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,
36 dataConInstPat, dataConOccInstPat
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 OccName ( OccName, occNameFS, mkVarOcc )
50 import VarSet ( unionVarSet )
52 import Name ( hashName, mkSysTvName )
54 import Packages ( isDllName )
56 import Literal ( hashLiteral, literalType, litIsDupable,
57 litIsTrivial, isZeroLit, Literal( MachLabel ) )
58 import DataCon ( DataCon, dataConRepArity, eqSpecPreds,
59 isVanillaDataCon, dataConTyCon, dataConRepArgTys,
60 dataConUnivTyVars, dataConExTyVars, dataConEqSpec )
61 import PrimOp ( PrimOp(..), primOpOkForSpeculation, primOpIsCheap )
62 import Id ( Id, idType, globalIdDetails, idNewStrictness,
63 mkWildId, idArity, idName, idUnfolding, idInfo,
64 isOneShotBndr, isStateHackType, isDataConWorkId_maybe, mkSysLocal,
65 isDataConWorkId, isBottomingId, isDictId
67 import IdInfo ( GlobalIdDetails(..), megaSeqIdInfo )
68 import NewDemand ( appIsBottom )
69 import Type ( Type, mkFunTy, mkForAllTy, splitFunTy_maybe,
70 splitFunTy, tcEqTypeX,
71 applyTys, isUnLiftedType, seqType, mkTyVarTy,
72 splitForAllTy_maybe, isForAllTy, splitRecNewType_maybe,
73 splitTyConApp_maybe, coreEqType, funResultTy, applyTy,
76 import Coercion ( Coercion, mkTransCoercion, coercionKind,
77 splitNewTypeRepCo_maybe, mkSymCoercion, mkLeftCoercion,
78 mkRightCoercion, decomposeCo, coercionKindPredTy,
79 splitCoercionKind, mkEqPred )
80 import TyCon ( tyConArity )
81 import TysWiredIn ( boolTy, trueDataCon, falseDataCon )
82 import CostCentre ( CostCentre )
83 import BasicTypes ( Arity )
84 import PackageConfig ( PackageId )
85 import Unique ( Unique )
87 import DynFlags ( DynFlags, DynFlag(Opt_DictsCheap), dopt )
88 import TysPrim ( alphaTy ) -- Debugging only
89 import Util ( equalLength, lengthAtLeast, foldl2 )
90 import FastString ( mkFastString )
94 %************************************************************************
96 \subsection{Find the type of a Core atom/expression}
98 %************************************************************************
101 exprType :: CoreExpr -> Type
103 exprType (Var var) = idType var
104 exprType (Lit lit) = literalType lit
105 exprType (Let _ body) = exprType body
106 exprType (Case _ _ ty alts) = ty
108 = let (_, ty) = coercionKind co in ty
109 exprType (Note other_note e) = exprType e
110 exprType (Lam binder expr) = mkPiType binder (exprType expr)
112 = case collectArgs e of
113 (fun, args) -> applyTypeToArgs e (exprType fun) args
115 exprType other = pprTrace "exprType" (pprCoreExpr other) alphaTy
117 coreAltType :: CoreAlt -> Type
118 coreAltType (_,_,rhs) = exprType rhs
121 @mkPiType@ makes a (->) type or a forall type, depending on whether
122 it is given a type variable or a term variable. We cleverly use the
123 lbvarinfo field to figure out the right annotation for the arrove in
124 case of a term variable.
127 mkPiType :: Var -> Type -> Type -- The more polymorphic version
128 mkPiTypes :: [Var] -> Type -> Type -- doesn't work...
130 mkPiTypes vs ty = foldr mkPiType ty vs
133 | isId v = mkFunTy (idType v) ty
134 | otherwise = mkForAllTy v ty
138 applyTypeToArg :: Type -> CoreExpr -> Type
139 applyTypeToArg fun_ty (Type arg_ty) = applyTy fun_ty arg_ty
140 applyTypeToArg fun_ty other_arg = funResultTy fun_ty
142 applyTypeToArgs :: CoreExpr -> Type -> [CoreExpr] -> Type
143 -- A more efficient version of applyTypeToArg
144 -- when we have several args
145 -- The first argument is just for debugging
146 applyTypeToArgs e op_ty [] = op_ty
148 applyTypeToArgs e op_ty (Type ty : args)
149 = -- Accumulate type arguments so we can instantiate all at once
152 go rev_tys (Type ty : args) = go (ty:rev_tys) args
153 go rev_tys rest_args = applyTypeToArgs e op_ty' rest_args
155 op_ty' = applyTys op_ty (reverse rev_tys)
157 applyTypeToArgs e op_ty (other_arg : args)
158 = case (splitFunTy_maybe op_ty) of
159 Just (_, res_ty) -> applyTypeToArgs e res_ty args
160 Nothing -> pprPanic "applyTypeToArgs" (pprCoreExpr e $$ ppr op_ty)
165 %************************************************************************
167 \subsection{Attaching notes}
169 %************************************************************************
171 mkNote removes redundant coercions, and SCCs where possible
175 mkNote :: Note -> CoreExpr -> CoreExpr
176 mkNote (SCC cc) expr = mkSCC cc expr
177 mkNote InlineMe expr = mkInlineMe expr
178 mkNote note expr = Note note expr
182 Drop trivial InlineMe's. This is somewhat important, because if we have an unfolding
183 that looks like (Note InlineMe (Var v)), the InlineMe doesn't go away because it may
184 not be *applied* to anything.
186 We don't use exprIsTrivial here, though, because we sometimes generate worker/wrapper
189 f = inline_me (coerce t fw)
190 As usual, the inline_me prevents the worker from getting inlined back into the wrapper.
191 We want the split, so that the coerces can cancel at the call site.
193 However, we can get left with tiresome type applications. Notably, consider
194 f = /\ a -> let t = e in (t, w)
195 Then lifting the let out of the big lambda gives
197 f = /\ a -> let t = inline_me (t' a) in (t, w)
198 The inline_me is to stop the simplifier inlining t' right back
199 into t's RHS. In the next phase we'll substitute for t (since
200 its rhs is trivial) and *then* we could get rid of the inline_me.
201 But it hardly seems worth it, so I don't bother.
204 mkInlineMe (Var v) = Var v
205 mkInlineMe e = Note InlineMe e
211 mkCoerce :: Coercion -> CoreExpr -> CoreExpr
212 mkCoerce co (Cast expr co2)
213 = ASSERT(let { (from_ty, to_ty) = coercionKind co;
214 (from_ty2, to_ty2) = coercionKind co2} in
215 from_ty `coreEqType` to_ty2 )
216 mkCoerce (mkTransCoercion co2 co) expr
219 = let (from_ty, to_ty) = coercionKind co in
220 -- if to_ty `coreEqType` from_ty
223 ASSERT2(from_ty `coreEqType` (exprType expr), text "Trying to coerce" <+> text "(" <> ppr expr $$ text "::" <+> ppr (exprType expr) <> text ")" $$ ppr co $$ ppr (coercionKindPredTy co))
228 mkSCC :: CostCentre -> Expr b -> Expr b
229 -- Note: Nested SCC's *are* preserved for the benefit of
230 -- cost centre stack profiling
231 mkSCC cc (Lit lit) = Lit lit
232 mkSCC cc (Lam x e) = Lam x (mkSCC cc e) -- Move _scc_ inside lambda
233 mkSCC cc (Note (SCC cc') e) = Note (SCC cc) (Note (SCC cc') e)
234 mkSCC cc (Note n e) = Note n (mkSCC cc e) -- Move _scc_ inside notes
235 mkSCC cc (Cast e co) = Cast (mkSCC cc e) co -- Move _scc_ inside cast
236 mkSCC cc expr = Note (SCC cc) expr
240 %************************************************************************
242 \subsection{Other expression construction}
244 %************************************************************************
247 bindNonRec :: Id -> CoreExpr -> CoreExpr -> CoreExpr
248 -- (bindNonRec x r b) produces either
251 -- case r of x { _DEFAULT_ -> b }
253 -- depending on whether x is unlifted or not
254 -- It's used by the desugarer to avoid building bindings
255 -- that give Core Lint a heart attack. Actually the simplifier
256 -- deals with them perfectly well.
