What’s New In Python 3.11

Release

3.11.0

Date

十一月 10, 2022

Editor

Pablo Galindo Salgado

This article explains the new features in Python 3.11, compared to 3.10.

For full details, see the changelog.

Summary — Release highlights

  • Python 3.11 is between 10-60% faster than Python 3.10. On average, we measured a 1.25x speedup on the standard benchmark suite. See Faster CPython for details.

New syntax features:

New built-in features:

New standard library modules:

Interpreter improvements:

New typing features:

Important deprecations, removals and restrictions:

New Features

PEP 657: Fine-grained error locations in tracebacks

When printing tracebacks, the interpreter will now point to the exact expression that caused the error, instead of just the line. For example:

  1. Traceback (most recent call last):
  2.   File "distance.py", line 11, in <module>
  3.     print(manhattan_distance(p1, p2))
  4.           ^^^^^^^^^^^^^^^^^^^^^^^^^^
  5.   File "distance.py", line 6, in manhattan_distance
  6.     return abs(point_1.x - point_2.x) + abs(point_1.y - point_2.y)
  7.                            ^^^^^^^^^
  8. AttributeError: 'NoneType' object has no attribute 'x'

Previous versions of the interpreter would point to just the line, making it ambiguous which object was None. These enhanced errors can also be helpful when dealing with deeply nested dict objects and multiple function calls:

  1. Traceback (most recent call last):
  2.   File "query.py", line 37, in <module>
  3.     magic_arithmetic('foo')
  4.   File "query.py", line 18, in magic_arithmetic
  5.     return add_counts(x) / 25
  6.            ^^^^^^^^^^^^^
  7.   File "query.py", line 24, in add_counts
  8.     return 25 + query_user(user1) + query_user(user2)
  9.                 ^^^^^^^^^^^^^^^^^
  10.   File "query.py", line 32, in query_user
  11.     return 1 + query_count(db, response['a']['b']['c']['user'], retry=True)
  12.                                ~~~~~~~~~~~~~~~~~~^^^^^
  13. TypeError: 'NoneType' object is not subscriptable

As well as complex arithmetic expressions:

  1. Traceback (most recent call last):
  2.   File "calculation.py", line 54, in <module>
  3.     result = (x / y / z) * (a / b / c)
  4.               ~~~~~~^~~
  5. ZeroDivisionError: division by zero

Additionally, the information used by the enhanced traceback feature is made available via a general API, that can be used to correlate bytecode instructions with source code location. This information can be retrieved using:

See PEP 657 for more details. (Contributed by Pablo Galindo, Batuhan Taskaya and Ammar Askar in bpo-43950.)

备注

This feature requires storing column positions in 代码对象, which may result in a small increase in interpreter memory usage and disk usage for compiled Python files. To avoid storing the extra information and deactivate printing the extra traceback information, use the -X no_debug_ranges command line option or the PYTHONNODEBUGRANGES environment variable.

PEP 654: Exception Groups and except*

PEP 654 introduces language features that enable a program to raise and handle multiple unrelated exceptions simultaneously. The builtin types ExceptionGroup and BaseExceptionGroup make it possible to group exceptions and raise them together, and the new except* syntax generalizes except to match subgroups of exception groups.

See PEP 654 for more details.

(Contributed by Irit Katriel in bpo-45292. PEP written by Irit Katriel, Yury Selivanov and Guido van Rossum.)

PEP 678: Exceptions can be enriched with notes

The add_note() method is added to BaseException. It can be used to enrich exceptions with context information that is not available at the time when the exception is raised. The added notes appear in the default traceback.

See PEP 678 for more details.

(Contributed by Irit Katriel in bpo-45607. PEP written by Zac Hatfield-Dodds.)

Windows py.exe launcher improvements

The copy of the 适用于Windows的Python启动器 included with Python 3.11 has been significantly updated. It now supports company/tag syntax as defined in PEP 514 using the -V:<company>/<tag> argument instead of the limited -<major>.<minor>. This allows launching distributions other than PythonCore, the one hosted on python.org.

When using -V: selectors, either company or tag can be omitted, but all installs will be searched. For example, -V:OtherPython/ will select the “best” tag registered for OtherPython, while -V:3.11 or -V:/3.11 will select the “best” distribution with tag 3.11.

When using the legacy -<major>, -<major>.<minor>, -<major>-<bitness> or -<major>.<minor>-<bitness> arguments, all existing behaviour should be preserved from past versions, and only releases from PythonCore will be selected. However, the -64 suffix now implies “not 32-bit” (not necessarily x86-64), as there are multiple supported 64-bit platforms. 32-bit runtimes are detected by checking the runtime’s tag for a -32 suffix. All releases of Python since 3.5 have included this in their 32-bit builds.

This section covers major changes affecting PEP 484 type hints and the typing module.

PEP 646: Variadic generics

PEP 484 previously introduced TypeVar, enabling creation of generics parameterised with a single type. PEP 646 adds TypeVarTuple, enabling parameterisation with an arbitrary number of types. In other words, a TypeVarTuple is a variadic type variable, enabling variadic generics.

This enables a wide variety of use cases. In particular, it allows the type of array-like structures in numerical computing libraries such as NumPy and TensorFlow to be parameterised with the array shape. Static type checkers will now be able to catch shape-related bugs in code that uses these libraries.

See PEP 646 for more details.

(Contributed by Matthew Rahtz in bpo-43224, with contributions by Serhiy Storchaka and Jelle Zijlstra. PEP written by Mark Mendoza, Matthew Rahtz, Pradeep Kumar Srinivasan, and Vincent Siles.)

PEP 655: Marking individual TypedDict items as required or not-required

Required and NotRequired provide a straightforward way to mark whether individual items in a TypedDict must be present. Previously, this was only possible using inheritance.

All fields are still required by default, unless the total parameter is set to False, in which case all fields are still not-required by default. For example, the following specifies a TypedDict with one required and one not-required key:

  1. class Movie(TypedDict):
  2.    title: str
  3.    year: NotRequired[int]
  5. m1: Movie = {"title": "Black Panther", "year": 2018}  # OK
  6. m2: Movie = {"title": "Star Wars"}  # OK (year is not required)
  7. m3: Movie = {"year": 2022}  # ERROR (missing required field title)

The following definition is equivalent:

  1. class Movie(TypedDict, total=False):
  2.    title: Required[str]
  3.    year: int

See PEP 655 for more details.

(Contributed by David Foster and Jelle Zijlstra in bpo-47087. PEP written by David Foster.)

PEP 673: Self type

The new Self annotation provides a simple and intuitive way to annotate methods that return an instance of their class. This behaves the same as the TypeVar-based approach specified in PEP 484, but is more concise and easier to follow.

Common use cases include alternative constructors provided as classmethods, and __enter__() methods that return self:

  1. class MyLock:
  2.     def __enter__(self) -> Self:
  3.         self.lock()
  4.         return self
  6.     ...
  8. class MyInt:
  9.     @classmethod
  10.     def fromhex(cls, s: str) -> Self:
  11.         return cls(int(s, 16))
  13.     ...

Self can also be used to annotate method parameters or attributes of the same type as their enclosing class.

See PEP 673 for more details.

