Agner says that AAA has latency 5 on Cannon Lake, so using that instruction is a bit faster than doing the operations manually. But if you vectorize (or use SWA

Agner says that AAA has latency 5 on Cannon Lake, so using that instruction is a bit faster than doing the operations manually. But if you vectorize (or use SWAR) I imagine you can start to beat the legacy instructions with larger numbers.