
900 where f has arity 1. Then: etaExpandArity e = 1;
901 that is, etaExpandArity looks through the coerce.
903 When we eta-expand e to arity 1: eta_expand 1 e T
904 we want to get: coerce T (\x::[T] -> (coerce ([T]->Int) e) x)
906 HOWEVER, note that if you use coerce bogusly you can ge
908 And since negate has arity 2, you might try to eta expand. But you can't
909 decopose Int to a function type. Hence the final case in eta_expand.
913 exprEtaExpandArity dflags e = arityDepth (arityType dflags e)
915 -- A limited sort of function type
916 data ArityType = AFun Bool ArityType -- True <=> one-shot
917 | ATop -- Know nothing
920 arityDepth :: ArityType -> Arity
921 arityDepth (AFun _ ty) = 1 + arityDepth ty
924 andArityType ABot at2 = at2
925 andArityType ATop at2 = ATop
926 andArityType (AFun t1 at1) (AFun t2 at2) = AFun (t1 && t2) (andArityType at1 at2)
927 andArityType at1 at2 = andArityType at2 at1
929 arityType :: DynFlags -> CoreExpr -> ArityType
930 -- (go1 e) = [b1,..,bn]
931 -- means expression can be rewritten \x_b1 -> ... \x_bn -> body
932 -- where bi is True <=> the lambda is one-shot
934 arityType dflags (Note n e) = arityType dflags e
935 -- Not needed any more: etaExpand is cleverer
936 -- | ok_note n = arityType dflags e
937 -- | otherwise = ATop
939 arityType dflags (Cast e co) = arityType dflags e
941 arityType dflags (Var v)
942 = mk (idArity v) (arg_tys (idType v))
944 mk :: Arity -> [Type] -> ArityType
945 -- The argument types are only to steer the "state hack"
946 -- Consider case x of
948 -- False -> \(s:RealWorld) -> e
949 -- where foo has arity 1. Then we want the state hack to
950 -- apply to foo too, so we can eta expand the case.
951 mk 0 tys | isBottomingId v = ABot
952 | (ty:tys) <- tys, isStateHackType ty = AFun True ATop
954 mk n (ty:tys) = AFun (isStateHackType ty) (mk (n-1) tys)
955 mk n [] = AFun False (mk (n-1) [])
957 arg_tys :: Type -> [Type] -- Ignore for-alls
959 | Just (_, ty') <- splitForAllTy_maybe ty = arg_tys ty'
960 | Just (arg,res) <- splitFunTy_maybe ty = arg : arg_tys res
963 -- Lambdas; increase arity
964 arityType dflags (Lam x e)
965 | isId x = AFun (isOneShotBndr x) (arityType dflags e)
966 | otherwise = arityType dflags e
968 -- Applications; decrease arity
969 arityType dflags (App f (Type _)) = arityType dflags f
970 arityType dflags (App f a) = case arityType dflags f of
971 AFun one_shot xs | exprIsCheap a -> xs
974 -- Case/Let; keep arity if either the expression is cheap
975 -- or it's a 1-shot lambda
976 -- The former is not really right for Haskell
977 -- f x = case x of { (a,b) -> \y. e }
979 -- f x y = case x of { (a,b) -> e }
980 -- The difference is observable using 'seq'
981 arityType dflags (Case scrut _ _ alts)
982 = case foldr1 andArityType [arityType dflags rhs | (_,_,rhs) <- alts] of
983 xs | exprIsCheap scrut -> xs
984 xs@(AFun one_shot _) | one_shot -> AFun True ATop
987 arityType dflags (Let b e)
988 = case arityType dflags e of
989 xs | cheap_bind b -> xs
990 xs@(AFun one_shot _) | one_shot -> AFun True ATop
993 cheap_bind (NonRec b e) = is_cheap (b,e)
994 cheap_bind (Rec prs) = all is_cheap prs
995 is_cheap (b,e) = (dopt Opt_DictsCheap dflags && isDictId b)
997 -- If the experimental -fdicts-cheap flag is on, we eta-expand through
998 -- dictionary bindings. This improves arities. Thereby, it also
999 -- means that full laziness is less prone to floating out the
1000 -- application of a function to its dictionary arguments, which
1001 -- can thereby lose opportunities for fusion. Example:
1002 -- foo :: Ord a => a -> ...
