Add a warn-by-default lint for items inside functions or expressions that implement methods or traits that are visible outside the function or expression. Consider ramping that lint to deny-by-default for Rust 2024, and evaluating a hard error for 2027.

Motivation

Currently, tools cross-referencing uses and definitions (such as IDEs) must either search inside all function bodies and other expression-containing items to find potential definitions corresponding to uses within another function, or not cross-reference those definitions at all.

Humans cross-referencing such uses and definitions may find themselves similarly baffled by code such as the following:

trait Trait<T> {
    fn method(&self) {}
}

struct Foo;

fn _foo() {
    struct Bar;
    impl Trait<Bar> for Foo {}
}

fn main() {
    Foo.method();
}

This change helps humans limit the scope of their search and avoid looking for definitions inside other functions or items, without missing any relevant definitions. If in the future we manage to forbid it entirely within a subsequent Rust edtion, tools will be able to rely on this as well.

Explanation

An "expression-containing item" is defined as any expression where an item may be defined. For example:

Functions
Closures
The values assigned to static items or non-anonymous const items.
The discriminant values assigned to enum variants

As an exception, anonymous const items are excluded from this definition for the purposes of this RFC (see unresolved-questions), as they're currently commonly used to define items within macros.

Rust will emit a warn-by-default lint when encountering an impl nested inside an expression-containing item (through any level of nesting), unless any of the following are true:

For impl T or impl TraitPath for T, the definition of type T is also nested inside the same expression-containing item.
For impl T<X, Y, ..> or impl TraitPath for T<X, Y, ..>, the definition of at least one of T, X, Y, .. is also nested inside the same expression- containing item.
For impl Trait for TypePath or impl Trait<X, Y, ..> for TypePath, the definition of Trait is also nested inside the same expression-containing item.
For impl Trait<X, Y, ..> for TypePath, the definition of at least one of X, Y, .. is also nested inside the same expression-containing item, and the full path Trait<X, Y, ..> may not be inferred from outside the expression-containing item, and further that such inferences on Trait would not be permitted even in the absense of all impls nested in expression- containing items.

For the purposes of the above rules, fundamental types such as &X and &mut X are considered parameterized types (T<X>).

Rust will emit a warn-by-default lint when encountering an exported macro (e.g. using #[macro_export]) nested inside an expression-containing item (through any level of nesting).

In a future edition, we may consider making this lint deny-by-default, or eventually making it a hard error. We'll evaluate the impact on the ecosystem and existing use cases before doing so.

The lint is considered to attach to the impl token of an impl block, or the macro_rules! token of a macro definition.

Drawbacks

Some existing code makes use of this pattern, and would need to migrate to a different pattern. In particular, this pattern may occur in macro-generated code, or in code generated by tools like rustdoc. Making this change would require such code and tools to restructure to meet this requirement.

Prior art

Other aspects of Rust's design attempt to enable local reasoning and avoid global reasoning, including non-inference of function signatures, and not having the function body affect non-opaque properties of impl Trait uses in the signature without reflecting those properties in the signature.

Unresolved questions

Should we flag these definitions in anonymous const items as well? This is used in some macro expansions for compatibility reasons.

Is the last rule regarding parameterized trait impl items viable to implement?

Future possibilities