Bug 97038

Could you bisect LLVM to find out where the degraded performance comes from?

 It is major PITA to do that on this low power APU... but if someone can't reproduce it or figure it out without bisecting i might do that i have time.

 BTW not all but many apps runs in slowmo mode with llvm 3.9, not just ioquake3 based games.

 Should i start llvm bisect or not i dunno, as 3.9 enable GL 4.3 instead of 4.1 with llvm 3.8, i guess it might be mesa just set up something wrong too :) I might start banning some of those new extensions to found offender or so.

 OK got it, here is a llvm bisect result - svn272292

 AMDGPU/SI: Fix 32-bit fdiv lowering

Is it better on trunk since r276051?

Which version of Mesa are you using?

 Mesa was latest git at moment, but i also tried something around that time and same problem.

 Will try later against r276051 i am not on that machine currently...

 Yup, on r276051 perf is fine.

(In reply to smoki from comment #8)
>  Yup, on r276051 perf is fine.

Marking as fixed.

(In reply to Vedran Miletić from comment #9)
> (In reply to smoki from comment #8)
> >  Yup, on r276051 perf is fine.
> 
> Marking as fixed.

 If you say so, OK.

 I think i will reopen this one until it is fixed with llvm 3.9 or if someone commented that it would not be fixed in llvm 3.9.

 Current mesa git and llvm 3.9 now works fine... bug somehow getting fixed by mesa instead.