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, mergeAlts,
16 -- Properties of expressions
17 exprType, coreAltType,
18 exprIsDupable, exprIsTrivial, exprIsCheap,
19 exprIsHNF,exprOkForSpeculation, exprIsBig,
20 exprIsConApp_maybe, exprIsBottom,
23 -- Arity and eta expansion
24 manifestArity, exprArity,
25 exprEtaExpandArity, etaExpand,
34 cheapEqExpr, tcEqExpr, tcEqExprX, applyTypeToArgs, applyTypeToArg
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 )
50 import Packages ( isDllName )
52 import Literal ( hashLiteral, literalType, litIsDupable,
53 litIsTrivial, isZeroLit, Literal( MachLabel ) )
54 import DataCon ( DataCon, dataConRepArity, dataConInstArgTys,
55 isVanillaDataCon, dataConTyCon )
56 import PrimOp ( PrimOp(..), primOpOkForSpeculation, primOpIsCheap )
57 import Id ( Id, idType, globalIdDetails, idNewStrictness,
58 mkWildId, idArity, idName, idUnfolding, idInfo,
59 isOneShotBndr, isStateHackType, isDataConWorkId_maybe, mkSysLocal,
60 isDataConWorkId, isBottomingId, isDictId
62 import IdInfo ( GlobalIdDetails(..), megaSeqIdInfo )
63 import NewDemand ( appIsBottom )
64 import Type ( Type, mkFunTy, mkForAllTy, splitFunTy_maybe,
65 splitFunTy, tcEqTypeX,
66 applyTys, isUnLiftedType, seqType, mkTyVarTy,
67 splitForAllTy_maybe, isForAllTy, splitRecNewType_maybe,
68 splitTyConApp_maybe, coreEqType, funResultTy, applyTy
70 import TyCon ( tyConArity )
71 import TysWiredIn ( boolTy, trueDataCon, falseDataCon )
72 import CostCentre ( CostCentre )
73 import BasicTypes ( Arity )
74 import PackageConfig ( PackageId )
75 import Unique ( Unique )
77 import DynFlags ( DynFlags, DynFlag(Opt_DictsCheap), dopt )
78 import TysPrim ( alphaTy ) -- Debugging only
79 import Util ( equalLength, lengthAtLeast, foldl2 )
83 %************************************************************************
85 \subsection{Find the type of a Core atom/expression}
87 %************************************************************************
90 exprType :: CoreExpr -> Type
92 exprType (Var var) = idType var
93 exprType (Lit lit) = literalType lit
94 exprType (Let _ body) = exprType body
95 exprType (Case _ _ ty alts) = ty
96 exprType (Note (Coerce ty _) e) = ty -- **! should take usage from e
97 exprType (Note other_note e) = exprType e
98 exprType (Lam binder expr) = mkPiType binder (exprType expr)
100 = case collectArgs e of
101 (fun, args) -> applyTypeToArgs e (exprType fun) args
103 exprType other = pprTrace "exprType" (pprCoreExpr other) alphaTy
105 coreAltType :: CoreAlt -> Type
106 coreAltType (_,_,rhs) = exprType rhs
109 @mkPiType@ makes a (->) type or a forall type, depending on whether
110 it is given a type variable or a term variable. We cleverly use the
111 lbvarinfo field to figure out the right annotation for the arrove in
112 case of a term variable.
115 mkPiType :: Var -> Type -> Type -- The more polymorphic version
116 mkPiTypes :: [Var] -> Type -> Type -- doesn't work...
118 mkPiTypes vs ty = foldr mkPiType ty vs
121 | isId v = mkFunTy (idType v) ty
122 | otherwise = mkForAllTy v ty
126 applyTypeToArg :: Type -> CoreExpr -> Type
127 applyTypeToArg fun_ty (Type arg_ty) = applyTy fun_ty arg_ty
128 applyTypeToArg fun_ty other_arg = funResultTy fun_ty
130 applyTypeToArgs :: CoreExpr -> Type -> [CoreExpr] -> Type
131 -- A more efficient version of applyTypeToArg
132 -- when we have several args
133 -- The first argument is just for debugging
134 applyTypeToArgs e op_ty [] = op_ty
136 applyTypeToArgs e op_ty (Type ty : args)
137 = -- Accumulate type arguments so we can instantiate all at once
140 go rev_tys (Type ty : args) = go (ty:rev_tys) args
141 go rev_tys rest_args = applyTypeToArgs e op_ty' rest_args
143 op_ty' = applyTys op_ty (reverse rev_tys)
145 applyTypeToArgs e op_ty (other_arg : args)
146 = case (splitFunTy_maybe op_ty) of
147 Just (_, res_ty) -> applyTypeToArgs e res_ty args
148 Nothing -> pprPanic "applyTypeToArgs" (pprCoreExpr e)
153 %************************************************************************
155 \subsection{Attaching notes}
157 %************************************************************************
159 mkNote removes redundant coercions, and SCCs where possible
163 mkNote :: Note -> CoreExpr -> CoreExpr
164 mkNote (Coerce to_ty from_ty) expr = mkCoerce2 to_ty from_ty expr
165 mkNote (SCC cc) expr = mkSCC cc expr
166 mkNote InlineMe expr = mkInlineMe expr
167 mkNote note expr = Note note expr
171 Drop trivial InlineMe's. This is somewhat important, because if we have an unfolding
172 that looks like (Note InlineMe (Var v)), the InlineMe doesn't go away because it may
173 not be *applied* to anything.
175 We don't use exprIsTrivial here, though, because we sometimes generate worker/wrapper
178 f = inline_me (coerce t fw)
179 As usual, the inline_me prevents the worker from getting inlined back into the wrapper.
180 We want the split, so that the coerces can cancel at the call site.
182 However, we can get left with tiresome type applications. Notably, consider
183 f = /\ a -> let t = e in (t, w)
184 Then lifting the let out of the big lambda gives
186 f = /\ a -> let t = inline_me (t' a) in (t, w)
187 The inline_me is to stop the simplifier inlining t' right back
188 into t's RHS. In the next phase we'll substitute for t (since
189 its rhs is trivial) and *then* we could get rid of the inline_me.
