In the testings, if the initialization of StarPU fails, starpu_shutdown() is still called, which makes StarPU joining the communication progression thread, which has not been launched! As a consequence, it either deadlocks after displaying CHAMELEON_Init() failed or prints a backtrace.
The problem can be reproduced with the following command:
mpirun -x STARPU_MAIN_THREAD_BIND=1 ./testing/chameleon_stesting -o potrf -t 1(I noticed the problem with NewMadeleine, but it seems it is also present with MPI).
This problem seems to be a regression introduced by the commit 39dbaf81.
Note that I just pushed a commit in StarPU that raises an error if the user calls starpu_mpi_shutdown() after a failed starpu_mpi_init(): starpu/starpu@d988deb7