Rustdoc treats primitive disambiguators incorrectly

[jyn514]

I tried this code:

//! [str::trim()]

I expected to see this happen: Rustdoc links to https://doc.rust-lang.org/nightly/std/primitive.str.html#method.trim.

Instead, this happened: Rustdoc gets confused that () means a function and thinks that the link is a primitive.

error: incompatible link kind for `str::trim`
   |
   |
57 | /// Using the [`str::trim()`] method ensures that no whitespace remains before parsing.
   |                ^^^^^^^^^^^^^ help: to link to the builtin type, prefix with `prim@`: ``prim@str::trim``
   |
   = note: `-D broken-intra-doc-links` implied by `-D warnings`
   = note: this link resolved to a builtin type, which is not a function

Meta

rustdoc --version:

rustdoc 1.50.0-beta.1 (05b602367 2020-12-29)

Relevant code:

rust/src/librustdoc/passes/collect_intra_doc_links.rs

Line 1181 in a6bd524

report_mismatch(other, Disambiguator::Primitive);

I think the fix is just to add if self.kind_side_channel.is_none() as a match guard.

[jyn514]

This is blocking #75807 and made #77875 (comment) unnecessarily difficult.

[max-heller]

@rustbot claim

[max-heller]

@jyn514 which arm do you mean to add the match guard to?

[jyn514]

@max-heller the one for telling if the resolved item is a primitive:

rust/src/librustdoc/passes/collect_intra_doc_links.rs

Line 1176 in a6bd524

Res::Primitive(_) => match disambiguator {

[max-heller]

@jyn514 It looks like self.kind_side_channel is in fact None for str::trim()

[jyn514]

Oh hmm - the issue is that for primitives, rustdoc generates the anchor manually:

rust/src/librustdoc/passes/collect_intra_doc_links.rs

Lines 401 to 402 in a6bd524

	Res::Primitive(prim_ty),
	Some(format!("{}#{}.{}", prim_ty.as_str(), out, item_str)),

. I would try storing the def_id of the associated item in kind_side_channel instead, so you know later whether the anchor is a manual link (str#method.trim) or whether it was generated by rustdoc.

[max-heller]

Adding self.kind_side_channel.set(Some((item.kind.as_def_kind(), item.def_id))); in here

rust/src/librustdoc/passes/collect_intra_doc_links.rs

Lines 394 to 398 in a6bd524

	.map(|item| match item.kind {
	ty::AssocKind::Fn => "method",
	ty::AssocKind::Const => "associatedconstant",
	ty::AssocKind::Type => "associatedtype",
	})

and adding the match guard seems to fix things, but also feels like a hack.

[jyn514]

I mean, all of kind_side_channel is a hack because other parts of rustdoc break when you store the associated item instead: 0c99d80. If you can figure out a way to get rid of it, I'd love to do that, but it will require some pretty major refactors I think.