The ltl_to_tgba_fm() translation function was using a hash_map of
maps (ugh!) to merge transitions on output. However recent libstd++
changed the implementation of hash_map (a.k.a. unordered_map) causing
transitions to be output in a different order. This
implementation-dependent order caused the ltl2ta.test to fail because
the BA->TA transformation can produce TA of different sizes if you
simply change the order of transitions in the input BA! This does not
sound like a nice property for the BA->TA transformation, but Ala Eddine
isn't sure how to fix it yet. In the meantime, this patch makes sure
ltl_to_tgba_fm() will return the same output regardless of the
implementation of hash_map.
The ltl2ta.test failure has been observed with g++ 4.9.2 on Arch Linux,
and with gcc-snapshot (5.0.0 20141016) on Debian.
* src/tgbaalgos/ltl2tgba_fm.cc: Rewrite the transition merging
using a std::vector and std::sort instead of nested maps tables.
* src/tgbatest/ltl2ta.test: Adjust sizes to the new order.
* NEWS: Mention the fix.
* src/ltlast/formula.hh: Specialize std::hash<>.
* src/ltlvisit/contain.hh, src/ltlvisit/relabel.cc,
src/tgba/taatgba.hh, src/tgbaalgos/ltl2tgba_fm.cc: Do
not pass formula_ptr_hash to unordered_map.
This is a huge patch, that took over a month to complete. The bit sets
are currently restricted to what 'unsigned can store', but it should be
easy to extend it to 'uint64_t' should we need it.
* NEWS: Update.
* src/tgba/acc.hh: New file.
* src/tgbatest/acc.cc, src/tgbatest/acc.test: Test it.
* src/tgba/tgbakvcomplement.cc, src/tgba/tgbakvcomplement.hh,
src/tgba/tgbasgba.cc, src/tgba/tgbasgba.hh: Delete. The KV
complementation is too slow to be used in practice, and I somehow broke
it during the conversion to bitsets. The tgba->sgba conversion was only
used for the KV complementation, and should be better redone on
tgba_digraph_ptr should it be needed again.
* src/bin/ltlcross.cc, src/dstarparse/dra2ba.cc,
src/dstarparse/nsa2tgba.cc, src/graphtest/tgbagraph.cc,
src/graphtest/tgbagraph.test, src/kripke/fairkripke.cc,
src/kripke/fairkripke.hh, src/kripke/kripke.cc, src/kripke/kripke.hh,
src/kripke/kripkeexplicit.cc, src/kripke/kripkeexplicit.hh,
src/misc/hash.hh, src/neverparse/neverclaimparse.yy, src/priv/accmap.hh,
src/ta/ta.cc, src/ta/ta.hh, src/ta/taexplicit.cc, src/ta/taexplicit.hh,
src/ta/taproduct.cc, src/ta/taproduct.hh, src/ta/tgta.cc,
src/ta/tgta.hh, src/ta/tgtaexplicit.cc, src/ta/tgtaexplicit.hh,
src/ta/tgtaproduct.cc, src/ta/tgtaproduct.hh, src/taalgos/dotty.cc,
src/taalgos/emptinessta.cc, src/taalgos/minimize.cc,
src/taalgos/tgba2ta.cc, src/tgba/Makefile.am, src/tgba/fwd.hh,
src/tgba/taatgba.cc, src/tgba/taatgba.hh, src/tgba/tgba.cc,
src/tgba/tgba.hh, src/tgba/tgbagraph.cc, src/tgba/tgbagraph.hh,
src/tgba/tgbamask.cc, src/tgba/tgbamask.hh, src/tgba/tgbaproduct.cc,
src/tgba/tgbaproduct.hh, src/tgba/tgbaproxy.cc, src/tgba/tgbaproxy.hh,
src/tgba/tgbasafracomplement.cc, src/tgba/tgbasafracomplement.hh,
src/tgbaalgos/bfssteps.cc, src/tgbaalgos/complete.cc,
src/tgbaalgos/compsusp.cc, src/tgbaalgos/degen.cc,
