Comment # 4 on bug 102955 from Elliot Thomas

I can confirm Marek's workaround works.

For what it's worth, I couldn't really measure a difference in performance
between the explicit DB flush/invalidation patch and using R600_DEBUG=nohyperz.