Measured through:
for ((n=0;n<10;n++)); do ./gum-tests -p /Core/Interceptor/cydia_substrate_replace_performance -m slow; done
for ((n=0;n<10;n++)); do ./gum-tests -p /Core/Interceptor/replace_performance -m slow; done
UPDATE: I published these results right after landing some massive architectural changes, and I got so excited that I published them immediately. Just took the time to profile the new code, and discovered an easy optimization to make it even faster. Check the gist's history to see the difference.