I've started to write a Generalized Suffix Tree implementation. The overall code is still in an experimental state, though I think it's mature enough to ask for a little review.

I'd like to know how I can make it better, especially on the use of the standard library, the consistency of the code and the design.

A quick example can be found running Live on Coliru.

#include <iostream>
#include <unordered_map>
#include <list>
#include <utility>
#include <limits>

template <typename S, typename C=int>
class SuffixTree {
    // Forward declarations of inner classes
    struct Node;

    typedef S string;
    typedef C character;
    typedef std::pair<Node*, std::pair<int, int>> ReferencePoint;


    // Node class:
    // Contains the suffix link to another Node
    // The Transitions "g(s,(k,i)) = s'" to children nodes
    // Note:
    // Transitions are stored in a hashtable indexed by the first substring
    // character. A given character can only have at most one Transition in a
    // node.

    // A Generalized Suffix Tree can contain more than one string at a time
    // Each string is labeled with an int. Thus each substring is related to
    // an appropriate reference string:
    // (ref string id, left ptr, right ptr)
    struct MappedSubstring {
        int ref_str;
        // A substring is a pair of integer (left ptr, right ptr)
        // To denote an empty substring, set right ptr < left ptr.
        int l;
        int r;
        MappedSubstring() : ref_str(0), l(0), r(0) {}
        MappedSubstring(int ref, int left, int right) :
        bool empty() {
            return (this->l > this->r);

    struct Transition {
        MappedSubstring sub;
        Node *tgt;
        Transition() : sub(), tgt(nullptr) {}
        Transition(MappedSubstring s, Node *t) : sub(s), tgt(t) {}

    struct Node {
        std::unordered_map<C, Transition> g;
        Node *suffix_link;
        virtual Transition find_alpha_transition(C alpha) {
            auto it = g.find(alpha);
            if (g.end() == it) {
                return Transition(MappedSubstring(0, 0, -1), nullptr);
            return it->second;

        Node() : suffix_link(nullptr) {}
        virtual ~Node() {}

    // Simple workaround for the specific sink node
    // This node must have a transition for every char of the input alphabet.
    // Instead of creating such transitions, we just make them up through
    // an override of `find_alpha_transition`
    struct SinkNode : public Node {
        virtual Transition find_alpha_transition(C alpha) override {
            return Transition(MappedSubstring(0, 0, 0), this->suffix_link);

    // Leaf nodes:
    // Leaves must contain an explicit reference to the suffix they represent
    // Some strings might have common suffixes, hence the map.
    // The suffix link **remains** UNIQUE nonetheless.
    struct Leaf : public Node {
        // TODO

    // Base - A tree nested base class
    // This clase is here to hide implementation details
    // And to handle destruction properly.
    // The processing (insertion, deletion of strings) is done by SuffixTree,
    // Base handles the cleanup.
    struct Base {
        SinkNode sink;
        Node root;
        Base() {
            root.suffix_link = &sink;
            sink.suffix_link = &root;
        virtual ~Base() {
        void clean() {
            std::list<Node*> del_list {&root};
            while (!del_list.empty()) {
                Node *current = del_list.front();
                for (auto it : current->g) {
                if (&root != current) {
                    delete current;


    Base tree;
    C end_token;

    std::unordered_map<int, S> haystack;
    std::unordered_map<int, Node*> borderpath_map;
    int last_index;

    // Given a Node n, a substring kp and a character t,
    // test_and_split must return if (n, kp) is the end point.
    // If not, and we are in an implicit state, a new state is created.
    bool test_and_split(Node *n, MappedSubstring kp, C t, const S& w, Node **r) {
        C tk = w[kp.l];
        int delta = kp.r - kp.l;
        if (0 <= delta) {
            Transition tk_trans = n->find_alpha_transition(tk);
            MappedSubstring kp_prime = tk_trans.sub;
            auto str_prime = haystack.find(kp_prime.ref_str);
            if (str_prime->second[kp_prime.l + delta + 1] == t) {
                *r = n;
                return true;
            *r = new Node();
            Transition new_t = tk_trans;
            new_t.sub.l += delta+1;
            (*r)->g.insert(std::pair<C, Transition>(
                str_prime->second[new_t.sub.l], new_t));
            tk_trans.sub.r = tk_trans.sub.l + delta;
            tk_trans.tgt = *r;
            n->g[tk] = tk_trans;
            return false;

