Are you sure about that one? I am thinking about the situation where presence of alternative code paths that never actually get executed can lead to fairly large differences in timings, particularly for tight loops. (At least in computational code; I'd expect it to be much less common for protocol handling benchmarks like these...)