Assigning `__conditional_annotations__` crashes the interpreter

[Lydxn]

Crash report

What happened?

In CPython 3.14 and later, assigning to the global __conditional_annotations__ before a variable annotation produces a segfault.

Here is a minimal reproducible example:

__conditional_annotations__ = 0
a: 1

Analyzing the bytecode, the compiler first initializes __conditional_annotations__ to an empty set. Whenever an annotation is declared, a SET_ADD is emitted under the assumption that the underlying type is still a set. But we can re-assign __conditional_annotations__ to another object in between, leading to a type confusion:

BUILD_SET                0
STORE_NAME               0 (__conditional_annotations__)
LOAD_SMALL_INT           0
STORE_NAME               0 (__conditional_annotations__)

LOAD_NAME                0 (__conditional_annotations__)
LOAD_SMALL_INT           0
SET_ADD                  1

CPython versions tested on:

3.14, 3.15

Operating systems tested on:

Linux

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

No response

Linked PRs

• gh-154902: Add an intrinsic for conditional annotations #155026
• [3.15] gh-154902: Type-check the SET_ADD operand #155071
• [3.14] gh-154902: Type-check the SET_ADD operand #155072

[deadlovelll]

hey @Lydxn! If you don't mind, can i pick this?

[johnslavik]

cc @JelleZijlstra

[picnixz]

@BHUVANSH855 Avoid low quality comments that restate what the OP already stated especially if it is an agent that generated the output. Check our AI policy in the devguide.

[JelleZijlstra]

PRs for this are welcome but it's helpful first to discuss how we should fix this, subject to the constraint that we cannot change the generated bytecode.

[deadlovelll]

PRs for this are welcome but it's helpful first to discuss how we should fix this, subject to the constraint that we cannot change the generated bytecode.

I can take a look and let you know later what the options for fixing it are

[deadlovelll]

hey @JelleZijlstra! I looked at the code, and what i found:

__conditional_annotations__ is a set and assigning any type instead of set is unacceptable. So, i think the proper solution here is to add PySet_Check in inst(SET_ADD) at Python/bytecodes.c

Like this:

            if (!PySet_Check(set_o)) {
                _PyErr_Format(tstate, PyExc_TypeError,
                              "__conditional_annotations__ must be a set, not %T",
                              set_o);
                PyStackRef_CLOSE(v);
                ERROR_IF(true);
            }

During the bc inspection bytecode doesn changed

You can also check it with this script:

import dis

src = "__conditional_annotations__ = 0\na: 1\n"
dis.dis(compile(src, "repro.py", "exec"))

If it looks good to you i'll create the pr, otherwise the solution can be discussed =)

[JelleZijlstra]

The trouble is that SET_ADD is also used elsewhere (I presume in set comprehensions), so we risk slowing down their code. Maybe that's OK, but let's do some more thinking about what we can do that fixes this crash without slowing down or breaking anything else.

[Nico-Posada]

Why can’t we change the generated bytecode? That seems like the easiest way to fix it without adding the set check in SET_ADD