ltl2tgba_fm.cc 24.9 KB
Newer Older
1
2
3
// Copyright (C) 2008, 2009, 2010 Laboratoire de Recherche et
// Dveloppement de l'Epita (LRDE).
// Copyright (C) 2003, 2004, 2005, 2006 Laboratoire
4
5
// d'Informatique de Paris 6 (LIP6), dpartement Systmes Rpartis
// Coopratifs (SRC), Universit Pierre et Marie Curie.
Alexandre Duret-Lutz's avatar
Alexandre Duret-Lutz committed
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
//
// 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 2 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 Spot; see the file COPYING.  If not, write to the Free
// Software Foundation, Inc., 59 Temple Place - Suite 330, Boston, MA
// 02111-1307, USA.

24
#include "misc/hash.hh"
25
#include "misc/bddalloc.hh"
26
#include "misc/bddlt.hh"
27
#include "misc/minato.hh"
28
29
#include "ltlast/visitor.hh"
#include "ltlast/allnodes.hh"
30
31
32
#include "ltlvisit/lunabbrev.hh"
#include "ltlvisit/nenoform.hh"
#include "ltlvisit/tostring.hh"
33
#include "ltlvisit/postfix.hh"
34
#include "ltlvisit/apcollect.hh"
35
#include <cassert>
36
#include <memory>
37
#include "ltl2tgba_fm.hh"
38
#include "ltlvisit/contain.hh"
39
40
41
42
43
44
45
46

namespace spot
{
  using namespace ltl;

  namespace
  {

47
48
    // Helper dictionary.  We represent formulae using BDDs to
    // simplify them, and then translate BDDs back into formulae.
49
50
51
52
53
    //
    // The name of the variables are inspired from Couvreur's FM paper.
    //   "a" variables are promises (written "a" in the paper)
    //   "next" variables are X's operands (the "r_X" variables from the paper)
    //   "var" variables are atomic propositions.
54
    class translate_dict
55
56
57
    {
    public:

58
59
      translate_dict(bdd_dict* dict)
	: dict(dict),
60
61
62
63
64
65
66
67
68
69
	  a_set(bddtrue),
	  var_set(bddtrue),
	  next_set(bddtrue)
      {
      }

      ~translate_dict()
      {
	fv_map::iterator i;
	for (i = next_map.begin(); i != next_map.end(); ++i)
70
	  i->first->destroy();
71
	dict->unregister_all_my_variables(this);
72
73
      }

74
75
      bdd_dict* dict;

76
77
      typedef bdd_dict::fv_map fv_map;
      typedef bdd_dict::vf_map vf_map;
78
79
80
81
82
83
84
85
86

      fv_map next_map;	       ///< Maps "Next" variables to BDD variables
      vf_map next_formula_map; ///< Maps BDD variables to "Next" variables

      bdd a_set;
      bdd var_set;
      bdd next_set;

      int
87
      register_proposition(const formula* f)
88
      {
89
	int num = dict->register_proposition(f, this);
90
91
92
93
94
	var_set &= bdd_ithvar(num);
	return num;
      }

      int
95
      register_a_variable(const formula* f)
96
      {
97
	int num = dict->register_acceptance_variable(f, this);
98
99
100
101
102
	a_set &= bdd_ithvar(num);
	return num;
      }

      int
103
      register_next_variable(const formula* f)
104
105
106
107
108
109
110
111
112
113
      {
	int num;
	// Do not build a Next variable that already exists.
	fv_map::iterator sii = next_map.find(f);
	if (sii != next_map.end())
	  {
	    num = sii->second;
	  }
	else
	  {
114
	    f = f->clone();
115
	    num = dict->register_anonymous_variables(1, this);
116
117
118
119
120
121
122
	    next_map[f] = num;
	    next_formula_map[num] = f;
	  }
	next_set &= bdd_ithvar(num);
	return num;
      }

