-module Vectorise( vectorise )
-where
-
-#include "HsVersions.h"
-
-import VectMonad
-import VectUtils
-import VectType
-import VectCore
+{-# OPTIONS -fno-warn-missing-signatures -fno-warn-unused-do-bind #-}
-import DynFlags
-import HscTypes
+module Vectorise ( vectorise )
+where
-import CoreLint ( showPass, endPass )
-import CoreSyn
-import CoreUtils
+import Vectorise.Type.Env
+import Vectorise.Type.Type
+import Vectorise.Convert
+import Vectorise.Utils.Hoisting
+import Vectorise.Exp
+import Vectorise.Vect
+import Vectorise.Env
+import Vectorise.Monad
+
+import HscTypes hiding ( MonadThings(..) )
+import CoreUnfold ( mkInlineUnfolding )
import CoreFVs
-import SimplMonad ( SimplCount, zeroSimplCount )
-import Rules ( RuleBase )
-import DataCon
-import TyCon
+import PprCore
+import CoreSyn
+import CoreMonad ( CoreM, getHscEnv )
import Type
-import FamInstEnv ( extendFamInstEnvList )
-import InstEnv ( extendInstEnvList )
-import Var
-import VarEnv
-import VarSet
-import Name ( Name, mkSysTvName, getName )
-import NameEnv
import Id
-import MkId ( unwrapFamInstScrut )
import OccName
-import Module ( Module )
+import DynFlags
+import BasicTypes ( isLoopBreaker )
+import Outputable
+import Util ( zipLazy )
+import MonadUtils
-import DsMonad hiding (mapAndUnzipM)
-import DsUtils ( mkCoreTup, mkCoreTupTy )
+import Control.Monad
-import Literal ( Literal )
-import PrelNames
-import TysWiredIn
-import TysPrim ( intPrimTy )
-import BasicTypes ( Boxity(..) )
-import Outputable
-import FastString
-import Control.Monad ( liftM, liftM2, zipWithM, mapAndUnzipM )
+-- | Vectorise a single module.
+--
+vectorise :: ModGuts -> CoreM ModGuts
+vectorise guts
+ = do { hsc_env <- getHscEnv
+ ; liftIO $ vectoriseIO hsc_env guts
+ }
-builtin_PAs :: [(Name, Module, FastString)]
-builtin_PAs = [
- mk closureTyConName FSLIT("dPA_Clo")
- , mk intTyConName FSLIT("dPA_Int")
- ]
- ++ tups
- where
- mk name fs = (name, nDP_INSTANCES, fs)
+-- | Vectorise a single monad, given the dynamic compiler flags and HscEnv.
+--
+vectoriseIO :: HscEnv -> ModGuts -> IO ModGuts
+vectoriseIO hsc_env guts
+ = do { -- Get information about currently loaded external packages.
+ ; eps <- hscEPS hsc_env
- tups = mk_tup 0 : map mk_tup [2..3]
- mk_tup n = (getName $ tupleTyCon Boxed n, nDP_INSTANCES,
- mkFastString $ "dPA_" ++ show n)
+ -- Combine vectorisation info from the current module, and external ones.
+ ; let info = hptVectInfo hsc_env `plusVectInfo` eps_vect_info eps
-vectorise :: HscEnv -> UniqSupply -> RuleBase -> ModGuts
- -> IO (SimplCount, ModGuts)
-vectorise 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
+ -- Run the main VM computation.
+ ; Just (info', guts') <- initV hsc_env guts info (vectModule guts)
+ ; return (guts' { mg_vect_info = info' })
+ }
+-- | Vectorise a single module, in the VM monad.
+--
vectModule :: ModGuts -> VM ModGuts
-vectModule guts
- = do
- defTyConBuiltinPAs builtin_PAs
- (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')
-
+vectModule guts@(ModGuts { mg_types = types
+ , mg_binds = binds
+ , mg_fam_insts = fam_insts
+ })
+ = do { dumpOptVt Opt_D_dump_vt_trace "Before vectorisation" $
+ pprCoreBindings binds
+
+ -- Vectorise the type environment.
