2 % (c) The GRASP/AQUA Project, Glasgow University, 1992-1998
4 \section[CoreUtils]{Utility functions on @Core@ syntax}
9 mkNote, mkInlineMe, mkSCC, mkCoerce, mkCoerce2,
10 bindNonRec, needsCaseBinding,
11 mkIfThenElse, mkAltExpr, mkPiType, mkPiTypes,
13 -- Taking expressions apart
14 findDefault, findAlt, hasDefault,
16 -- Properties of expressions
17 exprType, coreAltsType,
18 exprIsBottom, 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, isLitLitLit )
49 import DataCon ( DataCon, dataConRepArity, dataConArgTys,
50 isExistentialDataCon, dataConTyCon, dataConName )
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, splitNewType_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
157 mkNote :: Note -> CoreExpr -> CoreExpr
158 mkNote (Coerce to_ty from_ty) expr = mkCoerce2 to_ty from_ty expr
159 mkNote (SCC cc) expr = mkSCC cc expr
160 mkNote InlineMe expr = mkInlineMe expr
161 mkNote note expr = Note note expr
163 -- Slide InlineCall in around the function
164 -- No longer necessary I think (SLPJ Apr 99)
165 -- mkNote InlineCall (App f a) = App (mkNote InlineCall f) a
166 -- mkNote InlineCall (Var v) = Note InlineCall (Var v)
167 -- mkNote InlineCall expr = expr
170 Drop trivial InlineMe's. This is somewhat important, because if we have an unfolding
171 that looks like (Note InlineMe (Var v)), the InlineMe doesn't go away because it may
172 not be *applied* to anything.
174 We don't use exprIsTrivial here, though, because we sometimes generate worker/wrapper
177 f = inline_me (coerce t fw)
178 As usual, the inline_me prevents the worker from getting inlined back into the wrapper.
179 We want the split, so that the coerces can cancel at the call site.
181 However, we can get left with tiresome type applications. Notably, consider
182 f = /\ a -> let t = e in (t, w)
183 Then lifting the let out of the big lambda gives
185 f = /\ a -> let t = inline_me (t' a) in (t, w)
186 The inline_me is to stop the simplifier inlining t' right back
187 into t's RHS. In the next phase we'll substitute for t (since
188 its rhs is trivial) and *then* we could get rid of the inline_me.
189 But it hardly seems worth it, so I don't bother.
192 mkInlineMe (Var v) = Var v
193 mkInlineMe e = Note InlineMe e
199 mkCoerce :: Type -> CoreExpr -> CoreExpr
200 mkCoerce to_ty expr = mkCoerce2 to_ty (exprType expr) expr
202 mkCoerce2 :: Type -> Type -> CoreExpr -> CoreExpr
203 mkCoerce2 to_ty from_ty (Note (Coerce to_ty2 from_ty2) expr)
204 = ASSERT( from_ty `eqType` to_ty2 )
205 mkCoerce2 to_ty from_ty2 expr
207 mkCoerce2 to_ty from_ty expr
208 | to_ty `eqType` from_ty = expr
209 | otherwise = ASSERT( from_ty `eqType` exprType expr )
210 Note (Coerce to_ty from_ty) expr
214 mkSCC :: CostCentre -> Expr b -> Expr b
215 -- Note: Nested SCC's *are* preserved for the benefit of
216 -- cost centre stack profiling
217 -- Note2: We throw away an SCC on a single variable. If the
218 -- variable is a value, then there is no work to do in
219 -- evaluating it, and if it is a thunk, then it will be
220 -- attributed to its own CCS anyhow.
221 mkSCC cc (Lit lit) = Lit lit
222 mkSCC cc (Var v) = Var v
223 mkSCC cc (Lam x e) = Lam x (mkSCC cc e) -- Move _scc_ inside lambda
224 mkSCC cc (Note (SCC cc') e) = Note (SCC cc) (Note (SCC cc') e)
225 mkSCC cc (Note n e) = Note n (mkSCC cc e) -- Move _scc_ inside notes
226 mkSCC cc expr = Note (SCC cc) expr
230 %************************************************************************
232 \subsection{Other expression construction}
234 %************************************************************************
237 bindNonRec :: Id -> CoreExpr -> CoreExpr -> CoreExpr
238 -- (bindNonRec x r b) produces either
241 -- case r of x { _DEFAULT_ -> b }
243 -- depending on whether x is unlifted or not
244 -- It's used by the desugarer to avoid building bindings
245 -- that give Core Lint a heart attack. Actually the simplifier
246 -- deals with them perfectly well.
