A saddle—point condition, that is both necessary and sufficient for a feasible point to be globally optimal, is given for a large class of nonconvex programming problems. The condition assumes lower semicontinuity of the feasible set point-to-set mapping.