X-Git-Url: http://git.megacz.com/?p=ghc-hetmet.git;a=blobdiff_plain;f=compiler%2Fspecialise%2FSpecConstr.lhs;h=b8394126055761237b636a3c6292fc23a6a0d993;hp=f80b3205b8f89326b3a0b61f59e2d2a101981c3b;hb=90686adf9d3dc7a09a51853df051bc4ea472d840;hpb=44d4bf2c3eff873d18e683c0629f17a228e9dfe0 diff --git a/compiler/specialise/SpecConstr.lhs b/compiler/specialise/SpecConstr.lhs index f80b320..b839412 100644 --- a/compiler/specialise/SpecConstr.lhs +++ b/compiler/specialise/SpecConstr.lhs @@ -11,7 +11,7 @@ -- for details module SpecConstr( - specConstrProgram + specConstrProgram, SpecConstrAnnotation(..) ) where #include "HsVersions.h" @@ -20,33 +20,42 @@ import CoreSyn import CoreSubst import CoreUtils import CoreUnfold ( couldBeSmallEnoughToInline ) -import CoreLint ( showPass, endPass ) import CoreFVs ( exprsFreeVars ) -import CoreTidy ( tidyRules ) -import PprCore ( pprRules ) +import CoreMonad +import HscTypes ( ModGuts(..) ) import WwLib ( mkWorkerArgs ) -import DataCon ( dataConRepArity, dataConUnivTyVars ) +import DataCon ( dataConTyCon, dataConRepArity, dataConUnivTyVars ) +import TyCon ( TyCon ) +import Literal ( literalType ) import Coercion +import Rules import Type hiding( substTy ) -import Id ( Id, idName, idType, isDataConWorkId_maybe, idArity, - mkUserLocal, mkSysLocal, idUnfolding, isLocalId ) +import Id +import MkId ( mkImpossibleExpr ) import Var import VarEnv import VarSet import Name -import Rules ( addIdSpecialisations, mkLocalRule, rulesOfBinds ) -import OccName ( mkSpecOcc ) -import ErrUtils ( dumpIfSet_dyn ) -import DynFlags ( DynFlags(..), DynFlag(..) ) -import StaticFlags ( opt_SpecInlineJoinPoints ) -import BasicTypes ( Activation(..) ) -import Maybes ( orElse, catMaybes, isJust ) +import DynFlags ( DynFlags(..) ) +import StaticFlags ( opt_PprStyle_Debug ) +import Maybes ( orElse, catMaybes, isJust, isNothing ) +import Demand +import DmdAnal ( both ) +import Serialized ( deserializeWithData ) import Util -import List ( nubBy, partition ) import UniqSupply import Outputable import FastString import UniqFM +import qualified LazyUniqFM as L +import MonadUtils +import Control.Monad ( zipWithM ) +import Data.List +#if __GLASGOW_HASKELL__ > 609 +import Data.Data ( Data, Typeable ) +#else +import Data.Generics ( Data, Typeable ) +#endif \end{code} ----------------------------------------------------- @@ -343,6 +352,42 @@ The recursive call ends up looking like So we want to spot the construtor application inside the cast. That's why we have the Cast case in argToPat +Note [Local recursive groups] +~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ +For a *local* recursive group, we can see all the calls to the +function, so we seed the specialisation loop from the calls in the +body, not from the calls in the RHS. Consider: + + bar m n = foo n (n,n) (n,n) (n,n) (n,n) + where + foo n p q r s + | n == 0 = m + | n > 3000 = case p of { (p1,p2) -> foo (n-1) (p2,p1) q r s } + | n > 2000 = case q of { (q1,q2) -> foo (n-1) p (q2,q1) r s } + | n > 1000 = case r of { (r1,r2) -> foo (n-1) p q (r2,r1) s } + | otherwise = case s of { (s1,s2) -> foo (n-1) p q r (s2,s1) } + +If we start with the RHSs of 'foo', we get lots and lots of specialisations, +most of which are not needed. But if we start with the (single) call +in the rhs of 'bar' we get exactly one fully-specialised copy, and all +the recursive calls go to this fully-specialised copy. Indeed, the original +function is later collected as dead code. This is very important in +specialising the loops arising from stream fusion, for example in NDP where +we were getting literally hundreds of (mostly unused) specialisations of +a local function. + +Note [Do not specialise diverging functions] +~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ +Specialising a function that just diverges is a waste of code. +Furthermore, it broke GHC (simpl014) thus: + {-# STR Sb #-} + f = \x. case x of (a,b) -> f x +If we specialise f we get + f = \x. case x of (a,b) -> fspec a b +But fspec doesn't have decent strictnes info. As it happened, +(f x) :: IO t, so the state hack applied and we eta expanded fspec, +and hence f. But now f's strictness is less than its arity, which +breaks an invariant. ----------------------------------------------------- Stuff not yet handled @@ -419,7 +464,19 @@ But perhaps the first one isn't good. After all, we know that tpl_B2 is a T (I# x) really, because T is strict and Int has one constructor. (We can't unbox the strict fields, becuase T is polymorphic!) +%************************************************************************ +%* * +\subsection{Annotations} +%* * +%************************************************************************ +Annotating a type with NoSpecConstr will make SpecConstr not specialise +for arguments of that type. + +\begin{code} +data SpecConstrAnnotation = NoSpecConstr | ForceSpecConstr + deriving( Data, Typeable, Eq ) +\end{code} %************************************************************************ %* * @@ -428,24 +485,19 @@ unbox the strict fields, becuase T is polymorphic!) %************************************************************************ \begin{code} -specConstrProgram :: DynFlags -> UniqSupply -> [CoreBind] -> IO [CoreBind] -specConstrProgram dflags us binds +specConstrProgram :: ModGuts -> CoreM ModGuts +specConstrProgram guts = do - showPass dflags "SpecConstr" - - let (binds', _) = initUs us (go (initScEnv dflags) binds) - - endPass dflags "SpecConstr" Opt_D_dump_spec binds' - - dumpIfSet_dyn dflags Opt_D_dump_rules "Top-level specialisations" - (pprRules (tidyRules emptyTidyEnv (rulesOfBinds binds'))) - - return binds' + dflags <- getDynFlags + us <- getUniqueSupplyM + annos <- getFirstAnnotations deserializeWithData guts + let binds' = fst $ initUs us (go (initScEnv dflags annos) (mg_binds guts)) + return (guts { mg_binds = binds' }) where - go _ [] = returnUs [] - go env (bind:binds) = scBind env bind `thenUs` \ (env', _, bind') -> - go env' binds `thenUs` \ binds' -> - returnUs (bind' : binds') + go _ [] = return [] + go env (bind:binds) = do (env', bind') <- scTopBind env bind + binds' <- go env' binds + return (bind' : binds') \end{code} @@ -456,7 +508,9 @@ specConstrProgram dflags us binds %************************************************************************ \begin{code} -data ScEnv = SCE { sc_size :: Maybe Int, -- Size threshold +data ScEnv = SCE { sc_size :: Maybe Int, -- Size threshold + sc_count :: Maybe Int, -- Max # of specialisations for any one fn + -- See Note [Avoiding exponential blowup] sc_subst :: Subst, -- Current substitution -- Maps InIds to OutExprs @@ -465,16 +519,19 @@ data ScEnv = SCE { sc_size :: Maybe Int, -- Size threshold -- Binds interesting non-top-level variables -- Domain is OutVars (*after* applying the substitution) - sc_vals :: ValueEnv + sc_vals :: ValueEnv, -- Domain is OutIds (*after* applying the substitution) -- Used even for top-level bindings (but not imported ones) + + sc_annotations :: L.UniqFM SpecConstrAnnotation } --------------------- -- As we go, we apply a substitution (sc_subst) to the current term -type InExpr = CoreExpr -- *Before* applying the subst +type InExpr = CoreExpr -- _Before_ applying the subst +type InVar = Var -type OutExpr = CoreExpr -- *After* applying the subst +type OutExpr = CoreExpr -- _After_ applying the subst type OutId = Id type OutVar = Var @@ -483,20 +540,22 @@ type HowBoundEnv = VarEnv HowBound -- Domain is OutVars --------------------- type ValueEnv = IdEnv Value -- Domain is OutIds -data Value = ConVal AltCon [CoreArg] -- *Saturated* constructors +data Value = ConVal AltCon [CoreArg] -- _Saturated_ constructors | LambdaVal -- Inlinable lambdas or PAPs instance Outputable Value where ppr (ConVal con args) = ppr con <+> interpp'SP args - ppr LambdaVal = ptext SLIT("") + ppr LambdaVal = ptext (sLit "") --------------------- -initScEnv :: DynFlags -> ScEnv -initScEnv dflags +initScEnv :: DynFlags -> L.UniqFM SpecConstrAnnotation -> ScEnv +initScEnv dflags anns = SCE { sc_size = specConstrThreshold dflags, + sc_count = specConstrCount dflags, sc_subst = emptySubst, sc_how_bound = emptyVarEnv, - sc_vals = emptyVarEnv } + sc_vals = emptyVarEnv, + sc_annotations = anns } data HowBound = RecFun -- These are the recursive functions for which -- we seek interesting call patterns @@ -512,7 +571,7 @@ lookupHowBound :: ScEnv -> Id -> Maybe HowBound lookupHowBound env id = lookupVarEnv (sc_how_bound env) id scSubstId :: ScEnv -> Id -> CoreExpr -scSubstId env v = lookupIdSubst (sc_subst env) v +scSubstId env v = lookupIdSubst (text "scSubstId") (sc_subst env) v scSubstTy :: ScEnv -> Type -> Type scSubstTy env ty = substTy (sc_subst env) ty @@ -565,17 +624,28 @@ extendValEnv :: ScEnv -> Id -> Maybe Value -> ScEnv extendValEnv env _ Nothing = env extendValEnv env id (Just cv) = env { sc_vals = extendVarEnv (sc_vals env) id cv } -extendCaseBndrs :: ScEnv -> CoreExpr -> Id -> AltCon -> [Var] -> ScEnv +extendCaseBndrs :: ScEnv -> Id -> AltCon -> [Var] -> (ScEnv, [Var]) -- When we encounter -- case scrut of b -- C x y -> ... --- we want to bind b, and perhaps scrut too, to (C x y) --- NB: Extends only the sc_vals part of the envt -extendCaseBndrs env scrut case_bndr con alt_bndrs - = case scrut of - Var v -> extendValEnv env1 v cval - _other -> env1 +-- we want to bind b, to (C x y) +-- NB1: Extends only the sc_vals part of the envt +-- NB2: Kill the dead-ness info on the pattern binders x,y, since +-- they are potentially made alive by the [b -> C x y] binding +extendCaseBndrs env case_bndr con alt_bndrs + | isDeadBinder case_bndr + = (env, alt_bndrs) + | otherwise + = (env1, map zap alt_bndrs) + -- NB: We used to bind v too, if scrut = (Var v); but + -- the simplifer has already done this so it seems + -- redundant to do so here + -- case scrut of + -- Var v -> extendValEnv env1 v cval + -- _other -> env1 where + zap v | isTyVar v = v -- See NB2 above + | otherwise = zapIdOccInfo v env1 = extendValEnv env case_bndr cval cval = case con of DEFAULT -> Nothing @@ -584,8 +654,72 @@ extendCaseBndrs env scrut case_bndr con alt_bndrs where vanilla_args = map Type (tyConAppArgs (idType case_bndr)) ++ varsToCoreExprs alt_bndrs + +ignoreTyCon :: ScEnv -> TyCon -> Bool +ignoreTyCon env tycon + = L.lookupUFM (sc_annotations env) tycon == Just NoSpecConstr + +ignoreType :: ScEnv -> Type -> Bool +ignoreType env ty + = case splitTyConApp_maybe ty of + Just (tycon, _) -> ignoreTyCon env tycon + _ -> False + +ignoreAltCon :: ScEnv -> AltCon -> Bool +ignoreAltCon env (DataAlt dc) = ignoreTyCon env (dataConTyCon dc) +ignoreAltCon