2 % (c) The University of Glasgow 2006
3 % (c) The GRASP/AQUA Project, Glasgow University, 1992-1998
6 Utility functions on @Core@ syntax
11 mkInlineMe, mkSCC, mkCoerce,
12 bindNonRec, needsCaseBinding,
13 mkIfThenElse, mkAltExpr, mkPiType, mkPiTypes,
15 -- Taking expressions apart
16 findDefault, findAlt, isDefaultAlt, mergeAlts,
18 -- Properties of expressions
19 exprType, coreAltType,
20 exprIsDupable, exprIsTrivial, exprIsCheap,
21 exprIsHNF,exprOkForSpeculation, exprIsBig,
22 exprIsConApp_maybe, exprIsBottom,
25 -- Arity and eta expansion
26 manifestArity, exprArity,
27 exprEtaExpandArity, etaExpand,
36 cheapEqExpr, tcEqExpr, tcEqExprX, applyTypeToArgs, applyTypeToArg,
38 dataConOrigInstPat, dataConRepInstPat, dataConRepFSInstPat
41 #include "HsVersions.h"
77 import GHC.Exts -- For `xori`
81 %************************************************************************
83 \subsection{Find the type of a Core atom/expression}
85 %************************************************************************
88 exprType :: CoreExpr -> Type
90 exprType (Var var) = idType var
91 exprType (Lit lit) = literalType lit
92 exprType (Let _ body) = exprType body
93 exprType (Case _ _ ty alts) = ty
95 = let (_, ty) = coercionKind co in ty
96 exprType (Note other_note e) = exprType e
97 exprType (Lam binder expr) = mkPiType binder (exprType expr)
99 = case collectArgs e of
100 (fun, args) -> applyTypeToArgs e (exprType fun) args
102 exprType other = pprTrace "exprType" (pprCoreExpr other) alphaTy
104 coreAltType :: CoreAlt -> Type
105 coreAltType (_,_,rhs) = exprType rhs
108 @mkPiType@ makes a (->) type or a forall type, depending on whether
109 it is given a type variable or a term variable. We cleverly use the
110 lbvarinfo field to figure out the right annotation for the arrove in
111 case of a term variable.
114 mkPiType :: Var -> Type -> Type -- The more polymorphic version
115 mkPiTypes :: [Var] -> Type -> Type -- doesn't work...
117 mkPiTypes vs ty = foldr mkPiType ty vs
120 | isId v = mkFunTy (idType v) ty
121 | otherwise = mkForAllTy v ty
125 applyTypeToArg :: Type -> CoreExpr -> Type
126 applyTypeToArg fun_ty (Type arg_ty) = applyTy fun_ty arg_ty
127 applyTypeToArg fun_ty other_arg = funResultTy fun_ty
129 applyTypeToArgs :: CoreExpr -> Type -> [CoreExpr] -> Type
130 -- A more efficient version of applyTypeToArg
131 -- when we have several args
132 -- The first argument is just for debugging
133 applyTypeToArgs e op_ty [] = op_ty
135 applyTypeToArgs e op_ty (Type ty : args)
136 = -- Accumulate type arguments so we can instantiate all at once
139 go rev_tys (Type ty : args) = go (ty:rev_tys) args
140 go rev_tys rest_args = applyTypeToArgs e op_ty' rest_args
142 op_ty' = applyTys op_ty (reverse rev_tys)
144 applyTypeToArgs e op_ty (other_arg : args)
145 = case (splitFunTy_maybe op_ty) of
146 Just (_, res_ty) -> applyTypeToArgs e res_ty args
147 Nothing -> pprPanic "applyTypeToArgs" (pprCoreExpr e $$ ppr op_ty)
152 %************************************************************************
154 \subsection{Attaching notes}
156 %************************************************************************
158 mkNote removes redundant coercions, and SCCs where possible
162 mkNote :: Note -> CoreExpr -> CoreExpr
163 mkNote (SCC cc) expr = mkSCC cc expr
164 mkNote InlineMe expr = mkInlineMe expr
165 mkNote note expr = Note note expr
169 Drop trivial InlineMe's. This is somewhat important, because if we have an unfolding
170 that looks like (Note InlineMe (Var v)), the InlineMe doesn't go away because it may
171 not be *applied* to anything.
173 We don't use exprIsTrivial here, though, because we sometimes generate worker/wrapper
176 f = inline_me (coerce t fw)
177 As usual, the inline_me prevents the worker from getting inlined back into the wrapper.
178 We want the split, so that the coerces can cancel at the call site.
180 However, we can get left with tiresome type applications. Notably, consider
181 f = /\ a -> let t = e in (t, w)
182 Then lifting the let out of the big lambda gives
184 f = /\ a -> let t = inline_me (t' a) in (t, w)
185 The inline_me is to stop the simplifier inlining t' right back
186 into t's RHS. In the next phase we'll substitute for t (since
187 its rhs is trivial) and *then* we could get rid of the inline_me.
188 But it hardly seems worth it, so I don't bother.
191 mkInlineMe (Var v) = Var v
192 mkInlineMe e = Note InlineMe e
198 mkCoerce :: Coercion -> CoreExpr -> CoreExpr
199 mkCoerce co (Cast expr co2)
200 = ASSERT(let { (from_ty, _to_ty) = coercionKind co;
201 (_from_ty2, to_ty2) = coercionKind co2} in
202 from_ty `coreEqType` to_ty2 )
203 mkCoerce (mkTransCoercion co2 co) expr
206 = let (from_ty, to_ty) = coercionKind co in
207 -- if to_ty `coreEqType` from_ty
210 ASSERT2(from_ty `coreEqType` (exprType expr), text "Trying to coerce" <+> text "(" <> ppr expr $$ text "::" <+> ppr (exprType expr) <> text ")" $$ ppr co $$ ppr (coercionKindPredTy co))
215 mkSCC :: CostCentre -> Expr b -> Expr b
216 -- Note: Nested SCC's *are* preserved for the benefit of
217 -- cost centre stack profiling
218 mkSCC cc (Lit lit) = Lit lit
219 mkSCC cc (Lam x e) = Lam x (mkSCC cc e) -- Move _scc_ inside lambda
220 mkSCC cc (Note (SCC cc') e) = Note (SCC cc) (Note (SCC cc') e)
221 mkSCC cc (Note n e) = Note n (mkSCC cc e) -- Move _scc_ inside notes
222 mkSCC cc (Cast e co) = Cast (mkSCC cc e) co -- Move _scc_ inside cast
223 mkSCC cc expr = Note (SCC cc) expr
227 %************************************************************************
229 \subsection{Other expression construction}
231 %************************************************************************
234 bindNonRec :: Id -> CoreExpr -> CoreExpr -> CoreExpr
235 -- (bindNonRec x r b) produces either
238 -- case r of x { _DEFAULT_ -> b }
240 -- depending on whether x is unlifted or not
241 -- It's used by the desugarer to avoid building bindings
242 -- that give Core Lint a heart attack. Actually the simplifier
243 -- deals with them perfectly well.
