X-Git-Url: http://git.megacz.com/?a=blobdiff_plain;f=ghc%2Fcompiler%2FsimplCore%2FSetLevels.lhs;h=f8ab29dcd593189f77eb18929143305d51ae4b5a;hb=9d7da331989abcd1844e9d03b8d1e4163796fa85;hp=2ff47547dbb1b79a7f364c9387e7c7db7a92849d;hpb=30b5ebe424ebae69b162ac3fc547eb14d898535f;p=ghc-hetmet.git diff --git a/ghc/compiler/simplCore/SetLevels.lhs b/ghc/compiler/simplCore/SetLevels.lhs index 2ff4754..f8ab29d 100644 --- a/ghc/compiler/simplCore/SetLevels.lhs +++ b/ghc/compiler/simplCore/SetLevels.lhs @@ -7,62 +7,75 @@ 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} %************************************************************************ @@ -72,7 +85,9 @@ import List ( nub ) %************************************************************************ \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} @@ -97,43 +112,95 @@ The main function @lvlExpr@ carries a ``context level'' (@ctxt_lvl@). 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 "" 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} @@ -141,11 +208,12 @@ instance Outputable Level where %************************************************************************ \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 @@ -155,16 +223,18 @@ setLevels binds us 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} %************************************************************************ @@ -196,21 +266,24 @@ don't want @lvlExpr@ to turn the scrutinee of the @case@ into an MFE 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) @@ -225,61 +298,69 @@ 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 @@ -290,34 +371,62 @@ lvlMFE :: Bool -- True <=> strict context [body of case or let] 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} @@ -338,54 +447,75 @@ lvlBind :: TopLevelFlag -- Used solely to decide whether to clone -> 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 @@ -396,20 +526,18 @@ lvlBind top_lvl ctxt_lvl env (AnnRec 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} @@ -420,45 +548,65 @@ lvl_poly_rhs tyvars_w_lvls rhs_lvl rhs_env rhs %************************************************************************ \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} @@ -469,13 +617,18 @@ maxIdLvl (lvl_env,_,_) var lvl | isTyVar var = lvl %************************************************************************ \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 @@ -487,50 +640,143 @@ type LevelEnv = (VarEnv Level, SubstEnv, IdEnv ([TyVar], LevelledExpr)) -- 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} @@ -543,43 +789,59 @@ mapLvl = mapUs \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} +