[red-knot] Fallback to attributes on types.ModuleType if a symbol can't be found in locals or globals #13904

AlexWaygood · 2024-10-24T11:02:02Z

Summary

All modules in Python are instances of types.ModuleType. Various attributes on types.ModuleType in typeshed represent "implicit globals" that are part of the namespace of every module in Python, even if not explicitly defined. As such, before declaring that a symbol is unbound (and in fact, before looking the symbol up in the builtins scope) we should look the symbol up as an attribute on types.ModuleType. The only attributes on types.ModuleType that are not present as implicit globals are __dict__, __getattr__ and __init__.

Test Plan

Markdown test added as crates/red_knot_python_semantic/resources/mdtest/scopes/moduletype_attrs.md. We also get to remove a false-positive error from the tomllib benchmark 🎉

github-actions · 2024-10-24T11:15:35Z

`ruff-ecosystem` results

Linter (stable)

✅ ecosystem check detected no linter changes.

Linter (preview)

✅ ecosystem check detected no linter changes.

codspeed-hq · 2024-10-24T11:27:43Z

CodSpeed Performance Report

Merging #13904 will not alter performance

_{Comparing moduletype-attrs (eee3e80) with main (7dd0c7f)}

Summary

✅ 32 untouched benchmarks

crates/red_knot_python_semantic/src/types/infer.rs

carljm

Looks good, but I do think we need to resolve the regression here, if at all possible; this is too much of an edge-case to pay that much perf for.

crates/red_knot_python_semantic/src/types/infer.rs

carljm

Looks great! I assume you are still planning to try replacing the hard coded allowlist with a salsa query based on actual attributes found in typeshed, just noticed one minor terminology nit.

crates/red_knot_python_semantic/resources/mdtest/scopes/moduletype_attrs.md

AlexWaygood · 2024-10-27T19:16:19Z

Okay, I've switched from the hardcoded allowlist to a Salsa-cached generated allowlist! Locally I'm consistently seeing a regression of 1-2% on the incremental benchmark, but Codspeed reports the benchmarks are neutral on this PR, so maybe what I'm seeing locally is just noise.

In order to obtain a list of symbols that typeshed declares in the body of types.ModuleType, I had to add a new SymbolsFlag member, IS_DECLARED. We had ways of querying whether symbols in a SymbolTable had been defined in that scope, but no way of querying whether symbols in a SymbolTable had been declared in that scope -- and for a stub file, the latter is what's important.

I've also fixed several edge cases regarding accessing members as attributes on imported modules. In some cases, whether a module "defines" a member can vary depending on whether you access the member as a global variable from inside the module or as an attribute on the module object after it's been imported elsewhere.

TL;DR: ready for re-review!

MichaReiser

It seems unfortunate to me that we now have to track DECLARED for all symbols just for this one use case.

MichaReiser · 2024-10-28T07:04:54Z

crates/red_knot_python_semantic/src/semantic_index/symbol.rs

 }

 bitflags! {
    #[derive(Copy, Clone, Debug, Eq, PartialEq)]
    struct SymbolFlags: u8 {
        const IS_USED         = 1 << 0;
-        const IS_BOUND      = 1 << 1;
+        const IS_BOUND        = 1 << 1;
+        const IS_DECLARED     = 1 << 2;


Could you add some comment explaining IS_DECLARED's meaning?

I added a link to Carl's essay at the top of https://github.com/astral-sh/ruff/blob/main/crates/red_knot_python_semantic/src/semantic_index/use_def.rs

crates/red_knot_python_semantic/src/types.rs

AlexWaygood · 2024-10-28T11:47:09Z

It seems unfortunate to me that we now have to track DECLARED for all symbols just for this one use case.

I would wager that this information will come in useful in other contexts in the future (though I can't right now say what -- it's just an intuition I have).

However, one alternative way of tackling this would be to write a Python script that inspects the AST of the typeshed stub for types.ModuleType and generates a Rust file that contains a function that returns true if a given symbol is declared in the body of types.ModuleType. (This would be similar to the approach we take in https://github.com/astral-sh/ruff/blob/main/scripts/generate_known_standard_library.py.) We could run the script as part of every typeshed sync, to make sure that the generated function is always up to date with the typeshed stubs.

