[RFA] The Integrated Register Allocator

[Mark Mitchell]

Vladimir Makarov wrote:

> The single requirement is to define macro IR_COVER_CLASSES in the 
> machine description file.   There are a lot of variants for the choice 
> and IRA is very sensitive to this decision (I tried long ago to make the 
> choice automatically but failed to do it even for the targets it 
> currently works for).  It means that the quality will be very strong 
> dependent from the choice.  The code performance for a wrong choice may 
> be even much worse than one for the old register allocator.

Is there any way to compute this from the given macros, in a vaguely 
sensible way?

> But you are right to make IRA just working for a port is easy.  I think 
> that IRA is pretty stable there will be no test regression for most ports.

That's good news.

-- 
Mark Mitchell
CodeSourcery
mark@codesourcery.com
(650) 331-3385 x713