123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
      std::ostream&
      dump(std::ostream& os) const
      {
	fv_map::const_iterator fi;
	os << "Next Variables:" << std::endl;
	for (fi = next_map.begin(); fi != next_map.end(); ++fi)
	{
	  os << "  " << fi->second << ": Next[";
	  to_string(fi->first, os) << "]" << std::endl;
	}
	os << "Shared Dict:" << std::endl;
	dict->dump(os);
	return os;
      }

138
      formula*
139
140
141
142
      var_to_formula(int var) const
      {
	vf_map::const_iterator isi = next_formula_map.find(var);
	if (isi != next_formula_map.end())
143
	  return isi->second->clone();
144
145
	isi = dict->acc_formula_map.find(var);
	if (isi != dict->acc_formula_map.end())
146
	  return isi->second->clone();
147
148
	isi = dict->var_formula_map.find(var);
	if (isi != dict->var_formula_map.end())
149
	  return isi->second->clone();
150
	assert(0);
151
152
153
	// Never reached, but some GCC versions complain about
	// a missing return otherwise.
	return 0;
154
155
      }

156
      formula*
157
158
159
      conj_bdd_to_formula(bdd b)
      {
	if (b == bddfalse)
160
161
	  return constant::false_instance();
	multop::vec* v = new multop::vec;
162
163
164
	while (b != bddtrue)
	  {
	    int var = bdd_var(b);
165
	    formula* res = var_to_formula(var);
166
167
168
	    bdd high = bdd_high(b);
	    if (high == bddfalse)
	      {
169
		res = unop::instance(unop::Not, res);
170
171
172
173
		b = bdd_low(b);
	      }
	    else
	      {
174
		assert(bdd_low(b) == bddfalse);
175
176
177
178
179
		b = high;
	      }
	    assert(b != bddfalse);
	    v->push_back(res);
	  }
180
	return multop::instance(multop::And, v);
181
182
      }

183
184
      const formula*
      bdd_to_formula(bdd f)
185
      {
186
	if (f == bddfalse)
187
	  return constant::false_instance();
188

189
190
191
192
193
194
195
196
197
	multop::vec* v = new multop::vec;

	minato_isop isop(f);
	bdd cube;
	while ((cube = isop.next()) != bddfalse)
	  v->push_back(conj_bdd_to_formula(cube));

	return multop::instance(multop::Or, v);
      }
198
199

      void
200
      conj_bdd_to_acc(tgba_explicit_formula* a, bdd b,
201
		      tgba_explicit::transition* t)
202
203
204
205
206
207
208
209
      {
	assert(b != bddfalse);
	while (b != bddtrue)
	  {
	    int var = bdd_var(b);
	    bdd high = bdd_high(b);
	    if (high == bddfalse)
	      {
210
		// Simply ignore negated acceptance variables.
211
212
213
214
		b = bdd_low(b);
	      }
	    else
	      {
215
		formula* ac = var_to_formula(var);
216

217
		if (!a->has_acceptance_condition(ac))
218
		  a->declare_acceptance_condition(ac->clone());
219
		a->add_acceptance_condition(t, ac);
220
221
222
223
224
225
226
227
		b = high;
	      }
	    assert(b != bddfalse);
	  }
      }
    };


228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271

    // Gather all promises of a formula.  These are the
    // right-hand sides of U or F operators.
    class ltl_promise_visitor: public postfix_visitor
    {
    public:
      ltl_promise_visitor(translate_dict& dict)
	: dict_(dict), res_(bddtrue)
      {
      }

      virtual
      ~ltl_promise_visitor()
      {
      }

      bdd
      result() const
      {
	return res_;
      }

      using postfix_visitor::doit;

      virtual void
      doit(unop* node)
      {
	if (node->op() == unop::F)
	  res_ &= bdd_ithvar(dict_.register_a_variable(node->child()));
      }

      virtual void
      doit(binop* node)
      {
	if (node->op() == binop::U)
	  res_ &= bdd_ithvar(dict_.register_a_variable(node->second()));
      }

    private:
      translate_dict& dict_;
      bdd res_;
    };


