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
440 = go other_expr 0 True
442 go (Var f) n_args args_cheap
443 = (idAppIsCheap f n_args && args_cheap)
444 -- A constructor, cheap primop, or partial application
446 || idAppIsBottom f n_args
447 -- Application of a function which
448 -- always gives bottom; we treat this as cheap
449 -- because it certainly doesn't need to be shared!
451 go (App f a) n_args args_cheap
452 | not (isRuntimeArg a) = go f n_args args_cheap
453 | otherwise = go f (n_args + 1) (exprIsCheap a && args_cheap)
455 go other n_args args_cheap = False
457 idAppIsCheap :: Id -> Int -> Bool
458 idAppIsCheap id n_val_args
459 | n_val_args == 0 = True -- Just a type application of
460 -- a variable (f t1 t2 t3)
463 = case globalIdDetails id of
464 DataConWorkId _ -> True
465 RecordSelId {} -> n_val_args == 1 -- I'm experimenting with making record selection
466 ClassOpId _ -> n_val_args == 1 -- look cheap, so we will substitute it inside a
467 -- lambda. Particularly for dictionary field selection.
468 -- BUT: Take care with (sel d x)! The (sel d) might be cheap, but
469 -- there's no guarantee that (sel d x) will be too. Hence (n_val_args == 1)
471 PrimOpId op -> primOpIsCheap op -- In principle we should worry about primops
472 -- that return a type variable, since the result
473 -- might be applied to something, but I'm not going
474 -- to bother to check the number of args
475 other -> n_val_args < idArity id
478 exprOkForSpeculation returns True of an expression that it is
480 * safe to evaluate even if normal order eval might not
481 evaluate the expression at all, or
483 * safe *not* to evaluate even if normal order would do so
487 the expression guarantees to terminate,
489 without raising an exception,
490 without causing a side effect (e.g. writing a mutable variable)
492 let x = case y# +# 1# of { r# -> I# r# }
495 case y# +# 1# of { r# ->
500 We can only do this if the (y+1) is ok for speculation: it has no
501 side effects, and can't diverge or raise an exception.
504 exprOkForSpeculation :: CoreExpr -> Bool
505 exprOkForSpeculation (Lit _) = True
506 exprOkForSpeculation (Type _) = True
507 exprOkForSpeculation (Var v) = isUnLiftedType (idType v)
508 exprOkForSpeculation (Note _ e) = exprOkForSpeculation e
509 exprOkForSpeculation other_expr
510 = case collectArgs other_expr of
511 (Var f, args) -> spec_ok (globalIdDetails f) args
515 spec_ok (DataConWorkId _) args
516 = True -- The strictness of the constructor has already
517 -- been expressed by its "wrapper", so we don't need
518 -- to take the arguments into account
520 spec_ok (PrimOpId op) args
521 | isDivOp op, -- Special case for dividing operations that fail
522 [arg1, Lit lit] <- args -- only if the divisor is zero
523 = not (isZeroLit lit) && exprOkForSpeculation arg1
524 -- Often there is a literal divisor, and this
525 -- can get rid of a thunk in an inner looop
528 = primOpOkForSpeculation op &&
529 all exprOkForSpeculation args
530 -- A bit conservative: we don't really need
531 -- to care about lazy arguments, but this is easy
533 spec_ok other args = False
535 isDivOp :: PrimOp -> Bool
536 -- True of dyadic operators that can fail
537 -- only if the second arg is zero
538 -- This function probably belongs in PrimOp, or even in
539 -- an automagically generated file.. but it's such a
540 -- special case I thought I'd leave it here for now.