        } else {
            // kp represents an empty substring
            Transition t_Transition = n->find_alpha_transition(t);
            *r = n;
            return (t_Transition.tgt != nullptr);

    // update performs the heart of an iteration:
    // It walks the border path from the active point to the end point
    // and adds the required Transitions brought by the insertion of
    // the string's i-th character.
    // It returns the end point.
    ReferencePoint update(Node *s, MappedSubstring ki) {
        Node *oldr = &tree.root;
        Node *r = nullptr;
        bool is_endpoint = false;
        MappedSubstring ki1 = ki;
        auto ref_str_it = haystack.find(ki.ref_str);
        S w = ref_str_it->second;
        ReferencePoint sk(s, std::pair<int,int>(ki.ref_str, ki.l));
        ki1.r = ki.r-1;
        is_endpoint = test_and_split(s, ki1, w[ki.r], w, &r);
        while (!is_endpoint) {
            Leaf *r_prime = new Leaf();
              w[ki.r], Transition(MappedSubstring(
              ki.ref_str, ki.r, std::numeric_limits<int>::max()), r_prime)));
            if (&tree.root != oldr) {
                oldr->suffix_link = r;
            oldr = r;
            sk = canonize(sk.first->suffix_link, ki1);
            ki1.l = ki.l = sk.second.second;
            is_endpoint = test_and_split(sk.first, ki1, w[ki.r], w, &r); 
        if (&tree.root != oldr) {
            oldr->suffix_link = sk.first;
        return sk;

    // canonize - Get canonical pair
    // Given a Node and a substring,
    // This returns the canonical pair for this particular combination
    ReferencePoint canonize(Node *s, MappedSubstring kp) {
        if (kp.r < kp.l)
            return ReferencePoint(s, std::pair<int,int>(kp.ref_str, kp.l));
        auto kp_ref_str = haystack.find(kp.ref_str);
        int delta;
        Transition tk_trans = s->find_alpha_transition(kp_ref_str->second[kp.l]);
        while ((delta = tk_trans.sub.r - tk_trans.sub.l) <= kp.r - kp.l) {
            kp.l += 1 + delta;
            s = tk_trans.tgt;
            if (kp.l <= kp.r)
                tk_trans = s->find_alpha_transition(kp_ref_str->second[kp.l]);
        return ReferencePoint(s, std::pair<int,int>(kp.ref_str, kp.l));

    // get_starting_node - Find the starting node
    // @s[in]: The string to insert
    // @r[in/out]: The walk starting/ending point
    // get_starting_node walks down the tree until s does not match anymore
    // character.
    // @r is updated through this process and contains the reference pair of the
    // diverging point between @s and the tree.
    // The result '(s,k)' of this function may then be used to resume the Ukkonen's
    // algorithm.
    int get_starting_node(const S& s, ReferencePoint *r) {
        int k = r->second.second;
        int s_len = s.length();
        bool s_runout = false;
        while (!s_runout) {
            Transition t = r->first->find_alpha_transition(s[k]);
            if (nullptr != t.tgt) {
                int i;
                auto ref_str = haystack.find(t.sub.ref_str);
                for (i=1; (i <= t.sub.r - t.sub.l); ++i) {
                    if (k+i >= s_len) {
                        s_runout = true;
                    if (s[k+i] != ref_str->second[t.sub.l+i]) {
                        r->second.second = k;
                        return k+i;
                if (!s_runout) {
                    r->first = t.tgt;
                    k += i;
            } else {
                return k;
        r->second.second = std::numeric_limits<int>::max();
        return std::numeric_limits<int>::max();

