-- The continuation type
SimplCont(..), DupFlag(..), contIsDupable, contResultType,
- pushArgs, discardCont, countValArgs, countArgs,
- analyseCont, discardInline
+ countValArgs, countArgs, mkRhsStop, mkStop,
+ getContArgs, interestingCallContext, interestingArg, isStrictType, discardInline
) where
#include "HsVersions.h"
-import BinderInfo
-import CmdLineOpts ( opt_SimplDoLambdaEtaExpansion, opt_SimplCaseMerge )
+import CmdLineOpts ( switchIsOn, SimplifierSwitch(..),
+ opt_SimplDoLambdaEtaExpansion, opt_SimplCaseMerge, opt_DictsStrict,
+ opt_UF_UpdateInPlace
+ )
import CoreSyn
-import CoreUnfold ( isValueUnfolding )
-import CoreFVs ( exprFreeVars )
-import CoreUtils ( exprIsTrivial, cheapEqExpr, exprType, exprIsCheap, exprEtaExpandArity )
-import Subst ( InScopeSet, mkSubst, substBndrs, substBndr, substIds, lookupIdSubst )
-import Id ( Id, idType, isId, idName,
- idOccInfo, idUnfolding,
- idDemandInfo, mkId, idInfo
+import CoreUtils ( exprIsTrivial, cheapEqExpr, exprType, exprIsCheap, exprEtaExpandArity, bindNonRec )
+import Subst ( InScopeSet, mkSubst, substBndrs, substBndr, substIds, substExpr )
+import Id ( idType, idName,
+ idUnfolding, idStrictness,
+ mkVanillaId, idInfo
)
-import IdInfo ( arityLowerBound, setOccInfo, vanillaIdInfo )
+import IdInfo ( StrictnessInfo(..), ArityInfo, atLeastArity, vanillaIdInfo )
import Maybes ( maybeToBool, catMaybes )
-import Name ( isLocalName, setNameUnique )
+import Name ( setNameUnique )
+import Demand ( isStrict )
import SimplMonad
-import Type ( Type, tyVarsOfType, tyVarsOfTypes, mkForAllTys, seqType,
- splitTyConApp_maybe, splitAlgTyConApp_maybe, mkTyVarTys, applyTys, splitFunTys, mkFunTys
+import Type ( Type, mkForAllTys, seqType, repType,
+ splitTyConApp_maybe, mkTyVarTys, splitFunTys,
+ isDictTy, isDataType, isUnLiftedType,
+ splitRepFunTys
)
+import TyCon ( tyConDataConsIfAvailable )
import DataCon ( dataConRepArity )
-import TysPrim ( statePrimTyCon )
-import Var ( setVarUnique )
-import VarSet
-import VarEnv ( SubstEnv, SubstResult(..) )
-import UniqSupply ( splitUniqSupply, uniqFromSupply )
-import Util ( zipWithEqual, mapAccumL )
+import VarEnv ( SubstEnv )
+import Util ( lengthExceeds )
import Outputable
\end{code}
\begin{code}
data SimplCont -- Strict contexts
- = Stop OutType -- Type of the result
+ = 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
| ArgOf DupFlag -- An arbitrary strict context: the argument
-- of a strict function, or a primitive-arg fn
-- or a PrimOp
- OutType -- The type of the expression being sought by the context
+ 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 (Stop _ _) = ptext SLIT("Stop")
ppr (ApplyTo dup arg se cont) = (ptext SLIT("ApplyTo") <+> ppr dup <+> ppr arg) $$ ppr cont
ppr (ArgOf dup _ _) = ptext SLIT("ArgOf...") <+> ppr dup
ppr (Select dup bndr alts se cont) = (ptext SLIT("Select") <+> ppr dup <+> ppr bndr) $$
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 (Stop _ _) = True
contIsDupable (ApplyTo OkToDup _ _ _) = True
contIsDupable (ArgOf OkToDup _ _) = True
contIsDupable (Select OkToDup _ _ _ _) = True
contIsDupable (InlinePlease cont) = contIsDupable cont
contIsDupable other = False
-pushArgs :: SubstEnv -> [InExpr] -> SimplCont -> SimplCont
-pushArgs se [] cont = cont
-pushArgs se (arg:args) cont = ApplyTo NoDup arg se (pushArgs se args cont)
+-------------------
+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 (Stop to_ty) = Stop to_ty
-discardCont cont = CoerceIt to_ty (Stop to_ty)
- where
- to_ty = contResultType cont
+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 (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
\end{code}
-Comment about analyseCont
-~~~~~~~~~~~~~~~~~~~~~~~~~
+\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 idStrictness fun of
+ StrictnessInfo demands result_bot
