[libcu++] Skip __fp_set_exp fpclassify tests on denormals

[davebayer]

Fixes nvbug 6303102

[coderabbitai]

No actionable comments were generated in the recent review. 🎉

ℹ️ Recent review info

⚙️ Run configuration

Configuration used: Path: .coderabbit.yaml

Review profile: CHILL

Plan: Enterprise

Run ID: 11f86303-62a5-446c-aae1-0414ece45df4

📥 Commits

Reviewing files that changed from the base of the PR and between 1c753b7 and ebe765c.

📒 Files selected for processing (1)

• libcudacxx/test/libcudacxx/libcxx/numerics/floating.point/decompose/set_exp.pass.cpp

---

Note: CodeRabbit is enabled on this repository as a convenience for maintainers
and contributors. Use your best judgment when considering its review comments and
suggestions — a suggested change may be inadequate, unnecessary, or safe to ignore.
Contributors are not expected to address every comment. Human reviews are what
ultimately matter for merging.

Summary

This PR updates the __fp_set_exp test to skip fpclassify assertions that expect denormal/subnormal results to classify as FP_SUBNORMAL on NVHPC builds.

Changes

libcudacxx/test/libcudacxx/libcxx/numerics/floating.point/decompose/set_exp.pass.cpp

In the test_fp_set_exp denormal/subnormal coverage:

• For the positive-denormal case (T{3}) and the signed-denormal case (T{-3}), the cuda::std::fpclassify(res) == FP_SUBNORMAL check is now guarded with #if !_CCCL_COMPILER(NVHPC).
• The test keeps the other assertions unchanged (e.g., checks for isnormal, isnan, isinf, isfinite behavior).

Rationale

On some NVHPC configurations/architectures, subnormal values may be flushed to 0, and fpclassify is implemented via __builtin_fpclassify, which may therefore not reliably report FP_SUBNORMAL for denormal inputs. This targeted guard prevents spurious test failures while preserving the remaining correctness checks.

Walkthrough

Two fpclassify(res) == FP_SUBNORMAL assertions in the positive-denormal and negative-denormal subtests of test_fp_set_exp are wrapped with #if !TEST_COMPILER(NVHPC) guards. Comments explain that NVHPC flushes subnormals to zero via __builtin_fpclassify on some architectures.

Changes

Conditional FP_SUBNORMAL assertions for NVHPC compatibility

Layer / File(s)	Summary
Disable FP_SUBNORMAL assertions for positive and negative denormal cases libcudacxx/test/libcudacxx/libcxx/numerics/floating.point/decompose/set_exp.pass.cpp	Both fpclassify(res) == FP_SUBNORMAL assertions wrap with #if !TEST_COMPILER(NVHPC) guards and include notes that NVHPC flushes subnormals to zero via __builtin_fpclassify on some architectures.

---

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

[coderabbitai]

🧹 Nitpick comments (1)

libcudacxx/test/libcudacxx/libcxx/numerics/floating.point/decompose/set_exp.pass.cpp (1)

68-70: 🧹 Nitpick | 🔵 Trivial | ⚡ Quick win

suggestion: avoid disabling these checks for both targets; gate them so host skips the flaky fpclassify expectation while device still asserts FP_SUBNORMAL. This keeps coverage for __fp_set_exp denormal behavior without reintroducing host-compiler noise. As per coding guidelines, “Guard host-only or device-only behavior with NV_IF_TARGET(NV_IS_HOST, (...)) and NV_IF_TARGET(NV_IS_DEVICE, (...)) respectively.”

Also applies to: 80-82

Source: Coding guidelines

---

ℹ️ Review info

⚙️ Run configuration

Configuration used: Path: .coderabbit.yaml

Review profile: CHILL

Plan: Enterprise

Run ID: bc450af0-5bd9-4e2f-a9b6-6c593e405120

📥 Commits

Reviewing files that changed from the base of the PR and between f52338f and 679576e.

📒 Files selected for processing (1)

• libcudacxx/test/libcudacxx/libcxx/numerics/floating.point/decompose/set_exp.pass.cpp

[github-actions]

🥳 CI Workflow Results

🟩 Finished in 2h 48m: Pass: 100%/69 | Total: 1d 20h | Max: 1h 07m | Hits: 99%/312003

See results here.

[github-actions]

Successfully created backport PR for branch/3.4.x:

• [Backport branch/3.4.x] [libcu++] Skip __fp_set_exp fpclassify tests on denormals #9557