haskell - 长度索引链表的融合

标签 haskell optimization

我有标准Vect类型:

module Vect where

data Nat = Z | S Nat
data Vect (n :: Nat) (a :: Type) where
  VNil :: Vect Z a
  VCons :: a -> Vect n a -> Vect (S n) a

我有这些功能:
foldVect :: forall (p :: Nat -> Type) n a.
            p Z ->
            (forall m. a -> p m -> p (S m)) ->
            Vect n a -> p n
foldVect n c = go
  where go :: forall l. Vect l a -> p l
        go VNil = n
        go (VCons x xs) = x `c` go xs

newtype FVect a n = FVect { unFVect :: Vect n a }
buildVect :: forall n a.
             (forall (p :: Nat -> Type).
              p Z ->
              (forall m. a -> p m -> p (S m)) ->
              p n
             ) -> Vect n a
buildVect f = unFVect $ f (FVect VNil) $ \x (FVect xs) -> FVect $ x `VCons` xs

我试图从 base 重新创建(部分)机器允许列表融合:
instance Functor (Vect n) where
    fmap = mapVect
    {-# INLINE fmap #-}

mapVect :: forall n a b. (a -> b) -> (Vect n a -> Vect n b)
mapVect _f VNil = VNil
mapVect f (VCons x xs) = f x `VCons` mapVect f xs
mapFB :: forall (p :: Nat -> Type) n a b. (forall m. b -> p m -> p (S m)) -> (a -> b) -> a -> p n -> p (S n)
mapFB cons f = \x ys -> cons (f x) ys
{-# INLINE [0] mapFB #-}
{-# NOINLINE [0] mapVect #-}
{-# RULES "mapVect" [~1] forall f xs. mapVect f xs = buildVect (\nil cons -> foldVect nil (mapFB cons f) xs) #-}

{-# INLINE [0] foldVect #-}
-- base has this; I don't think it does anything without a "refolding" rule on mapVect
{-# INLINE [0] buildVect #-}
{-# RULES "foldVect/buildVect" forall (nil :: p Z)
                                      (cons :: forall m. a -> p m -> p (S m))
                                      (f :: forall (q :: Nat -> Type).
                                            q Z ->
                                            (forall m. a -> q m -> q (S m)) ->
                                            q n
                                      ).
          foldVect nil cons (buildVect f) = f nil cons
  #-}

进而
module Test where
import Vect

test :: Vect n Int -> Vect n Int
test = fmap (*5) . fmap (+2)

不会发生融合。如果我通过 -ddump-simpl ,我看到了 foldVectbuildVect已经被内联了,但是......
  • INLINE s 是分阶段的,因此它们无论如何都不会干扰融合。 (毕竟,这就是 base[] 所做的)
  • 更换 INLINE [0] s 与 NOINLINE描绘了一个相当惊人的图像:

    test
      = \ (@ (n_a141 :: Nat)) (x_X1lK :: Vect n_a141 Int) ->
          buildVect
            @ n_a141
            @ Int
            (\ (@ (p_X1jl :: Nat -> *))
               (nil_X11K [OS=OneShot] :: p_X1jl 'Z)
               (cons_X11M [OS=OneShot]
                  :: forall (m :: Nat). Int -> p_X1jl m -> p_X1jl ('S m)) ->
               foldVect
                 @ p_X1jl
                 @ n_a141
                 @ Int
                 nil_X11K
                 (\ (@ (m_a1i5 :: Nat))
                    (x1_aYI :: Int)
                    (ys_aYJ [OS=OneShot] :: p_X1jl m_a1i5) ->
                    cons_X11M
                      @ m_a1i5
                      (case x1_aYI of { GHC.Types.I# x2_a1l5 ->
                       GHC.Types.I# (GHC.Prim.*# x2_a1l5 5#)
                       })
                      ys_aYJ)
                 (buildVect
                    @ n_a141
                    @ Int
                    (\ (@ (p1_a1i0 :: Nat -> *))
                       (nil1_a10o [OS=OneShot] :: p1_a1i0 'Z)
                       (cons1_a10p [OS=OneShot]
                          :: forall (m :: Nat). Int -> p1_a1i0 m -> p1_a1i0 ('S m)) ->
                       foldVect
                         @ p1_a1i0
                         @ n_a141
                         @ Int
                         nil1_a10o
                         (\ (@ (m_a1i5 :: Nat))
                            (x1_aYI :: Int)
                            (ys_aYJ [OS=OneShot] :: p1_a1i0 m_a1i5) ->
                            cons1_a10p
                              @ m_a1i5
                              (case x1_aYI of { GHC.Types.I# x2_a1lh ->
                               GHC.Types.I# (GHC.Prim.+# x2_a1lh 2#)
                               })
                              ys_aYJ)
                         x_X1lK)))
    

    一切都在那里,但简化器只是没有它。

  • 如果我检查规则本身,我会看到这个

    "foldVect/buildVect"
        forall (@ (p_aYG :: Nat -> *))
               (@ (n_aYJ :: Nat))
               (@ a_aYH)
               (nil_aYD :: p_aYG 'Z)
               (cons_aYE :: forall (m :: Nat). a_aYH -> p_aYG m -> p_aYG ('S m))
               (f_aYF
                  :: forall (q :: Nat -> *).
                     q 'Z -> (forall (m :: Nat). a_aYH -> q m -> q ('S m)) -> q n_aYJ).
          foldVect @ p_aYG
                   @ n_aYJ
                   @ a_aYH
                   nil_aYD
                   cons_aYE
                   (buildVect
                      @ n_aYJ
                      @ a_aYH
                      (\ (@ (p1_a156 :: Nat -> *))
                         (ds_d1io :: p1_a156 'Z)
                         (ds1_d1ip
                            :: forall (m :: Nat). a_aYH -> p1_a156 m -> p1_a156 ('S m)) ->
                         f_aYF @ p1_a156 ds_d1io ds1_d1ip))
          = f_aYF @ p_aYG nil_aYD cons_aYE
    

    看来问题在于 buildVect 的论点需要是一个非常特定形式的 lambda 抽象,我在构建一个最终发生的重写系统时遇到了麻烦。

    我如何让融合工作?

