Exceptions are represented in the interpreter as (type, val, tb) triplets which most of the time contain redundant information (the type is the type of val and the tb is also on the exception). This complicates the code and is inefficient as opcodes that manage exceptions push and pop 3 items for each exception.

We will change the internal representation to be (1) just the exception value if it is normalised and (2) a tuple of the 3 values for the uncommon case where they are all needed.

See also faster-cpython/ideas#106.

Would there be any change at the Python level?

Initially not, neither in python nor in the c api.

It would be nice to replace PyErr_Fetch/Restore by a version that takes just an exception but that’s a long deprecation.

New changeset 05fbd60 by Irit Katriel in branch 'main':
bpo-45711: Use _PyErr_ClearExcState instead of setting only exc_value to NULL (GH-29404)
05fbd60

New changeset 4cdeee5 by Irit Katriel in branch 'main':
bpo-45711: remove unnecessary DUP_TOP and POP in exception handling (GH-29495)
4cdeee5

New changeset 8f1b71d by Brandt Bucher in branch 'main':
bpo-45711: Re-bump the magic number and update doc (GH-29528)
8f1b71d

New changeset de3db14 by Irit Katriel in branch 'main':
bpo-45711: assert that the type of exc_info is redundant (GH-29518)
de3db14

New changeset c456dfa by Irit Katriel in branch 'main':
bpo-45711: use exc_value instead of exc_type to determine if exc_info is valid. Add more assertions. (GH-29627)
c456dfa

Following the analysis/discussion on faster-cpython/ideas#106:

• exc_info is always normalized, so we actually will never need to create a tuple of these three values (exc_curinfo, the exception raised but not yet caught can be unnormalized, but it is not pushed/popped on the stack).

• We will reduce the interpreter's exc_info representation to just the exception instance.

• There are two APIs that are impacted, both in non-documented edge cases:

1. sys.exc_info()[2] can currently be different from sys.exc_info()[1].__traceback__ because changes to the latter (while an except clause is executing) don't show up in the former. This is arguably a bug, we will change it so that the type and traceback are always consistent with the exception.

2. PyErr_SetExcInfo does no arg checking, and will set exc_info to an inconsistent triplet if you ask it to. However, the exc_value arg must be an exception instance so the only thing you can do is pass in nonsensical args where the type/traceback do not match the exception. This function's purpose is to save/restore exc_info. We will make it ignore the type and traceback and document that change.

New changeset 8a45ca5 by Irit Katriel in branch 'main':
bpo-45711: Change exc_info related APIs to derive type and traceback from the exception instance (GH-29780)
8a45ca5

New changeset 2ff758b by Irit Katriel in branch 'main':
bpo-45711: [asyncio] Normalize exceptions immediately after Fetch, before they are stored as StackItem, which should be normalized (GH-29890)
2ff758b

New changeset 2109f78 by Irit Katriel in branch 'main':
bpo-45711: Remove unnecessary normalization of exc_info (GH-29922)
2109f78

FYI, we track the Cython side of this in
cython/cython#4500

New changeset 396b583 by Irit Katriel in branch 'main':
bpo-45711: Remove type and traceback from exc_info (GH-30122)
396b583

This is done, there will be more cleanups that this enables but they can have their own issues.

New changeset 80e1def by Irit Katriel in branch 'main':
bpo-45711: move whatsnew entries which are incorrectly listed under New Features (GH-30849)
80e1def

Would it be possible to revert the change until Cython and numpy are ready for it?

bpo-45711: Remove type and traceback from exc_info (GH-30122)

This change broke numpy which uses C code generated by Cython for coroutines in numpy/random/_mt19937.c (file generated from numpy/random/_mt19937.pyx).

Example of Cython code which no fails with a compiler error since exc_info->exc_type has been removed:

/* GetTopmostException */
#if CYTHON_USE_EXC_INFO_STACK
static _PyErr_StackItem *
__Pyx_PyErr_GetTopmostException(PyThreadState *tstate)
{
    _PyErr_StackItem *exc_info = tstate->exc_info;
    while ((exc_info->exc_type == NULL || exc_info->exc_type == Py_None) &&
           exc_info->previous_item != NULL)
    {
        exc_info = exc_info->previous_item;
    }
    return exc_info;
}
#endif

I propose this plan:

• (1) Revert this specific change
• (2) Wait until a change is merged in Cython 0.29.x
• (3) Wait until a new Cython 0.29.x version is released
• (4) Wait until numpy is released with regenerated code compatible with this change
• (5) Apply again the change

It should increase the number of packages when Python 3.11 will be released.

In Cython, the issue is tracked as: cython/cython#4500

I can probably just put back the two fields in _PyErr_StackItem and make sure they get updated when exc_value is set.

Reverting the whole thing would include big changes in compile.c and ceval.c, so I'd rather avoid that if we can.

Does unsetting CYTHON_USE_EXC_INFO_STACK still work?

