2 % (c) The GRASP/AQUA Project, Glasgow University, 1992-1998
4 \section[CoreUtils]{Utility functions on @Core@ syntax}
9 mkInlineMe, mkSCC, mkCoerce, mkCoerce2,
10 bindNonRec, needsCaseBinding,
11 mkIfThenElse, mkAltExpr, mkPiType, mkPiTypes,
13 -- Taking expressions apart
14 findDefault, findAlt, isDefaultAlt,
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
37 #include "HsVersions.h"
40 import GLAEXTS -- For `xori`
43 import CoreFVs ( exprFreeVars )
44 import PprCore ( pprCoreExpr )
46 import VarSet ( unionVarSet )
48 import Name ( hashName )
49 import Packages ( HomeModules )
51 import Packages ( isDllName )
53 import Literal ( hashLiteral, literalType, litIsDupable,
54 litIsTrivial, isZeroLit, Literal( MachLabel ) )
55 import DataCon ( DataCon, dataConRepArity, dataConInstArgTys,
56 isVanillaDataCon, dataConTyCon )
57 import PrimOp ( PrimOp(..), primOpOkForSpeculation, primOpIsCheap )
58 import Id ( Id, idType, globalIdDetails, idNewStrictness,
59 mkWildId, idArity, idName, idUnfolding, idInfo,
60 isOneShotBndr, isStateHackType, isDataConWorkId_maybe, mkSysLocal,
61 isDataConWorkId, isBottomingId
63 import IdInfo ( GlobalIdDetails(..), megaSeqIdInfo )
64 import NewDemand ( appIsBottom )
65 import Type ( Type, mkFunTy, mkForAllTy, splitFunTy_maybe,
66 splitFunTy, tcEqTypeX,
67 applyTys, isUnLiftedType, seqType, mkTyVarTy,
68 splitForAllTy_maybe, isForAllTy, splitRecNewType_maybe,
69 splitTyConApp_maybe, coreEqType, funResultTy, applyTy
71 import TyCon ( tyConArity )
72 import TysWiredIn ( boolTy, trueDataCon, falseDataCon )
73 import CostCentre ( CostCentre )
74 import BasicTypes ( Arity )
75 import Unique ( Unique )
77 import TysPrim ( alphaTy ) -- Debugging only
78 import Util ( equalLength, lengthAtLeast, foldl2 )
82 %************************************************************************
84 \subsection{Find the type of a Core atom/expression}
86 %************************************************************************
89 exprType :: CoreExpr -> Type
91 exprType (Var var) = idType var
92 exprType (Lit lit) = literalType lit
93 exprType (Let _ body) = exprType body
94 exprType (Case _ _ ty alts) = ty
95 exprType (Note (Coerce ty _) e) = ty -- **! should take usage from e
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)
152 %************************************************************************
154 \subsection{Attaching notes}
156 %************************************************************************
158 mkNote removes redundant coercions, and SCCs where possible
162 mkNote :: Note -> CoreExpr -> CoreExpr
163 mkNote (Coerce to_ty from_ty) expr = mkCoerce2 to_ty from_ty expr
164 mkNote (SCC cc) expr = mkSCC cc expr
165 mkNote InlineMe expr = mkInlineMe expr
166 mkNote note expr = Note note expr
169 -- Slide InlineCall in around the function
170 -- No longer necessary I think (SLPJ Apr 99)
171 -- mkNote InlineCall (App f a) = App (mkNote InlineCall f) a
172 -- mkNote InlineCall (Var v) = Note InlineCall (Var v)
173 -- mkNote InlineCall expr = expr
176 Drop trivial InlineMe's. This is somewhat important, because if we have an unfolding
177 that looks like (Note InlineMe (Var v)), the InlineMe doesn't go away because it may
178 not be *applied* to anything.
180 We don't use exprIsTrivial here, though, because we sometimes generate worker/wrapper
183 f = inline_me (coerce t fw)
184 As usual, the inline_me prevents the worker from getting inlined back into the wrapper.
185 We want the split, so that the coerces can cancel at the call site.
187 However, we can get left with tiresome type applications. Notably, consider
188 f = /\ a -> let t = e in (t, w)
189 Then lifting the let out of the big lambda gives
191 f = /\ a -> let t = inline_me (t' a) in (t, w)
192 The inline_me is to stop the simplifier inlining t' right back
193 into t's RHS. In the next phase we'll substitute for t (since
194 its rhs is trivial) and *then* we could get rid of the inline_me.
195 But it hardly seems worth it, so I don't bother.
198 mkInlineMe (Var v) = Var v
199 mkInlineMe e = Note InlineMe e
205 mkCoerce :: Type -> CoreExpr -> CoreExpr
206 mkCoerce to_ty expr = mkCoerce2 to_ty (exprType expr) expr
208 mkCoerce2 :: Type -> Type -> CoreExpr -> CoreExpr
209 mkCoerce2 to_ty from_ty (Note (Coerce to_ty2 from_ty2) expr)
210 = ASSERT( from_ty `coreEqType` to_ty2 )
211 mkCoerce2 to_ty from_ty2 expr
213 mkCoerce2 to_ty from_ty expr
214 | to_ty `coreEqType` from_ty = expr
215 | otherwise = ASSERT( from_ty `coreEqType` exprType expr )
216 Note (Coerce to_ty from_ty) expr
220 mkSCC :: CostCentre -> Expr b -> Expr b
221 -- Note: Nested SCC's *are* preserved for the benefit of
222 -- cost centre stack profiling
223 mkSCC cc (Lit lit) = Lit lit
224 mkSCC cc (Lam x e) = Lam x (mkSCC cc e) -- Move _scc_ inside lambda
225 mkSCC cc (Note (SCC cc') e) = Note (SCC cc) (Note (SCC cc') e)
226 mkSCC cc (Note n e) = Note n (mkSCC cc e) -- Move _scc_ inside notes
227 mkSCC cc expr = Note (SCC cc) expr
231 %************************************************************************
233 \subsection{Other expression construction}
235 %************************************************************************
238 bindNonRec :: Id -> CoreExpr -> CoreExpr -> CoreExpr
239 -- (bindNonRec x r b) produces either
242 -- case r of x { _DEFAULT_ -> b }
244 -- depending on whether x is unlifted or not
245 -- It's used by the desugarer to avoid building bindings
246 -- that give Core Lint a heart attack. Actually the simplifier
247 -- deals with them perfectly well.