258 bindNonRec bndr rhs body
259 | needsCaseBinding (idType bndr) rhs = Case rhs bndr (exprType body) [(DEFAULT,[],body)]
260 | otherwise = Let (NonRec bndr rhs) body
262 needsCaseBinding ty rhs = isUnLiftedType ty && not (exprOkForSpeculation rhs)
263 -- Make a case expression instead of a let
264 -- These can arise either from the desugarer,
265 -- or from beta reductions: (\x.e) (x +# y)
269 mkAltExpr :: AltCon -> [CoreBndr] -> [Type] -> CoreExpr
270 -- This guy constructs the value that the scrutinee must have
271 -- when you are in one particular branch of a case
272 mkAltExpr (DataAlt con) args inst_tys
273 = mkConApp con (map Type inst_tys ++ varsToCoreExprs args)
274 mkAltExpr (LitAlt lit) [] []
277 mkIfThenElse :: CoreExpr -> CoreExpr -> CoreExpr -> CoreExpr
278 mkIfThenElse guard then_expr else_expr
279 -- Not going to be refining, so okay to take the type of the "then" clause
280 = Case guard (mkWildId boolTy) (exprType then_expr)
281 [ (DataAlt falseDataCon, [], else_expr), -- Increasing order of tag!
282 (DataAlt trueDataCon, [], then_expr) ]
286 %************************************************************************
288 \subsection{Taking expressions apart}
290 %************************************************************************
292 The default alternative must be first, if it exists at all.
293 This makes it easy to find, though it makes matching marginally harder.
296 findDefault :: [CoreAlt] -> ([CoreAlt], Maybe CoreExpr)
297 findDefault ((DEFAULT,args,rhs) : alts) = ASSERT( null args ) (alts, Just rhs)
298 findDefault alts = (alts, Nothing)
300 findAlt :: AltCon -> [CoreAlt] -> CoreAlt
303 (deflt@(DEFAULT,_,_):alts) -> go alts deflt
304 other -> go alts panic_deflt
306 panic_deflt = pprPanic "Missing alternative" (ppr con $$ vcat (map ppr alts))
309 go (alt@(con1,_,_) : alts) deflt
310 = case con `cmpAltCon` con1 of
311 LT -> deflt -- Missed it already; the alts are in increasing order
313 GT -> ASSERT( not (con1 == DEFAULT) ) go alts deflt
315 isDefaultAlt :: CoreAlt -> Bool
316 isDefaultAlt (DEFAULT, _, _) = True
317 isDefaultAlt other = False
319 ---------------------------------
320 mergeAlts :: [CoreAlt] -> [CoreAlt] -> [CoreAlt]
321 -- Merge preserving order; alternatives in the first arg
322 -- shadow ones in the second
323 mergeAlts [] as2 = as2
324 mergeAlts as1 [] = as1
325 mergeAlts (a1:as1) (a2:as2)
326 = case a1 `cmpAlt` a2 of
327 LT -> a1 : mergeAlts as1 (a2:as2)
328 EQ -> a1 : mergeAlts as1 as2 -- Discard a2
329 GT -> a2 : mergeAlts (a1:as1) as2
333 %************************************************************************
335 \subsection{Figuring out things about expressions}
337 %************************************************************************
339 @exprIsTrivial@ is true of expressions we are unconditionally happy to
340 duplicate; simple variables and constants, and type
341 applications. Note that primop Ids aren't considered
344 @exprIsBottom@ is true of expressions that are guaranteed to diverge
347 There used to be a gruesome test for (hasNoBinding v) in the
349 exprIsTrivial (Var v) | hasNoBinding v = idArity v == 0
350 The idea here is that a constructor worker, like $wJust, is
351 really short for (\x -> $wJust x), becuase $wJust has no binding.
352 So it should be treated like a lambda. Ditto unsaturated primops.
353 But now constructor workers are not "have-no-binding" Ids. And
354 completely un-applied primops and foreign-call Ids are sufficiently
355 rare that I plan to allow them to be duplicated and put up with
358 SCC notes. We do not treat (_scc_ "foo" x) as trivial, because
359 a) it really generates code, (and a heap object when it's
360 a function arg) to capture the cost centre
361 b) see the note [SCC-and-exprIsTrivial] in Simplify.simplLazyBind
364 exprIsTrivial (Var v) = True -- See notes above
365 exprIsTrivial (Type _) = True
366 exprIsTrivial (Lit lit) = litIsTrivial lit
367 exprIsTrivial (App e arg) = not (isRuntimeArg arg) && exprIsTrivial e
368 exprIsTrivial (Note (SCC _) e) = False -- See notes above
369 exprIsTrivial (Note _ e) = exprIsTrivial e
370 exprIsTrivial (Cast e co) = exprIsTrivial e
371 exprIsTrivial (Lam b body) = not (isRuntimeVar b) && exprIsTrivial body
372 exprIsTrivial other = False
376 @exprIsDupable@ is true of expressions that can be duplicated at a modest
377 cost in code size. This will only happen in different case
378 branches, so there's no issue about duplicating work.
380 That is, exprIsDupable returns True of (f x) even if
381 f is very very expensive to call.
383 Its only purpose is to avoid fruitless let-binding
384 and then inlining of case join points
388 exprIsDupable (Type _) = True
389 exprIsDupable (Var v) = True
390 exprIsDupable (Lit lit) = litIsDupable lit
391 exprIsDupable (Note InlineMe e) = True
392 exprIsDupable (Note _ e) = exprIsDupable e
393 exprIsDupable (Cast e co) = exprIsDupable e
397 go (Var v) n_args = True
398 go (App f a) n_args = n_args < dupAppSize
401 go other n_args = False
404 dupAppSize = 4 -- Size of application we are prepared to duplicate
407 @exprIsCheap@ looks at a Core expression and returns \tr{True} if
408 it is obviously in weak head normal form, or is cheap to get to WHNF.
409 [Note that that's not the same as exprIsDupable; an expression might be
410 big, and hence not dupable, but still cheap.]
