X-Git-Url: http://git.megacz.com/?a=blobdiff_plain;f=ghc%2Fcompiler%2Fstranal%2FDmdAnal.lhs;h=3759fe7ef9ff58e50c60b5109f8600e6f2782d60;hb=5ad7caa537b7c7d873308f4548f7c40164b65074;hp=3f84afd564c9ce8a175a8ab3136a45a34584a21a;hpb=dd27c95844ede8ea93b871eb7f0489f09f5020ff;p=ghc-hetmet.git diff --git a/ghc/compiler/stranal/DmdAnal.lhs b/ghc/compiler/stranal/DmdAnal.lhs index 3f84afd..3759fe7 100644 --- a/ghc/compiler/stranal/DmdAnal.lhs +++ b/ghc/compiler/stranal/DmdAnal.lhs @@ -20,19 +20,26 @@ import PprCore import CoreUtils ( exprIsValue, exprArity ) import DataCon ( dataConTyCon ) import TyCon ( isProductTyCon, isRecursiveTyCon ) -import Id ( Id, idType, idDemandInfo, - isDataConId, isImplicitId, isGlobalId, - idNewStrictness, idNewStrictness_maybe, getNewStrictness, setIdNewStrictness, - idNewDemandInfo, setIdNewDemandInfo, newStrictnessFromOld ) -import IdInfo ( newDemand ) +import Id ( Id, idType, idInlinePragma, + isDataConId, isGlobalId, idArity, +#ifdef DEBUG + idDemandInfo, idStrictness, idCprInfo, +#endif + idNewStrictness, idNewStrictness_maybe, + setIdNewStrictness, idNewDemandInfo, + setIdNewDemandInfo, idName + ) +#ifdef DEBUG +import IdInfo ( newStrictnessFromOld, newDemand ) +#endif import Var ( Var ) import VarEnv import UniqFM ( plusUFM_C, addToUFM_Directly, lookupUFM_Directly, keysUFM, minusUFM, ufmToList, filterUFM ) import Type ( isUnLiftedType ) import CoreLint ( showPass, endPass ) -import Util ( mapAndUnzip, mapAccumL, mapAccumR ) -import BasicTypes ( Arity, TopLevelFlag(..), isTopLevel ) +import Util ( mapAndUnzip, mapAccumL, mapAccumR, lengthIs ) +import BasicTypes ( Arity, TopLevelFlag(..), isTopLevel, isNeverActive ) import Maybes ( orElse, expectJust ) import Outputable \end{code} @@ -60,12 +67,13 @@ dmdAnalPgm :: DynFlags -> [CoreBind] -> IO [CoreBind] dmdAnalPgm dflags binds = do { showPass dflags "Demand analysis" ; - let { binds_plus_dmds = do_prog binds ; - dmd_changes = get_changes binds_plus_dmds } ; + let { binds_plus_dmds = do_prog binds } ; endPass dflags "Demand analysis" Opt_D_dump_stranal binds_plus_dmds ; #ifdef DEBUG - -- Only if DEBUG is on, because only then is the old strictness analyser run + -- Only if DEBUG is on, because only then is the old + -- strictness analyser run + let { dmd_changes = get_changes binds_plus_dmds } ; printDump (text "Changes in demands" $$ dmd_changes) ; #endif return binds_plus_dmds @@ -78,17 +86,18 @@ dmdAnalTopBind :: SigEnv -> CoreBind -> (SigEnv, CoreBind) dmdAnalTopBind sigs (NonRec id rhs) - | isImplicitId id -- Don't touch the info on constructors, selectors etc - = (sigs, NonRec id rhs) -- It's pre-computed in MkId.lhs - | otherwise = let - (sigs', _, (id', rhs')) = downRhs TopLevel sigs (id, rhs) + ( _, _, (_, rhs1)) = dmdAnalRhs TopLevel sigs (id, rhs) + (sigs2, _, (id2, rhs2)) = dmdAnalRhs TopLevel sigs (id, rhs1) + -- Do two passes to improve CPR information + -- See the comments with mkSigTy.ignore_cpr_info below in - (sigs', NonRec id' rhs') + (sigs2, NonRec id2 rhs2) dmdAnalTopBind sigs (Rec pairs) = let (sigs', _, pairs') = dmdFix TopLevel sigs pairs + -- We get two iterations automatically in (sigs', Rec pairs') \end{code} @@ -103,7 +112,7 @@ dmdAnalTopRhs rhs where arity = exprArity rhs (rhs_ty, rhs') = dmdAnal emptySigEnv (vanillaCall arity) rhs - (_, sig) = mkSigTy rhs rhs_ty + sig = mkTopSigTy rhs rhs_ty \end{code} %************************************************************************ @@ -116,12 +125,13 @@ dmdAnalTopRhs rhs dmdAnal :: SigEnv -> Demand -> CoreExpr -> (DmdType, CoreExpr) dmdAnal sigs Abs e = (topDmdType, e) -dmdAnal sigs Bot e = (botDmdType, e) -dmdAnal sigs Lazy e = let - (res_ty, e') = dmdAnal sigs Eval e - in - (deferType res_ty, e') +dmdAnal sigs dmd e + | not (isStrictDmd dmd) + = let + (res_ty, e') = dmdAnal sigs evalDmd e + in + (deferType res_ty, e') -- It's important not to analyse e with a lazy demand because -- a) When we encounter case s of (a,b) -> -- we demand s with U(d1d2)... but if the overall demand is lazy @@ -132,6 +142,9 @@ dmdAnal sigs Lazy e = let -- We still want to mark x as