While on a fuzzer-found test case (added) that loop was faster, on real fonts, including NotoNastaliq in our benchmark, it was actually slower, which intuitively I would have expected. Still no idea why on that fuzzer case it's faster though. :(