Merge pull request #140 from paulz/godep

add Go dependencies by godep

23 files changed, 6591 insertions, 0 deletions

diff --git a/vendor/github.com/soniakeys/graph/.gitignore b/vendor/github.com/soniakeys/graph/.gitignore
new file mode 100644
index 0000000..3be6158
--- /dev/null
+++ b/vendor/github.com/soniakeys/graph/.gitignore
@@ -0,0 +1,2 @@
+*.dot
+
diff --git a/vendor/github.com/soniakeys/graph/.travis.yml b/vendor/github.com/soniakeys/graph/.travis.yml
new file mode 100644
index 0000000..bcc4f9f
--- /dev/null
+++ b/vendor/github.com/soniakeys/graph/.travis.yml
@@ -0,0 +1,8 @@
+sudo: false
+language: go
+# update travis.sh when changing version number here
+go:
+ - 1.2.1
+ - 1.6
+install: go get -t ./...
+script: ./travis.sh
diff --git a/vendor/github.com/soniakeys/graph/adj.go b/vendor/github.com/soniakeys/graph/adj.go
new file mode 100644
index 0000000..165f365
--- /dev/null
+++ b/vendor/github.com/soniakeys/graph/adj.go
@@ -0,0 +1,325 @@
+// Copyright 2014 Sonia Keys
+// License MIT: https://opensource.org/licenses/MIT
+
+package graph
+
+// adj.go contains methods on AdjacencyList and LabeledAdjacencyList.
+//
+// AdjacencyList methods are placed first and are alphabetized.
+// LabeledAdjacencyList methods follow, also alphabetized.
+// Only exported methods need be alphabetized; non-exported methods can
+// be left near their use.
+
+import (
+	"math"
+	"sort"
+)
+
+// HasParallelSort identifies if a graph contains parallel arcs, multiple arcs
+// that lead from a node to the same node.
+//
+// If the graph has parallel arcs, the results fr and to represent an example
+// where there are parallel arcs from node fr to node to.
+//
+// If there are no parallel arcs, the method returns false -1 -1.
+//
+// Multiple loops on a node count as parallel arcs.
+//
+// "Sort" in the method name indicates that sorting is used to detect parallel
+// arcs.  Compared to method HasParallelMap, this may give better performance
+// for small or sparse graphs but will have asymtotically worse performance for
+// large dense graphs.
+func (g AdjacencyList) HasParallelSort() (has bool, fr, to NI) {
+	var t NodeList
+	for n, to := range g {
+		if len(to) == 0 {
+			continue
+		}
+		// different code in the labeled version, so no code gen.
+		t = append(t[:0], to...)
+		sort.Sort(t)
+		t0 := t[0]
+		for _, to := range t[1:] {
+			if to == t0 {
+				return true, NI(n), t0
+			}
+			t0 = to
+		}
+	}
+	return false, -1, -1
+}
+
+// IsUndirected returns true if g represents an undirected graph.
+//
+// Returns true when all non-loop arcs are paired in reciprocal pairs.
+// Otherwise returns false and an example unpaired arc.
+func (g AdjacencyList) IsUndirected() (u bool, from, to NI) {
+	// similar code in dot/writeUndirected
+	unpaired := make(AdjacencyList, len(g))
+	for fr, to := range g {
+	arc: // for each arc in g
+		for _, to := range to {
+			if to == NI(fr) {
+				continue // loop
+			}
+			// search unpaired arcs
+			ut := unpaired[to]
+			for i, u := range ut {
+				if u == NI(fr) { // found reciprocal
+					last := len(ut) - 1
+					ut[i] = ut[last]
+					unpaired[to] = ut[:last]
+					continue arc
+				}
+			}
+			// reciprocal not found
+			unpaired[fr] = append(unpaired[fr], to)
+		}
+	}
+	for fr, to := range unpaired {
+		if len(to) > 0 {
+			return false, NI(fr), to[0]
+		}
+	}
+	return true, -1, -1
+}
+
+// Edgelist constructs the edge list rerpresentation of a graph.
+//
+// An edge is returned for each arc of the graph.  For undirected graphs
+// this includes reciprocal edges.
+//
+// See also WeightedEdgeList method.
+func (g LabeledAdjacencyList) EdgeList() (el []LabeledEdge) {
+	for fr, to := range g {
+		for _, to := range to {
+			el = append(el, LabeledEdge{Edge{NI(fr), to.To}, to.Label})
+		}
+	}
+	return
+}
+
+// FloydWarshall finds all pairs shortest distances for a simple weighted
+// graph without negative cycles.
+//
+// In result array d, d[i][j] will be the shortest distance from node i
+// to node j.  Any diagonal element < 0 indicates a negative cycle exists.
+//
+// If g is an undirected graph with no negative edge weights, the result
+// array will be a distance matrix, for example as used by package
+// github.com/soniakeys/cluster.
+func (g LabeledAdjacencyList) FloydWarshall(w WeightFunc) (d [][]float64) {
+	d = newFWd(len(g))
+	for fr, to := range g {
+		for _, to := range to {
+			d[fr][to.To] = w(to.Label)
+		}
+	}
+	solveFW(d)
+	return
+}
+
+// little helper function, makes a blank matrix for FloydWarshall.
+func newFWd(n int) [][]float64 {
+	d := make([][]float64, n)
+	for i := range d {
+		di := make([]float64, n)
+		for j := range di {
+			if j != i {
+				di[j] = math.Inf(1)
+			}
+		}
+		d[i] = di
+	}
+	return d
+}
+
+// Floyd Warshall solver, once the matrix d is initialized by arc weights.
+func solveFW(d [][]float64) {
+	for k, dk := range d {
+		for _, di := range d {
+			dik := di[k]
+			for j := range d {
+				if d2 := dik + dk[j]; d2 < di[j] {
+					di[j] = d2
+				}
+			}
+		}
+	}
+}
+
+// HasArcLabel returns true if g has any arc from node fr to node to
+// with label l.
+//
+// Also returned is the index within the slice of arcs from node fr.
+// If no arc from fr to to is present, HasArcLabel returns false, -1.
+func (g LabeledAdjacencyList) HasArcLabel(fr, to NI, l LI) (bool, int) {
+	t := Half{to, l}
+	for x, h := range g[fr] {
+		if h == t {
+			return true, x
+		}
+	}
+	return false, -1
+}
+
+// HasParallelSort identifies if a graph contains parallel arcs, multiple arcs
+// that lead from a node to the same node.
+//
+// If the graph has parallel arcs, the results fr and to represent an example
+// where there are parallel arcs from node fr to node to.
+//
+// If there are no parallel arcs, the method returns -1 -1.
+//
+// Multiple loops on a node count as parallel arcs.
+//
+// "Sort" in the method name indicates that sorting is used to detect parallel
+// arcs.  Compared to method HasParallelMap, this may give better performance
+// for small or sparse graphs but will have asymtotically worse performance for
+// large dense graphs.
+func (g LabeledAdjacencyList) HasParallelSort() (has bool, fr, to NI) {
+	var t NodeList
+	for n, to := range g {
+		if len(to) == 0 {
+			continue
+		}
+		// slightly different code needed here compared to AdjacencyList
+		t = t[:0]
+		for _, to := range to {
+			t = append(t, to.To)
+		}
+		sort.Sort(t)
+		t0 := t[0]
+		for _, to := range t[1:] {
+			if to == t0 {
+				return true, NI(n), t0
+			}
+			t0 = to
+		}
+	}
+	return false, -1, -1
+}
+
+// IsUndirected returns true if g represents an undirected graph.
+//
+// Returns true when all non-loop arcs are paired in reciprocal pairs with
+// matching labels.  Otherwise returns false and an example unpaired arc.
+//
+// Note the requirement that reciprocal pairs have matching labels is
+// an additional test not present in the otherwise equivalent unlabeled version
+// of IsUndirected.
+func (g LabeledAdjacencyList) IsUndirected() (u bool, from NI, to Half) {
+	unpaired := make(LabeledAdjacencyList, len(g))
+	for fr, to := range g {
+	arc: // for each arc in g
+		for _, to := range to {
+			if to.To == NI(fr) {
+				continue // loop
+			}
+			// search unpaired arcs
+			ut := unpaired[to.To]
+			for i, u := range ut {
+				if u.To == NI(fr) && u.Label == to.Label { // found reciprocal
+					last := len(ut) - 1
+					ut[i] = ut[last]
+					unpaired[to.To] = ut[:last]
+					continue arc
+				}
+			}
+			// reciprocal not found
+			unpaired[fr] = append(unpaired[fr], to)
+		}
+	}
+	for fr, to := range unpaired {
+		if len(to) > 0 {
+			return false, NI(fr), to[0]
+		}
+	}
+	return true, -1, to
+}
+
+// NegativeArc returns true if the receiver graph contains a negative arc.
+func (g LabeledAdjacencyList) NegativeArc(w WeightFunc) bool {
+	for _, nbs := range g {
+		for _, nb := range nbs {
+			if w(nb.Label) < 0 {
+				return true
+			}
+		}
+	}
+	return false
+}
+
+// Unlabeled constructs the unlabeled graph corresponding to g.
+func (g LabeledAdjacencyList) Unlabeled() AdjacencyList {
+	a := make(AdjacencyList, len(g))
+	for n, nbs := range g {
+		to := make([]NI, len(nbs))
+		for i, nb := range nbs {
+			to[i] = nb.To
+		}
+		a[n] = to
+	}
+	return a
+}
+
+// WeightedEdgeList constructs a WeightedEdgeList object from a
+// LabeledAdjacencyList.
+//
+// Internally it calls g.EdgeList() to obtain the Edges member.
+// See LabeledAdjacencyList.EdgeList().
+func (g LabeledAdjacencyList) WeightedEdgeList(w WeightFunc) *WeightedEdgeList {
+	return &WeightedEdgeList{
+		Order:      len(g),
+		WeightFunc: w,
+		Edges:      g.EdgeList(),
+	}
+}
+
+// WeightedInDegree computes the weighted in-degree of each node in g
+// for a given weight function w.
+//
+// The weighted in-degree of a node is the sum of weights of arcs going to
+// the node.
+//
+// A weighted degree of a node is often termed the "strength" of a node.
+//
+// See note for undirected graphs at LabeledAdjacencyList.WeightedOutDegree.
+func (g LabeledAdjacencyList) WeightedInDegree(w WeightFunc) []float64 {
+	ind := make([]float64, len(g))
+	for _, to := range g {
+		for _, to := range to {
+			ind[to.To] += w(to.Label)
+		}
+	}
+	return ind
+}
+
+// WeightedOutDegree computes the weighted out-degree of the specified node
+// for a given weight function w.
+//
+// The weighted out-degree of a node is the sum of weights of arcs going from
+// the node.
+//
+// A weighted degree of a node is often termed the "strength" of a node.
+//
+// Note for undirected graphs, the WeightedOutDegree result for a node will
+// equal the WeightedInDegree for the node.  You can use WeightedInDegree if
+// you have need for the weighted degrees of all nodes or use WeightedOutDegree
+// to compute the weighted degrees of individual nodes.  In either case loops
+// are counted just once, unlike the (unweighted) UndirectedDegree methods.
+func (g LabeledAdjacencyList) WeightedOutDegree(n NI, w WeightFunc) (d float64) {
+	for _, to := range g[n] {
+		d += w(to.Label)
+	}
+	return
+}
+
+// More about loops and strength:  I didn't see consensus on this especially
+// in the case of undirected graphs.  Some sources said to add in-degree and
+// out-degree, which would seemingly double both loops and non-loops.
+// Some said to double loops.  Some said sum the edge weights and had no
+// comment on loops.  R of course makes everything an option.  The meaning
+// of "strength" where loops exist is unclear.  So while I could write an
+// UndirectedWeighted degree function that doubles loops but not edges,
+// I'm going to just leave this for now.
diff --git a/vendor/github.com/soniakeys/graph/adj_RO.go b/vendor/github.com/soniakeys/graph/adj_RO.go
new file mode 100644
index 0000000..1d37d14
--- /dev/null
+++ b/vendor/github.com/soniakeys/graph/adj_RO.go
@@ -0,0 +1,387 @@
+// Copyright 2014 Sonia Keys
+// License MIT: http://opensource.org/licenses/MIT
+
+package graph
+
+// adj_RO.go is code generated from adj_cg.go by directives in graph.go.
+// Editing adj_cg.go is okay.
+// DO NOT EDIT adj_RO.go.  The RO is for Read Only.
+
+import (
+	"math/rand"
+	"time"
+)
+
+// ArcSize returns the number of arcs in g.
+//
+// Note that for an undirected graph without loops, the number of undirected
+// edges -- the traditional meaning of graph size -- will be ArcSize()/2.
+// On the other hand, if g is an undirected graph that has or may have loops,
+// g.ArcSize()/2 is not a meaningful quantity.
+//
+// There are equivalent labeled and unlabeled versions of this method.
+func (g AdjacencyList) ArcSize() int {
+	m := 0
+	for _, to := range g {
+		m += len(to)
+	}
+	return m
+}
+
+// BoundsOk validates that all arcs in g stay within the slice bounds of g.
+//
+// BoundsOk returns true when no arcs point outside the bounds of g.
+// Otherwise it returns false and an example arc that points outside of g.
+//
+// Most methods of this package assume the BoundsOk condition and may
+// panic when they encounter an arc pointing outside of the graph.  This
+// function can be used to validate a graph when the BoundsOk condition
+// is unknown.
+//
+// There are equivalent labeled and unlabeled versions of this method.
+func (g AdjacencyList) BoundsOk() (ok bool, fr NI, to NI) {
+	for fr, to := range g {
+		for _, to := range to {
+			if to < 0 || to >= NI(len(g)) {
+				return false, NI(fr), to
+			}
+		}
+	}
+	return true, -1, to
+}
+
+// BreadthFirst traverses a directed or undirected graph in breadth first order.
+//
+// Argument start is the start node for the traversal.  If r is nil, nodes are
+// visited in deterministic order.  If a random number generator is supplied,
+// nodes at each level are visited in random order.
+//
+// Argument f can be nil if you have no interest in the FromList path result.
+// If FromList f is non-nil, the method populates f.Paths and sets f.MaxLen.
+// It does not set f.Leaves.  For convenience argument f can be a zero value
+// FromList.  If f.Paths is nil, the FromList is initialized first.  If f.Paths
+// is non-nil however, the FromList is  used as is.  The method uses a value of
+// PathEnd.Len == 0 to indentify unvisited nodes.  Existing non-zero values
+// will limit the traversal.
+//
+// Traversal calls the visitor function v for each node starting with node
+// start.  If v returns true, traversal continues.  If v returns false, the
+// traversal terminates immediately.  PathEnd Len and From values are updated
+// before calling the visitor function.
+//
+// On return f.Paths and f.MaxLen are set but not f.Leaves.
+//
+// Returned is the number of nodes visited and ok = true if the traversal
+// ran to completion or ok = false if it was terminated by the visitor
+// function returning false.
+//
+// There are equivalent labeled and unlabeled versions of this method.
+func (g AdjacencyList) BreadthFirst(start NI, r *rand.Rand, f *FromList, v OkNodeVisitor) (visited int, ok bool) {
+	switch {
+	case f == nil:
+		e := NewFromList(len(g))
+		f = &e
+	case f.Paths == nil:
+		*f = NewFromList(len(g))
+	}
+	rp := f.Paths
+	// the frontier consists of nodes all at the same level
+	frontier := []NI{start}
+	level := 1
+	// assign path when node is put on frontier,
+	rp[start] = PathEnd{Len: level, From: -1}
+	for {
+		f.MaxLen = level
+		level++
+		var next []NI
+		if r == nil {
+			for _, n := range frontier {
+				visited++
+				if !v(n) { // visit nodes as they come off frontier
+					return
+				}
+				for _, nb := range g[n] {
+					if rp[nb].Len == 0 {
+						next = append(next, nb)
+						rp[nb] = PathEnd{From: n, Len: level}
+					}
+				}
+			}
+		} else { // take nodes off frontier at random
+			for _, i := range r.Perm(len(frontier)) {
+				n := frontier[i]
+				// remainder of block same as above
+				visited++
+				if !v(n) {
+					return
+				}
+				for _, nb := range g[n] {
+					if rp[nb].Len == 0 {
+						next = append(next, nb)
+						rp[nb] = PathEnd{From: n, Len: level}
+					}
+				}
+			}
+		}
+		if len(next) == 0 {
+			break
+		}
+		frontier = next
+	}
+	return visited, true
+}
+
+// BreadthFirstPath finds a single path from start to end with a minimum
+// number of nodes.
+//
+// Returned is the path as list of nodes.
+// The result is nil if no path was found.
+//
+// There are equivalent labeled and unlabeled versions of this method.
+func (g AdjacencyList) BreadthFirstPath(start, end NI) []NI {
+	var f FromList
+	g.BreadthFirst(start, nil, &f, func(n NI) bool { return n != end })
+	return f.PathTo(end, nil)
+}
+
+// Copy makes a deep copy of g.
+// Copy also computes the arc size ma, the number of arcs.
+//
+// There are equivalent labeled and unlabeled versions of this method.
+func (g AdjacencyList) Copy() (c AdjacencyList, ma int) {
+	c = make(AdjacencyList, len(g))
+	for n, to := range g {
+		c[n] = append([]NI{}, to...)
+		ma += len(to)
+	}
+	return
+}
+
+// DepthFirst traverses a graph depth first.
+//
+// As it traverses it calls visitor function v for each node.  If v returns
+// false at any point, the traversal is terminated immediately and DepthFirst
+// returns false.  Otherwise DepthFirst returns true.
+//
+// DepthFirst uses argument bm is used as a bitmap to guide the traversal.
+// For a complete traversal, bm should be 0 initially.  During the
+// traversal, bits are set corresponding to each node visited.
+// The bit is set before calling the visitor function.
+//
+// Argument bm can be nil if you have no need for it.
+// In this case a bitmap is created internally for one-time use.
+//
+// Alternatively v can be nil.  In this case traversal still proceeds and
+// updates the bitmap, which can be a useful result.
+// DepthFirst always returns true in this case.
+//
+// It makes no sense for both bm and v to be nil.  In this case DepthFirst
+// returns false immediately.
+//
+// There are equivalent labeled and unlabeled versions of this method.
+func (g AdjacencyList) DepthFirst(start NI, bm *Bits, v OkNodeVisitor) (ok bool) {
+	if bm == nil {
+		if v == nil {
+			return false
+		}
+		bm = &Bits{}
+	}
+	var df func(n NI) bool
+	df = func(n NI) bool {
+		if bm.Bit(n) == 1 {
+			return true
+		}
+		bm.SetBit(n, 1)
+		if v != nil && !v(n) {
+			return false
+		}
+		for _, nb := range g[n] {
+			if !df(nb) {
+				return false
+			}
+		}
+		return true
+	}
+	return df(start)
+}
+
+// DepthFirstRandom traverses a graph depth first, but following arcs in
+// random order among arcs from a single node.
+//
+// If Rand r is nil, the method creates a new source and generator for
+// one-time use.
+//
+// Usage is otherwise like the DepthFirst method.  See DepthFirst.
+//
+// There are equivalent labeled and unlabeled versions of this method.
+func (g AdjacencyList) DepthFirstRandom(start NI, bm *Bits, v OkNodeVisitor, r *rand.Rand) (ok bool) {
+	if bm == nil {
+		if v == nil {
+			return false
+		}
+		bm = &Bits{}
+	}
+	if r == nil {
+		r = rand.New(rand.NewSource(time.Now().UnixNano()))
+	}
+	var df func(n NI) bool
+	df = func(n NI) bool {
+		if bm.Bit(n) == 1 {
+			return true
+		}
+		bm.SetBit(n, 1)
+		if v != nil && !v(n) {
+			return false
+		}
+		to := g[n]
+		for _, i := range r.Perm(len(to)) {
+			if !df(to[i]) {
+				return false
+			}
+		}
+		return true
+	}
+	return df(start)
+}
+
+// HasArc returns true if g has any arc from node fr to node to.
+//
+// Also returned is the index within the slice of arcs from node fr.
+// If no arc from fr to to is present, HasArc returns false, -1.
+//
+// There are equivalent labeled and unlabeled versions of this method.
+func (g AdjacencyList) HasArc(fr, to NI) (bool, int) {
+	for x, h := range g[fr] {
+		if h == to {
+			return true, x
+		}
+	}
+	return false, -1
+}
+
+// HasLoop identifies if a graph contains a loop, an arc that leads from a
+// a node back to the same node.
+//
+// If the graph has a loop, the result is an example node that has a loop.
+//
+// If g contains a loop, the method returns true and an example of a node
+// with a loop.  If there are no loops in g, the method returns false, -1.
+//
+// There are equivalent labeled and unlabeled versions of this method.
+func (g AdjacencyList) HasLoop() (bool, NI) {
+	for fr, to := range g {
+		for _, to := range to {
+			if NI(fr) == to {
+				return true, to
+			}
+		}
+	}
+	return false, -1
+}
+
+// HasParallelMap identifies if a graph contains parallel arcs, multiple arcs
+// that lead from a node to the same node.
+//
+// If the graph has parallel arcs, the method returns true and
+// results fr and to represent an example where there are parallel arcs
+// from node fr to node to.
+//
+// If there are no parallel arcs, the method returns false, -1 -1.
+//
+// Multiple loops on a node count as parallel arcs.
+//
+// "Map" in the method name indicates that a Go map is used to detect parallel
+// arcs.  Compared to method HasParallelSort, this gives better asymtotic
+// performance for large dense graphs but may have increased overhead for
+// small or sparse graphs.
+//
+// There are equivalent labeled and unlabeled versions of this method.
+func (g AdjacencyList) HasParallelMap() (has bool, fr, to NI) {
+	for n, to := range g {
+		if len(to) == 0 {
+			continue
+		}
+		m := map[NI]struct{}{}
+		for _, to := range to {
+			if _, ok := m[to]; ok {
+				return true, NI(n), to
+			}
+			m[to] = struct{}{}
+		}
+	}
+	return false, -1, -1
+}
+
+// IsSimple checks for loops and parallel arcs.
+//
+// A graph is "simple" if it has no loops or parallel arcs.
+//
+// IsSimple returns true, -1 for simple graphs.  If a loop or parallel arc is
+// found, simple returns false and a node that represents a counterexample
+// to the graph being simple.
+//
+// See also separate methods HasLoop and HasParallel.
+//
+// There are equivalent labeled and unlabeled versions of this method.
+func (g AdjacencyList) IsSimple() (ok bool, n NI) {
+	if lp, n := g.HasLoop(); lp {
+		return false, n
+	}
+	if pa, n, _ := g.HasParallelSort(); pa {
+		return false, n
+	}
+	return true, -1
+}
+
+// IsolatedNodes returns a bitmap of isolated nodes in receiver graph g.
+//
+// An isolated node is one with no arcs going to or from it.
+//
+// There are equivalent labeled and unlabeled versions of this method.
+func (g AdjacencyList) IsolatedNodes() (i Bits) {
+	i.SetAll(len(g))
+	for fr, to := range g {
+		if len(to) > 0 {
+			i.SetBit(NI(fr), 0)
+			for _, to := range to {
+				i.SetBit(to, 0)
+			}
+		}
+	}
+	return
+}
+
+/*
+MaxmimalClique finds a maximal clique containing the node n.
+
+Not sure this is good for anything.  It produces a single maximal clique
+but there can be multiple maximal cliques containing a given node.
+This algorithm just returns one of them, not even necessarily the
+largest one.
+
+func (g LabeledAdjacencyList) MaximalClique(n int) []int {
+	c := []int{n}
+	var m bitset.BitSet
+	m.Set(uint(n))
+	for fr, to := range g {
+		if fr == n {
+			continue
+		}
+		if len(to) < len(c) {
+			continue
+		}
+		f := 0
+		for _, to := range to {
+			if m.Test(uint(to.To)) {
+				f++
+				if f == len(c) {
+					c = append(c, to.To)
+					m.Set(uint(to.To))
+					break
+				}
+			}
+		}
+	}
+	return c
+}
+*/
diff --git a/vendor/github.com/soniakeys/graph/adj_cg.go b/vendor/github.com/soniakeys/graph/adj_cg.go
new file mode 100644
index 0000000..a484ee0
--- /dev/null
+++ b/vendor/github.com/soniakeys/graph/adj_cg.go
@@ -0,0 +1,387 @@
+// Copyright 2014 Sonia Keys
+// License MIT: http://opensource.org/licenses/MIT
+
+package graph
+
+// adj_RO.go is code generated from adj_cg.go by directives in graph.go.
+// Editing adj_cg.go is okay.
+// DO NOT EDIT adj_RO.go.  The RO is for Read Only.
+
+import (
+	"math/rand"
+	"time"
+)
+
+// ArcSize returns the number of arcs in g.
+//
+// Note that for an undirected graph without loops, the number of undirected
+// edges -- the traditional meaning of graph size -- will be ArcSize()/2.
+// On the other hand, if g is an undirected graph that has or may have loops,
+// g.ArcSize()/2 is not a meaningful quantity.
+//
+// There are equivalent labeled and unlabeled versions of this method.
+func (g LabeledAdjacencyList) ArcSize() int {
+	m := 0
+	for _, to := range g {
+		m += len(to)
+	}
+	return m
+}
+
+// BoundsOk validates that all arcs in g stay within the slice bounds of g.
+//
+// BoundsOk returns true when no arcs point outside the bounds of g.
+// Otherwise it returns false and an example arc that points outside of g.
+//
+// Most methods of this package assume the BoundsOk condition and may
+// panic when they encounter an arc pointing outside of the graph.  This
+// function can be used to validate a graph when the BoundsOk condition
+// is unknown.
+//
+// There are equivalent labeled and unlabeled versions of this method.
+func (g LabeledAdjacencyList) BoundsOk() (ok bool, fr NI, to Half) {
+	for fr, to := range g {
+		for _, to := range to {
+			if to.To < 0 || to.To >= NI(len(g)) {
+				return false, NI(fr), to
+			}
+		}
+	}
+	return true, -1, to
+}
+
+// BreadthFirst traverses a directed or undirected graph in breadth first order.
+//
+// Argument start is the start node for the traversal.  If r is nil, nodes are
+// visited in deterministic order.  If a random number generator is supplied,
+// nodes at each level are visited in random order.
+//
+// Argument f can be nil if you have no interest in the FromList path result.
+// If FromList f is non-nil, the method populates f.Paths and sets f.MaxLen.
+// It does not set f.Leaves.  For convenience argument f can be a zero value
+// FromList.  If f.Paths is nil, the FromList is initialized first.  If f.Paths
+// is non-nil however, the FromList is  used as is.  The method uses a value of
+// PathEnd.Len == 0 to indentify unvisited nodes.  Existing non-zero values
+// will limit the traversal.
+//
+// Traversal calls the visitor function v for each node starting with node
+// start.  If v returns true, traversal continues.  If v returns false, the
+// traversal terminates immediately.  PathEnd Len and From values are updated
+// before calling the visitor function.
+//
+// On return f.Paths and f.MaxLen are set but not f.Leaves.
+//
+// Returned is the number of nodes visited and ok = true if the traversal
+// ran to completion or ok = false if it was terminated by the visitor
+// function returning false.
+//
+// There are equivalent labeled and unlabeled versions of this method.
