Feels like the useful benchmark would measure code size and i-cache effects together, since a bit of block cloning could beat a dispatcher until larger hot loops flip that tradeoff.