What if numerator // denominator is an Integral, but not int

Aargh; good point. I'll fix, and add a test for that case.

I'll fix, and add a test for that case.

Hmm. The fix is easy; the test less so. The most obvious way to test is to actually create a non-int-subclass that implements numbers.Integral, but that takes quite a lot of code.

I think that we can use an int subclass for simple testing:

class I(int):
    def __floordiv__(self, other):
        return I(int(self) // other)

I think that we can use an int subclass for simple testing:

Yes, that's probably good enough. Updated.

What is the performance impact of this change on int()?

What is the performance impact of this change on int()?

No idea, but I wouldn't expect it to be significant in either direction. I can do some testing, but performance doesn't seem like it ought to be a major concern here.

No idea, but I wouldn't expect it to be significant in either direction

Hmm; I was wrong. I do see a significant difference in casual timings, apparently arising from the slowness of the 'int' call.

On this branch:

lovelace:cpython mdickinson$ ./python.exe -m timeit -s "from fractions import Fraction; f = Fraction(1001, 65)" "int(f)"
2000000 loops, best of 5: 193 nsec per loop

On the main branch:

lovelace:cpython mdickinson$ ./python.exe -m timeit -s "from fractions import Fraction; f = Fraction(1001, 65)" "int(f)"
2000000 loops, best of 5: 166 nsec per loop

Using operator.index is faster here: if I use

    def __int__(a, _index=operator.index):
        """int(a)"""
        if a._numerator < 0:
            return _index(-(-a._numerator // a._denominator))
        else:
            return _index(a._numerator // a._denominator)

I get something that's consistently faster than either of the above:

lovelace:cpython mdickinson$ ./python.exe -m timeit -s "from fractions import Fraction; f = Fraction(1001, 65)" "int(f)"
2000000 loops, best of 5: 156 nsec per loop

Just wondering, what is the performance if remove the index call?

Should this change be backported? If no, it should be documented as a new feature (versionchanged and What's New).

Definitely an enhancement so no backports.