env (LitAlt lit) = ignoreType env (literalType lit) +ignoreAltCon _ DEFAULT = True + +forceSpecBndr :: ScEnv -> Var -> Bool +forceSpecBndr env var = forceSpecFunTy env . varType $ var + +forceSpecFunTy :: ScEnv -> Type -> Bool +forceSpecFunTy env = any (forceSpecArgTy env) . fst . splitFunTys + +forceSpecArgTy :: ScEnv -> Type -> Bool +forceSpecArgTy env ty + | Just ty' <- coreView ty = forceSpecArgTy env ty' + +forceSpecArgTy env ty + | Just (tycon, tys) <- splitTyConApp_maybe ty + , tycon /= funTyCon + = L.lookupUFM (sc_annotations env) tycon == Just ForceSpecConstr + || any (forceSpecArgTy env) tys + +forceSpecArgTy _ _ = False + +decreaseSpecCount :: ScEnv -> Int -> ScEnv +-- See Note [Avoiding exponential blowup] +decreaseSpecCount env n_specs + = env { sc_count = case sc_count env of + Nothing -> Nothing + Just n -> Just (n `div` (n_specs + 1)) } + -- The "+1" takes account of the original function; + -- See Note [Avoiding exponential blowup] \end{code} +Note [Avoiding exponential blowup] +~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ +The sc_count field of the ScEnv says how many times we are prepared to +duplicate a single function. But we must take care with recursive +specialiations. Consider + + let $j1 = let $j2 = let $j3 = ... + in + ...$j3... + in + ...$j2... + in + ...$j1... + +If we specialise $j1 then in each specialisation (as well as the original) +we can specialise $j2, and similarly $j3. Even if we make just *one* +specialisation of each, becuase we also have the original we'll get 2^n +copies of $j3, which is not good. + +So when recursively specialising we divide the sc_count by the number of +copies we are making at this level, including the original. + %************************************************************************ %* * @@ -658,10 +792,10 @@ A pattern binds b, x::a, y::b, z::b->a, but not 'a'! -} instance Outputable ArgOcc where - ppr (ScrutOcc xs) = ptext SLIT("scrut-occ") <> ppr xs - ppr UnkOcc = ptext SLIT("unk-occ") - ppr BothOcc = ptext SLIT("both-occ") - ppr NoOcc = ptext SLIT("no-occ") + ppr (ScrutOcc xs) = ptext (sLit "scrut-occ") <> ppr xs + ppr UnkOcc = ptext (sLit "unk-occ") + ppr BothOcc = ptext (sLit "both-occ") + ppr NoOcc = ptext (sLit "no-occ") -- Experimentally, this vesion of combineOcc makes ScrutOcc "win", so -- that if the thing is scrutinised anywhere then we get to see that @@ -680,7 +814,7 @@ combineOccs :: [ArgOcc] -> [ArgOcc] -> [ArgOcc] combineOccs xs ys = zipWithEqual "combineOccs" combineOcc xs ys setScrutOcc :: ScEnv -> ScUsage -> OutExpr -> ArgOcc -> ScUsage --- *Overwrite* the occurrence info for the scrutinee, if the scrutinee +-- _Overwrite_ the occurrence info for the scrutinee, if the scrutinee -- is a variable, and an interesting variable setScrutOcc env usg (Cast e _) occ = setScrutOcc env usg e occ setScrutOcc env usg (Note _ e) occ = setScrutOcc env usg e occ @@ -719,19 +853,19 @@ scExpr env e = scExpr' env e scExpr' env (Var v) = case scSubstId env v of - Var v' -> returnUs (varUsage env v' UnkOcc, Var v') + Var v' -> return (varUsage env v' UnkOcc, Var v') e' -> scExpr (zapScSubst env) e' -scExpr' env (Type t) = returnUs (nullUsage, Type (scSubstTy env t)) -scExpr' _ e@(Lit {}) = returnUs (nullUsage, e) -scExpr' env (Note n e) = do { (usg,e') <- scExpr env e - ; return (usg, Note n e') } -scExpr' env (Cast e co) = do { (usg, e') <- scExpr env e - ; return (usg, Cast e' (scSubstTy env co)) } +scExpr' env (Type t) = return (nullUsage, Type (scSubstTy env t)) +scExpr' _ e@(Lit {}) = return (nullUsage, e) +scExpr' env (Note n e) = do (usg,e') <- scExpr env e + return (usg, Note n e') +scExpr' env (Cast e co) = do (usg, e') <- scExpr env e + return (usg, Cast e' (scSubstTy env co)) scExpr' env e@(App _ _) = scApp env (collectArgs e) -scExpr' env (Lam b e) = do { let (env', b') = extendBndr env b - ; (usg, e') <- scExpr env' e - ; return (usg, Lam b' e') } +scExpr' env (Lam b e) = do let (env', b') = extendBndr env b + (usg, e') <- scExpr env' e + return (usg, Lam b' e') scExpr' env (Case scrut b ty alts) = do { (scrut_usg, scrut') <- scExpr env scrut @@ -742,7 +876,8 @@ scExpr' env (Case scrut b ty alts) where sc_con_app con args scrut' -- Known constructor; simplify = do { let (_, bs, rhs) = findAlt con alts - alt_env' = extendScSubstList env ((b,scrut') : bs `zip` trimConArgs con args) + `orElse` (DEFAULT, [], mkImpossibleExpr (coreAltsType alts)) + alt_env' = extendScSubstList env ((b,scrut') : bs `zip` trimConArgs con args) ; scExpr alt_env' rhs } sc_vanilla scrut_usg scrut' -- Normal case @@ -750,7 +885,7 @@ scExpr' env (Case scrut b ty alts) -- Record RecArg for the components ; (alt_usgs, alt_occs, alts') - <- mapAndUnzip3Us (sc_alt alt_env scrut' b') alts + <- mapAndUnzip3M (sc_alt alt_env scrut' b') alts ; let (alt_usg, b_occ) = lookupOcc (combineUsages alt_usgs) b' scrut_occ = foldr combineOcc b_occ alt_occs @@ -762,64 +897,87 @@ scExpr' env (Case scrut b ty alts) ; return (alt_usg `combineUsage` scrut_usg', Case scrut' b' (scSubstTy env ty) alts') } - sc_alt env scrut' b' (con,bs,rhs) - = do { let (env1, bs') = extendBndrsWith RecArg env bs - env2 = extendCaseBndrs env1 scrut' b' con bs' + sc_alt env _scrut' b' (con,bs,rhs) + = do { let (env1, bs1) = extendBndrsWith RecArg env bs + (env2, bs2) = extendCaseBndrs