+func (g LabeledAdjacencyList) BreadthFirst(start NI, r *rand.Rand, f *FromList, v OkNodeVisitor) (visited int, ok bool) {
+	switch {
+	case f == nil:
+		e := NewFromList(len(g))
+		f = &e
+	case f.Paths == nil:
+		*f = NewFromList(len(g))
+	}
+	rp := f.Paths
+	// the frontier consists of nodes all at the same level
+	frontier := []NI{start}
+	level := 1
+	// assign path when node is put on frontier,
+	rp[start] = PathEnd{Len: level, From: -1}
+	for {
+		f.MaxLen = level
+		level++
+		var next []NI
+		if r == nil {
+			for _, n := range frontier {
+				visited++
+				if !v(n) { // visit nodes as they come off frontier
+					return
+				}
+				for _, nb := range g[n] {
+					if rp[nb.To].Len == 0 {
+						next = append(next, nb.To)
+						rp[nb.To] = PathEnd{From: n, Len: level}
+					}
+				}
+			}
+		} else { // take nodes off frontier at random
+			for _, i := range r.Perm(len(frontier)) {
+				n := frontier[i]
+				// remainder of block same as above
+				visited++
+				if !v(n) {
+					return
+				}
+				for _, nb := range g[n] {
+					if rp[nb.To].Len == 0 {
+						next = append(next, nb.To)
+						rp[nb.To] = PathEnd{From: n, Len: level}
+					}
+				}
+			}
+		}
+		if len(next) == 0 {
+			break
+		}
+		frontier = next
+	}
+	return visited, true
+}
+
+// BreadthFirstPath finds a single path from start to end with a minimum
+// number of nodes.
+//
+// Returned is the path as list of nodes.
+// The result is nil if no path was found.
+//
+// There are equivalent labeled and unlabeled versions of this method.
+func (g LabeledAdjacencyList) BreadthFirstPath(start, end NI) []NI {
+	var f FromList
+	g.BreadthFirst(start, nil, &f, func(n NI) bool { return n != end })
+	return f.PathTo(end, nil)
+}
+
+// Copy makes a deep copy of g.
+// Copy also computes the arc size ma, the number of arcs.
+//
+// There are equivalent labeled and unlabeled versions of this method.
+func (g LabeledAdjacencyList) Copy() (c LabeledAdjacencyList, ma int) {
+	c = make(LabeledAdjacencyList, len(g))
+	for n, to := range g {
+		c[n] = append([]Half{}, to...)
+		ma += len(to)
+	}
+	return
+}
+
+// DepthFirst traverses a graph depth first.
+//
+// As it traverses it calls visitor function v for each node.  If v returns
+// false at any point, the traversal is terminated immediately and DepthFirst
+// returns false.  Otherwise DepthFirst returns true.
+//
+// DepthFirst uses argument bm is used as a bitmap to guide the traversal.
+// For a complete traversal, bm should be 0 initially.  During the
+// traversal, bits are set corresponding to each node visited.
+// The bit is set before calling the visitor function.
+//
+// Argument bm can be nil if you have no need for it.
+// In this case a bitmap is created internally for one-time use.
+//
+// Alternatively v can be nil.  In this case traversal still proceeds and
+// updates the bitmap, which can be a useful result.
+// DepthFirst always returns true in this case.
+//
+// It makes no sense for both bm and v to be nil.  In this case DepthFirst
+// returns false immediately.
+//
+// There are equivalent labeled and unlabeled versions of this method.
+func (g LabeledAdjacencyList) DepthFirst(start NI, bm *Bits, v OkNodeVisitor) (ok bool) {
+	if bm == nil {
+		if v == nil {
+			return false
+		}
+		bm = &Bits{}
+	}
+	var df func(n NI) bool
+	df = func(n NI) bool {
+		if bm.Bit(n) == 1 {
+			return true
+		}
+		bm.SetBit(n, 1)
+		if v != nil && !v(n) {
+			return false
+		}
+		for _, nb := range g[n] {
+			if !df(nb.To) {
+				return false
+			}
+		}
+		return true
+	}
+	return df(start)
+}
+
+// DepthFirstRandom traverses a graph depth first, but following arcs in
+// random order among arcs from a single node.
+//
+// If Rand r is nil, the method creates a new source and generator for
+// one-time use.
+//
+// Usage is otherwise like the DepthFirst method.  See DepthFirst.
+//
+// There are equivalent labeled and unlabeled versions of this method.
+func (g LabeledAdjacencyList) DepthFirstRandom(start NI, bm *Bits, v OkNodeVisitor, r *rand.Rand) (ok bool) {
+	if bm == nil {
+		if v == nil {
+			return false
+		}
+		bm = &Bits{}
+	}
+	if r == nil {
+		r = rand.New(rand.NewSource(time.Now().UnixNano()))
+	}
+	var df func(n NI) bool
+	df = func(n NI) bool {
+		if bm.Bit(n) == 1 {
+			return true
+		}
+		bm.SetBit(n, 1)
+		if v != nil && !v(n) {
+			return false
+		}
+		to := g[n]
+		for _, i := range r.Perm(len(to)) {
+			if !df(to[i].To) {
+				return false
+			}
+		}
+		return true
+	}
+	return df(start)
+}
+
+// HasArc returns true if g has any arc from node fr to node to.
+//
+// Also returned is the index within the slice of arcs from node fr.
+// If no arc from fr to to is present, HasArc returns false, -1.
+//
+// There are equivalent labeled and unlabeled versions of this method.
+func (g LabeledAdjacencyList) HasArc(fr, to NI) (bool, int) {
+	for x, h := range g[fr] {
+		if h.To == to {
+			return true, x
+		}
+	}
+	return false, -1
+}
+
+// HasLoop identifies if a graph contains a loop, an arc that leads from a
+// a node back to the same node.
+//
+// If the graph has a loop, the result is an example node that has a loop.
+//
+// If g contains a loop, the method returns true and an example of a node
+// with a loop.  If there are no loops in g, the method returns false, -1.
+//
+// There are equivalent labeled and unlabeled versions of this method.
+func (g LabeledAdjacencyList) HasLoop() (bool, NI) {
+	for fr, to := range g {
+		for _, to := range to {
+			if NI(fr) == to.To {
+				return true, to.To
+			}
+		}
+	}
+	return false, -1
+}
+
+// HasParallelMap identifies if a graph contains parallel arcs, multiple arcs
+// that lead from a node to the same node.
+//
+// If the graph has parallel arcs, the method returns true and
+// results fr and to represent an example where there are parallel arcs
+// from node fr to node to.
+//
+// If there are no parallel arcs, the method returns false, -1 -1.
+//
+// Multiple loops on a node count as parallel arcs.
+//
+// "Map" in the method name indicates that a Go map is used to detect parallel
+// arcs.  Compared to method HasParallelSort, this gives better asymtotic
+// performance for large dense graphs but may have increased overhead for
+// small or sparse graphs.
+//
+// There are equivalent labeled and unlabeled versions of this method.
+func (g LabeledAdjacencyList) HasParallelMap() (has bool, fr, to NI) {
+	for n, to := range g {
+		if len(to) == 0 {
+			continue
+		}
+		m := map[NI]struct{}{}
+		for _, to := range to {
+			if _, ok := m[to.To]; ok {
+				return true, NI(n), to.To
+			}
+			m[to.To] = struct{}{}
+		}
+	}
+	return false, -1, -1
+}
+
+// IsSimple checks for loops and parallel arcs.
+//
+// A graph is "simple" if it has no loops or parallel arcs.
+//
+// IsSimple returns true, -1 for simple graphs.  If a loop or parallel arc is
+// found, simple returns false and a node that represents a counterexample
+// to the graph being simple.
+//
+// See also separate methods HasLoop and HasParallel.
+//
+// There are equivalent labeled and unlabeled versions of this method.
+func (g LabeledAdjacencyList) IsSimple() (ok bool, n NI) {
+	if lp, n := g.HasLoop(); lp {
+		return false, n
+	}
+	if pa, n, _ := g.HasParallelSort(); pa {
+		return false, n
+	}
+	return true, -1
+}
+
+// IsolatedNodes returns a bitmap of isolated nodes in receiver graph g.
+//
+// An isolated node is one with no arcs going to or from it.
+//
+// There are equivalent labeled and unlabeled versions of this method.
+func (g LabeledAdjacencyList) IsolatedNodes() (i Bits) {
+	i.SetAll(len(g))
+	for fr, to := range g {
+		if len(to) > 0 {
+			i.SetBit(NI(fr), 0)
+			for _, to := range to {
+				i.SetBit(to.To, 0)
+			}
+		}
+	}
+	return
+}
+
+/*
+MaxmimalClique finds a maximal clique containing the node n.
+
+Not sure this is good for anything.  It produces a single maximal clique
+but there can be multiple maximal cliques containing a given node.
+This algorithm just returns one of them, not even necessarily the
+largest one.
+
+func (g LabeledAdjacencyList) MaximalClique(n int) []int {
+	c := []int{n}
+	var m bitset.BitSet
+	m.Set(uint(n))
+	for fr, to := range g {
+		if fr == n {
+			continue
+		}
+		if len(to) < len(c) {
+			continue
+		}
+		f := 0
+		for _, to := range to {
+			if m.Test(uint(to.To)) {
+				f++
+				if f == len(c) {
+					c = append(c, to.To)
+					m.Set(uint(to.To))
+					break
+				}
+			}
+		}
+	}
+	return c
+}
+*/
diff --git a/vendor/github.com/soniakeys/graph/bits.go b/vendor/github.com/soniakeys/graph/bits.go
new file mode 100644
index 0000000..b86703c
--- /dev/null
+++ b/vendor/github.com/soniakeys/graph/bits.go
@@ -0,0 +1,207 @@
+// Copyright 2014 Sonia Keys
+// License MIT: http://opensource.org/licenses/MIT
+
+package graph
+
+import (
+	"fmt"
+	"math/big"
+)
+
+// Bits is bitmap, or bitset, intended to store a single bit of information
+// per node of a graph.
+//
+// The current implementation is backed by a big.Int and so is a reference
+// type in the same way a big.Int is.
+type Bits struct {
+	i big.Int
+}
+
+// NewBits constructs a Bits value with the bits ns set to 1.
+func NewBits(ns ...NI) (b Bits) {
+	for _, n := range ns {
+		b.SetBit(n, 1)
+	}
+	return
+}
+
+// AllNot sets n bits of z to the complement of x.
+//
+// It is a convenience method for SetAll followed by AndNot.
+func (z *Bits) AllNot(n int, x Bits) {
+	var y Bits
+	y.SetAll(n)
+	z.AndNot(y, x)
+}
+
+// And sets z = x & y.
+func (z *Bits) And(x, y Bits) {
+	z.i.And(&x.i, &y.i)
+}
+
+// AndNot sets z = x &^ y.
+func (z *Bits) AndNot(x, y Bits) {
+	z.i.AndNot(&x.i, &y.i)
+}
+
+// Bit returns the value of the n'th bit of x.
+func (b Bits) Bit(n NI) uint {
+	return b.i.Bit(int(n))
+}
+
+// Clear sets all bits to 0.
+func (z *Bits) Clear() {
+	*z = Bits{}
+}
+
+// Format satisfies fmt.Formatter for fmt.Printf and related methods.
+//
+// graph.Bits format exactly like big.Ints.
+func (b Bits) Format(s fmt.State, ch rune) {
+	b.i.Format(s, ch)
+}
+
+// From returns the position of the first 1 bit at or after (from) position n.
+//
+// It returns -1 if there is no one bit at or after position n.
+//
+// This provides one way to iterate over one bits.
+// To iterate over the one bits, call with n = 0 to get the the first
+// one bit, then call with the result + 1 to get successive one bits.
+// Unlike the Iterate method, this technique is stateless and so allows
+// bits to be changed between successive calls.
+//
+// See also Iterate.
+//
+// (From is just a short word that means "at or after" here;
+// it has nothing to do with arc direction.)
+func (b Bits) From(n NI) NI {
+	words := b.i.Bits()
+	i := int(n)
+	x := i >> wordExp // x now index of word containing bit i.
+	if x >= len(words) {
+		return -1
+	}
+	// test for 1 in this word at or after n
+	if wx := words[x] >> (uint(i) & (wordSize - 1)); wx != 0 {
+		return n + NI(trailingZeros(wx))
+	}
+	x++
+	for y, wy := range words[x:] {
+		if wy != 0 {
+			return NI((x+y)<<wordExp | trailingZeros(wy))
+		}
+	}
+	return -1
+}
+
+// Iterate calls Visitor v for each bit with a value of 1, in order
+// from lowest bit to highest bit.
+//
+// Iteration continues to the highest bit as long as v returns true.
+// It stops if v returns false.
+//
+// Iterate returns true normally.  It returns false if v returns false.
+//
+// Bit values should not be modified during iteration, by the visitor function
+// for example.  See From for an iteration method that allows modification.
+func (b Bits) Iterate(v OkNodeVisitor) bool {
+	for x, w := range b.i.Bits() {
+		if w != 0 {
+			t := trailingZeros(w)
+			i := t // index in w of next 1 bit
+			for {
+				if !v(NI(x<<wordExp | i)) {
+					return false
+				}
+				w >>= uint(t + 1)
+				if w == 0 {
+					break
+				}
+				t = trailingZeros(w)
+				i += 1 + t
+			}
+		}
+	}
+	return true
+}
+
+// Or sets z = x | y.
+func (z *Bits) Or(x, y Bits) {
+	z.i.Or(&x.i, &y.i)
+}
+
+// PopCount returns the number of 1 bits.
+func (b Bits) PopCount() (c int) {
+	// algorithm selected to be efficient for sparse bit sets.
+	for _, w := range b.i.Bits() {
+		for w != 0 {
+			w &= w - 1
+			c++
+		}
+	}
+	return
+}
+
+// Set sets the bits of z to the bits of x.
+func (z *Bits) Set(x Bits) {
+	z.i.Set(&x.i)
+}
+
+var one = big.NewInt(1)
+
+// SetAll sets z to have n 1 bits.
+//
+// It's useful for initializing z to have a 1 for each node of a graph.
+func (z *Bits) SetAll(n int) {
+	z.i.Sub(z.i.Lsh(one, uint(n)), one)
+}
+
+// SetBit sets the n'th bit to b, where be is a 0 or 1.
+func (z *Bits) SetBit(n NI, b uint) {
+	z.i.SetBit(&z.i, int(n), b)
+}
+
+// Single returns true if b has exactly one 1 bit.
+func (b Bits) Single() bool {
+	// like PopCount, but stop as soon as two are found
+	c := 0
+	for _, w := range b.i.Bits() {
+		for w != 0 {
+			w &= w - 1
+			c++
+			if c == 2 {
+				return false
+			}
+		}
+	}
+	return c == 1
+}
+
+// Slice returns a slice with the positions of each 1 bit.
+func (b Bits) Slice() (s []NI) {
+	// (alternative implementation might use Popcount and make to get the
+	// exact cap slice up front.  unclear if that would be better.)
+	b.Iterate(func(n NI) bool {
+		s = append(s, n)
+		return true
+	})
+	return
+}
+
+// Xor sets z = x ^ y.
+func (z *Bits) Xor(x, y Bits) {
+	z.i.Xor(&x.i, &y.i)
+}
+
+// Zero returns true if there are no 1 bits.
+func (b Bits) Zero() bool {
+	return len(b.i.Bits()) == 0
+}
+
+// trailingZeros returns the number of trailing 0 bits in v.
+//
+// If v is 0, it returns 0.
+func trailingZeros(v big.Word) int {
+	return deBruijnBits[v&-v*deBruijnMultiple>>deBruijnShift]
+}
diff --git a/vendor/github.com/soniakeys/graph/bits32.go b/vendor/github.com/soniakeys/graph/bits32.go
new file mode 100644
index 0000000..18e07f9
--- /dev/null
+++ b/vendor/github.com/soniakeys/graph/bits32.go
@@ -0,0 +1,23 @@
+// Copyright 2014 Sonia Keys
+// License MIT: http://opensource.org/licenses/MIT
+
+// +build 386 arm
+
+package graph
+
+// "word" here is math/big.Word
+const (
+	wordSize = 32
+	wordExp  = 5 // 2^5 = 32
+)
+
+// deBruijn magic numbers used in trailingZeros()
+//
+// reference: http://graphics.stanford.edu/~seander/bithacks.html
+const deBruijnMultiple = 0x077CB531
+const deBruijnShift = 27
+
+var deBruijnBits = []int{
+	0, 1, 28, 2, 29, 14, 24, 3, 30, 22, 20, 15, 25, 17, 4, 8,
+	31, 27, 13, 23, 21, 19, 16, 7, 26, 12, 18, 6, 11, 5, 10, 9,
+}
diff --git a/vendor/github.com/soniakeys/graph/bits64.go b/vendor/github.com/soniakeys/graph/bits64.go
new file mode 100644
index 0000000..ab601dd
--- /dev/null
+++ b/vendor/github.com/soniakeys/graph/bits64.go
@@ -0,0 +1,22 @@
+// Copyright 2014 Sonia Keys
+// License MIT: http://opensource.org/licenses/MIT
+
+// +build !386,!arm
+
+package graph
+
+const (
+	wordSize = 64
+	wordExp  = 6 // 2^6 = 64
+)
+
+// reference: http://graphics.stanford.edu/~seander/bithacks.html
+const deBruijnMultiple = 0x03f79d71b4ca8b09
+const deBruijnShift = 58
+
+var deBruijnBits = []int{
+	0, 1, 56, 2, 57, 49, 28, 3, 61, 58, 42, 50, 38, 29, 17, 4,
+	62, 47, 59, 36, 45, 43, 51, 22, 53, 39, 33, 30, 24, 18, 12, 5,
+	63, 55, 48, 27, 60, 41, 37, 16, 46, 35, 44, 21, 52, 32, 23, 11,
+	54, 26, 40, 15, 34, 20, 31, 10, 25, 14, 19, 9, 13, 8, 7, 6,
+}
diff --git a/vendor/github.com/soniakeys/graph/dir.go b/vendor/github.com/soniakeys/graph/dir.go
new file mode 100644
index 0000000..508306d
--- /dev/null
+++ b/vendor/github.com/soniakeys/graph/dir.go
@@ -0,0 +1,538 @@
+// Copyright 2014 Sonia Keys
+// License MIT: http://opensource.org/licenses/MIT
+
+package graph
+
+// dir.go has methods specific to directed graphs, types Directed and
+// LabeledDirected.
+//
+// Methods on Directed are first, with exported methods alphabetized.
+
+import "errors"
+
+// DAGMaxLenPath finds a maximum length path in a directed acyclic graph.
+//
+// Argument ordering must be a topological ordering of g.
+func (g Directed) DAGMaxLenPath(ordering []NI) (path []NI) {
+	// dynamic programming. visit nodes in reverse order. for each, compute
+	// longest path as one plus longest of 'to' nodes.
+	// Visits each arc once.  O(m).
+	//
+	// Similar code in label.go
+	var n NI
+	mlp := make([][]NI, len(g.AdjacencyList)) // index by node number
+	for i := len(ordering) - 1; i >= 0; i-- {
+		fr := ordering[i] // node number
+		to := g.AdjacencyList[fr]
+		if len(to) == 0 {
+			continue
+		}
+		mt := to[0]
+		for _, to := range to[1:] {
+			if len(mlp[to]) > len(mlp[mt]) {
+				mt = to
+			}
+		}
+		p := append([]NI{mt}, mlp[mt]...)
+		mlp[fr] = p
+		if len(p) > len(path) {
+			n = fr
+			path = p
+		}
+	}
+	return append([]NI{n}, path...)
+}
+
+// EulerianCycle finds an Eulerian cycle in a directed multigraph.
+//
+// * If g has no nodes, result is nil, nil.
+//
+// * If g is Eulerian, result is an Eulerian cycle with err = nil.
+// The cycle result is a list of nodes, where the first and last
+// nodes are the same.
+//
+// * Otherwise, result is nil, error
+//
+// Internally, EulerianCycle copies the entire graph g.
+// See EulerianCycleD for a more space efficient version.
+func (g Directed) EulerianCycle() ([]NI, error) {
+	c, m := g.Copy()
+	return c.EulerianCycleD(m)
+}
+
+// EulerianCycleD finds an Eulerian cycle in a directed multigraph.
+//
+// EulerianCycleD is destructive on its receiver g.  See EulerianCycle for
+// a non-destructive version.
+//
+// Argument ma must be the correct arc size, or number of arcs in g.
+//
+// * If g has no nodes, result is nil, nil.
+//
+// * If g is Eulerian, result is an Eulerian cycle with err = nil.
+// The cycle result is a list of nodes, where the first and last
+// nodes are the same.
+//
+// * Otherwise, result is nil, error
+func (g Directed) EulerianCycleD(ma int) ([]NI, error) {
+	if len(g.AdjacencyList) == 0 {
+		return nil, nil
+	}
+	e := newEulerian(g.AdjacencyList, ma)
+	for e.s >= 0 {
+		v := e.top() // v is node that starts cycle
+		e.push()
+		// if Eulerian, we'll always come back to starting node
+		if e.top() != v {
+			return nil, errors.New("not balanced")
+		}
+		e.keep()
+	}
+	if !e.uv.Zero() {
+		return nil, errors.New("not strongly connected")
+	}
+	return e.p, nil
+}
+
+// EulerianPath finds an Eulerian path in a directed multigraph.
+//
+// * If g has no nodes, result is nil, nil.
+//
+// * If g has an Eulerian path, result is an Eulerian path with err = nil.
+// The path result is a list of nodes, where the first node is start.
+//
+// * Otherwise, result is nil, error
+//
+// Internally, EulerianPath copies the entire graph g.
+// See EulerianPathD for a more space efficient version.
+func (g Directed) EulerianPath() ([]NI, error) {
+	ind := g.InDegree()
+	var start NI
+	for n, to := range g.AdjacencyList {
+		if len(to) > ind[n] {
+			start = NI(n)
+			break
+		}
+	}
+	c, m := g.Copy()
+	return c.EulerianPathD(m, start)
+}
+
+// EulerianPathD finds an Eulerian path in a directed multigraph.
+//
+// EulerianPathD is destructive on its receiver g.  See EulerianPath for
+// a non-destructive version.
+//
+// Argument ma must be the correct arc size, or number of arcs in g.
+// Argument start must be a valid start node for the path.
+//
+// * If g has no nodes, result is nil, nil.
+//
+// * If g has an Eulerian path, result is an Eulerian path with err = nil.
+// The path result is a list of nodes, where the first node is start.
+//
+// * Otherwise, result is nil, error
+func (g Directed) EulerianPathD(ma int, start NI) ([]NI, error) {
+	if len(g.AdjacencyList) == 0 {
+		return nil, nil
+	}
+	e := newEulerian(g.AdjacencyList, ma)
+	e.p[0] = start
+	// unlike EulerianCycle, the first path doesn't have be a cycle.
+	e.push()
+	e.keep()
+	for e.s >= 0 {
+		start = e.top()
+		e.push()
+		// paths after the first must be cycles though
+		// (as long as there are nodes on the stack)
+		if e.top() != start {
+			return nil, errors.New("no Eulerian path")
+		}
+		e.keep()
+	}
+	if !e.uv.Zero() {
+		return nil, errors.New("no Eulerian path")
+	}
+	return e.p, nil
+}
+
+// starting at the node on the top of the stack, follow arcs until stuck.
+// mark nodes visited, push nodes on stack, remove arcs from g.
+func (e *eulerian) push() {
+	for u := e.top(); ; {
+		e.uv.SetBit(u, 0) // reset unvisited bit
+		arcs := e.g[u]
+		if len(arcs) == 0 {
+			return // stuck
+		}
+		w := arcs[0] // follow first arc
+		e.s++        // push followed node on stack
+		e.p[e.s] = w
+		e.g[u] = arcs[1:] // consume arc
+		u = w
+	}
+}
+
+// like push, but for for undirected graphs.
+func (e *eulerian) pushUndir() {
+	for u := e.top(); ; {
+		e.uv.SetBit(u, 0)
+		arcs := e.g[u]
+		if len(arcs) == 0 {
+			return
+		}
+		w := arcs[0]
+		e.s++
+		e.p[e.s] = w
+		e.g[u] = arcs[1:] // consume arc
+		// here is the only difference, consume reciprocal arc as well:
+		a2 := e.g[w]
+		for x, rx := range a2 {
+			if rx == u { // here it is
+				last := len(a2) - 1
+				a2[x] = a2[last]   // someone else gets the seat
+				e.g[w] = a2[:last] // and it's gone.
+				break
+			}
+		}
+		u = w
+	}
+}
+
+// starting with the node on top of the stack, move nodes with no arcs.
+func (e *eulerian) keep() {
+	for e.s >= 0 {
+		n := e.top()
+		if len(e.g[n]) > 0 {
+			break
+		}
+		e.p[e.m] = n
+		e.s--
+		e.m--
+	}
+}
+
+type eulerian struct {
+	g  AdjacencyList // working copy of graph, it gets consumed
+	m  int           // number of arcs in g, updated as g is consumed
+	uv Bits          // unvisited
+	// low end of p is stack of unfinished nodes
+	// high end is finished path
+	p []NI // stack + path
+	s int  // stack pointer
+}
+
+func (e *eulerian) top() NI {
+	return e.p[e.s]
+}
+
+func newEulerian(g AdjacencyList, m int) *eulerian {
+	e := &eulerian{
+		g: g,
+		m: m,
+		p: make([]NI, m+1),
+	}
+	e.uv.SetAll(len(g))
+	return e
+}
+
+// MaximalNonBranchingPaths finds all paths in a directed graph that are
+// "maximal" and "non-branching".
+//
+// A non-branching path is one where path nodes other than the first and last
+// have exactly one arc leading to the node and one arc leading from the node,
+// thus there is no possibility to branch away to a different path.
+//
+// A maximal non-branching path cannot be extended to a longer non-branching
+// path by including another node at either end.
+//
+// In the case of a cyclic non-branching path, the first and last elements
+// of the path will be the same node, indicating an isolated cycle.
+//
+// The method calls the emit argument for each path or isolated cycle in g,
+// as long as emit returns true.  If emit returns false,
+// MaximalNonBranchingPaths returns immediately.
+func (g Directed) MaximalNonBranchingPaths(emit func([]NI) bool) {
+	ind := g.InDegree()
+	var uv Bits
+	uv.SetAll(len(g.AdjacencyList))
+	for v, vTo := range g.AdjacencyList {
+		if !(ind[v] == 1 && len(vTo) == 1) {
+			for _, w := range vTo {
+				n := []NI{NI(v), w}
+				uv.SetBit(NI(v), 0)
+				uv.SetBit(w, 0)
+				wTo := g.AdjacencyList[w]
+				for ind[w] == 1 && len(wTo) == 1 {
+					u := wTo[0]
+					n = append(n, u)
+					uv.SetBit(u, 0)
+					w = u
+					wTo = g.AdjacencyList[w]
+				}
+				if !emit(n) { // n is a path
+					return
+				}
+			}
+		}
+	}
+	// use uv.From rather than uv.Iterate.
+	// Iterate doesn't work here because we're modifying uv
+	for b := uv.From(0); b >= 0; b = uv.From(b + 1) {
+		v := NI(b)
+		n := []NI{v}
+		for w := v; ; {
+			w = g.AdjacencyList[w][0]
+			uv.SetBit(w, 0)
+			n = append(n, w)
+			if w == v {
+				break
+			}
+		}
+		if !emit(n) { // n is an isolated cycle
+			return
+		}
+	}
+}
+
+// Undirected returns copy of g augmented as needed to make it undirected.
+func (g Directed) Undirected() Undirected {
+	c, _ := g.AdjacencyList.Copy()                  // start with a copy
+	rw := make(AdjacencyList, len(g.AdjacencyList)) // "reciprocals wanted"
+	for fr, to := range g.AdjacencyList {
+	arc: // for each arc in g
+		for _, to := range to {
+			if to == NI(fr) {
+				continue // loop
+			}
+			// search wanted arcs
+			wf := rw[fr]
+			for i, w := range wf {
+				if w == to { // found, remove
+					last := len(wf) - 1
+					wf[i] = wf[last]
+					rw[fr] = wf[:last]
+					continue arc
+				}
+			}
+			// arc not found, add to reciprocal to wanted list
+			rw[to] = append(rw[to], NI(fr))
+		}
+	}
+	// add missing reciprocals
+	for fr, to := range rw {
+		c[fr] = append(c[fr], to...)
