%
-% (c) The AQUA Project, Glasgow University, 1993-1996
+% (c) The AQUA Project, Glasgow University, 1993-1998
%
\section[SimplUtils]{The simplifier utilities}
\begin{code}
module SimplUtils (
+ simplBinder, simplBinders, simplRecIds, simplLetId,
+ tryRhsTyLam, tryEtaExpansion,
+ mkCase,
- newId, newIds,
+ -- The continuation type
+ SimplCont(..), DupFlag(..), contIsDupable, contResultType,
+ countValArgs, countArgs, mkRhsStop, mkStop,
+ getContArgs, interestingCallContext, interestingArg, isStrictType, discardInline
- floatExposesHNF,
-
- etaCoreExpr, mkRhsTyLam,
-
- etaExpandCount,
-
- simplIdWantsToBeINLINEd,
-
- singleConstructorType, typeOkForCase,
-
- substSpecEnvRhs
) where
#include "HsVersions.h"
-import BinderInfo
-import CmdLineOpts ( opt_DoEtaReduction, SimplifierSwitch(..) )
+import CmdLineOpts ( switchIsOn, SimplifierSwitch(..),
+ opt_SimplDoLambdaEtaExpansion, opt_SimplCaseMerge,
+ opt_UF_UpdateInPlace
+ )
import CoreSyn
-import CoreUnfold ( mkFormSummary, exprIsTrivial, FormSummary(..) )
-import MkId ( mkSysLocal )
-import Id ( idType, isBottomingId, getIdArity,
- addInlinePragma, addIdDemandInfo,
- idWantsToBeINLINEd, dataConArgTys, Id,
- lookupIdEnv, delOneFromIdEnv
+import CoreUtils ( exprIsTrivial, cheapEqExpr, exprType, exprIsCheap,
+ etaExpand, exprEtaExpandArity, bindNonRec, mkCoerce,
+ findDefault
)
-import IdInfo ( ArityInfo(..), DemandInfo )
-import Maybes ( maybeToBool )
-import PrelVals ( augmentId, buildId )
-import PrimOp ( primOpIsCheap )
-import SimplEnv
-import SimplMonad
-import Type ( tyVarsOfType, tyVarsOfTypes, mkForAllTys, mkTyVarTys, getTyVar_maybe,
- splitAlgTyConApp_maybe, instantiateTy, Type
+import Subst ( InScopeSet, mkSubst, substExpr )
+import qualified Subst ( simplBndrs, simplBndr, simplLetId )
+import Id ( idType, idName,
+ idUnfolding, idNewStrictness,
+ mkLocalId, idInfo
)
-import TyCon ( isDataTyCon )
-import TyVar ( mkTyVarSet, intersectTyVarSets, elementOfTyVarSet, tyVarSetToList,
- delFromTyVarEnv
+import IdInfo ( StrictnessInfo(..) )
+import Maybes ( maybeToBool, catMaybes )
+import Name ( setNameUnique )
+import NewDemand ( isStrictDmd, isBotRes, splitStrictSig )
+import SimplMonad
+import Type ( Type, mkForAllTys, seqType,
+ splitTyConApp_maybe, tyConAppArgs, mkTyVarTys,
+ isUnLiftedType,
+ splitRepFunTys
)
-import SrcLoc ( noSrcLoc )
-import Util ( isIn, zipWithEqual, panic, assertPanic )
-
+import TcType ( isStrictType )
+import TyCon ( tyConDataConsIfAvailable )
+import DataCon ( dataConRepArity )
+import VarEnv ( SubstEnv )
+import Util ( lengthExceeds, mapAccumL )
+import Outputable
\end{code}
%************************************************************************
%* *
-\subsection{New ids}
+\subsection{The continuation data type}
%* *
%************************************************************************
\begin{code}
-newId :: Type -> SmplM Id
-newId ty
- = getUniqueSmpl `thenSmpl` \ uniq ->
- returnSmpl (mkSysLocal SLIT("s") uniq ty noSrcLoc)
-
-newIds :: [Type] -> SmplM [Id]
-newIds tys
- = getUniquesSmpl (length tys) `thenSmpl` \ uniqs ->
- returnSmpl (zipWithEqual "newIds" mk_id tys uniqs)
+data SimplCont -- Strict contexts
+ = Stop OutType -- Type of the result
+ Bool -- True => This is the RHS of a thunk whose type suggests
+ -- that update-in-place would be possible
+ -- (This makes the inliner a little keener.)
+
+ | CoerceIt OutType -- The To-type, simplified
+ SimplCont
+
+ | InlinePlease -- This continuation makes a function very
+ SimplCont -- keen to inline itelf
+
+ | ApplyTo DupFlag
+ InExpr SubstEnv -- The argument, as yet unsimplified,
+ SimplCont -- and its subst-env
+
+ | Select DupFlag
+ InId [InAlt] SubstEnv -- The case binder, alts, and subst-env
+ SimplCont
+
+ | ArgOf DupFlag -- An arbitrary strict context: the argument
+ -- of a strict function, or a primitive-arg fn
+ -- or a PrimOp
+ OutType -- cont_ty: the type of the expression being sought by the context
+ -- f (error "foo") ==> coerce t (error "foo")
+ -- when f is strict
+ -- We need to know the type t, to which to coerce.
