module CoreUtils (
-- Construction
mkNote, mkInlineMe, mkSCC, mkCoerce,
- bindNonRec, mkIfThenElse, mkAltExpr,
- mkPiType,
+ bindNonRec, needsCaseBinding,
+ mkIfThenElse, mkAltExpr, mkPiType, mkPiTypes,
+
+ -- Taking expressions apart
+ findDefault, findAlt, hasDefault,
-- Properties of expressions
exprType, coreAltsType,
exprIsBottom, exprIsDupable, exprIsTrivial, exprIsCheap,
exprIsValue,exprOkForSpeculation, exprIsBig,
- exprIsConApp_maybe,
+ exprIsConApp_maybe, exprIsAtom,
idAppIsBottom, idAppIsCheap,
- exprArity,
- -- Expr transformation
- etaReduce, etaExpand,
- exprArity, exprEtaExpandArity,
+
+ -- Arity and eta expansion
+ manifestArity, exprArity,
+ exprEtaExpandArity, etaExpand,
-- Size
coreBindsSize,
hashExpr,
-- Equality
- cheapEqExpr, eqExpr, applyTypeToArgs
+ cheapEqExpr, eqExpr, applyTypeToArgs, applyTypeToArg
) where
#include "HsVersions.h"
import GlaExts -- For `xori`
import CoreSyn
-import CoreFVs ( exprFreeVars )
import PprCore ( pprCoreExpr )
import Var ( Var, isId, isTyVar )
-import VarSet
import VarEnv
import Name ( hashName )
-import Literal ( hashLiteral, literalType, litIsDupable )
-import DataCon ( DataCon, dataConRepArity )
-import PrimOp ( primOpOkForSpeculation, primOpIsCheap,
- primOpIsDupable )
-import Id ( Id, idType, idFlavour, idStrictness, idLBVarInfo,
- mkWildId, idArity, idName, idUnfolding, idInfo,
- isDataConId_maybe, isPrimOpId_maybe, mkSysLocal
+import Literal ( hashLiteral, literalType, litIsDupable, isZeroLit )
+import DataCon ( DataCon, dataConRepArity, dataConArgTys, isExistentialDataCon, 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 IdInfo ( LBVarInfo(..),
- IdFlavour(..),
+ GlobalIdDetails(..),
megaSeqIdInfo )
-import Demand ( appIsBottom )
-import Type ( Type, mkFunTy, mkForAllTy, splitFunTy_maybe,
- applyTys, isUnLiftedType, seqType, mkUTy, mkTyVarTy,
- splitForAllTy_maybe, splitNewType_maybe
+import NewDemand ( appIsBottom )
+import Type ( Type, mkFunTy, mkForAllTy, splitFunTy_maybe, splitFunTy,
+ applyTys, isUnLiftedType, seqType, mkTyVarTy,
+ splitForAllTy_maybe, isForAllTy, splitNewType_maybe,
+ splitTyConApp_maybe, eqType, funResultTy, applyTy,
+ funResultTy, applyTy
)
+import TyCon ( tyConArity )
import TysWiredIn ( boolTy, trueDataCon, falseDataCon )
import CostCentre ( CostCentre )
-import UniqSupply ( UniqSupply, splitUniqSupply, uniqFromSupply )
-import Maybes ( maybeToBool )
+import BasicTypes ( Arity )
+import Unique ( Unique )
import Outputable
import TysPrim ( alphaTy ) -- Debugging only
+import Util ( equalLength, lengthAtLeast )
\end{code}
case of a term variable.
\begin{code}
-mkPiType :: Var -> Type -> Type -- The more polymorphic version doesn't work...
-mkPiType v ty | isId v = (case idLBVarInfo v of
- LBVarInfo u -> mkUTy u
- otherwise -> id) $
- mkFunTy (idType v) ty
- | isTyVar v = mkForAllTy v ty
+mkPiType :: Var -> Type -> Type -- The more polymorphic version
+mkPiTypes :: [Var] -> Type -> Type -- doesn't work...
