I wonder if I should drop PR 4682.  I spent some more time working on it today.  I switched to the same scheme as Serhiy for the no-exception case, i.e. push a single NULL value, rather than six NULLs.  In ceval, we need to handle the non-exception case specially anyhow so I think it better to avoid the "stack churn".  With the micro-benchmarks that Serhiy posted, PR 4682 is as fast or slightly faster than PR 5006.  Doesn't really matter in real code though.

I did not push my latest changes as I did not fix the frame.f_lineno issue yet.  I think it is fixable without major changes.  Should I bother though?  Is there some other reason to prefer duplicating the final bodies rather than using a subroutine jump (as in 5006)?  A very minor speedup is not worth it I think because the compiler is a bit more complicated.  The settrace logic is more complicated too.