I finished up some cleanups of EXA today. It started with changes to fakexa, the fake EXA implementation in Xephyr, that lets me rapidly test EXA changes in a reliable environment. I made it so that anybody accessing pixmaps in framebuffer outside of the Prepare/FinishAccess wrappers ends up doing a null dereference. This was shockingly "successful" -- putimage, most GC ops, trapezoids, triangles, GC validation, core fonts, pretty much everything was doing invalid access. Before, you might get stale data in your software rendering, or stuff that wasn't un-tiled by the hardware. This could account for any of the "EXA renders wrong" bugs that we've all encountered. Especially for big-endian platforms.
A lot of this was exposed by xtest. I had failed to compile it myself (it hung!), but keithp handed me a compiled tarball that I've been using. Shockingly, fakexa is passing tests that Xephyr without fakexa is failing. That makes me pretty happy.
Now I'm feeling like EXA is pretty solid. I've been using xcompmgr -c on my desktop for over a day straight now and I'm using EXA on the laptop all the time, and it's probably time to start breaking things again. Hmm, what next?
![[info]](http://l-stat.livejournal.com/img/userinfo.gif){kind=small}
anholt's journal