X-Git-Url: http://git.megacz.com/?a=blobdiff_plain;f=compiler%2Fvectorise%2FVectorise.hs;h=59fded3c4f01c552c081869fd34c7698b2e285d6;hb=222415a5b658e737a0a1f2c980c6f80635289f75;hp=bbfa562f31396ccd5d19ecbf2db4d5d0e9217093;hpb=fc21a2bcf35337942156d246f0c1c1b7072f3f91;p=ghc-hetmet.git diff --git a/compiler/vectorise/Vectorise.hs b/compiler/vectorise/Vectorise.hs index bbfa562..59fded3 100644 --- a/compiler/vectorise/Vectorise.hs +++ b/compiler/vectorise/Vectorise.hs @@ -1,118 +1,126 @@ + module Vectorise( vectorise ) where -#include "HsVersions.h" - import VectMonad import VectUtils import VectType import VectCore -import DynFlags -import HscTypes +import HscTypes hiding ( MonadThings(..) ) -import CoreLint ( showPass, endPass ) +import Module ( PackageId ) import CoreSyn import CoreUtils +import CoreUnfold ( mkInlineRule ) +import MkCore ( mkWildCase ) import CoreFVs -import SimplMonad ( SimplCount, zeroSimplCount ) -import Rules ( RuleBase ) +import CoreMonad ( CoreM, getHscEnv ) import DataCon import TyCon import Type import FamInstEnv ( extendFamInstEnvList ) -import InstEnv ( extendInstEnvList ) import Var import VarEnv import VarSet -import Name ( mkSysTvName, getName ) -import NameEnv import Id -import MkId ( unwrapFamInstScrut ) import OccName +import BasicTypes ( isLoopBreaker ) -import DsMonad hiding (mapAndUnzipM) -import DsUtils ( mkCoreTup, mkCoreTupTy ) - -import Literal ( Literal ) -import PrelNames +import Literal ( Literal, mkMachInt ) import TysWiredIn import TysPrim ( intPrimTy ) -import BasicTypes ( Boxity(..) ) import Outputable import FastString -import Control.Monad ( liftM, liftM2, zipWithM, mapAndUnzipM ) +import Util ( zipLazy ) +import Control.Monad +import Data.List ( sortBy, unzip4 ) + +vectorise :: PackageId -> ModGuts -> CoreM ModGuts +vectorise backend guts = do + hsc_env <- getHscEnv + liftIO $ vectoriseIO backend hsc_env guts -vectorise :: HscEnv -> UniqSupply -> RuleBase -> ModGuts - -> IO (SimplCount, ModGuts) -vectorise hsc_env _ _ guts +vectoriseIO :: PackageId -> HscEnv -> ModGuts -> IO ModGuts +vectoriseIO backend hsc_env guts = do - showPass dflags "Vectorisation" eps <- hscEPS hsc_env let info = hptVectInfo hsc_env `plusVectInfo` eps_vect_info eps - Just (info', guts') <- initV hsc_env guts info (vectModule guts) - endPass dflags "Vectorisation" Opt_D_dump_vect (mg_binds guts') - return (zeroSimplCount dflags, guts' { mg_vect_info = info' }) - where - dflags = hsc_dflags hsc_env + Just (info', guts') <- initV backend hsc_env guts info (vectModule guts) + return (guts' { mg_vect_info = info' }) vectModule :: ModGuts -> VM ModGuts vectModule guts = do - (types', fam_insts, pa_insts) <- vectTypeEnv (mg_types guts) - - let insts = map painstInstance pa_insts - fam_inst_env' = extendFamInstEnvList (mg_fam_inst_env guts) fam_insts - inst_env' = extendInstEnvList (mg_inst_env guts) insts - updGEnv (setInstEnvs