[C API] Add Py_Is(x, y) and Py_IsNone(x) functions

[vstinner]

BPO	43753
Nosy	@tim-one, @rhettinger, @mdickinson, @vstinner, @cfbolz, @markshannon, @pablogsal, @shihai1991, @erlend-aasland
PRs	bpo-43753: Add Py_Is() and Py_IsNone() functions #25227 bpo-43753: _operator.is_() uses Py_Is() #28641

^{Note: these values reflect the state of the issue at the time it was migrated and might not reflect the current state.}

Show more details

GitHub fields:

assignee = None
closed_at = <Date 2021-05-18.20:44:33.857>
created_at = <Date 2021-04-06.18:38:58.346>
labels = ['expert-C-API', '3.10']
title = '[C API] Add Py_Is(x, y) and Py_IsNone(x) functions'
updated_at = <Date 2021-09-29.23:28:18.199>
user = 'https://github.com/vstinner'

bugs.python.org fields:

activity = <Date 2021-09-29.23:28:18.199>
actor = 'vstinner'
assignee = 'none'
closed = True
closed_date = <Date 2021-05-18.20:44:33.857>
closer = 'vstinner'
components = ['C API']
creation = <Date 2021-04-06.18:38:58.346>
creator = 'vstinner'
dependencies = []
files = []
hgrepos = []
issue_num = 43753
keywords = ['patch']
message_count = 19.0
messages = ['390358', '390360', '390364', '390379', '390408', '390411', '390412', '390435', '390456', '390469', '390526', '390613', '390614', '390616', '390628', '390750', '390751', '393907', '402922']
nosy_count = 9.0
nosy_names = ['tim.peters', 'rhettinger', 'mark.dickinson', 'vstinner', 'Carl.Friedrich.Bolz', 'Mark.Shannon', 'pablogsal', 'shihai1991', 'erlendaasland']
pr_nums = ['25227', '28641']
priority = 'normal'
resolution = 'fixed'
stage = 'resolved'
status = 'closed'
superseder = None
type = None
url = 'https://bugs.python.org/issue43753'
versions = ['Python 3.10']

[vstinner]

I propose to add at least Py_Is(x, y) function to the Python C API which would be simply implemented as:

    static inline int Py_Is(PyObject *x, PyObject *y) { return (x == y); }

Right now, there is no benefit for CPython. The idea is to prepare the CPython code base for future optimization, like tagged pointers. It may help PyPy. It can also help to prepare C extensions to migrate to HPy, since HPy requires to use HPy_Is(ctx, x, y): "x == y" (where x and y types is HPy) fails with a compiler error with HPy.

--

CPython uses reference counting and singletons like None and False. The "x is y" operator in Python is implemented with the IS_OP opcode implemented in Python/ceval.c as:

        case TARGET(IS_OP): {
            PyObject *right = POP();
            PyObject *left = TOP();
            int res = (left == right)^oparg;
            PyObject *b = res ? Py_True : Py_False;
            Py_INCREF(b);
            SET_TOP(b);
            Py_DECREF(left);
            Py_DECREF(right);
            PREDICT(POP_JUMP_IF_FALSE);
            PREDICT(POP_JUMP_IF_TRUE);
            DISPATCH();
        }

In short, "x is y" in Python simply compares directly PyObject* pointer values in C: "x == y" where x and y are PyObject* pointers.

PyPy doesn't use reference counting and so implements "x is y" differently: id(x) == id(y) if I understood correctly. In PyPy, id(obj) is more complex than simply getting the object memory address. For example, id(1) is always 17 in PyPy (on Linux/x86-64 at least).

At the C API level, using "x == y" to check if y object "is" x object doesn't work well with PyPy object models. Moreover, "x == Py_None" doesn't work if None is stored as a tagged pointer in CPython.

--

Maybe we can also add functions to check if an object is a singleton:

• Py_IsNone(x): x == Py_None
• Py_IsTrue(x): x == Py_True
• Py_IsFalse(x): x == Py_False

See also bpo-39511 "[subinterpreters] Per-interpreter singletons (None, True, False, etc.)" which may need to replace Py_None singleton with Py_GetNone(). IMO it makes more sense to call Py_IS_NONE(x) function to express that code is run, rather than writing "x == Py_None" as if Py_None is and will always be a variable directly accesssible.