src/tgbaalgos/dotty.cc, src/tgbaalgos/dtbasat.cc,
src/tgbaalgos/dtgbacomp.cc, src/tgbaalgos/dtgbasat.cc,
src/tgbaalgos/dupexp.cc, src/tgbaalgos/emptiness.cc,
src/tgbaalgos/emptiness.hh, src/tgbaalgos/gtec/ce.cc,
src/tgbaalgos/gtec/gtec.cc, src/tgbaalgos/gtec/gtec.hh,
src/tgbaalgos/gtec/sccstack.cc, src/tgbaalgos/gtec/sccstack.hh,
src/tgbaalgos/gv04.cc, src/tgbaalgos/hoaf.cc,
src/tgbaalgos/isweakscc.cc, src/tgbaalgos/lbtt.cc,
src/tgbaalgos/ltl2tgba_fm.cc, src/tgbaalgos/magic.cc,
src/tgbaalgos/ndfs_result.hxx, src/tgbaalgos/neverclaim.cc,
src/tgbaalgos/postproc.cc, src/tgbaalgos/powerset.cc,
src/tgbaalgos/randomgraph.cc, src/tgbaalgos/randomgraph.hh,
src/tgbaalgos/reducerun.cc, src/tgbaalgos/replayrun.cc,
src/tgbaalgos/safety.cc, src/tgbaalgos/save.cc, src/tgbaalgos/scc.cc,
src/tgbaalgos/scc.hh, src/tgbaalgos/sccfilter.cc,
src/tgbaalgos/sccinfo.cc, src/tgbaalgos/sccinfo.hh,
src/tgbaalgos/se05.cc, src/tgbaalgos/simulation.cc,
src/tgbaalgos/simulation.hh, src/tgbaalgos/stats.cc,
src/tgbaalgos/stripacc.cc, src/tgbaalgos/tau03.cc,
src/tgbaalgos/tau03opt.cc, src/tgbaalgos/weight.cc,
src/tgbaalgos/weight.hh, src/tgbaparse/tgbaparse.yy,
src/tgbatest/Makefile.am, src/tgbatest/complementation.cc,
src/tgbatest/complementation.test, src/tgbatest/degenlskip.test,
src/tgbatest/det.test, src/tgbatest/dstar.test, src/tgbatest/emptchk.cc,
src/tgbatest/explpro2.test, src/tgbatest/explpro3.test,
src/tgbatest/explpro4.test, src/tgbatest/explprod.test,
src/tgbatest/ltl2tgba.cc, src/tgbatest/ltl2tgba.test,
src/tgbatest/maskacc.cc, src/tgbatest/maskacc.test,
src/tgbatest/neverclaimread.test, src/tgbatest/randtgba.cc,
src/tgbatest/readsave.test, src/tgbatest/sim.test,
src/tgbatest/sim2.test, src/tgbatest/spotlbtt.test,
src/tgbatest/tgbaread.test, src/tgbatest/tripprod.test,
iface/dve2/dve2.cc: Adjust or use to the new acceptance interface.
The bug was reported by Joachim Klein.
* src/tgbaalgos/ltl2tgba_fm.cc (translate_dict::register_a_variable):
Reduce P(a M b) to P(a & b), not to P(a).
* src/tgbatest/ltlcross.test: Test Joachim's formula.
* src/tgbatest/ltl2ta.test: Adjust some expected values.
* NEWS: Mention the bug.
The conversion is not complete, because the conversion from SERE to DRA
used for the closure operator is still building a tgba_explicit_formula.
* src/tgbaalgos/ltl2tgba_fm.cc, src/tgbaalgos/ltl2tgba_fm.hh: Return
a tgba_digraph.
* src/priv/acccompl.cc: Simplify.
* src/graph/ngraph.hh: Add a way to iterate over all names.
* src/tgba/tgbagraph.hh (compute_support_conditions): Return something
useful. It's actually used by the constructor of testing automata.
* src/tgbatest/wdba.test: Adjust to the fact that state are not
labeled by formulas anymore.
* src/bin/ltl2tgba.cc, src/bin/ltl2tgta.cc: Do not try to enable
UTF8 on automata anymore.
This was only used in ELTL stuff, which I just removed because it was
unused.
* src/ltlast/automatop.cc, src/ltlast/automatop.hh,
src/ltlast/formula_tree.cc, src/ltlast/formula_tree.hh,
src/ltlast/nfa.cc, src/ltlast/nfa.hh: Delete.
* src/ltlast/Makefile.am: Adjust.
