{llvm,triton-llvm}: fix nondeterministic hang

[stephen-huan]

I'm experiencing a nondeterministic hang while running llvm's tests (it gets stuck on a lit test and doesn't finish, even after waiting for a long time). Other times it finishes fine. Seems to be more stable with 4 cores or less.

Possibly caused by llvm/llvm-project#56336 (see llvm/llvm-project@61708ec and the following comment)

# FIXME: This test is flaky and hangs randomly on multi-core systems.
# See https://github.com/llvm/llvm-project/issues/56336 for more
# details.
# REQUIRES:  less-than-4-cpu-cores-in-parallel

Not sure if this is the only flaky test as the hang is nondeterministic and llvm takes a long time to build.

Of course this could be fixed by --cores 4 but that would slow down the build as well (and not just the tests).

Happy to change the PR to just disable this test (rm llvm/utils/lit/tests/max-failures.py) if that would be cleaner.

Things done

• Built on platform(s)
  • x86_64-linux
  • aarch64-linux
  • x86_64-darwin
  • aarch64-darwin
• For non-Linux: Is sandboxing enabled in nix.conf? (See Nix manual)
  • sandbox = relaxed
  • sandbox = true
• Tested, as applicable:
  • NixOS test(s) (look inside nixos/tests)
  • and/or package tests
  • or, for functions and "core" functionality, tests in lib/tests or pkgs/test
  • made sure NixOS tests are linked to the relevant packages
• Tested compilation of all packages that depend on this change using nix-shell -p nixpkgs-review --run "nixpkgs-review rev HEAD". Note: all changes have to be committed, also see nixpkgs-review usage
• Tested basic functionality of all binary files (usually in ./result/bin/)
• 25.05 Release Notes (or backporting 24.11 and 25.05 Release notes)
  • (Package updates) Added a release notes entry if the change is major or breaking
  • (Module updates) Added a release notes entry if the change is significant
  • (Module addition) Added a release notes entry if adding a new NixOS module
• Fits CONTRIBUTING.md.

---

Add a 👍 reaction to pull requests you find important.

[RossComputerGuy]

I haven't been able to reproduce this issue. Tests pass after a few minutes. Limiting the lit jobs for me would severely slow LLVM builds down.

[stephen-huan]

I haven't been able to reproduce this issue. Tests pass after a few minutes.

What degree of parallelism are you using? It hangs for me with reasonably high probability (around one in two/three) at 8 cores, and I suspect the more cores the more likely. I haven't tested > 8 since memory becomes the bottleneck.

Limiting the lit jobs for me would severely slow LLVM builds down.

I noticed this as well. Would disabling the test instead be preferred (rm llvm/utils/lit/tests/max-failures.py)? I didn't originally write this PR in this way because I was concerned there may be other flaky tests, but I can disable them as they arise. (It's hard to test because llvm takes a long time to build + would have to build repeatedly to be confident.)

[RossComputerGuy]

What degree of parallelism are you using? It hangs for me with reasonably high probability (around one in two/three) at 8 cores, and I suspect the more cores the more likely. I haven't tested > 8 since memory becomes the bottleneck.

My system's default of 64 cores, I've been packaging LLVM since 17 released and I've never experienced this issue.

I noticed this as well. Would disabling the test instead be preferred (rm llvm/utils/lit/tests/max-failures.py)? I didn't originally write this PR in this way because I was concerned there may be other flaky tests, but I can disable them as they arise. (It's hard to test because llvm takes a long time to build + would have to build repeatedly to be confident.)

That is the correct approach here, the other thing is it might not be a bad idea to investigate into what exactly triggers this. I've never experienced this natively on aarch64, x86_64, and riscv64.

[stephen-huan]

the other thing is it might not be a bad idea to investigate into what exactly triggers this.

Thanks for the tip. It turns out the lit test was a red herring as the REQUIRES already correctly disables the test (!).

# REQUIRES:  less-than-4-cpu-cores-in-parallel

The real cause is a random hang in loading the configuration llvm/test/tools/llvm-exegesis/lit.local.cfg before the tests are even ran (see llvm/llvm-project#132861).

[stephen-huan]

I have a simpler reproduction, just using llvm-exegesis from pkgs.llvmPackages.llvm (no need to rebuild llvm). Run

llvm-exegesis -mode latency -opcode-name=ADD64rr -x86-lbr-sample-period 123 -repetition-mode loop

On the problematic desktop, it gives

---
mode:            latency
key:
  instructions:
    - 'ADD64rr RAX RAX R12'
  config:          ''
  register_initial_values:
    - 'RAX=0x0'
    - 'R12=0x0'
cpu_name:        alderlake
llvm_triple:     x86_64-unknown-linux-gnu
min_instructions: 10000
measurements:
  - { key: latency, value: 0.0001, per_snippet_value: 0.0001, validation_counters: {} }
error:           ''
info:            Repeating a single implicitly serial instruction
assembled_snippet: 415448B8000000000000000049BC000000000000000049B802000000000000004C01E04C01E04983C0FF75F4415CC3
...

or hangs. On my laptop (with basically the same nixos configuration) it gives

llvm-exegesis error: LBR not supported on this kernel and/or platform

without hanging. Seems that a hardware difference determines whether LBR is available or not.