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
16 -- Properties of expressions
18 exprIsDupable, exprIsTrivial, exprIsCheap,
19 exprIsValue,exprOkForSpeculation, exprIsBig,
23 -- Arity and eta expansion
24 manifestArity, exprArity,
25 exprEtaExpandArity, etaExpand,
34 cheapEqExpr, eqExpr, applyTypeToArgs, applyTypeToArg
37 #include "HsVersions.h"
40 import GLAEXTS -- For `xori`
43 import PprCore ( pprCoreExpr )
44 import Var ( Var, isId, isTyVar )
46 import Name ( hashName, isDllName )
47 import Literal ( hashLiteral, literalType, litIsDupable,
48 litIsTrivial, isZeroLit )
49 import DataCon ( DataCon, dataConRepArity, dataConArgTys,
50 isExistentialDataCon, dataConTyCon )
51 import PrimOp ( PrimOp(..), primOpOkForSpeculation, primOpIsCheap )
52 import Id ( Id, idType, globalIdDetails, idNewStrictness,
53 mkWildId, idArity, idName, idUnfolding, idInfo,
54 isOneShotLambda, isDataConWorkId_maybe, mkSysLocal,
55 isDataConWorkId, isBottomingId
57 import IdInfo ( GlobalIdDetails(..), megaSeqIdInfo )
58 import NewDemand ( appIsBottom )
59 import Type ( Type, mkFunTy, mkForAllTy, splitFunTy_maybe,
61 applyTys, isUnLiftedType, seqType, mkTyVarTy,
62 splitForAllTy_maybe, isForAllTy, splitRecNewType_maybe,
63 splitTyConApp_maybe, eqType, funResultTy, applyTy,
66 import TyCon ( tyConArity )
67 import TysWiredIn ( boolTy, trueDataCon, falseDataCon )
68 import CostCentre ( CostCentre )
69 import BasicTypes ( Arity )
70 import Unique ( Unique )
72 import TysPrim ( alphaTy ) -- Debugging only
73 import Util ( equalLength, lengthAtLeast )
74 import TysPrim ( statePrimTyCon )
78 %************************************************************************
80 \subsection{Find the type of a Core atom/expression}
82 %************************************************************************
85 exprType :: CoreExpr -> Type
87 exprType (Var var) = idType var
88 exprType (Lit lit) = literalType lit
89 exprType (Let _ body) = exprType body
90 exprType (Case _ _ alts) = coreAltsType alts
91 exprType (Note (Coerce ty _) e) = ty -- **! should take usage from e
92 exprType (Note other_note e) = exprType e
93 exprType (Lam binder expr) = mkPiType binder (exprType expr)
95 = case collectArgs e of
96 (fun, args) -> applyTypeToArgs e (exprType fun) args
98 exprType other = pprTrace "exprType" (pprCoreExpr other) alphaTy
100 coreAltsType :: [CoreAlt] -> Type
101 coreAltsType ((_,_,rhs) : _) = exprType rhs
104 @mkPiType@ makes a (->) type or a forall type, depending on whether
105 it is given a type variable or a term variable. We cleverly use the
106 lbvarinfo field to figure out the right annotation for the arrove in
107 case of a term variable.
110 mkPiType :: Var -> Type -> Type -- The more polymorphic version
111 mkPiTypes :: [Var] -> Type -> Type -- doesn't work...
113 mkPiTypes vs ty = foldr mkPiType ty vs
116 | isId v = mkFunTy (idType v) ty
117 | otherwise = mkForAllTy v ty
121 applyTypeToArg :: Type -> CoreExpr -> Type
122 applyTypeToArg fun_ty (Type arg_ty) = applyTy fun_ty arg_ty
123 applyTypeToArg fun_ty other_arg = funResultTy fun_ty
125 applyTypeToArgs :: CoreExpr -> Type -> [CoreExpr] -> Type
126 -- A more efficient version of applyTypeToArg
127 -- when we have several args
128 -- The first argument is just for debugging
129 applyTypeToArgs e op_ty [] = op_ty
131 applyTypeToArgs e op_ty (Type ty : args)
132 = -- Accumulate type arguments so we can instantiate all at once
135 go rev_tys (Type ty : args) = go (ty:rev_tys) args
136 go rev_tys rest_args = applyTypeToArgs e op_ty' rest_args
138 op_ty' = applyTys op_ty (reverse rev_tys)
140 applyTypeToArgs e op_ty (other_arg : args)
141 = case (splitFunTy_maybe op_ty) of
142 Just (_, res_ty) -> applyTypeToArgs e res_ty args
143 Nothing -> pprPanic "applyTypeToArgs" (pprCoreExpr e)
148 %************************************************************************
150 \subsection{Attaching notes}
152 %************************************************************************
154 mkNote removes redundant coercions, and SCCs where possible
158 mkNote :: Note -> CoreExpr -> CoreExpr
159 mkNote (Coerce to_ty from_ty) expr = mkCoerce2 to_ty from_ty expr
160 mkNote (SCC cc) expr = mkSCC cc expr
161 mkNote InlineMe expr = mkInlineMe expr
162 mkNote note expr = Note note expr
165 -- Slide InlineCall in around the function
166 -- No longer necessary I think (SLPJ Apr 99)
167 -- mkNote InlineCall (App f a) = App (mkNote InlineCall f) a
168 -- mkNote InlineCall (Var v) = Note InlineCall (Var v)
169 -- mkNote InlineCall expr = expr
172 Drop trivial InlineMe's. This is somewhat important, because if we have an unfolding
173 that looks like (Note InlineMe (Var v)), the InlineMe doesn't go away because it may
174 not be *applied* to anything.