+ (OutExpr -> SimplM OutExprStuff) -- What to do with the result
+ -- The result expression in the OutExprStuff has type cont_ty
+
+instance Outputable SimplCont where
+ ppr (Stop _ _) = ptext SLIT("Stop")
+ ppr (ApplyTo dup arg se cont) = (ptext SLIT("ApplyTo") <+> ppr dup <+> ppr arg) $$ ppr cont
+ ppr (ArgOf dup _ _) = ptext SLIT("ArgOf...") <+> ppr dup
+ ppr (Select dup bndr alts se cont) = (ptext SLIT("Select") <+> ppr dup <+> ppr bndr) $$
+ (nest 4 (ppr alts)) $$ ppr cont
+ ppr (CoerceIt ty cont) = (ptext SLIT("CoerceIt") <+> ppr ty) $$ ppr cont
+ ppr (InlinePlease cont) = ptext SLIT("InlinePlease") $$ ppr cont
+
+data DupFlag = OkToDup | NoDup
+
+instance Outputable DupFlag where
+ ppr OkToDup = ptext SLIT("ok")
+ ppr NoDup = ptext SLIT("nodup")
+
+
+-------------------
+mkRhsStop, mkStop :: OutType -> SimplCont
+mkStop ty = Stop ty False
+mkRhsStop ty = Stop ty (canUpdateInPlace ty)
+
+
+-------------------
+contIsDupable :: SimplCont -> Bool
+contIsDupable (Stop _ _) = True
+contIsDupable (ApplyTo OkToDup _ _ _) = True
+contIsDupable (ArgOf OkToDup _ _) = True
+contIsDupable (Select OkToDup _ _ _ _) = True
+contIsDupable (CoerceIt _ cont) = contIsDupable cont
+contIsDupable (InlinePlease cont) = contIsDupable cont
+contIsDupable other = False
+
+-------------------
+discardInline :: SimplCont -> SimplCont
+discardInline (InlinePlease cont) = cont
+discardInline (ApplyTo d e s cont) = ApplyTo d e s (discardInline cont)
+discardInline cont = cont
+
+-------------------
+discardableCont :: SimplCont -> Bool
+discardableCont (Stop _ _) = False
+discardableCont (CoerceIt _ cont) = discardableCont cont
+discardableCont (InlinePlease cont) = discardableCont cont
+discardableCont other = True
+
+discardCont :: SimplCont -- A continuation, expecting
+ -> SimplCont -- Replace the continuation with a suitable coerce
+discardCont cont = case cont of
+ Stop to_ty _ -> cont
+ other -> CoerceIt to_ty (mkStop to_ty)
+ where
+ to_ty = contResultType cont
+
+-------------------
+contResultType :: SimplCont -> OutType
+contResultType (Stop to_ty _) = to_ty
+contResultType (ArgOf _ to_ty _) = to_ty
+contResultType (ApplyTo _ _ _ cont) = contResultType cont
+contResultType (CoerceIt _ cont) = contResultType cont
+contResultType (InlinePlease cont) = contResultType cont
+contResultType (Select _ _ _ _ cont) = contResultType cont
+
+-------------------
+countValArgs :: SimplCont -> Int
+countValArgs (ApplyTo _ (Type ty) se cont) = countValArgs cont
+countValArgs (ApplyTo _ val_arg se cont) = 1 + countValArgs cont
+countValArgs other = 0
+
+countArgs :: SimplCont -> Int
+countArgs (ApplyTo _ arg se cont) = 1 + countArgs cont
+countArgs other = 0
+\end{code}
+
+
+\begin{code}
+getContArgs :: OutId -> SimplCont
+ -> SimplM ([(InExpr, SubstEnv, Bool)], -- Arguments; the Bool is true for strict args
+ SimplCont, -- Remaining continuation
+ Bool) -- Whether we came across an InlineCall
+-- getContArgs id k = (args, k', inl)
+-- args are the leading ApplyTo items in k
+-- (i.e. outermost comes first)
+-- augmented with demand info from the functionn
+getContArgs fun orig_cont
+ = getSwitchChecker `thenSmpl` \ chkr ->
+ let
+ -- Ignore strictness info if the no-case-of-case
+ -- flag is on. Strictness changes evaluation order
+ -- and that can change full laziness
+ stricts | switchIsOn chkr NoCaseOfCase = vanilla_stricts
+ | otherwise = computed_stricts
+ in
+ go [] stricts False orig_cont
+ where
+ ----------------------------
+
+ -- Type argument
+ go acc ss inl (ApplyTo _ arg@(Type _) se cont)
+ = go ((arg,se,False) : acc) ss inl cont
+ -- NB: don't bother to instantiate the function type
+
+ -- Value argument
+ go acc (s:ss) inl (ApplyTo _ arg se cont)
+ = go ((arg,se,s) : acc) ss inl cont
+
+ -- An Inline continuation
+ go acc ss inl (InlinePlease cont)
+ = go acc ss True cont
+
+ -- We're run out of arguments, or else we've run out of demands
+ -- The latter only happens if the result is guaranteed bottom
+ -- This is the case for
+ -- * case (error "hello") of { ... }
+ -- * (error "Hello") arg
+ -- * f (error "Hello") where f is strict
+ -- etc
+ go acc ss inl cont
+ | null ss && discardableCont cont = tick BottomFound `thenSmpl_`
+ returnSmpl (reverse acc, discardCont cont, inl)
+ | otherwise = returnSmpl (reverse acc, cont, inl)
+
+ ----------------------------
+ vanilla_stricts, computed_stricts :: [Bool]
+ vanilla_stricts = repeat False
+ computed_stricts = zipWith (||) fun_stricts arg_stricts
+
+ ----------------------------
+ (val_arg_tys, _) = splitRepFunTys (idType fun)
+ arg_stricts = map isStrictType val_arg_tys ++ repeat False
+ -- These argument types are used as a cheap and cheerful way to find
+ -- unboxed arguments, which must be strict. But it's an InType
+ -- and so there might be a type variable where we expect a function
+ -- type (the substitution hasn't happened yet). And we don't bother
+ -- doing the type applications for a polymorphic function.
+ -- Hence the split*Rep*FunTys
+
+ ----------------------------
+ -- If fun_stricts is finite, it means the function returns bottom
+ -- after that number of value args have been consumed
+ -- Otherwise it's infinite, extended with False
+ fun_stricts
+ = case splitStrictSig (idNewStrictness fun) of
+ (demands, result_info)
+ | not (demands `lengthExceeds` countValArgs orig_cont)
+ -> -- Enough args, use the strictness given.
+ -- For bottoming functions we used to pretend that the arg
+ -- is lazy, so that we don't treat the arg as an
+ -- interesting context. This avoids substituting
+ -- top-level bindings for (say) strings into
+ -- calls to error. But now we are more careful about
+ -- inlining lone variables, so its ok (see SimplUtils.analyseCont)
+ if isBotRes result_info then
+ map isStrictDmd demands -- Finite => result is bottom
+ else
+ map isStrictDmd demands ++ vanilla_stricts
+
+ other -> vanilla_stricts -- Not enough args, or no strictness
+
+-------------------
+interestingArg :: InScopeSet -> InExpr -> SubstEnv -> Bool
+ -- An argument is interesting if it has *some* structure
+ -- We are here trying to avoid unfolding a function that
+ -- is applied only to variables that have no unfolding
+ -- (i.e. they are probably lambda bound): f x y z
+ -- There is little point in inlining f here.
