Warnings

[aaadelmann]

Warning Summary

This summarizes warnings observed in two PR-544 builds.

• GH200 CUDA build: https://my.cdash.org/builds/3766785/build
• Eiger OpenMP build: https://my.cdash.org/builds/3766775/build

Build Summary

Build	Site	Configure warnings	Build warning rows	Unique grouped rows
3766785 / PR-544-gh200-Release-1-rank	Daint gh200 CUDA 12.9.1	0	38	4
3766775 / PR-544-openmp-Release-4-ranks	Eiger OpenMP	0	50	2

Grouped Warnings

Count in 3766785 GH200 CUDA	Count in 3766775 Eiger OpenMP	File	Line	Warning
10	0	/user-environment/env/default/include/experimental/__p0009_bits/utility.hpp	201	CUDA warning #20013-D: calling constexpr host function min from __host__ __device__ function in_range
10	0	/user-environment/env/default/include/experimental/__p0009_bits/utility.hpp	202	CUDA warning #20013-D: calling constexpr host function max from __host__ __device__ function in_range
10	0	/user-environment/env/default/include/experimental/__p0009_bits/utility.hpp	219	CUDA warning #20013-D: calling constexpr host function max from __host__ __device__ function check_mul_result_is_nonnegative_and_representable
8	0	unit_tests/Particle/ParticleUpdateNonuniform.cpp	881	CUDA warning #128-D: loop is not reachable
0	25	src/Communicate/Archive.hpp	343	GCC/OpenMP warning: 'value' may be used uninitialized [-Wmaybe-uninitialized]
0	25	src/Communicate/Archive.hpp	340	GCC note: 'value' declared here

Interpretation

The GH200 CUDA build has 38 warning. These reduce to two issues:

1. 30 warning come from the environment mdspan implementation:

   /user-environment/env/default/include/experimental/__p0009_bits/utility.hpp
   

   The CUDA compiler warns that std::min / std::max-like constexpr host functions are called from __host__ __device__ functions. The compiler suggests --expt-relaxed-constexpr. These warnings are external to IPPL source, but they may be triggered through IPPL/Kokkos use of mdspan.

2. 8 warning rows come from:

   unit_tests/Particle/ParticleUpdateNonuniform.cpp:881
   

   The warning is:

   warning #128-D: loop is not reachable
   

   This points at:

   REQUIRE_RANKS(PREF_RANKS);

   inside TYPED_TEST(TestParticleUpdateORB, ThreeDCornerMigrationAfterOrb), after:

   if constexpr (TestFixture::Dim != 3) {
       GTEST_SKIP() << "3-D specific test";
   }

   Likely cause: CUDA sees unreachable control flow after template instantiation / macro expansion for the non-3D typed-test variants.

The Eiger OpenMP build has 50 warning. These reduce to one real warning plus its repeated note:

1. 25 warning rows come from:

   src/Communicate/Archive.hpp:343
   

   The warning is:

   'value' may be used uninitialized [-Wmaybe-uninitialized]
   

   The relevant code is:

   T value;
   char* dst = reinterpret_cast<char*>(&value);
   copyBytes(dst, src, size);
   view.data()[i](d) = value;

   GCC does not prove that copyBytes(dst, src, size) initializes all bytes of value before the assignment. This is probably a false positive if copyBytes always copies sizeof(T) bytes, but the warning points at a fragile pattern.

2. 25 additional warnings are the associated GCC note:

   src/Communicate/Archive.hpp:340:23: note: 'value' declared here
   

Possible Follow-Up

• For the Archive.hpp warning, initialize value defensively:

  T value{};

  This should silence GCC and avoid any real uninitialized-read risk if the copy path ever changes.

• For the ParticleUpdateNonuniform.cpp CUDA warning, avoid placing REQUIRE_RANKS(PREF_RANKS) in a path that CUDA sees after a compile-time skipped branch. One option is to move the body into an else branch:

  if constexpr (TestFixture::Dim != 3) {
      GTEST_SKIP() << "3-D specific test";
  } else {
      REQUIRE_RANKS(PREF_RANKS);
      ...
  }

• For the external mdspan CUDA warnings, options are:

  • enable --expt-relaxed-constexpr for CUDA builds if acceptable,
  • check whether a newer Kokkos/mdspan combination avoids the warning,
  • suppress this specific CUDA diagnostic for external/system headers if the project policy allows it.