Overview
***************************
-* We attach binding levels to Core bindings, in preparation for floating
- outwards (@FloatOut@).
+1. We attach binding levels to Core bindings, in preparation for floating
+ outwards (@FloatOut@).
-* We also let-ify many expressions (notably case scrutinees), so they
- will have a fighting chance of being floated sensible.
+2. We also let-ify many expressions (notably case scrutinees), so they
+ will have a fighting chance of being floated sensible.
-* We clone the binders of any floatable let-binding, so that when it is
- floated out it will be unique. (This used to be done by the simplifier
- but the latter now only ensures that there's no shadowing.)
- NOTE: Very tiresomely, we must apply this substitution to
- the rules stored inside a variable too.
+3. We clone the binders of any floatable let-binding, so that when it is
+ floated out it will be unique. (This used to be done by the simplifier
+ but the latter now only ensures that there's no shadowing; indeed, even
+ that may not be true.)
- We do *not* clone top-level bindings, because some of them must not change,
- but we *do* clone bindings that are heading for the top level
+ NOTE: this can't be done using the uniqAway idea, because the variable
+ must be unique in the whole program, not just its current scope,
+ because two variables in different scopes may float out to the
+ same top level place
-* In the expression
+ NOTE: Very tiresomely, we must apply this substitution to
+ the rules stored inside a variable too.
+
+ We do *not* clone top-level bindings, because some of them must not change,
+ but we *do* clone bindings that are heading for the top level
+
+4. In the expression
case x of wild { p -> ...wild... }
- we substitute x for wild in the RHS of the case alternatives:
+ we substitute x for wild in the RHS of the case alternatives:
case x of wild { p -> ...x... }
- This means that a sub-expression involving x is not "trapped" inside the RHS.
- And it's not inconvenient because we already have a substitution.
+ This means that a sub-expression involving x is not "trapped" inside the RHS.
+ And it's not inconvenient because we already have a substitution.
+
+ Note that this is EXACTLY BACKWARDS from the what the simplifier does.
+ The simplifier tries to get rid of occurrences of x, in favour of wild,
+ in the hope that there will only be one remaining occurrence of x, namely
+ the scrutinee of the case, and we can inline it.
\begin{code}
module SetLevels (
- setLevels,
+ setLevels,
Level(..), tOP_LEVEL,
+ LevelledBind, LevelledExpr,
- incMinorLvl, ltMajLvl, ltLvl, isTopLvl
+ incMinorLvl, ltMajLvl, ltLvl, isTopLvl, isInlineCtxt
) where
#include "HsVersions.h"
import CoreSyn
-import CoreUtils ( coreExprType, exprIsTrivial, exprIsBottom )
+import DynFlags ( FloatOutSwitches(..) )
+import CoreUtils ( exprType, exprIsTrivial, exprIsCheap, mkPiTypes )
import CoreFVs -- all of it
-import Id ( Id, idType, mkSysLocal, isOneShotLambda, modifyIdInfo,
- getIdSpecialisation, getIdWorkerInfo
+import CoreSubst ( Subst, emptySubst, extendInScope, extendIdSubst,
+ cloneIdBndr, cloneRecIdBndrs )
+import Id ( Id, idType, mkSysLocal, isOneShotLambda,
+ zapDemandIdInfo,
+ idSpecialisation, idWorkerInfo, setIdInfo
)
-import IdInfo ( workerExists )
-import Var ( IdOrTyVar, Var, TyVar, setVarUnique )
-import VarEnv
-import Subst
+import IdInfo ( workerExists, vanillaIdInfo, isEmptySpecInfo )
+import Var ( Var )
import VarSet
+import VarEnv
import Name ( getOccName )
-import OccName ( occNameUserString )
-import Type ( isUnLiftedType, mkTyVarTy, mkForAllTys, Type )
+import OccName ( occNameString )
+import Type ( isUnLiftedType, Type )
import BasicTypes ( TopLevelFlag(..) )
-import VarSet
-import VarEnv
import UniqSupply
-import Maybes ( maybeToBool )
-import Util ( zipWithEqual, zipEqual )
+import Util ( sortLe, isSingleton, count )
import Outputable
-import List ( nub )
+import FastString
\end{code}
%************************************************************************
%************************************************************************
\begin{code}
-data Level = Level Int -- Level number of enclosing lambdas
+data Level = InlineCtxt -- A level that's used only for
+ -- the context parameter ctxt_lvl
+ | Level Int -- Level number of enclosing lambdas
Int -- Number of big-lambda and/or case expressions between
-- here and the nearest enclosing lambda
\end{code}
That's meant to be the level number of the enclosing binder in the
final (floated) program. If the level number of a sub-expression is
less than that of the context, then it might be worth let-binding the
-sub-expression so that it will indeed float. This context level starts
-at @Level 0 0@.
+sub-expression so that it will indeed float.
+
+If you can float to level @Level 0 0@ worth doing so because then your
+allocation becomes static instead of dynamic. We always start with
+context @Level 0 0@.
+
+
+InlineCtxt
+~~~~~~~~~~
+@InlineCtxt@ very similar to @Level 0 0@, but is used for one purpose:
+to say "don't float anything out of here". That's exactly what we
+want for the body of an INLINE, where we don't want to float anything
+out at all. See notes with lvlMFE below.
+
+But, check this out:
+
+-- At one time I tried the effect of not float anything out of an InlineMe,
+-- but it sometimes works badly. For example, consider PrelArr.done. It
+-- has the form __inline (\d. e)
+-- where e doesn't mention d. If we float this to
+-- __inline (let x = e in \d. x)
+-- things are bad. The inliner doesn't even inline it because it doesn't look
+-- like a head-normal form. So it seems a lesser evil to let things float.
+-- In SetLevels we do set the context to (Level 0 0) when we get to an InlineMe
+-- which discourages floating out.
+
+So the conclusion is: don't do any floating at all inside an InlineMe.
+(In the above example, don't float the {x=e} out of the \d.)
