After some kwalitee time with gdb, I isolated the problem in dotneato/dotgen/rank.c, the function collapse_sets() was using a Agraph_t (subg) to find the size of the cluster. This struct didn't have subg->u.drawing set up properly, so that when we got all the way down to actualwidth(), it was trying to find the width of the lable... and dumping core.

Here's a small patch that seems to fix the problem.

-philip
Input:

digraph JAAS {
    fontname=arial;
    fontsize=20;
    subgraph cluster0 {
	label="Foo";
        JOC [ label="JAAS" ];
    };
}

--- gv1.7.6/dotneato/dotgen/rank.c  Wed May  2 15:00:46 2001
+++ gv1.7.6.PG/dotneato/dotgen/rank.c   Sat Jul  7 03:29:09 2001
@@ -95,6 +95,7 @@
    graph_t     *mg,*subg;
    node_t      *mn,*n;
    edge_t      *me;
+   layout_t    *tl;

    mg = g->meta_node->graph;
    for (me = agfstout(mg,g->meta_node); me; me = agnxtout(mg,me)) {
@@ -103,7 +104,14 @@

        c = rank_set_class(subg);
        if (c) {
-           if ((c == CLUSTER) && CL_type == LOCAL) collapse_cluster(g,subg);
+           if ((c == CLUSTER) && CL_type == LOCAL) {
+               tl=subg->u.drawing;
+               if(subg->u.drawing==NULL)
+               subg->u.drawing=g->u.drawing;
+               collapse_cluster(g,subg);
+               subg->u.drawing=tl;
+
+           }
            else collapse_rankset(g,subg,c);
        }