249 bindNonRec bndr rhs body
250 | needsCaseBinding (idType bndr) rhs = Case rhs bndr (exprType body) [(DEFAULT,[],body)]
251 | otherwise = Let (NonRec bndr rhs) body
253 needsCaseBinding ty rhs = isUnLiftedType ty && not (exprOkForSpeculation rhs)
254 -- Make a case expression instead of a let
255 -- These can arise either from the desugarer,
256 -- or from beta reductions: (\x.e) (x +# y)
260 mkAltExpr :: AltCon -> [CoreBndr] -> [Type] -> CoreExpr
261 -- This guy constructs the value that the scrutinee must have
262 -- when you are in one particular branch of a case
263 mkAltExpr (DataAlt con) args inst_tys
264 = mkConApp con (map Type inst_tys ++ map varToCoreExpr args)
265 mkAltExpr (LitAlt lit) [] []
268 mkIfThenElse :: CoreExpr -> CoreExpr -> CoreExpr -> CoreExpr
269 mkIfThenElse guard then_expr else_expr
270 -- Not going to be refining, so okay to take the type of the "then" clause
271 = Case guard (mkWildId boolTy) (exprType then_expr)
272 [ (DataAlt falseDataCon, [], else_expr), -- Increasing order of tag!
273 (DataAlt trueDataCon, [], then_expr) ]
277 %************************************************************************
279 \subsection{Taking expressions apart}
281 %************************************************************************
283 The default alternative must be first, if it exists at all.
284 This makes it easy to find, though it makes matching marginally harder.
287 findDefault :: [CoreAlt] -> ([CoreAlt], Maybe CoreExpr)
288 findDefault ((DEFAULT,args,rhs) : alts) = ASSERT( null args ) (alts, Just rhs)
289 findDefault alts = (alts, Nothing)
291 findAlt :: AltCon -> [CoreAlt] -> CoreAlt
294 (deflt@(DEFAULT,_,_):alts) -> go alts deflt
295 other -> go alts panic_deflt
297 panic_deflt = pprPanic "Missing alternative" (ppr con $$ vcat (map ppr alts))
300 go (alt@(con1,_,_) : alts) deflt
301 = case con `cmpAltCon` con1 of
302 LT -> deflt -- Missed it already; the alts are in increasing order
304 GT -> ASSERT( not (con1 == DEFAULT) ) go alts deflt
306 isDefaultAlt :: CoreAlt -> Bool
307 isDefaultAlt (DEFAULT, _, _) = True
308 isDefaultAlt other = False
312 %************************************************************************
314 \subsection{Figuring out things about expressions}
316 %************************************************************************
318 @exprIsTrivial@ is true of expressions we are unconditionally happy to
319 duplicate; simple variables and constants, and type
320 applications. Note that primop Ids aren't considered
323 @exprIsBottom@ is true of expressions that are guaranteed to diverge
326 There used to be a gruesome test for (hasNoBinding v) in the
328 exprIsTrivial (Var v) | hasNoBinding v = idArity v == 0
329 The idea here is that a constructor worker, like $wJust, is
330 really short for (\x -> $wJust x), becuase $wJust has no binding.
331 So it should be treated like a lambda. Ditto unsaturated primops.
332 But now constructor workers are not "have-no-binding" Ids. And
333 completely un-applied primops and foreign-call Ids are sufficiently
334 rare that I plan to allow them to be duplicated and put up with
337 SCC notes. We do not treat (_scc_ "foo" x) as trivial, because
338 a) it really generates code, (and a heap object when it's
339 a function arg) to capture the cost centre
340 b) see the note [SCC-and-exprIsTrivial] in Simplify.simplLazyBind
343 exprIsTrivial (Var v) = True -- See notes above
344 exprIsTrivial (Type _) = True
345 exprIsTrivial (Lit lit) = litIsTrivial lit
346 exprIsTrivial (App e arg) = not (isRuntimeArg arg) && exprIsTrivial e
347 exprIsTrivial (Note (SCC _) e) = False -- See notes above
348 exprIsTrivial (Note _ e) = exprIsTrivial e
349 exprIsTrivial (Lam b body) = not (isRuntimeVar b) && exprIsTrivial body
350 exprIsTrivial other = False
354 @exprIsDupable@ is true of expressions that can be duplicated at a modest
355 cost in code size. This will only happen in different case
356 branches, so there's no issue about duplicating work.
358 That is, exprIsDupable returns True of (f x) even if
359 f is very very expensive to call.
361 Its only purpose is to avoid fruitless let-binding
362 and then inlining of case join points
366 exprIsDupable (Type _) = True
367 exprIsDupable (Var v) = True
368 exprIsDupable (Lit lit) = litIsDupable lit
369 exprIsDupable (Note InlineMe e) = True
370 exprIsDupable (Note _ e) = exprIsDupable e
374 go (Var v) n_args = True
375 go (App f a) n_args = n_args < dupAppSize
378 go other n_args = False
381 dupAppSize = 4 -- Size of application we are prepared to duplicate
384 @exprIsCheap@ looks at a Core expression and returns \tr{True} if
385 it is obviously in weak head normal form, or is cheap to get to WHNF.
386 [Note that that's not the same as exprIsDupable; an expression might be
387 big, and hence not dupable, but still cheap.]
