src/HOL/Tools/Sledgehammer/sledgehammer_fact_filter.ML
author blanchet
Thu Jun 24 10:38:01 2010 +0200 (2010-06-24 ago)
changeset 37538 97ab019d5ac8
parent 37537 8e56d1ccf189
child 37543 2e733b0a963c
permissions -rw-r--r--
make sure that theorems passed using "add:" to Sledgehammer are not eliminated on heuristic grounds
blanchet@35826
     1
(*  Title:      HOL/Tools/Sledgehammer/sledgehammer_fact_filter.ML
wenzelm@33309
     2
    Author:     Jia Meng, Cambridge University Computer Laboratory, NICTA
blanchet@36393
     3
    Author:     Jasmin Blanchette, TU Muenchen
wenzelm@33309
     4
*)
paulson@15452
     5
blanchet@35826
     6
signature SLEDGEHAMMER_FACT_FILTER =
wenzelm@16802
     7
sig
blanchet@37500
     8
  type cnf_thm = Sledgehammer_Fact_Preprocessor.cnf_thm
blanchet@35865
     9
  type classrel_clause = Sledgehammer_FOL_Clause.classrel_clause
blanchet@35865
    10
  type arity_clause = Sledgehammer_FOL_Clause.arity_clause
blanchet@35865
    11
  type hol_clause = Sledgehammer_HOL_Clause.hol_clause
blanchet@36473
    12
blanchet@35966
    13
  type relevance_override =
blanchet@35966
    14
    {add: Facts.ref list,
blanchet@35966
    15
     del: Facts.ref list,
blanchet@35966
    16
     only: bool}
blanchet@35966
    17
blanchet@37505
    18
  val use_natural_form : bool Unsynchronized.ref
blanchet@37344
    19
  val name_thms_pair_from_ref :
blanchet@37344
    20
    Proof.context -> thm list -> Facts.ref -> string * thm list
paulson@22989
    21
  val tvar_classes_of_terms : term list -> string list
paulson@22989
    22
  val tfree_classes_of_terms : term list -> string list
paulson@22989
    23
  val type_consts_of_terms : theory -> term list -> string list
blanchet@37538
    24
  val is_quasi_fol_theorem : theory -> thm -> bool
blanchet@37347
    25
  val relevant_facts :
blanchet@37347
    26
    bool -> bool -> real -> real -> bool -> int -> bool -> relevance_override
blanchet@37500
    27
    -> Proof.context * (thm list * 'a) -> thm list -> cnf_thm list
blanchet@36473
    28
  val prepare_clauses :
blanchet@37500
    29
    bool -> thm list -> cnf_thm list -> cnf_thm list -> theory
blanchet@37500
    30
    -> string vector
blanchet@37500
    31
       * (hol_clause list * hol_clause list * hol_clause list
blanchet@37500
    32
          * hol_clause list * classrel_clause list * arity_clause list)
paulson@15347
    33
end;
paulson@15347
    34
blanchet@35826
    35
structure Sledgehammer_Fact_Filter : SLEDGEHAMMER_FACT_FILTER =
paulson@15347
    36
struct
paulson@15347
    37
blanchet@35865
    38
open Sledgehammer_FOL_Clause
blanchet@35865
    39
open Sledgehammer_Fact_Preprocessor
blanchet@35865
    40
open Sledgehammer_HOL_Clause
blanchet@35826
    41
blanchet@37505
    42
(* Experimental feature: Filter theorems in natural form or in CNF? *)
blanchet@37506
    43
val use_natural_form = Unsynchronized.ref false
blanchet@37505
    44
blanchet@35966
    45
type relevance_override =
blanchet@35966
    46
  {add: Facts.ref list,
blanchet@35966
    47
   del: Facts.ref list,
blanchet@35966
    48
   only: bool}
paulson@21070
    49
wenzelm@28477
    50
(***************************************************************)
wenzelm@28477
    51
(* Relevance Filtering                                         *)
wenzelm@28477
    52
(***************************************************************)
mengj@19194
    53
blanchet@35865
    54
fun strip_Trueprop (@{const Trueprop} $ t) = t
paulson@24958
    55
  | strip_Trueprop t = t;
mengj@19194
    56
paulson@24287
    57
(*** constants with types ***)
paulson@24287
    58
paulson@24287
    59
(*An abstraction of Isabelle types*)
paulson@24287
    60
datatype const_typ =  CTVar | CType of string * const_typ list
paulson@24287
    61
paulson@24287
    62
(*Is the second type an instance of the first one?*)
blanchet@37505
    63
fun match_type (CType(con1,args1)) (CType(con2,args2)) =
paulson@24287
    64
      con1=con2 andalso match_types args1 args2
paulson@24287
    65
  | match_type CTVar _ = true
paulson@24287
    66
  | match_type _ CTVar = false
paulson@24287
    67
and match_types [] [] = true
paulson@24287
    68
  | match_types (a1::as1) (a2::as2) = match_type a1 a2 andalso match_types as1 as2;
paulson@24287
    69
paulson@24287
    70
(*Is there a unifiable constant?*)
blanchet@37505
    71
fun uni_mem goal_const_tab (c, c_typ) =
blanchet@37505
    72
  exists (match_types c_typ) (these (Symtab.lookup goal_const_tab c))
blanchet@37505
    73
paulson@24287
    74
(*Maps a "real" type to a const_typ*)
blanchet@37505
    75
fun const_typ_of (Type (c,typs)) = CType (c, map const_typ_of typs)
paulson@24287
    76
  | const_typ_of (TFree _) = CTVar
paulson@24287
    77
  | const_typ_of (TVar _) = CTVar
paulson@24287
    78
paulson@24287
    79
(*Pairs a constant with the list of its type instantiations (using const_typ)*)
blanchet@37505
    80
fun const_with_typ thy (c,typ) =
paulson@24287
    81
    let val tvars = Sign.const_typargs thy (c,typ)
paulson@24287
    82
    in (c, map const_typ_of tvars) end
blanchet@37505
    83
    handle TYPE _ => (c,[]);   (*Variable (locale constant): monomorphic*)
paulson@24287
    84
paulson@24287
    85
(*Add a const/type pair to the table, but a [] entry means a standard connective,
paulson@24287
    86
  which we ignore.*)
blanchet@37502
    87
fun add_const_type_to_table (c, ctyps) =
blanchet@37502
    88
  Symtab.map_default (c, [ctyps])
blanchet@37502
    89
                     (fn [] => [] | ctypss => insert (op =) ctyps ctypss)
paulson@24287
    90
blanchet@37537
    91
val fresh_sko_prefix = "Sledgehammer.Skox."
