import CmdLineOpts ( switchIsOn, opt_SimplDoEtaReduction,
opt_SimplNoPreInlining,
+ dopt, DynFlag(Opt_D_dump_inlinings),
SimplifierSwitch(..)
)
import SimplMonad
-import SimplUtils ( mkCase, transformRhs, findAlt,
- simplBinder, simplBinders, simplIds, findDefault,
- SimplCont(..), DupFlag(..),
+import SimplUtils ( mkCase, tryRhsTyLam, tryEtaExpansion,
+ simplBinder, simplBinders, simplIds,
+ SimplCont(..), DupFlag(..), mkStop, mkRhsStop,
contResultType, discardInline, countArgs, contIsDupable,
getContArgs, interestingCallContext, interestingArg, isStrictType
)
import Var ( mkSysTyVar, tyVarKind )
import VarEnv
-import Id ( Id, idType, idInfo, isDataConId,
+import Id ( Id, idType, idInfo, isDataConId, hasNoBinding,
idUnfolding, setIdUnfolding, isExportedId, isDeadBinder,
idDemandInfo, setIdInfo,
- idOccInfo, setIdOccInfo,
+ idOccInfo, setIdOccInfo,
zapLamIdInfo, setOneShotLambda,
)
-import IdInfo ( OccInfo(..), ArityInfo(..),
- setArityInfo, setUnfoldingInfo,
+import IdInfo ( OccInfo(..), isDeadOcc, isLoopBreaker,
+ setArityInfo,
+ setUnfoldingInfo, atLeastArity,
occInfo
)
-import Demand ( Demand, isStrict )
+import Demand ( isStrict )
import DataCon ( dataConNumInstArgs, dataConRepStrictness,
dataConSig, dataConArgTys
)
import CoreSyn
+import PprCore ( pprParendExpr, pprCoreExpr )
import CoreFVs ( mustHaveLocalBinding )
import CoreUnfold ( mkOtherCon, mkUnfolding, otherCons,
callSiteInline
)
-import CoreUtils ( cheapEqExpr, exprIsDupable, exprIsTrivial, exprIsConApp_maybe,
- exprType, coreAltsType, exprArity, exprIsValue, idAppIsCheap,
- exprOkForSpeculation, etaReduceExpr,
+import CoreUtils ( cheapEqExpr, exprIsDupable, exprIsTrivial,
+ exprIsConApp_maybe, mkPiType, findAlt, findDefault,
+ exprType, coreAltsType, exprIsValue,
+ exprOkForSpeculation, exprArity, exprIsCheap,
mkCoerce, mkSCC, mkInlineMe, mkAltExpr
)
import Rules ( lookupRule )
import CostCentre ( currentCCS )
import Type ( mkTyVarTys, isUnLiftedType, seqType,
- mkFunTy, splitFunTy, splitTyConApp_maybe,
+ mkFunTy, splitTyConApp_maybe, tyConAppArgs,
funResultTy
)
-import Subst ( mkSubst, substTy, substExpr,
+import Subst ( mkSubst, substTy,
isInScope, lookupIdSubst, substIdInfo
)
import TyCon ( isDataTyCon, tyConDataConsIfAvailable )
import TysPrim ( realWorldStatePrimTy )
import PrelInfo ( realWorldPrimId )
-import BasicTypes ( isLoopBreaker )
+import OrdList
import Maybes ( maybeToBool )
import Util ( zipWithEqual )
import Outputable
simplIds (bindersOfBinds binds) $ \ bndrs' ->
simpl_binds binds bndrs' `thenSmpl` \ (binds', _) ->
freeTick SimplifierDone `thenSmpl_`
- returnSmpl binds'
+ returnSmpl (fromOL binds')
where
-- We need to track the zapped top-level binders, because
-- they should have their fragile IdInfo zapped (notably occurrence info)
- simpl_binds [] bs = ASSERT( null bs ) returnSmpl ([], panic "simplTopBinds corner")
+ simpl_binds [] bs = ASSERT( null bs ) returnSmpl (nilOL, panic "simplTopBinds corner")
simpl_binds (NonRec bndr rhs : binds) (b:bs) = simplLazyBind True bndr b rhs (simpl_binds binds bs)
simpl_binds (Rec pairs : binds) bs = simplRecBind True pairs (take n bs) (simpl_binds binds (drop n bs))
where
simplRecBind :: Bool -> [(InId, InExpr)] -> [OutId]
-> SimplM (OutStuff a) -> SimplM (OutStuff a)
simplRecBind top_lvl pairs bndrs' thing_inside
- = go pairs bndrs' `thenSmpl` \ (binds', (binds'', res)) ->
- returnSmpl (Rec (flattenBinds binds') : binds'', res)
+ = go pairs bndrs' `thenSmpl` \ (binds', (_, (binds'', res))) ->
