Relabeling Formulas
Table of Contents
The task is to read an LTL formula, relabel all (possibly
double-quoted) atomic propositions, and provide #define statements
for each of these renamings, writing everything in Spin's syntax.
Shell
ltlfilt -ps --relabel=pnn --define -f '"Proc@Here" U ("var > 10" | "var < 4")'
#define p0 (Proc@Here) #define p1 (var < 4) #define p2 (var > 10) (p0) U ((p1) || (p2))
When is this output interesting, you may ask? It is useful for
instance if you want to call ltl2ba (or any other LTL-to-Büchi
translator) using a formula with complex atomic propositions it cannot
parse. Then you can pass the rewritten formula to ltl2ba, and
prepend all those #define to its output. For instance:
ltlfilt -ps --relabel=pnn --define=tmp.defs -f '"Proc@Here" U ("var > 10" | "var < 4")' >tmp.ltl
cat tmp.defs; ltl2ba -F tmp.ltl
rm tmp.defs tmp.ltl
#define p0 (Proc@Here)
#define p1 (var < 4)
#define p2 (var > 10)
never { /* (p0) U ((p1) || (p2))
*/
T0_init:
if
:: (p0) -> goto T0_init
:: (p1) || (p2) -> goto accept_all
fi;
accept_all:
skip
}
Aside: another way to work around syntax limitations of tools is to
use ltldo. On the above example, ltldo ltl2ba -f '"Proc@Here" U
("var > 10" | "var < 4")' -s would produce a never clam with the
correct atomic proposition, even though ltl2ba cannot parse them.
Python
The spot.relabel function takes an optional third parameter that
should be a relabeling_map. If supplied, this map is filled with
pairs of atomic propositions of the form (new-name, old-name).
import spot
f = spot.formula('"Proc@Here" U ("var > 10" | "var < 4")')
m = spot.relabeling_map()
g = spot.relabel(f, spot.Pnn, m)
for newname, oldname in m.items():
print("#define {} ({})".format(newname.to_str(), oldname.to_str('spin', True)))
print(g.to_str('spin', True))
C++
The spot::relabeling_map is just implemented as a std::map.
#include <string>
#include <iostream>
#include <spot/tl/parse.hh>
#include <spot/tl/print.hh>
#include <spot/tl/relabel.hh>
int main()
{
std::string input = "\"Proc@Here\" U (\"var > 10\" | \"var < 4\")";
spot::parsed_formula pf = spot::parse_infix_psl(input);
if (pf.format_errors(std::cerr))
return 1;
spot::formula f = pf.f;
spot::relabeling_map m;
f = spot::relabel(f, spot::Pnn, &m);
for (auto& i: m)
{
std::cout << "#define " << i.first << " (";
print_spin_ltl(std::cout, i.second, true) << ")\n";
}
print_spin_ltl(std::cout, f, true) << '\n';
return 0;
}
#define p0 (Proc@Here) #define p1 (var < 4) #define p2 (var > 10) (p0) U ((p1) || (p2))
Additional comments
Two ways to name atomic propositions
Instead of --relabel=pnn (or spot.Pnn, or spot::Pnn), you can
actually use --relabel=abc (or spot.Abc, or spot::Abc) to have
the atomic propositions named a, b, c, etc.
Relabeling Boolean sub-expressions
Instead of relabeling each atomic proposition, you could decide to relabel each Boolean sub-expression:
ltlfilt -ps --relabel-bool=pnn --define -f '"Proc@Here" U ("var > 10" | "var < 4")'
#define p0 (Proc@Here) #define p1 ((var < 4) || (var > 10)) (p0) U (p1)
The relabeling routine is smart enough to not give different names to Boolean expressions that have some sub-expression in common.
For instance a U (a & b) will not be relabeled into (p0) U (p1)
because that would hide the fact that both p0 and p1 check for
a. Instead we get this:
ltlfilt -ps --relabel-bool=pnn --define -f 'a U (a & b)'
#define p0 (a) #define p1 (b) (p0) U ((p0) && (p1))
This "Boolean sub-expression" relabeling is available in Python and
C++ via the relabel_bse function. The interface is identical to
relabel.