Py_IS_TRUE() name sounds like PyObject_IsTrue(). Can it be an issue? Can someone come with a better name?

[vstinner]

Python has more singletons. In C:

• Py_NotImplemented: NotImplemented
• Py_Ellipsis: Ellipsis

But I don't think that "x == NotImplemented" is common enough to justify a new function.

[vstinner]

Py_IS_TYPE(obj, type) was added to Python 3.9 by bpo-39573:
https://docs.python.org/dev/c-api/structures.html#c.Py_IS_TYPE

commit d905df7
Author: Dong-hee Na <donghee.na92@gmail.com>
Date: Fri Feb 14 02:37:17 2020 +0900

bpo-39573: Add Py_IS_TYPE() function (GH-18488)

Co-Author: Neil Schemenauer <nas-github@arctrix.com>

It's currently implemented as:

static inline int _Py_IS_TYPE(const PyObject *ob, const PyTypeObject *type) {
    return Py_TYPE(ob) == type;
}
#define Py_IS_TYPE(ob, type) _Py_IS_TYPE(_PyObject_CAST_CONST(ob), type)

[rhettinger]

Right now, there is no benefit for CPython.

Please don't this until we have a clear demonstrable benefit. As it stands now, this is all cost and no benefit.

Adding unnecessary abstraction layers just makes it more difficult for people to learn to be core devs. Also, it will likely result in a lot of pointless code churn where each change carries a risk of a new bug being introduced.

The notion of pointer comparison is so fundamental to C code that it is counter productive to try to abstract it away. It isn't much different than saying that every instance of "a + b" should be abstracted to "add(a, b)" and every "a[b]" should be abstracted to "index_lookup(a, b)" on the hope that maybe it might someday be helpful for PyPy or HPy even though they have never requested such a change.

From my own point of view, these need abstractions just make it harder to tell what code is actually doing. Further, it will just lead to wasting everyone's time in code reviews where the reviewer insists on the applying the new inline function and there is confusion about whether two pointers are generic pointers or python object pointers, each with their own comparison technique.

Also, there is too much faith in functions marked as "inline" always being inlined. Compilers get to make their own choices and under some circumstances will not inline, especially for cross module calls. This risks taking code that is currently obviously fast and occasionally, invisibility slowing it down massively — from a step that is 1 cycle at most and is sometimes zero cost to a step that involves an actual function call, possibly needing to save and restore registers.

Lastly, the API changes aren't just for you or the standard library. In effect, you're telling the entire ecosystem of C extensions that they are doing it wrong. Almost certainly, some will follow this path and some won't, further fracturing the ecosystem.

[markshannon]

For any sane design of tagged pointers, x == y (in C) will work fine.

is is not well defined except for a small set of values, so the docs for Py_Is would have to so vague as to be worthless, IMO.

[vstinner]

For any sane design of tagged pointers, x == y (in C) will work fine.

I wrote a proof-of-concept for using tagged pointners in CPython:
vstinner#6

"The PR is large because of the first changes which add Py_IS_NONE(), Py_IS_TRUE() and Py_IS_FALSE() macros."

For backward compatibility, I added a function to get a concrete Python object from a tagged pointer: the True as a tagged pointer becomes (PyObject *)&_Py_TrueStruct concrete object. Py_IS_TRUE() has to check for both values: tagged pointer or &_Py_TrueStruct (don't look at my exact implementation, it's wrong ;-)).

In a perfect world, there would be no backward compatibility and you would never have to create Python objects from a tagged pointer.

The other problem is that the stable ABI exposes "Py_True" as "&_Py_TrueStruct" and so C extensions built with the stable ABI uses "&_Py_TrueStruct", not the tagged pointer.

See also bpo-39511 which discuss solutions for these problems.

[vstinner]

Also, there is too much faith in functions marked as "inline" always being inlined.

I proposed to declare it as a "static inline" function, but I'm fine with a macro as well. I mostly care about the API: call "Py_Is()", not really about the exact implementation. In practice, for now, Py_Is(x, y) will continue to be compiled as "x == y".