272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
    // The rewrite rules used here are adapted from Jean-Michel
    // Couvreur's FM paper.
    class ltl_trad_visitor: public const_visitor
    {
    public:
      ltl_trad_visitor(translate_dict& dict)
	: dict_(dict)
      {
      }

      virtual
      ~ltl_trad_visitor()
      {
      }

287
288
      bdd
      result() const
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
      {
	return res_;
      }

      void
      visit(const atomic_prop* node)
      {
	res_ = bdd_ithvar(dict_.register_proposition(node));
      }

      void
      visit(const constant* node)
      {
	switch (node->val())
	  {
	  case constant::True:
	    res_ = bddtrue;
	    return;
	  case constant::False:
	    res_ = bddfalse;
	    return;
	  }
	/* Unreachable code.  */
	assert(0);
      }

      void
      visit(const unop* node)
      {
	switch (node->op())
	  {
	  case unop::F:
	    {
	      // r(Fy) = r(y) + a(y)r(XFy)
323
324
325
	      const formula* child = node->child();
	      bdd y = recurse(child);
	      int a = dict_.register_a_variable(child);
326
327
328
329
330
331
	      int x = dict_.register_next_variable(node);
	      res_ = y | (bdd_ithvar(a) & bdd_ithvar(x));
	      return;
	    }
	  case unop::G:
	    {
332
333
334
335
336
337
338
339
340
341
342
343
	      // The paper suggests that we optimize GFy
	      // as
	      //   r(GFy) = (r(y) + a(y))r(XGFy)
	      // instead of
	      //   r(GFy) = (r(y) + a(y)r(XFy)).r(XGFy)
	      // but this is just a particular case
	      // of the "merge all states with the same
	      // symbolic rewriting" optimization we do later.
	      // (r(Fy).r(GFy) and r(GFy) have the same symbolic
	      // rewriting.)  Let's keep things simple here.

	      // r(Gy) = r(y)r(XGy)
344
	      const formula* child = node->child();
345
	      int x = dict_.register_next_variable(node);
346
347
	      bdd y = recurse(child);
	      res_ = y & bdd_ithvar(x);
348
349
350
351
	      return;
	    }
	  case unop::Not:
	    {
352
	      // r(!y) = !r(y)
353
354
355
356
357
	      res_ = bdd_not(recurse(node->child()));
	      return;
	    }
	  case unop::X:
	    {
358
	      // r(Xy) = Next[y]
359
360
361
362
	      int x = dict_.register_next_variable(node->child());
	      res_ = bdd_ithvar(x);
	      return;
	    }
363
364
	  case unop::Finish:
	    assert(!"unsupported operator");
365
366
367
368
369
370
371
372
373
374
375
376
377
	  }
	/* Unreachable code.  */
	assert(0);
      }

      void
      visit(const binop* node)
      {
	bdd f1 = recurse(node->first());
	bdd f2 = recurse(node->second());

	switch (node->op())
	  {
378
	    // r(f1 logical-op f2) = r(f1) logical-op r(f2)
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
	  case binop::Xor:
	    res_ = bdd_apply(f1, f2, bddop_xor);
	    return;
	  case binop::Implies:
	    res_ = bdd_apply(f1, f2, bddop_imp);
	    return;
	  case binop::Equiv:
	    res_ = bdd_apply(f1, f2, bddop_biimp);
	    return;
	  case binop::U:
	    {
	      // r(f1 U f2) = r(f2) + a(f2)r(f1)r(X(f1 U f2))
	      int a = dict_.register_a_variable(node->second());
	      int x = dict_.register_next_variable(node);
	      res_ = f2 | (bdd_ithvar(a) & f1 & bdd_ithvar(x));
	      return;
	    }
	  case binop::R:
	    {
	      // r(f1 R f2) = r(f1)r(f2) + r(f2)r(X(f1 U f2))
	      int x = dict_.register_next_variable(node);
	      res_ = (f1 & f2) | (f2 & bdd_ithvar(x));
	      return;
	    }
	  }
	/* Unreachable code.  */
	assert(0);
      }