(Contributed by James Hilton-Balfe in bpo-46534. PEP written by Pradeep Kumar Srinivasan and James Hilton-Balfe.)

PEP 675: Arbitrary literal string type

The new LiteralString annotation may be used to indicate that a function parameter can be of any literal string type. This allows a function to accept arbitrary literal string types, as well as strings created from other literal strings. Type checkers can then enforce that sensitive functions, such as those that execute SQL statements or shell commands, are called only with static arguments, providing protection against injection attacks.

For example, a SQL query function could be annotated as follows:

  1. def run_query(sql: LiteralString) -> ...
  2.     ...
  4. def caller(
  5.     arbitrary_string: str,
  6.     query_string: LiteralString,
  7.     table_name: LiteralString,
  8. ) -> None:
  9.     run_query("SELECT * FROM students")       # ok
  10.     run_query(query_string)                   # ok
  11.     run_query("SELECT * FROM " + table_name)  # ok
  12.     run_query(arbitrary_string)               # type checker error
  13.     run_query(                                # type checker error
  14.         f"SELECT * FROM students WHERE name = {arbitrary_string}"
  15.     )

See PEP 675 for more details.

(Contributed by Jelle Zijlstra in bpo-47088. PEP written by Pradeep Kumar Srinivasan and Graham Bleaney.)

PEP 681: Data class transforms

dataclass_transform may be used to decorate a class, metaclass, or a function that is itself a decorator. The presence of @dataclass_transform() tells a static type checker that the decorated object performs runtime “magic” that transforms a class, giving it dataclass-like behaviors.

For example:

  1. # The create_model decorator is defined by a library.
  2. @typing.dataclass_transform()
  3. def create_model(cls: Type[T]) -> Type[T]:
  4.     cls.__init__ = ...
  5.     cls.__eq__ = ...
  6.     cls.__ne__ = ...
  7.     return cls
  9. # The create_model decorator can now be used to create new model classes:
  10. @create_model
  11. class CustomerModel:
  12.     id: int
  13.     name: str
  15. c = CustomerModel(id=327, name="Eric Idle")

See PEP 681 for more details.

(Contributed by Jelle Zijlstra in gh-91860. PEP written by Erik De Bonte and Eric Traut.)

PEP 563 may not be the future

PEP 563 Postponed Evaluation of Annotations (the from __future__ import annotations future statement) that was originally planned for release in Python 3.10 has been put on hold indefinitely. See this message from the Steering Council for more information.

Other Language Changes

Other CPython Implementation Changes

  • The special methods __complex__() for complex and __bytes__() for bytes are implemented to support the typing.SupportsComplex and typing.SupportsBytes protocols. (Contributed by Mark Dickinson and Dong-hee Na in bpo-24234.)

  • siphash13 is added as a new internal hashing algorithm. It has similar security properties as siphash24, but it is slightly faster for long inputs. str, bytes, and some other types now use it as the default algorithm for hash(). PEP 552 hash-based .pyc files now use siphash13 too. (Contributed by Inada Naoki in bpo-29410.)

  • When an active exception is re-raised by a raise statement with no parameters, the traceback attached to this exception is now always sys.exc_info()[1].__traceback__. This means that changes made to the traceback in the current except clause are reflected in the re-raised exception. (Contributed by Irit Katriel in bpo-45711.)

  • The interpreter state’s representation of handled exceptions (aka exc_info or _PyErr_StackItem) now only has the exc_value field; exc_type and exc_traceback have been removed, as they can be derived from exc_value. (Contributed by Irit Katriel in bpo-45711.)

  • A new command line option, AppendPath, has been added for the Windows installer. It behaves similarly to PrependPath, but appends the install and scripts directories instead of prepending them. (Contributed by Bastian Neuburger in bpo-44934.)

  • The PyConfig.module_search_paths_set field must now be set to 1 for initialization to use PyConfig.module_search_paths to initialize sys.path. Otherwise, initialization will recalculate the path and replace any values added to module_search_paths.

  • The output of the --help option now fits in 50 lines/80 columns. Information about Python environment variables and -X options is now available using the respective --help-env and --help-xoptions flags, and with the new --help-all. (Contributed by Éric Araujo in bpo-46142.)

  • Converting between int and str in bases other than 2 (binary), 4, 8 (octal), 16 (hexadecimal), or 32 such as base 10 (decimal) now raises a ValueError if the number of digits in string form is above a limit to avoid potential denial of service attacks due to the algorithmic complexity. This is a mitigation for CVE-2020-10735. This limit can be configured or disabled by environment variable, command line flag, or sys APIs. See the integer string conversion length limitation documentation. The default limit is 4300 digits in string form.

New Modules

Improved Modules

asyncio

contextlib

  • Added non parallel-safe chdir() context manager to change the current working directory and then restore it on exit. Simple wrapper around chdir(). (Contributed by Filipe Laíns in bpo-25625)

dataclasses

  • Change field default mutability check, allowing only defaults which are hashable instead of any object which is not an instance of dict, list or set. (Contributed by Eric V. Smith in bpo-44674.)

datetime

enum

  • Renamed EnumMeta to EnumType (EnumMeta kept as an alias).

  • Added StrEnum, with members that can be used as (and must be) strings.

  • Added ReprEnum, which only modifies the __repr__() of members while returning their literal values (rather than names) for __str__() and __format__() (used by str(), format() and f-strings).

  • Changed IntEnum, IntFlag and StrEnum to now inherit from ReprEnum, so their str() output now matches format() (both str(AnIntEnum.ONE) and format(AnIntEnum.ONE) return '1', whereas before str(AnIntEnum.ONE) returned 'AnIntEnum.ONE'.

  • Changed Enum.__format__() (the default for format(), str.format() and f-strings) of enums with mixed-in types (e.g. int, str) to also include the class name in the output, not just the member’s key. This matches the existing behavior of enum.Enum.__str__(), returning e.g. 'AnEnum.MEMBER' for an enum AnEnum(str, Enum) instead of just 'MEMBER'.

  • Added a new boundary class parameter to Flag enums and the FlagBoundary enum with its options, to control how to handle out-of-range flag values.

  • Added the verify() enum decorator and the EnumCheck enum with its options, to check enum classes against several specific constraints.

  • Added the member() and nonmember() decorators, to ensure the decorated object is/is not converted to an enum member.

  • Added the property() decorator, which works like property() except for enums. Use this instead of types.DynamicClassAttribute().

  • Added the global_enum() enum decorator, which adjusts __repr__() and __str__() to show values as members of their module rather than the enum class. For example, 're.ASCII' for the ASCII member of re.RegexFlag rather than 'RegexFlag.ASCII'.

  • Enhanced Flag to support len(), iteration and in/not in on its members. For example, the following now works: len(AFlag(3)) == 2 and list(AFlag(3)) == (AFlag.ONE, AFlag.TWO)

  • Changed Enum and Flag so that members are now defined before __init_subclass__() is called; dir() now includes methods, etc., from mixed-in data types.