247 bindNonRec bndr rhs body
248 | needsCaseBinding (idType bndr) rhs = Case rhs bndr [(DEFAULT,[],body)]
249 | otherwise = Let (NonRec bndr rhs) body
251 needsCaseBinding ty rhs = isUnLiftedType ty && not (exprOkForSpeculation rhs)
252 -- Make a case expression instead of a let
253 -- These can arise either from the desugarer,
254 -- or from beta reductions: (\x.e) (x +# y)
258 mkAltExpr :: AltCon -> [CoreBndr] -> [Type] -> CoreExpr
259 -- This guy constructs the value that the scrutinee must have
260 -- when you are in one particular branch of a case
261 mkAltExpr (DataAlt con) args inst_tys
262 = mkConApp con (map Type inst_tys ++ map varToCoreExpr args)
263 mkAltExpr (LitAlt lit) [] []
266 mkIfThenElse :: CoreExpr -> CoreExpr -> CoreExpr -> CoreExpr
267 mkIfThenElse guard then_expr else_expr
268 = Case guard (mkWildId boolTy)
269 [ (DataAlt trueDataCon, [], then_expr),
270 (DataAlt falseDataCon, [], else_expr) ]
274 %************************************************************************
276 \subsection{Taking expressions apart}
278 %************************************************************************
280 The default alternative must be first, if it exists at all.
281 This makes it easy to find, though it makes matching marginally harder.
284 hasDefault :: [CoreAlt] -> Bool
285 hasDefault ((DEFAULT,_,_) : alts) = True
288 findDefault :: [CoreAlt] -> ([CoreAlt], Maybe CoreExpr)
289 findDefault ((DEFAULT,args,rhs) : alts) = ASSERT( null args ) (alts, Just rhs)
290 findDefault alts = (alts, Nothing)
292 findAlt :: AltCon -> [CoreAlt] -> CoreAlt
295 (deflt@(DEFAULT,_,_):alts) -> go alts deflt
296 other -> go alts panic_deflt
299 panic_deflt = pprPanic "Missing alternative" (ppr con $$ vcat (map ppr alts))
302 go (alt@(con1,_,_) : alts) deflt | con == con1 = alt
303 | otherwise = ASSERT( not (con1 == DEFAULT) )
308 %************************************************************************
310 \subsection{Figuring out things about expressions}
312 %************************************************************************
314 @exprIsTrivial@ is true of expressions we are unconditionally happy to
315 duplicate; simple variables and constants, and type
316 applications. Note that primop Ids aren't considered
319 @exprIsBottom@ is true of expressions that are guaranteed to diverge
322 There used to be a gruesome test for (hasNoBinding v) in the
324 exprIsTrivial (Var v) | hasNoBinding v = idArity v == 0
325 The idea here is that a constructor worker, like $wJust, is
326 really short for (\x -> $wJust x), becuase $wJust has no binding.
327 So it should be treated like a lambda. Ditto unsaturated primops.
328 But now constructor workers are not "have-no-binding" Ids. And
329 completely un-applied primops and foreign-call Ids are sufficiently
330 rare that I plan to allow them to be duplicated and put up with
334 exprIsTrivial (Var v) = True -- See notes above
335 exprIsTrivial (Type _) = True
336 exprIsTrivial (Lit lit) = litIsTrivial lit
337 exprIsTrivial (App e arg) = not (isRuntimeArg arg) && exprIsTrivial e
338 exprIsTrivial (Note _ e) = exprIsTrivial e
339 exprIsTrivial (Lam b body) = not (isRuntimeVar b) && exprIsTrivial body
340 exprIsTrivial other = False
344 @exprIsDupable@ is true of expressions that can be duplicated at a modest
345 cost in code size. This will only happen in different case
346 branches, so there's no issue about duplicating work.
348 That is, exprIsDupable returns True of (f x) even if
349 f is very very expensive to call.
351 Its only purpose is to avoid fruitless let-binding
352 and then inlining of case join points
356 exprIsDupable (Type _) = True
357 exprIsDupable (Var v) = True
358 exprIsDupable (Lit lit) = litIsDupable lit
359 exprIsDupable (Note InlineMe e) = True
360 exprIsDupable (Note _ e) = exprIsDupable e
364 go (Var v) n_args = True
365 go (App f a) n_args = n_args < dupAppSize
368 go other n_args = False
371 dupAppSize = 4 -- Size of application we are prepared to duplicate
374 @exprIsCheap@ looks at a Core expression and returns \tr{True} if
375 it is obviously in weak head normal form, or is cheap to get to WHNF.
376 [Note that that's not the same as exprIsDupable; an expression might be
377 big, and hence not dupable, but still cheap.]
379 By ``cheap'' we mean a computation we're willing to:
380 push inside a lambda, or
381 inline at more than one place
382 That might mean it gets evaluated more than once, instead of being
383 shared. The main examples of things which aren't WHNF but are
388 (where e, and all the ei are cheap)
391 (where e and b are cheap)
394 (where op is a cheap primitive operator)
397 (because we are happy to substitute it inside a lambda)
399 Notice that a variable is considered 'cheap': we can push it inside a lambda,
400 because sharing will make sure it is only evaluated once.
403 exprIsCheap :: CoreExpr -> Bool
404 exprIsCheap (Lit lit) = True
405 exprIsCheap (Type _) = True
406 exprIsCheap (Var _) = True
407 exprIsCheap (Note InlineMe e) = True
408 exprIsCheap (Note _ e) = exprIsCheap e
409 exprIsCheap (Lam x e) = isRuntimeVar x || exprIsCheap e
410 exprIsCheap (Case e _ alts) = exprIsCheap e &&
411 and [exprIsCheap rhs | (_,_,rhs) <- alts]
412 -- Experimentally, treat (case x of ...) as cheap
413 -- (and case __coerce x etc.)