408
409
410
411
412
413
      void
      visit(const automatop*)
      {
	assert(!"unsupported operator");
      }

414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
      void
      visit(const multop* node)
      {
	int op = -1;
	switch (node->op())
	  {
	  case multop::And:
	    op = bddop_and;
	    res_ = bddtrue;
	    break;
	  case multop::Or:
	    op = bddop_or;
	    res_ = bddfalse;
	    break;
	  }
	assert(op != -1);
	unsigned s = node->size();
	for (unsigned n = 0; n < s; ++n)
	  {
	    res_ = bdd_apply(res_, recurse(node->nth(n)), op);
	  }
      }

      bdd
      recurse(const formula* f)
      {
	ltl_trad_visitor v(dict_);
	f->accept(v);
	return v.result();
      }


    private:
      translate_dict& dict_;
      bdd res_;
    };

451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512

    // Check whether a formula has a R or G operator at its top-level
    // (preceding logical operators do not count).
    class ltl_possible_fair_loop_visitor: public const_visitor
    {
    public:
      ltl_possible_fair_loop_visitor()
	: res_(false)
      {
      }

      virtual
      ~ltl_possible_fair_loop_visitor()
      {
      }

      bool
      result() const
      {
	return res_;
      }

      void
      visit(const atomic_prop*)
      {
      }

      void
      visit(const constant*)
      {
      }

      void
      visit(const unop* node)
      {
	if (node->op() == unop::G)
	  res_ = true;
      }

      void
      visit(const binop* node)
      {
	switch (node->op())
	  {
	    // r(f1 logical-op f2) = r(f1) logical-op r(f2)
	  case binop::Xor:
	  case binop::Implies:
	  case binop::Equiv:
	    node->first()->accept(*this);
	    if (!res_)
	      node->second()->accept(*this);
	    return;
	  case binop::U:
	    return;
	  case binop::R:
	    res_ = true;
	    return;
	  }
	/* Unreachable code.  */
	assert(0);
      }

513
514
515
516
517
518
      void
      visit(const automatop*)
      {
	assert(!"unsupported operator");
      }

519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
      void
      visit(const multop* node)
      {
	unsigned s = node->size();
	for (unsigned n = 0; n < s && !res_; ++n)
	  {
	    node->nth(n)->accept(*this);
	  }
      }

    private:
      bool res_;
    };

    // Check whether a formula can be part of a fair loop.
    // Cache the result for efficiency.
    class possible_fair_loop_checker
    {
    public:
      bool
      check(const formula* f)
      {
541
542
	pfl_map::const_iterator i = pfl_.find(f);
	if (i != pfl_.end())
543
544
545
546
	  return i->second;
	ltl_possible_fair_loop_visitor v;
	f->accept(v);
	bool rel = v.result();
547
	pfl_[f] = rel;
548
549
550
551
	return rel;
      }

    private:
552
      typedef Sgi::hash_map<const formula*, bool, formula_ptr_hash> pfl_map;
553
      pfl_map pfl_;
554
555
    };

556
557
558
    class formula_canonizer
    {
    public:
559
      formula_canonizer(translate_dict& d,
560
			bool fair_loop_approx, bdd all_promises)
561
562
	: v_(d),
	  fair_loop_approx_(fair_loop_approx),
563
	  all_promises_(all_promises)
564
565
566
567
568
      {
	// For cosmetics, register 1 initially, so the algorithm will
	// not register an equivalent formula first.
	b2f_[bddtrue] = constant::true_instance();
      }
569