+interestingArg in_scope arg subst
+ = analyse (substExpr (mkSubst in_scope subst) arg)
+ -- 'analyse' only looks at the top part of the result
+ -- and substExpr is lazy, so this isn't nearly as brutal
+ -- as it looks.
+ where
+ analyse (Var v) = hasSomeUnfolding (idUnfolding v)
+ -- Was: isValueUnfolding (idUnfolding v')
+ -- But that seems over-pessimistic
+ analyse (Type _) = False
+ analyse (App fn (Type _)) = analyse fn
+ analyse (Note _ a) = analyse a
+ analyse other = True
+ -- Consider let x = 3 in f x
+ -- The substitution will contain (x -> ContEx 3), and we want to
+ -- to say that x is an interesting argument.
+ -- But consider also (\x. f x y) y
+ -- The substitution will contain (x -> ContEx y), and we want to say
+ -- that x is not interesting (assuming y has no unfolding)
+\end{code}
+
+Comment about interestingCallContext
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+We want to avoid inlining an expression where there can't possibly be
+any gain, such as in an argument position. Hence, if the continuation
+is interesting (eg. a case scrutinee, application etc.) then we
+inline, otherwise we don't.
+
+Previously some_benefit used to return True only if the variable was
+applied to some value arguments. This didn't work:
+
+ let x = _coerce_ (T Int) Int (I# 3) in
+ case _coerce_ Int (T Int) x of
+ I# y -> ....
+
+we want to inline x, but can't see that it's a constructor in a case
+scrutinee position, and some_benefit is False.
+
+Another example:
+
+dMonadST = _/\_ t -> :Monad (g1 _@_ t, g2 _@_ t, g3 _@_ t)
+
+.... case dMonadST _@_ x0 of (a,b,c) -> ....
+
+we'd really like to inline dMonadST here, but we *don't* want to
+inline if the case expression is just
+
+ case x of y { DEFAULT -> ... }
+
+since we can just eliminate this case instead (x is in WHNF). Similar
+applies when x is bound to a lambda expression. Hence
+contIsInteresting looks for case expressions with just a single
+default case.
+
+\begin{code}
+interestingCallContext :: Bool -- False <=> no args at all
+ -> Bool -- False <=> no value args
+ -> SimplCont -> Bool
+ -- The "lone-variable" case is important. I spent ages
+ -- messing about with unsatisfactory varaints, but this is nice.
+ -- The idea is that if a variable appear all alone
+ -- as an arg of lazy fn, or rhs Stop
+ -- as scrutinee of a case Select
+ -- as arg of a strict fn ArgOf
+ -- then we should not inline it (unless there is some other reason,
+ -- e.g. is is the sole occurrence). We achieve this by making
+ -- interestingCallContext return False for a lone variable.
+ --
+ -- Why? At least in the case-scrutinee situation, turning
+ -- let x = (a,b) in case x of y -> ...
+ -- into
+ -- let x = (a,b) in case (a,b) of y -> ...
+ -- and thence to
+ -- let x = (a,b) in let y = (a,b) in ...
+ -- is bad if the binding for x will remain.
+ --
+ -- Another example: I discovered that strings
+ -- were getting inlined straight back into applications of 'error'
+ -- because the latter is strict.
+ -- s = "foo"
+ -- f = \x -> ...(error s)...
+
+ -- Fundamentally such contexts should not ecourage inlining becuase
+ -- the context can ``see'' the unfolding of the variable (e.g. case or a RULE)
+ -- so there's no gain.
+ --
+ -- However, even a type application or coercion isn't a lone variable.
+ -- Consider
+ -- case $fMonadST @ RealWorld of { :DMonad a b c -> c }
+ -- We had better inline that sucker! The case won't see through it.
+ --
+ -- For now, I'm treating treating a variable applied to types
+ -- in a *lazy* context "lone". The motivating example was
+ -- f = /\a. \x. BIG
+ -- g = /\a. \y. h (f a)
+ -- There's no advantage in inlining f here, and perhaps
+ -- a significant disadvantage. Hence some_val_args in the Stop case
+
+interestingCallContext some_args some_val_args cont
+ = interesting cont
where
- mk_id ty uniq = mkSysLocal SLIT("s") uniq ty noSrcLoc
+ interesting (InlinePlease _) = True
+ interesting (Select _ _ _ _ _) = some_args
+ interesting (ApplyTo _ _ _ _) = True -- Can happen if we have (coerce t (f x)) y
+ -- Perhaps True is a bit over-keen, but I've
+ -- seen (coerce f) x, where f has an INLINE prag,
+ -- So we have to give some motivaiton for inlining it
+ interesting (ArgOf _ _ _) = some_val_args
+ interesting (Stop ty upd_in_place) = some_val_args && upd_in_place
+ interesting (CoerceIt _ cont) = interesting cont
+ -- If this call is the arg of a strict function, the context
+ -- is a bit interesting. If we inline here, we may get useful
+ -- evaluation information to avoid repeated evals: e.g.
+ -- x + (y * z)
+ -- Here the contIsInteresting makes the '*' keener to inline,
+ -- which in turn exposes a constructor which makes the '+' inline.
+ -- Assuming that +,* aren't small enough to inline regardless.
+ --
+ -- It's also very important to inline in a strict context for things
+ -- like
+ -- foldr k z (f x)
+ -- Here, the context of (f x) is strict, and if f's unfolding is
+ -- a build it's *great* to inline it here. So we must ensure that
+ -- the context for (f x) is not totally uninteresting.
+
+
+-------------------
+canUpdateInPlace :: Type -> Bool
+-- Consider let x = <wurble> in ...
+-- If <wurble> returns an explicit constructor, we might be able
+-- to do update in place. So we treat even a thunk RHS context
+-- as interesting if update in place is possible. We approximate
+-- this by seeing if the type has a single constructor with a
+-- small arity. But arity zero isn't good -- we share the single copy
+-- for that case, so no point in sharing.