389 By ``cheap'' we mean a computation we're willing to:
390 push inside a lambda, or
391 inline at more than one place
392 That might mean it gets evaluated more than once, instead of being
393 shared. The main examples of things which aren't WHNF but are
398 (where e, and all the ei are cheap)
401 (where e and b are cheap)
404 (where op is a cheap primitive operator)
407 (because we are happy to substitute it inside a lambda)
409 Notice that a variable is considered 'cheap': we can push it inside a lambda,
410 because sharing will make sure it is only evaluated once.
413 exprIsCheap :: CoreExpr -> Bool
414 exprIsCheap (Lit lit) = True
415 exprIsCheap (Type _) = True
416 exprIsCheap (Var _) = True
417 exprIsCheap (Note InlineMe e) = True
418 exprIsCheap (Note _ e) = exprIsCheap e
419 exprIsCheap (Lam x e) = isRuntimeVar x || exprIsCheap e
420 exprIsCheap (Case e _ _ alts) = exprIsCheap e &&
421 and [exprIsCheap rhs | (_,_,rhs) <- alts]
422 -- Experimentally, treat (case x of ...) as cheap
423 -- (and case __coerce x etc.)
424 -- This improves arities of overloaded functions where
425 -- there is only dictionary selection (no construction) involved
426 exprIsCheap (Let (NonRec x _) e)
427 | isUnLiftedType (idType x) = exprIsCheap e
429 -- strict lets always have cheap right hand sides, and
432 exprIsCheap other_expr
433 = go other_expr 0 True
435 go (Var f) n_args args_cheap
436 = (idAppIsCheap f n_args && args_cheap)
437 -- A constructor, cheap primop, or partial application
439 || idAppIsBottom f n_args
440 -- Application of a function which
441 -- always gives bottom; we treat this as cheap
442 -- because it certainly doesn't need to be shared!
444 go (App f a) n_args args_cheap
445 | not (isRuntimeArg a) = go f n_args args_cheap
446 | otherwise = go f (n_args + 1) (exprIsCheap a && args_cheap)
448 go other n_args args_cheap = False
450 idAppIsCheap :: Id -> Int -> Bool
451 idAppIsCheap id n_val_args
452 | n_val_args == 0 = True -- Just a type application of
453 -- a variable (f t1 t2 t3)
456 = case globalIdDetails id of
457 DataConWorkId _ -> True
458 RecordSelId {} -> n_val_args == 1 -- I'm experimenting with making record selection
459 ClassOpId _ -> n_val_args == 1 -- look cheap, so we will substitute it inside a
460 -- lambda. Particularly for dictionary field selection.
461 -- BUT: Take care with (sel d x)! The (sel d) might be cheap, but
462 -- there's no guarantee that (sel d x) will be too. Hence (n_val_args == 1)
464 PrimOpId op -> primOpIsCheap op -- In principle we should worry about primops
465 -- that return a type variable, since the result
466 -- might be applied to something, but I'm not going
467 -- to bother to check the number of args
468 other -> n_val_args < idArity id
471 exprOkForSpeculation returns True of an expression that it is
473 * safe to evaluate even if normal order eval might not
474 evaluate the expression at all, or
476 * safe *not* to evaluate even if normal order would do so
480 the expression guarantees to terminate,
482 without raising an exception,
483 without causing a side effect (e.g. writing a mutable variable)
486 let x = case y# +# 1# of { r# -> I# r# }
489 case y# +# 1# of { r# ->
494 We can only do this if the (y+1) is ok for speculation: it has no
495 side effects, and can't diverge or raise an exception.
498 exprOkForSpeculation :: CoreExpr -> Bool
499 exprOkForSpeculation (Lit _) = True
500 exprOkForSpeculation (Type _) = True
501 exprOkForSpeculation (Var v) = isUnLiftedType (idType v)
502 exprOkForSpeculation (Note _ e) = exprOkForSpeculation e
503 exprOkForSpeculation other_expr
504 = case collectArgs other_expr of
505 (Var f, args) -> spec_ok (globalIdDetails f) args
509 spec_ok (DataConWorkId _) args
510 = True -- The strictness of the constructor has already
511 -- been expressed by its "wrapper", so we don't need
512 -- to take the arguments into account
514 spec_ok (PrimOpId op) args
515 | isDivOp op, -- Special case for dividing operations that fail
516 [arg1, Lit lit] <- args -- only if the divisor is zero
517 = not (isZeroLit lit) && exprOkForSpeculation arg1
518 -- Often there is a literal divisor, and this
519 -- can get rid of a thunk in an inner looop
522 = primOpOkForSpeculation op &&
523 all exprOkForSpeculation args
524 -- A bit conservative: we don't really need
525 -- to care about lazy arguments, but this is easy
527 spec_ok other args = False
529 isDivOp :: PrimOp -> Bool
530 -- True of dyadic operators that can fail
531 -- only if the second arg is zero
532 -- This function probably belongs in PrimOp, or even in
533 -- an automagically generated file.. but it's such a
534 -- special case I thought I'd leave it here for now.
