I don't think it's that important overall but it seems to be a decent percentage
of the time to enter the interpreter/bjit; this commit saves about 10% in a silly
ubenchmark.