Bug#776911: gnome-shell: fails to start on i386 when [Mesa was] built with llvm-3.5

[Simon McVittie]

Adding mesa at packages to Cc since I suspect #775235 needs reassigning
to Mesa, perhaps along with its merged bugs #770130, #776911.

Context for Mesa maintainers: #775235 is that gnome-shell crashes in
an i386 VM with its default choice of emulated CPU, with
"LLVM ERROR: Do not know how to split the result of this operator!".
See https://bugs.debian.org/cgi-bin/bugreport.cgi?bug=775235#10 for full
backtrace.

The merged bug #770130 is that gnome-shell crashes on an unknown i386
(probably relatively old, since it has 82830M/MG integrated graphics,
which Wikipedia says is a Pentium III-M chipset) with
"Cogl-ERROR **: Failed to create texture 2d due to size/format constraints".
Oddly, this only happens for non-power-of-two textures in the gnome-shell
run by the actual user, not by the gnome-shell used for the gdm login
prompt (even though both use non-power-of-two textures). Michael Gilbert
merged this with #775235 and #776911 without comment; I'm not sure of
the reasoning for believing that #775235 and #770130 is the same thing.

The other merged bug #776911 is that gnome-shell crashes on an
unknown i386 with "Cogl-ERROR **: Failed to create texture 2d due to
size/format constraints"; that looks like the same thing as #770130.

On Sat, 21 Feb 2015 at 21:46:01 -0500, Michael Gilbert wrote:
> On Fri, Feb 6, 2015 at 1:20 PM, Simon McVittie wrote:
> > Michael, I see you've found a solution or workaround: is there
> > anything you'd like Rafal to try?
> 
> Try gnome-shell built with llvm-3.4 instead of 3.5.

gnome-shell doesn't build-depend on any llvm version so I don't think
that's going to make any difference. Do you mean mesa built with
llvm-3.4 instead of 3.5, as seen in
<https://bugs.debian.org/cgi-bin/bugreport.cgi?bug=775235#15>?

Looking at the Ubuntu changelog, it seems that they rebuilt mesa with
llvm-3.5, then reverted to llvm-3.4 because "too many regressions" (e.g.
https://bugs.launchpad.net/ubuntu/+source/llvm-toolchain-3.5/+bug/1360241),
but then switched to llvm-3.6 for Mesa 10.4 in 15.04/vivid.

Mesa 10.4.2 seems to have been accidentally uploaded to unstable
instead of experimental. I infer from the git repository that
10.3.2 (via t-p-u) is still the version targeted for jessie: is this correct?

Would use of llvm-3.4 for jessie be acceptable to the Mesa maintainers?
According to https://bugs.debian.org/cgi-bin/bugreport.cgi?bug=775235#15
that fixes at least one of these merged bugs, perhaps all three.

Alternatively, would it be useful information for the Mesa maintainers
if people tried Mesa 10.4.2 on affected systems?

#775235 should be easy to reproduce on an i386 VM on an amd64 host
(I assume that's how Steve did the original failed test), but #770130
and #776911 seem to require access to real hardware. I have a Pentium IV
system which I might be able to resurrect if that would help, although
if the relevant CPU feature is SSE2, I think that's too new.

    S