+	}
+	return Undirected{c}
+}
+
+// StronglyConnectedComponents identifies strongly connected components
+// in a directed graph.
+//
+// Algorithm by David J. Pearce, from "An Improved Algorithm for Finding the
+// Strongly Connected Components of a Directed Graph".  It is algorithm 3,
+// PEA_FIND_SCC2 in
+// http://homepages.mcs.vuw.ac.nz/~djp/files/P05.pdf, accessed 22 Feb 2015.
+//
+// Returned is a list of components, each component is a list of nodes.
+/*
+func (g Directed) StronglyConnectedComponents() []int {
+	rindex := make([]int, len(g))
+	S := []int{}
+	index := 1
+	c := len(g) - 1
+	visit := func(v int) {
+		root := true
+		rindex[v] = index
+		index++
+		for _, w := range g[v] {
+			if rindex[w] == 0 {
+				visit(w)
+			}
+			if rindex[w] < rindex[v] {
+				rindex[v] = rindex[w]
+				root = false
+			}
+		}
+		if root {
+			index--
+			for top := len(S) - 1; top >= 0 && rindex[v] <= rindex[top]; top-- {
+				w = rindex[top]
+				S = S[:top]
+				rindex[w] = c
+				index--
+			}
+			rindex[v] = c
+			c--
+		} else {
+			S = append(S, v)
+		}
+	}
+	for v := range g {
+		if rindex[v] == 0 {
+			visit(v)
+		}
+	}
+	return rindex
+}
+*/
+
+// Transpose constructs a new adjacency list with all arcs reversed.
+//
+// For every arc from->to of g, the result will have an arc to->from.
+// Transpose also counts arcs as it traverses and returns ma the number of arcs
+// in g (equal to the number of arcs in the result.)
+func (g Directed) Transpose() (t Directed, ma int) {
+	ta := make(AdjacencyList, len(g.AdjacencyList))
+	for n, nbs := range g.AdjacencyList {
+		for _, nb := range nbs {
+			ta[nb] = append(ta[nb], NI(n))
+			ma++
+		}
+	}
+	return Directed{ta}, ma
+}
+
+// DAGMaxLenPath finds a maximum length path in a directed acyclic graph.
+//
+// Length here means number of nodes or arcs, not a sum of arc weights.
+//
+// Argument ordering must be a topological ordering of g.
+//
+// Returned is a node beginning a maximum length path, and a path of arcs
+// starting from that node.
+func (g LabeledDirected) DAGMaxLenPath(ordering []NI) (n NI, path []Half) {
+	// dynamic programming. visit nodes in reverse order. for each, compute
+	// longest path as one plus longest of 'to' nodes.
+	// Visits each arc once.  Time complexity O(m).
+	//
+	// Similar code in dir.go.
+	mlp := make([][]Half, len(g.LabeledAdjacencyList)) // index by node number
+	for i := len(ordering) - 1; i >= 0; i-- {
+		fr := ordering[i] // node number
+		to := g.LabeledAdjacencyList[fr]
+		if len(to) == 0 {
+			continue
+		}
+		mt := to[0]
+		for _, to := range to[1:] {
+			if len(mlp[to.To]) > len(mlp[mt.To]) {
+				mt = to
+			}
+		}
+		p := append([]Half{mt}, mlp[mt.To]...)
+		mlp[fr] = p
+		if len(p) > len(path) {
+			n = fr
+			path = p
+		}
+	}
+	return
+}
+
+// FromListLabels transposes a labeled graph into a FromList and associated
+// list of labels.
+//
+// Receiver g should be connected as a tree or forest.  Specifically no node
+// can have multiple incoming arcs.  If any node n in g has multiple incoming
+// arcs, the method returns (nil, nil, n) where n is a node with multiple
+// incoming arcs.
+//
+// Otherwise (normally) the method populates the From members in a
+// FromList.Path, populates a slice of labels, and returns the FromList,
+// labels, and -1.
+//
+// Other members of the FromList are left as zero values.
+// Use FromList.RecalcLen and FromList.RecalcLeaves as needed.
+func (g LabeledDirected) FromListLabels() (*FromList, []LI, NI) {
+	labels := make([]LI, len(g.LabeledAdjacencyList))
+	paths := make([]PathEnd, len(g.LabeledAdjacencyList))
+	for i := range paths {
+		paths[i].From = -1
+	}
+	for fr, to := range g.LabeledAdjacencyList {
+		for _, to := range to {
+			if paths[to.To].From >= 0 {
+				return nil, nil, to.To
+			}
+			paths[to.To].From = NI(fr)
+			labels[to.To] = to.Label
+		}
+	}
+	return &FromList{Paths: paths}, labels, -1
+}
+
+// Transpose constructs a new adjacency list that is the transpose of g.
+//
+// For every arc from->to of g, the result will have an arc to->from.
+// Transpose also counts arcs as it traverses and returns ma the number of
+// arcs in g (equal to the number of arcs in the result.)
+func (g LabeledDirected) Transpose() (t LabeledDirected, ma int) {
+	ta := make(LabeledAdjacencyList, len(g.LabeledAdjacencyList))
+	for n, nbs := range g.LabeledAdjacencyList {
+		for _, nb := range nbs {
+			ta[nb.To] = append(ta[nb.To], Half{To: NI(n), Label: nb.Label})
+			ma++
+		}
+	}
+	return LabeledDirected{ta}, ma
+}
+
+// Undirected returns a new undirected graph derived from g, augmented as
+// needed to make it undirected, with reciprocal arcs having matching labels.
+func (g LabeledDirected) Undirected() LabeledUndirected {
+	c, _ := g.LabeledAdjacencyList.Copy() // start with a copy
+	// "reciprocals wanted"
+	rw := make(LabeledAdjacencyList, len(g.LabeledAdjacencyList))
+	for fr, to := range g.LabeledAdjacencyList {
+	arc: // for each arc in g
+		for _, to := range to {
+			if to.To == NI(fr) {
+				continue // arc is a loop
+			}
+			// search wanted arcs
+			wf := rw[fr]
+			for i, w := range wf {
+				if w == to { // found, remove
+					last := len(wf) - 1
+					wf[i] = wf[last]
+					rw[fr] = wf[:last]
+					continue arc
+				}
+			}
+			// arc not found, add to reciprocal to wanted list
+			rw[to.To] = append(rw[to.To], Half{To: NI(fr), Label: to.Label})
+		}
+	}
+	// add missing reciprocals
+	for fr, to := range rw {
+		c[fr] = append(c[fr], to...)
+	}
+	return LabeledUndirected{c}
+}
+
+// Unlabeled constructs the unlabeled directed graph corresponding to g.
+func (g LabeledDirected) Unlabeled() Directed {
+	return Directed{g.LabeledAdjacencyList.Unlabeled()}
+}
+
+// UnlabeledTranspose constructs a new adjacency list that is the unlabeled
+// transpose of g.
+//
+// For every arc from->to of g, the result will have an arc to->from.
+// Transpose also counts arcs as it traverses and returns ma, the number of
+// arcs in g (equal to the number of arcs in the result.)
+//
+// It is equivalent to g.Unlabeled().Transpose() but constructs the result
+// directly.
+func (g LabeledDirected) UnlabeledTranspose() (t Directed, ma int) {
+	ta := make(AdjacencyList, len(g.LabeledAdjacencyList))
+	for n, nbs := range g.LabeledAdjacencyList {
+		for _, nb := range nbs {
+			ta[nb.To] = append(ta[nb.To], NI(n))
+			ma++
+		}
+	}
+	return Directed{ta}, ma
+}
diff --git a/vendor/github.com/soniakeys/graph/dir_RO.go b/vendor/github.com/soniakeys/graph/dir_RO.go
new file mode 100644
index 0000000..77558a9
--- /dev/null
+++ b/vendor/github.com/soniakeys/graph/dir_RO.go
@@ -0,0 +1,395 @@
+// Copyright 2014 Sonia Keys
+// License MIT: http://opensource.org/licenses/MIT
+
+package graph
+
+// dir_RO.go is code generated from dir_cg.go by directives in graph.go.
+// Editing dir_cg.go is okay.  It is the code generation source.
+// DO NOT EDIT dir_RO.go.
+// The RO means read only and it is upper case RO to slow you down a bit
+// in case you start to edit the file.
+
+// Balanced returns true if for every node in g, in-degree equals out-degree.
+//
+// There are equivalent labeled and unlabeled versions of this method.
+func (g Directed) Balanced() bool {
+	for n, in := range g.InDegree() {
+		if in != len(g.AdjacencyList[n]) {
+			return false
+		}
+	}
+	return true
+}
+
+// Copy makes a deep copy of g.
+// Copy also computes the arc size ma, the number of arcs.
+//
+// There are equivalent labeled and unlabeled versions of this method.
+func (g Directed) Copy() (c Directed, ma int) {
+	l, s := g.AdjacencyList.Copy()
+	return Directed{l}, s
+}
+
+// Cyclic determines if g contains a cycle, a non-empty path from a node
+// back to itself.
+//
+// Cyclic returns true if g contains at least one cycle.  It also returns
+// an example of an arc involved in a cycle.
+// Cyclic returns false if g is acyclic.
+//
+// Also see Topological, which detects cycles.
+//
+// There are equivalent labeled and unlabeled versions of this method.
+func (g Directed) Cyclic() (cyclic bool, fr NI, to NI) {
+	a := g.AdjacencyList
+	fr, to = -1, -1
+	var temp, perm Bits
+	var df func(NI)
+	df = func(n NI) {
+		switch {
+		case temp.Bit(n) == 1:
+			cyclic = true
+			return
+		case perm.Bit(n) == 1:
+			return
+		}
+		temp.SetBit(n, 1)
+		for _, nb := range a[n] {
+			df(nb)
+			if cyclic {
+				if fr < 0 {
+					fr, to = n, nb
+				}
+				return
+			}
+		}
+		temp.SetBit(n, 0)
+		perm.SetBit(n, 1)
+	}
+	for n := range a {
+		if perm.Bit(NI(n)) == 1 {
+			continue
+		}
+		if df(NI(n)); cyclic { // short circuit as soon as a cycle is found
+			break
+		}
+	}
+	return
+}
+
+// FromList transposes a labeled graph into a FromList.
+//
+// Receiver g should be connected as a tree or forest.  Specifically no node
+// can have multiple incoming arcs.  If any node n in g has multiple incoming
+// arcs, the method returns (nil, n) where n is a node with multiple
+// incoming arcs.
+//
+// Otherwise (normally) the method populates the From members in a
+// FromList.Path and returns the FromList and -1.
+//
+// Other members of the FromList are left as zero values.
+// Use FromList.RecalcLen and FromList.RecalcLeaves as needed.
+//
+// Unusual cases are parallel arcs and loops.  A parallel arc represents
+// a case of multiple arcs going to some node and so will lead to a (nil, n)
+// return, even though a graph might be considered a multigraph tree.
+// A single loop on a node that would otherwise be a root node, though,
+// is not a case of multiple incoming arcs and so does not force a (nil, n)
+// result.
+//
+// There are equivalent labeled and unlabeled versions of this method.
+func (g Directed) FromList() (*FromList, NI) {
+	paths := make([]PathEnd, len(g.AdjacencyList))
+	for i := range paths {
+		paths[i].From = -1
+	}
+	for fr, to := range g.AdjacencyList {
+		for _, to := range to {
+			if paths[to].From >= 0 {
+				return nil, to
+			}
+			paths[to].From = NI(fr)
+		}
+	}
+	return &FromList{Paths: paths}, -1
+}
+
+// InDegree computes the in-degree of each node in g
+//
+// There are equivalent labeled and unlabeled versions of this method.
+func (g Directed) InDegree() []int {
+	ind := make([]int, len(g.AdjacencyList))
+	for _, nbs := range g.AdjacencyList {
+		for _, nb := range nbs {
+			ind[nb]++
+		}
+	}
+	return ind
+}
+
+// IsTree identifies trees in directed graphs.
+//
+// Return value isTree is true if the subgraph reachable from root is a tree.
+// Further, return value allTree is true if the entire graph g is reachable
+// from root.
+//
+// There are equivalent labeled and unlabeled versions of this method.
+func (g Directed) IsTree(root NI) (isTree, allTree bool) {
+	a := g.AdjacencyList
+	var v Bits
+	v.SetAll(len(a))
+	var df func(NI) bool
+	df = func(n NI) bool {
+		if v.Bit(n) == 0 {
+			return false
+		}
+		v.SetBit(n, 0)
+		for _, to := range a[n] {
+			if !df(to) {
+				return false
+			}
+		}
+		return true
+	}
+	isTree = df(root)
+	return isTree, isTree && v.Zero()
+}
+
+// Tarjan identifies strongly connected components in a directed graph using
+// Tarjan's algorithm.
+//
+// The method calls the emit argument for each component identified.  Each
+// component is a list of nodes.  A property of the algorithm is that
+// components are emitted in reverse topological order of the condensation.
+// (See https://en.wikipedia.org/wiki/Strongly_connected_component#Definitions
+// for description of condensation.)
+//
+// There are equivalent labeled and unlabeled versions of this method.
+//
+// See also TarjanForward and TarjanCondensation.
+func (g Directed) Tarjan(emit func([]NI) bool) {
+	// See "Depth-first search and linear graph algorithms", Robert Tarjan,
+	// SIAM J. Comput. Vol. 1, No. 2, June 1972.
+	//
+	// Implementation here from Wikipedia pseudocode,
+	// http://en.wikipedia.org/w/index.php?title=Tarjan%27s_strongly_connected_components_algorithm&direction=prev&oldid=647184742
+	var indexed, stacked Bits
+	a := g.AdjacencyList
+	index := make([]int, len(a))
+	lowlink := make([]int, len(a))
+	x := 0
+	var S []NI
+	var sc func(NI) bool
+	sc = func(n NI) bool {
+		index[n] = x
+		indexed.SetBit(n, 1)
+		lowlink[n] = x
+		x++
+		S = append(S, n)
+		stacked.SetBit(n, 1)
+		for _, nb := range a[n] {
+			if indexed.Bit(nb) == 0 {
+				if !sc(nb) {
+					return false
+				}
+				if lowlink[nb] < lowlink[n] {
+					lowlink[n] = lowlink[nb]
+				}
+			} else if stacked.Bit(nb) == 1 {
+				if index[nb] < lowlink[n] {
+					lowlink[n] = index[nb]
+				}
+			}
+		}
+		if lowlink[n] == index[n] {
+			var c []NI
+			for {
+				last := len(S) - 1
+				w := S[last]
+				S = S[:last]
+				stacked.SetBit(w, 0)
+				c = append(c, w)
+				if w == n {
+					if !emit(c) {
+						return false
+					}
+					break
+				}
+			}
+		}
+		return true
+	}
+	for n := range a {
+		if indexed.Bit(NI(n)) == 0 && !sc(NI(n)) {
+			return
+		}
+	}
+}
+
+// TarjanForward returns strongly connected components.
+//
+// It returns components in the reverse order of Tarjan, for situations
+// where a forward topological ordering is easier.
+func (g Directed) TarjanForward() [][]NI {
+	var r [][]NI
+	g.Tarjan(func(c []NI) bool {
+		r = append(r, c)
+		return true
+	})
+	scc := make([][]NI, len(r))
+	last := len(r) - 1
+	for i, ci := range r {
+		scc[last-i] = ci
+	}
+	return scc
+}
+
+// TarjanCondensation returns strongly connected components and their
+// condensation graph.
+//
+// Components are ordered in a forward topological ordering.
+func (g Directed) TarjanCondensation() (scc [][]NI, cd AdjacencyList) {
+	scc = g.TarjanForward()
+	cd = make(AdjacencyList, len(scc))       // return value
+	cond := make([]NI, len(g.AdjacencyList)) // mapping from g node to cd node
+	for cn := NI(len(scc) - 1); cn >= 0; cn-- {
+		c := scc[cn]
+		for _, n := range c {
+			cond[n] = NI(cn) // map g node to cd node
+		}
+		var tos []NI // list of 'to' nodes
+		var m Bits   // tos map
+		m.SetBit(cn, 1)
+		for _, n := range c {
+			for _, to := range g.AdjacencyList[n] {
+				if ct := cond[to]; m.Bit(ct) == 0 {
+					m.SetBit(ct, 1)
+					tos = append(tos, ct)
+				}
+			}
+		}
+		cd[cn] = tos
+	}
+	return
+}
+
+// Topological computes a topological ordering of a directed acyclic graph.
+//
+// For an acyclic graph, return value ordering is a permutation of node numbers
+// in topologically sorted order and cycle will be nil.  If the graph is found
+// to be cyclic, ordering will be nil and cycle will be the path of a found
+// cycle.
+//
+// There are equivalent labeled and unlabeled versions of this method.
+func (g Directed) Topological() (ordering, cycle []NI) {
+	a := g.AdjacencyList
+	ordering = make([]NI, len(a))
+	i := len(ordering)
+	var temp, perm Bits
+	var cycleFound bool
+	var cycleStart NI
+	var df func(NI)
+	df = func(n NI) {
+		switch {
+		case temp.Bit(n) == 1:
+			cycleFound = true
+			cycleStart = n
+			return
+		case perm.Bit(n) == 1:
+			return
+		}
+		temp.SetBit(n, 1)
+		for _, nb := range a[n] {
+			df(nb)
+			if cycleFound {
+				if cycleStart >= 0 {
+					// a little hack: orderng won't be needed so repurpose the
+					// slice as cycle.  this is read out in reverse order
+					// as the recursion unwinds.
+					x := len(ordering) - 1 - len(cycle)
+					ordering[x] = n
+					cycle = ordering[x:]
+					if n == cycleStart {
+						cycleStart = -1
+					}
+				}
+				return
+			}
+		}
+		temp.SetBit(n, 0)
+		perm.SetBit(n, 1)
+		i--
+		ordering[i] = n
+	}
+	for n := range a {
+		if perm.Bit(NI(n)) == 1 {
+			continue
+		}
+		df(NI(n))
+		if cycleFound {
+			return nil, cycle
+		}
+	}
+	return ordering, nil
+}
+
+// TopologicalKahn computes a topological ordering of a directed acyclic graph.
+//
+// For an acyclic graph, return value ordering is a permutation of node numbers
+// in topologically sorted order and cycle will be nil.  If the graph is found
+// to be cyclic, ordering will be nil and cycle will be the path of a found
+// cycle.
+//
+// This function is based on the algorithm by Arthur Kahn and requires the
+// transpose of g be passed as the argument.
+//
+// There are equivalent labeled and unlabeled versions of this method.
+func (g Directed) TopologicalKahn(tr Directed) (ordering, cycle []NI) {
+	// code follows Wikipedia pseudocode.
+	var L, S []NI
+	// rem for "remaining edges," this function makes a local copy of the
+	// in-degrees and consumes that instead of consuming an input.
+	rem := make([]int, len(g.AdjacencyList))
+	for n, fr := range tr.AdjacencyList {
+		if len(fr) == 0 {
+			// accumulate "set of all nodes with no incoming edges"
+			S = append(S, NI(n))
+		} else {
+			// initialize rem from in-degree
+			rem[n] = len(fr)
+		}
+	}
+	for len(S) > 0 {
+		last := len(S) - 1 // "remove a node n from S"
+		n := S[last]
+		S = S[:last]
+		L = append(L, n) // "add n to tail of L"
+		for _, m := range g.AdjacencyList[n] {
+			// WP pseudo code reads "for each node m..." but it means for each
+			// node m *remaining in the graph.*  We consume rem rather than
+			// the graph, so "remaining in the graph" for us means rem[m] > 0.
+			if rem[m] > 0 {
+				rem[m]--         // "remove edge from the graph"
+				if rem[m] == 0 { // if "m has no other incoming edges"
+					S = append(S, m) // "insert m into S"
+				}
+			}
+		}
+	}
+	// "If graph has edges," for us means a value in rem is > 0.
+	for c, in := range rem {
+		if in > 0 {
+			// recover cyclic nodes
+			for _, nb := range g.AdjacencyList[c] {
+				if rem[nb] > 0 {
+					cycle = append(cycle, NI(c))
+					break
+				}
+			}
+		}
+	}
+	if len(cycle) > 0 {
+		return nil, cycle
+	}
+	return L, nil
+}
diff --git a/vendor/github.com/soniakeys/graph/dir_cg.go b/vendor/github.com/soniakeys/graph/dir_cg.go
new file mode 100644
index 0000000..2b82f4f
--- /dev/null
+++ b/vendor/github.com/soniakeys/graph/dir_cg.go
@@ -0,0 +1,395 @@
+// Copyright 2014 Sonia Keys
+// License MIT: http://opensource.org/licenses/MIT
+
+package graph
+
+// dir_RO.go is code generated from dir_cg.go by directives in graph.go.
+// Editing dir_cg.go is okay.  It is the code generation source.
+// DO NOT EDIT dir_RO.go.
+// The RO means read only and it is upper case RO to slow you down a bit
+// in case you start to edit the file.
+
+// Balanced returns true if for every node in g, in-degree equals out-degree.
+//
+// There are equivalent labeled and unlabeled versions of this method.
+func (g LabeledDirected) Balanced() bool {
+	for n, in := range g.InDegree() {
+		if in != len(g.LabeledAdjacencyList[n]) {
+			return false
+		}
+	}
+	return true
+}
+
+// Copy makes a deep copy of g.
+// Copy also computes the arc size ma, the number of arcs.
+//
+// There are equivalent labeled and unlabeled versions of this method.
+func (g LabeledDirected) Copy() (c LabeledDirected, ma int) {
+	l, s := g.LabeledAdjacencyList.Copy()
+	return LabeledDirected{l}, s
+}
+
+// Cyclic determines if g contains a cycle, a non-empty path from a node
+// back to itself.
+//
+// Cyclic returns true if g contains at least one cycle.  It also returns
+// an example of an arc involved in a cycle.
+// Cyclic returns false if g is acyclic.
+//
+// Also see Topological, which detects cycles.
+//
+// There are equivalent labeled and unlabeled versions of this method.
+func (g LabeledDirected) Cyclic() (cyclic bool, fr NI, to Half) {
+	a := g.LabeledAdjacencyList
+	fr, to.To = -1, -1
+	var temp, perm Bits
+	var df func(NI)
+	df = func(n NI) {
+		switch {
+		case temp.Bit(n) == 1:
+			cyclic = true
+			return
+		case perm.Bit(n) == 1:
+			return
+		}
+		temp.SetBit(n, 1)
+		for _, nb := range a[n] {
+			df(nb.To)
+			if cyclic {
+				if fr < 0 {
+					fr, to = n, nb
+				}
+				return
+			}
+		}
+		temp.SetBit(n, 0)
+		perm.SetBit(n, 1)
+	}
+	for n := range a {
+		if perm.Bit(NI(n)) == 1 {
+			continue
+		}
+		if df(NI(n)); cyclic { // short circuit as soon as a cycle is found
+			break
+		}
+	}
+	return
+}
+
+// FromList transposes a labeled graph into a FromList.
+//
+// Receiver g should be connected as a tree or forest.  Specifically no node
+// can have multiple incoming arcs.  If any node n in g has multiple incoming
+// arcs, the method returns (nil, n) where n is a node with multiple
+// incoming arcs.
+//
+// Otherwise (normally) the method populates the From members in a
+// FromList.Path and returns the FromList and -1.
+//
+// Other members of the FromList are left as zero values.
+// Use FromList.RecalcLen and FromList.RecalcLeaves as needed.
+//
+// Unusual cases are parallel arcs and loops.  A parallel arc represents
+// a case of multiple arcs going to some node and so will lead to a (nil, n)
+// return, even though a graph might be considered a multigraph tree.
+// A single loop on a node that would otherwise be a root node, though,
+// is not a case of multiple incoming arcs and so does not force a (nil, n)
+// result.
+//
+// There are equivalent labeled and unlabeled versions of this method.
+func (g LabeledDirected) FromList() (*FromList, NI) {
+	paths := make([]PathEnd, len(g.LabeledAdjacencyList))
+	for i := range paths {
+		paths[i].From = -1
+	}
+	for fr, to := range g.LabeledAdjacencyList {
+		for _, to := range to {
+			if paths[to.To].From >= 0 {
+				return nil, to.To
+			}
+			paths[to.To].From = NI(fr)
+		}
+	}
+	return &FromList{Paths: paths}, -1
+}
+
+// InDegree computes the in-degree of each node in g
+//
+// There are equivalent labeled and unlabeled versions of this method.
+func (g LabeledDirected) InDegree() []int {
+	ind := make([]int, len(g.LabeledAdjacencyList))
+	for _, nbs := range g.LabeledAdjacencyList {
+		for _, nb := range nbs {
+			ind[nb.To]++
+		}
+	}
+	return ind
+}
+
+// IsTree identifies trees in directed graphs.
+//
+// Return value isTree is true if the subgraph reachable from root is a tree.
+// Further, return value allTree is true if the entire graph g is reachable
+// from root.
+//
+// There are equivalent labeled and unlabeled versions of this method.
+func (g LabeledDirected) IsTree(root NI) (isTree, allTree bool) {
+	a := g.LabeledAdjacencyList
+	var v Bits
+	v.SetAll(len(a))
+	var df func(NI) bool
+	df = func(n NI) bool {
+		if v.Bit(n) == 0 {
+			return false
+		}
+		v.SetBit(n, 0)
+		for _, to := range a[n] {
+			if !df(to.To) {
+				return false
+			}
+		}
+		return true
+	}
+	isTree = df(root)
+	return isTree, isTree && v.Zero()
+}
+
+// Tarjan identifies strongly connected components in a directed graph using
+// Tarjan's algorithm.
+//
+// The method calls the emit argument for each component identified.  Each
+// component is a list of nodes.  A property of the algorithm is that
+// components are emitted in reverse topological order of the condensation.
+// (See https://en.wikipedia.org/wiki/Strongly_connected_component#Definitions
+// for description of condensation.)
+//
+// There are equivalent labeled and unlabeled versions of this method.
+//
+// See also TarjanForward and TarjanCondensation.
+func (g LabeledDirected) Tarjan(emit func([]NI) bool) {
+	// See "Depth-first search and linear graph algorithms", Robert Tarjan,
+	// SIAM J. Comput. Vol. 1, No. 2, June 1972.
+	//
+	// Implementation here from Wikipedia pseudocode,
+	// http://en.wikipedia.org/w/index.php?title=Tarjan%27s_strongly_connected_components_algorithm&direction=prev&oldid=647184742
+	var indexed, stacked Bits
+	a := g.LabeledAdjacencyList
+	index := make([]int, len(a))
+	lowlink := make([]int, len(a))
+	x := 0
+	var S []NI
+	var sc func(NI) bool
+	sc = func(n NI) bool {
+		index[n] = x
+		indexed.SetBit(n, 1)
+		lowlink[n] = x
+		x++
+		S = append(S, n)
+		stacked.SetBit(n, 1)
+		for _, nb := range a[n] {
+			if indexed.Bit(nb.To) == 0 {
+				if !sc(nb.To) {
+					return false
+				}
+				if lowlink[nb.To] < lowlink[n] {
+					lowlink[n] = lowlink[nb.To]
+				}
+			} else if stacked.Bit(nb.To) == 1 {
+				if index[nb.To] < lowlink[n] {
+					lowlink[n] = index[nb.To]
+				}
+			}
+		}
+		if lowlink[n] == index[n] {
+			var c []NI
+			for {
+				last := len(S) - 1
+				w := S[last]
+				S = S[:last]
+				stacked.SetBit(w, 0)
+				c = append(c, w)
+				if w == n {
+					if !emit(c) {
+						return false
+					}
+					break
+				}
+			}
+		}
+		return true
+	}
+	for n := range a {
+		if indexed.Bit(NI(n)) == 0 && !sc(NI(n)) {
+			return
+		}
+	}
+}
+
+// TarjanForward returns strongly connected components.
