I was doing some work with depth_first_search and came across some issues that seem surprising with the implementation. First vis.finish_edge doesn't seem to ever get called on msvc. Second changing the code to call finish_edge directly instead of through the call_finish_edge machinery seems to result in incorrect edges being finished. Third the implementation of depth_first_visit_impl pushes into a std::vector stack even for vertices with no out_edges. If I have a graph with a large number of vertices but few or no edges this seems bad. This seems to force a dependency on optional as well. Finally the loop as written should probably be a do..while since the stack will never be empty on the first test. I attached a patch that addresses the second and third issues. The tests continue to pass for me with this change.