To create a new layout plugin called xxx, you first need
to provide two functions: xxx_layout and xxx_cleanup. The
semantics of these are described below.

========================

  void xxx_layout(Agraph_t * g)

    Initialize the graph.
       - If the algorithm will use the common edge routing code, it should
    call setEdgeType (g, ...);

       - For each node, call common_init_node and gv_nodesize.

         If the algorithm will use spline_edges() to route the edges, the
    node coordinates need to be stored in ND_pos, so this should be
    allocated here. This, and the two calls mentioned above, are all
    handled by a call to neato_init_node().

       - For each edge, call common_init_edge

       - The algorithm should allocate whatever other data structures it
    needs. This can involve fields in the A*info_t fields. In addition,
    each of these fields contains a void* alg; subfield that the algorithm
    can use the store additional data. 
         Once we move to cgraph, this will all be replace with 
    algorithm specific records.

    Layout the graph. When finished, each node should have its coordinates
  stored in points in ND_coord_i(n), each edge should have its layout
  described in ED_spl(e). 
  (N.B. As of version 2.21, ND_coord_i has been replaced by ND_coord,
  which are now floating point coordinates.)

    To add edges, there are 3 functions available:

      - spline_edges1 (Agraph_t*, int edgeType)
          Assumes the node coordinates are stored in ND_coord_i, and that
        GD_bb is set. For each edge, this function constructs the appropriate 
        data and stores it in ED_spl.
      - spline_edges0 (Agraph_t*)
          Assumes the node coordinates are stored in ND_pos, and that
        GD_bb is set. This function uses the ratio attribute if set, 
        copies the values in ND_pos to ND_coord_i (converting from 
        inches to points); and calls spline_edges1 using the edge type
        specified by setEdgeType().
      - spline_edges (Agraph_t*)
          Assumes the node coordinates are stored in ND_pos. This
        function calculates the bounding box of g and stores it in GD_bb,
        then calls spline_edges0().

    If the algorithm only works with connected components, the code can
  use the pack library to get components, lay them out individually, and
  pack them together based on user specifications. A typical schema is
  given below. One can look at the code for twopi, circo, neato or fdp
  for more detailed examples.

        Agraph_t **ccs;
        Agraph_t *sg;
        Agnode_t *c = NULL;
        int ncc;
        int i;

        ccs = ccomps(g, &ncc, 0);
        if (ncc == 1) {
            /* layout nodes of g */
            adjustNodes(g);  /* if you need to remove overlaps */
            spline_edges(g); /* generic edge routing code */

        } else {
            pack_info pinfo;
            pack_mode pmode = getPackMode(g, l_node);

            for (i = 0; i < ncc; i++) {
                sg = ccs[i];
                /* layout sg */
                adjustNodes(sg);  /* if you need to remove overlaps */
            }
            spline_edges(g);  /* generic edge routing */

            /* initialize packing info, e.g. */
            pinfo.margin = getPack(g, CL_OFFSET, CL_OFFSET);
            pinfo.doSplines = 1;
            pinfo.mode = pmode;
            pinfo.fixed = 0;
            packSubgraphs(ncc, ccs, g, &pinfo);
        }
        for (i = 0; i < ncc; i++) {
            agdelete(g, ccs[i]);
        }

        free(ccs);

    Be careful in laying of subgraphs if you rely on attributes that have
  only been set in the root graph. With connected components, edges can
  be added with each component, before packing (as above) or after the
  components have been packed (see circo).

    It good to check for trivial cases where the graph has 0 or 1 nodes,
  or no edges.

    At the end of xxx_layout, call

       dotneato_postprocess(g);

  The following template will work in most cases, ignoring the problems of
  handling disconnected graphs and removing node overlaps:

  static void
  xxx_init_node(node_t * n)
  {
    neato_init_node(n);
    /* add algorithm-specific data, if desired */
  }

  static void
  xxx_init_edge(edge_t * e)
  {
    common_init_edge(e);
    /* add algorithm-specific data, if desired */
  }

  static void
  xxx_init_node_edge(graph_t * g)
  {
    node_t *n;
    edge_t *e;

    for (n = agfstnode(g); n; n = agnxtnode(g, n)) {
        xxx_init_node(n);
    }
    for (n = agfstnode(g); n; n = agnxtnode(g, n)) {
        for (e = agfstout(g, n); e; e = agnxtout(g, e)){          
            xxx_init_edge(e);
        }
    }
  }

  void
  xxx_layout (Agraph_t* g)
  {
    xxx_init_node_edge(g);
    /* Set ND_pos(n) for each node n */
    spline_edges(g);
    dotneato_postprocess(g);
  }  

======================

  void xxx_cleanup(Agraph_t * g)

    Free up any resources allocated in the layout.

