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 dataConOrigInstPat, dataConRepInstPat, dataConRepFSInstPat
39 #include "HsVersions.h"
42 import GLAEXTS -- For `xori`
45 import CoreFVs ( exprFreeVars )
46 import PprCore ( pprCoreExpr )
47 import Var ( Var, TyVar, CoVar, tyVarKind, mkCoVar, mkTyVar )
48 import OccName ( mkVarOccFS )
49 import SrcLoc ( noSrcLoc )
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 dataConTyCon, dataConRepArgTys,
60 dataConUnivTyVars, dataConExTyVars, dataConEqSpec,
61 dataConOrigArgTys, dataConTheta )
62 import PrimOp ( PrimOp(..), primOpOkForSpeculation, primOpIsCheap )
63 import Id ( Id, idType, globalIdDetails, idNewStrictness,
64 mkWildId, idArity, idName, idUnfolding, idInfo,
65 isOneShotBndr, isStateHackType,
66 isDataConWorkId_maybe, mkSysLocal, mkUserLocal,
67 isDataConWorkId, isBottomingId, isDictId
69 import IdInfo ( GlobalIdDetails(..), megaSeqIdInfo )
70 import NewDemand ( appIsBottom )
71 import Type ( Type, mkFunTy, mkForAllTy, splitFunTy_maybe,
72 splitFunTy, tcEqTypeX,
73 applyTys, isUnLiftedType, seqType, mkTyVarTy,
74 splitForAllTy_maybe, isForAllTy,
75 splitTyConApp_maybe, splitTyConApp, coreEqType, funResultTy, applyTy,
76 substTyWith, mkPredTy, zipOpenTvSubst, substTy, substTyVar
78 import Coercion ( Coercion, mkTransCoercion, coercionKind,
79 splitNewTypeRepCo_maybe, mkSymCoercion,
80 decomposeCo, coercionKindPredTy )
81 import TyCon ( tyConArity )
82 import TysWiredIn ( boolTy, trueDataCon, falseDataCon )
83 import CostCentre ( CostCentre )
84 import BasicTypes ( Arity )
85 import PackageConfig ( PackageId )
86 import Unique ( Unique )
88 import DynFlags ( DynFlags, DynFlag(Opt_DictsCheap), dopt )
89 import TysPrim ( alphaTy ) -- Debugging only
90 import Util ( equalLength, lengthAtLeast, foldl2 )
91 import FastString ( FastString )
95 %************************************************************************
97 \subsection{Find the type of a Core atom/expression}
99 %************************************************************************
102 exprType :: CoreExpr -> Type
104 exprType (Var var) = idType var
105 exprType (Lit lit) = literalType lit
106 exprType (Let _ body) = exprType body
107 exprType (Case _ _ ty alts) = ty
109 = let (_, ty) = coercionKind co in ty
110 exprType (Note other_note e) = exprType e
111 exprType (Lam binder expr) = mkPiType binder (exprType expr)
113 = case collectArgs e of
114 (fun, args) -> applyTypeToArgs e (exprType fun) args
116 exprType other = pprTrace "exprType" (pprCoreExpr other) alphaTy
118 coreAltType :: CoreAlt -> Type
119 coreAltType (_,_,rhs) = exprType rhs
122 @mkPiType@ makes a (->) type or a forall type, depending on whether
123 it is given a type variable or a term variable. We cleverly use the
124 lbvarinfo field to figure out the right annotation for the arrove in
125 case of a term variable.
128 mkPiType :: Var -> Type -> Type -- The more polymorphic version
129 mkPiTypes :: [Var] -> Type -> Type -- doesn't work...
131 mkPiTypes vs ty = foldr mkPiType ty vs
134 | isId v = mkFunTy (idType v) ty
135 | otherwise = mkForAllTy v ty
139 applyTypeToArg :: Type -> CoreExpr -> Type
140 applyTypeToArg fun_ty (Type arg_ty) = applyTy fun_ty arg_ty
141 applyTypeToArg fun_ty other_arg = funResultTy fun_ty
143 applyTypeToArgs :: CoreExpr -> Type -> [CoreExpr] -> Type
144 -- A more efficient version of applyTypeToArg
145 -- when we have several args
146 -- The first argument is just for debugging
147 applyTypeToArgs e op_ty [] = op_ty
149 applyTypeToArgs e op_ty (Type ty : args)
150 = -- Accumulate type arguments so we can instantiate all at once
153 go rev_tys (Type ty : args) = go (ty:rev_tys) args
154 go rev_tys rest_args = applyTypeToArgs e op_ty' rest_args
156 op_ty' = applyTys op_ty (reverse rev_tys)
158 applyTypeToArgs e op_ty (other_arg : args)
159 = case (splitFunTy_maybe op_ty) of
160 Just (_, res_ty) -> applyTypeToArgs e res_ty args
161 Nothing -> pprPanic "applyTypeToArgs" (pprCoreExpr e $$ ppr op_ty)
166 %************************************************************************
168 \subsection{Attaching notes}
170 %************************************************************************
172 mkNote removes redundant coercions, and SCCs where possible
176 mkNote :: Note -> CoreExpr -> CoreExpr
177 mkNote (SCC cc) expr = mkSCC cc expr
178 mkNote InlineMe expr = mkInlineMe expr
179 mkNote note expr = Note note expr
183 Drop trivial InlineMe's. This is somewhat important, because if we have an unfolding
184 that looks like (Note InlineMe (Var v)), the InlineMe doesn't go away because it may
185 not be *applied* to anything.
187 We don't use exprIsTrivial here, though, because we sometimes generate worker/wrapper
190 f = inline_me (coerce t fw)
191 As usual, the inline_me prevents the worker from getting inlined back into the wrapper.
192 We want the split, so that the coerces can cancel at the call site.
194 However, we can get left with tiresome type applications. Notably, consider
195 f = /\ a -> let t = e in (t, w)
196 Then lifting the let out of the big lambda gives
198 f = /\ a -> let t = inline_me (t' a) in (t, w)
199 The inline_me is to stop the simplifier inlining t' right back
200 into t's RHS. In the next phase we'll substitute for t (since
201 its rhs is trivial) and *then* we could get rid of the inline_me.
