What’s New In Python 3.8

This article explains the new features in Python 3.8, compared to 3.7.

For full details, see the changelog.

Note

Prerelease users should be aware that this document is currently in draft form. It will be updated substantially as Python 3.8 moves towards release, so it’s worth checking back even after reading earlier versions.

Summary – Release highlights

New Features

Assignment expressions

There is new syntax (the “walrus operator”, :=) to assign values to variables as part of an expression. Example:

if (n := len(a)) > 10:
    print(f"List is too long ({n} elements, expected <= 10)")

See PEP 572 for a full description.

(Contributed by Emily Morehouse in bpo-35224.)

Positional-only parameters

There is new syntax (/) to indicate that some function parameters must be specified positionally (i.e., cannot be used as keyword arguments). This is the same notation as shown by help() for functions implemented in C (produced by Larry Hastings’ “Argument Clinic” tool). Example:

def pow(x, y, z=None, /):
    r = x**y
    if z is not None:
        r %= z
    return r

Now pow(2, 10) and pow(2, 10, 17) are valid calls, but pow(x=2, y=10) and pow(2, 10, z=17) are invalid.

See PEP 570 for a full description.

(Contributed by Pablo Galindo in bpo-36540.)

Parallel filesystem cache for compiled bytecode files

The new PYTHONPYCACHEPREFIX setting (also available as -X pycache_prefix) configures the implicit bytecode cache to use a separate parallel filesystem tree, rather than the default __pycache__ subdirectories within each source directory.

The location of the cache is reported in sys.pycache_prefix (None indicates the default location in __pycache__ subdirectories).

(Contributed by Carl Meyer in bpo-33499.)

Debug build uses the same ABI as release build

Python now uses the same ABI whether it built in release or debug mode. On Unix, when Python is built in debug mode, it is now possible to load C extensions built in release mode and C extensions built using the stable ABI.

Release builds and debug builds are now ABI compatible: defining the Py_DEBUG macro no longer implies the Py_TRACE_REFS macro, which introduces the only ABI incompatibility. The Py_TRACE_REFS macro, which adds the sys.getobjects() function and the PYTHONDUMPREFS environment variable, can be set using the new ./configure --with-trace-refs build option. (Contributed by Victor Stinner in bpo-36465.)

On Unix, C extensions are no longer linked to libpython. It is now possible for a statically linked Python to load a C extension built using a shared library Python. (Contributed by Victor Stinner in bpo-21536.)

On Unix, when Python is built in debug mode, import now also looks for C extensions compiled in release mode and for C extensions compiled with the stable ABI. (Contributed by Victor Stinner in bpo-36722.)

Other Language Changes

  • A continue statement was illegal in the finally clause due to a problem with the implementation. In Python 3.8 this restriction was lifted. (Contributed by Serhiy Storchaka in bpo-32489.)

  • The int type now has a new as_integer_ratio() method compatible with the existing float.as_integer_ratio() method. (Contributed by Lisa Roach in bpo-33073.)

  • Added support of \N{name} escapes in regular expressions. (Contributed by Jonathan Eunice and Serhiy Storchaka in bpo-30688.)

  • Dict and dictviews are now iterable in reversed insertion order using reversed(). (Contributed by Rémi Lapeyre in bpo-33462.)

  • The syntax allowed for keyword names in function calls was further restricted. In particular, f((keyword)=arg) is no longer allowed. It was never intended to permit more than a bare name on the left-hand side of a keyword argument assignment term. See bpo-34641.

  • Iterable unpacking is now allowed without parentheses in yield and return statements. (Contributed by David Cuthbert and Jordan Chapman in bpo-32117.)

  • A backslash-character pair that is not a valid escape sequence generates a DeprecationWarning since Python 3.6. In Python 3.8 it generates a SyntaxWarning instead. (Contributed by Serhiy Storchaka in bpo-32912.)

  • The compiler now produces a SyntaxWarning in some cases when a comma is missed before tuple or list. For example:

    data = [
    (1, 2, 3) # oops, missing comma!
    (4, 5, 6)
]

    (Contributed by Serhiy Storchaka in bpo-15248.)

