spot/spot/tl/relabel.cc

// -*- coding: utf-8 -*-
// Copyright (C) by the Spot authors, see the AUTHORS file for details.
//
// This file is part of Spot, a model checking library.
//
// Spot is free software; you can redistribute it and/or modify it
// under the terms of the GNU General Public License as published by
// the Free Software Foundation; either version 3 of the License, or
// (at your option) any later version.
//
// Spot is distributed in the hope that it will be useful, but WITHOUT
// ANY WARRANTY; without even the implied warranty of MERCHANTABILITY
// or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public
// License for more details.
//
// You should have received a copy of the GNU General Public License
// along with this program.  If not, see <http://www.gnu.org/licenses/>.

#include "config.h"
#include <spot/tl/relabel.hh>
#include <spot/tl/simplify.hh>
#include <sstream>
#include <spot/misc/hash.hh>
#include <map>
#include <set>
#include <stack>
#include <iostream>

namespace spot
{
  //////////////////////////////////////////////////////////////////////
  // Basic relabeler
  //////////////////////////////////////////////////////////////////////

  namespace
  {
    struct ap_generator
    {
      virtual formula next() = 0;
      virtual ~ap_generator() {}
    };

    struct pnn_generator final: ap_generator
    {
      unsigned nn;
      pnn_generator()
        : nn(0)
        {
        }

      virtual formula next() override
      {
        std::ostringstream s;
        s << 'p' << nn++;
        return formula::ap(s.str());
      }
    };

    struct abc_generator final: ap_generator
    {
    public:
      abc_generator()
        : nn(0)
        {
        }

      unsigned nn;

      virtual formula next() override
      {
        std::string s;
        unsigned n = nn++;
        do
          {
            s.push_back('a' + (n % 26));
            n /= 26;
          }
        while (n);
        return formula::ap(s);
      }
    };

    // if subexp == false, matches APs
    // if subexp == true, matches boolean subexps
    template <bool subexp, bool use_bdd = false>
    class relabeler
    {
    public:
      typedef std::unordered_map<formula, formula> map;
      map newname;
      ap_generator* gen;
      relabeling_map* oldnames;
      tl_simplifier tl;

      relabeler(ap_generator* gen, relabeling_map* m)
        : gen(gen), oldnames(m)
      {
      }

      ~relabeler()
      {
        delete gen;
      }

      formula rename(formula old)
      {
        if constexpr (subexp)
          {
            // have we given a name to the negation of this formula?
            auto neg = newname.find(formula::Not(old));
            if (neg != newname.end())
              return formula::Not(neg->second);
          }

        auto r = newname.emplace(old, nullptr);
        if (!r.second)
          return r.first->second;

        if constexpr (use_bdd)
          if (!old.is(op::ap))
            {
              bdd b = tl.as_bdd(old);
              if (b == bddtrue)
                return r.first->second = formula::tt();
              if (b == bddfalse)
                return r.first->second = formula::ff();
            }

        formula res = gen->next();
        r.first->second = res;
        if (oldnames)
          (*oldnames)[res] = old;
        return res;
      }

      formula
      visit(formula f)
      {
        if ((!subexp && f.is(op::ap))
            || (subexp && f.is_boolean()))
          {
            return rename(f);
          }
        if (subexp && f.is(op::Or, op::And) && f[0].is_boolean())
          {
            // Boolean terms are always beginning of And and Or, so
            // the above test capture Or/And that some Boolean arguments
            // and some non-Boolean arguments.
            unsigned i = 0;
            formula b = f.boolean_operands(&i);
            unsigned sz = f.size();
            std::vector<formula> res;
            res.reserve(sz - i + 1);
            res.emplace_back(visit(b));
            for (; i < sz; ++i)
              res.emplace_back(visit(f[i]));
            return formula::multop(f.kind(), res);
          }
        else
          {
            return f.map([this](formula f)
            {
              return this->visit(f);
            });
          }
      }