412 By ``cheap'' we mean a computation we're willing to:
413 push inside a lambda, or
414 inline at more than one place
415 That might mean it gets evaluated more than once, instead of being
416 shared. The main examples of things which aren't WHNF but are
421 (where e, and all the ei are cheap)
424 (where e and b are cheap)
427 (where op is a cheap primitive operator)
430 (because we are happy to substitute it inside a lambda)
432 Notice that a variable is considered 'cheap': we can push it inside a lambda,
433 because sharing will make sure it is only evaluated once.
436 exprIsCheap :: CoreExpr -> Bool
437 exprIsCheap (Lit lit) = True
438 exprIsCheap (Type _) = True
439 exprIsCheap (Var _) = True
440 exprIsCheap (Note InlineMe e) = True
441 exprIsCheap (Note _ e) = exprIsCheap e
442 exprIsCheap (Cast e co) = exprIsCheap e
443 exprIsCheap (Lam x e) = isRuntimeVar x || exprIsCheap e
444 exprIsCheap (Case e _ _ alts) = exprIsCheap e &&
445 and [exprIsCheap rhs | (_,_,rhs) <- alts]
446 -- Experimentally, treat (case x of ...) as cheap
447 -- (and case __coerce x etc.)
448 -- This improves arities of overloaded functions where
449 -- there is only dictionary selection (no construction) involved
450 exprIsCheap (Let (NonRec x _) e)
451 | isUnLiftedType (idType x) = exprIsCheap e
453 -- strict lets always have cheap right hand sides,
454 -- and do no allocation.
456 exprIsCheap other_expr -- Applications and variables
459 -- Accumulate value arguments, then decide
460 go (App f a) val_args | isRuntimeArg a = go f (a:val_args)
461 | otherwise = go f val_args
463 go (Var f) [] = True -- Just a type application of a variable
464 -- (f t1 t2 t3) counts as WHNF
466 = case globalIdDetails f of
467 RecordSelId {} -> go_sel args
468 ClassOpId _ -> go_sel args
469 PrimOpId op -> go_primop op args
471 DataConWorkId _ -> go_pap args
472 other | length args < idArity f -> go_pap args
474 other -> isBottomingId f
475 -- Application of a function which
476 -- always gives bottom; we treat this as cheap
477 -- because it certainly doesn't need to be shared!
479 go other args = False
482 go_pap args = all exprIsTrivial args
483 -- For constructor applications and primops, check that all
484 -- the args are trivial. We don't want to treat as cheap, say,
486 -- We'll put up with one constructor application, but not dozens
489 go_primop op args = primOpIsCheap op && all exprIsCheap args
490 -- In principle we should worry about primops
491 -- that return a type variable, since the result
492 -- might be applied to something, but I'm not going
493 -- to bother to check the number of args
496 go_sel [arg] = exprIsCheap arg -- I'm experimenting with making record selection
497 go_sel other = False -- look cheap, so we will substitute it inside a
498 -- lambda. Particularly for dictionary field selection.
499 -- BUT: Take care with (sel d x)! The (sel d) might be cheap, but
500 -- there's no guarantee that (sel d x) will be too. Hence (n_val_args == 1)
503 exprOkForSpeculation returns True of an expression that it is
505 * safe to evaluate even if normal order eval might not
506 evaluate the expression at all, or
508 * safe *not* to evaluate even if normal order would do so
512 the expression guarantees to terminate,
514 without raising an exception,
515 without causing a side effect (e.g. writing a mutable variable)
517 NB: if exprIsHNF e, then exprOkForSpecuation e
520 let x = case y# +# 1# of { r# -> I# r# }
523 case y# +# 1# of { r# ->
528 We can only do this if the (y+1) is ok for speculation: it has no
529 side effects, and can't diverge or raise an exception.
532 exprOkForSpeculation :: CoreExpr -> Bool
533 exprOkForSpeculation (Lit _) = True
534 exprOkForSpeculation (Type _) = True
535 exprOkForSpeculation (Var v) = isUnLiftedType (idType v)
536 exprOkForSpeculation (Note _ e) = exprOkForSpeculation e
537 exprOkForSpeculation (Cast e co) = exprOkForSpeculation e
538 exprOkForSpeculation other_expr
539 = case collectArgs other_expr of
540 (Var f, args) -> spec_ok (globalIdDetails f) args
544 spec_ok (DataConWorkId _) args
545 = True -- The strictness of the constructor has already
546 -- been expressed by its "wrapper", so we don't need
547 -- to take the arguments into account
549 spec_ok (PrimOpId op) args
550 | isDivOp op, -- Special case for dividing operations that fail
551 [arg1, Lit lit] <- args -- only if the divisor is zero
552 = not (isZeroLit lit) && exprOkForSpeculation arg1
553 -- Often there is a literal divisor, and this
554 -- can get rid of a thunk in an inner looop
557 = primOpOkForSpeculation op &&
558 all exprOkForSpeculation args
559 -- A bit conservative: we don't really need
560 -- to care about lazy arguments, but this is easy
562 spec_ok other args = False
564 isDivOp :: PrimOp -> Bool
565 -- True of dyadic operators that can fail
566 -- only if the second arg is zero
567 -- This function probably belongs in PrimOp, or even in
568 -- an automagically generated file.. but it's such a
569 -- special case I thought I'd leave it here for now.
570 isDivOp IntQuotOp = True
571 isDivOp IntRemOp = True
572 isDivOp WordQuotOp = True
573 isDivOp WordRemOp = True
574 isDivOp IntegerQuotRemOp = True
575 isDivOp IntegerDivModOp = True
576 isDivOp FloatDivOp = True
577 isDivOp DoubleDivOp = True
578 isDivOp other = False
583 exprIsBottom :: CoreExpr -> Bool -- True => definitely bottom
584 exprIsBottom e = go 0 e
586 -- n is the number of args
587 go n (Note _ e) = go n e
588 go n (Cast e co) = go n e
589 go n (Let _ e) = go n e
590 go n (Case e _ _ _) = go 0 e -- Just check the scrut
591 go n (App e _) = go (n+1) e
592 go n (Var v) = idAppIsBottom v n
594 go n (Lam _ _) = False
595 go n (Type _) = False
597 idAppIsBottom :: Id -> Int -> Bool
598 idAppIsBottom id n_val_args = appIsBottom (idNewStrictness id) n_val_args
601 @exprIsHNF@ returns true for expressions that are certainly *already*
602 evaluated to *head* normal form. This is used to decide whether it's ok
605 case x of _ -> e ===> e
607 and to decide whether it's safe to discard a `seq`
609 So, it does *not* treat variables as evaluated, unless they say they are.
611 But it *does* treat partial applications and constructor applications
612 as values, even if their arguments are non-trivial, provided the argument
614 e.g. (:) (f x) (map f xs) is a value
615 map (...redex...) is a value
616 Because `seq` on such things completes immediately
618 For unlifted argument types, we have to be careful:
620 Suppose (f x) diverges; then C (f x) is not a value. True, but
621 this form is illegal (see the invariants in CoreSyn). Args of unboxed
622 type must be ok-for-speculation (or trivial).