541 isDivOp IntQuotOp = True
542 isDivOp IntRemOp = True
543 isDivOp WordQuotOp = True
544 isDivOp WordRemOp = True
545 isDivOp IntegerQuotRemOp = True
546 isDivOp IntegerDivModOp = True
547 isDivOp FloatDivOp = True
548 isDivOp DoubleDivOp = True
549 isDivOp other = False
554 exprIsBottom :: CoreExpr -> Bool -- True => definitely bottom
555 exprIsBottom e = go 0 e
557 -- n is the number of args
558 go n (Note _ e) = go n e
559 go n (Let _ e) = go n e
560 go n (Case e _ _ _) = go 0 e -- Just check the scrut
561 go n (App e _) = go (n+1) e
562 go n (Var v) = idAppIsBottom v n
564 go n (Lam _ _) = False
565 go n (Type _) = False
567 idAppIsBottom :: Id -> Int -> Bool
568 idAppIsBottom id n_val_args = appIsBottom (idNewStrictness id) n_val_args
571 @exprIsHNF@ returns true for expressions that are certainly *already*
572 evaluated to *head* normal form. This is used to decide whether it's ok
575 case x of _ -> e ===> e
577 and to decide whether it's safe to discard a `seq`
579 So, it does *not* treat variables as evaluated, unless they say they are.
581 But it *does* treat partial applications and constructor applications
582 as values, even if their arguments are non-trivial, provided the argument
584 e.g. (:) (f x) (map f xs) is a value
585 map (...redex...) is a value
586 Because `seq` on such things completes immediately
588 For unlifted argument types, we have to be careful:
590 Suppose (f x) diverges; then C (f x) is not a value. True, but
591 this form is illegal (see the invariants in CoreSyn). Args of unboxed
592 type must be ok-for-speculation (or trivial).
595 exprIsHNF :: CoreExpr -> Bool -- True => Value-lambda, constructor, PAP
596 exprIsHNF (Var v) -- NB: There are no value args at this point
597 = isDataConWorkId v -- Catches nullary constructors,
598 -- so that [] and () are values, for example
599 || idArity v > 0 -- Catches (e.g.) primops that don't have unfoldings
600 || isEvaldUnfolding (idUnfolding v)
601 -- Check the thing's unfolding; it might be bound to a value
602 -- A worry: what if an Id's unfolding is just itself:
603 -- then we could get an infinite loop...
605 exprIsHNF (Lit l) = True
606 exprIsHNF (Type ty) = True -- Types are honorary Values;
607 -- we don't mind copying them
608 exprIsHNF (Lam b e) = isRuntimeVar b || exprIsHNF e
609 exprIsHNF (Note _ e) = exprIsHNF e
610 exprIsHNF (App e (Type _)) = exprIsHNF e
611 exprIsHNF (App e a) = app_is_value e [a]
612 exprIsHNF other = False
614 -- There is at least one value argument
615 app_is_value (Var fun) args
616 | isDataConWorkId fun -- Constructor apps are values
617 || idArity fun > valArgCount args -- Under-applied function
618 = check_args (idType fun) args
619 app_is_value (App f a) as = app_is_value f (a:as)
620 app_is_value other as = False
622 -- 'check_args' checks that unlifted-type args
623 -- are in fact guaranteed non-divergent
624 check_args fun_ty [] = True
625 check_args fun_ty (Type _ : args) = case splitForAllTy_maybe fun_ty of
626 Just (_, ty) -> check_args ty args
627 check_args fun_ty (arg : args)
628 | isUnLiftedType arg_ty = exprOkForSpeculation arg
629 | otherwise = check_args res_ty args
631 (arg_ty, res_ty) = splitFunTy fun_ty
635 exprIsConApp_maybe :: CoreExpr -> Maybe (DataCon, [CoreExpr])
636 exprIsConApp_maybe (Note (Coerce to_ty from_ty) expr)
637 = -- Maybe this is over the top, but here we try to turn
638 -- coerce (S,T) ( x, y )
640 -- ( coerce S x, coerce T y )
641 -- This happens in anger in PrelArrExts which has a coerce
642 -- case coerce memcpy a b of
644 -- where the memcpy is in the IO monad, but the call is in
646 case exprIsConApp_maybe expr of {
650 case splitTyConApp_maybe to_ty of {
652 Just (tc, tc_arg_tys) | tc /= dataConTyCon dc -> Nothing
653 | not (isVanillaDataCon dc) -> Nothing
655 -- Type constructor must match
656 -- We knock out existentials to keep matters simple(r)
658 arity = tyConArity tc
659 val_args = drop arity args
660 to_arg_tys = dataConInstArgTys dc tc_arg_tys
661 mk_coerce ty arg = mkCoerce ty arg
662 new_val_args = zipWith mk_coerce to_arg_tys val_args
664 ASSERT( all isTypeArg (take arity args) )
665 ASSERT( equalLength val_args to_arg_tys )
666 Just (dc, map Type tc_arg_tys ++ new_val_args)
669 exprIsConApp_maybe (Note _ expr)
670 = exprIsConApp_maybe expr
671 -- We ignore InlineMe notes in case we have
672 -- x = __inline_me__ (a,b)
673 -- All part of making sure that INLINE pragmas never hurt
674 -- Marcin tripped on this one when making dictionaries more inlinable
676 -- In fact, we ignore all notes. For example,
677 -- case _scc_ "foo" (C a b) of
679 -- should be optimised away, but it will be only if we look
680 -- through the SCC note.
