I noticed our C++ test suite got 50% slower moving from iomp 2023 -> iomp 2026.
We use std::thread to run code which uses OMP parallel regions.
It looks like that libiomp5 does not clean up the the worker threads properly when the std thread joins. The result is that thousands of threads are being kept around during our test suite.
Using the iomp library of OneAPI 2025 the expected behavior is restored (the number of observed threads is equal to the OMP_NUM_THREADS specified).
A minimal reproducer and write-up can be found here: SoShiny/thread_and_omp_issue: Minimal reproducer for std::tread + OMP issue.
Please advise.