+ -- This may add new TyCons and DataCons.
+ ; (types', new_fam_insts, tc_binds) <- vectTypeEnv types
+
+ ; (_, fam_inst_env) <- readGEnv global_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 tc_binds : binds'
- , mg_fam_inst_env = fam_inst_env'
- , mg_fam_insts = mg_fam_insts guts ++ fam_insts
- }
+ -- Vectorise all the top level bindings.
+ ; binds' <- mapM vectTopBind binds
+
+ ; return $ guts { mg_types = types'
+ , mg_binds = Rec tc_binds : binds'
+ , mg_fam_inst_env = fam_inst_env
+ , mg_fam_insts = fam_insts ++ new_fam_insts
+ }
+ }
+
+-- | Try to vectorise a top-level binding.
+-- If it doesn't vectorise then return it unharmed.
+--
+-- For example, for the binding
+--
+-- @
+-- foo :: Int -> Int
+-- foo = \x -> x + x
+-- @
+--
+-- we get
+-- @
+-- foo :: Int -> Int
+-- foo = \x -> vfoo $: x
+--
+-- v_foo :: Closure void vfoo lfoo
+-- v_foo = closure vfoo lfoo void
+--
+-- vfoo :: Void -> Int -> Int
+-- vfoo = ...
+--
+-- lfoo :: PData Void -> PData Int -> PData Int
+-- lfoo = ...
+-- @
+--
+-- @vfoo@ is the "vectorised", or scalar, version that does the same as the original
+-- function foo, but takes an explicit environment.
+--
+-- @lfoo@ is the "lifted" version that works on arrays.
+--
+-- @v_foo@ combines both of these into a `Closure` that also contains the
+-- environment.
+--
+-- The original binding @foo@ is rewritten to call the vectorised version
+-- present in the closure.
+--
vectTopBind :: CoreBind -> VM CoreBind
vectTopBind b@(NonRec var expr)
- = do
- var' <- vectTopBinder var
- expr' <- vectTopRhs var expr
- hs <- takeHoisted
- return . Rec $ (var, expr) : (var', expr') : hs
+ = do { -- Vectorise the right-hand side, create an appropriate top-level binding and add it to
+ -- the vectorisation map.
+ ; (inline, isScalar, expr') <- vectTopRhs [] var expr
+ ; var' <- vectTopBinder var inline expr'
+ ; when isScalar $
+ addGlobalScalar var
+
+ -- We replace the original top-level binding by a value projected from the vectorised
+ -- closure and add any newly created hoisted top-level bindings.
+ ; cexpr <- tryConvert var var' expr
+ ; hs <- takeHoisted
+ ; return . Rec $ (var, cexpr) : (var', expr') : hs
+ }
`orElseV`
return b
-
vectTopBind b@(Rec bs)
- = do
- vars' <- mapM vectTopBinder vars
- exprs' <- zipWithM vectTopRhs vars exprs
- hs <- takeHoisted
- return . Rec $ bs ++ zip vars' exprs' ++ hs
+ = let (vars, exprs) = unzip bs
+ in
+ do { (vars', _, exprs', hs) <- fixV $
+ \ ~(_, inlines, rhss, _) ->
+ do { -- Vectorise the right-hand sides, create an appropriate top-level bindings and
+ -- add them to the vectorisation map.
+ ; vars' <- sequence [vectTopBinder var inline rhs
+ | (var, ~(inline, rhs)) <- zipLazy vars (zip inlines rhss)]
+ ; (inlines, areScalars, exprs') <- mapAndUnzip3M (uncurry $ vectTopRhs vars) bs
+ ; hs <- takeHoisted
+ ; if and areScalars
+ then -- (1) Entire recursive group is scalar
+ -- => add all variables to the global set of scalars
+ do { mapM addGlobalScalar vars
+ ; return (vars', inlines, exprs', hs)
+ }
+ else -- (2) At least one binding is not scalar
+ -- => vectorise again with empty set of local scalars
+ do { (inlines, _, exprs') <- mapAndUnzip3M (uncurry $ vectTopRhs []) bs
+ ; hs <- takeHoisted
+ ; return (vars', inlines, exprs', hs)
+ }
+ }
+
+ -- Replace the original top-level bindings by a values projected from the vectorised
+ -- closures and add any newly created hoisted top-level bindings to the group.
