With these changes, I can successfully build PIO. I tested a standalone
PIO build (i.e., outside of an ESMF build) with:

CC=clang FC=gfortran cmake -DPIO_USE_MPISERIAL=ON -DPIO_ENABLE_TIMING=OFF -DPIO_ENABLE_EXAMPLES=OFF -DPIO_ENABLE_FORTRAN=OFF -DMPISERIAL_PATH=/Users/sacks/esmf/esmf1/src/Infrastructure/stubs/mpiuni -DPIO_ENABLE_TESTS=OFF .

make VERBOSE=1

According to MPI documentation, most MPI routines check sendbuf, but
some (MPI_Scatter, MPI_Scatterv and maybe others; note that these are
not yet implemented yet in mpiuni) check recvbuf, and some (e.g.,
MPI_Sendrecv_replace) don't check for MPI_IN_PLACE at all. To make
mpiuni respect the MPI standard in this respect, I have added an
argument to MPIUNI_Memcpy saying whether to check the source (sendbuf),
dest (recvbuf) or neither for equality with MPI_IN_PLACE.

(We could probably get away with keeping things simpler by always
checking both a and b for equality with MPI_IN_PLACE, but following the
MPI standard in this respect seems marginally safer.)

This is needed for the PIO build. I'm using the same definition as is
used in mpi-serial.

This no longer depends on whether we're using mpiuni

I thought that we might need to add some logic to handle the case where
the user explicitly sets ESMF_PIO=OFF: I thought that would appear in
the esmf.mk file. But it turns out that ESMF_PIO doesn't appear in the
esmf.mk file in that case, so the logic here seems correct.

Implement MPI_Alltoallw, MPI_Type_create_indexed_block and MPI_Type_size.

These are needed for a srcfield.read call in ESMPy to work. With these
changes, the following python program appears to work correctly:

import os
import esmpy

from esmpy.util.cache_data import DATA_DIR
from esmpy.util.exceptions import DataMissing

datafile = os.path.join(DATA_DIR, "so_Omon_GISS-E2.nc")
meshfile = os.path.join(DATA_DIR, "mpas_uniform_10242_dual_counterclockwise.nc")

grid = esmpy.Grid(filename=os.path.join(DATA_DIR, datafile),
				  filetype=esmpy.FileFormat.GRIDSPEC)
srcfield = esmpy.Field(grid, staggerloc=esmpy.StaggerLoc.CENTER, ndbounds=[33, 2])

srcfield.read(filename=os.path.join(DATA_DIR, datafile),
	      variable="so", timeslice=2)

print("Worked!")

- Implement MPI_Scatterv; this is needed for PIO's subset rearranger,
which is used to read a Mesh (and is the only rearranger allowed by
PIOc_InitDecomp_ReadOnly)

- Generalize the implementation of MPI_Type_create_indexed_block to
allow count > 1

With these changes, the example in
https://earthsystemmodeling.org/esmpy_doc/nightly/develop/html/examples.html#grid-mesh-and-field-created-from-file
runs to completion

I had a check that the displacements go in increments of blocklength,
but from some experimentation with checking the MPI_Type_size of the new
type returned from this function from openmpi, it seems that the
resulting type's size is independent of the displacements. So I don't
think we need this check.

and MPI_Type_hvector

These are needed in PIO. I am implementing both MPI_Type_create_hvector
and MPI_Type_hvector because PIO potentially uses the latter when it
thinks it's using mpi-serial.

mpiuni's mpirun was always prepending './' to the command, even for a
command in your path. This prevented running python3 via mpirun (needed
in the ESMPy testing), for example.

This change only prepends './' if the given command isn't in your PATH.
This seems to be the behavior for openmpi's mpirun.