env1 b' con bs1 ; (usg,rhs') <- scExpr env2 rhs - ; let (usg', arg_occs) = lookupOccs usg bs' + ; let (usg', arg_occs) = lookupOccs usg bs2 scrut_occ = case con of DataAlt dc -> ScrutOcc (unitUFM dc arg_occs) - _ofther -> ScrutOcc emptyUFM - ; return (usg', scrut_occ, (con,bs',rhs')) } + _ -> ScrutOcc emptyUFM + ; return (usg', scrut_occ, (con, bs2, rhs')) } scExpr' env (Let (NonRec bndr rhs) body) | isTyVar bndr -- Type-lets may be created by doBeta = scExpr' (extendScSubst env bndr rhs) body - | otherwise + + | otherwise -- Note [Local let bindings] = do { let (body_env, bndr') = extendBndr env bndr - ; (rhs_usg, (_, args', rhs_body', _)) <- scRecRhs env (bndr',rhs) - ; let rhs' = mkLams args' rhs_body' - - ; if not opt_SpecInlineJoinPoints || null args' || isEmptyVarEnv (scu_calls rhs_usg) then do - do { -- Vanilla case - let body_env2 = extendValEnv body_env bndr' (isValue (sc_vals env) rhs') - -- Record if the RHS is a value - ; (body_usg, body') <- scExpr body_env2 body - ; return (body_usg `combineUsage` rhs_usg, Let (NonRec bndr' rhs') body') } - else -- For now, just brutally inline the join point - do { let body_env2 = extendScSubst env bndr rhs' - ; scExpr body_env2 body } } - + body_env2 = extendHowBound body_env [bndr'] RecFun + ; (body_usg, body') <- scExpr body_env2 body -{- Old code - do { -- Join-point case - let body_env2 = extendHowBound body_env [bndr'] RecFun - -- If the RHS of this 'let' contains calls - -- to recursive functions that we're trying - -- to specialise, then treat this let too - -- as one to specialise - ; (body_usg, body') <- scExpr body_env2 body + ; (rhs_usg, rhs_info) <- scRecRhs env (bndr',rhs) - ; (spec_usg, _, specs) <- specialise env (scu_calls body_usg) ([], rhs_info) + ; let force_spec = False + ; (spec_usg, specs) <- specialise env force_spec + (scu_calls body_usg) + rhs_info + (SI [] 0 (Just rhs_usg)) - ; return (body_usg { scu_calls = scu_calls body_usg `delVarEnv` bndr' } - `combineUsage` rhs_usg `combineUsage` spec_usg, - mkLets [NonRec b r | (b,r) <- addRules rhs_info specs] body') + ; return (body_usg { scu_calls = scu_calls body_usg `delVarEnv` bndr' } + `combineUsage` spec_usg, + mkLets [NonRec b r | (b,r) <- specInfoBinds rhs_info specs] body') } --} + +-- A *local* recursive group: see Note [Local recursive groups] scExpr' env (Let (Rec prs) body) - = do { (env', bind_usg, bind') <- scBind env (Rec prs) - ; (body_usg, body') <- scExpr env' body - ; return (bind_usg `combineUsage` body_usg, Let bind' body') } + = do { let (bndrs,rhss) = unzip prs + (rhs_env1,bndrs') = extendRecBndrs env bndrs + rhs_env2 = extendHowBound rhs_env1 bndrs' RecFun + force_spec = any (forceSpecBndr env) bndrs' + + ; (rhs_usgs, rhs_infos) <- mapAndUnzipM (scRecRhs rhs_env2) (bndrs' `zip` rhss) + ; (body_usg, body') <- scExpr rhs_env2 body + + -- NB: start specLoop from body_usg + ; (spec_usg, specs) <- specLoop rhs_env2 force_spec + (scu_calls body_usg) rhs_infos nullUsage + [SI [] 0 (Just usg) | usg <- rhs_usgs] + -- Do not unconditionally use rhs_usgs. + -- Instead use them only if we find an unspecialised call + -- See Note [Local recursive groups] + + ; let all_usg = spec_usg `combineUsage` body_usg + bind' = Rec (concat (zipWith specInfoBinds rhs_infos specs)) + + ; return (all_usg { scu_calls = scu_calls all_usg `delVarEnvList` bndrs' }, + Let bind' body') } +\end{code} + +Note [Local let bindings] +~~~~~~~~~~~~~~~~~~~~~~~~~ +It is not uncommon to find this + + let $j = \x. in ...$j True...$j True... + +Here $j is an arbitrary let-bound function, but it often comes up for +join points. We might like to specialise $j for its call patterns. +Notice the difference from a letrec, where we look for call patterns +in the *RHS* of the function. Here we look for call patterns in the +*body* of the let. + +At one point I predicated this on the RHS mentioning the outer +recursive function, but that's not essential and might even be +harmful. I'm not sure. ------------------------------------ +\begin{code} scApp :: ScEnv -> (InExpr, [InExpr]) -> UniqSM (ScUsage, CoreExpr) scApp env (Var fn, args) -- Function is a variable = ASSERT( not (null args) ) - do { args_w_usgs <- mapUs (scExpr env) args + do { args_w_usgs <- mapM (scExpr env) args ; let (arg_usgs, args') = unzip args_w_usgs arg_usg = combineUsages arg_usgs ; case scSubstId env fn of @@ -852,57 +1010,41 @@ scApp env (Var fn, args) -- Function is a variable -- (let f = ...f... in f) arg1 arg2 scApp env (other_fn, args) = do { (fn_usg, fn') <- scExpr env other_fn - ; (arg_usgs, args') <- mapAndUnzipUs (scExpr env) args + ; (arg_usgs, args') <- mapAndUnzipM (scExpr env) args ; return (combineUsages arg_usgs `combineUsage` fn_usg, mkApps fn' args') } ---------------------- -scBind :: ScEnv -> CoreBind -> UniqSM (ScEnv, ScUsage, CoreBind) -scBind env (Rec prs) +scTopBind :: ScEnv -> CoreBind -> UniqSM (ScEnv, CoreBind) +scTopBind env (Rec prs) | Just threshold <- sc_size env + , not force_spec , not (all (couldBeSmallEnoughToInline threshold) rhss) -- No specialisation = do { let (rhs_env,bndrs') = extendRecBndrs env bndrs - ; (rhs_usgs, rhss') <- mapAndUnzipUs (scExpr rhs_env) rhss - ; return (rhs_env, combineUsages rhs_usgs, Rec (bndrs' `zip` rhss')) } + ; (_, rhss') <- mapAndUnzipM (scExpr rhs_env) rhss + ; return (rhs_env, Rec (bndrs' `zip` rhss')) } | otherwise -- Do specialisation = do { let (rhs_env1,bndrs') = extendRecBndrs env bndrs rhs_env2 = extendHowBound rhs_env1 bndrs' RecFun - ; (rhs_usgs, rhs_infos) <- mapAndUnzipUs (scRecRhs rhs_env2) (bndrs' `zip` rhss) + ; (rhs_usgs, rhs_infos) <- mapAndUnzipM (scRecRhs rhs_env2) (bndrs' `zip` rhss) ; let rhs_usg = combineUsages rhs_usgs - ; (spec_usg, specs) <- spec_loop rhs_env2 (scu_calls rhs_usg) - (repeat [] `zip` rhs_infos) - - ; let all_usg = rhs_usg `combineUsage` spec_usg + ; (_, specs) <- specLoop rhs_env2 force_spec + (scu_calls rhs_usg) rhs_infos nullUsage + [SI [] 0 Nothing | _ <- bndrs] ; return (rhs_env1, -- For the body of the letrec, delete the RecFun business - all_usg { scu_calls = scu_calls rhs_usg `delVarEnvList` bndrs' }, - Rec (concat (zipWith addRules rhs_infos specs))) } + Rec (concat (zipWith specInfoBinds rhs_infos specs))) } where (bndrs,rhss) = unzip prs + force_spec = any (forceSpecBndr env) bndrs - spec_loop :: ScEnv - -> CallEnv - -> [([CallPat], RhsInfo)] -- One per binder - -> UniqSM (ScUsage, [[SpecInfo]]) -- One list per binder - spec_loop env all_calls rhs_stuff - = do { (spec_usg_s, new_pats_s, specs) <- mapAndUnzip3Us (specialise env all_calls) rhs_stuff - ; let spec_usg = combineUsages spec_usg_s - ; if all null new_pats_s then - return (spec_usg, specs) else do - { (spec_usg1, specs1) <- spec_loop env (scu_calls spec_usg) - (zipWith add_pats new_pats_s rhs_stuff) - ; return (spec_usg `combineUsage` spec_usg1, zipWith (++) specs specs1) } } - - add_pats :: [CallPat] -> ([CallPat], RhsInfo) -> ([CallPat], RhsInfo) - add_pats new_pats (done_pats, rhs_info) = (done_pats ++ new_pats, rhs_info) - -scBind env (NonRec bndr rhs) - = do { (usg, rhs') <- scExpr env rhs +scTopBind env (NonRec bndr rhs) + = do { (_, rhs') <- scExpr env rhs ; let (env1, bndr') = extendBndr env bndr env2 = extendValEnv env1 bndr' (isValue (sc_vals env) rhs') - ; return (env2, usg, NonRec bndr' rhs') } + ; return (env2, NonRec bndr' rhs') } ---------------------- scRecRhs :: ScEnv -> (OutId, InExpr) -> UniqSM (ScUsage, RhsInfo) @@ -911,20 +1053,20 @@ scRecRhs env (bndr,rhs) (body_env, arg_bndrs') = extendBndrsWith RecArg env arg_bndrs ; (body_usg, body') <- scExpr body_env body ; let (rhs_usg, arg_occs) = lookupOccs body_usg arg_bndrs' - ; return (rhs_usg, (bndr, arg_bndrs', body', arg_occs)) } - + ; return (rhs_usg, RI bndr (mkLams arg_bndrs' body') + arg_bndrs body arg_occs) } -- The arg_occs says how the visible, -- lambda-bound binders of the RHS are used -- (including the TyVar binders) -- Two pats are the same if they match both ways ---------------------- -addRules :: RhsInfo -> [SpecInfo] -> [(Id,CoreExpr)] -addRules (fn, args, body, _) specs - = [(id,rhs) | (_,id,rhs) <- specs] ++ - [(fn `addIdSpecialisations` rules, mkLams args body)] +specInfoBinds :: RhsInfo -> SpecInfo -> [(Id,CoreExpr)] +specInfoBinds (RI fn new_rhs _ _ _) (SI specs _ _) + = [(id,rhs) | OS _ _ id rhs <- specs] ++ + [(fn `addIdSpecialisations` rules, new_rhs)] where - rules = [r | (r,_,_) <- specs] + rules = [r | OS _ r _ _ <- specs] ---------------------- varUsage :: ScEnv -> OutVar -> ArgOcc -> ScUsage @@ -942,49 +1084,111 @@ varUsage env v use %************************************************************************ \begin{code} -type RhsInfo = (OutId, [OutVar], OutExpr, [ArgOcc]) - -- Info about the *original* RHS of a binding we are specialising - -- Original binding f = \xs.body - -- Plus info about usage of arguments +data RhsInfo = RI OutId -- The binder + OutExpr -- The new RHS + [InVar] InExpr -- The *original* RHS (\xs.body) + -- Note [Specialise original body] + [ArgOcc] -- Info on how the xs occur in body -type SpecInfo = (CoreRule, OutId, OutExpr) - -- One specialisation: Rule plus definition +data SpecInfo = SI [OneSpec] -- The specialisations we have generated + + Int -- Length of specs; used for numbering them + (Maybe ScUsage) -- Nothing => we have generated specialisations + -- from calls in the *original* RHS + -- Just cs => we haven't, and this is the usage + -- of the original RHS + -- See Note [Local recursive groups] + + -- One specialisation: Rule plus definition +data OneSpec = OS CallPat -- Call pattern that generated this specialisation + CoreRule -- Rule connecting original id with the specialisation + OutId OutExpr -- Spec id + its rhs + + +specLoop :: ScEnv + -> Bool -- force specialisation? + -> CallEnv + -> [RhsInfo] + -> ScUsage -> [SpecInfo] -- One per binder; acccumulating parameter + -> UniqSM (ScUsage, [SpecInfo]) -- ...ditto... +specLoop env force_spec all_calls rhs_infos usg_so_far specs_so_far + = do { specs_w_usg <- zipWithM (specialise env force_spec all_calls) rhs_infos specs_so_far + ; let (new_usg_s, all_specs) = unzip specs_w_usg + new_usg = combineUsages new_usg_s + new_calls = scu_calls new_usg + all_usg = usg_so_far `combineUsage` new_usg + ; if isEmptyVarEnv new_calls then + return (all_usg, all_specs) + else + specLoop env force_spec new_calls rhs_infos all_usg all_specs } specialise :: ScEnv + -> Bool -- force specialisation? -> CallEnv -- Info on calls - -> ([CallPat], RhsInfo) -- Original RHS plus patterns dealt with - -> UniqSM (ScUsage, [CallPat], [SpecInfo]) -- Specialised calls + -> RhsInfo + -> SpecInfo -- Original RHS