`make check` fails in test/CXX with GCC 15: `enum {false=0, true=1}` in testutils.h

[xylar]

Summary

With GCC 15 (which defaults to C23), ./configure reports checking for stdbool.h that conforms to C99... no. As a result HAVE_STDBOOL_H is never defined and test/common/testutils.h falls back to enum {false=0, true=1};, which is a hard error when the header is included from C++ in test/CXX:

In file included from nctst.cpp:20:
./../common/testutils.h:26:7: error: expected identifier before 'false'
   26 | enum {false=0, true=1};

Seen with PnetCDF 1.15.1 (also affects 1.15.0) on Linux x86_64 and aarch64 with GCC 15 and MPICH 5, OpenMPI 5 and MVAPICH, while building the conda-forge package (conda-forge/libpnetcdf-feedstock#14). The same fallback is present in src/utils/ncmpidump/ncmpidump.h, so ncmpidump is silently built with typedef int boolean instead of <stdbool.h> under C23 as well.

Cause

The shipped configure is generated by Autoconf 2.71, whose AC_HEADER_STDBOOL test contains #ifndef bool / #error "bool is not defined" for non-C++ compilers. In C23, bool, true and false are keywords rather than macros, so the test fails even though <stdbool.h> is fully conforming. Autoconf 2.72 fixed this test. The check can be reproduced by compiling the conftest.c from the configure log with gcc -std=gnu2x.

Suggested fixes

• Regenerate configure with Autoconf >= 2.72, and/or
• Make the fallback in testutils.h and ncmpidump.h robust: skip the enum when __cplusplus is defined (C++ already has true/false), e.g. #if !defined(__cplusplus) && (!defined(__STDC_VERSION__) || __STDC_VERSION__ < 202311L).

Workaround

Passing ac_cv_header_stdbool_h=yes in the environment when running ./configure restores the intended behaviour.

[wkliao]

Thanks for reporting.
However, I do not have an access to GCC 15.
Can you confirm that the solution of using Autoconf 2.72 alone fixes the problem?

[xylar]

Confirmed: regenerating configure with Autoconf 2.72 alone fixes it. Tested on the pnetcdf-1.15.1 tarball with GCC 15.3.0 and MPICH 5 on Linux x86_64, no other changes.

	checking for stdbool.h ...	config.h	make -C test/CXX check
shipped configure (Autoconf 2.71)	no	/* #undef HAVE_STDBOOL_H */	fails: testutils.h:26:7: error: expected identifier before 'false'
after autoreconf -fi (Autoconf 2.72)	yes	#define HAVE_STDBOOL_H 1	PASS: nctst, PASS: test_classic

The full library builds fine in both cases; only the C++ tests break with the shipped script.

If you don't have GCC 15 handy, the easiest way to reproduce is with a conda environment, which provides GCC 15, MPICH 5 and Autoconf 2.72 without touching the system. With miniforge (or any conda/mamba) installed:

# one-off environment: GCC 15 toolchain, MPICH 5, Autoconf 2.72
conda create -y -p ./gcc15env -c conda-forge \
    "gcc_linux-64=15" "gxx_linux-64=15" "gfortran_linux-64=15" \
    "mpich=5" "autoconf>=2.72" automake libtool m4 make
conda activate ./gcc15env

# the conda mpicc/mpicxx/mpifort wrappers pick up GCC 15 via $CC/$CXX/$FC
export MPICC=mpicc MPICXX=mpicxx MPIF77=mpifort MPIF90=mpifort

# A: shipped configure -> reproduces the failure
tar xzf pnetcdf-1.15.1.tar.gz && mv pnetcdf-1.15.1 srcA
(cd srcA && ./configure --enable-shared --disable-static | grep stdbool && make -j8 && make -C test/CXX check)

# B: regenerate with Autoconf 2.72 -> passes
tar xzf pnetcdf-1.15.1.tar.gz && mv pnetcdf-1.15.1 srcB
(cd srcB && autoreconf -fi && ./configure --enable-shared --disable-static | grep stdbool && make -j8 && make -C test/CXX check)

(On aarch64 replace linux-64 with linux-aarch64 in the package names.)

The enum {false=0, true=1}; fallback in testutils.h and ncmpidump.h is still not valid C++, so a #ifndef __cplusplus guard there would make the headers robust regardless of which Autoconf produced configure, but it is not needed to fix this report.

---

Posted by Claude Code on behalf of @xylar.

[xylar]

@wkliao, I let Claude Code to the testing but I stand by the finding. Thanks for having a look!