blanchet@37537
    92
blanchet@37537
    93
val flip = Option.map not
paulson@24287
    94
blanchet@37537
    95
(* Including equality in this list might be expected to stop rules like
blanchet@37537
    96
   subset_antisym from being chosen, but for some reason filtering works better
blanchet@37537
    97
   with them listed. The logical signs All, Ex, &, and --> are omitted for CNF
blanchet@37537
    98
   because any remaining occurrences must be within comprehensions. *)
blanchet@37537
    99
val boring_cnf_consts =
blanchet@37537
   100
  [@{const_name Trueprop}, @{const_name "==>"}, @{const_name all},
blanchet@37537
   101
   @{const_name "=="}, @{const_name "op |"}, @{const_name Not},
blanchet@37537
   102
   @{const_name "op ="}]
blanchet@37537
   103
blanchet@37537
   104
fun get_consts_typs thy pos ts =
blanchet@37505
   105
  let
blanchet@37505
   106
    (* Free variables are included, as well as constants, to handle locales.
blanchet@37505
   107
       "skolem_id" is included to increase the complexity of theorems containing
blanchet@37505
   108
       Skolem functions. In non-CNF form, "Ex" is included but each occurrence
blanchet@37505
   109
       is considered fresh, to simulate the effect of Skolemization. *)
blanchet@37537
   110
    fun do_term t =
blanchet@37537
   111
      case t of
blanchet@37537
   112
        Const x => add_const_type_to_table (const_with_typ thy x)
blanchet@37537
   113
      | Free x => add_const_type_to_table (const_with_typ thy x)
blanchet@37537
   114
      | (t as Const (@{const_name skolem_id}, _)) $ _ => do_term t
blanchet@37537
   115
      | t1 $ t2 => do_term t1 #> do_term t2
blanchet@37537
   116
      | Abs (_, _, t) => do_term t
blanchet@37537
   117
      | _ => I
blanchet@37537
   118
    fun do_quantifier sweet_pos pos body_t =
blanchet@37537
   119
      do_formula pos body_t
blanchet@37537
   120
      #> (if pos = SOME sweet_pos then I
blanchet@37537
   121
          else add_const_type_to_table (gensym fresh_sko_prefix, []))
blanchet@37537
   122
    and do_equality T t1 t2 =
blanchet@37537
   123
      fold (if T = @{typ bool} orelse T = @{typ prop} then do_formula NONE
blanchet@37537
   124
            else do_term) [t1, t2]
blanchet@37537
   125
    and do_formula pos t =
blanchet@37537
   126
      case t of
blanchet@37537
   127
        Const (@{const_name all}, _) $ Abs (_, _, body_t) =>
blanchet@37537
   128
        do_quantifier false pos body_t
blanchet@37537
   129
      | @{const "==>"} $ t1 $ t2 =>
blanchet@37537
   130
        do_formula (flip pos) t1 #> do_formula pos t2
blanchet@37537
   131
      | Const (@{const_name "=="}, Type (_, [T, _])) $ t1 $ t2 =>
blanchet@37537
   132
        do_equality T t1 t2
blanchet@37537
   133
      | @{const Trueprop} $ t1 => do_formula pos t1
blanchet@37537
   134
      | @{const Not} $ t1 => do_formula (flip pos) t1
blanchet@37537
   135
      | Const (@{const_name All}, _) $ Abs (_, _, body_t) =>
blanchet@37537
   136
        do_quantifier false pos body_t
blanchet@37537
   137
      | Const (@{const_name Ex}, _) $ Abs (_, _, body_t) =>
blanchet@37537
   138
        do_quantifier true pos body_t
blanchet@37537
   139
      | @{const "op &"} $ t1 $ t2 => fold (do_formula pos) [t1, t2]
blanchet@37537
   140
      | @{const "op |"} $ t1 $ t2 => fold (do_formula pos) [t1, t2]
blanchet@37537
   141
      | @{const "op -->"} $ t1 $ t2 =>
blanchet@37537
   142
        do_formula (flip pos) t1 #> do_formula pos t2
blanchet@37537
   143
      | Const (@{const_name "op ="}, Type (_, [T, _])) $ t1 $ t2 =>
blanchet@37537
   144
        do_equality T t1 t2
blanchet@37537
   145
      | (t0 as Const (_, @{typ bool})) $ t1 =>
blanchet@37537
   146
        do_term t0 #> do_formula pos t1  (* theory constant *)
blanchet@37537
   147
      | _ => do_term t
blanchet@37505
   148
  in
blanchet@37537
   149
    Symtab.empty
blanchet@37537
   150
    |> (if !use_natural_form then
blanchet@37537
   151
          fold (do_formula pos) ts
blanchet@37537
   152
        else
blanchet@37537
   153
          fold (Symtab.update o rpair []) boring_cnf_consts
blanchet@37537
   154
          #> fold do_term ts)
blanchet@37505
   155
  end
paulson@24287
   156
paulson@24287
   157
(*Inserts a dummy "constant" referring to the theory name, so that relevance
paulson@24287
   158
  takes the given theory into account.*)
blanchet@36220
   159
fun const_prop_of theory_relevant th =
blanchet@37505
   160
  if theory_relevant then
blanchet@37505
   161
    let
blanchet@37505
   162
      val name = Context.theory_name (theory_of_thm th)
blanchet@37505
   163
      val t = Const (name ^ ". 1", @{typ bool})
blanchet@37505
   164
    in t $ prop_of th end
blanchet@37505
   165
  else
blanchet@37505
   166
    prop_of th
blanchet@37505
   167
blanchet@37505
   168
fun appropriate_prop_of theory_relevant (cnf_thm, (_, orig_thm)) =
blanchet@37505
   169
  (if !use_natural_form then orig_thm else cnf_thm)
blanchet@37505
   170
  |> const_prop_of theory_relevant
paulson@24287
   171
paulson@24287
   172
(**** Constant / Type Frequencies ****)
paulson@24287