414 -- This improves arities of overloaded functions where
415 -- there is only dictionary selection (no construction) involved
416 exprIsCheap (Let (NonRec x _) e)
417 | isUnLiftedType (idType x) = exprIsCheap e
419 -- strict lets always have cheap right hand sides, and
422 exprIsCheap other_expr
423 = go other_expr 0 True
425 go (Var f) n_args args_cheap
426 = (idAppIsCheap f n_args && args_cheap)
427 -- A constructor, cheap primop, or partial application
429 || idAppIsBottom f n_args
430 -- Application of a function which
431 -- always gives bottom; we treat this as cheap
432 -- because it certainly doesn't need to be shared!
434 go (App f a) n_args args_cheap
435 | not (isRuntimeArg a) = go f n_args args_cheap
436 | otherwise = go f (n_args + 1) (exprIsCheap a && args_cheap)
438 go other n_args args_cheap = False
440 idAppIsCheap :: Id -> Int -> Bool
441 idAppIsCheap id n_val_args
442 | n_val_args == 0 = True -- Just a type application of
443 -- a variable (f t1 t2 t3)
445 | otherwise = case globalIdDetails id of
446 DataConWorkId _ -> True
447 RecordSelId _ -> True -- I'm experimenting with making record selection
448 ClassOpId _ -> True -- look cheap, so we will substitute it inside a
449 -- lambda. Particularly for dictionary field selection
451 PrimOpId op -> primOpIsCheap op -- In principle we should worry about primops
452 -- that return a type variable, since the result
453 -- might be applied to something, but I'm not going
454 -- to bother to check the number of args
455 other -> n_val_args < idArity id
458 exprOkForSpeculation returns True of an expression that it is
460 * safe to evaluate even if normal order eval might not
461 evaluate the expression at all, or
463 * safe *not* to evaluate even if normal order would do so
467 the expression guarantees to terminate,
469 without raising an exception,
470 without causing a side effect (e.g. writing a mutable variable)
473 let x = case y# +# 1# of { r# -> I# r# }
476 case y# +# 1# of { r# ->
481 We can only do this if the (y+1) is ok for speculation: it has no
482 side effects, and can't diverge or raise an exception.
485 exprOkForSpeculation :: CoreExpr -> Bool
486 exprOkForSpeculation (Lit _) = True
487 exprOkForSpeculation (Type _) = True
488 exprOkForSpeculation (Var v) = isUnLiftedType (idType v)
489 exprOkForSpeculation (Note _ e) = exprOkForSpeculation e
490 exprOkForSpeculation other_expr
491 = case collectArgs other_expr of
492 (Var f, args) -> spec_ok (globalIdDetails f) args
496 spec_ok (DataConWorkId _) args
497 = True -- The strictness of the constructor has already
498 -- been expressed by its "wrapper", so we don't need
499 -- to take the arguments into account
501 spec_ok (PrimOpId op) args
502 | isDivOp op, -- Special case for dividing operations that fail
503 [arg1, Lit lit] <- args -- only if the divisor is zero
504 = not (isZeroLit lit) && exprOkForSpeculation arg1
505 -- Often there is a literal divisor, and this
506 -- can get rid of a thunk in an inner looop
509 = primOpOkForSpeculation op &&
510 all exprOkForSpeculation args
511 -- A bit conservative: we don't really need
512 -- to care about lazy arguments, but this is easy
514 spec_ok other args = False
516 isDivOp :: PrimOp -> Bool
517 -- True of dyadic operators that can fail
518 -- only if the second arg is zero
519 -- This function probably belongs in PrimOp, or even in
520 -- an automagically generated file.. but it's such a
521 -- special case I thought I'd leave it here for now.
522 isDivOp IntQuotOp = True
523 isDivOp IntRemOp = True
524 isDivOp WordQuotOp = True
525 isDivOp WordRemOp = True
526 isDivOp IntegerQuotRemOp = True
527 isDivOp IntegerDivModOp = True
528 isDivOp FloatDivOp = True
529 isDivOp DoubleDivOp = True
530 isDivOp other = False
535 exprIsBottom :: CoreExpr -> Bool -- True => definitely bottom
536 exprIsBottom e = go 0 e
538 -- n is the number of args
539 go n (Note _ e) = go n e
540 go n (Let _ e) = go n e
541 go n (Case e _ _) = go 0 e -- Just check the scrut
542 go n (App e _) = go (n+1) e
543 go n (Var v) = idAppIsBottom v n
545 go n (Lam _ _) = False
547 idAppIsBottom :: Id -> Int -> Bool
548 idAppIsBottom id n_val_args = appIsBottom (idNewStrictness id) n_val_args
551 @exprIsValue@ returns true for expressions that are certainly *already*
552 evaluated to *head* normal form. This is used to decide whether it's ok
555 case x of _ -> e ===> e
557 and to decide whether it's safe to discard a `seq`
559 So, it does *not* treat variables as evaluated, unless they say they are.
561 But it *does* treat partial applications and constructor applications
562 as values, even if their arguments are non-trivial, provided the argument
564 e.g. (:) (f x) (map f xs) is a value
565 map (...redex...) is a value
566 Because `seq` on such things completes immediately
568 For unlifted argument types, we have to be careful:
570 Suppose (f x) diverges; then C (f x) is not a value. True, but
571 this form is illegal (see the invariants in CoreSyn). Args of unboxed
572 type must be ok-for-speculation (or trivial).