+ | 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 result_bot then
+ map isStrict demands -- Finite => result is bottom
+ else
+ map isStrict demands ++ vanilla_stricts
+
+ other -> vanilla_stricts -- Not enough args, or no strictness
+
+
+-------------------
+isStrictType :: Type -> Bool
+ -- isStrictType computes whether an argument (or let RHS) should
+ -- be computed strictly or lazily, based only on its type
+isStrictType ty
+ | isUnLiftedType ty = True
+ | opt_DictsStrict && isDictTy ty && isDataType ty = True
+ | otherwise = False
+ -- Return true only for dictionary types where the dictionary
+ -- has more than one component (else we risk poking on the component
+ -- of a newtype dictionary)
+
+-------------------
+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
default case.
\begin{code}
-analyseCont :: InScopeSet -> SimplCont
- -> ([Bool], -- Arg-info flags; one for each value argument
- Bool, -- Context of the result of the call is interesting
- Bool) -- There was an InlinePlease
-
-analyseCont in_scope cont
- = case cont of
+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 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).
- -- Why not? At least in the case-scrutinee situation, turning
- -- case x of y -> ...
+ -- 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 y = (a,b) in ...
+ -- 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
-- 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 isn't a lone variable. Consider
+ -- 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.
-
- (Stop _) -> boring_result -- Don't inline a lone variable
- (Select _ _ _ _ _) -> boring_result -- Ditto
- (ArgOf _ _ _) -> boring_result -- Ditto
- (ApplyTo _ (Type _) _ cont) -> analyse_ty_app cont
- other -> analyse_app cont
+ --
+ -- 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
- boring_result = ([], False, False)
-
- -- For now, I'm treating not treating a variable applied to types as
- -- "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.
- analyse_ty_app (Stop _) = boring_result
- analyse_ty_app (ArgOf _ _ _) = boring_result
- analyse_ty_app (Select _ _ _ _ _) = ([], True, False) -- See the $fMonadST example above
- analyse_ty_app (ApplyTo _ (Type _) _ cont) = analyse_ty_app cont
- analyse_ty_app cont = analyse_app cont
-
- analyse_app (InlinePlease cont)
- = case analyse_app cont of
- (infos, icont, inline) -> (infos, icont, True)
-
- analyse_app (ApplyTo _ arg subst cont)
- | isValArg arg = case analyse_app cont of
- (infos, icont, inline) -> (analyse_arg subst arg : infos, icont, inline)
- | otherwise = analyse_app cont
-
- analyse_app cont = ([], interesting_call_context cont, False)
-
- -- 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.
- analyse_arg :: SubstEnv -> InExpr -> Bool
- analyse_arg subst (Var v) = case lookupIdSubst (mkSubst in_scope subst) v of
- DoneId v' _ -> isValueUnfolding (idUnfolding v')
- other -> False
- analyse_arg subst (Type _) = False
- analyse_arg subst (App fn (Type _)) = analyse_arg subst fn
- analyse_arg subst (Note _ a) = analyse_arg subst a
- analyse_arg subst other = True
-
- interesting_call_context (Stop ty) = canUpdateInPlace ty
- interesting_call_context (InlinePlease _) = True
- interesting_call_context (Select _ _ _ _ _) = True
- interesting_call_context (CoerceIt _ cont) = interesting_call_context cont
- interesting_call_context (ApplyTo _ (Type _) _ cont) = interesting_call_context cont
- interesting_call_context (ApplyTo _ _ _ _) = True
- interesting_call_context (ArgOf _ _ _) = True
+ interesting (InlinePlease _) = True
+ interesting (Select _ _ _ _ _) = some_args
+ interesting (ApplyTo _ _ _ _) = some_args -- Can happen if we have (coerce t (f x)) y
+ 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.