+//
+// It returns components in the reverse order of Tarjan, for situations
+// where a forward topological ordering is easier.
+func (g LabeledDirected) TarjanForward() [][]NI {
+	var r [][]NI
+	g.Tarjan(func(c []NI) bool {
+		r = append(r, c)
+		return true
+	})
+	scc := make([][]NI, len(r))
+	last := len(r) - 1
+	for i, ci := range r {
+		scc[last-i] = ci
+	}
+	return scc
+}
+
+// TarjanCondensation returns strongly connected components and their
+// condensation graph.
+//
+// Components are ordered in a forward topological ordering.
+func (g LabeledDirected) TarjanCondensation() (scc [][]NI, cd AdjacencyList) {
+	scc = g.TarjanForward()
+	cd = make(AdjacencyList, len(scc))              // return value
+	cond := make([]NI, len(g.LabeledAdjacencyList)) // mapping from g node to cd node
+	for cn := NI(len(scc) - 1); cn >= 0; cn-- {
+		c := scc[cn]
+		for _, n := range c {
+			cond[n] = NI(cn) // map g node to cd node
+		}
+		var tos []NI // list of 'to' nodes
+		var m Bits   // tos map
+		m.SetBit(cn, 1)
+		for _, n := range c {
+			for _, to := range g.LabeledAdjacencyList[n] {
+				if ct := cond[to.To]; m.Bit(ct) == 0 {
+					m.SetBit(ct, 1)
+					tos = append(tos, ct)
+				}
+			}
+		}
+		cd[cn] = tos
+	}
+	return
+}
+
+// Topological computes a topological ordering of a directed acyclic graph.
+//
+// For an acyclic graph, return value ordering is a permutation of node numbers
+// in topologically sorted order and cycle will be nil.  If the graph is found
+// to be cyclic, ordering will be nil and cycle will be the path of a found
+// cycle.
+//
+// There are equivalent labeled and unlabeled versions of this method.
+func (g LabeledDirected) Topological() (ordering, cycle []NI) {
+	a := g.LabeledAdjacencyList
+	ordering = make([]NI, len(a))
+	i := len(ordering)
+	var temp, perm Bits
+	var cycleFound bool
+	var cycleStart NI
+	var df func(NI)
+	df = func(n NI) {
+		switch {
+		case temp.Bit(n) == 1:
+			cycleFound = true
+			cycleStart = n
+			return
+		case perm.Bit(n) == 1:
+			return
+		}
+		temp.SetBit(n, 1)
+		for _, nb := range a[n] {
+			df(nb.To)
+			if cycleFound {
+				if cycleStart >= 0 {
+					// a little hack: orderng won't be needed so repurpose the
+					// slice as cycle.  this is read out in reverse order
+					// as the recursion unwinds.
+					x := len(ordering) - 1 - len(cycle)
+					ordering[x] = n
+					cycle = ordering[x:]
+					if n == cycleStart {
+						cycleStart = -1
+					}
+				}
+				return
+			}
+		}
+		temp.SetBit(n, 0)
+		perm.SetBit(n, 1)
+		i--
+		ordering[i] = n
+	}
+	for n := range a {
+		if perm.Bit(NI(n)) == 1 {
+			continue
+		}
+		df(NI(n))
+		if cycleFound {
+			return nil, cycle
+		}
+	}
+	return ordering, nil
+}
+
+// TopologicalKahn computes a topological ordering of a directed acyclic graph.
+//
+// For an acyclic graph, return value ordering is a permutation of node numbers
+// in topologically sorted order and cycle will be nil.  If the graph is found
+// to be cyclic, ordering will be nil and cycle will be the path of a found
+// cycle.
+//
+// This function is based on the algorithm by Arthur Kahn and requires the
+// transpose of g be passed as the argument.
+//
+// There are equivalent labeled and unlabeled versions of this method.
+func (g LabeledDirected) TopologicalKahn(tr Directed) (ordering, cycle []NI) {
+	// code follows Wikipedia pseudocode.
+	var L, S []NI
+	// rem for "remaining edges," this function makes a local copy of the
+	// in-degrees and consumes that instead of consuming an input.
+	rem := make([]int, len(g.LabeledAdjacencyList))
+	for n, fr := range tr.AdjacencyList {
+		if len(fr) == 0 {
+			// accumulate "set of all nodes with no incoming edges"
+			S = append(S, NI(n))
+		} else {
+			// initialize rem from in-degree
+			rem[n] = len(fr)
+		}
+	}
+	for len(S) > 0 {
+		last := len(S) - 1 // "remove a node n from S"
+		n := S[last]
+		S = S[:last]
+		L = append(L, n) // "add n to tail of L"
+		for _, m := range g.LabeledAdjacencyList[n] {
+			// WP pseudo code reads "for each node m..." but it means for each
+			// node m *remaining in the graph.*  We consume rem rather than
+			// the graph, so "remaining in the graph" for us means rem[m] > 0.
+			if rem[m.To] > 0 {
+				rem[m.To]--         // "remove edge from the graph"
+				if rem[m.To] == 0 { // if "m has no other incoming edges"
+					S = append(S, m.To) // "insert m into S"
+				}
+			}
+		}
+	}
+	// "If graph has edges," for us means a value in rem is > 0.
+	for c, in := range rem {
+		if in > 0 {
+			// recover cyclic nodes
+			for _, nb := range g.LabeledAdjacencyList[c] {
+				if rem[nb.To] > 0 {
+					cycle = append(cycle, NI(c))
+					break
+				}
+			}
+		}
+	}
+	if len(cycle) > 0 {
+		return nil, cycle
+	}
+	return L, nil
+}
diff --git a/vendor/github.com/soniakeys/graph/doc.go b/vendor/github.com/soniakeys/graph/doc.go
new file mode 100644
index 0000000..6d07278
--- /dev/null
+++ b/vendor/github.com/soniakeys/graph/doc.go
@@ -0,0 +1,128 @@
+// Copyright 2014 Sonia Keys
+// License MIT: http://opensource.org/licenses/MIT
+
+// Graph algorithms: Dijkstra, A*, Bellman Ford, Floyd Warshall;
+// Kruskal and Prim minimal spanning tree; topological sort and DAG longest
+// and shortest paths; Eulerian cycle and path; degeneracy and k-cores;
+// Bron Kerbosch clique finding; connected components; and others.
+//
+// This is a graph library of integer indexes.  To use it with application
+// data, you associate data with integer indexes, perform searches or other
+// operations with the library, and then use the integer index results to refer
+// back to your application data.
+//
+// Thus it does not store application data, pointers to application data,
+// or require you to implement an interface on your application data.
+// The idea is to keep the library methods fast and lean.
+//
+// Representation overview
+//
+// The package defines a type for a node index (NI) which is just an integer
+// type.  It defines types for a number of number graph representations using
+// NI.  The fundamental graph type is AdjacencyList, which is the
+// common "list of lists" graph representation.  It is a list as a slice
+// with one element for each node of the graph.  Each element is a list
+// itself, a list of neighbor nodes, implemented as an NI slice.  Methods
+// on an AdjacencyList generally work on any representable graph, including
+// directed or undirected graphs, simple graphs or multigraphs.
+//
+// The type Undirected embeds an AdjacencyList adding methods specific to
+// undirected graphs.  Similarly the type Directed adds methods meaningful
+// for directed graphs.
+//
+// Similar to NI, the type LI is a "label index" which labels a
+// node-to-neighbor "arc" or edge.  Just as an NI can index arbitrary node
+// data, an LI can index arbitrary arc or edge data.  A number of algorithms
+// use a "weight" associated with an arc.  This package does not represent
+// weighted arcs explicitly, but instead uses the LI as a more general
+// mechanism allowing not only weights but arbitrary data to be associated
+// with arcs.  While AdjacencyList represents an arc with simply an NI,
+// the type LabeledAdjacencyList uses a type that pairs an NI with an LI.
+// This type is named Half, for half-arc.  (A full arc would represent
+// both ends.)  Types LabeledDirected and LabeledUndirected embed a
+// LabeledAdjacencyList.
+//
+// In contrast to Half, the type Edge represents both ends of an edge (but
+// no label.)  The type LabeledEdge adds the label.  The type WeightedEdgeList
+// bundles a list of LabeledEdges with a WeightFunc.  WeightedEdgeList is
+// currently only used by Kruskal methods.
+//
+// FromList is a compact rooted tree (or forest) respresentation.  Like
+// AdjacencyList and LabeledAdjacencyList, it is a list with one element for
+// each node of the graph.  Each element contains only a single neighbor
+// however, its parent in the tree, the "from" node.
+//
+// Code generation
+//
+// A number of methods on AdjacencyList, Directed, and Undirected are
+// applicable to LabeledAdjacencyList, LabeledDirected, and LabeledUndirected
+// simply by ignoring the label.  In these cases code generation provides
+// methods on both types from a single source implementation. These methods
+// are documented with the sentence "There are equivalent labeled and unlabeled
+// versions of this method" and examples are provided only for the unlabeled
+// version.
+//
+// Terminology
+//
+// This package uses the term "node" rather than "vertex."  It uses "arc"
+// to mean a directed edge, and uses "from" and "to" to refer to the ends
+// of an arc.  It uses "start" and "end" to refer to endpoints of a search
+// or traversal.
+//
+// The usage of "to" and "from" is perhaps most strange.  In common speech
+// they are prepositions, but throughout this package they are used as
+// adjectives, for example to refer to the "from node" of an arc or the
+// "to node".  The type "FromList" is named to indicate it stores a list of
+// "from" values.
+//
+// A "half arc" refers to just one end of an arc, either the to or from end.
+//
+// Two arcs are "reciprocal" if they connect two distinct nodes n1 and n2,
+// one arc leading from n1 to n2 and the other arc leading from n2 to n1.
+// Undirected graphs are represented with reciprocal arcs.
+//
+// A node with an arc to itself represents a "loop."  Duplicate arcs, where
+// a node has multiple arcs to another node, are termed "parallel arcs."
+// A graph with no loops or parallel arcs is "simple."  A graph that allows
+// parallel arcs is a "multigraph"
+//
+// The "size" of a graph traditionally means the number of undirected edges.
+// This package uses "arc size" to mean the number of arcs in a graph.  For an
+// undirected graph without loops, arc size is 2 * size.
+//
+// The "order" of a graph is the number of nodes.  An "ordering" though means
+// an ordered list of nodes.
+//
+// A number of graph search algorithms use a concept of arc "weights."
+// The sum of arc weights along a path is a "distance."  In contrast, the
+// number of nodes in a path, including start and end nodes, is the path's
+// "length."  (Yes, mixing weights and lengths would be nonsense physically,
+// but the terms used here are just distinct terms for abstract values.
+// The actual meaning to an application is likely to be something else
+// entirely and is not relevant within this package.)
+//
+// Finally, this package documentation takes back the word "object" in some
+// places to refer to a Go value, especially a value of a type with methods.
+//
+// Shortest path searches
+//
+// This package implements a number of shortest path searches.  Most work
+// with weighted graphs that are directed or undirected, and with graphs
+// that may have loops or parallel arcs.  For weighted graphs, "shortest"
+// is defined as the path distance (sum of arc weights) with path length
+// (number of nodes) breaking ties.  If multiple paths have the same minimum
+// distance with the same minimum length, search methods are free to return
+// any of them.
+//
+//  Type name      Description, methods
+//  BreadthFirst   Unweigted arcs, traversal, single path search or all paths.
+//  BreadthFirst2  Direction-optimizing variant of BreadthFirst.
+//  Dijkstra       Non-negative arc weights, single or all paths.
+//  AStar          Non-negative arc weights, heuristic guided, single path.
+//  BellmanFord    Negative arc weights allowed, no negative cycles, all paths.
+//  DAGPath        O(n) algorithm for DAGs, arc weights of any sign.
+//  FloydWarshall  all pairs distances, no negative cycles.
+//
+// These searches typically have one method that is full-featured and
+// then a convenience method with a simpler API targeting a simpler use case.
+package graph
diff --git a/vendor/github.com/soniakeys/graph/fromlist.go b/vendor/github.com/soniakeys/graph/fromlist.go
new file mode 100644
index 0000000..31d41fa
--- /dev/null
+++ b/vendor/github.com/soniakeys/graph/fromlist.go
@@ -0,0 +1,418 @@
+// Copyright 2014 Sonia Keys
+// License MIT: http://opensource.org/licenses/MIT
+
+package graph
+
+// FromList represents a rooted tree (or forest) where each node is associated
+// with a half arc identifying an arc "from" another node.
+//
+// Other terms for this data structure include "parent list",
+// "predecessor list", "in-tree", "inverse arborescence", and
+// "spaghetti stack."
+//
+// The Paths member represents the tree structure.  Leaves and MaxLen are
+// not always needed.  Where Leaves is used it serves as a bitmap where
+// Leaves.Bit(n) == 1 for each leaf n of the tree.  Where MaxLen is used it is
+// provided primarily as a convenience for functions that might want to
+// anticipate the maximum path length that would be encountered traversing
+// the tree.
+//
+// Various graph search methods use a FromList to returns search results.
+// For a start node of a search, From will be -1 and Len will be 1. For other
+// nodes reached by the search, From represents a half arc in a path back to
+// start and Len represents the number of nodes in the path.  For nodes not
+// reached by the search, From will be -1 and Len will be 0.
+//
+// A single FromList can also represent a forest.  In this case paths from
+// all leaves do not return to a single root node, but multiple root nodes.
+//
+// While a FromList generally encodes a tree or forest, it is technically
+// possible to encode a cyclic graph.  A number of FromList methods require
+// the receiver to be acyclic.  Graph methods documented to return a tree or
+// forest will never return a cyclic FromList.  In other cases however,
+// where a FromList is not known to by cyclic, the Cyclic method can be
+// useful to validate the acyclic property.
+type FromList struct {
+	Paths  []PathEnd // tree representation
+	Leaves Bits      // leaves of tree
+	MaxLen int       // length of longest path, max of all PathEnd.Len values
+}
+
+// PathEnd associates a half arc and a path length.
+//
+// A PathEnd list is an element type of FromList.
+type PathEnd struct {
+	From NI  // a "from" half arc, the node the arc comes from
+	Len  int // number of nodes in path from start
+}
+
+// NewFromList creates a FromList object of given order.
+//
+// The Paths member is allocated to length n but there is no other
+// initialization.
+func NewFromList(n int) FromList {
+	return FromList{Paths: make([]PathEnd, n)}
+}
+
+// BoundsOk validates the "from" values in the list.
+//
+// Negative values are allowed as they indicate root nodes.
+//
+// BoundsOk returns true when all from values are less than len(t).
+// Otherwise it returns false and a node with a from value >= len(t).
+func (f FromList) BoundsOk() (ok bool, n NI) {
+	for n, e := range f.Paths {
+		if int(e.From) >= len(f.Paths) {
+			return false, NI(n)
+		}
+	}
+	return true, -1
+}
+
+// CommonStart returns the common start node of minimal paths to a and b.
+//
+// It returns -1 if a and b cannot be traced back to a common node.
+//
+// The method relies on populated PathEnd.Len members.  Use RecalcLen if
+// the Len members are not known to be present and correct.
+func (f FromList) CommonStart(a, b NI) NI {
+	p := f.Paths
+	if p[a].Len < p[b].Len {
+		a, b = b, a
+	}
+	for bl := p[b].Len; p[a].Len > bl; {
+		a = p[a].From
+		if a < 0 {
+			return -1
+		}
+	}
+	for a != b {
+		a = p[a].From
+		if a < 0 {
+			return -1
+		}
+		b = p[b].From
+	}
+	return a
+}
+
+// Cyclic determines if f contains a cycle, a non-empty path from a node
+// back to itself.
+//
+// Cyclic returns true if g contains at least one cycle.  It also returns
+// an example of a node involved in a cycle.
+//
+// Cyclic returns (false, -1) in the normal case where f is acyclic.
+// Note that the bool is not an "ok" return.  A cyclic FromList is usually
+// not okay.
+func (f FromList) Cyclic() (cyclic bool, n NI) {
+	var vis Bits
+	p := f.Paths
+	for i := range p {
+		var path Bits
+		for n := NI(i); vis.Bit(n) == 0; {
+			vis.SetBit(n, 1)
+			path.SetBit(n, 1)
+			if n = p[n].From; n < 0 {
+				break
+			}
+			if path.Bit(n) == 1 {
+				return true, n
+			}
+		}
+	}
+	return false, -1
+}
+
+// IsolatedNodeBits returns a bitmap of isolated nodes in receiver graph f.
+//
+// An isolated node is one with no arcs going to or from it.
+func (f FromList) IsolatedNodes() (iso Bits) {
+	p := f.Paths
+	iso.SetAll(len(p))
+	for n, e := range p {
+		if e.From >= 0 {
+			iso.SetBit(NI(n), 0)
+			iso.SetBit(e.From, 0)
+		}
+	}
+	return
+}
+
+// PathTo decodes a FromList, recovering a single path.
+//
+// The path is returned as a list of nodes where the first element will be
+// a root node and the last element will be the specified end node.
+//
+// Only the Paths member of the receiver is used.  Other members of the
+// FromList do not need to be valid, however the MaxLen member can be useful
+// for allocating argument p.
+//
+// Argument p can provide the result slice.  If p has capacity for the result
+// it will be used, otherwise a new slice is created for the result.
+//
+// See also function PathTo.
+func (f FromList) PathTo(end NI, p []NI) []NI {
+	return PathTo(f.Paths, end, p)
+}
+
+// PathTo decodes a single path from a PathEnd list.
+//
+// A PathEnd list is the main data representation in a FromList.  See FromList.
+//
+// PathTo returns a list of nodes where the first element will be
+// a root node and the last element will be the specified end node.
+//
+// Argument p can provide the result slice.  If p has capacity for the result
+// it will be used, otherwise a new slice is created for the result.
+//
+// See also method FromList.PathTo.
+func PathTo(paths []PathEnd, end NI, p []NI) []NI {
+	n := paths[end].Len
+	if n == 0 {
+		return nil
+	}
+	if cap(p) >= n {
+		p = p[:n]
+	} else {
+		p = make([]NI, n)
+	}
+	for {
+		n--
+		p[n] = end
+		if n == 0 {
+			return p
+		}
+		end = paths[end].From
+	}
+}
+
+// Preorder traverses f calling Visitor v in preorder.
+//
+// Nodes are visited in order such that for any node n with from node fr,
+// fr is visited before n.  Where f represents a tree, the visit ordering
+// corresponds to a preordering, or depth first traversal of the tree.
+// Where f represents a forest, the preorderings of the trees can be
+// intermingled.
+//
+// Leaves must be set correctly first.  Use RecalcLeaves if leaves are not
+// known to be set correctly.  FromList f cannot be cyclic.
+//
+// Traversal continues while v returns true.  It terminates if v returns false.
+// Preorder returns true if it completes without v returning false.  Preorder
+// returns false if traversal is terminated by v returning false.
+func (f FromList) Preorder(v OkNodeVisitor) bool {
+	p := f.Paths
+	var done Bits
+	var df func(NI) bool
+	df = func(n NI) bool {
+		done.SetBit(n, 1)
+		if fr := p[n].From; fr >= 0 && done.Bit(fr) == 0 {
+			df(fr)
+		}
+		return v(n)
+	}
+	for n := range f.Paths {
+		p[n].Len = 0
+	}
+	return f.Leaves.Iterate(func(n NI) bool {
+		return df(n)
+	})
+}
+
+// RecalcLeaves recomputes the Leaves member of f.
+func (f *FromList) RecalcLeaves() {
+	p := f.Paths
+	lv := &f.Leaves
+	lv.SetAll(len(p))
+	for n := range f.Paths {
+		if fr := p[n].From; fr >= 0 {
+			lv.SetBit(fr, 0)
+		}
+	}
+}
+
+// RecalcLen recomputes Len for each path end, and recomputes MaxLen.
+//
+// RecalcLen relies on the Leaves member being valid.  If it is not known
+// to be valid, call RecalcLeaves before calling RecalcLen.
+func (f *FromList) RecalcLen() {
+	p := f.Paths
+	var setLen func(NI) int
+	setLen = func(n NI) int {
+		switch {
+		case p[n].Len > 0:
+			return p[n].Len
+		case p[n].From < 0:
+			p[n].Len = 1
+			return 1
+		}
+		l := 1 + setLen(p[n].From)
+		p[n].Len = l
+		return l
+	}
+	for n := range f.Paths {
+		p[n].Len = 0
+	}
+	f.MaxLen = 0
+	f.Leaves.Iterate(func(n NI) bool {
+		if l := setLen(NI(n)); l > f.MaxLen {
+			f.MaxLen = l
+		}
+		return true
+	})
+}
+
+// ReRoot reorients the tree containing n to make n the root node.
+//
+// It keeps the tree connected by "reversing" the path from n to the old root.
+//
+// After ReRoot, the Leaves and Len members are invalid.
+// Call RecalcLeaves or RecalcLen as needed.
+func (f *FromList) ReRoot(n NI) {
+	p := f.Paths
+	fr := p[n].From
+	if fr < 0 {
+		return
+	}
+	p[n].From = -1
+	for {
+		ff := p[fr].From
+		p[fr].From = n
+		if ff < 0 {
+			return
+		}
+		n = fr
+		fr = ff
+	}
+}
+
+// Root finds the root of a node in a FromList.
+func (f FromList) Root(n NI) NI {
+	for p := f.Paths; ; {
+		fr := p[n].From
+		if fr < 0 {
+			return n
+		}
+		n = fr
+	}
+}
+
+// Transpose constructs the directed graph corresponding to FromList f
+// but with arcs in the opposite direction.  That is, from roots toward leaves.
+//
+// The method relies only on the From member of f.Paths.  Other members of
+// the FromList are not used.
+//
+// See FromList.TransposeRoots for a version that also accumulates and returns
+// information about the roots.
+func (f FromList) Transpose() Directed {
+	g := make(AdjacencyList, len(f.Paths))
+	for n, p := range f.Paths {
+		if p.From == -1 {
+			continue
+		}
+		g[p.From] = append(g[p.From], NI(n))
+	}
+	return Directed{g}
+}
+
+// TransposeLabeled constructs the directed labeled graph corresponding
+// to FromList f but with arcs in the opposite direction.  That is, from
+// roots toward leaves.
+//
+// The argument labels can be nil.  In this case labels are generated matching
+// the path indexes.  This corresponds to the "to", or child node.
+//
+// If labels is non-nil, it must be the same length as f.Paths and is used
+// to look up label numbers by the path index.
+//
+// The method relies only on the From member of f.Paths.  Other members of
+// the FromList are not used.
+//
+// See FromList.TransposeLabeledRoots for a version that also accumulates
+// and returns information about the roots.
+func (f FromList) TransposeLabeled(labels []LI) LabeledDirected {
+	g := make(LabeledAdjacencyList, len(f.Paths))
+	for n, p := range f.Paths {
+		if p.From == -1 {
+			continue
+		}
+		l := LI(n)
+		if labels != nil {
+			l = labels[n]
+		}
+		g[p.From] = append(g[p.From], Half{NI(n), l})
+	}
+	return LabeledDirected{g}
+}
+
+// TransposeLabeledRoots constructs the labeled directed graph corresponding
+// to FromList f but with arcs in the opposite direction.  That is, from
+// roots toward leaves.
+//
+// TransposeLabeledRoots also returns a count of roots of the resulting forest
+// and a bitmap of the roots.
+//
+// The argument labels can be nil.  In this case labels are generated matching
+// the path indexes.  This corresponds to the "to", or child node.
+//
+// If labels is non-nil, it must be the same length as t.Paths and is used
+// to look up label numbers by the path index.
+//
+// The method relies only on the From member of f.Paths.  Other members of
+// the FromList are not used.
+//
+// See FromList.TransposeLabeled for a simpler verstion that returns the
+// forest only.
+func (f FromList) TransposeLabeledRoots(labels []LI) (forest LabeledDirected, nRoots int, roots Bits) {
+	p := f.Paths
+	nRoots = len(p)
+	roots.SetAll(len(p))
+	g := make(LabeledAdjacencyList, len(p))
+	for i, p := range f.Paths {
+		if p.From == -1 {
+			continue
+		}
+		l := LI(i)
+		if labels != nil {
+			l = labels[i]
+		}
+		n := NI(i)
+		g[p.From] = append(g[p.From], Half{n, l})
+		if roots.Bit(n) == 1 {
+			roots.SetBit(n, 0)
+			nRoots--
+		}
+	}
+	return LabeledDirected{g}, nRoots, roots
+}
+
+// TransposeRoots constructs the directed graph corresponding to FromList f
+// but with arcs in the opposite direction.  That is, from roots toward leaves.
+//
+// TransposeRoots also returns a count of roots of the resulting forest and
+// a bitmap of the roots.
+//
+// The method relies only on the From member of f.Paths.  Other members of
+// the FromList are not used.
+//
+// See FromList.Transpose for a simpler verstion that returns the forest only.
+func (f FromList) TransposeRoots() (forest Directed, nRoots int, roots Bits) {
+	p := f.Paths
+	nRoots = len(p)
+	roots.SetAll(len(p))
+	g := make(AdjacencyList, len(p))
+	for i, e := range p {
+		if e.From == -1 {
+			continue
+		}
+		n := NI(i)
+		g[e.From] = append(g[e.From], n)
+		if roots.Bit(n) == 1 {
+			roots.SetBit(n, 0)
+			nRoots--
+		}
+	}
+	return Directed{g}, nRoots, roots
+}
diff --git a/vendor/github.com/soniakeys/graph/graph.go b/vendor/github.com/soniakeys/graph/graph.go
new file mode 100644
index 0000000..a2044e9
--- /dev/null
+++ b/vendor/github.com/soniakeys/graph/graph.go
@@ -0,0 +1,181 @@
+// Copyright 2014 Sonia Keys
+// License MIT: http://opensource.org/licenses/MIT
+
+package graph
+
+// graph.go contains type definitions for all graph types and components.
+// Also, go generate directives for source transformations.
+//
+// For readability, the types are defined in a dependency order:
+//
+//  NI
+//  NodeList
+//  AdjacencyList
+//  Directed
+//  Undirected
+//  LI
+//  Half
+//  LabeledAdjacencyList
+//  LabeledDirected
+//  LabeledUndirected
+//  Edge
+//  LabeledEdge
+//  WeightFunc
+//  WeightedEdgeList
+
+//go:generate cp adj_cg.go adj_RO.go
+//go:generate gofmt -r "LabeledAdjacencyList -> AdjacencyList" -w adj_RO.go
+//go:generate gofmt -r "n.To -> n" -w adj_RO.go
+//go:generate gofmt -r "Half -> NI" -w adj_RO.go
+
+//go:generate cp dir_cg.go dir_RO.go
+//go:generate gofmt -r "LabeledDirected -> Directed" -w dir_RO.go
+//go:generate gofmt -r "LabeledAdjacencyList -> AdjacencyList" -w dir_RO.go
+//go:generate gofmt -r "n.To -> n" -w dir_RO.go
+//go:generate gofmt -r "Half -> NI" -w dir_RO.go
+
+//go:generate cp undir_cg.go undir_RO.go
+//go:generate gofmt -r "LabeledUndirected -> Undirected" -w undir_RO.go
+//go:generate gofmt -r "LabeledAdjacencyList -> AdjacencyList" -w undir_RO.go
+//go:generate gofmt -r "n.To -> n" -w undir_RO.go
+//go:generate gofmt -r "Half -> NI" -w undir_RO.go
+
+// NI is a "node int"
+//
+// It is a node number or node ID.  NIs are used extensively as slice indexes.
+// NIs typically account for a significant fraction of the memory footprint of
+// a graph.
+type NI int32
+
+// NodeList satisfies sort.Interface.