+
+One particular case is that of workers: we don't want to float the
+call to the worker outside the wrapper, otherwise the worker might get
+inlined into the floated expression, and an importing module won't see
+the worker at all.
\begin{code}
type LevelledExpr = TaggedExpr Level
-type LevelledArg = TaggedArg Level
type LevelledBind = TaggedBind Level
-tOP_LEVEL = Level 0 0
+tOP_LEVEL = Level 0 0
+iNLINE_CTXT = InlineCtxt
incMajorLvl :: Level -> Level
+-- For InlineCtxt we ignore any inc's; we don't want
+-- to do any floating at all; see notes above
+incMajorLvl InlineCtxt = InlineCtxt
incMajorLvl (Level major minor) = Level (major+1) 0
incMinorLvl :: Level -> Level
+incMinorLvl InlineCtxt = InlineCtxt
incMinorLvl (Level major minor) = Level major (minor+1)
maxLvl :: Level -> Level -> Level
+maxLvl InlineCtxt l2 = l2
+maxLvl l1 InlineCtxt = l1
maxLvl l1@(Level maj1 min1) l2@(Level maj2 min2)
| (maj1 > maj2) || (maj1 == maj2 && min1 > min2) = l1
| otherwise = l2
ltLvl :: Level -> Level -> Bool
+ltLvl any_lvl InlineCtxt = False
+ltLvl InlineCtxt (Level _ _) = True
ltLvl (Level maj1 min1) (Level maj2 min2)
= (maj1 < maj2) || (maj1 == maj2 && min1 < min2)
ltMajLvl :: Level -> Level -> Bool
-- Tells if one level belongs to a difft *lambda* level to another
+ltMajLvl any_lvl InlineCtxt = False
+ltMajLvl InlineCtxt (Level maj2 _) = 0 < maj2
ltMajLvl (Level maj1 _) (Level maj2 _) = maj1 < maj2
isTopLvl :: Level -> Bool
isTopLvl (Level 0 0) = True
-isTopLvl other = False
+isTopLvl other = False
+
+isInlineCtxt :: Level -> Bool
+isInlineCtxt InlineCtxt = True
+isInlineCtxt other = False
instance Outputable Level where
+ ppr InlineCtxt = text "<INLINE>"
ppr (Level maj min) = hcat [ char '<', int maj, char ',', int min, char '>' ]
+
+instance Eq Level where
+ InlineCtxt == InlineCtxt = True
+ (Level maj1 min1) == (Level maj2 min2) = maj1==maj2 && min1==min2
+ l1 == l2 = False
\end{code}
+
%************************************************************************
%* *
\subsection{Main level-setting code}
%************************************************************************
\begin{code}
-setLevels :: [CoreBind]
+setLevels :: FloatOutSwitches
+ -> [CoreBind]
-> UniqSupply
-> [LevelledBind]
-setLevels binds us
+setLevels float_lams binds us
= initLvl us (do_them binds)
where
-- "do_them"'s main business is to thread the monad along
do_them [] = returnLvl []
do_them (b:bs)
- = lvlTopBind b `thenLvl` \ (lvld_bind, _) ->
- do_them bs `thenLvl` \ lvld_binds ->
+ = lvlTopBind init_env b `thenLvl` \ (lvld_bind, _) ->
+ do_them bs `thenLvl` \ lvld_binds ->
returnLvl (lvld_bind : lvld_binds)
-lvlTopBind (NonRec binder rhs)
- = lvlBind TopLevel tOP_LEVEL initialEnv (AnnNonRec binder (freeVars rhs))
+ init_env = initialEnv float_lams
+
+lvlTopBind env (NonRec binder rhs)
+ = lvlBind TopLevel tOP_LEVEL env (AnnNonRec binder (freeVars rhs))
-- Rhs can have no free vars!
-lvlTopBind (Rec pairs)
- = lvlBind TopLevel tOP_LEVEL initialEnv (AnnRec [(b,freeVars rhs) | (b,rhs) <- pairs])
+lvlTopBind env (Rec pairs)
+ = lvlBind TopLevel tOP_LEVEL env (AnnRec [(b,freeVars rhs) | (b,rhs) <- pairs])
\end{code}
%************************************************************************
If there were another lambda in @r@'s rhs, it would get level-2 as well.
\begin{code}
-lvlExpr _ _ (_, AnnType ty) = returnLvl (Type ty)
-lvlExpr _ env (_, AnnVar v) = returnLvl (lookupVar env v)
-
-lvlExpr ctxt_lvl env (_, AnnCon con args)
- = mapLvl (lvlExpr ctxt_lvl env) args `thenLvl` \ args' ->
- returnLvl (Con con args')
+lvlExpr _ _ (_, AnnType ty) = returnLvl (Type ty)
+lvlExpr _ env (_, AnnVar v) = returnLvl (lookupVar env v)
+lvlExpr _ env (_, AnnLit lit) = returnLvl (Lit lit)
lvlExpr ctxt_lvl env (_, AnnApp fun arg)
- = lvlExpr ctxt_lvl env fun `thenLvl` \ fun' ->
+ = lvl_fun fun `thenLvl` \ fun' ->
lvlMFE False ctxt_lvl env arg `thenLvl` \ arg' ->
returnLvl (App fun' arg')
+ where
+-- gaw 2004
+ lvl_fun (_, AnnCase _ _ _ _) = lvlMFE True ctxt_lvl env fun
+ lvl_fun other = lvlExpr ctxt_lvl env fun
+ -- We don't do MFE on partial applications generally,
+ -- but we do if the function is big and hairy, like a case
lvlExpr ctxt_lvl env (_, AnnNote InlineMe expr)
- -- Don't float anything out of an InlineMe
- = lvlExpr tOP_LEVEL env expr `thenLvl` \ expr' ->
+-- Don't float anything out of an InlineMe; hence the iNLINE_CTXT
+ = lvlExpr iNLINE_CTXT env expr `thenLvl` \ expr' ->
returnLvl (Note InlineMe expr')
lvlExpr ctxt_lvl env (_, AnnNote note expr)
-- lambdas makes them more expensive.
