I don't have much faith in microbenchmarks. Usually all they measure is how much effort the author put into overoptimizing code.