202 But it hardly seems worth it, so I don't bother.
205 mkInlineMe (Var v) = Var v
206 mkInlineMe e = Note InlineMe e
212 mkCoerce :: Coercion -> CoreExpr -> CoreExpr
213 mkCoerce co (Cast expr co2)
214 = ASSERT(let { (from_ty, _to_ty) = coercionKind co;
215 (_from_ty2, to_ty2) = coercionKind co2} in
216 from_ty `coreEqType` to_ty2 )
217 mkCoerce (mkTransCoercion co2 co) expr
220 = let (from_ty, to_ty) = coercionKind co in
221 -- if to_ty `coreEqType` from_ty
224 ASSERT2(from_ty `coreEqType` (exprType expr), text "Trying to coerce" <+> text "(" <> ppr expr $$ text "::" <+> ppr (exprType expr) <> text ")" $$ ppr co $$ ppr (coercionKindPredTy co))
229 mkSCC :: CostCentre -> Expr b -> Expr b
230 -- Note: Nested SCC's *are* preserved for the benefit of
231 -- cost centre stack profiling
232 mkSCC cc (Lit lit) = Lit lit
233 mkSCC cc (Lam x e) = Lam x (mkSCC cc e) -- Move _scc_ inside lambda
234 mkSCC cc (Note (SCC cc') e) = Note (SCC cc) (Note (SCC cc') e)
235 mkSCC cc (Note n e) = Note n (mkSCC cc e) -- Move _scc_ inside notes
236 mkSCC cc (Cast e co) = Cast (mkSCC cc e) co -- Move _scc_ inside cast
237 mkSCC cc expr = Note (SCC cc) expr
241 %************************************************************************
243 \subsection{Other expression construction}
245 %************************************************************************
248 bindNonRec :: Id -> CoreExpr -> CoreExpr -> CoreExpr
249 -- (bindNonRec x r b) produces either
252 -- case r of x { _DEFAULT_ -> b }
254 -- depending on whether x is unlifted or not
255 -- It's used by the desugarer to avoid building bindings
256 -- that give Core Lint a heart attack. Actually the simplifier
257 -- deals with them perfectly well.
259 bindNonRec bndr rhs body
260 | needsCaseBinding (idType bndr) rhs = Case rhs bndr (exprType body) [(DEFAULT,[],body)]
261 | otherwise = Let (NonRec bndr rhs) body
263 needsCaseBinding ty rhs = isUnLiftedType ty && not (exprOkForSpeculation rhs)
264 -- Make a case expression instead of a let
265 -- These can arise either from the desugarer,
266 -- or from beta reductions: (\x.e) (x +# y)
270 mkAltExpr :: AltCon -> [CoreBndr] -> [Type] -> CoreExpr
271 -- This guy constructs the value that the scrutinee must have
272 -- when you are in one particular branch of a case
273 mkAltExpr (DataAlt con) args inst_tys
274 = mkConApp con (map Type inst_tys ++ varsToCoreExprs args)
275 mkAltExpr (LitAlt lit) [] []
278 mkIfThenElse :: CoreExpr -> CoreExpr -> CoreExpr -> CoreExpr
279 mkIfThenElse guard then_expr else_expr
280 -- Not going to be refining, so okay to take the type of the "then" clause
281 = Case guard (mkWildId boolTy) (exprType then_expr)
282 [ (DataAlt falseDataCon, [], else_expr), -- Increasing order of tag!
283 (DataAlt trueDataCon, [], then_expr) ]
287 %************************************************************************
289 \subsection{Taking expressions apart}
291 %************************************************************************
293 The default alternative must be first, if it exists at all.
294 This makes it easy to find, though it makes matching marginally harder.
297 findDefault :: [CoreAlt] -> ([CoreAlt], Maybe CoreExpr)
298 findDefault ((DEFAULT,args,rhs) : alts) = ASSERT( null args ) (alts, Just rhs)
299 findDefault alts = (alts, Nothing)
301 findAlt :: AltCon -> [CoreAlt] -> CoreAlt
304 (deflt@(DEFAULT,_,_):alts) -> go alts deflt
305 other -> go alts panic_deflt
307 panic_deflt = pprPanic "Missing alternative" (ppr con $$ vcat (map ppr alts))
310 go (alt@(con1,_,_) : alts) deflt
311 = case con `cmpAltCon` con1 of
312 LT -> deflt -- Missed it already; the alts are in increasing order
314 GT -> ASSERT( not (con1 == DEFAULT) ) go alts deflt
316 isDefaultAlt :: CoreAlt -> Bool
317 isDefaultAlt (DEFAULT, _, _) = True
318 isDefaultAlt other = False
320 ---------------------------------
321 mergeAlts :: [CoreAlt] -> [CoreAlt] -> [CoreAlt]
322 -- Merge preserving order; alternatives in the first arg
323 -- shadow ones in the second
324 mergeAlts [] as2 = as2
325 mergeAlts as1 [] = as1
326 mergeAlts (a1:as1) (a2:as2)
327 = case a1 `cmpAlt` a2 of
328 LT -> a1 : mergeAlts as1 (a2:as2)
329 EQ -> a1 : mergeAlts as1 as2 -- Discard a2
330 GT -> a2 : mergeAlts (a1:as1) as2
334 %************************************************************************
336 \subsection{Figuring out things about expressions}
338 %************************************************************************
340 @exprIsTrivial@ is true of expressions we are unconditionally happy to
341 duplicate; simple variables and constants, and type
342 applications. Note that primop Ids aren't considered
345 @exprIsBottom@ is true of expressions that are guaranteed to diverge
348 There used to be a gruesome test for (hasNoBinding v) in the
350 exprIsTrivial (Var v) | hasNoBinding v = idArity v == 0
351 The idea here is that a constructor worker, like $wJust, is
352 really short for (\x -> $wJust x), becuase $wJust has no binding.
353 So it should be treated like a lambda. Ditto unsaturated primops.
354 But now constructor workers are not "have-no-binding" Ids. And
355 completely un-applied primops and foreign-call Ids are sufficiently
356 rare that I plan to allow them to be duplicated and put up with
359 SCC notes. We do not treat (_scc_ "foo" x) as trivial, because
360 a) it really generates code, (and a heap object when it's
361 a function arg) to capture the cost centre
362 b) see the note [SCC-and-exprIsTrivial] in Simplify.simplLazyBind
365 exprIsTrivial (Var v) = True -- See notes above
366 exprIsTrivial (Type _) = True
367 exprIsTrivial (Lit lit) = litIsTrivial lit
368 exprIsTrivial (App e arg) = not (isRuntimeArg arg) && exprIsTrivial e
369 exprIsTrivial (Note (SCC _) e) = False -- See notes above
370 exprIsTrivial (Note _ e) = exprIsTrivial e
371 exprIsTrivial (Cast e co) = exprIsTrivial e
372 exprIsTrivial (Lam b body) = not (isRuntimeVar b) && exprIsTrivial body
373 exprIsTrivial other = False
377 @exprIsDupable@ is true of expressions that can be duplicated at a modest
378 cost in code size. This will only happen in different case
379 branches, so there's no issue about duplicating work.
381 That is, exprIsDupable returns True of (f x) even if
382 f is very very expensive to call.
384 Its only purpose is to avoid fruitless let-binding
385 and then inlining of case join points
389 exprIsDupable (Type _) = True
390 exprIsDupable (Var v) = True
391 exprIsDupable (Lit lit) = litIsDupable lit
392 exprIsDupable (Note InlineMe e) = True
393 exprIsDupable (Note _ e) = exprIsDupable e
394 exprIsDupable (Cast e co) = exprIsDupable e
398 go (Var v) n_args = True
399 go (App f a) n_args = n_args < dupAppSize
402 go other n_args = False
405 dupAppSize = 4 -- Size of application we are prepared to duplicate
408 @exprIsCheap@ looks at a Core expression and returns \tr{True} if
409 it is obviously in weak head normal form, or is cheap to get to WHNF.
410 [Note that that's not the same as exprIsDupable; an expression might be
411 big, and hence not dupable, but still cheap.]
