NVIDIA cannot verify contiguity of tmpptr in xgrid OpenMP blocks

[marshallward]

Describe the bug
When building the newest xgrid with NVIDIA and OpenMP enabled, I get the following error:

NVFORTRAN-S-0155-Illegal POINTER assignment - pointer target must be simply contiguous (/scratch4/GFDL/gfdloceans/Marshall.Ward/m6e/src/FMS2_gpu/exchange/xgrid.F90: 4084)

The code block is

FMS/exchange/xgrid.F90

Lines 4080 to 4097 in 043ab2f

	!$OMP parallel do default(none) shared(lsize,shape_d,d_addrs,xmap,recv_buffer,pl,ml) &
	!$OMP private(d,tmpptr,grid,i,j,p,pos)
	do l = 1, lsize
	call c_f_pointer(d_addrs(l), tmpptr, shape=shape_d)
	d(xmap%grids(1)%is_me:xmap%grids(1)%ie_me, xmap%grids(1)%js_me:xmap%grids(1)%je_me) => tmpptr
	do g=2,size(xmap%grids(:))
	grid => xmap%grids(g)
	do l3=1,grid%size_repro ! index into side1 grid's patterns
	i = grid%x_repro(l3)%i1
	j = grid%x_repro(l3)%j1
	p = grid%x_repro(l3)%pe-xmap%root_pe
	pos = pl(p) + (l-1)*ml(p) + grid%x_repro(l3)%recv_pos
	d(i,j) = d(i,j) + recv_buffer(pos)
	end do
	end do
	! normalize with side 1 grid cell areas
	d = d * xmap%grids(1)%area_inv
	enddo

Since tmpptr has been declared as contiguous, this could be an NVIDIA problem.

To Reproduce
Build with nvfortran with OpenMP enabled (e.g. -mp=gpu)

Expected behavior
It should compile!

System Environment
Describe the system environment, include:

• OS:

  $ uname -a
  Linux ufe03 5.14.0-570.60.1+4.1.el9_6_ciq.x86_64 #1 SMP PREEMPT_DYNAMIC Fri Feb 6 20:25:28 UTC 2026 x86_64 x86_64 x86_64 GNU/Linux
  

• Compiler(s): Type and version [e.g. Intel 19.1]

  $ nvfortran --version
  
  nvfortran 26.1-0 64-bit target on x86-64 Linux -tp znver4 
  NVIDIA Compilers and Tools
  

• MPI type, and version (hpc-x)

  $ mpiexec --version
  mpiexec (OpenRTE) 4.1.9a1
  

• netCDF Version: For both C and Fortran
  (can probably skip this one)

• Configure options: This is what I used, but probably more

  FCFLAGS = -g -O0 -mp=gpu -acc -gpu=cc90,mem:separate -Mnofma
  LDFLAGS = -mp=gpu
  

Additional context
If I remove the OpenMP directive on the block then it works. That's probably what I will do until this is resolved.

The issue seems to have come from this commit: 301b4fd

[marshallward]

This is also happening for the NVIDIA CPU OpenMP build (so just -mp).