Hi, On Mon, Aug 23 2021, Martin Jambor wrote: > When profile feedback is available, IPA-CP takes the count of the > hottest node and then evaluates all call contexts relative to it. > This means that typically almost no clones for specialized contexts > are ever created because the maximum is some special function, called > from everywhere (that is likely to get inlined anyway) and all the > examined edges look cold compared to it. > > This patch changes the selection. It simply sorts counts of all edges > eligible for cloning in a vector and then picks the count in 90th > percentile (the actual number is configurable via a parameter). > > I also tried more complex approaches which were summing the counts and > picking the edge which together with all hotter edges accounted for a > given portion of the total sum of all edge counts. But first it was > not apparently clear to me that they make more logical sense that the > simple method and practically I always also had to ignore a few > percent of the hottest edges with really extreme counts (looking at > bash and python). And when I had to do that anyway, it seemed simpler > to just "ignore" more and take the first non-ignored count as the > base. > > Nevertheless, if people think some more sophisticated method should be > used anyway, I am willing to be persuaded. But this patch is a clear > improvement over the current situation. > > gcc/ChangeLog: > > 2021-08-23 Martin Jambor <mjam...@suse.cz> > > * params.opt (param_ipa_cp_profile_count_base): New parameter. > * ipa-cp.c (max_count): Replace with base_count, replace all > occurrences too, unless otherwise stated. > (ipcp_cloning_candidate_p): identify mostly-directly called > functions based on their counts, not max_count. > (compare_edge_profile_counts): New function. > (ipcp_propagate_stage): Instead of setting max_count, find the > appropriate edge count in a sorted vector of counts of eligible > edges and make it the base_count.
Honza approved this patch in a private conversation but then I noticed I forgot to add an entry for the new parameter into invoke.texi, so I fixed that problem (and checked the result with make info and make pdf) and pushed the patch to master as commit ab1008255e37b5b51a433ed69e04c06300543799. Thanks, Martin
