Community
cancel
Showing results for 
Search instead for 
Did you mean: 
Highlighted
Beginner
35 Views

Opcode semantics

Hi,


Firstly, I apologize if this is the wrong forum; I could not find any other more relevant.

I'm looking for clarification in regards to a statement made that asserts there is a 1-cycle difference between the instructions:

0x3B (cmp reg, mem)
0x39 (cmp mem, reg)

As the two are functionally equivalent, I assume it would have to have something to do with the decoding circuit logic, but would like clarification if this statement reigns true in the first place.

Additionally, if this is true, where would I be able to find documentation of such details? All of the manuals I've read (even the IA32 optimization manual) does not mention these things.


Thanks,

Matt.
0 Kudos
3 Replies
Highlighted
Beginner
35 Views

Any ideas, anyone?
0 Kudos
Highlighted
New Contributor II
35 Views

Quoting - matt.j
Any ideas, anyone?

no idea, but here is the best place I know forthis kind of subtleties:

http://www.agner.org/optimize/

0 Kudos
Highlighted
New Contributor I
35 Views

Quoting - matt.j
Hi,


Firstly, I apologize if this is the wrong forum; I could not find any other more relevant.

I'm looking for clarification in regards to a statement made that asserts there is a 1-cycle difference between the instructions:

0x3B (cmp reg, mem)
0x39 (cmp mem, reg)

As the two are functionally equivalent, I assume it would have to have something to do with the decoding circuit logic, but would like clarification if this statement reigns true in the first place.

Additionally, if this is true, where would I be able to find documentation of such details? All of the manuals I've read (even the IA32 optimization manual) does not mention these things.


Thanks,

Matt.


Just a guess, but normally an instruction with a memory location as the 'destination' operand takes an extra cycle for the write operation. cmp doesn't actually write anything to the destination, but it might simplify the logic to handle these instructions uniformly and skip the write at a later point (where latencymight also beless of an issue). Compilers shouldn't emit this anyway, so they can make compromises like these.
0 Kudos