big slowdown in egcs-1.1.2->gcc-2.95 on alpha

[Joern Rennecke]

> If I read this correctly, it's spending a lot of time in reload
> (global_alloc isn't in the call graph, and reload is; in toplev.c, you
> see that one or the other is called, but the time is reported above as
> global_alloc either way).  Perhaps there's a problem in reload.

That's the price we pay for doing register spilling on a per-instruction
basis, and calling compute_use_by_pseudos twice for every instruction
and reload pass.

I think we could fix this by using pseudo register birth / death lists
instead of complete register sets.