# -*- coding: utf-8 -*-
#+TITLE: Constructing and transforming formulas
#+DESCRIPTION: Code example for constructing and transforming formulas in Spot
#+SETUPFILE: setup.org
#+HTML_LINK_UP: tut.html

This page explains how to build formulas and how to iterate over their
syntax trees.

We'll first describe how to build a formula from scratch, by using the
constructor functions associated to each operators, and show that
basic accessor methods for formulas.  We will do that for C++ first,
and then Python.  Once these basics are covered, we will show examples
for traversing and transforming formulas (again in C++ then Python).

* Constructing formulas

** C++

The =spot::formula= class contains static methods that act as
constructors for each supported operator.

The Boolean constants true and false are returned by =formula::tt()=
and =formula:ff()=.  Atomic propositions can be built with
=formula::ap("name")=.  Unary and binary operators use a
straighforward syntax like =formula::F(arg)= or =formula::U(first,
second)=, while n-ary operators take an initializer list as argument
as in =formula::And({arg1, arg2, arg3})=.

Here is the list of supported operators:

#+BEGIN_SRC C++
  // atomic proposition
  formula::ap(string)
  // constants
  formula::ff();
  formula::tt();
  formula::eword();               // empty word (for regular expressions)
  // unary operators
  formula::Not(arg);
  formula::X(arg);
  formula::F(arg);
  formula::G(arg);
  formula::Closure(arg);
  formula::NegClosure(arg);
  // binary operators
  formula::Xor(left, right);
  formula::Implies(left, right);
  formula::Equiv(left, right);
  formula::U(left, right);        // (strong) until
  formula::R(left, right);        // (weak) release
  formula::W(left, right);        // weak until
  formula::M(left, right);        // strong release
  formula::EConcat(left, right);  // Seq
  formula::UConcat(left, right);  // Triggers
  // n-ary operators
  formula::Or({args,...});        // omega-rational Or
  formula::OrRat({args,...});     // rational Or (for regular expressions)
  formula::And({args,...});       // omega-rational And
  formula::AndRat({args,...});    // rational And (for regular expressions)
  formula::AndNLM({args,...});    // non-length-matching rational And (for r.e.)
  formula::Concat({args,...});    // concatenation (for regular expressions)
  formula::Fusion({args,...});    // concatenation (for regular expressions)
  // star-like operators
  formula::Star(arg, min, max);   // Star (for a Kleene star, set min=0 and omit max)
  formula::FStar(arg, min, max);  // Fusion Star
#+END_SRC

These functions implement some very limited type of automatic
simplifications called /trivial identities/.  For instance
=formula::F(formula::X(formula::tt()))= will return the same formula
as =formula::tt()=.  These simplifications are those that involve the
true and false constants, impotence (=F(F(e))=F(e)=), involutions
(=Not(Not(e)=e=), associativity
(=And({And({e1,e2},e3})=And({e1,e2,e3})=).  See [[https://spot.lrde.epita.fr/tl.pdf][tl.pdf]] for a list of
these /trivial identities/.

In addition, the arguments of commutative operators
(e.g. =Xor(e1,e2)=Xor(e2,e1)=) are always reordered.  The order used
always put the Boolean subformulas before the temporal subformulas,
sorts the atomic propositions in alphabetic order, and otherwise order
subformulas by their unique identifier (a constant incremented each
time a new subformula is created).  This reordering is useful to favor
sharing of subformulas, but also helps algorithms that perform
memoization.

Building a formula using these operators is quite straightforward.
The second part of the following example shows how to print some
detail of the top-level oeprator in the formula.