245 bindNonRec bndr rhs body
246 | needsCaseBinding (idType bndr) rhs = Case rhs bndr (exprType body) [(DEFAULT,[],body)]
247 | otherwise = Let (NonRec bndr rhs) body
249 needsCaseBinding ty rhs = isUnLiftedType ty && not (exprOkForSpeculation rhs)
250 -- Make a case expression instead of a let
251 -- These can arise either from the desugarer,
252 -- or from beta reductions: (\x.e) (x +# y)
256 mkAltExpr :: AltCon -> [CoreBndr] -> [Type] -> CoreExpr
257 -- This guy constructs the value that the scrutinee must have
258 -- when you are in one particular branch of a case
259 mkAltExpr (DataAlt con) args inst_tys
260 = mkConApp con (map Type inst_tys ++ varsToCoreExprs args)
261 mkAltExpr (LitAlt lit) [] []
264 mkIfThenElse :: CoreExpr -> CoreExpr -> CoreExpr -> CoreExpr
265 mkIfThenElse guard then_expr else_expr
266 -- Not going to be refining, so okay to take the type of the "then" clause
267 = Case guard (mkWildId boolTy) (exprType then_expr)
268 [ (DataAlt falseDataCon, [], else_expr), -- Increasing order of tag!
269 (DataAlt trueDataCon, [], then_expr) ]
273 %************************************************************************
275 \subsection{Taking expressions apart}
277 %************************************************************************
279 The default alternative must be first, if it exists at all.
280 This makes it easy to find, though it makes matching marginally harder.
283 findDefault :: [CoreAlt] -> ([CoreAlt], Maybe CoreExpr)
284 findDefault ((DEFAULT,args,rhs) : alts) = ASSERT( null args ) (alts, Just rhs)
285 findDefault alts = (alts, Nothing)
287 findAlt :: AltCon -> [CoreAlt] -> CoreAlt
290 (deflt@(DEFAULT,_,_):alts) -> go alts deflt
291 other -> go alts panic_deflt
293 panic_deflt = pprPanic "Missing alternative" (ppr con $$ vcat (map ppr alts))
296 go (alt@(con1,_,_) : alts) deflt
297 = case con `cmpAltCon` con1 of
298 LT -> deflt -- Missed it already; the alts are in increasing order
300 GT -> ASSERT( not (con1 == DEFAULT) ) go alts deflt
302 isDefaultAlt :: CoreAlt -> Bool
303 isDefaultAlt (DEFAULT, _, _) = True
304 isDefaultAlt other = False
306 ---------------------------------
307 mergeAlts :: [CoreAlt] -> [CoreAlt] -> [CoreAlt]
308 -- Merge preserving order; alternatives in the first arg
309 -- shadow ones in the second
310 mergeAlts [] as2 = as2
311 mergeAlts as1 [] = as1
312 mergeAlts (a1:as1) (a2:as2)
313 = case a1 `cmpAlt` a2 of
314 LT -> a1 : mergeAlts as1 (a2:as2)
315 EQ -> a1 : mergeAlts as1 as2 -- Discard a2
316 GT -> a2 : mergeAlts (a1:as1) as2
320 %************************************************************************
322 \subsection{Figuring out things about expressions}
324 %************************************************************************
326 @exprIsTrivial@ is true of expressions we are unconditionally happy to
327 duplicate; simple variables and constants, and type
328 applications. Note that primop Ids aren't considered
331 @exprIsBottom@ is true of expressions that are guaranteed to diverge
334 There used to be a gruesome test for (hasNoBinding v) in the
336 exprIsTrivial (Var v) | hasNoBinding v = idArity v == 0
337 The idea here is that a constructor worker, like $wJust, is
338 really short for (\x -> $wJust x), becuase $wJust has no binding.
339 So it should be treated like a lambda. Ditto unsaturated primops.
340 But now constructor workers are not "have-no-binding" Ids. And
341 completely un-applied primops and foreign-call Ids are sufficiently
342 rare that I plan to allow them to be duplicated and put up with
345 SCC notes. We do not treat (_scc_ "foo" x) as trivial, because
346 a) it really generates code, (and a heap object when it's
347 a function arg) to capture the cost centre
348 b) see the note [SCC-and-exprIsTrivial] in Simplify.simplLazyBind
351 exprIsTrivial (Var v) = True -- See notes above
352 exprIsTrivial (Type _) = True
353 exprIsTrivial (Lit lit) = litIsTrivial lit
354 exprIsTrivial (App e arg) = not (isRuntimeArg arg) && exprIsTrivial e
355 exprIsTrivial (Note (SCC _) e) = False -- See notes above
356 exprIsTrivial (Note _ e) = exprIsTrivial e
357 exprIsTrivial (Cast e co) = exprIsTrivial e
358 exprIsTrivial (Lam b body) = not (isRuntimeVar b) && exprIsTrivial body
359 exprIsTrivial other = False
363 @exprIsDupable@ is true of expressions that can be duplicated at a modest
364 cost in code size. This will only happen in different case
365 branches, so there's no issue about duplicating work.
367 That is, exprIsDupable returns True of (f x) even if
368 f is very very expensive to call.
370 Its only purpose is to avoid fruitless let-binding
371 and then inlining of case join points
375 exprIsDupable (Type _) = True
376 exprIsDupable (Var v) = True
377 exprIsDupable (Lit lit) = litIsDupable lit
378 exprIsDupable (Note InlineMe e) = True
379 exprIsDupable (Note _ e) = exprIsDupable e
380 exprIsDupable (Cast e co) = exprIsDupable e
384 go (Var v) n_args = True
385 go (App f a) n_args = n_args < dupAppSize
388 go other n_args = False
391 dupAppSize = 4 -- Size of application we are prepared to duplicate
394 @exprIsCheap@ looks at a Core expression and returns \tr{True} if
395 it is obviously in weak head normal form, or is cheap to get to WHNF.
396 [Note that that's not the same as exprIsDupable; an expression might be
397 big, and hence not dupable, but still cheap.]
