cargo test with a proc_macro tries to run a host test on the target

[tkilbourn]

When I cross-compile the attached crate's "proc-macro-test-derive" subcrate, cargo attempts to run tests compiled for the host. (This crate attempts to mirror the layout of the rand and rand-derive crates, which is how I discovered the issue.)

Setup

Running rustc 1.19.0 (0ade33941 2017-07-17), cargo 0.20.0 (a60d185 2017-07-13)

1. Install a cross-compiler (e.g., on ubuntu trusty: apt-get install libc6-dev-arm64-cross gcc-aarch64-linux-gnu)
2. Install qemu (ubuntu trusty: apt-get install qemu-user-static)
3. Set a linker and runner for the target in ~/.cargo/config:

[target.aarch64-unknown-linux-gnu]
linker = "aarch64-linux-gnu-gcc"
runner = "qemu-aarch64-static -L /usr/aarch64-linux-gnu"

Repro

cd proc-macro-test-derive
cargo test --target aarch64-unknown-linux-gnu

This will attempt to run target/debug/deps/proc_macro_test_derive-1f6a3cff88feb964, which is compiled for the host rather than the target. When cargo test is run without a --target, this (empty) test succeeds and the intended test target/debug/deps/macros-659b3b1ec96599d4 is run.

For --target aarch64-unknown-linux-gnu, I expect target/aarch64-unknown-linux-gnu/debug/macros-da045a5e3c0e63ff to be run.

proc-macro-test.zip

[alexcrichton]

Ah yeah I think, like doctests, we need to just disable unit tests on proc-macro crates for cross-compiled situations.

[rtsuk]

It it possible to run these tests on the host even when cross-compiling? Otherwise it seems like for crates that are only ever cross compiled these sorts of tests would never be run.

[alexcrichton]

Hm true! Although I'd imagine that's covered by cargo test, right?

[rtsuk]

If you can invoke cargo test on the host. For magenta-rs I don't think that will work, as the host won't have the shared library needed to link. One could certainly work around it with config so that your crate can compile and run some tests on the host, but if it isn't hard to tell when a test is intended for the host even when cross compiling it seems better to not use that custom runner for those tests.

[alexcrichton]

Ah sorry I think I misinterpreted!