inst_env' fam_inst_env') - - dicts <- mapM buildPADict pa_insts - workers <- mapM vectDataConWorkers pa_insts + (types', fam_insts, tc_binds) <- vectTypeEnv (mg_types guts) + + let fam_inst_env' = extendFamInstEnvList (mg_fam_inst_env guts) fam_insts + updGEnv (setFamInstEnv fam_inst_env') + + -- dicts <- mapM buildPADict pa_insts + -- workers <- mapM vectDataConWorkers pa_insts binds' <- mapM vectTopBind (mg_binds guts) return $ guts { mg_types = types' - , mg_binds = Rec (concat workers ++ concat dicts) : binds' - , mg_inst_env = inst_env' + , mg_binds = Rec tc_binds : binds' , mg_fam_inst_env = fam_inst_env' - , mg_insts = mg_insts guts ++ insts , mg_fam_insts = mg_fam_insts guts ++ fam_insts } vectTopBind :: CoreBind -> VM CoreBind vectTopBind b@(NonRec var expr) = do - var' <- vectTopBinder var - expr' <- vectTopRhs var expr + (inline, expr') <- vectTopRhs var expr + var' <- vectTopBinder var inline expr' hs <- takeHoisted - return . Rec $ (var, expr) : (var', expr') : hs + cexpr <- tryConvert var var' expr + return . Rec $ (var, cexpr) : (var', expr') : hs `orElseV` return b vectTopBind b@(Rec bs) = do - vars' <- mapM vectTopBinder vars - exprs' <- zipWithM vectTopRhs vars exprs + (vars', _, exprs') <- fixV $ \ ~(_, inlines, rhss) -> + do + vars' <- sequence [vectTopBinder var inline rhs + | (var, ~(inline, rhs)) + <- zipLazy vars (zip inlines rhss)] + (inlines', exprs') <- mapAndUnzipM (uncurry vectTopRhs) bs + return (vars', inlines', exprs') hs <- takeHoisted - return . Rec $ bs ++ zip vars' exprs' ++ hs + cexprs <- sequence $ zipWith3 tryConvert vars vars' exprs + return . Rec $ zip vars cexprs ++ zip vars' exprs' ++ hs `orElseV` return b where (vars, exprs) = unzip bs -vectTopBinder :: Var -> VM Var -vectTopBinder var +-- NOTE: vectTopBinder *MUST* be lazy in inline and expr because of how it is +-- used inside of fixV in vectTopBind +vectTopBinder :: Var -> Inline -> CoreExpr -> VM Var +vectTopBinder var inline expr = do vty <- vectType (idType var) - var' <- cloneId mkVectOcc var vty + var' <- liftM (`setIdUnfolding` unfolding) $ cloneId mkVectOcc var vty defGlobalVar var var' return var' - -vectTopRhs :: Var -> CoreExpr -> VM CoreExpr + where + unfolding = case inline of + Inline arity -> mkInlineRule InlSat expr arity + DontInline -> noUnfolding + +vectTopRhs :: Var -> CoreExpr -> VM (Inline, CoreExpr) vectTopRhs var expr - = do - closedV . liftM vectorised - . inBind var - $ vectPolyExpr (freeVars expr) + = closedV + $ do + (inline, vexpr) <- inBind var + $ vectPolyExpr (isLoopBreaker $ idOccInfo var) + (freeVars expr) + return (inline, vectorised vexpr) + +tryConvert :: Var -> Var -> CoreExpr -> VM CoreExpr +tryConvert var vect_var rhs + = fromVect (idType var) (Var vect_var) `orElseV` return rhs -- ---------------------------------------------------------------------------- -- Bindings @@ -120,8 +128,7 @@ vectTopRhs var expr vectBndr :: Var -> VM