Upgrading from Spot 1.2.6

Let's begin by just trying to summarize what has changed between Spot 1.2.6 and Spot 2.x, just to get an idea of what will need to be updated.

1. Spot now compiles using the C++17 standard. Compliant compilers are sufficiently widespread now that this should not be an issue.
2. The layout of the source-tree and the layout of the installed header files have been completely overhauled, and several header files have been renamed. An early step of upgrading existing code dependent on Spot will therefore be to update all the #include.
3. The customized version of BuDDy distributed with Spot has been renamed to not clash with any regular version installed on the system.

4. The implementation of LTL formulas has been rewritten.

   They are no longer pointers but plain objects that performs their own reference counting, freeing the programmer from this tedious and error-prone task. They could be handled as if they were shared pointer, with the small difference that they are not using C++'s standard shared pointers.

   All operators are now implemented using a single type of node, with a field that can be used to make a switch instead of what used to require visitors.

5. Allocated object that are large or expected to have a long life (such as automata, BDD dictionaries, accepting runs) are now returned using shared pointers.

6. Spot used to be centered around the concept of TGBA (Transition-based Generalized Büchi Automata), i.e., ω-automata in which a run is accepting if it visits infinitely often one transition in each defined accepting set of transitions. This was implemented as a tgba class.

   In Spot 2.0, each automaton can have a different acceptance condition: this includes well known acceptance conditions such as Büchi, co-Büchi, Rabin, Streett, or parity, but it could also be an arbitrarily complex Boolean combination of those. The central automaton concept is now called TωA (for Transition-based ω-Automata), and implemented as a twa class. The tgba class does not exist anymore: a TGBA is just a TωA in which the acceptance condition has been set to "generalized Büchi".

   While the gained generality on acceptance conditions allows us to write algorithms that may work on any acceptance condition, Spot still contains algorithms that work only on a particular acceptance condition. Those restrictions have to be ensured using preconditions: the acceptance condition does not appear in the type of the C++ object representing the automaton.

7. Several class, functions, and methods, have been renamed. Some have been completely reimplemented, with different interfaces. In particular the tgba_explicit_* family of classes (tgba_explicit_formula, tgba_explicit_number, tgba_explicit_string used to encode a TGBA using a graph whose state was named using LTL formulas, integers, or strings) are replaced by a twa_graph class in which the representation is more compact and efficient, but where states are necessarily numbered. Many algorithms have been rewritten on top of this twa_graph class, and this simplified them a lot.

Because Spot now relies on C++17 features, programs that use Spot should at least be compiled using this version (or a later one) of the language.

This is usually done by passing the option -std=c++17 to g++ or clang++.

Upgrading from C++98, C++03, C++11, or C++14 to C++17 should be relatively smooth as the language is mostly backward compatible.

Spot install a customized version of the BDD library BuDDy.

As shown in the table of renamed header files, the header files for this library were all renamed: they have an x appended (bddx.h, bvecx.h, fddx.h). The name of the library was renamed as well: libbdd.so and libbdd.a are now respectively libbddx.so and libbddx.a. This means one should now pass -lspot -lbddx to the linker (instead of -lspot -lbdd) when linking against Spot.

As can be guessed from the the table of renamed header files, the class hierarchy for temporal formulas has been entirely rewritten. This change is actually quite massive (~13200 lines removed, ~8200 lines added), and brings some nice benefits:

• LTL/PSL formulas are now represented by lightweight formula objects (instead of pointers to children of an abstract formula class) that perform reference counting automatically.
• There is no hierarchy anymore: all operators are represented by a single type of node in the syntax tree, and an enumerator is used to distinguish between operators.
• Visitors have been replaced by member functions such as map() or traverse(), that take a function (usually written as a lambda function) and apply it to the nodes of the tree.
• As a consequence, writing algorithms that manipulate formula is more friendly, and several algorithms that spanned a few pages have been reduced to a few lines. This page illustrates the new interface.

Also the spot::ltl namespace has been removed: everything is directly in spot now.

In code where formulas are just parsed from input string, and then passed on to other algorithms (e.g., translation to automata): it will suffice to change all occurrences of const spot::ltl::formula* into a plain spot::formula, and remove all calls to ->destroy() or ->clone() (that were used to perform explicit reference counting).