535 isDivOp IntQuotOp = True
536 isDivOp IntRemOp = True
537 isDivOp WordQuotOp = True
538 isDivOp WordRemOp = True
539 isDivOp IntegerQuotRemOp = True
540 isDivOp IntegerDivModOp = True
541 isDivOp FloatDivOp = True
542 isDivOp DoubleDivOp = True
543 isDivOp other = False
548 exprIsBottom :: CoreExpr -> Bool -- True => definitely bottom
549 exprIsBottom e = go 0 e
551 -- n is the number of args
552 go n (Note _ e) = go n e
553 go n (Let _ e) = go n e
554 go n (Case e _ _ _) = go 0 e -- Just check the scrut
555 go n (App e _) = go (n+1) e
556 go n (Var v) = idAppIsBottom v n
558 go n (Lam _ _) = False
559 go n (Type _) = False
561 idAppIsBottom :: Id -> Int -> Bool
562 idAppIsBottom id n_val_args = appIsBottom (idNewStrictness id) n_val_args
565 @exprIsHNF@ returns true for expressions that are certainly *already*
566 evaluated to *head* normal form. This is used to decide whether it's ok
569 case x of _ -> e ===> e
571 and to decide whether it's safe to discard a `seq`
573 So, it does *not* treat variables as evaluated, unless they say they are.
575 But it *does* treat partial applications and constructor applications
576 as values, even if their arguments are non-trivial, provided the argument
578 e.g. (:) (f x) (map f xs) is a value
579 map (...redex...) is a value
580 Because `seq` on such things completes immediately
582 For unlifted argument types, we have to be careful:
584 Suppose (f x) diverges; then C (f x) is not a value. True, but
585 this form is illegal (see the invariants in CoreSyn). Args of unboxed
586 type must be ok-for-speculation (or trivial).
589 exprIsHNF :: CoreExpr -> Bool -- True => Value-lambda, constructor, PAP
590 exprIsHNF (Var v) -- NB: There are no value args at this point
591 = isDataConWorkId v -- Catches nullary constructors,
592 -- so that [] and () are values, for example
593 || idArity v > 0 -- Catches (e.g.) primops that don't have unfoldings
594 || isEvaldUnfolding (idUnfolding v)
595 -- Check the thing's unfolding; it might be bound to a value
596 -- A worry: what if an Id's unfolding is just itself:
597 -- then we could get an infinite loop...
599 exprIsHNF (Lit l) = True
600 exprIsHNF (Type ty) = True -- Types are honorary Values;
601 -- we don't mind copying them
602 exprIsHNF (Lam b e) = isRuntimeVar b || exprIsHNF e
603 exprIsHNF (Note _ e) = exprIsHNF e
604 exprIsHNF (App e (Type _)) = exprIsHNF e
605 exprIsHNF (App e a) = app_is_value e [a]
606 exprIsHNF other = False
608 -- There is at least one value argument
609 app_is_value (Var fun) args
610 | isDataConWorkId fun -- Constructor apps are values
611 || idArity fun > valArgCount args -- Under-applied function
612 = check_args (idType fun) args
613 app_is_value (App f a) as = app_is_value f (a:as)
614 app_is_value other as = False
616 -- 'check_args' checks that unlifted-type args
617 -- are in fact guaranteed non-divergent
618 check_args fun_ty [] = True
619 check_args fun_ty (Type _ : args) = case splitForAllTy_maybe fun_ty of
620 Just (_, ty) -> check_args ty args
621 check_args fun_ty (arg : args)
622 | isUnLiftedType arg_ty = exprOkForSpeculation arg
623 | otherwise = check_args res_ty args
625 (arg_ty, res_ty) = splitFunTy fun_ty
629 exprIsConApp_maybe :: CoreExpr -> Maybe (DataCon, [CoreExpr])
630 exprIsConApp_maybe (Note (Coerce to_ty from_ty) expr)
631 = -- Maybe this is over the top, but here we try to turn
632 -- coerce (S,T) ( x, y )
634 -- ( coerce S x, coerce T y )
635 -- This happens in anger in PrelArrExts which has a coerce
636 -- case coerce memcpy a b of
638 -- where the memcpy is in the IO monad, but the call is in
640 case exprIsConApp_maybe expr of {
644 case splitTyConApp_maybe to_ty of {
646 Just (tc, tc_arg_tys) | tc /= dataConTyCon dc -> Nothing
647 | not (isVanillaDataCon dc) -> Nothing
649 -- Type constructor must match
650 -- We knock out existentials to keep matters simple(r)
652 arity = tyConArity tc
653 val_args = drop arity args
654 to_arg_tys = dataConInstArgTys dc tc_arg_tys
655 mk_coerce ty arg = mkCoerce ty arg
656 new_val_args = zipWith mk_coerce to_arg_tys val_args
658 ASSERT( all isTypeArg (take arity args) )
659 ASSERT( equalLength val_args to_arg_tys )
660 Just (dc, map Type tc_arg_tys ++ new_val_args)
663 exprIsConApp_maybe (Note _ expr)
664 = exprIsConApp_maybe expr
665 -- We ignore InlineMe notes in case we have
666 -- x = __inline_me__ (a,b)
667 -- All part of making sure that INLINE pragmas never hurt
668 -- Marcin tripped on this one when making dictionaries more inlinable
670 -- In fact, we ignore all notes. For example,
671 -- case _scc_ "foo" (C a b) of
673 -- should be optimised away, but it will be only if we look
674 -- through the SCC note.
676 exprIsConApp_maybe expr = analyse (collectArgs expr)
678 analyse (Var fun, args)
679 | Just con <- isDataConWorkId_maybe fun,
680 args `lengthAtLeast` dataConRepArity con
681 -- Might be > because the arity excludes type args
684 -- Look through unfoldings, but only cheap ones, because
685 -- we are effectively duplicating the unfolding
686 analyse (Var fun, [])
687 | let unf = idUnfolding fun,
689 = exprIsConApp_maybe (unfoldingTemplate unf)
691 analyse other = Nothing
696 %************************************************************************
698 \subsection{Eta reduction and expansion}
700 %************************************************************************
703 exprEtaExpandArity :: CoreExpr -> Arity
704 {- The Arity returned is the number of value args the
705 thing can be applied to without doing much work
707 exprEtaExpandArity is used when eta expanding
710 It returns 1 (or more) to:
711 case x of p -> \s -> ...
712 because for I/O ish things we really want to get that \s to the top.