+
+canUpdateInPlace ty
+ | not opt_UF_UpdateInPlace = False
+ | otherwise
+ = case splitTyConApp_maybe ty of
+ Nothing -> False
+ Just (tycon, _) -> case tyConDataConsIfAvailable tycon of
+ [dc] -> arity == 1 || arity == 2
+ where
+ arity = dataConRepArity dc
+ other -> False
\end{code}
+
%************************************************************************
%* *
-\subsection{Floating}
+\section{Dealing with a single binder}
%* *
%************************************************************************
-The function @floatExposesHNF@ tells whether let/case floating will
-expose a head normal form. It is passed booleans indicating the
-desired strategy.
-
\begin{code}
-floatExposesHNF
- :: Bool -- Float let(rec)s out of rhs
- -> Bool -- Float cheap primops out of rhs
- -> GenCoreExpr bdr Id flexi
- -> Bool
-
-floatExposesHNF float_lets float_primops rhs
- = try rhs
- where
- try (Case (Prim _ _) (PrimAlts alts deflt) )
- | float_primops && null alts
- = or (try_deflt deflt : map try_alt alts)
-
- try (Let bind body) | float_lets = try body
-
- -- `build g'
- -- is like a HNF,
- -- because it *will* become one.
- -- likewise for `augment g h'
- --
- try (App (App (Var bld) _) _) | bld == buildId = True
- try (App (App (App (Var aug) _) _) _) | aug == augmentId = True
-
- try other = case mkFormSummary other of
- VarForm -> True
- ValueForm -> True
- other -> False
- {- but *not* necessarily "BottomForm"...
-
- We may want to float a let out of a let to expose WHNFs,
- but to do that to expose a "bottom" is a Bad Idea:
- let x = let y = ...
- in ...error ...y... -- manifestly bottom using y
- in ...
- =/=>
- let y = ...
- in let x = ...error ...y...
- in ...
-
- as y is only used in case of an error, we do not want
- to allocate it eagerly as that's a waste.
- -}
-
- try_alt (lit,rhs) = try rhs
-
- try_deflt NoDefault = False
- try_deflt (BindDefault _ rhs) = try rhs
+simplBinders :: [InBinder] -> ([OutBinder] -> SimplM a) -> SimplM a
+simplBinders bndrs thing_inside
+ = getSubst `thenSmpl` \ subst ->
+ let
+ (subst', bndrs') = Subst.simplBndrs subst bndrs
+ in
+ seqBndrs bndrs' `seq`
+ setSubst subst' (thing_inside bndrs')
+
+simplBinder :: InBinder -> (OutBinder -> SimplM a) -> SimplM a
+simplBinder bndr thing_inside
+ = getSubst `thenSmpl` \ subst ->
+ let
+ (subst', bndr') = Subst.simplBndr subst bndr
+ in
+ seqBndr bndr' `seq`
+ setSubst subst' (thing_inside bndr')
+
+
+simplRecIds :: [InBinder] -> ([OutBinder] -> SimplM a) -> SimplM a
+simplRecIds ids thing_inside
+ = getSubst `thenSmpl` \ subst ->
+ let
+ (subst', ids') = mapAccumL Subst.simplLetId subst ids
+ in
+ seqBndrs ids' `seq`
+ setSubst subst' (thing_inside ids')
+
+simplLetId :: InBinder -> (OutBinder -> SimplM a) -> SimplM a
+simplLetId id thing_inside
+ = getSubst `thenSmpl` \ subst ->
+ let
+ (subst', id') = Subst.simplLetId subst id
+ in
+ seqBndr id' `seq`
+ setSubst subst' (thing_inside id')
+
+seqBndrs [] = ()
+seqBndrs (b:bs) = seqBndr b `seq` seqBndrs bs
+
+seqBndr b | isTyVar b = b `seq` ()
+ | otherwise = seqType (idType b) `seq`
+ idInfo b `seq`
+ ()
\end{code}
-Local tyvar-lifting
-~~~~~~~~~~~~~~~~~~~
+%************************************************************************
+%* *
+\subsection{Local tyvar-lifting}
+%* *
+%************************************************************************
+
mkRhsTyLam tries this transformation, when the big lambda appears as
the RHS of a let(rec) binding:
into
letrec g' = /\a -> ... g' a ...
in
- let f = /\ a -> f a
+ let f = /\ a -> g' a
which is better. In effect, it means that big lambdas don't impede
let-floating.
This optimisation is CRUCIAL in eliminating the junk introduced by
desugaring mutually recursive definitions. Don't eliminate it lightly!
-So far as the implemtation is concerned:
+So far as the implementation is concerned:
Invariant: go F e = /\tvs -> F e
where
G = F . Let {xi = xi' tvs}
+[May 1999] If we do this transformation *regardless* then we can
+end up with some pretty silly stuff. For example,
+
+ let
+ st = /\ s -> let { x1=r1 ; x2=r2 } in ...
+ in ..
+becomes
+ let y1 = /\s -> r1
+ y2 = /\s -> r2
+ st = /\s -> ...[y1 s/x1, y2 s/x2]
+ in ..
+
+Unless the "..." is a WHNF there is really no point in doing this.
+Indeed it can make things worse. Suppose x1 is used strictly,
+and is of the form
+
+ x1* = case f y of { (a,b) -> e }
+
+If we abstract this wrt the tyvar we then can't do the case inline
+as we would normally do.