lvlExpr ctxt_lvl env expr@(_, AnnLam bndr rhs)
- = go (incMinorLvl ctxt_lvl) env False {- Havn't bumped major level in this group -} expr
+ = lvlMFE True new_lvl new_env body `thenLvl` \ new_body ->
+ returnLvl (mkLams new_bndrs new_body)
where
- go lvl env bumped_major (_, AnnLam bndr body)
- = go new_lvl new_env new_bumped_major body `thenLvl` \ new_body ->
- returnLvl (Lam lvld_bndr new_body)
- where
- -- Go to the next major level if this is a value binder,
- -- and we havn't already gone to the next level (one jump per group)
- -- and it isn't a one-shot lambda
- (new_lvl, new_bumped_major)
- | isId bndr &&
- not bumped_major &&
- not (isOneShotLambda bndr) = (incMajorLvl ctxt_lvl, True)
- | otherwise = (lvl, bumped_major)
- new_env = extendLvlEnv env [lvld_bndr]
- lvld_bndr = (bndr, new_lvl)
-
- -- Ignore notes, because we don't want to split
+ (bndrs, body) = collectAnnBndrs expr
+ (new_lvl, new_bndrs) = lvlLamBndrs ctxt_lvl bndrs
+ new_env = extendLvlEnv env new_bndrs
+ -- At one time we called a special verion of collectBinders,
+ -- which ignored coercions, because we don't want to split
-- a lambda like this (\x -> coerce t (\s -> ...))
- -- This happens quite a bit in state-transformer programs
- go lvl env bumped_major (_, AnnNote note body)
- = go lvl env bumped_major body `thenLvl` \ new_body ->
- returnLvl (Note note new_body)
-
- go lvl env bumped_major body
- = lvlMFE True lvl env body
-
+ -- This used to happen quite a bit in state-transformer programs,
+ -- but not nearly so much now non-recursive newtypes are transparent.
+ -- [See SetLevels rev 1.50 for a version with this approach.]
+
+lvlExpr ctxt_lvl env (_, AnnLet (AnnNonRec bndr rhs) body)
+ | isUnLiftedType (idType bndr)
+ -- Treat unlifted let-bindings (let x = b in e) just like (case b of x -> e)
+ -- That is, leave it exactly where it is
+ -- We used to float unlifted bindings too (e.g. to get a cheap primop
+ -- outside a lambda (to see how, look at lvlBind in rev 1.58)
+ -- but an unrelated change meant that these unlifed bindings
+ -- could get to the top level which is bad. And there's not much point;
+ -- unlifted bindings are always cheap, and so hardly worth floating.
+ = lvlExpr ctxt_lvl env rhs `thenLvl` \ rhs' ->
+ lvlExpr incd_lvl env' body `thenLvl` \ body' ->
+ returnLvl (Let (NonRec bndr' rhs') body')
+ where
+ incd_lvl = incMinorLvl ctxt_lvl
+ bndr' = TB bndr incd_lvl
+ env' = extendLvlEnv env [bndr']
lvlExpr ctxt_lvl env (_, AnnLet bind body)
= lvlBind NotTopLevel ctxt_lvl env bind `thenLvl` \ (bind', new_env) ->
lvlExpr ctxt_lvl new_env body `thenLvl` \ body' ->
returnLvl (Let bind' body')
-lvlExpr ctxt_lvl env (_, AnnCase expr case_bndr alts)
+lvlExpr ctxt_lvl env (_, AnnCase expr case_bndr ty alts)
= lvlMFE True ctxt_lvl env expr `thenLvl` \ expr' ->
let
alts_env = extendCaseBndrLvlEnv env expr' case_bndr incd_lvl
in
mapLvl (lvl_alt alts_env) alts `thenLvl` \ alts' ->
- returnLvl (Case expr' (case_bndr, incd_lvl) alts')
+ returnLvl (Case expr' (TB case_bndr incd_lvl) ty alts')
where
- expr_type = coreExprType (deAnnotate expr)
incd_lvl = incMinorLvl ctxt_lvl
lvl_alt alts_env (con, bs, rhs)
= lvlMFE True incd_lvl new_env rhs `thenLvl` \ rhs' ->
returnLvl (con, bs', rhs')
where
- bs' = [ (b, incd_lvl) | b <- bs ]
+ bs' = [ TB b incd_lvl | b <- bs ]
new_env = extendLvlEnv alts_env bs'
\end{code}
@lvlMFE@ is just like @lvlExpr@, except that it might let-bind
the expression, so that it can itself be floated.
+[NOTE: unlifted MFEs]
+We don't float unlifted MFEs, which potentially loses big opportunites.
+For example:
+ \x -> f (h y)
+where h :: Int -> Int# is expensive. We'd like to float the (h y) outside
+the \x, but we don't because it's unboxed. Possible solution: box it.
+
\begin{code}
lvlMFE :: Bool -- True <=> strict context [body of case or let]
-> Level -- Level of innermost enclosing lambda/tylam
lvlMFE strict_ctxt ctxt_lvl env (_, AnnType ty)
= returnLvl (Type ty)
+
lvlMFE strict_ctxt ctxt_lvl env ann_expr@(fvs, _)
- | isUnLiftedType ty -- Can't let-bind it
- || not (dest_lvl `ltMajLvl` ctxt_lvl) -- Does not escape a value lambda
- -- A decision to float entails let-binding this thing, and we only do
- -- that if we'll escape a value lambda. I considered doing it if it
- -- would make the thing go to top level, but I found things like
- -- concat = /\ a -> foldr ..a.. (++) []
- -- was getting turned into
- -- concat = /\ a -> lvl a
- -- lvl = /\ a -> foldr ..a.. (++) []
- -- which is pretty stupid. So for now at least, I don't let-bind things
- -- simply because they could go to top level.
- || exprIsTrivial expr -- Is trivial
- || (strict_ctxt && exprIsBottom expr) -- Strict context and is bottom
+ | isUnLiftedType ty -- Can't let-bind it; see [NOTE: unlifted MFEs]
+ || isInlineCtxt ctxt_lvl -- Don't float out of an __inline__ context
+ || exprIsTrivial expr -- Never float if it's trivial
+ || not good_destination
= -- Don't float it out
lvlExpr ctxt_lvl env ann_expr
| otherwise -- Float it out!