    Finish with calls to gv_cleanup_node and gv_cleanup_edge for
  each node and edge. This cleans up splines labels, ND_pos, shapes
  and 0's out the A*info_t, so these have to occur last, but could be
  part of explicit xxx_cleanup_node and xxx_cleanup_edge, if desired.
    At the end, you should do

  if (g != g->root) memset(&(g->u), 0, sizeof(Agraphinfo_t));

  This is necessary for the graph to be laid out again, as the layout
  code assumes this structure is clean.

  libgvc does a final cleanup to the root graph, freeing any drawing,
  freeing its label, and zeroing out Agraphinfo_t of the root graph.

  The following template will work in most cases:

  static void xxx_cleanup_graph(Agraph_t * g)
  {
    /* Free any algorithm-specific data attached to the graph */
    if (g != g->root) memset(&(g->u), 0, sizeof(Agraphinfo_t));
  }

  static void xxx_cleanup_edge (Agedge_t* e)
  {
    /* Free any algorithm-specific data attached to the edge */
    gv_cleanup_edge(e);
  }

  static void xxx_cleanup_node (Agnode_t* n)
  {
    /* Free any algorithm-specific data attached to the node */
    gv_cleanup_node(e);
  }

  void xxx_cleanup(Agraph_t * g)
  {
    Agnode_t *n;
    Agedge_t *e;

    for (n = agfstnode(g); n; n = agnxtnode(g, n)) {
        for (e = agfstout(g, n); e; e = agnxtout(g, e)) {
            xxx_cleanup_edge(e);
        }
        xxx_cleanup_node(n);
    }
    xxx_cleanup_graph(g);
  }   

==================

Most layouts use auxiliary routines similar to neato, so
the entry points can be added in plugin/neato_layout

Add to gvlayout_neato_layout.c:

gvlayout_engine_t xxxgen_engine = {
    xxx_layout,
    xxx_cleanup,
};

and the line

    {LAYOUT_XXX, "xxx", 0, &xxxgen_engine, &neatogen_features},

to gvlayout_neato_types and a new emum

    LAYOUT_XXX

to layout_type in that file.

The above allows the new layout to piggyback on top of the neato
plugin, but requires rebuilding the plugin. In general, a user
can (and probably should) build a layout plugin totally separately. 

To do this, after writing xxx_layout and xxx_cleanup, it is necessary to:

  - add the types and data structures

typedef enum { LAYOUT_XXX } layout_type;

static gvlayout_features_t xxxgen_features = {
    0
};
gvlayout_engine_t xxxgen_engine = {
    xxx_layout,
    xxx_cleanup,
};
static gvplugin_installed_t gvlayout_xxx_types[] = {
    {LAYOUT_XXX, "xxx", 0, &xxxgen_engine, &xxxgen_features},
    {0, NULL, 0, NULL, NULL}
};
static gvplugin_api_t apis[] = {
    {API_layout, &gvlayout_xxx_types},
    {(api_t)0, 0},
};
gvplugin_library_t gvplugin_xxx_layout_LTX_library = { "xxx_layout", apis };

  - combine all of this into a dynamic library whose name contains the 
  string "gvplugin_" and install the library in the same directory as the 
  other Graphviz plugins. For example, on Linux systems, the dot layout 
  plugin is in the library libgvplugin_dot_layout.so.

  - run
      dot -c
  to regenerate the config file.

NOTES:
  - Additional layouts can be added as extra lines in gvlayout_xxx_types.
  - Obviously, most of the names and strings can be arbitrary. One
  constraint is that external identifier for the gvplugin_library_t
  type must end in "_LTX_library". In addition, the string "xxx" in
  each entry of gvlayout_xxx_types is the name used to identify the
  layout algorithm, so needs to be distinct from any other layout name.
  - The features of a layout algorithm are currently limited to a 
  flag of bits, and the only flag supported is LAYOUT_USES_RANKDIR,
  which enables the layout to the rankdir attribute.

Changes need to be made to any applications that
statically know about layout algorithms.

==================

Software configuration - automake

If you want to integrate your code into the Graphviz software
and use its build system, follow the instructions below. 
You can certainly build and install your plugin using your own
build software.

0. Put your software in lib/xxxgen, and added the hooks describe above
into gvlayout_neato_layout.c
1. In lib/xxxgen, provide a Makefile.am (based on a simple example
like lib/fdpgen/Makefile.am)
3. In lib/Makefile.am, add xxxgen to SUBDIRS
2. In configure.ac, add lib/xxxgen/Makefile to AC_CONFIG_FILES.
4. In lib/plugin/neato_layout/Makefile.am, insert
	$(top_builddir)/lib/xxxgen/libxxxgen_C.la 
	in libgvplugin_neato_layout_C_la_LIBADD
5. Remember to run autogen.sh because on its own configure can guess wrong.

This also assumes you have a good version of the various automake tools
on your system.