Something Alex and I threw together during our 1:1 this morning. Allows
us to collect statistics on the prevalence of various types in a file,
most usefully TODO types or other dynamic types.
`FlowSnapshot` now tracks a `reachable` bool, which indicates whether we
have encountered a terminal statement on that control flow path. When
merging flow states together, we skip any that have been marked
unreachable. This ensures that bindings that can only be reached through
unreachable paths are not considered visible.
## Test Plan
The new mdtests failed (with incorrect `reveal_type` results, and
spurious `possibly-unresolved-reference` errors) before adding the new
visibility constraints.
---------
Co-authored-by: Carl Meyer <carl@astral.sh>
## Summary
When we discussed the plan on how to proceed with instance attributes,
we said that we should first extend our research into the behavior of
existing type checkers. The result of this research is summarized in the
newly added / modified tests in this PR. The TODO comments align with
existing behavior of other type checkers. If we deviate from the
behavior, it is described in a comment.
## Summary
Adds a slightly more comprehensive documentation of our behavior
regarding type inference for public uses of symbols. In particular:
- What public type do we infer for `x: int = any()`?
- What public type do we infer for `x: Unknown = 1`?
## Summary
On `main`, red-knot:
- Considers `P | Q` equivalent to `Q | P`
- Considered `tuple[P | Q]` equivalent to `tuple[Q | P]`
- Considers `tuple[P | tuple[P | Q]]` equivalent to `tuple[tuple[Q | P]
| P]`
- ‼️ Does _not_ consider `tuple[tuple[P | Q]]` equivalent to
`tuple[tuple[Q | P]]`
The key difference for the last one of these is that the union appears
inside a tuple that is directly nested inside another tuple.
This PR fixes this so that differently ordered unions are considered
equivalent even when they appear inside arbitrarily nested tuple types.
## Test Plan
- Added mdtests that fails on `main`
- Checked that all property tests continue to pass with this PR
This is a follow-up to #15702 that hopefully claws back the 1%
performance regression. Assuming it works, the trick is to iterate over
the constraints vectors via mut reference (aka a single pointer), so
that we're not copying `BitSet`s into and out of the zip tuples as we
iterate. We use `std::mem::take` as a poor-man's move constructor only
at the very end, when we're ready to emplace it into the result. (C++
idioms intended! 😄)
With local testing via hyperfine, I'm seeing this be 1-3% faster than
`main` most of the time — though a small number of runs (1 in 10,
maybe?) are a wash or have `main` faster. Codspeed reports a 2%
gain.
## Summary
Use `Unknown | T_inferred` as the type for *undeclared* public symbols.
## Test Plan
- Updated existing tests
- New test for external `__slots__` modifications.
- New tests for external modifications of public symbols.
## Summary
Another small PR to focus #15674 solely on the relevant changes. This
makes our Markdown tests less dependent on precise types of public
symbols, without actually changing anything semantically in these tests.
Best reviewed using ignore-whitespace-mode.
## Test Plan
Tested these changes on `main` and on the branch from #15674.
## Summary
Make the remaining `infer.rs` unit tests independent from public symbol
type inference decisions (see upcoming change in #15674).
## Test Plan
- Made sure that the unit tests actually fail if one of the
`assert_type` assertions is changed.
## Summary
Port comprehension tests from Rust to Markdown
I don' think the remaining tests in `infer.rs` should be ported to
Markdown, maybe except for the incremental-checking tests when (if ever)
we have support for that in the MD tests.
closes#13696
## Summary
- Port "deferred annotations" unit tests to Markdown
- Port `implicit_global_in_function` unit test to Markdown
- Removed `resolve_method` and `local_inference` unit tests. These seem
like relics from a time where type inference was in it's early stages.
There is no way that these tests would fail today without lots of other
things going wrong as well.
part of #13696
based on #15683
## Test Plan
New MD tests for existing Rust unit tests.
## Summary
- Add feature to specify a custom typeshed from within Markdown-based
tests
- Port "builtins" unit tests from `infer.rs` to Markdown tests, part of
#13696
## Test Plan
- Tests for the custom typeshed feature
- New Markdown tests for deleted Rust unit tests
## Summary
Somehow, I managed to crash the `mdtest` runner today. I struggled to
reproduce this again to see if it's actually fixed (even with an
artificial `sleep` between the two `cargo test` invocations), but the
original backtrace clearly showed that this is where the problem
originated from. And it seems like a clear TOCTOU problem.