713 We are prepared to evaluate x each time round the loop in order to get that
715 It's all a bit more subtle than it looks:
719 Consider one-shot lambdas
720 let x = expensive in \y z -> E
721 We want this to have arity 2 if the \y-abstraction is a 1-shot lambda
722 Hence the ArityType returned by arityType
724 2. The state-transformer hack
726 The one-shot lambda special cause is particularly important/useful for
727 IO state transformers, where we often get
728 let x = E in \ s -> ...
730 and the \s is a real-world state token abstraction. Such abstractions
731 are almost invariably 1-shot, so we want to pull the \s out, past the
732 let x=E, even if E is expensive. So we treat state-token lambdas as
733 one-shot even if they aren't really. The hack is in Id.isOneShotBndr.
735 3. Dealing with bottom
738 f = \x -> error "foo"
739 Here, arity 1 is fine. But if it is
743 then we want to get arity 2. Tecnically, this isn't quite right, because
745 should diverge, but it'll converge if we eta-expand f. Nevertheless, we
746 do so; it improves some programs significantly, and increasing convergence
747 isn't a bad thing. Hence the ABot/ATop in ArityType.
749 Actually, the situation is worse. Consider
753 Can we eta-expand here? At first the answer looks like "yes of course", but
756 This should diverge! But if we eta-expand, it won't. Again, we ignore this
757 "problem", because being scrupulous would lose an important transformation for
763 Non-recursive newtypes are transparent, and should not get in the way.
764 We do (currently) eta-expand recursive newtypes too. So if we have, say
766 newtype T = MkT ([T] -> Int)
770 where f has arity 1. Then: etaExpandArity e = 1;
771 that is, etaExpandArity looks through the coerce.
773 When we eta-expand e to arity 1: eta_expand 1 e T
774 we want to get: coerce T (\x::[T] -> (coerce ([T]->Int) e) x)
776 HOWEVER, note that if you use coerce bogusly you can ge
778 And since negate has arity 2, you might try to eta expand. But you can't
779 decopose Int to a function type. Hence the final case in eta_expand.
783 exprEtaExpandArity e = arityDepth (arityType e)
785 -- A limited sort of function type
786 data ArityType = AFun Bool ArityType -- True <=> one-shot
787 | ATop -- Know nothing
790 arityDepth :: ArityType -> Arity
791 arityDepth (AFun _ ty) = 1 + arityDepth ty
794 andArityType ABot at2 = at2
795 andArityType ATop at2 = ATop
796 andArityType (AFun t1 at1) (AFun t2 at2) = AFun (t1 && t2) (andArityType at1 at2)
797 andArityType at1 at2 = andArityType at2 at1
799 arityType :: CoreExpr -> ArityType
800 -- (go1 e) = [b1,..,bn]
801 -- means expression can be rewritten \x_b1 -> ... \x_bn -> body
802 -- where bi is True <=> the lambda is one-shot
804 arityType (Note n e) = arityType e
805 -- Not needed any more: etaExpand is cleverer
806 -- | ok_note n = arityType e
807 -- | otherwise = ATop
810 = mk (idArity v) (arg_tys (idType v))
812 mk :: Arity -> [Type] -> ArityType
813 -- The argument types are only to steer the "state hack"
814 -- Consider case x of
816 -- False -> \(s:RealWorld) -> e
817 -- where foo has arity 1. Then we want the state hack to
818 -- apply to foo too, so we can eta expand the case.
819 mk 0 tys | isBottomingId v = ABot
821 mk n (ty:tys) = AFun (isStateHackType ty) (mk (n-1) tys)
822 mk n [] = AFun False (mk (n-1) [])
824 arg_tys :: Type -> [Type] -- Ignore for-alls
826 | Just (_, ty') <- splitForAllTy_maybe ty = arg_tys ty'
827 | Just (arg,res) <- splitFunTy_maybe ty = arg : arg_tys res
830 -- Lambdas; increase arity
831 arityType (Lam x e) | isId x = AFun (isOneShotBndr x) (arityType e)
832 | otherwise = arityType e
834 -- Applications; decrease arity
835 arityType (App f (Type _)) = arityType f
836 arityType (App f a) = case arityType f of
837 AFun one_shot xs | exprIsCheap a -> xs
840 -- Case/Let; keep arity if either the expression is cheap
841 -- or it's a 1-shot lambda
842 -- The former is not really right for Haskell
843 -- f x = case x of { (a,b) -> \y. e }
845 -- f x y = case x of { (a,b) -> e }
846 -- The difference is observable using 'seq'
847 arityType (Case scrut _ _ alts) = case foldr1 andArityType [arityType rhs | (_,_,rhs) <- alts] of
848 xs@(AFun one_shot _) | one_shot -> xs
849 xs | exprIsCheap scrut -> xs
852 arityType (Let b e) = case arityType e of
853 xs@(AFun one_shot _) | one_shot -> xs
854 xs | all exprIsCheap (rhssOfBind b) -> xs
857 arityType other = ATop
859 {- NOT NEEDED ANY MORE: etaExpand is cleverer
860 ok_note InlineMe = False
862 -- Notice that we do not look through __inline_me__
863 -- This may seem surprising, but consider
864 -- f = _inline_me (\x -> e)
865 -- We DO NOT want to eta expand this to
866 -- f = \x -> (_inline_me (\x -> e)) x
867 -- because the _inline_me gets dropped now it is applied,
876 etaExpand :: Arity -- Result should have this number of value args
878 -> CoreExpr -> Type -- Expression and its type
880 -- (etaExpand n us e ty) returns an expression with
881 -- the same meaning as 'e', but with arity 'n'.
