During XFree86 compilation at least on Alpha -ffast-math optimization flag is used. During recent timing runs with various application codes on Alpha and with current releases of compilers from gcc-2.96 family it was noted that results were consistently _slower_ when -ffast-math was used. Differences were never substantial but noticeable. It is possible that the same applies to XFree/Mesa. It should be stressed that I do not have any comparative performance numbers for this software on Alpha compiled with different optimization flags. This is only to note that maybe it would be worthwhile to look closer at the issue at some moment.
I will look into optimizing Alpha flags in the next run.
I'm currently reviewing all X related bugs in "DEFERRED" state. This one is rather old, and we no longer ship Alpha releases. I suppose it's kindof moot now though. ;)