-- the context for (f x) is not totally uninteresting.
-discardInline :: SimplCont -> SimplCont
-discardInline (InlinePlease cont) = cont
-discardInline (ApplyTo d e s cont) = ApplyTo d e s (discardInline cont)
-discardInline cont = cont
-
+-------------------
+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
-- small arity. But arity zero isn't good -- we share the single copy
-- for that case, so no point in sharing.
-canUpdateInPlace ty = case splitAlgTyConApp_maybe ty of
- Just (_, _, [dc]) -> arity == 1 || arity == 2
- where
- arity = dataConRepArity dc
+-- Note the repType: we want to look through newtypes for this purpose
+
+canUpdateInPlace ty
+ | not opt_UF_UpdateInPlace = False
+ | otherwise
+ = case splitTyConApp_maybe (repType ty) of {
+ Nothing -> False ;
+ Just (tycon, _) ->
+
+ case tyConDataConsIfAvailable tycon of
+ [dc] -> arity == 1 || arity == 2
+ where
+ arity = dataConRepArity dc
other -> False
+ }
\end{code}
(b) type-lambda swizzling
\begin{code}
-transformRhs :: InExpr -> SimplM InExpr
-transformRhs rhs
- = tryEtaExpansion body `thenSmpl` \ body' ->
- mkRhsTyLam tyvars body'
- where
- (tyvars, body) = collectTyBinders rhs
+transformRhs :: OutExpr
+ -> (ArityInfo -> OutExpr -> SimplM (OutStuff a))
+ -> SimplM (OutStuff a)
+
+transformRhs rhs thing_inside
+ = tryRhsTyLam rhs $ \ rhs1 ->
+ tryEtaExpansion rhs1 thing_inside
\end{code}
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
\begin{code}
-mkRhsTyLam tyvars body -- Only does something if there's a let
+tryRhsTyLam rhs thing_inside -- Only does something if there's a let
| null tyvars || not (worth_it body) -- inside a type lambda, and a WHNF inside that
- = returnSmpl (mkLams tyvars body)
+ = thing_inside rhs
| otherwise
- = go (\x -> x) body
+ = go (\x -> x) body $ \ body' ->
+ thing_inside (mkLams tyvars body')
+
where
+ (tyvars, body) = collectTyBinders rhs
+
worth_it (Let _ e) = whnf_in_middle e
worth_it other = False
whnf_in_middle (Let _ e) = whnf_in_middle e
whnf_in_middle e = exprIsCheap e
- main_tyvar_set = mkVarSet tyvars
- go fn (Let bind@(NonRec var rhs) body) | exprIsTrivial rhs
- = go (fn . Let bind) body
+ go fn (Let bind@(NonRec var rhs) body) thing_inside
+ | exprIsTrivial rhs
+ = go (fn . Let bind) body thing_inside
+
+ go fn (Let bind@(NonRec var rhs) body) thing_inside
+ = mk_poly tyvars_here var `thenSmpl` \ (var', rhs') ->
+ addAuxiliaryBind (NonRec var' (mkLams tyvars_here (fn rhs))) $
+ go (fn . Let (mk_silly_bind var rhs')) body thing_inside
- go fn (Let bind@(NonRec var rhs) body)
- = mk_poly tyvars_here var `thenSmpl` \ (var', rhs') ->
- go (fn . Let (mk_silly_bind var rhs')) body `thenSmpl` \ body' ->
- returnSmpl (Let (NonRec var' (mkLams tyvars_here (fn rhs))) body')
where
tyvars_here = tyvars
+ -- 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
-- abstracting wrt *all* the tyvars. We'll see if that
-- gives rise to problems. SLPJ June 98
- var_ty = idType var
-
- go fn (Let (Rec prs) body)
+ go fn (Let (Rec prs) body) thing_inside
= 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'))
in
- go gn body `thenSmpl` \ body' ->
- returnSmpl (Let (Rec (vars' `zip` [mkLams tyvars_here (gn rhs) | rhs <- rhss])) body')
+ addAuxiliaryBind (Rec (vars' `zip` [mkLams tyvars_here (gn rhs) | rhs <- rhss])) $
+ go gn body thing_inside
where
(vars,rhss) = unzip prs
tyvars_here = tyvars
-- varSetElems (main_tyvar_set `intersectVarSet` tyVarsOfTypes var_tys)
+ -- var_tys = map idType vars
-- See notes with tyvars_here above
- var_tys = map idType vars
- go fn body = returnSmpl (mkLams tyvars (fn body))
+ go fn body thing_inside = thing_inside (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 = mkVanillaId poly_name poly_ty
- -- It's crucial to copy the occInfo of the original var, because
- -- we're looking at occurrence-analysed but as yet unsimplified code!
