## Summary
Another salsa upgrade.
The main motivation is to stay on a recent salsa version because there
are still a lot of breaking changes happening.
The most significant changes in this update:
* Salsa no longer derives `Debug` by default. It now requires
`interned(debug)` (or similar)
* This version ships the foundation for garbage collecting interned
values. However, this comes at the cost that queries now track which
interned values they created (or read). The micro benchmarks in the
salsa repo showed a significant perf regression. Will see if this also
visible in our benchmarks.
## Test Plan
`cargo test`
## Summary
This PR closes#16248.
If the return type of the function isn't assignable to the one
specified, an `invalid-return-type` error occurs.
I thought it would be better to report this as a different kind of error
than the `invalid-assignment` error, so I defined this as a new error.
## Test Plan
All type inconsistencies in the test cases have been replaced with
appropriate ones.
---------
Co-authored-by: Carl Meyer <carl@astral.sh>
## Summary
* Attributes/method are now properly looked up on metaclasses, when
called on class objects
* We properly distinguish between data descriptors and non-data
descriptors (but we do not yet support them in store-context, i.e.
`obj.data_descr = …`)
* The descriptor protocol is now implemented in a single unified place
for instances, classes and dunder-calls. Unions and possibly-unbound
symbols are supported in all possible stages of the process by creating
union types as results.
* In general, the handling of "possibly-unbound" symbols has been
improved in a lot of places: meta-class attributes, attributes,
descriptors with possibly-unbound `__get__` methods, instance
attributes, …
* We keep track of type qualifiers in a lot more places. I anticipate
that this will be useful if we import e.g. `Final` symbols from other
modules (see relevant change to typing spec:
https://github.com/python/typing/pull/1937).
* Detection and special-casing of the `typing.Protocol` special form in
order to avoid lots of changes in the test suite due to new `@Todo`
types when looking up attributes on builtin types which have `Protocol`
in their MRO. We previously
looked up attributes in a wrong way, which is why this didn't come up
before.
closes#16367closes#15966
## Context
The way attribute lookup in `Type::member` worked before was simply
wrong (mostly my own fault). The whole instance-attribute lookup should
probably never have been integrated into `Type::member`. And the
`Type::static_member` function that I introduced in my last descriptor
PR was the wrong abstraction. It's kind of fascinating how far this
approach took us, but I am pretty confident that the new approach
proposed here is what we need to model this correctly.
There are three key pieces that are required to implement attribute
lookups:
- **`Type::class_member`**/**`Type::find_in_mro`**: The
`Type::find_in_mro` method that can look up attributes on class bodies
(and corresponding bases). This is a partial function on types, as it
can not be called on instance types like`Type::Instance(…)` or
`Type::IntLiteral(…)`. For this reason, we usually call it through
`Type::class_member`, which is essentially just
`type.to_meta_type().find_in_mro(…)` plus union/intersection handling.
- **`Type::instance_member`**: This new function is basically the
type-level equivalent to `obj.__dict__[name]` when called on
`Type::Instance(…)`. We use this to discover instance attributes such as
those that we see as declarations on class bodies or as (annotated)
assignments to `self.attr` in methods of a class.
- The implementation of the descriptor protocol. It works slightly
different for instances and for class objects, but it can be described
by the general framework:
- Call `type.class_member("attribute")` to look up "attribute" in the
MRO of the meta type of `type`. Call the resulting `Symbol` `meta_attr`
(even if it's unbound).
- Use `meta_attr.class_member("__get__")` to look up `__get__` on the
*meta type* of `meta_attr`. Call it with `__get__(meta_attr, self,
self.to_meta_type())`. If this fails (either the lookup or the call),
just proceed with `meta_attr`. Otherwise, replace `meta_attr` in the
following with the return type of `__get__`. In this step, we also probe
if a `__set__` or `__delete__` method exists and store it in
`meta_attr_kind` (can be either "data descriptor" or "normal attribute
or non-data descriptor").
- Compute a `fallback` type.
- For instances, we use `self.instance_member("attribute")`
- For class objects, we use `class_attr =
self.find_in_mro("attribute")`, and then try to invoke the descriptor
protocol on `class_attr`, i.e. we look up `__get__` on the meta type of
`class_attr` and call it with `__get__(class_attr, None, self)`. This
additional invocation of the descriptor protocol on the fallback type is
one major asymmetry in the otherwise universal descriptor protocol
implementation.
- Finally, we look at `meta_attr`, `meta_attr_kind` and `fallback`, and
handle various cases of (possible) unboundness of these symbols.