* src/ltlast/allnodes.hh, src/ltlast/formula.hh, src/ltlast/predecl.hh,
src/ltlast/visitor.hh, src/ltltest/equals.cc, src/ltltest/ltlrel.cc,
src/ltltest/reduc.cc, src/ltlvisit/clone.cc, src/ltlvisit/clone.hh,
src/ltlvisit/dotty.cc, src/ltlvisit/lbt.cc, src/ltlvisit/mark.cc,
src/ltlvisit/postfix.cc, src/ltlvisit/postfix.hh,
src/ltlvisit/relabel.cc, src/ltlvisit/simplify.cc, src/ltlvisit/snf.cc,
src/ltlvisit/tostring.cc, src/tgba/formula2bdd.cc,
src/tgbaalgos/ltl2taa.cc, src/tgbaalgos/ltl2tgba_fm.cc,
src/tgbaalgos/ltl2tgba_lacim.cc, src/tgbatest/ltl2tgba.cc,
iface/dve2/dve2check.cc: Remove all references to automatop.
Instead of "delete iter;" we now do "aut->release_iter(iter);" to
give the iterator back to the automaton. The TGBA classes now
reuse a previously returned tgba_succ_iterator to answer a succ_iter()
call, therefore avoiding (1) memory allocation, as well as (2) vtable
and other constant member initialization.
* src/tgba/tgba.hh, src/tgba/tgba.cc (release_iter, iter_cache_):
Implement a release_iter() that stores the released iterator
in iter_cache_.
* src/tgba/succiter.hh (internal::succ_iterable): Move...
* src/tgba/tgba.hh (tgba::succ_iterable): ... here. And use
release_iter().
* iface/dve2/dve2.cc, src/kripke/kripke.cc, src/kripke/kripke.hh,
src/tgba/succiterconcrete.cc, src/tgba/succiterconcrete.hh,
src/tgba/taatgba.hh, src/tgba/tgbabddconcrete.cc,
src/tgba/tgbaexplicit.hh, src/tgba/tgbamask.cc, src/tgba/tgbaproduct.cc,
src/tgba/tgbaproxy.cc, src/tgba/tgbascc.cc, src/tgba/tgbatba.cc,
src/tgba/tgbaunion.cc, src/tgba/tgbaunion.hh, src/tgba/wdbacomp.cc,
src/tgbaalgos/bfssteps.cc, src/tgbaalgos/compsusp.cc,
src/tgbaalgos/cycles.cc, src/tgbaalgos/dtbasat.cc,
src/tgbaalgos/dtgbasat.cc, src/tgbaalgos/gtec/gtec.cc,
src/tgbaalgos/gv04.cc, src/tgbaalgos/isweakscc.cc,
src/tgbaalgos/lbtt.cc, src/tgbaalgos/ltl2tgba_fm.cc,
src/tgbaalgos/magic.cc, src/tgbaalgos/ndfs_result.hxx,
src/tgbaalgos/neverclaim.cc, src/tgbaalgos/reachiter.cc,
src/tgbaalgos/replayrun.cc, src/tgbaalgos/safety.cc,
src/tgbaalgos/scc.cc, src/tgbaalgos/se05.cc,
src/tgbaalgos/simulation.cc, src/tgbaalgos/tau03.cc,
src/tgbaalgos/tau03opt.cc: Use release_iter() instead of deleting
iterators, and used recycle iter_cache_ in implementations of
tgba::succ_iter().
This follows from a discussion with Ernesto Posse.
The semantics for the {...} operator we use in Spot comes from the
cl(...) operator defined by Dax et al. (ATVA'09). This is slightly
different from the the way the PSL spec interprets a SERE used in the
context of a temporal formula (appendix B.3.1.1.2, item 7).
cl({a;b}[*]) would match any infinite word that starts with a;b, while
in PSL {a;b}[*] would match any infinite word that alternates a and b.
Spot documents that {SERE} in a temporal formula is interpreted like
cl(SERE) however it failed to ignore the empty prefix of SERE. So
{{a;b}[*]} would match anything, because the empty word is a prefix of
any word, and is also accepted by {a;b}[*]. Some trivial identities
and basic rewritings were also wrongly considering these empty
prefixes as well.
This patch therefore fixes the translation and syntactic
simplification rules, to really ignore these empty prefixes.
