Re: [Caml-list] Type That's Concrete From Within A Library Abstract From Without

[Jun Inoue <jun.lambda@gmail.com>]

Hi Jacques,

OCaml gives a type error if a public type in b.ml references a
non-trivial type in a.ml.  Is there a way around this?

$ cat > a.ml
type t = Foo of int
let x : t = Foo 3
$ cat > b.ml
type t = A.t
let x2 = A.x
$ cat > p.mli
module B : sig type t = A.t val x2 : t end
$ ocamlc -for-pack P -c a.ml b.ml
$ ocamlc -c p.mli
$ ocamlc -pack -o p.cmo a.cmo b.cmo
File "_none_", line 1:
Error: The implementation (obtained by packing)
       does not match the interface p.mli:
       In module B:
       Modules do not match:
         sig type t = A.t val x2 : A.t end
       is not included in
         sig type t = A.t val x2 : t end
       In module B:
       Type declarations do not match:
         type t = A.t
       is not included in
         type t = A.t


On Fri, Apr 27, 2018 at 3:05 PM, Jacques Garrigue
<garrigue@math.nagoya-u.ac.jp> wrote:
> You can provide a mli for the -pack.
> Just compile it before.
>
> $ cat > a.ml
> type t = int
> let x : int = 3
> $ cat > b.ml
> let x2 = A.x * A.x
> $ ocamlc -for-pack P a.ml b.ml
> $ cat > p.mli
> module A : sig type t val x : t end
> module B : sig val x2 : int end
> $ ocamlc -c p.mli
> $ ocamlc -pack -o p.cmo a.cmo b.cmo
>
> Now, if you use your library with only p.cmo and p.cmi available, you will
> only be able to access it through the interface you provided.
>
> Also, the method using module aliases can work too: you just have
> to use longer file names for the internal modules, to reduce the risk of
> conflicts. But this is more involved than using -pack with a mli.
>
> Jacques Garrigue
>
> On 2018/04/27 14:48, Jun Inoue wrote:
>>
>> Hi Ivan,
>>
>> That's basically our current approach, but it doesn't solve the
>> namespace pollution problem.  In your example, when someone installs a
>> file named b.cmi (whose interface is unrelated to your b.ml), the name
>> conflict prevents loading the std.cma file at all:
>>
>> $ ocaml
>>        OCaml version 4.04.0
>>
>> # #show B;;
>> module B : sig val foo : int end
>> # #load "std.cma";;
>> The files std.cma and b.cmi disagree over interface B
>>
>> So the technique makes B inaccessible but doesn't remove it from the
>> namespace.  This is why we want to -pack things, because our analogue
>> of b.ml is named matrix.ml, and there's no other sensible name for it.
>>
>> This technique doesn't work with -pack because that option demands all
>> .cmi's, including b.cmi.  I guess we could rename matrix.ml to
>> matrix_internal_dont_touch.ml, but we wanted to know if there's a
>> cleaner approach.  I wish we could supply a .mli file to the product
>> of -pack, but that also doesn't work...
>>
>> On Fri, Apr 27, 2018 at 12:06 AM, Ivan Gotovchits <ivg@ieee.org> wrote:
>>> Hi Jun,
>>>
>>> You can achieve this by implying an extra layer of indirection, i.e., by
>>> having two levels of interfaces. For example,
>>>
>>>   * A.ml - implementation of module A
>>>   * A.mli - private interface of module A
>>>   * B.ml  - implementation of module B that may rely on anything in A.mli
>>>   * Std.ml - a set of modules that you would like to import, e.g., `module
>>> A = A`, `module B = B`
>>>   * Std.mli - public interface specification
>>>
>>>
>>> Next, you deploy `std.cmxa` and `std.cmi` but keep `a.cmi` and `b.cmi` to
>>> yourself. This will prevent users from accessing your private modules A and
>>> B directly. (In oasis you can use PrivateModules stanza for this)
>>>
>>> Now you will have `Std.A` and `Std.B` that exposes as much as you want. Not
>>> sure whether it will work with the `-pack`, but you can use this approach
>>> instead of it. This is how we address the same issue in [BAP][1]
>>>
>>> Cheers,
>>> Ivan
>>>
>>> [1]: https://github.com/BinaryAnalysisPlatform/bap
>>>
>>>
>>> On Thu, Apr 26, 2018 at 10:18 AM, Jun Inoue <jun.lambda@gmail.com> wrote:
>>>>
>>>> Dear list,
>>>>
>>>> Is there a way to make a type concrete inside a library, yet opaque to
>>>> library users, preferably in a way that works with -pack?  This is a
>>>> nagging issue in our sundials package
>>>> (http://inria-parkas.github.io/sundialsml/).
>>>>
>>>> Basically, we have a type declared in one module of the library that
>>>> is pattern-matched upon in other modules, like:
>>>>
>>>> (* private.ml *)
>>>> type opaque_type = Foo | Bar
>>>>
>>>> (* public.ml *)
>>>> let f : opaque_type -> int = function
>>>>  | Foo -> 0
>>>>  | Bar -> 1
>>>>
>>>> There are a few constraints:
>>>> - We don't want users to be able to pattern-match on opaque_type.
>>>> - We need multiple modules in the library to pattern-match on
>>>> opaque-type (so moving opaque_typ e to public.ml is not an option).
>>>> - To avoid namespace pollution, we want to pack the whole library
>>>> (with ocamlc -pack) as a single Sundials module, so the user sees a
>>>> Sundials.Public module instead of just Public.
>>>>
>>>> Is this possible?  Right now, we just collect public.cmo and
>>>> private.cmo into sundials.cma and throw away private.cmi.  But this
>>>> doesn't work with packing:
>>>>
>>>> $ ocamlc -pack -o sundials.cmo private.cmo public.cmo
>>>>
>>>> demands that there be a private.cmi.
>>>>
>>>> --
>>>> Jun Inoue
>>>>
>>>> --
>>>> Caml-list mailing list.  Subscription management and archives:
>>>> https://sympa.inria.fr/sympa/arc/caml-list
>>>> Beginner's list: http://groups.yahoo.com/group/ocaml_beginners
>>>> Bug reports: http://caml.inria.fr/bin/caml-bugs
>
>
>



-- 
Jun Inoue

-- 
Caml-list mailing list.  Subscription management and archives:
https://sympa.inria.fr/sympa/arc/caml-list
Beginner's list: http://groups.yahoo.com/group/ocaml_beginners
Bug reports: http://caml.inria.fr/bin/caml-bugs