190 But it hardly seems worth it, so I don't bother.
193 mkInlineMe (Var v) = Var v
194 mkInlineMe e = Note InlineMe e
200 mkCoerce :: Type -> CoreExpr -> CoreExpr
201 mkCoerce to_ty expr = mkCoerce2 to_ty (exprType expr) expr
203 mkCoerce2 :: Type -> Type -> CoreExpr -> CoreExpr
204 mkCoerce2 to_ty from_ty (Note (Coerce to_ty2 from_ty2) expr)
205 = ASSERT( from_ty `coreEqType` to_ty2 )
206 mkCoerce2 to_ty from_ty2 expr
208 mkCoerce2 to_ty from_ty expr
209 | to_ty `coreEqType` from_ty = expr
210 | otherwise = ASSERT( from_ty `coreEqType` exprType expr )
211 Note (Coerce to_ty from_ty) expr
215 mkSCC :: CostCentre -> Expr b -> Expr b
216 -- Note: Nested SCC's *are* preserved for the benefit of
217 -- cost centre stack profiling
218 mkSCC cc (Lit lit) = Lit lit
219 mkSCC cc (Lam x e) = Lam x (mkSCC cc e) -- Move _scc_ inside lambda
220 mkSCC cc (Note (SCC cc') e) = Note (SCC cc) (Note (SCC cc') e)
221 mkSCC cc (Note n e) = Note n (mkSCC cc e) -- Move _scc_ inside notes
222 mkSCC cc expr = Note (SCC cc) expr
226 %************************************************************************
228 \subsection{Other expression construction}
230 %************************************************************************
233 bindNonRec :: Id -> CoreExpr -> CoreExpr -> CoreExpr
234 -- (bindNonRec x r b) produces either
237 -- case r of x { _DEFAULT_ -> b }
239 -- depending on whether x is unlifted or not
240 -- It's used by the desugarer to avoid building bindings
241 -- that give Core Lint a heart attack. Actually the simplifier
242 -- deals with them perfectly well.
244 bindNonRec bndr rhs body
245 | needsCaseBinding (idType bndr) rhs = Case rhs bndr (exprType body) [(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 -- Not going to be refining, so okay to take the type of the "then" clause
266 = Case guard (mkWildId boolTy) (exprType then_expr)
267 [ (DataAlt falseDataCon, [], else_expr), -- Increasing order of tag!
268 (DataAlt trueDataCon, [], then_expr) ]
272 %************************************************************************
274 \subsection{Taking expressions apart}
276 %************************************************************************
278 The default alternative must be first, if it exists at all.
279 This makes it easy to find, though it makes matching marginally harder.
282 findDefault :: [CoreAlt] -> ([CoreAlt], Maybe CoreExpr)
283 findDefault ((DEFAULT,args,rhs) : alts) = ASSERT( null args ) (alts, Just rhs)
284 findDefault alts = (alts, Nothing)
286 findAlt :: AltCon -> [CoreAlt] -> CoreAlt
289 (deflt@(DEFAULT,_,_):alts) -> go alts deflt
290 other -> go alts panic_deflt
292 panic_deflt = pprPanic "Missing alternative" (ppr con $$ vcat (map ppr alts))
295 go (alt@(con1,_,_) : alts) deflt
296 = case con `cmpAltCon` con1 of
297 LT -> deflt -- Missed it already; the alts are in increasing order
299 GT -> ASSERT( not (con1 == DEFAULT) ) go alts deflt
301 isDefaultAlt :: CoreAlt -> Bool
302 isDefaultAlt (DEFAULT, _, _) = True
303 isDefaultAlt other = False
305 ---------------------------------
306 mergeAlts :: [CoreAlt] -> [CoreAlt] -> [CoreAlt]
307 -- Merge preserving order; alternatives in the first arg
308 -- shadow ones in the second
309 mergeAlts [] as2 = as2
310 mergeAlts as1 [] = as1
311 mergeAlts (a1:as1) (a2:as2)
312 = case a1 `cmpAlt` a2 of
313 LT -> a1 : mergeAlts as1 (a2:as2)
314 EQ -> a1 : mergeAlts as1 as2 -- Discard a2
315 GT -> a2 : mergeAlts (a1:as1) as2
319 %************************************************************************
321 \subsection{Figuring out things about expressions}
323 %************************************************************************
325 @exprIsTrivial@ is true of expressions we are unconditionally happy to
326 duplicate; simple variables and constants, and type
327 applications. Note that primop Ids aren't considered
330 @exprIsBottom@ is true of expressions that are guaranteed to diverge
333 There used to be a gruesome test for (hasNoBinding v) in the
335 exprIsTrivial (Var v) | hasNoBinding v = idArity v == 0
336 The idea here is that a constructor worker, like $wJust, is
337 really short for (\x -> $wJust x), becuase $wJust has no binding.
338 So it should be treated like a lambda. Ditto unsaturated primops.
339 But now constructor workers are not "have-no-binding" Ids. And
340 completely un-applied primops and foreign-call Ids are sufficiently
341 rare that I plan to allow them to be duplicated and put up with
344 SCC notes. We do not treat (_scc_ "foo" x) as trivial, because
345 a) it really generates code, (and a heap object when it's
346 a function arg) to capture the cost centre
347 b) see the note [SCC-and-exprIsTrivial] in Simplify.simplLazyBind
350 exprIsTrivial (Var v) = True -- See notes above
351 exprIsTrivial (Type _) = True
352 exprIsTrivial (Lit lit) = litIsTrivial lit
353 exprIsTrivial (App e arg) = not (isRuntimeArg arg) && exprIsTrivial e
354 exprIsTrivial (Note (SCC _) e) = False -- See notes above
355 exprIsTrivial (Note _ e) = exprIsTrivial e
356 exprIsTrivial (Lam b body) = not (isRuntimeVar b) && exprIsTrivial body
357 exprIsTrivial other = False
361 @exprIsDupable@ is true of expressions that can be duplicated at a modest
362 cost in code size. This will only happen in different case
363 branches, so there's no issue about duplicating work.
365 That is, exprIsDupable returns True of (f x) even if
366 f is very very expensive to call.