176 We don't use exprIsTrivial here, though, because we sometimes generate worker/wrapper
179 f = inline_me (coerce t fw)
180 As usual, the inline_me prevents the worker from getting inlined back into the wrapper.
181 We want the split, so that the coerces can cancel at the call site.
183 However, we can get left with tiresome type applications. Notably, consider
184 f = /\ a -> let t = e in (t, w)
185 Then lifting the let out of the big lambda gives
187 f = /\ a -> let t = inline_me (t' a) in (t, w)
188 The inline_me is to stop the simplifier inlining t' right back
189 into t's RHS. In the next phase we'll substitute for t (since
190 its rhs is trivial) and *then* we could get rid of the inline_me.
191 But it hardly seems worth it, so I don't bother.
194 mkInlineMe (Var v) = Var v
195 mkInlineMe e = Note InlineMe e
201 mkCoerce :: Type -> CoreExpr -> CoreExpr
202 mkCoerce to_ty expr = mkCoerce2 to_ty (exprType expr) expr
204 mkCoerce2 :: Type -> Type -> CoreExpr -> CoreExpr
205 mkCoerce2 to_ty from_ty (Note (Coerce to_ty2 from_ty2) expr)
206 = ASSERT( from_ty `eqType` to_ty2 )
207 mkCoerce2 to_ty from_ty2 expr
209 mkCoerce2 to_ty from_ty expr
210 | to_ty `eqType` from_ty = expr
211 | otherwise = ASSERT( from_ty `eqType` exprType expr )
212 Note (Coerce to_ty from_ty) expr
216 mkSCC :: CostCentre -> Expr b -> Expr b
217 -- Note: Nested SCC's *are* preserved for the benefit of
218 -- cost centre stack profiling
219 mkSCC cc (Lit lit) = Lit lit
220 mkSCC cc (Lam x e) = Lam x (mkSCC cc e) -- Move _scc_ inside lambda
221 mkSCC cc (Note (SCC cc') e) = Note (SCC cc) (Note (SCC cc') e)
222 mkSCC cc (Note n e) = Note n (mkSCC cc e) -- Move _scc_ inside notes
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.
244 bindNonRec bndr rhs body
245 | needsCaseBinding (idType bndr) rhs = Case rhs bndr [(DEFAULT,[],body)]
246 | otherwise = Let (NonRec bndr rhs) body
248 needsCaseBinding ty rhs = isUnLiftedType ty && not (exprOkForSpeculation rhs)
249 -- Make a case expression instead of a let
250 -- These can arise either from the desugarer,
251 -- or from beta reductions: (\x.e) (x +# y)
255 mkAltExpr :: AltCon -> [CoreBndr] -> [Type] -> CoreExpr
256 -- This guy constructs the value that the scrutinee must have
257 -- when you are in one particular branch of a case
258 mkAltExpr (DataAlt con) args inst_tys
259 = mkConApp con (map Type inst_tys ++ map varToCoreExpr args)
260 mkAltExpr (LitAlt lit) [] []
263 mkIfThenElse :: CoreExpr -> CoreExpr -> CoreExpr -> CoreExpr
264 mkIfThenElse guard then_expr else_expr
265 = Case guard (mkWildId boolTy)
266 [ (DataAlt trueDataCon, [], then_expr),
267 (DataAlt falseDataCon, [], else_expr) ]
271 %************************************************************************
273 \subsection{Taking expressions apart}
275 %************************************************************************
277 The default alternative must be first, if it exists at all.
278 This makes it easy to find, though it makes matching marginally harder.
281 findDefault :: [CoreAlt] -> ([CoreAlt], Maybe CoreExpr)
282 findDefault ((DEFAULT,args,rhs) : alts) = ASSERT( null args ) (alts, Just rhs)
283 findDefault alts = (alts, Nothing)
285 findAlt :: AltCon -> [CoreAlt] -> CoreAlt
288 (deflt@(DEFAULT,_,_):alts) -> go alts deflt
289 other -> go alts panic_deflt
292 panic_deflt = pprPanic "Missing alternative" (ppr con $$ vcat (map ppr alts))
295 go (alt@(con1,_,_) : alts) deflt | con == con1 = alt
296 | otherwise = ASSERT( not (con1 == DEFAULT) )
301 %************************************************************************
303 \subsection{Figuring out things about expressions}
305 %************************************************************************
307 @exprIsTrivial@ is true of expressions we are unconditionally happy to
308 duplicate; simple variables and constants, and type
309 applications. Note that primop Ids aren't considered
312 @exprIsBottom@ is true of expressions that are guaranteed to diverge
315 There used to be a gruesome test for (hasNoBinding v) in the
317 exprIsTrivial (Var v) | hasNoBinding v = idArity v == 0
318 The idea here is that a constructor worker, like $wJust, is
319 really short for (\x -> $wJust x), becuase $wJust has no binding.
320 So it should be treated like a lambda. Ditto unsaturated primops.
321 But now constructor workers are not "have-no-binding" Ids. And
322 completely un-applied primops and foreign-call Ids are sufficiently
323 rare that I plan to allow them to be duplicated and put up with
326 SCC notes. We do not treat (_scc_ "foo" x) as trivial, because
327 a) it really generates code, (and a heap object when it's
328 a function arg) to capture the cost centre
329 b) see the note [SCC-and-exprIsTrivial] in Simplify.simplLazyBind
332 exprIsTrivial (Var v) = True -- See notes above
333 exprIsTrivial (Type _) = True
334 exprIsTrivial (Lit lit) = litIsTrivial lit
335 exprIsTrivial (App e arg) = not (isRuntimeArg arg) && exprIsTrivial e
336 exprIsTrivial (Note (SCC _) e) = False -- See notes above
337 exprIsTrivial (Note _ e) = exprIsTrivial e
338 exprIsTrivial (Lam b body) = not (isRuntimeVar b) && exprIsTrivial body
339 exprIsTrivial other = False
343 @exprIsDupable@ is true of expressions that can be duplicated at a modest
344 cost in code size. This will only happen in different case
345 branches, so there's no issue about duplicating work.
