ltl2tgba_fm.cc 24.7 KB
Newer Older
1
2
3
// Copyright (C) 2003, 2004, 2005, 2006 Laboratoire 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
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
//
// 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.

22
#include "misc/hash.hh"
23
#include "misc/bddalloc.hh"
24
#include "misc/bddlt.hh"
25
#include "misc/minato.hh"
26
27
28
29
30
31
#include "ltlast/visitor.hh"
#include "ltlast/allnodes.hh"
#include "ltlvisit/lunabbrev.hh"
#include "ltlvisit/nenoform.hh"
#include "ltlvisit/destroy.hh"
#include "ltlvisit/tostring.hh"
32
#include "ltlvisit/postfix.hh"
33
#include "ltlvisit/apcollect.hh"
34
#include <cassert>
35
#include <memory>
36
#include "ltl2tgba_fm.hh"
37
#include "ltlvisit/contain.hh"
38
39
40
41
42
43
44
45

namespace spot
{
  using namespace ltl;

  namespace
  {

46
47
    // Helper dictionary.  We represent formulae using BDDs to
    // simplify them, and then translate BDDs back into formulae.
48
49
50
51
52
    //
    // 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.
53
    class translate_dict
54
55
56
    {
    public:

57
58
      translate_dict(bdd_dict* dict)
	: dict(dict),
59
60
61
62
63
64
65
66
67
68
	  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)
69
	  destroy(i->first);
70
	dict->unregister_all_my_variables(this);
71
72
      }

73
74
      bdd_dict* dict;

75
76
      typedef bdd_dict::fv_map fv_map;
      typedef bdd_dict::vf_map vf_map;
77
78
79
80
81
82
83
84
85

      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
86
      register_proposition(const formula* f)
87
      {
88
	int num = dict->register_proposition(f, this);
89
90
91
92
93
	var_set &= bdd_ithvar(num);
	return num;
      }

      int
94
      register_a_variable(const formula* f)
95
      {
96
	int num = dict->register_acceptance_variable(f, this);
97
98
99
100
101
	a_set &= bdd_ithvar(num);
	return num;
      }

      int
102
      register_next_variable(const formula* f)
103
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
	  {
	    f = clone(f);
114
	    num = dict->register_anonymous_variables(1, this);
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
	    next_map[f] = num;
	    next_formula_map[num] = f;
	  }
	next_set &= bdd_ithvar(num);
	return num;
      }

      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;
	}
132
133
	os << "Shared Dict:" << std::endl;
	dict->dump(os);
134
135
136
	return os;
      }

137
      formula*
138
139
140
141
      var_to_formula(int var) const
      {
	vf_map::const_iterator isi = next_formula_map.find(var);
	if (isi != next_formula_map.end())
142
	  return clone(isi->second);
143
144
	isi = dict->acc_formula_map.find(var);
	if (isi != dict->acc_formula_map.end())
145
	  return clone(isi->second);
146
147
	isi = dict->var_formula_map.find(var);
	if (isi != dict->var_formula_map.end())
148
	  return clone(isi->second);
149
	assert(0);
150
151
152
	// Never reached, but some GCC versions complain about
	// a missing return otherwise.
	return 0;
153
154
      }

155
      formula*
156
157
158
      conj_bdd_to_formula(bdd b)
      {
	if (b == bddfalse)
159
160
	  return constant::false_instance();
	multop::vec* v = new multop::vec;
161
162
163
	while (b != bddtrue)
	  {
	    int var = bdd_var(b);
164
	    formula* res = var_to_formula(var);
165
166
167
	    bdd high = bdd_high(b);
	    if (high == bddfalse)
	      {
168
		res = unop::instance(unop::Not, res);
169
170
171
172
		b = bdd_low(b);
	      }
	    else
	      {
173
		assert(bdd_low(b) == bddfalse);
174
175
176
177
178
		b = high;
	      }
	    assert(b != bddfalse);
	    v->push_back(res);
	  }
179
	return multop::instance(multop::And, v);
180
181
      }

182
183
      const formula*
      bdd_to_formula(bdd f)
184
      {
185
	if (f == bddfalse)
186
	  return constant::false_instance();
187

188
189
190
191
192
193
194
195
196
	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);
      }
197
198
199
200
201
202
203
204
205
206
207

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

215
		if (!a->has_acceptance_condition(ac))
216
217
		  a->declare_acceptance_condition(clone(ac));
		a->add_acceptance_condition(t, ac);
218
219
220
221
222
223
224
225
		b = high;
	      }
	    assert(b != bddfalse);
	  }
      }
    };


