The problem varied with xserver commits, I did post all the details on #radeonhd but never actually filed a bug on the grounds that running git xserver under ~/ with old server still installed was probably not the sort of reporter they wanted. Also the suse people already said -ENOTIMEFORAGP when r600 demo failed some tests for me.
I guess if someone using PCIE and a stock distro reports it they'll take a look.
It was because the screen xy dimensions the vsync code gets were either 0 = novsync or big junk = vsync by luck.
Last time I used HD I also had worse issues than that - screen not updating properly, and dmesg getting errors when it happened but I don't know if that one has surfaced for anyone else yet. I haven't actually run it since kms arrived.