399 By ``cheap'' we mean a computation we're willing to:
400 push inside a lambda, or
401 inline at more than one place
402 That might mean it gets evaluated more than once, instead of being
403 shared. The main examples of things which aren't WHNF but are
408 (where e, and all the ei are cheap)
411 (where e and b are cheap)
414 (where op is a cheap primitive operator)
417 (because we are happy to substitute it inside a lambda)
419 Notice that a variable is considered 'cheap': we can push it inside a lambda,
420 because sharing will make sure it is only evaluated once.
423 exprIsCheap :: CoreExpr -> Bool
424 exprIsCheap (Lit lit) = True
425 exprIsCheap (Type _) = True
426 exprIsCheap (Var _) = True
427 exprIsCheap (Note InlineMe e) = True
428 exprIsCheap (Note _ e) = exprIsCheap e
429 exprIsCheap (Cast e co) = exprIsCheap e
430 exprIsCheap (Lam x e) = isRuntimeVar x || exprIsCheap e
431 exprIsCheap (Case e _ _ alts) = exprIsCheap e &&
432 and [exprIsCheap rhs | (_,_,rhs) <- alts]
433 -- Experimentally, treat (case x of ...) as cheap
434 -- (and case __coerce x etc.)
435 -- This improves arities of overloaded functions where
436 -- there is only dictionary selection (no construction) involved
437 exprIsCheap (Let (NonRec x _) e)
438 | isUnLiftedType (idType x) = exprIsCheap e
440 -- strict lets always have cheap right hand sides,
441 -- and do no allocation.
443 exprIsCheap other_expr -- Applications and variables
446 -- Accumulate value arguments, then decide
447 go (App f a) val_args | isRuntimeArg a = go f (a:val_args)
448 | otherwise = go f val_args
450 go (Var f) [] = True -- Just a type application of a variable
451 -- (f t1 t2 t3) counts as WHNF
453 = case globalIdDetails f of
454 RecordSelId {} -> go_sel args
455 ClassOpId _ -> go_sel args
456 PrimOpId op -> go_primop op args
458 DataConWorkId _ -> go_pap args
459 other | length args < idArity f -> go_pap args
461 other -> isBottomingId f
462 -- Application of a function which
463 -- always gives bottom; we treat this as cheap
464 -- because it certainly doesn't need to be shared!
466 go other args = False
469 go_pap args = all exprIsTrivial args
470 -- For constructor applications and primops, check that all
471 -- the args are trivial. We don't want to treat as cheap, say,
473 -- We'll put up with one constructor application, but not dozens
476 go_primop op args = primOpIsCheap op && all exprIsCheap args
477 -- In principle we should worry about primops
478 -- that return a type variable, since the result
479 -- might be applied to something, but I'm not going
480 -- to bother to check the number of args
483 go_sel [arg] = exprIsCheap arg -- I'm experimenting with making record selection
484 go_sel other = False -- look cheap, so we will substitute it inside a
485 -- lambda. Particularly for dictionary field selection.
486 -- BUT: Take care with (sel d x)! The (sel d) might be cheap, but
487 -- there's no guarantee that (sel d x) will be too. Hence (n_val_args == 1)
490 exprOkForSpeculation returns True of an expression that it is
492 * safe to evaluate even if normal order eval might not
493 evaluate the expression at all, or
495 * safe *not* to evaluate even if normal order would do so
499 the expression guarantees to terminate,
501 without raising an exception,
502 without causing a side effect (e.g. writing a mutable variable)
504 NB: if exprIsHNF e, then exprOkForSpecuation e
507 let x = case y# +# 1# of { r# -> I# r# }
510 case y# +# 1# of { r# ->
515 We can only do this if the (y+1) is ok for speculation: it has no
516 side effects, and can't diverge or raise an exception.
519 exprOkForSpeculation :: CoreExpr -> Bool
520 exprOkForSpeculation (Lit _) = True
521 exprOkForSpeculation (Type _) = True
522 -- Tick boxes are *not* suitable for speculation
523 exprOkForSpeculation (Var v) = isUnLiftedType (idType v)
524 && not (isTickBoxOp v)
525 exprOkForSpeculation (Note _ e) = exprOkForSpeculation e
526 exprOkForSpeculation (Cast e co) = exprOkForSpeculation e
527 exprOkForSpeculation other_expr
528 = case collectArgs other_expr of
529 (Var f, args) -> spec_ok (globalIdDetails f) args
533 spec_ok (DataConWorkId _) args
534 = True -- The strictness of the constructor has already
535 -- been expressed by its "wrapper", so we don't need
536 -- to take the arguments into account
538 spec_ok (PrimOpId op) args
539 | isDivOp op, -- Special case for dividing operations that fail
540 [arg1, Lit lit] <- args -- only if the divisor is zero
541 = not (isZeroLit lit) && exprOkForSpeculation arg1
542 -- Often there is a literal divisor, and this
543 -- can get rid of a thunk in an inner looop
546 = primOpOkForSpeculation op &&
547 all exprOkForSpeculation args
548 -- A bit conservative: we don't really need
549 -- to care about lazy arguments, but this is easy
551 spec_ok other args = False
553 isDivOp :: PrimOp -> Bool
554 -- True of dyadic operators that can fail
555 -- only if the second arg is zero
556 -- This function probably belongs in PrimOp, or even in
557 -- an automagically generated file.. but it's such a
558 -- special case I thought I'd leave it here for now.
559 isDivOp IntQuotOp = True
560 isDivOp IntRemOp = True
561 isDivOp WordQuotOp = True
562 isDivOp WordRemOp = True
563 isDivOp IntegerQuotRemOp = True
564 isDivOp IntegerDivModOp = True
565 isDivOp FloatDivOp = True
566 isDivOp DoubleDivOp = True
567 isDivOp other = False
572 exprIsBottom :: CoreExpr -> Bool -- True => definitely bottom
573 exprIsBottom e = go 0 e
575 -- n is the number of args
576 go n (Note _ e) = go n e
577 go n (Cast e co) = go n e
578 go n (Let _ e) = go n e
579 go n (Case e _ _ _) = go 0 e -- Just check the scrut
580 go n (App e _) = go (n+1) e
581 go n (Var v) = idAppIsBottom v n
583 go n (Lam _ _) = False
584 go n (Type _) = False
586 idAppIsBottom :: Id -> Int -> Bool
587 idAppIsBottom id n_val_args = appIsBottom (idNewStrictness id) n_val_args
590 @exprIsHNF@ returns true for expressions that are certainly *already*
591 evaluated to *head* normal form. This is used to decide whether it's ok
594 case x of _ -> e ===> e
596 and to decide whether it's safe to discard a `seq`
598 So, it does *not* treat variables as evaluated, unless they say they are.
600 But it *does* treat partial applications and constructor applications
601 as values, even if their arguments are non-trivial, provided the argument
603 e.g. (:) (f x) (map f xs) is a value
604 map (...redex...) is a value
605 Because `seq` on such things completes immediately
607 For unlifted argument types, we have to be careful:
609 Suppose (f x) diverges; then C (f x) is not a value. True, but
610 this form is illegal (see the invariants in CoreSyn). Args of unboxed
611 type must be ok-for-speculation (or trivial).
