Just looked it up. Yes, from Broadwell forward a reg to reg cmov has latency 1. On Atom processors it is still 6.

If you decrement your array index you don't even need the cmp instruction. The compiler could probably gen so good code.