[Mesa-dev] static version of osmesa is missing llvm symbols, bug?

[Brian Paul]

On 03/11/2014 06:48 PM, Emil Velikov wrote:
> On 12/03/14 00:08, Burlen Loring wrote:
>> yep, I'm using 10.1.0 and also noticed the same in 9.2.2.
>>
> In that case "it wasn't me" ;-)
>
>> in short static linking is essential for reasonable performance of
>> parallel applications on Cray systems. In this scenario the application
>> is duplicated across 100s/1000s of compute nodes which share a
>> remote-parallel file system (Lustre/GPFS). As a dynamically linked
>> application starts, shared library loading hit's the file system hard.
>> However, Cray suggests users run static linked executables. Their
>> application launcher bypasses the file system and copies statically
>> linked executable directly into a ram fs onto the set of compute nodes
>> that a given job is scheduled to run on. This avoids file system
>> contention that would be introduced by a parallel application's shared
>> library loading.
>>
>> I maintain and support ParaView and interactive OpenGL based rendering
>> application on NERSC's Crays. Given that this is an interactive
>> application, having a reasonable startup time is important. Users have
>> complained and I've observed times when using shared libraries when app
>> startup takes more than 30 min! if we build fully static executable
>> startup time is reduced to seconds.
>>
>> For me statically linked mesa is important, which is why I want to
>> report this. Without OSMesa we can't use these systems, and the new
>> llvmpipe OSMesa state tracker has been really awesome, allowing us to do
>> things we couldn't in the past without GPUs. I hope that you don't
>> discontinue static link support.
>>
> Had some additional patches that cleans up mesa's build a bit further
> but I guess I'll bring back (classic and gallium) static osmesa first.
> Especially considering the numbers you've quoted.
>
>> We have experienced some segfaults in llvmpipe OSMesa in the static
>> linked version (bt shows SEGV at
>> create_jit_texture_type.isra.1 at draw_llvm.c:128). If I can reproduce this
>> off the Cray I will file a bug report. Thought I'd mention in case you
>> can confirm that there is a known issue with the static link.
>>
> AFAIK static linking with llvm is a bit of a hit and miss. For that
> reason mesa 10.2 may be building against shared llvm. Then again I do
> not work on that part of mesa, so take this with a pinch of salt.
>
> Brian Paul is the only person that I've seen working/bug fixing osmesa
> so I'm guessing that he'll take a look in due time. Meanwhile it may be
> worth filing bug report(s) for these two boogers :-)

Yes, please file bugs with as much info as possible.  I don't think I 
can help with the static build issue (my autocon/make knowledge is 
pretty thin).  I'd be happy to look into the LLVM segfault though (when 
I have time).

-Brian