226
227
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

    // 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_;
    };


270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
    // 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()
      {
      }

285
286
      bdd
      result() const
287
288
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
      {
	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)
321
322
323
	      const formula* child = node->child();
	      bdd y = recurse(child);
	      int a = dict_.register_a_variable(child);
324
325
326
327
328
329
	      int x = dict_.register_next_variable(node);
	      res_ = y | (bdd_ithvar(a) & bdd_ithvar(x));
	      return;
	    }
	  case unop::G:
	    {
330
331
332
333
334
335
336
337
338
339
340
341
	      // 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)
342
	      const formula* child = node->child();
343
	      int x = dict_.register_next_variable(node);
344
345
	      bdd y = recurse(child);
	      res_ = y & bdd_ithvar(x);
346
347
348
349
	      return;
	    }
	  case unop::Not:
	    {
350
	      // r(!y) = !r(y)
351
352
353
354
355
	      res_ = bdd_not(recurse(node->child()));
	      return;
	    }
	  case unop::X:
	    {
356
	      // r(Xy) = Next[y]
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
	      int x = dict_.register_next_variable(node->child());
	      res_ = bdd_ithvar(x);
	      return;
	    }
	  }
	/* Unreachable code.  */
	assert(0);
      }

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

	switch (node->op())
	  {
374
	    // r(f1 logical-op f2) = r(f1) logical-op r(f2)
375
376
377
378
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
408
409
410
411
412
413
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
	  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);
      }

      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_;
    };

441
442
443
444
445
446
447
448
449
450
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
513
514
515
516
517
518
519
520
521
522
523
524

    // 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);
      }

      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)
      {
525
526
	pfl_map::const_iterator i = pfl_.find(f);
	if (i != pfl_.end())
527
528
529
530
	  return i->second;
	ltl_possible_fair_loop_visitor v;
	f->accept(v);
	bool rel = v.result();
531
	pfl_[f] = rel;
532
533
534
535
	return rel;
      }

    private:
536
      typedef Sgi::hash_map<const formula*, bool, formula_ptr_hash> pfl_map;
537
      pfl_map pfl_;
538
539
    };

540
541
542
    class formula_canonizer
    {
    public:
543
      formula_canonizer(translate_dict& d,
544
545
			bool fair_loop_approx, bdd all_promises,
			language_containment_checker* lcc)
546
547
	: v_(d),
	  fair_loop_approx_(fair_loop_approx),
548
549
	  all_promises_(all_promises),
	  lcc_(lcc)
550
551
552
553
554
      {
	// For cosmetics, register 1 initially, so the algorithm will
	// not register an equivalent formula first.
	b2f_[bddtrue] = constant::true_instance();
      }
555

556
557
      ~formula_canonizer()
      {
558
	while (!f2b_.empty())
559
	  {
560
561
562
563
	    formula_to_bdd_map::iterator i = f2b_.begin();
	    const formula* f = i->first;
	    f2b_.erase(i);
	    destroy(f);
564
	  }
565
566
567
      }

      bdd
568
      translate(const formula* f, bool* new_flag = 0)
569
570
571
572
573
574
      {
	// Use the cached result if available.
	formula_to_bdd_map::const_iterator i = f2b_.find(f);
	if (i != f2b_.end())
	  return i->second;

575
576
577
	if (new_flag)
	  *new_flag = true;

578
579
580
	// Perform the actual translation.
	f->accept(v_);
	bdd res = v_.result();
581
582
583
584
585
586
587
588
589
590
591
592

	// 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_;
	  }

593
594
595
596
597
598
599
600
601
602
603
	f2b_[clone(f)] = res;

	// 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)
      {
604
605
	bool new_variable = false;
	bdd b = translate(f, &new_variable);
606
607

	bdd_to_formula_map::iterator i = b2f_.find(b);
608
609
	// Since we have just translated the formula, it is
	// necessarily in b2f_.
610
611
612
	assert(i != b2f_.end());

	if (i->second != f)
613
	  {
614
	    // The translated bdd maps to an already seen formula.
615
616
	    destroy(f);
	    f = clone(i->second);
617
	  }
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
	else if (new_variable && lcc_)
	  {
	    // It's a new bdd for a new formula.  Let's see if we can
	    // find an equivalent formula with language containment
	    // checks.
	    for (formula_to_bdd_map::const_iterator j = f2b_.begin();
		 j != f2b_.end(); ++j)
	      if (f != j->first && lcc_->equal(f, j->first))
		{
		  f2b_[f] = j->second;
		  i->second = j->first;
		  destroy(f);
		  f = clone(i->second);
		  break;
		}
	  }
634
	return f;
635
636
      }

