Fixup and re-enable many tests for NVPTX

[kulst]

While working with NVPTX, I found that many of its tests are currently disabled. They are configured as only-nvptx64 and ignore-nvptx64 at the same time, so they did not run at all. (For tests-suites not running in CI for NVPTX only-nvptx64 alone is sufficient to prevent a test from running)
Some of the tests also failed when I forced them to run.

I would like to improve the test coverage for NVPTX with this PR as a starting point. It

• fixes and re-enables some tests for NVPTX
• adds minicore, where needed.
• adds linker-asm as an assembly-output directive for linkers that emit assembly by default (like llbc) or which can be explicitly instructed to do so.

This PR also adds to the rustc-dev-guide best-practices for target-specific tests and a warning mentioning the interactions of only-X directives with our CI.

cc: @kjetilkjeka
@rustbot label +O-NVPTX
r? @ZuseZ4

[rustbot]

Some changes occurred in src/tools/compiletest

cc @jieyouxu

[ZuseZ4]

Nice find. I'm very busy till June 12th, but happy to check then if you don't mind waiting a bit, otherwise please feel free to reroll.

[ZuseZ4]

re target-linker-default, do you maybe want to wait with that for a later PR where you'll actually use it in a test?
While going through the list, I also saw https://github.com/rust-lang/rust/blob/main/src/doc/rustc-dev-guide/src/tests/directives.md#assembly still mentioning the ptx-linker, can you add another commit to remove it, now that we dropped support for it?

[rustbot]

rustc-dev-guide is developed in its own repository. If possible, consider making this change to rust-lang/rustc-dev-guide instead.

cc @BoxyUwU, @tshepang

[kulst]

re target-linker-default, do you maybe want to wait with that for a later PR where you'll actually use it in a test? While going through the list, I also saw https://github.com/rust-lang/rust/blob/main/src/doc/rustc-dev-guide/src/tests/directives.md#assembly still mentioning the ptx-linker, can you add another commit to remove it, now that we dropped support for it?

Not sure about that, since tests/assembly-llvm/nvptx-arch-default.rs and tests/assembly-llvm/nvptx-linking-cdylib.rs already use it?

My intention was to have an assembly-output directive, where the final linked artifact already is assembly. This allows assembly tests to consider linked code, too. This of course requires a linker that emits assembly by default or is at least able to when explicitly instructed to do so.
Considering this, probably linker-asm sounds more suitable.

I've rebased on top of the current main branch, added linker-asm to the documentation and removed ptx-linker from it.
Also I have enabled tests/assembly-llvm/nvptx-linking-cdylib.rs which is another test that requires the linker-asm directive.

[ZuseZ4]

Thanks, the new name and note in the docs imho give a much better explanation of the change.
Let's verify

@bors try jobs=test-various

otherwise lgtm.

[rust-bors]

💔 Test for cfc5c92 failed: CI. Failed job:

• try - test-various (web logs, enhanced plaintext logs)

[kulst]

Hm, interesting.
x86_64-unknown-linux-musl overwrites the linker in CI. So tests that need llvm-bitcode-linker fail here. Maybe I could overwrite the linker in these tests by using -Clinker=llvm-bitcode-linker but not sure if this is the right approach. Probably making these tests only-nvptx64 would be better.
Another solution would be to make compiletest not specify -Clinker if a custom target is used.

Anyway, it made me look deeper into how our test infrastructure works.
Turns out, in CI only run-make, run-make-cargo and assembly-llvm test-suites are run with ./x.py test --target=nvptx64-nvidia-cuda exactly in test-various. So in all other test-suites only-nvptx64 leads to a test not being run in CI at all (like this one). This is a bit surprising to a user that writes a test.

I was tempted to simply not restrict which test-suites are run in CI targeting nvptx64. However, many tests in UI require std without specifying needs-target-std, which makes them already fail locally.

This leaves us with the current approach, but we probably should document it properly:

• how only-nvptx64 works in the different test-suites
• that especially UI tests must be written with --target=nvptx64-nvidia-cuda and not only-nvptx64

[rustbot]

This PR was rebased onto a different main commit. Here's a range-diff highlighting what actually changed.

Rebasing is a normal part of keeping PRs up to date, so no action is needed—this note is just to help reviewers.

[kulst]

I created #159059 for the linker issue. This will allow us to stick to cross-build tests for NVPTX for all test-suites.

I also modified this PR a little:

• removed unnecessary ignore-backends: gcc directives in codegen-llvm and assembly-llvm
• added best-practices for target-specific tests to the dev guide
• added a warning in the dev-guide mentioning the interactions of only-X directives with our CI
• removed nvptx-linking-binary.rs, as --crate-type bin makes no sense for NVPTX

Since test-various does not run successfully without the linker fix:
@rustbot blocked

[rust-bors]

☔ The latest upstream changes (presumably #159579) made this pull request unmergeable. Please resolve the merge conflicts by rebasing.