[Pkg-julia-devel] Need help for packaging Julia 0.4

[Peter Colberg]

Hi Sébastien,

On Mon, Oct 19, 2015 at 02:45:21PM +0200, Sébastien Villemot wrote:
> I would not bother packaging utf8proc. It is very small and already
> embedded in a lot of packages (see http://codesearch.debian.net/results
> /utf8proc/).

The library is actually quite large with ~700K for the shared variant.
Would the goal not be to replace all the embedded copies of utf8proc
in other packages?

I think it would be good to package utf8proc not only for avoiding
the many embedded copies, but also to test and potentially improve
the upstream build files. By packaging JuliaDoc I could fix a few
oversights in setup.py that would otherwise not have been found.

> dsfmt is also embedded in several packages (like xmds2, shogun,
> openturns, as reported by http://codesearch.debian.net/results/dsfmt/).
> And my impression was that the patch for array-size arguments would
> break compatibility with the implementation embedded in other packages,
> thus defeating the purpose of having a separate package. So I did not
> bother packaging it separately.

The patch breaks ABI compatibility, but it preserves the API and could
be applied to a dsfmt package without affecting its usability by other
packages. I wrote autoconf/automake rules to build and test libraries,
which is straightforward with libtool.

Fedora has packaged dSFMT separately:

https://admin.fedoraproject.org/pkgdb/package/dSFMT/

One question is whether to package only one library for the most
frequently used exponent M=19937 as done in Fedora. I tend towards
testing and shipping all of the exponents (in one package). There
may be scientific use cases specifically for the larger exponents.

Another question is whether to enable SSE2 by default on i386. Given
that the library explicitly targets SIMD-enabled hardware, this seems
a sensible choice.

> For libuv, it would be indeed great to use the official Debian package.
> I don't know if this is now possible, but in the past I was told by the
> julia developers that it was not. See the discussion in this issue:
> https://github.com/JuliaLang/libuv/issues/2

Unfortunately a build against libuv1-dev fails due to missing
constants (UV_RAW_HANDLE, UV_RAW_FD). A search reveils that
these constants are only found in the Julia-patched libuv.

Regards,
Peter