883 -- Given e' = etaExpand n us e ty
885 -- ty = exprType e = exprType e'
887 -- Note that SCCs are not treated specially. If we have
888 -- etaExpand 2 (\x -> scc "foo" e)
889 -- = (\xy -> (scc "foo" e) y)
890 -- So the costs of evaluating 'e' (not 'e y') are attributed to "foo"
892 etaExpand n us expr ty
893 | manifestArity expr >= n = expr -- The no-op case
894 | otherwise = eta_expand n us expr ty
897 -- manifestArity sees how many leading value lambdas there are
898 manifestArity :: CoreExpr -> Arity
899 manifestArity (Lam v e) | isId v = 1 + manifestArity e
900 | otherwise = manifestArity e
901 manifestArity (Note _ e) = manifestArity e
904 -- etaExpand deals with for-alls. For example:
906 -- where E :: forall a. a -> a
908 -- (/\b. \y::a -> E b y)
910 -- It deals with coerces too, though they are now rare
911 -- so perhaps the extra code isn't worth it
913 eta_expand n us expr ty
915 -- The ILX code generator requires eta expansion for type arguments
916 -- too, but alas the 'n' doesn't tell us how many of them there
917 -- may be. So we eagerly eta expand any big lambdas, and just
918 -- cross our fingers about possible loss of sharing in the ILX case.
919 -- The Right Thing is probably to make 'arity' include
920 -- type variables throughout the compiler. (ToDo.)
922 -- Saturated, so nothing to do
925 -- Short cut for the case where there already
926 -- is a lambda; no point in gratuitously adding more
927 eta_expand n us (Lam v body) ty
929 = Lam v (eta_expand n us body (applyTy ty (mkTyVarTy v)))
932 = Lam v (eta_expand (n-1) us body (funResultTy ty))
934 -- We used to have a special case that stepped inside Coerces here,
935 -- thus: eta_expand n us (Note note@(Coerce _ ty) e) _
936 -- = Note note (eta_expand n us e ty)
937 -- BUT this led to an infinite loop
938 -- Example: newtype T = MkT (Int -> Int)
939 -- eta_expand 1 (coerce (Int->Int) e)
940 -- --> coerce (Int->Int) (eta_expand 1 T e)
942 -- --> coerce (Int->Int) (coerce T
943 -- (\x::Int -> eta_expand 1 (coerce (Int->Int) e)))
944 -- by the splitNewType_maybe case below
947 eta_expand n us expr ty
948 = case splitForAllTy_maybe ty of {
949 Just (tv,ty') -> Lam tv (eta_expand n us (App expr (Type (mkTyVarTy tv))) ty')
953 case splitFunTy_maybe ty of {
954 Just (arg_ty, res_ty) -> Lam arg1 (eta_expand (n-1) us2 (App expr (Var arg1)) res_ty)
956 arg1 = mkSysLocal FSLIT("eta") uniq arg_ty
962 -- newtype T = MkT ([T] -> Int)
963 -- Consider eta-expanding this
966 -- coerce T (\x::[T] -> (coerce ([T]->Int) e) x)
967 -- Only try this for recursive newtypes; the non-recursive kind
968 -- are transparent anyway
970 case splitRecNewType_maybe ty of {
971 Just ty' -> mkCoerce2 ty ty' (eta_expand n us (mkCoerce2 ty' ty expr) ty') ;
974 -- We have an expression of arity > 0, but its type isn't a function
975 -- This *can* legitmately happen: e.g. coerce Int (\x. x)
976 -- Essentially the programmer is playing fast and loose with types
977 -- (Happy does this a lot). So we simply decline to eta-expand.
982 exprArity is a cheap-and-cheerful version of exprEtaExpandArity.
983 It tells how many things the expression can be applied to before doing
984 any work. It doesn't look inside cases, lets, etc. The idea is that
985 exprEtaExpandArity will do the hard work, leaving something that's easy
986 for exprArity to grapple with. In particular, Simplify uses exprArity to
987 compute the ArityInfo for the Id.
989 Originally I thought that it was enough just to look for top-level lambdas, but
990 it isn't. I've seen this
992 foo = PrelBase.timesInt
994 We want foo to get arity 2 even though the eta-expander will leave it
995 unchanged, in the expectation that it'll be inlined. But occasionally it
996 isn't, because foo is blacklisted (used in a rule).
998 Similarly, see the ok_note check in exprEtaExpandArity. So
999 f = __inline_me (\x -> e)
1000 won't be eta-expanded.
1002 And in any case it seems more robust to have exprArity be a bit more intelligent.
1003 But note that (\x y z -> f x y z)
1004 should have arity 3, regardless of f's arity.
1007 exprArity :: CoreExpr -> Arity
1010 go (Var v) = idArity v
1011 go (Lam x e) | isId x = go e + 1
1013 go (Note n e) = go e
1014 go (App e (Type t)) = go e
1015 go (App f a) | exprIsCheap a = (go f - 1) `max` 0
1016 -- NB: exprIsCheap a!
1017 -- f (fac x) does not have arity 2,
1018 -- even if f has arity 3!
1019 -- NB: `max 0`! (\x y -> f x) has arity 2, even if f is
1020 -- unknown, hence arity 0
1024 %************************************************************************
1026 \subsection{Equality}
1028 %************************************************************************
1030 @cheapEqExpr@ is a cheap equality test which bales out fast!
