`./x.py test` for `thumb*-none-*` targets fails with "cargo must succeed"

[sekineh]

Updates (2018-07-18)

• We'll enable ./x.py test src/test/run-make for no_std targets.

Environment

• Ubuntu 16.04 (x86_64)
• ARM cross compiler is installed by sudo apt install gcc-arm-none-eabi libnewlib-arm-none-eabi

Problem

For cross target, we can't run the following tests:

# Both fails
./x.py test --target thumbv7m-none-eabi src/test/ui
./x.py test --target thumbv7m-none-eabi src/test/run-make

Typical error we got is cargo must succeed:

Caused by:
  process didn't exit successfully: `/home/sekineh/rustme/build/bootstrap/debug/rustc --crate-name term libterm/lib.rs --error-format json --crate-type dylib --crate-type rlib --emit=dep-info,link -C prefer-dynamic -C opt-level=2 -C metadata=b839a5796ea44350 -C extra-filename=-b839a5796ea44350 --out-dir /home/sekineh/rustme/build/x86_64-unknown-linux-gnu/stage2-test/thumbv7m-none-eabi/release/deps --target thumbv7m-none-eabi -L dependency=/home/sekineh/rustme/build/x86_64-unknown-linux-gnu/stage2-test/thumbv7m-none-eabi/release/deps -L dependency=/home/sekineh/rustme/build/x86_64-unknown-linux-gnu/stage2-test/release/deps` (exit code: 101)
command did not execute successfully: "/home/sekineh/rustme/build/x86_64-unknown-linux-gnu/stage0/bin/cargo" "build" "--target" "thumbv7m-none-eabi" "-j" "8" "--release" "--manifest-path" "/home/sekineh/rustme/src/libtest/Cargo.toml" "--message-format" "json"
expected success, got: exit code: 101
thread 'main' panicked at 'cargo must succeed', bootstrap/compile.rs:1091:9

The full log is posted on gist:

• https://gist.github.com/sekineh/8fdf934f42382bcd3c75de360da5a8f6

Notes

• I don't believe ui tests and run-make tests require a simulator to be executed. The whole tests are run on host only (host: Linux x86_64 in this case).

Background

We need to land CI for thumb* targets before Rust 2018 is released.
The fact the test harness cannot be run is the major blocker for the CI addition.

CC @jamesmunns @nerdyvaishali @kennytm

[jamesmunns]

Here's a little more background for reference:

What we are trying to do/Context

We would like to run tests for the thumbv7m-none-eabi target from a x86_64-unknown-linux-gnu host (on CI). At the moment, we are trying to get a "base case" of getting existing compiler-only tests to run in this configuration. This presents a problem, as the thumbv7m-none-eabi target is a bare metal/no_std target, and currently the CI fails while attempting to build the libtest crate as part of the "Building stage2 std artifacts (x86_64-unknown-linux-gnu -> thumbv7m-none-eabi)" test process. The testing process fails at this section of the bootstrap process in particular.

We are also working on adding a Dockerfile to support this CI testing.

In the future, we would like to add additional tests in this style - where we would like the CI to build a binary only for our target where we can perform static analysis (e.g. check the binary output for symbols, size, etc), or dynamic analysis (running the resulting binary in QEMU).

We have tried to model our testing after the existing CI tests based on the wasm32-unknown target, which is also a no_std target, however we are unable to understand why thumbv7m-none-eabi fails and wasm32-unknown testing fails.

What we need help with

At the moment, we are not sure whether this process should be failing here. Based on wasm32-unknown tests passing the stage where our tests fail, there seems to be some existing ability to run tests solely on the host system. Is the failure we are seeing a bug (and something needs to be added/fixed in configuration or the bootstrap testing), or is this expected behavior (and larger changes may be necessary to support this configuration)?

Support for this target in dist builds was recently added, however we are also unsure how to map these changes to a test build.

[kennytm]

The actual error from the gist is failure to build stage2-libtest for E0463 "can't find crate for std" 😕

Building stage2 test artifacts (x86_64-unknown-linux-gnu -> thumbv7m-none-eabi)
   Compiling term v0.0.0 (file:///home/sekineh/rustme/src/libterm)
   Compiling getopts v0.2.17
warning: dropping unsupported crate type `dylib` for target `thumbv7m-none-eabi`

error[E0463]: can't find crate for `std`
  |
  = note: the `thumbv7m-none-eabi` target may not be installed

error: aborting due to previous error

For more information about this error, try `rustc --explain E0463`.
error[E0463]: can't find crate for `std`
  |
  = note: the `thumbv7m-none-eabi` target may not be installed

error: aborting due to previous error

For more information about this error, try `rustc --explain E0463`.
error: Could not compile `getopts`.

[sekineh]

Yes, but the ‘std’ can be successfully compiled via ‘./x.py dist’ for that target.

[jamesmunns]

I wanted to expand on my previous comment, and perhaps provide a little more context/detail. Thanks to @kennytm for taking a look already :)

The issue we are seeing is that the test system x.py test ... is attempting to build the libtest crate for the target system, because that is the usual behavior of running x.py test (you want to run unit tests for/on your target).