  • Changed Flag to only consider primary values (power of two) canonical while composite values (3, 6, 10, etc.) are considered aliases; inverted flags are coerced to their positive equivalent.

fcntl

  • On FreeBSD, the F_DUP2FD and F_DUP2FD_CLOEXEC flags respectively are supported, the former equals to dup2 usage while the latter set the FD_CLOEXEC flag in addition.

fractions

  • Support PEP 515-style initialization of Fraction from string. (Contributed by Sergey B Kirpichev in bpo-44258.)

  • Fraction now implements an __int__ method, so that an isinstance(some_fraction, typing.SupportsInt) check passes. (Contributed by Mark Dickinson in bpo-44547.)

functools

  • functools.singledispatch() now supports types.UnionType and typing.Union as annotations to the dispatch argument.:

    1. >>> from functools import singledispatch
    2. >>> @singledispatch
    3. ... def fun(arg, verbose=False):
    4. ...     if verbose:
    5. ...         print("Let me just say,", end=" ")
    6. ...     print(arg)
    7. ...
    8. >>> @fun.register
    9. ... def _(arg: int | float, verbose=False):
    10. ...     if verbose:
    11. ...         print("Strength in numbers, eh?", end=" ")
    12. ...     print(arg)
    13. ...
    14. >>> from typing import Union
    15. >>> @fun.register
    16. ... def _(arg: Union[list, set], verbose=False):
    17. ...     if verbose:
    18. ...         print("Enumerate this:")
    19. ...     for i, elem in enumerate(arg):
    20. ...         print(i, elem)
    21. ...

    (Contributed by Yurii Karabas in bpo-46014.)

hashlib

  • hashlib.blake2b() and hashlib.blake2s() now prefer libb2 over Python’s vendored copy. (Contributed by Christian Heimes in bpo-47095.)

  • The internal _sha3 module with SHA3 and SHAKE algorithms now uses tiny_sha3 instead of the Keccak Code Package to reduce code and binary size. The hashlib module prefers optimized SHA3 and SHAKE implementations from OpenSSL. The change affects only installations without OpenSSL support. (Contributed by Christian Heimes in bpo-47098.)

  • Add hashlib.file_digest(), a helper function for efficient hashing of files or file-like objects. (Contributed by Christian Heimes in gh-89313.)

IDLE and idlelib

  • Apply syntax highlighting to .pyi files. (Contributed by Alex Waygood and Terry Jan Reedy in bpo-45447.)

  • Include prompts when saving Shell with inputs and outputs. (Contributed by Terry Jan Reedy in gh-95191.)

inspect

locale

logging

math

  • Add math.exp2(): return 2 raised to the power of x. (Contributed by Gideon Mitchell in bpo-45917.)

  • Add math.cbrt(): return the cube root of x. (Contributed by Ajith Ramachandran in bpo-44357.)

  • The behaviour of two math.pow() corner cases was changed, for consistency with the IEEE 754 specification. The operations math.pow(0.0, -math.inf) and math.pow(-0.0, -math.inf) now return inf. Previously they raised ValueError. (Contributed by Mark Dickinson in bpo-44339.)

  • The math.nan value is now always available. (Contributed by Victor Stinner in bpo-46917.)

operator

  • A new function operator.call has been added, such that operator.call(obj, *args, **kwargs) == obj(*args, **kwargs). (Contributed by Antony Lee in bpo-44019.)

os

  • On Windows, os.urandom() now uses BCryptGenRandom(), instead of CryptGenRandom() which is deprecated. (Contributed by Dong-hee Na in bpo-44611.)

pathlib

re

  • Atomic grouping ((?>...)) and possessive quantifiers (*+, ++, ?+, {m,n}+) are now supported in regular expressions. (Contributed by Jeffrey C. Jacobs and Serhiy Storchaka in bpo-433030.)

shutil

socket

  • Add CAN Socket support for NetBSD. (Contributed by Thomas Klausner in bpo-30512.)

  • create_connection() has an option to raise, in case of failure to connect, an ExceptionGroup containing all errors instead of only raising the last error. (Contributed by Irit Katriel in bpo-29980.)

sqlite3

string

sys

  • sys.exc_info() now derives the type and traceback fields from the value (the exception instance), so when an exception is modified while it is being handled, the changes are reflected in the results of subsequent calls to exc_info(). (Contributed by Irit Katriel in bpo-45711.)

  • Add sys.exception() which returns the active exception instance (equivalent to sys.exc_info()[1]). (Contributed by Irit Katriel in bpo-46328.)

  • Add the sys.flags.safe_path flag. (Contributed by Victor Stinner in gh-57684.)

sysconfig

  • Three new installation schemes (posix_venv, nt_venv and venv) were added and are used when Python creates new virtual environments or when it is running from a virtual environment. The first two schemes (posix_venv and nt_venv) are OS-specific for non-Windows and Windows, the venv is essentially an alias to one of them according to the OS Python runs on. This is useful for downstream distributors who modify sysconfig.get_preferred_scheme(). Third party code that creates new virtual environments should use the new venv installation scheme to determine the paths, as does venv. (Contributed by Miro Hrončok in bpo-45413.)

tempfile

threading

time

  • On Unix, time.sleep() now uses the clock_nanosleep() or nanosleep() function, if available, which has a resolution of 1 nanosecond (10-9 seconds), rather than using select() which has a resolution of 1 microsecond (10-6 seconds). (Contributed by Benjamin Szőke and Victor Stinner in bpo-21302.)

  • On Windows 8.1 and newer, time.sleep() now uses a waitable timer based on high-resolution timers which has a resolution of 100 nanoseconds (10-7 seconds). Previously, it had a resolution of 1 millisecond (10-3 seconds). (Contributed by Benjamin Szőke, Dong-hee Na, Eryk Sun and Victor Stinner in bpo-21302 and bpo-45429.)

tkinter

  • Added method info_patchlevel() which returns the exact version of the Tcl library as a named tuple similar to sys.version_info. (Contributed by Serhiy Storchaka in gh-91827.)

traceback

typing

For major changes, see New Features Related to Type Hints.

  • Add typing.assert_never() and typing.Never. typing.assert_never() is useful for asking a type checker to confirm that a line of code is not reachable. At runtime, it raises an AssertionError. (Contributed by Jelle Zijlstra in gh-90633.)

  • Add typing.reveal_type(). This is useful for asking a type checker what type it has inferred for a given expression. At runtime it prints the type of the received value. (Contributed by Jelle Zijlstra in gh-90572.)

  • Add typing.assert_type(). This is useful for asking a type checker to confirm that the type it has inferred for a given expression matches the given type. At runtime it simply returns the received value. (Contributed by Jelle Zijlstra in gh-90638.)

  • typing.TypedDict types can now be generic. (Contributed by Samodya Abeysiriwardane in gh-89026.)

  • NamedTuple types can now be generic. (Contributed by Serhiy Storchaka in bpo-43923.)

  • Allow subclassing of typing.Any. This is useful for avoiding type checker errors related to highly dynamic class, such as mocks. (Contributed by Shantanu Jain in gh-91154.)

  • The typing.final() decorator now sets the __final__ attributed on the decorated object. (Contributed by Jelle Zijlstra in gh-90500.)

  • The typing.get_overloads() function can be used for introspecting the overloads of a function. typing.clear_overloads() can be used to clear all registered overloads of a function. (Contributed by Jelle Zijlstra in gh-89263.)

