[Libmesh-devel] Snooping PETSc configuration

[Roy S. <roy...@ic...>]

I wrote the following to Ben and John when I was just concerned about
seeing if we needed to update our Wiki instructions for the lonestar
supercomputer, but I'm starting to think we might be able to use a
more general fix in the library:

On Mon, 3 Dec 2007, Roy Stogner wrote:

> I'm getting undefined references to PETSc dependencies whenever I try
> to link the libmesh examples, e.g.:
>
> /opt/apps/petsc-intel9-2.3.3/2.3.3/lib/em64t/libpetscksp.a(ml.o): In
> function `PCSetUp_ML':
> (.text+0x1de): undefined reference to `ML_Destroy'
>
> /opt/apps/petsc-intel9-2.3.3/2.3.3/lib/em64t/libpetscmat.a(chaco.o):
> In function `MatPartitioningApply_Chaco':
> (.text+0x14): undefined reference to `FREE_GRAPH'
>
> /opt/apps/petsc-intel9-2.3.3/2.3.3/lib/em64t/libpetscmat.a(mumps.o):
> In function `MatDestroy_MUMPS':
> (.text+0x137): undefined reference to `dmumps_c'
>
> ML, Chaco, and Mumps seem to be built in
> /opt/apps/petsc-intel9-2.3.3/2.3.3/externalpackages/, and I guess I
> can hunt down the right -L/-l options to add to my Make.common by
> hand, but I thought our config scripts were supposed to be snooping
> PETSc's configuration somehow to set up the necessary linker options.

What's going on is that we're snooping PETSc's configuration for a
bunch of individual library options: $(SUPERLU_DIST_LIB),
$(BLOPEX_LIB), etc, and we're including those options in our
Make.common.  Instead of trying to include every dependency one at a
time, shouldn't we be getting PETSc's whole list and using that?  What
worked for me on lonestar was to change this:

     libmesh_LIBS += $(libs-PETSC) \
                     $(BLOCKSOLVE_LIB) $(SPOOLES_LIB) $(HYPRE_LIB) \
                     $(SUPERLU_DIST_LIB) $(UMFPACK_LIB) $(SUPERLU_LIB) \
                     $(BLOPEX_LIB)  $(SPAI_LIB) \
                     $(BLASLAPACK_LIB) $(X11_LIB) $(MATLAB_LIB) \
                     $(libmesh_FLIBS)

Into this:

     libmesh_LIBS += $(libs-PETSC) \
                     $(PACKAGES_LIBS) \
                     $(libmesh_FLIBS)

It's a lot simpler, but more importantly I *think* it's more robust.  Unless
someone screams that I'm going to break their build, I'll probably
check the corresponding changes to Make.common.in into subversion soon.
---
Roy