Filter the files to run pytype tests on by stdlib/VERSIONS file. This
becomes important for Python 3.12, where e.g. checking asynchat.pyi
requires asyncore.pyi, both of which have been removed in 3.12.
- Add several type hints to untyped functions.
- While we're at it, upgrade several annotations to use modern syntax by using `from __future__ import annotations`. This means that the tests can't be run on Python 3.6, but 3.6 is EOL, and it is already the case that some scripts in this directory can only be run on more recent Python versions. E.g. `check_new_syntax.py` uses `ast.unparse`, which is only available in Python 3.9+.
- Fix a few pieces of code that didn't type check.
For now, the test requires changes that have been merged into the pytype master branch but not yet released. I'll update requirements-tests-py3.txt again once I've cut a new release (hopefully later this week).
* Fix pytype_test.
* Re-enable pytype_test.
* Remove continue-on-error from pytype_test.
* Temporarily install pytype from its master branch, as the code to
support the new directory structure hasn't been released yet.
* Set TYPESHED_HOME before using pytype.pytd.Typeshed().
I noticed that the pytype parse test was getting quite slow again. This
is a somewhat hacky change that caches pytype's internal pyi loader to
avoid parsing files multiple times. On my machine, the test goes from
taking ~6m to complete to ~30s.
[tests/pytype_test.py](./tests/pytype_test.py#L176) is still using
Python `3.6`:
> python_version="2.7" if version == 2 else "3.6",
The test fails to run if `python3.6` is unavailable (for example on
Debian 10 Buster, which has 3.7).
According to (https://devguide.python.org/#status-of-python-branches
`3.6` is still maintained until end of 2021.
`pytype` itself [claims](https://github.com/google/pytype#requirements)
to run on `2.7`, `3.5`-`3.8`.
Change the script to use the version of Python from the invoking
interpreter.
This replaces all uses of ‘blacklist’ with ‘exclude list’. Benefits:
- It is racially neutral terminology; see e.g. [1]
- It makes the meaning more clear. In fact, with the popular Python
autoformatter called ‘black’, also used by this project, files can be
‘blackened’ which is something completely different from them being on
a blacklist.
[1] https://chromium.googlesource.com/chromium/src/+/master/styleguide/inclusive_code.md#racially-neutral
This is a small cosmetic change. I want to encourage use of the
nicer-lookng pytype.config.Options.create() rather than
direct construction of the command-line args, so I'm changing
over all occurrences of the latter that I can find.
This also makes the flask stubs that depend on click parseable.
* In pytype_test, only parse files with the .pyi extension.
(There's a README in third_party/2and3/click/.)
* Annotate with ContextManager rather than using the
contextmanager decorator. This is shorter, removes a
dependency, and gets rid of the slight weirdness of a
decorator needing to be evaluated in a stub.
* Fix non-stub things.
During the PyCon sprints, I'm planning to bash the pytype pyi
parser into shape so we can handle all of third_party. As a first
step, create an explicit blacklist of everything pytype chokes on.
With today's release, pytype is able to fully load
stdlib/3/collections/__init__.pyi, so the test no longer needs the
ability to partially parse stubs using the pytd tool. Removing this
functionality allows the test code to be simplified considerably.
Speeds up pytype considerably by directly calling pytype.io.parse_pyi
rather than running `pytype --parse-pyi` in a subprocess.
The subprocess logic is still required to handle the one pyi file
marked `# parse only` in the pytype blacklist. Once
https://github.com/google/pytype/issues/242 is fixed, the test code
can be cleaned up considerably.
