Here's a patch:

 - rename _PyLong_AsScaledDouble to _PyLong_Frexp (for the benefit of
   anyone foolish enough to be using an undocumented private API
   function)
 - make the exponent type Py_ssize_t instead of int
 - do the scaling by PyLong_SHIFT in _PyLong_Frexp instead of in the
   functions that call it (e.g., loghelper in the math module)
 - remove longintrepr.h include from math module, since it's no longer
   needed
 - change _PyLong_Frexp algorithm to do correct rounding
 - refactor PyLong_Double to use _PyLong_Frexp.