Configuration problems with 1.7.1 and not 1.7.0

[drwells]

Perhaps related to #3357 (hi @amneetb)

At Amneet's request I checked out 1.7.1 and there is some issue with it detecting a system installation of MPI (i.e., I want to use /usr/include/mpi.h). I see some interesting lines in config.log:

configure:48297: checking for MPI_Init in -lmpi
configure:48322: mpic++ -o conftest -Wno-implicit-fallthrough   conftest.cpp -lmpi  -L/lib  >&5
configure:48322: $? = 0
configure:48331: result: yes
configure:48338: result: Found valid MPI installation...
configure:48639: result: Could not find MPI header <mpi.h>...

I don't really understand the logic but the check in acsm_mpi.m4 requires that MPI_INCLUDES_PATH get set at some point - it doesn't look like that happens, which causes MPI detection to fail at that point.

This issue doesn't exist with 1.7.0.

[roystgnr]

Could you double-check that you've done a recursive init on the submodules? #3357 looks like it might have just been a mismatch there.

the check in acsm_mpi.m4 requires that MPI_INCLUDES_PATH get set at some point - it doesn't look like that happens

That line is inside the [test -n "$MPI_LIBS_PATH" -a -n "$MPI_INCLUDES_PATH"] condition's true block, isn't it? So it should never be invoked unless we have something set there, either from your environment or from acsm_compiler_control_args.m4 (which looks for $MPI in your environment, --with-mpi, or --with-mpi-include, IIRC in the opposite order).

[drwells]

Yup - if I try things on branch_v1.7.1 then things work with updated submodules. They do not work in the release tarball.

[jwpeterson]

It's possible the directory that I ran make dist in was not recursively up-to-date? Our buildbox that usually makes the release tarballs currently has an error, so I made them by hand. I can try again...

[jwpeterson]

OK, I deleted the original 1.7.1 tarballs and regenerated/reuploaded them after making sure all the submodules were recursively up-to-date. The md5 sums of the files changed, so presumably something is actually different this time :grimacing:

a86901c65f640d841b5a63289ea93840  (old libmesh-1.7.1.tar.bz2)
4dcf325975db5e60abe1384ab5fb8542  (old libmesh-1.7.1.tar.gz)
0fc304be568db558f4f6256771823799  (old libmesh-1.7.1.tar.xz)

39acfb2492a4e0563226191ae98b10f2  (new libmesh-1.7.1.tar.bz2)
4af9ee0bebfb67f9d944d9e0585af068  (new libmesh-1.7.1.tar.gz)
0cb49ffede7a1bef572dfa431aef922f  (new libmesh-1.7.1.tar.xz)

[jwpeterson]

I also just tested running ./configure from one of the new tarballs and it seems to work fine for me... so hopefully this fixes the problem for you as well, and sorry about the mixup.

[drwells]

This still doesn't work with my setup and I get the same problem on the development branch too. Here's how I call configure:

set -e

export CC=mpicc
export CXX=mpic++
export FC=mpifort
export F77=mpifort

export PETSC_DIR=$HOME/Applications/petsc-3.7.7/x86_64-debug

unset SLEPC_DIR

mkdir -p objs-debug
cd objs-debug
../configure                                   \
    CCFLAGS="-Wno-implicit-fallthrough"        \
    CXXFLAGS="-Wno-implicit-fallthrough"       \
  --prefix=$HOME/Applications/libmesh-dev.g/   \
  --with-methods=dbg                           \
  --enable-triangle                            \
  --enable-exodus=yes                          \
  --enable-timestamps=no                       \
  --disable-glibcxx-debugging                  \
  --enable-poly2tri=no                         \
  --disable-openmp                             \
  --disable-perflog                            \
  --disable-pthreads                           \
  --disable-tbb                                \
  --disable-trilinos                           \
  --disable-boost                              \
  --with-thread-model=none                     \
  --disable-reference-counting                 \
  --disable-slepc                              \
  --disable-strict-lgpl                        \
  --disable-vtk                                \
  --disable-eigen                              \
  --with-metis=PETSc                           \
  --disable-deprecated                         \
  --disable-hdf5

make -j4
make install

and here's the log (also config.log

Things work correctly on my mac - I suspect the MPI detection is still getting confused over paths (I have /usr/include/mpi.h and /usr/bin/mpic++ - i.e., its not in some custom directory).

[roystgnr]

and I get the same problem on the development branch too

If branch_v1.7.1 is working but devel is failing, is it something you can bisect? I'm really not sure how to debug this remotely.