347 That is, exprIsDupable returns True of (f x) even if
348 f is very very expensive to call.
350 Its only purpose is to avoid fruitless let-binding
351 and then inlining of case join points
355 exprIsDupable (Type _) = True
356 exprIsDupable (Var v) = True
357 exprIsDupable (Lit lit) = litIsDupable lit
358 exprIsDupable (Note InlineMe e) = True
359 exprIsDupable (Note _ e) = exprIsDupable e
363 go (Var v) n_args = True
364 go (App f a) n_args = n_args < dupAppSize
367 go other n_args = False
370 dupAppSize = 4 -- Size of application we are prepared to duplicate
373 @exprIsCheap@ looks at a Core expression and returns \tr{True} if
374 it is obviously in weak head normal form, or is cheap to get to WHNF.
375 [Note that that's not the same as exprIsDupable; an expression might be
376 big, and hence not dupable, but still cheap.]
378 By ``cheap'' we mean a computation we're willing to:
379 push inside a lambda, or
380 inline at more than one place
381 That might mean it gets evaluated more than once, instead of being
382 shared. The main examples of things which aren't WHNF but are
387 (where e, and all the ei are cheap)
390 (where e and b are cheap)
393 (where op is a cheap primitive operator)
396 (because we are happy to substitute it inside a lambda)
398 Notice that a variable is considered 'cheap': we can push it inside a lambda,
399 because sharing will make sure it is only evaluated once.
402 exprIsCheap :: CoreExpr -> Bool
403 exprIsCheap (Lit lit) = True
404 exprIsCheap (Type _) = True
405 exprIsCheap (Var _) = True
406 exprIsCheap (Note InlineMe e) = True
407 exprIsCheap (Note _ e) = exprIsCheap e
408 exprIsCheap (Lam x e) = isRuntimeVar x || exprIsCheap e
409 exprIsCheap (Case e _ alts) = exprIsCheap e &&
410 and [exprIsCheap rhs | (_,_,rhs) <- alts]
411 -- Experimentally, treat (case x of ...) as cheap
412 -- (and case __coerce x etc.)
413 -- This improves arities of overloaded functions where
414 -- there is only dictionary selection (no construction) involved
415 exprIsCheap (Let (NonRec x _) e)
416 | isUnLiftedType (idType x) = exprIsCheap e
418 -- strict lets always have cheap right hand sides, and
421 exprIsCheap other_expr
422 = go other_expr 0 True
424 go (Var f) n_args args_cheap
425 = (idAppIsCheap f n_args && args_cheap)
426 -- A constructor, cheap primop, or partial application
428 || idAppIsBottom f n_args
429 -- Application of a function which
430 -- always gives bottom; we treat this as cheap
431 -- because it certainly doesn't need to be shared!
433 go (App f a) n_args args_cheap
434 | not (isRuntimeArg a) = go f n_args args_cheap
435 | otherwise = go f (n_args + 1) (exprIsCheap a && args_cheap)
437 go other n_args args_cheap = False
439 idAppIsCheap :: Id -> Int -> Bool
440 idAppIsCheap id n_val_args
441 | n_val_args == 0 = True -- Just a type application of
442 -- a variable (f t1 t2 t3)
444 | otherwise = case globalIdDetails id of
445 DataConWorkId _ -> True
446 RecordSelId _ -> True -- I'm experimenting with making record selection
447 ClassOpId _ -> True -- look cheap, so we will substitute it inside a
448 -- lambda. Particularly for dictionary field selection
450 PrimOpId op -> primOpIsCheap op -- In principle we should worry about primops
451 -- that return a type variable, since the result
452 -- might be applied to something, but I'm not going
453 -- to bother to check the number of args
454 other -> n_val_args < idArity id
457 exprOkForSpeculation returns True of an expression that it is
459 * safe to evaluate even if normal order eval might not
460 evaluate the expression at all, or
462 * safe *not* to evaluate even if normal order would do so
466 the expression guarantees to terminate,
468 without raising an exception,
469 without causing a side effect (e.g. writing a mutable variable)
472 let x = case y# +# 1# of { r# -> I# r# }
475 case y# +# 1# of { r# ->
480 We can only do this if the (y+1) is ok for speculation: it has no
481 side effects, and can't diverge or raise an exception.
484 exprOkForSpeculation :: CoreExpr -> Bool
485 exprOkForSpeculation (Lit _) = True
486 exprOkForSpeculation (Type _) = True
487 exprOkForSpeculation (Var v) = isUnLiftedType (idType v)
488 exprOkForSpeculation (Note _ e) = exprOkForSpeculation e
489 exprOkForSpeculation other_expr
490 = case collectArgs other_expr of
491 (Var f, args) -> spec_ok (globalIdDetails f) args
495 spec_ok (DataConWorkId _) args
496 = True -- The strictness of the constructor has already
497 -- been expressed by its "wrapper", so we don't need
498 -- to take the arguments into account
500 spec_ok (PrimOpId op) args
501 | isDivOp op, -- Special case for dividing operations that fail
502 [arg1, Lit lit] <- args -- only if the divisor is zero
503 = not (isZeroLit lit) && exprOkForSpeculation arg1
504 -- Often there is a literal divisor, and this
505 -- can get rid of a thunk in an inner looop
508 = primOpOkForSpeculation op &&
509 all exprOkForSpeculation args
510 -- A bit conservative: we don't really need
511 -- to care about lazy arguments, but this is easy
513 spec_ok other args = False
515 isDivOp :: PrimOp -> Bool
516 -- True of dyadic operators that can fail
517 -- only if the second arg is zero
518 -- This function probably belongs in PrimOp, or even in
519 -- an automagically generated file.. but it's such a
520 -- special case I thought I'd leave it here for now.
