Re: [PATCH] Add Julia.

[Ricardo Wurmus]

Mark H Weaver writes:

> Mark H Weaver <address@hidden> writes:
>
>> Ricardo Wurmus <address@hidden> writes:
>>
>>> From 41550ffafa8da05b2cf8dfad66d44e6f43c9042c Mon Sep 17 00:00:00 2001
>>> From: Ricardo Wurmus <address@hidden>
>>> Date: Fri, 20 Mar 2015 11:26:23 +0100
>>> Subject: [PATCH 2/2] gnu: Add Julia.
>>>
>>> * gnu/packages/julia.scm: New file.
>>> * gnu-system.am (GNU_SYSTEM_MODULES): Add it.
>>
>> Hydra was unable to successfully build Julia on any platform.  See:
>>
>>   http://hydra.gnu.org/build/339221
>>   http://hydra.gnu.org/build/339192
>
> The illegal instruction errors are in openblas, which is marked
> non-substitutable because it apparently builds itself customized to the
> CPU on the build machine.  So, I guess openblas might have been built on
> a build slave with more processor features than the one used to build
> Julia.
>
> Perhaps the answer is that Julia shouldn't be built by Hydra, because
> one of its transitive inputs is non-substitutable.

I encountered build failures trying to install openblas on a virtual
machine (works fine on my workstation).  While looking for a solution I
stumbled upon the openblas make flag "DYNAMIC_ARCH=1" which asks
openblas to build modules for all(?) supported CPUs.

After modifying the openblas package in this manner it was built on the
virtual machine without issues and I could use it with Julia (so far
without problems).

May I suggest adding passing the "DYNAMIC_ARCH=1" flag in openblas
instead of propagating non-substitutability to Julia?  When built with
"DYNAMIC_ARCH=1" the CPU type can be picked with the environment
variable OPENBLAS_CORETYPE=<type>.  If I'm not mistaken this makes
substitutions for openblas possible.

What do you think?

~~ Ricardo