+
+mkPiTypes vs ty = foldr mkPiType ty vs
+
+mkPiType v ty
+ | isId v = mkFunTy (idType v) ty
+ | otherwise = mkForAllTy v ty
\end{code}
\begin{code}
--- The first argument is just for debugging
+applyTypeToArg :: Type -> CoreExpr -> Type
+applyTypeToArg fun_ty (Type arg_ty) = applyTy fun_ty arg_ty
+applyTypeToArg fun_ty other_arg = funResultTy fun_ty
+
applyTypeToArgs :: CoreExpr -> Type -> [CoreExpr] -> Type
+-- A more efficient version of applyTypeToArg
+-- when we have several args
+-- The first argument is just for debugging
applyTypeToArgs e op_ty [] = op_ty
applyTypeToArgs e op_ty (Type ty : args)
= -- Accumulate type arguments so we can instantiate all at once
- applyTypeToArgs e (applyTys op_ty tys) rest_args
+ go [ty] args
where
- (tys, rest_args) = go [ty] args
- go tys (Type ty : args) = go (ty:tys) args
- go tys rest_args = (reverse tys, rest_args)
+ go rev_tys (Type ty : args) = go (ty:rev_tys) args
+ go rev_tys rest_args = applyTypeToArgs e op_ty' rest_args
+ where
+ op_ty' = applyTys op_ty (reverse rev_tys)
applyTypeToArgs e op_ty (other_arg : args)
= case (splitFunTy_maybe op_ty) of
As usual, the inline_me prevents the worker from getting inlined back into the wrapper.
We want the split, so that the coerces can cancel at the call site.
+However, we can get left with tiresome type applications. Notably, consider
+ f = /\ a -> let t = e in (t, w)
+Then lifting the let out of the big lambda gives
+ t' = /\a -> e
+ f = /\ a -> let t = inline_me (t' a) in (t, w)
+The inline_me is to stop the simplifier inlining t' right back
+into t's RHS. In the next phase we'll substitute for t (since
+its rhs is trivial) and *then* we could get rid of the inline_me.
+But it hardly seems worth it, so I don't bother.
+
\begin{code}
mkInlineMe (Var v) = Var v
mkInlineMe e = Note InlineMe e
mkCoerce :: Type -> Type -> CoreExpr -> CoreExpr
mkCoerce to_ty from_ty (Note (Coerce to_ty2 from_ty2) expr)
- = ASSERT( from_ty == to_ty2 )
+ = ASSERT( from_ty `eqType` to_ty2 )
mkCoerce to_ty from_ty2 expr
mkCoerce to_ty from_ty expr
- | to_ty == from_ty = expr
- | otherwise = ASSERT( from_ty == exprType expr )
- Note (Coerce to_ty from_ty) expr
+ | to_ty `eqType` from_ty = expr
+ | otherwise = ASSERT( from_ty `eqType` exprType expr )
+ Note (Coerce to_ty from_ty) expr
\end{code}
\begin{code}
mkSCC :: CostCentre -> Expr b -> Expr b
-- Note: Nested SCC's *are* preserved for the benefit of
- -- cost centre stack profiling (Durham)
-
-mkSCC cc (Lit lit) = Lit lit
-mkSCC cc (Lam x e) = Lam x (mkSCC cc e) -- Move _scc_ inside lambda
-mkSCC cc expr = Note (SCC cc) expr
+ -- cost centre stack profiling
+mkSCC cc (Lit lit) = Lit lit
+mkSCC cc (Lam x e) = Lam x (mkSCC cc e) -- Move _scc_ inside lambda
+mkSCC cc (Note (SCC cc') e) = Note (SCC cc) (Note (SCC cc') e)
+mkSCC cc (Note n e) = Note n (mkSCC cc e) -- Move _scc_ inside notes
+mkSCC cc expr = Note (SCC cc) expr
\end{code}
-- that give Core Lint a heart attack. Actually the simplifier
-- deals with them perfectly well.
