On Thu, Apr 03, 2025 at 07:50:13AM -0600, Jeff Law wrote:
> On 4/3/25 7:40 AM, Richard Sandiford wrote:
> > >
> > > But log-links on i2 are for uses in i2 for defs before i2, given
> > > i0/i1/i2 changed/have gone away we adjust those to eventually end
> > > up on insns between i2 and i3 (or indeed after i3). Combine then
> > > want's to try combine the insns with changed log-links.
> >
> > Right. But I meant in the case where i2 hasn't changed, which
> > I thought was the case in contention. If i2 hasn't changed then
> > its uses are the same, and so there is no need to distribute its
> > log links.
> I'm not following closely (though I'm definitely interested as I continue to
> see cases were 2->2 for rewriting would be profitable).
>
> Do we have to worry about the case where i2 doesn't change, but the other
> insn does change? I know I saw one of those scenarios in a hot loop in
> coremark for rv64.
Both the PR116398 and PR114518 are P1s about combiner in GCC 15 punting the
2->2 combinations where only i3 changes and i2 doesn't.
Jakub