    };

    template<bool subexp>
    formula
    relabel_do(formula f, relabeling_style style, relabeling_map* m)
    {
      ap_generator* gen = nullptr;
      switch (style)
        {
        case Pnn:
          gen = new pnn_generator;
          break;
        case Abc:
          gen = new abc_generator;
          break;
        }

      relabeler<subexp> r(gen, m);
      return r.visit(f);
    }
  }

  formula
  relabel(formula f, relabeling_style style, relabeling_map* m)
  {
    return relabel_do<false>(f, style, m);
  }

  formula
  relabel_overlapping_bse(formula f, relabeling_style style, relabeling_map* m)
  {
    return relabel_do<true>(f, style, m);
  }

  namespace
  {
    typedef std::map<formula, int> sub_formula_count_t;

    static void
    sub_formula_collect(formula f, sub_formula_count_t* s)
    {
      assert(s);
      f.traverse([&](const formula& f)
                 {
                   auto p = s->emplace(f, 1);
                   if (p.second)
                     return false;
                   p.first->second += 1;
                   return true;
                 });
    }

    static std::pair<formula, formula>
    split_used_once(formula f, const sub_formula_count_t& subcount)
    {
      assert(f.is_boolean());
      unsigned sz = f.size();
      if (sz <= 2)
        return {f, nullptr};
      // If we have a Boolean formula with more than two
      // children, like (a & b & c & d) where some children
      // (assume {a,b}) are used only once, but some other
      // (assume {c,d}) are used multiple time in the formula,
      // then split that into ((a & b) & (c & d)) to give
      // (a & b) a chance to be relabeled as a whole.
      bool has_once = false;
      bool has_mult = false;
      for (unsigned j = 0; j < sz; ++j)
        {
          auto p = subcount.find(f[j]);
          assert(p != subcount.end());
          unsigned sc = p->second;
          assert(sc > 0);
          if (sc == 1)
            has_once = true;
          else
            has_mult = true;
          if (has_once && has_mult)
            {
              std::vector<formula> once;
              std::vector<formula> mult;
              for (unsigned i = 0; i < j; ++i)
                mult.push_back(f[i]);
              once.push_back(f[j]);
              if (sc > 1)
                std::swap(once, mult);
              for (++j; j < sz; ++j)
                {
                  auto p = subcount.find(f[j]);
                  assert(p != subcount.end());
                  unsigned sc = p->second;
                  ((sc == 1) ? once : mult).push_back(f[j]);
                }
              formula f1 = formula::multop(f.kind(), std::move(once));
              formula f2 = formula::multop(f.kind(), std::move(mult));
              return { f1, f2 };
            }
        }
      return {f, nullptr};
    }
  }


  //////////////////////////////////////////////////////////////////////
  // Boolean-subexpression relabeler
  //////////////////////////////////////////////////////////////////////