VVar vectBndr v = do - vty <- vectType (idType v) - lty <- mkPArrayType vty + (vty, lty) <- vectAndLiftType (idType v) let vv = v `Id.setIdType` vty lv = v `Id.setIdType` lty updLEnv (mapTo vv lv) @@ -129,6 +136,16 @@ vectBndr v where mapTo vv lv env = env { local_vars = extendVarEnv (local_vars env) v (vv, lv) } +vectBndrNew :: Var -> FastString -> VM VVar +vectBndrNew v fs + = do + vty <- vectType (idType v) + vv <- newLocalVVar fs vty + updLEnv (upd vv) + return vv + where + upd vv env = env { local_vars = extendVarEnv (local_vars env) v vv } + vectBndrIn :: Var -> VM a -> VM (VVar, a) vectBndrIn v p = localV @@ -137,6 +154,14 @@ vectBndrIn v p x <- p return (vv, x) +vectBndrNewIn :: Var -> FastString -> VM a -> VM (VVar, a) +vectBndrNewIn v fs p + = localV + $ do + vv <- vectBndrNew v fs + x <- p + return (vv, x) + vectBndrsIn :: [Var] -> VM a -> VM ([VVar], a) vectBndrsIn vs p = localV @@ -156,7 +181,7 @@ vectVar v Local (vv,lv) -> return (Var vv, Var lv) Global vv -> do let vexpr = Var vv - lexpr <- liftPA vexpr + lexpr <- liftPD vexpr return (vexpr, lexpr) vectPolyVar :: Var -> [Type] -> VM VExpr @@ -169,24 +194,31 @@ vectPolyVar v tys (polyApply (Var lv) vtys) Global poly -> do vexpr <- polyApply (Var poly) vtys - lexpr <- liftPA vexpr + lexpr <- liftPD vexpr return (vexpr, lexpr) vectLiteral :: Literal -> VM VExpr vectLiteral lit = do - lexpr <- liftPA (Lit lit) + lexpr <- liftPD (Lit lit) return (Lit lit, lexpr) -vectPolyExpr :: CoreExprWithFVs -> VM VExpr -vectPolyExpr expr - = polyAbstract tvs $ \abstract -> - do - mono' <- vectExpr mono - return $ mapVect abstract mono' +vectPolyExpr :: Bool -> CoreExprWithFVs -> VM (Inline, VExpr) +vectPolyExpr loop_breaker (_, AnnNote note expr) + = do + (inline, expr') <- vectPolyExpr loop_breaker expr + return (inline, vNote note expr') +vectPolyExpr loop_breaker expr + = do + arity <- polyArity tvs + polyAbstract tvs $ \args -> + do + (inline, mono') <- vectFnExpr False loop_breaker mono + return (addInlineArity inline arity, + mapVect (mkLams $ tvs ++ args) mono') where - (tvs, mono) = collectAnnTypeBinders expr - + (tvs, mono) = collectAnnTypeBinders expr + vectExpr :: CoreExprWithFVs -> VM VExpr vectExpr (_, AnnType ty) = liftM vType (vectType ty) @@ -204,18 +236,37 @@ vectExpr e@(_, AnnApp _ arg) where (fn, tys) = collectAnnTypeArgs e +vectExpr (_, AnnApp (_, AnnVar v) (_, AnnLit lit)) + | Just con <- isDataConId_maybe v + , is_special_con con + = do + let vexpr = App (Var v) (Lit lit) + lexpr <- liftPD vexpr + return (vexpr, lexpr) + where + is_special_con con = con `elem` [intDataCon, floatDataCon, doubleDataCon] + + vectExpr (_, AnnApp fn arg) = do - fn' <- vectExpr fn - arg' <- vectExpr arg - mkClosureApp fn' arg' + arg_ty' <- vectType arg_ty + res_ty' <- vectType res_ty + fn' <- vectExpr fn + arg' <- vectExpr arg + mkClosureApp arg_ty' res_ty' fn' arg' + where + (arg_ty, res_ty) = splitFunTy . exprType $ deAnnotate fn -vectExpr (_, AnnCase expr bndr ty alts) - = panic "vectExpr: case" +vectExpr (_, AnnCase scrut bndr ty alts) + | Just (tycon, ty_args) <- splitTyConApp_maybe scrut_ty + , isAlgTyCon tycon + = vectAlgCase tycon ty_args scrut bndr ty alts + where + scrut_ty = exprType (deAnnotate scrut) vectExpr (_, AnnLet (AnnNonRec bndr rhs) body) = do - vrhs <- localV . inBind bndr $ vectPolyExpr rhs + vrhs <- localV . inBind bndr . liftM snd $ vectPolyExpr False rhs (vbndr, vbody) <- vectBndrIn bndr (vectExpr body) return $ vLet (vNonRec vbndr vrhs) vbody @@ -224,23 +275,72 @@ vectExpr (_, AnnLet (AnnRec bs) body) (vbndrs, (vrhss, vbody)) <- vectBndrsIn bndrs $ liftM2 (,) (zipWithM vect_rhs bndrs rhss) - (vectPolyExpr body) + (vectExpr body) return $ vLet (vRec vbndrs vrhss) vbody where (bndrs, rhss) = unzip bs vect_rhs bndr rhs = localV . inBind bndr - $ vectExpr rhs + . liftM snd + $ vectPolyExpr (isLoopBreaker $ idOccInfo bndr) rhs + +vectExpr e@(_, AnnLam bndr _) + | isId bndr = liftM snd $ vectFnExpr True False e +{- +onlyIfV (isEmptyVarSet fvs) (vectScalarLam bs $ deAnnotate body) + `orElseV` vectLam True fvs bs body + where + (bs,body) = collectAnnValBinders e +-} -vectExpr e@(fvs, AnnLam bndr _) - | not (isId bndr) = pprPanic "vectExpr" (ppr $ deAnnotate e) - | otherwise = vectLam fvs bs body +vectExpr e = cantVectorise "Can't vectorise expression" (ppr $ deAnnotate e) + +vectFnExpr :: Bool -> Bool -> CoreExprWithFVs -> VM (Inline, VExpr) +vectFnExpr inline loop_breaker e@(fvs, AnnLam bndr _) + | isId bndr = onlyIfV (isEmptyVarSet fvs) + (mark DontInline . vectScalarLam bs $ deAnnotate body) + `orElseV` mark inlineMe (vectLam inline loop_breaker fvs bs body) where (bs,body) = collectAnnValBinders e +vectFnExpr _ _ e = mark DontInline $ vectExpr e + +mark :: Inline -> VM a -> VM (Inline, a) +mark b p = do { x <- p; return (b,x) } + +vectScalarLam :: [Var] -> CoreExpr -> VM VExpr +vectScalarLam args body + = do + scalars <- globalScalars + onlyIfV (all is_scalar_ty arg_tys + && is_scalar_ty res_ty + && is_scalar (extendVarSetList scalars args) body) + $ do + fn_var <- hoistExpr (fsLit "fn") (mkLams args body) DontInline + zipf <- zipScalars arg_tys res_ty + clo <- scalarClosure arg_tys res_ty (Var fn_var) + (zipf `App` Var fn_var) + clo_var <- hoistExpr (fsLit "clo") clo DontInline + lclo <- liftPD (Var clo_var) + return (Var clo_var, lclo) + where + arg_tys = map idType args + res_ty = exprType body + + is_scalar_ty ty | Just (tycon, []) <- splitTyConApp_maybe ty + = tycon == intTyCon + || tycon == floatTyCon + || tycon == doubleTyCon + + | otherwise = False -vectLam :: VarSet -> [Var] -> CoreExprWithFVs -> VM VExpr -vectLam fvs bs body + is_scalar vs (Var v) = v `elemVarSet` vs + is_scalar _ e@(Lit _) = is_scalar_ty $ exprType e + is_scalar vs (App e1 e2) = is_scalar vs e1 && is_scalar vs e2 + is_scalar _ _ = False + +vectLam :: Bool -> Bool -> VarSet -> [Var] -> CoreExprWithFVs -> VM VExpr +vectLam inline loop_breaker fvs bs body = do tyvars <- localTyVars (vs, vvs) <- readLEnv $ \env -> @@ -251,14 +351,179 @@ vectLam fvs bs body res_ty <- vectType (exprType $ deAnnotate body) buildClosures tyvars vvs arg_tys res_ty - . hoistPolyVExpr tyvars + . hoistPolyVExpr tyvars (maybe_inline (length vs + length bs)) $ do lc <- builtin liftingContext (vbndrs, vbody) <- vectBndrsIn (vs ++ bs) (vectExpr body) - return $ vLams lc vbndrs vbody - + vbody' <- break_loop lc res_ty vbody + return $ vLams lc vbndrs vbody' + where + maybe_inline n | inline = Inline n + | otherwise = DontInline + + break_loop lc ty (ve, le) + | loop_breaker + = do + empty <- emptyPD ty + lty <- mkPDataType ty + return (ve, mkWildCase (Var lc) intPrimTy lty + [(DEFAULT, [], le), + (LitAlt (mkMachInt 0), [], empty)]) + + | otherwise = return (ve, le) + + vectTyAppExpr :: CoreExprWithFVs -> [Type] -> VM VExpr vectTyAppExpr (_, AnnVar v) tys = vectPolyVar v tys -vectTyAppExpr e tys = pprPanic "vectTyAppExpr" (ppr $ deAnnotate e) +vectTyAppExpr e tys = cantVectorise "Can't vectorise expression" + (ppr $ deAnnotate e `mkTyApps` tys) + +-- We convert +-- +-- case e :: t of v { ... } +-- +-- to +-- +-- V: let v' = e in case v' of _ { ... } +-- L: let v' = e in case v' `cast` ... of _ { ... } +-- +-- When lifting, we have to do it this way because v must have the type +-- [:V(T):] but the scrutinee must be cast to the representation type. We also +-- have to handle the case where v is a wild var correctly. +-- + +-- FIXME: this is too lazy +vectAlgCase :: TyCon -> [Type] -> CoreExprWithFVs -> Var -> Type + -> [(AltCon, [Var], CoreExprWithFVs)] + -> VM VExpr +vectAlgCase _tycon _ty_args scrut bndr ty [(DEFAULT, [], body)] + = do + vscrut <- vectExpr scrut + (vty, lty) <- vectAndLiftType ty + (vbndr, vbody) <- vectBndrIn bndr (vectExpr body) + return $ vCaseDEFAULT vscrut vbndr vty lty vbody + +vectAlgCase _tycon _ty_args scrut bndr ty [(DataAlt _, [], body)] + = do + vscrut <- vectExpr scrut + (vty, lty) <- vectAndLiftType ty + (vbndr, vbody) <- vectBndrIn bndr (vectExpr body) + return $ vCaseDEFAULT vscrut vbndr vty lty vbody + +vectAlgCase _tycon _ty_args scrut bndr ty [(DataAlt dc, bndrs, body)] + = do + (vty, lty) <- vectAndLiftType ty + vexpr <- vectExpr scrut + (vbndr, (vbndrs, (vect_body, lift_body))) + <- vect_scrut_bndr + . vectBndrsIn bndrs + $ vectExpr body + let (vect_bndrs, lift_bndrs) = unzip vbndrs + (vscrut, lscrut, pdata_tc, _arg_tys) <- mkVScrut (vVar vbndr) + vect_dc <- maybeV (lookupDataCon dc) + let [pdata_dc] = tyConDataCons pdata_tc + + let vcase = mk_wild_case vscrut vty vect_dc vect_bndrs vect_body + lcase = mk_wild_case lscrut lty pdata_dc lift_bndrs lift_body + + return $ vLet (vNonRec vbndr vexpr) (vcase, lcase) + where + vect_scrut_bndr | isDeadBinder bndr = vectBndrNewIn bndr (fsLit "scrut") + | otherwise = vectBndrIn bndr + + mk_wild_case expr ty dc bndrs body + = mkWildCase expr (exprType expr) ty [(DataAlt dc, bndrs, body)] + +vectAlgCase tycon _ty_args scrut bndr ty alts + = do + vect_tc <- maybeV (lookupTyCon tycon) + (vty, lty) <- vectAndLiftType ty + + let arity = length (tyConDataCons vect_tc) + sel_ty <- builtin (selTy arity) + sel_bndr <- newLocalVar (fsLit "sel") sel_ty + let sel = Var sel_bndr + + (vbndr, valts) <- vect_scrut_bndr + $ mapM (proc_alt arity sel vty lty) alts' + let (vect_dcs, vect_bndrss, lift_bndrss, vbodies) = unzip4 valts + + vexpr <- vectExpr scrut + (vect_scrut, lift_scrut, pdata_tc, _arg_tys) <- mkVScrut (vVar vbndr) + let [pdata_dc] = tyConDataCons pdata_tc + + let (vect_bodies, lift_bodies) = unzip vbodies + + vdummy <- newDummyVar (exprType vect_scrut) + ldummy <- newDummyVar (exprType lift_scrut) + let vect_case = Case vect_scrut vdummy vty + (zipWith3 mk_vect_alt vect_dcs vect_bndrss vect_bodies) + + lc <- builtin liftingContext + lbody <- combinePD vty (Var lc) sel lift_bodies + let lift_case = Case lift_scrut ldummy lty + [(DataAlt pdata_dc, sel_bndr : concat lift_bndrss, + lbody)] + + return . vLet (vNonRec vbndr vexpr) + $ (vect_case, lift_case) + where + vect_scrut_bndr | isDeadBinder bndr = vectBndrNewIn bndr (fsLit "scrut") + | otherwise = vectBndrIn bndr + + alts' = sortBy (\(alt1, _, _) (alt2, _, _) -> cmp alt1 alt2) alts + + cmp (DataAlt dc1) (DataAlt dc2) = dataConTag dc1 `compare` dataConTag dc2 + cmp DEFAULT DEFAULT = EQ + cmp DEFAULT _ = LT + cmp _ DEFAULT = GT + cmp _ _ = panic "vectAlgCase/cmp" + + proc_alt arity sel _ lty (DataAlt dc, bndrs, body) + = do + vect_dc <- maybeV (lookupDataCon dc) + let ntag = dataConTagZ vect_dc + tag = mkDataConTag vect_dc + fvs = freeVarsOf body `delVarSetList` bndrs + + sel_tags <- liftM (`App` sel) (builtin (selTags arity)) + lc <- builtin liftingContext + elems <- builtin (selElements arity ntag) + + (vbndrs, vbody) + <- vectBndrsIn bndrs + . localV + $ do + binds <- mapM (pack_var (Var lc) sel_tags tag) + . filter isLocalId + $ varSetElems fvs + (ve, le) <- vectExpr body + return (ve, Case (elems `App` sel) lc lty + [(DEFAULT, [], (mkLets (concat binds) le))]) + -- empty <- emptyPD vty + -- return (ve, Case (elems `App` sel) lc lty + -- [(DEFAULT, [], Let (NonRec flags_var flags_expr) + -- $ mkLets (concat binds) le), + -- (LitAlt (mkMachInt 0), [], empty)]) + let (vect_bndrs, lift_bndrs) = unzip vbndrs + return (vect_dc, vect_bndrs, lift_bndrs, vbody) + + proc_alt _ _ _ _ _ = panic "vectAlgCase/proc_alt" + + mk_vect_alt vect_dc bndrs body = (DataAlt vect_dc, bndrs, body) + + pack_var len tags t v + = do + r <- lookupVar v + case r of + Local (vv, lv) -> + do + lv' <- cloneVar lv + expr <- packByTagPD (idType vv) (Var lv) len tags t + updLEnv (\env -> env { local_vars = extendVarEnv + (local_vars env) v (vv, lv') }) + return [(NonRec lv' expr)] + + _ -> return []