Mostly comes from detecting the case of runs of ASCII printable characters and using a special fast path for them. Improves benchmarks as per benchmark.py by 15-30% for text heavy workloads