Hi, On Wed, Sep 30 2026, Kugan Vivekanandarajah wrote: > Hi Martin, > >> On 29 Sep 2026, at 8:28 pm, Kugan Vivekanandarajah <[email protected]> >> wrote: >> >> Hi Martin, >>> >>> On Tue, Sep 29 2026, Kugan Vivekanandarajah wrote: >>>> Ping ? >>> >>> can you please have a look at >>> https://inbox.sourceware.org/gcc-patches/[email protected]/ >>> and to what extent this will avoid (or aggravate=) the issues you are >>> having? >>> ] >> >> Since count_sum is scaled by afdo_count_scale and hot threshold is also >> multiplied by afdo_count_scale, maybe_hot_count_p stays consistent for >> AutoFDO. >> I am running CPU2026 benchmarks to check the thresholds for AutoFDO. Sorry I >> didn’t to get to it sooner. >> > > I tried it on intrate of spec2026 with AutoFDO and it helps overall.
Thanks, so let's ambush Honza on Friday and have the patch approved :-) > > So, The patch divides the IPA-CP cloning score by the scaled total > function cost so that profile feedback no longer pushes almost every > hot callsite above the threshold of 1100 ? Basically yes, though the latter is a consequence of doing things (more) properly. Martin