+ returnSmpl (unitOL (Rec (flattenBinds (fromOL binds'))) `appOL` binds'', res)
where
go [] _ = thing_inside `thenSmpl` \ stuff ->
- returnSmpl ([], stuff)
+ returnOutStuff stuff
go ((bndr, rhs) : pairs) (bndr' : bndrs')
= simplLazyBind top_lvl bndr bndr' rhs (go pairs bndrs')
%* *
%************************************************************************
-\begin{code}
-addLetBind :: OutId -> OutExpr -> SimplM (OutStuff a) -> SimplM (OutStuff a)
-addLetBind bndr rhs thing_inside
- = thing_inside `thenSmpl` \ (binds, res) ->
- returnSmpl (NonRec bndr rhs : binds, res)
-
-addLetBinds :: [CoreBind] -> SimplM (OutStuff a) -> SimplM (OutStuff a)
-addLetBinds binds1 thing_inside
- = thing_inside `thenSmpl` \ (binds2, res) ->
- returnSmpl (binds1 ++ binds2, res)
-
-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)
-
-addCaseBind bndr rhs thing_inside
- = getInScope `thenSmpl` \ in_scope ->
- thing_inside `thenSmpl` \ (floats, (_, body)) ->
- returnSmpl ([], (in_scope, Case rhs bndr [(DEFAULT, [], mkLets floats body)]))
-
-addNonRecBind bndr rhs thing_inside
- -- Checks for needing a case binding
- | needsCaseBinding (idType bndr) rhs = addCaseBind bndr rhs thing_inside
- | otherwise = addLetBind bndr rhs thing_inside
-\end{code}
-
The reason for this OutExprStuff stuff is that we want to float *after*
simplifying a RHS, not before. If we do so naively we get quadratic
behaviour as things float out.
\begin{code}
simplExpr :: CoreExpr -> SimplM CoreExpr
simplExpr expr = getSubst `thenSmpl` \ subst ->
- simplExprC expr (Stop (substTy subst (exprType expr)))
+ simplExprC expr (mkStop (substTy subst (exprType expr)))
-- The type in the Stop continuation is usually not used
-- It's only needed when discarding continuations after finding
-- a function that returns bottom.
-- Simplify an expression, given a continuation
simplExprC expr cont = simplExprF expr cont `thenSmpl` \ (floats, (_, body)) ->
- returnSmpl (mkLets floats body)
+ returnSmpl (wrapFloats floats body)
simplExprF :: InExpr -> SimplCont -> SimplM OutExprStuff
-- Simplify an expression, returning floated binds
-- If case-of-case is off, simply simplify the case expression
-- in a vanilla Stop context, and rebuild the result around it
simplExprC scrut (Select NoDup bndr alts subst_env
- (Stop (contResultType cont))) `thenSmpl` \ case_expr' ->
+ (mkStop (contResultType cont))) `thenSmpl` \ case_expr' ->
rebuild case_expr' cont
simplExprF expr@(Lam _ _) cont = simplLam expr cont
simplExprF (Type ty) cont
- = ASSERT( case cont of { Stop _ -> True; ArgOf _ _ _ -> True; other -> False } )
+ = ASSERT( case cont of { Stop _ _ -> True; ArgOf _ _ _ -> True; other -> False } )
simplType ty `thenSmpl` \ ty' ->
rebuild (Type ty') cont
simplExprF (Note InlineCall e) cont
= simplExprF e (InlinePlease cont)
--- Comments about the InlineMe case
--- ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+-- Comments about the InlineMe case
+-- ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
-- Don't inline in the RHS of something that has an
-- inline pragma. But be careful that the InScopeEnv that
-- we return does still have inlinings on!
-- the specialised version of g when f is inlined at some call site
-- (perhaps in some other module).
-simplExprF (Note InlineMe e) cont
- = case cont of
- Stop _ -> -- Totally boring continuation
- -- Don't inline inside an INLINE expression
- setBlackList noInlineBlackList (simplExpr e) `thenSmpl` \ e' ->
- rebuild (mkInlineMe e') cont
+-- It's also important not to inline a worker back into a wrapper.
