DFS in boost::graph with changing the graphs content

Question

Minimal example:

#include <boost/graph/graph_traits.hpp>
#include <boost/graph/adjacency_list.hpp>
#include <boost/graph/depth_first_search.hpp>

struct vertex
{
    int number;
};
struct edge {};

typedef boost::adjacency_list<boost::listS, boost::vecS, boost::directedS, vertex, edge> graph_t;
typedef boost::graph_traits<graph_t>::vertex_descriptor vertex_t;
typedef boost::graph_traits<graph_t>::edge_descriptor edge_t;

struct vertex_visitor : public boost::default_dfs_visitor
{
    void discover_vertex(vertex_t v, graph_t& g)
    {
        g[v].number = 42;
    }
};

int main()
{
    graph_t g;
    vertex_t v1 = boost::add_vertex(g);
    vertex_t v2 = boost::add_vertex(g);
    boost::add_edge(v1, v2, g);

    vertex_visitor vis;
    boost::depth_first_search(g, boost::visitor(vis));

    return 0;
}

It does not work because the graph needs to be referenced as const in vertex_visitor::discover_vertex().

Is there any better method to do what I want than writing my own DFS algorithm (or using const_cast)? Also, does your solution allow to add/delete edges and vertices while discovering a vertex?

Answer 1

Have a look at how Boost implements connected_components. In order to store the component id the following visitor is used:

// This visitor is used both in the connected_components algorithm
// and in the kosaraju strong components algorithm during the
// second DFS traversal.
template <class ComponentsMap>
class components_recorder : public dfs_visitor<>
{
  typedef typename property_traits<ComponentsMap>::value_type comp_type;
public:
  components_recorder(ComponentsMap c, 
                      comp_type& c_count)
    : m_component(c), m_count(c_count) {}

  template <class Vertex, class Graph>
  void start_vertex(Vertex, Graph&) {
    if (m_count == (std::numeric_limits<comp_type>::max)())
      m_count = 0; // start counting components at zero
    else
      ++m_count;
  }
  template <class Vertex, class Graph>
  void discover_vertex(Vertex u, Graph&) {
    put(m_component, u, m_count);
  }
protected:
  ComponentsMap m_component;
  comp_type& m_count;
};

The idea is that a property map is passed to the constructor of the visitor, and is later used to update the data. Relating to your example, the vertex_visitor could be rewritten in the following way:

template <class PropertyMap>
struct vertex_visitor : public boost::dfs_visitor<>
{
  PropertyMap m_pmap;
  vertex_visitor(PropertyMap pmap) : m_pmap(pmap) {}
  template <class Vertex, class Graph>
  void discover_vertex(Vertex v, const Graph& g)
  {
    boost::put(m_pmap, v, 42);
  }
};

Instantiation of this visitor is a bit convoluted because we need to specify the property map type explicitly:

typedef boost::property_map<graph_t, int vertex::*>::type NumbersProperty;
vertex_visitor<NumbersProperty> vis(boost::get(&vertex::number, g));

As per the last part of the question, mutation of graph structure (i.e. addition or deletion of vertices and edges) invalidates interators, so this would break the DFS algorithm. I think this is precisely the reason why the graph is passed by const-reference.