BUG: fix leak in reductions when a ufunc override errors or is not found - #31731

PR summary

Closes #31677

Just another thing I noticed while reading reduction code and realised there's an open issue for it already.
One can see this memory leak by doing something like the following and memory profiling that (with memray for example) and will notice the memory is just growing.

import numpy as np

class Array:
    data = np.array([1,2,3])
    def __array_ufunc__(ufunc, method, *inputs, **kwargs):
        return NotImplemented

arr = Array()

if __name__ == "__main__":
    for _ in range(1_000_000):
        try:
            np.add.reduce(arr)
        except Exception as e:
            pass

I think goto fail is probably nicer than

Py_XDECREF(full_args.in);
Py_XDECREF(full_args.out);

and I believe it's safe. Only the else if (override) branch right bellow cannot use goto fail as it needs to return the override and not NULL.

I don't think there is a way to have a test regarding the ref count in the python side because full_args is created internally in the function in C. But we could have a test that is just the above without the loop and a sanitizer would probably capture the leak. Let me know if you want that. cc @seberg because you wanted a similar test in #31699.

First time committer introduction

N/A

AI Disclosure

No AI