682 exprIsConApp_maybe expr = analyse (collectArgs expr)
684 analyse (Var fun, args)
685 | Just con <- isDataConWorkId_maybe fun,
686 args `lengthAtLeast` dataConRepArity con
687 -- Might be > because the arity excludes type args
690 -- Look through unfoldings, but only cheap ones, because
691 -- we are effectively duplicating the unfolding
692 analyse (Var fun, [])
693 | let unf = idUnfolding fun,
695 = exprIsConApp_maybe (unfoldingTemplate unf)
697 analyse other = Nothing
702 %************************************************************************
704 \subsection{Eta reduction and expansion}
706 %************************************************************************
709 exprEtaExpandArity :: DynFlags -> CoreExpr -> Arity
710 {- The Arity returned is the number of value args the
711 thing can be applied to without doing much work
713 exprEtaExpandArity is used when eta expanding
716 It returns 1 (or more) to:
717 case x of p -> \s -> ...
718 because for I/O ish things we really want to get that \s to the top.
719 We are prepared to evaluate x each time round the loop in order to get that
721 It's all a bit more subtle than it looks:
725 Consider one-shot lambdas
726 let x = expensive in \y z -> E
727 We want this to have arity 2 if the \y-abstraction is a 1-shot lambda
728 Hence the ArityType returned by arityType
730 2. The state-transformer hack
732 The one-shot lambda special cause is particularly important/useful for
733 IO state transformers, where we often get
734 let x = E in \ s -> ...
736 and the \s is a real-world state token abstraction. Such abstractions
737 are almost invariably 1-shot, so we want to pull the \s out, past the
738 let x=E, even if E is expensive. So we treat state-token lambdas as
739 one-shot even if they aren't really. The hack is in Id.isOneShotBndr.
741 3. Dealing with bottom
744 f = \x -> error "foo"
745 Here, arity 1 is fine. But if it is
749 then we want to get arity 2. Tecnically, this isn't quite right, because
751 should diverge, but it'll converge if we eta-expand f. Nevertheless, we
752 do so; it improves some programs significantly, and increasing convergence
753 isn't a bad thing. Hence the ABot/ATop in ArityType.
755 Actually, the situation is worse. Consider
759 Can we eta-expand here? At first the answer looks like "yes of course", but
762 This should diverge! But if we eta-expand, it won't. Again, we ignore this
763 "problem", because being scrupulous would lose an important transformation for
769 Non-recursive newtypes are transparent, and should not get in the way.
770 We do (currently) eta-expand recursive newtypes too. So if we have, say
772 newtype T = MkT ([T] -> Int)
776 where f has arity 1. Then: etaExpandArity e = 1;
777 that is, etaExpandArity looks through the coerce.
779 When we eta-expand e to arity 1: eta_expand 1 e T
780 we want to get: coerce T (\x::[T] -> (coerce ([T]->Int) e) x)
782 HOWEVER, note that if you use coerce bogusly you can ge
784 And since negate has arity 2, you might try to eta expand. But you can't
785 decopose Int to a function type. Hence the final case in eta_expand.
789 exprEtaExpandArity dflags e = arityDepth (arityType dflags e)
791 -- A limited sort of function type
792 data ArityType = AFun Bool ArityType -- True <=> one-shot
793 | ATop -- Know nothing
796 arityDepth :: ArityType -> Arity
797 arityDepth (AFun _ ty) = 1 + arityDepth ty
800 andArityType ABot at2 = at2
801 andArityType ATop at2 = ATop
802 andArityType (AFun t1 at1) (AFun t2 at2) = AFun (t1 && t2) (andArityType at1 at2)
803 andArityType at1 at2 = andArityType at2 at1
805 arityType :: DynFlags -> CoreExpr -> ArityType
806 -- (go1 e) = [b1,..,bn]
807 -- means expression can be rewritten \x_b1 -> ... \x_bn -> body
808 -- where bi is True <=> the lambda is one-shot
810 arityType dflags (Note n e) = arityType dflags e
811 -- Not needed any more: etaExpand is cleverer
812 -- | ok_note n = arityType dflags e
813 -- | otherwise = ATop
815 arityType dflags (Var v)
816 = mk (idArity v) (arg_tys (idType v))
818 mk :: Arity -> [Type] -> ArityType
819 -- The argument types are only to steer the "state hack"
820 -- Consider case x of
822 -- False -> \(s:RealWorld) -> e
823 -- where foo has arity 1. Then we want the state hack to
824 -- apply to foo too, so we can eta expand the case.
