See bug 1390527 for example. Right now checking weak pointers through do_QueryReferent() takes 2 virtual calls. It'd be nice if we made nsQueryReferent not be an nsCOMPtr_helper and eliminate the virtual nsQueryReferent::operator() call by making nsCOMPtr and friends know about nsQueryReferent in the same way that they know about nsCOMPtr_helper. That way they can call its operator() without having to make an indirect call.

nsQueryReferent is marked as `final` under the assumption that the compiler is able to devirtualize. But nsCOMPtr::assign_from_helper, which handles general assignment from nsCOMPtr_helper subclasses, isn't inlined (at least not with GCC and I can see other compilers not inlining it as well), so the compiler doesn't get a chance to see the opportunity for devirtualization. So `final` doesn't really help in this case, and we'd have to special-case assigning from nsQueryReferent in nsCOMPtr.

nsQueryReferent is defined as an nsCOMPtr_helper, which implies that calling its operator() method requires a virtual call. While nsQueryReferent is marked `final`, compiler inlining decisions make it impossible to de-virtualize the call to operator(). However, we have many other classes returned by do_* functions that nsCOMPtr handles directly, requiring no extra virtual calls, and we can give nsQueryReferent the same treatment. It's stupid that the ->QueryReferent method call inside nsQueryReferent::operator() is going to require a virtual call too, but maybe removing a virtual call from the call chain is virtuous in and of itself? We can address deCOMing nsIWeakReference (as a step to merging it with mozilla::WeakPtr or whatever) in a followup bug).

Comment on attachment 8900860 [details] [diff] [review] manually de-virtualize do_QueryReferent Review of attachment 8900860 [details] [diff] [review]: ----------------------------------------------------------------- ::: mfbt/RefPtr.h @@ +148,5 @@ > // construct from |Move(RefPtr<SomeSubclassOfT>)|. > { > } > > + MOZ_IMPLICIT RefPtr(const nsQueryReferent& aHelper); Out of curiosity, why are these not marked as inline? Do we want them to be? I am totally unsure how much compilers care about this of course. (Could be follow-up work of course.)

Pushed by nfroyd@mozilla.com: https://hg.mozilla.org/integration/mozilla-inbound/rev/20e6ab8c91bc manually de-virtualize do_QueryReferent; r=ehsan