  // Here we want to rewrite a formula such as
  //   "a & b & X(c & d) & GF(c & d)" into "p0 & Xp1 & GFp1"
  // where Boolean subexpressions are replaced by fresh propositions.
  //
  // Detecting Boolean subexpressions is not a problem.
  // Furthermore, because we are already representing LTL formulas
  // with sharing of identical sub-expressions we can easily rename
  // a subexpression (such as c&d above) only once.  However this
  // scheme (done by relabel_overlapping_bse()) has two problems:
  //
  //   A. It will not detect inter-dependent Boolean subexpressions.
  //      For instance it will mistakenly relabel "(a & b) U (a & !b)"
  //      as "p0 U p1", hiding the dependency between a&b and a&!b.
  //
  //   B. Because of our n-ary operators, it will fail to
  //      notice that (a & b) is a sub-expression of (a & b & c).
  //
  // The way we compute the subexpressions that can be relabeled is by
  // transforming the formula syntax tree into an undirected graph,
  // and computing the cut-points of this graph.  The cut-points (or
  // articulation points) are the nodes whose removal would split the
  // graph in two components; in our case, we extend this definition to
  // also consider the leaves as cut-points.
  //
  // For instance ((a|b)&c&d)U(!d&e&f) is represented by
  // the following graph, were cut-points are marked with *.
  //
  //             ((a|b)&c&d)U(!d&e&f)
  //                 ╱           ╲
  //            ((a|b)&c&d)*   (!d&e&f)*
  //             ╱   │   ╲     ╱  │  ╲
  //           a|b*  │    ╲   !   │   ╲
  //           ╱ ╲   │     ╲ ╱    │    ╲
  //          a*  b* c*     d     e*    f*
  //
  // The relabeling of a formula is done in 3 passes:
  //  1. Convert the formula's syntax tree into an undirected graph.
  //  2. Compute the (Boolean) cut points of that graph, using the
  //     Hopcroft-Tarjan algorithm (see below for a reference).
  //  3. Recursively scan the formula's tree until we reach
  //     a (Boolean) cut-point.  If all the children of this node
  //     are cut-points, rename the node with a fresh label.
  //     If it's a n-ary operator, group all children that are
  //     and cut-points relabel them as a whole.
  //
  // On the above example, when processing the cut-point
  // ((a|b)&c&d) we group its children that are cut-points
  // (a|b)&c and rename this group as p0.   Then d gets
  // his own name p1, and when processing (!d&e&f) we group
  // e&f because they are both cut-points, are rename them p1.
  // The result is (p0 & p1) U (!p1 & p2).
  //
  // Problem #B above (handling of n-ary expression) need some
  // additional tricks.  Consider (a&b&c&d) U X(c&d), and assume
  // {a,b,c,d} are Boolean subformulas.  The construction, as we have
  // presented it, would interconnect all of {a,b,c,d}, preventing c&d
  // from being relabeled together.  To help with that, we count the
  // number of time of each subformula is used (or how many parents
  // its has in the syntax DAG), and use that to split (a&b&c&d) into
  // (a&b)&(c&d), separating subformulas that are used only once.  The
  // counting is done by sub_formula_collect(), and the split by
  // split_used_once().
  namespace
  {
    typedef std::vector<formula> succ_vec;
    typedef std::map<formula, succ_vec> fgraph;

    // Convert the formula's syntax tree into an undirected graph
    // labeled by subformulas.
    class formula_to_fgraph final
    {
    public:
      fgraph& g;
      std::stack<formula> s;
      sub_formula_count_t& subcount;

      formula_to_fgraph(fgraph& g, sub_formula_count_t& subcount)
        : g(g), subcount(subcount)
      {
      }

      void visit(formula f)
      {
        {
          // Connect to parent
          auto in = g.emplace(f, succ_vec());
          if (!s.empty())
            {
              formula top = s.top();
              in.first->second.emplace_back(top);
              g[top].emplace_back(f);
              if (!in.second)
                return;
            }
          else
            {
              assert(in.second);
            }
        }
        s.push(f);

        unsigned sz = f.size();
        unsigned i = 0;
        if (sz > 2 && f.is_boolean())
          {
            // If we have a Boolean formula with more than two
            // children, like (a & b & c & d) where some children
            // (assume {a,b}) are used only once, but some other
            // (assume {c,d}) are used multiple time in the formula,
            // then split that into ((a & b) & (c & d)) to give
            // (a & b) a chance to be relabeled as a whole.
            auto pair = split_used_once(f, subcount);
            if (pair.second)
              {
                visit(pair.first);
                visit(pair.second);
                g[pair.first].emplace_back(pair.second);
                g[pair.second].emplace_back(pair.first);
                goto done;
              }
          }
        if (sz > 2 && !f.is_boolean() && f.is(op::And, op::Or))
          {
            /// If we have a formula like (a & b & Xc), consider
            /// it as ((a & b) & Xc) in the graph to isolate the
            /// Boolean operands as a single node.
            formula b = f.boolean_operands(&i);
            if (b)
              visit(b);
          }
        for (; i < sz; ++i)
          visit(f[i]);
      done:
        s.pop();
      }
    };