   173
paulson@24287
   174
(*A two-dimensional symbol table counts frequencies of constants. It's keyed first by
paulson@24287
   175
  constant name and second by its list of type instantiations. For the latter, we need
paulson@24287
   176
  a linear ordering on type const_typ list.*)
blanchet@37505
   177
paulson@24287
   178
local
paulson@24287
   179
paulson@24287
   180
fun cons_nr CTVar = 0
paulson@24287
   181
  | cons_nr (CType _) = 1;
paulson@24287
   182
paulson@24287
   183
in
paulson@24287
   184
paulson@24287
   185
fun const_typ_ord TU =
paulson@24287
   186
  case TU of
paulson@24287
   187
    (CType (a, Ts), CType (b, Us)) =>
paulson@24287
   188
      (case fast_string_ord(a,b) of EQUAL => dict_ord const_typ_ord (Ts,Us) | ord => ord)
paulson@24287
   189
  | (T, U) => int_ord (cons_nr T, cons_nr U);
paulson@24287
   190
paulson@24287
   191
end;
paulson@24287
   192
wenzelm@31971
   193
structure CTtab = Table(type key = const_typ list val ord = dict_ord const_typ_ord);
paulson@24287
   194
blanchet@37505
   195
fun count_axiom_consts theory_relevant thy axiom =
blanchet@37503
   196
  let
blanchet@37503
   197
    fun do_const (a, T) =
blanchet@37503
   198
      let val (c, cts) = const_with_typ thy (a,T) in
blanchet@37503
   199
        (* Two-dimensional table update. Constant maps to types maps to
blanchet@37503
   200
           count. *)
blanchet@37503
   201
        CTtab.map_default (cts, 0) (Integer.add 1)
blanchet@37503
   202
        |> Symtab.map_default (c, CTtab.empty)
blanchet@37503
   203
      end
blanchet@37503
   204
    fun do_term (Const x) = do_const x
blanchet@37503
   205
      | do_term (Free x) = do_const x
blanchet@37515
   206
      | do_term (Const (x as (@{const_name skolem_id}, _)) $ _) = do_const x
blanchet@37503
   207
      | do_term (t $ u) = do_term t #> do_term u
blanchet@37503
   208
      | do_term (Abs (_, _, t)) = do_term t
blanchet@37503
   209
      | do_term _ = I
blanchet@37505
   210
  in axiom |> appropriate_prop_of theory_relevant |> do_term end
paulson@24287
   211
paulson@24287
   212
paulson@24287
   213
(**** Actual Filtering Code ****)
paulson@24287
   214
paulson@24287
   215
(*The frequency of a constant is the sum of those of all instances of its type.*)
blanchet@37505
   216
fun const_frequency const_tab (c, cts) =
blanchet@36185
   217
  CTtab.fold (fn (cts', m) => match_types cts cts' ? Integer.add m)
blanchet@37505
   218
             (the (Symtab.lookup const_tab c)
blanchet@37505
   219
              handle Option.Option => raise Fail ("Const: " ^ c)) 0
paulson@24287
   220
blanchet@37503
   221
(*A surprising number of theorems contain only a few significant constants.
blanchet@37503
   222
  These include all induction rules, and other general theorems. Filtering
blanchet@37505
   223
  theorems in clause form reveals these complexities in the form of Skolem
blanchet@37503
   224
  functions. If we were instead to filter theorems in their natural form,
blanchet@37503
   225
  some other method of measuring theorem complexity would become necessary.*)
blanchet@37503
   226
blanchet@37503
   227
(* "log" seems best in practice. A constant function of one ignores the constant
blanchet@37503
   228
   frequencies. *)
blanchet@37503
   229
fun log_weight2 (x:real) = 1.0 + 2.0 / Math.ln (x + 1.0)
blanchet@37503
   230
blanchet@37503
   231
(* Computes a constant's weight, as determined by its frequency. *)
blanchet@37503
   232
val ct_weight = log_weight2 o real oo const_frequency
paulson@24287
   233
blanchet@37505
   234
(*Relevant constants are weighted according to frequency,
paulson@24287
   235
  but irrelevant constants are simply counted. Otherwise, Skolem functions,
paulson@24287
   236
  which are rare, would harm a clause's chances of being picked.*)
blanchet@37505
   237
fun clause_weight const_tab gctyps consts_typs =
paulson@24287
   238
    let val rel = filter (uni_mem gctyps) consts_typs
blanchet@37505
   239
        val rel_weight = fold (curry Real.+ o ct_weight const_tab) rel 0.0
paulson@24287
   240
    in
wenzelm@32960
   241
        rel_weight / (rel_weight + real (length consts_typs - length rel))
paulson@24287
   242
    end;
blanchet@37505
   243
paulson@24287
   244
(*Multiplies out to a list of pairs: 'a * 'b list -> ('a * 'b) list -> ('a * 'b) list*)
wenzelm@30190
   245
fun add_expand_pairs (x,ys) xys = List.foldl (fn (y,acc) => (x,y)::acc) xys ys;
paulson@24287
   246
blanchet@37505
   247
fun consts_typs_of_term thy t =
blanchet@37537
   248
  Symtab.fold add_expand_pairs (get_consts_typs thy (SOME false) [t]) []
paulson@24287
   249
blanchet@37505
   250
fun pair_consts_typs_axiom theory_relevant thy axiom =
blanchet@37505
   251
  (axiom, axiom |> appropriate_prop_of theory_relevant
blanchet@37505
   252
                |> consts_typs_of_term thy)
paulson@24287
   253
blanchet@37505
   254
exception CONST_OR_FREE of unit
blanchet@37505
   255
blanchet@37505
   256
fun dest_Const_or_Free (Const x) = x
blanchet@37505
   257
  | dest_Const_or_Free (Free x) = x
blanchet@37505
   258
  | dest_Const_or_Free _ = raise CONST_OR_FREE ()
