I benchmarked this out of curiosity. Even using the bitshift and eliding the bounds checks (unsoundly), it's 22% slower than my optimized version.