368 Its only purpose is to avoid fruitless let-binding
369 and then inlining of case join points
373 exprIsDupable (Type _) = True
374 exprIsDupable (Var v) = True
375 exprIsDupable (Lit lit) = litIsDupable lit
376 exprIsDupable (Note InlineMe e) = True
377 exprIsDupable (Note _ e) = exprIsDupable e
381 go (Var v) n_args = True
382 go (App f a) n_args = n_args < dupAppSize
385 go other n_args = False
388 dupAppSize = 4 -- Size of application we are prepared to duplicate
391 @exprIsCheap@ looks at a Core expression and returns \tr{True} if
392 it is obviously in weak head normal form, or is cheap to get to WHNF.
393 [Note that that's not the same as exprIsDupable; an expression might be
394 big, and hence not dupable, but still cheap.]
396 By ``cheap'' we mean a computation we're willing to:
397 push inside a lambda, or
398 inline at more than one place
399 That might mean it gets evaluated more than once, instead of being
400 shared. The main examples of things which aren't WHNF but are
405 (where e, and all the ei are cheap)
408 (where e and b are cheap)
411 (where op is a cheap primitive operator)
414 (because we are happy to substitute it inside a lambda)
416 Notice that a variable is considered 'cheap': we can push it inside a lambda,
417 because sharing will make sure it is only evaluated once.
420 exprIsCheap :: CoreExpr -> Bool
421 exprIsCheap (Lit lit) = True
422 exprIsCheap (Type _) = True
423 exprIsCheap (Var _) = True
424 exprIsCheap (Note InlineMe e) = True
425 exprIsCheap (Note _ e) = exprIsCheap e
426 exprIsCheap (Lam x e) = isRuntimeVar x || exprIsCheap e
427 exprIsCheap (Case e _ _ alts) = exprIsCheap e &&
428 and [exprIsCheap rhs | (_,_,rhs) <- alts]
429 -- Experimentally, treat (case x of ...) as cheap
430 -- (and case __coerce x etc.)
431 -- This improves arities of overloaded functions where
432 -- there is only dictionary selection (no construction) involved
433 exprIsCheap (Let (NonRec x _) e)
434 | isUnLiftedType (idType x) = exprIsCheap e
436 -- strict lets always have cheap right hand sides,
437 -- and do no allocation.
439 exprIsCheap other_expr -- Applications and variables
442 -- Accumulate value arguments, then decide
443 go (App f a) val_args | isRuntimeArg a = go f (a:val_args)
444 | otherwise = go f val_args
446 go (Var f) [] = True -- Just a type application of a variable
447 -- (f t1 t2 t3) counts as WHNF
449 = case globalIdDetails f of
450 RecordSelId {} -> go_sel args
451 ClassOpId _ -> go_sel args
452 PrimOpId op -> go_primop op args
454 DataConWorkId _ -> go_pap args
455 other | length args < idArity f -> go_pap args
457 other -> isBottomingId f
458 -- Application of a function which
459 -- always gives bottom; we treat this as cheap
460 -- because it certainly doesn't need to be shared!
462 go other args = False
465 go_pap args = all exprIsTrivial args
466 -- For constructor applications and primops, check that all
467 -- the args are trivial. We don't want to treat as cheap, say,
469 -- We'll put up with one constructor application, but not dozens
472 go_primop op args = primOpIsCheap op && all exprIsCheap args
473 -- In principle we should worry about primops
474 -- that return a type variable, since the result
475 -- might be applied to something, but I'm not going
476 -- to bother to check the number of args
479 go_sel [arg] = exprIsCheap arg -- I'm experimenting with making record selection
480 go_sel other = False -- look cheap, so we will substitute it inside a
481 -- lambda. Particularly for dictionary field selection.
482 -- BUT: Take care with (sel d x)! The (sel d) might be cheap, but
483 -- there's no guarantee that (sel d x) will be too. Hence (n_val_args == 1)
486 exprOkForSpeculation returns True of an expression that it is
488 * safe to evaluate even if normal order eval might not
489 evaluate the expression at all, or
491 * safe *not* to evaluate even if normal order would do so
495 the expression guarantees to terminate,
497 without raising an exception,
498 without causing a side effect (e.g. writing a mutable variable)
500 NB: if exprIsHNF e, then exprOkForSpecuation e
503 let x = case y# +# 1# of { r# -> I# r# }
506 case y# +# 1# of { r# ->
511 We can only do this if the (y+1) is ok for speculation: it has no
512 side effects, and can't diverge or raise an exception.
515 exprOkForSpeculation :: CoreExpr -> Bool
516 exprOkForSpeculation (Lit _) = True
517 exprOkForSpeculation (Type _) = True
518 exprOkForSpeculation (Var v) = isUnLiftedType (idType v)
519 exprOkForSpeculation (Note _ e) = exprOkForSpeculation e
520 exprOkForSpeculation other_expr
521 = case collectArgs other_expr of
522 (Var f, args) -> spec_ok (globalIdDetails f) args
526 spec_ok (DataConWorkId _) args
527 = True -- The strictness of the constructor has already
528 -- been expressed by its "wrapper", so we don't need
529 -- to take the arguments into account
531 spec_ok (PrimOpId op) args
532 | isDivOp op, -- Special case for dividing operations that fail
533 [arg1, Lit lit] <- args -- only if the divisor is zero
534 = not (isZeroLit lit) && exprOkForSpeculation arg1
535 -- Often there is a literal divisor, and this
536 -- can get rid of a thunk in an inner looop
539 = primOpOkForSpeculation op &&
540 all exprOkForSpeculation args
541 -- A bit conservative: we don't really need
542 -- to care about lazy arguments, but this is easy
544 spec_ok other args = False
546 isDivOp :: PrimOp -> Bool
547 -- True of dyadic operators that can fail
548 -- only if the second arg is zero
549 -- This function probably belongs in PrimOp, or even in
550 -- an automagically generated file.. but it's such a
551 -- special case I thought I'd leave it here for now.