570
571
      ~formula_canonizer()
      {
572
	while (!f2b_.empty())
573
	  {
574
575
576
	    formula_to_bdd_map::iterator i = f2b_.begin();
	    const formula* f = i->first;
	    f2b_.erase(i);
577
	    f->destroy();
578
	  }
579
580
581
      }

      bdd
582
      translate(const formula* f, bool* new_flag = 0)
583
584
585
586
587
588
      {
	// Use the cached result if available.
	formula_to_bdd_map::const_iterator i = f2b_.find(f);
	if (i != f2b_.end())
	  return i->second;

589
590
591
	if (new_flag)
	  *new_flag = true;

592
593
594
	// Perform the actual translation.
	f->accept(v_);
	bdd res = v_.result();
595
596
597
598
599
600
601
602
603
604
605
606

	// Apply the fair-loop approximation if requested.
	if (fair_loop_approx_)
	  {
	    // If the source cannot possibly be part of a fair
	    // loop, make all possible promises.
	    if (fair_loop_approx_
		&& f != constant::true_instance()
		&& !pflc_.check(f))
	      res &= all_promises_;
	  }

607
	f2b_[f->clone()] = res;
608
609
610
611
612
613
614
615
616
617

	// Register the reverse mapping if it is not already done.
	if (b2f_.find(res) == b2f_.end())
	  b2f_[res] = f;
	return res;
      }

      const formula*
      canonize(const formula* f)
      {
618
619
	bool new_variable = false;
	bdd b = translate(f, &new_variable);
620
621

	bdd_to_formula_map::iterator i = b2f_.find(b);
622
623
	// Since we have just translated the formula, it is
	// necessarily in b2f_.
624
625
626
	assert(i != b2f_.end());

	if (i->second != f)
627
	  {
628
	    // The translated bdd maps to an already seen formula.
629
	    f->destroy();
630
	    f = i->second->clone();
631
	  }
632
	return f;
633
634
      }

635
636
637
638
639
640
641
642
643
644
645
646
647
    private:
      ltl_trad_visitor v_;
      // Map each formula to its associated bdd.  This speed things up when
      // the same formula is translated several times, which especially
      // occurs when canonize() is called repeatedly inside exprop.
      typedef std::map<bdd, const formula*, bdd_less_than> bdd_to_formula_map;
      bdd_to_formula_map b2f_;
      // Map a representation of successors to a canonical formula.
      // We do this because many formulae (such as `aR(bRc)' and
      // `aR(bRc).(bRc)') are equivalent, and are trivially identified
      // by looking at the set of successors.
      typedef std::map<const formula*, bdd> formula_to_bdd_map;
      formula_to_bdd_map f2b_;
648
649
650
651

      possible_fair_loop_checker pflc_;
      bool fair_loop_approx_;
      bdd all_promises_;
652
653
654
655
656
    };

  }

  typedef std::map<bdd, bdd, bdd_less_than> prom_map;
657
  typedef Sgi::hash_map<const formula*, prom_map, formula_ptr_hash> dest_map;
658
659
660
661

  static void
  fill_dests(translate_dict& d, dest_map& dests, bdd label, const formula* dest)
  {
662
    bdd conds = bdd_existcomp(label, d.var_set);
663
664
    bdd promises = bdd_existcomp(label, d.a_set);

665
666
667
668
669
670
671
672
    dest_map::iterator i = dests.find(dest);
    if (i == dests.end())
      {
	dests[dest][promises] = conds;
      }
    else
      {
	i->second[promises] |= conds;
673
	dest->destroy();
674
675
676
677
      }
  }


678
  tgba_explicit*
679
  ltl_to_tgba_fm(const formula* f, bdd_dict* dict,
680
		 bool exprop, bool symb_merge, bool branching_postponement,
681
		 bool fair_loop_approx, const atomic_prop_set* unobs,
682
		 int reduce_ltl)
683
684
685
686
687
  {
    // Normalize the formula.  We want all the negations on
    // the atomic propositions.  We also suppress logic
    // abbreviations such as <=>, =>, or XOR, since they
    // would involve negations at the BDD level.
688
689
    formula* f1 = unabbreviate_logic(f);
    formula* f2 = negative_normal_form(f1);
690
    f1->destroy();
691

