Would be good to also do plain text benchmarks, not simple regexes. As I assume this might be the best optimized case