demanded, because it will be when we -- enter the let. If we analyse f's arg with a Lazy demand, we'll -- just mark x as Lazy + -- c) The application rule wouldn't be right either + -- Evaluating (f x) in a L demand does *not* cause + -- evaluation of f in a C(L) demand! dmdAnal sigs dmd (Lit lit) @@ -145,18 +158,20 @@ dmdAnal sigs dmd (Note n e) where (dmd_ty, e') = dmdAnal sigs dmd' e dmd' = case n of - Coerce _ _ -> Eval -- This coerce usually arises from a recursive - other -> dmd -- newtype, and we don't want to look inside them - -- for exactly the same reason that we don't look - -- inside recursive products -- we might not reach - -- a fixpoint. So revert to a vanilla Eval demand + Coerce _ _ -> evalDmd -- This coerce usually arises from a recursive + other -> dmd -- newtype, and we don't want to look inside them + -- for exactly the same reason that we don't look + -- inside recursive products -- we might not reach + -- a fixpoint. So revert to a vanilla Eval demand dmdAnal sigs dmd (App fun (Type ty)) = (fun_ty, App fun' (Type ty)) where (fun_ty, fun') = dmdAnal sigs dmd fun -dmdAnal sigs dmd (App fun arg) -- Non-type arguments +-- Lots of the other code is there to make this +-- beautiful, compositional, application rule :-) +dmdAnal sigs dmd e@(App fun arg) -- Non-type arguments = let -- [Type arg handled above] (fun_ty, fun') = dmdAnal sigs (Call dmd) fun (arg_ty, arg') = dmdAnal sigs arg_dmd arg @@ -180,7 +195,7 @@ dmdAnal sigs dmd (Lam var body) | otherwise -- Not enough demand on the lambda; but do the body = let -- anyway to annotate it and gather free var info - (body_ty, body') = dmdAnal sigs Eval body + (body_ty, body') = dmdAnal sigs evalDmd body (lam_ty, var') = annotateLamIdBndr body_ty var in (deferType lam_ty, Lam var' body') @@ -190,9 +205,21 @@ dmdAnal sigs dmd (Case scrut case_bndr [alt@(DataAlt dc,bndrs,rhs)]) isProductTyCon tycon, not (isRecursiveTyCon tycon) = let - (alt_ty, alt') = dmdAnalAlt sigs dmd alt - (alt_ty1, case_bndr') = annotateBndr alt_ty case_bndr - (_, bndrs', _) = alt' + sigs_alt = extendSigEnv NotTopLevel sigs case_bndr case_bndr_sig + (alt_ty, alt') = dmdAnalAlt sigs_alt dmd alt + (alt_ty1, case_bndr') = annotateBndr alt_ty case_bndr + (_, bndrs', _) = alt' + case_bndr_sig = StrictSig (mkDmdType emptyVarEnv [] RetCPR) + -- Inside the alternative, the case binder has the CPR property. + -- Meaning that a case on it will successfully cancel. + -- Example: + -- f True x = case x of y { I# x' -> if x' ==# 3 then y else I# 8 } + -- f False x = I# 3 + -- + -- We want f to have the CPR property: + -- f b x = case fw b x of { r -> I# r } + -- fw True x = case x of y { I# x' -> if x' ==# 3 then x' else 8 } + -- fw False x = 3 -- Figure out whether the demand on the case binder is used, and use -- that to set the scrut_dmd. This is utterly essential. @@ -213,18 +240,18 @@ dmdAnal sigs dmd (Case scrut case_bndr [alt@(DataAlt dc,bndrs,rhs)]) -- The insight is, of course, that a demand on y is a demand on the -- scrutinee, so we need to `both` it with the scrut demand - scrut_dmd = mkSeq Drop [idNewDemandInfo b | b <- bndrs', isId b] + scrut_dmd = Eval (Prod [idNewDemandInfo b | b <- bndrs', isId b]) `both` - idNewDemandInfo case_bndr' + idNewDemandInfo case_bndr' - (scrut_ty, scrut') = dmdAnal sigs scrut_dmd scrut + (scrut_ty, scrut') = dmdAnal sigs scrut_dmd scrut in (alt_ty1 `bothType` scrut_ty, Case scrut' case_bndr' [alt']) dmdAnal sigs dmd (Case scrut case_bndr alts) = let (alt_tys, alts') = mapAndUnzip (dmdAnalAlt sigs dmd) alts - (scrut_ty, scrut') = dmdAnal sigs Eval scrut + (scrut_ty, scrut') = dmdAnal sigs evalDmd scrut (alt_ty, case_bndr') = annotateBndr (foldr1 lubType alt_tys) case_bndr in -- pprTrace "dmdAnal:Case" (ppr alts $$ ppr alt_tys) @@ -232,12 +259,23 @@ dmdAnal sigs dmd (Case scrut case_bndr alts) dmdAnal sigs dmd (Let (NonRec id rhs) body) = let - (sigs', lazy_fv, (id1, rhs')) = downRhs NotTopLevel sigs (id, rhs) + (sigs', lazy_fv, (id1, rhs')) = dmdAnalRhs NotTopLevel sigs (id, rhs) (body_ty, body') = dmdAnal sigs' dmd body (body_ty1, id2) = annotateBndr