+type NodeList []NI
+
+func (l NodeList) Len() int           { return len(l) }
+func (l NodeList) Less(i, j int) bool { return l[i] < l[j] }
+func (l NodeList) Swap(i, j int)      { l[i], l[j] = l[j], l[i] }
+
+// An AdjacencyList represents a graph as a list of neighbors for each node.
+// The "node ID" of a node is simply it's slice index in the AdjacencyList.
+// For an AdjacencyList g, g[n] represents arcs going from node n to nodes
+// g[n].
+//
+// Adjacency lists are inherently directed but can be used to represent
+// directed or undirected graphs.  See types Directed and Undirected.
+type AdjacencyList [][]NI
+
+// Directed represents a directed graph.
+//
+// Directed methods generally rely on the graph being directed, specifically
+// that arcs do not have reciprocals.
+type Directed struct {
+	AdjacencyList // embedded to include AdjacencyList methods
+}
+
+// Undirected represents an undirected graph.
+//
+// In an undirected graph, for each arc between distinct nodes there is also
+// a reciprocal arc, an arc in the opposite direction.  Loops do not have
+// reciprocals.
+//
+// Undirected methods generally rely on the graph being undirected,
+// specifically that every arc between distinct nodes has a reciprocal.
+type Undirected struct {
+	AdjacencyList // embedded to include AdjacencyList methods
+}
+
+// LI is a label integer, used for associating labels with arcs.
+type LI int32
+
+// Half is a half arc, representing a labeled arc and the "neighbor" node
+// that the arc leads to.
+//
+// Halfs can be composed to form a labeled adjacency list.
+type Half struct {
+	To    NI // node ID, usable as a slice index
+	Label LI // half-arc ID for application data, often a weight
+}
+
+// A LabeledAdjacencyList represents a graph as a list of neighbors for each
+// node, connected by labeled arcs.
+//
+// Arc labels are not necessarily unique arc IDs.  Different arcs can have
+// the same label.
+//
+// Arc labels are commonly used to assocate a weight with an arc.  Arc labels
+// are general purpose however and can be used to associate arbitrary
+// information with an arc.
+//
+// Methods implementing weighted graph algorithms will commonly take a
+// weight function that turns a label int into a float64 weight.
+//
+// If only a small amount of information -- such as an integer weight or
+// a single printable character -- needs to be associated, it can sometimes
+// be possible to encode the information directly into the label int.  For
+// more generality, some lookup scheme will be needed.
+//
+// In an undirected labeled graph, reciprocal arcs must have identical labels.
+// Note this does not preclude parallel arcs with different labels.
+type LabeledAdjacencyList [][]Half
+
+// LabeledDirected represents a directed labeled graph.
+//
+// This is the labeled version of Directed.  See types LabeledAdjacencyList
+// and Directed.
+type LabeledDirected struct {
+	LabeledAdjacencyList // embedded to include LabeledAdjacencyList methods
+}
+
+// LabeledUndirected represents an undirected labeled graph.
+//
+// This is the labeled version of Undirected.  See types LabeledAdjacencyList
+// and Undirected.
+type LabeledUndirected struct {
+	LabeledAdjacencyList // embedded to include LabeledAdjacencyList methods
+}
+
+// Edge is an undirected edge between nodes N1 and N2.
+type Edge struct{ N1, N2 NI }
+
+// LabeledEdge is an undirected edge with an associated label.
+type LabeledEdge struct {
+	Edge
+	LI
+}
+
+// WeightFunc returns a weight for a given label.
+//
+// WeightFunc is a parameter type for various search functions.  The intent
+// is to return a weight corresponding to an arc label.  The name "weight"
+// is an abstract term.  An arc "weight" will typically have some application
+// specific meaning other than physical weight.
+type WeightFunc func(label LI) (weight float64)
+
+// WeightedEdgeList is a graph representation.
+//
+// It is a labeled edge list, with an associated weight function to return
+// a weight given an edge label.
+//
+// Also associated is the order, or number of nodes of the graph.
+// All nodes occurring in the edge list must be strictly less than Order.
+//
+// WeigtedEdgeList sorts by weight, obtained by calling the weight function.
+// If weight computation is expensive, consider supplying a cached or
+// memoized version.
+type WeightedEdgeList struct {
+	Order int
+	WeightFunc
+	Edges []LabeledEdge
+}
+
+// Len implements sort.Interface.
+func (l WeightedEdgeList) Len() int { return len(l.Edges) }
+
+// Less implements sort.Interface.
+func (l WeightedEdgeList) Less(i, j int) bool {
+	return l.WeightFunc(l.Edges[i].LI) < l.WeightFunc(l.Edges[j].LI)
+}
+
+// Swap implements sort.Interface.
+func (l WeightedEdgeList) Swap(i, j int) {
+	l.Edges[i], l.Edges[j] = l.Edges[j], l.Edges[i]
+}
diff --git a/vendor/github.com/soniakeys/graph/hacking.md b/vendor/github.com/soniakeys/graph/hacking.md
new file mode 100644
index 0000000..30d2d7c
--- /dev/null
+++ b/vendor/github.com/soniakeys/graph/hacking.md
@@ -0,0 +1,37 @@
+#Hacking
+
+Basic use of the package is just go get, or git clone; go install.  There are
+no dependencies outside the standard library.
+
+The primary to-do list is the issue tracker on Github.  I maintained a
+journal on google drive for a while but at some point filed issues for all
+remaining ideas in that document that still seemed relevant.  So currently
+there is no other roadmap or planning document.
+
+CI is currently on travis-ci.org.  The .travis.yml builds for go 1.2.1
+following https://github.com/soniakeys/graph/issues/49, and it currently builds
+for go 1.6 as well.  The travis script calls a shell script right away because
+I didn’t see a way to get it to do different steps for the different go
+versions.  For 1.2.1, I just wanted the basic tests.  For a current go version
+such as 1.6, there’s a growing list of checks.
+
+The GOARCH=386 test is for https://github.com/soniakeys/graph/issues/41.
+The problem is the architecture specific code in bits32.go and bits64.go.
+Yes, there are architecture independent algorithms.  There is also assembly
+to access machine instructions.  Anyway, it’s the way it is for now.
+
+Im not big on making go vet happy just for a badge but I really like the
+example check that I believe appeared with go 1.6.  (I think it will be a
+standard check with 1.7, so the test script will have to change then.)
+
+https://github.com/client9/misspell has been valuable.
+
+Also I wrote https://github.com/soniakeys/vetc to validate that each source
+file has copyright/license statement.
+
+Then, it’s not in the ci script, but I wrote https://github.com/soniakeys/rcv
+to put coverage stats in the readme.  Maybe it could be commit hook or
+something but for now I’ll try just running it manually now and then.
+
+Go fmt is not in the ci script, but I have at least one editor set up to run
+it on save, so code should stay formatted pretty well.
diff --git a/vendor/github.com/soniakeys/graph/mst.go b/vendor/github.com/soniakeys/graph/mst.go
new file mode 100644
index 0000000..028e680
--- /dev/null
+++ b/vendor/github.com/soniakeys/graph/mst.go
@@ -0,0 +1,244 @@
+// Copyright 2014 Sonia Keys
+// License MIT: http://opensource.org/licenses/MIT
+
+package graph
+
+import (
+	"container/heap"
+	"sort"
+)
+
+type dsElement struct {
+	from NI
+	rank int
+}
+
+type disjointSet struct {
+	set []dsElement
+}
+
+func newDisjointSet(n int) disjointSet {
+	set := make([]dsElement, n)
+	for i := range set {
+		set[i].from = -1
+	}
+	return disjointSet{set}
+}
+
+// return true if disjoint trees were combined.
+// false if x and y were already in the same tree.
+func (ds disjointSet) union(x, y NI) bool {
+	xr := ds.find(x)
+	yr := ds.find(y)
+	if xr == yr {
+		return false
+	}
+	switch xe, ye := &ds.set[xr], &ds.set[yr]; {
+	case xe.rank < ye.rank:
+		xe.from = yr
+	case xe.rank == ye.rank:
+		xe.rank++
+		fallthrough
+	default:
+		ye.from = xr
+	}
+	return true
+}
+
+func (ds disjointSet) find(n NI) NI {
+	// fast paths for n == root or from root.
+	// no updates need in these cases.
+	s := ds.set
+	fr := s[n].from
+	if fr < 0 { // n is root
+		return n
+	}
+	n, fr = fr, s[fr].from
+	if fr < 0 { // n is from root
+		return n
+	}
+	// otherwise updates needed.
+	// two iterative passes (rather than recursion or stack)
+	// pass 1: find root
+	r := fr
+	for {
+		f := s[r].from
+		if f < 0 {
+			break
+		}
+		r = f
+	}
+	// pass 2: update froms
+	for {
+		s[n].from = r
+		if fr == r {
+			return r
+		}
+		n = fr
+		fr = s[n].from
+	}
+}
+
+// Kruskal implements Kruskal's algorithm for constructing a minimum spanning
+// forest on an undirected graph.
+//
+// While the input graph is interpreted as undirected, the receiver edge list
+// does not actually need to contain reciprocal arcs.  A property of the
+// algorithm is that arc direction is ignored.  Thus only a single arc out of
+// a reciprocal pair must be present in the edge list.  Reciprocal arcs (and
+// parallel arcs) are allowed though, and do not affect the result.
+//
+// The forest is returned as an undirected graph.
+//
+// Also returned is a total distance for the returned forest.
+//
+// The edge list of the receiver is sorted as a side effect of this method.
+// See KruskalSorted for a version that relies on the edge list being already
+// sorted.
+func (l WeightedEdgeList) Kruskal() (g LabeledUndirected, dist float64) {
+	sort.Sort(l)
+	return l.KruskalSorted()
+}
+
+// KruskalSorted implements Kruskal's algorithm for constructing a minimum
+// spanning tree on an undirected graph.
+//
+// While the input graph is interpreted as undirected, the receiver edge list
+// does not actually need to contain reciprocal arcs.  A property of the
+// algorithm is that arc direction is ignored.  Thus only a single arc out of
+// a reciprocal pair must be present in the edge list.  Reciprocal arcs (and
+// parallel arcs) are allowed though, and do not affect the result.
+//
+// When called, the edge list of the receiver must be already sorted by weight.
+// See Kruskal for a version that accepts an unsorted edge list.
+//
+// The forest is returned as an undirected graph.
+//
+// Also returned is a total distance for the returned forest.
+func (l WeightedEdgeList) KruskalSorted() (g LabeledUndirected, dist float64) {
+	ds := newDisjointSet(l.Order)
+	g.LabeledAdjacencyList = make(LabeledAdjacencyList, l.Order)
+	for _, e := range l.Edges {
+		if ds.union(e.N1, e.N2) {
+			g.AddEdge(Edge{e.N1, e.N2}, e.LI)
+			dist += l.WeightFunc(e.LI)
+		}
+	}
+	return
+}
+
+// Prim implements the Jarník-Prim-Dijkstra algorithm for constructing
+// a minimum spanning tree on an undirected graph.
+//
+// Prim computes a minimal spanning tree on the connected component containing
+// the given start node.  The tree is returned in FromList f.  Argument f
+// cannot be a nil pointer although it can point to a zero value FromList.
+//
+// If the passed FromList.Paths has the len of g though, it will be reused.
+// In the case of a graph with multiple connected components, this allows a
+// spanning forest to be accumulated by calling Prim successively on
+// representative nodes of the components.  In this case if leaves for
+// individual trees are of interest, pass a non-nil zero-value for the argument
+// componentLeaves and it will be populated with leaves for the single tree
+// spanned by the call.
+//
+// If argument labels is non-nil, it must have the same length as g and will
+// be populated with labels corresponding to the edges of the tree.
+//
+// Returned are the number of nodes spanned for the single tree (which will be
+// the order of the connected component) and the total spanned distance for the
+// single tree.
+func (g LabeledUndirected) Prim(start NI, w WeightFunc, f *FromList, labels []LI, componentLeaves *Bits) (numSpanned int, dist float64) {
+	al := g.LabeledAdjacencyList
+	if len(f.Paths) != len(al) {
+		*f = NewFromList(len(al))
+	}
+	b := make([]prNode, len(al)) // "best"
+	for n := range b {
+		b[n].nx = NI(n)
+		b[n].fx = -1
+	}
+	rp := f.Paths
+	var frontier prHeap
+	rp[start] = PathEnd{From: -1, Len: 1}
+	numSpanned = 1
+	fLeaves := &f.Leaves
+	fLeaves.SetBit(start, 1)
+	if componentLeaves != nil {
+		componentLeaves.SetBit(start, 1)
+	}
+	for a := start; ; {
+		for _, nb := range al[a] {
+			if rp[nb.To].Len > 0 {
+				continue // already in MST, no action
+			}
+			switch bp := &b[nb.To]; {
+			case bp.fx == -1: // new node for frontier
+				bp.from = fromHalf{From: a, Label: nb.Label}
+				bp.wt = w(nb.Label)
+				heap.Push(&frontier, bp)
+			case w(nb.Label) < bp.wt: // better arc
+				bp.from = fromHalf{From: a, Label: nb.Label}
+				bp.wt = w(nb.Label)
+				heap.Fix(&frontier, bp.fx)
+			}
+		}
+		if len(frontier) == 0 {
+			break // done
+		}
+		bp := heap.Pop(&frontier).(*prNode)
+		a = bp.nx
+		rp[a].Len = rp[bp.from.From].Len + 1
+		rp[a].From = bp.from.From
+		if len(labels) != 0 {
+			labels[a] = bp.from.Label
+		}
+		dist += bp.wt
+		fLeaves.SetBit(bp.from.From, 0)
+		fLeaves.SetBit(a, 1)
+		if componentLeaves != nil {
+			componentLeaves.SetBit(bp.from.From, 0)
+			componentLeaves.SetBit(a, 1)
+		}
+		numSpanned++
+	}
+	return
+}
+
+// fromHalf is a half arc, representing a labeled arc and the "neighbor" node
+// that the arc originates from.
+//
+// (This used to be exported when there was a LabeledFromList.  Currently
+// unexported now that it seems to have much more limited use.)
+type fromHalf struct {
+	From  NI
+	Label LI
+}
+
+type prNode struct {
+	nx   NI
+	from fromHalf
+	wt   float64 // p.Weight(from.Label)
+	fx   int
+}
+
+type prHeap []*prNode
+
+func (h prHeap) Len() int           { return len(h) }
+func (h prHeap) Less(i, j int) bool { return h[i].wt < h[j].wt }
+func (h prHeap) Swap(i, j int) {
+	h[i], h[j] = h[j], h[i]
+	h[i].fx = i
+	h[j].fx = j
+}
+func (p *prHeap) Push(x interface{}) {
+	nd := x.(*prNode)
+	nd.fx = len(*p)
+	*p = append(*p, nd)
+}
+func (p *prHeap) Pop() interface{} {
+	r := *p
+	last := len(r) - 1
+	*p = r[:last]
+	return r[last]
+}
diff --git a/vendor/github.com/soniakeys/graph/random.go b/vendor/github.com/soniakeys/graph/random.go
new file mode 100644
index 0000000..99f0445
--- /dev/null
+++ b/vendor/github.com/soniakeys/graph/random.go
@@ -0,0 +1,325 @@
+// Copyright 2016 Sonia Keys
+// License MIT: https://opensource.org/licenses/MIT
+
+package graph
+
+import (
+	"errors"
+	"math"
+	"math/rand"
+	"time"
+)
+
+// Euclidean generates a random simple graph on the Euclidean plane.
+//
+// Nodes are associated with coordinates uniformly distributed on a unit
+// square.  Arcs are added between random nodes with a bias toward connecting
+// nearer nodes.
+//
+// Unfortunately the function has a few "knobs".
+// The returned graph will have order nNodes and arc size nArcs.  The affinity
+// argument controls the bias toward connecting nearer nodes.  The function
+// selects random pairs of nodes as a candidate arc then rejects the candidate
+// if the nodes fail an affinity test.  Also parallel arcs are rejected.
+// As more affine or denser graphs are requested, rejections increase,
+// increasing run time.  The patience argument controls the number of arc
+// rejections allowed before the function gives up and returns an error.
+// Note that higher affinity will require more patience and that some
+// combinations of nNodes and nArcs cannot be achieved with any amount of
+// patience given that the returned graph must be simple.
+//
+// If Rand r is nil, the method creates a new source and generator for
+// one-time use.
+//
+// Returned is a directed simple graph and associated positions indexed by
+// node number.
+//
+// See also LabeledEuclidean.
+func Euclidean(nNodes, nArcs int, affinity float64, patience int, r *rand.Rand) (g Directed, pos []struct{ X, Y float64 }, err error) {
+	a := make(AdjacencyList, nNodes) // graph
+	// generate random positions
+	if r == nil {
+		r = rand.New(rand.NewSource(time.Now().UnixNano()))
+	}
+	pos = make([]struct{ X, Y float64 }, nNodes)
+	for i := range pos {
+		pos[i].X = r.Float64()
+		pos[i].Y = r.Float64()
+	}
+	// arcs
+	var tooFar, dup int
+arc:
+	for i := 0; i < nArcs; {
+		if tooFar == nArcs*patience {
+			err = errors.New("affinity not found")
+			return
+		}
+		if dup == nArcs*patience {
+			err = errors.New("overcrowding")
+			return
+		}
+		n1 := NI(r.Intn(nNodes))
+		var n2 NI
+		for {
+			n2 = NI(r.Intn(nNodes))
+			if n2 != n1 { // no graph loops
+				break
+			}
+		}
+		c1 := &pos[n1]
+		c2 := &pos[n2]
+		dist := math.Hypot(c2.X-c1.X, c2.Y-c1.Y)
+		if dist*affinity > r.ExpFloat64() { // favor near nodes
+			tooFar++
+			continue
+		}
+		for _, nb := range a[n1] {
+			if nb == n2 { // no parallel arcs
+				dup++
+				continue arc
+			}
+		}
+		a[n1] = append(a[n1], n2)
+		i++
+	}
+	g = Directed{a}
+	return
+}
+
+// LabeledEuclidean generates a random simple graph on the Euclidean plane.
+//
+// Arc label values in the returned graph g are indexes into the return value
+// wt.  Wt is the Euclidean distance between the from and to nodes of the arc.
+//
+// Otherwise the function arguments and return values are the same as for
+// function Euclidean.  See Euclidean.
+func LabeledEuclidean(nNodes, nArcs int, affinity float64, patience int, r *rand.Rand) (g LabeledDirected, pos []struct{ X, Y float64 }, wt []float64, err error) {
+	a := make(LabeledAdjacencyList, nNodes) // graph
+	wt = make([]float64, nArcs)             // arc weights
+	// generate random positions
+	if r == nil {
+		r = rand.New(rand.NewSource(time.Now().UnixNano()))
+	}
+	pos = make([]struct{ X, Y float64 }, nNodes)
+	for i := range pos {
+		pos[i].X = r.Float64()
+		pos[i].Y = r.Float64()
+	}
+	// arcs
+	var tooFar, dup int
+arc:
+	for i := 0; i < nArcs; {
+		if tooFar == nArcs*patience {
+			err = errors.New("affinity not found")
+			return
+		}
+		if dup == nArcs*patience {
+			err = errors.New("overcrowding")
+			return
+		}
+		n1 := NI(r.Intn(nNodes))
+		var n2 NI
+		for {
+			n2 = NI(r.Intn(nNodes))
+			if n2 != n1 { // no graph loops
+				break
+			}
+		}
+		c1 := &pos[n1]
+		c2 := &pos[n2]
+		dist := math.Hypot(c2.X-c1.X, c2.Y-c1.Y)
+		if dist*affinity > r.ExpFloat64() { // favor near nodes
+			tooFar++
+			continue
+		}
+		for _, nb := range a[n1] {
+			if nb.To == n2 { // no parallel arcs
+				dup++
+				continue arc
+			}
+		}
+		wt[i] = dist
+		a[n1] = append(a[n1], Half{n2, LI(i)})
+		i++
+	}
+	g = LabeledDirected{a}
+	return
+}
+
+// Geometric generates a random geometric graph (RGG) on the Euclidean plane.
+//
+// An RGG is an undirected simple graph.  Nodes are associated with coordinates
+// uniformly distributed on a unit square.  Edges are added between all nodes
+// falling within a specified distance or radius of each other.
+//
+// The resulting number of edges is somewhat random but asymptotically
+// approaches m = πr²n²/2.   The method accumulates and returns the actual
+// number of edges constructed.
+//
+// If Rand r is nil, the method creates a new source and generator for
+// one-time use.
+//
+// See also LabeledGeometric.
+func Geometric(nNodes int, radius float64, r *rand.Rand) (g Undirected, pos []struct{ X, Y float64 }, m int) {
+	// Expected degree is approximately nπr².
+	a := make(AdjacencyList, nNodes)
+	if r == nil {
+		r = rand.New(rand.NewSource(time.Now().UnixNano()))
+	}
+	pos = make([]struct{ X, Y float64 }, nNodes)
+	for i := range pos {
+		pos[i].X = r.Float64()
+		pos[i].Y = r.Float64()
+	}
+	for u, up := range pos {
+		for v := u + 1; v < len(pos); v++ {
+			vp := pos[v]
+			if math.Hypot(up.X-vp.X, up.Y-vp.Y) < radius {
+				a[u] = append(a[u], NI(v))
+				a[v] = append(a[v], NI(u))
+				m++
+			}
+		}
+	}
+	g = Undirected{a}
+	return
+}
+
+// LabeledGeometric generates a random geometric graph (RGG) on the Euclidean
+// plane.
+//
+// Edge label values in the returned graph g are indexes into the return value
+// wt.  Wt is the Euclidean distance between nodes of the edge.  The graph
+// size m is len(wt).
+//
+// See Geometric for additional description.
+func LabeledGeometric(nNodes int, radius float64, r *rand.Rand) (g LabeledUndirected, pos []struct{ X, Y float64 }, wt []float64) {
+	a := make(LabeledAdjacencyList, nNodes)
+	if r == nil {
+		r = rand.New(rand.NewSource(time.Now().UnixNano()))
+	}
+	pos = make([]struct{ X, Y float64 }, nNodes)
+	for i := range pos {
+		pos[i].X = r.Float64()
+		pos[i].Y = r.Float64()
+	}
+	for u, up := range pos {
+		for v := u + 1; v < len(pos); v++ {
+			vp := pos[v]
+			if w := math.Hypot(up.X-vp.X, up.Y-vp.Y); w < radius {
+				a[u] = append(a[u], Half{NI(v), LI(len(wt))})
+				a[v] = append(a[v], Half{NI(u), LI(len(wt))})
+				wt = append(wt, w)
+			}
+		}
+	}
+	g = LabeledUndirected{a}
+	return
+}
+
+// KroneckerDirected generates a Kronecker-like random directed graph.
+//
+// The returned graph g is simple and has no isolated nodes but is not
+// necessarily fully connected.  The number of of nodes will be <= 2^scale,
+// and will be near 2^scale for typical values of arcFactor, >= 2.
+// ArcFactor * 2^scale arcs are generated, although loops and duplicate arcs
+// are rejected.
+//
+// If Rand r is nil, the method creates a new source and generator for
+// one-time use.
+//
+// Return value ma is the number of arcs retained in the result graph.
+func KroneckerDirected(scale uint, arcFactor float64, r *rand.Rand) (g Directed, ma int) {
+	a, m := kronecker(scale, arcFactor, true, r)
+	return Directed{a}, m
+}
+
+// KroneckerUndirected generates a Kronecker-like random undirected graph.
+//
+// The returned graph g is simple and has no isolated nodes but is not
+// necessarily fully connected.  The number of of nodes will be <= 2^scale,
+// and will be near 2^scale for typical values of edgeFactor, >= 2.
+// EdgeFactor * 2^scale edges are generated, although loops and duplicate edges
+// are rejected.
+//
+// If Rand r is nil, the method creates a new source and generator for
+// one-time use.
+//
+// Return value m is the true number of edges--not arcs--retained in the result
+// graph.
+func KroneckerUndirected(scale uint, edgeFactor float64, r *rand.Rand) (g Undirected, m int) {
+	al, s := kronecker(scale, edgeFactor, false, r)
+	return Undirected{al}, s
+}
+
+// Styled after the Graph500 example code.  Not well tested currently.
+// Graph500 example generates undirected only.  No idea if the directed variant
+// here is meaningful or not.
+//
+// note mma returns arc size ma for dir=true, but returns size m for dir=false
+func kronecker(scale uint, edgeFactor float64, dir bool, r *rand.Rand) (g AdjacencyList, mma int) {
+	if r == nil {
+		r = rand.New(rand.NewSource(time.Now().UnixNano()))
+	}
+	N := NI(1 << scale)                  // node extent
+	M := int(edgeFactor*float64(N) + .5) // number of arcs/edges to generate
+	a, b, c := 0.57, 0.19, 0.19          // initiator probabilities
+	ab := a + b
+	cNorm := c / (1 - ab)
+	aNorm := a / ab
+	ij := make([][2]NI, M)
+	var bm Bits
+	var nNodes int
+	for k := range ij {
+		var i, j NI
+		for b := NI(1); b < N; b <<= 1 {
+			if r.Float64() > ab {
+				i |= b
+				if r.Float64() > cNorm {
+					j |= b
+				}
+			} else if r.Float64() > aNorm {
+				j |= b
+			}
+		}
+		if bm.Bit(i) == 0 {
+			bm.SetBit(i, 1)
+			nNodes++
+		}
+		if bm.Bit(j) == 0 {
+			bm.SetBit(j, 1)
+			nNodes++
+		}
+		r := r.Intn(k + 1) // shuffle edges as they are generated
+		ij[k] = ij[r]
+		ij[r] = [2]NI{i, j}
+	}
+	p := r.Perm(nNodes) // mapping to shuffle IDs of non-isolated nodes
+	px := 0
+	rn := make([]NI, N)
+	for i := range rn {
+		if bm.Bit(NI(i)) == 1 {
+			rn[i] = NI(p[px]) // fill lookup table
+			px++
+		}
+	}
+	g = make(AdjacencyList, nNodes)
+ij:
+	for _, e := range ij {
+		if e[0] == e[1] {
+			continue // skip loops
+		}
+		ri, rj := rn[e[0]], rn[e[1]]
+		for _, nb := range g[ri] {
+			if nb == rj {
+				continue ij // skip parallel edges
+			}
+		}
+		g[ri] = append(g[ri], rj)
+		mma++
+		if !dir {
+			g[rj] = append(g[rj], ri)
+		}
+	}
+	return
+}
diff --git a/vendor/github.com/soniakeys/graph/readme.md b/vendor/github.com/soniakeys/graph/readme.md
new file mode 100644
index 0000000..539670f
--- /dev/null
+++ b/vendor/github.com/soniakeys/graph/readme.md
@@ -0,0 +1,38 @@
+#Graph
+
+A graph library with goals of speed and simplicity, Graph implements
+graph algorithms on graphs of zero-based integer node IDs.
+
+[![GoDoc](https://godoc.org/github.com/soniakeys/graph?status.svg)](https://godoc.org/github.com/soniakeys/graph) [![Go Walker](http://gowalker.org/api/v1/badge)](https://gowalker.org/github.com/soniakeys/graph) [![GoSearch](http://go-search.org/badge?id=github.com%2Fsoniakeys%2Fgraph)](http://go-search.org/view?id=github.com%2Fsoniakeys%2Fgraph)[![Build Status](https://travis-ci.org/soniakeys/graph.svg?branch=master)](https://travis-ci.org/soniakeys/graph)
+
+Status, 4 Apr 2016:  The repo has benefitted recently from being included
+in another package.  In response to users of that package, this repo now
+builds for 32 bit Windows and ARM, and for Go versions back to 1.2.1.
+Thank you all who have filed issues.
+
+###Non-source files of interest
+
+The directory [tutorials](tutorials) is a work in progress - there are only
+a couple of tutorials there yet - but the concept is to provide some topical
+walk-throughs to supplement godoc.  The source-based godoc documentation
+remains the primary documentation.
