When reading the issue #27725, I saw that x_sub() of Objects/longobject.c is too careful just to change the sign of the newly created number: z cannot be shared at the end of the function, before z is normalized.

Attached patch simplifies the code.

See also the issue #27073, another similar simplification.

I would add a comment as to why the assertion is there, otherwise it seems somewhat random that it exists.

> I would add a comment as to why the assertion is there, otherwise it seems somewhat random that it exists.

Hum. Maybe it's even better to remove the assertion :-)

That works too. :)

Updated patch.

LGTM.

Maybe use size_a instead of Py_SIZE(z)?

And look at "sign". Currently it takes values 1 and -1. Is it worth to replace it with boolean variable "negative"? Or Py_ssize_t variable size_z that takes values size_a and -size_a?

New changeset be9dc240bf28 by Victor Stinner in branch 'default':
Issue #27786: Simplify x_sub()
https://hg.python.org/cpython/rev/be9dc240bf28

Thanks for the review Brett, I pushed my fix.

Serhiy Storchaka: "LGTM."

Oh, you posted your comment while I was pushing the patch after Brett wrote LGTM on the review (not on the bug tracker).

> Maybe use size_a instead of Py_SIZE(z)?
> And look at "sign". Currently it takes values 1 and -1. Is it worth to replace it with boolean variable "negative"? Or Py_ssize_t variable size_z that takes values size_a and -size_a?

Hum, I don't know what is the best. I don't think that it has an impact on performance. Feel free to modify directly the code, your proposed changes look safe and simple enough.