body_ty id1 body_ty2 = addLazyFVs body_ty1 lazy_fv in --- pprTrace "dmdLet" (ppr id <+> ppr (sig,rhs_env)) +#ifdef DEBUG + -- If the actual demand is better than the vanilla + -- demand, we might do better to re-analyse with the + -- stronger demand. + (let vanilla_dmd = vanillaCall (idArity id) + actual_dmd = idNewDemandInfo id2 + in + if actual_dmd `betterDemand` vanilla_dmd && actual_dmd /= vanilla_dmd then + pprTrace "dmdLet: better demand" (ppr id <+> vcat [text "vanilla" <+> ppr vanilla_dmd, + text "actual" <+> ppr actual_dmd]) + else \x -> x) +#endif (body_ty2, Let (NonRec id2 rhs') body') dmdAnal sigs dmd (Let (Rec pairs) body) @@ -315,7 +353,7 @@ dmdFix top_lvl sigs orig_pairs ((sigs', lazy_fv'), pair') -- ) where - (sigs', lazy_fv1, pair') = downRhs top_lvl sigs (id,rhs) + (sigs', lazy_fv1, pair') = dmdAnalRhs top_lvl sigs (id,rhs) lazy_fv' = plusUFM_C both lazy_fv lazy_fv1 -- old_sig = lookup sigs id -- new_sig = lookup sigs' id @@ -331,21 +369,22 @@ dmdFix top_lvl sigs orig_pairs lookup sigs var = case lookupVarEnv sigs var of Just (sig,_) -> sig -downRhs :: TopLevelFlag +dmdAnalRhs :: TopLevelFlag -> SigEnv -> (Id, CoreExpr) -> (SigEnv, DmdEnv, (Id, CoreExpr)) -- Process the RHS of the binding, add the strictness signature -- to the Id, and augment the environment with the signature as well. -downRhs top_lvl sigs (id, rhs) +dmdAnalRhs top_lvl sigs (id, rhs) = (sigs', lazy_fv, (id', rhs')) where - arity = exprArity rhs -- The idArity may not be up to date - (rhs_ty, rhs') = dmdAnal sigs (vanillaCall arity) rhs - (lazy_fv, sig_ty) = WARN( arity /= dmdTypeDepth rhs_ty, ppr id ) - mkSigTy rhs rhs_ty - id' = id `setIdNewStrictness` sig_ty - sigs' = extendSigEnv top_lvl sigs id sig_ty + arity = idArity id -- The idArity should be up to date + -- The simplifier was run just beforehand + (rhs_dmd_ty, rhs') = dmdAnal sigs (vanillaCall arity) rhs + (lazy_fv, sig_ty) = WARN( arity /= dmdTypeDepth rhs_dmd_ty, ppr id ) + mkSigTy id rhs rhs_dmd_ty + id' = id `setIdNewStrictness` sig_ty + sigs' = extendSigEnv top_lvl sigs id sig_ty \end{code} %************************************************************************ @@ -355,9 +394,45 @@ downRhs top_lvl sigs (id, rhs) %************************************************************************ \begin{code} -mkSigTy :: CoreExpr -> DmdType -> (DmdEnv, StrictSig) --- Take a DmdType and turn it into a StrictSig -mkSigTy rhs (DmdType fv dmds res) +mkTopSigTy :: CoreExpr -> DmdType -> StrictSig + -- Take a DmdType and turn it into a StrictSig + -- NB: not used for never-inline things; hence False +mkTopSigTy rhs dmd_ty = snd (mk_sig_ty False False rhs dmd_ty) + +mkSigTy :: Id -> CoreExpr -> DmdType -> (DmdEnv, StrictSig) +mkSigTy id rhs dmd_ty = mk_sig_ty (isNeverActive (idInlinePragma id)) + (isStrictDmd (idNewDemandInfo id)) + rhs dmd_ty + +mk_sig_ty never_inline strictly_demanded rhs (DmdType fv dmds res) + | never_inline && not (isBotRes res) + -- HACK ALERT + -- Don't strictness-analyse NOINLINE things. Why not? Because + -- the NOINLINE says "don't expose any of the inner workings at the call + -- site" and the strictness is certainly an inner working. + -- + -- More concretely, the demand analyser discovers the following strictness + -- for unsafePerformIO: C(U(AV)) + -- But then consider + -- unsafePerformIO (\s -> let r = f x in + -- case writeIORef v r s of (# s1, _ #) -> + -- (# s1, r #) + -- The strictness analyser will find that the binding for r is strict, + -- (becuase of uPIO's strictness sig), and so it'll evaluate it before + -- doing the writeIORef. This actually makes tests/lib/should_run/memo002 + -- get a deadlock! + -- + -- Solution: don't expose the strictness of unsafePerformIO. + -- + -- But we do want to expose the strictness of error functions, + -- which are also often marked NOINLINE + -- {-# NOINLINE foo #-} + -- foo x = error ("wubble buggle" ++ x) + -- So (hack, hack) we only drop the strictness for non-bottom things + -- This is all very unsatisfactory. + = (deferEnv fv, topSig) + + | otherwise = (lazy_fv, mkStrictSig dmd_ty) where