Hi, > > We can look into this. We do compare manually the IR dumps from both and it > is not ideal. > What we should do is an additional (optional) pass that runs after > auto-profile to compare the annotations > using the profile-use. We will have to filter out any functions/path that > runs less than a threshold to reduce noise. > Functions that are fully inlined are also not having any profile.
With -fprofile-use there is already logic to compare static profile prediction heuristics with the real data. This is implemented by running static profile predictor whithin profile pass if -fdump-ipa-profile-details is enabled and dump staistics which can be then summarized by contrib/analyze_brprob.py pass. We could use same strategy here. combine_predictions_for_bb can dump banch probabilities with AFDO quality as a fake predictor which will give us idea about coverage of AFDO data and its quality. We probably want to compare the AFDO counts to real counts and quantify how far are from precise data. I would start by remembering info about basic blocks being hot/cold by AFDO and dumping info when real data show that block is hot even if count is divided say by 1000 and AFDO still marked it cold. Grepping the dump file can then point us to most important profile problems. Eventually we can get something more fancy... With -fno-auto-profile-inlining it should be possible to produce two binaries. One with -g for auto-profile and one with -fprofile-generate and then load both profiles to the compiler with -fauto-profile=xxx -fprofile-use I think in this case read FDO should just overwrite auto-profile giving us place to dump stats. Honza > > Thanks, > Kugan > > > > > > Honza > >
