What considerations go into predicting latency for operations on modern superscalar processors and how can I calculate them by hand?