614 exprIsHNF :: CoreExpr -> Bool -- True => Value-lambda, constructor, PAP
615 exprIsHNF (Var v) -- NB: There are no value args at this point
616 = isDataConWorkId v -- Catches nullary constructors,
617 -- so that [] and () are values, for example
618 || idArity v > 0 -- Catches (e.g.) primops that don't have unfoldings
619 || isEvaldUnfolding (idUnfolding v)
620 -- Check the thing's unfolding; it might be bound to a value
621 -- A worry: what if an Id's unfolding is just itself:
622 -- then we could get an infinite loop...
624 exprIsHNF (Lit l) = True
625 exprIsHNF (Type ty) = True -- Types are honorary Values;
626 -- we don't mind copying them
627 exprIsHNF (Lam b e) = isRuntimeVar b || exprIsHNF e
628 exprIsHNF (Note _ e) = exprIsHNF e
629 exprIsHNF (Cast e co) = exprIsHNF e
630 exprIsHNF (App e (Type _)) = exprIsHNF e
631 exprIsHNF (App e a) = app_is_value e [a]
632 exprIsHNF other = False
634 -- There is at least one value argument
635 app_is_value (Var fun) args
636 | isDataConWorkId fun -- Constructor apps are values
637 || idArity fun > valArgCount args -- Under-applied function
638 = check_args (idType fun) args
639 app_is_value (App f a) as = app_is_value f (a:as)
640 app_is_value other as = False
642 -- 'check_args' checks that unlifted-type args
643 -- are in fact guaranteed non-divergent
644 check_args fun_ty [] = True
645 check_args fun_ty (Type _ : args) = case splitForAllTy_maybe fun_ty of
646 Just (_, ty) -> check_args ty args
647 check_args fun_ty (arg : args)
648 | isUnLiftedType arg_ty = exprOkForSpeculation arg
649 | otherwise = check_args res_ty args
651 (arg_ty, res_ty) = splitFunTy fun_ty
655 -- These InstPat functions go here to avoid circularity between DataCon and Id
656 dataConRepInstPat = dataConInstPat dataConRepArgTys (repeat (FSLIT("ipv")))
657 dataConRepFSInstPat = dataConInstPat dataConRepArgTys
658 dataConOrigInstPat = dataConInstPat dc_arg_tys (repeat (FSLIT("ipv")))
660 dc_arg_tys dc = map mkPredTy (dataConTheta dc) ++ dataConOrigArgTys dc
661 -- Remember to include the existential dictionaries
663 dataConInstPat :: (DataCon -> [Type]) -- function used to find arg tys
664 -> [FastString] -- A long enough list of FSs to use for names
665 -> [Unique] -- An equally long list of uniques, at least one for each binder
667 -> [Type] -- Types to instantiate the universally quantified tyvars
668 -> ([TyVar], [CoVar], [Id]) -- Return instantiated variables
669 -- dataConInstPat arg_fun fss us con inst_tys returns a triple
670 -- (ex_tvs, co_tvs, arg_ids),
672 -- ex_tvs are intended to be used as binders for existential type args
674 -- co_tvs are intended to be used as binders for coercion args and the kinds
675 -- of these vars have been instantiated by the inst_tys and the ex_tys
677 -- arg_ids are indended to be used as binders for value arguments, including
678 -- dicts, and their types have been instantiated with inst_tys and ex_tys
681 -- The following constructor T1
684 -- T1 :: forall b. Int -> b -> T(a,b)
687 -- has representation type
688 -- forall a. forall a1. forall b. (a :=: (a1,b)) =>
691 -- dataConInstPat fss us T1 (a1',b') will return
693 -- ([a1'', b''], [c :: (a1', b'):=:(a1'', b'')], [x :: Int, y :: b''])
695 -- where the double-primed variables are created with the FastStrings and
696 -- Uniques given as fss and us
697 dataConInstPat arg_fun fss uniqs con inst_tys
698 = (ex_bndrs, co_bndrs, id_bndrs)
700 univ_tvs = dataConUnivTyVars con
701 ex_tvs = dataConExTyVars con
702 arg_tys = arg_fun con
703 eq_spec = dataConEqSpec con
704 eq_preds = eqSpecPreds eq_spec
707 n_co = length eq_spec
709 -- split the Uniques and FastStrings
710 (ex_uniqs, uniqs') = splitAt n_ex uniqs
711 (co_uniqs, id_uniqs) = splitAt n_co uniqs'
713 (ex_fss, fss') = splitAt n_ex fss
714 (co_fss, id_fss) = splitAt n_co fss'
716 -- Make existential type variables
717 ex_bndrs = zipWith3 mk_ex_var ex_uniqs ex_fss ex_tvs
718 mk_ex_var uniq fs var = mkTyVar new_name kind
720 new_name = mkSysTvName uniq fs
723 -- Make the instantiating substitution
724 subst = zipOpenTvSubst (univ_tvs ++ ex_tvs) (inst_tys ++ map mkTyVarTy ex_bndrs)
726 -- Make new coercion vars, instantiating kind
727 co_bndrs = zipWith3 mk_co_var co_uniqs co_fss eq_preds
728 mk_co_var uniq fs eq_pred = mkCoVar new_name co_kind
730 new_name = mkSysTvName uniq fs
731 co_kind = substTy subst (mkPredTy eq_pred)
733 -- make value vars, instantiating types
734 mk_id_var uniq fs ty = mkUserLocal (mkVarOccFS fs) uniq (substTy subst ty) noSrcLoc
735 id_bndrs = zipWith3 mk_id_var id_uniqs id_fss arg_tys
737 exprIsConApp_maybe :: CoreExpr -> Maybe (DataCon, [CoreExpr])
738 -- Returns (Just (dc, [x1..xn])) if the argument expression is
739 -- a constructor application of the form (dc x1 .. xn)
740 exprIsConApp_maybe (Cast expr co)
741 = -- Here we do the PushC reduction rule as described in the FC paper
742 case exprIsConApp_maybe expr of {
744 Just (dc, dc_args) ->
746 -- The transformation applies iff we have
747 -- (C e1 ... en) `cast` co
748 -- where co :: (T t1 .. tn) :=: (T s1 ..sn)
749 -- That is, with a T at the top of both sides
750 -- The left-hand one must be a T, because exprIsConApp returned True
751 -- but the right-hand one might not be. (Though it usually will.)