575 exprIsValue :: CoreExpr -> Bool -- True => Value-lambda, constructor, PAP
576 exprIsValue (Var v) -- NB: There are no value args at this point
577 = isDataConWorkId v -- Catches nullary constructors,
578 -- so that [] and () are values, for example
579 || idArity v > 0 -- Catches (e.g.) primops that don't have unfoldings
580 || isEvaldUnfolding (idUnfolding v)
581 -- Check the thing's unfolding; it might be bound to a value
582 -- A worry: what if an Id's unfolding is just itself:
583 -- then we could get an infinite loop...
585 exprIsValue (Lit l) = True
586 exprIsValue (Type ty) = True -- Types are honorary Values;
587 -- we don't mind copying them
588 exprIsValue (Lam b e) = isRuntimeVar b || exprIsValue e
589 exprIsValue (Note _ e) = exprIsValue e
590 exprIsValue (App e (Type _)) = exprIsValue e
591 exprIsValue (App e a) = app_is_value e [a]
592 exprIsValue other = False
594 -- There is at least one value argument
595 app_is_value (Var fun) args
596 | isDataConWorkId fun -- Constructor apps are values
597 || idArity fun > valArgCount args -- Under-applied function
598 = check_args (idType fun) args
599 app_is_value (App f a) as = app_is_value f (a:as)
600 app_is_value other as = False
602 -- 'check_args' checks that unlifted-type args
603 -- are in fact guaranteed non-divergent
604 check_args fun_ty [] = True
605 check_args fun_ty (Type _ : args) = case splitForAllTy_maybe fun_ty of
606 Just (_, ty) -> check_args ty args
607 check_args fun_ty (arg : args)
608 | isUnLiftedType arg_ty = exprOkForSpeculation arg
609 | otherwise = check_args res_ty args
611 (arg_ty, res_ty) = splitFunTy fun_ty
615 exprIsConApp_maybe :: CoreExpr -> Maybe (DataCon, [CoreExpr])
616 exprIsConApp_maybe (Note (Coerce to_ty from_ty) expr)
617 = -- Maybe this is over the top, but here we try to turn
618 -- coerce (S,T) ( x, y )
620 -- ( coerce S x, coerce T y )
621 -- This happens in anger in PrelArrExts which has a coerce
622 -- case coerce memcpy a b of
624 -- where the memcpy is in the IO monad, but the call is in
626 case exprIsConApp_maybe expr of {
630 case splitTyConApp_maybe to_ty of {
632 Just (tc, tc_arg_tys) | tc /= dataConTyCon dc -> Nothing
633 | isExistentialDataCon dc -> Nothing
635 -- Type constructor must match
636 -- We knock out existentials to keep matters simple(r)
638 arity = tyConArity tc
639 val_args = drop arity args
640 to_arg_tys = dataConArgTys dc tc_arg_tys
641 mk_coerce ty arg = mkCoerce ty arg
642 new_val_args = zipWith mk_coerce to_arg_tys val_args
644 ASSERT( all isTypeArg (take arity args) )
645 ASSERT( equalLength val_args to_arg_tys )
646 Just (dc, map Type tc_arg_tys ++ new_val_args)
649 exprIsConApp_maybe (Note _ expr)
650 = exprIsConApp_maybe expr
651 -- We ignore InlineMe notes in case we have
652 -- x = __inline_me__ (a,b)
653 -- All part of making sure that INLINE pragmas never hurt
654 -- Marcin tripped on this one when making dictionaries more inlinable
656 -- In fact, we ignore all notes. For example,
657 -- case _scc_ "foo" (C a b) of
659 -- should be optimised away, but it will be only if we look
660 -- through the SCC note.
662 exprIsConApp_maybe expr = analyse (collectArgs expr)
664 analyse (Var fun, args)
665 | Just con <- isDataConWorkId_maybe fun,
666 args `lengthAtLeast` dataConRepArity con
667 -- Might be > because the arity excludes type args
670 -- Look through unfoldings, but only cheap ones, because
671 -- we are effectively duplicating the unfolding
672 analyse (Var fun, [])
673 | let unf = idUnfolding fun,
675 = exprIsConApp_maybe (unfoldingTemplate unf)
677 analyse other = Nothing
682 %************************************************************************
684 \subsection{Eta reduction and expansion}
686 %************************************************************************
689 exprEtaExpandArity :: CoreExpr -> Arity
690 {- The Arity returned is the number of value args the
691 thing can be applied to without doing much work
693 exprEtaExpandArity is used when eta expanding
696 It returns 1 (or more) to:
697 case x of p -> \s -> ...
698 because for I/O ish things we really want to get that \s to the top.
699 We are prepared to evaluate x each time round the loop in order to get that
701 It's all a bit more subtle than it looks:
705 Consider one-shot lambdas
706 let x = expensive in \y z -> E
707 We want this to have arity 2 if the \y-abstraction is a 1-shot lambda
708 Hence the ArityType returned by arityType
710 2. The state-transformer hack
712 The one-shot lambda special cause is particularly important/useful for
713 IO state transformers, where we often get
714 let x = E in \ s -> ...
716 and the \s is a real-world state token abstraction. Such abstractions
717 are almost invariably 1-shot, so we want to pull the \s out, past the
718 let x=E, even if E is expensive. So we treat state-token lambdas as
719 one-shot even if they aren't really. The hack is in Id.isOneShotLambda.
