Libabigail compatibility

[bkmgit]

Libabigail enables checking of abi changes in libraries and whether applications are compatible making calls to a particular shared library. If a function has not changed, this can enable automated checking of whether a particular library version is compatible with a specific application even if the abi has changed since the functions used by the application will not have changed. Can the stub implementation be used to check for such changes, and would results of such a check be valid for standardized parts of other implementations?

[jeffhammond]

Good question. I don't know enough of how this works.

@raffenet @hzhou is MPICH covered by libabigail? How does it work?

[raffenet]

I haven't used libabigail before. Will check it out. MPICH has used https://lvc.github.io/abi-compliance-checker/ thus far to detect ABI issues.

[bkmgit]

Many projects now use particular commits of dependencies, Rust and Go projects in particular do this to increase developer velocity, though Swift has better dynamic linking features. MPI is relatively stable at present, but it may be helpful in future to be able to run a program on a program that uses MPI as a shared library and report back what ABIs it can use.

[raffenet]

I tried using the abidiff tool from libabigail. It did a good job finding which symbols were different between MPICH 4.3.0 and the stubs library. However I was surprised that it didn't report the signature difference MPI_Psend_init. Any abidiff users know if that can be detected?

[dalcinl]

However I was surprised that it didn't report the signature difference MPI_Psend_init

Under ELF and Mach-O, as far as I can tell, no tool could possible detect such ABI incompatibilities in a C library. The symbol names in shared libraries and executables are only built out of the function names, and the argument types are not used in any way for mangling. C++ would be a different story.

[dodjiseketeli]

I tried using the abidiff tool from libabigail. It did a good job finding which symbols were different between MPICH 4.3.0 and the stubs library. However I was surprised that it didn't report the signature difference MPI_Psend_init. Any abidiff users know if that can be detected?

Hello, have you compiled the binaries with debug info enabled? If yes, then abidiff should report any signature difference in C or C++, regardless. Otherwise, it's a bug and I'd be interested in getting my hands on the binaries.