+
+* [Dijkstra's algorithm](tutorials/dijkstra.md)
+* [AdjacencyList types](tutorials/adjacencylist.md)
+
+The directory [bench](bench) is another work in progress.  The concept is
+to present some plots showing benchmark performance approaching some
+theoretical asymptote.
+
+[hacking.md](hacking.md) has some information about how the library is
+developed, built, and tested.  It might be of interest if for example you
+plan to fork or contribute to the the repository.
+
+###Test coverage
+8 Apr 2016
+```
+graph          95.3%
+graph/df       20.7%
+graph/dot      77.5%
+graph/treevis  79.4%
+```
diff --git a/vendor/github.com/soniakeys/graph/sssp.go b/vendor/github.com/soniakeys/graph/sssp.go
new file mode 100644
index 0000000..32cc192
--- /dev/null
+++ b/vendor/github.com/soniakeys/graph/sssp.go
@@ -0,0 +1,881 @@
+// Copyright 2013 Sonia Keys
+// License MIT: http://opensource.org/licenses/MIT
+
+package graph
+
+import (
+	"container/heap"
+	"fmt"
+	"math"
+)
+
+// rNode holds data for a "reached" node
+type rNode struct {
+	nx    NI
+	state int8    // state constants defined below
+	f     float64 // "g+h", path dist + heuristic estimate
+	fx    int     // heap.Fix index
+}
+
+// for rNode.state
+const (
+	unreached = 0
+	reached   = 1
+	open      = 1
+	closed    = 2
+)
+
+type openHeap []*rNode
+
+// A Heuristic is defined on a specific end node.  The function
+// returns an estimate of the path distance from node argument
+// "from" to the end node.  Two subclasses of heuristics are "admissible"
+// and "monotonic."
+//
+// Admissible means the value returned is guaranteed to be less than or
+// equal to the actual shortest path distance from the node to end.
+//
+// An admissible estimate may further be monotonic.
+// Monotonic means that for any neighboring nodes A and B with half arc aB
+// leading from A to B, and for heuristic h defined on some end node, then
+// h(A) <= aB.ArcWeight + h(B).
+//
+// See AStarA for additional notes on implementing heuristic functions for
+// AStar search methods.
+type Heuristic func(from NI) float64
+
+// Admissible returns true if heuristic h is admissible on graph g relative to
+// the given end node.
+//
+// If h is inadmissible, the string result describes a counter example.
+func (h Heuristic) Admissible(g LabeledAdjacencyList, w WeightFunc, end NI) (bool, string) {
+	// invert graph
+	inv := make(LabeledAdjacencyList, len(g))
+	for from, nbs := range g {
+		for _, nb := range nbs {
+			inv[nb.To] = append(inv[nb.To],
+				Half{To: NI(from), Label: nb.Label})
+		}
+	}
+	// run dijkstra
+	// Dijkstra.AllPaths takes a start node but after inverting the graph
+	// argument end now represents the start node of the inverted graph.
+	f, dist, _ := inv.Dijkstra(end, -1, w)
+	// compare h to found shortest paths
+	for n := range inv {
+		if f.Paths[n].Len == 0 {
+			continue // no path, any heuristic estimate is fine.
+		}
+		if !(h(NI(n)) <= dist[n]) {
+			return false, fmt.Sprintf("h(%d) = %g, "+
+				"required to be <= found shortest path (%g)",
+				n, h(NI(n)), dist[n])
+		}
+	}
+	return true, ""
+}
+
+// Monotonic returns true if heuristic h is monotonic on weighted graph g.
+//
+// If h is non-monotonic, the string result describes a counter example.
+func (h Heuristic) Monotonic(g LabeledAdjacencyList, w WeightFunc) (bool, string) {
+	// precompute
+	hv := make([]float64, len(g))
+	for n := range g {
+		hv[n] = h(NI(n))
+	}
+	// iterate over all edges
+	for from, nbs := range g {
+		for _, nb := range nbs {
+			arcWeight := w(nb.Label)
+			if !(hv[from] <= arcWeight+hv[nb.To]) {
+				return false, fmt.Sprintf("h(%d) = %g, "+
+					"required to be <= arc weight + h(%d) (= %g + %g = %g)",
+					from, hv[from],
+					nb.To, arcWeight, hv[nb.To], arcWeight+hv[nb.To])
+			}
+		}
+	}
+	return true, ""
+}
+
+// AStarA finds a path between two nodes.
+//
+// AStarA implements both algorithm A and algorithm A*.  The difference in the
+// two algorithms is strictly in the heuristic estimate returned by argument h.
+// If h is an "admissible" heuristic estimate, then the algorithm is termed A*,
+// otherwise it is algorithm A.
+//
+// Like Dijkstra's algorithm, AStarA with an admissible heuristic finds the
+// shortest path between start and end.  AStarA generally runs faster than
+// Dijkstra though, by using the heuristic distance estimate.
+//
+// AStarA with an inadmissible heuristic becomes algorithm A.  Algorithm A
+// will find a path, but it is not guaranteed to be the shortest path.
+// The heuristic still guides the search however, so a nearly admissible
+// heuristic is likely to find a very good path, if not the best.  Quality
+// of the path returned degrades gracefully with the quality of the heuristic.
+//
+// The heuristic function h should ideally be fairly inexpensive.  AStarA
+// may call it more than once for the same node, especially as graph density
+// increases.  In some cases it may be worth the effort to memoize or
+// precompute values.
+//
+// Argument g is the graph to be searched, with arc weights returned by w.
+// As usual for AStar, arc weights must be non-negative.
+// Graphs may be directed or undirected.
+//
+// If AStarA finds a path it returns a FromList encoding the path, the arc
+// labels for path nodes, the total path distance, and ok = true.
+// Otherwise it returns ok = false.
+func (g LabeledAdjacencyList) AStarA(w WeightFunc, start, end NI, h Heuristic) (f FromList, labels []LI, dist float64, ok bool) {
+	// NOTE: AStarM is largely duplicate code.
+
+	f = NewFromList(len(g))
+	labels = make([]LI, len(g))
+	d := make([]float64, len(g))
+	r := make([]rNode, len(g))
+	for i := range r {
+		r[i].nx = NI(i)
+	}
+	// start node is reached initially
+	cr := &r[start]
+	cr.state = reached
+	cr.f = h(start) // total path estimate is estimate from start
+	rp := f.Paths
+	rp[start] = PathEnd{Len: 1, From: -1} // path length at start is 1 node
+	// oh is a heap of nodes "open" for exploration.  nodes go on the heap
+	// when they get an initial or new "g" path distance, and therefore a
+	// new "f" which serves as priority for exploration.
+	oh := openHeap{cr}
+	for len(oh) > 0 {
+		bestPath := heap.Pop(&oh).(*rNode)
+		bestNode := bestPath.nx
+		if bestNode == end {
+			return f, labels, d[end], true
+		}
+		bp := &rp[bestNode]
+		nextLen := bp.Len + 1
+		for _, nb := range g[bestNode] {
+			alt := &r[nb.To]
+			ap := &rp[alt.nx]
+			// "g" path distance from start
+			g := d[bestNode] + w(nb.Label)
+			if alt.state == reached {
+				if g > d[nb.To] {
+					// candidate path to nb is longer than some alternate path
+					continue
+				}
+				if g == d[nb.To] && nextLen >= ap.Len {
+					// candidate path has identical length of some alternate
+					// path but it takes no fewer hops.
+					continue
+				}
+				// cool, we found a better way to get to this node.
+				// record new path data for this node and
+				// update alt with new data and make sure it's on the heap.
+				*ap = PathEnd{From: bestNode, Len: nextLen}
+				labels[nb.To] = nb.Label
+				d[nb.To] = g
+				alt.f = g + h(nb.To)
+				if alt.fx < 0 {
+					heap.Push(&oh, alt)
+				} else {
+					heap.Fix(&oh, alt.fx)
+				}
+			} else {
+				// bestNode being reached for the first time.
+				*ap = PathEnd{From: bestNode, Len: nextLen}
+				labels[nb.To] = nb.Label
+				d[nb.To] = g
+				alt.f = g + h(nb.To)
+				alt.state = reached
+				heap.Push(&oh, alt) // and it's now open for exploration
+			}
+		}
+	}
+	return // no path
+}
+
+// AStarAPath finds a shortest path using the AStarA algorithm.
+//
+// This is a convenience method with a simpler result than the AStarA method.
+// See documentation on the AStarA method.
+//
+// If a path is found, the non-nil node path is returned with the total path
+// distance.  Otherwise the returned path will be nil.
+func (g LabeledAdjacencyList) AStarAPath(start, end NI, h Heuristic, w WeightFunc) ([]NI, float64) {
+	f, _, d, _ := g.AStarA(w, start, end, h)
+	return f.PathTo(end, nil), d
+}
+
+// AStarM is AStarA optimized for monotonic heuristic estimates.
+//
+// Note that this function requires a monotonic heuristic.  Results will
+// not be meaningful if argument h is non-monotonic.
+//
+// See AStarA for general usage.  See Heuristic for notes on monotonicity.
+func (g LabeledAdjacencyList) AStarM(w WeightFunc, start, end NI, h Heuristic) (f FromList, labels []LI, dist float64, ok bool) {
+	// NOTE: AStarM is largely code duplicated from AStarA.
+	// Differences are noted in comments in this method.
+
+	f = NewFromList(len(g))
+	labels = make([]LI, len(g))
+	d := make([]float64, len(g))
+	r := make([]rNode, len(g))
+	for i := range r {
+		r[i].nx = NI(i)
+	}
+	cr := &r[start]
+
+	// difference from AStarA:
+	// instead of a bit to mark a reached node, there are two states,
+	// open and closed. open marks nodes "open" for exploration.
+	// nodes are marked open as they are reached, then marked
+	// closed as they are found to be on the best path.
+	cr.state = open
+
+	cr.f = h(start)
+	rp := f.Paths
+	rp[start] = PathEnd{Len: 1, From: -1}
+	oh := openHeap{cr}
+	for len(oh) > 0 {
+		bestPath := heap.Pop(&oh).(*rNode)
+		bestNode := bestPath.nx
+		if bestNode == end {
+			return f, labels, d[end], true
+		}
+
+		// difference from AStarA:
+		// move nodes to closed list as they are found to be best so far.
+		bestPath.state = closed
+
+		bp := &rp[bestNode]
+		nextLen := bp.Len + 1
+		for _, nb := range g[bestNode] {
+			alt := &r[nb.To]
+
+			// difference from AStarA:
+			// Monotonicity means that f cannot be improved.
+			if alt.state == closed {
+				continue
+			}
+
+			ap := &rp[alt.nx]
+			g := d[bestNode] + w(nb.Label)
+
+			// difference from AStarA:
+			// test for open state, not just reached
+			if alt.state == open {
+
+				if g > d[nb.To] {
+					continue
+				}
+				if g == d[nb.To] && nextLen >= ap.Len {
+					continue
+				}
+				*ap = PathEnd{From: bestNode, Len: nextLen}
+				labels[nb.To] = nb.Label
+				d[nb.To] = g
+				alt.f = g + h(nb.To)
+
+				// difference from AStarA:
+				// we know alt was on the heap because we found it marked open
+				heap.Fix(&oh, alt.fx)
+			} else {
+				*ap = PathEnd{From: bestNode, Len: nextLen}
+				labels[nb.To] = nb.Label
+				d[nb.To] = g
+				alt.f = g + h(nb.To)
+
+				// difference from AStarA:
+				// nodes are opened when first reached
+				alt.state = open
+				heap.Push(&oh, alt)
+			}
+		}
+	}
+	return
+}
+
+// AStarMPath finds a shortest path using the AStarM algorithm.
+//
+// This is a convenience method with a simpler result than the AStarM method.
+// See documentation on the AStarM and AStarA methods.
+//
+// If a path is found, the non-nil node path is returned with the total path
+// distance.  Otherwise the returned path will be nil.
+func (g LabeledAdjacencyList) AStarMPath(start, end NI, h Heuristic, w WeightFunc) ([]NI, float64) {
+	f, _, d, _ := g.AStarM(w, start, end, h)
+	return f.PathTo(end, nil), d
+}
+
+// implement container/heap
+func (h openHeap) Len() int           { return len(h) }
+func (h openHeap) Less(i, j int) bool { return h[i].f < h[j].f }
+func (h openHeap) Swap(i, j int) {
+	h[i], h[j] = h[j], h[i]
+	h[i].fx = i
+	h[j].fx = j
+}
+func (p *openHeap) Push(x interface{}) {
+	h := *p
+	fx := len(h)
+	h = append(h, x.(*rNode))
+	h[fx].fx = fx
+	*p = h
+}
+
+func (p *openHeap) Pop() interface{} {
+	h := *p
+	last := len(h) - 1
+	*p = h[:last]
+	h[last].fx = -1
+	return h[last]
+}
+
+// BellmanFord finds shortest paths from a start node in a weighted directed
+// graph using the Bellman-Ford-Moore algorithm.
+//
+// WeightFunc w must translate arc labels to arc weights.
+// Negative arc weights are allowed but not negative cycles.
+// Loops and parallel arcs are allowed.
+//
+// If the algorithm completes without encountering a negative cycle the method
+// returns shortest paths encoded in a FromList, path distances indexed by
+// node, and return value end = -1.
+//
+// If it encounters a negative cycle reachable from start it returns end >= 0.
+// In this case the cycle can be obtained by calling f.BellmanFordCycle(end).
+//
+// Negative cycles are only detected when reachable from start.  A negative
+// cycle not reachable from start will not prevent the algorithm from finding
+// shortest paths from start.
+//
+// See also NegativeCycle to find a cycle anywhere in the graph, and see
+// HasNegativeCycle for lighter-weight negative cycle detection,
+func (g LabeledDirected) BellmanFord(w WeightFunc, start NI) (f FromList, dist []float64, end NI) {
+	a := g.LabeledAdjacencyList
+	f = NewFromList(len(a))
+	dist = make([]float64, len(a))
+	inf := math.Inf(1)
+	for i := range dist {
+		dist[i] = inf
+	}
+	rp := f.Paths
+	rp[start] = PathEnd{Len: 1, From: -1}
+	dist[start] = 0
+	for _ = range a[1:] {
+		imp := false
+		for from, nbs := range a {
+			fp := &rp[from]
+			d1 := dist[from]
+			for _, nb := range nbs {
+				d2 := d1 + w(nb.Label)
+				to := &rp[nb.To]
+				// TODO improve to break ties
+				if fp.Len > 0 && d2 < dist[nb.To] {
+					*to = PathEnd{From: NI(from), Len: fp.Len + 1}
+					dist[nb.To] = d2
+					imp = true
+				}
+			}
+		}
+		if !imp {
+			break
+		}
+	}
+	for from, nbs := range a {
+		d1 := dist[from]
+		for _, nb := range nbs {
+			if d1+w(nb.Label) < dist[nb.To] {
+				// return nb as end of a path with negative cycle at root
+				return f, dist, NI(from)
+			}
+		}
+	}
+	return f, dist, -1
+}
+
+// BellmanFordCycle decodes a negative cycle detected by BellmanFord.
+//
+// Receiver f and argument end must be results returned from BellmanFord.
+func (f FromList) BellmanFordCycle(end NI) (c []NI) {
+	p := f.Paths
+	var b Bits
+	for b.Bit(end) == 0 {
+		b.SetBit(end, 1)
+		end = p[end].From
+	}
+	for b.Bit(end) == 1 {
+		c = append(c, end)
+		b.SetBit(end, 0)
+		end = p[end].From
+	}
+	for i, j := 0, len(c)-1; i < j; i, j = i+1, j-1 {
+		c[i], c[j] = c[j], c[i]
+	}
+	return
+}
+
+// HasNegativeCycle returns true if the graph contains any negative cycle.
+//
+// HasNegativeCycle uses a Bellman-Ford-like algorithm, but finds negative
+// cycles anywhere in the graph.  Also path information is not computed,
+// reducing memory use somewhat compared to BellmanFord.
+//
+// See also NegativeCycle to obtain the cycle, and see BellmanFord for
+// single source shortest path searches.
+func (g LabeledDirected) HasNegativeCycle(w WeightFunc) bool {
+	a := g.LabeledAdjacencyList
+	dist := make([]float64, len(a))
+	for _ = range a[1:] {
+		imp := false
+		for from, nbs := range a {
+			d1 := dist[from]
+			for _, nb := range nbs {
+				d2 := d1 + w(nb.Label)
+				if d2 < dist[nb.To] {
+					dist[nb.To] = d2
+					imp = true
+				}
+			}
+		}
+		if !imp {
+			break
+		}
+	}
+	for from, nbs := range a {
+		d1 := dist[from]
+		for _, nb := range nbs {
+			if d1+w(nb.Label) < dist[nb.To] {
+				return true // negative cycle
+			}
+		}
+	}
+	return false
+}
+
+// NegativeCycle finds a negative cycle if one exists.
+//
+// NegativeCycle uses a Bellman-Ford-like algorithm, but finds negative
+// cycles anywhere in the graph.  If a negative cycle exists, one will be
+// returned.  The result is nil if no negative cycle exists.
+//
+// See also HasNegativeCycle for lighter-weight cycle detection, and see
+// BellmanFord for single source shortest paths.
+func (g LabeledDirected) NegativeCycle(w WeightFunc) (c []NI) {
+	a := g.LabeledAdjacencyList
+	f := NewFromList(len(a))
+	p := f.Paths
+	for n := range p {
+		p[n] = PathEnd{From: -1, Len: 1}
+	}
+	dist := make([]float64, len(a))
+	for _ = range a {
+		imp := false
+		for from, nbs := range a {
+			fp := &p[from]
+			d1 := dist[from]
+			for _, nb := range nbs {
+				d2 := d1 + w(nb.Label)
+				to := &p[nb.To]
+				if fp.Len > 0 && d2 < dist[nb.To] {
+					*to = PathEnd{From: NI(from), Len: fp.Len + 1}
+					dist[nb.To] = d2
+					imp = true
+				}
+			}
+		}
+		if !imp {
+			return nil
+		}
+	}
+	var vis Bits
+a:
+	for n := range a {
+		end := NI(n)
+		var b Bits
+		for b.Bit(end) == 0 {
+			if vis.Bit(end) == 1 {
+				continue a
+			}
+			vis.SetBit(end, 1)
+			b.SetBit(end, 1)
+			end = p[end].From
+			if end < 0 {
+				continue a
+			}
+		}
+		for b.Bit(end) == 1 {
+			c = append(c, end)
+			b.SetBit(end, 0)
+			end = p[end].From
+		}
+		for i, j := 0, len(c)-1; i < j; i, j = i+1, j-1 {
+			c[i], c[j] = c[j], c[i]
+		}
+		return c
+	}
+	return nil // no negative cycle
+}
+
+// A NodeVisitor is an argument to some graph traversal methods.
+//
+// Graph traversal methods call the visitor function for each node visited.
+// Argument n is the node being visited.
+type NodeVisitor func(n NI)
+
+// An OkNodeVisitor function is an argument to some graph traversal methods.
+//
+// Graph traversal methods call the visitor function for each node visited.
+// The argument n is the node being visited.  If the visitor function
+// returns true, the traversal will continue.  If the visitor function
+// returns false, the traversal will terminate immediately.
+type OkNodeVisitor func(n NI) (ok bool)
+
+// BreadthFirst2 traverses a graph breadth first using a direction
+// optimizing algorithm.
+//
+// The code is experimental and currently seems no faster than the
+// conventional breadth first code.
+//
+// Use AdjacencyList.BreadthFirst instead.
+func BreadthFirst2(g, tr AdjacencyList, ma int, start NI, f *FromList, v OkNodeVisitor) int {
+	if tr == nil {
+		var d Directed
+		d, ma = Directed{g}.Transpose()
+		tr = d.AdjacencyList
+	}
+	switch {
+	case f == nil:
+		e := NewFromList(len(g))
+		f = &e
+	case f.Paths == nil:
+		*f = NewFromList(len(g))
+	}
+	if ma <= 0 {
+		ma = g.ArcSize()
+	}
+	rp := f.Paths
+	level := 1
+	rp[start] = PathEnd{Len: level, From: -1}
+	if !v(start) {
+		f.MaxLen = level
+		return -1
+	}
+	nReached := 1 // accumulated for a return value
+	// the frontier consists of nodes all at the same level
+	frontier := []NI{start}
+	mf := len(g[start])     // number of arcs leading out from frontier
+	ctb := ma / 10          // threshold change from top-down to bottom-up
+	k14 := 14 * ma / len(g) // 14 * mean degree
+	cbt := len(g) / k14     // threshold change from bottom-up to top-down
+	//	var fBits, nextb big.Int
+	fBits := make([]bool, len(g))
+	nextb := make([]bool, len(g))
+	zBits := make([]bool, len(g))
+	for {
+		// top down step
+		level++
+		var next []NI
+		for _, n := range frontier {
+			for _, nb := range g[n] {
+				if rp[nb].Len == 0 {
+					rp[nb] = PathEnd{From: n, Len: level}
+					if !v(nb) {
+						f.MaxLen = level
+						return -1
+					}
+					next = append(next, nb)
+					nReached++
+				}
+			}
+		}
+		if len(next) == 0 {
+			break
+		}
+		frontier = next
+		if mf > ctb {
+			// switch to bottom up!
+		} else {
+			// stick with top down
+			continue
+		}
+		// convert frontier representation
+		nf := 0 // number of vertices on the frontier
+		for _, n := range frontier {
+			//			fBits.SetBit(&fBits, n, 1)
+			fBits[n] = true
+			nf++
+		}
+	bottomUpLoop:
+		level++
+		nNext := 0
+		for n := range tr {
+			if rp[n].Len == 0 {
+				for _, nb := range tr[n] {
+					//					if fBits.Bit(nb) == 1 {
+					if fBits[nb] {
+						rp[n] = PathEnd{From: nb, Len: level}
+						if !v(nb) {
+							f.MaxLen = level
+							return -1
+						}
+						//						nextb.SetBit(&nextb, n, 1)
+						nextb[n] = true
+						nReached++
+						nNext++
+						break
+					}
+				}
+			}
+		}
+		if nNext == 0 {
+			break
+		}
+		fBits, nextb = nextb, fBits
+		//		nextb.SetInt64(0)
+		copy(nextb, zBits)
+		nf = nNext
+		if nf < cbt {
+			// switch back to top down!
+		} else {
+			// stick with bottom up
+			goto bottomUpLoop
+		}
+		// convert frontier representation
+		mf = 0
+		frontier = frontier[:0]
+		for n := range g {
+			//			if fBits.Bit(n) == 1 {
+			if fBits[n] {
+				frontier = append(frontier, NI(n))
+				mf += len(g[n])
+				fBits[n] = false
+			}
+		}
+		//		fBits.SetInt64(0)
+	}
+	f.MaxLen = level - 1
+	return nReached
+}
+
+// DAGMinDistPath finds a single shortest path.
+//
+// Shortest means minimum sum of arc weights.
+//
+// Returned is the path and distance as returned by FromList.PathTo.
+//
+// This is a convenience method.  See DAGOptimalPaths for more options.
+func (g LabeledDirected) DAGMinDistPath(start, end NI, w WeightFunc) ([]NI, float64, error) {
+	return g.dagPath(start, end, w, false)
+}
+
+// DAGMaxDistPath finds a single longest path.
+//
+// Longest means maximum sum of arc weights.
+//
+// Returned is the path and distance as returned by FromList.PathTo.
+//
+// This is a convenience method.  See DAGOptimalPaths for more options.
+func (g LabeledDirected) DAGMaxDistPath(start, end NI, w WeightFunc) ([]NI, float64, error) {
+	return g.dagPath(start, end, w, true)
+}
+
+func (g LabeledDirected) dagPath(start, end NI, w WeightFunc, longest bool) ([]NI, float64, error) {
+	o, _ := g.Topological()
+	if o == nil {
+		return nil, 0, fmt.Errorf("not a DAG")
+	}
+	f, dist, _ := g.DAGOptimalPaths(start, end, o, w, longest)
+	if f.Paths[end].Len == 0 {
+		return nil, 0, fmt.Errorf("no path from %d to %d", start, end)
+	}
+	return f.PathTo(end, nil), dist[end], nil
+}
+
+// DAGOptimalPaths finds either longest or shortest distance paths in a
+// directed acyclic graph.
+//
+// Path distance is the sum of arc weights on the path.
+// Negative arc weights are allowed.
+// Where multiple paths exist with the same distance, the path length
+// (number of nodes) is used as a tie breaker.
+//
+// Receiver g must be a directed acyclic graph.  Argument o must be either nil
+// or a topological ordering of g.  If nil, a topologcal ordering is
+// computed internally.  If longest is true, an optimal path is a longest
+// distance path.  Otherwise it is a shortest distance path.
+//
+// Argument start is the start node for paths, end is the end node.  If end
+// is a valid node number, the method returns as soon as the optimal path
+// to end is found.  If end is -1, all optimal paths from start are found.
+//
+// Paths and path distances are encoded in the returned FromList and dist
+// slice.   The number of nodes reached is returned as nReached.
+func (g LabeledDirected) DAGOptimalPaths(start, end NI, ordering []NI, w WeightFunc, longest bool) (f FromList, dist []float64, nReached int) {
+	a := g.LabeledAdjacencyList
+	f = NewFromList(len(a))
+	dist = make([]float64, len(a))
+	if ordering == nil {
+		ordering, _ = g.Topological()
+	}
+	// search ordering for start
+	o := 0
+	for ordering[o] != start {
+		o++
+	}
+	var fBetter func(cand, ext float64) bool
+	var iBetter func(cand, ext int) bool
+	if longest {
+		fBetter = func(cand, ext float64) bool { return cand > ext }
+		iBetter = func(cand, ext int) bool { return cand > ext }
+	} else {
+		fBetter = func(cand, ext float64) bool { return cand < ext }
+		iBetter = func(cand, ext int) bool { return cand < ext }
+	}
+	p := f.Paths
+	p[start] = PathEnd{From: -1, Len: 1}
+	f.MaxLen = 1
+	leaves := &f.Leaves
+	leaves.SetBit(start, 1)
+	nReached = 1
+	for n := start; n != end; n = ordering[o] {
+		if p[n].Len > 0 && len(a[n]) > 0 {
+			nDist := dist[n]
+			candLen := p[n].Len + 1 // len for any candidate arc followed from n
+			for _, to := range a[n] {
+				leaves.SetBit(to.To, 1)
+				candDist := nDist + w(to.Label)
+				switch {
+				case p[to.To].Len == 0: // first path to node to.To
+					nReached++
+				case fBetter(candDist, dist[to.To]): // better distance
+				case candDist == dist[to.To] && iBetter(candLen, p[to.To].Len): // same distance but better path length
+				default:
+					continue
+				}
+				dist[to.To] = candDist
+				p[to.To] = PathEnd{From: n, Len: candLen}
+				if candLen > f.MaxLen {
+					f.MaxLen = candLen
+				}
+			}
+			leaves.SetBit(n, 0)
+		}
+		o++
+		if o == len(ordering) {
+			break
+		}
+	}
+	return
+}
+
+// Dijkstra finds shortest paths by Dijkstra's algorithm.
+//
+// Shortest means shortest distance where distance is the
+// sum of arc weights.  Where multiple paths exist with the same distance,
+// a path with the minimum number of nodes is returned.
+//
+// As usual for Dijkstra's algorithm, arc weights must be non-negative.
+// Graphs may be directed or undirected.  Loops and parallel arcs are
+// allowed.
+func (g LabeledAdjacencyList) Dijkstra(start, end NI, w WeightFunc) (f FromList, dist []float64, reached int) {
+	r := make([]tentResult, len(g))
+	for i := range r {
+		r[i].nx = NI(i)
+	}
+	f = NewFromList(len(g))
+	dist = make([]float64, len(g))
+	current := start
+	rp := f.Paths
+	rp[current] = PathEnd{Len: 1, From: -1} // path length at start is 1 node
+	cr := &r[current]
+	cr.dist = 0    // distance at start is 0.