  • Arithmetic operations between subclasses of datetime.date or datetime.datetime and datetime.timedelta objects now return an instance of the subclass, rather than the base class. This also affects the return type of operations whose implementation (directly or indirectly) uses datetime.timedelta arithmetic, such as datetime.datetime.astimezone(). (Contributed by Paul Ganssle in bpo-32417.)

  • When the Python interpreter is interrupted by Ctrl-C (SIGINT) and the resulting KeyboardInterrupt exception is not caught, the Python process now exits via a SIGINT signal or with the correct exit code such that the calling process can detect that it died due to a Ctrl-C. Shells on POSIX and Windows use this to properly terminate scripts in interactive sessions. (Contributed by Google via Gregory P. Smith in bpo-1054041.)

New Modules

  • None yet.

Improved Modules

asyncio

On Windows, the default event loop is now ProactorEventLoop.

collections

The _asdict() method for collections.namedtuple() now returns a dict instead of a collections.OrderedDict. This works because regular dicts have guaranteed ordering since Python 3.7. If the extra features of OrderedDict are required, the suggested remediation is to cast the result to the desired type: OrderedDict(nt._asdict()). (Contributed by Raymond Hettinger in bpo-35864.)

ctypes

On Windows, CDLL and subclasses now accept a winmode parameter to specify flags for the underlying LoadLibraryEx call. The default flags are set to only load DLL dependencies from trusted locations, including the path where the DLL is stored (if a full or partial path is used to load the initial DLL) and paths added by add_dll_directory().

datetime

Added new alternate constructors datetime.date.fromisocalendar() and datetime.datetime.fromisocalendar(), which construct date and datetime objects respectively from ISO year, week number and weekday; these are the inverse of each class’s isocalendar method. (Contributed by Paul Ganssle in bpo-36004.)

gettext

Added pgettext() and its variants. (Contributed by Franz Glasner, Éric Araujo, and Cheryl Sabella in bpo-2504.)

inspect

The inspect.getdoc() function can now find docstrings for __slots__ if that attribute is a dict where the values are docstrings. This provides documentation options similar to what we already have for property(), classmethod(), and staticmethod():

class AudioClip:
    __slots__ = {'bit_rate': 'expressed in kilohertz to one decimal place',
                 'duration': 'in seconds, rounded up to an integer'}
    def __init__(self, bit_rate, duration):
        self.bit_rate = round(bit_rate / 1000.0, 1)
        self.duration = ceil(duration)

gc

get_objects() can now receive an optional generation parameter indicating a generation to get objects from. Contributed in bpo-36016 by Pablo Galindo.

gzip

Added the mtime parameter to gzip.compress() for reproducible output. (Contributed by Guo Ci Teo in bpo-34898.)

idlelib and IDLE

Output over N lines (50 by default) is squeezed down to a button. N can be changed in the PyShell section of the General page of the Settings dialog. Fewer, but possibly extra long, lines can be squeezed by right clicking on the output. Squeezed output can be expanded in place by double-clicking the button or into the clipboard or a separate window by right-clicking the button. (Contributed by Tal Einat in bpo-1529353.)

The changes above have been backported to 3.7 maintenance releases.

json.tool

Add option --json-lines to parse every input line as separate JSON object. (Contributed by Weipeng Hong in bpo-31553.)

math

Added new function math.dist() for computing Euclidean distance between two points. (Contributed by Raymond Hettinger in bpo-33089.)

Expanded the math.hypot() function to handle multiple dimensions. Formerly, it only supported the 2-D case. (Contributed by Raymond Hettinger in bpo-33089.)

Added new function, math.prod(), as analogous function to sum() that returns the product of a ‘start’ value (default: 1) times an iterable of numbers. (Contributed by Pablo Galindo in bpo-35606)

os

Added new function add_dll_directory() on Windows for providing additional search paths for native dependencies when importing extension modules or loading DLLs using ctypes.

os.path

os.path functions that return a boolean result like exists(), lexists(), isdir(), isfile(), islink(), and ismount() now return False instead of raising ValueError or its subclasses UnicodeEncodeError and UnicodeDecodeError for paths that contain characters or bytes unrepresentable at the OS level. (Contributed by Serhiy Storchaka in bpo-33721.)

expanduser() on Windows now prefers the USERPROFILE environment variable and does not use HOME, which is not normally set for regular user accounts.

ncurses

Added a new variable holding structured version information for the underlying ncurses library: ncurses_version. (Contributed by Serhiy Storchaka in bpo-31680.)

pathlib

pathlib.Path methods that return a boolean result like exists(), is_dir(), is_file(), is_mount(), is_symlink(), is_block_device(), is_char_device(), is_fifo(), is_socket() now return False instead of raising ValueError or its subclass UnicodeEncodeError for paths that contain characters unrepresentable at the OS level. (Contributed by Serhiy Storchaka in bpo-33721.)