721 3. Dealing with bottom
724 f = \x -> error "foo"
725 Here, arity 1 is fine. But if it is
729 then we want to get arity 2. Tecnically, this isn't quite right, because
731 should diverge, but it'll converge if we eta-expand f. Nevertheless, we
732 do so; it improves some programs significantly, and increasing convergence
733 isn't a bad thing. Hence the ABot/ATop in ArityType.
735 Actually, the situation is worse. Consider
739 Can we eta-expand here? At first the answer looks like "yes of course", but
742 This should diverge! But if we eta-expand, it won't. Again, we ignore this
743 "problem", because being scrupulous would lose an important transformation for
748 exprEtaExpandArity e = arityDepth (arityType e)
750 -- A limited sort of function type
751 data ArityType = AFun Bool ArityType -- True <=> one-shot
752 | ATop -- Know nothing
755 arityDepth :: ArityType -> Arity
756 arityDepth (AFun _ ty) = 1 + arityDepth ty
759 andArityType ABot at2 = at2
760 andArityType ATop at2 = ATop
761 andArityType (AFun t1 at1) (AFun t2 at2) = AFun (t1 && t2) (andArityType at1 at2)
762 andArityType at1 at2 = andArityType at2 at1
764 arityType :: CoreExpr -> ArityType
765 -- (go1 e) = [b1,..,bn]
766 -- means expression can be rewritten \x_b1 -> ... \x_bn -> body
767 -- where bi is True <=> the lambda is one-shot
769 arityType (Note n e) = arityType e
770 -- Not needed any more: etaExpand is cleverer
771 -- | ok_note n = arityType e
772 -- | otherwise = ATop
777 mk :: Arity -> ArityType
778 mk 0 | isBottomingId v = ABot
780 mk n = AFun False (mk (n-1))
782 -- When the type of the Id encodes one-shot-ness,
783 -- use the idinfo here
785 -- Lambdas; increase arity
786 arityType (Lam x e) | isId x = AFun (isOneShotLambda x || isStateHack x) (arityType e)
787 | otherwise = arityType e
789 -- Applications; decrease arity
790 arityType (App f (Type _)) = arityType f
791 arityType (App f a) = case arityType f of
792 AFun one_shot xs | exprIsCheap a -> xs
795 -- Case/Let; keep arity if either the expression is cheap
796 -- or it's a 1-shot lambda
797 arityType (Case scrut _ alts) = case foldr1 andArityType [arityType rhs | (_,_,rhs) <- alts] of
798 xs@(AFun one_shot _) | one_shot -> xs
799 xs | exprIsCheap scrut -> xs
802 arityType (Let b e) = case arityType e of
803 xs@(AFun one_shot _) | one_shot -> xs
804 xs | all exprIsCheap (rhssOfBind b) -> xs
807 arityType other = ATop
809 isStateHack id = case splitTyConApp_maybe (idType id) of
810 Just (tycon,_) | tycon == statePrimTyCon -> True
813 -- The last clause is a gross hack. It claims that
814 -- every function over realWorldStatePrimTy is a one-shot
815 -- function. This is pretty true in practice, and makes a big
816 -- difference. For example, consider
817 -- a `thenST` \ r -> ...E...
818 -- The early full laziness pass, if it doesn't know that r is one-shot
819 -- will pull out E (let's say it doesn't mention r) to give
820 -- let lvl = E in a `thenST` \ r -> ...lvl...
821 -- When `thenST` gets inlined, we end up with
822 -- let lvl = E in \s -> case a s of (r, s') -> ...lvl...
823 -- and we don't re-inline E.
825 -- It would be better to spot that r was one-shot to start with, but
826 -- I don't want to rely on that.
828 -- Another good example is in fill_in in PrelPack.lhs. We should be able to
829 -- spot that fill_in has arity 2 (and when Keith is done, we will) but we can't yet.
831 {- NOT NEEDED ANY MORE: etaExpand is cleverer
832 ok_note InlineMe = False
834 -- Notice that we do not look through __inline_me__
835 -- This may seem surprising, but consider
836 -- f = _inline_me (\x -> e)
837 -- We DO NOT want to eta expand this to
838 -- f = \x -> (_inline_me (\x -> e)) x
839 -- because the _inline_me gets dropped now it is applied,
848 etaExpand :: Arity -- Result should have this number of value args
850 -> CoreExpr -> Type -- Expression and its type
852 -- (etaExpand n us e ty) returns an expression with
853 -- the same meaning as 'e', but with arity 'n'.
855 -- Given e' = etaExpand n us e ty
857 -- ty = exprType e = exprType e'
859 -- Note that SCCs are not treated specially. If we have
860 -- etaExpand 2 (\x -> scc "foo" e)
861 -- = (\xy -> (scc "foo" e) y)
862 -- So the costs of evaluating 'e' (not 'e y') are attributed to "foo"
864 etaExpand n us expr ty
865 | manifestArity expr >= n = expr -- The no-op case
866 | otherwise = eta_expand n us expr ty
869 -- manifestArity sees how many leading value lambdas there are
870 manifestArity :: CoreExpr -> Arity
871 manifestArity (Lam v e) | isId v = 1 + manifestArity e
872 | otherwise = manifestArity e
873 manifestArity (Note _ e) = manifestArity e
876 -- etaExpand deals with for-alls. For example:
878 -- where E :: forall a. a -> a
880 -- (/\b. \y::a -> E b y)
882 -- It deals with coerces too, though they are now rare
883 -- so perhaps the extra code isn't worth it
885 eta_expand n us expr ty
887 -- The ILX code generator requires eta expansion for type arguments
888 -- too, but alas the 'n' doesn't tell us how many of them there
889 -- may be. So we eagerly eta expand any big lambdas, and just
890 -- cross our fingers about possible loss of sharing in the ILX case.