413 By ``cheap'' we mean a computation we're willing to:
414 push inside a lambda, or
415 inline at more than one place
416 That might mean it gets evaluated more than once, instead of being
417 shared. The main examples of things which aren't WHNF but are
422 (where e, and all the ei are cheap)
425 (where e and b are cheap)
428 (where op is a cheap primitive operator)
431 (because we are happy to substitute it inside a lambda)
433 Notice that a variable is considered 'cheap': we can push it inside a lambda,
434 because sharing will make sure it is only evaluated once.
437 exprIsCheap :: CoreExpr -> Bool
438 exprIsCheap (Lit lit) = True
439 exprIsCheap (Type _) = True
440 exprIsCheap (Var _) = True
441 exprIsCheap (Note InlineMe e) = True
442 exprIsCheap (Note _ e) = exprIsCheap e
443 exprIsCheap (Cast e co) = exprIsCheap e
444 exprIsCheap (Lam x e) = isRuntimeVar x || exprIsCheap e
445 exprIsCheap (Case e _ _ alts) = exprIsCheap e &&
446 and [exprIsCheap rhs | (_,_,rhs) <- alts]
447 -- Experimentally, treat (case x of ...) as cheap
448 -- (and case __coerce x etc.)
449 -- This improves arities of overloaded functions where
450 -- there is only dictionary selection (no construction) involved
451 exprIsCheap (Let (NonRec x _) e)
452 | isUnLiftedType (idType x) = exprIsCheap e
454 -- strict lets always have cheap right hand sides,
455 -- and do no allocation.
457 exprIsCheap other_expr -- Applications and variables
460 -- Accumulate value arguments, then decide
461 go (App f a) val_args | isRuntimeArg a = go f (a:val_args)
462 | otherwise = go f val_args
464 go (Var f) [] = True -- Just a type application of a variable
465 -- (f t1 t2 t3) counts as WHNF
467 = case globalIdDetails f of
468 RecordSelId {} -> go_sel args
469 ClassOpId _ -> go_sel args
470 PrimOpId op -> go_primop op args
472 DataConWorkId _ -> go_pap args
473 other | length args < idArity f -> go_pap args
475 other -> isBottomingId f
476 -- Application of a function which
477 -- always gives bottom; we treat this as cheap
478 -- because it certainly doesn't need to be shared!
480 go other args = False
483 go_pap args = all exprIsTrivial args
484 -- For constructor applications and primops, check that all
485 -- the args are trivial. We don't want to treat as cheap, say,
487 -- We'll put up with one constructor application, but not dozens
490 go_primop op args = primOpIsCheap op && all exprIsCheap args
491 -- In principle we should worry about primops
492 -- that return a type variable, since the result
493 -- might be applied to something, but I'm not going
494 -- to bother to check the number of args
497 go_sel [arg] = exprIsCheap arg -- I'm experimenting with making record selection
498 go_sel other = False -- look cheap, so we will substitute it inside a
499 -- lambda. Particularly for dictionary field selection.
500 -- BUT: Take care with (sel d x)! The (sel d) might be cheap, but
501 -- there's no guarantee that (sel d x) will be too. Hence (n_val_args == 1)
504 exprOkForSpeculation returns True of an expression that it is
506 * safe to evaluate even if normal order eval might not
507 evaluate the expression at all, or
509 * safe *not* to evaluate even if normal order would do so
513 the expression guarantees to terminate,
515 without raising an exception,
516 without causing a side effect (e.g. writing a mutable variable)
518 NB: if exprIsHNF e, then exprOkForSpecuation e
521 let x = case y# +# 1# of { r# -> I# r# }
524 case y# +# 1# of { r# ->
529 We can only do this if the (y+1) is ok for speculation: it has no
530 side effects, and can't diverge or raise an exception.
533 exprOkForSpeculation :: CoreExpr -> Bool
534 exprOkForSpeculation (Lit _) = True
535 exprOkForSpeculation (Type _) = True
536 exprOkForSpeculation (Var v) = isUnLiftedType (idType v)
537 exprOkForSpeculation (Note _ e) = exprOkForSpeculation e
538 exprOkForSpeculation (Cast e co) = exprOkForSpeculation e
539 exprOkForSpeculation other_expr
540 = case collectArgs other_expr of
541 (Var f, args) -> spec_ok (globalIdDetails f) args
545 spec_ok (DataConWorkId _) args
546 = True -- The strictness of the constructor has already
547 -- been expressed by its "wrapper", so we don't need
548 -- to take the arguments into account
550 spec_ok (PrimOpId op) args
551 | isDivOp op, -- Special case for dividing operations that fail
552 [arg1, Lit lit] <- args -- only if the divisor is zero
553 = not (isZeroLit lit) && exprOkForSpeculation arg1
554 -- Often there is a literal divisor, and this
555 -- can get rid of a thunk in an inner looop
558 = primOpOkForSpeculation op &&
559 all exprOkForSpeculation args
560 -- A bit conservative: we don't really need
561 -- to care about lazy arguments, but this is easy
563 spec_ok other args = False
565 isDivOp :: PrimOp -> Bool
566 -- True of dyadic operators that can fail
567 -- only if the second arg is zero
568 -- This function probably belongs in PrimOp, or even in
569 -- an automagically generated file.. but it's such a
570 -- special case I thought I'd leave it here for now.
571 isDivOp IntQuotOp = True
572 isDivOp IntRemOp = True
573 isDivOp WordQuotOp = True
574 isDivOp WordRemOp = True
575 isDivOp IntegerQuotRemOp = True
576 isDivOp IntegerDivModOp = True
577 isDivOp FloatDivOp = True
578 isDivOp DoubleDivOp = True
579 isDivOp other = False
584 exprIsBottom :: CoreExpr -> Bool -- True => definitely bottom
585 exprIsBottom e = go 0 e
587 -- n is the number of args
588 go n (Note _ e) = go n e
589 go n (Cast e co) = go n e
590 go n (Let _ e) = go n e
591 go n (Case e _ _ _) = go 0 e -- Just check the scrut
592 go n (App e _) = go (n+1) e
593 go n (Var v) = idAppIsBottom v n
595 go n (Lam _ _) = False
596 go n (Type _) = False
598 idAppIsBottom :: Id -> Int -> Bool
599 idAppIsBottom id n_val_args = appIsBottom (idNewStrictness id) n_val_args
602 @exprIsHNF@ returns true for expressions that are certainly *already*
603 evaluated to *head* normal form. This is used to decide whether it's ok
606 case x of _ -> e ===> e
608 and to decide whether it's safe to discard a `seq`
610 So, it does *not* treat variables as evaluated, unless they say they are.
612 But it *does* treat partial applications and constructor applications
613 as values, even if their arguments are non-trivial, provided the argument
615 e.g. (:) (f x) (map f xs) is a value
616 map (...redex...) is a value
617 Because `seq` on such things completes immediately
619 For unlifted argument types, we have to be careful:
621 Suppose (f x) diverges; then C (f x) is not a value. True, but
622 this form is illegal (see the invariants in CoreSyn). Args of unboxed
623 type must be ok-for-speculation (or trivial).