Added pathlib.Path.link_to() which creates a hard link pointing to a path. (Contributed by Joannah Nanjekye in bpo-26978)

socket

Added create_server() and has_dualstack_ipv6() convenience functions to automate the necessary tasks usually involved when creating a server socket, including accepting both IPv4 and IPv6 connections on the same socket. (Contributed by Giampaolo Rodola in bpo-17561.)

shutil

shutil.copytree() now accepts a new dirs_exist_ok keyword argument. (Contributed by Josh Bronson in bpo-20849.)

shutil.make_archive() now defaults to the modern pax (POSIX.1-2001) format for new archives to improve portability and standards conformance, inherited from the corresponding change to the tarfile module. (Contributed by C.A.M. Gerlach in bpo-30661.)

ssl

Added SSLContext.post_handshake_auth to enable and ssl.SSLSocket.verify_client_post_handshake() to initiate TLS 1.3 post-handshake authentication. (Contributed by Christian Heimes in bpo-34670.)

statistics

Added statistics.fmean() as a faster, floating point variant of statistics.mean(). (Contributed by Raymond Hettinger and Steven D’Aprano in bpo-35904.)

Added statistics.geometric_mean() (Contributed by Raymond Hettinger in bpo-27181.)

Added statistics.multimode() that returns a list of the most common values. (Contributed by Raymond Hettinger in bpo-35892.)

Added statistics.quantiles() that divides data or a distribution in to equiprobable intervals (e.g. quartiles, deciles, or percentiles). (Contributed by Raymond Hettinger in bpo-36546.)

Added statistics.NormalDist, a tool for creating and manipulating normal distributions of a random variable. (Contributed by Raymond Hettinger in bpo-36018.)

>>> temperature_feb = NormalDist.from_samples([4, 12, -3, 2, 7, 14])
>>> temperature_feb.mean
6.0
>>> temperature_feb.stdev
6.356099432828281

>>> temperature_feb.cdf(3)            # Chance of being under 3 degrees
0.3184678262814532
>>> # Relative chance of being 7 degrees versus 10 degrees
>>> temperature_feb.pdf(7) / temperature_feb.pdf(10)
1.2039930378537762

>>> el_niño = NormalDist(4, 2.5)
>>> temperature_feb += el_niño        # Add in a climate effect
>>> temperature_feb
NormalDist(mu=10.0, sigma=6.830080526611674)

>>> temperature_feb * (9/5) + 32      # Convert to Fahrenheit
NormalDist(mu=50.0, sigma=12.294144947901014)
>>> temperature_feb.samples(3)        # Generate random samples
[7.672102882379219, 12.000027119750287, 4.647488369766392]

tarfile

The tarfile module now defaults to the modern pax (POSIX.1-2001) format for new archives, instead of the previous GNU-specific one. This improves cross-platform portability with a consistent encoding (UTF-8) in a standardized and extensible format, and offers several other benefits. (Contributed by C.A.M. Gerlach in bpo-36268.)

tokenize

The tokenize module now implicitly emits a NEWLINE token when provided with input that does not have a trailing new line. This behavior now matches what the C tokenizer does internally. (Contributed by Ammar Askar in bpo-33899.)

tkinter

Added methods selection_from(), selection_present(), selection_range() and selection_to() in the tkinter.Spinbox class. (Contributed by Juliette Monsel in bpo-34829.)

Added method moveto() in the tkinter.Canvas class. (Contributed by Juliette Monsel in bpo-23831.)