By convention foo_ptr is a typedef for std::share_ptr<foo>. The following typedefs are provided:

typedef std::shared_ptr<aig> aig_ptr;
typedef std::shared_ptr<bdd_dict> bdd_dict_ptr;
typedef std::shared_ptr<const aig> const_aig_ptr;
typedef std::shared_ptr<const fair_kripke> const_fair_kripke_ptr;
typedef std::shared_ptr<const kripke> const_kripke_ptr;
typedef std::shared_ptr<const kripke_explicit> const_kripke_explicit_ptr;
typedef std::shared_ptr<const parsed_aut> const_parsed_aut_ptr;
typedef std::shared_ptr<const taa_tgba_formula> const_taa_tgba_formula_ptr;
typedef std::shared_ptr<const taa_tgba_string> const_taa_tgba_string_ptr;
typedef std::shared_ptr<const ta> const_ta_ptr;
typedef std::shared_ptr<const ta_explicit> const_ta_explicit_ptr;
typedef std::shared_ptr<const ta_product> const_ta_product_ptr;
typedef std::shared_ptr<const tgta> const_tgta_ptr;
typedef std::shared_ptr<const tgta_explicit> const_tgta_explicit_ptr;
typedef std::shared_ptr<const twa> const_twa_ptr;
typedef std::shared_ptr<const twacube> const_twacube_ptr;
typedef std::shared_ptr<const twa_graph> const_twa_graph_ptr;
typedef std::shared_ptr<const twa_product> const_twa_product_ptr;
typedef std::shared_ptr<const twa_run> const_twa_run_ptr;
typedef std::shared_ptr<emptiness_check> emptiness_check_ptr;
typedef std::shared_ptr<emptiness_check_result> emptiness_check_result_ptr;
typedef std::shared_ptr<fair_kripke> fair_kripke_ptr;
typedef std::shared_ptr<kripke_explicit> kripke_explicit_ptr;
typedef std::shared_ptr<kripke_graph> kripke_graph_ptr;
typedef std::shared_ptr<kripke> kripke_ptr;
typedef std::shared_ptr<parsed_aut> parsed_aut_ptr;
typedef std::shared_ptr<taa_tgba_formula> taa_tgba_formula_ptr;
typedef std::shared_ptr<taa_tgba_string> taa_tgba_string_ptr;
typedef std::shared_ptr<ta_explicit> ta_explicit_ptr;
typedef std::shared_ptr<ta_product> ta_product_ptr;
typedef std::shared_ptr<ta> ta_ptr;
typedef std::shared_ptr<tgta_explicit> tgta_explicit_ptr;
typedef std::shared_ptr<tgta> tgta_ptr;
typedef std::shared_ptr<twacube> twacube_ptr;
typedef std::shared_ptr<twa_graph> twa_graph_ptr;
typedef std::shared_ptr<twa_product> twa_product_ptr;
typedef std::shared_ptr<twa_run> twa_run_ptr;
typedef std::shared_ptr<twa> twa_ptr;
typedef std::shared_ptr<twa_univ_remover> twa_univ_remover_ptr;
typedef std::shared_ptr<twa_word> twa_word_ptr;

You should update any foo* to foo_ptr if it appears in this list, and remove any explicit call to delete. Additionally, if foo is a class, there is usually a make_foo(...) function that calls std::make_shared<foo>(...).

As an example of usage, see this example of creating an automaton explicitly where spot::make_bdd_dict() is used to create a BDD dictionary that is then passed to spot::make_twa_graph().

