`#[cfg_attr(..., path = "submod_gated.rs")] mod submod;` makes RA recognize submod_gated.rs, discarding cfg_attr and submod.rs

[zjp-CN]

rust-analyzer version: rust-analyzer 1.80.0-nightly (7c52d2d 2024-06-03)

rustc version: rustc 1.80.0-nightly (7c52d2db6 2024-06-03)

editor or extension: neovim

relevant settings: None

repository link (if public, optional): None. But I found the problem from embassy-executor. For reproducible code, see below.

code snippet to reproduce:

// src/main.rs
fn main() {
    submod::f();
}

#[cfg_attr(feature = "gated", path = "submod_gated.rs")]
mod submod;

// src/submod.rs
pub fn f() {
    println!("from submod.rs");
}

// src/submod_gated.rs
pub fn f() {
    println!("from submod_gated.rs");
}

# Cargo.toml
[features]
gated = []

Currently, RA marks submod.rs as file not included in crate hierarchy when gated feature is nowhere enabled, and treats submod_gated.rs as the default submod file, discarding cfg_attr.

[Veykril]

Hmm, the attributes are configured at that point, so the issue is probably in

rust-analyzer/crates/hir-expand/src/attrs.rs

Lines 127 to 156 in 0c795a7

	let is_cfg_attr =
	attr.path.as_ident().map_or(false, |name| *name == crate::name![cfg_attr]);
	if !is_cfg_attr {
	return smallvec![attr.clone()];
	}
	let subtree = match attr.token_tree_value() {
	Some(it) => it,
	_ => return smallvec![attr.clone()],
	};
	let (cfg, parts) = match parse_cfg_attr_input(subtree) {
	Some(it) => it,
	None => return smallvec![attr.clone()],
	};
	let index = attr.id;
	let attrs = parts.enumerate().take(1 << AttrId::CFG_ATTR_BITS).filter_map(
	|(idx, attr)| Attr::from_tt(db, attr, index.with_cfg_attr(idx)),
	);
	let cfg_options = &crate_graph[krate].cfg_options;
	let cfg = Subtree { delimiter: subtree.delimiter, token_trees: Box::from(cfg) };
	let cfg = CfgExpr::parse(&cfg);
	if cfg_options.check(&cfg) == Some(false) {
	smallvec![]
	} else {
	cov_mark::hit!(cfg_attr_active);
	attrs.collect()
	}

somewhere

[zjp-CN]

As a sidenote, I find embassy codebase uses conditional compilation a lot by combining cfg_attr and path attribute macro .

So a similar case worth considering is like this:

#[cfg(feature = "_arch")]
#[cfg_attr(feature = "arch-avr", path = "arch/avr.rs")]
#[cfg_attr(feature = "arch-cortex-m", path = "arch/cortex_m.rs")]
#[cfg_attr(feature = "arch-riscv32", path = "arch/riscv32.rs")]
#[cfg_attr(feature = "arch-std", path = "arch/std.rs")]
#[cfg_attr(feature = "arch-wasm", path = "arch/wasm.rs")]
mod arch;

RA mistakenly recognizes arch/avr.rs as arch module file, when arch-riscv32 feature is enabled in the RA configuration file.

[zjp-CN]

Wow, this pattern causes 1K+ errors in embassy-hal-internal without enabling any feature in it.

(I could try to solve this issue, but don't have the time until next month.)

Edit: wait, cfg_attr is not the evil, they are caused by cfg attributes instead 😮

[Veykril]

Hmm those are rustc errors though, not rust-analyzer? This sounds like your cargo check for rust-analyzer is setup wrongly

[Veykril]

Do you have "rust-analyzer.cargo.features" / "rust-analyzer.check.features" set to "all"?

[zjp-CN]

Do you have "rust-analyzer.cargo.features" / "rust-analyzer.check.features" set to "all"?

No, I never have them in configuration file. That's what surprises me.

[zjp-CN]

Weird, these cfg error doesn't exist on vscode. embassy has own vscode settings which I didn't use in neovim.

I have to double check on my RA settings.

[zjp-CN]

Well, there are some interesting things going on:

• it's true that cargo.allFeatures is enabled by neovim RA plugin mrcjkb/rustaceanvim
  • if I set "rust-analyzer.cargo.allFeatures": false, the problem in the original issue on #[cfg_attr(feature = "gated", path = "submod_gated.rs")] is gone. I believe the 1K+ errors in embassy-hal-internal will also disappear, though I haven't tested it.
  • note that the original issue doesn't actually exist: I've confirmed it because RA jumps to correct submod.rs since no all features are enable by default: Disable rust-analyzer.{cargo,checkOnSave}.allFeatures by default #4711
• but I accidentally find a bug in RA: cargo.allFeatures seems deprecated since Config revamp #12010 , but it's still working. What if we both have a setting like this?

{
    "rust-analyzer.cargo.features": [],
    "rust-analyzer.cargo.allFeatures": true,
}

It's valid, and RA only respects the second and causes the issue.

allFeatures is not mentioned in the manual page any more, can RA stops recognizing it?
If allFeatures has to be kept, I think the documentation needs improvement which I can help.

[Veykril]

can RA stops recognizing it?

I'd love to do that but some of the third party clients have complained about us fixing up the config so we had to still accept that setting. We could maybe try removing that compat layer now. Accepting that is just a backwards compat change.

The bigger issue here is that https://github.com/mrcjkb/rustaceanvim/tree/master has a very bad default for that config here. There is a reason why allFeatures is disabled by default because crates tend to use non-additive features.