Fortran headers have new location in PETSc 3.6

[jkollasch]

From PETSc 3.5 to 3.6 my Fortran code stopped compiling, even with the
updated FindPETSc.cmake. I figured out it was because the Fortran
headers moved from $PETSC_DIR/include/finclude to
$PETSC_DIR/include/petsc/finclude. Therefore any F90 code with a
statement like

#include <finclude/petscsys.h>

would need to be changed to

#include <petsc/finclude/petscsys.h>

in the source files which is a bit of work (hypothetically more for
people with bigger codes than mine) and breaks the source from working
with PETSc <=3.5. I attached a minimal example of a code that breaks
from the 3.5 to 3.6 PETSc update. Note: I found a work around by
changing the attached CMakeLists.txt by editing the line

include_directories(${PETSC_INCLUDES})

to become

include_directories("${PETSC_INCLUDES}/petsc" "${PETSC_INCLUDES}")

which works for both 3.5 and 3.6 without editing my F90 source files
by simply exhaustively putting all the options of where finclude might
be in the include path. While I am fine with this, I thought I'd post
something to draw attention to the issue if you want to change it. For
example, defining a PETSC_Fortran_INCLUDES variable in FindPETSc.cmake
that automatically points to the fincludes directory regardless of the
PETSc version would avoid the need for the above trick (though it does
work).

Here is the example: main.F90

PROGRAM main
#include <finclude/petscsys.h>
PetscInt :: i
END PROGRAM

and CMakeLists.txt

set(CMAKE_MODULE_PATH ${CMAKE_MODULE_PATH} "${CMAKE_SOURCE_DIR}/cmake/Modules/") # Where I have my FindPETSc.cmake

cmake_minimum_required` (VERSION 2.8.8)
project (HELLO)
enable_language (Fortran)

find_package(PETSc REQUIRED)
include_directories(${PETSC_INCLUDES})

find_package(MPI REQUIRED)
add_definitions(${MPI_Fortran_COMPILE_FLAGS})
include_directories(${MPI_Fortran_INCLUDE_PATH})
link_directories(${MPI_Fortran_LIBRARIES})

add_executable(main.x main.F90)
target_link_libraries(main.x ${PETSC_LIBRARIES}  ${MPI_Fortran_LIBRARIES})

[jedbrown]

My opinion, and that of PETSc upstream, is that you should change your code to include Fortran files using the new names. Adding that include path with compiler flags is basically counteracting the value of namespacing. If you want to do it in your code, that's your business, but I can't recommend it for users in general. I know it's annoying.

[jkollasch]

Thanks. I'll keep the compiler flag thing since leaving "finclude" in the source files prevents namespacing issues (I believe), while letting me use older PETSc's.

[jedbrown]

Well, finclude is not a PETSc-specific namespace, but chances are you'll be fine. In any case, it's your problem if you do find a conflict.