It already does (SSE, that is). For example, just recently Cristian Packmann successfully optimized bilinear filtering with MMX/SSE routines.
64bit will be considered after R1 is out. So, not before at least R2.
Officially, for R1 gcc2.95 is used, which keeps binary compatibility to BeOS R5. There are also solutions for a hybrid gcc2/4 which can also execute gcc4-compiled binaries. These will, however, probably break in the future and will have to be re-compiled.
Haiku uses software rendering for the time being (MESA 7.4.4). Artur Wyszynski is working on a Gallium3D port which will pave the road to hardware accelleration, if I interpret that correctly.
I hope I got all the above correct, I’m no developer and therefore not necessarily adequately informed on all accounts.