Thank you Daniel. But profiling is not good for benchmarking, it adds too much overhead. It can help to find narrow places, but for measuring effect of optimization you need to measure the raw time on non-debug build.

Semantically your changes are correct. I just want to know whether this adds a significant overhead. If it is, we should consider some kind of optimizations. It does not matter for Python implementation of BytesIO because in normal we use the C implementation.