825 mk 0 tys | isBottomingId v = ABot
826 | (ty:tys) <- tys, isStateHackType ty = AFun True ATop
828 mk n (ty:tys) = AFun (isStateHackType ty) (mk (n-1) tys)
829 mk n [] = AFun False (mk (n-1) [])
831 arg_tys :: Type -> [Type] -- Ignore for-alls
833 | Just (_, ty') <- splitForAllTy_maybe ty = arg_tys ty'
834 | Just (arg,res) <- splitFunTy_maybe ty = arg : arg_tys res
837 -- Lambdas; increase arity
838 arityType dflags (Lam x e)
839 | isId x = AFun (isOneShotBndr x) (arityType dflags e)
840 | otherwise = arityType dflags e
842 -- Applications; decrease arity
843 arityType dflags (App f (Type _)) = arityType dflags f
844 arityType dflags (App f a) = case arityType dflags f of
845 AFun one_shot xs | exprIsCheap a -> xs
848 -- Case/Let; keep arity if either the expression is cheap
849 -- or it's a 1-shot lambda
850 -- The former is not really right for Haskell
851 -- f x = case x of { (a,b) -> \y. e }
853 -- f x y = case x of { (a,b) -> e }
854 -- The difference is observable using 'seq'
855 arityType dflags (Case scrut _ _ alts)
856 = case foldr1 andArityType [arityType dflags rhs | (_,_,rhs) <- alts] of
857 xs | exprIsCheap scrut -> xs
858 xs@(AFun one_shot _) | one_shot -> AFun True ATop
861 arityType dflags (Let b e)
862 = case arityType dflags e of
863 xs | cheap_bind b -> xs
864 xs@(AFun one_shot _) | one_shot -> AFun True ATop
867 cheap_bind (NonRec b e) = is_cheap (b,e)
868 cheap_bind (Rec prs) = all is_cheap prs
869 is_cheap (b,e) = (dopt Opt_DictsCheap dflags && isDictId b)
871 -- If the experimental -fdicts-cheap flag is on, we eta-expand through
872 -- dictionary bindings. This improves arities. Thereby, it also
873 -- means that full laziness is less prone to floating out the
874 -- application of a function to its dictionary arguments, which
875 -- can thereby lose opportunities for fusion. Example:
876 -- foo :: Ord a => a -> ...
877 -- foo = /\a \(d:Ord a). let d' = ...d... in \(x:a). ....
878 -- -- So foo has arity 1
880 -- f = \x. foo dInt $ bar x
882 -- The (foo DInt) is floated out, and makes ineffective a RULE
885 -- One could go further and make exprIsCheap reply True to any
886 -- dictionary-typed expression, but that's more work.
888 arityType dflags other = ATop
890 {- NOT NEEDED ANY MORE: etaExpand is cleverer
891 ok_note InlineMe = False
893 -- Notice that we do not look through __inline_me__
894 -- This may seem surprising, but consider
895 -- f = _inline_me (\x -> e)
896 -- We DO NOT want to eta expand this to
897 -- f = \x -> (_inline_me (\x -> e)) x
898 -- because the _inline_me gets dropped now it is applied,
907 etaExpand :: Arity -- Result should have this number of value args
909 -> CoreExpr -> Type -- Expression and its type
911 -- (etaExpand n us e ty) returns an expression with
912 -- the same meaning as 'e', but with arity 'n'.