dmd_ty = DmdType strict_fv final_dmds res' @@ -394,28 +469,47 @@ mkSigTy rhs (DmdType fv dmds res) -- DmdType, because that makes fixpointing very slow --- the -- DmdType gets full of lazy demands that are slow to converge. - lazified_dmds = map funArgDemand dmds - -- Get rid of defers in the arguments - final_dmds = setUnpackStrategy lazified_dmds + final_dmds = setUnpackStrategy dmds -- Set the unpacking strategy res' = case res of - RetCPR | not (exprIsValue rhs) -> TopRes - other -> res + RetCPR | ignore_cpr_info -> TopRes + other -> res + ignore_cpr_info = is_thunk && not strictly_demanded + is_thunk = not (exprIsValue rhs) -- If the rhs is a thunk, we forget the CPR info, because -- it is presumably shared (else it would have been inlined, and -- so we'd lose sharing if w/w'd it into a function. -- - -- DONE IN OLD CPR ANALYSER, BUT NOT YET HERE - -- Also, if the strictness analyser has figured out that it's strict, - -- the let-to-case transformation will happen, so again it's good. - -- (CPR analysis runs before the simplifier has had a chance to do - -- the let-to-case transform.) + -- Also, if the strictness analyser has figured out (in a previous iteration) + -- that it's strict, the let-to-case transformation will happen, so again + -- it's good. -- This made a big difference to PrelBase.modInt, which had something like -- modInt = \ x -> let r = ... -> I# v in -- ...body strict in r... -- r's RHS isn't a value yet; but modInt returns r in various branches, so -- if r doesn't have the CPR property then neither does modInt + -- Another case I found in practice (in Complex.magnitude), looks like this: + -- let k = if ... then I# a else I# b + -- in ... body strict in k .... + -- (For this example, it doesn't matter whether k is returned as part of + -- the overall result.) Left to itself, the simplifier will make a join + -- point thus: + -- let $j k = ...body strict in k... + -- if ... then $j (I# a) else $j (I# b) + -- + -- + -- The difficulty with this is that we need the strictness type to + -- look at the body... but we now need the body to calculate the demand + -- on the variable, so we can decide whether its strictness type should + -- have a CPR in it or not. Simple solution: + -- a) use strictness info from the previous iteration + -- b) make sure we do at least 2 iterations, by doing a second + -- round for top-level non-recs. Top level recs will get at + -- least 2 iterations except for totally-bottom functions + -- which aren't very interesting anyway. + -- + -- NB: strictly_demanded is never true of a top-level Id, or of a recursive Id. \end{code} The unpack strategy determines whether we'll *really* unpack the argument, @@ -431,15 +525,13 @@ setUnpackStrategy ds -> [Demand] -> (Int, [Demand]) -- Args remaining after subcomponents of [Demand] are unpacked - go n (Seq keep cs : ds) - | n' >= 0 = Seq keep cs' `cons` go n'' ds - | otherwise = Eval `cons` go n ds + go n (Eval (Prod cs) : ds) + | n' >= 0 = Eval (Prod cs') `cons` go n'' ds + | otherwise = Box (Eval (Prod cs)) `cons` go n ds where (n'',cs') = go n' cs - n' = n + box - non_abs_args - box = case keep of - Keep -> 0 - Drop -> 1 -- Add one to the budget if we drop the top-level arg + n' = n + 1 - non_abs_args + -- Add one to the budget 'cos we drop the top-level arg non_abs_args = nonAbsentArgs cs -- Delete # of non-absent args to which we'll now be committed @@ -464,13 +556,10 @@ nonAbsentArgs (d : ds) = 1 + nonAbsentArgs ds \begin{code} splitDmdTy :: DmdType -> (Demand, DmdType) -- Split off one function argument +-- We already have a suitable demand on all +-- free vars, so no need to add more! splitDmdTy (DmdType fv (dmd:dmds) res_ty) = (dmd, DmdType fv dmds res_ty) -splitDmdTy ty@(DmdType fv [] TopRes) = (Lazy, ty) -splitDmdTy ty@(DmdType fv [] BotRes) = (Bot, ty) - -- NB: Bot not Abs -splitDmdTy (DmdType fv [] RetCPR) = panic "splitDmdTy" - -- We already have a suitable demand on all - -- free vars, so no need to add more! +splitDmdTy ty@(DmdType fv [] res_ty) = (resTypeArgDmd res_ty, ty) \end{code} \begin{code} @@ -517,11 +606,9 @@ annotateBndr :: DmdType -> Var -> (DmdType, Var) -- No effect on the argument demands annotateBndr dmd_ty@(DmdType fv ds res) var | isTyVar var = (dmd_ty, var) - | otherwise = (DmdType fv' ds res, setIdNewDemandInfo var hacked_dmd) + | otherwise = (DmdType fv' ds res, setIdNewDemandInfo var dmd) where (fv', dmd) = removeFV fv var res - hacked_dmd | isUnLiftedType (idType var) = unliftedDemand dmd - | otherwise = dmd annotateBndrs = mapAccumR annotateBndr @@ -532,9 +619,8 @@ annotateLamIdBndr dmd_ty@(DmdType fv ds res) id (DmdType fv' (hacked_dmd:ds) res, setIdNewDemandInfo id hacked_dmd) where (fv', dmd) = removeFV fv id res - hacked_dmd | isUnLiftedType (idType id) = unliftedDemand dmd - | otherwise = funArgDemand dmd - -- This call to funArgDemand is vital, because otherwise we label + hacked_dmd = argDemand dmd + -- This call to argDemand is vital, because otherwise we label -- a lambda binder with demand 'B'. But in terms of calling -- conventions that's Abs, because we don't pass it. But -- when we do a w/w split we get @@ -542,12 +628,19 @@ annotateLamIdBndr dmd_ty@(DmdType fv ds res) id -- And then the simplifier things the 'B' is a strict demand -- and evaluates the (error "oops"). Sigh -removeFV fv var res = (fv', dmd) +removeFV fv id res = (fv', zapUnlifted id dmd) where - fv' = fv `delVarEnv` var - dmd = lookupVarEnv fv var `orElse` deflt + fv' = fv `delVarEnv` id + dmd = lookupVarEnv fv id `orElse` deflt deflt | isBotRes res = Bot | otherwise = Abs + +-- For unlifted-type variables, we are only +-- interested in Bot/Abs/Box Abs +zapUnlifted is Bot = Bot +zapUnlifted id Abs = Abs +zapUnlifted id dmd | isUnLiftedType (idType id) = lazyDmd + | otherwise = dmd \end{code} %************************************************************************ @@ -582,23 +675,31 @@ dmdTransform :: SigEnv -- The strictness environment dmdTransform sigs var dmd ------ DATA CONSTRUCTOR - | isDataConId var, -- Data constructor - Seq k ds <- res_dmd, -- and the demand looks inside its fields - let StrictSig dmd_ty = idNewStrictness var, -- It must have a strictness sig - let DmdType _ con_ds con_res = dmd_ty - = if length con_ds == length ds then -- Saturated, so unleash the demand - -- ds can be empty, when we are just seq'ing the thing + | isDataConId var -- Data constructor + = let + StrictSig dmd_ty = idNewStrictness var -- It must have a strictness sig + DmdType _ _ con_res = dmd_ty + arity = idArity var + in + if arity == call_depth then -- Saturated, so unleash the demand let - arg_ds = case k of - Keep -> zipWith lub ds con_ds - Drop -> ds - Defer -> ds -- Important! If we Keep the constructor application, then - -- we need the demands the constructor places (usually lazy) + -- we need the demands the constructor places (always lazy) -- If not, we don't need to. For example: -- f p@(x,y) = (p,y) -- S(AL) -- g a b = f (a,b) -- It's vital that we don't calculate Absent for a! + dmd_ds = case res_dmd of + Box (Eval ds) -> mapDmds box ds + Eval ds -> ds + other -> Poly Top + + -- ds can be empty, when we are just seq'ing the thing + -- If so we must make up a suitable bunch of demands + arg_ds = case dmd_ds of + Poly d -> replicate arity d + Prod ds -> ASSERT( ds `lengthIs` arity ) ds + in mkDmdType emptyDmdEnv arg_ds con_res -- Must remember whether it's a product, hence con_res, not TopRes @@ -607,7 +708,7 @@ dmdTransform sigs var dmd ------ IMPORTED FUNCTION | isGlobalId var, -- Imported function - let StrictSig dmd_ty = getNewStrictness var + let StrictSig dmd_ty = idNewStrictness var = if dmdTypeDepth dmd_ty <= call_depth then -- Saturated, so unleash the demand dmd_ty else @@ -647,36 +748,31 @@ splitCallDmd (Call d) = case splitCallDmd d of splitCallDmd d = (0, d) vanillaCall :: Arity -> Demand -vanillaCall 0 = Eval +vanillaCall 0 = evalDmd vanillaCall n = Call (vanillaCall (n-1)) deferType :: DmdType -> DmdType -deferType (DmdType fv _ _) = DmdType (mapVarEnv defer fv) [] TopRes +deferType (DmdType fv _ _) = DmdType (deferEnv fv) [] TopRes -- Notice that we throw away info about both arguments and