552 isDivOp IntQuotOp = True
553 isDivOp IntRemOp = True
554 isDivOp WordQuotOp = True
555 isDivOp WordRemOp = True
556 isDivOp IntegerQuotRemOp = True
557 isDivOp IntegerDivModOp = True
558 isDivOp FloatDivOp = True
559 isDivOp DoubleDivOp = True
560 isDivOp other = False
565 exprIsBottom :: CoreExpr -> Bool -- True => definitely bottom
566 exprIsBottom e = go 0 e
568 -- n is the number of args
569 go n (Note _ e) = go n e
570 go n (Let _ e) = go n e
571 go n (Case e _ _ _) = go 0 e -- Just check the scrut
572 go n (App e _) = go (n+1) e
573 go n (Var v) = idAppIsBottom v n
575 go n (Lam _ _) = False
576 go n (Type _) = False
578 idAppIsBottom :: Id -> Int -> Bool
579 idAppIsBottom id n_val_args = appIsBottom (idNewStrictness id) n_val_args
582 @exprIsHNF@ returns true for expressions that are certainly *already*
583 evaluated to *head* normal form. This is used to decide whether it's ok
586 case x of _ -> e ===> e
588 and to decide whether it's safe to discard a `seq`
590 So, it does *not* treat variables as evaluated, unless they say they are.
592 But it *does* treat partial applications and constructor applications
593 as values, even if their arguments are non-trivial, provided the argument
595 e.g. (:) (f x) (map f xs) is a value
596 map (...redex...) is a value
597 Because `seq` on such things completes immediately
599 For unlifted argument types, we have to be careful:
601 Suppose (f x) diverges; then C (f x) is not a value. True, but
602 this form is illegal (see the invariants in CoreSyn). Args of unboxed
603 type must be ok-for-speculation (or trivial).
606 exprIsHNF :: CoreExpr -> Bool -- True => Value-lambda, constructor, PAP
607 exprIsHNF (Var v) -- NB: There are no value args at this point
608 = isDataConWorkId v -- Catches nullary constructors,
609 -- so that [] and () are values, for example
610 || idArity v > 0 -- Catches (e.g.) primops that don't have unfoldings
611 || isEvaldUnfolding (idUnfolding v)
612 -- Check the thing's unfolding; it might be bound to a value
613 -- A worry: what if an Id's unfolding is just itself:
614 -- then we could get an infinite loop...
616 exprIsHNF (Lit l) = True
617 exprIsHNF (Type ty) = True -- Types are honorary Values;
618 -- we don't mind copying them
619 exprIsHNF (Lam b e) = isRuntimeVar b || exprIsHNF e
620 exprIsHNF (Note _ e) = exprIsHNF e
621 exprIsHNF (App e (Type _)) = exprIsHNF e
622 exprIsHNF (App e a) = app_is_value e [a]
623 exprIsHNF other = False
625 -- There is at least one value argument
626 app_is_value (Var fun) args
627 | isDataConWorkId fun -- Constructor apps are values
628 || idArity fun > valArgCount args -- Under-applied function
629 = check_args (idType fun) args
630 app_is_value (App f a) as = app_is_value f (a:as)
631 app_is_value other as = False
633 -- 'check_args' checks that unlifted-type args
634 -- are in fact guaranteed non-divergent
635 check_args fun_ty [] = True
636 check_args fun_ty (Type _ : args) = case splitForAllTy_maybe fun_ty of
637 Just (_, ty) -> check_args ty args
638 check_args fun_ty (arg : args)
639 | isUnLiftedType arg_ty = exprOkForSpeculation arg
640 | otherwise = check_args res_ty args
642 (arg_ty, res_ty) = splitFunTy fun_ty
646 exprIsConApp_maybe :: CoreExpr -> Maybe (DataCon, [CoreExpr])
647 exprIsConApp_maybe (Note (Coerce to_ty from_ty) expr)
648 = -- Maybe this is over the top, but here we try to turn
649 -- coerce (S,T) ( x, y )
651 -- ( coerce S x, coerce T y )
652 -- This happens in anger in PrelArrExts which has a coerce
653 -- case coerce memcpy a b of
655 -- where the memcpy is in the IO monad, but the call is in
657 case exprIsConApp_maybe expr of {
661 case splitTyConApp_maybe to_ty of {
663 Just (tc, tc_arg_tys) | tc /= dataConTyCon dc -> Nothing
664 | not (isVanillaDataCon dc) -> Nothing
666 -- Type constructor must match
667 -- We knock out existentials to keep matters simple(r)
669 arity = tyConArity tc
670 val_args = drop arity args
671 to_arg_tys = dataConInstArgTys dc tc_arg_tys
672 mk_coerce ty arg = mkCoerce ty arg
673 new_val_args = zipWith mk_coerce to_arg_tys val_args
675 ASSERT( all isTypeArg (take arity args) )
676 ASSERT( equalLength val_args to_arg_tys )
677 Just (dc, map Type tc_arg_tys ++ new_val_args)
680 exprIsConApp_maybe (Note _ expr)
681 = exprIsConApp_maybe expr
682 -- We ignore InlineMe notes in case we have
683 -- x = __inline_me__ (a,b)
684 -- All part of making sure that INLINE pragmas never hurt
685 -- Marcin tripped on this one when making dictionaries more inlinable
687 -- In fact, we ignore all notes. For example,
688 -- case _scc_ "foo" (C a b) of
690 -- should be optimised away, but it will be only if we look
691 -- through the SCC note.
693 exprIsConApp_maybe expr = analyse (collectArgs expr)
695 analyse (Var fun, args)
696 | Just con <- isDataConWorkId_maybe fun,
697 args `lengthAtLeast` dataConRepArity con
698 -- Might be > because the arity excludes type args
701 -- Look through unfoldings, but only cheap ones, because
702 -- we are effectively duplicating the unfolding
703 analyse (Var fun, [])
704 | let unf = idUnfolding fun,
706 = exprIsConApp_maybe (unfoldingTemplate unf)
708 analyse other = Nothing
713 %************************************************************************
715 \subsection{Eta reduction and expansion}
717 %************************************************************************
720 exprEtaExpandArity :: DynFlags -> CoreExpr -> Arity
721 {- The Arity returned is the number of value args the
722 thing can be applied to without doing much work
724 exprEtaExpandArity is used when eta expanding
727 It returns 1 (or more) to:
728 case x of p -> \s -> ...