  • The __init__() method of Protocol subclasses is now preserved. (Contributed by Adrian Garcia Badarasco in gh-88970.)

  • The representation of empty tuple types (Tuple[()]) is simplified. This affects introspection, e.g. get_args(Tuple[()]) now evaluates to () instead of ((),). (Contributed by Serhiy Storchaka in gh-91137.)

  • Loosen runtime requirements for type annotations by removing the callable check in the private typing._type_check function. (Contributed by Gregory Beauregard in gh-90802.)

  • typing.get_type_hints() now supports evaluating strings as forward references in PEP 585 generic aliases. (Contributed by Niklas Rosenstein in gh-85542.)

  • typing.get_type_hints() no longer adds Optional to parameters with None as a default. (Contributed by Nikita Sobolev in gh-90353.)

  • typing.get_type_hints() now supports evaluating bare stringified ClassVar annotations. (Contributed by Gregory Beauregard in gh-90711.)

  • typing.no_type_check() no longer modifies external classes and functions. It also now correctly marks classmethods as not to be type checked. (Contributed by Nikita Sobolev in gh-90729.)

unicodedata

  • The Unicode database has been updated to version 14.0.0. (Contributed by Benjamin Peterson in bpo-45190).

unittest

venv

  • When new Python virtual environments are created, the venv sysconfig installation scheme is used to determine the paths inside the environment. When Python runs in a virtual environment, the same installation scheme is the default. That means that downstream distributors can change the default sysconfig install scheme without changing behavior of virtual environments. Third party code that also creates new virtual environments should do the same. (Contributed by Miro Hrončok in bpo-45413.)

warnings

zipfile

  • Added support for specifying member name encoding for reading metadata in a ZipFile‘s directory and file headers. (Contributed by Stephen J. Turnbull and Serhiy Storchaka in bpo-28080.)

  • Added ZipFile.mkdir() for creating new directories inside ZIP archives. (Contributed by Sam Ezeh in gh-49083.)

  • Added stem, suffix and suffixes to zipfile.Path. (Contributed by Miguel Brito in gh-88261.)

Optimizations

This section covers specific optimizations independent of the Faster CPython project, which is covered in its own section.

  • The compiler now optimizes simple printf-style % formatting on string literals containing only the format codes %s, %r and %a and makes it as fast as a corresponding f-string expression. (Contributed by Serhiy Storchaka in bpo-28307.)

  • Integer division (//) is better tuned for optimization by compilers. It is now around 20% faster on x86-64 when dividing an int by a value smaller than 2**30. (Contributed by Gregory P. Smith and Tim Peters in gh-90564.)

  • sum() is now nearly 30% faster for integers smaller than 2**30. (Contributed by Stefan Behnel in gh-68264.)

  • Resizing lists is streamlined for the common case, speeding up list.append() by ≈15% and simple list comprehensions by up to 20-30% (Contributed by Dennis Sweeney in gh-91165.)

  • Dictionaries don’t store hash values when all keys are Unicode objects, decreasing dict size. For example, sys.getsizeof(dict.fromkeys("abcdefg")) is reduced from 352 bytes to 272 bytes (23% smaller) on 64-bit platforms. (Contributed by Inada Naoki in bpo-46845.)

  • Using asyncio.DatagramProtocol is now orders of magnitude faster when transferring large files over UDP, with speeds over 100 times higher for a ≈60 MiB file. (Contributed by msoxzw in gh-91487.)

  • math functions comb() and perm() are now ≈10 times faster for large arguments (with a larger speedup for larger k). (Contributed by Serhiy Storchaka in bpo-37295.)

  • The statistics functions mean(), variance() and stdev() now consume iterators in one pass rather than converting them to a list first. This is twice as fast and can save substantial memory. (Contributed by Raymond Hettinger in gh-90415.)

  • unicodedata.normalize() now normalizes pure-ASCII strings in constant time. (Contributed by Dong-hee Na in bpo-44987.)

Faster CPython

CPython 3.11 is on average 25% faster than CPython 3.10 when measured with the pyperformance benchmark suite, and compiled with GCC on Ubuntu Linux. Depending on your workload, the speedup could be up to 10-60% faster.

This project focuses on two major areas in Python: faster startup and faster runtime. Other optimizations not under this project are listed in Optimizations.

Faster Startup

Frozen imports / Static code objects

Python caches bytecode in the __pycache__ directory to speed up module loading.

Previously in 3.10, Python module execution looked like this:

  1. Read __pycache__ -> Unmarshal -> Heap allocated code object -> Evaluate

In Python 3.11, the core modules essential for Python startup are “frozen”. This means that their code objects (and bytecode) are statically allocated by the interpreter. This reduces the steps in module execution process to this:

  1. Statically allocated code object -> Evaluate

Interpreter startup is now 10-15% faster in Python 3.11. This has a big impact for short-running programs using Python.

(Contributed by Eric Snow, Guido van Rossum and Kumar Aditya in numerous issues.)

Faster Runtime

Cheaper, lazy Python frames

Python frames are created whenever Python calls a Python function. This frame holds execution information. The following are new frame optimizations:

  • Streamlined the frame creation process.

  • Avoided memory allocation by generously re-using frame space on the C stack.

  • Streamlined the internal frame struct to contain only essential information. Frames previously held extra debugging and memory management information.

Old-style frame objects are now created only when requested by debuggers or by Python introspection functions such as sys._getframe or inspect.currentframe. For most user code, no frame objects are created at all. As a result, nearly all Python functions calls have sped up significantly. We measured a 3-7% speedup in pyperformance.

(Contributed by Mark Shannon in bpo-44590.)

Inlined Python function calls

During a Python function call, Python will call an evaluating C function to interpret that function’s code. This effectively limits pure Python recursion to what’s safe for the C stack.

In 3.11, when CPython detects Python code calling another Python function, it sets up a new frame, and “jumps” to the new code inside the new frame. This avoids calling the C interpreting function altogether.

Most Python function calls now consume no C stack space. This speeds up most of such calls. In simple recursive functions like fibonacci or factorial, a 1.7x speedup was observed. This also means recursive functions can recurse significantly deeper (if the user increases the recursion limit). We measured a 1-3% improvement in pyperformance.

(Contributed by Pablo Galindo and Mark Shannon in bpo-45256.)

PEP 659: Specializing Adaptive Interpreter

PEP 659 is one of the key parts of the faster CPython project. The general idea is that while Python is a dynamic language, most code has regions where objects and types rarely change. This concept is known as type stability.

At runtime, Python will try to look for common patterns and type stability in the executing code. Python will then replace the current operation with a more specialized one. This specialized operation uses fast paths available only to those use cases/types, which generally outperform their generic counterparts. This also brings in another concept called inline caching, where Python caches the results of expensive operations directly in the bytecode.

The specializer will also combine certain common instruction pairs into one superinstruction. This reduces the overhead during execution.

Python will only specialize when it sees code that is “hot” (executed multiple times). This prevents Python from wasting time for run-once code. Python can also de-specialize when code is too dynamic or when the use changes. Specialization is attempted periodically, and specialization attempts are not too expensive. This allows specialization to adapt to new circumstances.

