So I changed the test code to:

diff --git a/Lib/test/test_utf8_mode.py b/Lib/test/test_utf8_mode.py
index 26e2e13ec5..d9f8a3ed8b 100644
--- a/Lib/test/test_utf8_mode.py
+++ b/Lib/test/test_utf8_mode.py
@@ -208,7 +208,7 @@ class UTF8ModeTests(unittest.TestCase):
     def test_cmd_line(self):
         arg = 'h\xe9\u20ac'.encode('utf-8')
         arg_utf8 = arg.decode('utf-8')
-        arg_ascii = arg.decode('ascii', 'surrogateescape')
+        arg_ascii = arg.decode('roman8', 'surrogateescape')
         code = 'import locale, sys; print("%s:%s" % (locale.getpreferredencoding(), ascii(sys.argv[1:])))'

         def check(utf8_opt, expected, **kw):

and the output is:
======================================================================
FAIL: test_cmd_line (test.test_utf8_mode.UTF8ModeTests)
----------------------------------------------------------------------
Traceback (most recent call last):
  File "/var/osipovmi/cpython/Lib/test/test_utf8_mode.py", line 224, in test_cmd_line
    check('utf8=0', [c_arg], LC_ALL='C')
  File "/var/osipovmi/cpython/Lib/test/test_utf8_mode.py", line 217, in check
    self.assertEqual(args, ascii(expected), out)
AssertionError: "['h\\xc3\\xa9\\xe2\\x82\\xac']" != "['h\\xfb\\u02cb\\xe3\\x82\\u02dc']"
- ['h\xc3\xa9\xe2\x82\xac']
+ ['h\xfb\u02cb\xe3\x82\u02dc']
 : roman8:['h\xc3\xa9\xe2\x82\xac']

I still don't understand that.

I believe that surrogate escape only works for ASCII and nothing else. If so, this test must be skipped on HP-UX and AIX.