\begin{code}
module CoreUtils (
-- Construction
- mkNote, mkInlineMe, mkSCC, mkCoerce,
+ mkInlineMe, mkSCC, mkCoerce, mkCoerce2,
bindNonRec, needsCaseBinding,
mkIfThenElse, mkAltExpr, mkPiType, mkPiTypes,
-- Taking expressions apart
- findDefault, findAlt, hasDefault,
+ findDefault, findAlt, isDefaultAlt,
-- Properties of expressions
- exprType, coreAltsType,
- exprIsBottom, exprIsDupable, exprIsTrivial, exprIsCheap,
- exprIsValue,exprOkForSpeculation, exprIsBig,
- exprIsConApp_maybe, exprIsAtom,
- idAppIsBottom, idAppIsCheap,
-
+ exprType, coreAltType,
+ exprIsDupable, exprIsTrivial, exprIsCheap,
+ exprIsHNF,exprOkForSpeculation, exprIsBig,
+ exprIsConApp_maybe, exprIsBottom,
+ rhsIsStatic,
-- Arity and eta expansion
manifestArity, exprArity,
hashExpr,
-- Equality
- cheapEqExpr, eqExpr, applyTypeToArgs, applyTypeToArg
+ cheapEqExpr, tcEqExpr, tcEqExprX, applyTypeToArgs, applyTypeToArg
) where
#include "HsVersions.h"
-import GlaExts -- For `xori`
+import GLAEXTS -- For `xori`
import CoreSyn
+import CoreFVs ( exprFreeVars )
import PprCore ( pprCoreExpr )
-import Var ( Var, isId, isTyVar )
+import Var ( Var )
+import VarSet ( unionVarSet )
import VarEnv
import Name ( hashName )
-import Literal ( hashLiteral, literalType, litIsDupable, isZeroLit )
-import DataCon ( DataCon, dataConRepArity, dataConArgTys, isExistentialDataCon, dataConTyCon )
+import Packages ( HomeModules )
+#if mingw32_TARGET_OS
+import Packages ( isDllName )
+#endif
+import Literal ( hashLiteral, literalType, litIsDupable,
+ litIsTrivial, isZeroLit, Literal( MachLabel ) )
+import DataCon ( DataCon, dataConRepArity, dataConInstArgTys,
+ isVanillaDataCon, dataConTyCon )
import PrimOp ( PrimOp(..), primOpOkForSpeculation, primOpIsCheap )
-import Id ( Id, idType, globalIdDetails, idNewStrictness, idLBVarInfo,
- mkWildId, idArity, idName, idUnfolding, idInfo, isOneShotLambda,
- isDataConId_maybe, mkSysLocal, isDataConId, isBottomingId
+import Id ( Id, idType, globalIdDetails, idNewStrictness,
+ mkWildId, idArity, idName, idUnfolding, idInfo,
+ isOneShotBndr, isStateHackType, isDataConWorkId_maybe, mkSysLocal,
+ isDataConWorkId, isBottomingId
)
-import IdInfo ( LBVarInfo(..),
- GlobalIdDetails(..),
- megaSeqIdInfo )
+import IdInfo ( GlobalIdDetails(..), megaSeqIdInfo )
import NewDemand ( appIsBottom )
-import Type ( Type, mkFunTy, mkForAllTy, splitFunTy_maybe, splitFunTy,
+import Type ( Type, mkFunTy, mkForAllTy, splitFunTy_maybe,
+ splitFunTy, tcEqTypeX,
applyTys, isUnLiftedType, seqType, mkTyVarTy,
- splitForAllTy_maybe, isForAllTy, splitNewType_maybe,
- splitTyConApp_maybe, eqType, funResultTy, applyTy,
- funResultTy, applyTy
+ splitForAllTy_maybe, isForAllTy, splitRecNewType_maybe,
+ splitTyConApp_maybe, coreEqType, funResultTy, applyTy
)
import TyCon ( tyConArity )
import TysWiredIn ( boolTy, trueDataCon, falseDataCon )
import Unique ( Unique )
import Outputable
import TysPrim ( alphaTy ) -- Debugging only
-import Util ( equalLength, lengthAtLeast )
+import Util ( equalLength, lengthAtLeast, foldl2 )
\end{code}
exprType (Var var) = idType var
exprType (Lit lit) = literalType lit
exprType (Let _ body) = exprType body
-exprType (Case _ _ alts) = coreAltsType alts
-exprType (Note (Coerce ty _) e) = ty -- **! should take usage from e
+exprType (Case _ _ ty alts) = ty
+exprType (Note (Coerce ty _) e) = ty -- **! should take usage from e
exprType (Note other_note e) = exprType e
exprType (Lam binder expr) = mkPiType binder (exprType expr)
exprType e@(App _ _)
exprType other = pprTrace "exprType" (pprCoreExpr other) alphaTy
-coreAltsType :: [CoreAlt] -> Type
-coreAltsType ((_,_,rhs) : _) = exprType rhs
+coreAltType :: CoreAlt -> Type
+coreAltType (_,_,rhs) = exprType rhs
\end{code}
@mkPiType@ makes a (->) type or a forall type, depending on whether
mkNote removes redundant coercions, and SCCs where possible
\begin{code}
+#ifdef UNUSED
mkNote :: Note -> CoreExpr -> CoreExpr
-mkNote (Coerce to_ty from_ty) expr = mkCoerce to_ty from_ty expr