521 isDivOp IntQuotOp = True
522 isDivOp IntRemOp = True
523 isDivOp WordQuotOp = True
524 isDivOp WordRemOp = True
525 isDivOp IntegerQuotRemOp = True
526 isDivOp IntegerDivModOp = True
527 isDivOp FloatDivOp = True
528 isDivOp DoubleDivOp = True
529 isDivOp other = False
534 exprIsBottom :: CoreExpr -> Bool -- True => definitely bottom
535 exprIsBottom e = go 0 e
537 -- n is the number of args
538 go n (Note _ e) = go n e
539 go n (Let _ e) = go n e
540 go n (Case e _ _) = go 0 e -- Just check the scrut
541 go n (App e _) = go (n+1) e
542 go n (Var v) = idAppIsBottom v n
544 go n (Lam _ _) = False
546 idAppIsBottom :: Id -> Int -> Bool
547 idAppIsBottom id n_val_args = appIsBottom (idNewStrictness id) n_val_args
550 @exprIsValue@ returns true for expressions that are certainly *already*
551 evaluated to *head* normal form. This is used to decide whether it's ok
554 case x of _ -> e ===> e
556 and to decide whether it's safe to discard a `seq`
558 So, it does *not* treat variables as evaluated, unless they say they are.
560 But it *does* treat partial applications and constructor applications
561 as values, even if their arguments are non-trivial, provided the argument
563 e.g. (:) (f x) (map f xs) is a value
564 map (...redex...) is a value
565 Because `seq` on such things completes immediately
567 For unlifted argument types, we have to be careful:
569 Suppose (f x) diverges; then C (f x) is not a value. True, but
570 this form is illegal (see the invariants in CoreSyn). Args of unboxed
571 type must be ok-for-speculation (or trivial).
574 exprIsValue :: CoreExpr -> Bool -- True => Value-lambda, constructor, PAP
575 exprIsValue (Var v) -- NB: There are no value args at this point
576 = isDataConWorkId v -- Catches nullary constructors,
577 -- so that [] and () are values, for example
578 || idArity v > 0 -- Catches (e.g.) primops that don't have unfoldings
579 || isEvaldUnfolding (idUnfolding v)
580 -- Check the thing's unfolding; it might be bound to a value
581 -- A worry: what if an Id's unfolding is just itself:
582 -- then we could get an infinite loop...
584 exprIsValue (Lit l) = True
585 exprIsValue (Type ty) = True -- Types are honorary Values;
586 -- we don't mind copying them
587 exprIsValue (Lam b e) = isRuntimeVar b || exprIsValue e
588 exprIsValue (Note _ e) = exprIsValue e
589 exprIsValue (App e (Type _)) = exprIsValue e
590 exprIsValue (App e a) = app_is_value e [a]
591 exprIsValue other = False
593 -- There is at least one value argument
594 app_is_value (Var fun) args
595 | isDataConWorkId fun -- Constructor apps are values
596 || idArity fun > valArgCount args -- Under-applied function
597 = check_args (idType fun) args
598 app_is_value (App f a) as = app_is_value f (a:as)
599 app_is_value other as = False
601 -- 'check_args' checks that unlifted-type args
602 -- are in fact guaranteed non-divergent
603 check_args fun_ty [] = True
604 check_args fun_ty (Type _ : args) = case splitForAllTy_maybe fun_ty of
605 Just (_, ty) -> check_args ty args
606 check_args fun_ty (arg : args)
607 | isUnLiftedType arg_ty = exprOkForSpeculation arg
608 | otherwise = check_args res_ty args
610 (arg_ty, res_ty) = splitFunTy fun_ty
614 exprIsConApp_maybe :: CoreExpr -> Maybe (DataCon, [CoreExpr])
615 exprIsConApp_maybe (Note (Coerce to_ty from_ty) expr)
616 = -- Maybe this is over the top, but here we try to turn
617 -- coerce (S,T) ( x, y )
619 -- ( coerce S x, coerce T y )
620 -- This happens in anger in PrelArrExts which has a coerce
621 -- case coerce memcpy a b of
623 -- where the memcpy is in the IO monad, but the call is in
625 case exprIsConApp_maybe expr of {
629 case splitTyConApp_maybe to_ty of {
631 Just (tc, tc_arg_tys) | tc /= dataConTyCon dc -> Nothing
632 | isExistentialDataCon dc -> Nothing
634 -- Type constructor must match
635 -- We knock out existentials to keep matters simple(r)
637 arity = tyConArity tc
638 val_args = drop arity args
639 to_arg_tys = dataConArgTys dc tc_arg_tys
640 mk_coerce ty arg = mkCoerce ty arg
641 new_val_args = zipWith mk_coerce to_arg_tys val_args
643 ASSERT( all isTypeArg (take arity args) )
644 ASSERT( equalLength val_args to_arg_tys )
645 Just (dc, map Type tc_arg_tys ++ new_val_args)
648 exprIsConApp_maybe (Note _ expr)
649 = exprIsConApp_maybe expr
650 -- We ignore InlineMe notes in case we have
651 -- x = __inline_me__ (a,b)
652 -- All part of making sure that INLINE pragmas never hurt
653 -- Marcin tripped on this one when making dictionaries more inlinable
655 -- In fact, we ignore all notes. For example,
656 -- case _scc_ "foo" (C a b) of
658 -- should be optimised away, but it will be only if we look
659 -- through the SCC note.