- = lvlExpr expr_lvl expr_env ann_expr `thenLvl` \ expr' ->
- newLvlVar "lvl" (mkForAllTys tyvars ty) `thenLvl` \ var ->
- returnLvl (Let (NonRec (var,dest_lvl) (mkLams tyvars_w_lvls expr'))
- (mkTyVarApps var tyvars))
+ = lvlFloatRhs abs_vars dest_lvl env ann_expr `thenLvl` \ expr' ->
+ newLvlVar "lvl" abs_vars ty `thenLvl` \ var ->
+ returnLvl (Let (NonRec (TB var dest_lvl) expr')
+ (mkVarApps (Var var) abs_vars))
where
expr = deAnnotate ann_expr
- ty = coreExprType expr
- dest_lvl = destLevel env fvs
- (tyvars, tyvars_w_lvls, expr_lvl) = abstractTyVars dest_lvl env fvs
- expr_env = extendLvlEnv env tyvars_w_lvls
+ ty = exprType expr
+ dest_lvl = destLevel env fvs (isFunction ann_expr)
+ abs_vars = abstractVars dest_lvl env fvs
+
+ -- A decision to float entails let-binding this thing, and we only do
+ -- that if we'll escape a value lambda, or will go to the top level.
+ good_destination
+ | dest_lvl `ltMajLvl` ctxt_lvl -- Escapes a value lambda
+ = not (exprIsCheap expr) || isTopLvl dest_lvl
+ -- Even if it escapes a value lambda, we only
+ -- float if it's not cheap (unless it'll get all the
+ -- way to the top). I've seen cases where we
+ -- float dozens of tiny free expressions, which cost
+ -- more to allocate than to evaluate.
+ -- NB: exprIsCheap is also true of bottom expressions, which
+ -- is good; we don't want to share them
+ --
+ -- It's only Really Bad to float a cheap expression out of a
+ -- strict context, because that builds a thunk that otherwise
+ -- would never be built. So another alternative would be to
+ -- add
+ -- || (strict_ctxt && not (exprIsBottom expr))
+ -- to the condition above. We should really try this out.
+
+ | otherwise -- Does not escape a value lambda
+ = isTopLvl dest_lvl -- Only float if we are going to the top level
+ && floatConsts env -- and the floatConsts flag is on
+ && not strict_ctxt -- Don't float from a strict context
+ -- We are keen to float something to the top level, even if it does not
+ -- escape a lambda, because then it needs no allocation. But it's controlled
+ -- by a flag, because doing this too early loses opportunities for RULES
+ -- which (needless to say) are important in some nofib programs
+ -- (gcd is an example).
+ --
+ -- Beware:
+ -- concat = /\ a -> foldr ..a.. (++) []
+ -- was getting turned into
+ -- concat = /\ a -> lvl a
+ -- lvl = /\ a -> foldr ..a.. (++) []
+ -- which is pretty stupid. Hence the strict_ctxt test
\end{code}
-> LvlM (LevelledBind, LevelEnv)
lvlBind top_lvl ctxt_lvl env (AnnNonRec bndr rhs@(rhs_fvs,_))
- | null tyvars
+ | isInlineCtxt ctxt_lvl -- Don't do anything inside InlineMe
+ = lvlExpr ctxt_lvl env rhs `thenLvl` \ rhs' ->
+ returnLvl (NonRec (TB bndr ctxt_lvl) rhs', env)
+
+ | null abs_vars
= -- No type abstraction; clone existing binder
- lvlExpr rhs_lvl rhs_env rhs `thenLvl` \ rhs' ->
- cloneVar top_lvl env bndr dest_lvl `thenLvl` \ (env', bndr') ->
- returnLvl (NonRec (bndr', dest_lvl) rhs', env')
+ lvlExpr dest_lvl env rhs `thenLvl` \ rhs' ->
+ cloneVar top_lvl env bndr ctxt_lvl dest_lvl `thenLvl` \ (env', bndr') ->
+ returnLvl (NonRec (TB bndr' dest_lvl) rhs', env')
| otherwise
= -- Yes, type abstraction; create a new binder, extend substitution, etc
- WARN( workerExists (getIdWorkerInfo bndr)
- || not (isEmptyCoreRules (getIdSpecialisation bndr)),
- text "lvlBind: discarding info on" <+> ppr bndr )
-
- lvl_poly_rhs tyvars_w_lvls rhs_lvl rhs_env rhs `thenLvl` \ rhs' ->
- new_poly_bndr tyvars bndr `thenLvl` \ bndr' ->
- let
- env' = extendPolyLvlEnv env dest_lvl tyvars [(bndr, bndr')]
- in
- returnLvl (NonRec (bndr', dest_lvl) rhs', env')
+ lvlFloatRhs abs_vars dest_lvl env rhs `thenLvl` \ rhs' ->
+ newPolyBndrs dest_lvl env abs_vars [bndr] `thenLvl` \ (env', [bndr']) ->
+ returnLvl (NonRec (TB bndr' dest_lvl) rhs', env')
where
bind_fvs = rhs_fvs `unionVarSet` idFreeVars bndr
-
- dest_lvl | isUnLiftedType (idType bndr) = destLevel env bind_fvs `maxLvl` Level 1 0
- | otherwise = destLevel env bind_fvs
- -- Hack alert! We do have some unlifted bindings, for cheap primops, and
- -- it is ok to float them out; but not to the top level. If they would otherwise
- -- go to the top level, we pin them inside the topmost lambda
-
- (tyvars, tyvars_w_lvls, rhs_lvl) = abstractTyVars dest_lvl env bind_fvs
- rhs_env = extendLvlEnv env tyvars_w_lvls
+ abs_vars = abstractVars dest_lvl env bind_fvs
+ dest_lvl = destLevel env bind_fvs (isFunction rhs)
\end{code}
\begin{code}
lvlBind top_lvl ctxt_lvl env (AnnRec pairs)
- | null tyvars
- = cloneVars top_lvl env bndrs dest_lvl `thenLvl` \ (new_env, new_bndrs) ->
- mapLvl (lvlExpr rhs_lvl new_env) rhss `thenLvl` \ new_rhss ->
