Add PyList_GetItemRef and use it in list.get_item

[ngoldbaum]

Refs #4265 #4355

See the C API docs.

Both the new shim and the bindings to the CPython C API on 3.13 should be covered by existing tests in PyO3.

[davidhewitt]

Thanks! Just a few small observations to round this one off...

[davidhewitt]

Looks great, thanks very much! 👍

[davidhewitt]

Hmm the debug builds failing on 3.12. If I had to guess, it's an issue in the definition of Py_XINCREF just due to the way we set up the debug interpreter. I'll try to investigate tomorrow.

[ngoldbaum]

The debug build is hitting an assert inside CPython, running test_gevent.py.

Separately, it looks like there have also been some upstream changes for subinterpreters in 3.13. that aren't captured in this file (the module is still private and named _interpreters now). I'll issue a separate PR for that.

[ngoldbaum]

(this is on main btw)

frame #7: 0x00000001011a6b68 libpython3.12d.dylib`gc_decref(g=0x00000001045397b0) at gcmodule.c:113:5
   110 	static inline void
   111 	gc_decref(PyGC_Head *g)
   112 	{
-> 113 	    _PyObject_ASSERT_WITH_MSG(FROM_GC(g),
   114 	                              gc_get_refs(g) > 0,
   115 	                              "refcount is too small");
   116 	    g->_gc_prev -= 1 << _PyGC_PREV_SHIFT;

* thread #1, queue = 'com.apple.main-thread', stop reason = signal SIGABRT
    frame #0: 0x000000018425aa60 libsystem_kernel.dylib`__pthread_kill + 8
    frame #1: 0x0000000184292c20 libsystem_pthread.dylib`pthread_kill + 288
    frame #2: 0x000000018419fa30 libsystem_c.dylib`abort + 180
    frame #3: 0x000000010114713c libpython3.12d.dylib`fatal_error_exit(status=-1) at pylifecycle.c:2735:9
    frame #4: 0x0000000101147000 libpython3.12d.dylib`fatal_error(fd=2, header=1, prefix="_PyObject_AssertFailed", msg="_PyObject_AssertFailed", status=-1) at pylifecycle.c:2916:5
    frame #5: 0x00000001011468dc libpython3.12d.dylib`_Py_FatalErrorFunc(func="_PyObject_AssertFailed", msg="_PyObject_AssertFailed") at pylifecycle.c:2932:5
    frame #6: 0x0000000100f6e714 libpython3.12d.dylib`_PyObject_AssertFailed(obj=[], expr="gc_get_refs(g) > 0", msg="refcount is too small", file="Modules/gcmodule.c", line=115, function="gc_decref") at object.c:2603:5
  * frame #7: 0x00000001011a6b68 libpython3.12d.dylib`gc_decref(g=0x00000001045397b0) at gcmodule.c:113:5
    frame #8: 0x00000001011a6ad0 libpython3.12d.dylib`visit_decref(op=[], parent=0x000000010474d850) at gcmodule.c:472:13
    frame #9: 0x0000000100f4e850 libpython3.12d.dylib`dict_traverse(op=0x000000010474d850, visit=(libpython3.12d.dylib`visit_decref at gcmodule.c:462), arg=0x000000010474d850) at dictobject.c:3545:17
    frame #10: 0x00000001011a6768 libpython3.12d.dylib`subtract_refs(containers=0x00000001014fa4a8) at gcmodule.c:491:16
    frame #11: 0x00000001011a5860 libpython3.12d.dylib`deduce_unreachable(base=0x00000001014fa4a8, unreachable=0x000000016fdcdda8) at gcmodule.c:1116:5
    frame #12: 0x00000001011a1760 libpython3.12d.dylib`gc_collect_main(tstate=0x0000000101557d80, generation=0, n_collected=0x000000016fdcde20, n_uncollectable=0x000000016fdcde18, nofail=0) at gcmodule.c:1242:5
    frame #13: 0x00000001011a13c4 libpython3.12d.dylib`gc_collect_with_callback(tstate=0x0000000101557d80, generation=0) at gcmodule.c:1426:14
    frame #14: 0x00000001011a27f8 libpython3.12d.dylib`gc_collect_generations(tstate=0x0000000101557d80) at gcmodule.c:1481:17
    frame #15: 0x00000001011a270c libpython3.12d.dylib`_Py_RunGC(tstate=0x0000000101557d80) at gcmodule.c:2295:5
    frame #16: 0x000000010110a644 libpython3.12d.dylib`_Py_HandlePending(tstate=0x0000000101557d80) at ceval_gil.c:1045:9
    frame #17: 0x00000001010908cc libpython3.12d.dylib`_PyEval_EvalFrameDefault(tstate=0x0000000101557d80, frame=0x000000010010f0e0, throwflag=0) at ceval.c:834:9

[ngoldbaum]

Odd, if I go back to a version of PyO3 from last month the issue is still there on the debug python build, so it's not a new problem. I have no idea why it's suddenly showing up in CI.

[davidhewitt]

The bug with test_gevent is zopefoundation/zope.interface#316 (comment) - I can fix locally if I install zope.interface<7. Will push a PR to fix.