When micro-benchmarking on relatively short ASCII strings, the new implementation was about 30% faster than the old one.