[MLton] more FFI revisions

[Stephen Weeks]

> Looking at that, I think we ought to unify the treatment of the "*"
> forms.  Here is one proposal, which I think has the nice property
> that applying a "*" form to a pointer yields precisely the type of
> the corresponding "non-*" form:

OK.

> I would way way prefer if you just specified the full type you get,
> even though it involves some duplication.  After all, if that is
> painful then I can (or should) be able to use a type polymorphic
> type abbreviation for that.

Good point.  Certainly have the ":" be like a type annotation will be
less confusing.

Combining these ideas together, one gets the following.

  _address "symbol" : ptrTy;
  _import "symbol" [cdecl | stdcall] : cfTy;
  _import * [cdecl | stdcall] : ptrTy -> cfTy;
  _export "symbol" [cdecl | stdcall] : cfTy -> unit;
  _symbol "symbol" [alloc] : (unit -> cbTy) * (cbTy -> unit);
  _symbol * : ptrTy -> (unit -> cbTy) * (cbTy -> unit);

This does break backward compatibility with _export, but I could see
that being reasonable.  We could even detect, allow, and deprecate the
old usage for this release.  So we would also have

  _import "symbol" : cbTy;  ==>  cbTy  (* DEPRECATED *)
  _export "symbol" [cdecl | stdcall] : cfTy;  ==> cfTy -> unit  (* DEPRECATED *)

It doesn't seem like it would be hard for people to fix these in their
code.