bindNonRec bndr rhs body
- | isUnLiftedType (idType bndr) = Case rhs bndr [(DEFAULT,[],body)]
- | otherwise = Let (NonRec bndr rhs) body
+ | needsCaseBinding (idType bndr) rhs = Case rhs bndr [(DEFAULT,[],body)]
+ | otherwise = Let (NonRec bndr rhs) body
+
+needsCaseBinding ty rhs = isUnLiftedType ty && not (exprOkForSpeculation rhs)
+ -- Make a case expression instead of a let
+ -- These can arise either from the desugarer,
+ -- or from beta reductions: (\x.e) (x +# y)
\end{code}
\begin{code}
(DataAlt falseDataCon, [], else_expr) ]
\end{code}
+
+%************************************************************************
+%* *
+\subsection{Taking expressions apart}
+%* *
+%************************************************************************
+
+The default alternative must be first, if it exists at all.
+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)
+
+findAlt :: AltCon -> [CoreAlt] -> CoreAlt
+findAlt con alts
+ = 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
+\end{code}
+
+
%************************************************************************
%* *
\subsection{Figuring out things about expressions}
@exprIsBottom@ is true of expressions that are guaranteed to diverge
+There used to be a gruesome test for (hasNoBinding v) in the
+Var case:
+ exprIsTrivial (Var v) | hasNoBinding v = idArity v == 0
+The idea here is that a constructor worker, like $wJust, is
+really short for (\x -> $wJust x), becuase $wJust has no binding.
+So it should be treated like a lambda. Ditto unsaturated primops.
+But now constructor workers are not "have-no-binding" Ids. And
+completely un-applied primops and foreign-call Ids are sufficiently
+rare that I plan to allow them to be duplicated and put up with
+saturating them.
+
\begin{code}
-exprIsTrivial (Var v)
- | Just op <- isPrimOpId_maybe v = primOpIsDupable op
- | otherwise = True
-exprIsTrivial (Type _) = True
-exprIsTrivial (Lit lit) = True
-exprIsTrivial (App e arg) = isTypeArg arg && exprIsTrivial e
-exprIsTrivial (Note _ e) = exprIsTrivial e
-exprIsTrivial (Lam b body) | isTyVar b = exprIsTrivial body
-exprIsTrivial other = False
+exprIsTrivial (Var v) = True -- See notes above
+exprIsTrivial (Type _) = True
+exprIsTrivial (Lit lit) = True
+exprIsTrivial (App e arg) = not (isRuntimeArg arg) && exprIsTrivial e
+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}
-exprIsDupable (Type _) = True
-exprIsDupable (Var v) = True
-exprIsDupable (Lit lit) = litIsDupable lit
-exprIsDupable (Note _ e) = exprIsDupable e
+exprIsDupable (Type _) = True
+exprIsDupable (Var v) = True
+exprIsDupable (Lit lit) = litIsDupable lit
+exprIsDupable (Note InlineMe e) = True
+exprIsDupable (Note _ e) = exprIsDupable e
exprIsDupable expr
= go expr 0
where
exprIsCheap (Lit lit) = True
exprIsCheap (Type _) = True
exprIsCheap (Var _) = True
+exprIsCheap (Note InlineMe e) = True
exprIsCheap (Note _ e) = exprIsCheap e
-exprIsCheap (Lam x e) = if isId x then True else 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
-- because it certainly doesn't need to be shared!