- If `meta_attr` is bound and a data descriptor, just return `meta_attr`
- If `meta_attr` is not a data descriptor, and `fallback` is bound, just
return `fallback`
- If `meta_attr` is not a data descriptor, and `fallback` is unbound,
return `meta_attr`
- Return unions of these three possibilities for partially-bound
symbols.
This allows us to handle class objects and instances within the same
framework. There is a minor additional detail where for instances, we do
not allow the fallback type (the instance attribute) to completely
shadow the non-data descriptor. We do this because we (currently) don't
want to pretend that we can statically infer that an instance attribute
is always set.
Dunder method calls can also be embedded into this framework. The only
thing that changes is that *there is no fallback type*. If a dunder
method is called on an instance, we do not fall back to instance
variables. If a dunder method is called on a class object, we only look
it up on the meta class, never on the class itself.
## Test Plan
New Markdown tests.
We now resolve references in "eager" scopes correctly — using the
bindings and declarations that are visible at the point where the eager
scope is created, not the "public" type of the symbol (typically the
bindings visible at the end of the scope).
---------
Co-authored-by: Alex Waygood <alex.waygood@gmail.com>
## Summary
Transition to using coarse-grained tracked structs (depends on
https://github.com/salsa-rs/salsa/pull/657). For now, this PR doesn't
add any `#[tracked]` fields, meaning that any changes cause the entire
struct to be invalidated. It also changes `AstNodeRef` to be
compared/hashed by pointer address, instead of performing a deep AST
comparison.
## Test Plan
This yields a 10-15% improvement on my machine (though weirdly some runs
were 5-10% without being flagged as inconsistent by criterion, is there
some non-determinism involved?). It's possible that some of this is
unrelated, I'll try applying the patch to the current salsa version to
make sure.
---------
Co-authored-by: Micha Reiser <micha@reiser.io>
## Summary
Add support for implicitly-defined instance attributes, i.e. support
type inference for cases like this:
```py
class C:
def __init__(self) -> None:
self.x: int = 1
self.y = None
reveal_type(C().x) # int
reveal_type(C().y) # Unknown | None
```
## Benchmarks
Codspeed reports no change in a cold-cache benchmark, and a -1%
regression in the incremental benchmark. On `black`'s `src` folder, I
don't see a statistically significant difference between the branches:
| Command | Mean [ms] | Min [ms] | Max [ms] | Relative |
|:---|---:|---:|---:|---:|
| `./red_knot_main check --project /home/shark/black/src` | 133.7 ± 9.5 | 126.7 | 164.7 | 1.01 ± 0.08 |
| `./red_knot_feature check --project /home/shark/black/src` | 132.2 ± 5.1 | 118.1 | 140.9 | 1.00 |
## Test Plan
Updated and new Markdown tests
## Summary
Add support for (non-generic) type aliases. The main motivation behind
this was to get rid of panics involving expressions in (generic) type
aliases. But it turned out the best way to fix it was to implement
(partial) support for type aliases.
```py
type IntOrStr = int | str
reveal_type(IntOrStr) # revealed: typing.TypeAliasType
reveal_type(IntOrStr.__name__) # revealed: Literal["IntOrStr"]
x: IntOrStr = 1
reveal_type(x) # revealed: Literal[1]
def f() -> None:
reveal_type(x) # revealed: int | str
```
## Test Plan
- Updated corpus test allow list to reflect that we don't panic anymore.
- Added Markdown-based test for type aliases (`type_alias.md`)
Add support for declared types to the semantic index. This involves a
lot of renaming to clarify the distinction between bindings and
declarations. The Definition (or more specifically, the DefinitionKind)
becomes responsible for determining which definitions are bindings,
which are declarations, and which are both, and the symbol table
building is refactored a bit so that the `IS_BOUND` (renamed from
`IS_DEFINED` for consistent terminology) flag is always set when a
binding is added, rather than being set separately (and requiring us to
ensure it is set properly).
The `SymbolState` is split into two parts, `SymbolBindings` and
`SymbolDeclarations`, because we need to store live bindings for every
declaration and live declarations for every binding; the split lets us
do this without storing more than we need.
The massive doc comment in `use_def.rs` is updated to reflect bindings
vs declarations.
The `UseDefMap` gains some new APIs which are allow-unused for now,
since this PR doesn't yet update type inference to take declarations
into account.
Add support for non-local name lookups.
There's one TODO around annotated assignments without a RHS; these need
a fair amount of attention, which they'll get in an upcoming PR about
declared vs inferred types.
