Atomic::add_and_fetch currently doesn't honor the relaxed memory order hint. I suggest that we change the code to do so. We initially found this issue when looking into why one of our parallel iterators were slower than expected. We have been using this for a while in the Generational ZGC repository.