This is because Maverick wants to discount the size of ret

No, it's not that! I've tried everything and got timings like 32cycles per instruction for general registers!!!
You can get real timings when you push all your registers *changed* in the code.
i.e. add ebx,0 took me 1 cycle; add ebx,1 and others that change ebx took 32 cycles.
I came to a final code like:

I use these macros to get a cycle count in edx:eax, simply sandwhich you code between the STARTTIMER, and STOPTIMER macros and printout the result with whatever you use to printout results. And don't forget to run it empty and subtract the result from the main result.
MadMatt

macro STARTTIMER
{
push ecx
push ebx
rdtsc
push edx
push eax
}

macro STOPTIMER ;edx:eax = cpu cycles ellapsed
{
rdtsc
pop ebx
pop ecx
clc
sbb eax,ebx
sbb edx,ecx
pop ebx
pop ecx
}

madmatt, nice work

Very nice, Thanks

If you have AMD processor, you could try free AMD CodeAnalist Performance Analyser tool - it allows CPU pipeline analysis and many other things. It requires debug information in PBD file to run, please take look at this thread for information how to build it.

Hey,

Just a few more things I didn't see here:

rdtsc is not a serializing instruction (it doesn't sync the instructions currently executing out of order) so you should prefix it with an instruction that does such as cpuid. You can also do a 'warmup' to make sure the cache works in your favour by executing a few cpuid/rdtsc first.

Also, time slicing will probably interfere with your code. You can write a quick driver that will cli first to prevent this, or keep your code small, execute it a few times and take the lowest result.

Performance mesurements with RDTSC