You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
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).
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).
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.
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).
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.
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.
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.
Activity
vicuna commentedon Aug 26, 2013
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).
vicuna commentedon Aug 26, 2013
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.
vicuna commentedon Aug 27, 2013
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).
vicuna commentedon Aug 28, 2013
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.
vicuna commentedon Nov 4, 2013
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.
vicuna commentedon Nov 6, 2013
Comment author: @alainfrisch
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.
vicuna commentedon Nov 6, 2013
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.
vicuna commentedon Nov 6, 2013
Comment author: @dbuenzli
Also in favour of the last two comments. Let's put OCaml's type system and ocamlot to good use.
vicuna commentedon May 8, 2014
Comment author: @dbuenzli
Has this a chance of occurring for 4.02 ?
vicuna commentedon Dec 18, 2015
Comment author: @johnwhitington
Was consensus reached on this? It would be a useful improvement.
vicuna commentedon Dec 24, 2015
Comment author: @alainfrisch
I don't think there has been any more discussion on this. Probably not for 4.03.
vicuna commentedon Jan 5, 2017
Comment author: @dbuenzli
There is now a GPR #997
vicuna commentedon Mar 3, 2017
Comment author: @dbuenzli
And another one here #1077
vicuna commentedon Apr 9, 2018
Comment author: @gasche
Solved by merging #1685