#+BEGIN_SRC C++ :results verbatim :exports both
  #include <iostream>
  #include <spot/tl/formula.hh>
  #include <spot/tl/print.hh>

  int main()
  {
    // Build FGa -> (GFb & GFc)
    spot::formula fga = spot::formula::F(spot::formula::G(spot::formula::ap("a")));
    spot::formula gfb = spot::formula::G(spot::formula::F(spot::formula::ap("b")));
    spot::formula gfc = spot::formula::G(spot::formula::F(spot::formula::ap("c")));
    spot::formula f = spot::formula::Implies(fga, spot::formula::And({gfb, gfc}));

    std::cout << f << '\n';

    // kindstar() prints the name of the operator
    // size() return the number of operands of the operators
    std::cout << f.kindstr() << ", " << f.size() << " children\n";
    // operator[] accesses each operand
    std::cout << "left: " << f[0] << ", right: " << f[1] << '\n';
    // you can also iterate over all operands using a for loop
    for (auto child: f)
      std::cout << "  * " << child << '\n';
    // the type of the operator can be accessed with kind(), which
    // return an element of the spot::op enum.
    std::cout << f[1][0]
              << (f[1][0].kind() == spot::op::F ? " is F\n" : " is not F\n");
    // however because writing f.kind() == spot::op::XXX is quite common, there
    // is also a is() shortcut:
    std::cout << f[1][1]
              << (f[1][1].is(spot::op::G) ? " is G\n" : " is not G\n");
    return 0;
  }
#+END_SRC

#+RESULTS:
: FGa -> (GFb & GFc)
: Implies, 2 children
: left: FGa, right: GFb & GFc
:   * FGa
:   * GFb & GFc
: GFb is not F
: GFc is G

** Python

The Python equivalent is similar:

#+BEGIN_SRC python :results output :exports both
  import spot

  # Build FGa -> (GFb & GFc)
  fga = spot.formula.F(spot.formula.G(spot.formula.ap("a")))
  gfb = spot.formula.G(spot.formula.F(spot.formula.ap("b")));
  gfc = spot.formula.G(spot.formula.F(spot.formula.ap("c")));
  f = spot.formula.Implies(fga, spot.formula.And([gfb, gfc]));

  print(f)

  # kindstar() prints the name of the operator
  # size() return the number of operands of the operators
  print("{}, {} children".format(f.kindstr(), f.size()))
  # [] accesses each operand
  print("left: {f[0]}, right: {f[1]}".format(f=f))
  # you can also iterate over all operands using a for loop
  for child in f:
     print("  *", child)
  # the type of the operator can be accessed with kind(), which returns
  # an op_XXX constant (corresponding the the spot::op enum of C++)
  print(f[1][0], "is F" if f[1][0].kind() == spot.op_F else "is not F")
  # "is" is keyword in Python, the so shortcut is called _is:
  print(f[1][1], "is G" if f[1][1]._is(spot.op_G) else "is not G")
#+END_SRC

#+RESULTS:
: FGa -> (GFb & GFc)
: Implies, 2 children
: left: FGa, right: GFb & GFc
:   * FGa
:   * GFb & GFc
: GFb is not F
: GFc is G

* Transforming formulas

** C++

In Spot, Formula objects are immutable: this allows identical subtrees
to be shared among multiple formulas.  Algorithm that "transform"
formulas (for instance the [[file:tut02.org][relabeling function]]) actually recursively
traverses the input formula to construct the output formula.

Using the operators described in the previous section is enough to
write algorithms on formulas.  However there are two special methods
that makes it a lot easier: =traverse= and =map=.

=traverse= takes a function =fun=, and applies it to a subformulas of
a formula, including the formula itself.  The formula is explored in a
DFS fashion (without skipping subformula that appear twice).  The
children of a formula are explored only if =fun= returns =false=.  If
=fun= returns =true=, that indicates to stop the recursion.

In the following we use a lambda function to count the number of =G=
in the formula.  We also print each subformula to show the recursion,
and stop the recursion as soon as we encounter a subformula without
sugar (the =is_sugar_free_ltl()= method is a constant-time operation,
that tells whether a formulas contains a =F= or =G= operator) to save
time time by not exploring further.