+-- A wrapper looks like
+-- wraper = inline_me (\x -> ...worker... )
+-- Normally, the inline_me prevents the worker getting inlined into
+-- the wrapper (initially, the worker's only call site!). But,
+-- if the wrapper is sure to be called, the strictness analyser will
+-- mark it 'demanded', so when the RHS is simplified, it'll get an ArgOf
+-- continuation. That's why the keep_inline predicate returns True for
+-- ArgOf continuations. It shouldn't do any harm not to dissolve the
+-- inline-me note under these circumstances
- other -> -- Dissolve the InlineMe note if there's
- -- an interesting context of any kind to combine with
- -- (even a type application -- anything except Stop)
- simplExprF e cont
+simplExprF (Note InlineMe e) cont
+ | keep_inline cont -- Totally boring continuation
+ = -- Don't inline inside an INLINE expression
+ setBlackList noInlineBlackList (simplExpr e) `thenSmpl` \ e' ->
+ rebuild (mkInlineMe e') cont
+
+ | otherwise -- Dissolve the InlineMe note if there's
+ -- an interesting context of any kind to combine with
+ -- (even a type application -- anything except Stop)
+ = simplExprF e cont
+ where
+ keep_inline (Stop _ _) = True -- See notes above
+ keep_inline (ArgOf _ _ _) = True -- about this predicate
+ keep_inline other = False
-- A non-recursive let is dealt with by simplBeta
simplExprF (Let (NonRec bndr rhs) body) cont
go expr cont = simplExprF expr cont
-- completeLam deals with the case where a lambda doesn't have an ApplyTo
--- continuation.
--- We used to try for eta reduction here, but I found that this was
--- eta reducing things like
--- f = \x -> (coerce (\x -> e))
--- This made f's arity reduce, which is a bad thing, so I removed the
--- eta reduction at this point, and now do it only when binding
--- (at the call to postInlineUnconditionally)
-
-completeLam acc (Lam bndr body) cont
+-- continuation, so there are real lambdas left to put in the result
+
+-- We try for eta reduction here, but *only* if we get all the
+-- way to an exprIsTrivial expression.
+-- We don't want to remove extra lambdas unless we are going
+-- to avoid allocating this thing altogether
+
+completeLam rev_bndrs (Lam bndr body) cont
= simplBinder bndr $ \ bndr' ->
- completeLam (bndr':acc) body cont
+ completeLam (bndr':rev_bndrs) body cont
-completeLam acc body cont
+completeLam rev_bndrs body cont
= simplExpr body `thenSmpl` \ body' ->
- rebuild (foldl (flip Lam) body' acc) cont
- -- Remember, acc is the *reversed* binders
+ case try_eta body' of
+ Just etad_lam -> tick (EtaReduction (head rev_bndrs)) `thenSmpl_`
+ rebuild etad_lam cont
+
+ Nothing -> rebuild (foldl (flip Lam) body' rev_bndrs) cont
+ where
+ -- We don't use CoreUtils.etaReduce, because we can be more
+ -- efficient here:
+ -- (a) we already have the binders,
+ -- (b) we can do the triviality test before computing the free vars
+ -- [in fact I take the simple path and look for just a variable]
+ -- (c) we don't want to eta-reduce a data con worker or primop
+ -- because we only have to eta-expand them later when we saturate
+ try_eta body | not opt_SimplDoEtaReduction = Nothing
+ | otherwise = go rev_bndrs body
+
+ go (b : bs) (App fun arg) | ok_arg b arg = go bs fun -- Loop round
+ go [] body | ok_body body = Just body -- Success!
+ go _ _ = Nothing -- Failure!
+
+ ok_body (Var v) = not (v `elem` rev_bndrs) && not (hasNoBinding v)
+ ok_body other = False
+ ok_arg b arg = varToCoreExpr b `cheapEqExpr` arg
mkLamBndrZapper :: CoreExpr -- Function
-> SimplCont -- The context
simplValArg arg_ty is_strict arg arg_se cont_ty thing_inside
| is_strict
- = transformRhs arg `thenSmpl` \ t_arg ->
- getEnv `thenSmpl` \ env ->
+ = getEnv `thenSmpl` \ env ->
setSubstEnv arg_se $
- simplExprF t_arg (ArgOf NoDup cont_ty $ \ rhs' ->
+ simplExprF arg (ArgOf NoDup cont_ty $ \ rhs' ->
setAllExceptInScope env $
- thing_inside (etaFirst rhs'))
+ thing_inside rhs')
| otherwise
= simplRhs False {- Not top level -}
True {- OK to float unboxed -}
arg_ty arg arg_se
thing_inside
-
--- Do eta-reduction on the simplified RHS, if eta reduction is on
--- But *only* if we get all the way to an exprIsTrivial expression.
--- We don't want to remove extra lambdas unless we are going
--- to avoid allocating this thing altogether
-etaFirst rhs
- | opt_SimplDoEtaReduction && exprIsTrivial rhs' = rhs'
- | otherwise = rhs
- where
- rhs' = etaReduceExpr rhs
\end{code}
-> SimplM (OutStuff a)
completeBinding old_bndr new_bndr top_lvl black_listed new_rhs thing_inside
- | (case occ_info of -- This happens; for example, the case_bndr during case of
- IAmDead -> True -- known constructor: case (a,b) of x { (p,q) -> ... }
- other -> False) -- Here x isn't mentioned in the RHS, so we don't want to
+ | isDeadOcc occ_info -- This happens; for example, the case_bndr during case of
+ -- known constructor: case (a,b) of x { (p,q) -> ... }
+ -- Here x isn't mentioned in the RHS, so we don't want to
-- create the (dead) let-binding let x = (a,b) in ...
