4 ms·

These days, it’s not only compilers that are optimizing for the particulars of the CPU, but CPUs optimizing for compiled code. My favorite examples of this are

These days, it’s not only compilers that are optimizing for the particulars of the CPU, but CPUs optimizing for compiled code. My favorite examples of this are the inconsistent and sometimes downright crappy performance of “rep” string instructions and the “loop” instruction. Both seem like ideal easy things to write for handwritten assembly, yet the performance of both constructs can be quite awful on certain modern CPUs compared to the naive loops that compilers output. Much of this can be blamed on the fact that compilers rarely use either construct, so chipmakers had no reason to make either of them efficient (and, indeed, seem to have actually pessimized them on some chips!).

userbinator 7y ago
REP MOVS is still extremely fast --- and also small. It'll copy cacheline-sized chunks if the size is large enough. LOOP is a bit of a weird case. I've seen it benchmark both slower and faster than dec/jnz depending on the surrounding instructions.
- netrikare 7y ago
  loopXX instructions do not use CPU LSD (Loop Stream Detector) while cmp/jnz construct takes advantage of it. This speeds up some small loops. Also, there are some rules in intel manuals for instructions within cmp/jnz loop like no mismatched push/pop, etc. now, does anyone know why?
  - userbinator 7y ago
    like no mismatched push/pop, etc. My guess is virtual stack pointer update prediction latency. To expand on that, Intel's CPUs have had for a long time a separate piece of hardware dedicated to a "virtual" stack which speeds up push/pop instructions. If pushes and pops are not mismatched, then all stack operations can stay entirely within that and there's no need to update the "real" stack pointer nor stack entries upon leaving the loop.
    - netrikare 7y ago
      Thank you for your answer! Any idea why would loops not use LSD when programmed using loopXX instructions but would use when cmp/jnX is used?

userbinator 7y ago

REP MOVS is still extremely fast --- and also small. It'll copy cacheline-sized chunks if the size is large enough. LOOP is a bit of a weird case. I've seen it benchmark both slower and faster than dec/jnz depending on the surrounding instructions.

netrikare 7y ago
loopXX instructions do not use CPU LSD (Loop Stream Detector) while cmp/jnz construct takes advantage of it. This speeds up some small loops. Also, there are some rules in intel manuals for instructions within cmp/jnz loop like no mismatched push/pop, etc. now, does anyone know why?
- userbinator 7y ago
  like no mismatched push/pop, etc. My guess is virtual stack pointer update prediction latency. To expand on that, Intel's CPUs have had for a long time a separate piece of hardware dedicated to a "virtual" stack which speeds up push/pop instructions. If pushes and pops are not mismatched, then all stack operations can stay entirely within that and there's no need to update the "real" stack pointer nor stack entries upon leaving the loop.
  - netrikare 7y ago
    Thank you for your answer! Any idea why would loops not use LSD when programmed using loopXX instructions but would use when cmp/jnX is used?