+
+
\begin{code}
-mkRhsTyLam [] body = returnSmpl body
+tryRhsTyLam :: OutExpr -> SimplM ([OutBind], OutExpr)
+
+tryRhsTyLam rhs -- Only does something if there's a let
+ | null tyvars || not (worth_it body) -- inside a type lambda,
+ = returnSmpl ([], rhs) -- and a WHNF inside that
+
+ | otherwise
+ = go (\x -> x) body `thenSmpl` \ (binds, body') ->
+ returnSmpl (binds, mkLams tyvars body')
-mkRhsTyLam tyvars body
- = go (\x -> x) body
where
- main_tyvar_set = mkTyVarSet tyvars
+ (tyvars, body) = collectTyBinders rhs
- go fn (Let bind@(NonRec var rhs) body) | exprIsTrivial rhs
- = go (fn . Let bind) body
+ worth_it e@(Let _ _) = whnf_in_middle e
+ worth_it e = False
+
+ whnf_in_middle (Let (NonRec x rhs) e) | isUnLiftedType (idType x) = False
+ whnf_in_middle (Let _ e) = whnf_in_middle e
+ whnf_in_middle e = exprIsCheap e
go fn (Let bind@(NonRec var rhs) body)
- = mk_poly tyvars_here var_ty `thenSmpl` \ (var', rhs') ->
- go (fn . Let (mk_silly_bind var rhs')) body `thenSmpl` \ body' ->
- returnSmpl (Let (NonRec var' (mkTyLam tyvars_here (fn rhs))) body')
+ | exprIsTrivial rhs
+ = go (fn . Let bind) body
+
+ go fn (Let (NonRec var rhs) body)
+ = mk_poly tyvars_here var `thenSmpl` \ (var', rhs') ->
+ go (fn . Let (mk_silly_bind var rhs')) body `thenSmpl` \ (binds, body') ->
+ returnSmpl (NonRec var' (mkLams tyvars_here (fn rhs)) : binds, body')
+
where
- var_ty = idType var
tyvars_here = tyvars
- -- tyVarSetToList (main_tyvar_set `intersectTyVarSets` tyVarsOfTypes var_ty)
+ -- main_tyvar_set = mkVarSet tyvars
+ -- var_ty = idType var
+ -- varSetElems (main_tyvar_set `intersectVarSet` tyVarsOfType var_ty)
-- tyvars_here was an attempt to reduce the number of tyvars
-- wrt which the new binding is abstracted. But the naive
-- approach of abstract wrt the tyvars free in the Id's type
-- /\ a b -> let t :: (a,b) = (e1, e2)
-- x :: a = fst t
-- in ...
- -- Here, b isn't free in a's type, but we must nevertheless
+ -- Here, b isn't free in x's type, but we must nevertheless
-- abstract wrt b as well, because t's type mentions b.
-- Since t is floated too, we'd end up with the bogus:
-- poly_t = /\ a b -> (e1, e2)
-- gives rise to problems. SLPJ June 98
go fn (Let (Rec prs) body)
- = mapAndUnzipSmpl (mk_poly tyvars_here) var_tys `thenSmpl` \ (vars', rhss') ->
+ = mapAndUnzipSmpl (mk_poly tyvars_here) vars `thenSmpl` \ (vars', rhss') ->
let
- gn body = fn $ foldr Let body (zipWith mk_silly_bind vars rhss')
+ gn body = fn (foldr Let body (zipWith mk_silly_bind vars rhss'))
+ new_bind = Rec (vars' `zip` [mkLams tyvars_here (gn rhs) | rhs <- rhss])
in
- go gn body `thenSmpl` \ body' ->
- returnSmpl (Let (Rec (vars' `zip` [mkTyLam tyvars_here (gn rhs) | rhs <- rhss])) body')
+ go gn body `thenSmpl` \ (binds, body') ->
+ returnSmpl (new_bind : binds, body')
where
(vars,rhss) = unzip prs
- var_tys = map idType vars
- tyvars_here = tyvars -- See notes on tyvars_here above
-
- go fn body = returnSmpl (mkTyLam tyvars (fn body))
-
- mk_poly tyvars_here var_ty
- = newId (mkForAllTys tyvars_here var_ty) `thenSmpl` \ poly_id ->
- returnSmpl (poly_id, mkTyApp (Var poly_id) (mkTyVarTys tyvars_here))
+ tyvars_here = tyvars
+ -- varSetElems (main_tyvar_set `intersectVarSet` tyVarsOfTypes var_tys)
+ -- var_tys = map idType vars
+ -- See notes with tyvars_here above
+
+ go fn body = returnSmpl ([], fn body)
+
+ mk_poly tyvars_here var
+ = getUniqueSmpl `thenSmpl` \ uniq ->
+ let
+ poly_name = setNameUnique (idName var) uniq -- Keep same name
+ poly_ty = mkForAllTys tyvars_here (idType var) -- But new type of course
+ poly_id = mkLocalId poly_name poly_ty
+
+ -- In the olden days, it was crucial to copy the occInfo of the original var,
+ -- because we were looking at occurrence-analysed but as yet unsimplified code!
+ -- In particular, we mustn't lose the loop breakers. BUT NOW we are looking
+ -- at already simplified code, so it doesn't matter
+ --
+ -- It's even right to retain single-occurrence or dead-var info:
+ -- Suppose we started with /\a -> let x = E in B
+ -- where x occurs once in B. Then we transform to:
+ -- let x' = /\a -> E in /\a -> let x* = x' a in B
+ -- where x* has an INLINE prag on it. Now, once x* is inlined,
+ -- the occurrences of x' will be just the occurrences originally
+ -- pinned on x.
+ in
+ returnSmpl (poly_id, mkTyApps (Var poly_id) (mkTyVarTys tyvars_here))
- mk_silly_bind var rhs = NonRec (addInlinePragma var) rhs
- -- The addInlinePragma is really important! If we don't say
- -- INLINE on these silly little bindings then look what happens!
+ mk_silly_bind var rhs = NonRec var (Note InlineMe rhs)
-- Suppose we start with:
--
- -- x = let g = /\a -> \x -> f x x
- -- in
- -- /\ b -> let g* = g b in E
+ -- x = /\ a -> let g = G in E
+ --
+ -- Then we'll float to get
+ --
+ -- x = let poly_g = /\ a -> G
+ -- in /\ a -> let g = poly_g a in E
--
- -- Then: * the binding for g gets floated out
- -- * but then it gets inlined into the rhs of g*
- -- * then the binding for g* is floated out of the /\b
- -- * so we're back to square one
- -- The silly binding for g* must be INLINE, so that no inlining
- -- will happen in its RHS.
+ -- But now the occurrence analyser will see just one occurrence
+ -- of poly_g, not inside a lambda, so the simplifier will
+ -- PreInlineUnconditionally poly_g back into g! Badk to square 1!
+ -- (I used to think that the "don't inline lone occurrences" stuff
+ -- would stop this happening, but since it's the *only* occurrence,
+ -- PreInlineUnconditionally kicks in first!)