626 exprIsHNF :: CoreExpr -> Bool -- True => Value-lambda, constructor, PAP
627 exprIsHNF (Var v) -- NB: There are no value args at this point
628 = isDataConWorkId v -- Catches nullary constructors,
629 -- so that [] and () are values, for example
630 || idArity v > 0 -- Catches (e.g.) primops that don't have unfoldings
631 || isEvaldUnfolding (idUnfolding v)
632 -- Check the thing's unfolding; it might be bound to a value
633 -- A worry: what if an Id's unfolding is just itself:
634 -- then we could get an infinite loop...
636 exprIsHNF (Lit l) = True
637 exprIsHNF (Type ty) = True -- Types are honorary Values;
638 -- we don't mind copying them
639 exprIsHNF (Lam b e) = isRuntimeVar b || exprIsHNF e
640 exprIsHNF (Note _ e) = exprIsHNF e
641 exprIsHNF (Cast e co) = exprIsHNF e
642 exprIsHNF (App e (Type _)) = exprIsHNF e
643 exprIsHNF (App e a) = app_is_value e [a]
644 exprIsHNF other = False
646 -- There is at least one value argument
647 app_is_value (Var fun) args
648 | isDataConWorkId fun -- Constructor apps are values
649 || idArity fun > valArgCount args -- Under-applied function
650 = check_args (idType fun) args
651 app_is_value (App f a) as = app_is_value f (a:as)
652 app_is_value other as = False
654 -- 'check_args' checks that unlifted-type args
655 -- are in fact guaranteed non-divergent
656 check_args fun_ty [] = True
657 check_args fun_ty (Type _ : args) = case splitForAllTy_maybe fun_ty of
658 Just (_, ty) -> check_args ty args
659 check_args fun_ty (arg : args)
660 | isUnLiftedType arg_ty = exprOkForSpeculation arg
661 | otherwise = check_args res_ty args
663 (arg_ty, res_ty) = splitFunTy fun_ty
667 -- These InstPat functions go here to avoid circularity between DataCon and Id
668 dataConRepInstPat = dataConInstPat dataConRepArgTys (repeat (FSLIT("ipv")))
669 dataConRepFSInstPat = dataConInstPat dataConRepArgTys
670 dataConOrigInstPat = dataConInstPat dc_arg_tys (repeat (FSLIT("ipv")))
672 dc_arg_tys dc = map mkPredTy (dataConTheta dc) ++ dataConOrigArgTys dc
673 -- Remember to include the existential dictionaries
675 dataConInstPat :: (DataCon -> [Type]) -- function used to find arg tys
676 -> [FastString] -- A long enough list of FSs to use for names
677 -> [Unique] -- An equally long list of uniques, at least one for each binder
679 -> [Type] -- Types to instantiate the universally quantified tyvars
680 -> ([TyVar], [CoVar], [Id]) -- Return instantiated variables
681 -- dataConInstPat arg_fun fss us con inst_tys returns a triple
682 -- (ex_tvs, co_tvs, arg_ids),
684 -- ex_tvs are intended to be used as binders for existential type args
686 -- co_tvs are intended to be used as binders for coercion args and the kinds
687 -- of these vars have been instantiated by the inst_tys and the ex_tys
689 -- arg_ids are indended to be used as binders for value arguments, including
690 -- dicts, and their types have been instantiated with inst_tys and ex_tys
693 -- The following constructor T1
696 -- T1 :: forall b. Int -> b -> T(a,b)
699 -- has representation type
700 -- forall a. forall a1. forall b. (a :=: (a1,b)) =>
703 -- dataConInstPat fss us T1 (a1',b') will return
705 -- ([a1'', b''], [c :: (a1', b'):=:(a1'', b'')], [x :: Int, y :: b''])
707 -- where the double-primed variables are created with the FastStrings and
708 -- Uniques given as fss and us
709 dataConInstPat arg_fun fss uniqs con inst_tys
710 = (ex_bndrs, co_bndrs, id_bndrs)
712 univ_tvs = dataConUnivTyVars con
713 ex_tvs = dataConExTyVars con
714 arg_tys = arg_fun con
715 eq_spec = dataConEqSpec con
716 eq_preds = eqSpecPreds eq_spec
719 n_co = length eq_spec
721 -- split the Uniques and FastStrings
722 (ex_uniqs, uniqs') = splitAt n_ex uniqs
723 (co_uniqs, id_uniqs) = splitAt n_co uniqs'
725 (ex_fss, fss') = splitAt n_ex fss
726 (co_fss, id_fss) = splitAt n_co fss'
728 -- Make existential type variables
729 ex_bndrs = zipWith3 mk_ex_var ex_uniqs ex_fss ex_tvs
730 mk_ex_var uniq fs var = mkTyVar new_name kind
732 new_name = mkSysTvName uniq fs
735 -- Make the instantiating substitution
736 subst = zipOpenTvSubst (univ_tvs ++ ex_tvs) (inst_tys ++ map mkTyVarTy ex_bndrs)
738 -- Make new coercion vars, instantiating kind
739 co_bndrs = zipWith3 mk_co_var co_uniqs co_fss eq_preds
740 mk_co_var uniq fs eq_pred = mkCoVar new_name co_kind
742 new_name = mkSysTvName uniq fs
743 co_kind = substTy subst (mkPredTy eq_pred)
745 -- make value vars, instantiating types
746 mk_id_var uniq fs ty = mkUserLocal (mkVarOccFS fs) uniq (substTy subst ty) noSrcLoc
747 id_bndrs = zipWith3 mk_id_var id_uniqs id_fss arg_tys
749 exprIsConApp_maybe :: CoreExpr -> Maybe (DataCon, [CoreExpr])
750 -- Returns (Just (dc, [x1..xn])) if the argument expression is
751 -- a constructor application of the form (dc x1 .. xn)
752 exprIsConApp_maybe (Cast expr co)
753 = -- Here we do the PushC reduction rule as described in the FC paper
754 case exprIsConApp_maybe expr of {
756 Just (dc, dc_args) ->
758 -- The transformation applies iff we have
759 -- (C e1 ... en) `cast` co
760 -- where co :: (T t1 .. tn) :=: (T s1 ..sn)
761 -- That is, with a T at the top of both sides
762 -- The left-hand one must be a T, because exprIsConApp returned True
763 -- but the right-hand one might not be. (Though it usually will.)
