The following patch should solve problem 
http://gcc.gnu.org/bugzilla/show_bug.cgi?id=48435.  It might solve other 
latest IRA problems too including performance related ones.  The patch 
is for targets which require some values to be placed in hard registers 
starting with an even (odd) hard registers.

I already addressed an analogous problem recently and the patch I sent 
that time although solved some problems it created even more new 
problems.  I should acknowledge this.

The problem was in that profitable hard regs were used for colorability 
criterion and finding hard registers where allocno values can be resided 
but *also* as starting allocno hard registers.  It resulted in spilling 
allocnos which should be placed in multi-registers starting on a 
specific border because profitable hard registers were only starting 
registers and when we calculated number of available hard registers 
multi-register allocnos can not fit only in their starting hard registers.

The following patch was successfully bootstrapped on x86/x86-64 and on 
i686 with H.J.'s autotester options.

OK to commit?

2011-04-07  Vladimir Makarov <vmakarov@redhat.com>

     PR 4435
     * ira-color.c (setup_profitable_hard_regs): Add comments.
     Don't take prohibited hard regs into account.
     (setup_conflict_profitable_regs): Rename to
     get_conflict_profitable_regs.
     (check_hard_reg_p): Check prohibited hard regs.