Subtooting somewhat: the problem with relying too much on benchmark-driven optimization of individual functions, in any moderately complex system, is it usually gives you the wrong answer. Worse, it systemically drives poor performance of the whole. That's because for a small program, it drives you to increase code and data size (e.g loop unroll), to increase inner loop performance. In a system, however, almost always the best thing you can do is *smaller size* because cache and I/O dominate.