765 let (from_ty, to_ty) = coercionKind co
766 (from_tc, from_tc_arg_tys) = splitTyConApp from_ty
767 -- The inner one must be a TyConApp
769 case splitTyConApp_maybe to_ty of {
771 Just (to_tc, to_tc_arg_tys)
772 | from_tc /= to_tc -> Nothing
773 -- These two Nothing cases are possible; we might see
774 -- (C x y) `cast` (g :: T a ~ S [a]),
775 -- where S is a type function. In fact, exprIsConApp
776 -- will probably not be called in such circumstances,
777 -- but there't nothing wrong with it
781 tc_arity = tyConArity from_tc
783 (univ_args, rest1) = splitAt tc_arity dc_args
784 (ex_args, rest2) = splitAt n_ex_tvs rest1
785 (co_args, val_args) = splitAt n_cos rest2
787 arg_tys = dataConRepArgTys dc
788 dc_univ_tyvars = dataConUnivTyVars dc
789 dc_ex_tyvars = dataConExTyVars dc
790 dc_eq_spec = dataConEqSpec dc
791 dc_tyvars = dc_univ_tyvars ++ dc_ex_tyvars
792 n_ex_tvs = length dc_ex_tyvars
793 n_cos = length dc_eq_spec
795 -- Make the "theta" from Fig 3 of the paper
796 gammas = decomposeCo tc_arity co
797 new_tys = gammas ++ map (\ (Type t) -> t) ex_args
798 theta = zipOpenTvSubst dc_tyvars new_tys
800 -- First we cast the existential coercion arguments
801 cast_co (tv,ty) (Type co) = Type $ mkSymCoercion (substTyVar theta tv)
803 `mkTransCoercion` (substTy theta ty)
804 new_co_args = zipWith cast_co dc_eq_spec co_args
806 -- ...and now value arguments
807 new_val_args = zipWith cast_arg arg_tys val_args
808 cast_arg arg_ty arg = mkCoerce (substTy theta arg_ty) arg
811 ASSERT( length univ_args == tc_arity )
812 ASSERT( from_tc == dataConTyCon dc )
813 ASSERT( and (zipWith coreEqType [t | Type t <- univ_args] from_tc_arg_tys) )
814 ASSERT( all isTypeArg (univ_args ++ ex_args) )
815 ASSERT2( equalLength val_args arg_tys, ppr dc $$ ppr dc_tyvars $$ ppr dc_ex_tyvars $$ ppr arg_tys $$ ppr dc_args $$ ppr univ_args $$ ppr ex_args $$ ppr val_args $$ ppr arg_tys )
817 Just (dc, map Type to_tc_arg_tys ++ ex_args ++ new_co_args ++ new_val_args)
820 exprIsConApp_maybe (Note _ expr)
821 = exprIsConApp_maybe expr
822 -- We ignore InlineMe notes in case we have
823 -- x = __inline_me__ (a,b)
824 -- All part of making sure that INLINE pragmas never hurt
825 -- Marcin tripped on this one when making dictionaries more inlinable
827 -- In fact, we ignore all notes. For example,
828 -- case _scc_ "foo" (C a b) of
830 -- should be optimised away, but it will be only if we look
831 -- through the SCC note.
833 exprIsConApp_maybe expr = analyse (collectArgs expr)
835 analyse (Var fun, args)
836 | Just con <- isDataConWorkId_maybe fun,
837 args `lengthAtLeast` dataConRepArity con
838 -- Might be > because the arity excludes type args
841 -- Look through unfoldings, but only cheap ones, because
842 -- we are effectively duplicating the unfolding
843 analyse (Var fun, [])
844 | let unf = idUnfolding fun,
846 = exprIsConApp_maybe (unfoldingTemplate unf)
848 analyse other = Nothing
853 %************************************************************************
855 \subsection{Eta reduction and expansion}
857 %************************************************************************
860 exprEtaExpandArity :: DynFlags -> CoreExpr -> Arity
861 {- The Arity returned is the number of value args the
862 thing can be applied to without doing much work
864 exprEtaExpandArity is used when eta expanding
867 It returns 1 (or more) to:
868 case x of p -> \s -> ...
869 because for I/O ish things we really want to get that \s to the top.
870 We are prepared to evaluate x each time round the loop in order to get that
872 It's all a bit more subtle than it looks:
876 Consider one-shot lambdas
877 let x = expensive in \y z -> E
878 We want this to have arity 2 if the \y-abstraction is a 1-shot lambda
879 Hence the ArityType returned by arityType
881 2. The state-transformer hack
883 The one-shot lambda special cause is particularly important/useful for
884 IO state transformers, where we often get
885 let x = E in \ s -> ...
887 and the \s is a real-world state token abstraction. Such abstractions
888 are almost invariably 1-shot, so we want to pull the \s out, past the
889 let x=E, even if E is expensive. So we treat state-token lambdas as
890 one-shot even if they aren't really. The hack is in Id.isOneShotBndr.
892 3. Dealing with bottom
895 f = \x -> error "foo"
896 Here, arity 1 is fine. But if it is
900 then we want to get arity 2. Tecnically, this isn't quite right, because
902 should diverge, but it'll converge if we eta-expand f. Nevertheless, we
903 do so; it improves some programs significantly, and increasing convergence
904 isn't a bad thing. Hence the ABot/ATop in ArityType.
906 Actually, the situation is worse. Consider
910 Can we eta-expand here? At first the answer looks like "yes of course", but
913 This should diverge! But if we eta-expand, it won't. Again, we ignore this
914 "problem", because being scrupulous would lose an important transformation for
920 Non-recursive newtypes are transparent, and should not get in the way.
921 We do (currently) eta-expand recursive newtypes too. So if we have, say
923 newtype T = MkT ([T] -> Int)
927 where f has arity 1. Then: etaExpandArity e = 1;
928 that is, etaExpandArity looks through the coerce.
930 When we eta-expand e to arity 1: eta_expand 1 e T
931 we want to get: coerce T (\x::[T] -> (coerce ([T]->Int) e) x)
933 HOWEVER, note that if you use coerce bogusly you can ge
935 And since negate has arity 2, you might try to eta expand. But you can't
936 decopose Int to a function type. Hence the final case in eta_expand.
940 exprEtaExpandArity dflags e = arityDepth (arityType dflags e)
942 -- A limited sort of function type
943 data ArityType = AFun Bool ArityType -- True <=> one-shot
944 | ATop -- Know nothing
947 arityDepth :: ArityType -> Arity
948 arityDepth (AFun _ ty) = 1 + arityDepth ty
951 andArityType ABot at2 = at2
952 andArityType ATop at2 = ATop
953 andArityType (AFun t1 at1) (AFun t2 at2) = AFun (t1 && t2) (andArityType at1 at2)
954 andArityType at1 at2 = andArityType at2 at1
956 arityType :: DynFlags -> CoreExpr -> ArityType
957 -- (go1 e) = [b1,..,bn]
958 -- means expression can be rewritten \x_b1 -> ... \x_bn -> body
959 -- where bi is True <=> the lambda is one-shot
961 arityType dflags (Note n e) = arityType dflags e
962 -- Not needed any more: etaExpand is cleverer
963 -- | ok_note n = arityType dflags e
964 -- | otherwise = ATop
966 arityType dflags (Cast e co) = arityType dflags e
968 arityType dflags (Var v)
969 = mk (idArity v) (arg_tys (idType v))
971 mk :: Arity -> [Type] -> ArityType
972 -- The argument types are only to steer the "state hack"
973 -- Consider case x of
975 -- False -> \(s:RealWorld) -> e
976 -- where foo has arity 1. Then we want the state hack to
977 -- apply to foo too, so we can eta expand the case.