paulson@24287
   259
paulson@24287
   260
(*Look for definitions of the form f ?x1 ... ?xn = t, but not reversed.*)
wenzelm@32994
   261
fun defines thy thm gctypes =
paulson@24287
   262
    let val tm = prop_of thm
wenzelm@32960
   263
        fun defs lhs rhs =
paulson@24287
   264
            let val (rator,args) = strip_comb lhs
blanchet@37505
   265
                val ct = const_with_typ thy (dest_Const_or_Free rator)
haftmann@33037
   266
            in
haftmann@33037
   267
              forall is_Var args andalso uni_mem gctypes ct andalso
haftmann@33038
   268
                subset (op =) (Term.add_vars rhs [], Term.add_vars lhs [])
paulson@24287
   269
            end
blanchet@37505
   270
            handle CONST_OR_FREE () => false
blanchet@37505
   271
    in
blanchet@35963
   272
        case tm of
blanchet@37505
   273
          @{const Trueprop} $ (Const (@{const_name "op ="}, _) $ lhs $ rhs) =>
blanchet@37505
   274
            defs lhs rhs
blanchet@35963
   275
        | _ => false
paulson@24287
   276
    end;
paulson@24287
   277
blanchet@37500
   278
type annotated_clause = cnf_thm * ((string * const_typ list) list)
blanchet@37505
   279
paulson@24287
   280
(*For a reverse sort, putting the largest values first.*)
blanchet@37500
   281
fun compare_pairs ((_, w1), (_, w2)) = Real.compare (w2, w1)
paulson@24287
   282
paulson@24287
   283
(*Limit the number of new clauses, to prevent runaway acceptance.*)
blanchet@37500
   284
fun take_best max_new (newpairs : (annotated_clause * real) list) =
paulson@24287
   285
  let val nnew = length newpairs
paulson@24287
   286
  in
wenzelm@28477
   287
    if nnew <= max_new then (map #1 newpairs, [])
blanchet@37505
   288
    else
paulson@24287
   289
      let val cls = sort compare_pairs newpairs
wenzelm@28477
   290
          val accepted = List.take (cls, max_new)
paulson@24287
   291
      in
blanchet@37505
   292
        trace_msg (fn () => ("Number of candidates, " ^ Int.toString nnew ^
wenzelm@32960
   293
                       ", exceeds the limit of " ^ Int.toString (max_new)));
blanchet@35865
   294
        trace_msg (fn () => ("Effective pass mark: " ^ Real.toString (#2 (List.last accepted))));
blanchet@35865
   295
        trace_msg (fn () => "Actually passed: " ^
blanchet@37500
   296
          space_implode ", " (map (fn (((_,((name,_), _)),_),_) => name) accepted));
paulson@24287
   297
wenzelm@32960
   298
        (map #1 accepted, map #1 (List.drop (cls, max_new)))
paulson@24287
   299
      end
paulson@24287
   300
  end;
paulson@24287
   301
blanchet@36922
   302
fun relevant_clauses ctxt relevance_convergence defs_relevant max_new
blanchet@37505
   303
                     ({add, del, ...} : relevance_override) const_tab =
blanchet@36182
   304
  let
blanchet@37344
   305
    val thy = ProofContext.theory_of ctxt
blanchet@37501
   306
    val add_thms = maps (ProofContext.get_fact ctxt) add
blanchet@37501
   307
    val del_thms = maps (ProofContext.get_fact ctxt) del
blanchet@37505
   308
    fun iter threshold rel_const_tab =
blanchet@36182
   309
      let
blanchet@36182
   310
        fun relevant ([], _) [] = []  (* Nothing added this iteration *)
blanchet@37344
   311
          | relevant (newpairs, rejects) [] =
blanchet@36182
   312
            let
blanchet@36182
   313
              val (newrels, more_rejects) = take_best max_new newpairs
blanchet@36182
   314
              val new_consts = maps #2 newrels
blanchet@37505
   315
              val rel_const_tab =
blanchet@37505
   316
                rel_const_tab |> fold add_const_type_to_table new_consts
blanchet@37344
   317
              val threshold =
blanchet@37344
   318
                threshold + (1.0 - threshold) / relevance_convergence
wenzelm@32960
   319
            in
blanchet@36182
   320
              trace_msg (fn () => "relevant this iteration: " ^
blanchet@36182
   321
                                  Int.toString (length newrels));
blanchet@37505
   322
              map #1 newrels @ iter threshold rel_const_tab
blanchet@37344
   323
                  (more_rejects @ rejects)
wenzelm@32960
   324
            end
blanchet@36182
   325
          | relevant (newrels, rejects)
blanchet@37503
   326
                     ((ax as (clsthm as (_, ((name, n), orig_th)),
blanchet@37501
   327
                              consts_typs)) :: axs) =
blanchet@35966
   328
            let
blanchet@37505
   329
              val weight =
blanchet@37505
   330
                if member Thm.eq_thm add_thms orig_th then 1.0
blanchet@37505
   331
                else if member Thm.eq_thm del_thms orig_th then 0.0
blanchet@37505
   332
                else clause_weight const_tab rel_const_tab consts_typs
wenzelm@32960
   333
            in
blanchet@37344
   334
              if weight >= threshold orelse
blanchet@37505
   335
                 (defs_relevant andalso
blanchet@37505
   336
                  defines thy (#1 clsthm) rel_const_tab) then
blanchet@37537
   337
                (trace_msg (fn () =>
blanchet@37537
   338
                     name ^ " clause " ^ Int.toString n ^