753 let (from_ty, to_ty) = coercionKind co
754 (from_tc, from_tc_arg_tys) = splitTyConApp from_ty
755 -- The inner one must be a TyConApp
757 case splitTyConApp_maybe to_ty of {
759 Just (to_tc, to_tc_arg_tys)
760 | from_tc /= to_tc -> Nothing
761 -- These two Nothing cases are possible; we might see
762 -- (C x y) `cast` (g :: T a ~ S [a]),
763 -- where S is a type function. In fact, exprIsConApp
764 -- will probably not be called in such circumstances,
765 -- but there't nothing wrong with it
769 tc_arity = tyConArity from_tc
771 (univ_args, rest1) = splitAt tc_arity dc_args
772 (ex_args, rest2) = splitAt n_ex_tvs rest1
773 (co_args, val_args) = splitAt n_cos rest2
775 arg_tys = dataConRepArgTys dc
776 dc_univ_tyvars = dataConUnivTyVars dc
777 dc_ex_tyvars = dataConExTyVars dc
778 dc_eq_spec = dataConEqSpec dc
779 dc_tyvars = dc_univ_tyvars ++ dc_ex_tyvars
780 n_ex_tvs = length dc_ex_tyvars
781 n_cos = length dc_eq_spec
783 -- Make the "theta" from Fig 3 of the paper
784 gammas = decomposeCo tc_arity co
785 new_tys = gammas ++ map (\ (Type t) -> t) ex_args
786 theta = zipOpenTvSubst dc_tyvars new_tys
788 -- First we cast the existential coercion arguments
789 cast_co (tv,ty) (Type co) = Type $ mkSymCoercion (substTyVar theta tv)
791 `mkTransCoercion` (substTy theta ty)
792 new_co_args = zipWith cast_co dc_eq_spec co_args
794 -- ...and now value arguments
795 new_val_args = zipWith cast_arg arg_tys val_args
796 cast_arg arg_ty arg = mkCoerce (substTy theta arg_ty) arg
799 ASSERT( length univ_args == tc_arity )
800 ASSERT( from_tc == dataConTyCon dc )
801 ASSERT( and (zipWith coreEqType [t | Type t <- univ_args] from_tc_arg_tys) )
802 ASSERT( all isTypeArg (univ_args ++ ex_args) )
803 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 )
805 Just (dc, map Type to_tc_arg_tys ++ ex_args ++ new_co_args ++ new_val_args)
809 -- We do not want to tell the world that we have a
810 -- Cons, to *stop* Case of Known Cons, which removes
812 exprIsConApp_maybe (Note (TickBox {}) expr)
814 exprIsConApp_maybe (Note (BinaryTickBox {}) expr)
818 exprIsConApp_maybe (Note _ expr)
819 = exprIsConApp_maybe expr
820 -- We ignore InlineMe notes in case we have
821 -- x = __inline_me__ (a,b)
822 -- All part of making sure that INLINE pragmas never hurt
823 -- Marcin tripped on this one when making dictionaries more inlinable
825 -- In fact, we ignore all notes. For example,
826 -- case _scc_ "foo" (C a b) of
828 -- should be optimised away, but it will be only if we look
829 -- through the SCC note.
831 exprIsConApp_maybe expr = analyse (collectArgs expr)
833 analyse (Var fun, args)
834 | Just con <- isDataConWorkId_maybe fun,
835 args `lengthAtLeast` dataConRepArity con
836 -- Might be > because the arity excludes type args
839 -- Look through unfoldings, but only cheap ones, because
840 -- we are effectively duplicating the unfolding
841 analyse (Var fun, [])
842 | let unf = idUnfolding fun,
844 = exprIsConApp_maybe (unfoldingTemplate unf)
846 analyse other = Nothing
851 %************************************************************************
853 \subsection{Eta reduction and expansion}
855 %************************************************************************
858 exprEtaExpandArity :: DynFlags -> CoreExpr -> Arity
859 {- The Arity returned is the number of value args the
860 thing can be applied to without doing much work
862 exprEtaExpandArity is used when eta expanding
865 It returns 1 (or more) to:
866 case x of p -> \s -> ...
867 because for I/O ish things we really want to get that \s to the top.
868 We are prepared to evaluate x each time round the loop in order to get that
870 It's all a bit more subtle than it looks:
874 Consider one-shot lambdas
875 let x = expensive in \y z -> E
876 We want this to have arity 2 if the \y-abstraction is a 1-shot lambda
877 Hence the ArityType returned by arityType
879 2. The state-transformer hack
881 The one-shot lambda special cause is particularly important/useful for
882 IO state transformers, where we often get
883 let x = E in \ s -> ...
885 and the \s is a real-world state token abstraction. Such abstractions
886 are almost invariably 1-shot, so we want to pull the \s out, past the
887 let x=E, even if E is expensive. So we treat state-token lambdas as
888 one-shot even if they aren't really. The hack is in Id.isOneShotBndr.
890 3. Dealing with bottom
893 f = \x -> error "foo"
894 Here, arity 1 is fine. But if it is
898 then we want to get arity 2. Tecnically, this isn't quite right, because
900 should diverge, but it'll converge if we eta-expand f. Nevertheless, we
901 do so; it improves some programs significantly, and increasing convergence
902 isn't a bad thing. Hence the ABot/ATop in ArityType.
904 Actually, the situation is worse. Consider
908 Can we eta-expand here? At first the answer looks like "yes of course", but
911 This should diverge! But if we eta-expand, it won't. Again, we ignore this
912 "problem", because being scrupulous would lose an important transformation for
918 Non-recursive newtypes are transparent, and should not get in the way.
919 We do (currently) eta-expand recursive newtypes too. So if we have, say
921 newtype T = MkT ([T] -> Int)
925 where f has arity 1. Then: etaExpandArity e = 1;
926 that is, etaExpandArity looks through the coerce.
928 When we eta-expand e to arity 1: eta_expand 1 e T
929 we want to get: coerce T (\x::[T] -> (coerce ([T]->Int) e) x)
931 HOWEVER, note that if you use coerce bogusly you can ge
933 And since negate has arity 2, you might try to eta expand. But you can't
934 decopose Int to a function type. Hence the final case in eta_expand.
