enum.Flag should be more set-like

[belm0]

BPO	38250
Nosy	@belm0, @ethanfurman, @vedgar, @hroncok, @Zheaoli, @belm0, @hauntsaninja
PRs	bpo-32218: make Flag and IntFlag members iterable #22221 bpo-38250: minor Flag refactor #22734 bpo-38250: [Enum] single-bit flags are canonical #24215 bpo-38250: [Enum] only include .rst test if file available #24342 bpo-38250: add version added for FlagBoundary #25820

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

Show more details

GitHub fields:

assignee = 'https://github.com/ethanfurman'
closed_at = <Date 2021-01-26.22:15:42.201>
created_at = <Date 2019-09-22.11:14:02.624>
labels = ['type-feature', 'library', '3.10']
title = 'enum.Flag should be more set-like'
updated_at = <Date 2021-05-26.19:32:15.310>
user = 'https://github.com/belm0'

bugs.python.org fields:

activity = <Date 2021-05-26.19:32:15.310>
actor = 'hroncok'
assignee = 'ethan.furman'
closed = True
closed_date = <Date 2021-01-26.22:15:42.201>
closer = 'vstinner'
components = ['Library (Lib)']
creation = <Date 2019-09-22.11:14:02.624>
creator = 'John Belmonte'
dependencies = []
files = []
hgrepos = []
issue_num = 38250
keywords = ['patch']
message_count = 19.0
messages = ['352967', '365051', '365072', '378436', '378439', '378455', '378456', '378461', '378552', '378553', '378555', '378606', '385062', '385672', '385677', '385701', '385735', '385737', '394458']
nosy_count = 7.0
nosy_names = ['jbelmonte', 'ethan.furman', 'veky', 'hroncok', 'Manjusaka', 'John Belmonte', 'hauntsaninja']
pr_nums = ['22221', '22734', '24215', '24342', '25820']
priority = 'normal'
resolution = 'fixed'
stage = 'resolved'
status = 'closed'
superseder = None
type = 'enhancement'
url = 'https://bugs.python.org/issue38250'
versions = ['Python 3.10']

[belm0]

I would like Flag class instances to have more set-like abilities:

1. iteration, to walk through each set bit of the value
2. len corresponding to Support "bpo-" in Misc/NEWS #1
3. subset operator

I may be told "implement it yourself as instance methods", or that #3 has an idiom (a & b is b). Ideally though, every Flag user should be able to rely on these being implemented consistently.

When trying to implement #1 without devolving into bit fiddling, naturally one might try to use the class iterator. Unfortunately the semantics of that enumeration include 0, aliases, and compound values. I've used Flag in several situations and projects, and so far there hasn't been a case where that was the desired semantics. Interesting though, if #1 were implemented in the standard library, then we could enumerate all bits of the Flag via iteration of ~MyFlag(0)... though that's obscuring things behind another idiom.

Thank you for considering.

[vedgar]

1. +0 _if_ the implementation is easy to explain. If backward compatibility is an issue, we can always add a property:
   for flag in flags.set:
   (though set might imply unorderedness:)
2. -1. Guido said long ago that all lens should be O(1).
   (Of course, if you do make it O(1), I have no objection.)
3. +1, absolutely.

[Zheaoli]

1. not sure I gett the Point

2. not sure

3. absolutely yes

[belm0]

Part of this issue (#1) was intended to be addressed by #22221 which added an __iter__ implementation to Flag and IntFlag. (The PR did not reference this issue, and was already merged last month.)

However that PR seems problematic on several counts:

1. __iter__ diverges from the existing __contains__. The latter includes 0 and compound values
2. the implementation is fairly heavy
3. len() on an enum instance is going to be O(n)

I've put post-merge comments on the PR.

I think it would be safer to have an explicitly named bits() iterator on flag instances, rather than use __iter__().

[vedgar]

Just a comment, (1) is analogous to str. iter('abc') gives only 'a', 'b' and 'c', while contains accepts '', 'ab', 'bc', and 'abc' too. At least in my mind, it's a pretty strong analogy.

[belm0]

Just a comment, (1) is analogous to str. iter('abc') gives only 'a', 'b' and 'c', while contains accepts '', 'ab', 'bc', and 'abc' too. At least in my mind, it's a pretty strong analogy.

