JIT (correctness): Different `UnboundLocalError` for same code with JIT on or off

[devdanzin]

Bug report

Bug description:

The following code results in two different UnboundLocalError exceptions depending on whether it's run with the JIT on or off:

def f():

    if 0:
        class Class:
            pass

    if 0:
        instance = Class()

    for _ in range(5000):
        try:
            _ = instance.f()
        except Exception:
            pass
    Class.__bases__ = (object,)

f()

With JIT on, the UnboundLocalError points to 'instance', while with JIT off it points to 'Class'.

Traceback with JIT on:

Traceback (most recent call last):
  File "/home/danzin/divergences/divergence_child_6010_3_1.py", line 17, in <module>
    f()
    ~^^
  File "/home/danzin/divergences/divergence_child_6010_3_1.py", line 12, in f
    _ = instance.f()
        ^^^^^^^^
UnboundLocalError: cannot access local variable 'instance' where it is not associated with a value

Traceback with JIT off:

Traceback (most recent call last):
  File "/home/danzin/divergences/divergence_child_6010_3_1.py", line 17, in <module>
    f()
    ~^^
  File "/home/danzin/divergences/divergence_child_6010_3_1.py", line 15, in f
    Class.__bases__ = (object,)
    ^^^^^
UnboundLocalError: cannot access local variable 'Class' where it is not associated with a value

Output for PYTHON_LLTRACE=4:
divergence_01_lltrace.txt

Output for PYTHON_OPT_DEBUG=4:
divergence_01_opt_debug.txt

Please let me know whether you're able to reproduce this issue.

Found using lafleur.

CPython versions tested on:

CPython main branch

Operating systems tested on:

Linux

Linked PRs

• gh-140104: Set next_instr properly in the JIT during exceptions #140233
• [3.14] gh-140104: Set next_instr properly in the JIT during exceptions (GH-140233) #140687

[Fidget-Spinner]

It's possible #139109 might end up fixing this inadvertently, so let's not fix this for now maybe? Or we could fix it with the knowledge the code will be ripped out soon.

[devdanzin]

Here's a simpler reproducer for what seems to be the same issue:

def f():
    for i in range(500):
        try:
            g(i)
        except Exception:
            pass

for _ in range(5000):
    f()

This one runs to completion without the JIT and raises an exception with it.

[efimov-mikhail]

We need a backport for 3.14, don't we?
Or we ignore backports since JIT is in expiremental stage?

[Fidget-Spinner]

Let's backport to 3.14 but not 3.13.

[Fidget-Spinner]

Backport failed, maybe I should open up the backport to anyone who wants to try please?

[efimov-mikhail]

Backport failed, maybe I should open up the backport to anyone who wants to try please?

I'd like to provide this backport.