625 exprIsHNF :: CoreExpr -> Bool -- True => Value-lambda, constructor, PAP
626 exprIsHNF (Var v) -- NB: There are no value args at this point
627 = isDataConWorkId v -- Catches nullary constructors,
628 -- so that [] and () are values, for example
629 || idArity v > 0 -- Catches (e.g.) primops that don't have unfoldings
630 || isEvaldUnfolding (idUnfolding v)
631 -- Check the thing's unfolding; it might be bound to a value
632 -- A worry: what if an Id's unfolding is just itself:
633 -- then we could get an infinite loop...
635 exprIsHNF (Lit l) = True
636 exprIsHNF (Type ty) = True -- Types are honorary Values;
637 -- we don't mind copying them
638 exprIsHNF (Lam b e) = isRuntimeVar b || exprIsHNF e
639 exprIsHNF (Note _ e) = exprIsHNF e
640 exprIsHNF (Cast e co) = exprIsHNF e
641 exprIsHNF (App e (Type _)) = exprIsHNF e
642 exprIsHNF (App e a) = app_is_value e [a]
643 exprIsHNF other = False
645 -- There is at least one value argument
646 app_is_value (Var fun) args
647 | isDataConWorkId fun -- Constructor apps are values
648 || idArity fun > valArgCount args -- Under-applied function
649 = check_args (idType fun) args
650 app_is_value (App f a) as = app_is_value f (a:as)
651 app_is_value other as = False
653 -- 'check_args' checks that unlifted-type args
654 -- are in fact guaranteed non-divergent
655 check_args fun_ty [] = True
656 check_args fun_ty (Type _ : args) = case splitForAllTy_maybe fun_ty of
657 Just (_, ty) -> check_args ty args
658 check_args fun_ty (arg : args)
659 | isUnLiftedType arg_ty = exprOkForSpeculation arg
660 | otherwise = check_args res_ty args
662 (arg_ty, res_ty) = splitFunTy fun_ty
666 -- deep applies a TyConApp coercion as a substitution to a reflexive coercion
667 -- deepCast t [a1,...,an] co corresponds to deep(t, [a1,...,an], co) from
669 deepCast :: Type -> [TyVar] -> Coercion -> Coercion
670 deepCast ty tyVars co
671 = ASSERT( let {(lty, rty) = coercionKind co;
672 Just (tc1, lArgs) = splitTyConApp_maybe lty;
673 Just (tc2, rArgs) = splitTyConApp_maybe rty}
675 tc1 == tc2 && length lArgs == length rArgs &&
676 length lArgs == length tyVars )
677 substTyWith tyVars coArgs ty
679 -- coArgs = [right (left (left co)), right (left co), right co]
680 coArgs = decomposeCo (length tyVars) co
682 -- This goes here to avoid circularity between DataCon and Id
683 dataConInstPat :: [Unique] -- A long enough list of uniques, at least one for each binder
685 -> [Type] -- Types to instantiate the universally quantified tyvars
686 -> ([TyVar], [CoVar], [Id]) -- Return instantiated variables
687 -- dataConInstPat us con inst_tys returns a triple (ex_tvs, co_tvs, arg_ids),
689 -- ex_tvs are intended to be used as binders for existential type args
691 -- co_tvs are intended to be used as binders for coercion args and the kinds
692 -- of these vars have been instantiated by the inst_tys and the ex_tys
694 -- arg_ids are indended to be used as binders for value arguments, including
695 -- dicts, and have their types instantiated with inst_tys and ex_tys
698 -- The following constructor T1
701 -- T1 :: forall b. Int -> b -> T(a,b)
704 -- has representation type
705 -- forall a. forall a1. forall a2. forall b. (a :=: (a1,a2)) =>
708 -- dataConInstPat us T1 (a1',a2') will return
710 -- ([a1'', a2'', b''],[c :: (a1',a2'):=:(a1'',a2'')],[x :: Int,y :: b''])
712 -- where the double-primed variables are created from the unique list input
713 dataConInstPat uniqs con inst_tys
714 = dataConOccInstPat uniqs occs con inst_tys
716 -- dataConOccInstPat doesn't actually make use of the OccName directly for
717 -- existential and coercion variable binders, so it is right to just
718 -- use the VarName namespace for all of the OccNames
720 mk_occs n = mkVarOcc ("ipv" ++ show n) : mk_occs (n+1)
722 dataConOccInstPat :: [Unique] -- A long enough list of uniques, at least one for each binder
723 -> [OccName] -- An equally long list of OccNames to use
725 -> [Type] -- Types to instantiate the universally quantified tyvars
726 -> ([TyVar], [CoVar], [Id]) -- Return instantiated variables
727 -- This function actually does the job specified in the comment for
728 -- dataConInstPat, but uses the specified list of OccNames. This is
729 -- is necessary for use in e.g. tcIfaceDataAlt
730 dataConOccInstPat uniqs occs con inst_tys
731 = (ex_bndrs, co_bndrs, id_bndrs)
733 univ_tvs = dataConUnivTyVars con
734 ex_tvs = dataConExTyVars con
735 arg_tys = dataConRepArgTys con
736 eq_spec = dataConEqSpec con
737 eq_preds = eqSpecPreds eq_spec
740 n_co = length eq_spec
741 n_id = length arg_tys
743 -- split the Uniques and OccNames
744 (ex_uniqs, uniqs') = splitAt n_ex uniqs
745 (co_uniqs, id_uniqs) = splitAt n_co uniqs'
747 (ex_occs, occs') = splitAt n_ex occs
748 (co_occs, id_occs) = splitAt n_co occs'
750 -- make existential type variables
751 mk_ex_var uniq occ var = mkTyVar new_name kind
753 new_name = mkSysTvName uniq (occNameFS occ)
756 ex_bndrs = zipWith3 mk_ex_var ex_uniqs ex_occs ex_tvs
758 -- make the instantiation substitution
759 inst_subst = substTyWith (univ_tvs ++ ex_tvs) (inst_tys ++ map mkTyVarTy ex_bndrs)
761 -- make new coercion vars, instantiating kind
762 mk_co_var uniq occ eq_pred = mkCoVar new_name (inst_subst (mkPredTy eq_pred))
764 new_name = mkSysTvName uniq (occNameFS occ)
766 co_bndrs = zipWith3 mk_co_var co_uniqs co_occs eq_preds
768 -- make value vars, instantiating types
769 mk_id_var uniq occ ty = mkUserLocal occ uniq (inst_subst ty) noSrcLoc
770 id_bndrs = zipWith3 mk_id_var id_uniqs id_occs arg_tys
772 exprIsConApp_maybe :: CoreExpr -> Maybe (DataCon, [CoreExpr])
773 -- Returns (Just (dc, [x1..xn])) if the argument expression is
774 -- a constructor application of the form (dc x1 .. xn)
776 exprIsConApp_maybe (Cast expr co)
777 = -- Maybe this is over the top, but here we try to turn
778 -- coerce (S,T) ( x, y )
780 -- ( coerce S x, coerce T y )
781 -- This happens in anger in PrelArrExts which has a coerce
782 -- case coerce memcpy a b of
784 -- where the memcpy is in the IO monad, but the call is in
786 case exprIsConApp_maybe expr of {
790 let (from_ty, to_ty) = coercionKind co in
792 case splitTyConApp_maybe to_ty of {
794 Just (tc, tc_arg_tys) | tc /= dataConTyCon dc -> Nothing