- returnLvl (Rec ((new_bndrs `zip` repeat dest_lvl) `zip` new_rhss), new_env)
-
- | otherwise
- = mapLvl (new_poly_bndr tyvars) bndrs `thenLvl` \ new_bndrs ->
+ | isInlineCtxt ctxt_lvl -- Don't do anything inside InlineMe
+ = mapLvl (lvlExpr ctxt_lvl env) rhss `thenLvl` \ rhss' ->
+ returnLvl (Rec ([TB b ctxt_lvl | b <- bndrs] `zip` rhss'), env)
+
+ | null abs_vars
+ = cloneRecVars top_lvl env bndrs ctxt_lvl dest_lvl `thenLvl` \ (new_env, new_bndrs) ->
+ mapLvl (lvlExpr ctxt_lvl new_env) rhss `thenLvl` \ new_rhss ->
+ returnLvl (Rec ([TB b dest_lvl | b <- new_bndrs] `zip` new_rhss), new_env)
+
+ | isSingleton pairs && count isId abs_vars > 1
+ = -- Special case for self recursion where there are
+ -- several variables carried around: build a local loop:
+ -- poly_f = \abs_vars. \lam_vars . letrec f = \lam_vars. rhs in f lam_vars
+ -- This just makes the closures a bit smaller. If we don't do
+ -- this, allocation rises significantly on some programs
+ --
+ -- We could elaborate it for the case where there are several
+ -- mutually functions, but it's quite a bit more complicated
+ --
+ -- This all seems a bit ad hoc -- sigh
+ let
+ (bndr,rhs) = head pairs
+ (rhs_lvl, abs_vars_w_lvls) = lvlLamBndrs dest_lvl abs_vars
+ rhs_env = extendLvlEnv env abs_vars_w_lvls
+ in
+ cloneVar NotTopLevel rhs_env bndr rhs_lvl rhs_lvl `thenLvl` \ (rhs_env', new_bndr) ->
let
- new_env = extendPolyLvlEnv env dest_lvl tyvars (bndrs `zip` new_bndrs)
- rhs_env = extendLvlEnv new_env tyvars_w_lvls
- in
- mapLvl (lvl_poly_rhs tyvars_w_lvls rhs_lvl rhs_env) rhss `thenLvl` \ new_rhss ->
- returnLvl (Rec ((new_bndrs `zip` repeat dest_lvl) `zip` new_rhss), new_env)
+ (lam_bndrs, rhs_body) = collectAnnBndrs rhs
+ (body_lvl, new_lam_bndrs) = lvlLamBndrs rhs_lvl lam_bndrs
+ body_env = extendLvlEnv rhs_env' new_lam_bndrs
+ in
+ lvlExpr body_lvl body_env rhs_body `thenLvl` \ new_rhs_body ->
+ newPolyBndrs dest_lvl env abs_vars [bndr] `thenLvl` \ (poly_env, [poly_bndr]) ->
+ returnLvl (Rec [(TB poly_bndr dest_lvl,
+ mkLams abs_vars_w_lvls $
+ mkLams new_lam_bndrs $
+ Let (Rec [(TB new_bndr rhs_lvl, mkLams new_lam_bndrs new_rhs_body)])
+ (mkVarApps (Var new_bndr) lam_bndrs))],
+ poly_env)
+
+ | otherwise -- Non-null abs_vars
+ = newPolyBndrs dest_lvl env abs_vars bndrs `thenLvl` \ (new_env, new_bndrs) ->
+ mapLvl (lvlFloatRhs abs_vars dest_lvl new_env) rhss `thenLvl` \ new_rhss ->
+ returnLvl (Rec ([TB b dest_lvl | b <- new_bndrs] `zip` new_rhss), new_env)
where
(bndrs,rhss) = unzip pairs
`minusVarSet`
mkVarSet bndrs
- dest_lvl = destLevel env bind_fvs
-
- (tyvars, tyvars_w_lvls, rhs_lvl) = abstractTyVars dest_lvl env bind_fvs
+ dest_lvl = destLevel env bind_fvs (all isFunction rhss)
+ abs_vars = abstractVars dest_lvl env bind_fvs
----------------------------------------------------
--- Three help functons Stuff for the type-abstraction case
+-- Three help functons for the type-abstraction case
-new_poly_bndr tyvars bndr
- = newLvlVar ("poly_" ++ occNameUserString (getOccName bndr))
- (mkForAllTys tyvars (idType bndr))
-
-lvl_poly_rhs tyvars_w_lvls rhs_lvl rhs_env rhs
- = lvlExpr rhs_lvl rhs_env rhs `thenLvl` \ rhs' ->
- returnLvl (mkLams tyvars_w_lvls rhs')
+lvlFloatRhs abs_vars dest_lvl env rhs
+ = lvlExpr rhs_lvl rhs_env rhs `thenLvl` \ rhs' ->
+ returnLvl (mkLams abs_vars_w_lvls rhs')
+ where
+ (rhs_lvl, abs_vars_w_lvls) = lvlLamBndrs dest_lvl abs_vars
+ rhs_env = extendLvlEnv env abs_vars_w_lvls
\end{code}
%************************************************************************
\begin{code}
-abstractTyVars :: Level -> LevelEnv -> VarSet
- -> ([TyVar], [(TyVar,Level)], Level)
- -- Find the tyvars whose level is higher than the supplied level
- -- There should be no Ids with this property
-abstractTyVars lvl env fvs
- | null tyvars = ([], [], lvl) -- Don't increment level
-
- | otherwise
- = ASSERT( not (any bad fv_list) )
- (tyvars, tyvars_w_lvls, incd_lvl)
+lvlLamBndrs :: Level -> [CoreBndr] -> (Level, [TaggedBndr Level])
+-- Compute the levels for the binders of a lambda group
+-- The binders returned are exactly the same as the ones passed,
+-- but they are now paired with a level
+lvlLamBndrs lvl []
+ = (lvl, [])
+
+lvlLamBndrs lvl bndrs
+ = go (incMinorLvl lvl)
+ False -- Havn't bumped major level in this group
+ [] bndrs
where
- bad v = isId v && lvl `ltLvl` varLevel env v
- fv_list = varSetElems fvs
- tyvars = nub [tv | v <- fv_list, tv <- tvs_of v, abstract_tv tv]
+ go old_lvl bumped_major rev_lvld_bndrs (bndr:bndrs)
+ | isId bndr && -- Go to the next major level if this is a value binder,
+ not bumped_major && -- and we havn't already gone to the next level (one jump per group)
+ not (isOneShotLambda bndr) -- and it isn't a one-shot lambda