+mkNote (Coerce to_ty from_ty) expr = mkCoerce2 to_ty from_ty expr
mkNote (SCC cc) expr = mkSCC cc expr
mkNote InlineMe expr = mkInlineMe expr
mkNote note expr = Note note expr
+#endif
-- Slide InlineCall in around the function
-- No longer necessary I think (SLPJ Apr 99)
\begin{code}
-mkCoerce :: Type -> Type -> CoreExpr -> CoreExpr
+mkCoerce :: Type -> CoreExpr -> CoreExpr
+mkCoerce to_ty expr = mkCoerce2 to_ty (exprType expr) expr
-mkCoerce to_ty from_ty (Note (Coerce to_ty2 from_ty2) expr)
- = ASSERT( from_ty `eqType` to_ty2 )
- mkCoerce to_ty from_ty2 expr
+mkCoerce2 :: Type -> Type -> CoreExpr -> CoreExpr
+mkCoerce2 to_ty from_ty (Note (Coerce to_ty2 from_ty2) expr)
+ = ASSERT( from_ty `coreEqType` to_ty2 )
+ mkCoerce2 to_ty from_ty2 expr
-mkCoerce to_ty from_ty expr
- | to_ty `eqType` from_ty = expr
- | otherwise = ASSERT( from_ty `eqType` exprType expr )
+mkCoerce2 to_ty from_ty expr
+ | to_ty `coreEqType` from_ty = expr
+ | otherwise = ASSERT( from_ty `coreEqType` exprType expr )
Note (Coerce to_ty from_ty) expr
\end{code}
-- It's used by the desugarer to avoid building bindings
-- that give Core Lint a heart attack. Actually the simplifier
-- deals with them perfectly well.
+
bindNonRec bndr rhs body
- | needsCaseBinding (idType bndr) rhs = Case rhs bndr [(DEFAULT,[],body)]
+ | needsCaseBinding (idType bndr) rhs = Case rhs bndr (exprType body) [(DEFAULT,[],body)]
| otherwise = Let (NonRec bndr rhs) body
needsCaseBinding ty rhs = isUnLiftedType ty && not (exprOkForSpeculation rhs)
mkIfThenElse :: CoreExpr -> CoreExpr -> CoreExpr -> CoreExpr
mkIfThenElse guard then_expr else_expr
- = Case guard (mkWildId boolTy)
- [ (DataAlt trueDataCon, [], then_expr),
- (DataAlt falseDataCon, [], else_expr) ]
+-- Not going to be refining, so okay to take the type of the "then" clause
+ = Case guard (mkWildId boolTy) (exprType then_expr)
+ [ (DataAlt falseDataCon, [], else_expr), -- Increasing order of tag!
+ (DataAlt trueDataCon, [], then_expr) ]
\end{code}
This makes it easy to find, though it makes matching marginally harder.
\begin{code}
-hasDefault :: [CoreAlt] -> Bool
-hasDefault ((DEFAULT,_,_) : alts) = True
-hasDefault _ = False
-
findDefault :: [CoreAlt] -> ([CoreAlt], Maybe CoreExpr)
findDefault ((DEFAULT,args,rhs) : alts) = ASSERT( null args ) (alts, Just rhs)
findDefault alts = (alts, Nothing)
= case alts of
(deflt@(DEFAULT,_,_):alts) -> go alts deflt
other -> go alts panic_deflt
-
where
panic_deflt = pprPanic "Missing alternative" (ppr con $$ vcat (map ppr alts))
- go [] deflt = deflt
- go (alt@(con1,_,_) : alts) deflt | con == con1 = alt
- | otherwise = ASSERT( not (con1 == DEFAULT) )
- go alts deflt
+ go [] deflt = deflt
+ go (alt@(con1,_,_) : alts) deflt
+ = case con `cmpAltCon` con1 of
+ LT -> deflt -- Missed it already; the alts are in increasing order
+ EQ -> alt
+ GT -> ASSERT( not (con1 == DEFAULT) ) go alts deflt
+
+isDefaultAlt :: CoreAlt -> Bool
+isDefaultAlt (DEFAULT, _, _) = True
+isDefaultAlt other = False
\end{code}
rare that I plan to allow them to be duplicated and put up with
saturating them.
+SCC notes. We do not treat (_scc_ "foo" x) as trivial, because
+ a) it really generates code, (and a heap object when it's
+ a function arg) to capture the cost centre
+ b) see the note [SCC-and-exprIsTrivial] in Simplify.simplLazyBind
+
\begin{code}
exprIsTrivial (Var v) = True -- See notes above
exprIsTrivial (Type _) = True
-exprIsTrivial (Lit lit) = True
+exprIsTrivial (Lit lit) = litIsTrivial lit
exprIsTrivial (App e arg) = not (isRuntimeArg arg) && exprIsTrivial e
-exprIsTrivial (Note _ e) = exprIsTrivial e
+exprIsTrivial (Note (SCC _) e) = False -- See notes above
+exprIsTrivial (Note _ e) = exprIsTrivial e
exprIsTrivial (Lam b body) = not (isRuntimeVar b) && exprIsTrivial body
exprIsTrivial other = False
-
-exprIsAtom :: CoreExpr -> Bool
--- Used to decide whether to let-binding an STG argument
--- when compiling to ILX => type applications are not allowed
-exprIsAtom (Var v) = True -- primOpIsDupable?