1003 -- foo = /\a \(d:Ord a). let d' = ...d... in \(x:a). ....
1004 -- -- So foo has arity 1
1006 -- f = \x. foo dInt $ bar x
1008 -- The (foo DInt) is floated out, and makes ineffective a RULE
1009 -- foo (bar x) = ...
1011 -- One could go further and make exprIsCheap reply True to any
1012 -- dictionary-typed expression, but that's more work.
1014 arityType dflags other = ATop
1016 {- NOT NEEDED ANY MORE: etaExpand is cleverer
1017 ok_note InlineMe = False
1018 ok_note other = True
1019 -- Notice that we do not look through __inline_me__
1020 -- This may seem surprising, but consider
1021 -- f = _inline_me (\x -> e)
1022 -- We DO NOT want to eta expand this to
1023 -- f = \x -> (_inline_me (\x -> e)) x
1024 -- because the _inline_me gets dropped now it is applied,
1033 etaExpand :: Arity -- Result should have this number of value args
1035 -> CoreExpr -> Type -- Expression and its type
1037 -- (etaExpand n us e ty) returns an expression with
1038 -- the same meaning as 'e', but with arity 'n'.
1040 -- Given e' = etaExpand n us e ty
1042 -- ty = exprType e = exprType e'
1044 -- Note that SCCs are not treated specially. If we have
1045 -- etaExpand 2 (\x -> scc "foo" e)
1046 -- = (\xy -> (scc "foo" e) y)
1047 -- So the costs of evaluating 'e' (not 'e y') are attributed to "foo"
1049 etaExpand n us expr ty
1050 | manifestArity expr >= n = expr -- The no-op case
1052 = eta_expand n us expr ty
1055 -- manifestArity sees how many leading value lambdas there are
1056 manifestArity :: CoreExpr -> Arity
1057 manifestArity (Lam v e) | isId v = 1 + manifestArity e
1058 | otherwise = manifestArity e
1059 manifestArity (Note _ e) = manifestArity e
1060 manifestArity (Cast e _) = manifestArity e
1063 -- etaExpand deals with for-alls. For example:
1065 -- where E :: forall a. a -> a
1067 -- (/\b. \y::a -> E b y)
1069 -- It deals with coerces too, though they are now rare
1070 -- so perhaps the extra code isn't worth it
1072 eta_expand n us expr ty
1074 -- The ILX code generator requires eta expansion for type arguments
1075 -- too, but alas the 'n' doesn't tell us how many of them there
1076 -- may be. So we eagerly eta expand any big lambdas, and just
1077 -- cross our fingers about possible loss of sharing in the ILX case.
1078 -- The Right Thing is probably to make 'arity' include
1079 -- type variables throughout the compiler. (ToDo.)