1031 True => definitely equal
1032 False => may or may not be equal
1035 cheapEqExpr :: Expr b -> Expr b -> Bool
1037 cheapEqExpr (Var v1) (Var v2) = v1==v2
1038 cheapEqExpr (Lit lit1) (Lit lit2) = lit1 == lit2
1039 cheapEqExpr (Type t1) (Type t2) = t1 `coreEqType` t2
1041 cheapEqExpr (App f1 a1) (App f2 a2)
1042 = f1 `cheapEqExpr` f2 && a1 `cheapEqExpr` a2
1044 cheapEqExpr _ _ = False
1046 exprIsBig :: Expr b -> Bool
1047 -- Returns True of expressions that are too big to be compared by cheapEqExpr
1048 exprIsBig (Lit _) = False
1049 exprIsBig (Var v) = False
1050 exprIsBig (Type t) = False
1051 exprIsBig (App f a) = exprIsBig f || exprIsBig a
1052 exprIsBig other = True
1057 tcEqExpr :: CoreExpr -> CoreExpr -> Bool
1058 -- Used in rule matching, so does *not* look through
1059 -- newtypes, predicate types; hence tcEqExpr
1061 tcEqExpr e1 e2 = tcEqExprX rn_env e1 e2
1063 rn_env = mkRnEnv2 (mkInScopeSet (exprFreeVars e1 `unionVarSet` exprFreeVars e2))
1065 tcEqExprX :: RnEnv2 -> CoreExpr -> CoreExpr -> Bool
1066 tcEqExprX env (Var v1) (Var v2) = rnOccL env v1 == rnOccR env v2
1067 tcEqExprX env (Lit lit1) (Lit lit2) = lit1 == lit2
1068 tcEqExprX env (App f1 a1) (App f2 a2) = tcEqExprX env f1 f2 && tcEqExprX env a1 a2
1069 tcEqExprX env (Lam v1 e1) (Lam v2 e2) = tcEqExprX (rnBndr2 env v1 v2) e1 e2
1070 tcEqExprX env (Let (NonRec v1 r1) e1)
1071 (Let (NonRec v2 r2) e2) = tcEqExprX env r1 r2
1072 && tcEqExprX (rnBndr2 env v1 v2) e1 e2
1073 tcEqExprX env (Let (Rec ps1) e1)
1074 (Let (Rec ps2) e2) = equalLength ps1 ps2
1075 && and (zipWith eq_rhs ps1 ps2)
1076 && tcEqExprX env' e1 e2
1078 env' = foldl2 rn_bndr2 env ps2 ps2
1079 rn_bndr2 env (b1,_) (b2,_) = rnBndr2 env b1 b2
1080 eq_rhs (_,r1) (_,r2) = tcEqExprX env' r1 r2
1081 tcEqExprX env (Case e1 v1 t1 a1)
1082 (Case e2 v2 t2 a2) = tcEqExprX env e1 e2
1083 && tcEqTypeX env t1 t2
1084 && equalLength a1 a2
1085 && and (zipWith (eq_alt env') a1 a2)
1087 env' = rnBndr2 env v1 v2
1089 tcEqExprX env (Note n1 e1) (Note n2 e2) = eq_note env n1 n2 && tcEqExprX env e1 e2
1090 tcEqExprX env (Type t1) (Type t2) = tcEqTypeX env t1 t2
1091 tcEqExprX env e1 e2 = False
1093 eq_alt env (c1,vs1,r1) (c2,vs2,r2) = c1==c2 && tcEqExprX (rnBndrs2 env vs1 vs2) r1 r2
1095 eq_note env (SCC cc1) (SCC cc2) = cc1 == cc2
1096 eq_note env (Coerce t1 f1) (Coerce t2 f2) = tcEqTypeX env t1 t2 && tcEqTypeX env f1 f2
1097 eq_note env InlineCall InlineCall = True
1098 eq_note env (CoreNote s1) (CoreNote s2) = s1 == s2
1099 eq_note env other1 other2 = False
1103 %************************************************************************
1105 \subsection{The size of an expression}
1107 %************************************************************************
1110 coreBindsSize :: [CoreBind] -> Int
1111 coreBindsSize bs = foldr ((+) . bindSize) 0 bs
1113 exprSize :: CoreExpr -> Int
1114 -- A measure of the size of the expressions
1115 -- It also forces the expression pretty drastically as a side effect
1116 exprSize (Var v) = v `seq` 1
1117 exprSize (Lit lit) = lit `seq` 1
1118 exprSize (App f a) = exprSize f + exprSize a
1119 exprSize (Lam b e) = varSize b + exprSize e
1120 exprSize (Let b e) = bindSize b + exprSize e
1121 exprSize (Case e b t as) = seqType t `seq` exprSize e + varSize b + 1 + foldr ((+) . altSize) 0 as
1122 exprSize (Note n e) = noteSize n + exprSize e
1123 exprSize (Type t) = seqType t `seq` 1
1125 noteSize (SCC cc) = cc `seq` 1
1126 noteSize (Coerce t1 t2) = seqType t1 `seq` seqType t2 `seq` 1
1127 noteSize InlineCall = 1
1128 noteSize InlineMe = 1
1129 noteSize (CoreNote s) = s `seq` 1 -- hdaume: core annotations
1131 varSize :: Var -> Int
1132 varSize b | isTyVar b = 1
1133 | otherwise = seqType (idType b) `seq`
1134 megaSeqIdInfo (idInfo b) `seq`
1137 varsSize = foldr ((+) . varSize) 0
1139 bindSize (NonRec b e) = varSize b + exprSize e
1140 bindSize (Rec prs) = foldr ((+) . pairSize) 0 prs
1142 pairSize (b,e) = varSize b + exprSize e
1144 altSize (c,bs,e) = c `seq` varsSize bs + exprSize e
1148 %************************************************************************
1150 \subsection{Hashing}
1152 %************************************************************************
1155 hashExpr :: CoreExpr -> Int
1156 hashExpr e | hash < 0 = 77 -- Just in case we hit -maxInt
1159 hash = abs (hash_expr e) -- Negative numbers kill UniqFM
1161 hash_expr (Note _ e) = hash_expr e
1162 hash_expr (Let (NonRec b r) e) = hashId b
1163 hash_expr (Let (Rec ((b,r):_)) e) = hashId b
1164 hash_expr (Case _ b _ _) = hashId b
1165 hash_expr (App f e) = hash_expr f * fast_hash_expr e
1166 hash_expr (Var v) = hashId v
1167 hash_expr (Lit lit) = hashLiteral lit
1168 hash_expr (Lam b _) = hashId b
1169 hash_expr (Type t) = trace "hash_expr: type" 1 -- Shouldn't happen
1171 fast_hash_expr (Var v) = hashId v
1172 fast_hash_expr (Lit lit) = hashLiteral lit
1173 fast_hash_expr (App f (Type _)) = fast_hash_expr f
1174 fast_hash_expr (App f a) = fast_hash_expr a
1175 fast_hash_expr (Lam b _) = hashId b
1176 fast_hash_expr other = 1
1179 hashId id = hashName (idName id)
1182 %************************************************************************
1184 \subsection{Determining non-updatable right-hand-sides}
1186 %************************************************************************
1188 Top-level constructor applications can usually be allocated
1189 statically, but they can't if the constructor, or any of the
1190 arguments, come from another DLL (because we can't refer to static
1191 labels in other DLLs).