blanchet@37537
   339
                     " passes: " ^ Real.toString weight
blanchet@37537
   340
                     (* ^ " consts: " ^ commas (map fst consts_typs) *));
blanchet@37537
   341
                 relevant ((ax, weight) :: newrels, rejects) axs)
blanchet@36182
   342
              else
blanchet@36182
   343
                relevant (newrels, ax :: rejects) axs
wenzelm@32960
   344
            end
blanchet@36182
   345
        in
blanchet@37344
   346
          trace_msg (fn () => "relevant_clauses, current threshold: " ^
blanchet@37344
   347
                              Real.toString threshold);
blanchet@36182
   348
          relevant ([], [])
blanchet@36182
   349
        end
blanchet@36182
   350
  in iter end
blanchet@37505
   351
blanchet@36922
   352
fun relevance_filter ctxt relevance_threshold relevance_convergence
blanchet@36922
   353
                     defs_relevant max_new theory_relevant relevance_override
blanchet@37505
   354
                     thy (axioms : cnf_thm list) goals =
blanchet@37538
   355
  if relevance_threshold > 1.0 then
blanchet@37538
   356
    []
blanchet@37538
   357
  else if relevance_threshold < 0.0 then
blanchet@37538
   358
    axioms
blanchet@37538
   359
  else
blanchet@35963
   360
    let
blanchet@37503
   361
      val const_tab = fold (count_axiom_consts theory_relevant thy) axioms
blanchet@37503
   362
                           Symtab.empty
blanchet@37537
   363
      val goal_const_tab = get_consts_typs thy (SOME true) goals
blanchet@35963
   364
      val _ =
blanchet@35963
   365
        trace_msg (fn () => "Initial constants: " ^
blanchet@35963
   366
                            commas (Symtab.keys goal_const_tab))
blanchet@35963
   367
      val relevant =
blanchet@36922
   368
        relevant_clauses ctxt relevance_convergence defs_relevant max_new
blanchet@37344
   369
                         relevance_override const_tab relevance_threshold
blanchet@36922
   370
                         goal_const_tab
blanchet@36220
   371
                         (map (pair_consts_typs_axiom theory_relevant thy)
blanchet@36220
   372
                              axioms)
blanchet@35963
   373
    in
blanchet@35963
   374
      trace_msg (fn () => "Total relevant: " ^ Int.toString (length relevant));
blanchet@35963
   375
      relevant
blanchet@35963
   376
    end
paulson@24287
   377
paulson@24287
   378
(***************************************************************)
mengj@19768
   379
(* Retrieving and filtering lemmas                             *)
mengj@19768
   380
(***************************************************************)
mengj@19768
   381
paulson@33022
   382
(*** retrieve lemmas and filter them ***)
mengj@19768
   383
mengj@19768
   384
(*Hashing to detect duplicate and variant clauses, e.g. from the [iff] attribute*)
mengj@19768
   385
paulson@22382
   386
fun setinsert (x,s) = Symtab.update (x,()) s;
mengj@19768
   387
paulson@20757
   388
(*Reject theorems with names like "List.filter.filter_list_def" or
paulson@21690
   389
  "Accessible_Part.acc.defs", as these are definitions arising from packages.*)
paulson@20757
   390
fun is_package_def a =
wenzelm@30364
   391
  let val names = Long_Name.explode a
paulson@21690
   392
  in
paulson@21690
   393
     length names > 2 andalso
paulson@21690
   394
     not (hd names = "local") andalso
paulson@21690
   395
     String.isSuffix "_def" a  orelse  String.isSuffix "_defs" a
paulson@21690
   396
  end;
paulson@20757
   397
blanchet@36061
   398
fun mk_clause_table xs =
blanchet@36061
   399
  fold (Termtab.update o `(prop_of o fst)) xs Termtab.empty
paulson@22382
   400
blanchet@36061
   401
fun make_unique xs =
blanchet@36061
   402
  Termtab.fold (cons o snd) (mk_clause_table xs) []
mengj@19768
   403
blanchet@36061
   404
(* Remove existing axiom clauses from the conjecture clauses, as this can
blanchet@36061
   405
   dramatically boost an ATP's performance (for some reason). *)
blanchet@36061
   406
fun subtract_cls ax_clauses =
blanchet@36061
   407
  filter_out (Termtab.defined (mk_clause_table ax_clauses) o prop_of)
mengj@19768
   408
blanchet@37345
   409
fun all_name_thms_pairs respect_no_atp ctxt chained_ths =
paulson@22382
   410
  let
wenzelm@26675
   411
    val global_facts = PureThy.facts_of (ProofContext.theory_of ctxt);
wenzelm@26278
   412
    val local_facts = ProofContext.facts_of ctxt;
wenzelm@33641
   413
    val full_space =
wenzelm@33641
   414
      Name_Space.merge (Facts.space_of global_facts, Facts.space_of local_facts);
wenzelm@33641
   415
wenzelm@33641
   416
    fun valid_facts facts =
wenzelm@33641
   417
      (facts, []) |-> Facts.fold_static (fn (name, ths0) =>
blanchet@37399
   418
        if Facts.is_concealed facts name orelse
blanchet@37399
   419
           (respect_no_atp andalso is_package_def name) orelse
blanchet@37399
   420
           member (op =) multi_base_blacklist (Long_Name.base_name name) then
blanchet@37399
   421
          I
blanchet@37399
   422
        else
blanchet@37399
   423
          let
blanchet@37399
   424
