https://gcc.gnu.org/bugzilla/show_bug.cgi?id=86957
--- Comment #3 from Indu Bhagat <ibhagatgnu at gmail dot com> --- Currently, GCC dumps some information in the dump files regarding functions not being executed in the training run (And, I agree its not satisfactorily direct). First, in the "Symbol table:" dump section in the .*i.profile dumpfile - a. functions which are not executed in the training run (hence, zero count profiles) are shown as "unlikely_executed". Thats the case when profiles are not estimated. b. If function profiles are estimated however (in -O1, -fguess-branch-probability is enabled), an "estimated_locally" appears alongside "unlikely_executed", e.g., ["Function flags: count: XXX (estimated locally, globally 0) body unlikely_executed"]. I see deeper classification in profile_count::dump (FILE *f) const API (which needs closer look), but I agree that we can improve the presented information regarding the quality of the profile. Perhaps abstracting out the somewhat superfluous information (in the current context), e.g., "globally 0" vs "globally 0 adjusted"... Why not additionally dump a clear "Profile Feedback" : "Run/Not run" per function in the dump file ? I need to take a closer look at the information being dumped currently before I make a proposal. But I agree that gcc can give clearer information. Will be happy to work on this as another CR.