I don't agree. The "zero" bit does not exist, so having contains return True on Foo(0) in x is misaligned with the iterator. And having contains return True for specific compound values just because they happen to be explicitly defined, while returning False for others, is arbitrary. contains seems to be of very little use, and moreover a trap for the unwary. Assuming we have to live with that until Python 4, it's better to make an explicit iterator like bits() so that the API doesn't contradict itself.

[vedgar]

Of course, if it returns True only on _some_ bits combinations it doesn't make sense. I thought every element of a Boolean span would be _in_ the Foo.

But about "zero bit", I still think it's perfectly analogous to '' in 'abc'.

[belm0]

I think #22221 should be reverted (considering the design issue, performance issue, and bugs), and lets have a proper design and review.

While just reading the code, I found an existing bug in Flag. And the new __iter__ uses the buggy internal function, and so itself has bugs.

#22221 (comment)

[belm0]

It's completely undocumented, but today I noticed that Flag.__contains__() is actually a subset operation.

    def __contains__(self, other):
        ...
        return other._value_ & self._value_ == other._value_

It's an unfortunate departure from the set type, which uses in for membership test and issubset() / <= for subset test.

For set operations, the Flag individual bits should be considered the members of a set (not Flag compound values, which are themselves equivalent to a set).

[vedgar]

Again, I disagree. str used to work like this in Py2.0 (or somewhere around then), only 'x' was in 'xyz', not 'xy'. Then Guido came to his senses. :-)

This is not set theory, this is mereology. You don't differentiate between a digit and a one-digit number, a char and a one-char string, and in the same way you shouldn't differentiate between a bit and a one-bit flag.

[belm0]

I agree that a bit and one-bit flag are the same.

only 'x' was in 'xyz', not 'xy

I don't understand the comparison, because str 'a in b' tests if 'a' is a subsequence of 'b'. It is not a subset operation ('xz' in 'xyz' is false).

I can understand the argument that Flag has a subset operator (currently __contains__), and given that one-bit flags can be used freely with the subset operator, there is no reason to add a bit membership operator.

However, since flag values are arguably a set of enabled bits, I think the use of in for subset is a confusing departure from the set API.

[vedgar]

Flag values are _not_ a set of enabled bits.

At least, it is a weird kind of set where a is the same as {a}. That's why I mentioned mereology... there is no reasonable "membership", just "inclusion". I understand that str is not a perfect analogy, but sets are even more wrong.

[ethanfurman]

The code sample:

    class Color(IntFlag):
        BLACK = 0
        RED = 1
        GREEN = 2
        BLUE = 4
        PURPLE = RED | BLUE
        WHITE = RED | GREEN | BLUE

Here's the summary of the changes:

• single-bit flags are canonical
• multi-bit and zero-bit flags are aliases
  + only canonical flags are returned during iteration

    >>> list(Color.WHITE)
    [<Color.RED: 1>, <Color.GREEN: 2>, <Color.BLUE: 4>]

• negating a flag or flag set returns a new flag/flag set with the
  corresponding positive integer value

  Color.GREEN
  <Color.GREEN: 2>

  ~Color.GREEN
  <Color.PURPLE: 5>

• names of pseudo-flags are constructed from their members' names

    >>> (Color.RED | Color.GREEN).name
    'RED|GREEN'

• multi-bit flags, aka aliases, can be returned from operations

    >>> Color.RED | Color.BLUE
    <Color.PURPLE: 5>

    >>> Color(7)  # or Color(-1)
    <Color.WHITE: 7>

• membership / containment checking has changed slightly -- zero valued flags
  are never considered to be contained:

    >>> Color.BLACK in Color.WHITE
    False

otherwise, if all bits of one flag are in the other flag, True is returned:

    >>> Color.PURPLE in Color.WHITE
    True

There is a new boundary mechanism that controls how out-of-range / invalid bits are handled: STRICT, CONFORM, EJECT', and KEEP':

STRICT --> raises an exception when presented with invalid values
CONFORM --> discards any invalid bits
EJECT --> lose Flag status and become a normal int with the given value
KEEP --> keep the extra bits
- keeps Flag status and extra bits
- they don't show up in iteration
- they do show up in repr() and str()

The default for Flag is STRICT, the default for IntFlag is DISCARD, and the default for _convert_ is KEEP (see ssl.Options for an example of when KEEP is needed).