795 -- | not (isVanillaDataCon dc) -> Nothing
797 -- Type constructor must match datacon
799 case splitTyConApp_maybe from_ty of {
801 Just (tc', tc_arg_tys') | tc /= tc' -> Nothing
802 -- Both sides of coercion must have the same type constructor
806 -- here we do the PushC reduction rule as described in the FC paper
807 arity = tyConArity tc
808 n_ex_tvs = length dc_ex_tyvars
810 (univ_args, rest) = splitAt arity args
811 (ex_args, val_args) = splitAt n_ex_tvs rest
813 arg_tys = dataConRepArgTys dc
814 dc_tyvars = dataConUnivTyVars dc
815 dc_ex_tyvars = dataConExTyVars dc
817 deep arg_ty = deepCast arg_ty dc_tyvars co
819 -- first we appropriately cast the value arguments
820 arg_cos = map deep arg_tys
821 new_val_args = zipWith mkCoerce (map deep arg_tys) val_args
823 -- then we cast the existential coercion arguments
824 orig_tvs = dc_tyvars ++ dc_ex_tyvars
825 gammas = decomposeCo arity co
826 new_tys = gammas ++ (map (\ (Type t) -> t) ex_args)
827 theta = substTyWith orig_tvs new_tys
830 , (ty1, ty2) <- splitCoercionKind (tyVarKind tv)
831 = Type $ mkTransCoercion (mkSymCoercion (theta ty1))
832 (mkTransCoercion ty (theta ty2))
835 new_ex_args = zipWith cast_ty dc_ex_tyvars ex_args
838 ASSERT( all isTypeArg (take arity args) )
839 ASSERT( equalLength val_args arg_tys )
840 Just (dc, map Type tc_arg_tys ++ new_ex_args ++ new_val_args)
843 exprIsConApp_maybe (Note _ expr)
844 = exprIsConApp_maybe expr
845 -- We ignore InlineMe notes in case we have
846 -- x = __inline_me__ (a,b)
847 -- All part of making sure that INLINE pragmas never hurt
848 -- Marcin tripped on this one when making dictionaries more inlinable
850 -- In fact, we ignore all notes. For example,
851 -- case _scc_ "foo" (C a b) of
853 -- should be optimised away, but it will be only if we look
854 -- through the SCC note.
856 exprIsConApp_maybe expr = analyse (collectArgs expr)
858 analyse (Var fun, args)
859 | Just con <- isDataConWorkId_maybe fun,
860 args `lengthAtLeast` dataConRepArity con
861 -- Might be > because the arity excludes type args
864 -- Look through unfoldings, but only cheap ones, because
865 -- we are effectively duplicating the unfolding
866 analyse (Var fun, [])
867 | let unf = idUnfolding fun,
869 = exprIsConApp_maybe (unfoldingTemplate unf)
871 analyse other = Nothing
876 %************************************************************************
878 \subsection{Eta reduction and expansion}
880 %************************************************************************
883 exprEtaExpandArity :: DynFlags -> CoreExpr -> Arity
884 {- The Arity returned is the number of value args the
885 thing can be applied to without doing much work
887 exprEtaExpandArity is used when eta expanding
890 It returns 1 (or more) to:
891 case x of p -> \s -> ...
892 because for I/O ish things we really want to get that \s to the top.
893 We are prepared to evaluate x each time round the loop in order to get that
895 It's all a bit more subtle than it looks:
899 Consider one-shot lambdas
900 let x = expensive in \y z -> E
901 We want this to have arity 2 if the \y-abstraction is a 1-shot lambda
902 Hence the ArityType returned by arityType
904 2. The state-transformer hack
906 The one-shot lambda special cause is particularly important/useful for
907 IO state transformers, where we often get
908 let x = E in \ s -> ...
910 and the \s is a real-world state token abstraction. Such abstractions
911 are almost invariably 1-shot, so we want to pull the \s out, past the
912 let x=E, even if E is expensive. So we treat state-token lambdas as
913 one-shot even if they aren't really. The hack is in Id.isOneShotBndr.
915 3. Dealing with bottom
918 f = \x -> error "foo"
919 Here, arity 1 is fine. But if it is
923 then we want to get arity 2. Tecnically, this isn't quite right, because
925 should diverge, but it'll converge if we eta-expand f. Nevertheless, we
926 do so; it improves some programs significantly, and increasing convergence
927 isn't a bad thing. Hence the ABot/ATop in ArityType.
929 Actually, the situation is worse. Consider
933 Can we eta-expand here? At first the answer looks like "yes of course", but
936 This should diverge! But if we eta-expand, it won't. Again, we ignore this
937 "problem", because being scrupulous would lose an important transformation for
943 Non-recursive newtypes are transparent, and should not get in the way.
944 We do (currently) eta-expand recursive newtypes too. So if we have, say
946 newtype T = MkT ([T] -> Int)
950 where f has arity 1. Then: etaExpandArity e = 1;
951 that is, etaExpandArity looks through the coerce.
953 When we eta-expand e to arity 1: eta_expand 1 e T
954 we want to get: coerce T (\x::[T] -> (coerce ([T]->Int) e) x)
956 HOWEVER, note that if you use coerce bogusly you can ge
958 And since negate has arity 2, you might try to eta expand. But you can't
959 decopose Int to a function type. Hence the final case in eta_expand.
963 exprEtaExpandArity dflags e = arityDepth (arityType dflags e)
965 -- A limited sort of function type
966 data ArityType = AFun Bool ArityType -- True <=> one-shot
967 | ATop -- Know nothing
970 arityDepth :: ArityType -> Arity
971 arityDepth (AFun _ ty) = 1 + arityDepth ty
974 andArityType ABot at2 = at2
975 andArityType ATop at2 = ATop
976 andArityType (AFun t1 at1) (AFun t2 at2) = AFun (t1 && t2) (andArityType at1 at2)
977 andArityType at1 at2 = andArityType at2 at1
979 arityType :: DynFlags -> CoreExpr -> ArityType
980 -- (go1 e) = [b1,..,bn]
981 -- means expression can be rewritten \x_b1 -> ... \x_bn -> body
982 -- where bi is True <=> the lambda is one-shot
984 arityType dflags (Note n e) = arityType dflags e
985 -- Not needed any more: etaExpand is cleverer
986 -- | ok_note n = arityType dflags e
987 -- | otherwise = ATop
989 arityType dflags (Cast e co) = arityType dflags e
991 arityType dflags (Var v)
992 = mk (idArity v) (arg_tys (idType v))
994 mk :: Arity -> [Type] -> ArityType
995 -- The argument types are only to steer the "state hack"
996 -- Consider case x of
998 -- False -> \(s:RealWorld) -> e
999 -- where foo has arity 1. Then we want the state hack to
1000 -- apply to foo too, so we can eta expand the case.
