I've gone back to kernel 3.6.11 and still get GPU locks when running this
program (and the min() CL builtin in piglit).

I'm leaning towards the issue being in mesa, which I'll have a much easier time
bisecting than kernel changes.

Tom, I've also verified that the LLVM .so version is correct (after deleting
the old one and rebuilding mesa).

I've tried Alex's patch as is, and also when swapping in the commented line in
evergreen_compute.c...  It still locks, but we'll see if I can bisect to find a
culprit.