gh-91851: Trivial optimizations in Fraction

[skirpichev]

NB: this is the first commit from #25518

• Issue: Optimize the Fraction class arithmetics for small components #91851

[hauntsaninja]

Thanks! Microbenchmarks confirm that this is about 2x as fast.

cc @JelleZijlstra

[mdickinson]

LGTM.

That's a pretty clunky implementation of round-ties-to-even in the ndigits is None case. Would something like the following be faster?

d=self._denominatorrounded, remainder=divmod(self._numerator+ (d>>1), d) ifnotremainderandnot (d&1): # tie caserounded&=-2returnrounded

EDIT: A bit more streamlined, and avoiding unnecessary not operations:

d=self._denominatorrounded, remainder=divmod(self._numerator+ (d>>1), d) returnroundedifremainderord&1elserounded&-2

[mdickinson]

Would something like the following be faster?

Turns out not, at least on my machine.