results -- For example, f = let ... in \x -> x -- We don't want to get a stricness type V->T for f. + -- Peter?? + +deferEnv :: DmdEnv -> DmdEnv +deferEnv fv = mapVarEnv defer fv ---------------- -bothLazy :: Demand -> Demand -bothLazy = both Lazy -bothLazy_s :: [Demand] -> [Demand] -bothLazy_s = map bothLazy -funArgDemand :: Demand -> Demand +---------------- +argDemand :: Demand -> Demand -- The 'Defer' demands are just Lazy at function boundaries -- Ugly! Ask John how to improve it. -funArgDemand (Seq Defer ds) = Lazy -funArgDemand (Seq k ds) = Seq k (map funArgDemand ds) -funArgDemand Err = Eval -- Args passed to a bottoming function -funArgDemand Bot = Abs -- Don't pass args that are consumed by bottom/err -funArgDemand d = d - -unliftedDemand :: Demand -> Demand --- Same idea, but for unlifted types the domain is much simpler: --- Either we use it (Lazy) or we don't (Abs) -unliftedDemand Bot = Abs -unliftedDemand Abs = Abs -unliftedDemand other = Lazy +argDemand Top = lazyDmd +argDemand (Defer d) = lazyDmd +argDemand (Eval ds) = Eval (mapDmds argDemand ds) +argDemand (Box Bot) = evalDmd +argDemand (Box d) = box (argDemand d) +argDemand Bot = Abs -- Don't pass args that are consumed by bottom/err +argDemand d = d \end{code} \begin{code} @@ -688,8 +784,6 @@ betterDmdType t1 t2 = (t1 `lubType` t2) == t2 betterDemand :: Demand -> Demand -> Bool -- If d1 `better` d2, and d2 `better` d2, then d1==d2 betterDemand d1 d2 = (d1 `lub` d2) == d2 - -squashDmdEnv (StrictSig (DmdType fv ds res)) = StrictSig (DmdType emptyDmdEnv ds res) \end{code} \begin{code} @@ -699,16 +793,25 @@ squashDmdEnv (StrictSig (DmdType fv ds res)) = StrictSig (DmdType emptyDmdEnv ds -- *implicitly* has {y->A}. So we must put {y->(V `lub` A)} -- in the result env. lubType (DmdType fv1 ds1 r1) (DmdType fv2 ds2 r2) - = DmdType lub_fv2 (zipWith lub ds1 ds2) (r1 `lubRes` r2) + = DmdType lub_fv2 (lub_ds ds1 ds2) (r1 `lubRes` r2) where lub_fv = plusUFM_C lub fv1 fv2 - lub_fv1 = modifyEnv (not (isBotRes r1)) defer fv2 fv1 lub_fv - lub_fv2 = modifyEnv (not (isBotRes r2)) defer fv1 fv2 lub_fv1 + lub_fv1 = modifyEnv (not (isBotRes r1)) absLub fv2 fv1 lub_fv + lub_fv2 = modifyEnv (not (isBotRes r2)) absLub fv1 fv2 lub_fv1 -- lub is the identity for Bot + -- Extend the shorter argument list to match the longer + lub_ds (d1:ds1) (d2:ds2) = lub d1 d2 : lub_ds ds1 ds2 + lub_ds [] [] = [] + lub_ds ds1 [] = map (`lub` resTypeArgDmd r2) ds1 + lub_ds [] ds2 = map (resTypeArgDmd r1 `lub`) ds2 + ----------------------------------- -- (t1 `bothType` t2) takes the argument/result info from t1, -- using t2 just for its free-var info +-- NB: Don't forget about r2! It might be BotRes, which is +-- a bottom demand on all the in-scope variables. +-- Peter: can this be done more neatly? bothType (DmdType fv1 ds1 r1) (DmdType fv2 ds2 r2) = DmdType both_fv2 ds1 (r1 `bothRes` r2) where @@ -727,22 +830,8 @@ lubRes r1 r2 = TopRes -- If either diverges, the whole thing does -- Otherwise take CPR info from the first -bothRes BotRes r2 = BotRes -bothRes r1 BotRes = BotRes -bothRes r1 r2 = r1 -\end{code} - -\begin{code} --- A Seq can have an empty list of demands, in the polymorphic case. -lubs [] ds2 = ds2 -lubs ds1 [] = ds1 -lubs ds1 ds2 = ASSERT( length ds1 == length ds2 ) zipWith lub ds1 ds2 - ------------------------------------ --- A Seq can have an empty list of demands, in the polymorphic case. -boths [] ds2 = ds2 -boths ds1 [] = ds1 -boths ds1 ds2 = ASSERT( length ds1 == length ds2 ) zipWith both ds1 ds2 +bothRes r1 BotRes = BotRes +bothRes r1 r2 = r1 \end{code} \begin{code} @@ -769,114 +858,144 @@ modifyEnv need_to_modify zapper env1 env2 env %* * %************************************************************************ - \begin{code} lub :: Demand -> Demand -> Demand -lub Bot d = d - -lub Err Bot = Err -lub Err d = d - -lub Lazy d = Lazy - -lub Abs d = defer d - -lub Eval Abs = Lazy -lub Eval Lazy = Lazy -lub Eval (Seq Drop ds) | not (null ds) = Seq Drop [Lazy | d <- ds] -lub Eval d = Eval - -- For the Seq case, consier - -- f n [] = n - -- f n (x:xs) = f (n+x) xs - -- Here we want to do better than just V for n. It's - -- unboxed in the (x:xs) case, and we might be prepared to - -- rebox it in the [] case. - -- But if we don't use *any* of the components, give up - -- and revert to V - -lub (Call d1) (Call d2) = Call (lub d1 d2) -lub d1@(Call _) d2 = d2 `lub` d1 - -lub (Seq k1 ds1) (Seq k2 ds2) - = Seq (k1 `lub_keep` k2) (lub_ds k1 ds1 k2 ds2) - where - ------------------ - lub_ds Keep ds1 Keep ds2 = ds1 `lubs` ds2 - lub_ds Keep ds1 non_keep ds2 | null ds1 = [Lazy | d <- ds2] - | otherwise = bothLazy_s ds1 `lubs` ds2 - - lub_ds non_keep ds1 Keep ds2 | null ds2 = [Lazy | d <- ds1] - | otherwise = ds1 `lubs` bothLazy_s ds2 - - lub_ds k1 ds1 k2 ds2 = ds1 `lubs` ds2 - - ------------------ - lub_keep Keep k = k - - lub_keep Drop Defer = Defer - lub_keep Drop k = Drop - - lub_keep Defer k = Defer +lub Bot d2 = d2 +lub Abs d2 = absLub d2 +lub Top d2 = Top +lub (Defer ds1) d2 = defer (Eval ds1 `lub` d2) + +lub (Call d1) (Call d2) = Call (d1 `lub` d2) +lub d1@(Call _) (Box d2) = d1 `lub` d2 -- Just strip the box +lub d1@(Call _) d2@(Eval _) = d2 -- Presumably seq or vanilla eval +lub d1@(Call _) d2 = d2 `lub` d1 -- Bot, Abs, Top + +-- For the Eval case, we use these approximation rules +-- Box Bot <= Eval (Box Bot ...) +-- Box Top <= Defer (Box Bot ...) +-- Box (Eval ds) <= Eval (map Box ds) +lub (Eval ds1) (Eval ds2) = Eval (ds1 `lubs` ds2) +lub (Eval ds1) (Box Bot) = Eval (mapDmds (`lub` Box Bot) ds1) +lub (Eval ds1) (Box (Eval ds2)) = Eval (ds1 `lubs` mapDmds box ds2) +lub (Eval ds1) (Box Abs) = deferEval (mapDmds (`lub` Box Bot) ds1) +lub d1@(Eval _) d2 = d2 `lub` d1 -- Bot,Abs,Top,Call,Defer + +lub (Box d1) (Box d2) = box (d1 `lub` d2) +lub d1@(Box _) d2 = d2 `lub` d1 + +lubs = zipWithDmds lub + +--------------------- +-- box is the smart constructor for Box +-- It computes & d +-- INVARIANT: (Box d) => d = Bot, Abs, Eval +-- Seems to be no point in allowing (Box (Call d)) +box (Call d) = Call d -- The odd man out. Why? +box (Box d) = Box d +box (Defer _) = lazyDmd +box Top = lazyDmd -- Box Abs and Box Top +box Abs = lazyDmd -- are the same +box d = Box d -- Bot, Eval -lub d1@(Seq _ _) d2 = d2 `lub` d1 +--------------- +defer :: Demand -> Demand + +-- defer is the smart constructor for Defer +-- The idea is that (Defer ds) = +-- +-- It specifies what happens at a lazy function argument +-- or a lambda; the L* operator +-- Set the strictness part to L, but leave +-- the boxity side unaffected +-- It also ensures that Defer (Eval [LLLL]) = L + +defer Bot = Abs +defer Abs = Abs +defer Top = Top +defer (Call _) = lazyDmd -- Approximation here? +defer (Box _) = lazyDmd +defer (Defer ds) = Defer ds +defer (Eval ds) = deferEval ds + +-- deferEval ds = defer (Eval ds) +deferEval ds | allTop ds = Top + | otherwise = Defer ds + +--------------------- +absLub :: Demand -> Demand +-- Computes (Abs `lub` d) +-- For the Bot case consider +-- f x y = if ... then x else error x +-- Then for y we get Abs `lub` Bot, and we really +-- want Abs overall +absLub Bot = Abs +absLub Abs = Abs +absLub Top = Top +absLub (Call _) = Top +absLub (Box _) = Top +absLub (Eval ds) = Defer (absLubs ds) -- Or (Defer ds)? +absLub (Defer ds) = Defer (absLubs ds) -- Or (Defer ds)? + +absLubs = mapDmds absLub --------------- both :: Demand -> Demand -> Demand -both Bot Bot = Bot -both Bot Abs = Bot -both Bot d = Err - -both Err d = Err - -both Abs d = d - -both Lazy Bot = Err -both Lazy Err = Err -both Lazy Eval = Eval -both Lazy (Call d) = Call d -both Lazy (Seq Defer ds) = Lazy -both Lazy (Seq k ds) = Seq Keep ds -both Lazy d = Lazy - --- For the (Eval `both` Bot) case, consider --- f x = error x --- From 'error' itself we get demand Bot on x --- From the arg demand on x we get Eval --- So we want Eval `both` Bot to be Err. --- That's what Err is *for* -both Eval Bot = Err -both Eval Err = Err -both Eval (Seq k ds) = Seq Keep ds -both Eval d = Eval - -both (Call d1) (Call d2) = Call (d1 `both` d2) -both d1@(Call _) d2 = d2 `both` d1 - -both (Seq k1 ds1) (Seq k2 ds2) - = Seq (k1 `both_keep` k2) (both_ds k1 ds1 