So, the approach of duplicating finally blocks tends to lead to a non-trivial bytecode increase.  There is even a potential combinatorial explosion with nested "try..finally" block:

def f():
    try:
        ...
    finally:
        try:
            ...
        finally:
            # etc.

Such a chain of N nested "finally"s will emit O(2**N) opcodes.

I'm not sure it's a pratical concern but I find it is detrimental to the readibility of bytecode (which is a nice thing to have when doing low-level debugging or tweaking).  I think we can massage the PR to remove the "finally" block duplication while keeping the predictable stack effect.