(PEP written by Mark Shannon, with ideas inspired by Stefan Brunthaler. See PEP 659 for more information. Implementation by Mark Shannon and Brandt Bucher, with additional help from Irit Katriel and Dennis Sweeney.)

Operation

Form

Specialization

Operation speedup (up to)

Contributor(s)

Binary operations

x+x; xx; x-x;

Binary add, multiply and subtract for common types such as int, float, and str take custom fast paths for their underlying types.

10%

Mark Shannon, Dong-hee Na, Brandt Bucher, Dennis Sweeney

Subscript

a[i]

Subscripting container types such as list, tuple and dict directly index the underlying data structures.

Subscripting custom getitem is also inlined similar to Inlined Python function calls.

10-25%

Irit Katriel, Mark Shannon

Store subscript

a[i] = z

Similar to subscripting specialization above.

10-25%

Dennis Sweeney

Calls

f(arg) C(arg)

Calls to common builtin (C) functions and types such as len and str directly call their underlying C version. This avoids going through the internal calling convention.

20%

Mark Shannon, Ken Jin

Load global variable

print len

The object’s index in the globals/builtins namespace is cached. Loading globals and builtins require zero namespace lookups.

1

Mark Shannon

Load attribute

o.attr

Similar to loading global variables. The attribute’s index inside the class/object’s namespace is cached. In most cases, attribute loading will require zero namespace lookups.

2

Mark Shannon

Load methods for call

o.meth()

The actual address of the method is cached. Method loading now has no namespace lookups — even for classes with long inheritance chains.

10-20%

Ken Jin, Mark Shannon

Store attribute

o.attr = z

Similar to load attribute optimization.

2% in pyperformance

Mark Shannon

Unpack Sequence

seq

Specialized for common containers such as list and tuple. Avoids internal calling convention.

8%

Brandt Bucher

1

A similar optimization already existed since Python 3.8. 3.11 specializes for more forms and reduces some overhead.

2

A similar optimization already existed since Python 3.10. 3.11 specializes for more forms. Furthermore, all attribute loads should be sped up by bpo-45947.

Misc

  • Objects now require less memory due to lazily created object namespaces. Their namespace dictionaries now also share keys more freely. (Contributed Mark Shannon in bpo-45340 and bpo-40116.)

  • A more concise representation of exceptions in the interpreter reduced the time required for catching an exception by about 10%. (Contributed by Irit Katriel in bpo-45711.)

FAQ

Q: How should I write my code to utilize these speedups?

A: You don’t have to change your code. Write Pythonic code that follows common best practices. The Faster CPython project optimizes for common code patterns we observe.

Q: Will CPython 3.11 use more memory?

A: Maybe not. We don’t expect memory use to exceed 20% more than 3.10. This is offset by memory optimizations for frame objects and object dictionaries as mentioned above.

Q: I don’t see any speedups in my workload. Why?

A: Certain code won’t have noticeable benefits. If your code spends most of its time on I/O operations, or already does most of its computation in a C extension library like numpy, there won’t be significant speedup. This project currently benefits pure-Python workloads the most.

Furthermore, the pyperformance figures are a geometric mean. Even within the pyperformance benchmarks, certain benchmarks have slowed down slightly, while others have sped up by nearly 2x!

Q: Is there a JIT compiler?

A: No. We’re still exploring other optimizations.

About

Faster CPython explores optimizations for CPython. The main team is funded by Microsoft to work on this full-time. Pablo Galindo Salgado is also funded by Bloomberg LP to work on the project part-time. Finally, many contributors are volunteers from the community.

CPython bytecode changes

The bytecode now contains inline cache entries, which take the form of the newly-added CACHE instructions. Many opcodes expect to be followed by an exact number of caches, and instruct the interpreter to skip over them at runtime. Populated caches can look like arbitrary instructions, so great care should be taken when reading or modifying raw, adaptive bytecode containing quickened data.

New opcodes

Replaced opcodes

Replaced Opcode(s)

New Opcode(s)

Notes

BINARY*
INPLACE

BINARYOP

Replaced all numeric binary/in-place opcodes with a single opcode

CALL_FUNCTION
CALL_FUNCTION_KW
CALL_METHOD

Decouples argument shifting for methods from handling of keyword arguments; allows better specialization of calls

DUP_TOP
DUP_TOP_TWO
ROT_TWO
ROT_THREE
ROT_FOUR
ROT_N

Stack manipulation instructions

JUMP_IF_NOT_EXC_MATCH

Now performs check but doesn’t jump

JUMP_ABSOLUTE
POP_JUMP_IF_FALSE
POP_JUMP_IF_TRUE

See 3; TRUE, FALSE, NONE and NOT_NONE variants for each direction

SETUP_WITH
SETUP_ASYNC_WITH

BEFORE_WITH

with block setup

3

All jump opcodes are now relative, including the existing JUMP_IF_TRUE_OR_POP and JUMP_IF_FALSE_OR_POP. The argument is now an offset from the current instruction rather than an absolute location.

Changed/removed opcodes

  • Changed MATCH_CLASS and MATCH_KEYS to no longer push an additional boolean value to indicate success/failure. Instead, None is pushed on failure in place of the tuple of extracted values.

  • Changed opcodes that work with exceptions to reflect them now being represented as one item on the stack instead of three (see gh-89874).

  • Removed COPY_DICT_WITHOUT_KEYS, GEN_START, POP_BLOCK, SETUP_FINALLY and YIELD_FROM.

Deprecated

This section lists Python APIs that have been deprecated in Python 3.11.

Deprecated C APIs are listed separately.

Language/Builtins

  • Chaining classmethod descriptors (introduced in bpo-19072) is now deprecated. It can no longer be used to wrap other descriptors such as property. The core design of this feature was flawed and caused a number of downstream problems. To “pass-through” a classmethod, consider using the __wrapped__ attribute that was added in Python 3.10. (Contributed by Raymond Hettinger in gh-89519.)

  • Octal escapes in string and bytes literals with values larger than 0o377 (255 in decimal) now produce a DeprecationWarning. In a future Python version, they will raise a SyntaxWarning and eventually a SyntaxError. (Contributed by Serhiy Storchaka in gh-81548.)

  • The delegation of int() to __trunc__() is now deprecated. Calling int(a) when type(a) implements __trunc__() but not __int__() or __index__() now raises a DeprecationWarning. (Contributed by Zackery Spytz in bpo-44977.)

Modules

Standard Library

Pending Removal in Python 3.12

The following Python APIs have been deprecated in earlier Python releases, and will be removed in Python 3.12.

C APIs pending removal are listed separately.

Removed

This section lists Python APIs that have been removed in Python 3.11.

Removed C APIs are listed separately.

  • Removed the @asyncio.coroutine() decorator enabling legacy generator-based coroutines to be compatible with async / await code. The function has been deprecated since Python 3.8 and the removal was initially scheduled for Python 3.10. Use async def instead. (Contributed by Illia Volochii in bpo-43216.)

  • Removed asyncio.coroutines.CoroWrapper used for wrapping legacy generator-based coroutine objects in the debug mode. (Contributed by Illia Volochii in bpo-43216.)

