"./x run miri" does not work with "--target"

[RalfJung]

Summary

Miri can run things for arbitrary targets, so ./x run miri --target <foo> should just work. However, something does not seem to be wired up correctly:

$ ./x run miri --stage 1 --target aarch64-apple-darwin -- src/tools/miri/tests/pass/hello.rs
Building bootstrap
    Finished `dev` profile [unoptimized] target(s) in 0.07s

thread 'main' panicked at src/bootstrap/src/core/build_steps/compile.rs:1841:13:
assertion `left == right` failed: Cross-compiling is not yet supported with `download-rustc`
  left: x86_64-unknown-linux-gnu
 right: aarch64-apple-darwin

This shouldn't be cross-compiling anything so the error makes little sense. But okay, let's disable download-rustc:

$ ./x run miri --stage 1 --target aarch64-apple-darwin --set rust.download-rustc=false -- src/tools/miri/tests/pass/hello.rs
Building bootstrap
    Finished `dev` profile [unoptimized] target(s) in 0.11s

thread 'main' panicked at src/bootstrap/src/core/builder/mod.rs:1436:17:
It is not possible to build the standard library for `aarch64-apple-darwin` using the stage0 compiler.
You have to build a stage1 compiler for `x86_64-unknown-linux-gnu` first, and then use it to build a standard library for `aarch64-apple-darwin`.

At the same time, ./x test miri --target ... works fine.

I guess the problem here is that --target is ambiguous: it could mean "build miri for that target and then run it", or it could mean"build miri for the host and then run it for the given target". But IMO it makes most sense to be consistent with test here, where it means the latter.

Operating system

Debian testing

HEAD

6ba0ce4

[Kobzol]

To clarify, you don't even need to build the standard library for the specified target in order to run Miri?

[RalfJung]

Miri has its own logic for building the standard library that should kick in here (it works for ./x run miri without --target anyway).

[Kobzol]

So the miri sysroot setup logic should be invoked for aarch64-apple-darwin, but miri itself and thus also the rustc rlib artifacts that it links to should always be built for the host in x run miri, is that correct?

[RalfJung]

Yes, exactly. That's what already happens in ./x test miri.