http://bugs.winehq.org/show_bug.cgi?id=24097
--- Comment #40 from Henri Verbeet hverbeet@gmail.com 2010-08-31 06:07:02 CDT --- (In reply to comment #39)
The question now is...WHY did it fix the problem?
It disables using the page attribute table for cache control of memory mappings, just using the (more coarse) MTRRs instead. For a more extensive description see e.g. http://www.intel.com/Assets/PDF/manual/253668.pdf, chapter 11, though I imagine you'll find it fairly technical. I guess the fact that nopat helps implies fglrx messing up cache control for some memory mapping, e.g. for a buffer object for a texture resource or color buffer. I'm obviously not familiar with fglrx internals though, so maybe I'm way off.