CI failure when testing with LLVM backend

[bjorn3]

https://github.com/rust-lang/rustc_codegen_cranelift/actions/runs/15097894001/job/42435005556

  ---- floats::f16::test_total_cmp stdout ----
  ---- floats::f16::test_total_cmp stderr ----
  
  thread 'main' panicked at coretests/tests/floats/f16.rs:628:5:
  assertion `left == right` failed
    left: Less
   right: Greater
  stack backtrace:
  note: Some details are omitted, run with `RUST_BACKTRACE=full` for a verbose backtrace.
  
  
  failures:
      floats::f16::test_total_cmp
  
  test result: FAILED. 1936 passed; 1 failed; 4 ignored; 0 measured; 0 filtered out; finished in 2.54s
  
  error: test failed, to rerun pass `-p coretests --test coretests`

[tgross35]

Is this LLVM-built tests against a cranelift-built std? I’m not sure how the test would be enabled otherwise, since it’s gated behind cfg(target_has_reliable_f16_math) which should be off for cranelift.

[bjorn3]

No, it builds the standard library using cg_llvm too. cg_clif doesn't get built at all on this CI job.

[bjorn3]

The standard library does get compiled with compiler-builtins-no-f16-f128 however. Maybe the test fails with the libgcc version of an intrinsic?

[tgross35]

Oh no, it looks like this test has been failing for a while https://rust.godbolt.org/z/cb5vjs4eY. I guess the test that failed was running without optimizations? It seems like the behavior gets const propagated out otherwise.

[tgross35]

Slightly reduced https://rust.godbolt.org/z/K35Gc86ad. I'm thinking the sNaN may be getting quieted from a roundtrip via f32. @beetrees do you happen to know if this matches a known bug anywhere?

[beetrees]

This is being caused by llvm/llvm-project#104915 (comment): this LLVM bug means that negation and .abs() will incorrectly silence signalling NaNs (via a roundtrip to f32) on multiple architectures, causing the .total_cmp() order to change.

[tgross35]

Oof, it looks like that is a hard LLVM bug to fix considering the state of llvm/llvm-project#106153. I guess the thing to do here is just comment out any test cases with a FIXME if they involve sNaNs with f16 and make sure tests pass without optimizations locally.

If either of you can put up a r-l/r patch doing that then feel free to r? me, otherwise I'll try to put one up ~tomorrow.

[rustbot]

Error: The feature assign is not enabled in this repository.
To enable it add its section in the triagebot.toml in the root of the repository.

Please file an issue on GitHub at triagebot if there's a problem with this bot, or reach out on #t-infra on Zulip.

[tgross35]

Put up rust-lang/rust#141676 to disable these