Yes, I know about that, but does it really break it? The status "warn" doesn't mean broken.

The commit also fixes other tests and it's just the correct thing to do.

Correction, meant e5f6b6d0feb9b9ad7132d5f236959ef088237347 as pointed out on irc.

Created attachment 99514 [details]
screenshot with commit reverted, appears to be identical to i915/llvmpipe/nouveau

Created attachment 99515 [details]
commit generating warnings

If you add "-auto", what's the piglit result?

Probe at (48,66)
  Expected: 0.423529 0.011765 0.313726 0.501961
  Observed: 0.501961 0.125490 0.125490 0.501961
PIGLIT: {'result': 'warn' }

"warn" is not a failure, like I said. I knew about this pass->warn change. It actually uncovers a potential issue with the hardware - or with the test. It depends on how you look at it.

From the test:

 * Note: some OpenGL implementations do not pass the "flat_first" and
 * "flat_last" tests when rendering quads or polygons.  That is, they
 * produce a tessellation which contains the correct vertices, but not
 * in the order required to preserve flat shaded colors.  This is
 * unlikely to cause problems for client programs, since client
 * programs that use new features like transform feedback are unlikely
 * to also use deprecated features like quads and polygons.  Also, it
 * is a matter of interpretation whether these tests are expected to
 * pass at all--after all, the spec does say that "the order of
 * tessellation within a primitive is undefined".  Accordingly, these
 * failures, should they occur, are flagged as warnings rather than
 * failures.

However, my Mesa commit which causes this test not to pass on nouveau, radeon, and probably other gallium drivers, and I did know about this, actually fixes real issues with flat shading and primitive restart and therefore cannot be reverted.