go (App f a) n_args args_cheap
- | isTypeArg a = go f n_args args_cheap
- | otherwise = go f (n_args + 1) (exprIsCheap a && args_cheap)
+ | not (isRuntimeArg a) = go f n_args args_cheap
+ | otherwise = go f (n_args + 1) (exprIsCheap a && args_cheap)
go other n_args args_cheap = False
| n_val_args == 0 = True -- Just a type application of
-- a variable (f t1 t2 t3)
-- counts as WHNF
- | otherwise = case idFlavour id of
+ | 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
\begin{code}
exprOkForSpeculation :: CoreExpr -> Bool
exprOkForSpeculation (Lit _) = True
+exprOkForSpeculation (Type _) = True
exprOkForSpeculation (Var v) = isUnLiftedType (idType v)
exprOkForSpeculation (Note _ e) = exprOkForSpeculation e
exprOkForSpeculation other_expr
- = go other_expr 0 True
+ = case collectArgs other_expr of
+ (Var f, args) -> spec_ok (globalIdDetails f) args
+ other -> False
+
where
- go (Var f) n_args args_ok
- = case idFlavour f of
- DataConId _ -> True -- The strictness of the constructor has already
- -- been expressed by its "wrapper", so we don't need
- -- to take the arguments into account
-
- PrimOpId op -> primOpOkForSpeculation op && args_ok
+ spec_ok (DataConId _) 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
+
+ spec_ok (PrimOpId op) args
+ | isDivOp op, -- Special case for dividing operations that fail
+ [arg1, Lit lit] <- args -- only if the divisor is zero
+ = not (isZeroLit lit) && exprOkForSpeculation arg1
+ -- Often there is a literal divisor, and this
+ -- can get rid of a thunk in an inner looop
+
+ | otherwise
+ = primOpOkForSpeculation op &&
+ all exprOkForSpeculation args
-- A bit conservative: we don't really need
-- to care about lazy arguments, but this is easy
- other -> False
-
- go (App f a) n_args args_ok
- | isTypeArg a = go f n_args args_ok
- | otherwise = go f (n_args + 1) (exprOkForSpeculation a && args_ok)
-
- go other n_args args_ok = False
+ spec_ok other args = False
+
+isDivOp :: PrimOp -> Bool
+-- True of dyadic operators that can fail
+-- only if the second arg is zero
+-- This function probably belongs in PrimOp, or even in
+-- an automagically generated file.. but it's such a
+-- special case I thought I'd leave it here for now.
+isDivOp IntQuotOp = True
+isDivOp IntRemOp = True
+isDivOp WordQuotOp = True
+isDivOp WordRemOp = True
+isDivOp IntegerQuotRemOp = True
+isDivOp IntegerDivModOp = True
+isDivOp FloatDivOp = True
+isDivOp DoubleDivOp = True
+isDivOp other = False
\end{code}
go n (Lam _ _) = False
idAppIsBottom :: Id -> Int -> Bool
-idAppIsBottom id n_val_args = appIsBottom (idStrictness id) n_val_args
+idAppIsBottom id n_val_args = appIsBottom (idNewStrictness id) n_val_args
\end{code}
@exprIsValue@ returns true for expressions that are certainly *already*
-evaluated to WHNF. This is used to decide wether it's ok to change
+evaluated to *head* normal form. This is used to decide whether it's ok
+to change
+
case x of _ -> e ===> e
and to decide whether it's safe to discard a `seq`
-So, it does *not* treat variables as evaluated, unless they say they are
+So, it does *not* treat variables as evaluated, unless they say they are.
