Created attachment 115745 [details] Good llvm r237140 bisected where Xonotic start to render mostly green artifacts on textures. Attached R600_DEBUG=ps,vs,gs outputs from good on llvm r237139
Created attachment 115747 [details] Bad And bad with llvm r237140
mesa is 58715b72396133350c1549381553121f936a198e, kernel 4.1-rc3, hardware Athlon 5350 (Kabini), etc...
You said on IRC that the same commit also broke some piglit tests. If so, please attach the good and bad R600_DEBUG=ps,vs,gs output for one of those.
Created attachment 115763 [details] piglit-good Yes, example output is from: R600_DEBUG=ps,vs,gs ./bin/copy-pixels -samples=8 -auto on r237139
Created attachment 115764 [details] piglit-bad and bad with r237140
Uploaded piglit comparison within these llvm commits https://dl.dropboxusercontent.com/u/74553632/bug90439.tar.bz2 So basicaly the same ~130 piglits fail, just like in bug 89034 if i enable sub-reg liveness.
Could you get good and bad dumps from the glsl-fs-abs-03 test ?
Created attachment 115962 [details] pass Hi Tom, good and bad diff taken from glsl-fs-abs-03 test. 121,125c121,125 < v_interp_p1_f32 v4, v0, 2, 0, [m0] ; C8100200 < v_interp_p2_f32 v4, [v4], v1, 2, 0, [m0] ; C8110201 < v_cvt_pkrtz_f16_f32_e64 v0, |v2|, |v3| ; D25E0300 00020702 < v_cvt_pkrtz_f16_f32_e64 v1, |v4|, 1.0 ; D25E0101 0001E504 < exp 15, 0, 1, 1, 1, v0, v1, v0, v1 ; F8001C0F 01000100 --- > v_interp_p1_f32 v0, v0, 2, 0, [m0] ; C8000200 > v_interp_p2_f32 v0, [v0], v1, 2, 0, [m0] ; C8010201 > v_cvt_pkrtz_f16_f32_e64 v1, |v2|, |v3| ; D25E0301 00020702 > v_cvt_pkrtz_f16_f32_e64 v0, |v0|, 1.0 ; D25E0100 0001E500 > exp 15, 0, 1, 1, 1, v1, v0, v1, v0 ; F8001C0F 00010001 130c130 < VGPRS: 8 --- > VGPRS: 4
Created attachment 115963 [details] fail fail on svn237140 attachment
Created attachment 115965 [details] [review] Make sure v_interp_p1_f32 doesn't use the source register as destination Does this proof-of-concept LLVM patch fix the problem?
(In reply to Michel Dänzer from comment #10) > Created attachment 115965 [details] [review] [review] > Make sure v_interp_p1_f32 doesn't use the source register as destination > > Does this proof-of-concept LLVM patch fix the problem? Bingo! That seems fixed everyting. On top of that also i can enable SubRegLiveness without a problem.
Can you try this branch: http://cgit.freedesktop.org/~tstellar/llvm/log/?h=vinterp-fix
(In reply to Tom Stellard from comment #12) > Can you try this branch: > http://cgit.freedesktop.org/~tstellar/llvm/log/?h=vinterp-fix Works fine, apps render fine and no piglit regression. With that mainlined, this one and bug 89034 can be marked as fixed.
Patches mainlined. Closing.
Use of freedesktop.org services, including Bugzilla, is subject to our Code of Conduct. How we collect and use information is described in our Privacy Policy.