891 -- The Right Thing is probably to make 'arity' include
892 -- type variables throughout the compiler. (ToDo.)
894 -- Saturated, so nothing to do
897 -- Short cut for the case where there already
898 -- is a lambda; no point in gratuitously adding more
899 eta_expand n us (Lam v body) ty
901 = Lam v (eta_expand n us body (applyTy ty (mkTyVarTy v)))
904 = Lam v (eta_expand (n-1) us body (funResultTy ty))
906 -- We used to have a special case that stepped inside Coerces here,
907 -- thus: eta_expand n us (Note note@(Coerce _ ty) e) _
908 -- = Note note (eta_expand n us e ty)
909 -- BUT this led to an infinite loop
910 -- Example: newtype T = MkT (Int -> Int)
911 -- eta_expand 1 (coerce (Int->Int) e)
912 -- --> coerce (Int->Int) (eta_expand 1 T e)
914 -- --> coerce (Int->Int) (coerce T
915 -- (\x::Int -> eta_expand 1 (coerce (Int->Int) e)))
916 -- by the splitNewType_maybe case below
919 eta_expand n us expr ty
920 = case splitForAllTy_maybe ty of {
921 Just (tv,ty') -> Lam tv (eta_expand n us (App expr (Type (mkTyVarTy tv))) ty')
925 case splitFunTy_maybe ty of {
926 Just (arg_ty, res_ty) -> Lam arg1 (eta_expand (n-1) us2 (App expr (Var arg1)) res_ty)
928 arg1 = mkSysLocal FSLIT("eta") uniq arg_ty
934 -- newtype T = MkT (Int -> Int)
935 -- Consider eta-expanding this
938 -- coerce T (\x::Int -> (coerce (Int->Int) e) x)
940 case splitNewType_maybe ty of {
941 Just ty' -> mkCoerce2 ty ty' (eta_expand n us (mkCoerce2 ty' ty expr) ty') ;
942 Nothing -> pprTrace "Bad eta expand" (ppr expr $$ ppr ty) expr
946 exprArity is a cheap-and-cheerful version of exprEtaExpandArity.
947 It tells how many things the expression can be applied to before doing
948 any work. It doesn't look inside cases, lets, etc. The idea is that
949 exprEtaExpandArity will do the hard work, leaving something that's easy
950 for exprArity to grapple with. In particular, Simplify uses exprArity to
951 compute the ArityInfo for the Id.
953 Originally I thought that it was enough just to look for top-level lambdas, but
954 it isn't. I've seen this
956 foo = PrelBase.timesInt
958 We want foo to get arity 2 even though the eta-expander will leave it
959 unchanged, in the expectation that it'll be inlined. But occasionally it
960 isn't, because foo is blacklisted (used in a rule).
962 Similarly, see the ok_note check in exprEtaExpandArity. So
963 f = __inline_me (\x -> e)
964 won't be eta-expanded.
966 And in any case it seems more robust to have exprArity be a bit more intelligent.
967 But note that (\x y z -> f x y z)
968 should have arity 3, regardless of f's arity.
971 exprArity :: CoreExpr -> Arity
974 go (Var v) = idArity v
975 go (Lam x e) | isId x = go e + 1
978 go (App e (Type t)) = go e
979 go (App f a) | exprIsCheap a = (go f - 1) `max` 0
980 -- NB: exprIsCheap a!
981 -- f (fac x) does not have arity 2,
982 -- even if f has arity 3!
983 -- NB: `max 0`! (\x y -> f x) has arity 2, even if f is
984 -- unknown, hence arity 0
988 %************************************************************************
990 \subsection{Equality}
992 %************************************************************************
994 @cheapEqExpr@ is a cheap equality test which bales out fast!
995 True => definitely equal
996 False => may or may not be equal
999 cheapEqExpr :: Expr b -> Expr b -> Bool
1001 cheapEqExpr (Var v1) (Var v2) = v1==v2
1002 cheapEqExpr (Lit lit1) (Lit lit2) = lit1 == lit2
1003 cheapEqExpr (Type t1) (Type t2) = t1 `eqType` t2
1005 cheapEqExpr (App f1 a1) (App f2 a2)
1006 = f1 `cheapEqExpr` f2 && a1 `cheapEqExpr` a2
1008 cheapEqExpr _ _ = False
1010 exprIsBig :: Expr b -> Bool
1011 -- Returns True of expressions that are too big to be compared by cheapEqExpr
1012 exprIsBig (Lit _) = False
1013 exprIsBig (Var v) = False
1014 exprIsBig (Type t) = False
1015 exprIsBig (App f a) = exprIsBig f || exprIsBig a
1016 exprIsBig other = True
1021 eqExpr :: CoreExpr -> CoreExpr -> Bool
1022 -- Works ok at more general type, but only needed at CoreExpr
1023 -- Used in rule matching, so when we find a type we use
1024 -- eqTcType, which doesn't look through newtypes
1025 -- [And it doesn't risk falling into a black hole either.]