938 exprEtaExpandArity dflags e = arityDepth (arityType dflags e)
940 -- A limited sort of function type
941 data ArityType = AFun Bool ArityType -- True <=> one-shot
942 | ATop -- Know nothing
945 arityDepth :: ArityType -> Arity
946 arityDepth (AFun _ ty) = 1 + arityDepth ty
949 andArityType ABot at2 = at2
950 andArityType ATop at2 = ATop
951 andArityType (AFun t1 at1) (AFun t2 at2) = AFun (t1 && t2) (andArityType at1 at2)
952 andArityType at1 at2 = andArityType at2 at1
954 arityType :: DynFlags -> CoreExpr -> ArityType
955 -- (go1 e) = [b1,..,bn]
956 -- means expression can be rewritten \x_b1 -> ... \x_bn -> body
957 -- where bi is True <=> the lambda is one-shot
959 arityType dflags (Note n e) = arityType dflags e
960 -- Not needed any more: etaExpand is cleverer
961 -- | ok_note n = arityType dflags e
962 -- | otherwise = ATop
964 arityType dflags (Cast e co) = arityType dflags e
966 arityType dflags (Var v)
967 = mk (idArity v) (arg_tys (idType v))
969 mk :: Arity -> [Type] -> ArityType
970 -- The argument types are only to steer the "state hack"
971 -- Consider case x of
973 -- False -> \(s:RealWorld) -> e
974 -- where foo has arity 1. Then we want the state hack to
975 -- apply to foo too, so we can eta expand the case.
976 mk 0 tys | isBottomingId v = ABot
977 | (ty:tys) <- tys, isStateHackType ty = AFun True ATop
979 mk n (ty:tys) = AFun (isStateHackType ty) (mk (n-1) tys)
980 mk n [] = AFun False (mk (n-1) [])
982 arg_tys :: Type -> [Type] -- Ignore for-alls
984 | Just (_, ty') <- splitForAllTy_maybe ty = arg_tys ty'
985 | Just (arg,res) <- splitFunTy_maybe ty = arg : arg_tys res
988 -- Lambdas; increase arity
989 arityType dflags (Lam x e)
990 | isId x = AFun (isOneShotBndr x) (arityType dflags e)
991 | otherwise = arityType dflags e
993 -- Applications; decrease arity
994 arityType dflags (App f (Type _)) = arityType dflags f
995 arityType dflags (App f a) = case arityType dflags f of
996 AFun one_shot xs | exprIsCheap a -> xs
999 -- Case/Let; keep arity if either the expression is cheap
1000 -- or it's a 1-shot lambda
1001 -- The former is not really right for Haskell
1002 -- f x = case x of { (a,b) -> \y. e }
1004 -- f x y = case x of { (a,b) -> e }
1005 -- The difference is observable using 'seq'
1006 arityType dflags (Case scrut _ _ alts)
1007 = case foldr1 andArityType [arityType dflags rhs | (_,_,rhs) <- alts] of
1008 xs | exprIsCheap scrut -> xs
1009 xs@(AFun one_shot _) | one_shot -> AFun True ATop
1012 arityType dflags (Let b e)
1013 = case arityType dflags e of
1014 xs | cheap_bind b -> xs
1015 xs@(AFun one_shot _) | one_shot -> AFun True ATop
1018 cheap_bind (NonRec b e) = is_cheap (b,e)
1019 cheap_bind (Rec prs) = all is_cheap prs
1020 is_cheap (b,e) = (dopt Opt_DictsCheap dflags && isDictId b)
1022 -- If the experimental -fdicts-cheap flag is on, we eta-expand through
1023 -- dictionary bindings. This improves arities. Thereby, it also
1024 -- means that full laziness is less prone to floating out the
1025 -- application of a function to its dictionary arguments, which
1026 -- can thereby lose opportunities for fusion. Example:
1027 -- foo :: Ord a => a -> ...
1028 -- foo = /\a \(d:Ord a). let d' = ...d... in \(x:a). ....
1029 -- -- So foo has arity 1
1031 -- f = \x. foo dInt $ bar x
1033 -- The (foo DInt) is floated out, and makes ineffective a RULE
1034 -- foo (bar x) = ...
1036 -- One could go further and make exprIsCheap reply True to any
1037 -- dictionary-typed expression, but that's more work.
1039 arityType dflags other = ATop
1041 {- NOT NEEDED ANY MORE: etaExpand is cleverer
1042 ok_note InlineMe = False
1043 ok_note other = True
1044 -- Notice that we do not look through __inline_me__
1045 -- This may seem surprising, but consider
1046 -- f = _inline_me (\x -> e)
1047 -- We DO NOT want to eta expand this to
1048 -- f = \x -> (_inline_me (\x -> e)) x
1049 -- because the _inline_me gets dropped now it is applied,
1058 etaExpand :: Arity -- Result should have this number of value args
1060 -> CoreExpr -> Type -- Expression and its type
1062 -- (etaExpand n us e ty) returns an expression with
1063 -- the same meaning as 'e', but with arity 'n'.
1065 -- Given e' = etaExpand n us e ty
1067 -- ty = exprType e = exprType e'
1069 -- Note that SCCs are not treated specially. If we have
1070 -- etaExpand 2 (\x -> scc "foo" e)
1071 -- = (\xy -> (scc "foo" e) y)
1072 -- So the costs of evaluating 'e' (not 'e y') are attributed to "foo"
1074 etaExpand n us expr ty
1075 | manifestArity expr >= n = expr -- The no-op case
1077 = eta_expand n us expr ty
1080 -- manifestArity sees how many leading value lambdas there are
1081 manifestArity :: CoreExpr -> Arity
1082 manifestArity (Lam v e) | isId v = 1 + manifestArity e
1083 | otherwise = manifestArity e
1084 manifestArity (Note _ e) = manifestArity e
1085 manifestArity (Cast e _) = manifestArity e
1088 -- etaExpand deals with for-alls. For example:
1090 -- where E :: forall a. a -> a
1092 -- (/\b. \y::a -> E b y)
1094 -- It deals with coerces too, though they are now rare
1095 -- so perhaps the extra code isn't worth it
1097 eta_expand n us expr ty
1099 -- The ILX code generator requires eta expansion for type arguments
1100 -- too, but alas the 'n' doesn't tell us how many of them there
1101 -- may be. So we eagerly eta expand any big lambdas, and just
1102 -- cross our fingers about possible loss of sharing in the ILX case.
1103 -- The Right Thing is probably to make 'arity' include
1104 -- type variables throughout the compiler. (ToDo.)
