Vectorise.hs 15.8 KB
Newer Older
1 2 3 4 5 6
-- Main entry point to the vectoriser.  It is invoked iff the option '-fvectorise' is passed.
--
-- This module provides the function 'vectorise', which vectorises an entire (desugared) module.
-- It vectorises all type declarations and value bindings.  It also processes all VECTORISE pragmas
-- (aka vectorisation declarations), which can lead to the vectorisation of imported data types
-- and the enrichment of imported functions with vectorised versions.
7

8
module Vectorise ( vectorise )
9 10
where

11 12 13
import Vectorise.Type.Env
import Vectorise.Type.Type
import Vectorise.Convert
14
import Vectorise.Utils.Hoisting
15
import Vectorise.Exp
16
import Vectorise.Vect
17
import Vectorise.Env
18
import Vectorise.Monad
19

20
import HscTypes hiding      ( MonadThings(..) )
21
import CoreUnfold           ( mkInlineUnfolding )
22
import CoreFVs
23 24
import PprCore
import CoreSyn
Ian Lynagh's avatar
Ian Lynagh committed
25
import CoreMonad            ( CoreM, getHscEnv )
26
import Type
rl@cse.unsw.edu.au's avatar
rl@cse.unsw.edu.au committed
27
import Id
28
import DynFlags
29
import BasicTypes           ( isStrongLoopBreaker )
30
import Outputable
31
import Util                 ( zipLazy )
32 33
import MonadUtils

34
import Control.Monad
35
import Data.Maybe
36 37


38
-- |Vectorise a single module.
39 40 41 42 43 44
--
vectorise :: ModGuts -> CoreM ModGuts
vectorise guts
 = do { hsc_env <- getHscEnv
      ; liftIO $ vectoriseIO hsc_env guts
      }
45

46
-- Vectorise a single monad, given the dynamic compiler flags and HscEnv.
47 48 49 50 51
--
vectoriseIO :: HscEnv -> ModGuts -> IO ModGuts
vectoriseIO hsc_env guts
 = do {   -- Get information about currently loaded external packages.
      ; eps <- hscEPS hsc_env
52

53 54
          -- Combine vectorisation info from the current module, and external ones.
      ; let info = hptVectInfo hsc_env `plusVectInfo` eps_vect_info eps
55

56 57 58 59
          -- Run the main VM computation.
      ; Just (info', guts') <- initV hsc_env guts info (vectModule guts)
      ; return (guts' { mg_vect_info = info' })
      }
60

61
-- Vectorise a single module, in the VM monad.
62
--
rl@cse.unsw.edu.au's avatar
rl@cse.unsw.edu.au committed
63
vectModule :: ModGuts -> VM ModGuts
64
vectModule guts@(ModGuts { mg_tcs        = tycons
65 66 67
                         , mg_binds      = binds
                         , mg_fam_insts  = fam_insts
                         , mg_vect_decls = vect_decls
68 69 70 71
                         })
 = do { dumpOptVt Opt_D_dump_vt_trace "Before vectorisation" $ 
          pprCoreBindings binds
 
72 73 74 75
          -- Pick out all 'VECTORISE type' and 'VECTORISE class' pragmas
      ; let ty_vect_decls  = [vd | vd@(VectType _ _ _) <- vect_decls]
            cls_vect_decls = [vd | vd@(VectClass _)    <- vect_decls]
      
76 77 78 79 80 81
          -- Vectorise the type environment.  This will add vectorised
          -- type constructors, their representaions, and the
          -- conrresponding data constructors.  Moreover, we produce
          -- bindings for dfuns and family instances of the classes
          -- and type families used in the DPH library to represent
          -- array types.
82
      ; (new_tycons, new_fam_insts, tc_binds) <- vectTypeEnv tycons ty_vect_decls cls_vect_decls
83

84 85
          -- Family instance environment for /all/ home-package modules including those instances
          -- generated by 'vectTypeEnv'.
86
      ; (_, fam_inst_env) <- readGEnv global_fam_inst_env
Ian Lynagh's avatar
Ian Lynagh committed
87

88
          -- Vectorise all the top level bindings and VECTORISE declarations on imported identifiers
89 90
      ; let impBinds = [imp_id | Vect          imp_id _ <- vect_decls, isGlobalId imp_id] ++
                       [imp_id | VectInst True imp_id   <- vect_decls, isGlobalId imp_id]
91
      ; binds_top <- mapM vectTopBind binds
92
      ; binds_imp <- mapM vectImpBind impBinds
rl@cse.unsw.edu.au's avatar
rl@cse.unsw.edu.au committed
93

