Created attachment 141980 [details]
dmesg from boot to the issue
This is a semi recent (within the last few months) bug on Polaris in NSMBU. Many polygons (maybe?) have major glitches across the screen, or don't show up at all. This doesn't occur on Tahiti or Pitcairn. I've gone back to 18.2.1 and verified it still exists, so it must be prior. It also exists on both linux 4.18.12 and the amdgpu-staging-drm-next branch. I'll continue attempting to bisect, unless someone else knows exactly what the issue is.
There are quite a few GPU faults listed in the attached dmesg, though I'm not sure if they're related to the issue. Glxinfo, Xorg.0.log, as well as a screenshot of the issue.
Created attachment 141981 [details]
Created attachment 141982 [details]
Created attachment 141983 [details]
screenshot of the issue
I forgot to mention, if IRC works better for communication for anyone, I'm TheRealJohnGalt on Freenode #radeon.
I was incorrect, this also happens on Tahiti now.
Seems more likely to be an LLVM or Mesa bug than a kernel one.
(In reply to Alex Deucher from comment #6)
> Seems more likely to be an LLVM or Mesa bug than a kernel one.
I went back to 4.18.7 and it persisted, so you're correct. It was okay early on at 18.2.0 with llvm svn. I'll go back to 18.2.0 to test, and if it's broken then we know it's llvm.
I attempted to bisect in Mesa, and found that I wasn't able to get rid of the glitch. The regression must be in LLVM.
R600_DEBUG=nir works around this issue. If requested, I can provide apidoc of this rendering issue.
Is this still an issue with the development version of LLVM 9.0 ? There have been many LLVM bug fixes since this issue was reported.