internal: Teach cargo about cfg(rust_analyzer)

[lnicola]

r? @Urgau is this a good idea?, CC @Veykril

[Veykril]

Is this so we can remove the cfg from bootstrap again?

[lnicola]

Maybe, but I'm not too keen on trying it right now.

See https://blog.rust-lang.org/inside-rust/2024/01/03/this-development-cycle-in-cargo-1-76.html#-zcheck-cfg, I don't know if it will be enabled by default in the future.

[Veykril]

I see, well fine with me either way.

[lnicola]

warning: proc-macro-srv@0.0.0: cargo:rustc-check-cfg requires -Zcheck-cfg flag

This is a bit annoying, I guess it will have to wait.

[Urgau]

This looks alright, but as noted it won't work if cargo -Zcheck-cfg isn't used.

See https://blog.rust-lang.org/inside-rust/2024/01/03/this-development-cycle-in-cargo-1-76.html#-zcheck-cfg, I don't know if it will be enabled by default in the future.

The goal is to have it enable by default (and therefore remove -Zcheck-cfg from cargo). We still still have somethings to iron out before stabilization can happen but I'm hoping it will be stabilized and enabled by default before mid-2024.

[Veykril]

Blocked on rust-lang/cargo#13571

[Urgau]

Blocked on rust-lang/cargo#13571

Note that since rust-lang/cargo#13438, Cargo no longer emits a warning when using cargo::rustc-check-cfg in stable/beta (>=1.77). The instruction will just be silently ignored.

[Veykril]

@bors r+

[bors]

📌 Commit 3a2ae64 has been approved by Veykril

It is now in the queue for this repository.

[lnicola]

Not sure why I added these, but.. 😅.

[bors]

⌛ Testing commit 3a2ae64 with merge af1fd88...

[bors]

☀️ Test successful - checks-actions
Approved by: Veykril
Pushing af1fd88 to master...

[epage]

To double check, is cfg(rust_analyzer) something used exclusively within rust-analyzer for its own development or does rust-analyzer always set it when it runs so code can detect when its being built by rust-analyzer to do things differently?

If its the latter, that sounds like a "well known cfg" and should likely be defined in rustc along with miri and others or cargo along with docsrs.

[Veykril]

rust-analyzer sets this for workspace crates of a loaded project when doing its own analysis, but not when invoking cargo or any other tool. It's strictly contained to its analysis only.

[Urgau]

I was under the impression that the rust_analyzer cfg was a internal cfg (due to being used for internal things), but now looking at the code I'm no longer certain, it seems to be unconditionally added to every Cargo project:

rust-analyzer/crates/project-model/src/workspace.rs

Lines 1002 to 1006 in af1fd88

	// Add test cfg for local crates
	if cargo[pkg].is_local {
	cfg_options.insert_atom("test".into());
	cfg_options.insert_atom("rust_analyzer".into());
	}

and also every "detached" Rust files

rust-analyzer/crates/project-model/src/workspace.rs

Line 1183 in af1fd88

cfg_options.insert_atom("rust_analyzer".into());

@Veykril Is my analysis right, that the rust_analyzer cfg is added to every compilation unit by rust analyzer? If yes, is that cfg considered for public use (in the sense that users can use it)?

EDIT: raced with @Veykril 😄

[Veykril]

Did that answer your question already? 😄 Effectively to cargo, the cfg will always be disabled (unless the user somehow passes it to cargo themselves), the cfgs linked in the code there are the ones that r-a sees, but not any invoked tooling.

[Urgau]

Did that answer your question already?

I think so. From what I understand during rust-analyzer analysis of the code r-a adds (an analysis-only) rust_analyzer cfg, which is never given to any tooling.

And then there is a second rust_analyzer cfg, the one used for the proc-macro-srv crates, which is "real" but for internal use only.

If my understanding is right, then every is good, rust_analyzer is not a well-known cfg.

[Veykril]

It is not well known as it's fairly recent, and mainly an escape hatch to be able to cfg out things that are very slow for the ide (like the quote crate).

And then there is a second rust_analyzer cfg, the one used for the proc-macro-srv crates, which is "real" but for internal use only.

There it's also only used by analysis, its just that those crates are the only ones in the r-a codebase that actually have a cfg(rust-analyzer) in it.

[Urgau]

It is not well known as it's fairly recent

My bad, I was referring to well-known in the sense of --check-cfg well known names and values.

There it's also only used by analysis, its just that those crates are the only ones in the r-a codebase that actually have a cfg(rust-analyzer) in it.

This seems very similar to the miri and clippy cfgs, which are used to prevent the tool from going somewhere they don't handle (well).

So users can expect the rust_analyzer cfg to be present and set by r-a?

[Veykril]

So users can expect the rust_analyzer cfg to be present and set by r-a?

Yes but only on workspace member crates (so not dependencies in the crates-io registry), to prevent libraries from special casing rust-analyzer for the time being.

[Urgau]

If you don't want users to "special case rust-analyzer for the time being", does that also mean that you don't want to advertise the cfg ? and if so, are fine with users having to expect the cfg in a build.rs ?

More broadly the question is: should rustc know (and by extension advertise) the rust_analyzer cfg ?

[Veykril]

Having it be known is fine I think, it's not going away I'd say. The restriction is only there to deter people from doing weird r-a specific things in their libraries for now