+	cr.done = true // mark start done.  it skips the heap.
+	nDone := 1     // accumulated for a return value
+	var t tent
+	for current != end {
+		nextLen := rp[current].Len + 1
+		for _, nb := range g[current] {
+			// d.arcVis++
+			hr := &r[nb.To]
+			if hr.done {
+				continue // skip nodes already done
+			}
+			dist := cr.dist + w(nb.Label)
+			vl := rp[nb.To].Len
+			visited := vl > 0
+			if visited {
+				if dist > hr.dist {
+					continue // distance is worse
+				}
+				// tie breaker is a nice touch and doesn't seem to
+				// impact performance much.
+				if dist == hr.dist && nextLen >= vl {
+					continue // distance same, but number of nodes is no better
+				}
+			}
+			// the path through current to this node is shortest so far.
+			// record new path data for this node and update tentative set.
+			hr.dist = dist
+			rp[nb.To].Len = nextLen
+			rp[nb.To].From = current
+			if visited {
+				heap.Fix(&t, hr.fx)
+			} else {
+				heap.Push(&t, hr)
+			}
+		}
+		//d.ndVis++
+		if len(t) == 0 {
+			return f, dist, nDone // no more reachable nodes. AllPaths normal return
+		}
+		// new current is node with smallest tentative distance
+		cr = heap.Pop(&t).(*tentResult)
+		cr.done = true
+		nDone++
+		current = cr.nx
+		dist[current] = cr.dist // store final distance
+	}
+	// normal return for single shortest path search
+	return f, dist, -1
+}
+
+// DijkstraPath finds a single shortest path.
+//
+// Returned is the path and distance as returned by FromList.PathTo.
+func (g LabeledAdjacencyList) DijkstraPath(start, end NI, w WeightFunc) ([]NI, float64) {
+	f, dist, _ := g.Dijkstra(start, end, w)
+	return f.PathTo(end, nil), dist[end]
+}
+
+// tent implements container/heap
+func (t tent) Len() int           { return len(t) }
+func (t tent) Less(i, j int) bool { return t[i].dist < t[j].dist }
+func (t tent) Swap(i, j int) {
+	t[i], t[j] = t[j], t[i]
+	t[i].fx = i
+	t[j].fx = j
+}
+func (s *tent) Push(x interface{}) {
+	nd := x.(*tentResult)
+	nd.fx = len(*s)
+	*s = append(*s, nd)
+}
+func (s *tent) Pop() interface{} {
+	t := *s
+	last := len(t) - 1
+	*s = t[:last]
+	return t[last]
+}
+
+type tentResult struct {
+	dist float64 // tentative distance, sum of arc weights
+	nx   NI      // slice index, "node id"
+	fx   int     // heap.Fix index
+	done bool
+}
+
+type tent []*tentResult
diff --git a/vendor/github.com/soniakeys/graph/travis.sh b/vendor/github.com/soniakeys/graph/travis.sh
new file mode 100644
index 0000000..5a8030a
--- /dev/null
+++ b/vendor/github.com/soniakeys/graph/travis.sh
@@ -0,0 +1,11 @@
+#!/bin/bash
+set -ex
+go test ./...
+if [ "$TRAVIS_GO_VERSION" = "1.6" ]; then
+ GOARCH=386 go test ./...
+ go tool vet -example .
+ go get github.com/client9/misspell/cmd/misspell
+ go get github.com/soniakeys/vetc
+ misspell -error * */* */*/*
+ vetc
+fi
diff --git a/vendor/github.com/soniakeys/graph/undir.go b/vendor/github.com/soniakeys/graph/undir.go
new file mode 100644
index 0000000..75a7f24
--- /dev/null
+++ b/vendor/github.com/soniakeys/graph/undir.go
@@ -0,0 +1,321 @@
+// Copyright 2014 Sonia Keys
+// License MIT: http://opensource.org/licenses/MIT
+
+package graph
+
+// undir.go has methods specific to undirected graphs, Undirected and
+// LabeledUndirected.
+
+import "errors"
+
+// AddEdge adds an edge to a graph.
+//
+// It can be useful for constructing undirected graphs.
+//
+// When n1 and n2 are distinct, it adds the arc n1->n2 and the reciprocal
+// n2->n1.  When n1 and n2 are the same, it adds a single arc loop.
+//
+// The pointer receiver allows the method to expand the graph as needed
+// to include the values n1 and n2.  If n1 or n2 happen to be greater than
+// len(*p) the method does not panic, but simply expands the graph.
+func (p *Undirected) AddEdge(n1, n2 NI) {
+	// Similar code in LabeledAdjacencyList.AddEdge.
+
+	// determine max of the two end points
+	max := n1
+	if n2 > max {
+		max = n2
+	}
+	// expand graph if needed, to include both
+	g := p.AdjacencyList
+	if int(max) >= len(g) {
+		p.AdjacencyList = make(AdjacencyList, max+1)
+		copy(p.AdjacencyList, g)
+		g = p.AdjacencyList
+	}
+	// create one half-arc,
+	g[n1] = append(g[n1], n2)
+	// and except for loops, create the reciprocal
+	if n1 != n2 {
+		g[n2] = append(g[n2], n1)
+	}
+}
+
+// EulerianCycleD for undirected graphs is a bit of an experiment.
+//
+// It is about the same as the directed version, but modified for an undirected
+// multigraph.
+//
+// Parameter m in this case must be the size of the undirected graph -- the
+// number of edges.  Use Undirected.Size if the size is unknown.
+//
+// It works, but contains an extra loop that I think spoils the time
+// complexity.  Probably still pretty fast in practice, but a different
+// graph representation might be better.
+func (g Undirected) EulerianCycleD(m int) ([]NI, error) {
+	if len(g.AdjacencyList) == 0 {
+		return nil, nil
+	}
+	e := newEulerian(g.AdjacencyList, m)
+	for e.s >= 0 {
+		v := e.top()
+		e.pushUndir() // call modified method
+		if e.top() != v {
+			return nil, errors.New("not balanced")
+		}
+		e.keep()
+	}
+	if !e.uv.Zero() {
+		return nil, errors.New("not strongly connected")
+	}
+	return e.p, nil
+}
+
+// TarjanBiconnectedComponents decomposes a graph into maximal biconnected
+// components, components for which if any node were removed the component
+// would remain connected.
+//
+// The receiver g must be a simple graph.  The method calls the emit argument
+// for each component identified, as long as emit returns true.  If emit
+// returns false, TarjanBiconnectedComponents returns immediately.
+//
+// See also the eqivalent labeled TarjanBiconnectedComponents.
+func (g Undirected) TarjanBiconnectedComponents(emit func([]Edge) bool) {
+	// Implemented closely to pseudocode in "Depth-first search and linear
+	// graph algorithms", Robert Tarjan, SIAM J. Comput. Vol. 1, No. 2,
+	// June 1972.
+	//
+	// Note Tarjan's "adjacency structure" is graph.AdjacencyList,
+	// His "adjacency list" is an element of a graph.AdjacencyList, also
+	// termed a "to-list", "neighbor list", or "child list."
+	number := make([]int, len(g.AdjacencyList))
+	lowpt := make([]int, len(g.AdjacencyList))
+	var stack []Edge
+	var i int
+	var biconnect func(NI, NI) bool
+	biconnect = func(v, u NI) bool {
+		i++
+		number[v] = i
+		lowpt[v] = i
+		for _, w := range g.AdjacencyList[v] {
+			if number[w] == 0 {
+				stack = append(stack, Edge{v, w})
+				if !biconnect(w, v) {
+					return false
+				}
+				if lowpt[w] < lowpt[v] {
+					lowpt[v] = lowpt[w]
+				}
+				if lowpt[w] >= number[v] {
+					var bcc []Edge
+					top := len(stack) - 1
+					for number[stack[top].N1] >= number[w] {
+						bcc = append(bcc, stack[top])
+						stack = stack[:top]
+						top--
+					}
+					bcc = append(bcc, stack[top])
+					stack = stack[:top]
+					top--
+					if !emit(bcc) {
+						return false
+					}
+				}
+			} else if number[w] < number[v] && w != u {
+				stack = append(stack, Edge{v, w})
+				if number[w] < lowpt[v] {
+					lowpt[v] = number[w]
+				}
+			}
+		}
+		return true
+	}
+	for w := range g.AdjacencyList {
+		if number[w] == 0 && !biconnect(NI(w), 0) {
+			return
+		}
+	}
+}
+
+/* half-baked.  Read the 72 paper.  Maybe revisit at some point.
+type BiconnectedComponents struct {
+	Graph  AdjacencyList
+	Start  int
+	Cuts   big.Int // bitmap of node cuts
+	From   []int   // from-tree
+	Leaves []int   // leaves of from-tree
+}
+
+func NewBiconnectedComponents(g Undirected) *BiconnectedComponents {
+	return &BiconnectedComponents{
+		Graph: g,
+		From:  make([]int, len(g)),
+	}
+}
+
+func (b *BiconnectedComponents) Find(start int) {
+	g := b.Graph
+	depth := make([]int, len(g))
+	low := make([]int, len(g))
+	// reset from any previous run
+	b.Cuts.SetInt64(0)
+	bf := b.From
+	for n := range bf {
+		bf[n] = -1
+	}
+	b.Leaves = b.Leaves[:0]
+	d := 1 // depth. d > 0 means visited
+	depth[start] = d
+	low[start] = d
+	d++
+	var df func(int, int)
+	df = func(from, n int) {
+		bf[n] = from
+		depth[n] = d
+		dn := d
+		l := d
+		d++
+		cut := false
+		leaf := true
+		for _, nb := range g[n] {
+			if depth[nb] == 0 {
+				leaf = false
+				df(n, nb)
+				if low[nb] < l {
+					l = low[nb]
+				}
+				if low[nb] >= dn {
+					cut = true
+				}
+			} else if nb != from && depth[nb] < l {
+				l = depth[nb]
+			}
+		}
+		low[n] = l
+		if cut {
+			b.Cuts.SetBit(&b.Cuts, n, 1)
+		}
+		if leaf {
+			b.Leaves = append(b.Leaves, n)
+		}
+		d--
+	}
+	nbs := g[start]
+	if len(nbs) == 0 {
+		return
+	}
+	df(start, nbs[0])
+	var rc uint
+	for _, nb := range nbs[1:] {
+		if depth[nb] == 0 {
+			rc = 1
+			df(start, nb)
+		}
+	}
+	b.Cuts.SetBit(&b.Cuts, start, rc)
+	return
+}
+*/
+
+// AddEdge adds an edge to a labeled graph.
+//
+// It can be useful for constructing undirected graphs.
+//
+// When n1 and n2 are distinct, it adds the arc n1->n2 and the reciprocal
+// n2->n1.  When n1 and n2 are the same, it adds a single arc loop.
+//
+// If the edge already exists in *p, a parallel edge is added.
+//
+// The pointer receiver allows the method to expand the graph as needed
+// to include the values n1 and n2.  If n1 or n2 happen to be greater than
+// len(*p) the method does not panic, but simply expands the graph.
+func (p *LabeledUndirected) AddEdge(e Edge, l LI) {
+	// Similar code in AdjacencyList.AddEdge.
+
+	// determine max of the two end points
+	max := e.N1
+	if e.N2 > max {
+		max = e.N2
+	}
+	// expand graph if needed, to include both
+	g := p.LabeledAdjacencyList
+	if max >= NI(len(g)) {
+		p.LabeledAdjacencyList = make(LabeledAdjacencyList, max+1)
+		copy(p.LabeledAdjacencyList, g)
+		g = p.LabeledAdjacencyList
+	}
+	// create one half-arc,
+	g[e.N1] = append(g[e.N1], Half{To: e.N2, Label: l})
+	// and except for loops, create the reciprocal
+	if e.N1 != e.N2 {
+		g[e.N2] = append(g[e.N2], Half{To: e.N1, Label: l})
+	}
+}
+
+// TarjanBiconnectedComponents decomposes a graph into maximal biconnected
+// components, components for which if any node were removed the component
+// would remain connected.
+//
+// The receiver g must be a simple graph.  The method calls the emit argument
+// for each component identified, as long as emit returns true.  If emit
+// returns false, TarjanBiconnectedComponents returns immediately.
+//
+// See also the eqivalent unlabeled TarjanBiconnectedComponents.
+func (g LabeledUndirected) TarjanBiconnectedComponents(emit func([]LabeledEdge) bool) {
+	// Implemented closely to pseudocode in "Depth-first search and linear
+	// graph algorithms", Robert Tarjan, SIAM J. Comput. Vol. 1, No. 2,
+	// June 1972.
+	//
+	// Note Tarjan's "adjacency structure" is graph.AdjacencyList,
+	// His "adjacency list" is an element of a graph.AdjacencyList, also
+	// termed a "to-list", "neighbor list", or "child list."
+	//
+	// Nearly identical code in undir.go.
+	number := make([]int, len(g.LabeledAdjacencyList))
+	lowpt := make([]int, len(g.LabeledAdjacencyList))
+	var stack []LabeledEdge
+	var i int
+	var biconnect func(NI, NI) bool
+	biconnect = func(v, u NI) bool {
+		i++
+		number[v] = i
+		lowpt[v] = i
+		for _, w := range g.LabeledAdjacencyList[v] {
+			if number[w.To] == 0 {
+				stack = append(stack, LabeledEdge{Edge{v, w.To}, w.Label})
+				if !biconnect(w.To, v) {
+					return false
+				}
+				if lowpt[w.To] < lowpt[v] {
+					lowpt[v] = lowpt[w.To]
+				}
+				if lowpt[w.To] >= number[v] {
+					var bcc []LabeledEdge
+					top := len(stack) - 1
+					for number[stack[top].N1] >= number[w.To] {
+						bcc = append(bcc, stack[top])
+						stack = stack[:top]
+						top--
+					}
+					bcc = append(bcc, stack[top])
+					stack = stack[:top]
+					top--
+					if !emit(bcc) {
+						return false
+					}
+				}
+			} else if number[w.To] < number[v] && w.To != u {
+				stack = append(stack, LabeledEdge{Edge{v, w.To}, w.Label})
+				if number[w.To] < lowpt[v] {
+					lowpt[v] = number[w.To]
+				}
+			}
+		}
+		return true
+	}
+	for w := range g.LabeledAdjacencyList {
+		if number[w] == 0 && !biconnect(NI(w), 0) {
+			return
+		}
+	}
+}
diff --git a/vendor/github.com/soniakeys/graph/undir_RO.go b/vendor/github.com/soniakeys/graph/undir_RO.go
new file mode 100644
index 0000000..fd8e377
--- /dev/null
+++ b/vendor/github.com/soniakeys/graph/undir_RO.go
@@ -0,0 +1,659 @@
+// Copyright 2014 Sonia Keys
+// License MIT: http://opensource.org/licenses/MIT
+
+package graph
+
+// undir_RO.go is code generated from undir_cg.go by directives in graph.go.
+// Editing undir_cg.go is okay.  It is the code generation source.
+// DO NOT EDIT undir_RO.go.
+// The RO means read only and it is upper case RO to slow you down a bit
+// in case you start to edit the file.
+
+// Bipartite determines if a connected component of an undirected graph
+// is bipartite, a component where nodes can be partitioned into two sets
+// such that every edge in the component goes from one set to the other.
+//
+// Argument n can be any representative node of the component.
+//
+// If the component is bipartite, Bipartite returns true and a two-coloring
+// of the component.  Each color set is returned as a bitmap.  If the component
+// is not bipartite, Bipartite returns false and a representative odd cycle.
+//
+// There are equivalent labeled and unlabeled versions of this method.
+func (g Undirected) Bipartite(n NI) (b bool, c1, c2 Bits, oc []NI) {
+	b = true
+	var open bool
+	var df func(n NI, c1, c2 *Bits)
+	df = func(n NI, c1, c2 *Bits) {
+		c1.SetBit(n, 1)
+		for _, nb := range g.AdjacencyList[n] {
+			if c1.Bit(nb) == 1 {
+				b = false
+				oc = []NI{nb, n}
+				open = true
+				return
+			}
+			if c2.Bit(nb) == 1 {
+				continue
+			}
+			df(nb, c2, c1)
+			if b {
+				continue
+			}
+			switch {
+			case !open:
+			case n == oc[0]:
+				open = false
+			default:
+				oc = append(oc, n)
+			}
+			return
+		}
+	}
+	df(n, &c1, &c2)
+	if b {
+		return b, c1, c2, nil
+	}
+	return b, Bits{}, Bits{}, oc
+}
+
+// BronKerbosch1 finds maximal cliques in an undirected graph.
+//
+// The graph must not contain parallel edges or loops.
+//
+// See https://en.wikipedia.org/wiki/Clique_(graph_theory) and
+// https://en.wikipedia.org/wiki/Bron%E2%80%93Kerbosch_algorithm for background.
+//
+// This method implements the BronKerbosch1 algorithm of WP; that is,
+// the original algorithm without improvements.
+//
+// The method calls the emit argument for each maximal clique in g, as long
+// as emit returns true.  If emit returns false, BronKerbosch1 returns
+// immediately.
+//
+// There are equivalent labeled and unlabeled versions of this method.
+//
+// See also more sophisticated variants BronKerbosch2 and BronKerbosch3.
+func (g Undirected) BronKerbosch1(emit func([]NI) bool) {
+	a := g.AdjacencyList
+	var f func(R, P, X *Bits) bool
+	f = func(R, P, X *Bits) bool {
+		switch {
+		case !P.Zero():
+			var r2, p2, x2 Bits
+			pf := func(n NI) bool {
+				r2.Set(*R)
+				r2.SetBit(n, 1)
+				p2.Clear()
+				x2.Clear()
+				for _, to := range a[n] {
+					if P.Bit(to) == 1 {
+						p2.SetBit(to, 1)
+					}
+					if X.Bit(to) == 1 {
+						x2.SetBit(to, 1)
+					}
+				}
+				if !f(&r2, &p2, &x2) {
+					return false
+				}
+				P.SetBit(n, 0)
+				X.SetBit(n, 1)
+				return true
+			}
+			if !P.Iterate(pf) {
+				return false
+			}
+		case X.Zero():
+			return emit(R.Slice())
+		}
+		return true
+	}
+	var R, P, X Bits
+	P.SetAll(len(a))
+	f(&R, &P, &X)
+}
+
+// BKPivotMaxDegree is a strategy for BronKerbosch methods.
+//
+// To use it, take the method value (see golang.org/ref/spec#Method_values)
+// and pass it as the argument to BronKerbosch2 or 3.
+//
+// The strategy is to pick the node from P or X with the maximum degree
+// (number of edges) in g.  Note this is a shortcut from evaluating degrees
+// in P.
+//
+// There are equivalent labeled and unlabeled versions of this method.
+func (g Undirected) BKPivotMaxDegree(P, X *Bits) (p NI) {
+	// choose pivot u as highest degree node from P or X
+	a := g.AdjacencyList
+	maxDeg := -1
+	P.Iterate(func(n NI) bool { // scan P
+		if d := len(a[n]); d > maxDeg {
+			p = n
+			maxDeg = d
+		}
+		return true
+	})
+	X.Iterate(func(n NI) bool { // scan X
+		if d := len(a[n]); d > maxDeg {
+			p = n
+			maxDeg = d
+		}
+		return true
+	})
+	return
+}
+
+// BKPivotMinP is a strategy for BronKerbosch methods.
+//
+// To use it, take the method value (see golang.org/ref/spec#Method_values)
+// and pass it as the argument to BronKerbosch2 or 3.
+//
+// The strategy is to simply pick the first node in P.
+//
+// There are equivalent labeled and unlabeled versions of this method.
+func (g Undirected) BKPivotMinP(P, X *Bits) NI {
+	return P.From(0)
+}
+
+// BronKerbosch2 finds maximal cliques in an undirected graph.
+//
+// The graph must not contain parallel edges or loops.
+//
+// See https://en.wikipedia.org/wiki/Clique_(graph_theory) and
+// https://en.wikipedia.org/wiki/Bron%E2%80%93Kerbosch_algorithm for background.
+//
+// This method implements the BronKerbosch2 algorithm of WP; that is,
+// the original algorithm plus pivoting.
+//
+// The argument is a pivot function that must return a node of P or X.
+// P is guaranteed to contain at least one node.  X is not.
+// For example see BKPivotMaxDegree.
+//
+// The method calls the emit argument for each maximal clique in g, as long
+// as emit returns true.  If emit returns false, BronKerbosch1 returns
+// immediately.
+//
+// There are equivalent labeled and unlabeled versions of this method.
+//
+// See also simpler variant BronKerbosch1 and more sophisticated variant
+// BronKerbosch3.
+func (g Undirected) BronKerbosch2(pivot func(P, X *Bits) NI, emit func([]NI) bool) {
+	a := g.AdjacencyList
+	var f func(R, P, X *Bits) bool
+	f = func(R, P, X *Bits) bool {
+		switch {
+		case !P.Zero():
+			var r2, p2, x2, pnu Bits
+			// compute P \ N(u).  next 5 lines are only difference from BK1
+			pnu.Set(*P)
+			for _, to := range a[pivot(P, X)] {
+				pnu.SetBit(to, 0)
+			}
+			// remaining code like BK1
+			pf := func(n NI) bool {
+				r2.Set(*R)
+				r2.SetBit(n, 1)
+				p2.Clear()
+				x2.Clear()
+				for _, to := range a[n] {
+					if P.Bit(to) == 1 {
+						p2.SetBit(to, 1)
+					}
+					if X.Bit(to) == 1 {
+						x2.SetBit(to, 1)
+					}
+				}
+				if !f(&r2, &p2, &x2) {
+					return false
+				}
+				P.SetBit(n, 0)
+				X.SetBit(n, 1)
+				return true
+			}
+			if !pnu.Iterate(pf) {
+				return false
+			}
+		case X.Zero():
+			return emit(R.Slice())
+		}
+		return true
+	}
+	var R, P, X Bits
+	P.SetAll(len(a))
+	f(&R, &P, &X)
+}
+
+// BronKerbosch3 finds maximal cliques in an undirected graph.
+//
+// The graph must not contain parallel edges or loops.
+//
+// See https://en.wikipedia.org/wiki/Clique_(graph_theory) and
+// https://en.wikipedia.org/wiki/Bron%E2%80%93Kerbosch_algorithm for background.
+//
+// This method implements the BronKerbosch3 algorithm of WP; that is,
+// the original algorithm with pivoting and degeneracy ordering.
+//
+// The argument is a pivot function that must return a node of P or X.
+// P is guaranteed to contain at least one node.  X is not.
+// For example see BKPivotMaxDegree.
+//
+// The method calls the emit argument for each maximal clique in g, as long
+// as emit returns true.  If emit returns false, BronKerbosch1 returns
+// immediately.
+//
+// There are equivalent labeled and unlabeled versions of this method.
+//
+// See also simpler variants BronKerbosch1 and BronKerbosch2.
+func (g Undirected) BronKerbosch3(pivot func(P, X *Bits) NI, emit func([]NI) bool) {
+	a := g.AdjacencyList
+	var f func(R, P, X *Bits) bool
+	f = func(R, P, X *Bits) bool {
+		switch {
+		case !P.Zero():
+			var r2, p2, x2, pnu Bits
+			// compute P \ N(u).  next lines are only difference from BK1
+			pnu.Set(*P)
+			for _, to := range a[pivot(P, X)] {
+				pnu.SetBit(to, 0)
+			}
+			// remaining code like BK2
+			pf := func(n NI) bool {
+				r2.Set(*R)
+				r2.SetBit(n, 1)
+				p2.Clear()
+				x2.Clear()
+				for _, to := range a[n] {
+					if P.Bit(to) == 1 {
+						p2.SetBit(to, 1)
+					}
+					if X.Bit(to) == 1 {
+						x2.SetBit(to, 1)
+					}
+				}
+				if !f(&r2, &p2, &x2) {
+					return false
+				}
+				P.SetBit(n, 0)
+				X.SetBit(n, 1)
+				return true
+			}
+			if !pnu.Iterate(pf) {
+				return false
+			}
+		case X.Zero():
+			return emit(R.Slice())
+		}
+		return true
+	}
+	var R, P, X Bits
+	P.SetAll(len(a))
+	// code above same as BK2
+	// code below new to BK3
+	_, ord, _ := g.Degeneracy()
+	var p2, x2 Bits
+	for _, n := range ord {
+		R.SetBit(n, 1)
+		p2.Clear()
+		x2.Clear()
+		for _, to := range a[n] {
+			if P.Bit(to) == 1 {
+				p2.SetBit(to, 1)
+			}
+			if X.Bit(to) == 1 {
+				x2.SetBit(to, 1)
+			}
+		}
+		if !f(&R, &p2, &x2) {
+			return
+		}
+		R.SetBit(n, 0)
+		P.SetBit(n, 0)
+		X.SetBit(n, 1)
+	}
+}
+
+// ConnectedComponentBits returns a function that iterates over connected
+// components of g, returning a member bitmap for each.
+//
+// Each call of the returned function returns the order (number of nodes)
+// and bits of a connected component.  The returned function returns zeros
+// after returning all connected components.
+//
+// There are equivalent labeled and unlabeled versions of this method.
+//
+// See also ConnectedComponentReps, which has lighter weight return values.
+func (g Undirected) ConnectedComponentBits() func() (order int, bits Bits) {
+	a := g.AdjacencyList
+	var vg Bits  // nodes visited in graph
+	var vc *Bits // nodes visited in current component
+	var nc int
+	var df func(NI)
+	df = func(n NI) {
+		vg.SetBit(n, 1)
+		vc.SetBit(n, 1)
+		nc++
+		for _, nb := range a[n] {
+			if vg.Bit(nb) == 0 {
+				df(nb)
+			}
+		}
+		return
+	}
+	var n NI
+	return func() (o int, bits Bits) {
+		for ; n < NI(len(a)); n++ {
+			if vg.Bit(n) == 0 {
+				vc = &bits
+				nc = 0
+				df(n)
+				return nc, bits
+			}
+		}
+		return
+	}
+}
+
+// ConnectedComponentLists returns a function that iterates over connected
+// components of g, returning the member list of each.
+//
+// Each call of the returned function returns a node list of a connected
+// component.  The returned function returns nil after returning all connected
+// components.
+//
+// There are equivalent labeled and unlabeled versions of this method.
+//
+// See also ConnectedComponentReps, which has lighter weight return values.
+func (g Undirected) ConnectedComponentLists() func() []NI {
+	a := g.AdjacencyList
+	var vg Bits // nodes visited in graph
+	var m []NI  // members of current component
+	var df func(NI)
+	df = func(n NI) {
+		vg.SetBit(n, 1)
+		m = append(m, n)
+		for _, nb := range a[n] {
+			if vg.Bit(nb) == 0 {
+				df(nb)
+			}
+		}
+		return
+	}
+	var n NI
+	return func() []NI {
+		for ; n < NI(len(a)); n++ {
+			if vg.Bit(n) == 0 {
+				m = nil
+				df(n)
+				return m
+			}
+		}
+		return nil
+	}
+}
+
+// ConnectedComponentReps returns a representative node from each connected
+// component of g.
+//
+// Returned is a slice with a single representative node from each connected
+// component and also a parallel slice with the order, or number of nodes,
+// in the corresponding component.
+//
+// This is fairly minimal information describing connected components.
+// From a representative node, other nodes in the component can be reached
+// by depth first traversal for example.
+//
+// There are equivalent labeled and unlabeled versions of this method.
+//
+// See also ConnectedComponentBits and ConnectedComponentLists which can
+// collect component members in a single traversal, and IsConnected which
+// is an even simpler boolean test.