661 exprIsConApp_maybe expr = analyse (collectArgs expr)
663 analyse (Var fun, args)
664 | Just con <- isDataConWorkId_maybe fun,
665 args `lengthAtLeast` dataConRepArity con
666 -- Might be > because the arity excludes type args
669 -- Look through unfoldings, but only cheap ones, because
670 -- we are effectively duplicating the unfolding
671 analyse (Var fun, [])
672 | let unf = idUnfolding fun,
674 = exprIsConApp_maybe (unfoldingTemplate unf)
676 analyse other = Nothing
681 %************************************************************************
683 \subsection{Eta reduction and expansion}
685 %************************************************************************
688 exprEtaExpandArity :: CoreExpr -> Arity
689 {- The Arity returned is the number of value args the
690 thing can be applied to without doing much work
692 exprEtaExpandArity is used when eta expanding
695 It returns 1 (or more) to:
696 case x of p -> \s -> ...
697 because for I/O ish things we really want to get that \s to the top.
698 We are prepared to evaluate x each time round the loop in order to get that
700 It's all a bit more subtle than it looks:
704 Consider one-shot lambdas
705 let x = expensive in \y z -> E
706 We want this to have arity 2 if the \y-abstraction is a 1-shot lambda
707 Hence the ArityType returned by arityType
709 2. The state-transformer hack
711 The one-shot lambda special cause is particularly important/useful for
712 IO state transformers, where we often get
713 let x = E in \ s -> ...
715 and the \s is a real-world state token abstraction. Such abstractions
716 are almost invariably 1-shot, so we want to pull the \s out, past the
717 let x=E, even if E is expensive. So we treat state-token lambdas as
718 one-shot even if they aren't really. The hack is in Id.isOneShotLambda.
720 3. Dealing with bottom
723 f = \x -> error "foo"
724 Here, arity 1 is fine. But if it is
728 then we want to get arity 2. Tecnically, this isn't quite right, because
730 should diverge, but it'll converge if we eta-expand f. Nevertheless, we
731 do so; it improves some programs significantly, and increasing convergence
732 isn't a bad thing. Hence the ABot/ATop in ArityType.
734 Actually, the situation is worse. Consider
738 Can we eta-expand here? At first the answer looks like "yes of course", but
741 This should diverge! But if we eta-expand, it won't. Again, we ignore this
742 "problem", because being scrupulous would lose an important transformation for
747 exprEtaExpandArity e = arityDepth (arityType e)
749 -- A limited sort of function type
750 data ArityType = AFun Bool ArityType -- True <=> one-shot
751 | ATop -- Know nothing
754 arityDepth :: ArityType -> Arity
755 arityDepth (AFun _ ty) = 1 + arityDepth ty
758 andArityType ABot at2 = at2
759 andArityType ATop at2 = ATop
760 andArityType (AFun t1 at1) (AFun t2 at2) = AFun (t1 && t2) (andArityType at1 at2)
761 andArityType at1 at2 = andArityType at2 at1
763 arityType :: CoreExpr -> ArityType
764 -- (go1 e) = [b1,..,bn]
765 -- means expression can be rewritten \x_b1 -> ... \x_bn -> body
766 -- where bi is True <=> the lambda is one-shot
768 arityType (Note n e) = arityType e
769 -- Not needed any more: etaExpand is cleverer
770 -- | ok_note n = arityType e
771 -- | otherwise = ATop
776 mk :: Arity -> ArityType
777 mk 0 | isBottomingId v = ABot
779 mk n = AFun False (mk (n-1))
781 -- When the type of the Id encodes one-shot-ness,
782 -- use the idinfo here
784 -- Lambdas; increase arity
785 arityType (Lam x e) | isId x = AFun (isOneShotLambda x || isStateHack x) (arityType e)
786 | otherwise = arityType e
788 -- Applications; decrease arity
789 arityType (App f (Type _)) = arityType f
790 arityType (App f a) = case arityType f of
791 AFun one_shot xs | exprIsCheap a -> xs
794 -- Case/Let; keep arity if either the expression is cheap
795 -- or it's a 1-shot lambda
796 -- The former is not really right for Haskell
797 -- f x = case x of { (a,b) -> \y. e }
799 -- f x y = case x of { (a,b) -> e }
800 -- The difference is observable using 'seq'
801 arityType (Case scrut _ alts) = case foldr1 andArityType [arityType rhs | (_,_,rhs) <- alts] of
802 xs@(AFun one_shot _) | one_shot -> xs
803 xs | exprIsCheap scrut -> xs
806 arityType (Let b e) = case arityType e of
807 xs@(AFun one_shot _) | one_shot -> xs
808 xs | all exprIsCheap (rhssOfBind b) -> xs
811 arityType other = ATop
813 isStateHack id = case splitTyConApp_maybe (idType id) of
814 Just (tycon,_) | tycon == statePrimTyCon -> True
817 -- The last clause is a gross hack. It claims that
818 -- every function over realWorldStatePrimTy is a one-shot
819 -- function. This is pretty true in practice, and makes a big
820 -- difference. For example, consider
821 -- a `thenST` \ r -> ...E...
822 -- The early full laziness pass, if it doesn't know that r is one-shot
823 -- will pull out E (let's say it doesn't mention r) to give
824 -- let lvl = E in a `thenST` \ r -> ...lvl...
825 -- When `thenST` gets inlined, we end up with
826 -- let lvl = E in \s -> case a s of (r, s') -> ...lvl...
827 -- and we don't re-inline E.
829 -- It would be better to spot that r was one-shot to start with, but
830 -- I don't want to rely on that.
832 -- Another good example is in fill_in in PrelPack.lhs. We should be able to
833 -- spot that fill_in has arity 2 (and when Keith is done, we will) but we can't yet.