+ --
+ -- Solution: put an INLINE note on g's RHS, so that poly_g seems
+ -- to appear many times. (NB: mkInlineMe eliminates
+ -- such notes on trivial RHSs, so do it manually.)
\end{code}
-Eta reduction
-~~~~~~~~~~~~~
-@etaCoreExpr@ trys an eta reduction at the top level of a Core Expr.
-
-e.g. \ x y -> f x y ===> f
-
-It is used
- a) Before constructing an Unfolding, to
- try to make the unfolding smaller;
- b) In tidyCoreExpr, which is done just before converting to STG.
-
-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.
-But if they expose a simple variable then we definitely win. Even
-if they expose a type application we win. So we check for this special
-case.
-
-It does arise:
-
- f xs = [y | (y,_) <- xs]
-
-gives rise to a recursive function for the list comprehension, and
-f turns out to be just a single call to this recursive function.
-
-Doing eta on type lambdas is useful too:
-
- /\a -> <expr> a ===> <expr>
-where <expr> doesn't mention a.
-This is sometimes quite useful, because we can get the sequence:
+%************************************************************************
+%* *
+\subsection{Eta expansion}
+%* *
+%************************************************************************
- f ab d = let d1 = ...d... in
- letrec f' b x = ...d...(f' b)... in
- f' b
-specialise ==>
+ Try eta expansion for RHSs
- f.Int b = letrec f' b x = ...dInt...(f' b)... in
- f' b
+We go for:
+ Case 1 f = \x1..xn -> N ==> f = \x1..xn y1..ym -> N y1..ym
+ (n >= 0)
+ OR
+ Case 2 f = N E1..En ==> z1=E1
+ (n > 0) ..
+ zn=En
+ f = \y1..ym -> N z1..zn y1..ym
-float ==>
+where (in both cases)
- f' b x = ...dInt...(f' b)...
- f.Int b = f' b
+ * The xi can include type variables
-Now we really want to simplify to
+ * The yi are all value variables
- f.Int = f'
+ * N is a NORMAL FORM (i.e. no redexes anywhere)
+ wanting a suitable number of extra args.
-and then replace all the f's with f.Ints.
+ * the Ei must not have unlifted type
-N.B. We are careful not to partially eta-reduce a sequence of type
-applications since this breaks the specialiser:
+There is no point in looking for a combination of the two, because
+that would leave use with some lets sandwiched between lambdas; that's
+what the final test in the first equation is for.
- /\ a -> f Char# a =NO=> f Char#
+In Case 1, we may have to sandwich some coerces between the lambdas
+to make the types work. exprEtaExpandArity looks through coerces
+when computing arity; and etaExpand adds the coerces as necessary when
+actually computing the expansion.
\begin{code}
-etaCoreExpr :: CoreExpr -> CoreExpr
-
-
-etaCoreExpr expr@(Lam bndr body)
- | opt_DoEtaReduction
- = case etaCoreExpr body of
- App fun arg | eta_match bndr arg &&
- residual_ok fun
- -> fun -- Eta
- other -> expr -- Can't eliminate it, so do nothing at all
+tryEtaExpansion :: OutExpr -> OutType -> SimplM ([OutBind], OutExpr)
+tryEtaExpansion rhs rhs_ty
+ | not opt_SimplDoLambdaEtaExpansion -- Not if switched off
+ || exprIsTrivial rhs -- Not if RHS is trivial
+ || final_arity == 0 -- Not if arity is zero
+ = returnSmpl ([], rhs)
+
+ | n_val_args == 0 && not arity_is_manifest
+ = -- Some lambdas but not enough: case 1
+ getUniqSupplySmpl `thenSmpl` \ us ->
+ returnSmpl ([], etaExpand final_arity us rhs rhs_ty)
+
+ | n_val_args > 0 && not (any cant_bind arg_infos)
+ = -- Partial application: case 2
+ mapAndUnzipSmpl bind_z_arg arg_infos `thenSmpl` \ (maybe_z_binds, z_args) ->
+ getUniqSupplySmpl `thenSmpl` \ us ->
+ returnSmpl (catMaybes maybe_z_binds,
+ etaExpand final_arity us (mkApps fun z_args) rhs_ty)
+
+ | otherwise
+ = returnSmpl ([], rhs)
where
- eta_match (ValBinder v) (VarArg v') = v == v'
- eta_match (TyBinder tv) (TyArg ty) = case getTyVar_maybe ty of
- Nothing -> False
- Just tv' -> tv == tv'
- eta_match bndr arg = False
-
- residual_ok :: CoreExpr -> Bool -- Checks for type application
- -- and function not one of the
- -- bound vars
-
- (VarArg v) `mentions` (ValBinder v') = v == v'
- (TyArg ty) `mentions` (TyBinder tv) = tv `elementOfTyVarSet` tyVarsOfType ty
- bndr `mentions` arg = False
-
- residual_ok (Var v)
- = not (VarArg v `mentions` bndr)
- residual_ok (App fun arg)
- | arg `mentions` bndr = False
- | otherwise = residual_ok fun
- residual_ok (Note (Coerce to_ty from_ty) body)
- | TyArg to_ty `mentions` bndr
- || TyArg from_ty `mentions` bndr = False
- | otherwise = residual_ok body
-
- residual_ok other = False -- Safe answer
- -- This last clause may seem conservative, but consider:
- -- primops, constructors, and literals, are impossible here
- -- let and case are unlikely (the argument would have been floated inside)
- -- SCCs we probably want to be conservative about (not sure, but it's safe to be)
-
-etaCoreExpr expr = expr -- The common case
+ (fun, args) = collectArgs rhs
+ n_val_args = valArgCount args
+ (fun_arity, arity_is_manifest) = exprEtaExpandArity fun
+ final_arity = 0 `max` (fun_arity - n_val_args)
+ arg_infos = [(arg, exprType arg, exprIsTrivial arg) | arg <- args]
+ cant_bind (_, ty, triv) = not triv && isUnLiftedType ty
+
+ bind_z_arg (arg, arg_ty, trivial_arg)
+ | trivial_arg = returnSmpl (Nothing, arg)
+ | otherwise = newId SLIT("z") arg_ty $ \ z ->
+ returnSmpl (Just (NonRec z arg), Var z)
\end{code}
-
-
-Eta expansion
-~~~~~~~~~~~~~
-@etaExpandCount@ takes an expression, E, and returns an integer n,
-such that
- E ===> (\x1::t1 x1::t2 ... xn::tn -> E x1 x2 ... xn)
-is a safe transformation. In particular, the transformation should
-not cause work to be duplicated, unless it is ``cheap'' (see
-@manifestlyCheap@ below).
