StaticSum
=========

While SML makes it impossible to write functions whose types would
depend on the values of their arguments, or so called dependently
typed functions, it is possible, and arguably commonplace, to write
functions whose types depend on the types of their arguments.  Indeed,
the types of parametrically polymorphic functions like `map` and
`foldl` can be said to depend on the types of their arguments.  What
is less commonplace, however, is to write functions whose behavior
would depend on the types of their arguments.  Nevertheless, there are
several techniques for writing such functions.
<:TypeIndexedValues:Type-indexed values> and <:Fold:fold> are two such
techniques.  This page presents another such technique dubbed static
sums.


== Ordinary Sums ==

Consider the sum type as defined below:
[source,sml]
----
structure Sum = struct
   datatype ('a, 'b) t = INL of 'a | INR of 'b
end
----

While a generic sum type such as defined above is very useful, it has
a number of limitations.  As an example, we could write the function
`out` to extract the value from a sum as follows:
[source,sml]
----
fun out (s : ('a, 'a) Sum.t) : 'a =
    case s
     of Sum.INL a => a
      | Sum.INR a => a
----

As can be seen from the type of `out`, it is limited in the sense that
it requires both variants of the sum to have the same type.  So, `out`
cannot be used to extract the value of a sum of two different types,
such as the type `(int, real) Sum.t`.  As another example of a
limitation, consider the following attempt at a `succ` function:
[source,sml]
----
fun succ (s : (int, real) Sum.t) : ??? =
    case s
     of Sum.INL i => i + 1
      | Sum.INR r => Real.nextAfter (r, Real.posInf)
----

The above definition of `succ` cannot be typed, because there is no
type for the codomain within SML.


== Static Sums ==

Interestingly, it is possible to define values `inL`, `inR`, and
`match` that satisfy the laws
----
match (inL x) (f, g) = f x
match (inR x) (f, g) = g x
----
and do not suffer from the same limitions.  The definitions are
actually quite trivial:
[source,sml]
----
structure StaticSum = struct
   fun inL x (f, _) = f x
   fun inR x (_, g) = g x
   fun match x = x
end
----

Now, given the `succ` function defined as
[source,sml]
----
fun succ s =
    StaticSum.match s
       (fn i => i + 1,
        fn r => Real.nextAfter (r, Real.posInf))
----
we get
[source,sml]
----
succ (StaticSum.inL 1) = 2
succ (StaticSum.inR Real.maxFinite) = Real.posInf
----

To better understand how this works, consider the following signature
for static sums:
[source,sml]
----
structure StaticSum :> sig
   type ('dL, 'cL, 'dR, 'cR, 'c) t
   val inL : 'dL -> ('dL, 'cL, 'dR, 'cR, 'cL) t
   val inR : 'dR -> ('dL, 'cL, 'dR, 'cR, 'cR) t
   val match : ('dL, 'cL, 'dR, 'cR, 'c) t -> ('dL -> 'cL) * ('dR -> 'cR) -> 'c
end = struct
   type ('dL, 'cL, 'dR, 'cR, 'c) t = ('dL -> 'cL) * ('dR -> 'cR) -> 'c
   open StaticSum
end
----

Above, `'d` stands for domain and `'c` for codomain.  The key
difference between an ordinary sum type, like `(int, real) Sum.t`, and
a static sum type, like `(int, real, real, int, real) StaticSum.t`, is
that the ordinary sum type says nothing about the type of the result
of deconstructing a sum while the static sum type specifies the type.

With the sealed static sum module, we get the type
[source,sml]
----
val succ : (int, int, real, real, 'a) StaticSum.t -> 'a
----
for the previously defined `succ` function.  The type specifies that
`succ` maps a left `int` to an `int` and a right `real` to a `real`.
For example, the type of `StaticSum.inL 1` is
`(int, 'cL, 'dR, 'cR, 'cL) StaticSum.t`.  Unifying this with the
argument type of `succ` gives the type `(int, int, real, real, int)
StaticSum.t -> int`.

The `out` function is quite useful on its own.  Here is how it can be
defined:
[source,sml]
----
structure StaticSum = struct
   open StaticSum
   val out : ('a, 'a, 'b, 'b, 'c) t -> 'c =
    fn s => match s (fn x => x, fn x => x)
end
----

Due to the value restriction, lack of first class polymorphism and
polymorphic recursion, the usefulness and convenience of static sums
is somewhat limited in SML.  So, don't throw away the ordinary sum
type just yet.  Static sums can nevertheless be quite useful.