1001 mk 0 tys | isBottomingId v = ABot
1002 | (ty:tys) <- tys, isStateHackType ty = AFun True ATop
1004 mk n (ty:tys) = AFun (isStateHackType ty) (mk (n-1) tys)
1005 mk n [] = AFun False (mk (n-1) [])
1007 arg_tys :: Type -> [Type] -- Ignore for-alls
1009 | Just (_, ty') <- splitForAllTy_maybe ty = arg_tys ty'
1010 | Just (arg,res) <- splitFunTy_maybe ty = arg : arg_tys res
1013 -- Lambdas; increase arity
1014 arityType dflags (Lam x e)
1015 | isId x = AFun (isOneShotBndr x) (arityType dflags e)
1016 | otherwise = arityType dflags e
1018 -- Applications; decrease arity
1019 arityType dflags (App f (Type _)) = arityType dflags f
1020 arityType dflags (App f a) = case arityType dflags f of
1021 AFun one_shot xs | exprIsCheap a -> xs
1024 -- Case/Let; keep arity if either the expression is cheap
1025 -- or it's a 1-shot lambda
1026 -- The former is not really right for Haskell
1027 -- f x = case x of { (a,b) -> \y. e }
1029 -- f x y = case x of { (a,b) -> e }
1030 -- The difference is observable using 'seq'
1031 arityType dflags (Case scrut _ _ alts)
1032 = case foldr1 andArityType [arityType dflags rhs | (_,_,rhs) <- alts] of
1033 xs | exprIsCheap scrut -> xs
1034 xs@(AFun one_shot _) | one_shot -> AFun True ATop
1037 arityType dflags (Let b e)
1038 = case arityType dflags e of
1039 xs | cheap_bind b -> xs
1040 xs@(AFun one_shot _) | one_shot -> AFun True ATop
1043 cheap_bind (NonRec b e) = is_cheap (b,e)
1044 cheap_bind (Rec prs) = all is_cheap prs
1045 is_cheap (b,e) = (dopt Opt_DictsCheap dflags && isDictId b)
1047 -- If the experimental -fdicts-cheap flag is on, we eta-expand through
1048 -- dictionary bindings. This improves arities. Thereby, it also
1049 -- means that full laziness is less prone to floating out the
1050 -- application of a function to its dictionary arguments, which
1051 -- can thereby lose opportunities for fusion. Example:
1052 -- foo :: Ord a => a -> ...
1053 -- foo = /\a \(d:Ord a). let d' = ...d... in \(x:a). ....
1054 -- -- So foo has arity 1
1056 -- f = \x. foo dInt $ bar x
1058 -- The (foo DInt) is floated out, and makes ineffective a RULE
1059 -- foo (bar x) = ...
1061 -- One could go further and make exprIsCheap reply True to any
1062 -- dictionary-typed expression, but that's more work.
1064 arityType dflags other = ATop
1066 {- NOT NEEDED ANY MORE: etaExpand is cleverer
1067 ok_note InlineMe = False
1068 ok_note other = True
1069 -- Notice that we do not look through __inline_me__
1070 -- This may seem surprising, but consider
1071 -- f = _inline_me (\x -> e)
1072 -- We DO NOT want to eta expand this to
1073 -- f = \x -> (_inline_me (\x -> e)) x
1074 -- because the _inline_me gets dropped now it is applied,
1083 etaExpand :: Arity -- Result should have this number of value args
1085 -> CoreExpr -> Type -- Expression and its type
1087 -- (etaExpand n us e ty) returns an expression with
1088 -- the same meaning as 'e', but with arity 'n'.
1090 -- Given e' = etaExpand n us e ty
1092 -- ty = exprType e = exprType e'
1094 -- Note that SCCs are not treated specially. If we have
1095 -- etaExpand 2 (\x -> scc "foo" e)
1096 -- = (\xy -> (scc "foo" e) y)
1097 -- So the costs of evaluating 'e' (not 'e y') are attributed to "foo"
1099 etaExpand n us expr ty
1100 | manifestArity expr >= n = expr -- The no-op case
1102 = eta_expand n us expr ty
1105 -- manifestArity sees how many leading value lambdas there are
1106 manifestArity :: CoreExpr -> Arity
1107 manifestArity (Lam v e) | isId v = 1 + manifestArity e
1108 | otherwise = manifestArity e
1109 manifestArity (Note _ e) = manifestArity e
1110 manifestArity (Cast e _) = manifestArity e
1113 -- etaExpand deals with for-alls. For example:
1115 -- where E :: forall a. a -> a
1117 -- (/\b. \y::a -> E b y)
1119 -- It deals with coerces too, though they are now rare
1120 -- so perhaps the extra code isn't worth it
1122 eta_expand n us expr ty
1124 -- The ILX code generator requires eta expansion for type arguments
1125 -- too, but alas the 'n' doesn't tell us how many of them there
1126 -- may be. So we eagerly eta expand any big lambdas, and just
1127 -- cross our fingers about possible loss of sharing in the ILX case.
1128 -- The Right Thing is probably to make 'arity' include
1129 -- type variables throughout the compiler. (ToDo.)
1131 -- Saturated, so nothing to do
1134 -- Short cut for the case where there already
1135 -- is a lambda; no point in gratuitously adding more
1136 eta_expand n us (Lam v body) ty
1138 = Lam v (eta_expand n us body (applyTy ty (mkTyVarTy v)))
1141 = Lam v (eta_expand (n-1) us body (funResultTy ty))
1143 -- We used to have a special case that stepped inside Coerces here,
1144 -- thus: eta_expand n us (Note note@(Coerce _ ty) e) _
1145 -- = Note note (eta_expand n us e ty)
1146 -- BUT this led to an infinite loop
1147 -- Example: newtype T = MkT (Int -> Int)
1148 -- eta_expand 1 (coerce (Int->Int) e)
1149 -- --> coerce (Int->Int) (eta_expand 1 T e)
1151 -- --> coerce (Int->Int) (coerce T
1152 -- (\x::Int -> eta_expand 1 (coerce (Int->Int) e)))
1153 -- by the splitNewType_maybe case below
1156 eta_expand n us expr ty
1157 = ASSERT2 (exprType expr `coreEqType` ty, ppr (exprType expr) $$ ppr ty)
1158 case splitForAllTy_maybe ty of {
1161 Lam lam_tv (eta_expand n us2 (App expr (Type (mkTyVarTy lam_tv))) (substTyWith [tv] [mkTyVarTy lam_tv] ty'))
1163 lam_tv = mkTyVar (mkSysTvName uniq FSLIT("etaT")) (tyVarKind tv)
1168 case splitFunTy_maybe ty of {
1169 Just (arg_ty, res_ty) -> Lam arg1 (eta_expand (n-1) us2 (App expr (Var arg1)) res_ty)
1171 arg1 = mkSysLocal FSLIT("eta") uniq arg_ty
1177 -- newtype T = MkT ([T] -> Int)
1178 -- Consider eta-expanding this
1181 -- coerce T (\x::[T] -> (coerce ([T]->Int) e) x)
1183 case splitNewTypeRepCo_maybe ty of {
1185 mkCoerce co (eta_expand n us (mkCoerce (mkSymCoercion co) expr) ty1) ;
1188 -- We have an expression of arity > 0, but its type isn't a function
1189 -- This *can* legitmately happen: e.g. coerce Int (\x. x)
1190 -- Essentially the programmer is playing fast and loose with types
1191 -- (Happy does this a lot). So we simply decline to eta-expand.
1196 exprArity is a cheap-and-cheerful version of exprEtaExpandArity.
