I don't understand this test failure. EXCEPTION_ACCESS_VIOLATION should be defined as 0xC0000005 on Windows, which isn't negative.

EXCEPTION_ACCESS_VIOLATION is >2**31 so when it's cast to a long, which is signed and only 32-bits on Windows, it becomes negative:

Ah, and it seems that PyArg_ParseTuple allows negative values for whatever reason (but AC doesn't):

static PyObject *
foo(PyObject *self, PyObject *args)
{
    unsigned int a;
    if (!PyArg_ParseTuple(args, "I", &a)) {
        return NULL;
    }
    printf("a: %u\n", a);

    Py_RETURN_NONE;
}

>>> foo(-1)
a: 4294967295

Yuck.

I fixed the Windows EXCEPTION constants by using PyLong_FromUnsignedLong(). Constants and _raise_exception() are only defined for test_faulthandler, so it's ok to no longer accept negative values.

Ok, the CI does now pass. The PR is now ready for a review :-) Sorry, I should have marked it as a draft in the beginning.

This PR includes also some behavior changes (constants are positive, negative arguments rejected), so I think it may need a NEWS entry.

These changes only concern private functions and private constants, so I consider that they should not be documented. End users are not impacted, these functions and constants are only used by tests.