https://gcc.gnu.org/bugzilla/show_bug.cgi?id=109210
--- Comment #2 from Jan Hubicka <hubicka at gcc dot gnu.org> --- Yep, I was thinking about this some time ago while reading the code. At the time we run this we already identified BBs with count 0 and edges with count 0. So I think when analyzing PHI we may want to see if any non-zero incomming edge exists and if so look if all non-zero edges has same prediction and if so combine it. If there is no non-zero edge in we can still combine same way the info over zero edges. There is also a paper on using VRP for branch prediction where value ranges are histograms anotated by probabilities.
