On Tue, Aug 14, 2012 at 10:45:18AM +0200, Uros Bizjak wrote:
> > The 46829 failure is due to combine pass blindly propagating r8 into
> > divmod instruction. This is invalid for divmod, which expects ax
> > there.
> > This can be fixed by introducing "ax_register_operand" predicate, but
> > I think that combine shouldn't propagate hard registers. This blocks
> > register allocator, and we have many passes that handle propagation of
> > hard registers after IRA much more effectively.
> >
> > Various "*_not_xmm0_*" predicates were introduced just to fight this
> > issue. They would be immediately obsolete with above combine change.
> >
> > IMO, preventing combine to propagate hard regs will fix 90% of spill
> > failures on x86.
>
> Probably, we can live with the check that propagated hard register
> really satisfies operand constraint.
>
> Jakub, do you have any opinion here?
The combiner doesn't propagate likely spilled hard registers (when they
appear in a pseudo <= hard reg or hard reg <= pseudo simple assignments),
see cant_combine_insn_p. It does propagate other hard registers, and
removing that could probably pessimize quite a lot of targets.
Jakub
