http://www.agner.org/optimize 的 2. 里有这么一段: Instructions with multiple prefixes can slow down decoding. The maximum number of prefixes that an instruction can have without slowing down decoding is 1 on 32-bit Intel processors, 2 on Intel P4E processors, 3 on AMD processors, and unlimited on Core2. Avoid address size prefixes. Avoid operand size prefixes on instructions with an immediate operand. For example, it is preferred to replace MOV AX,2 by MOV EAX,2.