This is the mail archive of the
mailing list for the GCC project.
Re: [patch] Fix computing address costs in ivopts
> The use of target_spill_cost in tree-ssa-loop-ivopts.c looks
> questionable to me. It appears that the algorithm adds the cost of a
> spill to many cost computations without any analysis of the need for a
> spill. That might be correct for a register-starved architecture, but
> exagerates the costs on architectures without that constraint.
the places where it happens are marked as FIXME, so (surprise :-) I am
aware about the problem. I have however did not meet so far an example where
this setting would cause any significant problems. I will check the apsi benchmark
you mention below, but I would be quite surprised if this would be the
cause of the regressions.
> In the generated assembly for 301.apsi, the current code with the
> inflated costs appears to be using more registers, causing more spills and
> more increments of the additional IVs.