Regression in pwd of cargo doc

[dtolnay]

For one of my projects I build and publish docs of the master branch from CI. I use the command RUSTDOC=./rustdoc cargo doc with a script that looks like:

#!/bin/sh

exec rustdoc "$@" --html-after-content script.js

This works as of stable rustc 1.23.0. But with rustc 1.24.0-beta.3 it no longer works because the RUSTDOC is resolved relative to /path/to/registry/src/github.com-$hash/$crate-$version instead of relative to the directory from which I ran cargo doc. This is different from how RUSTC=./foo cargo build is resolved so I believe it is a bug.

[dtolnay]

Mentioning some rustdoc contributors: @QuietMisdreavus @GuillaumeGomez

[QuietMisdreavus]

😕 If it's a problem of resolving the RUSTDOC environment variable, it seems more like Cargo being weird? @rust-lang/cargo What do y'all think?

[Mark-Simulacrum]

This looks like it's probably caused by rust-lang/cargo@8647a87, and as such probably intentional (cc @alexcrichton).

However, the behavior where a cargo doc invoked in the "root" package doesn't use that packages workspace root does seem rather odd, so perhaps we can make some changes there, or resolve RUSTDOC relative to the current working directory first, by calling canonicalize. It would be logical to do the same for RUSTC, then.

[dtolnay]

resolve RUSTDOC relative to the current working directory first

Haven't read through that Cargo commit, but this is what I would consider the correct behavior. From what I can tell, this is how it works for RUSTC.

Note that there are two relevant directory issues and I maybe should have mentioned the second one above:

• Directory where RUSTDOC is resolved
• Directory that the script given by RUSTDOC sees as its current directory

My shell script refers to script.js which exists in the directory from which cargo doc was invoked, and that used to work as of rustc 1.23.0. As of 1.24.0-beta.3 the script no longer sees that directory as its current directory.

[alexcrichton]

Ok so there's two things going on here I think. As @Mark-Simulacrum pointed out this is caused by rust-lang/cargo#4788, so strictly speaking it's a bug in Cargo, not rustc. In any case:

• It's a bug in Cargo that RUSTDOC=./foo doesn't work. That should be resolved relative to the invoking cwd and should be used as an absolute path from then on out.
• It's "expected breakage" that the cwd is changing during compiles which would break the original script here @dtolnay is it possible to use an absolute path to script.js instead?

[wycats]

@alexcrichton can we fix the Cargo bug?

[alexcrichton]

@wycats the bug about RUSTDOC=./foo certainly, the "bug" about changing cwd is much less clear.

[Mark-Simulacrum]

@rust-lang/cargo We need to discuss and resolve the relative rustdoc path issue RUSTDOC=./foo soon. The other discussion (cwd changing) is I believe expected, though may also be worth discussing. Nominating and setting P-high.