GH-141861: Fix invalid memory read in the ENTER_EXECUTOR

[sergey-miryanov]

• Issue: JIT: invalid memory read in _PyEval_EvalFrameDefault #141861

[sergey-miryanov]

@Fidget-Spinner Could you please take a look?

[Fidget-Spinner]

Thank you for reducing the test case.

[markshannon]

Does this really fix the issue?
Without a proper analysis of the issue, I don't see how we can tell if this is fixed.

Is this a recent issue? When did it start?

[sergey-miryanov]

@markshannon
Fix for this PR provided by @Fidget-Spinner

Failure bisected to a3b78a3

a3b78a3adecf794dc576404b14ed3f2c8401fdf1 is the first bad commit
commit a3b78a3adecf794dc576404b14ed3f2c8401fdf1
Author: Mikhail Efimov <efimov.mikhail@gmail.com>
Date:   Fri Nov 21 20:00:25 2025 +0300

    gh-141498: Change backoff counter to use prime numbers instead of powers of 2 (GH-141591)

 Include/internal/pycore_backoff.h                  | 68 +++++++++++++---------
 Lib/test/test_opcache.py                           |  2 +-
 .../2025-11-15-14-04-35.gh-issue-141589.VfdMDD.rst |  3 +
 3 files changed, 44 insertions(+), 29 deletions(-)
 create mode 100644 Misc/NEWS.d/next/Core_and_Builtins/2025-11-15-14-04-35.gh-issue-141589.VfdMDD.rst

[markshannon]

What's the root cause?
On the issue @Fidget-Spinner says "The problem is that it's possible for oparg to change under our feet".
What is causing the oparg to change and when?

[markshannon]

Since the opcode/oparg pair are written atomically, the fix looks wrong.
If ENTER_EXECUTOR is overwritten (by instrumentation presumably) then the new this_instr->op.arg could easily be an invalid index.

[sergey-miryanov]

I noticed that if I disable specialization (ENABLE_SPECIALIZATION=0) then failure is gone. However, I'm not sure if this is useful.