1081 -- Saturated, so nothing to do
1084 -- Short cut for the case where there already
1085 -- is a lambda; no point in gratuitously adding more
1086 eta_expand n us (Lam v body) ty
1088 = Lam v (eta_expand n us body (applyTy ty (mkTyVarTy v)))
1091 = Lam v (eta_expand (n-1) us body (funResultTy ty))
1093 -- We used to have a special case that stepped inside Coerces here,
1094 -- thus: eta_expand n us (Note note@(Coerce _ ty) e) _
1095 -- = Note note (eta_expand n us e ty)
1096 -- BUT this led to an infinite loop
1097 -- Example: newtype T = MkT (Int -> Int)
1098 -- eta_expand 1 (coerce (Int->Int) e)
1099 -- --> coerce (Int->Int) (eta_expand 1 T e)
1101 -- --> coerce (Int->Int) (coerce T
1102 -- (\x::Int -> eta_expand 1 (coerce (Int->Int) e)))
1103 -- by the splitNewType_maybe case below
1106 eta_expand n us expr ty
1107 = ASSERT2 (exprType expr `coreEqType` ty, ppr (exprType expr) $$ ppr ty)
1108 case splitForAllTy_maybe ty of {
1109 Just (tv,ty') -> Lam tv (eta_expand n us (App expr (Type (mkTyVarTy tv))) ty')
1113 case splitFunTy_maybe ty of {
1114 Just (arg_ty, res_ty) -> Lam arg1 (eta_expand (n-1) us2 (App expr (Var arg1)) res_ty)
1116 arg1 = mkSysLocal FSLIT("eta") uniq arg_ty
1122 -- newtype T = MkT ([T] -> Int)
1123 -- Consider eta-expanding this
1126 -- coerce T (\x::[T] -> (coerce ([T]->Int) e) x)
1128 case splitNewTypeRepCo_maybe ty of {
1130 mkCoerce co (eta_expand n us (mkCoerce (mkSymCoercion co) expr) ty1) ;
1133 -- We have an expression of arity > 0, but its type isn't a function
1134 -- This *can* legitmately happen: e.g. coerce Int (\x. x)
1135 -- Essentially the programmer is playing fast and loose with types
1136 -- (Happy does this a lot). So we simply decline to eta-expand.
1141 exprArity is a cheap-and-cheerful version of exprEtaExpandArity.
1142 It tells how many things the expression can be applied to before doing
1143 any work. It doesn't look inside cases, lets, etc. The idea is that
1144 exprEtaExpandArity will do the hard work, leaving something that's easy
1145 for exprArity to grapple with. In particular, Simplify uses exprArity to
1146 compute the ArityInfo for the Id.
1148 Originally I thought that it was enough just to look for top-level lambdas, but
1149 it isn't. I've seen this
1151 foo = PrelBase.timesInt
1153 We want foo to get arity 2 even though the eta-expander will leave it
1154 unchanged, in the expectation that it'll be inlined. But occasionally it
1155 isn't, because foo is blacklisted (used in a rule).
1157 Similarly, see the ok_note check in exprEtaExpandArity. So
1158 f = __inline_me (\x -> e)
1159 won't be eta-expanded.
1161 And in any case it seems more robust to have exprArity be a bit more intelligent.
1162 But note that (\x y z -> f x y z)
1163 should have arity 3, regardless of f's arity.
1166 exprArity :: CoreExpr -> Arity
1169 go (Var v) = idArity v
1170 go (Lam x e) | isId x = go e + 1
1172 go (Note n e) = go e
1173 go (Cast e _) = go e
1174 go (App e (Type t)) = go e
1175 go (App f a) | exprIsCheap a = (go f - 1) `max` 0
1176 -- NB: exprIsCheap a!
1177 -- f (fac x) does not have arity 2,
1178 -- even if f has arity 3!
1179 -- NB: `max 0`! (\x y -> f x) has arity 2, even if f is
1180 -- unknown, hence arity 0
1184 %************************************************************************
1186 \subsection{Equality}
1188 %************************************************************************
1190 @cheapEqExpr@ is a cheap equality test which bales out fast!