+
+But it *does* treat partial applications and constructor applications
+as values, even if their arguments are non-trivial, provided the argument
+type is lifted;
+ e.g. (:) (f x) (map f xs) is a value
+ map (...redex...) is a value
+Because `seq` on such things completes immediately
+
+For unlifted argument types, we have to be careful:
+ C (f x :: Int#)
+Suppose (f x) diverges; then C (f x) is not a value. True, but
+this form is illegal (see the invariants in CoreSyn). Args of unboxed
+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) = isId b || exprIsValue e
+exprIsValue (Lam b e) = isRuntimeVar b || exprIsValue e
exprIsValue (Note _ e) = exprIsValue e
-exprIsValue other_expr
- = go other_expr 0
- where
- go (Var f) n_args = idAppIsValue f n_args
-
- go (App f a) n_args
- | isTypeArg a = go f n_args
- | otherwise = go f (n_args + 1)
-
- go (Note _ f) n_args = go f n_args
-
- go other n_args = False
-
-idAppIsValue :: Id -> Int -> Bool
-idAppIsValue id n_val_args
- = case idFlavour id of
- DataConId _ -> True
- PrimOpId _ -> n_val_args < idArity id
- other | n_val_args == 0 -> isEvaldUnfolding (idUnfolding id)
- | otherwise -> n_val_args < idArity id
+exprIsValue (Var v) = idArity v > 0 || isEvaldUnfolding (idUnfolding v)
+ -- The idArity case catches data cons and primops that
+ -- don't have unfoldings
-- 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
+ 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
\end{code}
\begin{code}
exprIsConApp_maybe :: CoreExpr -> Maybe (DataCon, [CoreExpr])
-exprIsConApp_maybe expr
- = analyse (collectArgs expr)
+exprIsConApp_maybe (Note (Coerce to_ty from_ty) expr)
+ = -- Maybe this is over the top, but here we try to turn
+ -- coerce (S,T) ( x, y )
+ -- effectively into
+ -- ( coerce S x, coerce T y )
+ -- This happens in anger in PrelArrExts which has a coerce
+ -- case coerce memcpy a b of
+ -- (# r, s #) -> ...
+ -- where the memcpy is in the IO monad, but the call is in
+ -- the (ST s) monad
+ case exprIsConApp_maybe expr of {
+ Nothing -> Nothing ;
+ Just (dc, args) ->
+
+ case splitTyConApp_maybe to_ty of {
+ Nothing -> Nothing ;
+ Just (tc, tc_arg_tys) | tc /= dataConTyCon dc -> Nothing
+ | isExistentialDataCon 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
+ new_val_args = zipWith mk_coerce to_arg_tys val_args
+ in
+ ASSERT( all isTypeArg (take arity args) )
+ ASSERT( equalLength val_args to_arg_tys )
+ Just (dc, map Type tc_arg_tys ++ new_val_args)
+ }}
+
+exprIsConApp_maybe (Note _ expr)
+ = exprIsConApp_maybe expr
+ -- We ignore InlineMe notes in case we have
+ -- x = __inline_me__ (a,b)
+ -- All part of making sure that INLINE pragmas never hurt
+ -- Marcin tripped on this one when making dictionaries more inlinable
+ --
+ -- In fact, we ignore all notes. For example,
+ -- case _scc_ "foo" (C a b) of
+ -- C a b -> e
+ -- should be optimised away, but it will be only if we look
+ -- through the SCC note.
+
+exprIsConApp_maybe expr = analyse (collectArgs expr)
where
analyse (Var fun, args)
- | maybeToBool maybe_con_app = maybe_con_app
- where
- maybe_con_app = case isDataConId_maybe fun of
- Just con | length args >= dataConRepArity con
- -- Might be > because the arity excludes type args
- -> Just (con, args)
- other -> Nothing
+ | Just con <- isDataConId_maybe fun,
+ args `lengthAtLeast` dataConRepArity con
+ -- Might be > because the arity excludes type args
+ = Just (con,args)
+ -- Look through unfoldings, but only cheap ones, because
+ -- we are effectively duplicating the unfolding
analyse (Var fun, [])
- = case maybeUnfoldingTemplate (idUnfolding fun) of
- Nothing -> Nothing
- Just unf -> exprIsConApp_maybe unf
+ | let unf = idUnfolding fun,
+ isCheapUnfolding unf
+ = exprIsConApp_maybe (unfoldingTemplate unf)
analyse other = Nothing
\end{code}
-The arity of an expression (in the code-generator sense, i.e. the
-number of lambdas at the beginning).
-
-\begin{code}
-exprArity :: CoreExpr -> Int
-exprArity (Lam x e)
- | isTyVar x = exprArity e
- | otherwise = 1 + exprArity e
-exprArity (Note _ e)
- -- Ignore coercions. Top level sccs are removed by the final
- -- profiling pass, so we ignore those too.