The tkinter.PhotoImage class now has transparency_get() and transparency_set() methods. (Contributed by Zackery Spytz in bpo-25451.)

time

Added new clock CLOCK_UPTIME_RAW for macOS 10.12. (Contributed by Joannah Nanjekye in bpo-35702.)

unicodedata

unittest

venv

  • venv now includes an Activate.ps1 script on all platforms for activating virtual environments under PowerShell Core 6.1. (Contributed by Brett Cannon in bpo-32718.)

weakref

  • The proxy objects returned by weakref.proxy() now support the matrix multiplication operators @ and @= in addition to the other numeric operators. (Contributed by Mark Dickinson in bpo-36669.)

xml

  • As mitigation against DTD and external entity retrieval, the xml.dom.minidom and xml.sax modules no longer process external entities by default. (Contributed by Christian Heimes in bpo-17239.)
  • The .find*() methods in the xml.etree.ElementTree module support wildcard searches like {*}tag which ignores the namespace and {namespace}* which returns all tags in the given namespace. (Contributed by Stefan Behnel in bpo-28238.)
  • The xml.etree.ElementTree module provides a new function –xml.etree.ElementTree.canonicalize() that implements C14N 2.0. (Contributed by Stefan Behnel in bpo-13611.)

Optimizations

  • The subprocess module can now use the os.posix_spawn() function in some cases for better performance. Currently, it is only used on macOS and Linux (using glibc 2.24 or newer) if all these conditions are met:
    • close_fds is false;
    • preexec_fn, pass_fds, cwd and start_new_session parameters are not set;
    • the executable path contains a directory.
  • shutil.copyfile(), shutil.copy(), shutil.copy2(), shutil.copytree() and shutil.move() use platform-specific “fast-copy” syscalls on Linux, macOS and Solaris in order to copy the file more efficiently. “fast-copy” means that the copying operation occurs within the kernel, avoiding the use of userspace buffers in Python as in “outfd.write(infd.read())”. On Windows shutil.copyfile() uses a bigger default buffer size (1 MiB instead of 16 KiB) and a memoryview()-based variant of shutil.copyfileobj() is used. The speedup for copying a 512 MiB file within the same partition is about +26% on Linux, +50% on macOS and +40% on Windows. Also, much less CPU cycles are consumed. See Platform-dependent efficient copy operations section. (Contributed by Giampaolo Rodola’ in bpo-33671.)
  • shutil.copytree() uses os.scandir() function and all copy functions depending from it use cached os.stat() values. The speedup for copying a directory with 8000 files is around +9% on Linux, +20% on Windows and +30% on a Windows SMB share. Also the number of os.stat() syscalls is reduced by 38% making shutil.copytree() especially faster on network filesystems. (Contributed by Giampaolo Rodola’ in bpo-33695.)
  • The default protocol in the pickle module is now Protocol 4, first introduced in Python 3.4. It offers better performance and smaller size compared to Protocol 3 available since Python 3.0.
  • Removed one Py_ssize_t member from PyGC_Head. All GC tracked objects (e.g. tuple, list, dict) size is reduced 4 or 8 bytes. (Contributed by Inada Naoki in bpo-33597)
  • uuid.UUID now uses __slots__ to reduce its memory footprint.
  • Improved performance of operator.itemgetter() by 33%. Optimized argument handling and added a fast path for the common case of a single non-negative integer index into a tuple (which is the typical use case in the standard library). (Contributed by Raymond Hettinger in bpo-35664.)
  • Sped-up field lookups in collections.namedtuple(). They are now more than two times faster, making them the fastest form of instance variable lookup in Python. (Contributed by Raymond Hettinger, Pablo Galindo, and Joe Jevnik, Serhiy Storchaka in bpo-32492.)
  • The list constructor does not overallocate the internal item buffer if the input iterable has a known length (the input implements __len__). This makes the created list 12% smaller on average. (Contributed by Raymond Hettinger and Pablo Galindo in bpo-33234.)
  • Doubled the speed of class variable writes. When a non-dunder attribute was updated, there was an unnecessary call to update slots. (Contributed by Stefan Behnel, Pablo Galindo Salgado, Raymond Hettinger, Neil Schemenauer, and Serhiy Storchaka in bpo-36012.)
  • Reduced an overhead of converting arguments passed to many builtin functions and methods. This sped up calling some simple builtin functions and methods up to 20–50%. (Contributed by Serhiy Storchaka in bpo-23867, bpo-35582 and bpo-36127.)