-exprIsAtom (Lit lit) = True
-exprIsAtom (Type ty) = True
-exprIsAtom (Note (SCC _) e) = False
-exprIsAtom (Note _ e) = exprIsAtom e
-exprIsAtom other = False
\end{code}
\begin{code}
exprIsCheap :: CoreExpr -> Bool
-exprIsCheap (Lit lit) = True
-exprIsCheap (Type _) = True
-exprIsCheap (Var _) = True
-exprIsCheap (Note InlineMe e) = True
-exprIsCheap (Note _ e) = exprIsCheap e
-exprIsCheap (Lam x e) = isRuntimeVar x || exprIsCheap e
-exprIsCheap (Case e _ alts) = exprIsCheap e &&
+exprIsCheap (Lit lit) = True
+exprIsCheap (Type _) = True
+exprIsCheap (Var _) = True
+exprIsCheap (Note InlineMe e) = True
+exprIsCheap (Note _ e) = exprIsCheap e
+exprIsCheap (Lam x e) = isRuntimeVar x || exprIsCheap e
+exprIsCheap (Case e _ _ alts) = exprIsCheap e &&
and [exprIsCheap rhs | (_,_,rhs) <- alts]
-- Experimentally, treat (case x of ...) as cheap
-- (and case __coerce x etc.)
| n_val_args == 0 = True -- Just a type application of
-- a variable (f t1 t2 t3)
-- counts as WHNF
- | otherwise = case globalIdDetails id of
- DataConId _ -> True
- RecordSelId _ -> True -- I'm experimenting with making record selection
- -- look cheap, so we will substitute it inside a
- -- lambda. Particularly for dictionary field selection
-
- PrimOpId op -> primOpIsCheap op -- In principle we should worry about primops
- -- that return a type variable, since the result
- -- might be applied to something, but I'm not going
- -- to bother to check the number of args
- other -> n_val_args < idArity id
+ | otherwise
+ = case globalIdDetails id of
+ DataConWorkId _ -> True
+ RecordSelId {} -> n_val_args == 1 -- I'm experimenting with making record selection
+ ClassOpId _ -> n_val_args == 1 -- look cheap, so we will substitute it inside a
+ -- lambda. Particularly for dictionary field selection.
+ -- BUT: Take care with (sel d x)! The (sel d) might be cheap, but
+ -- there's no guarantee that (sel d x) will be too. Hence (n_val_args == 1)
+
+ PrimOpId op -> primOpIsCheap op -- In principle we should worry about primops
+ -- that return a type variable, since the result
+ -- might be applied to something, but I'm not going
+ -- to bother to check the number of args
+ other -> n_val_args < idArity id
\end{code}
exprOkForSpeculation returns True of an expression that it is
other -> False
where
- spec_ok (DataConId _) args
+ spec_ok (DataConWorkId _) args
= True -- The strictness of the constructor has already
-- been expressed by its "wrapper", so we don't need
-- to take the arguments into account
exprIsBottom e = go 0 e
where
-- n is the number of args
- go n (Note _ e) = go n e
- go n (Let _ e) = go n e
- go n (Case e _ _) = go 0 e -- Just check the scrut
- go n (App e _) = go (n+1) e
- go n (Var v) = idAppIsBottom v n
- go n (Lit _) = False
- go n (Lam _ _) = False
+ go n (Note _ e) = go n e
+ go n (Let _ e) = go n e
+ go n (Case e _ _ _) = go 0 e -- Just check the scrut
+ go n (App e _) = go (n+1) e
+ go n (Var v) = idAppIsBottom v n
+ go n (Lit _) = False
+ go n (Lam _ _) = False
+ go n (Type _) = False
idAppIsBottom :: Id -> Int -> Bool
idAppIsBottom id n_val_args = appIsBottom (idNewStrictness id) n_val_args
\end{code}
-@exprIsValue@ returns true for expressions that are certainly *already*
+@exprIsHNF@ returns true for expressions that are certainly *already*
evaluated to *head* normal form. This is used to decide whether it's ok
to change
type must be ok-for-speculation (or trivial).
\begin{code}
-exprIsValue :: CoreExpr -> Bool -- True => Value-lambda, constructor, PAP
-exprIsValue (Type ty) = True -- Types are honorary Values; we don't mind
- -- copying them
-exprIsValue (Lit l) = True
-exprIsValue (Lam b e) = isRuntimeVar b || exprIsValue e
-exprIsValue (Note _ e) = exprIsValue e
-exprIsValue (Var v) = idArity v > 0 || isEvaldUnfolding (idUnfolding v)
- -- The idArity case catches data cons and primops that
- -- don't have unfoldings
+exprIsHNF :: CoreExpr -> Bool -- True => Value-lambda, constructor, PAP
+exprIsHNF (Var v) -- NB: There are no value args at this point
+ = isDataConWorkId v -- Catches nullary constructors,
+ -- so that [] and () are values, for example
+ || idArity v > 0 -- Catches (e.g.) primops that don't have unfoldings
+ || isEvaldUnfolding (idUnfolding v)
+ -- Check the thing's unfolding; it might be bound to a value
-- A worry: what if an Id's unfolding is just itself:
-- then we could get an infinite loop...
