The same CI failure occurred in `test_tgsm()` when I dropped DXIL code in there, so the SM 6 shader was the issue. The CI runs on a 6600 XT and I have 6750 XT, so results should be the same, except the CI runs an old Mesa version. The failures were most likely caused by mishandling of barriers in the virtualised control flow, resulting in the UAV being written too early, but this is hard to prove.