#if CYTHON_USE_EXC_INFO_STACK
// See  https://bugs.python.org/issue25612
#define __Pyx_ExcInfoStruct  _PyErr_StackItem
#else
// Minimal replacement struct for Py<3.7, without the Py3.7 exception state stack.
typedef struct {
    PyObject *exc_type;
    PyObject *exc_value;
    PyObject *exc_traceback;
} __Pyx_ExcInfoStruct;
#endif

In Cython, the issue is tracked as: cython/cython#4500

I don't understand the Cython status. The code was updated in the master branch:

• Adapt to "exc_info" changes in CPython 3.11a4 cython/cython#4584
• cython/cython@7769570

cython/cython#4500 was closed even if the change was not backported to 0.29.x.

I understand that there is no released Cython 0.29.x version compatible with this change.

I commented the Cython issue:
cython/cython#4500 (comment)

I also opened a discussion on python-dev: "Please update Cython before introcuding C API incompatible changes in Python".
https://mail.python.org/archives/list/python-dev@python.org/thread/RS2C53LDZPXHRR2VCY2G2YSPDVA4LNQU/

Irit Katriel:

Reverting the whole thing would include big changes in compile.c and ceval.c, so I'd rather avoid that if we can.

It was a good idea to split changes of this issue into multiple commits :-) I'm only proposing to revert the one which introduces the C API incompatible changes: the commit 396b583 "bpo-45711: Remove type and traceback from exc_info (GH-30122)".

Maybe Cython can be quickly updated, but the situation seems to be complicated :-( cython/cython#4500

That commit has significant changes in ceval.c and compile.c. They don't need to be reverted to unbreak cython. I'm working on a PR for a simpler change.

Have you tried with CYTHON_USE_EXC_INFO_STACK undefined?

If this is still the position of cython maintainers:

cython/cython#4581 (comment)

then I will need to revert the change until 3.12.

Time to insist on directly communicating with the Cython team (esp. @scoder) and broker some kind of compromise.

__Pyx_PyErr_GetTopmostException(PyThreadState *tstate)

Python provides a *private* _PyErr_GetTopmostException(tstate) function, but Cython reimplements its own function. I'm not sure why.

#74716 proposes adding PyErr_GetActiveException() function which has no parameter, but Cython __Pyx_PyErr_GetTopmostException() has a tstate parameter.

Simplified example numpy/random/_mt19937.c code:

static CYTHON_INLINE void
__Pyx__ExceptionSave(PyThreadState *tstate, PyObject **type,
                     PyObject **value, PyObject **tb)
{
    _PyErr_StackItem *exc_info = __Pyx_PyErr_GetTopmostException(tstate);
    *type = exc_info->exc_type;
    *value = exc_info->exc_value;
    *tb = exc_info->exc_traceback;
    Py_XINCREF(*type);
    Py_XINCREF(*value);
    Py_XINCREF(*tb);
}

static CYTHON_INLINE void
__Pyx__ExceptionReset(PyThreadState *tstate, PyObject *type,
                      PyObject *value, PyObject *tb)
{
    PyObject *tmp_type, *tmp_value, *tmp_tb;
    _PyErr_StackItem *exc_info = tstate->exc_info;
    tmp_type = exc_info->exc_type;
    tmp_value = exc_info->exc_value;
    tmp_tb = exc_info->exc_traceback;
    exc_info->exc_type = type;
    exc_info->exc_value = value;
    exc_info->exc_traceback = tb;
    Py_XDECREF(tmp_type);
    Py_XDECREF(tmp_value);
    Py_XDECREF(tmp_tb);
}

Cython saves/restores the current exception of tstate. Maybe we need to provide a high-level API for that as well?

This is a backport of @scoder's patch to 0.29.x. (I don't know if this is helpful).

https://github.com/cython/cython/compare/master...iritkatriel:exc_info?expand=1

#74716 proposes adding PyErr_GetActiveException() function which has no parameter, but Cython __Pyx_PyErr_GetTopmostException() has a tstate parameter.

I've now updated it to follow the pattern of other functions, where the is a private function that takes tstate and the public function calls it.

So it adds in Include/pyerrors.h

PyAPI_FUNC(PyObject*) PyErr_GetActiveException(void);
PyAPI_FUNC(void) PyErr_SetActiveException(PyObject *);

and in Include/cpython/pyerrors.h

PyAPI_FUNC(PyObject*) _PyErr_GetActiveException(PyThreadState *);
PyAPI_FUNC(void) _PyErr_SetActiveException(PyThreadState *, PyObject *);

Just a quick comment on Cython and these changes:

Cython 0.29 can build itself on Python 3.11a4 with CFLAGS="-DCYTHON_FAST_THREAD_STATE=0 -DCYTHON_USE_EXC_INFO_STACK=0" python3.11 setup.py build_ext.

I think there's some coroutine code left that isn't fixed by those flags, but a large chunk of Cython libraries won't use coroutines and so will work fine with these flags. Hopefully that unblocks some stuff for you.

lxml does crash on the current Cython 0.29.x development branch.

We have agreed on python-dev [1] that the cpython changes will not be reverted, and the issue will be fixed in cython. So I am closing this again.

[1] https://mail.python.org/archives/list/python-dev@python.org/message/BHIQL4P6F7OPMCAP6U24XEZUPQKI62UT/