= thing_inside
- | postInlineUnconditionally black_listed occ_info old_bndr new_rhs
- -- Maybe we don't need a let-binding! Maybe we can just
- -- inline it right away. Unlike the preInlineUnconditionally case
- -- we are allowed to look at the RHS.
+ | trivial_rhs && not must_keep_binding
+ -- We're looking at a binding with a trivial RHS, so
+ -- perhaps we can discard it altogether!
--
- -- NB: a loop breaker never has postInlineUnconditionally True
+ -- NB: a loop breaker has must_keep_binding = True
-- and non-loop-breakers only have *forward* references
-- Hence, it's safe to discard the binding
--
- -- NB: You might think that postInlineUnconditionally is an optimisation,
- -- but if we have
- -- let x = f Bool in (x, y)
- -- then because of the constructor, x will not be *inlined* in the pair,
- -- so the trivial binding will stay. But in this postInlineUnconditionally
- -- gag we use the *substitution* to substitute (f Bool) for x, and that *will*
- -- happen.
- = tick (PostInlineUnconditionally old_bndr) `thenSmpl_`
- extendSubst old_bndr (DoneEx new_rhs)
- thing_inside
+ -- NOTE: This isn't our last opportunity to inline.
+ -- We're at the binding site right now, and
+ -- we'll get another opportunity when we get to the ocurrence(s)
+
+ -- Note that we do this unconditional inlining only for trival RHSs.
+ -- Don't inline even WHNFs inside lambdas; doing so may
+ -- simply increase allocation when the function is called
+ -- This isn't the last chance; see NOTE above.
+ --
+ -- NB: Even inline pragmas (e.g. IMustBeINLINEd) are ignored here
+ -- Why? Because we don't even want to inline them into the
+ -- RHS of constructor arguments. See NOTE above
+ --
+ -- NB: Even NOINLINEis ignored here: if the rhs is trivial
+ -- it's best to inline it anyway. We often get a=E; b=a
+ -- from desugaring, with both a and b marked NOINLINE.
+ = -- Drop the binding
+ extendSubst old_bndr (DoneEx new_rhs) $
+ -- Use the substitution to make quite, quite sure that the substitution
+ -- will happen, since we are going to discard the binding
+ tick (PostInlineUnconditionally old_bndr) `thenSmpl_`
+ thing_inside
+
+ | Note coercion@(Coerce _ inner_ty) inner_rhs <- new_rhs,
+ not trivial_rhs && not (isUnLiftedType inner_ty)
+ -- x = coerce t e ==> c = e; x = inline_me (coerce t c)
+ -- Now x can get inlined, which moves the coercion
+ -- to the usage site. This is a bit like worker/wrapper stuff,
+ -- but it's useful to do it very promptly, so that
+ -- x = coerce T (I# 3)
+ -- get's w/wd to
+ -- c = I# 3
+ -- x = coerce T c
+ -- This in turn means that
+ -- case (coerce Int x) of ...
+ -- will inline x.
+ -- Also the full-blown w/w thing isn't set up for non-functions
+ --
+ -- The (not (isUnLiftedType inner_ty)) avoids the nasty case of
+ -- x::Int = coerce Int Int# (foo y)
+ -- ==>
+ -- v::Int# = foo y
+ -- x::Int = coerce Int Int# v
+ -- which would be bogus because then v will be evaluated strictly.
+ -- How can this arise? Via
+ -- x::Int = case (foo y) of { ... }
+ -- followed by case elimination.
+ --
+ -- The inline_me note is so that the simplifier doesn't
+ -- just substitute c back inside x's rhs! (Typically, x will
+ -- get substituted away, but not if it's exported.)
+ = newId SLIT("c") inner_ty $ \ c_id ->
+ completeBinding c_id c_id top_lvl False inner_rhs $
+ completeBinding old_bndr new_bndr top_lvl black_listed
+ (Note InlineMe (Note coercion (Var c_id))) $
+ thing_inside
| otherwise
- = getSubst `thenSmpl` \ subst ->
- let
- -- We make new IdInfo for the new binder by starting from the old binder,
- -- doing appropriate substitutions.
- -- Then we add arity and unfolding info to get the new binder
- old_info = idInfo old_bndr
- new_bndr_info = substIdInfo subst old_info (idInfo new_bndr)
- `setArityInfo` ArityAtLeast (exprArity new_rhs)
-
- -- Add the unfolding *only* for non-loop-breakers
- -- Making loop breakers not have an unfolding at all
- -- means that we can avoid tests in exprIsConApp, for example.