=== Example: Send and Receive with Argument Type Dependent Result Types ===

In some situations it would seem useful to define functions whose
result type would depend on some of the arguments.  Traditionally such
functions have been thought to be impossible in SML and the solution
has been to define multiple functions.  For example, the
http://www.standardml.org/Basis/socket.html[`Socket` structure] of the
Basis library defines 16 `send` and 16 `recv` functions.  In contrast,
the Net structure
(<!ViewGitFile(mltonlib,master,com/sweeks/basic/unstable/net.sig)>) of the
Basic library designed by Stephen Weeks defines only a single `send`
and a single `receive` and the result types of the functions depend on
their arguments.  The implementation
(<!ViewGitFile(mltonlib,master,com/sweeks/basic/unstable/net.sml)>) uses
static sums (with a slighly different signature:
<!ViewGitFile(mltonlib,master,com/sweeks/basic/unstable/static-sum.sig)>).


=== Example: Picking Monad Results ===

Suppose that we need to write a parser that accepts a pair of integers
and returns their sum given a monadic parsing combinator library.  A
part of the signature of such library could look like this
[source,sml]
----
signature PARSING = sig
   include MONAD
   val int : int t
   val lparen : unit t
   val rparen : unit t
   val comma : unit t
   (* ... *)
end
----
where the `MONAD` signature could be defined as
[source,sml]
----
signature MONAD = sig
   type 'a t
   val return : 'a -> 'a t
   val >>= : 'a t * ('a -> 'b t) -> 'b t
end
infix >>=
----

The straightforward, but tedious, way to write the desired parser is:
[source,sml]
----
val p = lparen >>= (fn _ =>
        int    >>= (fn x =>
        comma  >>= (fn _ =>
        int    >>= (fn y =>
        rparen >>= (fn _ =>
        return (x + y))))))
----

In Haskell, the parser could be written using the `do` notation
considerably less verbosely as:
[source,haskell]
----
p = do { lparen ; x <- int ; comma ; y <- int ; rparen ; return $ x + y }
----

SML doesn't provide a `do` notation, so we need another solution.

Suppose we would have a "pick" notation for monads that would allows
us to write the parser as
[source,sml]
----
val p = `lparen ^ \int ^ `comma ^ \int ^ `rparen @ (fn x & y => x + y)
----
using four auxiliary combinators: +&grave;+, `\`, `^`, and `@`.

Roughly speaking

* +&grave;p+ means that the result of `p` is dropped,
* `\p` means that the result of `p` is taken,
* `p ^ q` means that results of `p` and `q` are taken as a product, and
* `p @ a` means that the results of `p` are passed to the function `a` and that result is returned.

The difficulty is in implementing the concatenation combinator `^`.
The type of the result of the concatenation depends on the types of
the arguments.

Using static sums and the <:ProductType:product type>, the pick
notation for monads can be implemented as follows:
[source,sml]
----
functor MkMonadPick (include MONAD) = let
   open StaticSum
in
   struct
      fun `a = inL (a >>= (fn _ => return ()))
      val \ = inR
      fun a @ f = out a >>= (return o f)
      fun a ^ b =
          (match b o match a)
             (fn a =>
                 (fn b => inL (a >>= (fn _ => b)),
                  fn b => inR (a >>= (fn _ => b))),
              fn a =>
                 (fn b => inR (a >>= (fn a => b >>= (fn _ => return a))),
                  fn b => inR (a >>= (fn a => b >>= (fn b => return (a & b))))))
   end
end
----

The above implementation is inefficient, however.  It uses many more
bind operations, `>>=`, than necessary.  That can be solved with an
additional level of abstraction:
[source,sml]
----
functor MkMonadPick (include MONAD) = let
   open StaticSum
in
   struct
      fun `a = inL (fn b => a >>= (fn _ => b ()))
      fun \a = inR (fn b => a >>= b)
      fun a @ f = out a (return o f)
      fun a ^ b =
          (match b o match a)
             (fn a => (fn b => inL (fn c => a (fn () => b c)),
                       fn b => inR (fn c => a (fn () => b c))),
              fn a => (fn b => inR (fn c => a (fn a => b (fn () => c a))),
                       fn b => inR (fn c => a (fn a => b (fn b => c (a & b))))))
   end
end
----

After instantiating and opening either of the above monad pick
implementations, the previously given definition of `p` can be
compiled and results in a parser whose result is of type `int`.  Here
is a functor to test the theory:
[source,sml]
----
functor Test (Arg : PARSING) = struct
   local
      structure Pick = MkMonadPick (Arg)
      open Pick Arg
   in
      val p : int t =
          `lparen ^ \int ^ `comma ^ \int ^ `rparen @ (fn x & y => x + y)
   end
end
----


== Also see ==

There are a number of related techniques.  Here are some of them.

* <:Fold:>
* <:TypeIndexedValues:>