Build and C API Changes

  • The PyByteArray_Init() and PyByteArray_Fini() functions have been removed. They did nothing since Python 2.7.4 and Python 3.2.0, were excluded from the limited API (stable ABI), and were not documented.

  • The result of PyExceptionClass_Name() is now of type const char * rather of char *. (Contributed by Serhiy Storchaka in bpo-33818.)

  • The duality of Modules/Setup.dist and Modules/Setup has been removed. Previously, when updating the CPython source tree, one had to manually copy Modules/Setup.dist (inside the source tree) to Modules/Setup (inside the build tree) in order to reflect any changes upstream. This was of a small benefit to packagers at the expense of a frequent annoyance to developers following CPython development, as forgetting to copy the file could produce build failures.

    Now the build system always reads from Modules/Setup inside the source tree. People who want to customize that file are encouraged to maintain their changes in a git fork of CPython or as patch files, as they would do for any other change to the source tree.

    (Contributed by Antoine Pitrou in bpo-32430.)

  • Functions that convert Python number to C integer like PyLong_AsLong() and argument parsing functions like PyArg_ParseTuple() with integer converting format units like 'i' will now use the __index__() special method instead of __int__(), if available. The deprecation warning will be emitted for objects with the __int__() method but without the __index__() method (like Decimal and Fraction). PyNumber_Check() will now return 1 for objects implementing __index__(). (Contributed by Serhiy Storchaka in bpo-36048.)

  • Heap-allocated type objects will now increase their reference count in PyObject_Init() (and its parallel macro PyObject_INIT) instead of in PyType_GenericAlloc(). Types that modify instance allocation or deallocation may need to be adjusted. (Contributed by Eddie Elizondo in bpo-35810.)

Deprecated

API and Feature Removals

The following features and APIs have been removed from Python 3.8:

  • The macpath module, deprecated in Python 3.7, has been removed. (Contributed by Victor Stinner in bpo-35471.)
  • The function platform.popen() has been removed, it was deprecated since Python 3.3: use os.popen() instead.
  • The pyvenv script has been removed in favor of python3.8 -m venv to help eliminate confusion as to what Python interpreter the pyvenv script is tied to. (Contributed by Brett Cannon in bpo-25427.)
  • parse_qs, parse_qsl, and escape are removed from cgi module. They are deprecated from Python 3.2 or older.
  • filemode function is removed from tarfile module. It is not documented and deprecated since Python 3.3.
  • The XMLParser constructor no longer accepts the html argument. It never had effect and was deprecated in Python 3.4. All other parameters are now keyword-only. (Contributed by Serhiy Storchaka in bpo-29209.)
  • Removed the doctype() method of XMLParser. (Contributed by Serhiy Storchaka in bpo-29209.)
  • “unicode_internal” codec is removed. (Contributed by Inada Naoki in bpo-36297.)

Porting to Python 3.8

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

Changes in Python behavior

  • Yield expressions (both yield and yield from clauses) are now disallowed in comprehensions and generator expressions (aside from the iterable expression in the leftmost for clause). (Contributed by Serhiy Storchaka in bpo-10544.)
  • The compiler now produces a SyntaxWarning when identity checks (is and is not) are used with certain types of literals (e.g. strings, ints). These can often work by accident in CPython, but are not guaranteed by the language spec. The warning advises users to use equality tests (== and !=) instead. (Contributed by Serhiy Storchaka in bpo-34850.)
  • The CPython interpreter can swallow exceptions in some circumstances. In Python 3.8 this happens in less cases. In particular, exceptions raised when getting the attribute from the type dictionary are no longer ignored. (Contributed by Serhiy Storchaka in bpo-35459.)
  • On AIX, sys.platform doesn’t contain the major version anymore. It is always 'aix', instead of 'aix3' .. 'aix7'. Since older Python versions include the version number, it is recommended to always use the sys.platform.startswith('aix'). (Contributed by M. Felt in bpo-36588.)
  • PyEval_AcquireLock() and PyEval_AcquireThread() now terminate the current thread if called while the interpreter is finalizing, making them consistent with PyEval_RestoreThread(), Py_END_ALLOW_THREADS(), and PyGILState_Ensure(). If this behaviour is not desired, guard the call by checking _Py_IsFinalizing() or sys.is_finalizing().