- -- This is important: if exprIsConApp says 'yes' for a recursive
- -- thing, then we can get into an infinite loop
- info_w_unf | isLoopBreaker (occInfo old_info) = new_bndr_info
- | otherwise = new_bndr_info `setUnfoldingInfo` mkUnfolding top_lvl new_rhs
+ = getSubst `thenSmpl` \ subst ->
+ let
+ -- We make new IdInfo for the new binder by starting from the old binder,
+ -- doing appropriate substitutions.
+ -- Then we add arity and unfolding info to get the new binder
+ new_bndr_info = substIdInfo subst old_info (idInfo new_bndr)
+ `setArityInfo` arity_info
+
+ -- Add the unfolding *only* for non-loop-breakers
+ -- Making loop breakers not have an unfolding at all
+ -- means that we can avoid tests in exprIsConApp, for example.
+ -- This is important: if exprIsConApp says 'yes' for a recursive
+ -- thing, then we can get into an infinite loop
+ info_w_unf | loop_breaker = new_bndr_info
+ | otherwise = new_bndr_info `setUnfoldingInfo` mkUnfolding top_lvl new_rhs
final_id = new_bndr `setIdInfo` info_w_unf
- in
- -- These seqs forces the Id, and hence its IdInfo,
- -- and hence any inner substitutions
- final_id `seq`
- addLetBind final_id new_rhs $
- modifyInScope new_bndr final_id thing_inside
+ in
+ -- These seqs forces the Id, and hence its IdInfo,
+ -- and hence any inner substitutions
+ final_id `seq`
+ addLetBind (NonRec final_id new_rhs) $
+ modifyInScope new_bndr final_id thing_inside
where
- occ_info = idOccInfo old_bndr
+ old_info = idInfo old_bndr
+ occ_info = occInfo old_info
+ loop_breaker = isLoopBreaker occ_info
+ trivial_rhs = exprIsTrivial new_rhs
+ must_keep_binding = black_listed || loop_breaker || isExportedId old_bndr
+ arity_info = atLeastArity (exprArity new_rhs)
\end{code}
+
%************************************************************************
%* *
\subsection{simplLazyBind}
\begin{code}
simplRhs :: Bool -- True <=> Top level
-> Bool -- True <=> OK to float unboxed (speculative) bindings
+ -- False for (a) recursive and (b) top-level bindings
-> OutType -- Type of RHS; used only occasionally
-> InExpr -> SubstEnv
-> (OutExpr -> SimplM (OutStuff a))
-> SimplM (OutStuff a)
simplRhs top_lvl float_ubx rhs_ty rhs rhs_se thing_inside
- = -- Swizzle the inner lets past the big lambda (if any)
- -- and try eta expansion
- transformRhs rhs `thenSmpl` \ t_rhs ->
-
- -- Simplify it
- setSubstEnv rhs_se (simplExprF t_rhs (Stop rhs_ty)) `thenSmpl` \ (floats, (in_scope', rhs')) ->
-
- -- Float lets out of RHS
+ = -- Simplify it
+ setSubstEnv rhs_se (simplExprF rhs (mkRhsStop rhs_ty)) `thenSmpl` \ (floats1, (rhs_in_scope, rhs1)) ->
let
- (floats_out, rhs'') = splitFloats float_ubx floats rhs'
+ (floats2, rhs2) = splitFloats float_ubx floats1 rhs1
in
- if (top_lvl || wantToExpose 0 rhs') && -- Float lets if (a) we're at the top level
- not (null floats_out) -- or (b) the resulting RHS is one we'd like to expose
- then
- tickLetFloat floats_out `thenSmpl_`
- -- Do the float
- --
-- There's a subtlety here. There may be a binding (x* = e) in the
-- floats, where the '*' means 'will be demanded'. So is it safe
-- to float it out? Answer no, but it won't matter because
-- we only float if arg' is a WHNF,
-- and so there can't be any 'will be demanded' bindings in the floats.
-- Hence the assert
- WARN( any demanded_float floats_out, ppr floats_out )
- addLetBinds floats_out $
- setInScope in_scope' $
- thing_inside (etaFirst rhs'')
- -- in_scope' may be excessive, but that's OK;
- -- it's a superset of what's in scope
+ WARN( any demanded_float (fromOL floats2), ppr (fromOL floats2) )
+
+ -- Transform the RHS
+ -- It's important that we do eta expansion on function *arguments* (which are
+ -- simplified with simplRhs), as well as let-bound right-hand sides.
+ -- Otherwise we find that things like
+ -- f (\x -> case x of I# x' -> coerce T (\ y -> ...))
