Re: [RFA] The Integrated Register Allocator

[Vladimir Makarov <vmakarov at redhat dot com>]

Andrew Pinski wrote:
> On Mon, Apr 7, 2008 at 5:50 PM, Andrew Pinski <pinskia@gmail.com> wrote:
>   
> >  But I am getting more register to register moves with IRA than with the old RA.
> >  For a simple testcase (which is internal, I will see if I can clean it
> >  up and let you have it), I am getting about 192 more register to
> >  register moves.
> >     
>
> After checking a newer trunk out to make a more fair comparison, we
> get slightly smaller code (6955 vs 7109) but still more register to
> register copies (352 vs 380) for this testcase (which I still don't
> know if I can give out).
>
> I should note sometimes we remove a register to register copy but that
> will not make the code smaller because we still need a nop for the
> branch hint.
>
> I will finish testing the testsuite with IRA turned on later today.
>
>   
Thanks, Andrew.  That would be interesting.  I'll add your 
IRA_COVER_CLASS to the branch.

Static count of spilled insns is not what IRA tries to minimize.  
Although often that is the result of IRA work.  -fira-algorithm=CB (that 
means one region allocation) gives smaller code in more cases than the 
default regional allocation.  The regional allocation can give smaller 
code because, even it generates ld/st/move insns on the region borders,  
less spills in the regions  can compensate  the size of the generated 
insns.  I'd also recommend to try -fcoalesce, it might  reduce the code too.