RUSTDOC_LOG does not invalidate output cache

[jyn514]

Problem
When I set RUSTDOC_LOG=info after previously running cargo doc, any time I run cargo doc in the future will not have any output from rustdoc. The same problem is also present for RUSTC_LOG.

Steps

1. cargo doc on any cargo project
2. RUSTDOC_LOG=info cargo doc

The second run will have the same output as the first. For comparison, the expected output is something like this:

RUSTDOC_LOG example

$ RUSTDOC_LOG=rustdoc cargo doc 
 Documenting hexponent v0.3.0 (/home/joshua/src/rust/hexponent)
[INFO  rustdoc] starting to run rustc
[INFO  rustdoc::core] Executing passes
[INFO  rustdoc] finished with rustc
[INFO  rustdoc] going to format
[INFO  rustdoc::html::sources] emitting source files
[INFO  rustdoc::html::render] Recursing into /home/joshua/.local/lib/cargo/target/doc/hexponent
[INFO  rustdoc::html::render] Documenting hexponent
[INFO  rustdoc::html::render] Recursed; leaving /home/joshua/.local/lib/cargo/target/doc/hexponent
[INFO  rustdoc::html::render] Documenting FPFormat
[INFO  rustdoc::html::render] Documenting from_literal
[INFO  rustdoc::html::render] Recursing into /home/joshua/.local/lib/cargo/target/doc/hexponent/fpformat
[INFO  rustdoc::html::render] Recursed; leaving /home/joshua/.local/lib/cargo/target/doc/hexponent/fpformat
[INFO  rustdoc::html::render] Recursing into /home/joshua/.local/lib/cargo/target/doc/hexponent/parse_utils
[INFO  rustdoc::html::render] Recursed; leaving /home/joshua/.local/lib/cargo/target/doc/hexponent/parse_utils
[INFO  rustdoc::html::render] Documenting ParseErrorKind
[INFO  rustdoc::html::render] Documenting MissingPrefix
[INFO  rustdoc::html::render] Documenting MissingDigits
[INFO  rustdoc::html::render] Documenting MissingExponent
[INFO  rustdoc::html::render] Documenting ExponentOverflow
[INFO  rustdoc::html::render] Documenting MissingEnd
[INFO  rustdoc::html::render] Documenting ConversionResult
[INFO  rustdoc::html::render] Documenting Precise
[INFO  rustdoc::html::render] Documenting Imprecise
[INFO  rustdoc::html::render] Documenting FloatLiteral
[INFO  rustdoc::html::render] Documenting ParseError
[INFO  rustdoc::html::render] Documenting kind
[INFO  rustdoc::html::render] Documenting index
    Finished dev [unoptimized + debuginfo] target(s) in 0.63s

Possible Solution(s)

Invalidate cargo's output cache whenever RUSTDOC_LOG is changed.

Notes

Output of cargo version:

cargo 1.44.0 (05d080faa 2020-05-06) (but the problem was present on nightly last I checked a few days ago).

[jyn514]

It turns out rustdoc can fix this on its end using env_depinfo: https://rust-lang.zulipchat.com/#narrow/stream/257328-clippy/topic/CLIPPY_ARGS.20is.20not.20tracked.20by.20Cargo

@rustbot claim

[jyn514]

It turns out rustdoc can fix this on its end using env_depinfo: https://rust-lang.zulipchat.com/#narrow/stream/257328-clippy/topic/CLIPPY_ARGS.20is.20not.20tracked.20by.20Cargo

This still doesn't help because Cargo doesn't track dep-info from rustdoc - and in fact, it can't:

$ cargo +rustc2 rustdoc -q -- --emit=dep-info
error: Unrecognized option: 'emit'

Is it intentional that cargo doesn't do this? If I added support for --emit to rustdoc, would cargo use it?

[jyn514]

I guess another alternative is to change rustdoc to use RUSTC_LOG instead, which should inherit the depinfo from when rustc is run. But that seems not ideal.

[jyn514]

This still doesn't help because Cargo doesn't track dep-info from rustdoc - and in fact, it can't:

Actually it's even more than that - cargo only uses the dep-info from rustc for rustc itself, not for rustdoc 😕

$ cat span.d 
span.d: src/lib.rs src/span.rs

src/lib.rs:
src/span.rs:

# env-dep:RUSTC_LOG
# env-dep:RUSTC_LOG_COLOR
# env-dep:RUSTDOC_LOG
# env-dep:RUSTDOC_LOG_COLOR

$ cargo +rustc2 rustdoc
 Documenting span v0.1.0 (/home/joshua/test-rustdoc/span)
    Finished dev [unoptimized + debuginfo] target(s) in 0.58s
$ RUSTDOC_LOG=info cargo +rustc2 rustdoc
    Finished dev [unoptimized + debuginfo] target(s) in 0.00s

[jyn514]

@ehuss why did you assign me?

[ehuss]

Oh, sorry. I was debugging why the rustbot claim above #8374 (comment) didn't work.

[jyn514]

Ah ok - I no longer want to tackle this, the pr fixing it was closed: rust-lang/rust#82763.

[jyn514]

hmm, I just had another idea for how to fix this: right now it looks like cargo unconditionally caches all rustc output — could we change it to only cache lints instead? I think lints were the original motivation for adding caching.

to avoid confusion we could say something like "note: 50 lines of non-lint output were not cached"?