835 {- NOT NEEDED ANY MORE: etaExpand is cleverer
836 ok_note InlineMe = False
838 -- Notice that we do not look through __inline_me__
839 -- This may seem surprising, but consider
840 -- f = _inline_me (\x -> e)
841 -- We DO NOT want to eta expand this to
842 -- f = \x -> (_inline_me (\x -> e)) x
843 -- because the _inline_me gets dropped now it is applied,
852 etaExpand :: Arity -- Result should have this number of value args
854 -> CoreExpr -> Type -- Expression and its type
856 -- (etaExpand n us e ty) returns an expression with
857 -- the same meaning as 'e', but with arity 'n'.
859 -- Given e' = etaExpand n us e ty
861 -- ty = exprType e = exprType e'
863 -- Note that SCCs are not treated specially. If we have
864 -- etaExpand 2 (\x -> scc "foo" e)
865 -- = (\xy -> (scc "foo" e) y)
866 -- So the costs of evaluating 'e' (not 'e y') are attributed to "foo"
868 etaExpand n us expr ty
869 | manifestArity expr >= n = expr -- The no-op case
870 | otherwise = eta_expand n us expr ty
873 -- manifestArity sees how many leading value lambdas there are
874 manifestArity :: CoreExpr -> Arity
875 manifestArity (Lam v e) | isId v = 1 + manifestArity e
876 | otherwise = manifestArity e
877 manifestArity (Note _ e) = manifestArity e
880 -- etaExpand deals with for-alls. For example:
882 -- where E :: forall a. a -> a
884 -- (/\b. \y::a -> E b y)
886 -- It deals with coerces too, though they are now rare
887 -- so perhaps the extra code isn't worth it
889 eta_expand n us expr ty
891 -- The ILX code generator requires eta expansion for type arguments
892 -- too, but alas the 'n' doesn't tell us how many of them there
893 -- may be. So we eagerly eta expand any big lambdas, and just
894 -- cross our fingers about possible loss of sharing in the ILX case.
895 -- The Right Thing is probably to make 'arity' include
896 -- type variables throughout the compiler. (ToDo.)
898 -- Saturated, so nothing to do
901 -- Short cut for the case where there already
902 -- is a lambda; no point in gratuitously adding more
903 eta_expand n us (Lam v body) ty
905 = Lam v (eta_expand n us body (applyTy ty (mkTyVarTy v)))
908 = Lam v (eta_expand (n-1) us body (funResultTy ty))
910 -- We used to have a special case that stepped inside Coerces here,
911 -- thus: eta_expand n us (Note note@(Coerce _ ty) e) _
912 -- = Note note (eta_expand n us e ty)
913 -- BUT this led to an infinite loop
914 -- Example: newtype T = MkT (Int -> Int)
915 -- eta_expand 1 (coerce (Int->Int) e)
916 -- --> coerce (Int->Int) (eta_expand 1 T e)
918 -- --> coerce (Int->Int) (coerce T
919 -- (\x::Int -> eta_expand 1 (coerce (Int->Int) e)))
920 -- by the splitNewType_maybe case below
923 eta_expand n us expr ty
924 = case splitForAllTy_maybe ty of {
925 Just (tv,ty') -> Lam tv (eta_expand n us (App expr (Type (mkTyVarTy tv))) ty')
929 case splitFunTy_maybe ty of {
930 Just (arg_ty, res_ty) -> Lam arg1 (eta_expand (n-1) us2 (App expr (Var arg1)) res_ty)
932 arg1 = mkSysLocal FSLIT("eta") uniq arg_ty
938 -- newtype T = MkT ([T] -> Int)
939 -- Consider eta-expanding this
942 -- coerce T (\x::[T] -> (coerce ([T]->Int) e) x)
943 -- Only try this for recursive newtypes; the non-recursive kind
944 -- are transparent anyway
946 case splitRecNewType_maybe ty of {
947 Just ty' -> mkCoerce2 ty ty' (eta_expand n us (mkCoerce2 ty' ty expr) ty') ;
948 Nothing -> pprTrace "Bad eta expand" (ppr n $$ ppr expr $$ ppr ty) expr
952 exprArity is a cheap-and-cheerful version of exprEtaExpandArity.
953 It tells how many things the expression can be applied to before doing
954 any work. It doesn't look inside cases, lets, etc. The idea is that
955 exprEtaExpandArity will do the hard work, leaving something that's easy
956 for exprArity to grapple with. In particular, Simplify uses exprArity to
957 compute the ArityInfo for the Id.
959 Originally I thought that it was enough just to look for top-level lambdas, but
960 it isn't. I've seen this
962 foo = PrelBase.timesInt
964 We want foo to get arity 2 even though the eta-expander will leave it
965 unchanged, in the expectation that it'll be inlined. But occasionally it
966 isn't, because foo is blacklisted (used in a rule).
968 Similarly, see the ok_note check in exprEtaExpandArity. So
969 f = __inline_me (\x -> e)
970 won't be eta-expanded.
972 And in any case it seems more robust to have exprArity be a bit more intelligent.
973 But note that (\x y z -> f x y z)
974 should have arity 3, regardless of f's arity.
977 exprArity :: CoreExpr -> Arity
980 go (Var v) = idArity v
981 go (Lam x e) | isId x = go e + 1
984 go (App e (Type t)) = go e
985 go (App f a) | exprIsCheap a = (go f - 1) `max` 0
986 -- NB: exprIsCheap a!
987 -- f (fac x) does not have arity 2,
988 -- even if f has arity 3!
989 -- NB: `max 0`! (\x y -> f x) has arity 2, even if f is
990 -- unknown, hence arity 0
994 %************************************************************************
996 \subsection{Equality}
998 %************************************************************************
1000 @cheapEqExpr@ is a cheap equality test which bales out fast!