729 because for I/O ish things we really want to get that \s to the top.
730 We are prepared to evaluate x each time round the loop in order to get that
732 It's all a bit more subtle than it looks:
736 Consider one-shot lambdas
737 let x = expensive in \y z -> E
738 We want this to have arity 2 if the \y-abstraction is a 1-shot lambda
739 Hence the ArityType returned by arityType
741 2. The state-transformer hack
743 The one-shot lambda special cause is particularly important/useful for
744 IO state transformers, where we often get
745 let x = E in \ s -> ...
747 and the \s is a real-world state token abstraction. Such abstractions
748 are almost invariably 1-shot, so we want to pull the \s out, past the
749 let x=E, even if E is expensive. So we treat state-token lambdas as
750 one-shot even if they aren't really. The hack is in Id.isOneShotBndr.
752 3. Dealing with bottom
755 f = \x -> error "foo"
756 Here, arity 1 is fine. But if it is
760 then we want to get arity 2. Tecnically, this isn't quite right, because
762 should diverge, but it'll converge if we eta-expand f. Nevertheless, we
763 do so; it improves some programs significantly, and increasing convergence
764 isn't a bad thing. Hence the ABot/ATop in ArityType.
766 Actually, the situation is worse. Consider
770 Can we eta-expand here? At first the answer looks like "yes of course", but
773 This should diverge! But if we eta-expand, it won't. Again, we ignore this
774 "problem", because being scrupulous would lose an important transformation for
780 Non-recursive newtypes are transparent, and should not get in the way.
781 We do (currently) eta-expand recursive newtypes too. So if we have, say
783 newtype T = MkT ([T] -> Int)
787 where f has arity 1. Then: etaExpandArity e = 1;
788 that is, etaExpandArity looks through the coerce.
790 When we eta-expand e to arity 1: eta_expand 1 e T
791 we want to get: coerce T (\x::[T] -> (coerce ([T]->Int) e) x)
793 HOWEVER, note that if you use coerce bogusly you can ge
795 And since negate has arity 2, you might try to eta expand. But you can't
796 decopose Int to a function type. Hence the final case in eta_expand.
800 exprEtaExpandArity dflags e = arityDepth (arityType dflags e)
802 -- A limited sort of function type
803 data ArityType = AFun Bool ArityType -- True <=> one-shot
804 | ATop -- Know nothing
807 arityDepth :: ArityType -> Arity
808 arityDepth (AFun _ ty) = 1 + arityDepth ty
811 andArityType ABot at2 = at2
812 andArityType ATop at2 = ATop
813 andArityType (AFun t1 at1) (AFun t2 at2) = AFun (t1 && t2) (andArityType at1 at2)
814 andArityType at1 at2 = andArityType at2 at1
816 arityType :: DynFlags -> CoreExpr -> ArityType
817 -- (go1 e) = [b1,..,bn]
818 -- means expression can be rewritten \x_b1 -> ... \x_bn -> body
819 -- where bi is True <=> the lambda is one-shot
821 arityType dflags (Note n e) = arityType dflags e
822 -- Not needed any more: etaExpand is cleverer
823 -- | ok_note n = arityType dflags e
824 -- | otherwise = ATop
826 arityType dflags (Var v)
827 = mk (idArity v) (arg_tys (idType v))
829 mk :: Arity -> [Type] -> ArityType
830 -- The argument types are only to steer the "state hack"
831 -- Consider case x of
833 -- False -> \(s:RealWorld) -> e
834 -- where foo has arity 1. Then we want the state hack to
835 -- apply to foo too, so we can eta expand the case.
836 mk 0 tys | isBottomingId v = ABot
837 | (ty:tys) <- tys, isStateHackType ty = AFun True ATop
839 mk n (ty:tys) = AFun (isStateHackType ty) (mk (n-1) tys)
840 mk n [] = AFun False (mk (n-1) [])
842 arg_tys :: Type -> [Type] -- Ignore for-alls
844 | Just (_, ty') <- splitForAllTy_maybe ty = arg_tys ty'
845 | Just (arg,res) <- splitFunTy_maybe ty = arg : arg_tys res
848 -- Lambdas; increase arity
849 arityType dflags (Lam x e)
850 | isId x = AFun (isOneShotBndr x) (arityType dflags e)
851 | otherwise = arityType dflags e
853 -- Applications; decrease arity
854 arityType dflags (App f (Type _)) = arityType dflags f
855 arityType dflags (App f a) = case arityType dflags f of
856 AFun one_shot xs | exprIsCheap a -> xs
859 -- Case/Let; keep arity if either the expression is cheap
860 -- or it's a 1-shot lambda
861 -- The former is not really right for Haskell
862 -- f x = case x of { (a,b) -> \y. e }
864 -- f x y = case x of { (a,b) -> e }
865 -- The difference is observable using 'seq'
866 arityType dflags (Case scrut _ _ alts)
867 = case foldr1 andArityType [arityType dflags rhs | (_,_,rhs) <- alts] of
868 xs | exprIsCheap scrut -> xs
869 xs@(AFun one_shot _) | one_shot -> AFun True ATop
872 arityType dflags (Let b e)
873 = case arityType dflags e of
874 xs | cheap_bind b -> xs
875 xs@(AFun one_shot _) | one_shot -> AFun True ATop
878 cheap_bind (NonRec b e) = is_cheap (b,e)
879 cheap_bind (Rec prs) = all is_cheap prs
880 is_cheap (b,e) = (dopt Opt_DictsCheap dflags && isDictId b)
882 -- If the experimental -fdicts-cheap flag is on, we eta-expand through
883 -- dictionary bindings. This improves arities. Thereby, it also
884 -- means that full laziness is less prone to floating out the
885 -- application of a function to its dictionary arguments, which
886 -- can thereby lose opportunities for fusion. Example:
887 -- foo :: Ord a => a -> ...
888 -- foo = /\a \(d:Ord a). let d' = ...d... in \(x:a). ....
889 -- -- So foo has arity 1
891 -- f = \x. foo dInt $ bar x
893 -- The (foo DInt) is floated out, and makes ineffective a RULE
896 -- One could go further and make exprIsCheap reply True to any
897 -- dictionary-typed expression, but that's more work.
