[awhitworth]

I don't know exactly why, but I have some theories. First, the two instructions you list there aren't parallelizable, because you have a write to ESP and then a read from it, so the first instruction needs to complete first (or, get far enough in the pipeline for the partial result to be usable). This is almost certainly going to cause some stalls.

Beyond that, the best explanations I've seen are that the push/pop instructions are "optimized" in some way. I don't know if that means they are more optimized than just making the inc/read pair on the ESP atomic so it doesn't stall or if there is more to it.