## Summary
Raise "invalid-assignment" diagnostics for incorrect assignments to
attributes, for example:
```py
class C:
var: str = "a"
C.var = 1 # error: "Object of type `Literal[1]` is not assignable to `str`"
```
closes#15456
## Test Plan
- Updated test assertions
- New test for assignments to module-attributes
## Summary
This PR generalizes some of the logic we have in `Type::is_subtype_of`
and `Type::is_disjoint_from` so that we fallback to the instance type of
the metaclass more often in `Type::ClassLiteral` and `Type::SubclassOf`
branches. This simplifies the code (we end up with one less branch in
`is_subtype_of`, and we can remove a helper method that's no longer
used), makes the code more robust (any fixes made to subtyping or
disjointness of instance types will automatically improve our
understanding of subtyping/disjointness for class-literal types and
`type[]` types) and more elegantly expresses the type-system invariants
encoded in these branches.
## Test Plan
No new tests added (it's a pure refactor, adding no new functionality).
All existing tests pass, however, including the property tests.
The AST generator creates a reference enum for each syntax group — an
enum where each variant contains a reference to the relevant syntax
node. Previously you could customize the name of the reference enum for
a group — primarily because there was an existing `ExpressionRef` type
that wouldn't have lined up with the auto-derived name `ExprRef`. This
follow-up PR is a simple search/replace to switch over to the
auto-derived name, so that we can remove this customization point.
## Summary
Test executables usually write failure messages (including panics) to
stdout, but I just managed to make a mdtest crash with
```
thread 'mdtest__unary_not' has overflowed its stack
fatal runtime error: stack overflow
```
which is printed to stderr. This test simply appends stderr to stdout
(`stderr=subprocess.STDOUT` can not be used with `capture_output`)
## Test Plan
Make sure that the error message is now visible in the output of `uv -q
run crates/red_knot_python_semantic/mdtest.py`
## Summary
The `Options` struct is intended to capture the user's configuration
options but
`EnvironmentOptions::venv_path` supports both a `SitePackages::Known`
and `SitePackages::Derived`.
Users should only be able to provide `SitePackages::Derived`—they
specify a path to a venv, and Red Knot derives the path to the
site-packages directory. We'll only use the `Known` variant once we
automatically discover the Python installation.
That's why this PR changes `EnvironmentOptions::venv_path` from
`Option<SitePackages>` to `Option<SystemPathBuf>`.
This requires making some changes to the file watcher test, and I
decided to use `extra_paths` over venv path
because our venv validation is annoyingly correct -- making mocking a
venv rather involved.
## Test Plan
`cargo test`
## Summary
As more and more tests move to Markdown, running the mdtest suite
becomes one of the most common tasks for developers working on Red Knot.
There are a few pain points when doing so, however:
- The `build.rs` script enforces recompilation (~five seconds) whenever
something changes in the `resource/mdtest` folder. This is strictly
necessary, because whenever files are added or removed, the test harness
needs to be updated. But this is very rarely the case! The most common
scenario is that a Markdown file has *changed*, and in this case, no
recompilation is necessary. It is currently not possible to distinguish
these two cases using `cargo::rerun-if-changed`. One can work around
this by running the test executable manually, but it requires finding
the path to the correct `mdtest-<random-hash>` executable.
- All Markdown tests are run by default. This is needed whenever Rust
code changes, but while working on the tests themselves, it is often
much more convenient to only run the tests for a single file. This can
be done by using a `mdtest__path_to_file` filter, but this needs to be
manually spelled out or copied from the test output.
- `cargo`s test output for a failing Markdown test is often
unnecessarily verbose. Unless there is an *actual* panic somewhere in
the code, mdtests usually fail with the explicit *"Some tests failed"*
panic in the mdtest suite. But in those cases, we are not interested in
the pointer to the source of this panic, but only in the mdtest suite
output.
This PR adds a Markdown test runner tool that attempts to make the
developer experience better.
Once it is started using
```bash
uv run -q crates/red_knot_python_semantic/mdtest.py
```
it will first recompile the tests once (if cargo requires it), find the
path to the `mdtest` executable, and then enter into a mode where it
watches for changes in the `red_knot_python_semantic` crate. Whenever …
* … a Markdown file changes, it will rerun the mdtest for this specific
file automatically (no recompilation!).