+ ; 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
- = do
- vty <- vectType (idType var)
- var' <- cloneId mkVectOcc var vty
- defGlobalVar var var'
- return var'
+ return b
-vectTopRhs :: Var -> CoreExpr -> VM CoreExpr
-vectTopRhs var expr
- = do
- closedV . liftM vectorised
- . inBind var
- $ vectPolyExpr (freeVars expr)
-
--- ----------------------------------------------------------------------------
--- Bindings
-
-vectBndr :: Var -> VM VVar
-vectBndr v
- = do
- vty <- vectType (idType v)
- lty <- mkPArrayType vty
- let vv = v `Id.setIdType` vty
- lv = v `Id.setIdType` lty
- updLEnv (mapTo vv lv)
- return (vv, lv)
- where
- mapTo vv lv env = env { local_vars = extendVarEnv (local_vars env) v (vv, lv) }
-
-vectBndrIn :: Var -> VM a -> VM (VVar, a)
-vectBndrIn v p
- = localV
- $ do
- vv <- vectBndr v
- x <- p
- return (vv, x)
-
-vectBndrIn' :: Var -> (VVar -> VM a) -> VM (VVar, a)
-vectBndrIn' v p
- = localV
- $ do
- vv <- vectBndr v
- x <- p vv
- return (vv, x)
-
-vectBndrsIn :: [Var] -> VM a -> VM ([VVar], a)
-vectBndrsIn vs p
- = localV
- $ do
- vvs <- mapM vectBndr vs
- x <- p
- return (vvs, x)
-
--- ----------------------------------------------------------------------------
--- Expressions
-
-vectVar :: Var -> VM VExpr
-vectVar v
- = do
- r <- lookupVar v
- case r of
- Local (vv,lv) -> return (Var vv, Var lv)
- Global vv -> do
- let vexpr = Var vv
- lexpr <- liftPA vexpr
- return (vexpr, lexpr)
-
-vectPolyVar :: Var -> [Type] -> VM VExpr
-vectPolyVar v tys
- = do
- vtys <- mapM vectType tys
- r <- lookupVar v
- case r of
- Local (vv, lv) -> liftM2 (,) (polyApply (Var vv) vtys)
- (polyApply (Var lv) vtys)
- Global poly -> do
- vexpr <- polyApply (Var poly) vtys
- lexpr <- liftPA vexpr
- return (vexpr, lexpr)
-
-vectLiteral :: Literal -> VM VExpr
-vectLiteral lit
- = do
- lexpr <- liftPA (Lit lit)
- return (Lit lit, lexpr)
-
-vectPolyExpr :: CoreExprWithFVs -> VM VExpr
-vectPolyExpr expr
- = polyAbstract tvs $ \abstract ->
- do
- mono' <- vectExpr mono
- return $ mapVect abstract mono'
- where
- (tvs, mono) = collectAnnTypeBinders expr
-
-vectExpr :: CoreExprWithFVs -> VM VExpr
-vectExpr (_, AnnType ty)
- = liftM vType (vectType ty)
-
-vectExpr (_, AnnVar v) = vectVar v
-
-vectExpr (_, AnnLit lit) = vectLiteral lit
-
-vectExpr (_, AnnNote note expr)
- = liftM (vNote note) (vectExpr expr)
-
-vectExpr e@(_, AnnApp _ arg)
- | isAnnTypeArg arg
- = vectTyAppExpr fn tys
- where
- (fn, tys) = collectAnnTypeArgs e
-
-vectExpr (_, AnnApp fn arg)
- = do
- fn' <- vectExpr fn
- arg' <- vectExpr arg
- mkClosureApp fn' arg'
-
-vectExpr (_, AnnCase scrut bndr ty alts)
- | isAlgType scrut_ty
- = vectAlgCase scrut bndr ty alts
- where
- scrut_ty = exprType (deAnnotate scrut)
-
-vectExpr (_, AnnCase expr bndr ty alts)
- = panic "vectExpr: case"
-
-vectExpr (_, AnnLet (AnnNonRec bndr rhs) body)
- = do
- vrhs <- localV . inBind bndr $ vectPolyExpr rhs