1197 It tells how many things the expression can be applied to before doing
1198 any work. It doesn't look inside cases, lets, etc. The idea is that
1199 exprEtaExpandArity will do the hard work, leaving something that's easy
1200 for exprArity to grapple with. In particular, Simplify uses exprArity to
1201 compute the ArityInfo for the Id.
1203 Originally I thought that it was enough just to look for top-level lambdas, but
1204 it isn't. I've seen this
1206 foo = PrelBase.timesInt
1208 We want foo to get arity 2 even though the eta-expander will leave it
1209 unchanged, in the expectation that it'll be inlined. But occasionally it
1210 isn't, because foo is blacklisted (used in a rule).
1212 Similarly, see the ok_note check in exprEtaExpandArity. So
1213 f = __inline_me (\x -> e)
1214 won't be eta-expanded.
1216 And in any case it seems more robust to have exprArity be a bit more intelligent.
1217 But note that (\x y z -> f x y z)
1218 should have arity 3, regardless of f's arity.
1221 exprArity :: CoreExpr -> Arity
1224 go (Var v) = idArity v
1225 go (Lam x e) | isId x = go e + 1
1227 go (Note n e) = go e
1228 go (Cast e _) = go e
1229 go (App e (Type t)) = go e
1230 go (App f a) | exprIsCheap a = (go f - 1) `max` 0
1231 -- NB: exprIsCheap a!
1232 -- f (fac x) does not have arity 2,
1233 -- even if f has arity 3!
1234 -- NB: `max 0`! (\x y -> f x) has arity 2, even if f is
1235 -- unknown, hence arity 0
1239 %************************************************************************
1241 \subsection{Equality}
1243 %************************************************************************
1245 @cheapEqExpr@ is a cheap equality test which bales out fast!
1246 True => definitely equal
1247 False => may or may not be equal
1250 cheapEqExpr :: Expr b -> Expr b -> Bool
1252 cheapEqExpr (Var v1) (Var v2) = v1==v2
1253 cheapEqExpr (Lit lit1) (Lit lit2) = lit1 == lit2
1254 cheapEqExpr (Type t1) (Type t2) = t1 `coreEqType` t2
1256 cheapEqExpr (App f1 a1) (App f2 a2)
1257 = f1 `cheapEqExpr` f2 && a1 `cheapEqExpr` a2
1259 cheapEqExpr _ _ = False
1261 exprIsBig :: Expr b -> Bool
1262 -- Returns True of expressions that are too big to be compared by cheapEqExpr
1263 exprIsBig (Lit _) = False
1264 exprIsBig (Var v) = False
1265 exprIsBig (Type t) = False
1266 exprIsBig (App f a) = exprIsBig f || exprIsBig a
1267 exprIsBig other = True
1272 tcEqExpr :: CoreExpr -> CoreExpr -> Bool
1273 -- Used in rule matching, so does *not* look through
1274 -- newtypes, predicate types; hence tcEqExpr
1276 tcEqExpr e1 e2 = tcEqExprX rn_env e1 e2
1278 rn_env = mkRnEnv2 (mkInScopeSet (exprFreeVars e1 `unionVarSet` exprFreeVars e2))
1280 tcEqExprX :: RnEnv2 -> CoreExpr -> CoreExpr -> Bool
1281 tcEqExprX env (Var v1) (Var v2) = rnOccL env v1 == rnOccR env v2
1282 tcEqExprX env (Lit lit1) (Lit lit2) = lit1 == lit2
1283 tcEqExprX env (App f1 a1) (App f2 a2) = tcEqExprX env f1 f2 && tcEqExprX env a1 a2
1284 tcEqExprX env (Lam v1 e1) (Lam v2 e2) = tcEqExprX (rnBndr2 env v1 v2) e1 e2
1285 tcEqExprX env (Let (NonRec v1 r1) e1)
1286 (Let (NonRec v2 r2) e2) = tcEqExprX env r1 r2
1287 && tcEqExprX (rnBndr2 env v1 v2) e1 e2
1288 tcEqExprX env (Let (Rec ps1) e1)
1289 (Let (Rec ps2) e2) = equalLength ps1 ps2
1290 && and (zipWith eq_rhs ps1 ps2)
1291 && tcEqExprX env' e1 e2
1293 env' = foldl2 rn_bndr2 env ps2 ps2
1294 rn_bndr2 env (b1,_) (b2,_) = rnBndr2 env b1 b2
1295 eq_rhs (_,r1) (_,r2) = tcEqExprX env' r1 r2
1296 tcEqExprX env (Case e1 v1 t1 a1)
1297 (Case e2 v2 t2 a2) = tcEqExprX env e1 e2
1298 && tcEqTypeX env t1 t2
1299 && equalLength a1 a2
1300 && and (zipWith (eq_alt env') a1 a2)
1302 env' = rnBndr2 env v1 v2
1304 tcEqExprX env (Note n1 e1) (Note n2 e2) = eq_note env n1 n2 && tcEqExprX env e1 e2
1305 tcEqExprX env (Cast e1 co1) (Cast e2 co2) = tcEqTypeX env co1 co2 && tcEqExprX env e1 e2
1306 tcEqExprX env (Type t1) (Type t2) = tcEqTypeX env t1 t2
1307 tcEqExprX env e1 e2 = False
1309 eq_alt env (c1,vs1,r1) (c2,vs2,r2) = c1==c2 && tcEqExprX (rnBndrs2 env vs1 vs2) r1 r2
1311 eq_note env (SCC cc1) (SCC cc2) = cc1 == cc2
1312 eq_note env (CoreNote s1) (CoreNote s2) = s1 == s2
1313 eq_note env other1 other2 = False
1317 %************************************************************************
1319 \subsection{The size of an expression}
1321 %************************************************************************
1324 coreBindsSize :: [CoreBind] -> Int
1325 coreBindsSize bs = foldr ((+) . bindSize) 0 bs
1327 exprSize :: CoreExpr -> Int
1328 -- A measure of the size of the expressions
1329 -- It also forces the expression pretty drastically as a side effect
1330 exprSize (Var v) = v `seq` 1
1331 exprSize (Lit lit) = lit `seq` 1
1332 exprSize (App f a) = exprSize f + exprSize a
1333 exprSize (Lam b e) = varSize b + exprSize e
1334 exprSize (Let b e) = bindSize b + exprSize e
1335 exprSize (Case e b t as) = seqType t `seq` exprSize e + varSize b + 1 + foldr ((+) . altSize) 0 as
1336 exprSize (Cast e co) = (seqType co `seq` 1) + exprSize e
1337 exprSize (Note n e) = noteSize n + exprSize e
1338 exprSize (Type t) = seqType t `seq` 1
1340 noteSize (SCC cc) = cc `seq` 1
1341 noteSize InlineMe = 1
1342 noteSize (CoreNote s) = s `seq` 1 -- hdaume: core annotations
1344 varSize :: Var -> Int
1345 varSize b | isTyVar b = 1
1346 | otherwise = seqType (idType b) `seq`
1347 megaSeqIdInfo (idInfo b) `seq`
1350 varsSize = foldr ((+) . varSize) 0
1352 bindSize (NonRec b e) = varSize b + exprSize e
1353 bindSize (Rec prs) = foldr ((+) . pairSize) 0 prs
1355 pairSize (b,e) = varSize b + exprSize e
1357 altSize (c,bs,e) = c `seq` varsSize bs + exprSize e
1361 %************************************************************************
1363 \subsection{Hashing}
1365 %************************************************************************
1368 hashExpr :: CoreExpr -> Int
1369 -- Two expressions that hash to the same Int may be equal (but may not be)
1370 -- Two expressions that hash to the different Ints are definitely unequal
1372 -- But "unequal" here means "not identical"; two alpha-equivalent
1373 -- expressions may hash to the different Ints
1375 -- The emphasis is on a crude, fast hash, rather than on high precision
1377 hashExpr e | hash < 0 = 77 -- Just in case we hit -maxInt
1380 hash = abs (hash_expr e) -- Negative numbers kill UniqFM
1382 hash_expr (Note _ e) = hash_expr e
1383 hash_expr (Cast e co) = hash_expr e
1384 hash_expr (Let (NonRec b r) e) = hashId b
1385 hash_expr (Let (Rec ((b,r):_)) e) = hashId b
1386 hash_expr (Case _ b _ _) = hashId b
1387 hash_expr (App f e) = hash_expr f * fast_hash_expr e
1388 hash_expr (Var v) = hashId v
1389 hash_expr (Lit lit) = hashLiteral lit
1390 hash_expr (Lam b _) = hashId b
1391 hash_expr (Type t) = trace "hash_expr: type" 1 -- Shouldn't happen
1393 fast_hash_expr (Var v) = hashId v
1394 fast_hash_expr (Lit lit) = hashLiteral lit
1395 fast_hash_expr (App f (Type _)) = fast_hash_expr f
1396 fast_hash_expr (App f a) = fast_hash_expr a
1397 fast_hash_expr (Lam b _) = hashId b
1398 fast_hash_expr other = 1
1401 hashId id = hashName (idName id)
1404 %************************************************************************
1406 \subsection{Determining non-updatable right-hand-sides}
1408 %************************************************************************
1410 Top-level constructor applications can usually be allocated
1411 statically, but they can't if the constructor, or any of the
1412 arguments, come from another DLL (because we can't refer to static
1413 labels in other DLLs).
