https://gcc.gnu.org/bugzilla/show_bug.cgi?id=120518
--- Comment #2 from sandra at gcc dot gnu.org --- OK, found it. The problem is the newly-added code that wraps the condition/device_num expr in a cleanup_point. This screws up the code in omp_device_num_check (in omp-general.cc) that detects special cases that refer to the host so that the selectors can be resolved statically. Not sure what the best solution is... either back out the cleanup_point addition entirely, or make it smarter, or make omp_device_num_check smarter. :-S