914 -- Given e' = etaExpand n us e ty
916 -- ty = exprType e = exprType e'
918 -- Note that SCCs are not treated specially. If we have
919 -- etaExpand 2 (\x -> scc "foo" e)
920 -- = (\xy -> (scc "foo" e) y)
921 -- So the costs of evaluating 'e' (not 'e y') are attributed to "foo"
923 etaExpand n us expr ty
924 | manifestArity expr >= n = expr -- The no-op case
925 | otherwise = eta_expand n us expr ty
928 -- manifestArity sees how many leading value lambdas there are
929 manifestArity :: CoreExpr -> Arity
930 manifestArity (Lam v e) | isId v = 1 + manifestArity e
931 | otherwise = manifestArity e
932 manifestArity (Note _ e) = manifestArity e
935 -- etaExpand deals with for-alls. For example:
937 -- where E :: forall a. a -> a
939 -- (/\b. \y::a -> E b y)
941 -- It deals with coerces too, though they are now rare
942 -- so perhaps the extra code isn't worth it
944 eta_expand n us expr ty
946 -- The ILX code generator requires eta expansion for type arguments
947 -- too, but alas the 'n' doesn't tell us how many of them there
948 -- may be. So we eagerly eta expand any big lambdas, and just
949 -- cross our fingers about possible loss of sharing in the ILX case.
950 -- The Right Thing is probably to make 'arity' include
951 -- type variables throughout the compiler. (ToDo.)
953 -- Saturated, so nothing to do
956 -- Short cut for the case where there already
957 -- is a lambda; no point in gratuitously adding more
958 eta_expand n us (Lam v body) ty
960 = Lam v (eta_expand n us body (applyTy ty (mkTyVarTy v)))
963 = Lam v (eta_expand (n-1) us body (funResultTy ty))
965 -- We used to have a special case that stepped inside Coerces here,
966 -- thus: eta_expand n us (Note note@(Coerce _ ty) e) _
967 -- = Note note (eta_expand n us e ty)
968 -- BUT this led to an infinite loop
969 -- Example: newtype T = MkT (Int -> Int)
970 -- eta_expand 1 (coerce (Int->Int) e)
971 -- --> coerce (Int->Int) (eta_expand 1 T e)
973 -- --> coerce (Int->Int) (coerce T
974 -- (\x::Int -> eta_expand 1 (coerce (Int->Int) e)))
975 -- by the splitNewType_maybe case below
978 eta_expand n us expr ty
979 = case splitForAllTy_maybe ty of {
980 Just (tv,ty') -> Lam tv (eta_expand n us (App expr (Type (mkTyVarTy tv))) ty')
984 case splitFunTy_maybe ty of {
985 Just (arg_ty, res_ty) -> Lam arg1 (eta_expand (n-1) us2 (App expr (Var arg1)) res_ty)
987 arg1 = mkSysLocal FSLIT("eta") uniq arg_ty
993 -- newtype T = MkT ([T] -> Int)
994 -- Consider eta-expanding this
997 -- coerce T (\x::[T] -> (coerce ([T]->Int) e) x)
998 -- Only try this for recursive newtypes; the non-recursive kind
999 -- are transparent anyway
1001 case splitRecNewType_maybe ty of {
1002 Just ty' -> mkCoerce2 ty ty' (eta_expand n us (mkCoerce2 ty' ty expr) ty') ;
1005 -- We have an expression of arity > 0, but its type isn't a function
1006 -- This *can* legitmately happen: e.g. coerce Int (\x. x)
1007 -- Essentially the programmer is playing fast and loose with types
1008 -- (Happy does this a lot). So we simply decline to eta-expand.
1013 exprArity is a cheap-and-cheerful version of exprEtaExpandArity.
1014 It tells how many things the expression can be applied to before doing
1015 any work. It doesn't look inside cases, lets, etc. The idea is that
1016 exprEtaExpandArity will do the hard work, leaving something that's easy
1017 for exprArity to grapple with. In particular, Simplify uses exprArity to
1018 compute the ArityInfo for the Id.
1020 Originally I thought that it was enough just to look for top-level lambdas, but
1021 it isn't. I've seen this
1023 foo = PrelBase.timesInt
1025 We want foo to get arity 2 even though the eta-expander will leave it
1026 unchanged, in the expectation that it'll be inlined. But occasionally it
1027 isn't, because foo is blacklisted (used in a rule).
1029 Similarly, see the ok_note check in exprEtaExpandArity. So
1030 f = __inline_me (\x -> e)
1031 won't be eta-expanded.
1033 And in any case it seems more robust to have exprArity be a bit more intelligent.
1034 But note that (\x y z -> f x y z)
1035 should have arity 3, regardless of f's arity.