692
693
694
695
    // Simplify the formula, if requested.
    if (reduce_ltl)
      {
	formula* tmp = reduce(f2, reduce_ltl);
696
	f2->destroy();
697
698
699
	f2 = tmp;
      }

700
701
    typedef std::set<const formula*, formula_ptr_less_than> set_type;
    set_type formulae_to_translate;
702

703
    translate_dict d(dict);
704

705
706
    // Compute the set of all promises that can possibly occurre
    // inside the formula.
707
    bdd all_promises = bddtrue;
708
    if (fair_loop_approx || unobs)
709
710
711
712
713
714
      {
	ltl_promise_visitor pv(d);
	f2->accept(pv);
	all_promises = pv.result();
      }

715
    formula_canonizer fc(d, fair_loop_approx, all_promises);
716

717
718
719
720
721
722
723
724
725
726
    // These are used when atomic propositions are interpreted as
    // events.  There are two kinds of events: observable events are
    // those used in the formula, and unobservable events or other
    // events that can occur at anytime.  All events exclude each
    // other.
    bdd observable_events = bddfalse;
    bdd unobservable_events = bddfalse;
    if (unobs)
      {
	bdd neg_events = bddtrue;
727
728
	std::auto_ptr<atomic_prop_set> aps(atomic_prop_collect(f));
	for (atomic_prop_set::const_iterator i = aps->begin();
729
730
731
732
733
734
735
736
	     i != aps->end(); ++i)
	  {
	    int p = d.register_proposition(*i);
	    bdd pos = bdd_ithvar(p);
	    bdd neg = bdd_nithvar(p);
	    observable_events = (observable_events & neg) | (neg_events & pos);
	    neg_events &= neg;
	  }
737
	for (atomic_prop_set::const_iterator i = unobs->begin();
738
739
740
741
742
743
744
745
746
747
748
749
750
	     i != unobs->end(); ++i)
	  {
	    int p = d.register_proposition(*i);
	    bdd pos = bdd_ithvar(p);
	    bdd neg = bdd_nithvar(p);
	    unobservable_events = ((unobservable_events & neg)
				   | (neg_events & pos));
	    observable_events &= neg;
	    neg_events &= neg;
	  }
      }
    bdd all_events = observable_events | unobservable_events;

751

752
    tgba_explicit_formula* a = new tgba_explicit_formula(dict);
753

754
755
    formulae_to_translate.insert(f2);
    a->set_init_state(f2);
756
757
758
759

    while (!formulae_to_translate.empty())
      {
	// Pick one formula.
760
	const formula* now = *formulae_to_translate.begin();
761
762
763
	formulae_to_translate.erase(formulae_to_translate.begin());

	// Translate it into a BDD to simplify it.
764
	bdd res = fc.translate(now);
765

766
767
768
769
	// Handle exclusive events.
	if (unobs)
	  {
	    res &= observable_events;
770
	    int n = d.register_next_variable(now);
771
772
773
	    res |= unobservable_events & bdd_ithvar(n) & all_promises;
	  }