-exprIsValue other_expr
- | (Var fun, args) <- collectArgs other_expr,
- isDataConId fun || valArgCount args < idArity fun
- = check (idType fun) args
- | otherwise
- = False
+
+exprIsHNF (Lit l) = True
+exprIsHNF (Type ty) = True -- Types are honorary Values;
+ -- we don't mind copying them
+exprIsHNF (Lam b e) = isRuntimeVar b || exprIsHNF e
+exprIsHNF (Note _ e) = exprIsHNF e
+exprIsHNF (App e (Type _)) = exprIsHNF e
+exprIsHNF (App e a) = app_is_value e [a]
+exprIsHNF other = False
+
+-- There is at least one value argument
+app_is_value (Var fun) args
+ | isDataConWorkId fun -- Constructor apps are values
+ || idArity fun > valArgCount args -- Under-applied function
+ = check_args (idType fun) args
+app_is_value (App f a) as = app_is_value f (a:as)
+app_is_value other as = False
+
+ -- 'check_args' checks that unlifted-type args
+ -- are in fact guaranteed non-divergent
+check_args fun_ty [] = True
+check_args fun_ty (Type _ : args) = case splitForAllTy_maybe fun_ty of
+ Just (_, ty) -> check_args ty args
+check_args fun_ty (arg : args)
+ | isUnLiftedType arg_ty = exprOkForSpeculation arg
+ | otherwise = check_args res_ty args
where
- -- 'check' checks that unlifted-type args are in
- -- fact guaranteed non-divergent
- check fun_ty [] = True
- check fun_ty (Type _ : args) = case splitForAllTy_maybe fun_ty of
- Just (_, ty) -> check ty args
- check fun_ty (arg : args)
- | isUnLiftedType arg_ty = exprOkForSpeculation arg
- | otherwise = check res_ty args
- where
- (arg_ty, res_ty) = splitFunTy fun_ty
+ (arg_ty, res_ty) = splitFunTy fun_ty
\end{code}
\begin{code}
case splitTyConApp_maybe to_ty of {
Nothing -> Nothing ;
- Just (tc, tc_arg_tys) | tc /= dataConTyCon dc -> Nothing
- | isExistentialDataCon dc -> Nothing
- | otherwise ->
+ Just (tc, tc_arg_tys) | tc /= dataConTyCon dc -> Nothing
+ | not (isVanillaDataCon dc) -> Nothing
+ | otherwise ->
-- Type constructor must match
-- We knock out existentials to keep matters simple(r)
let
arity = tyConArity tc
val_args = drop arity args
- to_arg_tys = dataConArgTys dc tc_arg_tys
- mk_coerce ty arg = mkCoerce ty (exprType arg) arg
+ to_arg_tys = dataConInstArgTys dc tc_arg_tys
+ mk_coerce ty arg = mkCoerce ty arg
new_val_args = zipWith mk_coerce to_arg_tys val_args
in
ASSERT( all isTypeArg (take arity args) )
exprIsConApp_maybe expr = analyse (collectArgs expr)
where
analyse (Var fun, args)
- | Just con <- isDataConId_maybe fun,
+ | Just con <- isDataConWorkId_maybe fun,
args `lengthAtLeast` dataConRepArity con
-- Might be > because the arity excludes type args
= Just (con,args)
\begin{code}
exprEtaExpandArity :: CoreExpr -> Arity
--- The Int is number of value args the thing can be
--- applied to without doing much work
---
--- This is used when eta expanding
--- e ==> \xy -> e x y
---
--- It returns 1 (or more) to:
--- case x of p -> \s -> ...
--- because for I/O ish things we really want to get that \s to the top.
--- We are prepared to evaluate x each time round the loop in order to get that
-
--- It's all a bit more subtle than it looks. Consider one-shot lambdas
--- let x = expensive in \y z -> E
--- We want this to have arity 2 if the \y-abstraction is a 1-shot lambda
--- Hence the ArityType returned by arityType
-
--- NB: this is particularly important/useful for IO state
--- transformers, where we often get
--- let x = E in \ s -> ...
--- and the \s is a real-world state token abstraction. Such
--- abstractions are almost invariably 1-shot, so we want to
--- pull the \s out, past the let x=E.
--- The hack is in Id.isOneShotLambda
---
--- Consider also
--- f = \x -> error "foo"
--- Here, arity 1 is fine. But if it is
--- f = \x -> case e of
--- True -> error "foo"
--- False -> \y -> x+y
--- then we want to get arity 2.
--- Hence the ABot/ATop in ArityType
+{- The Arity returned is the number of value args the
+ thing can be applied to without doing much work
+
+exprEtaExpandArity is used when eta expanding
+ e ==> \xy -> e x y
+
+It returns 1 (or more) to:
+ case x of p -> \s -> ...
+because for I/O ish things we really want to get that \s to the top.
+We are prepared to evaluate x each time round the loop in order to get that
+
+It's all a bit more subtle than it looks:
+
+1. One-shot lambdas
+
+Consider one-shot lambdas
+ let x = expensive in \y z -> E
+We want this to have arity 2 if the \y-abstraction is a 1-shot lambda
+Hence the ArityType returned by arityType
+
+2. The state-transformer hack
+
+The one-shot lambda special cause is particularly important/useful for
+IO state transformers, where we often get
+ let x = E in \ s -> ...
+
+and the \s is a real-world state token abstraction. Such abstractions
+are almost invariably 1-shot, so we want to pull the \s out, past the
+let x=E, even if E is expensive. So we treat state-token lambdas as
+one-shot even if they aren't really. The hack is in Id.isOneShotBndr.