#+BEGIN_SRC C++ :results verbatim :exports both
  #include <iostream>
  #include <spot/tl/formula.hh>
  #include <spot/tl/print.hh>
  #include <spot/tl/parse.hh>

  int main()
  {
    spot::formula f = spot::parse_formula("FGa -> (GFb & GF(c & b & d))");

    int gcount = 0;
    f.traverse([&gcount](spot::formula f)
               {
                 std::cout << f << '\n';
                 if (f.is(spot::op::G))
                   ++gcount;
                 return f.is_sugar_free_ltl();
               });
    std::cout << "=== " << gcount << " G seen ===\n";
    return 0;
  }
#+END_SRC

#+RESULTS:
#+begin_example
FGa -> (GFb & GF(b & c & d))
FGa
Ga
a
GFb & GF(b & c & d)
GFb
Fb
b
GF(b & c & d)
F(b & c & d)
b & c & d
=== 3 G seen ===
#+end_example


The other useful operation is =map=.  This also takes a functional
argument, but that function should input a formula and output a
replacement formula.  =f.map(fun)= applies =fun= to all children of
=f=, and rebuild a same formula as =f=.

Here is a demonstration of how to exchange all =F= and =G= operators
in a formula:

#+BEGIN_SRC C++ :results verbatim :exports both
  #include <iostream>
  #include <spot/tl/formula.hh>
  #include <spot/tl/print.hh>
  #include <spot/tl/parse.hh>

  spot::formula xchg_fg(spot::formula in)
  {
     if (in.is(spot::op::F))
       return spot::formula::G(xchg_fg(in[0]));
     if (in.is(spot::op::G))
       return spot::formula::F(xchg_fg(in[0]));
     // No need to transform Boolean subformulas
     if (in.is_sugar_free_ltl())
       return in;
     // Apply xchg_fg recursively on any other operator's children
     return in.map(xchg_fg);
  }

  int main()
  {
    spot::formula f = spot::parse_formula("FGa -> (GFb & GF(c & b & d))");
    std::cout << "before: " << f << '\n';
    std::cout << "after:  " << xchg_fg(f) << '\n';
    return 0;
  }
#+END_SRC

#+RESULTS:
: before: FGa -> (GFb & GF(b & c & d))
: after:  GFa -> (FGb & FG(b & c & d))


** Python

The Python version of the above two examples uses a very similar
syntax.  Python only supports a very limited form of lambda
expressions, so we have to write a standard function instead:

#+BEGIN_SRC python :results output :exports both
  import spot

  gcount = 0
  def countg(f):
      global gcount
      print(f)
      if f._is(spot.op_G):
          gcount += 1
      return f.is_sugar_free_ltl()

  f = spot.formula("FGa -> (GFb & GF(c & b & d))")
  f.traverse(countg)
  print("===", gcount, "G seen ===")
#+END_SRC

#+RESULTS:
#+begin_example
FGa -> (GFb & GF(b & c & d))
FGa
Ga
a
GFb & GF(b & c & d)
GFb
Fb
b
GF(b & c & d)
F(b & c & d)
b & c & d
=== 3 G seen ===
#+end_example

Here is the =F= and =G= exchange:

#+BEGIN_SRC python :results output :exports both
  import spot

  def xchg_fg(i):
     if i._is(spot.op_F):
        return spot.formula.G(xchg_fg(i[0]));
     if i._is(spot.op_G):
        return spot.formula.F(xchg_fg(i[0]));
     # No need to transform Boolean subformulas
     if i.is_sugar_free_ltl():
        return i;
     # Apply xchg_fg recursively on any other operator's children
     return i.map(xchg_fg);

  f = spot.formula("FGa -> (GFb & GF(c & b & d))")
  print("before:", f)
  print("after: ", xchg_fg(f))
#+END_SRC

#+RESULTS:
: before: FGa -> (GFb & GF(b & c & d))
: after:  GFa -> (FGb & FG(b & c & d))