[cmake-developers] FindMPI take 2

[Thompson, KT]

Brad King wrote:

>> 0002-first-try-to-see-if-what-user-provided-produces-an-m.patch

> The check for whether the CMAKE_<LANG>_COMPILER is already a
> MPI compiler requires a try_compile.  Is there some other way
> to check to see if there is any chance before the full test?

I have been looking into this.  For my setup, I found it useful to compare the CMAKE_<LANG>_COMPILER to the list of known mpi wrapper names and prepend _MPI_${lang}_COMPILER_NAMES with CMAKE_<LANG>_COMPILER so that the remaining logic in FindMPI.cmake chooses the user supplied compiler/mpi-wrapper over any of the other choices.  Something like this:

# append vendor-specific compilers to the list if we either don't know the compiler id,
# or if we know it matches the regular compiler.
foreach (lang C CXX Fortran)
  foreach (id GNU Intel PGI XL)
    if (NOT CMAKE_${lang}_COMPILER_ID OR CMAKE_${lang}_COMPILER_ID STREQUAL id)
      list(APPEND _MPI_${lang}_COMPILER_NAMES ${_MPI_${id}_${lang}_COMPILER_NAMES})
    endif()
    unset(_MPI_${id}_${lang}_COMPILER_NAMES)    # clean up the namespace here
  endforeach()

+  # If cmake_$lang_compiler matches a known mpi compiler wrapper name,
+  # prefer the provided value.
+  get_filename_component( compiler_wo_path "${CMAKE_${lang}_COMPILER}" NAME )
+  set( ${lang}_compiler_is_mpiwrapper false )
+  foreach( mpiwrapper ${_MPI_${lang}_COMPILER_NAMES} )
+    if( "${mpiwrapper}" STREQUAL "${compiler_wo_path}" )
+      set( ${lang}_compiler_is_mpiwrapper true )
+    endif()
+  endforeach()
+  if( ${lang}_compiler_is_mpiwrapper )
+    list( REMOVE_ITEM _MPI_${lang}_COMPILER_NAMES ${compiler_wo_path} )
+    list( INSERT _MPI_${lang}_COMPILER_NAMES 0 ${compiler_wo_path} )
+  endif()
endforeach()

There is probably a better way to do this, but you get the idea.  In my setup, I'm trying to get FindMPI.cmake to select mpiicpc over mpicc without manually setting the MPI_${lang}_COMPILER manually.

> Also I'm not sure we fully support using a MPI compiler 
> as the main compiler.  I can't think of anything that 
> definitely won't work, but we don't have nightly testing
> for the case.

My project does nightly regressions with and without the use of MPI compiler wrappers.  I have never had a problem using the compiler wrappers (e.g.: mpicc, mpiicpc, Cray's CC).  Additionally, I think Trilinos (trilinos.org) requires the use of compiler wrappers to build their code (cmake build system).

-kt