774
775
776
777
778
779
780
781
782
783
784
785
786
787
788
	// We used to factor only Next and A variables while computing
	// prime implicants, with
	//    minato_isop isop(res, d.next_set & d.a_set);
	// in order to obtain transitions with formulae of atomic
	// proposition directly, but unfortunately this led to strange
	// factorizations.  For instance f U g was translated as
	//     r(f U g) = g + a(g).r(X(f U g)).(f + g)
	// instead of just
	//     r(f U g) = g + a(g).r(X(f U g)).f
	// Of course both formulae are logically equivalent, but the
	// latter is "more deterministic" than the former, so it should
	// be preferred.
	//
	// Therefore we now factor all variables.  This may lead to more
	// transitions than necessary (e.g.,  r(f + g) = f + g  will be
789
790
791
	// coded as two transitions), but we later merge all transitions
	// with same source/destination and acceptance conditions.  This
	// is the goal of the `dests' hash.
792
	//
793
	// Note that this is still not optimal.  For instance it is
794
	// better to encode `f U g' as
795
	//     r(f U g) = g + a(g).r(X(f U g)).f.!g
796
797
798
799
	// because that leads to a deterministic automaton.  In order
	// to handle this, we take the conditions of any transition
	// going to true (it's `g' here), and remove it from the other
	// transitions.
800
801
802
	//
	// In `exprop' mode, considering all possible combinations of
	// outgoing propositions generalizes the above trick.
803
	dest_map dests;
804

805
	// Compute all outgoing arcs.
806
807
808
809
810

	// If EXPROP is set, we will refine the symbolic
	// representation of the successors for all combinations of
	// the atomic properties involved in the formula.
	// VAR_SET is the set of these properties.
811
	bdd var_set = bdd_existcomp(bdd_support(res), d.var_set);
812
813
814
815
816
	// ALL_PROPS is the combinations we have yet to consider.
	// We used to start with `all_props = bddtrue', but it is
	// more efficient to start with the set of all satisfiable
	// variables combinations.
	bdd all_props = bdd_existcomp(res, d.var_set);
817
	while (all_props != bddfalse)
818
	  {
819
820
821
	    bdd one_prop_set = bddtrue;
	    if (exprop)
	      one_prop_set = bdd_satoneset(all_props, var_set, bddtrue);
822
	    all_props -= one_prop_set;
823

824
	    typedef std::map<bdd, const formula*, bdd_less_than> succ_map;
825
826
	    succ_map succs;

827
828
829
830
831
832
833
834
835
836
837
838
839
840
	    // Compute prime implicants.
	    // The reason we use prime implicants and not bdd_satone()
	    // is that we do not want to get any negation in front of Next
	    // or Acc variables.  We wouldn't know what to do with these.
	    // We never added negations in front of these variables when
	    // we built the BDD, so prime implicants will not "invent" them.
	    //
	    // FIXME: minato_isop is quite expensive, and I (=adl)
	    // don't think we really care that much about getting the
	    // smalled sum of products that minato_isop strives to
	    // compute.  Given that Next and Acc variables should
	    // always be positive, maybe there is a faster way to
	    // compute the successors?  E.g. using bdd_satone() and
	    // ignoring negated Next and Acc variables.
841
842
843
	    minato_isop isop(res & one_prop_set);
	    bdd cube;
	    while ((cube = isop.next()) != bddfalse)
844
	      {
845
		bdd label = bdd_exist(cube, d.next_set);
846
		bdd dest_bdd = bdd_existcomp(cube, d.next_set);
847
848
		const formula* dest = d.conj_bdd_to_formula(dest_bdd);

849
850
851
852
		// Simplify the formula, if requested.
		if (reduce_ltl)
		  {
		    formula* tmp = reduce(dest, reduce_ltl);
853
		    dest->destroy();
854
855
856
857
858
859
		    dest = tmp;
		    // Ignore the arc if the destination reduces to false.
		    if (dest == constant::false_instance())
		      continue;
		  }

860
861
862
863
		// If we already know a state with the same
		// successors, use it in lieu of the current one.
		if (symb_merge)
		  dest = fc.canonize(dest);
864

