Raymond, I am super grateful that you are showing us the assembly language. That's the kind of thing we need to do more of if we want to be successful. (In this case I'm not surprised by the outcome, but I have never studied x86 assembly and this was a nice explanation.)

Regarding speeding up marshal, I am not at all convinced that this is still I/O bound. Raymond, do you have data about this? In my own informal experiments I believe the main cost I found was (a) stat() operations, and (b) the allocation and initialization of a large number of objects. In my (limited) understanding, "freezing" modules (like in bpo-45020) is mostly a win because it avoids stat() operations (and listdir(), too). But it would be nice to do a detailed experiment to get real facts here.