Fixes#11663
## Summary
This PR adds scope and definition for comprehension nodes. This includes
the following nodes:
* List comprehension
* Dictionary comprehension
* Set comprehension
* Generator expression
### Scope
Each expression here adds it's own scope with one caveat - the `iter`
expression of the first generator is part of the parent scope. For
example, in the following code snippet the `iter1` variable is evaluated
in the outer scope.
```py
[x for x in iter1]
```
> The iterable expression in the leftmost for clause is evaluated
directly in the enclosing scope and then passed as an argument to the
implicitly nested scope.
>
> Reference:
https://docs.python.org/3/reference/expressions.html#displays-for-lists-sets-and-dictionaries
There's another special case for assignment expressions:
> There is one special case: an assignment expression occurring in a
list, set or dict comprehension or in a generator expression (below
collectively referred to as “comprehensions”) binds the target in the
containing scope, honoring a nonlocal or global declaration for the
target in that scope, if one exists.
>
> Reference: https://peps.python.org/pep-0572/#scope-of-the-target
For example, in the following code snippet, the variables `a` and `b`
are available after the comprehension while `x` isn't:
```py
[a := 1 for x in range(2) if (b := 2)]
```
### Definition
Each comprehension node adds a single definition, the "target" variable
(`[_ for target in iter]`). This has been accounted for and a new
variant has been added to `DefinitionKind`.
### Type Inference
Currently, type inference is limited to a single scope. It doesn't
_enter_ in another scope to infer the types of the remaining expressions
of a node. To accommodate this, the type inference for a **scope**
requires new methods which _doesn't_ infer the type of the `iter`
expression of the leftmost outer generator (that's defined in the
enclosing scope).
The type inference for the scope region is split into two parts:
* `infer_generator_expression` (similarly for comprehensions) infers the
type of the `iter` expression of the leftmost outer generator
* `infer_generator_expression_scope` (similarly for comprehension)
infers the type of the remaining expressions except for the one
mentioned in the previous point
The type inference for the **definition** also needs to account for this
special case of leftmost generator. This is done by defining a `first`
boolean parameter which indicates whether this comprehension definition
occurs first in the enclosing expression.
## Test Plan
New test cases were added to validate multiple scenarios. Refer to the
documentation for each test case which explains what is being tested.
Extend red-knot type inference to cover all syntax, so that inferring
types for a scope gives all expressions a type. This means we can run
the red-knot semantic lint on all Python code without panics. It also
means we can infer types for `builtins.pyi` without panics.
To keep things simple, this PR intentionally doesn't add any new type
inference capabilities: the expanded coverage is all achieved with
`Type::Unknown`. But this puts the skeleton in place for adding better
inference of all these language features.
I also had to add basic Salsa cycle recovery (with just `Type::Unknown`
for now), because some `builtins.pyi` definitions are cyclic.
To test this, I added a comprehensive corpus of test snippets sourced
from Cinder under [MIT
license](https://github.com/facebookincubator/cinder/blob/cinder/3.10/cinderx/LICENSE),
which matches Ruff's license. I also added to this corpus some
additional snippets for newer language features: all the
`27_func_generic_*` and `73_class_generic_*` files, as well as
`20_lambda_default_arg.py`, and added a test which runs semantic-lint
over all these files. (The test doesn't assert the test-corpus files are
lint-free; just that they are able to lint without a panic.)
Per comments in https://github.com/astral-sh/ruff/pull/12269, "module
global" is kind of long, and arguably redundant.
I tried just using "module" but there were too many cases where I felt
this was ambiguous. I like the way "global" works out better, though it
does require an understanding that in Python "global" generally means
"module global" not "globally global" (though in a sense module globals
are also globally global since modules are singletons).
Support falling back to a global name lookup if a name isn't defined in
the local scope, in the cases where that is correct according to Python
semantics.
In class scopes, a name lookup checks the local namespace first, and if
the name isn't found there, looks it up in globals.
In function scopes (and type parameter scopes, which are function-like),
if a name has any definitions in the local scope, it is a local, and
accessing it when none of those definitions have executed yet just
results in an `UnboundLocalError`, it does not fall back to a global. If
the name does not have any definitions in the local scope, then it is an
implicit global.
Public symbol type lookups never include such a fall back. For example,
if a name is not defined in a class scope, it is not available as a
member on that class, even if a name lookup within the class scope would
have fallen back to a global lookup.
This PR makes the `@override` lint rule work again.