899 arityType dflags other = ATop
901 {- NOT NEEDED ANY MORE: etaExpand is cleverer
902 ok_note InlineMe = False
904 -- Notice that we do not look through __inline_me__
905 -- This may seem surprising, but consider
906 -- f = _inline_me (\x -> e)
907 -- We DO NOT want to eta expand this to
908 -- f = \x -> (_inline_me (\x -> e)) x
909 -- because the _inline_me gets dropped now it is applied,
918 etaExpand :: Arity -- Result should have this number of value args
920 -> CoreExpr -> Type -- Expression and its type
922 -- (etaExpand n us e ty) returns an expression with
923 -- the same meaning as 'e', but with arity 'n'.
925 -- Given e' = etaExpand n us e ty
927 -- ty = exprType e = exprType e'
929 -- Note that SCCs are not treated specially. If we have
930 -- etaExpand 2 (\x -> scc "foo" e)
931 -- = (\xy -> (scc "foo" e) y)
932 -- So the costs of evaluating 'e' (not 'e y') are attributed to "foo"
934 etaExpand n us expr ty
935 | manifestArity expr >= n = expr -- The no-op case
936 | otherwise = eta_expand n us expr ty
939 -- manifestArity sees how many leading value lambdas there are
940 manifestArity :: CoreExpr -> Arity
941 manifestArity (Lam v e) | isId v = 1 + manifestArity e
942 | otherwise = manifestArity e
943 manifestArity (Note _ e) = manifestArity e
946 -- etaExpand deals with for-alls. For example:
948 -- where E :: forall a. a -> a
950 -- (/\b. \y::a -> E b y)
952 -- It deals with coerces too, though they are now rare
953 -- so perhaps the extra code isn't worth it
955 eta_expand n us expr ty
957 -- The ILX code generator requires eta expansion for type arguments
958 -- too, but alas the 'n' doesn't tell us how many of them there
959 -- may be. So we eagerly eta expand any big lambdas, and just
960 -- cross our fingers about possible loss of sharing in the ILX case.
961 -- The Right Thing is probably to make 'arity' include
962 -- type variables throughout the compiler. (ToDo.)
964 -- Saturated, so nothing to do
967 -- Short cut for the case where there already
968 -- is a lambda; no point in gratuitously adding more
969 eta_expand n us (Lam v body) ty
971 = Lam v (eta_expand n us body (applyTy ty (mkTyVarTy v)))
974 = Lam v (eta_expand (n-1) us body (funResultTy ty))
976 -- We used to have a special case that stepped inside Coerces here,
977 -- thus: eta_expand n us (Note note@(Coerce _ ty) e) _
978 -- = Note note (eta_expand n us e ty)
979 -- BUT this led to an infinite loop
980 -- Example: newtype T = MkT (Int -> Int)
981 -- eta_expand 1 (coerce (Int->Int) e)
982 -- --> coerce (Int->Int) (eta_expand 1 T e)
984 -- --> coerce (Int->Int) (coerce T
985 -- (\x::Int -> eta_expand 1 (coerce (Int->Int) e)))
986 -- by the splitNewType_maybe case below
989 eta_expand n us expr ty
990 = case splitForAllTy_maybe ty of {
991 Just (tv,ty') -> Lam tv (eta_expand n us (App expr (Type (mkTyVarTy tv))) ty')
995 case splitFunTy_maybe ty of {
996 Just (arg_ty, res_ty) -> Lam arg1 (eta_expand (n-1) us2 (App expr (Var arg1)) res_ty)
998 arg1 = mkSysLocal FSLIT("eta") uniq arg_ty
1004 -- newtype T = MkT ([T] -> Int)
1005 -- Consider eta-expanding this
1008 -- coerce T (\x::[T] -> (coerce ([T]->Int) e) x)
1009 -- Only try this for recursive newtypes; the non-recursive kind
1010 -- are transparent anyway
1012 case splitRecNewType_maybe ty of {
1013 Just ty' -> mkCoerce2 ty ty' (eta_expand n us (mkCoerce2 ty' ty expr) ty') ;
1016 -- We have an expression of arity > 0, but its type isn't a function
1017 -- This *can* legitmately happen: e.g. coerce Int (\x. x)
1018 -- Essentially the programmer is playing fast and loose with types
1019 -- (Happy does this a lot). So we simply decline to eta-expand.
1024 exprArity is a cheap-and-cheerful version of exprEtaExpandArity.
1025 It tells how many things the expression can be applied to before doing
1026 any work. It doesn't look inside cases, lets, etc. The idea is that
1027 exprEtaExpandArity will do the hard work, leaving something that's easy
1028 for exprArity to grapple with. In particular, Simplify uses exprArity to
1029 compute the ArityInfo for the Id.
1031 Originally I thought that it was enough just to look for top-level lambdas, but
1032 it isn't. I've seen this
1034 foo = PrelBase.timesInt
1036 We want foo to get arity 2 even though the eta-expander will leave it
1037 unchanged, in the expectation that it'll be inlined. But occasionally it
1038 isn't, because foo is blacklisted (used in a rule).
1040 Similarly, see the ok_note check in exprEtaExpandArity. So
1041 f = __inline_me (\x -> e)
1042 won't be eta-expanded.
1044 And in any case it seems more robust to have exprArity be a bit more intelligent.
1045 But note that (\x y z -> f x y z)
1046 should have arity 3, regardless of f's arity.
1049 exprArity :: CoreExpr -> Arity
1052 go (Var v) = idArity v
1053 go (Lam x e) | isId x = go e + 1
1055 go (Note n e) = go e
1056 go (App e (Type t)) = go e
1057 go (App f a) | exprIsCheap a = (go f - 1) `max` 0
1058 -- NB: exprIsCheap a!
1059 -- f (fac x) does not have arity 2,
1060 -- even if f has arity 3!
1061 -- NB: `max 0`! (\x y -> f x) has arity 2, even if f is
1062 -- unknown, hence arity 0
1066 %************************************************************************
1068 \subsection{Equality}
1070 %************************************************************************
1072 @cheapEqExpr@ is a cheap equality test which bales out fast!