-
-@etaExpandCount@ errs on the conservative side. It is always safe to
-return 0.
-
-An application of @error@ is special, because it can absorb as many
-arguments as you care to give it. For this special case we return
-100, to represent "infinity", which is a bit of a hack.
+%************************************************************************
+%* *
+\subsection{Case absorption and identity-case elimination}
+%* *
+%************************************************************************
\begin{code}
-etaExpandCount :: GenCoreExpr bdr Id flexi
- -> Int -- Number of extra args you can safely abstract
-
-etaExpandCount (Lam (ValBinder _) body)
- = 1 + etaExpandCount body
-
-etaExpandCount (Let bind body)
- | all manifestlyCheap (rhssOfBind bind)
- = etaExpandCount body
+mkCase :: OutExpr -> OutId -> [OutAlt] -> SimplM OutExpr
+\end{code}
-etaExpandCount (Case scrut alts)
- | manifestlyCheap scrut
- = minimum [etaExpandCount rhs | rhs <- rhssOfAlts alts]
+@mkCase@ tries the following transformation (if possible):
+
+case e of b { ==> case e of b {
+ p1 -> rhs1 p1 -> rhs1
+ ... ...
+ pm -> rhsm pm -> rhsm
+ _ -> case b of b' { pn -> rhsn[b/b'] {or (alg) let b=b' in rhsn}
+ {or (prim) case b of b' { _ -> rhsn}}
+ pn -> rhsn ...
+ ... po -> rhso[b/b']
+ po -> rhso _ -> rhsd[b/b'] {or let b'=b in rhsd}
+ _ -> rhsd
+}
+
+which merges two cases in one case when -- the default alternative of
+the outer case scrutises the same variable as the outer case This
+transformation is called Case Merging. It avoids that the same
+variable is scrutinised multiple times.
-etaExpandCount fun@(Var _) = eta_fun fun
-etaExpandCount (App fun arg)
- | notValArg arg = eta_fun fun
- | otherwise = case etaExpandCount fun of
- 0 -> 0
- n -> n-1 -- Knock off one
+\begin{code}
+mkCase scrut outer_bndr outer_alts
+ | opt_SimplCaseMerge
+ && maybeToBool maybe_case_in_default
+
+ = tick (CaseMerge outer_bndr) `thenSmpl_`
+ returnSmpl (Case scrut outer_bndr new_alts)
+ -- Warning: don't call mkCase recursively!
+ -- Firstly, there's no point, because inner alts have already had
+ -- mkCase applied to them, so they won't have a case in their default
+ -- Secondly, if you do, you get an infinite loop, because the bindNonRec
+ -- in munge_rhs puts a case into the DEFAULT branch!
+ where
+ new_alts = add_default maybe_inner_default
+ (outer_alts_without_deflt ++ inner_con_alts)
-etaExpandCount other = 0 -- Give up
- -- Lit, Con, Prim,
- -- non-val Lam,
- -- Scc (pessimistic; ToDo),
- -- Let with non-whnf rhs(s),
- -- Case with non-whnf scrutinee
+ maybe_case_in_default = case findDefault outer_alts of
+ (outer_alts_without_default,
+ Just (Case (Var scrut_var) inner_bndr inner_alts))
+ | outer_bndr == scrut_var
+ -> Just (outer_alts_without_default, inner_bndr, inner_alts)
+ other -> Nothing
------------------------------
-eta_fun :: GenCoreExpr bdr Id flexi -- The function
- -> Int -- How many args it can safely be applied to
+ Just (outer_alts_without_deflt, inner_bndr, inner_alts) = maybe_case_in_default
-eta_fun (App fun arg) | notValArg arg = eta_fun fun
+ -- Eliminate any inner alts which are shadowed by the outer ones
+ outer_cons = [con | (con,_,_) <- outer_alts_without_deflt]
-eta_fun expr@(Var v)
- | isBottomingId v -- Bottoming ids have "infinite arity"
- = 10000 -- Blargh. Infinite enough!
+ munged_inner_alts = [ (con, args, munge_rhs rhs)
+ | (con, args, rhs) <- inner_alts,
+ not (con `elem` outer_cons) -- Eliminate shadowed inner alts
+ ]
+ munge_rhs rhs = bindNonRec inner_bndr (Var outer_bndr) rhs
-eta_fun expr@(Var v) = idMinArity v
+ (inner_con_alts, maybe_inner_default) = findDefault munged_inner_alts
-eta_fun other = 0 -- Give up
+ add_default (Just rhs) alts = (DEFAULT,[],rhs) : alts
+ add_default Nothing alts = alts
\end{code}
-@manifestlyCheap@ looks at a Core expression and returns \tr{True} if
-it is obviously in weak head normal form, or is cheap to get to WHNF.
-By ``cheap'' we mean a computation we're willing to duplicate in order
-to bring a couple of lambdas together. The main examples of things
-which aren't WHNF but are ``cheap'' are:
-
- * case e of
- pi -> ei
-
- where e, and all the ei are cheap; and
+Now the identity-case transformation:
- * let x = e
- in b
+ case e of ===> e
+ True -> True;
+ False -> False
- where e and b are cheap; and
-
- * op x1 ... xn
-
- where op is a cheap primitive operator
+and similar friends.
\begin{code}
-manifestlyCheap :: GenCoreExpr bndr Id flexi -> Bool
-
-manifestlyCheap (Var _) = True
-manifestlyCheap (Lit _) = True
-manifestlyCheap (Con _ _) = True
-manifestlyCheap (Note _ e) = manifestlyCheap e
-manifestlyCheap (Lam x e) = if isValBinder x then True else manifestlyCheap e
-manifestlyCheap (Prim op _) = primOpIsCheap op
-
-manifestlyCheap (Let bind body)
- = manifestlyCheap body && all manifestlyCheap (rhssOfBind bind)
-
-manifestlyCheap (Case scrut alts)
- = manifestlyCheap scrut && all manifestlyCheap (rhssOfAlts alts)
-
-manifestlyCheap other_expr -- look for manifest partial application
- = case (collectArgs other_expr) of { (fun, _, vargs) ->
- case fun of
-
- Var f | isBottomingId f -> True -- Application of a function which
- -- always gives bottom; we treat this as
- -- a WHNF, because it certainly doesn't
- -- need to be shared!