+ -- get right through to the code generator as two separate lambdas,
+ -- which is a Bad Thing
+ tryRhsTyLam rhs2 `thenSmpl` \ (floats3, rhs3) ->
+ tryEtaExpansion rhs3 rhs_ty `thenSmpl` \ (floats4, rhs4) ->
+
+ -- Float lets if (a) we're at the top level
+ -- or (b) the resulting RHS is one we'd like to expose
+ if (top_lvl || exprIsCheap rhs4) then
+ (if (isNilOL floats2 && null floats3 && null floats4) then
+ returnSmpl ()
+ else
+ tick LetFloatFromLet) `thenSmpl_`
+
+ addFloats floats2 rhs_in_scope $
+ addAuxiliaryBinds floats3 $
+ addAuxiliaryBinds floats4 $
+ thing_inside rhs4
else
-- Don't do the float
- thing_inside (etaFirst (mkLets floats rhs'))
+ thing_inside (wrapFloats floats1 rhs1)
--- In a let-from-let float, we just tick once, arbitrarily
--- choosing the first floated binder to identify it
-tickLetFloat (NonRec b r : fs) = tick (LetFloatFromLet b)
-tickLetFloat (Rec ((b,r):prs) : fs) = tick (LetFloatFromLet b)
-
demanded_float (NonRec b r) = isStrict (idDemandInfo b) && not (isUnLiftedType (idType b))
-- Unlifted-type (cheap-eagerness) lets may well have a demanded flag on them
demanded_float (Rec _) = False
-- can tolerate them.
splitFloats float_ubx floats rhs
| float_ubx = (floats, rhs) -- Float them all
- | otherwise = go floats
+ | otherwise = go (fromOL floats)
where
- go [] = ([], rhs)
- go (f:fs) | must_stay f = ([], mkLets (f:fs) rhs)
+ go [] = (nilOL, rhs)
+ go (f:fs) | must_stay f = (nilOL, mkLets (f:fs) rhs)
| otherwise = case go fs of
- (out, rhs') -> (f:out, rhs')
+ (out, rhs') -> (f `consOL` out, rhs')
must_stay (Rec prs) = False -- No unlifted bindings in here
must_stay (NonRec b r) = isUnLiftedType (idType b)
-
-wantToExpose :: Int -> CoreExpr -> Bool
--- True for expressions that we'd like to expose at the
--- top level of an RHS. This includes partial applications
--- even if the args aren't cheap; the next pass will let-bind the
--- args and eta expand the partial application. So exprIsCheap won't do.
--- Here's the motivating example:
--- z = letrec g = \x y -> ...g... in g E
--- Even though E is a redex we'd like to float the letrec to give
--- g = \x y -> ...g...
--- z = g E
--- Now the next use of SimplUtils.tryEtaExpansion will give
--- g = \x y -> ...g...
--- z = let v = E in \w -> g v w
--- And now we'll float the v to give
--- g = \x y -> ...g...
--- v = E
--- z = \w -> g v w
--- Which is what we want; chances are z will be inlined now.
---
--- This defn isn't quite like
--- exprIsCheap (it ignores non-cheap args)
--- exprIsValue (may not say True for a lone variable)
--- which is slightly weird
-wantToExpose n (Var v) = idAppIsCheap v n
-wantToExpose n (Lit l) = True
-wantToExpose n (Lam _ e) = True
-wantToExpose n (Note _ e) = wantToExpose n e
-wantToExpose n (App f (Type _)) = wantToExpose n f
-wantToExpose n (App f a) = wantToExpose (n+1) f
-wantToExpose n other = False -- There won't be any lets
\end{code}
---------------------------------------------------------
-- Dealing with a call
-completeCall var occ cont
+completeCall var occ_info cont
= getBlackList `thenSmpl` \ black_list_fn ->
getInScope `thenSmpl` \ in_scope ->
getContArgs var cont `thenSmpl` \ (args, call_cont, inline_call) ->
+ getDOptsSmpl `thenSmpl` \ dflags ->
let
black_listed = black_list_fn var
arg_infos = [ interestingArg in_scope arg subst
inline_cont | inline_call = discardInline cont
| otherwise = cont
- maybe_inline = callSiteInline black_listed inline_call occ
+ maybe_inline = callSiteInline dflags black_listed inline_call occ_info
var arg_infos interesting_cont
in
-- First, look for an inlining
-- But the black-listing mechanism means that inlining of the wrapper
-- won't occur for things that have specialisations till a later phase, so
-- it's ok to try for inlining first.
+ --
+ -- You might think that we shouldn't apply rules for a loop breaker:
+ -- doing so might give rise to an infinite loop, because a RULE is
+ -- rather like an extra equation for the function:
+ -- RULE: f (g x) y = x+y
+ -- Eqn: f a y = a-y
+ --
+ -- But it's too drastic to disable rules for loop breakers.