plus patterns dealt with + -> UniqSM (ScUsage, SpecInfo) -- New specialised versions and their usage -- Note: the rhs here is the optimised version of the original rhs -- So when we make a specialised copy of the RHS, we're starting -- from an RHS whose nested functions have been optimised already. -specialise env bind_calls (done_pats, (fn, arg_bndrs, body, arg_occs)) - | notNull arg_bndrs, -- Only specialise functions - Just all_calls <- lookupVarEnv bind_calls fn - = do { pats <- callsToPats env done_pats arg_occs all_calls --- ; pprTrace "specialise" (vcat [ppr fn <+> ppr arg_occs, --- text "calls" <+> ppr all_calls, --- text "good pats" <+> ppr pats]) $ +specialise env force_spec bind_calls (RI fn _ arg_bndrs body arg_occs) + spec_info@(SI specs spec_count mb_unspec) + | not (isBottomingId fn) -- Note [Do not specialise diverging functions] + , notNull arg_bndrs -- Only specialise functions + , Just all_calls <- lookupVarEnv bind_calls fn + = do { (boring_call, pats) <- callsToPats env specs arg_occs all_calls +-- ; pprTrace "specialise" (vcat [ ppr fn <+> text "with" <+> int (length pats) <+> text "good patterns" +-- , text "arg_occs" <+> ppr arg_occs +-- , text "calls" <+> ppr all_calls +-- , text "good pats" <+> ppr pats]) $ -- return () - ; (spec_usgs, specs) <- mapAndUnzipUs (spec_one env fn arg_bndrs body) - (pats `zip` [length done_pats..]) - - ; return (combineUsages spec_usgs, pats, specs) } + -- Bale out if too many specialisations + ; let n_pats = length pats + spec_count' = n_pats + spec_count + ; case sc_count env of + Just max | not force_spec && spec_count' > max + -> pprTrace "SpecConstr" msg $ + return (nullUsage, spec_info) + where + msg = vcat [ sep [ ptext (sLit "Function") <+> quotes (ppr fn) + , nest 2 (ptext (sLit "has") <+> + speakNOf spec_count' (ptext (sLit "call pattern")) <> comma <+> + ptext (sLit "but the limit is") <+> int max) ] + , ptext (sLit "Use -fspec-constr-count=n to set the bound") + , extra ] + extra | not opt_PprStyle_Debug = ptext (sLit "Use -dppr-debug to see specialisations") + | otherwise = ptext (sLit "Specialisations:") <+> ppr (pats ++ [p | OS p _ _ _ <- specs]) + + _normal_case -> do { + + let spec_env = decreaseSpecCount env n_pats + ; (spec_usgs, new_specs) <- mapAndUnzipM (spec_one spec_env fn arg_bndrs body) + (pats `zip` [spec_count..]) + -- See Note [Specialise original body] + + ; let spec_usg = combineUsages spec_usgs + (new_usg, mb_unspec') + = case mb_unspec of + Just rhs_usg | boring_call -> (spec_usg `combineUsage` rhs_usg, Nothing) + _ -> (spec_usg, mb_unspec) + + ; return (new_usg, SI (new_specs ++ specs) spec_count' mb_unspec') } } | otherwise - = return (nullUsage, [], []) -- The boring case + = return (nullUsage, spec_info) -- The boring case --------------------- spec_one :: ScEnv -> OutId -- Function - -> [Var] -- Lambda-binders of RHS; should match patterns - -> CoreExpr -- Body of the original function - -> (([Var], [CoreArg]), Int) - -> UniqSM (ScUsage, SpecInfo) -- Rule and binding + -> [InVar] -- Lambda-binders of RHS; should match patterns + -> InExpr -- Body of the original function + -> (CallPat, Int) + -> UniqSM (ScUsage, OneSpec) -- Rule and binding -- spec_one creates a specialised copy of the function, together -- with a rule for using it. I'm very proud of how short this @@ -1008,45 +1212,103 @@ spec_one :: ScEnv f (b,c) ((:) (a,(b,c)) (x,v) hw) = f_spec b c v hw -} -spec_one env fn arg_bndrs body ((qvars, pats), rule_number) - = do { -- Specialise the body - let spec_env = extendScSubstList (extendScInScope env qvars) +spec_one env fn arg_bndrs body (call_pat@(qvars, pats), rule_number) + = do { spec_uniq <- getUniqueUs + ; let spec_env = extendScSubstList (extendScInScope env qvars) (arg_bndrs `zip` pats) + fn_name = idName fn + fn_loc = nameSrcSpan fn_name + spec_occ = mkSpecOcc (nameOccName fn_name) + rule_name = mkFastString ("SC:" ++ showSDoc (ppr fn <> int rule_number)) + spec_name = mkInternalName spec_uniq spec_occ fn_loc +-- ; pprTrace "{spec_one" (ppr (sc_count env) <+> ppr fn <+> ppr pats <+> text "-->" <+> ppr spec_name) $ +-- return () + + -- Specialise the body ; (spec_usg, spec_body) <- scExpr spec_env body --- ; pprTrace "spec_one" (ppr fn <+> vcat [text "pats" <+> ppr pats, --- text "calls" <+> (ppr (scu_calls spec_usg))]) --- (return ()) +-- ; pprTrace "done spec_one}" (ppr fn) $ +-- return () -- And build the results - ; spec_uniq <- getUniqueUs - ; let (spec_lam_args, spec_call_args) = mkWorkerArgs qvars body_ty + ; let spec_id = mkLocalId spec_name (mkPiTypes spec_lam_args body_ty) + `setIdStrictness` spec_str -- See Note [Transfer strictness] + `setIdArity` count isId spec_lam_args + spec_str = calcSpecStrictness fn spec_lam_args pats + (spec_lam_args, spec_call_args) = mkWorkerArgs qvars body_ty -- Usual w/w hack to avoid generating -- a spec_rhs of unlifted type and no args - - fn_name = idName fn - fn_loc = nameSrcSpan fn_name - spec_occ = mkSpecOcc (nameOccName fn_name) - rule_name = mkFastString ("SC:" ++ showSDoc (ppr fn <> int rule_number)) - spec_rhs = mkLams spec_lam_args spec_body - spec_id = mkUserLocal spec_occ spec_uniq (mkPiTypes spec_lam_args body_ty) fn_loc - body_ty = exprType spec_body - rule_rhs = mkVarApps (Var spec_id) spec_call_args - rule = mkLocalRule rule_name specConstrActivation