Changes in the Python API

  • subprocess.Popen can now use os.posix_spawn() in some cases for better performance. On Windows Subsystem for Linux and QEMU User Emulation, Popen constructor using os.posix_spawn() no longer raise an exception on errors like missing program, but the child process fails with a non-zero returncode.
  • The imap.IMAP4.logout() method no longer ignores silently arbitrary exceptions.
  • The function platform.popen() has been removed, it was deprecated since Python 3.3: use os.popen() instead.
  • The statistics.mode() function no longer raises an exception when given multimodal data. Instead, it returns the first mode encountered in the input data. (Contributed by Raymond Hettinger in bpo-35892.)
  • The selection() method of the tkinter.ttk.Treeview class no longer takes arguments. Using it with arguments for changing the selection was deprecated in Python 3.6. Use specialized methods like selection_set() for changing the selection. (Contributed by Serhiy Storchaka in bpo-31508.)
  • The writexml(), toxml() and toprettyxml() methods of the xml.dom.minidom module, and xml.etree now preserve the attribute order specified by the user. (Contributed by Diego Rojas and Raymond Hettinger in bpo-34160.)
  • A dbm.dumb database opened with flags 'r' is now read-only. dbm.dumb.open() with flags 'r' and 'w' no longer creates a database if it does not exist. (Contributed by Serhiy Storchaka in bpo-32749.)
  • The doctype() method defined in a subclass of XMLParser will no longer be called and will cause emitting a RuntimeWarning instead of a DeprecationWarning. Define the doctype() method on a target for handling an XML doctype declaration. (Contributed by Serhiy Storchaka in bpo-29209.)
  • A RuntimeError is now raised when the custom metaclass doesn’t provide the __classcell__ entry in the namespace passed to type.__new__. A DeprecationWarning was emitted in Python 3.6–3.7. (Contributed by Serhiy Storchaka in bpo-23722.)
  • The cProfile.Profile class can now be used as a context manager. (Contributed by Scott Sanderson in bpo-29235.)
  • shutil.copyfile(), shutil.copy(), shutil.copy2(), shutil.copytree() and shutil.move() use platform-specific “fast-copy” syscalls (see Platform-dependent efficient copy operations section).
  • shutil.copyfile() default buffer size on Windows was changed from 16 KiB to 1 MiB.
  • PyGC_Head struct is changed completely. All code touched the struct member should be rewritten. (See bpo-33597)
  • The PyInterpreterState struct has been moved into the “internal” header files (specifically Include/internal/pycore_pystate.h). An opaque PyInterpreterState is still available as part of the public API (and stable ABI). The docs indicate that none of the struct’s fields are public, so we hope no one has been using them. However, if you do rely on one or more of those private fields and have no alternative then please open a BPO issue. We’ll work on helping you adjust (possibly including adding accessor functions to the public API). (See bpo-35886.)
  • Asyncio tasks can now be named, either by passing the name keyword argument to asyncio.create_task() or the create_task() event loop method, or by calling the set_name() method on the task object. The task name is visible in the repr() output of asyncio.Task and can also be retrieved using the get_name() method.
  • The mmap.flush() method now returns None on success and raises an exception on error under all platforms. Previously, its behavior was platform-depended: a nonzero value was returned on success; zero was returned on error under Windows. A zero value was returned on success; an exception was raised on error under Unix. (Contributed by Berker Peksag in bpo-2122.)
  • The function math.factorial() no longer accepts arguments that are not int-like. (Contributed by Pablo Galindo in bpo-33083.)
  • xml.dom.minidom and xml.sax modules no longer process external entities by default. (Contributed by Christian Heimes in bpo-17239.)
  • Deleting a key from a read-only dbm database (dbm.dumb, dbm.gnu or dbm.ndbm) raises error (dbm.dumb.error, dbm.gnu.error or dbm.ndbm.error) instead of KeyError. (Contributed by Xiang Zhang in bpo-33106.)
  • expanduser() on Windows now prefers the USERPROFILE environment variable and does not use HOME, which is not normally set for regular user accounts.
  • DLL dependencies for extension modules and DLLs loaded with ctypes on Windows are now resolved more securely. Only the system paths, the directory containing the DLL or PYD file, and directories added with add_dll_directory() are searched for load-time dependencies. Specifically, PATH and the current working directory are no longer used, and modifications to these will no longer have any effect on normal DLL resolution. If your application relies on these mechanisms, you should check for add_dll_directory() and if it exists, use it to add your DLLs directory while loading your library. Note that Windows 7 users will need to ensure that Windows Update KB2533625 has been installed (this is also verified by the installer). (See bpo-36085.)
  • The header files and functions related to pgen have been removed after its replacement by a pure Python implementation. (Contributed by Pablo Galindo in bpo-36623.)