+ -- Even the foldr/build rule would be disabled, because foldr
+ -- is recursive, and hence a loop breaker:
+ -- foldr k z (build g) = g k z
+ -- So it's up to the programmer: rules can cause divergence
getSwitchChecker `thenSmpl` \ chkr ->
let
case maybe_rule of {
Just (rule_name, rule_rhs) ->
tick (RuleFired rule_name) `thenSmpl_`
+#ifdef DEBUG
+ (if dopt Opt_D_dump_inlinings dflags then
+ pprTrace "Rule fired" (vcat [
+ text "Rule:" <+> ptext rule_name,
+ text "Before:" <+> ppr var <+> sep (map pprParendExpr args'),
+ text "After: " <+> pprCoreExpr rule_rhs])
+ else
+ id) $
+#endif
simplExprF rule_rhs call_cont ;
Nothing -> -- No rules
(a) some might appear as a function argument, so we simply
replace static allocation with dynamic allocation:
l = <...>
- x = f x
+ x = f l
becomes
x = f <...>
OneOcc in_lam once -> not in_lam && once
-- Not inside a lambda, one occurrence ==> safe!
other -> False
-
-
-postInlineUnconditionally :: Bool -- Black listed
- -> OccInfo
- -> InId -> OutExpr -> Bool
- -- Examines a (bndr = rhs) binding, AFTER the rhs has been simplified
- -- It returns True if it's ok to discard the binding and inline the
- -- RHS at every use site.
-
- -- NOTE: This isn't our last opportunity to inline.
- -- We're at the binding site right now, and
- -- we'll get another opportunity when we get to the ocurrence(s)
-
-postInlineUnconditionally black_listed occ_info bndr rhs
- | isExportedId bndr = False -- Don't inline these, ever
- | black_listed = False
- | isLoopBreaker occ_info = False
- | otherwise = exprIsTrivial rhs -- Duplicating is free
- -- Don't inline even WHNFs inside lambdas; doing so may
- -- simply increase allocation when the function is called
- -- This isn't the last chance; see NOTE above.
- --
- -- NB: Even inline pragmas (e.g. IMustBeINLINEd) are ignored here
- -- Why? Because we don't even want to inline them into the
- -- RHS of constructor arguments. See NOTE above
- --
- -- NB: Even NOINLINEis ignored here: if the rhs is trivial
- -- it's best to inline it anyway. We often get a=E; b=a
- -- from desugaring, with both a and b marked NOINLINE.
\end{code}
\begin{code}
-------------------------------------------------------------------
-- Finish rebuilding
-rebuild_done expr
- = getInScope `thenSmpl` \ in_scope ->
- returnSmpl ([], (in_scope, expr))
+rebuild_done expr = returnOutStuff expr
---------------------------------------------------------
rebuild :: OutExpr -> SimplCont -> SimplM OutExprStuff
-- Stop continuation
-rebuild expr (Stop _) = rebuild_done expr
+rebuild expr (Stop _ _) = rebuild_done expr
-- ArgOf continuation
rebuild expr (ArgOf _ _ cont_fn) = cont_fn expr
-> SimplM OutExprStuff
knownCon expr con args bndr alts se cont
- = tick (KnownBranch bndr) `thenSmpl_`
+ = -- Arguments should be atomic;
+ -- yell if not
+ WARN( not (all exprIsTrivial args),
+ text "knownCon" <+> ppr expr )
+ tick (KnownBranch bndr) `thenSmpl_`
setSubstEnv se (
simplBinder bndr $ \ bndr' ->
completeBinding bndr bndr' False False expr $
simplAlts zap_occ_info scrut_cons case_bndr' alts cont'
= mapSmpl simpl_alt alts
where
- inst_tys' = case splitTyConApp_maybe (idType case_bndr') of
- Just (tycon, inst_tys) -> inst_tys
+ inst_tys' = tyConAppArgs (idType case_bndr')
-- handled_cons is all the constructors that are dealt
-- with, either by being impossible, or by there being an alternative
mkDupableCont join_arg_ty (ArgOf _ cont_ty cont_fn) thing_inside
= -- Build the RHS of the join point
newId SLIT("a") join_arg_ty ( \ arg_id ->
- cont_fn (Var arg_id) `thenSmpl` \ (binds, (_, rhs)) ->
- returnSmpl (Lam (setOneShotLambda arg_id) (mkLets binds rhs))
+ cont_fn (Var arg_id) `thenSmpl` \ (floats, (_, rhs)) ->
+ returnSmpl (Lam (setOneShotLambda arg_id) (wrapFloats floats rhs))
) `thenSmpl` \ join_rhs ->
-- Build the join Id and continuation
-- Want to tick here so that we go round again,
-- and maybe copy or inline the code;
-- not strictly CaseOf Case
- addLetBind join_id join_rhs (thing_inside new_cont)
+ addLetBind (NonRec join_id join_rhs) $
+ thing_inside new_cont
mkDupableCont ty (ApplyTo _ arg se cont) thing_inside
= mkDupableCont (funResultTy ty) cont $ \ cont' ->
-- and maybe copy or inline the code;
-- not strictly CaseOf Case
- addLetBind bndr arg' $
+ addLetBind (NonRec bndr arg') $
-- But what if the arg should be case-bound? We can't use
-- addNonRecBind here because its type is too specific.