1193 If this happens we simply make the RHS into an updatable thunk,
1194 and 'exectute' it rather than allocating it statically.
1197 rhsIsStatic :: HomeModules -> CoreExpr -> Bool
1198 -- This function is called only on *top-level* right-hand sides
1199 -- Returns True if the RHS can be allocated statically, with
1200 -- no thunks involved at all.
1202 -- It's called (i) in TidyPgm.hasCafRefs to decide if the rhs is, or
1203 -- refers to, CAFs; and (ii) in CoreToStg to decide whether to put an
1204 -- update flag on it.
1206 -- The basic idea is that rhsIsStatic returns True only if the RHS is
1207 -- (a) a value lambda
1208 -- (b) a saturated constructor application with static args
1210 -- BUT watch out for
1211 -- (i) Any cross-DLL references kill static-ness completely
1212 -- because they must be 'executed' not statically allocated
1213 -- ("DLL" here really only refers to Windows DLLs, on other platforms,
1214 -- this is not necessary)
1216 -- (ii) We treat partial applications as redexes, because in fact we
1217 -- make a thunk for them that runs and builds a PAP
1218 -- at run-time. The only appliations that are treated as
1219 -- static are *saturated* applications of constructors.
1221 -- We used to try to be clever with nested structures like this:
1222 -- ys = (:) w ((:) w [])
1223 -- on the grounds that CorePrep will flatten ANF-ise it later.
1224 -- But supporting this special case made the function much more
1225 -- complicated, because the special case only applies if there are no
1226 -- enclosing type lambdas:
1227 -- ys = /\ a -> Foo (Baz ([] a))
1228 -- Here the nested (Baz []) won't float out to top level in CorePrep.
1230 -- But in fact, even without -O, nested structures at top level are
1231 -- flattened by the simplifier, so we don't need to be super-clever here.
1235 -- f = \x::Int. x+7 TRUE
1236 -- p = (True,False) TRUE
1238 -- d = (fst p, False) FALSE because there's a redex inside
1239 -- (this particular one doesn't happen but...)
1241 -- h = D# (1.0## /## 2.0##) FALSE (redex again)
1242 -- n = /\a. Nil a TRUE
1244 -- t = /\a. (:) (case w a of ...) (Nil a) FALSE (redex)
1247 -- This is a bit like CoreUtils.exprIsHNF, with the following differences:
1248 -- a) scc "foo" (\x -> ...) is updatable (so we catch the right SCC)
1250 -- b) (C x xs), where C is a contructors is updatable if the application is
1253 -- c) don't look through unfolding of f in (f x).
1255 -- When opt_RuntimeTypes is on, we keep type lambdas and treat
1256 -- them as making the RHS re-entrant (non-updatable).
1258 rhsIsStatic hmods rhs = is_static False rhs
1260 is_static :: Bool -- True <=> in a constructor argument; must be atomic
1263 is_static False (Lam b e) = isRuntimeVar b || is_static False e
1265 is_static in_arg (Note (SCC _) e) = False
1266 is_static in_arg (Note _ e) = is_static in_arg e
1268 is_static in_arg (Lit lit)
1270 MachLabel _ _ -> False
1272 -- A MachLabel (foreign import "&foo") in an argument
1273 -- prevents a constructor application from being static. The
1274 -- reason is that it might give rise to unresolvable symbols
1275 -- in the object file: under Linux, references to "weak"
1276 -- symbols from the data segment give rise to "unresolvable
1277 -- relocation" errors at link time This might be due to a bug
1278 -- in the linker, but we'll work around it here anyway.
1281 is_static in_arg other_expr = go other_expr 0
1283 go (Var f) n_val_args
1284 #if mingw32_TARGET_OS
1285 | not (isDllName hmods (idName f))
1287 = saturated_data_con f n_val_args
1288 || (in_arg && n_val_args == 0)
1289 -- A naked un-applied variable is *not* deemed a static RHS
1291 -- Reason: better to update so that the indirection gets shorted
1292 -- out, and the true value will be seen
1293 -- NB: if you change this, you'll break the invariant that THUNK_STATICs
1294 -- are always updatable. If you do so, make sure that non-updatable
1295 -- ones have enough space for their static link field!
1297 go (App f a) n_val_args
1298 | isTypeArg a = go f n_val_args
1299 | not in_arg && is_static True a = go f (n_val_args + 1)
1300 -- The (not in_arg) checks that we aren't in a constructor argument;
1301 -- if we are, we don't allow (value) applications of any sort
1303 -- NB. In case you wonder, args are sometimes not atomic. eg.
1304 -- x = D# (1.0## /## 2.0##)
1305 -- can't float because /## can fail.
1307 go (Note (SCC _) f) n_val_args = False
1308 go (Note _ f) n_val_args = go f n_val_args
1310 go other n_val_args = False
1312 saturated_data_con f n_val_args
1313 = case isDataConWorkId_maybe f of
1314 Just dc -> n_val_args == dataConRepArity dc