Skip to content

reversing the Unix and Bigarray dependency #6139

Closed

Description

@vicuna

Original bug ID: 6139
Reporter: @avsm
Assigned to: @gasche
Status: resolved (set by @gasche on 2018-04-09T14:37:32Z)
Resolution: fixed
Priority: normal
Severity: feature
Target version: later
Fixed in version: 4.07.0+dev/beta2/rc1/rc2
Category: otherlibs
Related to: #6885
Monitored by: @hhugo @gasche @diml @glondu @jmeber @hcarty @dbuenzli @Chris00

Bug description

With the splitting out otherlibs/ from the compiler distribution, I wanted to also see if we could remove the hard dependency on the Unix library from Bigarray. Currently, there is just one function (map_file) which causes Bigarray to depend on Unix.

Bigarray is increasingly used in the more exotic backends (js_of_ocaml, Mirage, ocamljava) as a convenient way to access externally allocated memory safely and fast. It would therefore make compilation much easier if the map_file were moved into the Unix module, thus introducing a Unix->Bigarray dependency rather than Bigarray->Unix. It's hard to override Bigarray via package management as it's installed in the standard library area and therefore picked up by default unless overridden for every case where it's used.

I'm happy to come up with a patch if there's general agreement that is a good idea. I can't see how not to break the existing interface, although telling people who to switch their existing use should be quite easy.

(two concrete reason I want this is to purge the last of POSIX from Mirage/Xen, which uses Bigarray, and also to more easily compile Core to Javascript for an interactive Real World OCaml toplevel).

Activity

  1. vicuna commented on Aug 26, 2013

    @vicuna
    Author

    Comment author: @alainfrisch

    I'd actually argue for making Bigarray part of the stdlib. AFAIK, it is supported on all platforms (at least for the core features, I'm unsure about map_file, precisely) and it has a special status w.r.t. the compiler (syntax and optimizations).

  2. vicuna commented on Aug 26, 2013

    @vicuna
    Author

    Comment author: @avsm

    Putting it in base, sans map_file, would be great. It's been very easy to port to all kinds of embedded platforms (as long as they have a malloc), and it has win32 support.

  3. vicuna commented on Aug 27, 2013

    @vicuna
    Author

    Comment author: @dbuenzli

    FWIW I'd also like to have them in the stdlib. I did sometimes avoid them at legitimate places because of their special status (namely sources/destinations for codecs, though in the current state it would still be hard to support an efficient interface that allows to e.g. source from either an in_channel, a string or a bigarray).

  4. vicuna commented on Aug 28, 2013

    @vicuna
    Author

    Comment author: @damiendoligez

    If you want to maintain compatibility at all costs, you can rename your new version of the Bigarray module to something like Bigarray_core, then make a new Bigarray that just reexports all the functions plus map_file.

  5. vicuna commented on Nov 4, 2013

    @vicuna
    Author

    Comment author: @johnwhitington

    Can I add my voice to the calls for reversing the dependency and putting Bigarray into the stdlib? It would be very useful.

    Not really a strong argument, but recall that Bigarray has special syntax even without loading bigarray.cma:

    $ ocaml
    OCaml version 4.01.0

    Bigarray.Array1.create;;

    Error: Reference to undefined global `Bigarray'

    let x = 0;;

    val x : int = 0

    x.{2};;

    Error: This expression has type int but an expression was expected of type
    ('a, 'b, 'c) Bigarray.Array1.t

    So in some sense, it's already in the stdlib.

  6. vicuna commented on Nov 6, 2013

    @vicuna
    Author

    Comment author: @alainfrisch

    If you want to maintain compatibility at all costs, you can rename your new
    version of the Bigarray module to something like Bigarray_core, then make a new
    Bigarray that just reexports all the functions plus map_file.

    What's the opinion of other proponents of the integration of Bigarray into stdlib? My preference would be to simply move the map_file functions into Unix (and not provide a compatibility module). It requires to fix client code, but the required change is trivial.

  7. vicuna commented on Nov 6, 2013

    @vicuna
    Author

    Comment author: @avsm

    Yes, I'd prefer the simpler approach of moving Bigarray into the stdlib, and the map_file function into Unix. I'm happy to help fix up the resulting problems with an OPAM bulk build over the 4.2 release cycle.

  8. vicuna commented on Nov 6, 2013

    @vicuna
    Author

    Comment author: @dbuenzli

    Also in favour of the last two comments. Let's put OCaml's type system and ocamlot to good use.

  9. vicuna commented on May 8, 2014

    @vicuna
    Author

    Comment author: @dbuenzli

    Has this a chance of occurring for 4.02 ?

  10. vicuna commented on Dec 18, 2015

    @vicuna
    Author

    Comment author: @johnwhitington

    Was consensus reached on this? It would be a useful improvement.

  11. vicuna commented on Dec 24, 2015

    @vicuna
    Author

    Comment author: @alainfrisch

    I don't think there has been any more discussion on this. Probably not for 4.03.

  12. vicuna commented on Jan 5, 2017

    @vicuna
    Author

    Comment author: @dbuenzli

    There is now a GPR #997

  13. vicuna commented on Mar 3, 2017

    @vicuna
    Author

    Comment author: @dbuenzli

    And another one here #1077

  14. vicuna commented on Apr 9, 2018

    @vicuna
    Author

    Comment author: @gasche

    Solved by merging #1685

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

Type

No type

Projects

No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions