Contrary to the comment in ceval.c, the f_lasti field
is not always correct because it is not updated by the
PREDICT / PREDICTED macros.  This means that when a
GET_ITER is followed by a FOR_ITER, f_lasti will be
left at the index of the GET_ITER the first time
FOR_ITER is executed.  I don't think this is a problem
for YIELD_VALUE because it's not predicted to follow
any other opcode.

I'm running into this when examining bytecode in
calling frames within a debugger callback.

I suggest either documenting that f_lasti may be
incorrect or adjusting it in the PREDICTED macro.