  • Due to significant security concerns, the reuse_address parameter of asyncio.loop.create_datagram_endpoint(), disabled in Python 3.9, is now entirely removed. This is because of the behavior of the socket option SO_REUSEADDR in UDP. (Contributed by Hugo van Kemenade in bpo-45129.)

  • Removed the binhex module, deprecated in Python 3.9. Also removed the related, similarly-deprecated binascii functions:

    • binascii.a2b_hqx()

    • binascii.b2a_hqx()

    • binascii.rlecode_hqx()

    • binascii.rldecode_hqx()

    The binascii.crc_hqx() function remains available.

    (Contributed by Victor Stinner in bpo-45085.)

  • Removed the distutils bdist_msi command deprecated in Python 3.9. Use bdist_wheel (wheel packages) instead. (Contributed by Hugo van Kemenade in bpo-45124.)

  • Removed the __getitem__() methods of xml.dom.pulldom.DOMEventStream, wsgiref.util.FileWrapper and fileinput.FileInput, deprecated since Python 3.9. (Contributed by Hugo van Kemenade in bpo-45132.)

  • Removed the deprecated gettext functions lgettext(), ldgettext(), lngettext() and ldngettext(). Also removed the bind_textdomain_codeset() function, the NullTranslations.output_charset() and NullTranslations.set_output_charset() methods, and the codeset parameter of translation() and install(), since they are only used for the l*gettext() functions. (Contributed by Dong-hee Na and Serhiy Storchaka in bpo-44235.)

  • Removed from the inspect module:

    (Contributed by Hugo van Kemenade in bpo-45320.)

  • Removed the __class_getitem__() method from pathlib.PurePath, because it was not used and added by mistake in previous versions. (Contributed by Nikita Sobolev in bpo-46483.)

  • Removed the MailmanProxy class in the smtpd module, as it is unusable without the external mailman package. (Contributed by Dong-hee Na in bpo-35800.)

  • Removed the deprecated split() method of _tkinter.TkappType. (Contributed by Erlend E. Aasland in bpo-38371.)

  • Removed namespace package support from unittest discovery. It was introduced in Python 3.4 but has been broken since Python 3.7. (Contributed by Inada Naoki in bpo-23882.)

  • Removed the undocumented private float.__set_format__() method, previously known as float.__setformat__() in Python 3.7. Its docstring said: “You probably don’t want to use this function. It exists mainly to be used in Python’s test suite.” (Contributed by Victor Stinner in bpo-46852.)

  • The --experimental-isolated-subinterpreters configure flag (and corresponding EXPERIMENTAL_ISOLATED_SUBINTERPRETERS macro) have been removed.

  • Pynche —- The Pythonically Natural Color and Hue Editor —- has been moved out of Tools/scripts and is being developed independently from the Python source tree.

Porting to Python 3.11

This section lists previously described changes and other bugfixes in the Python API that may require changes to your Python code.

Porting notes for the C API are listed separately.

  • open(), io.open(), codecs.open() and fileinput.FileInput no longer accept 'U' (“universal newline”) in the file mode. In Python 3, “universal newline” mode is used by default whenever a file is opened in text mode, and the 'U' flag has been deprecated since Python 3.3. The newline parameter to these functions controls how universal newlines work. (Contributed by Victor Stinner in bpo-37330.)

  • ast.AST node positions are now validated when provided to compile() and other related functions. If invalid positions are detected, a ValueError will be raised. (Contributed by Pablo Galindo in gh-93351)

  • Prohibited passing non-concurrent.futures.ThreadPoolExecutor executors to asyncio.loop.set_default_executor() following a deprecation in Python 3.8. (Contributed by Illia Volochii in bpo-43234.)

  • calendar: The calendar.LocaleTextCalendar and calendar.LocaleHTMLCalendar classes now use locale.getlocale(), instead of using locale.getdefaultlocale(), if no locale is specified. (Contributed by Victor Stinner in bpo-46659.)

  • The pdb module now reads the .pdbrc configuration file with the 'UTF-8' encoding. (Contributed by Srinivas Reddy Thatiparthy (శ్రీనివాస్ రెడ్డి తాటిపర్తి) in bpo-41137.)

  • The population parameter of random.sample() must be a sequence, and automatic conversion of sets to lists is no longer supported. Also, if the sample size is larger than the population size, a ValueError is raised. (Contributed by Raymond Hettinger in bpo-40465.)

  • The random optional parameter of random.shuffle() was removed. It was previously an arbitrary random function to use for the shuffle; now, random.random() (its previous default) will always be used.

  • In re 正则表达式语法, global inline flags (e.g. (?i)) can now only be used at the start of regular expressions. Using them elsewhere has been deprecated since Python 3.6. (Contributed by Serhiy Storchaka in bpo-47066.)

  • In the re module, several long-standing bugs where fixed that, in rare cases, could cause capture groups to get the wrong result. Therefore, this could change the captured output in these cases. (Contributed by Ma Lin in bpo-35859.)

Build Changes

  • CPython now has PEP 11 Tier 3 support for cross compiling to the WebAssembly platforms Emscripten (wasm32-unknown-emscripten, i.e. Python in the browser) and WebAssembly System Interface (WASI) (wasm32-unknown-wasi). The effort is inspired by previous work like Pyodide. These platforms provide a limited subset of POSIX APIs; Python standard libraries features and modules related to networking, processes, threading, signals, mmap, and users/groups are not available or don’t work. (Emscripten contributed by Christian Heimes and Ethan Smith in gh-84461 and WASI contributed by Christian Heimes in gh-90473; platforms promoted in gh-95085)

  • Building Python now requires:

    • A C11 compiler. Optional C11 features#Optional_features) are not required. (Contributed by Victor Stinner in bpo-46656.)

    • Support for IEEE 754 floating point numbers. (Contributed by Victor Stinner in bpo-46917.)

    • Support for floating point Not-a-Number (NaN), as the Py_NO_NAN macro has been removed. (Contributed by Victor Stinner in bpo-46656.)

    • A C99 <math.h> header file providing the copysign(), hypot(), isfinite(), isinf(), isnan(), and round() functions (contributed by Victor Stinner in bpo-45440); and a NAN constant or the __builtin_nan() function (Contributed by Victor Stinner in bpo-46640).

  • The tkinter package now requires Tcl/Tk version 8.5.12 or newer. (Contributed by Serhiy Storchaka in bpo-46996.)

  • Build dependencies, compiler flags, and linker flags for most stdlib extension modules are now detected by configure. libffi, libnsl, libsqlite3, zlib, bzip2, liblzma, libcrypt, Tcl/Tk, and uuid flags are detected by pkg-config (when available). tkinter now requires a pkg-config command to detect development settings for Tcl/Tk headers and libraries. (Contributed by Christian Heimes and Erlend Egeberg Aasland in bpo-45847, bpo-45747, and bpo-45763.)

  • libpython is no longer linked against libcrypt. (Contributed by Mike Gilbert in bpo-45433.)

  • CPython can now be built with the ThinLTO option via passing thin to --with-lto, i.e. --with-lto=thin. (Contributed by Dong-hee Na and Brett Holman in bpo-44340.)