- -- In particular, we mustn't lose the loop breakers.
+ -- 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 E. Then we transform to:
+ -- 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 originaly
+ -- the occurrences of x' will be just the occurrences originally
-- pinned on x.
- poly_info = vanillaIdInfo `setOccInfo` idOccInfo var
-
- poly_id = mkId poly_name poly_ty poly_info
in
returnSmpl (poly_id, mkTyApps (Var poly_id) (mkTyVarTys tyvars_here))
- mk_silly_bind var rhs = NonRec var rhs
- -- The Inline note 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 INLINEd, so that
- -- we simply substitute for g* throughout.
+ -- 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}
Try eta expansion for RHSs
We go for:
- \x1..xn -> N ==> \x1..xn y1..ym -> N y1..ym
- AND
- N E1..En ==> let z1=E1 .. zn=En in \y1..ym -> N z1..zn y1..ym
+ 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
+
+where (in both cases)
+
+ * The xi can include type variables
+
+ * The yi are all value variables
-where (in both cases) N is a NORMAL FORM (i.e. no redexes anywhere)
-wanting a suitable number of extra args.
+ * N is a NORMAL FORM (i.e. no redexes anywhere)
+ wanting a suitable number of extra args.
-NB: the Ei may have unlifted type, but the simplifier (which is applied
-to the result) deals OK with this.
+ * the Ei must not have unlifted type
-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.
+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.
\begin{code}
-tryEtaExpansion :: InExpr -> SimplM InExpr
-tryEtaExpansion rhs
+tryEtaExpansion :: OutExpr
+ -> (ArityInfo -> OutExpr -> SimplM (OutStuff a))
+ -> SimplM (OutStuff a)
+tryEtaExpansion rhs thing_inside
| not opt_SimplDoLambdaEtaExpansion
- || exprIsTrivial rhs -- Don't eta-expand a trival RHS
- || null y_tys -- No useful expansion
- || not (null x_bndrs || and trivial_args) -- Not (no x-binders or no z-binds)
- = returnSmpl rhs
-
- | otherwise -- Consider eta expansion
- = newIds y_tys $ ( \ y_bndrs ->
- tick (EtaExpansion (head y_bndrs)) `thenSmpl_`
- mapAndUnzipSmpl bind_z_arg (args `zip` trivial_args) `thenSmpl` (\ (maybe_z_binds, z_args) ->
- returnSmpl (mkLams x_bndrs $
- mkLets (catMaybes maybe_z_binds) $
- mkLams y_bndrs $
- mkApps (mkApps fun z_args) (map Var y_bndrs))))
+ || null y_tys -- No useful expansion
+ || not (is_case1 || is_case2) -- Neither case matches
+ = thing_inside final_arity rhs -- So, no eta expansion, but
+ -- return a good arity
+
+ | is_case1
+ = make_y_bndrs $ \ y_bndrs ->
+ thing_inside final_arity
+ (mkLams x_bndrs $ mkLams y_bndrs $
+ mkApps body (map Var y_bndrs))
+
+ | otherwise -- Must be case 2
+ = mapAndUnzipSmpl bind_z_arg arg_infos `thenSmpl` \ (maybe_z_binds, z_args) ->
+ addAuxiliaryBinds (catMaybes maybe_z_binds) $
+ make_y_bndrs $ \ y_bndrs ->
+ thing_inside final_arity
+ (mkLams y_bndrs $
+ mkApps (mkApps fun z_args) (map Var y_bndrs))
where
- (x_bndrs, body) = collectValBinders rhs
- (fun, args) = collectArgs body
- trivial_args = map exprIsTrivial args
- fun_arity = exprEtaExpandArity fun