1191 True => definitely equal
1192 False => may or may not be equal
1195 cheapEqExpr :: Expr b -> Expr b -> Bool
1197 cheapEqExpr (Var v1) (Var v2) = v1==v2
1198 cheapEqExpr (Lit lit1) (Lit lit2) = lit1 == lit2
1199 cheapEqExpr (Type t1) (Type t2) = t1 `coreEqType` t2
1201 cheapEqExpr (App f1 a1) (App f2 a2)
1202 = f1 `cheapEqExpr` f2 && a1 `cheapEqExpr` a2
1204 cheapEqExpr _ _ = False
1206 exprIsBig :: Expr b -> Bool
1207 -- Returns True of expressions that are too big to be compared by cheapEqExpr
1208 exprIsBig (Lit _) = False
1209 exprIsBig (Var v) = False
1210 exprIsBig (Type t) = False
1211 exprIsBig (App f a) = exprIsBig f || exprIsBig a
1212 exprIsBig other = True
1217 tcEqExpr :: CoreExpr -> CoreExpr -> Bool
1218 -- Used in rule matching, so does *not* look through
1219 -- newtypes, predicate types; hence tcEqExpr
1221 tcEqExpr e1 e2 = tcEqExprX rn_env e1 e2
1223 rn_env = mkRnEnv2 (mkInScopeSet (exprFreeVars e1 `unionVarSet` exprFreeVars e2))
1225 tcEqExprX :: RnEnv2 -> CoreExpr -> CoreExpr -> Bool
1226 tcEqExprX env (Var v1) (Var v2) = rnOccL env v1 == rnOccR env v2
1227 tcEqExprX env (Lit lit1) (Lit lit2) = lit1 == lit2
1228 tcEqExprX env (App f1 a1) (App f2 a2) = tcEqExprX env f1 f2 && tcEqExprX env a1 a2
1229 tcEqExprX env (Lam v1 e1) (Lam v2 e2) = tcEqExprX (rnBndr2 env v1 v2) e1 e2
1230 tcEqExprX env (Let (NonRec v1 r1) e1)
1231 (Let (NonRec v2 r2) e2) = tcEqExprX env r1 r2
1232 && tcEqExprX (rnBndr2 env v1 v2) e1 e2
1233 tcEqExprX env (Let (Rec ps1) e1)
1234 (Let (Rec ps2) e2) = equalLength ps1 ps2
1235 && and (zipWith eq_rhs ps1 ps2)
1236 && tcEqExprX env' e1 e2
1238 env' = foldl2 rn_bndr2 env ps2 ps2
1239 rn_bndr2 env (b1,_) (b2,_) = rnBndr2 env b1 b2
1240 eq_rhs (_,r1) (_,r2) = tcEqExprX env' r1 r2
1241 tcEqExprX env (Case e1 v1 t1 a1)
1242 (Case e2 v2 t2 a2) = tcEqExprX env e1 e2
1243 && tcEqTypeX env t1 t2
1244 && equalLength a1 a2
1245 && and (zipWith (eq_alt env') a1 a2)
1247 env' = rnBndr2 env v1 v2
1249 tcEqExprX env (Note n1 e1) (Note n2 e2) = eq_note env n1 n2 && tcEqExprX env e1 e2
1250 tcEqExprX env (Cast e1 co1) (Cast e2 co2) = tcEqTypeX env co1 co2 && tcEqExprX env e1 e2
1251 tcEqExprX env (Type t1) (Type t2) = tcEqTypeX env t1 t2
1252 tcEqExprX env e1 e2 = False
1254 eq_alt env (c1,vs1,r1) (c2,vs2,r2) = c1==c2 && tcEqExprX (rnBndrs2 env vs1 vs2) r1 r2
1256 eq_note env (SCC cc1) (SCC cc2) = cc1 == cc2
1257 eq_note env (CoreNote s1) (CoreNote s2) = s1 == s2
1258 eq_note env other1 other2 = False
1262 %************************************************************************
1264 \subsection{The size of an expression}
1266 %************************************************************************
1269 coreBindsSize :: [CoreBind] -> Int
1270 coreBindsSize bs = foldr ((+) . bindSize) 0 bs
1272 exprSize :: CoreExpr -> Int
1273 -- A measure of the size of the expressions
1274 -- It also forces the expression pretty drastically as a side effect
1275 exprSize (Var v) = v `seq` 1
1276 exprSize (Lit lit) = lit `seq` 1
1277 exprSize (App f a) = exprSize f + exprSize a
1278 exprSize (Lam b e) = varSize b + exprSize e
1279 exprSize (Let b e) = bindSize b + exprSize e
1280 exprSize (Case e b t as) = seqType t `seq` exprSize e + varSize b + 1 + foldr ((+) . altSize) 0 as
1281 exprSize (Cast e co) = (seqType co `seq` 1) + exprSize e
1282 exprSize (Note n e) = noteSize n + exprSize e
1283 exprSize (Type t) = seqType t `seq` 1
1285 noteSize (SCC cc) = cc `seq` 1
1286 noteSize InlineMe = 1
1287 noteSize (CoreNote s) = s `seq` 1 -- hdaume: core annotations
