@BarrySmith @jedbrown @ckhroulev

Summary

PISM fails to configure because PETSc has not stored path to MPI in its configuration files. This is probably caused by a change in the Spack package; specifically the loss of the --with-mpi-dir flag from d04ae9a2.

Details

When PETSc builds, it creates a file <prefix>/lib/petsc/conf/petscvariables, containing a number of key-value pairs describing the PETSc installation. This file is later read and relied upon by some programs that use PETSc; PISM, for example:
https://github.com/pism/pism

PISM's CMake build looks up the PETSc variable PETSC_CC_INCLUDES:
https://github.com/pism/pism/blob/master/CMake/FindPETSc.cmake#L159

It then uses this variable to construct correct command lines for small test programs used to deduce the configuration of PETSc:

https://github.com/pism/pism/blob/b17bcd0c0bce5720736e2dad8317a8c9cd7b5706/CMake/FindPETSc.cmake#L265

The Problem

At the point in the PISM CMake build where test programs are compiled, the variable ${includes} needs to contain an -I flag for MPI (in my case, OpenMPI). However, it does not. This is traced back to PETSC_CC_INCLUDES not containing a -I flag for MPI in the PETSc file petscvariables (above).

It is believed that OpenMPI is not showing up in PETSC_CC_INCLUDES because the Spack package no longer proviedes PETSc with the flag --with-mpi-dir as of d04ae9a2.

Proposed Fix

I will try restoring --with-mpi=1 and --with-mpi-dir, while keeping --with-cc, --with-cxx and --with-fc as well. Thoughts from @jedbrown and @BarrySmith on possible effects this might have --- or whether there's a better way involving PETSc internals --- would be appreciated. I will report back on what I find.

d04ae9a2