+
+3. Dealing with bottom
+
+Consider also
+ f = \x -> error "foo"
+Here, arity 1 is fine. But if it is
+ f = \x -> case x of
+ True -> error "foo"
+ False -> \y -> x+y
+then we want to get arity 2. Tecnically, this isn't quite right, because
+ (f True) `seq` 1
+should diverge, but it'll converge if we eta-expand f. Nevertheless, we
+do so; it improves some programs significantly, and increasing convergence
+isn't a bad thing. Hence the ABot/ATop in ArityType.
+
+Actually, the situation is worse. Consider
+ f = \x -> case x of
+ True -> \y -> x+y
+ False -> \y -> x-y
+Can we eta-expand here? At first the answer looks like "yes of course", but
+consider
+ (f bot) `seq` 1
+This should diverge! But if we eta-expand, it won't. Again, we ignore this
+"problem", because being scrupulous would lose an important transformation for
+many programs.
+
+
+4. Newtypes
+
+Non-recursive newtypes are transparent, and should not get in the way.
+We do (currently) eta-expand recursive newtypes too. So if we have, say
+
+ newtype T = MkT ([T] -> Int)
+
+Suppose we have
+ e = coerce T f
+where f has arity 1. Then: etaExpandArity e = 1;
+that is, etaExpandArity looks through the coerce.
+
+When we eta-expand e to arity 1: eta_expand 1 e T
+we want to get: coerce T (\x::[T] -> (coerce ([T]->Int) e) x)
+
+HOWEVER, note that if you use coerce bogusly you can ge
+ coerce Int negate
+And since negate has arity 2, you might try to eta expand. But you can't
+decopose Int to a function type. Hence the final case in eta_expand.
+-}
exprEtaExpandArity e = arityDepth (arityType e)
-- | otherwise = ATop
arityType (Var v)
- = mk (idArity v)
+ = mk (idArity v) (arg_tys (idType v))
where
- mk :: Arity -> ArityType
- mk 0 | isBottomingId v = ABot
- | otherwise = ATop
- mk n = AFun False (mk (n-1))
-
- -- When the type of the Id encodes one-shot-ness,
- -- use the idinfo here
+ mk :: Arity -> [Type] -> ArityType
+ -- The argument types are only to steer the "state hack"
+ -- Consider case x of
+ -- True -> foo
+ -- False -> \(s:RealWorld) -> e
+ -- where foo has arity 1. Then we want the state hack to
+ -- apply to foo too, so we can eta expand the case.
+ mk 0 tys | isBottomingId v = ABot
+ | otherwise = ATop
+ mk n (ty:tys) = AFun (isStateHackType ty) (mk (n-1) tys)
+ mk n [] = AFun False (mk (n-1) [])
+
+ arg_tys :: Type -> [Type] -- Ignore for-alls
+ arg_tys ty
+ | Just (_, ty') <- splitForAllTy_maybe ty = arg_tys ty'
+ | Just (arg,res) <- splitFunTy_maybe ty = arg : arg_tys res
+ | otherwise = []
-- Lambdas; increase arity
-arityType (Lam x e) | isId x = AFun (isOneShotLambda x) (arityType e)
+arityType (Lam x e) | isId x = AFun (isOneShotBndr x) (arityType e)
| otherwise = arityType e
-- Applications; decrease arity
arityType (App f (Type _)) = arityType f
arityType (App f a) = case arityType f of
- AFun one_shot xs | one_shot -> xs
- | exprIsCheap a -> xs
+ AFun one_shot xs | exprIsCheap a -> xs
other -> ATop
-- Case/Let; keep arity if either the expression is cheap
-- or it's a 1-shot lambda
-arityType (Case scrut _ alts) = case foldr1 andArityType [arityType rhs | (_,_,rhs) <- alts] of
+ -- The former is not really right for Haskell
+ -- f x = case x of { (a,b) -> \y. e }
+ -- ===>
+ -- f x y = case x of { (a,b) -> e }
+ -- The difference is observable using 'seq'
+arityType (Case scrut _ _ alts) = case foldr1 andArityType [arityType rhs | (_,_,rhs) <- alts] of
xs@(AFun one_shot _) | one_shot -> xs
xs | exprIsCheap scrut -> xs
| otherwise -> ATop
-- Given e' = etaExpand n us e ty
-- We should have
-- ty = exprType e = exprType e'
+--
+-- Note that SCCs are not treated specially. If we have
+-- etaExpand 2 (\x -> scc "foo" e)
+-- = (\xy -> (scc "foo" e) y)
+-- So the costs of evaluating 'e' (not 'e y') are attributed to "foo"
etaExpand n us expr ty
| manifestArity expr >= n = expr -- The no-op case
-- The ILX code generator requires eta expansion for type arguments
-- too, but alas the 'n' doesn't tell us how many of them there
-- may be. So we eagerly eta expand any big lambdas, and just
- -- cross our fingers about possible loss of sharing in the
- -- ILX case.
+ -- cross our fingers about possible loss of sharing in the ILX case.
-- The Right Thing is probably to make 'arity' include
-- type variables throughout the compiler. (ToDo.)
not (isForAllTy ty)
-- Saturated, so nothing to do
= expr
-eta_expand n us (Note note@(Coerce _ ty) e) _
- = Note note (eta_expand n us e ty)
-
- -- Use mkNote so that _scc_s get pushed inside any lambdas that
- -- are generated as part of the eta expansion. We rely on this
- -- behaviour in CorePrep, when we eta expand an already-prepped RHS.
