src/HOL/Tools/Sledgehammer/sledgehammer_fact_filter.ML
author blanchet
Tue Jun 22 18:31:49 2010 +0200 (2010-06-22 ago)
changeset 37505 d9af5c01dc4a
parent 37504 4308d2bbbca8
child 37506 32a1ee39c49b
permissions -rw-r--r--
added code to optionally perform fact filtering on the original (non-CNF) formulas
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@37399
    24
  val is_quasi_fol_term : theory -> term -> 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@37505
    43
val use_natural_form = 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@37505
    91
val fresh_Ex_prefix = "Sledgehammer.Ex."
paulson@24287
    92
blanchet@37502
    93
fun get_goal_consts_typs thy goals =
blanchet@37505
    94
  let
blanchet@37505
    95
    val use_natural_form = !use_natural_form
blanchet@37505
    96
    (* Free variables are included, as well as constants, to handle locales.
blanchet@37505
    97
       "skolem_id" is included to increase the complexity of theorems containing
blanchet@37505
    98
       Skolem functions. In non-CNF form, "Ex" is included but each occurrence
blanchet@37505
    99
       is considered fresh, to simulate the effect of Skolemization. *)
blanchet@37505
   100
    fun aux (Const (x as (s, _))) =
blanchet@37505
   101
        (if s = @{const_name Ex} andalso use_natural_form then
blanchet@37505
   102
           (gensym fresh_Ex_prefix, [])
blanchet@37505
   103
         else
blanchet@37505
   104
           (const_with_typ thy x))
blanchet@37505
   105
        |> add_const_type_to_table
blanchet@37505
   106
      | aux (Free x) = add_const_type_to_table (const_with_typ thy x)
blanchet@37505
   107
      | aux ((t as Const (@{const_name skolem_id}, _)) $ _) = aux t
blanchet@37505
   108
      | aux (t $ u) = aux t #> aux u
blanchet@37505
   109
      | aux (Abs (_, _, t)) = aux t
blanchet@37505
   110
      | aux _ = I
blanchet@37505
   111
    (* Including equality in this list might be expected to stop rules like
blanchet@37505
   112
       subset_antisym from being chosen, but for some reason filtering works better
blanchet@37505
   113
       with them listed. The logical signs All, Ex, &, and --> are omitted for CNF
blanchet@37505
   114
       because any remaining occurrences must be within comprehensions. *)
blanchet@37505
   115
    val standard_consts =
blanchet@37505
   116
      [@{const_name Trueprop}, @{const_name "==>"}, @{const_name all},
blanchet@37505
   117
       @{const_name "=="}, @{const_name "op |"}, @{const_name Not},
blanchet@37505
   118
       @{const_name "op ="}] @
blanchet@37505
   119
      (if use_natural_form then
blanchet@37505
   120
         [@{const_name All}, @{const_name Ex}, @{const_name "op &"},
blanchet@37505
   121
          @{const_name "op -->"}]
blanchet@37505
   122
       else
blanchet@37505
   123
         [])
blanchet@37505
   124
  in
blanchet@37505
   125
    Symtab.empty |> fold (Symtab.update o rpair []) standard_consts
blanchet@37505
   126
                 |> fold aux goals
blanchet@37505
   127
  end
paulson@24287
   128
paulson@24287
   129
(*Inserts a dummy "constant" referring to the theory name, so that relevance
paulson@24287
   130
  takes the given theory into account.*)
blanchet@36220
   131
fun const_prop_of theory_relevant th =
blanchet@37505
   132
  if theory_relevant then
blanchet@37505
   133
    let
blanchet@37505
   134
      val name = Context.theory_name (theory_of_thm th)
blanchet@37505
   135
      val t = Const (name ^ ". 1", @{typ bool})
blanchet@37505
   136
    in t $ prop_of th end
blanchet@37505
   137
  else
blanchet@37505
   138
    prop_of th
blanchet@37505
   139
blanchet@37505
   140
fun appropriate_prop_of theory_relevant (cnf_thm, (_, orig_thm)) =
blanchet@37505
   141
  (if !use_natural_form then orig_thm else cnf_thm)
blanchet@37505
   142
  |> const_prop_of theory_relevant
paulson@24287
   143
paulson@24287
   144
(**** Constant / Type Frequencies ****)
paulson@24287
   145
paulson@24287
   146
(*A two-dimensional symbol table counts frequencies of constants. It's keyed first by
paulson@24287
   147
  constant name and second by its list of type instantiations. For the latter, we need
