Total real-time persistent queues

问题:

Okasaki describes persistent real-time queues which can be realized in Haskell using the type

data Queue a = forall x . Queue
  { front :: [a]
  , rear :: [a]
  , schedule :: [x]
  }

where incremental rotations maintain the invariant

length schedule = length front - length rear

More details

If you're familiar with the queues involved, you can skip this section.

The rotation function looks like

rotate :: [a] -> [a] -> [a] -> [a]
rotate [] (y : _) a = y : a
rotate (x : xs) (y : ys) a =
  x : rotate xs ys (y : a)

and it's called by a smart constructor

exec :: [a] -> [a] -> [x] -> Queue a
exec f r (_ : s) = Queue f r s
exec f r [] = Queue f' [] f' where
  f' = rotate f r []

after each queue operation. The smart constructor is always called when length s = length f - length r + 1, ensuring that the pattern match in rotate will succeed.

The problem

I hate partial functions! I'd love to find a way to express the structural invariant in the types. The usual dependent vector seems a likely choice:

data Nat = Z | S Nat

data Vec n a where
  Nil :: Vec 'Z a
  Cons :: a -> Vec n a -> Vec ('S n) a

and then (perhaps)

data Queue a = forall x rl sl . Queue
  { front :: Vec (sl :+ rl) a
  , rear :: Vec rl a
  , schedule :: Vec sl x
  }

The trouble is that I haven't been able to figure out how to juggle the types. It seems extremely likely that some amount of unsafeCoerce will be needed to make this efficient. However, I haven't been able to come up with an approach that's even vaguely manageable. Is it possible to do this nicely in Haskell?

回答1:

Here is what I got:

open import Function
open import Data.Nat.Base
open import Data.Vec

grotate : ∀ {n m} {A : Set}
        -> (B : ℕ -> Set)
        -> (∀ {n} -> A -> B n -> B (suc n))
        -> Vec A n
        -> Vec A (suc n + m)
        -> B m
        -> B (suc n + m)
grotate B cons  []      (y ∷ ys) a = cons y a
grotate B cons (x ∷ xs) (y ∷ ys) a = grotate (B ∘ suc) cons xs ys (cons y a)

rotate : ∀ {n m} {A : Set} -> Vec A n -> Vec A (suc n + m) -> Vec A m -> Vec A (suc n + m)
rotate = grotate (Vec _) _∷_

record Queue (A : Set) : Set₁ where
  constructor queue
  field
    {X}      : Set
    {n m}    : ℕ
    front    : Vec A (n + m)
    rear     : Vec A m
    schedule : Vec X n

open import Relation.Binary.PropositionalEquality
open import Data.Nat.Properties.Simple

exec : ∀ {m n A} -> Vec A (n + m) -> Vec A (suc m) -> Vec A n -> Queue A
exec {m} {suc n} f r (_ ∷ s) = queue (subst (Vec _) (sym (+-suc n m)) f) r s
exec {m}         f r  []     = queue (with-zero f') [] f' where
 with-zero    = subst (Vec _ ∘ suc) (sym (+-right-identity m))
 without-zero = subst (Vec _ ∘ suc) (+-right-identity m)

 f' = without-zero (rotate f (with-zero r) [])

rotate is defined in terms of grotate for the same reason reverse is defined in terms of foldl (or enumerate in terms of genumerate): because Vec A (suc n + m) is not definitionally Vec A (n + suc m), while (B ∘ suc) m is definitionally B (suc m).

exec has the same implementation as you provided (modulo those substs), but I'm not sure about the types: is it OK that r must be non-empty?

回答2:

The other answer is super clever (please take a moment to upvote it), but as someone not familiar with Agda, how this would be implemented in Haskell was not obvious to me. Here's a full Haskell version. We'll need a whole slew of extensions, as well as Data.Type.Equality (since we will need to do some limited amount of type-proofs).