637
638
639
640
641
642
643
644
645
646
647
648
649
    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_;
650
651
652
653

      possible_fair_loop_checker pflc_;
      bool fair_loop_approx_;
      bdd all_promises_;
654
      language_containment_checker* lcc_;
655
656
657
658
659
    };

  }

  typedef std::map<bdd, bdd, bdd_less_than> prom_map;
660
  typedef Sgi::hash_map<const formula*, prom_map, formula_ptr_hash> dest_map;
661
662
663
664

  static void
  fill_dests(translate_dict& d, dest_map& dests, bdd label, const formula* dest)
  {
665
    bdd conds = bdd_existcomp(label, d.var_set);
666
667
    bdd promises = bdd_existcomp(label, d.a_set);

668
669
670
671
672
673
674
675
676
677
678
679
680
    dest_map::iterator i = dests.find(dest);
    if (i == dests.end())
      {
	dests[dest][promises] = conds;
      }
    else
      {
	i->second[promises] |= conds;
	destroy(dest);
      }
  }


681
  tgba_explicit*
682
  ltl_to_tgba_fm(const formula* f, bdd_dict* dict,
683
		 bool exprop, bool symb_merge, bool branching_postponement,
684
		 bool fair_loop_approx, const atomic_prop_set* unobs,
685
		 int reduce_ltl, bool containment_checks)
686
  {
687
688
    symb_merge |= containment_checks;

689
690
691
692
    // 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.
693
694
695
    formula* f1 = unabbreviate_logic(f);
    formula* f2 = negative_normal_form(f1);
    destroy(f1);
696

697
698
699
700
701
702
703
704
    // Simplify the formula, if requested.
    if (reduce_ltl)
      {
	formula* tmp = reduce(f2, reduce_ltl);
	destroy(f2);
	f2 = tmp;
      }

705
706
707
    typedef std::set<const formula*, formula_ptr_less_than> set_type;
    set_type formulae_seen;
    set_type formulae_to_translate;
708

709
    translate_dict d(dict);
710

711
712
    // Compute the set of all promises that can possibly occurre
    // inside the formula.
713
    bdd all_promises = bddtrue;
714
    if (fair_loop_approx || unobs)
715
716
717
718
719
720
      {
	ltl_promise_visitor pv(d);
	f2->accept(pv);
	all_promises = pv.result();
      }

721
722
723
724
725
726
    language_containment_checker lcc(dict, exprop, symb_merge,
				     branching_postponement,
				     fair_loop_approx);

    formula_canonizer fc(d, fair_loop_approx, all_promises,
			 containment_checks ? &lcc : 0);
727

728
729
730
731
732
733
734
735
736
737
    // 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;
738
739
	std::auto_ptr<atomic_prop_set> aps(atomic_prop_collect(f));
	for (atomic_prop_set::const_iterator i = aps->begin();
740
741
742
743
744
745
746
747
	     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;
	  }
748
	for (atomic_prop_set::const_iterator i = unobs->begin();
749
750
751
752
753
754
755
756
757
758
759
760
761
	     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;

762
763
764
765
766
767
768
769
770
771
    formulae_seen.insert(f2);
    formulae_to_translate.insert(f2);

    tgba_explicit* a = new tgba_explicit(dict);

    a->set_init_state(to_string(f2));

    while (!formulae_to_translate.empty())
      {
	// Pick one formula.
772
	const formula* f = *formulae_to_translate.begin();
773
774
775
	formulae_to_translate.erase(formulae_to_translate.begin());

	// Translate it into a BDD to simplify it.
776
	bdd res = fc.translate(f);
777

778
779
780
781
782
783
784
785
	// Handle exclusive events.
	if (unobs)
	  {
	    res &= observable_events;
	    int n = d.register_next_variable(f);
	    res |= unobservable_events & bdd_ithvar(n) & all_promises;
	  }

786
787
	std::string now = to_string(f);

788
789
790
791
792
793
794
795
796
797
798
799
800
801
802
	// 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
803
804
805
	// 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.
806
	//
807
	// Note that this is still not optimal.  For instance it is
808
	// better to encode `f U g' as
809
	//     r(f U g) = g + a(g).r(X(f U g)).f.!g
810
811
812
813
	// 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.
814
815
816
	//
	// In `exprop' mode, considering all possible combinations of
	// outgoing propositions generalizes the above trick.
817
	dest_map dests;
818

