https://gcc.gnu.org/bugzilla/show_bug.cgi?id=92311
--- Comment #2 from Jakub Jelinek <jakub at gcc dot gnu.org> --- Also, looking at that testcase, that doesn't look valid even in OpenMP 5.0. It is using use_device_ptr on an non-associated pointer, from that one gets a device pointer, but then it associates it with host variables and maps normally. I fail to see the point why it is trying to use use_device_ptr in this case at all, with use_device_ptr one gets something that can be passed through is_device_ptr or passed to specialized routines that can handle say CUDA device pointers.