+func (g Undirected) ConnectedComponentReps() (reps []NI, orders []int) {
+	a := g.AdjacencyList
+	var c Bits
+	var o int
+	var df func(NI)
+	df = func(n NI) {
+		c.SetBit(n, 1)
+		o++
+		for _, nb := range a[n] {
+			if c.Bit(nb) == 0 {
+				df(nb)
+			}
+		}
+		return
+	}
+	for n := range a {
+		if c.Bit(NI(n)) == 0 {
+			reps = append(reps, NI(n))
+			o = 0
+			df(NI(n))
+			orders = append(orders, o)
+		}
+	}
+	return
+}
+
+// Copy makes a deep copy of g.
+// Copy also computes the arc size ma, the number of arcs.
+//
+// There are equivalent labeled and unlabeled versions of this method.
+func (g Undirected) Copy() (c Undirected, ma int) {
+	l, s := g.AdjacencyList.Copy()
+	return Undirected{l}, s
+}
+
+// Degeneracy computes k-degeneracy, vertex ordering and k-cores.
+//
+// See Wikipedia https://en.wikipedia.org/wiki/Degeneracy_(graph_theory)
+//
+// There are equivalent labeled and unlabeled versions of this method.
+func (g Undirected) Degeneracy() (k int, ordering []NI, cores []int) {
+	a := g.AdjacencyList
+	// WP algorithm
+	ordering = make([]NI, len(a))
+	var L Bits
+	d := make([]int, len(a))
+	var D [][]NI
+	for v, nb := range a {
+		dv := len(nb)
+		d[v] = dv
+		for len(D) <= dv {
+			D = append(D, nil)
+		}
+		D[dv] = append(D[dv], NI(v))
+	}
+	for ox := range a {
+		// find a non-empty D
+		i := 0
+		for len(D[i]) == 0 {
+			i++
+		}
+		// k is max(i, k)
+		if i > k {
+			for len(cores) <= i {
+				cores = append(cores, 0)
+			}
+			cores[k] = ox
+			k = i
+		}
+		// select from D[i]
+		Di := D[i]
+		last := len(Di) - 1
+		v := Di[last]
+		// Add v to ordering, remove from Di
+		ordering[ox] = v
+		L.SetBit(v, 1)
+		D[i] = Di[:last]
+		// move neighbors
+		for _, nb := range a[v] {
+			if L.Bit(nb) == 1 {
+				continue
+			}
+			dn := d[nb]  // old number of neighbors of nb
+			Ddn := D[dn] // nb is in this list
+			// remove it from the list
+			for wx, w := range Ddn {
+				if w == nb {
+					last := len(Ddn) - 1
+					Ddn[wx], Ddn[last] = Ddn[last], Ddn[wx]
+					D[dn] = Ddn[:last]
+				}
+			}
+			dn-- // new number of neighbors
+			d[nb] = dn
+			// re--add it to it's new list
+			D[dn] = append(D[dn], nb)
+		}
+	}
+	cores[k] = len(ordering)
+	return
+}
+
+// Degree for undirected graphs, returns the degree of a node.
+//
+// The degree of a node in an undirected graph is the number of incident
+// edges, where loops count twice.
+//
+// If g is known to be loop-free, the result is simply equivalent to len(g[n]).
+// See handshaking lemma example at AdjacencyList.ArcSize.
+//
+// There are equivalent labeled and unlabeled versions of this method.
+func (g Undirected) Degree(n NI) int {
+	to := g.AdjacencyList[n]
+	d := len(to) // just "out" degree,
+	for _, to := range to {
+		if to == n {
+			d++ // except loops count twice
+		}
+	}
+	return d
+}
+
+// FromList constructs a FromList representing the tree reachable from
+// the given root.
+//
+// The connected component containing root should represent a simple graph,
+// connected as a tree.
+//
+// For nodes connected as a tree, the Path member of the returned FromList
+// will be populated with both From and Len values.  The MaxLen member will be
+// set but Leaves will be left a zero value.  Return value cycle will be -1.
+//
+// If the connected component containing root is not connected as a tree,
+// a cycle will be detected.  The returned FromList will be a zero value and
+// return value cycle will be a node involved in the cycle.
+//
+// Loops and parallel edges will be detected as cycles, however only in the
+// connected component containing root.  If g is not fully connected, nodes
+// not reachable from root will have PathEnd values of {From: -1, Len: 0}.
+//
+// There are equivalent labeled and unlabeled versions of this method.
+func (g Undirected) FromList(root NI) (f FromList, cycle NI) {
+	p := make([]PathEnd, len(g.AdjacencyList))
+	for i := range p {
+		p[i].From = -1
+	}
+	ml := 0
+	var df func(NI, NI) bool
+	df = func(fr, n NI) bool {
+		l := p[n].Len + 1
+		for _, to := range g.AdjacencyList[n] {
+			if to == fr {
+				continue
+			}
+			if p[to].Len > 0 {
+				cycle = to
+				return false
+			}
+			p[to] = PathEnd{From: n, Len: l}
+			if l > ml {
+				ml = l
+			}
+			if !df(n, to) {
+				return false
+			}
+		}
+		return true
+	}
+	p[root].Len = 1
+	if !df(-1, root) {
+		return
+	}
+	return FromList{Paths: p, MaxLen: ml}, -1
+}
+
+// IsConnected tests if an undirected graph is a single connected component.
+//
+// There are equivalent labeled and unlabeled versions of this method.
+//
+// See also ConnectedComponentReps for a method returning more information.
+func (g Undirected) IsConnected() bool {
+	a := g.AdjacencyList
+	if len(a) == 0 {
+		return true
+	}
+	var b Bits
+	b.SetAll(len(a))
+	var df func(NI)
+	df = func(n NI) {
+		b.SetBit(n, 0)
+		for _, to := range a[n] {
+			if b.Bit(to) == 1 {
+				df(to)
+			}
+		}
+	}
+	df(0)
+	return b.Zero()
+}
+
+// IsTree identifies trees in undirected graphs.
+//
+// Return value isTree is true if the connected component reachable from root
+// is a tree.  Further, return value allTree is true if the entire graph g is
+// connected.
+//
+// There are equivalent labeled and unlabeled versions of this method.
+func (g Undirected) IsTree(root NI) (isTree, allTree bool) {
+	a := g.AdjacencyList
+	var v Bits
+	v.SetAll(len(a))
+	var df func(NI, NI) bool
+	df = func(fr, n NI) bool {
+		if v.Bit(n) == 0 {
+			return false
+		}
+		v.SetBit(n, 0)
+		for _, to := range a[n] {
+			if to != fr && !df(n, to) {
+				return false
+			}
+		}
+		return true
+	}
+	v.SetBit(root, 0)
+	for _, to := range a[root] {
+		if !df(root, to) {
+			return false, false
+		}
+	}
+	return true, v.Zero()
+}
+
+// Size returns the number of edges in g.
+//
+// See also ArcSize and HasLoop.
+func (g Undirected) Size() int {
+	m2 := 0
+	for fr, to := range g.AdjacencyList {
+		m2 += len(to)
+		for _, to := range to {
+			if to == NI(fr) {
+				m2++
+			}
+		}
+	}
+	return m2 / 2
+}
diff --git a/vendor/github.com/soniakeys/graph/undir_cg.go b/vendor/github.com/soniakeys/graph/undir_cg.go
new file mode 100644
index 0000000..35b5b97
--- /dev/null
+++ b/vendor/github.com/soniakeys/graph/undir_cg.go
@@ -0,0 +1,659 @@
+// Copyright 2014 Sonia Keys
+// License MIT: http://opensource.org/licenses/MIT
+
+package graph
+
+// undir_RO.go is code generated from undir_cg.go by directives in graph.go.
+// Editing undir_cg.go is okay.  It is the code generation source.
+// DO NOT EDIT undir_RO.go.
+// The RO means read only and it is upper case RO to slow you down a bit
+// in case you start to edit the file.
+
+// Bipartite determines if a connected component of an undirected graph
+// is bipartite, a component where nodes can be partitioned into two sets
+// such that every edge in the component goes from one set to the other.
+//
+// Argument n can be any representative node of the component.
+//
+// If the component is bipartite, Bipartite returns true and a two-coloring
+// of the component.  Each color set is returned as a bitmap.  If the component
+// is not bipartite, Bipartite returns false and a representative odd cycle.
+//
+// There are equivalent labeled and unlabeled versions of this method.
+func (g LabeledUndirected) Bipartite(n NI) (b bool, c1, c2 Bits, oc []NI) {
+	b = true
+	var open bool
+	var df func(n NI, c1, c2 *Bits)
+	df = func(n NI, c1, c2 *Bits) {
+		c1.SetBit(n, 1)
+		for _, nb := range g.LabeledAdjacencyList[n] {
+			if c1.Bit(nb.To) == 1 {
+				b = false
+				oc = []NI{nb.To, n}
+				open = true
+				return
+			}
+			if c2.Bit(nb.To) == 1 {
+				continue
+			}
+			df(nb.To, c2, c1)
+			if b {
+				continue
+			}
+			switch {
+			case !open:
+			case n == oc[0]:
+				open = false
+			default:
+				oc = append(oc, n)
+			}
+			return
+		}
+	}
+	df(n, &c1, &c2)
+	if b {
+		return b, c1, c2, nil
+	}
+	return b, Bits{}, Bits{}, oc
+}
+
+// BronKerbosch1 finds maximal cliques in an undirected graph.
+//
+// The graph must not contain parallel edges or loops.
+//
+// See https://en.wikipedia.org/wiki/Clique_(graph_theory) and
+// https://en.wikipedia.org/wiki/Bron%E2%80%93Kerbosch_algorithm for background.
+//
+// This method implements the BronKerbosch1 algorithm of WP; that is,
+// the original algorithm without improvements.
+//
+// The method calls the emit argument for each maximal clique in g, as long
+// as emit returns true.  If emit returns false, BronKerbosch1 returns
+// immediately.
+//
+// There are equivalent labeled and unlabeled versions of this method.
+//
+// See also more sophisticated variants BronKerbosch2 and BronKerbosch3.
+func (g LabeledUndirected) BronKerbosch1(emit func([]NI) bool) {
+	a := g.LabeledAdjacencyList
+	var f func(R, P, X *Bits) bool
+	f = func(R, P, X *Bits) bool {
+		switch {
+		case !P.Zero():
+			var r2, p2, x2 Bits
+			pf := func(n NI) bool {
+				r2.Set(*R)
+				r2.SetBit(n, 1)
+				p2.Clear()
+				x2.Clear()
+				for _, to := range a[n] {
+					if P.Bit(to.To) == 1 {
+						p2.SetBit(to.To, 1)
+					}
+					if X.Bit(to.To) == 1 {
+						x2.SetBit(to.To, 1)
+					}
+				}
+				if !f(&r2, &p2, &x2) {
+					return false
+				}
+				P.SetBit(n, 0)
+				X.SetBit(n, 1)
+				return true
+			}
+			if !P.Iterate(pf) {
+				return false
+			}
+		case X.Zero():
+			return emit(R.Slice())
+		}
+		return true
+	}
+	var R, P, X Bits
+	P.SetAll(len(a))
+	f(&R, &P, &X)
+}
+
+// BKPivotMaxDegree is a strategy for BronKerbosch methods.
+//
+// To use it, take the method value (see golang.org/ref/spec#Method_values)
+// and pass it as the argument to BronKerbosch2 or 3.
+//
+// The strategy is to pick the node from P or X with the maximum degree
+// (number of edges) in g.  Note this is a shortcut from evaluating degrees
+// in P.
+//
+// There are equivalent labeled and unlabeled versions of this method.
+func (g LabeledUndirected) BKPivotMaxDegree(P, X *Bits) (p NI) {
+	// choose pivot u as highest degree node from P or X
+	a := g.LabeledAdjacencyList
+	maxDeg := -1
+	P.Iterate(func(n NI) bool { // scan P
+		if d := len(a[n]); d > maxDeg {
+			p = n
+			maxDeg = d
+		}
+		return true
+	})
+	X.Iterate(func(n NI) bool { // scan X
+		if d := len(a[n]); d > maxDeg {
+			p = n
+			maxDeg = d
+		}
+		return true
+	})
+	return
+}
+
+// BKPivotMinP is a strategy for BronKerbosch methods.
+//
+// To use it, take the method value (see golang.org/ref/spec#Method_values)
+// and pass it as the argument to BronKerbosch2 or 3.
+//
+// The strategy is to simply pick the first node in P.
+//
+// There are equivalent labeled and unlabeled versions of this method.
+func (g LabeledUndirected) BKPivotMinP(P, X *Bits) NI {
+	return P.From(0)
+}
+
+// BronKerbosch2 finds maximal cliques in an undirected graph.
+//
+// The graph must not contain parallel edges or loops.
+//
+// See https://en.wikipedia.org/wiki/Clique_(graph_theory) and
+// https://en.wikipedia.org/wiki/Bron%E2%80%93Kerbosch_algorithm for background.
+//
+// This method implements the BronKerbosch2 algorithm of WP; that is,
+// the original algorithm plus pivoting.
+//
+// The argument is a pivot function that must return a node of P or X.
+// P is guaranteed to contain at least one node.  X is not.
+// For example see BKPivotMaxDegree.
+//
+// The method calls the emit argument for each maximal clique in g, as long
+// as emit returns true.  If emit returns false, BronKerbosch1 returns
+// immediately.
+//
+// There are equivalent labeled and unlabeled versions of this method.
+//
+// See also simpler variant BronKerbosch1 and more sophisticated variant
+// BronKerbosch3.
+func (g LabeledUndirected) BronKerbosch2(pivot func(P, X *Bits) NI, emit func([]NI) bool) {
+	a := g.LabeledAdjacencyList
+	var f func(R, P, X *Bits) bool
+	f = func(R, P, X *Bits) bool {
+		switch {
+		case !P.Zero():
+			var r2, p2, x2, pnu Bits
+			// compute P \ N(u).  next 5 lines are only difference from BK1
+			pnu.Set(*P)
+			for _, to := range a[pivot(P, X)] {
+				pnu.SetBit(to.To, 0)
+			}
+			// remaining code like BK1
+			pf := func(n NI) bool {
+				r2.Set(*R)
+				r2.SetBit(n, 1)
+				p2.Clear()
+				x2.Clear()
+				for _, to := range a[n] {
+					if P.Bit(to.To) == 1 {
+						p2.SetBit(to.To, 1)
+					}
+					if X.Bit(to.To) == 1 {
+						x2.SetBit(to.To, 1)
+					}
+				}
+				if !f(&r2, &p2, &x2) {
+					return false
+				}
+				P.SetBit(n, 0)
+				X.SetBit(n, 1)
+				return true
+			}
+			if !pnu.Iterate(pf) {
+				return false
+			}
+		case X.Zero():
+			return emit(R.Slice())
+		}
+		return true
+	}
+	var R, P, X Bits
+	P.SetAll(len(a))
+	f(&R, &P, &X)
+}
+
+// BronKerbosch3 finds maximal cliques in an undirected graph.
+//
+// The graph must not contain parallel edges or loops.
+//
+// See https://en.wikipedia.org/wiki/Clique_(graph_theory) and
+// https://en.wikipedia.org/wiki/Bron%E2%80%93Kerbosch_algorithm for background.
+//
+// This method implements the BronKerbosch3 algorithm of WP; that is,
+// the original algorithm with pivoting and degeneracy ordering.
+//
+// The argument is a pivot function that must return a node of P or X.
+// P is guaranteed to contain at least one node.  X is not.
+// For example see BKPivotMaxDegree.
+//
+// The method calls the emit argument for each maximal clique in g, as long
+// as emit returns true.  If emit returns false, BronKerbosch1 returns
+// immediately.
+//
+// There are equivalent labeled and unlabeled versions of this method.
+//
+// See also simpler variants BronKerbosch1 and BronKerbosch2.
+func (g LabeledUndirected) BronKerbosch3(pivot func(P, X *Bits) NI, emit func([]NI) bool) {
+	a := g.LabeledAdjacencyList
+	var f func(R, P, X *Bits) bool
+	f = func(R, P, X *Bits) bool {
+		switch {
+		case !P.Zero():
+			var r2, p2, x2, pnu Bits
+			// compute P \ N(u).  next lines are only difference from BK1
+			pnu.Set(*P)
+			for _, to := range a[pivot(P, X)] {
+				pnu.SetBit(to.To, 0)
+			}
+			// remaining code like BK2
+			pf := func(n NI) bool {
+				r2.Set(*R)
+				r2.SetBit(n, 1)
+				p2.Clear()
+				x2.Clear()
+				for _, to := range a[n] {
+					if P.Bit(to.To) == 1 {
+						p2.SetBit(to.To, 1)
+					}
+					if X.Bit(to.To) == 1 {
+						x2.SetBit(to.To, 1)
+					}
+				}
+				if !f(&r2, &p2, &x2) {
+					return false
+				}
+				P.SetBit(n, 0)
+				X.SetBit(n, 1)
+				return true
+			}
+			if !pnu.Iterate(pf) {
+				return false
+			}
+		case X.Zero():
+			return emit(R.Slice())
+		}
+		return true
+	}
+	var R, P, X Bits
+	P.SetAll(len(a))
+	// code above same as BK2
+	// code below new to BK3
+	_, ord, _ := g.Degeneracy()
+	var p2, x2 Bits
+	for _, n := range ord {
+		R.SetBit(n, 1)
+		p2.Clear()
+		x2.Clear()
+		for _, to := range a[n] {
+			if P.Bit(to.To) == 1 {
+				p2.SetBit(to.To, 1)
+			}
+			if X.Bit(to.To) == 1 {
+				x2.SetBit(to.To, 1)
+			}
+		}
+		if !f(&R, &p2, &x2) {
+			return
+		}
+		R.SetBit(n, 0)
+		P.SetBit(n, 0)
+		X.SetBit(n, 1)
+	}
+}
+
+// ConnectedComponentBits returns a function that iterates over connected
+// components of g, returning a member bitmap for each.
+//
+// Each call of the returned function returns the order (number of nodes)
+// and bits of a connected component.  The returned function returns zeros
+// after returning all connected components.
+//
+// There are equivalent labeled and unlabeled versions of this method.
+//
+// See also ConnectedComponentReps, which has lighter weight return values.
+func (g LabeledUndirected) ConnectedComponentBits() func() (order int, bits Bits) {
+	a := g.LabeledAdjacencyList
+	var vg Bits  // nodes visited in graph
+	var vc *Bits // nodes visited in current component
+	var nc int
+	var df func(NI)
+	df = func(n NI) {
+		vg.SetBit(n, 1)
+		vc.SetBit(n, 1)
+		nc++
+		for _, nb := range a[n] {
+			if vg.Bit(nb.To) == 0 {
+				df(nb.To)
+			}
+		}
+		return
+	}
+	var n NI
+	return func() (o int, bits Bits) {
+		for ; n < NI(len(a)); n++ {
+			if vg.Bit(n) == 0 {
+				vc = &bits
+				nc = 0
+				df(n)
+				return nc, bits
+			}
+		}
+		return
+	}
+}
+
+// ConnectedComponentLists returns a function that iterates over connected
+// components of g, returning the member list of each.
+//
+// Each call of the returned function returns a node list of a connected
+// component.  The returned function returns nil after returning all connected
+// components.
+//
+// There are equivalent labeled and unlabeled versions of this method.
+//
+// See also ConnectedComponentReps, which has lighter weight return values.
+func (g LabeledUndirected) ConnectedComponentLists() func() []NI {
+	a := g.LabeledAdjacencyList
+	var vg Bits // nodes visited in graph
+	var m []NI  // members of current component
+	var df func(NI)
+	df = func(n NI) {
+		vg.SetBit(n, 1)
+		m = append(m, n)
+		for _, nb := range a[n] {
+			if vg.Bit(nb.To) == 0 {
+				df(nb.To)
+			}
+		}
+		return
+	}
+	var n NI
+	return func() []NI {
+		for ; n < NI(len(a)); n++ {
+			if vg.Bit(n) == 0 {
+				m = nil
+				df(n)
+				return m
+			}
+		}
+		return nil
+	}
+}
+
+// ConnectedComponentReps returns a representative node from each connected
+// component of g.
+//
+// Returned is a slice with a single representative node from each connected
+// component and also a parallel slice with the order, or number of nodes,
+// in the corresponding component.
+//
+// This is fairly minimal information describing connected components.
+// From a representative node, other nodes in the component can be reached
+// by depth first traversal for example.
+//
+// There are equivalent labeled and unlabeled versions of this method.
+//
+// See also ConnectedComponentBits and ConnectedComponentLists which can
+// collect component members in a single traversal, and IsConnected which
+// is an even simpler boolean test.
+func (g LabeledUndirected) ConnectedComponentReps() (reps []NI, orders []int) {
+	a := g.LabeledAdjacencyList
+	var c Bits
+	var o int
+	var df func(NI)
+	df = func(n NI) {
+		c.SetBit(n, 1)
+		o++
+		for _, nb := range a[n] {
+			if c.Bit(nb.To) == 0 {
+				df(nb.To)
+			}
+		}
+		return
+	}
+	for n := range a {
+		if c.Bit(NI(n)) == 0 {
+			reps = append(reps, NI(n))
+			o = 0
+			df(NI(n))
+			orders = append(orders, o)
+		}
+	}
+	return
+}
+
+// Copy makes a deep copy of g.
+// Copy also computes the arc size ma, the number of arcs.
+//
+// There are equivalent labeled and unlabeled versions of this method.
+func (g LabeledUndirected) Copy() (c LabeledUndirected, ma int) {
+	l, s := g.LabeledAdjacencyList.Copy()
+	return LabeledUndirected{l}, s
+}
+
+// Degeneracy computes k-degeneracy, vertex ordering and k-cores.
+//
+// See Wikipedia https://en.wikipedia.org/wiki/Degeneracy_(graph_theory)
+//
+// There are equivalent labeled and unlabeled versions of this method.
+func (g LabeledUndirected) Degeneracy() (k int, ordering []NI, cores []int) {
+	a := g.LabeledAdjacencyList
+	// WP algorithm
+	ordering = make([]NI, len(a))
+	var L Bits
+	d := make([]int, len(a))
+	var D [][]NI
+	for v, nb := range a {
+		dv := len(nb)
+		d[v] = dv
+		for len(D) <= dv {
+			D = append(D, nil)
+		}
+		D[dv] = append(D[dv], NI(v))
+	}
+	for ox := range a {
+		// find a non-empty D
+		i := 0
+		for len(D[i]) == 0 {
+			i++
+		}
+		// k is max(i, k)
+		if i > k {
+			for len(cores) <= i {
+				cores = append(cores, 0)
+			}
+			cores[k] = ox
+			k = i
+		}
+		// select from D[i]
+		Di := D[i]
+		last := len(Di) - 1
+		v := Di[last]
+		// Add v to ordering, remove from Di
+		ordering[ox] = v
+		L.SetBit(v, 1)
+		D[i] = Di[:last]
+		// move neighbors
+		for _, nb := range a[v] {
+			if L.Bit(nb.To) == 1 {
+				continue
+			}
+			dn := d[nb.To] // old number of neighbors of nb
+			Ddn := D[dn]   // nb is in this list
+			// remove it from the list
+			for wx, w := range Ddn {
+				if w == nb.To {
+					last := len(Ddn) - 1
+					Ddn[wx], Ddn[last] = Ddn[last], Ddn[wx]
+					D[dn] = Ddn[:last]
+				}
+			}
+			dn-- // new number of neighbors
+			d[nb.To] = dn
+			// re--add it to it's new list
+			D[dn] = append(D[dn], nb.To)
+		}
+	}
+	cores[k] = len(ordering)
+	return
+}
+
+// Degree for undirected graphs, returns the degree of a node.
+//
+// The degree of a node in an undirected graph is the number of incident
+// edges, where loops count twice.
+//
+// If g is known to be loop-free, the result is simply equivalent to len(g[n]).
+// See handshaking lemma example at AdjacencyList.ArcSize.
+//
+// There are equivalent labeled and unlabeled versions of this method.
+func (g LabeledUndirected) Degree(n NI) int {
+	to := g.LabeledAdjacencyList[n]
+	d := len(to) // just "out" degree,
+	for _, to := range to {
+		if to.To == n {
+			d++ // except loops count twice
+		}
+	}
+	return d
+}
+
+// FromList constructs a FromList representing the tree reachable from
+// the given root.
+//
+// The connected component containing root should represent a simple graph,
+// connected as a tree.
+//
+// For nodes connected as a tree, the Path member of the returned FromList
+// will be populated with both From and Len values.  The MaxLen member will be
+// set but Leaves will be left a zero value.  Return value cycle will be -1.
+//
+// If the connected component containing root is not connected as a tree,
+// a cycle will be detected.  The returned FromList will be a zero value and
+// return value cycle will be a node involved in the cycle.
+//
+// Loops and parallel edges will be detected as cycles, however only in the
+// connected component containing root.  If g is not fully connected, nodes
+// not reachable from root will have PathEnd values of {From: -1, Len: 0}.
+//
+// There are equivalent labeled and unlabeled versions of this method.
+func (g LabeledUndirected) FromList(root NI) (f FromList, cycle NI) {
+	p := make([]PathEnd, len(g.LabeledAdjacencyList))
+	for i := range p {
+		p[i].From = -1
+	}
+	ml := 0
+	var df func(NI, NI) bool
+	df = func(fr, n NI) bool {
+		l := p[n].Len + 1
+		for _, to := range g.LabeledAdjacencyList[n] {
+			if to.To == fr {
+				continue
+			}
+			if p[to.To].Len > 0 {
+				cycle = to.To
+				return false
+			}
+			p[to.To] = PathEnd{From: n, Len: l}
+			if l > ml {
+				ml = l
+			}
+			if !df(n, to.To) {
+				return false
+			}
+		}
+		return true
+	}
+	p[root].Len = 1
+	if !df(-1, root) {
+		return
+	}
+	return FromList{Paths: p, MaxLen: ml}, -1
+}
+
+// IsConnected tests if an undirected graph is a single connected component.
+//
+// There are equivalent labeled and unlabeled versions of this method.
+//
+// See also ConnectedComponentReps for a method returning more information.
+func (g LabeledUndirected) IsConnected() bool {
+	a := g.LabeledAdjacencyList
+	if len(a) == 0 {
+		return true
+	}
+	var b Bits
+	b.SetAll(len(a))
+	var df func(NI)
+	df = func(n NI) {
+		b.SetBit(n, 0)
+		for _, to := range a[n] {
+			if b.Bit(to.To) == 1 {
+				df(to.To)
+			}
+		}
+	}
+	df(0)
+	return b.Zero()
+}
+
+// IsTree identifies trees in undirected graphs.
+//
+// Return value isTree is true if the connected component reachable from root
+// is a tree.  Further, return value allTree is true if the entire graph g is
+// connected.
+//
+// There are equivalent labeled and unlabeled versions of this method.
+func (g LabeledUndirected) IsTree(root NI) (isTree, allTree bool) {
+	a := g.LabeledAdjacencyList
+	var v Bits
+	v.SetAll(len(a))
+	var df func(NI, NI) bool
+	df = func(fr, n NI) bool {
+		if v.Bit(n) == 0 {
+			return false
+		}
+		v.SetBit(n, 0)
+		for _, to := range a[n] {
+			if to.To != fr && !df(n, to.To) {
+				return false
+			}
+		}
+		return true
+	}
+	v.SetBit(root, 0)
+	for _, to := range a[root] {
+		if !df(root, to.To) {
+			return false, false
+		}
+	}
+	return true, v.Zero()
+}
+
+// Size returns the number of edges in g.
+//
+// See also ArcSize and HasLoop.
+func (g LabeledUndirected) Size() int {
+	m2 := 0
+	for fr, to := range g.LabeledAdjacencyList {
+		m2 += len(to)
+		for _, to := range to {
+			if to.To == NI(fr) {
+				m2++
+			}
+		}
+	}
+	return m2 / 2
+}