There are two problems with that:

1. That's not what we want. For our (initial) testing, we aren't running unit tests, only looking at the binary compiled for the target. Because of this, libtest is not needed at all. This is important because...
2. It is not possible to build libtest for a thumb target. libtest has a dependency on std, and there is no std crate that can be downloaded or built. In general, it doesn't make sense to build unit tests for a microcontroller (or at least, it doesn't make sense for what we are trying to do at the moment) - see the bottom of this comment for more detail on std for thumb* targets.

Issue 2 describes the symptom we are seeing, where the build process is failing because there is no std crate for the thumb target. By trying to run x.py test for a thumb target, the build system assumes we want to run unit tests ON that target, which requires libtest is built for the target, which is not possible at the moment. However, that is not what we want.

So, in my opinion, there are two possible ways forward:

1. We could teach the build system that not all cases of x.py test means "build unit tests for the target, and run them natively on that target (or in a QEMU instance)". This would mean introducing some kind of configurable option that we could disable for thumb tests. wasm targets may already have this implemented, but I am not sure.
2. We could introduce a new verb, something like x.py host-test ..., which does what we would like it to do (build a compiler for the host, and inspect binaries built for the target, do not expect to run the produced binary).

I still don't have a good understanding of why this is an issue for us (e.g. thumb* targets, and not wasm targets, but I haven't had time to dig deeper into what is going on yet.

Another note about @sekineh's comment that dist works for this target: This is a bit of a misnomer. The std dist package that is produced by x.py for thumb* targets doesn't include std, but is instead only a packaged, distributable version of core and some other miscellaneous "under the hood" things. The nomenclature was used to fit within the current build and distribution procedure when @japaric introduced pre-built core for embedded targets (see here).

[jamesmunns]

@alexcrichton - From looking at the git blame, it looks like you implemented most of the wasm tests in Rust's CI. If you have a minute, would you mind reading through this thread and chiming in? If it wasn't you who did it, would you mind pointing us in the correct direction on who we can talk to about the wasm CI tests? It would be greatly appreciated :)

[japaric]

wasm32 is not a no_std target; it supports std -- I don't know about the test crate. (I believe some parts of std just call unimplemented! on wasm so std supports wasm in the "at least compiles" sense).

[jamesmunns]

Thanks @japaric, I'm not sure where I got that it was a no_std target. Did it used to be, and I'm just out of the loop?

I'll update my other comments at some point today to take out the details about wasm, though I think the rest of the comments should still be fine.

Perhaps it would just make sense for us to call a shell script that does something like this instead of trying to use x.py test:

# Build a compiler for us to use, just for the host
x.py build

git clone <some embedded example> && cd $_
<path to built cargo> build --target thumbv7m-none-eabi ...

# some procedure to validate binary produced
verify_bin <path to binary>

exit $?

[sekineh]

I'm not sure where I got that it was a no_std target

I'm fairly confident that that misconception came from my edit of ci-plan.md. The target wasm is not strictly a no_std, just no_std'ish in some sense.
https://github.com/jamesmunns/irr-embedded-2018/blame/master/ci-plan.md

So the conclusion is that x.py test does not support no_std targets currently and it won't support them in near future?

[kennytm]

Would it work if we remove this line:

rust/src/bootstrap/test.rs

Line 969 in 1fa76a4

builder.ensure(compile::Test { compiler, target });

? Or change to target: compiler.host?

[sekineh]

The short answer is 'Yes'.

If I commented out the line 969, run-make runs and succeeds, and ui fails. I think the behavior is understandable.
If I change the line to like below, both tests are at least run. (results are: run-make: success, ui: failure)

        // builder.ensure(compile::Test { compiler, target });
        builder.ensure(compile::Test { compiler, target: compiler.host });

Huge hint. A possible hack is to add if to that line, ...... for now.

[alexcrichton]

Ah yes, as @japaric mentioned the wasm target does include both libtest and libstd, so it's not quite the same. In general I don't think there's any way to get ./x.py test meaningfully working for no_std targets, 99% of our test suite assumes it links to std and "does something" (either run a binary or use items defined in std).

I'd recommend that if we wanted to add CI for this target that we'd write one-off tests in src/test/run-make. That'll allow you to define a fully custom script which only works for that one particular target. We may need to tweak rustbuild to not compile libtest for some targets (or otherwise make libtest a noop for the target), but that'll allow flavorful tests as well.

[sekineh]

Thanks, @japaric , @kennytm and @alexcrichton for the help!
@jamesmunns , our initial task is basically OK if run-make runs. Do you agree with me?

For the record, I ended up with this snippet:

        if !target.contains("-none-") {
            builder.ensure(compile::Test { compiler, target });
        }

And it runs run-make tests.

running 3 tests
iii
test result: ok. 0 passed; 0 failed; 3 ignored; 0 measured; 0 filtered out

	finished in 0.052
Build completed successfully in 0:00:03

I could add this line into our (embedded CI) pull request.

[kennytm]

If we only run run-make tests, directly if builder.no_std(target) that line out seems better than changing it to target: compiler.host.

(cc @Mark-Simulacrum)