https://gcc.gnu.org/bugzilla/show_bug.cgi?id=89005
Bug ID: 89005 Summary: [nvptx, libgomp] Too much memory allocated in map_init Product: gcc Version: 9.0 Status: UNCONFIRMED Severity: enhancement Priority: P3 Component: target Assignee: unassigned at gcc dot gnu.org Reporter: vries at gcc dot gnu.org Target Milestone: --- Before "[nvptx] Remove use of CUDA unified memory in libgomp", in map_init we allocated a page-sized chunk of page-locked memory, which was used to provide the memory for all subsequent map_push calls. After mentioned commit, map_push allocates host and device memory, if needed. map_init though allocates an initial element, which happens to be ... paged-sized. While it's not wrong to allocate an initial element, although it's not clear to me whether it's beneficial, allocating one of this size is probably overdoing it. We should reduce the size of the initial element, or drop it altogether.