- (vbndr, vbody) <- vectBndrIn bndr (vectExpr body)
- return $ vLet (vNonRec vbndr vrhs) vbody
-
-vectExpr (_, AnnLet (AnnRec bs) body)
- = do
- (vbndrs, (vrhss, vbody)) <- vectBndrsIn bndrs
- $ liftM2 (,)
- (zipWithM vect_rhs bndrs rhss)
- (vectPolyExpr body)
- return $ vLet (vRec vbndrs vrhss) vbody
- where
- (bndrs, rhss) = unzip bs
-
- vect_rhs bndr rhs = localV
- . inBind bndr
- $ vectExpr rhs
-
-vectExpr e@(fvs, AnnLam bndr _)
- | not (isId bndr) = pprPanic "vectExpr" (ppr $ deAnnotate e)
- | otherwise = vectLam fvs bs body
+-- | Make the vectorised version of this top level binder, and add the mapping
+-- between it and the original to the state. For some binder @foo@ the vectorised
+-- version is @$v_foo@
+--
+-- NOTE: vectTopBinder *MUST* be lazy in inline and expr because of how it is
+-- used inside of fixV in vectTopBind
+--
+vectTopBinder :: Var -- ^ Name of the binding.
+ -> Inline -- ^ Whether it should be inlined, used to annotate it.
+ -> CoreExpr -- ^ RHS of binding, used to set the 'Unfolding' of the returned 'Var'.
+ -> VM Var -- ^ Name of the vectorised binding.
+vectTopBinder var inline expr
+ = do { -- Vectorise the type attached to the var.
+ ; vty <- vectType (idType var)
+
+ -- If there is a vectorisation declartion for this binding, make sure that its type
+ -- matches
+ ; vectDecl <- lookupVectDecl var
+ ; case vectDecl of
+ Nothing -> return ()
+ Just (vdty, _)
+ | eqType vty vdty -> return ()
+ | otherwise ->
+ cantVectorise ("Type mismatch in vectorisation pragma for " ++ show var) $
+ (text "Expected type" <+> ppr vty)
+ $$
+ (text "Inferred type" <+> ppr vdty)
+
+ -- Make the vectorised version of binding's name, and set the unfolding used for inlining
+ ; var' <- liftM (`setIdUnfoldingLazily` unfolding)
+ $ cloneId mkVectOcc var vty
+
+ -- Add the mapping between the plain and vectorised name to the state.
+ ; defGlobalVar var var'
+
+ ; return var'
+ }
where
- (bs,body) = collectAnnValBinders e
-
-vectLam :: VarSet -> [Var] -> CoreExprWithFVs -> VM VExpr
-vectLam fvs bs body
- = do
- tyvars <- localTyVars
- (vs, vvs) <- readLEnv $ \env ->
- unzip [(var, vv) | var <- varSetElems fvs
- , Just vv <- [lookupVarEnv (local_vars env) var]]
-
- arg_tys <- mapM (vectType . idType) bs
- res_ty <- vectType (exprType $ deAnnotate body)
+ unfolding = case inline of
+ Inline arity -> mkInlineUnfolding (Just arity) expr
+ DontInline -> noUnfolding
- buildClosures tyvars vvs arg_tys res_ty
- . hoistPolyVExpr tyvars
- $ do
- lc <- builtin liftingContext
- (vbndrs, vbody) <- vectBndrsIn (vs ++ bs)
- (vectExpr body)
- return $ vLams lc vbndrs vbody
-
-vectTyAppExpr :: CoreExprWithFVs -> [Type] -> VM VExpr
-vectTyAppExpr (_, AnnVar v) tys = vectPolyVar v tys
-vectTyAppExpr e tys = pprPanic "vectTyAppExpr" (ppr $ deAnnotate e)
-
-type CoreAltWithFVs = AnnAlt Id VarSet
-
--- We convert
+-- | Vectorise the RHS of a top-level binding, in an empty local environment.