94
      ; return $ guts { mg_tcs          = tycons ++ new_tycons
95 96
                        -- we produce no new classes or instances, only new class type constructors
                        -- and dfuns
97
                      , mg_binds        = Rec tc_binds : (binds_top ++ binds_imp)
98 99 100 101
                      , mg_fam_inst_env = fam_inst_env
                      , mg_fam_insts    = fam_insts ++ new_fam_insts
                      }
      }
102

103
-- Try to vectorise a top-level binding.  If it doesn't vectorise then return it unharmed.
104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142
--
-- 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.
--
-- Vectorisation may be surpressed by annotating a binding with a 'NOVECTORISE' pragma.  If this
-- pragma is used in a group of mutually recursive bindings, either all or no binding must have
-- the pragma.  If only some bindings are annotated, a fatal error is being raised.
-- FIXME: Once we support partial vectorisation, we may be able to vectorise parts of a group, or
--   we may emit a warning and refrain from vectorising the entire group.
143
--
144
vectTopBind :: CoreBind -> VM CoreBind
145
vectTopBind b@(NonRec var expr)
146 147 148 149 150 151 152 153 154 155 156 157 158 159
  = unlessNoVectDecl $
      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
         }
160 161 162 163
     `orElseErrV`
     do { emitVt "  Could NOT vectorise top-level binding" $ ppr var
        ; return b
        }
164 165 166 167 168 169 170
  where
    unlessNoVectDecl vectorise
      = do { hasNoVectDecl <- noVectDecl var
           ; when hasNoVectDecl $
               traceVt "NOVECTORISE" $ ppr var
           ; if hasNoVectDecl then return b else vectorise
           }
171
vectTopBind b@(Rec bs)
172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199
  = unlessSomeNoVectDecl $
      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
         }
200
     `orElseErrV`
201 202 203 204 205 206 207 208 209 210 211 212 213 214 215
       return b    
  where
    (vars, exprs) = unzip bs

    unlessSomeNoVectDecl vectorise
      = do { hasNoVectDecls <- mapM noVectDecl vars
           ; when (and hasNoVectDecls) $
               traceVt "NOVECTORISE" $ ppr vars
           ; if and hasNoVectDecls 
             then return b                              -- all bindings have 'NOVECTORISE'
             else if or hasNoVectDecls 
             then cantVectorise noVectoriseErr (ppr b)  -- some (but not all) have 'NOVECTORISE'
             else vectorise                             -- no binding has a 'NOVECTORISE' decl
           }
    noVectoriseErr = "NOVECTORISE must be used on all or no bindings of a recursive group"
216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234

-- Add a vectorised binding to an imported top-level variable that has a VECTORISE [SCALAR] pragma
-- in this module.
--
vectImpBind :: Id -> VM CoreBind
vectImpBind var
  = do {   -- Vectorise the right-hand side, create an appropriate top-level binding and add it
           -- to the vectorisation map.  For the non-lifted version, we refer to the original
           -- definition — i.e., 'Var var'.
       ; (inline, isScalar, expr') <- vectTopRhs [] var (Var var)
       ; var' <- vectTopBinder var inline expr'
       ; when isScalar $ 
           addGlobalScalar var

           -- We add any newly created hoisted top-level bindings.
       ; hs <- takeHoisted
       ; return . Rec $ (var', expr') : hs
       }

235 236 237 238
-- | 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@
--
239 240
--   NOTE: 'vectTopBinder' *MUST* be lazy in inline and expr because of how it is
--   used inside of 'fixV' in 'vectTopBind'.
241 242 243 244 245
--
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.
246
vectTopBinder var inline expr
247 248 249 250
 = 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
251
          -- matches
252 253
      ; vectDecl <- lookupVectDecl var
      ; case vectDecl of
254
          Nothing             -> return ()
255
          Just (vdty, _) 
256
            | eqType vty vdty -> return ()
257
            | otherwise       -> 
258 259 260 261 262 263 264
              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) 
265
                $  mkVectId var vty
266 267 268 269 270 271

          -- Add the mapping between the plain and vectorised name to the state.
      ; defGlobalVar var var'

