This removes the performance regression in 3.11, **at the expense of not fixing
the "bug" that allows accessing values from values** (e.g. `Color.RED.BLUE`).
Using the benchmark @markshannon [presented](https://github.com/python/cpython/issues/93910#issuecomment-1165503032), the results are:
| Version | Enum | Fast enum | Normal class |
| --- | --- | --- | --- |
| 3.10 | 2.04 | 0.59 | 0.56 |
| 3.11 | 2.78 | 0.31 | 0.15 |
| This PR | 1.30 | 0.32 | 0.16 |
I share this mostly as information about the source of the regression, as this may be useful. It may be that the lower-risk approach for the beta is just to revert to a previously-known working state.
When used with plain Enum, auto() returns the last numeric value assigned, skipping any incompatible member values (such as strings); starting in 3.13 the default auto() for plain Enums will require all the values to be of compatible types, and will return a new value that is 1 higher than any existing value.
Co-authored-by: Ethan Furman <ethan@stoneleaf.us>
GH-26658 introduced a regression in copy / pickle protocol for combined
`enum.Flag`s. `copy.copy(re.A | re.I)` would fail with
`AttributeError: ASCII|IGNORECASE`.
`enum.Flag` now has a `__reduce_ex__()` method that reduces flags by
combined value, not by combined name.
In previous versions of Python if an IntEnum member was combined with another integer type value using a bit-wise operation, the resulting value would still be the IntEnum type. This change restores that behavior.
`EnumType` attempts to create a custom docstring for each enum/flag, but that was failing with pathological flags that had no members (only multi-bit aliases).
- add member() and nonmember() functions
- add deprecation warning for internal classes in enums not
becoming members in 3.13
Co-authored-by: edwardcwang
Undo rejected PEP-663 changes:
- restore `repr()` to its 3.10 status
- restore `str()` to its 3.10 status
New changes:
- `IntEnum` and `IntFlag` now leave `__str__` as the original `int.__str__` so that str() and format() return the same result
- zero-valued flags without a name have a slightly changed repr(), e.g. `repr(Color(0)) == '<Color: 0>'`
- update `dir()` for mixed-in types to return all the methods and attributes of the mixed-in type
- added `_numeric_repr_` to `Flag` to control display of unnamed values
- enums without doc strings have a more comprehensive doc string added
- `ReprEnum` added -- inheriting from this makes it so only `__repr__` is replaced, not `__str__` nor `__format__`; `IntEnum`, `IntFlag`, and `StrEnum` all inherit from `ReprEnum`
Modify the ``EnumType.__dir__()`` and ``Enum.__dir__()`` to ensure
that user-defined methods and methods inherited from mixin classes always
show up in the output of `help()`. This change also makes it easier for
IDEs to provide auto-completion.
Creating an Enum exhibited quadratic behavior based on the number of members in three places:
- `EnumDict._member_names`: a list searched with each new member's name
- member creation: a `for` loop checking each existing member to see if new member was a duplicate
- `auto()` values: a list of all previous values in enum was copied before being sent to `_generate_next_value()`
Two of those issues have been resolved:
- `_EnumDict._member_names` is now a dictionary so lookups are fast
- member creation tries a fast value lookup before falling back to the slower `for` loop lookup
The third issue still remains, as `_generate_next_value_()` can be user-overridden and could corrupt the last values list if it were not copied.
* Fix typo in __repr__ code
* Add more tests for global int flag reprs
* use last module if multi-module string
- when an enum's `__module__` contains several module names, only
use the last one
Co-authored-by: Łukasz Langa <lukasz@langa.pl>
Co-authored-by: Ethan Furman <ethan@stoneleaf.us>
* [Enum] reduce scope of new format behavior
Instead of treating all Enums the same for format(), only user mixed-in
enums will be affected. In other words, IntEnum and IntFlag will not be
changing the format() behavior, due to the requirement that they be
drop-in replacements of existing integer constants.
If a user creates their own integer-based enum, then the new behavior
will apply:
class Grades(int, Enum):
A = 5
B = 4
C = 3
D = 2
F = 0
Now: format(Grades.B) -> DeprecationWarning and '4'
3.12: -> no warning, and 'B'
by-value lookups could fail on complex enums, necessitating a check for
__reduce__ and possibly sabotaging the final enum;
by-name lookups should never fail, and sabotaging is no longer necessary
for class-based enum creation.
This enables, for example, two base Enums to both inherit from `str`, and then both be mixed into the same final Enum:
class Str1Enum(str, Enum):
# some behavior here
class Str2Enum(str, Enum):
# some more behavior here
class FinalStrEnum(Str1Enum, Str2Enum):
# this now works
In 3.12 ``True`` or ``False`` will be returned for all containment checks,
with ``True`` being returned if the value is either a member of that enum
or one of its members' value.
In 3.12 the enum member, not the member's value, will be used for
format() calls. Format specifiers can be used to retain the current
display of enum members:
Example enumeration:
class Color(IntEnum):
RED = 1
GREEN = 2
BLUE = 3
Current behavior:
f'{Color.RED}' --> '1'
Future behavior:
f'{Color.RED}' --> 'RED'
Using d specifier:
f'{Color.RED:d}' --> '1'
Using specifiers can be done now and is future-compatible.
