This test seems to fail randomly (as example 3 times in 10 runs). Will try to bisect when this started.
I can reproduce this on Haswell with Mesa 10.5, 10.6 and current Mesa master (afa1efd) but not on Skylake. This might be HSW specific.
Were you ever able to figure anything out on this?
(In reply to Ian Romanick from comment #2)
> Were you ever able to figure anything out on this?
This issue has disappeared and I cannot reproduce this anymore even with older commits (like commit afa1efd mentioned). I'll resolved this as WORKSFORME.