`test.test_math.MathTests.testAtan2` failure on OpenIndiana/Solaris

[mtelka]

I do see the following failure with Python-3.14.0rc2 on OpenIndiana. This is a regression from Python-3.13.7.

======================================================================
ERROR: testAtan2 (test.test_math.MathTests.testAtan2)
----------------------------------------------------------------------
Traceback (most recent call last):
  File "/data/builds/ul-workspace/components/python/python-314/Python-3.14.0rc2/Lib/test/test_math.py", line 339, in testAtan2
    self.ftest('atan2(0., -0.)', math.atan2(0., -0.), math.pi)
                                 ~~~~~~~~~~^^^^^^^^^
ValueError: math domain error

----------------------------------------------------------------------

Linked PRs

• gh-138573: restrict testAtan2() the IEC 60559-compatible platforms #138812
• gh-138573: filter out failing math tests on Solaris #146402
• [3.14] gh-138573: Filter out failing math tests on Solaris (GH-146402) #146438

[StanFromIreland]

Also occurs on our Solaris buildbot.

I actually talked about this with @kulikjak, and this is an issue on their side. So there isn't anything for us to do.

[terryjreedy]

Does that mean we should close this?

[skirpichev]

math.atan2 is essentially a thin wrapper over libm's atan2:

cpython/Modules/mathmodule.c

Lines 1100 to 1103 in 8554c09

	FUNC2(atan2, atan2,
	"atan2($module, y, x, /)\n--\n\n"
	"Return the arc tangent (measured in radians) of y/x.\n\n"
	"Unlike atan(y/x), the signs of both x and y are considered.")

cpython/Modules/mathmodule.c

Lines 1014 to 1047 in 8554c09

	static PyObject *
	math_2(PyObject *const *args, Py_ssize_t nargs,
	double (*func) (double, double), const char *funcname)
	{
	double x, y, r;
	if (!_PyArg_CheckPositional(funcname, nargs, 2, 2))
	return NULL;
	x = PyFloat_AsDouble(args[0]);
	if (x == -1.0 && PyErr_Occurred()) {
	return NULL;
	}
	y = PyFloat_AsDouble(args[1]);
	if (y == -1.0 && PyErr_Occurred()) {
	return NULL;
	}
	errno = 0;
	r = (*func)(x, y);
	if (isnan(r)) {
	if (!isnan(x) && !isnan(y))
	errno = EDOM;
	else
	errno = 0;
	}
	else if (isinf(r)) {
	if (isfinite(x) && isfinite(y))
	errno = ERANGE;
	else
	errno = 0;
	}
	if (errno && is_error(r, 1))
	return NULL;
	else
	return PyFloat_FromDouble(r);
	}

I don't think something was changed here since v3.13 other than using libm's isinf/isnan macro directly.

It looks like in the given case platform's atan2 returns a nan. @mtelka, could you verify this?

Correct answer is pi, see https://en.cppreference.com/w/c/numeric/math/atan2.html (Assuming IEEE 754-compatible behavior.) @mtelka, did you test on same version of the OS? This looks as a regression for me, but not in the CPython.

Edit: I assume, that we won't add a workaround for unsupported broken platform. Or?

[StanFromIreland]

math.atan2 is essentially a thin wrapper over libm's atan2:

Yes, as I said, this is an issue in their library, which they hope to fix soon.

[skirpichev]

Ok, let's close. I don't think workaround is needed.

[mtelka]

I tested this simple C:

#include <math.h>
#include <stdio.h>

int main(void)
{
        printf("%f\n", atan2(0., -0.));
        return 0;
}

and here is the result:

$ gcc -Wall -o test test.c 
$ ./test 
3.141593
$

How can I prove the issue is with the platform?