RUSTFLAGS / RUSTDOCFLAGS are not being passed to rustdoc when target-applies-to-host nightly feature is used

[jplatte]

Problem

For two project I maintain, the docs generated by CI no longer includes an index file despite the appropriate nightly flags being used. I initially opened rust-lang/rust#97881 for this, but now found out that it's actually a bug in Cargo that doesn't pass RUSTDOCFLAGS on to rustdoc in the affected workspaces (though it doesn't seem to affect all workspaces).

Steps

• Clone https://github.com/ruma/ruma/tree/rustdoc-bug
• With the environment variable RUSTDOCFLAGS set to --cfg compile_error, run cargo +nightly doc -p ruma
• Observe no errors despite this in crates/ruma/src/lib.rs

  #[cfg(compile_error)]
  compile_error!("!!!");

Notes

• Reproducible with a recent version of cargo (1.63.0-nightly (85e457e15 2022-06-07))
• Reproducible with cargo 1.61.0-nightly (3d6970d 2022-02-28) (rustup nightly-2022-03-02)
• Not reproducible with cargo 1.61.0-nightly (d6cdde5 2022-02-22) (rustup nightly-2022-03-01).

Version

cargo 1.63.0-nightly (85e457e15 2022-06-07)
release: 1.63.0-nightly
commit-hash: 85e457e158db216a2938d51bc3b617a5a7fe6015
commit-date: 2022-06-07
host: x86_64-unknown-linux-gnu
libgit2: 1.4.2 (sys:0.14.2 vendored)
libcurl: 7.83.1-DEV (sys:0.4.55+curl-7.83.1 vendored ssl:OpenSSL/1.1.1n)
os: Arch Linux Rolling Release [64-bit]

[jplatte]

I just verified that this is definitely cargo not passing the flags, by looking at the --verbose output, and also trying to set RUSTFLAGS and RUSTDOCFLAGS to nonsensical values.

[jplatte]

I found out what is causing this / why we only saw it recently despite the bug being reproducible back to start of March: This only happens with

target-applies-to-host = false

[unstable]
target-applies-to-host = true

in the cargo configuration. Based on that and its timing, I suspect this bug was introduced in #10395; cc @jonhoo.

[jplatte]

Minimized reproduction: https://github.com/jplatte/rustflags-bug

[jplatte]

Okay, so this is definitely from #10395, had I read the comments there I would have known that this is expected behavior: #10395 (comment)

The way forward seems to be this: #10395 (comment)

[jonhoo]

See also #10462 for follow up in that direction.

[jplatte]

I'm not really clear on whether that would help for our main use case, which is

• we want to configure clippy through rustflags in .cargo/config.toml, because that's currently the best workaround to otherwise having no workspace-wide lint configuration
• we don't want changing rustflags for host dependencies to cause constant rebuilding when switching between non---target and --target builds