According to benchmarks, there is practically no difference in
performance. Only explicit unrolling is helpful, but that has been
implemented for the general algorithm in Reduction::reduce as well.