1027 = eq emptyVarEnv e1 e2
1029 -- The "env" maps variables in e1 to variables in ty2
1030 -- So when comparing lambdas etc,
1031 -- we in effect substitute v2 for v1 in e1 before continuing
1032 eq env (Var v1) (Var v2) = case lookupVarEnv env v1 of
1033 Just v1' -> v1' == v2
1036 eq env (Lit lit1) (Lit lit2) = lit1 == lit2
1037 eq env (App f1 a1) (App f2 a2) = eq env f1 f2 && eq env a1 a2
1038 eq env (Lam v1 e1) (Lam v2 e2) = eq (extendVarEnv env v1 v2) e1 e2
1039 eq env (Let (NonRec v1 r1) e1)
1040 (Let (NonRec v2 r2) e2) = eq env r1 r2 && eq (extendVarEnv env v1 v2) e1 e2
1041 eq env (Let (Rec ps1) e1)
1042 (Let (Rec ps2) e2) = equalLength ps1 ps2 &&
1043 and (zipWith eq_rhs ps1 ps2) &&
1046 env' = extendVarEnvList env [(v1,v2) | ((v1,_),(v2,_)) <- zip ps1 ps2]
1047 eq_rhs (_,r1) (_,r2) = eq env' r1 r2
1048 eq env (Case e1 v1 a1)
1049 (Case e2 v2 a2) = eq env e1 e2 &&
1050 equalLength a1 a2 &&
1051 and (zipWith (eq_alt env') a1 a2)
1053 env' = extendVarEnv env v1 v2
1055 eq env (Note n1 e1) (Note n2 e2) = eq_note env n1 n2 && eq env e1 e2
1056 eq env (Type t1) (Type t2) = t1 `eqType` t2
1057 eq env e1 e2 = False
1059 eq_list env [] [] = True
1060 eq_list env (e1:es1) (e2:es2) = eq env e1 e2 && eq_list env es1 es2
1061 eq_list env es1 es2 = False
1063 eq_alt env (c1,vs1,r1) (c2,vs2,r2) = c1==c2 &&
1064 eq (extendVarEnvList env (vs1 `zip` vs2)) r1 r2
1066 eq_note env (SCC cc1) (SCC cc2) = cc1 == cc2
1067 eq_note env (Coerce t1 f1) (Coerce t2 f2) = t1 `eqType` t2 && f1 `eqType` f2
1068 eq_note env InlineCall InlineCall = True
1069 eq_note env (CoreNote s1) (CoreNote s2) = s1 == s2
1070 eq_note env other1 other2 = False
1074 %************************************************************************
1076 \subsection{The size of an expression}
1078 %************************************************************************
1081 coreBindsSize :: [CoreBind] -> Int
1082 coreBindsSize bs = foldr ((+) . bindSize) 0 bs
1084 exprSize :: CoreExpr -> Int
1085 -- A measure of the size of the expressions
1086 -- It also forces the expression pretty drastically as a side effect
1087 exprSize (Var v) = v `seq` 1
1088 exprSize (Lit lit) = lit `seq` 1
1089 exprSize (App f a) = exprSize f + exprSize a
1090 exprSize (Lam b e) = varSize b + exprSize e
1091 exprSize (Let b e) = bindSize b + exprSize e
1092 exprSize (Case e b as) = exprSize e + varSize b + foldr ((+) . altSize) 0 as
1093 exprSize (Note n e) = noteSize n + exprSize e
1094 exprSize (Type t) = seqType t `seq` 1
1096 noteSize (SCC cc) = cc `seq` 1
1097 noteSize (Coerce t1 t2) = seqType t1 `seq` seqType t2 `seq` 1
1098 noteSize InlineCall = 1
1099 noteSize InlineMe = 1
1100 noteSize (CoreNote s) = s `seq` 1 -- hdaume: core annotations
1102 varSize :: Var -> Int
1103 varSize b | isTyVar b = 1
1104 | otherwise = seqType (idType b) `seq`
1105 megaSeqIdInfo (idInfo b) `seq`
1108 varsSize = foldr ((+) . varSize) 0
1110 bindSize (NonRec b e) = varSize b + exprSize e
1111 bindSize (Rec prs) = foldr ((+) . pairSize) 0 prs
1113 pairSize (b,e) = varSize b + exprSize e
1115 altSize (c,bs,e) = c `seq` varsSize bs + exprSize e
1119 %************************************************************************
1121 \subsection{Hashing}
1123 %************************************************************************
1126 hashExpr :: CoreExpr -> Int
1127 hashExpr e | hash < 0 = 77 -- Just in case we hit -maxInt
1130 hash = abs (hash_expr e) -- Negative numbers kill UniqFM
1132 hash_expr (Note _ e) = hash_expr e
1133 hash_expr (Let (NonRec b r) e) = hashId b
1134 hash_expr (Let (Rec ((b,r):_)) e) = hashId b
1135 hash_expr (Case _ b _) = hashId b
1136 hash_expr (App f e) = hash_expr f * fast_hash_expr e
1137 hash_expr (Var v) = hashId v
1138 hash_expr (Lit lit) = hashLiteral lit
1139 hash_expr (Lam b _) = hashId b
1140 hash_expr (Type t) = trace "hash_expr: type" 1 -- Shouldn't happen
1142 fast_hash_expr (Var v) = hashId v
1143 fast_hash_expr (Lit lit) = hashLiteral lit
1144 fast_hash_expr (App f (Type _)) = fast_hash_expr f
1145 fast_hash_expr (App f a) = fast_hash_expr a
1146 fast_hash_expr (Lam b _) = hashId b
1147 fast_hash_expr other = 1
1150 hashId id = hashName (idName id)
1153 %************************************************************************
1155 \subsection{Determining non-updatable right-hand-sides}
1157 %************************************************************************
1159 Top-level constructor applications can usually be allocated
1160 statically, but they can't if
1161 a) the constructor, or any of the arguments, come from another DLL
1162 b) any of the arguments are LitLits
1163 (because we can't refer to static labels in other DLLs).
