RUSTC_WRAPPER is applied even after being unset

[jplatte]

Problem

Recently, I ran cargo bloat --time on one a Cargo project. Afterwards, I noticed there are now json-time {...} output lines when running other cargo commands (check for example). It looks like these are how cargo-bloat collects the compilation times (https://github.com/RazrFalcon/cargo-bloat/blob/ddbb5f55088eb7e28c0cc80e7a1fdfb1310c19b5/src/main.rs#L399), through a RUSTC_WRAPPER executable.

I have verified that at the time I run cargo check, no RUSTC_WRAPPER is set so I'm assuming it's being cached by Cargo somewhere.

Steps

1. Install cargo-bloat
2. Run cargo bloat --time on some Cargo project
3. Run cargo check

Notes

Tested with both cargo 1.53.0-beta (4369396ce 2021-04-27) and cargo 1.54.0-nightly (0cecbd673 2021-06-01).

[gbaranski]

Same issue, solved with cargo clean in project where cargo bloat --time has been used.

Not sure if uninstalling cargo-bloat and removing anything related to it from ~/.cargo helped anything, but that might be worth mentioning in caes cargo clean didn't help.

[weihanglo]

Cargo indeed caches the stderr of rustc invocation and then replays if it's a fresh re-compilation.

cargo/src/cargo/core/compiler/mod.rs

Lines 178 to 186 in df8cda0

	let work = replay_output_cache(
	unit.pkg.package_id(),
	PathBuf::from(unit.pkg.manifest_path()),
	&unit.target,
	cx.files().message_cache_path(unit),
	cx.bcx.build_config.message_format,
	cx.bcx.config.shell().err_supports_color(),
	unit.show_warnings(bcx.config),
	);

In addition, lines do not look like JSON at start are also cached:

cargo/src/cargo/core/compiler/mod.rs

Lines 1302 to 1305 in df8cda0

	if !line.starts_with('{') {
	state.stderr(line.to_string())?;
	return Ok(true);
	}

In general, rustc should emit stderr in JSON, but I doubt that we can blindly disable caching on those not starting with a {.

[jplatte]

Cargo indeed caches the stderr of rustc invocation and then replays if it's a fresh re-compilation.

Shouldn't RUSTC_WRAPPER changing invalidate all cached rustc invocation output though?

[weihanglo]

It seems that fingerprint calculation does not include RUSTC_WRAPPER change. I don't know if there is a reason, but there are cases that cargo-fix and clippy leverage the other wrapper env RUSTC_WORKSPACE_WRAPPER. Invalidating all rustc invocation for that may lead to full recompilation.

[jplatte]

It seems like RUSTC_WRAPPERs should be able to decide whether they get the same cache as other invocations...

Alternatively, maybe cargo-bloat --time could be updated to set a different cache directory (and delete it after it is done), instead of cleaning the regular target dir as I think it does now. Would that solve this issue? cc @RazrFalcon

[RazrFalcon]

Yes, cargo-bloat has to be updated. I just didn't had time to do it.

[jplatte]

Alright, so is RUSTC_WRAPPER emitting extra (non-JSON) output just not meant to happen and this issue can be closed?

[RazrFalcon]

I still think this is a Cargo issue.