I think this is a combination of https://bugs.python.org/issue44297 and the PREDICT macros.

I don't have a  windows machine to confirm this on, but I suspect that if you rewrite `doit` as:

def doit():
    o = ((1,2), (3,4))
    o = (a for
         a in
         o)
    for tup in o:
        x = tup[0]
        y = tup[1]

then you should be able to observe a difference between Windows and Mac on 3.9 as well.