1038 exprArity :: CoreExpr -> Arity
1041 go (Var v) = idArity v
1042 go (Lam x e) | isId x = go e + 1
1044 go (Note n e) = go e
1045 go (App e (Type t)) = go e
1046 go (App f a) | exprIsCheap a = (go f - 1) `max` 0
1047 -- NB: exprIsCheap a!
1048 -- f (fac x) does not have arity 2,
1049 -- even if f has arity 3!
1050 -- NB: `max 0`! (\x y -> f x) has arity 2, even if f is
1051 -- unknown, hence arity 0
1055 %************************************************************************
1057 \subsection{Equality}
1059 %************************************************************************
1061 @cheapEqExpr@ is a cheap equality test which bales out fast!
1062 True => definitely equal
1063 False => may or may not be equal
1066 cheapEqExpr :: Expr b -> Expr b -> Bool
1068 cheapEqExpr (Var v1) (Var v2) = v1==v2
1069 cheapEqExpr (Lit lit1) (Lit lit2) = lit1 == lit2
1070 cheapEqExpr (Type t1) (Type t2) = t1 `coreEqType` t2
1072 cheapEqExpr (App f1 a1) (App f2 a2)
1073 = f1 `cheapEqExpr` f2 && a1 `cheapEqExpr` a2
1075 cheapEqExpr _ _ = False
1077 exprIsBig :: Expr b -> Bool
1078 -- Returns True of expressions that are too big to be compared by cheapEqExpr
1079 exprIsBig (Lit _) = False
1080 exprIsBig (Var v) = False
1081 exprIsBig (Type t) = False
1082 exprIsBig (App f a) = exprIsBig f || exprIsBig a
1083 exprIsBig other = True
1088 tcEqExpr :: CoreExpr -> CoreExpr -> Bool
1089 -- Used in rule matching, so does *not* look through
1090 -- newtypes, predicate types; hence tcEqExpr
1092 tcEqExpr e1 e2 = tcEqExprX rn_env e1 e2
1094 rn_env = mkRnEnv2 (mkInScopeSet (exprFreeVars e1 `unionVarSet` exprFreeVars e2))
1096 tcEqExprX :: RnEnv2 -> CoreExpr -> CoreExpr -> Bool
1097 tcEqExprX env (Var v1) (Var v2) = rnOccL env v1 == rnOccR env v2
1098 tcEqExprX env (Lit lit1) (Lit lit2) = lit1 == lit2
1099 tcEqExprX env (App f1 a1) (App f2 a2) = tcEqExprX env f1 f2 && tcEqExprX env a1 a2
1100 tcEqExprX env (Lam v1 e1) (Lam v2 e2) = tcEqExprX (rnBndr2 env v1 v2) e1 e2
1101 tcEqExprX env (Let (NonRec v1 r1) e1)
1102 (Let (NonRec v2 r2) e2) = tcEqExprX env r1 r2
1103 && tcEqExprX (rnBndr2 env v1 v2) e1 e2
1104 tcEqExprX env (Let (Rec ps1) e1)
1105 (Let (Rec ps2) e2) = equalLength ps1 ps2
1106 && and (zipWith eq_rhs ps1 ps2)
1107 && tcEqExprX env' e1 e2
1109 env' = foldl2 rn_bndr2 env ps2 ps2
1110 rn_bndr2 env (b1,_) (b2,_) = rnBndr2 env b1 b2
1111 eq_rhs (_,r1) (_,r2) = tcEqExprX env' r1 r2
1112 tcEqExprX env (Case e1 v1 t1 a1)
1113 (Case e2 v2 t2 a2) = tcEqExprX env e1 e2
1114 && tcEqTypeX env t1 t2
1115 && equalLength a1 a2
1116 && and (zipWith (eq_alt env') a1 a2)
1118 env' = rnBndr2 env v1 v2
1120 tcEqExprX env (Note n1 e1) (Note n2 e2) = eq_note env n1 n2 && tcEqExprX env e1 e2
1121 tcEqExprX env (Type t1) (Type t2) = tcEqTypeX env t1 t2
1122 tcEqExprX env e1 e2 = False
1124 eq_alt env (c1,vs1,r1) (c2,vs2,r2) = c1==c2 && tcEqExprX (rnBndrs2 env vs1 vs2) r1 r2
1126 eq_note env (SCC cc1) (SCC cc2) = cc1 == cc2
1127 eq_note env (Coerce t1 f1) (Coerce t2 f2) = tcEqTypeX env t1 t2 && tcEqTypeX env f1 f2
1128 eq_note env (CoreNote s1) (CoreNote s2) = s1 == s2
1129 eq_note env other1 other2 = False
1133 %************************************************************************
1135 \subsection{The size of an expression}
1137 %************************************************************************