[wkliao]

I have created PR #241 to address this issue.
It adopts your suggestion. Please let me know if it looks OK with you.

[wkliao]

@xylar
Also, thanks for providing the instructions for running and setting the condo environment.
I wonder if they can be done as a Github CI workflow.

[xylar]

@wkliao, yes for sure they can be.

[wkliao]

Do you have an example I can follow to write one for PnetCDF?

[xylar]

@wkliao, here are a few projects whose GitHub Actions CI builds and tests Fortran/MPI codes using the conda-forge toolchain. I have picked ones with short, readable workflows rather than the conda-forge feedstock machinery, which is more complex than you need.

• lfortran/fortran_mpi CI.yml: the closest match. A matrix over {ubuntu, macos} × {mpich, openmpi} using mamba-org/setup-micromamba, with each combination described by a 5-line environment file such as ci/environment_gfortran_mpich.yml. Compiler and MPI versions are pinned directly in that file.
• dftbplus/skprogs build.yml (and its sibling vanderhe/fortnet): a Fortran+MPI scientific code using conda-incubator/setup-miniconda with miniforge, a mpi: [nompi, mpich] matrix, and the c-compiler cxx-compiler fortran-compiler metapackages plus mpich mpich-mpifort, then a normal CMake build and ctest.
• gotm-model/code gotm.yml: a Fortran+netCDF model whose conda-forge job runs the same build on Ubuntu, macOS and Windows from conda install -c conda-forge fortran-compiler netcdf-fortran. It is a nice illustration of getting three platforms from one job definition.

For PnetCDF specifically, a minimal workflow following the same pattern would look like the one below. It is the GitHub Actions equivalent of the steps I posted above, and I have verified the underlying commands locally: on the 1.15.1 tarball with GCC 15.3 and MPICH 5 from conda-forge, autoreconf -fi && ./configure && make && make check passes all 115 tests. The point of the gcc matrix axis is that conda-forge lets you pick any recent GCC (13, 14, 15) on the same ubuntu-latest runner, independent of what Ubuntu ships, so you can catch issues like this one before they reach downstream packagers. The wrappers mpicc/mpicxx/mpifort from conda-forge's mpich and openmpi automatically use the conda compilers via $CC/$CXX/$FC, so no extra configuration is needed. Since the workflow regenerates configure with the environment's Autoconf, it also means the shipped configure is never what gets tested; if you want to test that too, simply drop the autoreconf step in one of the matrix entries.

name: CI - conda-forge toolchain

on: [push, pull_request, workflow_dispatch]

jobs:
  build:
    runs-on: ubuntu-latest
    strategy:
      fail-fast: false
      matrix:
        mpi: [mpich, openmpi]
        gcc: ["14", "15"]
    defaults:
      run:
        shell: bash -el {0}
    steps:
      - uses: actions/checkout@v4

      - uses: mamba-org/setup-micromamba@v2
        with:
          environment-name: ci
          create-args: >-
            gcc_linux-64=${{ matrix.gcc }}
            gxx_linux-64=${{ matrix.gcc }}
            gfortran_linux-64=${{ matrix.gcc }}
            ${{ matrix.mpi }}
            autoconf>=2.72 automake libtool m4 make

      - name: Build and test
        env:
          MPICC: mpicc
          MPICXX: mpicxx
          MPIF77: mpifort
          MPIF90: mpifort
          # let OpenMPI start more ranks than the runner has cores
          OMPI_MCA_rmaps_base_oversubscribe: "yes"
        run: |
          autoreconf -fi
          ./configure --enable-shared --disable-static
          make -j4
          make check

macOS can be added as a second runs-on value by swapping the three *_linux-64 packages for clang_osx-arm64 clangxx_osx-arm64 gfortran_osx-arm64, or by using the platform-neutral c-compiler cxx-compiler fortran-compiler metapackages if you do not need to pin a specific GCC version.

---

Posted by Claude Code on behalf of @xylar.

[wkliao]

Hi @xylar

PnetCDF requires libtool version 2.5.4.
How do I modify this line to specify the version number?

            autoconf>=2.72 automake libtool m4 make

I tried libtool=2.5.4, but it did not work.

[wkliao]

Please ignore my question. I misspelled the command.
Continue trying the CI using Conda-forge.