* … a Markdown file is added, it will recompile the tests and then run
the mdtest for the new file
* … Rust code is changed, it will recompile the tests and run all of
them
The tool also trims down `cargo test` output and only shows the actual
mdtest errors.
The tool will certainly require a few more iterations before it becomes
mature, but I'm curious to hear if there is any interest for something
like this.
## Test Plan
- Tested the new runner under various scenarios.
---------
Co-authored-by: Alex Waygood <Alex.Waygood@Gmail.com>
## Summary
Rename two functions with outdated names (they used to return `Type`s):
* `bindings_ty` => `symbol_from_bindings` (returns `Symbol`)
* `declarations_ty` => `symbol_from_declarations` (returns a
`SymbolAndQualifiers` result)
I chose `symbol_from_*` instead of `*_symbol` as I found the previous
name quite confusing. Especially since `binding_ty` and `declaration_ty`
also exist (singular).
## Test Plan
—
---------
Co-authored-by: Alex Waygood <Alex.Waygood@Gmail.com>
## Summary
Add support for `typing.ClassVar`, i.e. emit a diagnostic in this
scenario:
```py
from typing import ClassVar
class C:
x: ClassVar[int] = 1
c = C()
c.x = 3 # error: "Cannot assign to pure class variable `x` from an instance of type `C`"
```
## Test Plan
- New tests for the `typing.ClassVar` qualifier
- Fixed one TODO in `attributes.md`
## Summary
This is a small, tentative step towards the bigger goal of understanding
instance attributes.
- Adds partial support for pure instance variables declared in the class
body, i.e. this case:
```py
class C:
variable1: str = "a"
variable2 = "b"
reveal_type(C().variable1) # str
reveal_type(C().variable2) # Unknown | Literal["b"]
```
- Adds `property` as a known class to query for `@property` decorators
- Splits up various `@Todo(instance attributes)` cases into
sub-categories.
## Test Plan
Modified existing MD tests.
## Summary
This PR adds support for configuring Red Knot in the `tool.knot` section
of the project's
`pyproject.toml` section. Options specified on the CLI precede the
options in the configuration file.
This PR only supports the `environment` and the `src.root` options for
now.
Other options will be added as separate PRs.
There are also a few concerns that I intentionally ignored as part of
this PR:
* Handling of relative paths: We need to anchor paths relative to the
current working directory (CLI), or the project (`pyproject.toml` or
`knot.toml`)
* Tracking the source of a value. Diagnostics would benefit from knowing
from which configuration a value comes so that we can point the user to
the right configuration file (or CLI) if the configuration is invalid.
* Schema generation and there's a lot more; see
https://github.com/astral-sh/ruff/issues/15491
This PR changes the default for first party codes: Our existing default
was to only add the project root. Now, Red Knot adds the project root
and `src` (if such a directory exists).
Theoretically, we'd have to add a file watcher event that changes the
first-party search paths if a user later creates a `src` directory. I
think this is pretty uncommon, which is why I ignored the complexity for
now but I can be persuaded to handle it if it's considered important.
Part of https://github.com/astral-sh/ruff/issues/15491
## Test Plan
Existing tests, new file watching test demonstrating that changing the
python version and platform is correctly reflected.
## Summary
Closes https://github.com/astral-sh/ruff/issues/15508
For any two instance types `T` and `S`, we know they are disjoint if
either `T` is final and `T` is not a subclass of `S` or `S` is final and
`S` is not a subclass of `T`.
Correspondingly, for any two types `type[T]` and `S` where `S` is an
instance type, `type[T]` can be said to be disjoint from `S` if `S` is
disjoint from `U`, where `U` is the type that represents all instances
of `T`'s metaclass.
And a heterogeneous tuple type can be said to be disjoint from an
instance type if the instance type is disjoint from `tuple` (a type
representing all instances of the `tuple` class at runtime).
## Test Plan
- A new mdtest added. Most of our `is_disjoint_from()` tests are not
written as mdtests just yet, but it's pretty hard to test some of these
edge cases from a Rust unit test!
- Ran `QUICKCHECK_TESTS=1000000 cargo test --release -p
red_knot_python_semantic -- --ignored types::property_tests::stable`
---------
Co-authored-by: Carl Meyer <carl@astral.sh>
