I'm a bit surprised that no tests fail, because we rely on patching to fire c_return when we return to a generated C call while the event is on. Not sure how we're firing the event otherwise. If you look at the codegen, we don't put code for firing the event usually.

EDIT: Ah, ISEQ_TRACE_EVENTS includes C_RETURN now since 48c8df9. It looks like now the logic for the else that comes after the diff is outdated now because it assumes RUBY_EVENT_C_RETURN is not part of ISEQ_TRACE_EVENTS?

I think we might want to change only the if condition and keep the invalidation at the end of the function. Theoretically, when we release the VM lock inside rb_yjit_tracing_invalidate_all(), we could yield to another sleeping ractor. With this patch the lock release comes before the global events update, so there is a window where we could compile new code with outdated event info.

Addressed feedback from both of you.

It looks like now the logic for the else that comes after the diff is outdated now because it assumes RUBY_EVENT_C_RETURN is not part of ISEQ_TRACE_EVENTS?

Agreed. I wondered why it existed, and looking at it again, I'm sure we can clean it up. But that involves function removals, so I'll make it a separate PR.