978 mk 0 tys | isBottomingId v = ABot
979 | (ty:tys) <- tys, isStateHackType ty = AFun True ATop
981 mk n (ty:tys) = AFun (isStateHackType ty) (mk (n-1) tys)
982 mk n [] = AFun False (mk (n-1) [])
984 arg_tys :: Type -> [Type] -- Ignore for-alls
986 | Just (_, ty') <- splitForAllTy_maybe ty = arg_tys ty'
987 | Just (arg,res) <- splitFunTy_maybe ty = arg : arg_tys res
990 -- Lambdas; increase arity
991 arityType dflags (Lam x e)
992 | isId x = AFun (isOneShotBndr x) (arityType dflags e)
993 | otherwise = arityType dflags e
995 -- Applications; decrease arity
996 arityType dflags (App f (Type _)) = arityType dflags f
997 arityType dflags (App f a) = case arityType dflags f of
998 AFun one_shot xs | exprIsCheap a -> xs
1001 -- Case/Let; keep arity if either the expression is cheap
1002 -- or it's a 1-shot lambda
1003 -- The former is not really right for Haskell
1004 -- f x = case x of { (a,b) -> \y. e }
1006 -- f x y = case x of { (a,b) -> e }
1007 -- The difference is observable using 'seq'
1008 arityType dflags (Case scrut _ _ alts)
1009 = case foldr1 andArityType [arityType dflags rhs | (_,_,rhs) <- alts] of
1010 xs | exprIsCheap scrut -> xs
1011 xs@(AFun one_shot _) | one_shot -> AFun True ATop
1014 arityType dflags (Let b e)
1015 = case arityType dflags e of
1016 xs | cheap_bind b -> xs
1017 xs@(AFun one_shot _) | one_shot -> AFun True ATop
1020 cheap_bind (NonRec b e) = is_cheap (b,e)
1021 cheap_bind (Rec prs) = all is_cheap prs
1022 is_cheap (b,e) = (dopt Opt_DictsCheap dflags && isDictId b)
1024 -- If the experimental -fdicts-cheap flag is on, we eta-expand through
1025 -- dictionary bindings. This improves arities. Thereby, it also
1026 -- means that full laziness is less prone to floating out the
1027 -- application of a function to its dictionary arguments, which
1028 -- can thereby lose opportunities for fusion. Example:
1029 -- foo :: Ord a => a -> ...
1030 -- foo = /\a \(d:Ord a). let d' = ...d... in \(x:a). ....
1031 -- -- So foo has arity 1
1033 -- f = \x. foo dInt $ bar x
1035 -- The (foo DInt) is floated out, and makes ineffective a RULE
1036 -- foo (bar x) = ...
1038 -- One could go further and make exprIsCheap reply True to any
1039 -- dictionary-typed expression, but that's more work.
1041 arityType dflags other = ATop
1043 {- NOT NEEDED ANY MORE: etaExpand is cleverer
1044 ok_note InlineMe = False
1045 ok_note other = True
1046 -- Notice that we do not look through __inline_me__
1047 -- This may seem surprising, but consider
1048 -- f = _inline_me (\x -> e)
1049 -- We DO NOT want to eta expand this to
1050 -- f = \x -> (_inline_me (\x -> e)) x
1051 -- because the _inline_me gets dropped now it is applied,
1060 etaExpand :: Arity -- Result should have this number of value args
1062 -> CoreExpr -> Type -- Expression and its type
1064 -- (etaExpand n us e ty) returns an expression with
1065 -- the same meaning as 'e', but with arity 'n'.
1067 -- Given e' = etaExpand n us e ty
1069 -- ty = exprType e = exprType e'
1071 -- Note that SCCs are not treated specially. If we have
1072 -- etaExpand 2 (\x -> scc "foo" e)
1073 -- = (\xy -> (scc "foo" e) y)
1074 -- So the costs of evaluating 'e' (not 'e y') are attributed to "foo"
1076 etaExpand n us expr ty
1077 | manifestArity expr >= n = expr -- The no-op case
1079 = eta_expand n us expr ty
1082 -- manifestArity sees how many leading value lambdas there are
1083 manifestArity :: CoreExpr -> Arity
1084 manifestArity (Lam v e) | isId v = 1 + manifestArity e
1085 | otherwise = manifestArity e
1086 manifestArity (Note _ e) = manifestArity e
1087 manifestArity (Cast e _) = manifestArity e
1090 -- etaExpand deals with for-alls. For example:
1092 -- where E :: forall a. a -> a
1094 -- (/\b. \y::a -> E b y)
1096 -- It deals with coerces too, though they are now rare
1097 -- so perhaps the extra code isn't worth it
1099 eta_expand n us expr ty
1101 -- The ILX code generator requires eta expansion for type arguments
1102 -- too, but alas the 'n' doesn't tell us how many of them there
1103 -- may be. So we eagerly eta expand any big lambdas, and just
1104 -- cross our fingers about possible loss of sharing in the ILX case.
1105 -- The Right Thing is probably to make 'arity' include
1106 -- type variables throughout the compiler. (ToDo.)
1108 -- Saturated, so nothing to do
1111 -- Short cut for the case where there already
1112 -- is a lambda; no point in gratuitously adding more
1113 eta_expand n us (Lam v body) ty
1115 = Lam v (eta_expand n us body (applyTy ty (mkTyVarTy v)))
1118 = Lam v (eta_expand (n-1) us body (funResultTy ty))
1120 -- We used to have a special case that stepped inside Coerces here,
1121 -- thus: eta_expand n us (Note note@(Coerce _ ty) e) _
1122 -- = Note note (eta_expand n us e ty)
1123 -- BUT this led to an infinite loop
1124 -- Example: newtype T = MkT (Int -> Int)
1125 -- eta_expand 1 (coerce (Int->Int) e)
1126 -- --> coerce (Int->Int) (eta_expand 1 T e)
1128 -- --> coerce (Int->Int) (coerce T
1129 -- (\x::Int -> eta_expand 1 (coerce (Int->Int) e)))
1130 -- by the splitNewType_maybe case below
1133 eta_expand n us expr ty
1134 = ASSERT2 (exprType expr `coreEqType` ty, ppr (exprType expr) $$ ppr ty)
1135 case splitForAllTy_maybe ty of {
1138 Lam lam_tv (eta_expand n us2 (App expr (Type (mkTyVarTy lam_tv))) (substTyWith [tv] [mkTyVarTy lam_tv] ty'))
1140 lam_tv = mkTyVar (mkSysTvName uniq FSLIT("etaT")) (tyVarKind tv)
1144 case splitFunTy_maybe ty of {
1145 Just (arg_ty, res_ty) -> Lam arg1 (eta_expand (n-1) us2 (App expr (Var arg1)) res_ty)
1147 arg1 = mkSysLocal FSLIT("eta") uniq arg_ty
1153 -- newtype T = MkT ([T] -> Int)
1154 -- Consider eta-expanding this
1157 -- coerce T (\x::[T] -> (coerce ([T]->Int) e) x)
1159 case splitNewTypeRepCo_maybe ty of {
1161 mkCoerce (mkSymCoercion co) (eta_expand n us (mkCoerce co expr) ty1) ;
1164 -- We have an expression of arity > 0, but its type isn't a function
1165 -- This *can* legitmately happen: e.g. coerce Int (\x. x)
1166 -- Essentially the programmer is playing fast and loose with types
1167 -- (Happy does this a lot). So we simply decline to eta-expand.
1172 exprArity is a cheap-and-cheerful version of exprEtaExpandArity.