1415 If this happens we simply make the RHS into an updatable thunk,
1416 and 'exectute' it rather than allocating it statically.
1419 rhsIsStatic :: PackageId -> CoreExpr -> Bool
1420 -- This function is called only on *top-level* right-hand sides
1421 -- Returns True if the RHS can be allocated statically, with
1422 -- no thunks involved at all.
1424 -- It's called (i) in TidyPgm.hasCafRefs to decide if the rhs is, or
1425 -- refers to, CAFs; and (ii) in CoreToStg to decide whether to put an
1426 -- update flag on it.
1428 -- The basic idea is that rhsIsStatic returns True only if the RHS is
1429 -- (a) a value lambda
1430 -- (b) a saturated constructor application with static args
1432 -- BUT watch out for
1433 -- (i) Any cross-DLL references kill static-ness completely
1434 -- because they must be 'executed' not statically allocated
1435 -- ("DLL" here really only refers to Windows DLLs, on other platforms,
1436 -- this is not necessary)
1438 -- (ii) We treat partial applications as redexes, because in fact we
1439 -- make a thunk for them that runs and builds a PAP
1440 -- at run-time. The only appliations that are treated as
1441 -- static are *saturated* applications of constructors.
1443 -- We used to try to be clever with nested structures like this:
1444 -- ys = (:) w ((:) w [])
1445 -- on the grounds that CorePrep will flatten ANF-ise it later.
1446 -- But supporting this special case made the function much more
1447 -- complicated, because the special case only applies if there are no
1448 -- enclosing type lambdas:
1449 -- ys = /\ a -> Foo (Baz ([] a))
1450 -- Here the nested (Baz []) won't float out to top level in CorePrep.
1452 -- But in fact, even without -O, nested structures at top level are
1453 -- flattened by the simplifier, so we don't need to be super-clever here.
1457 -- f = \x::Int. x+7 TRUE
1458 -- p = (True,False) TRUE
1460 -- d = (fst p, False) FALSE because there's a redex inside
1461 -- (this particular one doesn't happen but...)
1463 -- h = D# (1.0## /## 2.0##) FALSE (redex again)
1464 -- n = /\a. Nil a TRUE
1466 -- t = /\a. (:) (case w a of ...) (Nil a) FALSE (redex)
1469 -- This is a bit like CoreUtils.exprIsHNF, with the following differences:
1470 -- a) scc "foo" (\x -> ...) is updatable (so we catch the right SCC)
1472 -- b) (C x xs), where C is a contructors is updatable if the application is
1475 -- c) don't look through unfolding of f in (f x).
1477 -- When opt_RuntimeTypes is on, we keep type lambdas and treat
1478 -- them as making the RHS re-entrant (non-updatable).
1480 rhsIsStatic this_pkg rhs = is_static False rhs
1482 is_static :: Bool -- True <=> in a constructor argument; must be atomic
1485 is_static False (Lam b e) = isRuntimeVar b || is_static False e
1487 is_static in_arg (Note (SCC _) e) = False
1488 is_static in_arg (Note _ e) = is_static in_arg e
1489 is_static in_arg (Cast e co) = is_static in_arg e
1491 is_static in_arg (Lit lit)
1493 MachLabel _ _ -> False
1495 -- A MachLabel (foreign import "&foo") in an argument
1496 -- prevents a constructor application from being static. The
1497 -- reason is that it might give rise to unresolvable symbols
1498 -- in the object file: under Linux, references to "weak"
1499 -- symbols from the data segment give rise to "unresolvable
1500 -- relocation" errors at link time This might be due to a bug
1501 -- in the linker, but we'll work around it here anyway.
1504 is_static in_arg other_expr = go other_expr 0
1506 go (Var f) n_val_args
1507 #if mingw32_TARGET_OS
1508 | not (isDllName this_pkg (idName f))
1510 = saturated_data_con f n_val_args
1511 || (in_arg && n_val_args == 0)
1512 -- A naked un-applied variable is *not* deemed a static RHS
1514 -- Reason: better to update so that the indirection gets shorted
1515 -- out, and the true value will be seen
1516 -- NB: if you change this, you'll break the invariant that THUNK_STATICs
1517 -- are always updatable. If you do so, make sure that non-updatable
1518 -- ones have enough space for their static link field!
1520 go (App f a) n_val_args
1521 | isTypeArg a = go f n_val_args
1522 | not in_arg && is_static True a = go f (n_val_args + 1)
1523 -- The (not in_arg) checks that we aren't in a constructor argument;
1524 -- if we are, we don't allow (value) applications of any sort
1526 -- NB. In case you wonder, args are sometimes not atomic. eg.
1527 -- x = D# (1.0## /## 2.0##)
1528 -- can't float because /## can fail.
1530 go (Note (SCC _) f) n_val_args = False
1531 go (Note _ f) n_val_args = go f n_val_args
1532 go (Cast e co) n_val_args = go e n_val_args
1534 go other n_val_args = False
1536 saturated_data_con f n_val_args
1537 = case isDataConWorkId_maybe f of
1538 Just dc -> n_val_args == dataConRepArity dc