New register allocator branch created

[Daniel Berlin]

On Sun, 28 Jan 2001, Richard Kenner wrote:

>     Indeed, one has to react partly that the new allocator looks good, but
>     partly also that the old one seems surprisingly bad in this case.
>
> My experience is that we do fairly well on machines with lots of registers,
> but quie poorly on machines with very few.

However, this example is on a powerpc, which has 77 hard registers (32
gpr, 32 floating).

That should have been plenty. I would imagine either the allocation order
is wrong or something, causing local/global-alloc to use a register it
shouldn't (the link register in this case), or it just felt it *had* to
use that register.
--dan