1001 True => definitely equal
1002 False => may or may not be equal
1005 cheapEqExpr :: Expr b -> Expr b -> Bool
1007 cheapEqExpr (Var v1) (Var v2) = v1==v2
1008 cheapEqExpr (Lit lit1) (Lit lit2) = lit1 == lit2
1009 cheapEqExpr (Type t1) (Type t2) = t1 `eqType` t2
1011 cheapEqExpr (App f1 a1) (App f2 a2)
1012 = f1 `cheapEqExpr` f2 && a1 `cheapEqExpr` a2
1014 cheapEqExpr _ _ = False
1016 exprIsBig :: Expr b -> Bool
1017 -- Returns True of expressions that are too big to be compared by cheapEqExpr
1018 exprIsBig (Lit _) = False
1019 exprIsBig (Var v) = False
1020 exprIsBig (Type t) = False
1021 exprIsBig (App f a) = exprIsBig f || exprIsBig a
1022 exprIsBig other = True
1027 eqExpr :: CoreExpr -> CoreExpr -> Bool
1028 -- Works ok at more general type, but only needed at CoreExpr
1029 -- Used in rule matching, so when we find a type we use
1030 -- eqTcType, which doesn't look through newtypes
1031 -- [And it doesn't risk falling into a black hole either.]
1033 = eq emptyVarEnv e1 e2
1035 -- The "env" maps variables in e1 to variables in ty2
1036 -- So when comparing lambdas etc,
1037 -- we in effect substitute v2 for v1 in e1 before continuing
1038 eq env (Var v1) (Var v2) = case lookupVarEnv env v1 of
1039 Just v1' -> v1' == v2
1042 eq env (Lit lit1) (Lit lit2) = lit1 == lit2
1043 eq env (App f1 a1) (App f2 a2) = eq env f1 f2 && eq env a1 a2
1044 eq env (Lam v1 e1) (Lam v2 e2) = eq (extendVarEnv env v1 v2) e1 e2
1045 eq env (Let (NonRec v1 r1) e1)
1046 (Let (NonRec v2 r2) e2) = eq env r1 r2 && eq (extendVarEnv env v1 v2) e1 e2
1047 eq env (Let (Rec ps1) e1)
1048 (Let (Rec ps2) e2) = equalLength ps1 ps2 &&
1049 and (zipWith eq_rhs ps1 ps2) &&
1052 env' = extendVarEnvList env [(v1,v2) | ((v1,_),(v2,_)) <- zip ps1 ps2]
1053 eq_rhs (_,r1) (_,r2) = eq env' r1 r2
1054 eq env (Case e1 v1 a1)
1055 (Case e2 v2 a2) = eq env e1 e2 &&
1056 equalLength a1 a2 &&
1057 and (zipWith (eq_alt env') a1 a2)
1059 env' = extendVarEnv env v1 v2
1061 eq env (Note n1 e1) (Note n2 e2) = eq_note env n1 n2 && eq env e1 e2
1062 eq env (Type t1) (Type t2) = t1 `eqType` t2
1063 eq env e1 e2 = False
1065 eq_list env [] [] = True
1066 eq_list env (e1:es1) (e2:es2) = eq env e1 e2 && eq_list env es1 es2
1067 eq_list env es1 es2 = False
1069 eq_alt env (c1,vs1,r1) (c2,vs2,r2) = c1==c2 &&
1070 eq (extendVarEnvList env (vs1 `zip` vs2)) r1 r2
1072 eq_note env (SCC cc1) (SCC cc2) = cc1 == cc2
1073 eq_note env (Coerce t1 f1) (Coerce t2 f2) = t1 `eqType` t2 && f1 `eqType` f2
1074 eq_note env InlineCall InlineCall = True
1075 eq_note env (CoreNote s1) (CoreNote s2) = s1 == s2
1076 eq_note env other1 other2 = False
1080 %************************************************************************
1082 \subsection{The size of an expression}
1084 %************************************************************************
1087 coreBindsSize :: [CoreBind] -> Int
1088 coreBindsSize bs = foldr ((+) . bindSize) 0 bs
1090 exprSize :: CoreExpr -> Int
1091 -- A measure of the size of the expressions
1092 -- It also forces the expression pretty drastically as a side effect
1093 exprSize (Var v) = v `seq` 1
1094 exprSize (Lit lit) = lit `seq` 1
1095 exprSize (App f a) = exprSize f + exprSize a
1096 exprSize (Lam b e) = varSize b + exprSize e
1097 exprSize (Let b e) = bindSize b + exprSize e
1098 exprSize (Case e b as) = exprSize e + varSize b + foldr ((+) . altSize) 0 as
1099 exprSize (Note n e) = noteSize n + exprSize e
1100 exprSize (Type t) = seqType t `seq` 1
1102 noteSize (SCC cc) = cc `seq` 1
1103 noteSize (Coerce t1 t2) = seqType t1 `seq` seqType t2 `seq` 1
1104 noteSize InlineCall = 1
1105 noteSize InlineMe = 1
1106 noteSize (CoreNote s) = s `seq` 1 -- hdaume: core annotations
1108 varSize :: Var -> Int
1109 varSize b | isTyVar b = 1
1110 | otherwise = seqType (idType b) `seq`
1111 megaSeqIdInfo (idInfo b) `seq`
1114 varsSize = foldr ((+) . varSize) 0
1116 bindSize (NonRec b e) = varSize b + exprSize e
1117 bindSize (Rec prs) = foldr ((+) . pairSize) 0 prs
1119 pairSize (b,e) = varSize b + exprSize e
1121 altSize (c,bs,e) = c `seq` varsSize bs + exprSize e
1125 %************************************************************************
1127 \subsection{Hashing}
1129 %************************************************************************
1132 hashExpr :: CoreExpr -> Int
1133 hashExpr e | hash < 0 = 77 -- Just in case we hit -maxInt
1136 hash = abs (hash_expr e) -- Negative numbers kill UniqFM
1138 hash_expr (Note _ e) = hash_expr e
1139 hash_expr (Let (NonRec b r) e) = hashId b
1140 hash_expr (Let (Rec ((b,r):_)) e) = hashId b
1141 hash_expr (Case _ b _) = hashId b
1142 hash_expr (App f e) = hash_expr f * fast_hash_expr e
1143 hash_expr (Var v) = hashId v
1144 hash_expr (Lit lit) = hashLiteral lit
1145 hash_expr (Lam b _) = hashId b
1146 hash_expr (Type t) = trace "hash_expr: type" 1 -- Shouldn't happen
1148 fast_hash_expr (Var v) = hashId v
1149 fast_hash_expr (Lit lit) = hashLiteral lit
1150 fast_hash_expr (App f (Type _)) = fast_hash_expr f
1151 fast_hash_expr (App f a) = fast_hash_expr a
1152 fast_hash_expr (Lam b _) = hashId b
1153 fast_hash_expr other = 1
1156 hashId id = hashName (idName id)
1159 %************************************************************************
1161 \subsection{Determining non-updatable right-hand-sides}
1163 %************************************************************************
1165 Top-level constructor applications can usually be allocated
1166 statically, but they can't if the constructor, or any of the
1167 arguments, come from another DLL (because we can't refer to static
1168 labels in other DLLs).
