Linking fails with AOCC-5.0.0 compilers

[bourdin]

Thank you for taking the time to submit an issue!

Background information

What version of Open MPI are you using? (e.g., v4.1.6, v5.0.1, git branch name and hash, etc.)

v5.0.7

Describe how Open MPI was installed (e.g., from a source/distribution tarball, from a git clone, from an operating system distribution package, etc.)

Trying to build from source.

If you are building/installing from a git clone, please copy-n-paste the output from git submodule status.

Please describe the system on which you are running

rockylinux 9.5 (RHEL 9.5)

---

Details of the problem

Running configure with the following options:

./configure --prefix=/2/sw/HPC/openmpi/5.0.7 CC=clang FC=flang CXX=clang++ --with-slurm

I get the following error message while trying to link libmpi.so:

ld.lld: error: undefined reference due to --no-allow-shlib-undefined: __extendhfsf2
>>> referenced by ../../../ompi/.libs/libmpi.so

ld.lld: error: undefined reference due to --no-allow-shlib-undefined: __truncsfhf2
>>> referenced by ../../../ompi/.libs/libmpi.so
clang: error: linker command failed with exit code 1 (use -v to see invocation)
make[2]: *** [Makefile:1478: ompi_info] Error 1
make[2]: Leaving directory '/2/sw/HPC/src/openmpi-5.0.7/ompi/tools/ompi_info'
make[1]: *** [Makefile:2710: all-recursive] Error 1
make[1]: Leaving directory '/2/sw/HPC/src/openmpi-5.0.7/ompi'
make: *** [Makefile:1539: all-recursive] Error 1

Full report is at https://macdrive.mcmaster.ca/f/5879ebda7cfe49c283fe/

[ggouaillardet]

TL;DR configure --disable-alt-short-float ... (you won't have support for half precision though)

These symbols are used when using half precision float numbers.
They are not available in a shared library but only in .../lib/clang/17/lib/linux/libclang_rt.builtins-x86_64.a
An other workaround is to manually link ompi_info (and other binaries depending on libmpi.so) with that file.
I'll try to check if this is an unresolved issue in the latest LLVM, meanwhile, feel free to report this issue to AMD.

[ggouaillardet]

This is a more general issue with clang (could reproduce with upstream 17.0.6 and latest 20.1.2
The detection of short float (aka FP16) is not robust enough, I will issue a PR tomorrow to fix that.
If you do need short float, you can
configure --with-wrapper-ldflags=--rtlib=compiler-rt LDFLAGS="--rtlib=compiler-rt -lunwind" ...

passing -lunwind is necessary when building the pthread test program with the C++ compiler (!).
At first glance this looks odd to me and will likely report it upstream after I give it some more thoughts.