+ = go new_lvl True (TB bndr new_lvl : rev_lvld_bndrs) bndrs
- -- If f is free in the exression, and f maps to poly_f a b c in the
- -- current substitution, then we must report a b c as candidate type
- -- variables
- tvs_of v | isId v = lookupTyVars env v
- | otherwise = [v]
-
- abstract_tv var | isId var = False
- | otherwise = lvl `ltLvl` varLevel env var
+ | otherwise
+ = go old_lvl bumped_major (TB bndr old_lvl : rev_lvld_bndrs) bndrs
- -- These defns are just like those in the TyLam case of lvlExpr
- incd_lvl = incMinorLvl lvl
- tyvars_w_lvls = [(tv,incd_lvl) | tv <- tyvars]
+ where
+ new_lvl = incMajorLvl old_lvl
+ go old_lvl _ rev_lvld_bndrs []
+ = (old_lvl, reverse rev_lvld_bndrs)
+ -- a lambda like this (\x -> coerce t (\s -> ...))
+ -- This happens quite a bit in state-transformer programs
+\end{code}
+\begin{code}
-- Destintion level is the max Id level of the expression
-- (We'll abstract the type variables, if any.)
-destLevel :: LevelEnv -> VarSet -> Level
-destLevel env fvs = foldVarSet (maxIdLvl env) tOP_LEVEL fvs
-
-maxIdLvl :: LevelEnv -> IdOrTyVar -> Level -> Level
-maxIdLvl (lvl_env,_,_) var lvl | isTyVar var = lvl
- | otherwise = case lookupVarEnv lvl_env var of
- Just lvl' -> maxLvl lvl' lvl
- Nothing -> lvl
+destLevel :: LevelEnv -> VarSet -> Bool -> Level
+destLevel env fvs is_function
+ | floatLams env
+ && is_function = tOP_LEVEL -- Send functions to top level; see
+ -- the comments with isFunction
+ | otherwise = maxIdLevel env fvs
+
+isFunction :: CoreExprWithFVs -> Bool
+-- The idea here is that we want to float *functions* to
+-- the top level. This saves no work, but
+-- (a) it can make the host function body a lot smaller,
+-- and hence inlinable.
+-- (b) it can also save allocation when the function is recursive:
+-- h = \x -> letrec f = \y -> ...f...y...x...
+-- in f x
+-- becomes
+-- f = \x y -> ...(f x)...y...x...
+-- h = \x -> f x x
+-- No allocation for f now.
+-- We may only want to do this if there are sufficiently few free
+-- variables. We certainly only want to do it for values, and not for
+-- constructors. So the simple thing is just to look for lambdas
+isFunction (_, AnnLam b e) | isId b = True
+ | otherwise = isFunction e
+isFunction (_, AnnNote n e) = isFunction e
+isFunction other = False
\end{code}
%************************************************************************
\begin{code}
-type LevelEnv = (VarEnv Level, SubstEnv, IdEnv ([TyVar], LevelledExpr))
+type LevelEnv = (FloatOutSwitches,
+ VarEnv Level, -- Domain is *post-cloned* TyVars and Ids
+ Subst, -- Domain is pre-cloned Ids; tracks the in-scope set
+ -- so that subtitution is capture-avoiding
+ IdEnv ([Var], LevelledExpr)) -- Domain is pre-cloned Ids
-- We clone let-bound variables so that they are still
-- distinct when floated out; hence the SubstEnv/IdEnv.
+ -- (see point 3 of the module overview comment).
-- We also use these envs when making a variable polymorphic
-- because we want to float it out past a big lambda.
--
- -- The two Envs always implement the same mapping, but the
+ -- The SubstEnv and IdEnv always implement the same mapping, but the
-- SubstEnv maps to CoreExpr and the IdEnv to LevelledExpr
-- Since the range is always a variable or type application,
-- there is never any difference between the two, but sadly
-- the type application repeatedly.
--
-- The domain of the both envs is *pre-cloned* Ids, though
+ --
+ -- The domain of the VarEnv Level is the *post-cloned* Ids
-initialEnv :: LevelEnv
-initialEnv = (emptyVarEnv, emptySubstEnv, emptyVarEnv)
+initialEnv :: FloatOutSwitches -> LevelEnv
+initialEnv float_lams = (float_lams, emptyVarEnv, emptySubst, emptyVarEnv)
-extendLvlEnv :: LevelEnv -> [(Var,Level)] -> LevelEnv
- -- Used when *not* cloning
-extendLvlEnv (lvl_env, subst_env, id_env) prs
- = (foldl add lvl_env prs, subst_env, id_env)
+floatLams :: LevelEnv -> Bool
+floatLams (FloatOutSw float_lams _, _, _, _) = float_lams
+
+floatConsts :: LevelEnv -> Bool
+floatConsts (FloatOutSw _ float_consts, _, _, _) = float_consts
+
+extendLvlEnv :: LevelEnv -> [TaggedBndr Level] -> LevelEnv
+-- Used when *not* cloning
+extendLvlEnv (float_lams, lvl_env, subst, id_env) prs
+ = (float_lams,
+ foldl add_lvl lvl_env prs,
+ foldl del_subst subst prs,
+ foldl del_id id_env prs)
where
- add env (v,l) = extendVarEnv env v l
+ add_lvl env (TB v l) = extendVarEnv env v l
+ del_subst env (TB v _) = extendInScope env v
+ del_id env (TB v _) = delVarEnv env v
+ -- We must remove any clone for this variable name in case of
+ -- shadowing. This bit me in the following case
+ -- (in nofib/real/gg/Spark.hs):
+ --
+ -- case ds of wild {
+ -- ... -> case e of wild {
+ -- ... -> ... wild ...