            fun check_thms a =
blanchet@37399
   425
              (case try (ProofContext.get_thms ctxt) a of
blanchet@37399
   426
                NONE => false
blanchet@37399
   427
              | SOME ths1 => Thm.eq_thms (ths0, ths1));
wenzelm@33641
   428
blanchet@37399
   429
            val name1 = Facts.extern facts name;
blanchet@37399
   430
            val name2 = Name_Space.extern full_space name;
blanchet@37399
   431
            val ths = filter_out is_theorem_bad_for_atps ths0;
blanchet@37399
   432
          in
blanchet@37399
   433
            case find_first check_thms [name1, name2, name] of
blanchet@37399
   434
              NONE => I
blanchet@37399
   435
            | SOME name' =>
blanchet@37399
   436
              cons (name' |> forall (member Thm.eq_thm chained_ths) ths
blanchet@37399
   437
                             ? prefix chained_prefix, ths)
blanchet@37399
   438
          end)
wenzelm@26675
   439
  in valid_facts global_facts @ valid_facts local_facts end;
paulson@21224
   440
wenzelm@33309
   441
fun multi_name a th (n, pairs) =
wenzelm@33309
   442
  (n + 1, (a ^ "(" ^ Int.toString n ^ ")", th) :: pairs);
paulson@21224
   443
blanchet@37498
   444
fun add_names (_, []) pairs = pairs
blanchet@37399
   445
  | add_names (a, [th]) pairs = (a, th) :: pairs
blanchet@37399
   446
  | add_names (a, ths) pairs = #2 (fold (multi_name a) ths (1, pairs))
paulson@21224
   447
paulson@21290
   448
fun is_multi (a, ths) = length ths > 1 orelse String.isSuffix ".axioms" a;
paulson@21290
   449
blanchet@36550
   450
(* The single-name theorems go after the multiple-name ones, so that single
blanchet@36550
   451
   names are preferred when both are available. *)
blanchet@37344
   452
fun name_thm_pairs respect_no_atp ctxt name_thms_pairs =
wenzelm@33309
   453
  let
blanchet@37344
   454
    val (mults, singles) = List.partition is_multi name_thms_pairs
blanchet@37399
   455
    val ps = [] |> fold add_names singles |> fold add_names mults
blanchet@36060
   456
  in ps |> respect_no_atp ? filter_out (No_ATPs.member ctxt o snd) end;
paulson@21224
   457
blanchet@37344
   458
fun is_named ("", th) =
blanchet@37344
   459
    (warning ("No name for theorem " ^
blanchet@37344
   460
              Display.string_of_thm_without_context th); false)
blanchet@37344
   461
  | is_named _ = true
blanchet@37344
   462
fun checked_name_thm_pairs respect_no_atp ctxt =
blanchet@37344
   463
  name_thm_pairs respect_no_atp ctxt
blanchet@37344
   464
  #> tap (fn ps => trace_msg
blanchet@37344
   465
                        (fn () => ("Considering " ^ Int.toString (length ps) ^
blanchet@37344
   466
                                   " theorems")))
blanchet@37344
   467
  #> filter is_named
paulson@19894
   468
blanchet@37344
   469
fun name_thms_pair_from_ref ctxt chained_ths xref =
blanchet@37344
   470
  let
blanchet@37344
   471
    val ths = ProofContext.get_fact ctxt xref
blanchet@37344
   472
    val name = Facts.string_of_ref xref
blanchet@37344
   473
               |> forall (member Thm.eq_thm chained_ths) ths
blanchet@37344
   474
                  ? prefix chained_prefix
blanchet@37344
   475
  in (name, ths) end
blanchet@37344
   476
mengj@19768
   477
paulson@21290
   478
(***************************************************************)
paulson@21290
   479
(* Type Classes Present in the Axiom or Conjecture Clauses     *)
paulson@21290
   480
(***************************************************************)
paulson@21290
   481
wenzelm@32952
   482
fun add_classes (sorts, cset) = List.foldl setinsert cset (flat sorts);
paulson@21290
   483
paulson@21290
   484
(*Remove this trivial type class*)
blanchet@35865
   485
fun delete_type cset = Symtab.delete_safe (the_single @{sort HOL.type}) cset;
paulson@21290
   486
paulson@21290
   487
fun tvar_classes_of_terms ts =
wenzelm@29270
   488
  let val sorts_list = map (map #2 o OldTerm.term_tvars) ts
wenzelm@30190
   489
  in  Symtab.keys (delete_type (List.foldl add_classes Symtab.empty sorts_list))  end;
paulson@21290
   490
paulson@21290
   491
fun tfree_classes_of_terms ts =
wenzelm@29270
   492
  let val sorts_list = map (map #2 o OldTerm.term_tfrees) ts
wenzelm@30190
   493
  in  Symtab.keys (delete_type (List.foldl add_classes Symtab.empty sorts_list))  end;
paulson@20526
   494
paulson@21373
   495
(*fold type constructors*)
paulson@21373
   496
fun fold_type_consts f (Type (a, Ts)) x = fold (fold_type_consts f) Ts (f (a,x))
wenzelm@32994
   497
  | fold_type_consts _ _ x = x;
paulson@21373
   498
paulson@21397
   499
(*Type constructors used to instantiate overloaded constants are the only ones needed.*)
paulson@21397
   500
fun add_type_consts_in_term thy =
blanchet@37504
   501
  let
blanchet@37504
   502
    val const_typargs = Sign.const_typargs thy
blanchet@37504
   503
    fun aux (Const cT) = fold (fold_type_consts setinsert) (const_typargs cT)
blanchet@37504
   504
      | aux (Abs (_, _, u)) = aux u
blanchet@37504
   505
      | aux (Const (@{const_name skolem_id}, _) $ _) = I
blanchet@37504
   506
      | aux (t $ u) = aux t #> aux u