1165 If this happens we simply make the RHS into an updatable thunk,
1166 and 'exectute' it rather than allocating it statically.
1169 rhsIsStatic :: CoreExpr -> Bool
1170 -- This function is called only on *top-level* right-hand sides
1171 -- Returns True if the RHS can be allocated statically, with
1172 -- no thunks involved at all.
1174 -- It's called (i) in TidyPgm.hasCafRefs to decide if the rhs is, or
1175 -- refers to, CAFs; and (ii) in CoreToStg to decide whether to put an
1176 -- update flag on it.
1178 -- The basic idea is that rhsIsStatic returns True only if the RHS is
1179 -- (a) a value lambda
1180 -- (b) a saturated constructor application with static args
1182 -- BUT watch out for
1183 -- (i) Any cross-DLL references kill static-ness completely
1184 -- because they must be 'executed' not statically allocated
1186 -- (ii) We treat partial applications as redexes, because in fact we
1187 -- make a thunk for them that runs and builds a PAP
1188 -- at run-time. The only appliations that are treated as
1189 -- static are *saturated* applications of constructors.
1191 -- We used to try to be clever with nested structures like this:
1192 -- ys = (:) w ((:) w [])
1193 -- on the grounds that CorePrep will flatten ANF-ise it later.
1194 -- But supporting this special case made the function much more
1195 -- complicated, because the special case only applies if there are no
1196 -- enclosing type lambdas:
1197 -- ys = /\ a -> Foo (Baz ([] a))
1198 -- Here the nested (Baz []) won't float out to top level in CorePrep.
1200 -- But in fact, even without -O, nested structures at top level are
1201 -- flattened by the simplifier, so we don't need to be super-clever here.
1205 -- f = \x::Int. x+7 TRUE
1206 -- p = (True,False) TRUE
1208 -- d = (fst p, False) FALSE because there's a redex inside
1209 -- (this particular one doesn't happen but...)
1211 -- h = D# (1.0## /## 2.0##) FALSE (redex again)
1212 -- n = /\a. Nil a TRUE
1214 -- t = /\a. (:) (case w a of ...) (Nil a) FALSE (redex)
1217 -- This is a bit like CoreUtils.exprIsValue, with the following differences:
1218 -- a) scc "foo" (\x -> ...) is updatable (so we catch the right SCC)
1220 -- b) (C x xs), where C is a contructors is updatable if the application is
1223 -- c) don't look through unfolding of f in (f x).
1225 -- When opt_RuntimeTypes is on, we keep type lambdas and treat
1226 -- them as making the RHS re-entrant (non-updatable).
1228 rhsIsStatic rhs = is_static False rhs
1230 is_static :: Bool -- True <=> in a constructor argument; must be atomic
1233 is_static False (Lam b e) = isRuntimeVar b || is_static False e
1235 is_static in_arg (Note (SCC _) e) = False
1236 is_static in_arg (Note _ e) = is_static in_arg e
1238 is_static in_arg (Lit lit) = not (isLitLitLit lit)
1239 -- lit-lit arguments cannot be used in static constructors either.
1240 -- (litlits are deprecated, so I'm not going to bother cleaning up this infelicity --SDM).
1242 is_static in_arg other_expr = go other_expr 0
1244 go (Var f) n_val_args
1245 | not (isDllName (idName f))
1246 = saturated_data_con f n_val_args
1247 || (in_arg && n_val_args == 0)
1248 -- A naked un-applied variable is *not* deemed a static RHS
1250 -- Reason: better to update so that the indirection gets shorted
1251 -- out, and the true value will be seen
1252 -- NB: if you change this, you'll break the invariant that THUNK_STATICs
1253 -- are always updatable. If you do so, make sure that non-updatable
1254 -- ones have enough space for their static link field!
1256 go (App f a) n_val_args
1257 | isTypeArg a = go f n_val_args
1258 | not in_arg && is_static True a = go f (n_val_args + 1)
1259 -- The (not in_arg) checks that we aren't in a constructor argument;
1260 -- if we are, we don't allow (value) applications of any sort
1262 -- NB. In case you wonder, args are sometimes not atomic. eg.
1263 -- x = D# (1.0## /## 2.0##)
1264 -- can't float because /## can fail.
1266 go (Note (SCC _) f) n_val_args = False
1267 go (Note _ f) n_val_args = go f n_val_args
1269 go other n_val_args = False
1271 saturated_data_con f n_val_args
1272 = case isDataConWorkId_maybe f of
1273 Just dc -> n_val_args == dataConRepArity dc