x.py test --stage 0 src/tools/clippy does not work

[jyn514]

$ x.py test --stage 0 src/tools/clippy 
Updating only changed submodules
Submodules updated in 0.01 seconds
    Finished dev [unoptimized + debuginfo] target(s) in 0.09s
Building stage0 std artifacts (x86_64-unknown-linux-gnu -> x86_64-unknown-linux-gnu)
    Finished release [optimized] target(s) in 0.09s
Copying stage0 std from stage0 (x86_64-unknown-linux-gnu -> x86_64-unknown-linux-gnu / x86_64-unknown-linux-gnu)
Building stage0 compiler artifacts (x86_64-unknown-linux-gnu -> x86_64-unknown-linux-gnu)
    Finished release [optimized] target(s) in 0.13s
Copying stage0 rustc from stage0 (x86_64-unknown-linux-gnu -> x86_64-unknown-linux-gnu / x86_64-unknown-linux-gnu)
Building stage0 tool clippy-driver (x86_64-unknown-linux-gnu)
   Compiling clippy v0.0.212 (/home/joshua/rustc/src/tools/clippy)
    Finished release [optimized] target(s) in 1.54s
   Compiling clippy v0.0.212 (/home/joshua/rustc/src/tools/clippy)
    Finished release [optimized] target(s) in 1.49s
     Running build/x86_64-unknown-linux-gnu/stage0-tools/x86_64-unknown-linux-gnu/release/deps/clippy_driver-c0ddca9b2ddf2dba

running 1 test
test test_arg_value ... ok

test result: ok. 1 passed; 0 failed; 0 ignored; 0 measured; 0 filtered out

     Running build/x86_64-unknown-linux-gnu/stage0-tools/x86_64-unknown-linux-gnu/release/deps/compile_test-93629616c1644211

running 1 test
test compile_test ... FAILED

failures:

---- compile_test stdout ----
thread 'compile_test' panicked at 'Found multiple rlibs for crate `serde`: `"/home/joshua/rustc/build/x86_64-unknown-linux-gnu/stage0-tools/x86_64-unknown-linux-gnu/release/deps/libserde-0cbb27a77c795012.rlib"` and `"/home/joshua/rustc/build/x86_64-unknown-linux-gnu/stage0-tools/x86_64-unknown-linux-gnu/release/deps/libserde-d748cd52d3612bbb.rlib"', src/tools/clippy/tests/compile-test.rs:46:25
note: run with `RUST_BACKTRACE=1` environment variable to display a backtrace

Meta

3478d7c

[jyn514]

The really strange thing here is that clippy is built twice:

Building stage0 tool clippy-driver (x86_64-unknown-linux-gnu)
   Compiling clippy v0.0.212 (/home/joshua/rustc/src/tools/clippy)
    Finished release [optimized] target(s) in 1.54s
   Compiling clippy v0.0.212 (/home/joshua/rustc/src/tools/clippy)
    Finished release [optimized] target(s) in 1.49s
     Running build/x86_64-unknown-linux-gnu/stage0-tools/x86_64-unknown-linux-gnu/release/deps/clippy_driver-c0ddca9b2ddf2dba

[jyn514]

This does work fine with --stage 1.

[jyn514]

Ok, I dug quite a bit into this. The main issue is that clippy is trying to run tests against the sysroot of the host compiler, not the target compiler. In particular, test --stage 0 is failing because it's trying to use the .rlibs in stage0-sysroot, instead of stage1.

I fixed that. The problem now is that clippy wants to link the test files to its own dependencies. That doesn't work while bootstrapping, because its dependencies were built by a different toolchain than the currently running clippy (stage0, not stage1).

@rust-lang/clippy why does clippy need to use its own dependencies in tests? Is it possible to write the tests without those? I can fix this, but it will make --stage 0 essentially useless because it will build the compiler twice (once to build clippy itself, and once with the fresh toolchain to build its dependencies).

If it's not possible to do that, can we instead add some dummy package that only depends on TEST_DEPENDENCIES, so we don't have to rebuild the whole compiler?

[Manishearth]

why does clippy need to use its own dependencies in tests

Because clippy contains lints for working on clippy itself, and they need to be tested too

[Manishearth]

Also tbh my experience with rustc is that stage0 is basically always useless and should never be used, so it doesn't seem too off for stage 0 to be useless for working on clippy. Maybe we can just disallow this for clippy.

thread 'compile_test' panicked at 'Found multiple rlibs for crate `serde`: `"/home/joshua/rustc/build/x86_64-unknown-linux-gnu/stage0-tools/x86_64-unknown-linux-gnu/release/deps/libserde-0cbb27a77c795012.rlib"` and `"/home/joshua/rustc/build/x86_64-unknown-linux-gnu/stage0-tools/x86_64-unknown-linux-gnu/release/deps/libserde-d748cd52d3612bbb.rlib"', src/tools/clippy/tests/compile-test.rs:46:25
note: run with `RUST_BACKTRACE=1` environment variable to display a backtrace

worth noting, this error is kinda also because compiletest doesn't quite know what crates to link to, this happens pretty often in many different scenarios.

[jyn514]

@Manishearth the numbering for clippy is inconsistent with the numbering for the rest of the compiler. test --stage 0 builds the compiler once; test --stage 1 builds it twice. That's very painful when trying to iterate quickly.

That error up there is outdated, it was fixed by rust-lang/rust-clippy#7631. The new error is

error[E0514]: found crate `std` compiled by an incompatible version of rustc
  |
  = help: please recompile that crate using this compiler (rustc 1.61.0-beta.5 (cf5fd8947 2022-05-04)) (consider running `cargo clean` first)
  = note: the following crate versions were found:
          crate `std` compiled by rustc 1.62.0-nightly (e7575f967 2022-04-14): /home/callie/.rustup/toolchains/nightly-x86_64-unknown-linux-gnu/lib/rustlib/x86_64-unknown-linux-gnu/lib/libstd_detect-939385b13e387309.rlib
          crate `std` compiled by rustc 1.62.0-nightly (e7575f967 2022-04-14): /home/callie/.rustup/toolchains/nightly-x86_64-unknown-linux-gnu/lib/rustlib/x86_64-unknown-linux-gnu/lib/libstd-443e20b28eeed844.rlib

or something similar - I don't have the exact output on hand, but it's related to stage0 vs stage1 sysroot.

[jyn514]

If it's not possible to do that, can we instead add some dummy package that only depends on TEST_DEPENDENCIES, so we don't have to rebuild the whole compiler?

This is hard to do, because cargo only allows one library per package, and doesn't allow building binaries without first building the library :( I guess we could hard-code the dependencies clippy uses in another package somewhere? nested workspaces are still a pipe dream unfortunately ...

[RalfJung]

Also tbh my experience with rustc is that stage0 is basically always useless and should never be used

FWIW Miri tests work fine in stage 0, and generally there are very few things that require building the compiler twice these days -- testing clippy is one of them. This makes clippy failures more painful to deal with than most other issues one runs into when working on rustc.