SH target - disable @(r0,rm) addressing mode?
Wed Feb 5 21:12:00 GMT 2003
On Wed, 5 Feb 2003, Joern Rennecke wrote:
> email@example.com wrote:
> > I don't think this allows the scheduler to move the loads, though?
> Well, if the bulk of the arithmetic doesn't use r0, its relative
> position to the loads can be changed. If that is moving the loads
> or moving the arithmetic is a matter of opinion ;-)
> > The primary problem (as I see it) is the dependency of each load on the r0
> > register, which limits the scheduling freedom.
> If r0 was only used by loads, that would be a non-issue.
> > Maybe reenabling the first instruction scheduling pass is a cleaner
> > solution?
> Yes, certainly. We'd have to get it to actually work and also prevent it
> from increasing register pressure too much, though.
Most of the extra register pressure seems to come from related addresses.
We may be able to reduce this additional register pressure by coaxing the
new-ra rematerialization to handle related addresses (which it currently
doesn't). Perhaps this would solve most of the problem.
More information about the Gcc-bugs