https://gcc.gnu.org/bugzilla/show_bug.cgi?id=68771
--- Comment #27 from Iain Sandoe <iains at gcc dot gnu.org> --- (In reply to Daniel Vollmer from comment #26) > (In reply to Iain Sandoe from comment #25) > > (In reply to Daniel Vollmer from comment #24) > > > (In reply to Iain Sandoe from comment #23) > > Sadly, without a reproducer, it's hard to make progress - possibly more > > complex control flow is needed - multiple function? > > I've tried but not managed a simple reproducer yet. > The "real" code definitely has more complex control flow, but is also in a > shared-library that is in turn exposed / used by a Python extension. So when you profile this, you have to profile it underneath Python? i.e. you have no way to exercise the library's API to train it? rather a lot of moving parts to nail down if it's the former.