865
866
867
868
869
870
		// If we are not postponing the branching, we can
		// declare the outgoing transitions immediately.
		// Otherwise, we merge transitions with identical
		// label, and declare the outgoing transitions in a
		// second loop.
		if (!branching_postponement)
871
		  {
872
		    fill_dests(d, dests, label, dest);
873
874
875
		  }
		else
		  {
876
877
878
879
		    succ_map::iterator si = succs.find(label);
		    if (si == succs.end())
		      succs[label] = dest;
		    else
880
881
882
883
		      si->second =
			multop::instance(multop::Or,
					 const_cast<formula*>(si->second),
					 const_cast<formula*>(dest));
884
885
		  }
	      }
886
887
888
	    if (branching_postponement)
	      for (succ_map::const_iterator si = succs.begin();
		   si != succs.end(); ++si)
889
		fill_dests(d, dests, si->first, si->second);
890
	  }
891

892
	// Check for an arc going to 1 (True).  Register it first, that
893
	// way it will be explored before others during model checking.
894
	dest_map::const_iterator i = dests.find(constant::true_instance());
895
	// COND_FOR_TRUE is the conditions of the True arc, so we
896
897
898
	// can remove them from all other arcs.  It might sounds that
	// this is not needed when exprop is used, but in fact it is
	// complementary.
899
900
	//
	// Consider
901
	//   f = r(X(1) R p) = p.(1 + r(X(1) R p))
902
	// with exprop the two outgoing arcs would be
903
904
        //         p               p
	//     f ----> 1       f ----> f
905
906
	//
	// where in fact we could output
907
908
        //         p
	//     f ----> 1
909
	//
910
	// because there is no point in looping on f if we can go to 1.
911
	bdd cond_for_true = bddfalse;
912
913
	if (i != dests.end())
	  {
914
	    // When translating LTL for an event-based logic with
915
916
	    // unobservable events, the 1 state should accept all events,
	    // even unobservable events.
917
	    if (unobs && now == constant::true_instance())
918
	      cond_for_true = all_events;
919
920
921
922
923
	    else
	      {
		// There should be only one transition going to 1 (true) ...
		assert(i->second.size() == 1);
		prom_map::const_iterator j = i->second.begin();
924
925
926
		// ... and it is not expected to make any promises (unless
		// fair loop approximations are used).
		assert(fair_loop_approx || j->first == bddtrue);
927
928
		cond_for_true = j->second;
	      }
929
930
	    if (!a->has_state(constant::true_instance()))
	      formulae_to_translate.insert(constant::true_instance());
931
	    tgba_explicit::transition* t =
932
	      a->create_transition(now, constant::true_instance());
933
	    a->add_condition(t, d.bdd_to_formula(cond_for_true));
934
935
936
937
938
	  }
	// Register other transitions.
	for (i = dests.begin(); i != dests.end(); ++i)
	  {
	    const formula* dest = i->first;
939
940
941
942
	    // The cond_for_true optimization can cause some
	    // transitions to be removed.  So we have to remember
	    // whether a formula is actually reachable.
	    bool reachable = false;
943

944
945
946
	    // Will this be a new state?
	    bool seen = a->has_state(dest);

947
948
949
950
951
	    if (dest != constant::true_instance())
	      {
		for (prom_map::const_iterator j = i->second.begin();
		     j != i->second.end(); ++j)
		  {
952
953
954
		    bdd cond = j->second - cond_for_true;
		    if (cond == bddfalse) // Skip false transitions.
		      continue;
955
		    tgba_explicit::transition* t =
956
		      a->create_transition(now, dest);
957
		    a->add_condition(t, d.bdd_to_formula(cond));
958
		    d.conj_bdd_to_acc(a, j->first, t);
959
		    reachable = true;
960
961
		  }
	      }
962
963
964
965
966
	    else
	      {
		// "1" is reachable.
		reachable = true;
	      }
967
968
	    if (reachable && !seen)
	      formulae_to_translate.insert(dest);
969
	    else
970
	      dest->destroy();
971
972
973
	  }
      }

974
975
    // Turn all promises into real acceptance conditions.
    a->complement_all_acceptance_conditions();
976
977
978
979
    return a;
  }

}