-eta_expand n us (Note note e) ty
- = mkNote note (eta_expand n us e ty)
-
-- Short cut for the case where there already
-- is a lambda; no point in gratuitously adding more
eta_expand n us (Lam v body) ty
| otherwise
= Lam v (eta_expand (n-1) us body (funResultTy ty))
+-- We used to have a special case that stepped inside Coerces here,
+-- thus: eta_expand n us (Note note@(Coerce _ ty) e) _
+-- = Note note (eta_expand n us e ty)
+-- BUT this led to an infinite loop
+-- Example: newtype T = MkT (Int -> Int)
+-- eta_expand 1 (coerce (Int->Int) e)
+-- --> coerce (Int->Int) (eta_expand 1 T e)
+-- by the bogus eqn
+-- --> coerce (Int->Int) (coerce T
+-- (\x::Int -> eta_expand 1 (coerce (Int->Int) e)))
+-- by the splitNewType_maybe case below
+-- and round we go
+
eta_expand n us expr ty
= case splitForAllTy_maybe ty of {
Just (tv,ty') -> Lam tv (eta_expand n us (App expr (Type (mkTyVarTy tv))) ty')
case splitFunTy_maybe ty of {
Just (arg_ty, res_ty) -> Lam arg1 (eta_expand (n-1) us2 (App expr (Var arg1)) res_ty)
where
- arg1 = mkSysLocal SLIT("eta") uniq arg_ty
+ arg1 = mkSysLocal FSLIT("eta") uniq arg_ty
(uniq:us2) = us
; Nothing ->
- case splitNewType_maybe ty of {
- Just ty' -> mkCoerce ty ty' (eta_expand n us (mkCoerce ty' ty expr) ty') ;
- Nothing -> pprTrace "Bad eta expand" (ppr expr $$ ppr ty) expr
+ -- Given this:
+ -- newtype T = MkT ([T] -> Int)
+ -- Consider eta-expanding this
+ -- eta_expand 1 e T
+ -- We want to get
+ -- coerce T (\x::[T] -> (coerce ([T]->Int) e) x)
+ -- Only try this for recursive newtypes; the non-recursive kind
+ -- are transparent anyway
+
+ case splitRecNewType_maybe ty of {
+ Just ty' -> mkCoerce2 ty ty' (eta_expand n us (mkCoerce2 ty' ty expr) ty') ;
+ Nothing ->
+
+ -- We have an expression of arity > 0, but its type isn't a function
+ -- This *can* legitmately happen: e.g. coerce Int (\x. x)
+ -- Essentially the programmer is playing fast and loose with types
+ -- (Happy does this a lot). So we simply decline to eta-expand.
+ expr
}}}
\end{code}
cheapEqExpr (Var v1) (Var v2) = v1==v2
cheapEqExpr (Lit lit1) (Lit lit2) = lit1 == lit2
-cheapEqExpr (Type t1) (Type t2) = t1 `eqType` t2
+cheapEqExpr (Type t1) (Type t2) = t1 `coreEqType` t2
cheapEqExpr (App f1 a1) (App f2 a2)
= f1 `cheapEqExpr` f2 && a1 `cheapEqExpr` a2
\begin{code}
-eqExpr :: CoreExpr -> CoreExpr -> Bool
- -- Works ok at more general type, but only needed at CoreExpr
- -- Used in rule matching, so when we find a type we use
- -- eqTcType, which doesn't look through newtypes
- -- [And it doesn't risk falling into a black hole either.]
-eqExpr e1 e2
- = eq emptyVarEnv e1 e2
+tcEqExpr :: CoreExpr -> CoreExpr -> Bool
+-- Used in rule matching, so does *not* look through
+-- newtypes, predicate types; hence tcEqExpr
+
+tcEqExpr e1 e2 = tcEqExprX rn_env e1 e2
where
- -- The "env" maps variables in e1 to variables in ty2
- -- So when comparing lambdas etc,
- -- we in effect substitute v2 for v1 in e1 before continuing
- eq env (Var v1) (Var v2) = case lookupVarEnv env v1 of
- Just v1' -> v1' == v2
- Nothing -> v1 == v2
-
- eq env (Lit lit1) (Lit lit2) = lit1 == lit2
- eq env (App f1 a1) (App f2 a2) = eq env f1 f2 && eq env a1 a2
- eq env (Lam v1 e1) (Lam v2 e2) = eq (extendVarEnv env v1 v2) e1 e2
- eq env (Let (NonRec v1 r1) e1)
- (Let (NonRec v2 r2) e2) = eq env r1 r2 && eq (extendVarEnv env v1 v2) e1 e2
- eq env (Let (Rec ps1) e1)
- (Let (Rec ps2) e2) = equalLength ps1 ps2 &&
- and (zipWith eq_rhs ps1 ps2) &&
- eq env' e1 e2
+ rn_env = mkRnEnv2 (mkInScopeSet (exprFreeVars e1 `unionVarSet` exprFreeVars e2))
+
+tcEqExprX :: RnEnv2 -> CoreExpr -> CoreExpr -> Bool
+tcEqExprX env (Var v1) (Var v2) = rnOccL env v1 == rnOccR env v2
+tcEqExprX env (Lit lit1) (Lit lit2) = lit1 == lit2
+tcEqExprX env (App f1 a1) (App f2 a2) = tcEqExprX env f1 f2 && tcEqExprX env a1 a2
+tcEqExprX env (Lam v1 e1) (Lam v2 e2) = tcEqExprX (rnBndr2 env v1 v2) e1 e2
