Fix --debug/--reldebug GPU builds

[sbryngelson]

Problem

--debug and --reldebug GPU builds (--gpu acc/--gpu mp) fail to compile with:

NVFORTRAN-S-1061-Procedures called in a compute region must have acc routine
information - s_mpi_abort
  (m_variables_conversion.fpp: 1335)

Root cause

s_compute_fast_magnetosonic_speed is a GPU device routine ($:GPU_ROUTINE(... parallelism='[seq]')), but its #ifdef MFC_DEBUG diagnostic block calls the host routine s_mpi_abort. That call is only legal in Release because IPO/-Minline inlines it away — and --debug/--reldebug disable IPO. With IPO off, the device routine is left calling a host MPI routine, which is illegal in device code.

This is why the failure is debug-only: in Release the entire #ifdef MFC_DEBUG block is compiled out.

Fix

Guard the host s_mpi_abort call with #ifndef MFC_GPU, keeping the device-legal diagnostic print:

#ifdef MFC_DEBUG
        if (disc < 0._wp) then
            print *, 'rho, c, Bx, By, Bz, h, term, disc:', ...
#ifndef MFC_GPU
            call s_mpi_abort('Error: negative discriminant in s_compute_fast_magnetosonic_speed')
#endif
        end if
#endif

A full-tree scan confirms this is the only device-region s_mpi_abort site for NVHPC/Cray (the 6 NaN-check aborts in m_mpi_common.fpp are inside #if defined(__INTEL_COMPILER) and so are not compiled for these backends). CPU and Release builds are byte-for-byte unchanged: off-GPU the guarded call is identical, and in Release the block is compiled out.

Verification

Built on NVHPC 24.5 (Quadro RTX 6000, MFC_CUDA_CC=75):

Build	Result
simulation --gpu acc --debug	compiles, 0× NVFORTRAN-S-1061
simulation --gpu acc --reldebug	compiles, 0× NVFORTRAN-S-1061
./mfc.sh precheck	all 7 checks pass
./mfc.sh format	clean

Verified on NVHPC only (the GPU compiler available on the test node); the fix is gated on MFC_GPU and is therefore compiler-agnostic (Cray/AMD untested here).

[Copilot]

Pull request overview

Fixes GPU Debug/RelDebug compilation failures by preventing a host-only MPI abort routine (s_mpi_abort) from being referenced inside a GPU device routine (s_compute_fast_magnetosonic_speed) when IPO/inlining is disabled.

Changes:

• Guarded the s_mpi_abort call in the MFC_DEBUG diagnostic path so it is not compiled for MFC_GPU builds.
• Added explanatory comments documenting why the guard is required for device compilation.

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 60.64%. Comparing base (08d12f8) to head (978ef64).
⚠️ Report is 1 commits behind head on master.

Additional details and impacted files

@@           Coverage Diff           @@
##           master    #1527   +/-   ##
=======================================
  Coverage   60.64%   60.64%           
=======================================
  Files          73       73           
  Lines       20213    20213           
  Branches     2936     2936           
=======================================
  Hits        12259    12259           
  Misses       5966     5966           
  Partials     1988     1988           

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.