x86-64汇编乘3操作：IMUL替换为SHL+ADD能否提升运行效率？

Great question—this is exactly the kind of micro-optimization that matters when you're squeezing every cycle out of low-level code. Let's break down the tradeoffs clearly:

Modern x86-64 CPUs (Intel Skylake+/AMD Zen+ and newer) #

For these architectures, stick with IMUL rdx, 3—it's faster and more efficient. Here's why:

• Single-cycle latency: Modern CPUs optimize small-immediate integer multiplies heavily. The IMUL r64, imm form has a latency of just 1 cycle, meaning the result is ready for use in the next cycle.
• Higher throughput: You can execute up to 2 of these IMUL instructions per cycle (they can run on multiple execution ports simultaneously).
• Smaller code footprint: IMUL rdx, 3 is only 3 bytes long, compared to 7 bytes for the MOV+SHL+ADD sequence. This reduces instruction cache pressure, which can help with overall performance in larger programs.

Older CPUs (Intel Nehalem/AMD Bulldozer and earlier) #

On these older architectures, the shift-add sequence might be faster:

• Legacy IMUL instructions had higher latency (3-4 cycles for immediate multiplies), while the shift-add chain has a total dependency latency of 2 cycles:
  1. MOV rcx, rdx (1 cycle, runs in parallel with the next instruction)
  2. SHL rdx, 1 (1 cycle, produces 2*rdx)
  3. ADD rdx, rcx (1 cycle, combines 2*rdx with original rdx to get 3*rdx)
• Simple ALU operations like SHL and ADD also had higher throughput on older cores, making the three-instruction sequence more efficient than a single higher-latency multiply.

Key Takeaway #

If your target is modern mainstream CPUs (the vast majority of systems today), IMUL rdx, 3 is the better choice—it's simpler, shorter, and faster. Only use the shift-add sequence if you specifically need to support older hardware where multiply instructions were less optimized.

As a side note: Compilers like GCC and Clang already make this decision automatically based on the target architecture—check their output for reference if you're curious!

内容的提问来源于stack exchange，提问作者Cosmin Aprodu