While working on another project I noticed that there's a cache of shared object handles kept inside _PyImport_FindSharedFuncptr. See https://github.com/python/cpython/blob/b2b6cd00c6329426fc3b34700f2e22155b44168c/Python/dynload_shlib.c#L51-L55. It appears to be an optimization to work around poor caching of shared object handles in old libc implementations. After some testing, I have been unable to find any meaningful performance difference from this cache, so I propose we remove it to save the space.

My initial tests were on Linux (Ubuntu 18.04). I saw no discernible difference in the time for running the Python test suite with a single thread. Running the test suite using a single thread shows a lot of variance, but after running with and without the cache 40 times the mean times with/without the cache was nearly the same. Interpreter startup time also appears to be unaffected. This was all with a debug build, so I'm in the process of collecting data with a release build to see if that changes anything.

Proposed patch is in https://github.com/python/cpython/pull/25487.

New changeset fed2fc4443235fa9669b73817827fd6da88e3417 by Ian Henriksen in branch 'main':
bpo-43895: Remove an unnecessary cache of shared object handles (GH-25487)
https://github.com/python/cpython/commit/fed2fc4443235fa9669b73817827fd6da88e3417

Thanks Ian!