r600: image/buffer stores vs barrier
In !14319 (merged) images got flaky in some tests. I noticed that SFN at a barrier makes the previous stores emit an ack and then it waits for ack just before the barrier. This would presumably hang if the store was under control flow that the barrier wasn't, but more importantly to me the non-SFN path doesn't have this behavior. I thought this might cause the flakes I was seeing, but https://gitlab.freedesktop.org/anholt/mesa/-/commits/r600-rat-wait didn't fix it.