819
	// Compute all outgoing arcs.
820
821
822
823
824

	// 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.
825
	bdd var_set = bdd_existcomp(bdd_support(res), d.var_set);
826
827
828
829
830
	// 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);
831
	while (all_props != bddfalse)
832
	  {
833
834
835
	    bdd one_prop_set =
	      exprop ? bdd_satoneset(all_props, var_set, bddtrue) : bddtrue;
	    all_props -= one_prop_set;
836

837
	    typedef std::map<bdd, const formula*, bdd_less_than> succ_map;
838
839
	    succ_map succs;

840
841
842
	    minato_isop isop(res & one_prop_set);
	    bdd cube;
	    while ((cube = isop.next()) != bddfalse)
843
	      {
844
		bdd label = bdd_exist(cube, d.next_set);
845
		bdd dest_bdd = bdd_existcomp(cube, d.next_set);
846
847
		const formula* dest = d.conj_bdd_to_formula(dest_bdd);

848
849
850
851
852
853
854
855
856
857
858
		// Simplify the formula, if requested.
		if (reduce_ltl)
		  {
		    formula* tmp = reduce(dest, reduce_ltl);
		    destroy(dest);
		    dest = tmp;
		    // Ignore the arc if the destination reduces to false.
		    if (dest == constant::false_instance())
		      continue;
		  }

859
860
861
862
		// 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);
863

864
865
866
867
868
869
		// 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)
870
		  {
871
		    fill_dests(d, dests, label, dest);
872
873
874
		  }
		else
		  {
875
876
877
878
		    succ_map::iterator si = succs.find(label);
		    if (si == succs.end())
		      succs[label] = dest;
		    else
879
880
881
882
		      si->second =
			multop::instance(multop::Or,
					 const_cast<formula*>(si->second),
					 const_cast<formula*>(dest));
883
884
		  }
	      }
885
886
887
	    if (branching_postponement)
	      for (succ_map::const_iterator si = succs.begin();
		   si != succs.end(); ++si)
888
		fill_dests(d, dests, si->first, si->second);
889
	  }
890

891
	// Check for an arc going to 1 (True).  Register it first, that
892
893
894
	// way it will be explored before the other during the model
	// checking.
	dest_map::const_iterator i = dests.find(constant::true_instance());
895
896
897
898
	// COND_FOR_TRUE is the conditions of the True arc, so when
	// 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 ----------> 1
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
918
	    if (unobs && f == constant::true_instance())
	      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
	    tgba_explicit::transition* t =
	      a->create_transition(now, constant::true_instance()->val_name());
931
	    a->add_condition(t, d.bdd_to_formula(cond_for_true));
932
933
934
935
936
	  }
	// Register other transitions.
	for (i = dests.begin(); i != dests.end(); ++i)
	  {
	    const formula* dest = i->first;
937
938
939
940
	    // 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;
941

942
943
944
945
946
947
	    if (dest != constant::true_instance())
	      {
		std::string next = to_string(dest);
		for (prom_map::const_iterator j = i->second.begin();
		     j != i->second.end(); ++j)
		  {
948
949
950
		    bdd cond = j->second - cond_for_true;
		    if (cond == bddfalse) // Skip false transitions.
		      continue;
951
952
		    tgba_explicit::transition* t =
		      a->create_transition(now, next);
953
		    a->add_condition(t, d.bdd_to_formula(cond));
954
		    d.conj_bdd_to_acc(a, j->first, t);
955
		    reachable = true;
956
957
		  }
	      }
958
959
960
961
962
963
964
	    else
	      {
		// "1" is reachable.
		reachable = true;
	      }
	    if (reachable
		&& formulae_seen.find(dest) == formulae_seen.end())
965
966
967
968
969
970
	      {
		formulae_seen.insert(dest);
		formulae_to_translate.insert(dest);
	      }
	    else
	      {
971
		destroy(dest);
972
973
974
975
976
	      }
	  }
      }

    // Free all formulae.
977
    for (std::set<const formula*>::iterator i = formulae_seen.begin();
978
	 i != formulae_seen.end(); ++i)
979
      destroy(*i);
980

981
982
    // Turn all promises into real acceptance conditions.
    a->complement_all_acceptance_conditions();
983
984
985
986
    return a;
  }

}