Re: [ira] Makes IRA default for some targets

[Vladimir Makarov <vmakarov at redhat dot com>]

Jeff Law wrote:
> Vladimir Makarov wrote:
> > The patch makes IRA default for targets for which IRA was ported.
> >
> > 2008-08-25  Vladimir Makarov  <vmakarov@redhat.com>
> >
> >    * opts.c (decode_options): Set up flag_ira.
> >
> >    * doc/invoke.texi (fira): Document it.
> >       * caller-save.c (save_call_clobbered_regs): Define and use
> >    do_placement_opt_p.
> >   
>
> +  bool do_placement_opt_p = 0 /* flag_ira && optimize */;
>
> Is this what you really wanted?  I don't see anywhere else where this
> variable is set...
There is a bootstrap failure on x86 (x86_64, ppc, itanium are ok) when 
the save/restore placement optimization is on.  I guess it does not do 
well with sharing save/restore slots.  So I switched it off temporarily 
on the branch.

I found that this optimization gives no visible improvement on SPEC on 
x86_64, ppc, itanium (only about 0.05% space saving on SPEC on x86_64).  
In general, this optimization permits to use saved hard-registers for 
reload registers.  It is not enough to get a visible improvement.  I 
wanted to implement using the hard-registers for pseudos spilled in RA.  
In this case, I expect a visible improvement.  But such thing is not 
easy to implement and it is definitely will be nor ready for gcc4.4.  I 
decided not to merge the optimization to the trunk.

I am still going to work on save/restore placement optimization and 
probably make it  available for gcc4.5.