    typedef std::set<formula> fset;
    struct data_entry // for each node of the graph
    {
      unsigned num; // serial number, in pre-order
      unsigned low; // lowest number accessible via unstacked descendants
      data_entry(unsigned num = 0, unsigned low = 0)
        : num(num), low(low)
      {
      }
    };
    typedef std::unordered_map<formula, data_entry> fmap_t;
    struct stack_entry
    {
      formula grand_parent;
      formula parent;        // current node
      succ_vec::const_iterator current_child;
      succ_vec::const_iterator last_child;
    };
    typedef std::stack<stack_entry> stack_t;

    // Fill c with the Boolean cutpoints of g, starting from start.
    //
    // This is based no "Efficient Algorithms for Graph
    // Manipulation", J. Hopcroft & R. Tarjan, in Communications of
    // the ACM, 16 (6), June 1973.
    //
    // It differs from the original algorithm by returning only the
    // Boolean cut-points, not dealing with the initial state
    // properly (our initial state will always be considered as a
    // cut-point, but since we only return Boolean cut-points it's
    // OK: if the top-most formula is Boolean we want to replace it
    // as a whole), and considering the atomic propositions that
    // are leaves as cutpoints too.
    void cut_points(const fgraph& g, fset& c, formula start)
    {
      stack_t s;

      unsigned num = 0;
      fmap_t data;
      data_entry d = { num, num };
      data[start] = d;
      ++num;
      const succ_vec& children = g.find(start)->second;
      stack_entry e = { start, start, children.begin(), children.end() };
      s.push(e);

      while (!s.empty())
        {
          stack_entry& e  = s.top();
          if (e.current_child != e.last_child)
            {
              // Skip the edge if it is just the reverse of the one
              // we took.
              formula child = *e.current_child;
              if (child == e.grand_parent)
                {
                  ++e.current_child;
                  continue;
                }
              auto i = data.emplace(std::piecewise_construct,
                                    std::forward_as_tuple(child),
                                    std::forward_as_tuple(num, num));
              if (i.second)        // New destination.
                {
                  ++num;
                  const succ_vec& children = g.find(child)->second;
                  stack_entry newe = { e.parent, child,
                                       children.begin(), children.end() };
                  s.push(newe);
                }
              else           // Destination exists.
                {
                  data_entry& dparent = data[e.parent];
                  data_entry& dchild = i.first->second;
                  // If this is a back-edge, update
                  // the low field of the parent.
                  if (dchild.num <= dparent.num)
                    if (dparent.low > dchild.num)
                      dparent.low = dchild.num;
                }
              ++e.current_child;
            }
          else
            {
              formula grand_parent = e.grand_parent;
              formula parent = e.parent;
              s.pop();
              if (!s.empty())
                {
                  data_entry& dparent = data[parent];
                  data_entry& dgrand_parent = data[grand_parent];
                  if (dparent.low >= dgrand_parent.num // cut-point
                      && grand_parent.is_boolean())
                    {
                      c.insert(grand_parent);
                      // Also consider atomic propositions as
                      // cut-points if they are leaves.
                      if (parent.is(op::ap)
                          && g.find(parent)->second.size() == 1)
                        c.insert(parent);
                    }
                  if (dparent.low < dgrand_parent.low)
                    dgrand_parent.low = dparent.low;
                }
            }
        }
    }


    class bse_relabeler final: public relabeler<false, true>
    {
    public:
      const fset& c;
      const sub_formula_count_t& subcount;

      bse_relabeler(ap_generator* gen, const fset& c,
                    relabeling_map* m, const sub_formula_count_t& subcount)
        : relabeler(gen, m), c(c), subcount(subcount)
      {
      }

      using relabeler::visit;

      formula visit(formula f)
      {
        if (f.is(op::ap))
          return rename(f);