1289 varSize :: Var -> Int
1290 varSize b | isTyVar b = 1
1291 | otherwise = seqType (idType b) `seq`
1292 megaSeqIdInfo (idInfo b) `seq`
1295 varsSize = foldr ((+) . varSize) 0
1297 bindSize (NonRec b e) = varSize b + exprSize e
1298 bindSize (Rec prs) = foldr ((+) . pairSize) 0 prs
1300 pairSize (b,e) = varSize b + exprSize e
1302 altSize (c,bs,e) = c `seq` varsSize bs + exprSize e
1306 %************************************************************************
1308 \subsection{Hashing}
1310 %************************************************************************
1313 hashExpr :: CoreExpr -> Int
1314 -- Two expressions that hash to the same Int may be equal (but may not be)
1315 -- Two expressions that hash to the different Ints are definitely unequal
1317 -- But "unequal" here means "not identical"; two alpha-equivalent
1318 -- expressions may hash to the different Ints
1320 -- The emphasis is on a crude, fast hash, rather than on high precision
1322 hashExpr e | hash < 0 = 77 -- Just in case we hit -maxInt
1325 hash = abs (hash_expr e) -- Negative numbers kill UniqFM
1327 hash_expr (Note _ e) = hash_expr e
1328 hash_expr (Cast e co) = hash_expr e
1329 hash_expr (Let (NonRec b r) e) = hashId b
1330 hash_expr (Let (Rec ((b,r):_)) e) = hashId b
1331 hash_expr (Case _ b _ _) = hashId b
1332 hash_expr (App f e) = hash_expr f * fast_hash_expr e
1333 hash_expr (Var v) = hashId v
1334 hash_expr (Lit lit) = hashLiteral lit
1335 hash_expr (Lam b _) = hashId b
1336 hash_expr (Type t) = trace "hash_expr: type" 1 -- Shouldn't happen
1338 fast_hash_expr (Var v) = hashId v
1339 fast_hash_expr (Lit lit) = hashLiteral lit
1340 fast_hash_expr (App f (Type _)) = fast_hash_expr f
1341 fast_hash_expr (App f a) = fast_hash_expr a
1342 fast_hash_expr (Lam b _) = hashId b
1343 fast_hash_expr other = 1
1346 hashId id = hashName (idName id)
1349 %************************************************************************
1351 \subsection{Determining non-updatable right-hand-sides}
1353 %************************************************************************
1355 Top-level constructor applications can usually be allocated
1356 statically, but they can't if the constructor, or any of the
1357 arguments, come from another DLL (because we can't refer to static
1358 labels in other DLLs).
1360 If this happens we simply make the RHS into an updatable thunk,
1361 and 'exectute' it rather than allocating it statically.
1364 rhsIsStatic :: PackageId -> CoreExpr -> Bool
1365 -- This function is called only on *top-level* right-hand sides
1366 -- Returns True if the RHS can be allocated statically, with
1367 -- no thunks involved at all.
1369 -- It's called (i) in TidyPgm.hasCafRefs to decide if the rhs is, or
1370 -- refers to, CAFs; and (ii) in CoreToStg to decide whether to put an
1371 -- update flag on it.
1373 -- The basic idea is that rhsIsStatic returns True only if the RHS is
1374 -- (a) a value lambda
1375 -- (b) a saturated constructor application with static args
1377 -- BUT watch out for
1378 -- (i) Any cross-DLL references kill static-ness completely
1379 -- because they must be 'executed' not statically allocated
1380 -- ("DLL" here really only refers to Windows DLLs, on other platforms,
1381 -- this is not necessary)
1383 -- (ii) We treat partial applications as redexes, because in fact we
1384 -- make a thunk for them that runs and builds a PAP
1385 -- at run-time. The only appliations that are treated as
1386 -- static are *saturated* applications of constructors.
