RFC 0003: Attribute usage lint

compiler (attributes)

Summary

Rust currently has an attribute usage lint but it does not work particularly well. This RFC proposes a new implementation strategy that should make it significantly more useful.

Motivation

The current implementation has two major issues:

#[deriving(Clone)]; // Shouldn't have put a ; here
struct Foo;

#[ignore(attribute-usage)] // Should have used #[allow(attribute-usage)] instead!
mod bar {
    //...
}
#[feature(phase)];
#[phase(syntax, link)]
extern crate some_orm;

#[ormify]
pub struct Foo {
    #[column(foo_)]
    #[primary_key]
    foo: int
}

Detailed design

The current implementation is implemented as a simple fold over the AST, comparing attributes against a whitelist. Crate-level attributes use a separate whitelist, but no other distinctions are made.

This RFC would change the implementation to actually track which attributes are used during the compilation process. syntax::ast::Attribute_ would be modified to add an ID field:

pub struct AttrId(uint);

pub struct Attribute_ {
    id: AttrId,
    style: AttrStyle,
    value: @MetaItem,
    is_sugared_doc: bool,
}

syntax::ast::parse::ParseSess will generate new AttrIds on demand. I believe that attributes will only be created during parsing and expansion, and the ParseSess is accessible in both.

The AttrIds will be used to create a side table of used attributes. This will most likely be a thread local to make it easily accessible during all stages of compilation by calling a function in syntax::attr:

fn mark_used(attr: &Attribute) { }

The attribute-usage lint would run at the end of compilation and warn on all attributes whose ID does not appear in the side table.

One interesting edge case is attributes like doc that are used, but not in the normal compilation process. There could either be a separate fold pass to mark all doc attributes as used or doc could simply be whitelisted in the attribute-usage lint.

Attributes in code that has been eliminated with #[cfg()] will not be linted, but I feel that this is consistent with the way #[cfg()] works in general (e.g. the code won't be type-checked either).

Alternatives

An alternative would be to rewrite rustc::middle::lint to robustly check that attributes are used where they're supposed to be. This will be fairly complex and be prone to failure if/when more nodes are added to the AST. This also doesn't solve motivation #2, which would require externally loaded lint support.

Unresolved questions