1106 -- Saturated, so nothing to do
1109 -- Short cut for the case where there already
1110 -- is a lambda; no point in gratuitously adding more
1111 eta_expand n us (Lam v body) ty
1113 = Lam v (eta_expand n us body (applyTy ty (mkTyVarTy v)))
1116 = Lam v (eta_expand (n-1) us body (funResultTy ty))
1118 -- We used to have a special case that stepped inside Coerces here,
1119 -- thus: eta_expand n us (Note note@(Coerce _ ty) e) _
1120 -- = Note note (eta_expand n us e ty)
1121 -- BUT this led to an infinite loop
1122 -- Example: newtype T = MkT (Int -> Int)
1123 -- eta_expand 1 (coerce (Int->Int) e)
1124 -- --> coerce (Int->Int) (eta_expand 1 T e)
1126 -- --> coerce (Int->Int) (coerce T
1127 -- (\x::Int -> eta_expand 1 (coerce (Int->Int) e)))
1128 -- by the splitNewType_maybe case below
1131 eta_expand n us expr ty
1132 = ASSERT2 (exprType expr `coreEqType` ty, ppr (exprType expr) $$ ppr ty)
1133 case splitForAllTy_maybe ty of {
1136 Lam lam_tv (eta_expand n us2 (App expr (Type (mkTyVarTy lam_tv))) (substTyWith [tv] [mkTyVarTy lam_tv] ty'))
1138 lam_tv = setVarName tv (mkSysTvName uniq FSLIT("etaT"))
1139 -- Using tv as a base retains its tyvar/covar-ness
1143 case splitFunTy_maybe ty of {
1144 Just (arg_ty, res_ty) -> Lam arg1 (eta_expand (n-1) us2 (App expr (Var arg1)) res_ty)
1146 arg1 = mkSysLocal FSLIT("eta") uniq arg_ty
1152 -- newtype T = MkT ([T] -> Int)
1153 -- Consider eta-expanding this
1156 -- coerce T (\x::[T] -> (coerce ([T]->Int) e) x)
1158 case splitNewTypeRepCo_maybe ty of {
1160 mkCoerce (mkSymCoercion co) (eta_expand n us (mkCoerce co expr) ty1) ;
1163 -- We have an expression of arity > 0, but its type isn't a function
1164 -- This *can* legitmately happen: e.g. coerce Int (\x. x)
1165 -- Essentially the programmer is playing fast and loose with types
1166 -- (Happy does this a lot). So we simply decline to eta-expand.
1171 exprArity is a cheap-and-cheerful version of exprEtaExpandArity.
1172 It tells how many things the expression can be applied to before doing
1173 any work. It doesn't look inside cases, lets, etc. The idea is that
1174 exprEtaExpandArity will do the hard work, leaving something that's easy
1175 for exprArity to grapple with. In particular, Simplify uses exprArity to
1176 compute the ArityInfo for the Id.
1178 Originally I thought that it was enough just to look for top-level lambdas, but
1179 it isn't. I've seen this
1181 foo = PrelBase.timesInt
1183 We want foo to get arity 2 even though the eta-expander will leave it
1184 unchanged, in the expectation that it'll be inlined. But occasionally it
1185 isn't, because foo is blacklisted (used in a rule).
1187 Similarly, see the ok_note check in exprEtaExpandArity. So
1188 f = __inline_me (\x -> e)
1189 won't be eta-expanded.
1191 And in any case it seems more robust to have exprArity be a bit more intelligent.
1192 But note that (\x y z -> f x y z)
1193 should have arity 3, regardless of f's arity.
1196 exprArity :: CoreExpr -> Arity
1199 go (Var v) = idArity v
1200 go (Lam x e) | isId x = go e + 1
1202 go (Note n e) = go e
1203 go (Cast e _) = go e
1204 go (App e (Type t)) = go e
1205 go (App f a) | exprIsCheap a = (go f - 1) `max` 0
1206 -- NB: exprIsCheap a!
1207 -- f (fac x) does not have arity 2,
1208 -- even if f has arity 3!
1209 -- NB: `max 0`! (\x y -> f x) has arity 2, even if f is
1210 -- unknown, hence arity 0
1214 %************************************************************************
1216 \subsection{Equality}
1218 %************************************************************************
1220 @cheapEqExpr@ is a cheap equality test which bales out fast!
1221 True => definitely equal
1222 False => may or may not be equal
1225 cheapEqExpr :: Expr b -> Expr b -> Bool
1227 cheapEqExpr (Var v1) (Var v2) = v1==v2
1228 cheapEqExpr (Lit lit1) (Lit lit2) = lit1 == lit2
1229 cheapEqExpr (Type t1) (Type t2) = t1 `coreEqType` t2
1231 cheapEqExpr (App f1 a1) (App f2 a2)
1232 = f1 `cheapEqExpr` f2 && a1 `cheapEqExpr` a2
1234 cheapEqExpr _ _ = False
1236 exprIsBig :: Expr b -> Bool
1237 -- Returns True of expressions that are too big to be compared by cheapEqExpr
1238 exprIsBig (Lit _) = False
1239 exprIsBig (Var v) = False
1240 exprIsBig (Type t) = False
1241 exprIsBig (App f a) = exprIsBig f || exprIsBig a
1242 exprIsBig (Cast e _) = exprIsBig e -- Hopefully coercions are not too big!
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 -- We must be careful that \x.x and \y.y map to the same hash code,
1354 -- (at least if we want the above invariant to be true)
1356 hashExpr e = fromIntegral (hash_expr (1,emptyVarEnv) e .&. 0x7fffffff)
1357 -- UniqFM doesn't like negative Ints
1359 type HashEnv = (Int, VarEnv Int) -- Hash code for bound variables
1361 hash_expr :: HashEnv -> CoreExpr -> Word32
1362 -- Word32, because we're expecting overflows here, and overflowing
1363 -- signed types just isn't cool. In C it's even undefined.
1364 hash_expr env (Note _ e) = hash_expr env e
1365 hash_expr env (Cast e co) = hash_expr env e
1366 hash_expr env (Var v) = hashVar env v
1367 hash_expr env (Lit lit) = fromIntegral (hashLiteral lit)
1368 hash_expr env (App f e) = hash_expr env f * fast_hash_expr env e
1369 hash_expr env (Let (NonRec b r) e) = hash_expr (extend_env env b) e * fast_hash_expr env r
1370 hash_expr env (Let (Rec ((b,r):_)) e) = hash_expr (extend_env env b) e
1371 hash_expr env (Case e _ _ _) = hash_expr env e
1372 hash_expr env (Lam b e) = hash_expr (extend_env env b) e
1373 hash_expr env (Type t) = WARN(True, text "hash_expr: type") 1
1374 -- Shouldn't happen. Better to use WARN than trace, because trace
1375 -- prevents the CPR optimisation kicking in for hash_expr.