This would be a more complicated than the current solution. But it would have the following advantages:

We could probably use the typeshed_client library in the Python script, which means that the script itself probably wouldn't be that complicated
The generated Rust function would likely be more performant than the current solution that uses a Salsa query, because we wouldn't have to look anything up in the cache: the specific symbols that are declared on types.ModuleType would just be hardcoded into the generated Rust function.
We wouldn't have to track DECLARED for all symbols just for this one use case

I'm not sure whether this would be better or worse on net than the solution I currently have in this PR -- but it's an alternative to consider, anyway!

MichaReiser · 2024-10-28T12:01:20Z

I would wager that this information will come in useful in other contexts in the future (though I can't right now say what -- it's just an intuition I have).

Yeah. I'm not a fan of it but I don't consider it blocking. I guess an alternative to this would be to look at the class type and iterate over its members. I don't think we support it today, but I suspect that's a capability we want long-term.

carljm

Looks great! A few comments that I think are worth at least testing and/or mentioning in TODOs, even if we don't make code changes for them right now.

an alternative to this would be to look at the class type and iterate over its members. I don't think we support it today, but I suspect that's a capability we want long-term.

I agree that we'll want this, and I think the best / most efficient way to implement this would be to iterate over symbols in the symbol table that are defined/declared in the class, using IS_DEFINED and IS_DECLARED flags. So I would flip it around and say this is not an alternative approach, rather it's another likely future use case for this flag.

crates/red_knot_python_semantic/resources/mdtest/scopes/moduletype_attrs.md

carljm · 2024-10-28T21:09:52Z

crates/red_knot_python_semantic/resources/mdtest/scopes/moduletype_attrs.md

+# TODO: this should probably be added to typeshed; not sure why it isn't?
+# error: [unresolved-reference]
+# revealed: Unbound
+reveal_type(__doc__)


Probably because it's on builtins.object?

The awkward thing is that I think __doc__ might be the only attribute on object that is accessible in every module as a global name.

I think all of them except for __module__ are accessible on module objects as an attribute.

You'll know better than I whether it's likely we can get __doc__ added specifically to ModuleType (despite it already being on object) to help support fixing this TODO with fewer special cases, or whether we should instead just bite the bullet and handle it as a special case.

Yeah... I think it would be fine to add __doc__ to ModuleType in typeshed if we add a comment saying why we have the "redundant" override from object... I'll try to put up a typeshed PR today...

Filed python/typeshed#12918

carljm · 2024-10-28T21:13:07Z

crates/red_knot_python_semantic/resources/mdtest/scopes/moduletype_attrs.md

+reveal_type(__init__)
+```
+
+## Accessed as attributes


__class__ is another one defined on object that is accessible on ModuleType instances, but not as a module global name.

__module__ is an odd special case in that typeshed says it exists on object but it does not actually exist on module objects. Probably not worth modeling this LSP violation, it's fine if we just let it exist on all objects?

__module__ is an odd special case in that typeshed says it exists on object but it does not actually exist on module objects. Probably not worth modeling this LSP violation, it's fine if we just let it exist on all objects?

Sigh. __module__ really shouldn't be listed as an instance attribute on object; it's just obviously incorrect. But I'm not going to restart that conversation 🙃

I think this is not worth special-casing in red-knot; we should just accept it as an inaccuracy in type inference that we have to live with unless and until it's fixed in typeshed.

crates/red_knot_python_semantic/resources/mdtest/scopes/moduletype_attrs.md

…'t be found in locals or globals

AlexWaygood added the red-knot Multi-file analysis & type inference label Oct 24, 2024

AlexWaygood requested review from carljm and MichaReiser as code owners October 24, 2024 11:02

AlexWaygood marked this pull request as draft October 24, 2024 11:17

AlexWaygood force-pushed the moduletype-attrs branch 2 times, most recently from dcccb7a to 22fc347 Compare October 24, 2024 11:22

AlexWaygood commented Oct 24, 2024

View reviewed changes

crates/red_knot_python_semantic/src/types/infer.rs Outdated Show resolved Hide resolved

AlexWaygood mentioned this pull request Oct 24, 2024

[red-knot] Infer Todo, not Unknown, for PEP-604 unions in annotations #13908

Merged

AlexWaygood force-pushed the moduletype-attrs branch from 22fc347 to 7dabd5c Compare October 24, 2024 12:40

AlexWaygood changed the base branch from main to alex/pep604-todo October 24, 2024 12:41