- = exprArity e
-exprArity _ = 0
-\end{code}
%************************************************************************
%* *
%************************************************************************
-@etaReduce@ trys an eta reduction at the top level of a Core Expr.
-
-e.g. \ x y -> f x y ===> f
-
-But we only do this if it gets rid of a whole lambda, not part.
-The idea is that lambdas are often quite helpful: they indicate
-head normal forms, so we don't want to chuck them away lightly.
-
-\begin{code}
-etaReduce :: CoreExpr -> CoreExpr
- -- ToDo: we should really check that we don't turn a non-bottom
- -- lambda into a bottom variable. Sigh
-
-etaReduce expr@(Lam bndr body)
- = check (reverse binders) body
- where
- (binders, body) = collectBinders expr
-
- check [] body
- | not (any (`elemVarSet` body_fvs) binders)
- = body -- Success!
- where
- body_fvs = exprFreeVars body
-
- check (b : bs) (App fun arg)
- | (varToCoreExpr b `cheapEqExpr` arg)
- = check bs fun
-
- check _ _ = expr -- Bale out
-
-etaReduce expr = expr -- The common case
-\end{code}
-
-
\begin{code}
-exprEtaExpandArity :: CoreExpr -> (Int, Bool)
+exprEtaExpandArity :: CoreExpr -> Arity
-- The Int is number of value args the thing can be
-- applied to without doing much work
--- The Bool is True iff there are enough explicit value lambdas
--- at the top to make this arity apparent
--- (but ignore it when arity==0)
-
+--
-- This is used when eta expanding
-- e ==> \xy -> e x y
--
-- 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
--- Hence "generous" arity
-exprEtaExpandArity e
- = go 0 e
+-- 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
+
+
+exprEtaExpandArity e = arityDepth (arityType e)
+
+-- A limited sort of function type
+data ArityType = AFun Bool ArityType -- True <=> one-shot
+ | ATop -- Know nothing
+ | ABot -- Diverges
+
+arityDepth :: ArityType -> Arity
+arityDepth (AFun _ ty) = 1 + arityDepth ty
+arityDepth ty = 0
+
+andArityType ABot at2 = at2
+andArityType ATop at2 = ATop
+andArityType (AFun t1 at1) (AFun t2 at2) = AFun (t1 && t2) (andArityType at1 at2)
+andArityType at1 at2 = andArityType at2 at1
+
+arityType :: CoreExpr -> ArityType
+ -- (go1 e) = [b1,..,bn]
+ -- means expression can be rewritten \x_b1 -> ... \x_bn -> body
+ -- where bi is True <=> the lambda is one-shot
+
+arityType (Note n e) = arityType e
+-- Not needed any more: etaExpand is cleverer
+-- | ok_note n = arityType e
+-- | otherwise = ATop
+
+arityType (Var v)
+ = mk (idArity v)
where
- go ar (Lam x e) | isId x = go (ar+1) e
- | otherwise = go ar e
- go ar (Note n e) | ok_note n = go ar e
- go ar other = (ar + ar', ar' == 0)
- where
- ar' = go1 other `max` 0
-
- go1 (Var v) = idArity v
- go1 (Lam x e) | isId x = go1 e + 1
- | otherwise = go1 e
- go1 (Note n e) | ok_note n = go1 e
- go1 (App f (Type _)) = go1 f
- go1 (App f a) | exprIsCheap a = go1 f - 1
- go1 (Case scrut _ alts)
- | exprIsCheap scrut = min_zero [go1 rhs | (_,_,rhs) <- alts]
- go1 (Let b e)
- | all exprIsCheap (rhssOfBind b) = go1 e
-
- go1 other = 0