1140 coreBindsSize :: [CoreBind] -> Int
1141 coreBindsSize bs = foldr ((+) . bindSize) 0 bs
1143 exprSize :: CoreExpr -> Int
1144 -- A measure of the size of the expressions
1145 -- It also forces the expression pretty drastically as a side effect
1146 exprSize (Var v) = v `seq` 1
1147 exprSize (Lit lit) = lit `seq` 1
1148 exprSize (App f a) = exprSize f + exprSize a
1149 exprSize (Lam b e) = varSize b + exprSize e
1150 exprSize (Let b e) = bindSize b + exprSize e
1151 exprSize (Case e b t as) = seqType t `seq` exprSize e + varSize b + 1 + foldr ((+) . altSize) 0 as
1152 exprSize (Note n e) = noteSize n + exprSize e
1153 exprSize (Type t) = seqType t `seq` 1
1155 noteSize (SCC cc) = cc `seq` 1
1156 noteSize (Coerce t1 t2) = seqType t1 `seq` seqType t2 `seq` 1
1157 noteSize InlineMe = 1
1158 noteSize (CoreNote s) = s `seq` 1 -- hdaume: core annotations
1160 varSize :: Var -> Int
1161 varSize b | isTyVar b = 1
1162 | otherwise = seqType (idType b) `seq`
1163 megaSeqIdInfo (idInfo b) `seq`
1166 varsSize = foldr ((+) . varSize) 0
1168 bindSize (NonRec b e) = varSize b + exprSize e
1169 bindSize (Rec prs) = foldr ((+) . pairSize) 0 prs
1171 pairSize (b,e) = varSize b + exprSize e
1173 altSize (c,bs,e) = c `seq` varsSize bs + exprSize e
1177 %************************************************************************
1179 \subsection{Hashing}
1181 %************************************************************************
1184 hashExpr :: CoreExpr -> Int
1185 hashExpr e | hash < 0 = 77 -- Just in case we hit -maxInt
1188 hash = abs (hash_expr e) -- Negative numbers kill UniqFM
1190 hash_expr (Note _ e) = hash_expr e
1191 hash_expr (Let (NonRec b r) e) = hashId b
1192 hash_expr (Let (Rec ((b,r):_)) e) = hashId b
1193 hash_expr (Case _ b _ _) = hashId b
1194 hash_expr (App f e) = hash_expr f * fast_hash_expr e
1195 hash_expr (Var v) = hashId v
1196 hash_expr (Lit lit) = hashLiteral lit
1197 hash_expr (Lam b _) = hashId b
1198 hash_expr (Type t) = trace "hash_expr: type" 1 -- Shouldn't happen
1200 fast_hash_expr (Var v) = hashId v
1201 fast_hash_expr (Lit lit) = hashLiteral lit
1202 fast_hash_expr (App f (Type _)) = fast_hash_expr f
1203 fast_hash_expr (App f a) = fast_hash_expr a
1204 fast_hash_expr (Lam b _) = hashId b
1205 fast_hash_expr other = 1
1208 hashId id = hashName (idName id)
1211 %************************************************************************
1213 \subsection{Determining non-updatable right-hand-sides}
1215 %************************************************************************
1217 Top-level constructor applications can usually be allocated
1218 statically, but they can't if the constructor, or any of the
1219 arguments, come from another DLL (because we can't refer to static
1220 labels in other DLLs).
1222 If this happens we simply make the RHS into an updatable thunk,
1223 and 'exectute' it rather than allocating it statically.
1226 rhsIsStatic :: PackageId -> CoreExpr -> Bool
1227 -- This function is called only on *top-level* right-hand sides
1228 -- Returns True if the RHS can be allocated statically, with
1229 -- no thunks involved at all.
1231 -- It's called (i) in TidyPgm.hasCafRefs to decide if the rhs is, or
1232 -- refers to, CAFs; and (ii) in CoreToStg to decide whether to put an
1233 -- update flag on it.
