avoid type redefine of MPI_Comm and MPI_Info

[wkliao]

MPI_Comm and MPI_Info should only be used when parallel I/O is enabled.
Compiling file libdispatch/dparallel.c is skipped when parallel I/O is disabled.

[edwardhartnett]

Thanks, this is a good solution!

[edwardhartnett]

Clever and simple! ;-)

[mathstuf]

Should netcdf_par.h be skipped from the install as well?

[wkliao]

netcdf_par.h is installed only if parallel I/O is enabled.

netcdf-c/include/Makefile.am

Lines 12 to 14 in 4366db8

	if BUILD_PARALLEL
	include_HEADERS += netcdf_par.h
	endif

netcdf-c/include/CMakeLists.txt

Lines 38 to 42 in 4366db8

	IF(ENABLE_PNETCDF OR ENABLE_PARALLEL4)
	INSTALL(FILES ${netCDF_SOURCE_DIR}/include/netcdf_par.h
	DESTINATION ${CMAKE_INSTALL_INCLUDEDIR}
	COMPONENT headers)
	ENDIF()

[WardF]

Fantastic, thank you!

[edwardhartnett]

Maybe we don't want to only build dparallel.c when parallel is enabled.

See Unidata/netcdf-fortran#344

[edwardhartnett]

I have submitted a PR to undo these changes (#2300).

If these kind of changes are needed, they must be done in a way that does not break netcdf-fortran.

[mathstuf]

Does this mean that #2187 is the way forward then?

[edwardhartnett]

I really don't know. ;-)

However, whatever is done must be done in such a way that netcdf-fortran does not break. That could mean we restore these changes and then change netcdf-fortran so that it too does not include or use the parallel functions except for parallel builds.

Or perhaps some other solution, like #2187 is best.

But whatever it is, it needs to not break netcdf-fortran.

I am wondering if netcdf-c shouldn't build and test netcdf-fortran as part of the netcdf-c CI run. That is, treat netcdf-fortran as an extra set of tests for netcdf-c. That way, we would not make the mistake of making changes to netcdf-c which are fine by themselves, but break the fortran build.

[edwardhartnett]

See #2301

[wkliao]

I suggest to fix netcdf-fortran.
If netcdf-fortran detects the parallel feature is disabled in netcdf-c,
it should skip defining/compiling those parallel APIs.
Luckily, there are only 3 parallel APIs:

• nf_create_par
• nf_open_par
• nf_var_par_accese

[edwardhartnett]

@wkliao I don't object to that solution, but I suspect it will have to wait until after the next netcdf-c/netcdf-fortran release, unless it is done immediately...

[wkliao]

It turns out the fix is simple. See #352

[WardF]

Thanks!