The x86 Gentoo Refleaks 3.x report that test_sys_setprofile is potentially leaking referenecs:

https://buildbot.python.org/all/#/builders/1/builds/292

----------------------------------------------------------------------
Ran 21 tests in 0.007s
OK
.
test_sys_setprofile leaked [1, 1, 1] references, sum=3
1 test failed again:
    test_sys_setprofile
== Tests result: FAILURE then FAILURE ==
408 tests OK.
10 slowest tests:
- test_asyncio: 28 min 19 sec
- test_concurrent_futures: 23 min 10 sec
- test_multiprocessing_spawn: 22 min 1 sec
- test_multiprocessing_forkserver: 11 min 24 sec
- test_multiprocessing_fork: 8 min 35 sec
- test_lib2to3: 6 min 39 sec
- test_zipfile: 6 min 21 sec
- test_decimal: 6 min 18 sec
- test_io: 5 min 44 sec
- test_gdb: 5 min 39 sec
1 test failed:
    test_sys_setprofile
8 tests skipped:
    test_devpoll test_kqueue test_msilib test_startfile
    test_winconsoleio test_winreg test_winsound test_zipfile64
1 re-run test:
    test_sys_setprofile
Total duration: 2 hour 10 min
Tests result: FAILURE then FAILURE

Same error in the x86 Gentoo Refleaks 3.7 buildbot:

https://buildbot.python.org/all/#/builders/114/builds/174

I'm able to reproduce the issue on master:

vstinner@apu$ ./python -m test -R 3:3 test_sys_setprofile -m test.test_sys_setprofile.ProfileSimulatorTestCase.test_unbound_method_invalid_args
Run tests sequentially
0:00:00 load avg: 0.82 [1/1] test_sys_setprofile
beginning 6 repetitions
123456
......
test_sys_setprofile leaked [1, 1, 1] references, sum=3
test_sys_setprofile failed

== Tests result: FAILURE ==

1 test failed:
    test_sys_setprofile

Total duration: 113 ms
Tests result: FAILURE

The test has been modified recently:

commit 56868f940e0cc0b35d33c0070107ff3bed2d8766
Author: jdemeyer <jdemeyer@cage.ugent.be>
Date:   Sat Jul 21 10:30:59 2018 +0200

    bpo-34126: Fix crashes while profiling invalid calls. (GH-8300)

Just to avoid duplicate work: I can have a look at this.

Ok, so it's a real bug, a reference leak, which have been introduced by:

commit 93fac8dd358cd0e85e7b59115db226ce685d3f6f
Author: INADA Naoki <methane@users.noreply.github.com>
Date:   Tue Mar 7 14:24:37 2017 +0900

    bpo-29676: fix lsprof can't profile C method call. (GH523)
    
    When LOAD_METHOD is used for calling C mehtod, PyMethodDescrObject
    was passed to profilefunc from 5566bbb.
    But lsprof traces only PyCFunctionObject. Additionally, there can be
    some third party extension which assumes passed arg is
    PyCFunctionObject without calling PyCFunction_Check().
    
    So make PyCFunctionObject from PyMethodDescrObject when
    tstate->c_profilefunc is set.

New changeset 147d95511f59cfdd2d522f9d736f2335457bae20 by Victor Stinner (jdemeyer) in branch 'master':
bpo-34190: Fix reference leak in call_function() (GH-8413)
https://github.com/python/cpython/commit/147d95511f59cfdd2d522f9d736f2335457bae20

New changeset 25b87d35675301a929982db6d37836666af4f715 by Victor Stinner (Miss Islington (bot)) in branch '3.7':
bpo-34190: Fix reference leak in call_function() (GH-8413) (GH-8418)
https://github.com/python/cpython/commit/25b87d35675301a929982db6d37836666af4f715

Thanks Jeroen Demeyer for the good and quick fix! I applied it to 3.7 and master branches.

stage: patch review -> resolved