+ -- }
+ -- }
+ --
+ -- The inside occurrence of @wild@ was being replaced with @ds@,
+ -- incorrectly, because the SubstEnv was still lying around. Ouch!
+ -- KSW 2000-07.
-- extendCaseBndrLvlEnv adds the mapping case-bndr->scrut-var if it can
-extendCaseBndrLvlEnv (lvl_env, subst_env, id_env) scrut case_bndr lvl
- = case scrut of
- Var v -> (new_lvl_env, extendSubstEnv subst_env case_bndr (DoneEx (Var v)),
- extendVarEnv id_env case_bndr ([], scrut))
- other -> (new_lvl_env, subst_env, id_env)
+-- (see point 4 of the module overview comment)
+extendCaseBndrLvlEnv (float_lams, lvl_env, subst, id_env) (Var scrut_var) case_bndr lvl
+ = (float_lams,
+ extendVarEnv lvl_env case_bndr lvl,
+ extendIdSubst subst case_bndr (Var scrut_var),
+ extendVarEnv id_env case_bndr ([scrut_var], Var scrut_var))
+
+extendCaseBndrLvlEnv env scrut case_bndr lvl
+ = extendLvlEnv env [TB case_bndr lvl]
+
+extendPolyLvlEnv dest_lvl (float_lams, lvl_env, subst, id_env) abs_vars bndr_pairs
+ = (float_lams,
+ foldl add_lvl lvl_env bndr_pairs,
+ foldl add_subst subst bndr_pairs,
+ foldl add_id id_env bndr_pairs)
+ where
+ add_lvl env (v,v') = extendVarEnv env v' dest_lvl
+ add_subst env (v,v') = extendIdSubst env v (mkVarApps (Var v') abs_vars)
+ add_id env (v,v') = extendVarEnv env v ((v':abs_vars), mkVarApps (Var v') abs_vars)
+
+extendCloneLvlEnv lvl (float_lams, lvl_env, _, id_env) new_subst bndr_pairs
+ = (float_lams,
+ foldl add_lvl lvl_env bndr_pairs,
+ new_subst,
+ foldl add_id id_env bndr_pairs)
where
- new_lvl_env = extendVarEnv lvl_env case_bndr lvl
+ add_lvl env (v,v') = extendVarEnv env v' lvl
+ add_id env (v,v') = extendVarEnv env v ([v'], Var v')
-extendPolyLvlEnv (lvl_env, subst_env, id_env) dest_lvl tyvars bndr_pairs
- = (foldl add_lvl lvl_env bndr_pairs,
- foldl add_subst subst_env bndr_pairs,
- foldl add_id id_env bndr_pairs)
+
+maxIdLevel :: LevelEnv -> VarSet -> Level
+maxIdLevel (_, lvl_env,_,id_env) var_set
+ = foldVarSet max_in tOP_LEVEL var_set
where
- add_lvl env (v,_ ) = extendVarEnv env v dest_lvl
- add_subst env (v,v') = extendSubstEnv env v (DoneEx (mkTyVarApps v' tyvars))
- add_id env (v,v') = extendVarEnv env v (tyvars, mkTyVarApps v' tyvars)
+ max_in in_var lvl = foldr max_out lvl (case lookupVarEnv id_env in_var of
+ Just (abs_vars, _) -> abs_vars
+ Nothing -> [in_var])
-varLevel :: LevelEnv -> IdOrTyVar -> Level
-varLevel (lvl_env, _, _) v
- = case lookupVarEnv lvl_env v of
- Just level -> level
- Nothing -> tOP_LEVEL
+ max_out out_var lvl
+ | isId out_var = case lookupVarEnv lvl_env out_var of
+ Just lvl' -> maxLvl lvl' lvl
+ Nothing -> lvl
+ | otherwise = lvl -- Ignore tyvars in *maxIdLevel*
lookupVar :: LevelEnv -> Id -> LevelledExpr
-lookupVar (_, _, id_env) v = case lookupVarEnv id_env v of
- Just (_, expr) -> expr
- other -> Var v
-
-lookupTyVars :: LevelEnv -> Id -> [TyVar]
-lookupTyVars (_, _, id_env) v = case lookupVarEnv id_env v of
- Just (tyvars, _) -> tyvars
- Nothing -> []
+lookupVar (_, _, _, id_env) v = case lookupVarEnv id_env v of
+ Just (_, expr) -> expr
+ other -> Var v
+
+abstractVars :: Level -> LevelEnv -> VarSet -> [Var]
+ -- Find the variables in fvs, free vars of the target expresion,
+ -- whose level is greater than the destination level
+ -- These are the ones we are going to abstract out
+abstractVars dest_lvl env fvs
+ = uniq (sortLe le [var | fv <- varSetElems fvs, var <- absVarsOf dest_lvl env fv])
+ where
+ -- Sort the variables so we don't get
+ -- mixed-up tyvars and Ids; it's just messy
+ v1 `le` v2 = case (isId v1, isId v2) of
+ (True, False) -> False
+ (False, True) -> True
+ other -> v1 <= v2 -- Same family
+
+ uniq :: [Var] -> [Var]
+ -- Remove adjacent duplicates; the sort will have brought them together
+ uniq (v1:v2:vs) | v1 == v2 = uniq (v2:vs)
+ | otherwise = v1 : uniq (v2:vs)
+ uniq vs = vs
+
+absVarsOf :: Level -> LevelEnv -> Var -> [Var]
+ -- If f is free in the expression, and f maps to poly_f a b c in the
+ -- current substitution, then we must report a b c as candidate type
+ -- variables
+absVarsOf dest_lvl (_, lvl_env, _, id_env) v
+ | isId v