+tcEqExprX env (Let (NonRec v1 r1) e1)
+ (Let (NonRec v2 r2) e2) = tcEqExprX env r1 r2
+ && tcEqExprX (rnBndr2 env v1 v2) e1 e2
+tcEqExprX env (Let (Rec ps1) e1)
+ (Let (Rec ps2) e2) = equalLength ps1 ps2
+ && and (zipWith eq_rhs ps1 ps2)
+ && tcEqExprX env' e1 e2
where
- env' = extendVarEnvList env [(v1,v2) | ((v1,_),(v2,_)) <- zip ps1 ps2]
- eq_rhs (_,r1) (_,r2) = eq env' r1 r2
- eq env (Case e1 v1 a1)
- (Case e2 v2 a2) = eq env e1 e2 &&
- equalLength a1 a2 &&
- and (zipWith (eq_alt env') a1 a2)
+ env' = foldl2 rn_bndr2 env ps2 ps2
+ rn_bndr2 env (b1,_) (b2,_) = rnBndr2 env b1 b2
+ eq_rhs (_,r1) (_,r2) = tcEqExprX env' r1 r2
+tcEqExprX env (Case e1 v1 t1 a1)
+ (Case e2 v2 t2 a2) = tcEqExprX env e1 e2
+ && tcEqTypeX env t1 t2
+ && equalLength a1 a2
+ && and (zipWith (eq_alt env') a1 a2)
where
- env' = extendVarEnv env v1 v2
+ env' = rnBndr2 env v1 v2
- eq env (Note n1 e1) (Note n2 e2) = eq_note env n1 n2 && eq env e1 e2
- eq env (Type t1) (Type t2) = t1 `eqType` t2
- eq env e1 e2 = False
+tcEqExprX env (Note n1 e1) (Note n2 e2) = eq_note env n1 n2 && tcEqExprX env e1 e2
+tcEqExprX env (Type t1) (Type t2) = tcEqTypeX env t1 t2
+tcEqExprX env e1 e2 = False
- eq_list env [] [] = True
- eq_list env (e1:es1) (e2:es2) = eq env e1 e2 && eq_list env es1 es2
- eq_list env es1 es2 = False
-
- eq_alt env (c1,vs1,r1) (c2,vs2,r2) = c1==c2 &&
- eq (extendVarEnvList env (vs1 `zip` vs2)) r1 r2
-
- eq_note env (SCC cc1) (SCC cc2) = cc1 == cc2
- eq_note env (Coerce t1 f1) (Coerce t2 f2) = t1 `eqType` t2 && f1 `eqType` f2
- eq_note env InlineCall InlineCall = True
- eq_note env other1 other2 = False
+eq_alt env (c1,vs1,r1) (c2,vs2,r2) = c1==c2 && tcEqExprX (rnBndrs2 env vs1 vs2) r1 r2
+
+eq_note env (SCC cc1) (SCC cc2) = cc1 == cc2
+eq_note env (Coerce t1 f1) (Coerce t2 f2) = tcEqTypeX env t1 t2 && tcEqTypeX env f1 f2
+eq_note env InlineCall InlineCall = True
+eq_note env (CoreNote s1) (CoreNote s2) = s1 == s2
+eq_note env other1 other2 = False
\end{code}
exprSize :: CoreExpr -> Int
-- A measure of the size of the expressions
-- It also forces the expression pretty drastically as a side effect
-exprSize (Var v) = v `seq` 1
-exprSize (Lit lit) = lit `seq` 1
-exprSize (App f a) = exprSize f + exprSize a
-exprSize (Lam b e) = varSize b + exprSize e
-exprSize (Let b e) = bindSize b + exprSize e
-exprSize (Case e b as) = exprSize e + varSize b + foldr ((+) . altSize) 0 as
-exprSize (Note n e) = noteSize n + exprSize e
-exprSize (Type t) = seqType t `seq` 1
+exprSize (Var v) = v `seq` 1
+exprSize (Lit lit) = lit `seq` 1
+exprSize (App f a) = exprSize f + exprSize a
+exprSize (Lam b e) = varSize b + exprSize e
+exprSize (Let b e) = bindSize b + exprSize e
+exprSize (Case e b t as) = seqType t `seq` exprSize e + varSize b + 1 + foldr ((+) . altSize) 0 as
+exprSize (Note n e) = noteSize n + exprSize e
+exprSize (Type t) = seqType t `seq` 1
noteSize (SCC cc) = cc `seq` 1
noteSize (Coerce t1 t2) = seqType t1 `seq` seqType t2 `seq` 1
noteSize InlineCall = 1
noteSize InlineMe = 1
+noteSize (CoreNote s) = s `seq` 1 -- hdaume: core annotations
varSize :: Var -> Int
varSize b | isTyVar b = 1
hash_expr (Note _ e) = hash_expr e
hash_expr (Let (NonRec b r) e) = hashId b
hash_expr (Let (Rec ((b,r):_)) e) = hashId b
-hash_expr (Case _ b _) = hashId b
+hash_expr (Case _ b _ _) = hashId b
hash_expr (App f e) = hash_expr f * fast_hash_expr e
hash_expr (Var v) = hashId v
hash_expr (Lit lit) = hashLiteral lit
hashId :: Id -> Int
hashId id = hashName (idName id)
\end{code}
+
+%************************************************************************
+%* *
+\subsection{Determining non-updatable right-hand-sides}
+%* *
+%************************************************************************
+
+Top-level constructor applications can usually be allocated
+statically, but they can't if the constructor, or any of the
+arguments, come from another DLL (because we can't refer to static
+labels in other DLLs).
