Which is quickest?

[Strobe Raver]

Moving



Or

Push-Pop

[Slugsnack]

moving

[sinsi]

Raspberry-Lemonade
Artificially flavored


hah!

[Neil]

How about XCHG REG,REG

[sinsi]

Isn't xchg eax,eax the exact same microcode for nop?

[Slugsnack]

it is indeed. regardless i think he wants to move/copy register contents as opposed to swap their values

[sinsi]

depends on the registers, the cpu mode, the cpu, the phase of the moon...

[dedndave]

with newer architectures, push/pop is often the best choice
push/pop reg32 's are single-byte instructions
that means they consume less cache space - more code fits in the cache

[Astro]

push x - 5 bytes
pop y - 5 bytes

10 bytes total.


mov reg,x - 5 bytes (4 bytes data, 1 byte op-code with dest encoded into op-code)
mov y,reg - 5 bytes

10 bytes total.

I think that is correct.

According to the Intel manual (Vol. 2. PDF Page 442):

A1 - MOV EAX,moffs32* Move doubleword at (seg:offset) to EAX

So A1 means MOVe to EAX in a single byte.

In terms of actual speed - great question. Even if you used more clock cycles in the mov, would the speed difference between register access vs. a memory access be quicker overall?

Best regards,
Astro.

[redskull]

If you run it on an OoO processor, the mov's should be faster overall.  All other things being equal, a PUSH/POP pair generates 6 uops, whereas the MOV only generates 3 (you take the extra hit for doing the stack pointer math).  Plus, since I believe they use different ports, the two MOV commands should be able to go through at the same time, assuming both values are in the cache.  The PUSH/POP use the same ones, so it would have to do them sequentially.  I'm not too sure about that, so don't quote me, I have to go look that up when I get a moment.

-r

[Damos]

I did loads of speed test on pushes and pops and found them to be slow, I posted the results here at the time.