docs.rs reuses crate-specific resources from `empty_library` for other crates

[coastalwhite]

Docs.rs shows 'empty_library' in the sidebar and has no search results other than empty_library.

Original bug report

When publishing a no_std crate without dependencies, it will add an empty_library to the crates on [docs.rs].

Reproduction steps:

1. Publish a no_std library without dependencies.
2. The docs.rs page will now have a empty_library in the crates.

This seems like a rather awkward way to tell the reader the crate has no dependencies. Now if you click any deeper into the documentation, on the left (under the crates header) the empty_library will be a clickable link that leads to a dead link.

This is what I mean: Link to CRC8-RS Docs

[Nemo157]

Not specific to no-std libraries, seems to have started happening within the last month or so, either a docs.rs, rustwide or rustdoc change. Skimming the docs.rs merge requests none stand out to me as potentially causing this.

[Nemo157]

More specifically it appears to have changed between

rustc 1.52.0-nightly (45b3c2851 2021-03-04) docsrs 0.6.0 (da1f362 2021-03-05)

and

rustc 1.52.0-nightly (caca2121f 2021-03-05) docsrs 0.6.0 (918ee5d 2021-03-06)

(from comparing https://docs.rs/syn/1.0.62/syn/ and https://docs.rs/syn/1.0.61/syn/)

docs.rs changes: da1f362...918ee5d
rustdoc changes: rust-lang/rust@45b3c28...caca212

[Nemo157]

To confirm I went back through the release pages and found the exact builds which changed behaviour, https://docs.rs/near-runtime/4.0.0-pre.1/near_runtime/ was the last build that didn't show empty_library, and https://docs.rs/mangadex-api/0.2.4/mangadex_api/ was the first build that showed it (currently on page https://docs.rs/releases/recent/59, but will be moving further back as more builds happen). These were both with docs.rs da1f362, but the rustc version changed.

[coastalwhite]

Would this then be a rustdoc issue, and should I open one over there?

I am not getting this behaviour, when just normally compiling documentation on my local machine. It could be that there are some flags of rustdoc that are bugged.

[Nemo157]

It seems like it's a change in rustdoc behaviour, but it is docs.rs specific, we use empty_library as a dummy library. It might be that we need to change something here to workaround whatever changed rather than changing rustdoc. (I skimmed the PRs, and couldn't see anything that would have changed rustdoc's behaviour in that release 🤷).

docs.rs/src/docbuilder/rustwide_builder.rs

Line 59 in 65e946c

const DUMMY_CRATE_NAME: &str = "empty-library";

[jyn514]

I'd expect this to be a regression from rust-lang/rust#80845, cc @GuillaumeGomez. Do we maybe keep the artifacts from the empty_library crate around when documenting other crates?

[Nemo157]

Huh, weird that that PR doesn't appear in the compare view though? Looks like that would be in the 2021-03-06/07 nightly, not the 2021-03-05 one.

[jyn514]

Oh oops I was looking at the wrong dates. These are the changes between the toolchains:

$ git log 45b3c2851..caca2121f --oneline src/librustdoc/
8fd946c63a6 Auto merge of #82795 - m-ou-se:rollup-uzx0b92, r=m-ou-se
ad919154557 Remove unused code from main.js
2f28361936f Rollup merge of #82088 - Nicholas-Baron:shorten_html_render, r=GuillaumeGomez
1683cb12e44 Use cache access levels
e4287992a7b No more need for borrow call
ad30c39918c Pass TyCtxt directly instead of DocContext in librustdoc::visit_ast::inherits_doc_hidden
186f13914a8 Move visibility check inside the should_have_doc_example function
4b30625094e Don't warn for `missing_doc_examples` when item is #[doc(hidden)]
afb8220d0af Corrected imports for render tests and mod files
fd14e386123 Moved `write_shared` to its own file
14983b9812e Moved the `make_item_keywords` function to `context.rs` as it is only used there
6c7d7a6bf4b Moved `print_item` and helpers to a separate file
48167c499aa Moved Context and its impls to a separate file

It would be helpful if someone could bisect them.

[GuillaumeGomez]

Checking what's wrong.

[GuillaumeGomez]

It's an interesting "bug" from rustdoc indeed. Sending a fix.

[GuillaumeGomez]

Actually no, it's not a bug from rustdoc. Here how it works for rustdoc: crates.js list the crates from search-index.js. Meaning that at some point, the file isn't empty as it should. The problem here is that the rustdoc update wasn't fully integrated into docs.rs.

[GuillaumeGomez]

So here, I see two possibilities:

1. We remove crates.js from the folder
2. We don't generate crates.js when the --disable-per-crate-search argument is passed in rustdoc.

Which solution do you prefer?