https://gcc.gnu.org/bugzilla/show_bug.cgi?id=123126
--- Comment #9 from Jan Hubicka <hubicka at ucw dot cz> --- > And -fprofile-partial-training helps a lot (15%) and resolves the regression. > I wonder if that should be enabled by default? It might also explain why > probably_never_executed_edge_p makes a difference here. > > IMO this is a non-bug and -fprofile-use should imply > -fprofile-partial-training? -fprofile-partial-training makes binaries a lot bigger, since nothing gets optimized for size. This is counter-productive for big programs where the train set is reasonable, but indeed is not a problem for small/medium sized benchmarks. Do we know if the train run is simply not representative for reference run or it is a profile updating problem? I will run it with --train_with=refrate
