CNF allows to transition from a free-form AST (with statements containing formulas as TypedSTerm.t) into an AST using clauses and without some constructs such as "if/then/else" or "match" or "let". The output is more suitable for a Superposition-like prover.
There also are conversion functions to go from clauses that use TypedSTerm.t, into clauses that use the Term.t (hashconsed, and usable in unification, indexing, etc.).
We follow chapter 6 "Computing small clause normal forms" of the "handbook of automated reasoning" for the theoretical part.
A few notes:
In worst case, normal CNF transformation can lead to an exponential number of clauses, which is prohibitive. To avoid that, we use the Tseitin trick to name some intermediate formulas by introducing fresh symbols (herein named "proxies") and defining them to be equivalent to the formula they define.
This is done only if we estimate that adding the proxy will reduce the final number of clauses (See Estimation module).
Before doing CNF we remove all the high-level constructs such as pattern-matching and "let" by introducing new symbols and defining the subterm to eliminate using this new symbol (See Flatten module). It is important to capture variables properly in this phase (as in closure conversion).
type term = TypedSTerm.ttype form = TypedSTerm.ttype type_ = TypedSTerm.ttype lit = term SLiteral.tSee "computing small normal forms", in the handbook of automated reasoning. All transformations are made on curried terms and formulas.
exception NotCNF of formApply miniscoping transformation to the term.
type options = | DistributeExistsif enabled, will distribute existential quantifiers over disjunctions. This can make skolem symbols smaller (smaller arity) but introduce more of them.
*)| DisableRenamingdisables formula renaming. Can re-introduce the worst-case exponential behavior of CNF.
*)| InitialProcessing of form -> formany processing, at the beginning, before CNF starts
*)| PostNNF of form -> formany processing that keeps negation at leaves, just after reduction to NNF. Its output must not break the NNF form (negation at root only).
*)| PostSkolem of form -> formtransformation applied just after skolemization. It must not break skolemization nor NNF (no quantifier, no non-leaf negation).
*)Options are used to tune the behavior of the CNF conversion.
type clause = lit listBasic clause representation, as list of literals
val clause_to_fo : ?ctx:Term.Conv.ctx -> clause -> Term.t SLiteral.t listtype f_statement = (term, term, type_) Statement.tA statement before CNF
type c_statement = (clause, term, type_) Statement.tA statement after CNF
val pp_f_statement : f_statement CCFormat.printerval pp_c_statement : c_statement CCFormat.printerval pp_fo_c_statement :
(Term.t SLiteral.t list, Term.t, Type.t) Statement.t CCFormat.printerval is_clause : form -> boolval is_cnf : form -> boolval cnf_of :
?opts:options list ->
?ctx:Skolem.ctx ->
f_statement ->
c_statement CCVector.ro_vectorTransform the statement into proper CNF; returns a list of statements, including type declarations for new Skolem symbols or formulas proxies. Options are used to tune the behavior.
val cnf_of_seq :
?opts:options list ->
?ctx:Skolem.ctx ->
f_statement Iter.t ->
c_statement CCVector.ro_vectorval type_declarations : c_statement Iter.t -> type_ ID.Map.tCompute the types declared in the statement sequence
val convert : c_statement Iter.t -> Statement.clause_t CCVector.ro_vectorConverts statements based on TypedSTerm into statements based on Term and Type