From mboxrd@z Thu Jan  1 00:00:00 1970
Received: (from weis@localhost) by pauillac.inria.fr (8.7.6/8.7.3) id VAA32575 for caml-red; Sun, 22 Oct 2000 21:41:54 +0200 (MET DST)
Received: from nez-perce.inria.fr (nez-perce.inria.fr [192.93.2.78]) by pauillac.inria.fr (8.7.6/8.7.3) with ESMTP id AAA23280 for <weis@pauillac.inria.fr>; Sun, 22 Oct 2000 00:36:51 +0200 (MET DST)
Received: from shell5.ba.best.com (shell5.ba.best.com [206.184.139.136])
	by nez-perce.inria.fr (8.11.1/8.10.0) with ESMTP id e9LManj21625;
	Sun, 22 Oct 2000 00:36:50 +0200 (MET DST)
Received: from localhost (bpr@localhost)
	by shell5.ba.best.com (8.9.3/8.9.2/best.sh) with ESMTP id PAA20590;
	Sat, 21 Oct 2000 15:36:43 -0700 (PDT)
Date: Sat, 21 Oct 2000 15:36:43 -0700 (PDT)
From: Brian Rogoff <bpr@best.com>
To: Chris Hecker <checker@d6.com>
cc: Pierre Weis <Pierre.Weis@inria.fr>, caml-list@inria.fr
Subject: Re: circular types?
In-Reply-To: <4.3.2.7.2.20001020142837.02ee31f0@shell16.ba.best.com>
Message-ID: <Pine.BSF.4.21.0010211513100.28299-100000@shell5.ba.best.com>
MIME-Version: 1.0
Content-Type: TEXT/PLAIN; charset=US-ASCII
Sender: weis@pauillac.inria.fr

On Fri, 20 Oct 2000, Chris Hecker wrote:
> >It's exactly similar to value definitions: put a ``and'' between the
> >two recursive definitions.
> >type foo = B of bar
> >and bar = F of foo
> Ah, thanks!  Is there any way to do it without the "and"?  In other
words, 
> what if I want to do this but the types are defined "far away" from each 
> other in the source.  The "and" solution works, but I'm looking for 
> something like forward declarations from C++:
> 
> class foo;
> class bar;
> class foo { bar *Bar; };
> // ...lots of stuff...
> class bar { foo *Foo; };

Nope, no forwards in Ocaml. At some point, you'll have to tie the recursive 
knot with an "and". If the problem you're having is that "far away" means 
"in different modules" then you bump into the lack of recursive modules, 
IMO the biggest PITA with ML style modules. This hurts most in a situation 
where you want something like the Composite pattern with modules  

type composite = { data : string ; children : CompositeSet.t }
and module CompositeSet = 
  Set.Make( struct 
              type t = composite 
              let compare c1 c2  = 
                Pervasives.compare (String.length c1.data) 
                                   (String.length c2.data) 
            end )

The first choice, which you don't like, is to put the mutually recursive
types into the same module. This would entail copying all of the source for
Set.Make into the module where you're defining the Composite, and using
"and" to tie the recursion together. I agree this is a poor workaround.

Another choice, which I saw on caml-list in a post from Benoit deBoursetty, 
is to make polymorphic versions of the Set functor by grab-and-hack on the
source of Set, and use the polymorphism to move the recursive dependence out 
of lift the module. So you make a 

module type PolyOrderedType =
  sig
    type 'a t
    val compare: 'a t -> 'a t -> int
  end;;

type 'a composite_fwd = { data : string ; forward : 'a };;

module OrderedComposites =
  struct
    type 'a t = 'a composite_fwd
    let compare cf1 cf2 = 
      Pervasives.compare (String.length cf1.data) (String.length cf2.data) 
  end;;

module CompositeSet = PolySet.Make(OrderedComposites);;

and given these polymorphic sets we can untie the mutual recursion with an 
"and" as follows 

type children = { children : composite CompositeSet.t }
and  composite = children composite_fwd;;

This is better, but involves an unpleasant extra level of indirection. You
can find the polymorphic sets at

http://www.stud.enst.fr/~debourse/caml.html

if this technique looks useful.

If the entities involved in the module spanning mutual recursion are
classes, and you *absolutely-don't-want-to-put-them-in-the-same-module*
you can create a dummy module in which the dependent classes are linked
and inherit from that. That moves the mutual recursion but doesn't break
it; there is no way I know of to break it short of doing a downcast or
narrowing-conversion on the classes, which is illegal in Ocaml. 

Here is a silly example; starting with these impossible signatures

module type DoctorType = 
  sig 
    class c : 
      object 
	method treat_patient : PatientType.c 
	method bill_patient : PatientType.c 
      end 
  end ;;

module type PatientType = 
  sig 
    class c : 
      object 
	method visit_doctor : DoctorType.c 
	method pay_doctor : DoctorType.c 
      end 
  end ;;

we add a parent module which expresses an interface dependency and write
the signatures in terms of that. 

module Forwards = 
  sig 
    class virtual patient_intf = 
      object 
	method virtual visit_doctor : doctor_intf  
	method virtual pay_doctor :   doctor_intf
      end 
    and virtual doctor_intf = 
      object 
	method virtual treat_patient : patient_intf
        method virtual bill_patient  : patient_intf
      end 
  end ;;

module type DoctorType = 
  sig 
    class c : 
      object 
	method treat_patient : Forwards.patient_intf
	method bill_patient : Forwards.patient_intf
      end 
  end ;;

module type PatientType = 
  sig 
    class c : 
      object 
	method visit_doctor : Forwards.doctor_intf
	method pay_doctor : Forwards.doctor_intf
      end 
  end ;;

-- Brian