In the paper, the authors claim that the additional pointer arithmetic consists only of register operations & so one wouldn't expect any overhead. Then, it is surprising that the paper mentions 35% overhead on integer benchmark. wondering where that comes from...