I guess this is the only case where we can end up with a bogus ps_delta. It should be quite rare.
The assignment should go here, not at the top of the loop. We only break from the loop when the atomic_fcmpset at the end succeeds. In one iteration we might set ps_delta to a nonzero value, fail the update, and then exit the loop through this case, in which case ps_delta will be left equal to the old and now-incorrect value.