blanchet@37504
   507
      | aux _ = I
blanchet@37504
   508
  in aux end
paulson@21373
   509
paulson@21397
   510
fun type_consts_of_terms thy ts =
paulson@21397
   511
  Symtab.keys (fold (add_type_consts_in_term thy) ts Symtab.empty);
paulson@21373
   512
paulson@21373
   513
mengj@19194
   514
(***************************************************************)
mengj@19194
   515
(* ATP invocation methods setup                                *)
mengj@19194
   516
(***************************************************************)
mengj@19194
   517
blanchet@37538
   518
fun is_quasi_fol_theorem thy =
blanchet@37538
   519
  Meson.is_fol_term thy o snd o conceal_skolem_somes ~1 [] o prop_of
paulson@20526
   520
blanchet@37347
   521
(**** Predicates to detect unwanted clauses (prolific or likely to cause
blanchet@37347
   522
      unsoundness) ****)
paulson@21470
   523
paulson@21470
   524
(** Too general means, positive equality literal with a variable X as one operand,
paulson@21470
   525
  when X does not occur properly in the other operand. This rules out clearly
paulson@21470
   526
  inconsistent clauses such as V=a|V=b, though it by no means guarantees soundness. **)
wenzelm@21588
   527
blanchet@37348
   528
fun var_occurs_in_term ix =
blanchet@37348
   529
  let
blanchet@37348
   530
    fun aux (Var (jx, _)) = (ix = jx)
blanchet@37348
   531
      | aux (t1 $ t2) = aux t1 orelse aux t2
blanchet@37348
   532
      | aux (Abs (_, _, t)) = aux t
blanchet@37348
   533
      | aux _ = false
blanchet@37348
   534
  in aux end
paulson@21470
   535
blanchet@37348
   536
fun is_record_type T = not (null (Record.dest_recTs T))
paulson@21470
   537
paulson@21470
   538
(*Unwanted equalities include
paulson@21470
   539
  (1) those between a variable that does not properly occur in the second operand,
paulson@21470
   540
  (2) those between a variable and a record, since these seem to be prolific "cases" thms
wenzelm@21588
   541
*)
blanchet@37348
   542
fun too_general_eqterms (Var (ix,T), t) =
blanchet@37348
   543
    not (var_occurs_in_term ix t) orelse is_record_type T
paulson@21470
   544
  | too_general_eqterms _ = false;
paulson@21470
   545
blanchet@35865
   546
fun too_general_equality (Const (@{const_name "op ="}, _) $ x $ y) =
paulson@21470
   547
      too_general_eqterms (x,y) orelse too_general_eqterms(y,x)
paulson@21470
   548
  | too_general_equality _ = false;
paulson@21470
   549
wenzelm@29267
   550
fun has_typed_var tycons = exists_subterm
wenzelm@29267
   551
  (fn Var (_, Type (a, _)) => member (op =) tycons a | _ => false);
paulson@21431
   552
blanchet@37347
   553
(* Clauses are forbidden to contain variables of these types. The typical reason
blanchet@37347
   554
   is that they lead to unsoundness. Note that "unit" satisfies numerous
blanchet@37347
   555
   equations like "?x = ()". The resulting clause will have no type constraint,
blanchet@37347
   556
   yielding false proofs. Even "bool" leads to many unsound proofs, though only
blanchet@37347
   557
   for higher-order problems. *)
blanchet@37347
   558
val dangerous_types = [@{type_name unit}, @{type_name bool}];
paulson@22217
   559
blanchet@37347
   560
(* Clauses containing variables of type "unit" or "bool" or of the form
blanchet@37347
   561
   "?x = A | ?x = B | ?x = C" are likely to lead to unsound proofs if types are
blanchet@37347
   562
   omitted. *)
blanchet@37347
   563
fun is_dangerous_term _ @{prop True} = true
blanchet@37347
   564
  | is_dangerous_term full_types t =
blanchet@37505
   565
    not full_types andalso
blanchet@37347
   566
    (has_typed_var dangerous_types t orelse
blanchet@37347
   567
     forall too_general_equality (HOLogic.disjuncts (strip_Trueprop t)))
paulson@21470
   568
blanchet@37399
   569
fun is_fol_goal thy = forall (Meson.is_fol_term thy) o map prop_of
immler@30536
   570
blanchet@37347
   571
fun relevant_facts full_types respect_no_atp relevance_threshold
blanchet@37347
   572
                   relevance_convergence defs_relevant max_new theory_relevant
blanchet@37347
   573
                   (relevance_override as {add, del, only})
blanchet@37347
   574
                   (ctxt, (chained_ths, _)) goal_cls =
blanchet@37538
   575
  let
blanchet@37538
   576
    val thy = ProofContext.theory_of ctxt
blanchet@37538
   577
    val add_thms = maps (ProofContext.get_fact ctxt) add
blanchet@37538
   578
    val has_override = not (null add) orelse not (null del)
blanchet@37538
   579
    val is_FO = is_fol_goal thy goal_cls
blanchet@37538
   580
    val axioms =
blanchet@37538
   581
      checked_name_thm_pairs (respect_no_atp andalso not only) ctxt
blanchet@37538
   582
          (map (name_thms_pair_from_ref ctxt chained_ths) add @
blanchet@37538
   583
           (if only then []
blanchet@37538
   584
            else all_name_thms_pairs respect_no_atp ctxt chained_ths))
blanchet@37538
   585
      |> cnf_rules_pairs thy
blanchet@37538
   586
      |> not has_override ? make_unique
blanchet@37538
   587
      |> filter (fn (cnf_thm, (_, orig_thm)) =>
blanchet@37538
   588
