On Fri, Feb 28, 2025 at 10:06:49AM +0100, Richard Biener wrote:
> > I've done 3 x86_64-linux and i686-linux bootstraps/regtests, each time
> > with the 3rd patch to gather statistics on number of successful ICF function
> > merges, and once with no further patches, once with the first patch and
> > once with the second patch instead of the first.
> >
> > The numbers of successful ICF function merges across the 2
> > bootstraps/regtests are
> > vanilla 175170
> > first 168617
> > second 168858
> > So, the second patch causes slightly more successful ICF merges over the
> > first one, but only tiny bit, for the first patch it is ~3.75% fewer
> > ICF merges, for the second patch ~3.6% fewer ICF merges.
> >
> > Guess another option would be to somehow try to be conservative about such
> > cases, but for ICF that sounds really hard, how do we figure out that we
> > need to adjust something in the chosen candidate and what exactly in it.
> > And for SCCVN how to arrange to modify the chosen winner so that it is
> > conservatively ok for both the merged cases.
>
> I prefer the first patch.
I know, the reason why I have tested the second patch at all was that at
first the 3.75% fewer merges looked like a big number to me and I wanted to
see if the second patch helps with that. It clearly does but almost in the
noise area.
Honza, your thoughts on this?
Jakub
