So I ran some benchmarks: just doing an empty lock with: no icache pollution artificial icache pollution articicial icache and branch prediction pollution comparing the new locks with the old locks. No difference outside run-to-run variability in either case. -Andi