As the rcpps instruction is 4-operand parallel, your finding doesn't surprise me. So, it may be a documentation bug, which you could submit on premier.intel.com. However, the documentation of such functions is intentionally incomplete; too many changes occur between compiler versions for them to be recommended for application code. In view of the improved performance of IEEE divide on Penryn and later CPUs, I have been trying to avoid rccps.
