rustdoc: Get rid of Item::Attributes

[jyn514]

This is a special case of #76382. Attributes is 72 bytes large and used in quite a few places:

• Item::attrs
• FnDecl::attrs
• ExternalCrate::attrs

This has two downsides:

1. Attributes takes up a lot of space in memory, so this bloats memory usage
2. It means attributes have to calculated ahead of time, instead of on-demand. In particular, this requires adding a new attrs field whenever render needs to look at the attrs of a specific clean type (Add stability tags to ImportItem. #83900 (comment)).

It would be better to calculate this on-demand, so it uses less memory and doesn't have to be calculated for each and every item.

[jyn514]

Steps to implement this:

• Change all the fields on Attributes to functions on AttributesExt.
  • These are calculated in from_ast, so you'll have to pass in some context. I wonder if reporting a diagnostic there is actually necessary - I'd expect rustc to verify that doc(cfg) has valid syntax, if it doesn't already, we could move it into check_doc_attrs:

    rust/compiler/rustc_passes/src/check_attr.rs

    Line 596 in b0c818c

    fn check_doc_attrs(&self, attr: &Attribute, hir_id: HirId, target: Target) -> bool {

  • The hard thing will be getting rid of additional_attrs - this is used for inlining:

    rust/src/librustdoc/clean/inline.rs

    Line 300 in b021bee

    Attributes::from_ast(diag, old_attrs, Some((inner, new_id)))

    . I wonder if it makes sense to have an inlined_attrs field on Cache - that would let render calculate all this on-demand without having to come up with some crazy scheme to update the attrs in the main TyCtxt. Hopefully a simple HashMap<DefId, (Attributes, /*parent_module: */ DefId)> mapping will be enough; I don't know if we ever need to deal with attributes for something without a DefId.
• Change all functions on Attributes to functions on AttributesExt; this is the easy part.

[tdelabro]

So the new Attributes fields should be diagnostic, attrs and additional_attrs, aka the arguements required by a call to from_ast ?

[jyn514]

There should be no diagnostic at all - I have a suspicion it's not doing anything currently, but in case it is, those checks should be moved out of rustdoc and into rustc.

Hmm, keeping Attributes but only storing attrs and additional_attrs is an interesting idea. I would be ok with that too - my original plan was to get rid of Attributes altogether, but that would be harder than holding on to additional_attrs.

[tdelabro]

@rustbot claim

[tdelabro]

I will give it a try. I'm not familiar with the codebase but keeping a lightened Attributes and compute its fields on the fly doesn't seem to hard of a refacto.

I will message you on Zulip if I have any questions.