`OUT_DIR` variable not set

[Razican]

Some Travis jobs are not compiling due to OUT_DIR not being present on nightly on my build.rs script. Is this an expected behaviour? how could it be fixed?

Example failure: https://travis-ci.org/SUPERAndroidAnalyzer/super/jobs/181113222#L146

[joshtriplett]

I can reproduce this with "cargo install ripgrep" using a cargo built from latest cargo git. Debugging...

[joshtriplett]

Looks like commit 41579ba caused this issue.

[upsuper]

This breaks building servo/rust-bindgen as well.

[JIghtuse]

cargo even has failing test for me with the same symptoms: http://paste.fedoraproject.org/500936/48110771/

Can't build ripgrep too.

[SimonSapin]

I’ve managed to build bindgen by replacing env! in the build script with std::env::var: rust-lang/rust-bindgen#323

[SimonSapin]

Still, it looks like this is a breaking change. So Cargo should probably revert to providing these variables when compiling a build script. At least for a while, with a warning that this will change.

[joshtriplett]

It looks like commit 95193ca clarified the documentation to state that Cargo doesn't set these environment variables when building the build script, only when running it, even though Cargo did actually set them when building the build script as well. A later commit in Cargo then made the behavior match the documentation.

This does indeed represent a breaking change from the existing behavior, and existing build.rs scripts relied on the existing behavior.

[alexcrichton]

Ah this is definitely an accidental regression! Yes env!("OUT_DIR") is likely not what you want but rather std::env::var is correct in build scripts. The env! version is the out dir that the build script was built with, and the env::var version is for when the build script is being run.

I'm kinda surprised that env! ever worked...

[alexcrichton]

Well... so this was a "regression" but also a bug fix at the same time. The OUT_DIR environment variable can be different when the build script is being compiled compared to when the build script is being run. Notably during cross compilation these two can be different. This means that crates which use env!("OUT_DIR") I believe are broken for cross compiles where crates which use env::var("OUT_DIR") will correctly work with cross compiles.

To those hitting this issue, could you confirm that the crate hasn't been used for cross compiles? And perhaps if it has have weird issues arisen?

We could add a fix for this but I'm tempted to avoid doing so if we can (as it's "incorrect" to rely on the old behavior). If it's possible to fixup the packages in question quickly that'd be best, but if there's ongoing pain then I can change Cargo back to the previous behavior.

[joshtriplett]

@alexcrichton I've seen various crates get fixed for this change (which, at a minimum, needs documentation in release notes).

It'd be good to be able to build existing packages with new Cargo. For the most part, Cargo has maintained backward compatibility across many versions. On the other hand, I do agree that crates shouldn't rely on the old behavior, and it would indeed have broken cross-compiles.

I don't see an obvious way to work-but-warn in this case, without awful hacks (such as somehow substituting in a version of env! that checks the environment variable name and produces a warning).

[alexcrichton]

Yes my only worry is if crates can't be updated. Crates can use std::env::var and work across all versions of Cargo, so if a crate can be proactively updated I'm not too worried about it.