I see some slowdowns with etqw (can't test heaven) on HD4890 that seem to be caused by this.

It's not constant - in fact mostly etqw with current head is the fastest I've ever seen it.

It may, however, during a short timenetdemo just suddenly drop to 4fps for a few seconds, or it may start slow ~10fps then go full speed.

With this reverted I haven't seen any slowdowns (yet).

Do you mean it's slower than without the commit or is it running mostly at the same speed when it's slowing down?

(In reply to comment #2)
> Do you mean it's slower than without the commit or is it running mostly at
> the same speed when it's slowing down?

It's way slower than without the commit, but only temporarily then it recovers.

In case you have etqw, I am testing @1920x1080 fullscreen with everything turned on/up as high as possible.

I've found another issue, but it's not related to this or llvm and could be quite old as I've not left etqw running for ~30 mins for some months.

The issue is I run out of mem (4 gig). I'll open a new bug when I've had more time to test with older kernels.

Also affects Xonotic.

Using kernel from linus's repo fixes this issue.
Not sure if it is a "proper" fix, so I'll let you close the repoer if it should be closed.

(now I'm getting 40 fps in heaven , nice boost if we forget about the 3 fps thing)

(In reply to comment #5)
> Using kernel from linus's repo fixes this issue.
> Not sure if it is a "proper" fix, so I'll let you close the repoer if it
> should be closed.
> 
> (now I'm getting 40 fps in heaven , nice boost if we forget about the 3 fps
> thing)

etqw is also running normally again with drm-next kernel and current mesa/llvm.

Using linux 3.8 fixes this so I'm closing.