+ all_trivial_args = all is_trivial arg_infos
+ is_case1 = all_trivial_args
+ is_case2 = null x_bndrs && not (any unlifted_non_trivial arg_infos)
- bind_z_arg (arg, trivial_arg)
+ (x_bndrs, body) = collectBinders rhs -- NB: x_bndrs can include type variables
+ x_arity = valBndrCount x_bndrs
+
+ (fun, args) = collectArgs body
+ arg_infos = [(arg, exprType arg, exprIsTrivial arg) | arg <- args]
+
+ is_trivial (_, _, triv) = triv
+ unlifted_non_trivial (_, ty, triv) = not triv && isUnLiftedType ty
+
+ fun_arity = exprEtaExpandArity fun
+
+ final_arity | all_trivial_args = atLeastArity (x_arity + extra_args_wanted)
+ | otherwise = atLeastArity x_arity
+ -- Arity can be more than the number of lambdas
+ -- because of coerces. E.g. \x -> coerce t (\y -> e)
+ -- will have arity at least 2
+ -- The worker/wrapper pass will bring the coerce out to the top
+
+ bind_z_arg (arg, arg_ty, trivial_arg)
| trivial_arg = returnSmpl (Nothing, arg)
- | otherwise = newId (exprType arg) $ \ z ->
+ | otherwise = newId SLIT("z") arg_ty $ \ z ->
returnSmpl (Just (NonRec z arg), Var z)
- -- Note: I used to try to avoid the exprType call by using
- -- the type of the binder. But this type doesn't necessarily
- -- belong to the same substitution environment as this rhs;
- -- and we are going to make extra term binders (y_bndrs) from the type
- -- which will be processed with the rhs substitution environment.
- -- This only went wrong in a mind bendingly complicated case.
- (potential_extra_arg_tys, inner_ty) = splitFunTys (exprType body)
+ make_y_bndrs thing_inside
+ = ASSERT( not (exprIsTrivial rhs) )
+ newIds SLIT("y") y_tys $ \ y_bndrs ->
+ tick (EtaExpansion (head y_bndrs)) `thenSmpl_`
+ thing_inside y_bndrs
+
+ (potential_extra_arg_tys, _) = splitFunTys (exprType body)
y_tys :: [InType]
- y_tys = take no_extras_wanted potential_extra_arg_tys
+ y_tys = take extra_args_wanted potential_extra_arg_tys
- no_extras_wanted :: Int
- no_extras_wanted = 0 `max`
+ extra_args_wanted :: Int -- Number of extra args we want
+ extra_args_wanted = 0 `max` (fun_arity - valArgCount args)
-- We used to expand the arity to the previous arity fo the
-- function; but this is pretty dangerous. Consdier
-- f = \xy -> let z = BIG in e
--
-- (bndr_arity - no_of_xs) `max`
-
- -- See if the body could obviously do with more args
- (fun_arity - valArgCount args)
-
--- This case is now deal with by exprEtaExpandArity
- -- Finally, see if it's a state transformer, and xs is non-null
- -- (so it's also a function not a thunk) in which
- -- case we eta-expand on principle! This can waste work,
- -- but usually doesn't.
- -- I originally checked for a singleton type [ty] in this case
- -- but then I found a situation in which I had
- -- \ x -> let {..} in \ s -> f (...) s
- -- AND f RETURNED A FUNCTION. That is, 's' wasn't the only
- -- potential extra arg.
--- case (x_bndrs, potential_extra_arg_tys) of
--- (_:_, ty:_) -> case splitTyConApp_maybe ty of
--- Just (tycon,_) | tycon == statePrimTyCon -> 1
--- other -> 0
--- other -> 0
\end{code}