1173 It tells how many things the expression can be applied to before doing
1174 any work. It doesn't look inside cases, lets, etc. The idea is that
1175 exprEtaExpandArity will do the hard work, leaving something that's easy
1176 for exprArity to grapple with. In particular, Simplify uses exprArity to
1177 compute the ArityInfo for the Id.
1179 Originally I thought that it was enough just to look for top-level lambdas, but
1180 it isn't. I've seen this
1182 foo = PrelBase.timesInt
1184 We want foo to get arity 2 even though the eta-expander will leave it
1185 unchanged, in the expectation that it'll be inlined. But occasionally it
1186 isn't, because foo is blacklisted (used in a rule).
1188 Similarly, see the ok_note check in exprEtaExpandArity. So
1189 f = __inline_me (\x -> e)
1190 won't be eta-expanded.
1192 And in any case it seems more robust to have exprArity be a bit more intelligent.
1193 But note that (\x y z -> f x y z)
1194 should have arity 3, regardless of f's arity.
1197 exprArity :: CoreExpr -> Arity
1200 go (Var v) = idArity v
1201 go (Lam x e) | isId x = go e + 1
1203 go (Note n e) = go e
1204 go (Cast e _) = go e
1205 go (App e (Type t)) = go e
1206 go (App f a) | exprIsCheap a = (go f - 1) `max` 0
1207 -- NB: exprIsCheap a!
1208 -- f (fac x) does not have arity 2,
1209 -- even if f has arity 3!
1210 -- NB: `max 0`! (\x y -> f x) has arity 2, even if f is
1211 -- unknown, hence arity 0
1215 %************************************************************************
1217 \subsection{Equality}
1219 %************************************************************************
1221 @cheapEqExpr@ is a cheap equality test which bales out fast!
1222 True => definitely equal
1223 False => may or may not be equal
1226 cheapEqExpr :: Expr b -> Expr b -> Bool
1228 cheapEqExpr (Var v1) (Var v2) = v1==v2
1229 cheapEqExpr (Lit lit1) (Lit lit2) = lit1 == lit2
1230 cheapEqExpr (Type t1) (Type t2) = t1 `coreEqType` t2
1232 cheapEqExpr (App f1 a1) (App f2 a2)
1233 = f1 `cheapEqExpr` f2 && a1 `cheapEqExpr` a2
1235 cheapEqExpr _ _ = False
1237 exprIsBig :: Expr b -> Bool
1238 -- Returns True of expressions that are too big to be compared by cheapEqExpr
1239 exprIsBig (Lit _) = False
1240 exprIsBig (Var v) = False
1241 exprIsBig (Type t) = False
1242 exprIsBig (App f a) = exprIsBig f || exprIsBig a
1243 exprIsBig other = True
1248 tcEqExpr :: CoreExpr -> CoreExpr -> Bool
1249 -- Used in rule matching, so does *not* look through
1250 -- newtypes, predicate types; hence tcEqExpr
1252 tcEqExpr e1 e2 = tcEqExprX rn_env e1 e2
1254 rn_env = mkRnEnv2 (mkInScopeSet (exprFreeVars e1 `unionVarSet` exprFreeVars e2))
1256 tcEqExprX :: RnEnv2 -> CoreExpr -> CoreExpr -> Bool
1257 tcEqExprX env (Var v1) (Var v2) = rnOccL env v1 == rnOccR env v2
1258 tcEqExprX env (Lit lit1) (Lit lit2) = lit1 == lit2
1259 tcEqExprX env (App f1 a1) (App f2 a2) = tcEqExprX env f1 f2 && tcEqExprX env a1 a2
1260 tcEqExprX env (Lam v1 e1) (Lam v2 e2) = tcEqExprX (rnBndr2 env v1 v2) e1 e2
1261 tcEqExprX env (Let (NonRec v1 r1) e1)
1262 (Let (NonRec v2 r2) e2) = tcEqExprX env r1 r2
1263 && tcEqExprX (rnBndr2 env v1 v2) e1 e2
1264 tcEqExprX env (Let (Rec ps1) e1)
1265 (Let (Rec ps2) e2) = equalLength ps1 ps2
1266 && and (zipWith eq_rhs ps1 ps2)
1267 && tcEqExprX env' e1 e2
1269 env' = foldl2 rn_bndr2 env ps2 ps2
1270 rn_bndr2 env (b1,_) (b2,_) = rnBndr2 env b1 b2
1271 eq_rhs (_,r1) (_,r2) = tcEqExprX env' r1 r2
1272 tcEqExprX env (Case e1 v1 t1 a1)
1273 (Case e2 v2 t2 a2) = tcEqExprX env e1 e2
1274 && tcEqTypeX env t1 t2
1275 && equalLength a1 a2
1276 && and (zipWith (eq_alt env') a1 a2)
1278 env' = rnBndr2 env v1 v2
1280 tcEqExprX env (Note n1 e1) (Note n2 e2) = eq_note env n1 n2 && tcEqExprX env e1 e2
1281 tcEqExprX env (Cast e1 co1) (Cast e2 co2) = tcEqTypeX env co1 co2 && tcEqExprX env e1 e2
1282 tcEqExprX env (Type t1) (Type t2) = tcEqTypeX env t1 t2
1283 tcEqExprX env e1 e2 = False
1285 eq_alt env (c1,vs1,r1) (c2,vs2,r2) = c1==c2 && tcEqExprX (rnBndrs2 env vs1 vs2) r1 r2
1287 eq_note env (SCC cc1) (SCC cc2) = cc1 == cc2
1288 eq_note env (CoreNote s1) (CoreNote s2) = s1 == s2
1289 eq_note env other1 other2 = False
1293 %************************************************************************
1295 \subsection{The size of an expression}
1297 %************************************************************************
1300 coreBindsSize :: [CoreBind] -> Int
1301 coreBindsSize bs = foldr ((+) . bindSize) 0 bs
1303 exprSize :: CoreExpr -> Int
1304 -- A measure of the size of the expressions
1305 -- It also forces the expression pretty drastically as a side effect
1306 exprSize (Var v) = v `seq` 1
1307 exprSize (Lit lit) = lit `seq` 1
1308 exprSize (App f a) = exprSize f + exprSize a
1309 exprSize (Lam b e) = varSize b + exprSize e
1310 exprSize (Let b e) = bindSize b + exprSize e
1311 exprSize (Case e b t as) = seqType t `seq` exprSize e + varSize b + 1 + foldr ((+) . altSize) 0 as
1312 exprSize (Cast e co) = (seqType co `seq` 1) + exprSize e
1313 exprSize (Note n e) = noteSize n + exprSize e
1314 exprSize (Type t) = seqType t `seq` 1
1316 noteSize (SCC cc) = cc `seq` 1
1317 noteSize InlineMe = 1
1318 noteSize (CoreNote s) = s `seq` 1 -- hdaume: core annotations
1320 varSize :: Var -> Int
1321 varSize b | isTyVar b = 1
1322 | otherwise = seqType (idType b) `seq`
1323 megaSeqIdInfo (idInfo b) `seq`
1326 varsSize = foldr ((+) . varSize) 0
1328 bindSize (NonRec b e) = varSize b + exprSize e
1329 bindSize (Rec prs) = foldr ((+) . pairSize) 0 prs
1331 pairSize (b,e) = varSize b + exprSize e
1333 altSize (c,bs,e) = c `seq` varsSize bs + exprSize e
1337 %************************************************************************
1339 \subsection{Hashing}
1341 %************************************************************************
1344 hashExpr :: CoreExpr -> Int
1345 -- Two expressions that hash to the same Int may be equal (but may not be)
1346 -- Two expressions that hash to the different Ints are definitely unequal
1348 -- But "unequal" here means "not identical"; two alpha-equivalent
1349 -- expressions may hash to the different Ints
1351 -- The emphasis is on a crude, fast hash, rather than on high precision
1353 hashExpr e | hash < 0 = 77 -- Just in case we hit -maxInt
1356 hash = abs (hash_expr e) -- Negative numbers kill UniqFM
1358 hash_expr (Note _ e) = hash_expr e
1359 hash_expr (Cast e co) = hash_expr e
1360 hash_expr (Let (NonRec b r) e) = hashId b
1361 hash_expr (Let (Rec ((b,r):_)) e) = hashId b
1362 hash_expr (Case _ b _ _) = hashId b
1363 hash_expr (App f e) = hash_expr f * fast_hash_expr e
1364 hash_expr (Var v) = hashId v
1365 hash_expr (Lit lit) = hashLiteral lit
1366 hash_expr (Lam b _) = hashId b
1367 hash_expr (Type t) = trace "hash_expr: type" 1 -- Shouldn't happen
1369 fast_hash_expr (Var v) = hashId v
1370 fast_hash_expr (Lit lit) = hashLiteral lit
1371 fast_hash_expr (App f (Type _)) = fast_hash_expr f
1372 fast_hash_expr (App f a) = fast_hash_expr a
1373 fast_hash_expr (Lam b _) = hashId b
1374 fast_hash_expr other = 1
1377 hashId id = hashName (idName id)
1380 %************************************************************************
1382 \subsection{Determining non-updatable right-hand-sides}
1384 %************************************************************************
1386 Top-level constructor applications can usually be allocated
1387 statically, but they can't if the constructor, or any of the
1388 arguments, come from another DLL (because we can't refer to static
1389 labels in other DLLs).