                    member Thm.eq_thm add_thms orig_thm orelse
blanchet@37538
   589
                    ((not is_FO orelse is_quasi_fol_theorem thy cnf_thm) andalso
blanchet@37538
   590
                     not (is_dangerous_term full_types (prop_of cnf_thm))))
blanchet@37538
   591
  in
blanchet@37538
   592
    relevance_filter ctxt relevance_threshold relevance_convergence
blanchet@37538
   593
                     defs_relevant max_new theory_relevant relevance_override
blanchet@37538
   594
                     thy axioms (map prop_of goal_cls)
blanchet@37538
   595
    |> has_override ? make_unique
blanchet@37538
   596
    |> sort (prod_ord string_ord int_ord o pairself (fst o snd))
blanchet@37538
   597
  end
immler@30536
   598
blanchet@37509
   599
(** Helper clauses **)
blanchet@37509
   600
blanchet@37509
   601
fun count_combterm (CombConst ((c, _), _, _)) =
blanchet@37509
   602
    Symtab.map_entry c (Integer.add 1)
blanchet@37509
   603
  | count_combterm (CombVar _) = I
blanchet@37509
   604
  | count_combterm (CombApp (t1, t2)) = count_combterm t1 #> count_combterm t2
blanchet@37509
   605
fun count_literal (Literal (_, t)) = count_combterm t
blanchet@37509
   606
fun count_clause (HOLClause {literals, ...}) = fold count_literal literals
blanchet@37509
   607
blanchet@37509
   608
val raw_cnf_rules_pairs = map (fn (name, thm) => (thm, ((name, 0), thm)))
blanchet@37509
   609
fun cnf_helper_thms thy raw =
blanchet@37509
   610
  map (`Thm.get_name_hint)
blanchet@37509
   611
  #> (if raw then raw_cnf_rules_pairs else cnf_rules_pairs thy)
blanchet@37509
   612
blanchet@37509
   613
val optional_helpers =
blanchet@37509
   614
  [(["c_COMBI", "c_COMBK"], (false, @{thms COMBI_def COMBK_def})),
blanchet@37509
   615
   (["c_COMBB", "c_COMBC"], (false, @{thms COMBB_def COMBC_def})),
blanchet@37509
   616
   (["c_COMBS"], (false, @{thms COMBS_def}))]
blanchet@37509
   617
val optional_typed_helpers =
blanchet@37509
   618
  [(["c_True", "c_False"], (true, @{thms True_or_False})),
blanchet@37509
   619
   (["c_If"], (true, @{thms if_True if_False True_or_False}))]
blanchet@37509
   620
val mandatory_helpers = @{thms fequal_imp_equal equal_imp_fequal}
blanchet@37509
   621
blanchet@37509
   622
val init_counters =
blanchet@37509
   623
  Symtab.make (maps (maps (map (rpair 0) o fst))
blanchet@37509
   624
                    [optional_helpers, optional_typed_helpers])
blanchet@37509
   625
blanchet@37509
   626
fun get_helper_clauses thy is_FO full_types conjectures axcls =
blanchet@37509
   627
  let
blanchet@37509
   628
    val axclauses = map snd (make_axiom_clauses thy axcls)
blanchet@37509
   629
    val ct = fold (fold count_clause) [conjectures, axclauses] init_counters
blanchet@37509
   630
    fun is_needed c = the (Symtab.lookup ct c) > 0
blanchet@37509
   631
    val cnfs =
blanchet@37509
   632
      (optional_helpers
blanchet@37509
   633
       |> full_types ? append optional_typed_helpers
blanchet@37509
   634
       |> maps (fn (ss, (raw, ths)) =>
blanchet@37509
   635
                   if exists is_needed ss then cnf_helper_thms thy raw ths
blanchet@37509
   636
                   else []))
blanchet@37509
   637
      @ (if is_FO then [] else cnf_helper_thms thy false mandatory_helpers)
blanchet@37509
   638
  in map snd (make_axiom_clauses thy cnfs) end
blanchet@37509
   639
immler@31752
   640
(* prepare for passing to writer,
immler@31752
   641
   create additional clauses based on the information from extra_cls *)
blanchet@37498
   642
fun prepare_clauses full_types goal_cls axcls extra_cls thy =
immler@31409
   643
  let
blanchet@37399
   644
    val is_FO = is_fol_goal thy goal_cls
blanchet@36061
   645
    val ccls = subtract_cls extra_cls goal_cls
blanchet@35865
   646
    val _ = app (fn th => trace_msg (fn _ => Display.string_of_thm_global thy th)) ccls
immler@30536
   647
    val ccltms = map prop_of ccls
immler@31752
   648
    and axtms = map (prop_of o #1) extra_cls
immler@30536
   649
    val subs = tfree_classes_of_terms ccltms
immler@30536
   650
    and supers = tvar_classes_of_terms axtms
blanchet@35865
   651
    and tycons = type_consts_of_terms thy (ccltms @ axtms)
immler@30536
   652
    (*TFrees in conjecture clauses; TVars in axiom clauses*)
blanchet@37498
   653
    val conjectures = make_conjecture_clauses thy ccls
blanchet@37498
   654
    val (_, extra_clauses) = ListPair.unzip (make_axiom_clauses thy extra_cls)
blanchet@37498
   655
    val (clnames, axiom_clauses) = ListPair.unzip (make_axiom_clauses thy axcls)
blanchet@37479
   656
    val helper_clauses =
blanchet@37498
   657
      get_helper_clauses thy is_FO full_types conjectures extra_cls
blanchet@37498
   658
    val (supers', arity_clauses) = make_arity_clauses thy tycons supers
blanchet@35865
   659
    val classrel_clauses = make_classrel_clauses thy subs supers'
immler@30536
   660
  in
immler@31752
   661
    (Vector.fromList clnames,
immler@31865
   662
      (conjectures, axiom_clauses, extra_clauses, helper_clauses, classrel_clauses, arity_clauses))
immler@31409
   663
  end
quigley@15644
   664
paulson@15347
   665
end;