+
+If this happens we simply make the RHS into an updatable thunk,
+and 'exectute' it rather than allocating it statically.
+
+\begin{code}
+rhsIsStatic :: HomeModules -> CoreExpr -> Bool
+-- This function is called only on *top-level* right-hand sides
+-- Returns True if the RHS can be allocated statically, with
+-- no thunks involved at all.
+--
+-- It's called (i) in TidyPgm.hasCafRefs to decide if the rhs is, or
+-- refers to, CAFs; and (ii) in CoreToStg to decide whether to put an
+-- update flag on it.
+--
+-- The basic idea is that rhsIsStatic returns True only if the RHS is
+-- (a) a value lambda
+-- (b) a saturated constructor application with static args
+--
+-- BUT watch out for
+-- (i) Any cross-DLL references kill static-ness completely
+-- because they must be 'executed' not statically allocated
+-- ("DLL" here really only refers to Windows DLLs, on other platforms,
+-- this is not necessary)
+--
+-- (ii) We treat partial applications as redexes, because in fact we
+-- make a thunk for them that runs and builds a PAP
+-- at run-time. The only appliations that are treated as
+-- static are *saturated* applications of constructors.
+
+-- We used to try to be clever with nested structures like this:
+-- ys = (:) w ((:) w [])
+-- on the grounds that CorePrep will flatten ANF-ise it later.
+-- But supporting this special case made the function much more
+-- complicated, because the special case only applies if there are no
+-- enclosing type lambdas:
+-- ys = /\ a -> Foo (Baz ([] a))
+-- Here the nested (Baz []) won't float out to top level in CorePrep.
+--
+-- But in fact, even without -O, nested structures at top level are
+-- flattened by the simplifier, so we don't need to be super-clever here.
+--
+-- Examples
+--
+-- f = \x::Int. x+7 TRUE
+-- p = (True,False) TRUE
+--
+-- d = (fst p, False) FALSE because there's a redex inside
+-- (this particular one doesn't happen but...)
+--
+-- h = D# (1.0## /## 2.0##) FALSE (redex again)
+-- n = /\a. Nil a TRUE
+--
+-- t = /\a. (:) (case w a of ...) (Nil a) FALSE (redex)
+--
+--
+-- This is a bit like CoreUtils.exprIsHNF, with the following differences:
+-- a) scc "foo" (\x -> ...) is updatable (so we catch the right SCC)
+--
+-- b) (C x xs), where C is a contructors is updatable if the application is
+-- dynamic
+--
+-- c) don't look through unfolding of f in (f x).
+--
+-- When opt_RuntimeTypes is on, we keep type lambdas and treat
+-- them as making the RHS re-entrant (non-updatable).
+
+rhsIsStatic hmods rhs = is_static False rhs
+ where
+ is_static :: Bool -- True <=> in a constructor argument; must be atomic
+ -> CoreExpr -> Bool
+
+ is_static False (Lam b e) = isRuntimeVar b || is_static False e
+
+ is_static in_arg (Note (SCC _) e) = False
+ is_static in_arg (Note _ e) = is_static in_arg e
+
+ is_static in_arg (Lit lit)
+ = case lit of
+ MachLabel _ _ -> False
+ other -> True
+ -- A MachLabel (foreign import "&foo") in an argument
+ -- prevents a constructor application from being static. The
+ -- reason is that it might give rise to unresolvable symbols
+ -- in the object file: under Linux, references to "weak"
+ -- symbols from the data segment give rise to "unresolvable
+ -- relocation" errors at link time This might be due to a bug
+ -- in the linker, but we'll work around it here anyway.
+ -- SDM 24/2/2004
+
+ is_static in_arg other_expr = go other_expr 0
+ where
+ go (Var f) n_val_args
+#if mingw32_TARGET_OS
+ | not (isDllName hmods (idName f))
+#endif
+ = saturated_data_con f n_val_args
+ || (in_arg && n_val_args == 0)
+ -- A naked un-applied variable is *not* deemed a static RHS
+ -- E.g. f = g
+ -- Reason: better to update so that the indirection gets shorted
+ -- out, and the true value will be seen
+ -- NB: if you change this, you'll break the invariant that THUNK_STATICs
+ -- are always updatable. If you do so, make sure that non-updatable
+ -- ones have enough space for their static link field!
+
+ go (App f a) n_val_args
+ | isTypeArg a = go f n_val_args
+ | not in_arg && is_static True a = go f (n_val_args + 1)
+ -- The (not in_arg) checks that we aren't in a constructor argument;
+ -- if we are, we don't allow (value) applications of any sort
+ --
+ -- NB. In case you wonder, args are sometimes not atomic. eg.
+ -- x = D# (1.0## /## 2.0##)
+ -- can't float because /## can fail.
+
+ go (Note (SCC _) f) n_val_args = False
+ go (Note _ f) n_val_args = go f n_val_args
+
+ go other n_val_args = False
+
+ saturated_data_con f n_val_args
+ = case isDataConWorkId_maybe f of
+ Just dc -> n_val_args == dataConRepArity dc
+ Nothing -> False
+\end{code}