-
- ok_note (Coerce _ _) = True
- ok_note InlineCall = True
- ok_note other = False
- -- Notice that we do not look through __inline_me__
- -- This one is a bit more surprising, but consider
- -- f = _inline_me (\x -> e)
- -- We DO NOT want to eta expand this to
- -- f = \x -> (_inline_me (\x -> e)) x
- -- because the _inline_me gets dropped now it is applied,
- -- giving just
- -- f = \x -> e
- -- A Bad Idea
-
-min_zero :: [Int] -> Int -- Find the minimum, but zero is the smallest
-min_zero (x:xs) = go x xs
- where
- go 0 xs = 0 -- Nothing beats zero
- go min [] = min
- go min (x:xs) | x < min = go x xs
- | otherwise = go min xs
-
+ 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
+
+ -- Lambdas; increase arity
+arityType (Lam x e) | isId x = AFun (isOneShotLambda 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
+ 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
+ xs@(AFun one_shot _) | one_shot -> xs
+ xs | exprIsCheap scrut -> xs
+ | otherwise -> ATop
+
+arityType (Let b e) = case arityType e of
+ xs@(AFun one_shot _) | one_shot -> xs
+ xs | all exprIsCheap (rhssOfBind b) -> xs
+ | otherwise -> ATop
+
+arityType other = ATop
+
+{- NOT NEEDED ANY MORE: etaExpand is cleverer
+ok_note InlineMe = False
+ok_note other = True
+ -- Notice that we do not look through __inline_me__
+ -- This may seem surprising, but consider
+ -- f = _inline_me (\x -> e)
+ -- We DO NOT want to eta expand this to
+ -- f = \x -> (_inline_me (\x -> e)) x
+ -- because the _inline_me gets dropped now it is applied,
+ -- giving just
+ -- f = \x -> e
+ -- A Bad Idea
+-}
\end{code}
\begin{code}
-etaExpand :: Int -- Add this number of value args
- -> UniqSupply
+etaExpand :: Arity -- Result should have this number of value args
+ -> [Unique]
-> CoreExpr -> Type -- Expression and its type
-> CoreExpr
-- (etaExpand n us e ty) returns an expression with
-- the same meaning as 'e', but with arity 'n'.
-
+--
-- Given e' = etaExpand n us e ty
-- We should have
-- ty = exprType e = exprType e'
---
--- etaExpand deals with for-alls and coerces. For example:
+
+etaExpand n us expr ty
+ | manifestArity expr >= n = expr -- The no-op case
+ | otherwise = eta_expand n us expr ty
+ where
+
+-- manifestArity sees how many leading value lambdas there are
+manifestArity :: CoreExpr -> Arity
+manifestArity (Lam v e) | isId v = 1 + manifestArity e
+ | otherwise = manifestArity e
+manifestArity (Note _ e) = manifestArity e
+manifestArity e = 0
+
+-- etaExpand deals with for-alls. For example:
-- etaExpand 1 E
--- where E :: forall a. T
--- newtype T = MkT (A -> B)
---
+-- where E :: forall a. a -> a
-- would return
--- (/\b. coerce T (\y::A -> (coerce (A->B) (E b) y)
+-- (/\b. \y::a -> E b y)
+--
+-- It deals with coerces too, though they are now rare
+-- so perhaps the extra code isn't worth it
+
+eta_expand n us expr ty
+ | n == 0 &&
+ -- 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.
+ -- The Right Thing is probably to make 'arity' include
+ -- type variables throughout the compiler. (ToDo.)