1073 True => definitely equal
1074 False => may or may not be equal
1077 cheapEqExpr :: Expr b -> Expr b -> Bool
1079 cheapEqExpr (Var v1) (Var v2) = v1==v2
1080 cheapEqExpr (Lit lit1) (Lit lit2) = lit1 == lit2
1081 cheapEqExpr (Type t1) (Type t2) = t1 `coreEqType` t2
1083 cheapEqExpr (App f1 a1) (App f2 a2)
1084 = f1 `cheapEqExpr` f2 && a1 `cheapEqExpr` a2
1086 cheapEqExpr _ _ = False
1088 exprIsBig :: Expr b -> Bool
1089 -- Returns True of expressions that are too big to be compared by cheapEqExpr
1090 exprIsBig (Lit _) = False
1091 exprIsBig (Var v) = False
1092 exprIsBig (Type t) = False
1093 exprIsBig (App f a) = exprIsBig f || exprIsBig a
1094 exprIsBig other = True
1099 tcEqExpr :: CoreExpr -> CoreExpr -> Bool
1100 -- Used in rule matching, so does *not* look through
1101 -- newtypes, predicate types; hence tcEqExpr
1103 tcEqExpr e1 e2 = tcEqExprX rn_env e1 e2
1105 rn_env = mkRnEnv2 (mkInScopeSet (exprFreeVars e1 `unionVarSet` exprFreeVars e2))
1107 tcEqExprX :: RnEnv2 -> CoreExpr -> CoreExpr -> Bool
1108 tcEqExprX env (Var v1) (Var v2) = rnOccL env v1 == rnOccR env v2
1109 tcEqExprX env (Lit lit1) (Lit lit2) = lit1 == lit2
1110 tcEqExprX env (App f1 a1) (App f2 a2) = tcEqExprX env f1 f2 && tcEqExprX env a1 a2
1111 tcEqExprX env (Lam v1 e1) (Lam v2 e2) = tcEqExprX (rnBndr2 env v1 v2) e1 e2
1112 tcEqExprX env (Let (NonRec v1 r1) e1)
1113 (Let (NonRec v2 r2) e2) = tcEqExprX env r1 r2
1114 && tcEqExprX (rnBndr2 env v1 v2) e1 e2
1115 tcEqExprX env (Let (Rec ps1) e1)
1116 (Let (Rec ps2) e2) = equalLength ps1 ps2
1117 && and (zipWith eq_rhs ps1 ps2)
1118 && tcEqExprX env' e1 e2
1120 env' = foldl2 rn_bndr2 env ps2 ps2
1121 rn_bndr2 env (b1,_) (b2,_) = rnBndr2 env b1 b2
1122 eq_rhs (_,r1) (_,r2) = tcEqExprX env' r1 r2
1123 tcEqExprX env (Case e1 v1 t1 a1)
1124 (Case e2 v2 t2 a2) = tcEqExprX env e1 e2
1125 && tcEqTypeX env t1 t2
1126 && equalLength a1 a2
1127 && and (zipWith (eq_alt env') a1 a2)
1129 env' = rnBndr2 env v1 v2
1131 tcEqExprX env (Note n1 e1) (Note n2 e2) = eq_note env n1 n2 && tcEqExprX env e1 e2
1132 tcEqExprX env (Type t1) (Type t2) = tcEqTypeX env t1 t2
1133 tcEqExprX env e1 e2 = False
1135 eq_alt env (c1,vs1,r1) (c2,vs2,r2) = c1==c2 && tcEqExprX (rnBndrs2 env vs1 vs2) r1 r2
1137 eq_note env (SCC cc1) (SCC cc2) = cc1 == cc2
1138 eq_note env (Coerce t1 f1) (Coerce t2 f2) = tcEqTypeX env t1 t2 && tcEqTypeX env f1 f2
1139 eq_note env (CoreNote s1) (CoreNote s2) = s1 == s2
1140 eq_note env other1 other2 = False
1144 %************************************************************************
1146 \subsection{The size of an expression}
1148 %************************************************************************
1151 coreBindsSize :: [CoreBind] -> Int
1152 coreBindsSize bs = foldr ((+) . bindSize) 0 bs
1154 exprSize :: CoreExpr -> Int
1155 -- A measure of the size of the expressions
1156 -- It also forces the expression pretty drastically as a side effect
1157 exprSize (Var v) = v `seq` 1
1158 exprSize (Lit lit) = lit `seq` 1
1159 exprSize (App f a) = exprSize f + exprSize a
1160 exprSize (Lam b e) = varSize b + exprSize e
1161 exprSize (Let b e) = bindSize b + exprSize e
1162 exprSize (Case e b t as) = seqType t `seq` exprSize e + varSize b + 1 + foldr ((+) . altSize) 0 as
1163 exprSize (Note n e) = noteSize n + exprSize e
1164 exprSize (Type t) = seqType t `seq` 1
1166 noteSize (SCC cc) = cc `seq` 1
1167 noteSize (Coerce t1 t2) = seqType t1 `seq` seqType t2 `seq` 1
1168 noteSize InlineMe = 1
1169 noteSize (CoreNote s) = s `seq` 1 -- hdaume: core annotations
1171 varSize :: Var -> Int
1172 varSize b | isTyVar b = 1
1173 | otherwise = seqType (idType b) `seq`
1174 megaSeqIdInfo (idInfo b) `seq`
1177 varsSize = foldr ((+) . varSize) 0
1179 bindSize (NonRec b e) = varSize b + exprSize e
1180 bindSize (Rec prs) = foldr ((+) . pairSize) 0 prs
1182 pairSize (b,e) = varSize b + exprSize e
1184 altSize (c,bs,e) = c `seq` varsSize bs + exprSize e
1188 %************************************************************************
1190 \subsection{Hashing}
1192 %************************************************************************
1195 hashExpr :: CoreExpr -> Int
1196 hashExpr e | hash < 0 = 77 -- Just in case we hit -maxInt
1199 hash = abs (hash_expr e) -- Negative numbers kill UniqFM
1201 hash_expr (Note _ e) = hash_expr e
1202 hash_expr (Let (NonRec b r) e) = hashId b
1203 hash_expr (Let (Rec ((b,r):_)) e) = hashId b
1204 hash_expr (Case _ b _ _) = hashId b
1205 hash_expr (App f e) = hash_expr f * fast_hash_expr e
1206 hash_expr (Var v) = hashId v
1207 hash_expr (Lit lit) = hashLiteral lit
1208 hash_expr (Lam b _) = hashId b
1209 hash_expr (Type t) = trace "hash_expr: type" 1 -- Shouldn't happen
1211 fast_hash_expr (Var v) = hashId v
1212 fast_hash_expr (Lit lit) = hashLiteral lit
1213 fast_hash_expr (App f (Type _)) = fast_hash_expr f
1214 fast_hash_expr (App f a) = fast_hash_expr a
1215 fast_hash_expr (Lam b _) = hashId b
1216 fast_hash_expr other = 1
1219 hashId id = hashName (idName id)
1222 %************************************************************************
1224 \subsection{Determining non-updatable right-hand-sides}
1226 %************************************************************************
1228 Top-level constructor applications can usually be allocated
1229 statically, but they can't if the constructor, or any of the
1230 arguments, come from another DLL (because we can't refer to static
1231 labels in other DLLs).