{-# LANGUAGE GADTs, ScopedTypeVariables,RankNTypes,
             TypeInType, TypeFamilies, TypeOperators #-}

import Data.Type.Equality

Defining Nat, Vec, and Queue

Next, we define the usual type-level natural numbers (this looks like just a regular data definition, but because we have TypeInType enabled, it will get automatically promoted when we use it in a type) and a type function (a type family) for addition. Note that although there are multiple ways of defining +, our choice here will impact what follows. We'll also define the usual Vec which is very much like a list except that it encodes its length in the phantom type n. With that, we can go ahead and define the type of our queue.

data Nat = Z | S Nat

type family n + m where
    Z   + m = m
    S n + m = S (n + m)

data Vec a n where
    Nil   :: Vec a Z
    (:::) :: a -> Vec a n -> Vec a (S n)

data Queue a where
    Queue :: { front :: Vec a (n + m)
             , rear :: Vec a m
             , schedule :: Vec x n } -> Queue a

Defining rotate

Now, things start to get hairier. We want to define a function rotate that has type rotate :: Vec a n -> Vec a (S n + m) -> Vec a m -> Vec a (S n + m), but you quickly run into a variety of proof related problems with just defining this recursively. The solution is instead to define a slightly more general grotate, which can be defined recursively, and for which rotate is a special case.

The point of Bump is to circumvent the fact that there is no such thing as type level composition in Haskell. There is no way of writing things an operator like (∘) such that (S ∘ S) x is S (S x). The workaround is to continuously wrap/unwrap with Bump/lower.

newtype Bump p n = Bump { lower :: p (S n) }

grotate :: forall p n m a.
           (forall n. a -> p n -> p (S n)) ->
           Vec a n ->
           Vec a (S n + m) ->
           p m ->
           p (S n + m)
grotate cons Nil        (y ::: _)  zs = cons y zs
grotate cons (x ::: xs) (y ::: ys) zs = lower (grotate consS xs ys (Bump (cons y zs))) 
  where
    consS :: forall n. a -> Bump p n -> Bump p (S n)
    consS = \a -> Bump . cons a . lower 

rotate :: Vec a n -> Vec a (S n + m) -> Vec a m -> Vec a (S n + m)
rotate = grotate (:::)

We need explicit foralls here to make it very clear which type variables are getting captured and which aren't, as well as to denote higher-rank types.

Singleton natural numbers SNat

Before we proceed to exec, we set up some machinery that will allow us to prove some type-level arithmetic claims (which we need to get exec to typecheck). We start by making an SNat type (which is a singleton type corresponding to Nat). SNat reflects its value in a phantom type variable.

data SNat n where
  SZero :: SNat Z
  SSucc :: SNat n -> SNat (S n)

We can then make a couple useful functions to do things with SNat.

sub1 :: SNat (S n) -> SNat n
sub1 (SSucc x) = x

size :: Vec a n -> SNat n
size Nil = SZero
size (_ ::: xs) = SSucc (size xs)

Finally, we are prepared to prove some arithmetic, namely that n + S m ~ S (n + m) and n + Z ~ n.

plusSucc :: (SNat n) -> (SNat m) -> (n + S m) :~: S (n + m)
plusSucc SZero _ = Refl
plusSucc (SSucc n) m = gcastWith (plusSucc n m) Refl

plusZero :: SNat n -> (n + Z) :~: n
plusZero SZero = Refl
plusZero (SSucc n) = gcastWith (plusZero n) Refl 

Defining exec

Now that we have rotate, we can define exec. This definition looks almost identical to the one in the question (with lists), except annotated with gcastWith <some-proof>.

exec :: Vec a (n + m) -> Vec a (S m) -> Vec a n -> Queue a
exec f r (_ ::: s) = gcastWith (plusSucc (size s) (sub1 (size r))) $ Queue f r s
exec f r Nil       = gcastWith (plusZero (sub1 (size r))) $
  let f' = rotate f r Nil in (Queue f' Nil f')

---

It is probably worth noting that we can get some stuff for free by using singletons. With the right extensions enabled, the following more readable code

import Data.Singletons.TH 

singletons [d|
    data Nat = Z | S Nat

    (+) :: Nat -> Nat -> Nat
    Z   + n = n
    S m + n = S (m + n)
  |]

defines, Nat, the type family :+ (equivalent to my +), and the singleton type SNat (with constructors SZ and SS equivalent to my SZero and SSucc) all in one.