`test_call.test_super_deep` started segfaulting with 3.14.0a4 freethreading

[mgorny]

Crash report

What happened?

test_super_deep (test.test_call.TestRecursion.test_super_deep) ... Fatal Python error: Segmentation fault

The test passes with a3 but segfaults with a4 — so I'm guessing stack frames grew or something. Not sure if this is the kind of regression in code that needs to be fixed, or merely recursion limit needs to be lowered.

This is Gentoo Linux amd64, CPython built with LTO + PGO, CFLAGS are -march=znver2 --param=l1-cache-size=32 --param=l1-cache-line-size=64 -O2 -pipe -frecord-gcc-switches.

Full log: dev-lang:python-3.14.0_alpha4-r100:20250117-163521.txt

CPython versions tested on:

3.14

Operating systems tested on:

Linux

Output from running 'python -VV' on the command line:

Python 3.14.0a4 experimental free-threading build (main, Jan 17 2025, 17:46:22) [GCC 14.2.1 20241221]

Linked PRs

• gh-128954: Reorder _PyInterpreterFrame fields for reduced memory usage #128958

[mgorny]

#127455 seems to indicate that they're hitting a similar problem on Windows.

[colesbury]

We can lower the recursion limit, but it seems like it'd be good to reduce the stack usage in the free-threading build:

cpython/Include/cpython/pystate.h

Line 232 in d95ba9f

# define Py_C_RECURSION_LIMIT 10000

[mgorny]

I'm going to try if bisect can tell what change caused the issue to appear with a4.

[mgorny]

Bisect points to 128cc47:

commit 128cc47fbd44e3e09c50d9674fe4a4bba3be450c (HEAD)
Author:     Mark Shannon <mark@hotpy.org>
AuthorDate: 2024-12-20 17:52:20 +0100
Commit:     GitHub <noreply@github.com>
CommitDate: 2024-12-20 17:52:20 +0100

    GH-127705: Add debug mode for `_PyStackRef`s inspired by HPy debug mode (GH-128121)

CC @markshannon