  • Freelists for object structs can now be disabled. A new configure option --without-freelists can be used to disable all freelists except empty tuple singleton. (Contributed by Christian Heimes in bpo-45522.)

  • Modules/Setup and Modules/makesetup have been improved and tied up. Extension modules can now be built through makesetup. All except some test modules can be linked statically into a main binary or library. (Contributed by Brett Cannon and Christian Heimes in bpo-45548, bpo-45570, bpo-45571, and bpo-43974.)

    备注

    Use the environment variables TCLTK_CFLAGS and TCLTK_LIBS to manually specify the location of Tcl/Tk headers and libraries. The configure options --with-tcltk-includes and --with-tcltk-libs have been removed.

    On RHEL 7 and CentOS 7 the development packages do not provide tcl.pc and tk.pc; use TCLTK_LIBS="-ltk8.5 -ltkstub8.5 -ltcl8.5". The directory Misc/rhel7 contains .pc files and instructions on how to build Python with RHEL 7’s and CentOS 7’s Tcl/Tk and OpenSSL.

  • CPython will now use 30-bit digits by default for the Python int implementation. Previously, the default was to use 30-bit digits on platforms with SIZEOF_VOID_P >= 8, and 15-bit digits otherwise. It’s still possible to explicitly request use of 15-bit digits via either the --enable-big-digits option to the configure script or (for Windows) the PYLONG_BITS_IN_DIGIT variable in PC/pyconfig.h, but this option may be removed at some point in the future. (Contributed by Mark Dickinson in bpo-45569.)

C API Changes

New Features

Porting to Python 3.11

  • Some macros have been converted to static inline functions to avoid macro pitfalls. The change should be mostly transparent to users, as the replacement functions will cast their arguments to the expected types to avoid compiler warnings due to static type checks. However, when the limited C API is set to >=3.11, these casts are not done, and callers will need to cast arguments to their expected types. See PEP 670 for more details. (Contributed by Victor Stinner and Erlend E. Aasland in gh-89653.)

  • PyErr_SetExcInfo() no longer uses the type and traceback arguments, the interpreter now derives those values from the exception instance (the value argument). The function still steals references of all three arguments. (Contributed by Irit Katriel in bpo-45711.)

  • PyErr_GetExcInfo() now derives the type and traceback fields of the result from the exception instance (the value field). (Contributed by Irit Katriel in bpo-45711.)

  • _frozen has a new is_package field to indicate whether or not the frozen module is a package. Previously, a negative value in the size field was the indicator. Now only non-negative values be used for size. (Contributed by Kumar Aditya in bpo-46608.)

  • _PyFrameEvalFunction() now takes _PyInterpreterFrame* as its second parameter, instead of PyFrameObject*. See PEP 523 for more details of how to use this function pointer type.

  • PyCode_New() and PyCode_NewWithPosOnlyArgs() now take an additional exception_table argument. Using these functions should be avoided, if at all possible. To get a custom code object: create a code object using the compiler, then get a modified version with the replace method.

  • PyCodeObject no longer has the co_code, co_varnames, co_cellvars and co_freevars fields. Instead, use PyCode_GetCode(), PyCode_GetVarnames(), PyCode_GetCellvars() and PyCode_GetFreevars() respectively to access them via the C API. (Contributed by Brandt Bucher in bpo-46841 and Ken Jin in gh-92154 and gh-94936.)

  • The old trashcan macros (Py_TRASHCAN_SAFE_BEGIN/Py_TRASHCAN_SAFE_END) are now deprecated. They should be replaced by the new macros Py_TRASHCAN_BEGIN and Py_TRASHCAN_END.

    A tp_dealloc function that has the old macros, such as:

    1. static void
    2. mytype_dealloc(mytype *p)
    3. {
    4.     PyObject_GC_UnTrack(p);
    5.     Py_TRASHCAN_SAFE_BEGIN(p);
    6.     ...
    7.     Py_TRASHCAN_SAFE_END
    8. }

    should migrate to the new macros as follows:

    1. static void
    2. mytype_dealloc(mytype *p)
    3. {
    4.     PyObject_GC_UnTrack(p);
    5.     Py_TRASHCAN_BEGIN(p, mytype_dealloc)
    6.     ...
    7.     Py_TRASHCAN_END
    8. }

    Note that Py_TRASHCAN_BEGIN has a second argument which should be the deallocation function it is in.

    To support older Python versions in the same codebase, you can define the following macros and use them throughout the code (credit: these were copied from the mypy codebase):

    1. #if PY_MAJOR_VERSION >= 3 && PY_MINOR_VERSION >= 8
    2. #  define CPy_TRASHCAN_BEGIN(op, dealloc) Py_TRASHCAN_BEGIN(op, dealloc)
    3. #  define CPy_TRASHCAN_END(op) Py_TRASHCAN_END
    4. #else
    5. #  define CPy_TRASHCAN_BEGIN(op, dealloc) Py_TRASHCAN_SAFE_BEGIN(op)
    6. #  define CPy_TRASHCAN_END(op) Py_TRASHCAN_SAFE_END(op)
    7. #endif

  • The PyType_Ready() function now raises an error if a type is defined with the Py_TPFLAGS_HAVE_GC flag set but has no traverse function (PyTypeObject.tp_traverse). (Contributed by Victor Stinner in bpo-44263.)

  • Heap types with the Py_TPFLAGS_IMMUTABLETYPE flag can now inherit the PEP 590 vectorcall protocol. Previously, this was only possible for static types. (Contributed by Erlend E. Aasland in bpo-43908)

  • Since Py_TYPE() is changed to a inline static function, Py_TYPE(obj) = new_type must be replaced with Py_SET_TYPE(obj, new_type): see the Py_SET_TYPE() function (available since Python 3.9). For backward compatibility, this macro can be used:

    1. #if PY_VERSION_HEX < 0x030900A4 && !defined(Py_SET_TYPE)
    2. static inline void _Py_SET_TYPE(PyObject *ob, PyTypeObject *type)
    3. { ob->ob_type = type; }
    4. #define Py_SET_TYPE(ob, type) _Py_SET_TYPE((PyObject*)(ob), type)
    5. #endif

    (Contributed by Victor Stinner in bpo-39573.)

  • Since Py_SIZE() is changed to a inline static function, Py_SIZE(obj) = new_size must be replaced with Py_SET_SIZE(obj, new_size): see the Py_SET_SIZE() function (available since Python 3.9). For backward compatibility, this macro can be used:

    1. #if PY_VERSION_HEX < 0x030900A4 && !defined(Py_SET_SIZE)
    2. static inline void _Py_SET_SIZE(PyVarObject *ob, Py_ssize_t size)
    3. { ob->ob_size = size; }
    4. #define Py_SET_SIZE(ob, size) _Py_SET_SIZE((PyVarObject*)(ob), size)
    5. #endif

    (Contributed by Victor Stinner in bpo-39573.)

  • <Python.h> no longer includes the header files <stdlib.h>, <stdio.h>, <errno.h> and <string.h> when the Py_LIMITED_API macro is set to 0x030b0000 (Python 3.11) or higher. C extensions should explicitly include the header files after #include <Python.h>. (Contributed by Victor Stinner in bpo-45434.)

