On Tue, Mar 09, 2021 at 05:11:09AM +0000, Tim Peters wrote:
> those for + and - are much subtler

In fact, these optimizations will payoff faster (wrt denominators
size), esp. due to gcd==1 branch.

Sorry for off-topic:

> WRT which, I added Python's Karatsuba implementation and regret doing so.
> I don't want to take it out now (it's already in ;-) ), but it added quite
> a pile of delicate code to what _was_ a much easier module to grasp.

(And was much more useless, even as a fallback.

But in the end - I agreed, you can't outperform professional bigint
implementations.  I think, you can _use_ them instead.)

> People who need fast multiplication are still far better off using gmpy2 anyway

(Another strange python "feature", IMHO.  Why the custom bigint
implementation, why not use the project, that run professionals in the
field?  Looking on the issue 21922 - it seems, that small ints
arithmetics can be almost as fast as for python ints.  Is the
memory handling - out-of-memory situation - the only severe problem?)