As it was found out while working on bug 149749, I had quite some trouble running the layout tests on a recently installed development machine with "standard" Debian Testing (no packages from other sources), where I would get ~230 tests crashing due to some kind of problem in GStreamer, as you can check here:
Results here: https://drive.google.com/file/d/0B6Gdj3EoWfFLRFNSejRIMlViY28/view?usp=sharing
After some investigation today, I finally found the problem (and why it was so hard to find in the first place):
The problem is that a newer versions of cpp (I have 5.2.1, while the bot has 4.9.2) produces a different output while processing the headers from libfaad2, causing gst-plugins-bad to be compiled without the FAAD2_MINOR_VERSION definition in config.h, which causes crashes and stack corruption on runtime.
This problem has been reported already for Fedora 22 in the RH's bugzilla:
Fortunately, this has been already fixed upstream in gst-plugins-bad:
I'm 100% sure that's the same issue I was seeing because I could double check myself that the output for f21 and f22 as mentioned in https://bugzilla.redhat.com/show_bug.cgi?id=1219320#c21 is exactly the same one that I'm getting in the bots (with help from aperezdc), with cpp 4.9.2, and my local Debian Testing machine, with cpp 5.2.1.
Besides, I've applied the patch locally myself to the gst-plugins-bad from the internal jhbuild, recompiled it together with the WebKit libraries and binaries... and it worked! No more weird crashes around running the tests!
So, I'll submit a new patch to backport that patch from upstream into gst-plugins-bad 1.4.4 from the jhbuild
Created attachment 262883 [details]
See the proposed patch attached, please review. Thanks!
Committed r190854: <http://trac.webkit.org/changeset/190854>