        // This is Boolean cut-point?
        // We can only relabel it if all its children are cut-points.
        if (c.find(f) != c.end())
          {
            unsigned fsz = f.size();
            assert(fsz > 0);    // A cut point has children
            if (fsz == 1
                || (fsz == 2
                    && ((c.find(f[0]) != c.end())
                        == (c.find(f[1]) != c.end()))))
              return rename(f);
            if (fsz > 2)
              {
                // cp[0] will contains non cut-points
                // cp[1] will contain cut-points or atomic propositions
                std::vector<formula> cp[2];
                cp[0].reserve(fsz);
                cp[1].reserve(fsz);
                for (unsigned i = 0; i < fsz; ++i)
                  {
                    formula cf = f[i];
                    cp[c.find(cf) != c.end()].push_back(cf);
                  }
                if (cp[0].empty()
                    || cp[1].empty())
                  // all children are cut-points or non-cut-points
                  return rename(f);
                formula cp1group = rename(formula::multop(f.kind(), cp[1]));
                formula cp0group = visit(formula::multop(f.kind(), cp[0]));
                return formula::multop(f.kind(), {cp1group, cp0group});
              }
          }

        // Not a cut-point, recurse
        unsigned sz = f.size();
        if (sz <= 2)
          return f.map([this](formula f)
                       {
                         return visit(f);
                       });

        if (f.is_boolean() && sz > 2)
          // If we have a Boolean formula with more than two
          // children, like (a & b & c & d) where some children
          // (assume {a,b}) are used only once, but some other
          // (assume {c,d}) are used multiple time in the formula,
          // then split that into ((a & b) & (c & d)) to give
          // (a & b) a chance to be relabeled as a whole.
          if (auto pair = split_used_once(f, subcount); pair.second)
            {
              formula left = visit(pair.first);
              formula right = visit(pair.second);
              return formula::multop(f.kind(), { left, right });
            }
        /// If we have a formula like (a & b & Xc), consider
        /// it as ((a & b) & Xc) in the graph to isolate the
        /// Boolean operands as a single node.
        unsigned i = 0;
        std::vector<formula> res;
        formula b = f.boolean_operands(&i);
        if (b && b != f)
          {
            res.reserve(sz - i + 1);
            res.emplace_back(visit(b));
          }
        else
          {
            i = 0;
            res.reserve(sz);
          }
        for (; i < sz; ++i)
          res.emplace_back(visit(f[i]));
        return formula::multop(f.kind(), res);
      }
    };
  }


  formula
  relabel_bse(formula f, relabeling_style style, relabeling_map* m)
  {
    fgraph g;
    sub_formula_count_t subcount;

    // Scan f for sub-formulas used once.
    sub_formula_collect(f, &subcount);

    // Build the graph g from the formula f.
    {
      formula_to_fgraph conv(g, subcount);
      conv.visit(f);
    }

    //// Uncomment to print the graph.
    // for (auto& [f, sv]: g)
    //   {
    //     std::cerr << f << ":\n";
    //     for (auto& s: sv)
    //       std::cerr << "  " << s << '\n';
    //   }

    // Compute its cut-points
    fset c;
    cut_points(g, c, f);

    // std::cerr << "cut-points\n";
    // for (formula cp: c)
    //   std::cerr << "  - " << cp << '\n';

    // Relabel the formula recursively, stopping
    // at cut-points or atomic propositions.
    ap_generator* gen = nullptr;
    switch (style)
      {
      case Pnn:
        gen = new pnn_generator;
        break;
      case Abc:
        gen = new abc_generator;
        break;
      }
    bse_relabeler rel(gen, c, m, subcount);
    return rel.visit(f);
  }

  formula
  relabel_apply(formula f, relabeling_map* m)
  {
    if (f.is(op::ap))
      {
        auto i = m->find(f);
        if (i != m->end())
          return i->second;
      }
    return f.map(relabel_apply, m);
  }

}