1170 If this happens we simply make the RHS into an updatable thunk,
1171 and 'exectute' it rather than allocating it statically.
1174 rhsIsStatic :: CoreExpr -> Bool
1175 -- This function is called only on *top-level* right-hand sides
1176 -- Returns True if the RHS can be allocated statically, with
1177 -- no thunks involved at all.
1179 -- It's called (i) in TidyPgm.hasCafRefs to decide if the rhs is, or
1180 -- refers to, CAFs; and (ii) in CoreToStg to decide whether to put an
1181 -- update flag on it.
1183 -- The basic idea is that rhsIsStatic returns True only if the RHS is
1184 -- (a) a value lambda
1185 -- (b) a saturated constructor application with static args
1187 -- BUT watch out for
1188 -- (i) Any cross-DLL references kill static-ness completely
1189 -- because they must be 'executed' not statically allocated
1191 -- (ii) We treat partial applications as redexes, because in fact we
1192 -- make a thunk for them that runs and builds a PAP
1193 -- at run-time. The only appliations that are treated as
1194 -- static are *saturated* applications of constructors.
1196 -- We used to try to be clever with nested structures like this:
1197 -- ys = (:) w ((:) w [])
1198 -- on the grounds that CorePrep will flatten ANF-ise it later.
1199 -- But supporting this special case made the function much more
1200 -- complicated, because the special case only applies if there are no
1201 -- enclosing type lambdas:
1202 -- ys = /\ a -> Foo (Baz ([] a))
1203 -- Here the nested (Baz []) won't float out to top level in CorePrep.
1205 -- But in fact, even without -O, nested structures at top level are
1206 -- flattened by the simplifier, so we don't need to be super-clever here.
1210 -- f = \x::Int. x+7 TRUE
1211 -- p = (True,False) TRUE
1213 -- d = (fst p, False) FALSE because there's a redex inside
1214 -- (this particular one doesn't happen but...)
1216 -- h = D# (1.0## /## 2.0##) FALSE (redex again)
1217 -- n = /\a. Nil a TRUE
1219 -- t = /\a. (:) (case w a of ...) (Nil a) FALSE (redex)
1222 -- This is a bit like CoreUtils.exprIsValue, with the following differences:
1223 -- a) scc "foo" (\x -> ...) is updatable (so we catch the right SCC)
1225 -- b) (C x xs), where C is a contructors is updatable if the application is
1228 -- c) don't look through unfolding of f in (f x).
1230 -- When opt_RuntimeTypes is on, we keep type lambdas and treat
1231 -- them as making the RHS re-entrant (non-updatable).
1233 rhsIsStatic rhs = is_static False rhs
1235 is_static :: Bool -- True <=> in a constructor argument; must be atomic
1238 is_static False (Lam b e) = isRuntimeVar b || is_static False e
1240 is_static in_arg (Note (SCC _) e) = False
1241 is_static in_arg (Note _ e) = is_static in_arg e
1242 is_static in_arg (Lit lit) = True
1244 is_static in_arg other_expr = go other_expr 0
1246 go (Var f) n_val_args
1247 | not (isDllName (idName f))
1248 = saturated_data_con f n_val_args
1249 || (in_arg && n_val_args == 0)
1250 -- A naked un-applied variable is *not* deemed a static RHS
1252 -- Reason: better to update so that the indirection gets shorted
1253 -- out, and the true value will be seen
1254 -- NB: if you change this, you'll break the invariant that THUNK_STATICs
1255 -- are always updatable. If you do so, make sure that non-updatable
1256 -- ones have enough space for their static link field!
1258 go (App f a) n_val_args
1259 | isTypeArg a = go f n_val_args
1260 | not in_arg && is_static True a = go f (n_val_args + 1)
1261 -- The (not in_arg) checks that we aren't in a constructor argument;
1262 -- if we are, we don't allow (value) applications of any sort
1264 -- NB. In case you wonder, args are sometimes not atomic. eg.
1265 -- x = D# (1.0## /## 2.0##)
1266 -- can't float because /## can fail.
1268 go (Note (SCC _) f) n_val_args = False
1269 go (Note _ f) n_val_args = go f n_val_args
1271 go other n_val_args = False
1273 saturated_data_con f n_val_args
1274 = case isDataConWorkId_maybe f of
1275 Just dc -> n_val_args == dataConRepArity dc