Depending on usage, it's possible for Flag members to have the _inverted_ attribute when they are testing, while the Flag being testing against will not have that attribute on its members -- so skip that comparison.
add:
* `_simple_enum` decorator to transform a normal class into an enum
* `_test_simple_enum` function to compare
* `_old_convert_` to enable checking `_convert_` generated enums
`_simple_enum` takes a normal class and converts it into an enum:
@simple_enum(Enum)
class Color:
RED = 1
GREEN = 2
BLUE = 3
`_old_convert_` works much like` _convert_` does, using the original logic:
# in a test file
import socket, enum
CheckedAddressFamily = enum._old_convert_(
enum.IntEnum, 'AddressFamily', 'socket',
lambda C: C.isupper() and C.startswith('AF_'),
source=_socket,
)
`_test_simple_enum` takes a traditional enum and a simple enum and
compares the two:
# in the REPL or the same module as Color
class CheckedColor(Enum):
RED = 1
GREEN = 2
BLUE = 3
_test_simple_enum(CheckedColor, Color)
_test_simple_enum(CheckedAddressFamily, socket.AddressFamily)
Any important differences will raise a TypeError
add:
_simple_enum decorator to transform a normal class into an enum
_test_simple_enum function to compare
_old_convert_ to enable checking _convert_ generated enums
_simple_enum takes a normal class and converts it into an enum:
@simple_enum(Enum)
class Color:
RED = 1
GREEN = 2
BLUE = 3
_old_convert_ works much like _convert_ does, using the original logic:
# in a test file
import socket, enum
CheckedAddressFamily = enum._old_convert_(
enum.IntEnum, 'AddressFamily', 'socket',
lambda C: C.isupper() and C.startswith('AF_'),
source=_socket,
)
test_simple_enum takes a traditional enum and a simple enum and
compares the two:
# in the REPL or the same module as Color
class CheckedColor(Enum):
RED = 1
GREEN = 2
BLUE = 3
_test_simple_enum(CheckedColor, Color)
_test_simple_enum(CheckedAddressFamily, socket.AddressFamily)
Any important differences will raise a TypeError
* Enum: streamline repr() and str(); improve docs
- repr() is now ``enum_class.member_name``
- stdlib global enums are ``module_name.member_name``
- str() is now ``member_name``
- add HOW-TO section for ``Enum``
- change main documentation to be an API reference
In 3.5 (?) a speed optimization made it possible to access members as
attributes of other members, i.e. ``Color.RED.BLUE``. This was always
discouraged in the docs, and other recent optimizations has made that
one no longer necessary. Because some may be relying on it anyway, it
is being deprecated in 3.10, and will be removed in 3.11.
Flag members are now divided by one-bit verses multi-bit, with multi-bit being treated as aliases. Iterating over a flag only returns the contained single-bit flags.
Iterating, repr(), and str() show members in definition order.
When constructing combined-member flags, any extra integer values are either discarded (CONFORM), turned into ints (EJECT) or treated as errors (STRICT). Flag classes can specify which of those three behaviors is desired:
>>> class Test(Flag, boundary=CONFORM):
... ONE = 1
... TWO = 2
...
>>> Test(5)
<Test.ONE: 1>
Besides the three above behaviors, there is also KEEP, which should not be used unless necessary -- for example, _convert_ specifies KEEP as there are flag sets in the stdlib that are incomplete and/or inconsistent (e.g. ssl.Options). KEEP will, as the name suggests, keep all bits; however, iterating over a flag with extra bits will only return the canonical flags contained, not the extra bits.
Iteration is now in member definition order. If member definition order
matches increasing value order, then a more efficient method of flag
decomposition is used; otherwise, sort() is called on the results of
that method to get definition order.
``re`` module:
repr() has been modified to support as closely as possible its previous
output; the big difference is that inverted flags cannot be output as
before because the inversion operation now always returns the comparable
positive result; i.e.
re.A|re.I|re.M|re.S is ~(re.L|re.U|re.S|re.T|re.DEBUG)
in both of the above terms, the ``value`` is 282.
re's tests have been updated to reflect the modifications to repr().
`type.__new__` calls `__set_name__` and `__init_subclass__`, which means
that any work metaclasses do after calling `super().__new__()` will not
be available to those two methods. In particular, `Enum` classes that
want to make use of `__init_subclass__` will not see any members.
Almost all customization is therefore moved to before the
`type.__new__()` call, including changing all members to a proto member
descriptor with a `__set_name__` that will do the final conversion of a
member to be an instance of the `Enum` class.
Michael Droettboom
Serhiy Storchaka
Sam Ezeh
Ethan Furman
Oscar R
Carl Bordum Hansen
Christian Heimes
Tobin Yehle
Kumar Aditya
Victor Stinner
Nikita Sobolev
Alex Waygood
Carl Friedrich Bolz-Tereick
Pablo Galindo Salgado
andrei kulakov