fn_name qvars pats rule_rhs - ; return (spec_usg, (rule, spec_id, spec_rhs)) } - --- In which phase should the specialise-constructor rules be active? --- Originally I made them always-active, but Manuel found that --- this defeated some clever user-written rules. So Plan B --- is to make them active only in Phase 0; after all, currently, --- the specConstr transformation is only run after the simplifier --- has reached Phase 0. In general one would want it to be --- flag-controllable, but for now I'm leaving it baked in --- [SLPJ Oct 01] -specConstrActivation :: Activation -specConstrActivation = ActiveAfter 0 -- Baked in; see comments above + + spec_rhs = mkLams spec_lam_args spec_body + body_ty = exprType spec_body + rule_rhs = mkVarApps (Var spec_id) spec_call_args + inline_act = idInlineActivation fn + rule = mkLocalRule rule_name inline_act fn_name qvars pats rule_rhs + ; return (spec_usg, OS call_pat rule spec_id spec_rhs) } + +calcSpecStrictness :: Id -- The original function + -> [Var] -> [CoreExpr] -- Call pattern + -> StrictSig -- Strictness of specialised thing +-- See Note [Transfer strictness] +calcSpecStrictness fn qvars pats + = StrictSig (mkTopDmdType spec_dmds TopRes) + where + spec_dmds = [ lookupVarEnv dmd_env qv `orElse` lazyDmd | qv <- qvars, isId qv ] + StrictSig (DmdType _ dmds _) = idStrictness fn + + dmd_env = go emptyVarEnv dmds pats + + go env ds (Type {} : pats) = go env ds pats + go env (d:ds) (pat : pats) = go (go_one env d pat) ds pats + go env _ _ = env + + go_one env d (Var v) = extendVarEnv_C both env v d + go_one env (Box d) e = go_one env d e + go_one env (Eval (Prod ds)) e + | (Var _, args) <- collectArgs e = go env ds args + go_one env _ _ = env + \end{code} +Note [Specialise original body] +~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ +The RhsInfo for a binding keeps the *original* body of the binding. We +must specialise that, *not* the result of applying specExpr to the RHS +(which is also kept in RhsInfo). Otherwise we end up specialising a +specialised RHS, and that can lead directly to exponential behaviour. + +Note [Transfer activation] +~~~~~~~~~~~~~~~~~~~~~~~~~~ +In which phase should the specialise-constructor rules be active? +Originally I made them always-active, but Manuel found that this +defeated some clever user-written rules. Then I made them active only +in Phase 0; after all, currently, the specConstr transformation is +only run after the simplifier has reached Phase 0, but that meant +that specialisations didn't fire inside wrappers; see test +simplCore/should_compile/spec-inline. + +So now I just use the inline-activation of the parent Id, as the +activation for the specialiation RULE, just like the main specialiser; +see Note [Auto-specialisation and RULES] in Specialise. + + +Note [Transfer strictness] +~~~~~~~~~~~~~~~~~~~~~~~~~~ +We must transfer strictness information from the original function to +the specialised one. Suppose, for example + + f has strictness SS + and a RULE f (a:as) b = f_spec a as b + +Now we want f_spec to have strictess LLS, otherwise we'll use call-by-need +when calling f_spec instead of call-by-value. And that can result in +unbounded worsening in space (cf the classic foldl vs foldl') + +See Trac #3437 for a good example. + +The function calcSpecStrictness performs the calculation. + + %************************************************************************ %* * \subsection{Argument analysis} @@ -1061,17 +1323,20 @@ they are constructor applications. type CallPat = ([Var], [CoreExpr]) -- Quantified variables and arguments -callsToPats :: ScEnv -> [CallPat] -> [ArgOcc] -> [Call] -> UniqSM [CallPat] +callsToPats :: ScEnv -> [OneSpec] -> [ArgOcc] -> [Call] -> UniqSM (Bool, [CallPat]) -- Result has no duplicate patterns, -- nor ones mentioned in done_pats -callsToPats env done_pats bndr_occs calls + -- Bool indicates that there was at least one boring pattern +callsToPats env done_specs bndr_occs calls = do { mb_pats <- mapM (callToPats env bndr_occs) calls ; let good_pats :: [([Var], [CoreArg])] good_pats = catMaybes mb_pats + done_pats = [p | OS p _ _ _ <- done_specs] is_done p = any (samePat p) done_pats - ; return (filterOut is_done (nubBy samePat good_pats)) } + ; return (any isNothing mb_pats, + filterOut is_done (nubBy samePat good_pats)) } callToPats :: ScEnv -> [ArgOcc] -> Call -> UniqSM (Maybe CallPat) -- The [Var] is the variables to quantify over in the rule @@ -1083,8 +1348,8 @@ callToPats env bndr_occs (con_env, args) = return Nothing | otherwise = do { let in_scope = substInScope (sc_subst env) - ; prs <- argsToPats in_scope con_env (args `zip` bndr_occs) - ; let (good_pats, pats) = unzip prs + ; prs <- argsToPats env in_scope con_env (args `zip` bndr_occs) + ; let (interesting_s, pats) = unzip prs pat_fvs = varSetElems (exprsFreeVars pats) qvars = filterOut (`elemInScopeSet` in_scope) pat_fvs -- Quantify over variables that are not in sccpe @@ -1097,7 +1362,7 @@ callToPats env bndr_occs (con_env, args) -- variable may mention a type variable ; -- pprTrace "callToPats" (ppr args $$ ppr prs $$ ppr bndr_occs) $ - if or good_pats + if or interesting_s then return (Just (qvars', pats)) else return Nothing } @@ -1107,7 +1372,8 @@ callToPats env bndr_occs (con_env, args) -- placeholder