  • The non-limited API files cellobject.h, classobject.h, code.h, context.h, funcobject.h, genobject.h and longintrepr.h have been moved to the Include/cpython directory. Moreover, the eval.h header file was removed. These files must not be included directly, as they are already included in Python.h: Include Files. If they have been included directly, consider including Python.h instead. (Contributed by Victor Stinner in bpo-35134.)

  • The PyUnicode_CHECK_INTERNED() macro has been excluded from the limited C API. It was never usable there, because it used internal structures which are not available in the limited C API. (Contributed by Victor Stinner in bpo-46007.)

  • The following frame functions and type are now directly available with #include <Python.h>, it’s no longer needed to add #include <frameobject.h>:

    (Contributed by Victor Stinner in gh-93937.)

  • The PyFrameObject structure members have been removed from the public C API.

    While the documentation notes that the PyFrameObject fields are subject to change at any time, they have been stable for a long time and were used in several popular extensions.

    In Python 3.11, the frame struct was reorganized to allow performance optimizations. Some fields were removed entirely, as they were details of the old implementation.

    PyFrameObject fields:

    The Python frame object is now created lazily. A side effect is that the f_back member must not be accessed directly, since its value is now also computed lazily. The PyFrame_GetBack() function must be called instead.

    Debuggers that accessed the f_locals directly must call PyFrame_GetLocals() instead. They no longer need to call PyFrame_FastToLocalsWithError() or PyFrame_LocalsToFast(), in fact they should not call those functions. The necessary updating of the frame is now managed by the virtual machine.

    Code defining PyFrame_GetCode() on Python 3.8 and older:

    1. #if PY_VERSION_HEX < 0x030900B1
    2. static inline PyCodeObject* PyFrame_GetCode(PyFrameObject *frame)
    3. {
    4.     Py_INCREF(frame->f_code);
    5.     return frame->f_code;
    6. }
    7. #endif

    Code defining PyFrame_GetBack() on Python 3.8 and older:

    1. #if PY_VERSION_HEX < 0x030900B1
    2. static inline PyFrameObject* PyFrame_GetBack(PyFrameObject *frame)
    3. {
    4.     Py_XINCREF(frame->f_back);
    5.     return frame->f_back;
    6. }
    7. #endif

    Or use the pythoncapi_compat project to get these two functions on older Python versions.

  • Changes of the PyThreadState structure members:

    Code defining PyThreadState_GetFrame() on Python 3.8 and older:

    1. #if PY_VERSION_HEX < 0x030900B1
    2. static inline PyFrameObject* PyThreadState_GetFrame(PyThreadState *tstate)
    3. {
    4.     Py_XINCREF(tstate->frame);
    5.     return tstate->frame;
    6. }
    7. #endif

    Code defining PyThreadState_EnterTracing() and PyThreadState_LeaveTracing() on Python 3.10 and older:

    1. #if PY_VERSION_HEX < 0x030B00A2
    2. static inline void PyThreadState_EnterTracing(PyThreadState *tstate)
    3. {
    4.     tstate->tracing++;
    5. #if PY_VERSION_HEX >= 0x030A00A1
    6.     tstate->cframe->use_tracing = 0;
    7. #else
    8.     tstate->use_tracing = 0;
    9. #endif
    10. }
    12. static inline void PyThreadState_LeaveTracing(PyThreadState *tstate)
    13. {
    14.     int use_tracing = (tstate->c_tracefunc != NULL || tstate->c_profilefunc != NULL);
    15.     tstate->tracing--;
    16. #if PY_VERSION_HEX >= 0x030A00A1
    17.     tstate->cframe->use_tracing = use_tracing;
    18. #else
    19.     tstate->use_tracing = use_tracing;
    20. #endif
    21. }
    22. #endif

    Or use the pythoncapi_compat project to get these functions on old Python functions.

  • Distributors are encouraged to build Python with the optimized Blake2 library libb2.

  • The PyConfig.module_search_paths_set field must now be set to 1 for initialization to use PyConfig.module_search_paths to initialize sys.path. Otherwise, initialization will recalculate the path and replace any values added to module_search_paths.

  • PyConfig_Read() no longer calculates the initial search path, and will not fill any values into PyConfig.module_search_paths. To calculate default paths and then modify them, finish initialization and use PySys_GetObject() to retrieve sys.path as a Python list object and modify it directly.

Deprecated

Pending Removal in Python 3.12

The following C APIs have been deprecated in earlier Python releases, and will be removed in Python 3.12.

Removed

  • PyFrame_BlockSetup() and PyFrame_BlockPop() have been removed. (Contributed by Mark Shannon in bpo-40222.)

  • Remove the following math macros using the errno variable:

    • Py_ADJUST_ERANGE1()

    • Py_ADJUST_ERANGE2()

    • Py_OVERFLOWED()

    • Py_SET_ERANGE_IF_OVERFLOW()

    • Py_SET_ERRNO_ON_MATH_ERROR()

    (Contributed by Victor Stinner in bpo-45412.)

  • Remove Py_UNICODE_COPY() and Py_UNICODE_FILL() macros, deprecated since Python 3.3. Use PyUnicode_CopyCharacters() or memcpy() (wchar_t* string), and PyUnicode_Fill() functions instead. (Contributed by Victor Stinner in bpo-41123.)

  • Remove the pystrhex.h header file. It only contains private functions. C extensions should only include the main <Python.h> header file. (Contributed by Victor Stinner in bpo-45434.)

  • Remove the Py_FORCE_DOUBLE() macro. It was used by the Py_IS_INFINITY() macro. (Contributed by Victor Stinner in bpo-45440.)

  • The following items are no longer available when Py_LIMITED_API is defined:

    These are not part of the limited API.

    (Contributed by Victor Stinner in bpo-45474.)

  • Exclude PyWeakref_GET_OBJECT() from the limited C API. It never worked since the PyWeakReference structure is opaque in the limited C API. (Contributed by Victor Stinner in bpo-35134.)

  • Remove the PyHeapType_GET_MEMBERS() macro. It was exposed in the public C API by mistake, it must only be used by Python internally. Use the PyTypeObject.tp_members member instead. (Contributed by Victor Stinner in bpo-40170.)

  • Remove the HAVE_PY_SET_53BIT_PRECISION macro (moved to the internal C API). (Contributed by Victor Stinner in bpo-45412.)

  • Remove the Py_UNICODE encoder APIs, as they have been deprecated since Python 3.3, are little used and are inefficient relative to the recommended alternatives.

    The removed functions are:

    • PyUnicode_Encode()

    • PyUnicode_EncodeASCII()

    • PyUnicode_EncodeLatin1()

    • PyUnicode_EncodeUTF7()

    • PyUnicode_EncodeUTF8()

    • PyUnicode_EncodeUTF16()

    • PyUnicode_EncodeUTF32()

    • PyUnicode_EncodeUnicodeEscape()

    • PyUnicode_EncodeRawUnicodeEscape()

    • PyUnicode_EncodeCharmap()

    • PyUnicode_TranslateCharmap()

    • PyUnicode_EncodeDecimal()

    • PyUnicode_TransformDecimalToASCII()

    See PEP 624 for details and migration guidance. (Contributed by Inada Naoki in bpo-44029.)