In some future version it should probably be wise to rename this {...}
operator as cl(...), and use {...} for the semantics given in appendix
B.3.1.1.2 (item 7) of the PSL specs.
* src/ltlast/unop.cc: Fix trivial identities. We have
{[*0]} = 0 and !{[*0]} = 1.
* src/ltlvisit/simplify.cc: Fix basic rewriting rules.
{e[*]} = {e} and !{e[*]} = !{e}.
* doc/tl/tl.tex: Adjust documentation.
* doc/tl/tl.bib (dax.09.atva): New entry.
* src/tgbaalgos/ltl2tgba_fm.cc: Do not accept any
infinite word for {e[*]} just because the empty
prefix is matched by e[*].
* src/tgbatest/ltl2tgba.test: Add a test case.
* NEWS: Mention it.
* THANKS: Add Ernesto.
This helps reducing (p&XF!p)|(!p&XFp)|X(Fp&F!p) to (p&XF!p)|(!p&XFp).
* src/tgbaalgos/ltl2tgba_fm.cc: Adjust rewriting rules of X.
* src/tgbatest/ltl2tgba.test: Add a test case.
* src/tgba/bdddict.hh, src/tgba/bdddict.cc: Store variable types and
associated formula in a vector indexed by BDD variable numbers,
instead of using several maps.
* src/evtgbaalgos/tgba2evtgba.cc, src/tgba/bddprint.cc,
src/tgba/formula2bdd.cc, src/tgbaalgos/ltl2tgba_fm.cc,
src/tgbaalgos/save.cc: Adjust usage.
* src/tgbaalgos/ltl2tgba_fm.cc (implied_subforfmulae): New function.
(ltl_trad_visitor::visit(const binop*)): Use it. This is an attempt
to correct the unoptimal translation of 'Fa & GFa' left by previous
patch. It still fails to optimize 'Fa & GF(a&b)', but this is not a
regression compared to previous version.
* src/tgbaalgos/ltl2tgba_fm.cc: Add a "recurring" mode for the
translation of the child of G. This generalizes Couvreur's original
trick to translate GFa as (a|Prom(a))&X(GFa) without generating an
intermediate GF(a)&F(a) state that would have to be merged with GF(a)
latter. This implementation will also speedup formulas such a G((a U
b) & (c M d)). With this patch, translating GF(p1) & GF(p2) &
... GF(p20) into a TGBA takes 57s instead of 128s on my computer.
However it alsos causes some formulas to be translated into larger
automata that are not immediately reduced (the simulation-reduction is
needed). For instance Fa & GFa now has a different signature than
GFa, so translating 'Fa & GFa' generates two states where is used to
generate only one.
We need a marked version of !{r} to perform breakpoint unroling.
* src/ltlast/unop.cc, src/ltlast/unop.hh: Declare a NegClosureMarked
operator.
* src/ltlvisit/mark.hh, src/ltlvisit/mark.cc,
src/tgbaalgos/ltl2tgba_fm.cc: Adjust to deal with NegClosureMarked
and NegClosure as apropriate.
* src/ltlvisit/simplify.cc, src/ltlvisit/tostring.cc,
src/ltlvisit/tunabbrev.cc, src/tgbaalgos/eltl2tgba_lacim.cc,
src/tgbaalgos/ltl2taa.cc, src/tgbaalgos/ltl2tgba_lacim.cc,
src/tgba/formula2bdd.cc: Deal with NegClosureMarked in the same way as
we deal with NegClosure.
* src/tgbatest/ltl2tgba.test: More tests.
* src/ltltest/kind.test: Adjust.
* doc/tl/tl.tex: Mention the marked negated closure.
It was a mistake to try to overload And/Or LTL operator for these when
trivial simplification are performed. The reason is so simple it is
embarassing: And(f,1)=f is a trivial identity that should not be
applied with AndRat. E.g. AndRat(a;b, 1) is equal to 0, not a;b.
* src/ltlast/multop.hh, src/ltlast/multop.cc: Add the AndRat and OrRat
operators.
* src/ltlparse/ltlparse.yy: Build them.