1235 -- The basic idea is that rhsIsStatic returns True only if the RHS is
1236 -- (a) a value lambda
1237 -- (b) a saturated constructor application with static args
1239 -- BUT watch out for
1240 -- (i) Any cross-DLL references kill static-ness completely
1241 -- because they must be 'executed' not statically allocated
1242 -- ("DLL" here really only refers to Windows DLLs, on other platforms,
1243 -- this is not necessary)
1245 -- (ii) We treat partial applications as redexes, because in fact we
1246 -- make a thunk for them that runs and builds a PAP
1247 -- at run-time. The only appliations that are treated as
1248 -- static are *saturated* applications of constructors.
1250 -- We used to try to be clever with nested structures like this:
1251 -- ys = (:) w ((:) w [])
1252 -- on the grounds that CorePrep will flatten ANF-ise it later.
1253 -- But supporting this special case made the function much more
1254 -- complicated, because the special case only applies if there are no
1255 -- enclosing type lambdas:
1256 -- ys = /\ a -> Foo (Baz ([] a))
1257 -- Here the nested (Baz []) won't float out to top level in CorePrep.
1259 -- But in fact, even without -O, nested structures at top level are
1260 -- flattened by the simplifier, so we don't need to be super-clever here.
1264 -- f = \x::Int. x+7 TRUE
1265 -- p = (True,False) TRUE
1267 -- d = (fst p, False) FALSE because there's a redex inside
1268 -- (this particular one doesn't happen but...)
1270 -- h = D# (1.0## /## 2.0##) FALSE (redex again)
1271 -- n = /\a. Nil a TRUE
1273 -- t = /\a. (:) (case w a of ...) (Nil a) FALSE (redex)
1276 -- This is a bit like CoreUtils.exprIsHNF, with the following differences:
1277 -- a) scc "foo" (\x -> ...) is updatable (so we catch the right SCC)
1279 -- b) (C x xs), where C is a contructors is updatable if the application is
1282 -- c) don't look through unfolding of f in (f x).
1284 -- When opt_RuntimeTypes is on, we keep type lambdas and treat
1285 -- them as making the RHS re-entrant (non-updatable).
1287 rhsIsStatic this_pkg rhs = is_static False rhs
1289 is_static :: Bool -- True <=> in a constructor argument; must be atomic
1292 is_static False (Lam b e) = isRuntimeVar b || is_static False e
1294 is_static in_arg (Note (SCC _) e) = False
1295 is_static in_arg (Note _ e) = is_static in_arg e
1297 is_static in_arg (Lit lit)
1299 MachLabel _ _ -> False
1301 -- A MachLabel (foreign import "&foo") in an argument
1302 -- prevents a constructor application from being static. The
1303 -- reason is that it might give rise to unresolvable symbols
1304 -- in the object file: under Linux, references to "weak"
1305 -- symbols from the data segment give rise to "unresolvable
1306 -- relocation" errors at link time This might be due to a bug
1307 -- in the linker, but we'll work around it here anyway.
1310 is_static in_arg other_expr = go other_expr 0
1312 go (Var f) n_val_args
1313 #if mingw32_TARGET_OS
1314 | not (isDllName this_pkg (idName f))
1316 = saturated_data_con f n_val_args
1317 || (in_arg && n_val_args == 0)
1318 -- A naked un-applied variable is *not* deemed a static RHS
1320 -- Reason: better to update so that the indirection gets shorted
1321 -- out, and the true value will be seen
1322 -- NB: if you change this, you'll break the invariant that THUNK_STATICs
1323 -- are always updatable. If you do so, make sure that non-updatable
1324 -- ones have enough space for their static link field!
1326 go (App f a) n_val_args
1327 | isTypeArg a = go f n_val_args
1328 | not in_arg && is_static True a = go f (n_val_args + 1)
1329 -- The (not in_arg) checks that we aren't in a constructor argument;
1330 -- if we are, we don't allow (value) applications of any sort
1332 -- NB. In case you wonder, args are sometimes not atomic. eg.
1333 -- x = D# (1.0## /## 2.0##)
1334 -- can't float because /## can fail.
1336 go (Note (SCC _) f) n_val_args = False
1337 go (Note _ f) n_val_args = go f n_val_args
1339 go other n_val_args = False
1341 saturated_data_con f n_val_args
1342 = case isDataConWorkId_maybe f of
1343 Just dc -> n_val_args == dataConRepArity dc