[wkliao]

One question. Once the Conda-forge environment is set, I can see the followings.

    gcc_impl_linux-64         14.4.0   h33e224c_5             conda-forge
    gcc_linux-64              14.4.0   hc469d41_2             conda-forge
    gfortran_impl_linux-64    14.4.0   he75539a_5             conda-forge
    gfortran_linux-64         14.4.0   ha2afc21_2             conda-forge
    gxx_impl_linux-64         14.4.0   hc436dd5_5             conda-forge
    gxx_linux-64              14.4.0   h0047a55_2             conda-forge

But command gcc is still pointing to /usr/bin/gcc.
What would be the right command for gcc?
Is it gcc_impl_linux-64 or gcc_linux-64 ?

[xylar]

Short answer: nothing is wrong, and you do not need to change anything. gcc still resolving to /usr/bin/gcc is intentional, and the conda-forge compiler is nevertheless the one your build uses, because PnetCDF is configured through the MPI wrappers. You can verify that directly with:

$ mpicc -show
x86_64-conda-linux-gnu-cc -I$CONDA_PREFIX/include ... -lmpi

$ mpifort -show
x86_64-conda-linux-gnu-gfortran -I$CONDA_PREFIX/include ... -lmpifort -lmpi

conda-forge deliberately does not install a plain gcc into the environment, so that the system compiler is not shadowed. The compilers are installed under their cross-compilation triplet names instead, and on linux-64 the commands are:

$ ls $CONDA_PREFIX/bin/x86_64-conda-linux-gnu-*
x86_64-conda-linux-gnu-gcc       x86_64-conda-linux-gnu-g++     x86_64-conda-linux-gnu-gfortran
x86_64-conda-linux-gnu-cc        x86_64-conda-linux-gnu-c++     x86_64-conda-linux-gnu-ar  (etc.)

To answer the literal question, neither of the two names you listed is a command. gcc_impl_linux-64 is the package holding the actual compiler; gcc_linux-64 is the thin "activation" package that you install, and it is what provides the triplet-prefixed wrappers and sets the environment variables below when the environment is activated. The same pattern applies to gxx_linux-64 and gfortran_linux-64.

Rather than spelling out the triplet, the portable way to refer to the compilers is through the variables that these packages export on activation. On my machine, with gcc_linux-64=15:

variable	value
CC	x86_64-conda-linux-gnu-cc
CXX	x86_64-conda-linux-gnu-c++
FC, F77, F90, F95	$CONDA_PREFIX/bin/x86_64-conda-linux-gnu-gfortran
GCC, GXX, GFORTRAN	the gcc / g++ / gfortran names
CPP, AR, RANLIB, LD, NM, AS, STRIP, OBJDUMP, READELF, ...	matching binutils
HOST, BUILD	x86_64-conda-linux-gnu
CFLAGS, CXXFLAGS, FFLAGS, CPPFLAGS, LDFLAGS	conda-forge's default optimization/hardening flags
DEBUG_CFLAGS, DEBUG_CXXFLAGS, DEBUG_FFLAGS, ...	the -Og -g -Wall counterparts

So $CC --version works anywhere, and in a GitHub Actions job you can assert the version you think you are testing:

- run: $CC --version   # x86_64-conda-linux-gnu-cc (conda-forge gcc 15.3.0-5) 15.3.0

To discover this for yourself in any environment, the activation scripts are plain shell and can be listed and read:

$ ls $CONDA_PREFIX/etc/conda/activate.d/
activate-binutils_linux-64.sh  activate-gcc_linux-64.sh  activate-gfortran_linux-64.sh  activate-gxx_linux-64.sh

and this prints exactly what activation changed:

$ env | sort > /tmp/before.txt
$ conda activate ./myenv
$ env | sort > /tmp/after.txt
$ diff /tmp/before.txt /tmp/after.txt | grep -v CONDA_BACKUP

The reference documentation is conda-forge's knowledge base on compilers and conda-build's compiler-tools page, which lists the full set of variables.

Two small clarifications, one of which corrects something I wrote earlier:

• I previously said the MPI wrappers pick up the conda compiler "via $CC". That is not the mechanism: conda-forge's mpicc has the compiler baked in at build time (CC="x86_64-conda-linux-gnu-cc" near the top of the script), which is why it keeps working even with CC unset. The practical outcome is the same, but if you ever do want to override it, the wrapper honours MPICH_CC / MPICH_FC (and OMPI_CC / OMPI_FC for OpenMPI), not CC.
• Activation also sets CPPFLAGS to include -DNDEBUG -O2, and CFLAGS to a portable baseline such as -march=nocona -mtune=haswell. configure will inherit these. That is usually what you want in CI, but if you want a debug build you should pass your own CFLAGS/CPPFLAGS to configure explicitly, or use the DEBUG_* variants.

