CODEWITHSUNDEEP
performanceoptimizationassemblyx86
R.. GitHub STOP HELPING ICE
R.. GitHub STOP HELPING ICE

Reputation: 215507

Reliable information about x86 string instruction performance?

Common widsom is that rep movsb is much slower than rep movsd (or on 64-bit, rep movsq) when performing identical operations. However, I've been testing on a few modern machines, and the run times are coming out identical (up to measurement noise) across a huge range of buffer sizes (10 bytes to 2 megs). So far I have just tested on 2 machines (32-bit Intel Atom D510 and 64-bit AMD FX 8120).

I'm asking this question from a standpoint of wanting to avoid cargo-culting a bunch of tests to break memory up into unaligned head/tail and aligned middle for the sake of using rep movsd or rep movsq if there's no actual benefit to doing this...

Upvotes: 14

Views: 1714

Answers (1)

user555045
user555045

Reputation: 64913

Lots of benchmarks here: instlatx64.atw.hu

For example (Intel Core 2 Duo E6700):

REP MOVSB   BW in L1D:13.04 B/c  34829MiB/s
REP MOVSW   BW in L1D:13.29 B/c  35493MiB/s
REP MOVSD   BW in L1D:13.40 B/c  35783MiB/s

Which shows that there is a difference, but it's tiny.

This one for SandyBridge is a little weird:

REP MOVSB   BW in L1D:25.50 B/c  86986MiB/s
REP MOVSW   BW in L1D:18.09 B/c  61721MiB/s
REP MOVSD   BW in L1D:27.47 B/c  93693MiB/s

Seems there is a big difference on some Atoms (seems to have disappeared with the D5xx, so you just missed it):

REP MOVSB   BW in L1D: 0.53 B/c    990MiB/s
REP MOVSW   BW in L1D: 1.93 B/c   3598MiB/s
REP MOVSD   BW in L1D: 3.74 B/c   6960MiB/s

I haven't found such big difference on anything else that can be considered new.

Upvotes: 16

Related Questions