AlexWaygood marked this pull request as ready for review October 24, 2024 12:41

carljm requested changes Oct 24, 2024

View reviewed changes

crates/red_knot_python_semantic/src/types/infer.rs Outdated Show resolved Hide resolved

crates/red_knot_python_semantic/src/types/infer.rs Outdated Show resolved Hide resolved

Base automatically changed from alex/pep604-todo to main October 25, 2024 18:21

AlexWaygood force-pushed the moduletype-attrs branch 2 times, most recently from 3739088 to ae8ab30 Compare October 26, 2024 16:19

carljm reviewed Oct 26, 2024

View reviewed changes

crates/red_knot_python_semantic/resources/mdtest/scopes/moduletype_attrs.md Outdated Show resolved Hide resolved

AlexWaygood force-pushed the moduletype-attrs branch 4 times, most recently from 2f14b09 to abb4375 Compare October 27, 2024 19:04

AlexWaygood requested a review from carljm October 27, 2024 19:16

MichaReiser reviewed Oct 28, 2024

View reviewed changes

AlexWaygood force-pushed the moduletype-attrs branch from 1e40cec to 9ff6f66 Compare October 28, 2024 11:29

carljm approved these changes Oct 28, 2024

View reviewed changes

[red-knot] Fallback to attributes on types.ModuleType if a symbol can…

b758655

…'t be found in locals or globals

AlexWaygood added 6 commits October 29, 2024 10:25

feedback

7ceac1a

Make it work for attributes on modules as well

939c9c5

hardcode an allowlist instead of a denylist

dfcd9a6

fix edge cases to do with accessing members as attributes

bc7cc49

Use a Salsa query

0a40501

feedback from Micha

eee3e80

AlexWaygood force-pushed the moduletype-attrs branch from 9ff6f66 to eee3e80 Compare October 29, 2024 10:47

AlexWaygood merged commit d2c9f5e into main Oct 29, 2024
20 checks passed

AlexWaygood deleted the moduletype-attrs branch October 29, 2024 10:59

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

[red-knot] Fallback to attributes on types.ModuleType if a symbol can't be found in locals or globals #13904

[red-knot] Fallback to attributes on types.ModuleType if a symbol can't be found in locals or globals #13904

AlexWaygood commented Oct 24, 2024 •

edited by carljm

Loading

github-actions bot commented Oct 24, 2024 •

edited

Loading

codspeed-hq bot commented Oct 24, 2024 •

edited

Loading

carljm left a comment

carljm left a comment

AlexWaygood commented Oct 27, 2024 •

edited

Loading

MichaReiser left a comment •

edited

Loading

MichaReiser Oct 28, 2024

AlexWaygood Oct 28, 2024

AlexWaygood commented Oct 28, 2024

MichaReiser commented Oct 28, 2024

carljm left a comment

carljm Oct 28, 2024

AlexWaygood Oct 29, 2024

AlexWaygood Oct 29, 2024

carljm Oct 28, 2024

AlexWaygood Oct 29, 2024

[red-knot] Fallback to attributes on types.ModuleType if a symbol can't be found in locals or globals #13904

[red-knot] Fallback to attributes on types.ModuleType if a symbol can't be found in locals or globals #13904

Conversation

AlexWaygood commented Oct 24, 2024 • edited by carljm Loading

Summary

Test Plan

github-actions bot commented Oct 24, 2024 • edited Loading

ruff-ecosystem results

Linter (stable)

Linter (preview)

codspeed-hq bot commented Oct 24, 2024 • edited Loading

Merging #13904 will not alter performance

Summary

carljm left a comment

Choose a reason for hiding this comment

carljm left a comment

Choose a reason for hiding this comment

AlexWaygood commented Oct 27, 2024 • edited Loading

MichaReiser left a comment • edited Loading

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

AlexWaygood commented Oct 28, 2024

MichaReiser commented Oct 28, 2024

carljm left a comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

AlexWaygood commented Oct 24, 2024 •

edited by carljm

Loading

github-actions bot commented Oct 24, 2024 •

edited

Loading

`ruff-ecosystem` results

codspeed-hq bot commented Oct 24, 2024 •

edited

Loading

AlexWaygood commented Oct 27, 2024 •

edited

Loading

MichaReiser left a comment •

edited

Loading