Created attachment 115622 [details]
Hack

Hardware state is not reset correctly when a process makes use of a new GL
context, after it has used and then destroyed the previous one.
This manifests in a (still uncommitted,
https://www.winehq.org/pipermail/wine-patches/2015-April/138974.html) Wine test
which fails on nv50. Actually commenting out stream_test(), which makes use of
instanced draws with ARB_instanced_arrays, workarounds the issue.
A single process executes both tests, destroying and then recreating a GL
context in between. AFAIU specifically the per-instance enable flags are never
reset on the HW and that breaks the latter test.

I'm attaching a patch which "fixes" the issue for me but it's not the correct
approach.