+ not (isForAllTy ty)
+ -- Saturated, so nothing to do
+ = expr
--- (case x of { I# x -> /\ a -> coerce T E)
+eta_expand n us (Note note@(Coerce _ ty) e) _
+ = Note note (eta_expand n us e ty)
-etaExpand n us expr ty
- | n == 0 -- Saturated, so nothing to do
- = expr
+ -- 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
+ | isTyVar v
+ = Lam v (eta_expand n us body (applyTy ty (mkTyVarTy v)))
+
+ | otherwise
+ = Lam v (eta_expand (n-1) us body (funResultTy ty))
- | otherwise -- An unsaturated constructor or primop; eta expand it
+eta_expand n us expr ty
= case splitForAllTy_maybe ty of {
- Just (tv,ty') -> Lam tv (etaExpand n us (App expr (Type (mkTyVarTy tv))) ty')
+ Just (tv,ty') -> Lam tv (eta_expand n us (App expr (Type (mkTyVarTy tv))) ty')
; Nothing ->
case splitFunTy_maybe ty of {
- Just (arg_ty, res_ty) -> Lam arg1 (etaExpand (n-1) us2 (App expr (Var arg1)) res_ty)
+ 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
- (us1, us2) = splitUniqSupply us
- uniq = uniqFromSupply us1
+ (uniq:us2) = us
- ; Nothing ->
-
+ ; Nothing ->
+
case splitNewType_maybe ty of {
- Just ty' -> mkCoerce ty ty' (etaExpand n us (mkCoerce ty' ty expr) ty') ;
-
- Nothing -> pprTrace "Bad eta expand" (ppr expr $$ ppr ty) expr
+ Just ty' -> mkCoerce ty ty' (eta_expand n us (mkCoerce ty' ty expr) ty') ;
+ Nothing -> pprTrace "Bad eta expand" (ppr expr $$ ppr ty) expr
}}}
\end{code}
+exprArity is a cheap-and-cheerful version of exprEtaExpandArity.
+It tells how many things the expression can be applied to before doing
+any work. It doesn't look inside cases, lets, etc. The idea is that
+exprEtaExpandArity will do the hard work, leaving something that's easy
+for exprArity to grapple with. In particular, Simplify uses exprArity to
+compute the ArityInfo for the Id.
+
+Originally I thought that it was enough just to look for top-level lambdas, but
+it isn't. I've seen this
+
+ foo = PrelBase.timesInt
+
+We want foo to get arity 2 even though the eta-expander will leave it
+unchanged, in the expectation that it'll be inlined. But occasionally it
+isn't, because foo is blacklisted (used in a rule).
+
+Similarly, see the ok_note check in exprEtaExpandArity. So
+ f = __inline_me (\x -> e)
+won't be eta-expanded.
+
+And in any case it seems more robust to have exprArity be a bit more intelligent.
+But note that (\x y z -> f x y z)
+should have arity 3, regardless of f's arity.
+
+\begin{code}
+exprArity :: CoreExpr -> Arity
+exprArity e = go e
+ where
+ go (Var v) = idArity v
+ go (Lam x e) | isId x = go e + 1
+ | otherwise = go e
+ go (Note n e) = go e
+ go (App e (Type t)) = go e
+ go (App f a) | exprIsCheap a = (go f - 1) `max` 0
+ -- NB: exprIsCheap a!
+ -- f (fac x) does not have arity 2,
+ -- even if f has arity 3!
+ -- NB: `max 0`! (\x y -> f x) has arity 2, even if f is
+ -- unknown, hence arity 0
+ go _ = 0
+\end{code}
%************************************************************************
%* *
cheapEqExpr (Var v1) (Var v2) = v1==v2
cheapEqExpr (Lit lit1) (Lit lit2) = lit1 == lit2
-cheapEqExpr (Type t1) (Type t2) = t1 == t2
+cheapEqExpr (Type t1) (Type t2) = t1 `eqType` 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
where
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) = length ps1 == length ps2 &&
+ (Let (Rec ps2) e2) = equalLength ps1 ps2 &&
and (zipWith eq_rhs ps1 ps2) &&
eq env' e1 e2
where
eq_rhs (_,r1) (_,r2) = eq env' r1 r2
eq env (Case e1 v1 a1)
(Case e2 v2 a2) = eq env e1 e2 &&
- length a1 == length a2 &&
+ equalLength a1 a2 &&
and (zipWith (eq_alt env') a1 a2)
where
env' = extendVarEnv 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 == t2
+ eq env (Type t1) (Type t2) = t1 `eqType` t2
eq env e1 e2 = False
eq_list env [] [] = True
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==t2 && f1==f2
+ 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
\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) = varSize v
+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