[ethanfurman]

New changeset 7aaeb2a by Ethan Furman in branch 'master':
bpo-38250: [Enum] single-bit flags are canonical (GH-24215)
7aaeb2a

[ethanfurman]

Thank you to everyone involved. :-)

To answer the first three points that started this issue:

1. iteration -> each single-bit flag in the entire flag, or a
   combinations of flags, is returned one at a time -- not the
   empty set, not other multi-bit values

2. length is implemented -> len(Color.BLUE | Color.RED) == 2

3. subset is implemented as containment checking:
   Color.BLUE in (Color.RED | Color.BLUE) is True

[vstinner]

test_enum fails when Python is installed:

PPC64LE Fedora Rawhide Clang Installed 3.x:
https://buildbot.python.org/all/#builders/312/builds/597

0:01:37 load avg: 8.99 [232/426/1] test_enum failed -- running: test_tokenize (1 min 37 sec), test_unparse (31.7 sec), test_concurrent_futures (37.5 sec)
Failed to call load_tests:
Traceback (most recent call last):
  File "/home/buildbot/buildarea/3.x.cstratak-fedora-rawhide-ppc64le.clang-installed/build/target/lib/python3.10/unittest/loader.py", line 130, in loadTestsFromModule
    return load_tests(self, tests, pattern)
  File "/home/buildbot/buildarea/3.x.cstratak-fedora-rawhide-ppc64le.clang-installed/build/target/lib/python3.10/test/test_enum.py", line 20, in load_tests
    tests.addTests(doctest.DocFileSuite(
  File "/home/buildbot/buildarea/3.x.cstratak-fedora-rawhide-ppc64le.clang-installed/build/target/lib/python3.10/doctest.py", line 2511, in DocFileSuite
    suite.addTest(DocFileTest(path, **kw))
  File "/home/buildbot/buildarea/3.x.cstratak-fedora-rawhide-ppc64le.clang-installed/build/target/lib/python3.10/doctest.py", line 2433, in DocFileTest
    doc, path = _load_testfile(path, package, module_relative,
  File "/home/buildbot/buildarea/3.x.cstratak-fedora-rawhide-ppc64le.clang-installed/build/target/lib/python3.10/doctest.py", line 231, in _load_testfile
    file_contents = loader.get_data(filename)
  File "<frozen importlib._bootstrap_external>", line 1023, in get_data
FileNotFoundError: [Errno 2] No such file or directory: '/home/buildbot/buildarea/3.x.cstratak-fedora-rawhide-ppc64le.clang-installed/build/target/lib/python3.10/test/../../Doc/library/enum.rst'

test test_enum crashed -- Traceback (most recent call last):
  File "/home/buildbot/buildarea/3.x.cstratak-fedora-rawhide-ppc64le.clang-installed/build/target/lib/python3.10/test/libregrtest/runtest.py", line 272, in _runtest_inner
    refleak = _runtest_inner2(ns, test_name)
  File "/home/buildbot/buildarea/3.x.cstratak-fedora-rawhide-ppc64le.clang-installed/build/target/lib/python3.10/test/libregrtest/runtest.py", line 236, in _runtest_inner2
    test_runner()
  File "/home/buildbot/buildarea/3.x.cstratak-fedora-rawhide-ppc64le.clang-installed/build/target/lib/python3.10/test/libregrtest/runtest.py", line 210, in _test_module
    raise Exception("errors while loading tests")
Exception: errors while loading tests

Tests are loaded by Lib/test/test_enum.py with:

def load_tests(loader, tests, ignore):
    tests.addTests(doctest.DocTestSuite(enum))
    tests.addTests(doctest.DocFileSuite(
            '../../Doc/library/enum.rst',
            optionflags=doctest.ELLIPSIS|doctest.NORMALIZE_WHITESPACE,
            ))
    return tests

[ethanfurman]

New changeset 01faf45 by Ethan Furman in branch 'master':
bpo-38250: [Enum] only include .rst test if file available (GH-24342)
01faf45

[vstinner]

PPC64LE Fedora Rawhide Clang Installed 3.x is back to green, thanks for the fix ;-)
https://buildbot.python.org/all/#/builders/312/builds/600

[hroncok]

I've reported https://bugs.python.org/issue44242 because I believe there is a regression in this change.