Workaround for __builtin_fmaxf/fminf in clang+Arm64 (nightly build failure)

[fbusato]

Description

Clang lowers fmax/fmin to AArch64 HW instructions, which returns quite NaN instead of signaling NaNs. Releted LLVM issue llvm/llvm-project#176624

[coderabbitai]

Understand this PR’s impact

Explore downstream dependencies and potential security impact with Blast Radius.

View blast radius →

No actionable comments were generated in the recent review. 🎉

ℹ️ Recent review info

⚙️ Run configuration

Configuration used: Repository: NVIDIA/cccl/.coderabbit.yaml

Review profile: CHILL

Plan: Enterprise

Run ID: 24ff7147-f857-4578-83ec-942789ade646

📥 Commits

Reviewing files that changed from the base of the PR and between 2c9db0f and 0915f91.

📒 Files selected for processing (1)

• libcudacxx/include/cuda/std/__cmath/min_max.h

Included review availability: Your plan provides up to 12 included reviews per hour; 11 remain after this review.

---

📝 Summary

Summary by CodeRabbit

• Bug Fixes
  • Improved floating-point minimum and maximum handling on ARM64 systems when one operand is NaN.
  • Ensured the non-NaN operand is returned in affected cases while preserving existing behavior for other platforms and inputs.

Walkthrough

The change adds host ARM64 Clang-specific NaN checks to __with_builtin_fmax and __with_builtin_fmin. Each function returns the other operand when one operand is NaN. Other configurations and non-NaN inputs retain the builtin path.

Changes

ARM64 NaN handling

Layer / File(s)	Summary
ARM64 Clang NaN guards libcudacxx/include/cuda/std/__cmath/min_max.h	__with_builtin_fmax and __with_builtin_fmin check operands for NaN on host ARM64 Clang. Each function returns the non-NaN operand before calling the builtin. Other paths retain the existing builtin implementation.

Priority: ⬇️ Low

Change: Bug fix

---

_{Comment @coderabbitai help to get the list of available commands.}

[coderabbitai]

Actionable comments posted: 1

---

ℹ️ Review info

⚙️ Run configuration

Configuration used: Path: .coderabbit.yaml

Review profile: CHILL

Plan: Enterprise

Run ID: 83d91acc-d041-483c-9ba8-a635fc1f7984

📥 Commits

Reviewing files that changed from the base of the PR and between 2fcc3f0 and 2c9db0f.

📒 Files selected for processing (1)

• libcudacxx/include/cuda/std/__cmath/min_max.h

Included review availability: Your plan provides up to 12 included reviews per hour; 11 remain after this review.

[miscco]

Question: Why is this not changing fmin when that is also affected?

[fbusato]

Question: Why is this not changing fmin when that is also affected?

the last commit was not pushed. Thanks for reminding me!

[github-actions]

🥳 CI Workflow Results

🟩 Finished in 3h 08m: Pass: 100%/221 | Total: 3d 02h | Max: 1h 57m | Hits: 99%/607373

See results here.