Also, from the list you posted the environment resolved GCC to 14.4.0 rather than 15. If you want to be sure you are exercising the C23 default that triggered this issue, pin all three to the same major version, e.g. gcc_linux-64=15 gxx_linux-64=15 gfortran_linux-64=15, and check with $CC --version as above.

---

Posted by Claude Code on behalf of @xylar.

[wkliao]

Thanks. These information are very helpful for a Conda-Forge beginner like me.

I encountered an error related to running MPI jobs.

[1790203248.379116] [runnervmtr4k5:56180:0]      ucp_worker.c:1492 UCX  ERROR uct_iface_open(ud_verbs/mana_0:1) failed: Address not valid
Abort(740443279): Fatal error in internal_Init: Other MPI error, error stack:
internal_Init(70)........: MPI_Init(argc=0x7fff8aaad0cc, argv=0x7fff8aaad0c0) failed

The Github Copilt suggests this solution.

        - name: Configure UCX for GitHub-hosted runners
          if: ${{ matrix.mpi_vendor == 'mpich' }}
          run: |
            {
              echo "UCX_TLS=tcp,sm,self"
              echo "UCX_LOG_LEVEL=error"
            } >> "$GITHUB_ENV"

It dose solve the problem, However, I found that running 6 MPI processes is every
slow, much slower than 4 processes. It is also much slower than not under Conda-Forge.
Is this a known issue, or is there a fix to improve the speed?
Many of PnetCDF tests (when running command make ptest) run more than 4 MPI processes.

[xylar]

I think this seems like a limitation of CI. running MPI on CI is very slow because it often only provides 2 cores (maybe 4 with threading) and 6 is badly over-subscribing. But I can only say this based on my vague intuition. I don't really know the details. Let's see what Claude has to say...

[xylar]

@wkliao, @xylar's intuition is right: this is oversubscription. The reason it is so much worse than your existing CI is the MPICH device, not conda-forge.

• Your current workflows build MPICH with --with-device=ch3:sock. A rank that is waiting blocks in poll() on a socket and gives up its core, so running more ranks than cores costs little.
• conda-forge's MPICH (on linux-64) is built with --with-device=ch4:ucx,ofi and uses ch4 shared memory within a node. Waiting ranks busy-poll. With 6 ranks on the runner's 4 vCPUs, the ranks that are spinning take CPU time away from the ranks they are waiting for, and every collective falls off a cliff. The UCX error was a separate problem: UCX finds the runner's Azure mana RDMA device, which it cannot use. UCX_TLS=tcp,sm,self is a fine fix for that, but it has no effect on the polling.

MPICH has a CVAR for exactly this. After a number of empty polls it calls usleep(1):

MPIR_CVAR_CH4_PROGRESS_THROTTLE: "1"
MPIR_CVAR_CH4_PROGRESS_THROTTLE_NO_PROGRESS_COUNT: "64"

I measured this locally with conda-forge MPICH 5.0.1. The test was a small loop of imbalanced compute plus Allreduce/Sendrecv/Barrier, with the ranks pinned to 4 cores using taskset to mimic the runner:

ranks on 4 cores	default	throttle only (count 4096)	throttle, count 64
4	0.29 s	0.31 s	0.30 s
6	14.7 s	7.2 s	0.71 s
8	22.0 s	16.5 s	0.89 s
10	–	–	0.99 s

6 ranks on 2 cores took 0.95 s with the same settings. Lowering the count is what matters. The default of 4096 only helps halfway, and 16–64 worked best. Throttling made no measurable difference when ranks ≤ cores. Two other settings did not help: MPIR_CVAR_POLLS_BEFORE_YIELD=10 (together with MPIR_CVAR_POSIX_POLL_FREQUENCY=10) still took 11 s at 6 ranks, and switching the netmod to OFI (MPIR_CVAR_CH4_NETMOD=ofi FI_PROVIDER=tcp) took 10 s.

So I would add these to the job's env: next to the UCX settings, only for the MPICH entries:

UCX_TLS: tcp,sm,self
UCX_LOG_LEVEL: error
MPIR_CVAR_CH4_PROGRESS_THROTTLE: "1"
MPIR_CVAR_CH4_PROGRESS_THROTTLE_NO_PROGRESS_COUNT: "64"

I have not run make ptest itself with these settings, so please let us know whether it gets back to the timings you see with ch3:sock.

---

Posted by Claude Code on behalf of @xylar.