variables. For example: -- C a (D (f x) (g y)) ==> C p1 (D p2 p3) -argToPat :: InScopeSet -- What's in scope at the fn defn site +argToPat :: ScEnv + -> InScopeSet -- What's in scope at the fn defn site -> ValueEnv -- ValueEnv at the call site -> CoreArg -- A call arg (or component thereof) -> ArgOcc @@ -1122,11 +1388,11 @@ argToPat :: InScopeSet -- What's in scope at the fn defn site -- lvl7 --> (True, lvl7) if lvl7 is bound -- somewhere further out -argToPat _in_scope _val_env arg@(Type {}) _arg_occ +argToPat _env _in_scope _val_env arg@(Type {}) _arg_occ = return (False, arg) -argToPat in_scope val_env (Note _ arg) arg_occ - = argToPat in_scope val_env arg arg_occ +argToPat env in_scope val_env (Note _ arg) arg_occ + = argToPat env in_scope val_env arg arg_occ -- Note [Notes in call patterns] -- ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ -- Ignore Notes. In particular, we want to ignore any InlineMe notes @@ -1134,24 +1400,28 @@ argToPat in_scope val_env (Note _ arg) arg_occ -- ride roughshod over them all for now. --- See Note [Notes in RULE matching] in Rules -argToPat in_scope val_env (Let _ arg) arg_occ - = argToPat in_scope val_env arg arg_occ +argToPat env in_scope val_env (Let _ arg) arg_occ + = argToPat env in_scope val_env arg arg_occ -- Look through let expressions -- e.g. f (let v = rhs in \y -> ...v...) -- Here we can specialise for f (\y -> ...) -- because the rule-matcher will look through the let. -argToPat in_scope val_env (Cast arg co) arg_occ - = do { (interesting, arg') <- argToPat in_scope val_env arg arg_occ - ; let (ty1,ty2) = coercionKind co +argToPat env in_scope val_env (Cast arg co) arg_occ + | not (ignoreType env ty2) + = do { (interesting, arg') <- argToPat env in_scope val_env arg arg_occ ; if not interesting then wildCardPat ty2 else do { -- Make a wild-card pattern for the coercion uniq <- getUniqueUs - ; let co_name = mkSysTvName uniq FSLIT("sg") + ; let co_name = mkSysTvName uniq (fsLit "sg") co_var = mkCoVar co_name (mkCoKind ty1 ty2) ; return (interesting, Cast arg' (mkTyVarTy co_var)) } } + where + (ty1, ty2) = coercionKind co + + {- Disabling lambda specialisation for now It's fragile, and the spec_loop can be infinite @@ -1167,15 +1437,16 @@ argToPat in_scope val_env arg arg_occ -- Check for a constructor application -- NB: this *precedes* the Var case, so that we catch nullary constrs -argToPat in_scope val_env arg arg_occ +argToPat env in_scope val_env arg arg_occ | Just (ConVal dc args) <- isValue val_env arg + , not (ignoreAltCon env dc) , case arg_occ of ScrutOcc _ -> True -- Used only by case scrutinee BothOcc -> case arg of -- Used elsewhere App {} -> True -- see Note [Reboxing] _other -> False _other -> False -- No point; the arg is not decomposed - = do { args' <- argsToPats in_scope val_env (args `zip` conArgOccs arg_occ dc) + = do { args' <- argsToPats env in_scope val_env (args `zip` conArgOccs arg_occ dc) ; return (True, mk_con_app dc (map snd args')) } -- Check if the argument is a variable that @@ -1183,9 +1454,10 @@ argToPat in_scope val_env arg arg_occ -- It's worth specialising on this if -- (a) it's used in an interesting way in the body -- (b) we know what its value is -argToPat in_scope val_env (Var v) arg_occ +argToPat env in_scope val_env (Var v) arg_occ | case arg_occ of { UnkOcc -> False; _other -> True }, -- (a) - is_value -- (b) + is_value, -- (b) + not (ignoreType env (varType v)) = return (True, Var v) where is_value @@ -1200,29 +1472,35 @@ argToPat in_scope val_env (Var v) arg_occ -- variables that are in soope, which in turn can -- expose the weakness in let-matching -- See Note [Matching lets] in Rules + -- Check for a variable bound inside the function. -- Don't make a wild-card, because we may usefully share -- e.g. f a = let x = ... in f (x,x) -- NB: this case follows the lambda and con-app cases!! -argToPat _in_scope _val_env (Var v) _arg_occ - = return (False, Var v) +-- argToPat _in_scope _val_env (Var v) _arg_occ +-- = return (False, Var v) + -- SLPJ : disabling this to avoid proliferation of versions + -- also works badly when thinking about seeding the loop + -- from the body of the let + -- f x y = letrec g z = ... in g (x,y) + -- We don't want to specialise for that *particular* x,y -- The default case: make a wild-card -argToPat _in_scope _val_env arg _arg_occ +argToPat _env _in_scope _val_env arg _arg_occ = wildCardPat (exprType arg) wildCardPat :: Type -> UniqSM (Bool, CoreArg) wildCardPat ty = do { uniq <- getUniqueUs - ; let id = mkSysLocal FSLIT("sc") uniq ty + ; let id = mkSysLocal (fsLit "sc") uniq ty ; return (False, Var id) } -argsToPats :: InScopeSet -> ValueEnv +argsToPats :: ScEnv -> InScopeSet -> ValueEnv -> [(CoreArg, ArgOcc)] -> UniqSM [(Bool, CoreArg)] -argsToPats in_scope val_env args - = mapUs do_one args +argsToPats env in_scope val_env args + = mapM do_one args where - do_one (arg,occ) = argToPat in_scope val_env arg occ + do_one (arg,occ) = argToPat env in_scope val_env arg occ \end{code} @@ -1245,7 +1523,9 @@ isValue env (Var v) -- as well, for let-bound constructors! isValue env (Lam b e) - | isTyVar b = isValue env e + | isTyVar b = case isValue env e of + Just _ -> Just LambdaVal + Nothing -> Nothing | otherwise = Just LambdaVal isValue _env expr -- Maybe it's a constructor application