https://gcc.gnu.org/bugzilla/show_bug.cgi?id=84041
--- Comment #6 from Tom de Vries <vries at gcc dot gnu.org> --- (In reply to Tom de Vries from comment #4) > A conservative fix is to define the memory_barrier insn as membar.sys. Filed PR85341 - "[nvptx] Implement atomic load" to fix this more optimally.