1391 If this happens we simply make the RHS into an updatable thunk,
1392 and 'exectute' it rather than allocating it statically.
1395 rhsIsStatic :: PackageId -> CoreExpr -> Bool
1396 -- This function is called only on *top-level* right-hand sides
1397 -- Returns True if the RHS can be allocated statically, with
1398 -- no thunks involved at all.
1400 -- It's called (i) in TidyPgm.hasCafRefs to decide if the rhs is, or
1401 -- refers to, CAFs; and (ii) in CoreToStg to decide whether to put an
1402 -- update flag on it.
1404 -- The basic idea is that rhsIsStatic returns True only if the RHS is
1405 -- (a) a value lambda
1406 -- (b) a saturated constructor application with static args
1408 -- BUT watch out for
1409 -- (i) Any cross-DLL references kill static-ness completely
1410 -- because they must be 'executed' not statically allocated
1411 -- ("DLL" here really only refers to Windows DLLs, on other platforms,
1412 -- this is not necessary)
1414 -- (ii) We treat partial applications as redexes, because in fact we
1415 -- make a thunk for them that runs and builds a PAP
1416 -- at run-time. The only appliations that are treated as
1417 -- static are *saturated* applications of constructors.
1419 -- We used to try to be clever with nested structures like this:
1420 -- ys = (:) w ((:) w [])
1421 -- on the grounds that CorePrep will flatten ANF-ise it later.
1422 -- But supporting this special case made the function much more
1423 -- complicated, because the special case only applies if there are no
1424 -- enclosing type lambdas:
1425 -- ys = /\ a -> Foo (Baz ([] a))
1426 -- Here the nested (Baz []) won't float out to top level in CorePrep.
1428 -- But in fact, even without -O, nested structures at top level are
1429 -- flattened by the simplifier, so we don't need to be super-clever here.
1433 -- f = \x::Int. x+7 TRUE
1434 -- p = (True,False) TRUE
1436 -- d = (fst p, False) FALSE because there's a redex inside
1437 -- (this particular one doesn't happen but...)
1439 -- h = D# (1.0## /## 2.0##) FALSE (redex again)
1440 -- n = /\a. Nil a TRUE
1442 -- t = /\a. (:) (case w a of ...) (Nil a) FALSE (redex)
1445 -- This is a bit like CoreUtils.exprIsHNF, with the following differences:
1446 -- a) scc "foo" (\x -> ...) is updatable (so we catch the right SCC)
1448 -- b) (C x xs), where C is a contructors is updatable if the application is
1451 -- c) don't look through unfolding of f in (f x).
1453 -- When opt_RuntimeTypes is on, we keep type lambdas and treat
1454 -- them as making the RHS re-entrant (non-updatable).
1456 rhsIsStatic this_pkg rhs = is_static False rhs
1458 is_static :: Bool -- True <=> in a constructor argument; must be atomic
1461 is_static False (Lam b e) = isRuntimeVar b || is_static False e
1463 is_static in_arg (Note (SCC _) e) = False
1464 is_static in_arg (Note _ e) = is_static in_arg e
1465 is_static in_arg (Cast e co) = is_static in_arg e
1467 is_static in_arg (Lit lit)
1469 MachLabel _ _ -> False
1471 -- A MachLabel (foreign import "&foo") in an argument
1472 -- prevents a constructor application from being static. The
1473 -- reason is that it might give rise to unresolvable symbols
1474 -- in the object file: under Linux, references to "weak"
1475 -- symbols from the data segment give rise to "unresolvable
1476 -- relocation" errors at link time This might be due to a bug
1477 -- in the linker, but we'll work around it here anyway.
1480 is_static in_arg other_expr = go other_expr 0
1482 go (Var f) n_val_args
1483 #if mingw32_TARGET_OS
1484 | not (isDllName this_pkg (idName f))
1486 = saturated_data_con f n_val_args
1487 || (in_arg && n_val_args == 0)
1488 -- A naked un-applied variable is *not* deemed a static RHS
1490 -- Reason: better to update so that the indirection gets shorted
1491 -- out, and the true value will be seen
1492 -- NB: if you change this, you'll break the invariant that THUNK_STATICs
1493 -- are always updatable. If you do so, make sure that non-updatable
1494 -- ones have enough space for their static link field!
1496 go (App f a) n_val_args
1497 | isTypeArg a = go f n_val_args
1498 | not in_arg && is_static True a = go f (n_val_args + 1)
1499 -- The (not in_arg) checks that we aren't in a constructor argument;
1500 -- if we are, we don't allow (value) applications of any sort
1502 -- NB. In case you wonder, args are sometimes not atomic. eg.
1503 -- x = D# (1.0## /## 2.0##)
1504 -- can't float because /## can fail.
1506 go (Note (SCC _) f) n_val_args = False
1507 go (Note _ f) n_val_args = go f n_val_args
1508 go (Cast e co) n_val_args = go e n_val_args
1510 go other n_val_args = False
1512 saturated_data_con f n_val_args
1513 = case isDataConWorkId_maybe f of
1514 Just dc -> n_val_args == dataConRepArity dc