Re: [Caml-list] Polymorphic Variants

[Jacques GARRIGUE <garrigue@math.nagoya-u.ac.jp>]

From: Tom <tom.primozic@gmail.com>
> So... why actually are polymorphic variants useful? Why can't they simply be
> implemented as normal, concrete (or how would you call them? ...) variants?

The original motivation was for the LablTk library, where some types
(index for instance) have lots of small variations. At that point
there where several options
* overloading (but ocaml loathes overloading, you could say that the
  total absence of overloading is essential to the language)
* refinement types: define a type with all constructors, and have
  restricted versions of it where only some constructors are allowed
* full polymorphism, i.e. polymorphic variants

If you eliminate the first option, then the choice is between
refinement types and polymorphic variants. Refinement types are rather
attractive: they combine precise typing with explicit declarations.
The arguments in favour of polymorphic variants are
* they were somehow easier to add, as they are an orthogonal extension
  of the type system
* one can add new cases afterwards.
* they are defined structurally: two identical definitions in two
  different modules are equivalent. This can be pretty handy at times.
* you don't need to open a module to use them: mixes nicely with ocaml
  programming style

>From modularity considerations, it ends up being nicer to write
     type t = [`A of int |`B of bool]
     type u = [t | `C of char]
than
     type u = A of int | B of bool | C of char
     type t = u[A B]

Afterwards I discovered that, using some idioms, they also allow
extensible programs in a very concise way. (Much more concise than
objects, when they are relevant.)

> Doesn't the use of polymorphic variants just mess up the function type?
What do you mean by "mess up the function type"?
If you mean that, without type annotations, types and errors become
very hard to read, this is true, but the manual explicitely encourages
to add type annotations :-)

> I'm not orthogonally against polymorphic variants, it's just that I am
> looking for an alternative concept that could be used instead... Maybe
> subtyped records?

In terms of expressiveness, you can simulate polymorphic variants
using objects (which are subtyped records,) but this is often much
more verbose, and requires higher order types. I have some code using
objects (visitor pattern), recursive modules, lazyness, and private
row types, in an utterly non trivial way, just to do what can be done
by standard recursive function definitions using polymorphic variants...

Jacques Garrigue