1377 fast_hash_expr env (Var v) = hashVar env v
1378 fast_hash_expr env (Type t) = fast_hash_type env t
1379 fast_hash_expr env (Lit lit) = fromIntegral (hashLiteral lit)
1380 fast_hash_expr env (Cast e co) = fast_hash_expr env e
1381 fast_hash_expr env (Note n e) = fast_hash_expr env e
1382 fast_hash_expr env (App f a) = fast_hash_expr env a -- A bit idiosyncratic ('a' not 'f')!
1383 fast_hash_expr env other = 1
1385 fast_hash_type :: HashEnv -> Type -> Word32
1386 fast_hash_type env ty
1387 | Just tv <- getTyVar_maybe ty = hashVar env tv
1388 | Just (tc,_) <- splitTyConApp_maybe ty
1389 = fromIntegral (hashName (tyConName tc))
1392 extend_env :: HashEnv -> Var -> (Int, VarEnv Int)
1393 extend_env (n,env) b = (n+1, extendVarEnv env b n)
1395 hashVar :: HashEnv -> Var -> Word32
1397 = fromIntegral (lookupVarEnv env v `orElse` hashName (idName v))
1400 %************************************************************************
1402 \subsection{Determining non-updatable right-hand-sides}
1404 %************************************************************************
1406 Top-level constructor applications can usually be allocated
1407 statically, but they can't if the constructor, or any of the
1408 arguments, come from another DLL (because we can't refer to static
1409 labels in other DLLs).
1411 If this happens we simply make the RHS into an updatable thunk,
1412 and 'exectute' it rather than allocating it statically.
1415 rhsIsStatic :: PackageId -> CoreExpr -> Bool
1416 -- This function is called only on *top-level* right-hand sides
1417 -- Returns True if the RHS can be allocated statically, with
1418 -- no thunks involved at all.
1420 -- It's called (i) in TidyPgm.hasCafRefs to decide if the rhs is, or
1421 -- refers to, CAFs; and (ii) in CoreToStg to decide whether to put an
1422 -- update flag on it.
1424 -- The basic idea is that rhsIsStatic returns True only if the RHS is
1425 -- (a) a value lambda
1426 -- (b) a saturated constructor application with static args
1428 -- BUT watch out for
1429 -- (i) Any cross-DLL references kill static-ness completely
1430 -- because they must be 'executed' not statically allocated
1431 -- ("DLL" here really only refers to Windows DLLs, on other platforms,
1432 -- this is not necessary)
1434 -- (ii) We treat partial applications as redexes, because in fact we
1435 -- make a thunk for them that runs and builds a PAP
1436 -- at run-time. The only appliations that are treated as
1437 -- static are *saturated* applications of constructors.
1439 -- We used to try to be clever with nested structures like this:
1440 -- ys = (:) w ((:) w [])
1441 -- on the grounds that CorePrep will flatten ANF-ise it later.
1442 -- But supporting this special case made the function much more
1443 -- complicated, because the special case only applies if there are no
1444 -- enclosing type lambdas:
1445 -- ys = /\ a -> Foo (Baz ([] a))
1446 -- Here the nested (Baz []) won't float out to top level in CorePrep.
1448 -- But in fact, even without -O, nested structures at top level are
1449 -- flattened by the simplifier, so we don't need to be super-clever here.
1453 -- f = \x::Int. x+7 TRUE
1454 -- p = (True,False) TRUE
1456 -- d = (fst p, False) FALSE because there's a redex inside
1457 -- (this particular one doesn't happen but...)
1459 -- h = D# (1.0## /## 2.0##) FALSE (redex again)
1460 -- n = /\a. Nil a TRUE
1462 -- t = /\a. (:) (case w a of ...) (Nil a) FALSE (redex)
1465 -- This is a bit like CoreUtils.exprIsHNF, with the following differences:
1466 -- a) scc "foo" (\x -> ...) is updatable (so we catch the right SCC)
1468 -- b) (C x xs), where C is a contructors is updatable if the application is
1471 -- c) don't look through unfolding of f in (f x).
1473 -- When opt_RuntimeTypes is on, we keep type lambdas and treat
1474 -- them as making the RHS re-entrant (non-updatable).
1476 rhsIsStatic this_pkg rhs = is_static False rhs
1478 is_static :: Bool -- True <=> in a constructor argument; must be atomic
1481 is_static False (Lam b e) = isRuntimeVar b || is_static False e
1483 is_static in_arg (Note (SCC _) e) = False
1484 is_static in_arg (Note _ e) = is_static in_arg e
1485 is_static in_arg (Cast e co) = is_static in_arg e
1487 is_static in_arg (Lit lit)
1489 MachLabel _ _ -> False
1491 -- A MachLabel (foreign import "&foo") in an argument
1492 -- prevents a constructor application from being static. The
1493 -- reason is that it might give rise to unresolvable symbols
1494 -- in the object file: under Linux, references to "weak"
1495 -- symbols from the data segment give rise to "unresolvable
1496 -- relocation" errors at link time This might be due to a bug
1497 -- in the linker, but we'll work around it here anyway.
1500 is_static in_arg other_expr = go other_expr 0
1502 go (Var f) n_val_args
1503 #if mingw32_TARGET_OS
1504 | not (isDllName this_pkg (idName f))
1506 = saturated_data_con f n_val_args
1507 || (in_arg && n_val_args == 0)
1508 -- A naked un-applied variable is *not* deemed a static RHS
1510 -- Reason: better to update so that the indirection gets shorted
1511 -- out, and the true value will be seen
1512 -- NB: if you change this, you'll break the invariant that THUNK_STATICs
1513 -- are always updatable. If you do so, make sure that non-updatable
1514 -- ones have enough space for their static link field!
1516 go (App f a) n_val_args
1517 | isTypeArg a = go f n_val_args
1518 | not in_arg && is_static True a = go f (n_val_args + 1)
1519 -- The (not in_arg) checks that we aren't in a constructor argument;
1520 -- if we are, we don't allow (value) applications of any sort
1522 -- NB. In case you wonder, args are sometimes not atomic. eg.
1523 -- x = D# (1.0## /## 2.0##)
1524 -- can't float because /## can fail.
1526 go (Note (SCC _) f) n_val_args = False
1527 go (Note _ f) n_val_args = go f n_val_args
1528 go (Cast e co) n_val_args = go e n_val_args
1530 go other n_val_args = False
1532 saturated_data_con f n_val_args
1533 = case isDataConWorkId_maybe f of
1534 Just dc -> n_val_args == dataConRepArity dc