> On Thu, Nov 15, 2018 at 7:14 AM Jan Hubicka <hubi...@ucw.cz> wrote: > > > > > > A warning seems appropriate. You think the front end is the right > > > > place for that? > > > > > > Probably yes. Note that middle-end can optimize about dead branches and > > > so that > > > theoretically one can end up with a branching where e.g. both branches > > > are [[likely]]. > > > I wouldn't bother users with these. > > > > Note that what really happens in this case is that if conditional is > > constant propagated and it has predict_expr, the predict_expr stays and > > will get assigned to the random control dependence edge which controls > > execution of the original statement. This is not very intuitive > > behaviour. Does C++ say what should happen in this case? > > No, it doesn't say much. > > > One option would be to deal with this gratefully at high level gimple > > and turn predict_exprs into edge probabilities eariler than we do normal > > branch prediction (which is intended to be later so profile does not end > > up unnecesarily inconsistent) > > Sounds reasonable. > > This additional patch implements the suggested diagnostics.
Looks good to me. Next stage1 I will experiment with moving path prediction up in the pass queue. It should be quite easy to do. Honza
