bpo-38302. __rpow__ now called when __ipow__ returns NotImplemented

[ashkop]

Basically copied the behavior of binary_iop for the ternary operator. Also made sure that op_name is used in error message.

https://bugs.python.org/issue38302

[brandtbucher]

Thanks, this looks good! Just noticed a couple of small things:

[brandtbucher]

PyNumber_InPlacePower can now just mirror the other PyNumber_InPlace* functions, right?

So this is the only line that's needed here.

[ashkop]

Ah yes, thank you.

[brandtbucher]

It looks like we're also lacking a test for when A doesn't define __ipow__. Do you mind adding one?

[ashkop]

Sure, added the test

[ammaraskar]

Could you add a test for the two fallbacks and ensure the same class is returning NotImplemented and implementing the regular power methods i.e something like:

class A:
    def __ipow__(self, other):
        return NotImplemented

class B(A):
    def __pow__(self, other):
        return 1
class C(A):
    def __rpow__(self, other):
        return 2

The linked bug is only exhibited when a class has an ipow slot but when called it returns NotImplemented.

[ashkop]

Not sure I understood correctly what you meant, but I updated the test to also test the __pow__ fallback.

[ammaraskar]

Thank you, what you have now looks good. What I meant was that to catch the original bug/properly be a regression test, the class has to look like:

class A:
    def __ipow__(self, _):
        return NotImplemented
    def __pow__(self, other):
        ...

if the class looks like

class A:
    def __pow__(self, other):
        ...

then the old logic returns a false here and correctly goes to the non-inplace version of power anyway:

cpython/Objects/abstract.c

Lines 1162 to 1163 in 8e19c8b

	if (Py_TYPE(v)->tp_as_number &&
	Py_TYPE(v)->tp_as_number->nb_inplace_power != NULL) {

[ashkop]

Ah yes, you are right. Thanks!

[brandtbucher]

@ashkop, one more small thing in the new test. Then it looks good!

[brandtbucher]

This can just be:

Suggested change

	a = A()
	a = object()

Then you don't need the class definition above!

[ashkop]

Done.

[serhiy-storchaka]

LGTM. 👍

[ammaraskar]

Could you add a test for the two fallbacks and ensure the same class is returning NotImplemented and implementing the regular power methods i.e something like:

class A:
    def __ipow__(self, other):
        return NotImplemented

class B(A):
    def __pow__(self, other):
        return 1
class C(A):
    def __rpow__(self, other):
        return 2

The linked bug is only exhibited when a class has an ipow slot but when called it returns NotImplemented.

[ammaraskar]

Just fyi, @brettcannon posted on the original bug pointing out that this would not call __pow__ either. I'd recommend changing this to

If :func:`object.__ipow__` returns :const:`NotImplemented`, the operator will correctly
fall back to :func:`object.__pow__` and :func:`object.__rpow__` as expected.

It might also be worth mentioning this in whatsnew, even though it's obscure it is potentially a change in behavior that people should be informed about.

[ashkop]

Updated the news entry, and also added a line to What's New

[ammaraskar]

Thanks for fixing this! Really weird that op_name went unused so long. Would you mind adding a test for this like:

        x = None
        with self.assertRaises(TypeError) as cm:
            x **= 2
        self.assertIn('unsupported operand type(s) for **=', str(cm.exception))

[ashkop]

Thanks, added a new test

[ammaraskar]

Thank you, what you have now looks good. What I meant was that to catch the original bug/properly be a regression test, the class has to look like:

class A:
    def __ipow__(self, _):
        return NotImplemented
    def __pow__(self, other):
        ...

if the class looks like

class A:
    def __pow__(self, other):
        ...

then the old logic returns a false here and correctly goes to the non-inplace version of power anyway:

cpython/Objects/abstract.c

Lines 1162 to 1163 in 8e19c8b

	if (Py_TYPE(v)->tp_as_number &&
	Py_TYPE(v)->tp_as_number->nb_inplace_power != NULL) {

[ashkop]

Merged upstream, added Py_TEST usage, removed TERNARY_OP definition since op_name is now used in the error message.

[brettcannon]

I also tested this PR against my own test suite that uncovered this issue and it passed!

[bedevere-bot]

@brettcannon: Please replace # with GH- in the commit message next time. Thanks!

[brettcannon]

@ashkop thanks for the PR!