Right, this is why you should use AutoFDO nowadays not PGO. With AutoFDO you occasionally (e.g. in prod this might be something like record for 1s on average every 300s) record what branches were taken using perf-record from prod binaries, and then feed this back to the compiler.