Actually, with this very code I got 15% variation between runs of identical code. I find this is often true when benchmarking nano-optimizations like this. That is why I consider 20% to be "indeterminate" (I add 5% to "be safe") for nano-optimization benchmarks.

Now, for macro-optimization benchmarks, I consider 5% to be "indeterminate". The problem with nano-optimization benchmarks is that the run time of the code being timed is so miniscule per iteration that slight variations in the "outside" code can make a relatively large impact on the timing results.

And, of course, the other problem with nano-optimization is that even if you find a 2-fold speed-up for one of these tiny, tiny operations, the actual change you end up seeing in how long it takes your script to run is usually a tiny fraction of that.

        - tye (but they are fun, aren't they?)