1233 If this happens we simply make the RHS into an updatable thunk,
1234 and 'exectute' it rather than allocating it statically.
1237 rhsIsStatic :: PackageId -> CoreExpr -> Bool
1238 -- This function is called only on *top-level* right-hand sides
1239 -- Returns True if the RHS can be allocated statically, with
1240 -- no thunks involved at all.
1242 -- It's called (i) in TidyPgm.hasCafRefs to decide if the rhs is, or
1243 -- refers to, CAFs; and (ii) in CoreToStg to decide whether to put an
1244 -- update flag on it.
1246 -- The basic idea is that rhsIsStatic returns True only if the RHS is
1247 -- (a) a value lambda
1248 -- (b) a saturated constructor application with static args
1250 -- BUT watch out for
1251 -- (i) Any cross-DLL references kill static-ness completely
1252 -- because they must be 'executed' not statically allocated
1253 -- ("DLL" here really only refers to Windows DLLs, on other platforms,
1254 -- this is not necessary)
1256 -- (ii) We treat partial applications as redexes, because in fact we
1257 -- make a thunk for them that runs and builds a PAP
1258 -- at run-time. The only appliations that are treated as
1259 -- static are *saturated* applications of constructors.
1261 -- We used to try to be clever with nested structures like this:
1262 -- ys = (:) w ((:) w [])
1263 -- on the grounds that CorePrep will flatten ANF-ise it later.
1264 -- But supporting this special case made the function much more
1265 -- complicated, because the special case only applies if there are no
1266 -- enclosing type lambdas:
1267 -- ys = /\ a -> Foo (Baz ([] a))
1268 -- Here the nested (Baz []) won't float out to top level in CorePrep.
1270 -- But in fact, even without -O, nested structures at top level are
1271 -- flattened by the simplifier, so we don't need to be super-clever here.
1275 -- f = \x::Int. x+7 TRUE
1276 -- p = (True,False) TRUE
1278 -- d = (fst p, False) FALSE because there's a redex inside
1279 -- (this particular one doesn't happen but...)
1281 -- h = D# (1.0## /## 2.0##) FALSE (redex again)
1282 -- n = /\a. Nil a TRUE
1284 -- t = /\a. (:) (case w a of ...) (Nil a) FALSE (redex)
1287 -- This is a bit like CoreUtils.exprIsHNF, with the following differences:
1288 -- a) scc "foo" (\x -> ...) is updatable (so we catch the right SCC)
1290 -- b) (C x xs), where C is a contructors is updatable if the application is
1293 -- c) don't look through unfolding of f in (f x).
1295 -- When opt_RuntimeTypes is on, we keep type lambdas and treat
1296 -- them as making the RHS re-entrant (non-updatable).
1298 rhsIsStatic this_pkg rhs = is_static False rhs
1300 is_static :: Bool -- True <=> in a constructor argument; must be atomic
1303 is_static False (Lam b e) = isRuntimeVar b || is_static False e
1305 is_static in_arg (Note (SCC _) e) = False
1306 is_static in_arg (Note _ e) = is_static in_arg e
1308 is_static in_arg (Lit lit)
1310 MachLabel _ _ -> False
1312 -- A MachLabel (foreign import "&foo") in an argument
1313 -- prevents a constructor application from being static. The
1314 -- reason is that it might give rise to unresolvable symbols
1315 -- in the object file: under Linux, references to "weak"
1316 -- symbols from the data segment give rise to "unresolvable
1317 -- relocation" errors at link time This might be due to a bug
1318 -- in the linker, but we'll work around it here anyway.
1321 is_static in_arg other_expr = go other_expr 0
1323 go (Var f) n_val_args
1324 #if mingw32_TARGET_OS
1325 | not (isDllName this_pkg (idName f))
1327 = saturated_data_con f n_val_args
1328 || (in_arg && n_val_args == 0)
1329 -- A naked un-applied variable is *not* deemed a static RHS
1331 -- Reason: better to update so that the indirection gets shorted
1332 -- out, and the true value will be seen
1333 -- NB: if you change this, you'll break the invariant that THUNK_STATICs
1334 -- are always updatable. If you do so, make sure that non-updatable
1335 -- ones have enough space for their static link field!
1337 go (App f a) n_val_args
1338 | isTypeArg a = go f n_val_args
1339 | not in_arg && is_static True a = go f (n_val_args + 1)
1340 -- The (not in_arg) checks that we aren't in a constructor argument;
1341 -- if we are, we don't allow (value) applications of any sort
1343 -- NB. In case you wonder, args are sometimes not atomic. eg.
1344 -- x = D# (1.0## /## 2.0##)
1345 -- can't float because /## can fail.
1347 go (Note (SCC _) f) n_val_args = False
1348 go (Note _ f) n_val_args = go f n_val_args
1350 go other n_val_args = False
1352 saturated_data_con f n_val_args
1353 = case isDataConWorkId_maybe f of
1354 Just dc -> n_val_args == dataConRepArity dc