-
- Var f -> let
- num_val_args = length vargs
- in
- num_val_args == 0 || -- Just a type application of
- -- a variable (f t1 t2 t3)
- -- counts as WHNF
- num_val_args < idMinArity f
-
- _ -> False
- }
-
+mkCase scrut case_bndr alts
+ | all identity_alt alts
+ = tick (CaseIdentity case_bndr) `thenSmpl_`
+ returnSmpl (re_note scrut)
+ where
+ identity_alt (con, args, rhs) = de_note rhs `cheapEqExpr` identity_rhs con args
+
+ identity_rhs (DataAlt con) args = mkConApp con (arg_tys ++ map varToCoreExpr args)
+ identity_rhs (LitAlt lit) _ = Lit lit
+ identity_rhs DEFAULT _ = Var case_bndr
+
+ arg_tys = map Type (tyConAppArgs (idType case_bndr))
+
+ -- We've seen this:
+ -- case coerce T e of x { _ -> coerce T' x }
+ -- And we definitely want to eliminate this case!
+ -- So we throw away notes from the RHS, and reconstruct
+ -- (at least an approximation) at the other end
+ de_note (Note _ e) = de_note e
+ de_note e = e
+
+ -- re_note wraps a coerce if it might be necessary
+ re_note scrut = case head alts of
+ (_,_,rhs1@(Note _ _)) -> mkCoerce (exprType rhs1) (idType case_bndr) scrut
+ other -> scrut
\end{code}
+The catch-all case. We do a final transformation that I've
+occasionally seen making a big difference:
-\begin{code}
-simplIdWantsToBeINLINEd :: Id -> SimplEnv -> Bool
+ case e of =====> case e of
+ C _ -> f x D v -> ....v....
+ D v -> ....v.... DEFAULT -> f x
+ DEFAULT -> f x
-simplIdWantsToBeINLINEd id env
- = {- We used to arrange that in the final simplification pass we'd switch
- off all INLINE pragmas, so that we'd inline workers back into the
- body of their wrapper if the wrapper hadn't itself been inlined by then.
- This occurred especially for methods in dictionaries.
+The point is that we merge common RHSs, at least for the DEFAULT case.
+[One could do something more elaborate but I've never seen it needed.]
+The case where this came up was like this (lib/std/PrelCError.lhs):
- We no longer do this:
- a) there's a good chance that the exported wrapper will get
- inlined in some importing scope, in which case we don't
- want to lose the w/w idea.
+ x | p `is` 1 -> e1
+ | p `is` 2 -> e2
+ ...etc...
- b) The occurrence analyser must agree about what has an
- INLINE pragma. Not hard, but delicate.
+where @is@ was something like
- c) if the worker gets inlined we have to tell the wrapepr
- that it's no longer a wrapper, else the interface file stuff
- asks for a worker that no longer exists.
-
- if switchIsSet env IgnoreINLINEPragma
- then False
- else
- -}
-
- idWantsToBeINLINEd id
-
-idMinArity id = case getIdArity id of
- UnknownArity -> 0
- ArityAtLeast n -> n
- ArityExactly n -> n
-
-singleConstructorType :: Type -> Bool
-singleConstructorType ty
- = case (splitAlgTyConApp_maybe ty) of
- Just (tycon, ty_args, [con]) | isDataTyCon tycon -> True
- other -> False
-
-typeOkForCase :: Type -> Bool
-typeOkForCase ty
- = case (splitAlgTyConApp_maybe ty) of
- Just (tycon, ty_args, []) -> False
- Just (tycon, ty_args, non_null_data_cons) | isDataTyCon tycon -> True
- other -> False
- -- Null data cons => type is abstract, which code gen can't
- -- currently handle. (ToDo: when return-in-heap is universal we
- -- don't need to worry about this.)
-\end{code}
-
+ p `is` n = p /= (-1) && p == n
+This gave rise to a horrible sequence of cases
-substSpecEnvRhs applies a substitution to the RHS's of a SpecEnv
-It exploits the known structure of a SpecEnv's RHS to have fewer
-equations.
+ case p of
+ (-1) -> $j p
+ 1 -> e1
+ DEFAULT -> $j p
+and similarly in cascade for all the join points!
+
\begin{code}
-substSpecEnvRhs te ve rhs
- = go te ve rhs
- where
- go te ve (App f (TyArg ty)) = App (go te ve f) (TyArg (instantiateTy te ty))
- go te ve (App f (VarArg v)) = App (go te ve f) (case lookupIdEnv ve v of
- Just (SubstVar v') -> VarArg v'
- Just (SubstLit l) -> LitArg l
- Nothing -> VarArg v)
- go te ve (Var v) = case lookupIdEnv ve v of
- Just (SubstVar v') -> Var v'
- Just (SubstLit l) -> Lit l
- Nothing -> Var v
-
- -- These equations are a bit half baked, because
- -- they don't deal properly wih capture.
- -- But I'm sure it'll never matter... sigh.
- go te ve (Lam b@(TyBinder tyvar) e) = Lam b (go te' ve e)
- where
- te' = delFromTyVarEnv te tyvar
-
- go te ve (Lam b@(ValBinder v) e) = Lam b (go te ve' e)
- where
- ve' = delOneFromIdEnv ve v
+mkCase other_scrut case_bndr other_alts
+ = returnSmpl (Case other_scrut case_bndr (mergeDefault other_alts))
+
+mergeDefault (deflt_alt@(DEFAULT,_,deflt_rhs) : con_alts)
+ = deflt_alt : [alt | alt@(con,_,rhs) <- con_alts, not (rhs `cheapEqExpr` deflt_rhs)]
+ -- NB: we can neglect the binders because we won't get equality if the
+ -- binders are mentioned in rhs (no shadowing)
+mergeDefault other_alts
+ = other_alts
\end{code}