-- This has been this way for a long time, so I'll leave it,
setSubstEnv se (
simplBinder case_bndr $ \ case_bndr' ->
prepareCaseCont alts cont $ \ cont' ->
- mapAndUnzipSmpl (mkDupableAlt case_bndr case_bndr' cont') alts `thenSmpl` \ (alt_binds_s, alts') ->
- returnSmpl (concat alt_binds_s, alts')
- ) `thenSmpl` \ (alt_binds, alts') ->
+ mkDupableAlts case_bndr case_bndr' cont' alts $ \ alts' ->
+ returnOutStuff alts'
+ ) `thenSmpl` \ (alt_binds, (in_scope, alts')) ->
- addNewInScopeIds [b | NonRec b _ <- alt_binds] $
+ addFloats alt_binds in_scope $
-- NB that the new alternatives, alts', are still InAlts, using the original
-- binders. That means we can keep the case_bndr intact. This is important
-- This is VITAL when the type of case_bndr is an unboxed pair (often the
-- case in I/O rich code. We aren't allowed a lambda bound
-- arg of unboxed tuple type, and indeed such a case_bndr is always dead
- addLetBinds alt_binds $
- thing_inside (Select OkToDup case_bndr alts' se (Stop (contResultType cont)))
-
-mkDupableAlt :: InId -> OutId -> SimplCont -> InAlt -> SimplM (OutStuff InAlt)
-mkDupableAlt case_bndr case_bndr' cont alt@(con, bndrs, rhs)
+ thing_inside (Select OkToDup case_bndr alts' se (mkStop (contResultType cont)))
+
+mkDupableAlts :: InId -> OutId -> SimplCont -> [InAlt]
+ -> ([InAlt] -> SimplM (OutStuff a))
+ -> SimplM (OutStuff a)
+mkDupableAlts case_bndr case_bndr' cont [] thing_inside
+ = thing_inside []
+mkDupableAlts case_bndr case_bndr' cont (alt:alts) thing_inside
+ = mkDupableAlt case_bndr case_bndr' cont alt $ \ alt' ->
+ mkDupableAlts case_bndr case_bndr' cont alts $ \ alts' ->
+ thing_inside (alt' : alts')
+
+mkDupableAlt case_bndr case_bndr' cont alt@(con, bndrs, rhs) thing_inside
= simplBinders bndrs $ \ bndrs' ->
simplExprC rhs cont `thenSmpl` \ rhs' ->
- if (case cont of { Stop _ -> exprIsDupable rhs'; other -> False}) then
+ if (case cont of { Stop _ _ -> exprIsDupable rhs'; other -> False}) then
-- It is worth checking for a small RHS because otherwise we
-- get extra let bindings that may cause an extra iteration of the simplifier to
-- inline back in place. Quite often the rhs is just a variable or constructor.
-- because otherwise we'd need to pair it up with an empty subst-env.
-- (Remember we must zap the subst-env before re-simplifying something).
-- Rather than do this we simply agree to re-simplify the original (small) thing later.
- returnSmpl ([], alt)
+ thing_inside alt
else
let
`thenSmpl` \ (final_bndrs', final_args) ->
-- See comment about "$j" name above
- newId SLIT("$j") (foldr (mkFunTy . idType) rhs_ty' final_bndrs') $ \ join_bndr ->
-
- -- Notice that we make the lambdas into one-shot-lambdas. The
+ newId SLIT("$j") (foldr mkPiType rhs_ty' final_bndrs') $ \ join_bndr ->
+ -- Notice the funky mkPiType. If the contructor has existentials
+ -- it's possible that the join point will be abstracted over
+ -- type varaibles as well as term variables.
+ -- Example: Suppose we have
+ -- data T = forall t. C [t]
+ -- Then faced with
+ -- case (case e of ...) of
+ -- C t xs::[t] -> rhs
+ -- We get the join point
+ -- let j :: forall t. [t] -> ...
+ -- j = /\t \xs::[t] -> rhs
+ -- in
+ -- case (case e of ...) of
+ -- C t xs::[t] -> j t xs
+
+ let
+ -- We make the lambdas into one-shot-lambdas. The
-- join point is sure to be applied at most once, and doing so
-- prevents the body of the join point being floated out by
-- the full laziness pass
- returnSmpl ([NonRec join_bndr (mkLams (map setOneShotLambda final_bndrs') rhs')],
- (con, bndrs, mkApps (Var join_bndr) final_args))
+ really_final_bndrs = map one_shot final_bndrs'
+ one_shot v | isId v = setOneShotLambda v
+ | otherwise = v
+ in
+ addLetBind (NonRec join_bndr (mkLams really_final_bndrs rhs')) $
+ thing_inside (con, bndrs, mkApps (Var join_bndr) final_args)
\end{code}