k2 ds2) - where - ---------------- - both_keep Keep k2 = Keep - - both_keep Drop Keep = Keep - both_keep Drop k2 = Drop - - both_keep Defer k2 = k2 - - ---------------- - both_ds Defer ds1 Defer ds2 = ds1 `boths` ds2 - both_ds Defer ds1 non_defer ds2 = map defer ds1 `boths` ds2 - - both_ds non_defer ds1 Defer ds2 = ds1 `boths` map defer ds2 - - both_ds k1 ds1 k2 ds2 = ds1 `boths` ds2 - -both d1@(Seq _ _) d2 = d2 `both` d1 +both Abs d2 = d2 + +both Bot Bot = Bot +both Bot Abs = Bot +both Bot (Eval ds) = Eval (mapDmds (`both` Bot) ds) + -- Consider + -- f x = error x + -- From 'error' itself we get demand Bot on x + -- From the arg demand on x we get + -- x :-> evalDmd = Box (Eval (Poly Abs)) + -- So we get Bot `both` Box (Eval (Poly Abs)) + -- = Seq Keep (Poly Bot) + -- + -- Consider also + -- f x = if ... then error (fst x) else fst x + -- Then we get (Eval (Box Bot, Bot) `lub` Eval (SA)) + -- = Eval (SA) + -- which is what we want. +both Bot d = errDmd + +both Top Bot = errDmd +both Top Abs = Top +both Top Top = Top +both Top (Box d) = Box d +both Top (Call d) = Call d +both Top (Eval ds) = Eval (mapDmds (`both` Top) ds) +both Top (Defer ds) -- = defer (Top `both` Eval ds) + -- = defer (Eval (mapDmds (`both` Top) ds)) + = deferEval (mapDmds (`both` Top) ds) + + +both (Box d1) (Box d2) = box (d1 `both` d2) +both (Box d1) d2@(Call _) = box (d1 `both` d2) +both (Box d1) d2@(Eval _) = box (d1 `both` d2) +both (Box d1) (Defer d2) = Box d1 +both d1@(Box _) d2 = d2 `both` d1 + +both (Call d1) (Call d2) = Call (d1 `both` d2) +both (Call d1) (Eval ds2) = Call d1 -- Could do better for (Poly Bot)? +both (Call d1) (Defer ds2) = Call d1 -- Ditto +both d1@(Call _) d2 = d1 `both` d1 + +both (Eval ds1) (Eval ds2) = Eval (ds1 `boths` ds2) +both (Eval ds1) (Defer ds2) = Eval (ds1 `boths` mapDmds defer ds2) +both d1@(Eval ds1) d2 = d2 `both` d1 + +both (Defer ds1) (Defer ds2) = deferEval (ds1 `boths` ds2) +both d1@(Defer ds1) d2 = d2 `both` d1 + +boths = zipWithDmds both \end{code} + %************************************************************************ %* * \subsection{Miscellaneous @@ -885,14 +1004,14 @@ both d1@(Seq _ _) d2 = d2 `both` d1 \begin{code} +#ifdef DEBUG get_changes binds = vcat (map get_changes_bind binds) get_changes_bind (Rec pairs) = vcat (map get_changes_pr pairs) get_changes_bind (NonRec id rhs) = get_changes_pr (id,rhs) get_changes_pr (id,rhs) - | isImplicitId id = empty -- We don't look inside these - | otherwise = get_changes_var id $$ get_changes_expr rhs + = get_changes_var id $$ get_changes_expr rhs get_changes_var var | isId var = get_changes_str var $$ get_changes_dmd var @@ -917,8 +1036,9 @@ get_changes_str id where message word = text word <+> text "strictness for" <+> ppr id <+> info info = (text "Old" <+> ppr old) $$ (text "New" <+> ppr new) - new = squashDmdEnv (idNewStrictness id) -- Don't report diffs in the env - old = newStrictnessFromOld id + new = squashSig (idNewStrictness id) -- Don't report spurious diffs that the old + -- strictness analyser can't track + old = newStrictnessFromOld (idName id) (idArity id) (idStrictness id) (idCprInfo id) old_better = old `betterStrictness` new new_better = new `betterStrictness` old @@ -931,8 +1051,21 @@ get_changes_dmd id where message word = text word <+> text "demand for" <+> ppr id <+> info info = (text "Old" <+> ppr old) $$ (text "New" <+> ppr new) - new = funArgDemand (idNewDemandInfo id) -- FunArgDemand to avoid spurious improvements + new = squashDmd (argDemand (idNewDemandInfo id)) -- To avoid spurious improvements + -- A bit of a hack old = newDemand (idDemandInfo id) new_better = new `betterDemand` old old_better = old `betterDemand` new +#endif + +squashSig (StrictSig (DmdType fv ds res)) + = StrictSig (DmdType emptyDmdEnv (map squashDmd ds) res) + where + -- squash just gets rid of call demands + -- which the old analyser doesn't track +squashDmd (Call d) = evalDmd +squashDmd (Box d) = Box (squashDmd d) +squashDmd (Eval ds) = Eval (mapDmds squashDmd ds) +squashDmd (Defer ds) = Defer (mapDmds squashDmd ds) +squashDmd d = d \end{code}