Re: [RFA] The Integrated Register Allocator

[Vladimir Makarov <vmakarov at redhat dot com>]

Paolo Bonzini wrote:
> > On gcc torture tests there is one additional regression for x86_64 and
> > one for ppc64 in comparison with the trunk.  These two regressions are
> > not critical: some particular code generation is expected.  The code
> > is different with IRA usage but it is still correct.  I'll see what
> > can I do with the tests.
>
> I think it's best if you make these tests use IRA, hoping that it 
> becomes the default more sooner than later.
>
> How are the compile times?
When  I checked last time, x86_64 bootstrap with IRA needed 0.4% more 
time than with the old register allocator (with 
--enable-checking=release).  SPECINT2000 compilation (practically for 
all targets) is in 3% slowdown frame (again with 
--enable-checking=release).    SPECFP2000 needs more time for IRA and 
the slowdown can be as worse as 8% for itanium.

When Chaitin-Briggs algorithm is used (instead of default and more 
complicated regional one) compiler slowdown is even less (about 1% for 
x86_64 on SPECINT2000).

I spent last half year mostly on speeding IRA up.  I've rewritten a big 
part of IRA for this and decreased compiler slowdown on SPECINT2000 from 
10% to 3%.

I think it is impossible to achieve better and faster RA using the 
reload than the old RA.  My observation is that better register 
allocation results in one more iteration of the reload because of more 
dense usage of hard-registers.