https://gcc.gnu.org/bugzilla/show_bug.cgi?id=69077
--- Comment #13 from Richard Biener <rguenth at gcc dot gnu.org> --- It's not referenced from the t2.cc cgraph at all, just from the WeightedStdDev constructor, taking its address via the vtable setting. So the possibly-inlined version is dropped on the floor and we keep the non-possibly-inlined variant. Which means this flag is quite bogus in the face of devirtualization (we'd basically need to set it everywhere?).