* src/ltlvisit/mark.cc, src/ltlvisit/simplify.cc,
src/ltlvisit/tostring.cc, src/tgba/formula2bdd.cc,
src/tgbaalgos/eltl2tgba_lacim.cc, src/tgbaalgos/ltl2taa.cc,
src/tgbaalgos/ltl2tgba_fm.cc, src/tgbaalgos/ltl2tgba_lacim.cc:
Adjust all switches.
* src/ltlast/bunop.hh, src/ltlast/bunop.cc, src/ltlvisit/randomltl.cc,
src/ltlvisit/simplify.cc, src/ltlvisit/tostring.cc,
src/tgbaalgos/ltl2tgba_fm.cc: Remove all traces of these two
operators since they are not handled like sugar.
* doc/tl/tl.tex: Adjust documentation to reflect the fact that these
two operators are sugar.
* src/ltlvisit/mark.hh, src/ltlvisit/mark.cc (mark_concat_ops,
simplify_mark): Rewrite these two functions as methods of
(mark_tools): this new class.
* src/ltlast/binop.cc, src/ltlast/unop.cc: Adjust computation
of not_marked to ignore marked operators that are not at
the top-level. I.e., something like X(!{a}) is not marked.
* src/tgbaalgos/ltl2tgba_fm.cc (translate_dict::mt): New
instance of mark_tools.
(formula_canonizer::translate) Adjust calls to
mark_concat_ops() and simplify_mark().
* src/tgbaalgos/ltl2tgba_fm.cc (ratexp_to_dfa::translate): Do not
translate a subformula if we have already proved it useless in
a previous rational expression.
* src/tgbatest/ltl2tgba.test: Add an example, although that
test does not ensure the subformula is ignored early in the
translation. I.e., it would still work without the patch.
* src/tgbaalgos/ltl2tgba_fm.cc (ratexp_to_dfa::translate,
ltl_to_tgba_fm): Do not convert labels as Boolean formulas before
creating transitions. Use the bdd directly, and register the used
transitions later.
We used to cache it only for formulas used as states.
* src/tgbaalgos/ltl2tgba_fm.cc (translate_dict::ltl_to_bdd): New method.
(ltl_trad_visitor::recurse): Use ltl_to_bdd().
(formula_canonizer): Likewise.
(ltl_to_tgba_fm): Adjust.
This fixes a bug where {(a&!a)[=2]} was translated either into an
universal automaton (with simplification turned off) or in an
empty automaton (with simplification turned on).
* src/tgbaalgos/ltl2tgba_fm.cc (ratexp_to_dfa::translate): Trim
the automaton.
(ratexp_to_dfa::succ, ratexp_to_dfa::get_label): Deal with trimed
states.
(ltl_trad_visitor::visit(unop::Closure)): Likewise.
* src/tgbatest/ltl2tgba.test, src/ltltest/reduccmp.test: New test
cases.
This is especially important when translating the Closure
operators, because normally we should only keep the satisfiable
formulae (i.e. co-accessible states), which seems hard to check on
the fly. After this patch we need to teach
ratexp_to_dfa::translate() how to trim (and then minimize) the DFA
to prune those useless (non co-accessible) states.
* src/tgbaalgos/ltl2tgba_fm.cc (ratexp_to_dfa): New class.
(translate_dict::transdfa): New member.
(ltl_trad_visitor::visit(unop::Closure)): Use transdfa.
* src/ltlvisit/simplify.hh, src/ltlvisit/simplify.cc
(ltl_simplifier::ltl_simplifier, ltl_simplifier::get_dict): Make
it possible to supply and retrieve the dictionary used.
(ltl_simplifier::as_bdd): New function, exported from the cache.
* src/tgbaalgos/ltl2tgba_fm.cc (translate_dict): Store the
ltl_simplifier object.
(translate_dict::boolean_to_bdd): Call ltl_simplifier::as_bdd.
(translate_ratexp): New wrapper around the ratexp_trad_visitor,
calling boolean_to_bdd whenever possible.
(ratexp_trad_visitor): Do not deal with negated formulae, there
are necessarily Boolean and handled by translate_ratexp().
(ltl_visitor): Adjust to call translate_ratexp.
(ltl_to_tgba_fm): Adjust passing of the ltl_simplifier to the
translate_dict, and make sure everybody is using the same
dictionary.
* src/tgbatest/ltl2tgba.cc: Pass the dictionary to the
ltl_simplifier.