      ; return var'
    }
272 273
  where
    unfolding = case inline of
274
                  Inline arity -> mkInlineUnfolding (Just arity) expr
275
                  DontInline   -> noUnfolding
276 277 278 279 280 281 282 283 284 285 286 287
{-
!!!TODO: dfuns and unfoldings:
           -- Do not inline the dfun; instead give it a magic DFunFunfolding
           -- See Note [ClassOp/DFun selection]
           -- See also note [Single-method classes]
        dfun_id_w_fun
           | isNewTyCon class_tc
           = dfun_id `setInlinePragma` alwaysInlinePragma { inl_sat = Just 0 }
           | otherwise
           = dfun_id `setIdUnfolding`  mkDFunUnfolding dfun_ty dfun_args
                     `setInlinePragma` dfunInlinePragma
 -}
Ian Lynagh's avatar
Ian Lynagh committed
288

289
-- | Vectorise the RHS of a top-level binding, in an empty local environment.
290
--
291
-- We need to distinguish four cases:
292 293 294 295 296
--
-- (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
-- 
297
-- (2) We have a scalar vectorisation declaration for a variable that is no dfun
298 299
--     => generate vectorised code that uses a scalar 'map'/'zipWith' to lift the computation
-- 
300 301 302 303
-- (3) We have a scalar vectorisation declaration for a variable that *is* a dfun
--     => generate vectorised code according to the the "Note [Scalar dfuns]" below
-- 
-- (4) There is no vectorisation declaration for the variable
304 305
--     => perform automatic vectorisation of the RHS
--
306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332
-- Note [Scalar dfuns]
-- ~~~~~~~~~~~~~~~~~~~
--
-- Here is the translation scheme for scalar dfuns — assume the instance declaration:
--
--   instance Num Int where
--     (+) = primAdd
--   {-# VECTORISE SCALAR instance Num Int #-}
--
-- It desugars to
--
--   $dNumInt :: Num Int
--   $dNumInt = D:Num primAdd
--
-- We vectorise it to
--
--   $v$dNumInt :: V:Num Int
--   $v$dNumInt = D:V:Num (closure2 ((+) $dNumInt) (scalar_zipWith ((+) $dNumInt))))
--
-- while adding the following entry to the vectorisation map: '$dNumInt' --> '$v$dNumInt'.
--
-- See "Note [Vectorising classes]" in 'Vectorise.Type.Env' for the definition of 'V:Num'.
--
-- NB: The outlined vectorisation scheme does not require the right-hand side of the original dfun.
--     In fact, we definitely want to refer to the dfn variable instead of the right-hand side to 
--     ensure that the dictionary selection rules fire.
--
333 334 335 336 337 338
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
339
vectTopRhs recFs var expr
340
  = closedV
341
  $ do { globalScalar <- isGlobalScalar var
342
       ; vectDecl     <- lookupVectDecl var
343
       ; let isDFun = isDFunId var
344

345
       ; traceVt ("vectTopRhs of " ++ show var ++ info globalScalar isDFun vectDecl) $ ppr expr
346

347
       ; rhs globalScalar isDFun vectDecl
348 349
       }
  where
350
    rhs _globalScalar _isDFun (Just (_, expr'))               -- Case (1)
351
      = return (inlineMe, False, expr')
352
    rhs True          False   Nothing                         -- Case (2)
353 354 355
      = do { expr' <- vectScalarFun True recFs expr
           ; return (inlineMe, True, vectorised expr')
           }
356 357 358 359 360
    rhs True          True    Nothing                         -- Case (3)
      = do { expr' <- vectScalarDFun var recFs
           ; return (DontInline, True, expr')
           }
    rhs False         _isDFun Nothing                         -- Case (4)
361
      = do { let fvs = freeVars expr
362 363 364
           ; (inline, isScalar, vexpr) 
               <- inBind var $
                    vectPolyExpr (isStrongLoopBreaker $ idOccInfo var) recFs fvs
365 366
           ; return (inline, isScalar, vectorised vexpr)
           }
367
    
368 369 370 371
    info True  False _                          = " [VECTORISE SCALAR]"
    info True  True  _                          = " [VECTORISE SCALAR instance]"
    info False _     vectDecl | isJust vectDecl = " [VECTORISE]"
                              | otherwise       = " (no pragma)"
372

373 374
-- |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. 
375 376 377 378 379
--
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
380
tryConvert var vect_var rhs
381 382 383 384 385
  = do { globalScalar <- isGlobalScalar var
       ; if globalScalar
         then
           return rhs
         else
386 387 388 389 390
           fromVect (idType var) (Var vect_var) 
           `orElseErrV` 
           do { emitVt "  Could NOT call vectorised from original version" $ ppr var
              ; return rhs
              }
391
       }