Not yet included/supported in this PR:
* Support for free variables / closures: a free symbol in a nested
function-like scope referring to a symbol in an outer function-like
scope.
* Support for `global` and `nonlocal` statements, which force a symbol
to be treated as global or nonlocal even if it has definitions in the
local scope.
* Module-global lookups should fall back to builtins if the name isn't
found in the module scope.
I would like to expose nicer APIs for the various kinds of symbols
(explicit global, implicit global, free, etc), but this will also wait
for a later PR, when more kinds of symbols are supported.
Implements definition-level type inference, with basic control flow
(only if statements and if expressions so far) in Salsa.
There are a couple key ideas here:
1) We can do type inference queries at any of three region
granularities: an entire scope, a single definition, or a single
expression. These are represented by the `InferenceRegion` enum, and the
entry points are the salsa queries `infer_scope_types`,
`infer_definition_types`, and `infer_expression_types`. Generally
per-scope will be used for scopes that we are directly checking and
per-definition will be used anytime we are looking up symbol types from
another module/scope. Per-expression should be uncommon: used only for
the RHS of an unpacking or multi-target assignment (to avoid
re-inferring the RHS once per symbol defined in the assignment) and for
test nodes in type narrowing (e.g. the `test` of an `If` node). All
three queries return a `TypeInference` with a map of types for all
definitions and expressions within their region. If you do e.g.
scope-level inference, when it hits a definition, or an
independently-inferable expression, it should use the relevant query
(which may already be cached) to get all types within the smaller
region. This avoids double-inferring smaller regions, even though larger
regions encompass smaller ones.
2) Instead of building a control-flow graph and lazily traversing it to
find definitions which reach a use of a name (which is O(n^2) in the
worst case), instead semantic indexing builds a use-def map, where every
use of a name knows which definitions can reach that use. We also no
longer track all definitions of a symbol in the symbol itself; instead
the use-def map also records which defs remain visible at the end of the
scope, and considers these the publicly-visible definitions of the
symbol (see below).
Major items left as TODOs in this PR, to be done in follow-up PRs:
1) Free/global references aren't supported yet (only lookup based on
definitions in current scope), which means the override-check example
doesn't currently work. This is the first thing I'll fix as follow-up to
this PR.
2) Control flow outside of if statements and expressions.
3) Type narrowing.
There are also some smaller relevant changes here:
1) Eliminate `Option` in the return type of member lookups; instead
always return `Type::Unbound` for a name we can't find. Also use
`Type::Unbound` for modules we can't resolve (not 100% sure about this
one yet.)
2) Eliminate the use of the terms "public" and "root" to refer to
module-global scope or symbols. Instead consistently use the term
"module-global". It's longer, but it's the clearest, and the most
consistent with typical Python terminology. In particular I don't like
"public" for this use because it has other implications around author
intent (is an underscore-prefixed module-global symbol "public"?). And
"root" is just not commonly used for this in Python.
3) Eliminate the `PublicSymbol` Salsa ingredient. Many non-module-global
symbols can also be seen from other scopes (e.g. by a free var in a
nested scope, or by class attribute access), and thus need to have a
"public type" (that is, the type not as seen from a particular use in
the control flow of the same scope, but the type as seen from some other
scope.) So all symbols need to have a "public type" (here I want to keep
the use of the term "public", unless someone has a better term to
suggest -- since it's "public type of a symbol" and not "public symbol"
the confusion with e.g. initial underscores is less of an issue.) At
least initially, I would like to try not having special handling for
module-global symbols vs other symbols.
4) Switch to using "definitions that reach end of scope" rather than
"all definitions" in determining the public type of a symbol. I'm
convinced that in general this is the right way to go. We may want to
refine this further in future for some free-variable cases, but it can
be changed purely by making changes to the building of the use-def map
(the `public_definitions` index in it), without affecting any other
code. One consequence of combining this with no control-flow support
(just last-definition-wins) is that some inference tests now give more
wrong-looking results; I left TODO comments on these tests to fix them
when control flow is added.
And some potential areas for consideration in the future:
1) Should `symbol_ty` be a Salsa query? This would require making all
symbols a Salsa ingredient, and tracking even more dependencies. But it
would save some repeated reconstruction of unions, for symbols with
multiple public definitions. For now I'm not making it a query, but open
to changing this in future with actual perf evidence that it's better.
Intern types using Salsa interning instead of in the `TypeInference`
result.
This eliminates the need for `TypingContext`, and also paves the way for
finer-grained type inference queries.