The change has several implications besides just the changing the name:

• twa can represent a larger class of automata than tgba could, since the acceptance condition can be arbitrarily complex. Algorithm that used to input tgba can be either generalized to handle those larger acceptance conditions, or restricted to generalized Büchi acceptance. The typical way to ensure that an input automaton has generalized Büchi acceptance is

if (!input->acc().is_generalized_buchi())
   throw std::runtime_error
     ("myalgorithm() can only works with generalized Büchi acceptance");

• Some methods of the tgba class have been removed, include some pure virtual methods. If you wrote a subclass, you may simply remove the following methods:
  • compute_support_conditions(),
  • compute_support_variables(),
  • transition_annotation(),
  • neg_acceptance_conditions().
• The succ_iter() method lost its last two (optional) arguments.
• The badly named number_of_acceptance_conditions() has been renamed to num_sets().
• The tgba_succ_iterator class was renamed to twa_succ_iterator, and its interface was changed:
  • current_state(), current_condition(), and current_acceptance_conditions() are now named respectively dst(), cond(), acc().
  • the first() and next() method now return a Boolean that indicates whether we moved to a new successor (true), or reached the end of the list of successors (false). This return value should be the same as what !done() would return, and can be used to avoid some costly calls to the done() virtual function.
• Membership to acceptance sets is now represented using a bit set (instead of BDDs). An TωA aut uses aut->num_sets() acceptance sets numbered from 0 to aut->num_sets()-1. The bit sets are implemented as instances of spot::acc_cond::mark_t, so they are also called "acceptance marks".

• The twa::release_iter() method allow iterators to be recycled. Each twa now contains a mutable field iter_cache_ where release_iter() stores the last returned iterator. If this field is not equal to nullptr, then succ_iter() can use it to allocate a new iterator more efficiently.

  In the code for succ_iter(), instead of:

  // ...
  return new my_iterator(arg1, arg2, arg3);
  

  we can now have

  // ...
  if (iter_cache_)
    {
       my_iterator* it = static_cast<my_iterator*>(iter_cache_);
       iter_cache_ = nullptr;
       // Some method to change the member that need changing.
       // (Here we assume that arg2 is the same for all iterators
       // built on this automaton.)
       it->recycle(arg1, arg3);
       return it;
    }
  return new my_iterator(arg1, arg2, arg3);
  

  So not only do we save the calls to new and delete, but we also save the time it takes to construct the objects (including setting up the virtual table), and via a recycle() method that has to be added to the iterator, we update only the attributes that needs to be updated (for instance if the iterator contains a pointer back to the automaton, this pointer requires no update when the iterator is recycled).

  The iterator stored in twa::iter_cache_ is destroyed by twa::~twa(). In some case, this is too late. For instance the class twa_product uses a custom memory pool to allocate new iterators more efficiently, and this memory pool is destroyed in twa_product::~twa_product(). In this case, referencing twa::iter_cache_ for its deletion in twa::~tgba() would be incorrect: it has to be done in twa_product::~twa_product() before destroying the pool.

For a TGBA aut, and a state s, the original way to loop over the successors of s was:

tgba_succ_iterator* i = aut->succ_iter(s);
for (i->first(); !i->done(); i->next())
  {
     // use i->current_state()
     //     i->current_condition()
     //     i->current_acceptance_conditions()
  }
delete i;

While the above would still run after a few renamings, it can now be written more efficiently as:

twa_succ_iterator* i = aut->succ_iter(s);
if (i->first())
  do
    {
      // use i->dst()
      //     i->cond()
      //     i->acc()
    }
  while (i->next());
aut->release_iter(i);

If the original code did \(1\) virtual call to first(), \(n+1\) virtual calls to done(), and \(n\) virtual calls to next(), the new version does as many calls to first() and done() while avoiding all the calls to done(). Furthermore the call release_iter() makes it possible for the next call to succ_iter() to reuse the same iterator.

Because the the above loop is quite common we also have some syntactic sugar via the new succ() method that returns a fake container with begin() and end() methods so that this works:

for (auto i: aut->succ(s))
  {
    // use i->dst()
    //     i->cond()
    //     i->acc()
  }

• Each twa now has a BDD dictionary, so the get_dict() method is implemented once for all in twa, and should not be implemented anymore in sub-classes.
• There should now be very few cases where it is necessary to call methods of the BDD dictionary attached to a twa. Registering the atomic proposition used by a twa should now be done via the twa::register_ap() or twa::copy_ap_of() methods. Accessing all registered propositions is achievable with twa::ap() or twa::ap_vars(). All propositions registered by an automaton are automatically unregistered when the automaton is destroyed.

The following is a non-exhaustive list of functions or classes that have been renamed.