    // deploy_suffixes - Deploy suffixes
    // @s[in]: The string to insert in the tree
    // @sindex[in]: The index id of @s
    // deploy_suffixes performs the Ukkonen's algorithm to inser @s into the
    // tree.
    int deploy_suffixes(const S& s, int sindex) {
        ReferencePoint active_point(&tree.root, std::pair<int,int>(sindex, 0));
        int i = get_starting_node(s, &active_point);
        if (std::numeric_limits<int>::max() == i) {
            return -1;
        for (; i < s.length(); ++i) {
            MappedSubstring ki(sindex, active_point.second.second, i);
            active_point = update(active_point.first, ki);
            ki.l = active_point.second.second;
            active_point = canonize(active_point.first, ki);
        return sindex;

    void dump_node(Node *n, bool same_line, int padding, MappedSubstring orig) {
        int delta = 0;
        if (!same_line) {
            for (int i = 0; i < padding; ++i) {
                std::cout << " ";
        if (!orig.empty()) {
            auto s = haystack.find(orig.ref_str);
            for (int i = orig.l; i <= orig.r && i <= s->second.length(); ++i) {
                std::cout << s->second[i];
            std::cout << "-";
            delta = orig.r - orig.l + 2;
            if (orig.r == std::numeric_limits<int>::max()) {
                delta = s->second.length() - orig.l + 2;
        same_line = true;
        for (auto t_it : n->g) {
            dump_node(t_it.second.tgt, same_line, padding + delta, t_it.second.sub);
            same_line = false;
        if (same_line) {
            std::cout << "##" << std::endl;
    SuffixTree() : end_token('$'), last_index(0) {
    int add_string(const S new_string) {
        haystack.insert(std::pair<int, S>(last_index, new_string));
        if (0 > deploy_suffixes(new_string, last_index)) {
            return -1;
        return last_index;

    virtual ~SuffixTree() {

    void dump_tree() {
        dump_node(&tree.root, true, 0, MappedSubstring(0,0,-1));
  • \$\begingroup\$ Welcome to Code Review! We're glad you found your way here, I hope you get some good reviews! \$\endgroup\$ – Phrancis Feb 13 '15 at 16:13

You've forgotten to implement the TODO comment in struct Leaf. This may be why your code segfaults with optimizations turned on; I don't know. You might try compiling with clang++ -std=c++14 -fsanitize=address (on a Linux box) to see if that helps track down the misbehavior.

Since the code segfaults, I don't feel guilty about giving merely a superficial review. Here are some style comments, though:

Never ever pass or return objects by const value.

int add_string(const S new_string)

should be either

int add_string(const S& new_string)


int add_string(S new_string)

Your code triggers at least two compiler warnings about "signed versus unsigned comparison." I know you've seen the warnings, because you posted the link to Coliru. Why didn't you fix those warnings as soon as the compiler pointed them out?

(The idiomatic way to fix them would be to go through your code and make sure you're using size_t for any variable that's intended to hold an index or byte-count. The lazy-practical way would be to add casts.)

// Base - A tree nested base class

You say that, but then when I look at the code I find that nothing inherits from Base. So it's not a base class.

Also, you should get rid of all the virtual destructors except those that are actually necessary. Classes that aren't used polymorphically shouldn't have any polymorphic methods. (This applies to Base and SuffixTree, obviously; but it also applies to Node and SinkNode as well, because you never use SinkNode polymorphically.)

In general, the code is just way too complicated for what it does. You shouldn't need so many different classes, or probably any virtual methods at all.

On the theme of simplification, your definition of ReferencePoint as std::pair<X, std::pair<Y,Z>> would be better expressed as std::tuple<X,Y,Z>.

MappedSubstring::empty() should be a const member function.

  • \$\begingroup\$ That's not what I would call a merely superficial review! Thanks. I applied the advice you gave me and tried the given example with optimizations (-o3) but did not get a segfault (with g++). I'll look on your example \$\endgroup\$ – Rerito Mar 3 '15 at 9:57
  • \$\begingroup\$ That's a bit WTFesque here, I compiled the code using clang++(3.4) and the same options you provided (except I used -std=c++11) and still no segfault?! \$\endgroup\$ – Rerito Mar 3 '15 at 10:21

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.