+ = [zap av2 | av1 <- lookup_avs v, av2 <- add_tyvars av1, abstract_me av2]
+
+ | otherwise
+ = if abstract_me v then [v] else []
+
+ where
+ abstract_me v = case lookupVarEnv lvl_env v of
+ Just lvl -> dest_lvl `ltLvl` lvl
+ Nothing -> False
+
+ lookup_avs v = case lookupVarEnv id_env v of
+ Just (abs_vars, _) -> abs_vars
+ Nothing -> [v]
+
+ add_tyvars v | isId v = v : varSetElems (idFreeTyVars v)
+ | otherwise = [v]
+
+ -- We are going to lambda-abstract, so nuke any IdInfo,
+ -- and add the tyvars of the Id (if necessary)
+ zap v | isId v = WARN( workerExists (idWorkerInfo v) ||
+ not (isEmptySpecInfo (idSpecialisation v)),
+ text "absVarsOf: discarding info on" <+> ppr v )
+ setIdInfo v vanillaIdInfo
+ | otherwise = v
\end{code}
\begin{code}
\end{code}
\begin{code}
-newLvlVar :: String -> Type -> LvlM Id
-newLvlVar str ty = getUniqueUs `thenLvl` \ uniq ->
- returnUs (mkSysLocal (_PK_ str) uniq ty)
+newPolyBndrs dest_lvl env abs_vars bndrs
+ = getUniquesUs `thenLvl` \ uniqs ->
+ let
+ new_bndrs = zipWith mk_poly_bndr bndrs uniqs
+ in
+ returnLvl (extendPolyLvlEnv dest_lvl env abs_vars (bndrs `zip` new_bndrs), new_bndrs)
+ where
+ mk_poly_bndr bndr uniq = mkSysLocal (mkFastString str) uniq poly_ty
+ where
+ str = "poly_" ++ occNameString (getOccName bndr)
+ poly_ty = mkPiTypes abs_vars (idType bndr)
+
+newLvlVar :: String
+ -> [CoreBndr] -> Type -- Abstract wrt these bndrs
+ -> LvlM Id
+newLvlVar str vars body_ty
+ = getUniqueUs `thenLvl` \ uniq ->
+ returnUs (mkSysLocal (mkFastString str) uniq (mkPiTypes vars body_ty))
+
-- The deeply tiresome thing is that we have to apply the substitution
-- to the rules inside each Id. Grr. But it matters.
-cloneVar :: TopLevelFlag -> LevelEnv -> Id -> Level -> LvlM (LevelEnv, Id)
-cloneVar TopLevel env v lvl
+cloneVar :: TopLevelFlag -> LevelEnv -> Id -> Level -> Level -> LvlM (LevelEnv, Id)
+cloneVar TopLevel env v ctxt_lvl dest_lvl
= returnUs (env, v) -- Don't clone top level things
-cloneVar NotTopLevel (lvl_env, subst_env, id_env) v lvl
- = getUniqueUs `thenLvl` \ uniq ->
+cloneVar NotTopLevel env@(_,_,subst,_) v ctxt_lvl dest_lvl
+ = ASSERT( isId v )
+ getUs `thenLvl` \ us ->
let
- subst = mkSubst emptyVarSet subst_env
- v' = setVarUnique v uniq
- v'' = modifyIdInfo (\info -> substIdInfo subst info info) v'
- subst_env' = extendSubstEnv subst_env v (DoneEx (Var v''))
- id_env' = extendVarEnv id_env v ([], Var v'')
- lvl_env' = extendVarEnv lvl_env v lvl
+ (subst', v1) = cloneIdBndr subst us v
+ v2 = zap_demand ctxt_lvl dest_lvl v1
+ env' = extendCloneLvlEnv dest_lvl env subst' [(v,v2)]
in
- returnUs ((lvl_env', subst_env', id_env'), v'')
+ returnUs (env', v2)
-cloneVars :: TopLevelFlag -> LevelEnv -> [Id] -> Level -> LvlM (LevelEnv, [Id])
-cloneVars TopLevel env vs lvl
+cloneRecVars :: TopLevelFlag -> LevelEnv -> [Id] -> Level -> Level -> LvlM (LevelEnv, [Id])
+cloneRecVars TopLevel env vs ctxt_lvl dest_lvl
= returnUs (env, vs) -- Don't clone top level things
-cloneVars NotTopLevel (lvl_env, subst_env, id_env) vs lvl
- = getUniquesUs (length vs) `thenLvl` \ uniqs ->
+cloneRecVars NotTopLevel env@(_,_,subst,_) vs ctxt_lvl dest_lvl
+ = ASSERT( all isId vs )
+ getUs `thenLvl` \ us ->
let
- subst = mkSubst emptyVarSet subst_env'
- vs' = zipWith setVarUnique vs uniqs
- vs'' = map (modifyIdInfo (\info -> substIdInfo subst info info)) vs'
- subst_env' = extendSubstEnvList subst_env vs [DoneEx (Var v'') | v'' <- vs'']
- id_env' = extendVarEnvList id_env (vs `zip` [([], Var v') | v' <- vs''])
- lvl_env' = extendVarEnvList lvl_env (vs `zip` repeat lvl)
+ (subst', vs1) = cloneRecIdBndrs subst us vs
+ vs2 = map (zap_demand ctxt_lvl dest_lvl) vs1
+ env' = extendCloneLvlEnv dest_lvl env subst' (vs `zip` vs2)
in
- returnUs ((lvl_env', subst_env', id_env'), vs'')
+ returnUs (env', vs2)
-mkTyVarApps var tyvars = foldl (\e tv -> App e (Type (mkTyVarTy tv)))
- (Var var) tyvars
+ -- VERY IMPORTANT: we must zap the demand info
+ -- if the thing is going to float out past a lambda
+zap_demand dest_lvl ctxt_lvl id
+ | ctxt_lvl == dest_lvl = id -- Stays put
+ | otherwise = zapDemandIdInfo id -- Floats out
\end{code}
+