1388 -- We used to try to be clever with nested structures like this:
1389 -- ys = (:) w ((:) w [])
1390 -- on the grounds that CorePrep will flatten ANF-ise it later.
1391 -- But supporting this special case made the function much more
1392 -- complicated, because the special case only applies if there are no
1393 -- enclosing type lambdas:
1394 -- ys = /\ a -> Foo (Baz ([] a))
1395 -- Here the nested (Baz []) won't float out to top level in CorePrep.
1397 -- But in fact, even without -O, nested structures at top level are
1398 -- flattened by the simplifier, so we don't need to be super-clever here.
1402 -- f = \x::Int. x+7 TRUE
1403 -- p = (True,False) TRUE
1405 -- d = (fst p, False) FALSE because there's a redex inside
1406 -- (this particular one doesn't happen but...)
1408 -- h = D# (1.0## /## 2.0##) FALSE (redex again)
1409 -- n = /\a. Nil a TRUE
1411 -- t = /\a. (:) (case w a of ...) (Nil a) FALSE (redex)
1414 -- This is a bit like CoreUtils.exprIsHNF, with the following differences:
1415 -- a) scc "foo" (\x -> ...) is updatable (so we catch the right SCC)
1417 -- b) (C x xs), where C is a contructors is updatable if the application is
1420 -- c) don't look through unfolding of f in (f x).
1422 -- When opt_RuntimeTypes is on, we keep type lambdas and treat
1423 -- them as making the RHS re-entrant (non-updatable).
1425 rhsIsStatic this_pkg rhs = is_static False rhs
1427 is_static :: Bool -- True <=> in a constructor argument; must be atomic
1430 is_static False (Lam b e) = isRuntimeVar b || is_static False e
1432 is_static in_arg (Note (SCC _) e) = False
1433 is_static in_arg (Note _ e) = is_static in_arg e
1434 is_static in_arg (Cast e co) = is_static in_arg e
1436 is_static in_arg (Lit lit)
1438 MachLabel _ _ -> False
1440 -- A MachLabel (foreign import "&foo") in an argument
1441 -- prevents a constructor application from being static. The
1442 -- reason is that it might give rise to unresolvable symbols
1443 -- in the object file: under Linux, references to "weak"
1444 -- symbols from the data segment give rise to "unresolvable
1445 -- relocation" errors at link time This might be due to a bug
1446 -- in the linker, but we'll work around it here anyway.
1449 is_static in_arg other_expr = go other_expr 0
1451 go (Var f) n_val_args
1452 #if mingw32_TARGET_OS
1453 | not (isDllName this_pkg (idName f))
1455 = saturated_data_con f n_val_args
1456 || (in_arg && n_val_args == 0)
1457 -- A naked un-applied variable is *not* deemed a static RHS
1459 -- Reason: better to update so that the indirection gets shorted
1460 -- out, and the true value will be seen
1461 -- NB: if you change this, you'll break the invariant that THUNK_STATICs
1462 -- are always updatable. If you do so, make sure that non-updatable
1463 -- ones have enough space for their static link field!
1465 go (App f a) n_val_args
1466 | isTypeArg a = go f n_val_args
1467 | not in_arg && is_static True a = go f (n_val_args + 1)
1468 -- The (not in_arg) checks that we aren't in a constructor argument;
1469 -- if we are, we don't allow (value) applications of any sort
1471 -- NB. In case you wonder, args are sometimes not atomic. eg.
1472 -- x = D# (1.0## /## 2.0##)
1473 -- can't float because /## can fail.
1475 go (Note (SCC _) f) n_val_args = False
1476 go (Note _ f) n_val_args = go f n_val_args
1477 go (Cast e co) n_val_args = go e n_val_args
1479 go other n_val_args = False
1481 saturated_data_con f n_val_args
1482 = case isDataConWorkId_maybe f of
1483 Just dc -> n_val_args == dataConRepArity dc