On the other hand, _PyLong_Format currently contains general machinery 
for integer -> string conversion in any base in the range [2, 36], but I 
don't think that machinery is ever used for bases other than 2, 8, 10 
and 16.  So ripping _PyLong_Format out and just leaving the binary base 
and the suggested base 10 code might actually make longobject.c shorter.

I'd be interested to know how much the conversion of two digits at one 
time instead of one is contributing to the speedup by itself.  For large 
integers, I'd suspect that this makes very little difference.