Changes in the C API

  • On Unix, C extensions are no longer linked to libpython except on Android. When Python is embedded, libpython must not be loaded with RTLD_LOCAL, but RTLD_GLOBAL instead. Previously, using RTLD_LOCAL, it was already not possible to load C extensions which were not linked to libpython, like C extensions of the standard library built by the *shared* section of Modules/Setup.

  • Use of # variants of formats in parsing or building value (e.g. PyArg_ParseTuple(), Py_BuildValue(), PyObject_CallFunction(), etc.) without PY_SSIZE_T_CLEAN defined raises DeprecationWarning now. It will be removed in 3.10 or 4.0. Read Parsing arguments and building values for detail. (Contributed by Inada Naoki in bpo-36381.)

  • Instances of heap-allocated types (such as those created with PyType_FromSpec()) hold a reference to their type object. Increasing the reference count of these type objects has been moved from PyType_GenericAlloc() to the more low-level functions, PyObject_Init() and PyObject_INIT(). This makes types created through PyType_FromSpec() behave like other classes in managed code.

    Statically allocated types are not affected.

    For the vast majority of cases, there should be no side effect. However, types that manually increase the reference count after allocating an instance (perhaps to work around the bug) may now become immortal. To avoid this, these classes need to call Py_DECREF on the type object during instance deallocation.

    To correctly port these types into 3.8, please apply the following changes:

    • Remove Py_INCREF on the type object after allocating an instance - if any. This may happen after calling PyObject_New(), PyObject_NewVar(), PyObject_GC_New(), PyObject_GC_NewVar(), or any other custom allocator that uses PyObject_Init() or PyObject_INIT().

      Example:

      static foo_struct *
foo_new(PyObject *type) {
    foo_struct *foo = PyObject_GC_New(foo_struct, (PyTypeObject *) type);
    if (foo == NULL)
        return NULL;
#if PY_VERSION_HEX < 0x03080000
    // Workaround for Python issue 35810; no longer necessary in Python 3.8
    PY_INCREF(type)
#endif
    return foo;
}

    • Ensure that all custom tp_dealloc functions of heap-allocated types decrease the type’s reference count.

      Example:

      static void
foo_dealloc(foo_struct *instance) {
    PyObject *type = Py_TYPE(instance);
    PyObject_GC_Del(instance);
#if PY_VERSION_HEX >= 0x03080000
    // This was not needed before Python 3.8 (Python issue 35810)
    Py_DECREF(type);
#endif
}

    (Contributed by Eddie Elizondo in bpo-35810.)

CPython bytecode changes

  • The interpreter loop has been simplified by moving the logic of unrolling the stack of blocks into the compiler. The compiler emits now explicit instructions for adjusting the stack of values and calling the cleaning-up code for break, continue and return.

    Removed opcodes BREAK_LOOP, CONTINUE_LOOP, SETUP_LOOP and SETUP_EXCEPT. Added new opcodes ROT_FOUR, BEGIN_FINALLY, CALL_FINALLY and POP_FINALLY. Changed the behavior of END_FINALLY and WITH_CLEANUP_START.

    (Contributed by Mark Shannon, Antoine Pitrou and Serhiy Storchaka in bpo-17611.)

  • Added new opcode END_ASYNC_FOR for handling exceptions raised when awaiting a next item in an async for loop. (Contributed by Serhiy Storchaka in bpo-33041.)

Demos and Tools

  • Added a benchmark script for timing various ways to access variables: Tools/scripts/var_access_benchmark.py. (Contributed by Raymond Hettinger in bpo-35884.)