--
--- case e :: t of v { ... }
+-- We need to distinguish three cases:
--
--- to
+-- (1) We have a (non-scalar) vectorisation declaration for the variable (which explicitly provides
+-- vectorised code implemented by the user)
+-- => no automatic vectorisation & instead use the user-supplied code
+--
+-- (2) We have a scalar vectorisation declaration for the variable
+-- => generate vectorised code that uses a scalar 'map'/'zipWith' to lift the computation
+--
+-- (3) There is no vectorisation declaration for the variable
+-- => perform automatic vectorisation of the RHS
--
--- V: let v = e in case v of _ { ... }
--- L: let v = e in case v `cast` ... of _ { ... }
+vectTopRhs :: [Var] -- ^ Names of all functions in the rec block
+ -> Var -- ^ Name of the binding.
+ -> CoreExpr -- ^ Body of the binding.
+ -> VM ( Inline -- (1) inline specification for the binding
+ , Bool -- (2) whether the right-hand side is a scalar computation
+ , CoreExpr) -- (3) the vectorised right-hand side
+vectTopRhs recFs var expr
+ = closedV
+ $ do { traceVt ("vectTopRhs of " ++ show var) $ ppr expr
+
+ ; globalScalar <- isGlobalScalar var
+ ; vectDecl <- lookupVectDecl var
+ ; rhs globalScalar vectDecl
+ }
+ where
+ rhs _globalScalar (Just (_, expr')) -- Case (1)
+ = return (inlineMe, False, expr')
+ rhs True Nothing -- Case (2)
+ = do { expr' <- vectScalarFun True recFs expr
+ ; return (inlineMe, True, vectorised expr')
+ }
+ rhs False Nothing -- Case (3)
+ = do { let fvs = freeVars expr
+ ; (inline, isScalar, vexpr) <- inBind var $
+ vectPolyExpr (isLoopBreaker $ idOccInfo var) recFs fvs
+ ; return (inline, isScalar, vectorised vexpr)
+ }
+
+-- | Project out the vectorised version of a binding from some closure,
+-- or return the original body if that doesn't work or the binding is scalar.
--
--- 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.
---
-
--- FIXME: this is too lazy
-vectAlgCase scrut bndr ty [(DEFAULT, [], body)]
- = do
- vscrut <- vectExpr scrut
- vty <- vectType ty
- lty <- mkPArrayType vty
- (vbndr, vbody) <- vectBndrIn bndr (vectExpr body)
- return $ vCaseDEFAULT vscrut vbndr vty lty vbody
-
-vectAlgCase scrut bndr ty [(DataAlt dc, bndrs, body)]
- = do
- vty <- vectType ty
- lty <- mkPArrayType vty
- vexpr <- vectExpr scrut
- (vbndr, (vbndrs, vbody)) <- vectBndrIn bndr
- . vectBndrsIn bndrs
- $ vectExpr body
-
- (vscrut, arr_tc, arg_tys) <- mkVScrut (vVar vbndr)
- vect_dc <- maybeV (lookupDataCon dc)
- let [arr_dc] = tyConDataCons arr_tc
- let shape_tys = take (dataConRepArity arr_dc - length bndrs)
- (dataConRepArgTys arr_dc)
- shape_bndrs <- mapM (newLocalVar FSLIT("s")) shape_tys
- return . vLet (vNonRec vbndr vexpr)
- $ vCaseProd vscrut vty lty vect_dc arr_dc shape_bndrs vbndrs vbody
+tryConvert :: Var -- ^ Name of the original binding (eg @foo@)
+ -> Var -- ^ Name of vectorised version of binding (eg @$vfoo@)
+ -> CoreExpr -- ^ The original body of the binding.
+ -> VM CoreExpr
+tryConvert var vect_var rhs
+ = do { globalScalar <- isGlobalScalar var
+ ; if globalScalar
+ then
+ return rhs
+ else
+ fromVect (idType var) (Var vect_var) `orElseV` return rhs
+ }