    (我不知道这是否有用,甚至是否正确;我只是想看看是否可以。)

    最佳答案

    像往常一样,newtype每当编译器被牛逼时,就可以节省一天:

    module Vect where
    -- everything else the same...
    newtype VectBuilder n a = VectBuilder { runVectBuilder :: forall (p :: Nat -> Type).
                                                              p Z ->
                                                              (forall m. a -> p m -> p (S m)) ->
                                                              p n
                                          }
    
    buildVect' :: forall n a. VectBuilder n a -> Vect n a
    buildVect' f = unFVect $
                    runVectBuilder f (FVect VNil) $ \x (FVect xs) -> FVect $ x `VCons` xs
    {-# INLINE [0] buildVect' #-}
    buildVect :: forall n a.
                 (forall (p :: Nat -> Type).
                  p Z ->
                  (forall m. a -> p m -> p (S m)) ->
                  p n
                 ) -> Vect n a
    buildVect f = buildVect' (VectBuilder f)
    {-# INLINE buildVect #-}
    
    {-# RULES "foldVect/buildVect'" forall (nil :: p Z)
                                           (cons :: forall m. a -> p m -> p (S m))
                                           (f :: VectBuilder n a).
                                    foldVect nil cons (buildVect' f) = runVectBuilder f nil cons
      #-}
    -- compiler no longer has a chance to muck up the LHS by eta expanding f because
    -- f "isn't" a function anymore
    
    -- rule for mapVect goes unchanged, so I guess that's evidence that this is totally transparent
    
    module Test where
    import Vect
    test :: Vect n Int -> Vect n Int
    test = fmap (*5) . fmap (+2)
    

    Rec {
    -- RHS size: {terms: 19, types: 31, coercions: 13, joins: 0/0}
    Test.test_go [Occ=LoopBreaker]
      :: forall (l :: Nat). Vect l Int -> FVect Int l
    [GblId, Arity=1, Caf=NoCafRefs, Str=<S,1*U>]
    Test.test_go
      = \ (@ (l_a14W :: Nat)) (ds_d1jk :: Vect l_a14W Int) ->
          case ds_d1jk of {
            VNil co_a14Y -> (Vect.$WVNil @ Int) `cast` <Co:4>;
            VCons @ n2_a151 co_a152 x_aYE xs_aYF ->
              (Vect.VCons
                 @ ('S n2_a151)
                 @ Int
                 @ n2_a151
                 @~ <Co:2>
                 (case x_aYE of { GHC.Types.I# x1_a1xr ->
                  GHC.Types.I# (GHC.Prim.*# (GHC.Prim.+# x1_a1xr 2#) 5#) -- success!
                  })
                 ((Test.test_go @ n2_a151 xs_aYF) `cast` <Co:3>))
              `cast` <Co:4>
          }
    end Rec }
    
    -- RHS size: {terms: 4, types: 5, coercions: 0, joins: 0/0}
    Test.test1 :: forall (n :: Nat). Vect n Int -> FVect Int n
    [GblId,
     Arity=1,
     Caf=NoCafRefs,
     Str=<S,1*U>,
     Unf=Unf{Src=<vanilla>, TopLvl=True, Value=True, ConLike=True,
             WorkFree=True, Expandable=True,
             Guidance=ALWAYS_IF(arity=1,unsat_ok=True,boring_ok=True)}]
    Test.test1
      = \ (@ (n_a1wd :: Nat)) (x_X1xa :: Vect n_a1wd Int) ->
          Test.test_go @ n_a1wd x_X1xa
    
    -- RHS size: {terms: 1, types: 0, coercions: 9, joins: 0/0}
    test :: forall (n :: Nat). Vect n Int -> Vect n Int
    [GblId,
     Arity=1,
     Caf=NoCafRefs,
     Str=<S,1*U>,
     Unf=Unf{Src=<vanilla>, TopLvl=True, Value=True, ConLike=True,
             WorkFree=True, Expandable=True,
             Guidance=ALWAYS_IF(arity=0,unsat_ok=True,boring_ok=True)}]
    test = Test.test1 `cast` <Co:9>
    

    故事寓意:排名足够高的类型使 RULES系统内爆,所以给 GHC 一些帮助 newtype s,即使它们不是必需的。

    关于haskell - 长度索引链表的融合,我们在Stack Overflow上找到一个类似的问题: https://stackoverflow.com/questions/50844444/

    相关文章:

    haskell - 将字符串列表连接到一个字符串 haskell

    haskell - 避免在 Snap 中进行链式查找的 case 表达式

    haskell - 为什么 IntSet 查找是 O(min(n,W)),而不是 O(1)?

    c++ - 在 C++ 中保存和加载大数组

    java - 从 Iterator<?> 返回相同的对象

    optimization - 使用-fast编译时,波特兰组FORTRAN pgf90程序失败,-fast -Mnounroll成功时,

    haskell - 解构存在类型

    haskell - 递归函数组合中的惰性

    python - 模拟退火 - 直觉

    c - 在 C 中读取一个巨大的矩阵文件