paulson@24287
   148
  a linear ordering on type const_typ list.*)
blanchet@37505
   149
paulson@24287
   150
local
paulson@24287
   151
paulson@24287
   152
fun cons_nr CTVar = 0
paulson@24287
   153
  | cons_nr (CType _) = 1;
paulson@24287
   154
paulson@24287
   155
in
paulson@24287
   156
paulson@24287
   157
fun const_typ_ord TU =
paulson@24287
   158
  case TU of
paulson@24287
   159
    (CType (a, Ts), CType (b, Us)) =>
paulson@24287
   160
      (case fast_string_ord(a,b) of EQUAL => dict_ord const_typ_ord (Ts,Us) | ord => ord)
paulson@24287
   161
  | (T, U) => int_ord (cons_nr T, cons_nr U);
paulson@24287
   162
paulson@24287
   163
end;
paulson@24287
   164
wenzelm@31971
   165
structure CTtab = Table(type key = const_typ list val ord = dict_ord const_typ_ord);
paulson@24287
   166
blanchet@37505
   167
fun count_axiom_consts theory_relevant thy axiom =
blanchet@37503
   168
  let
blanchet@37503
   169
    fun do_const (a, T) =
blanchet@37503
   170
      let val (c, cts) = const_with_typ thy (a,T) in
blanchet@37503
   171
        (* Two-dimensional table update. Constant maps to types maps to
blanchet@37503
   172
           count. *)
blanchet@37503
   173
        CTtab.map_default (cts, 0) (Integer.add 1)
blanchet@37503
   174
        |> Symtab.map_default (c, CTtab.empty)
blanchet@37503
   175
      end
blanchet@37503
   176
    fun do_term (Const x) = do_const x
blanchet@37503
   177
      | do_term (Free x) = do_const x
blanchet@37503
   178
      | do_term (Const (@{const_name skolem_id}, _) $ _) = I
blanchet@37503
   179
      | do_term (t $ u) = do_term t #> do_term u
blanchet@37503
   180
      | do_term (Abs (_, _, t)) = do_term t
blanchet@37503
   181
      | do_term _ = I
blanchet@37505
   182
  in axiom |> appropriate_prop_of theory_relevant |> do_term end
paulson@24287
   183
paulson@24287
   184
paulson@24287
   185
(**** Actual Filtering Code ****)
paulson@24287
   186
paulson@24287
   187
(*The frequency of a constant is the sum of those of all instances of its type.*)
blanchet@37505
   188
fun const_frequency const_tab (c, cts) =
blanchet@36185
   189
  CTtab.fold (fn (cts', m) => match_types cts cts' ? Integer.add m)
blanchet@37505
   190
             (the (Symtab.lookup const_tab c)
blanchet@37505
   191
              handle Option.Option => raise Fail ("Const: " ^ c)) 0
paulson@24287
   192
blanchet@37503
   193
(*A surprising number of theorems contain only a few significant constants.
blanchet@37503
   194
  These include all induction rules, and other general theorems. Filtering
blanchet@37505
   195
  theorems in clause form reveals these complexities in the form of Skolem
blanchet@37503
   196
  functions. If we were instead to filter theorems in their natural form,
blanchet@37503
   197
  some other method of measuring theorem complexity would become necessary.*)
blanchet@37503
   198
blanchet@37503
   199
(* "log" seems best in practice. A constant function of one ignores the constant
blanchet@37503
   200
   frequencies. *)
blanchet@37503
   201
fun log_weight2 (x:real) = 1.0 + 2.0 / Math.ln (x + 1.0)
blanchet@37503
   202
blanchet@37503
   203
(* Computes a constant's weight, as determined by its frequency. *)
blanchet@37503
   204
val ct_weight = log_weight2 o real oo const_frequency
paulson@24287
   205
blanchet@37505
   206
(*Relevant constants are weighted according to frequency,
paulson@24287
   207
  but irrelevant constants are simply counted. Otherwise, Skolem functions,
paulson@24287
   208
  which are rare, would harm a clause's chances of being picked.*)
blanchet@37505
   209
fun clause_weight const_tab gctyps consts_typs =
paulson@24287
   210
    let val rel = filter (uni_mem gctyps) consts_typs
blanchet@37505
   211
        val rel_weight = fold (curry Real.+ o ct_weight const_tab) rel 0.0
paulson@24287
   212
    in
wenzelm@32960
   213
        rel_weight / (rel_weight + real (length consts_typs - length rel))
paulson@24287
   214
    end;
blanchet@37505
   215
paulson@24287
   216
(*Multiplies out to a list of pairs: 'a * 'b list -> ('a * 'b) list -> ('a * 'b) list*)
wenzelm@30190
   217
fun add_expand_pairs (x,ys) xys = List.foldl (fn (y,acc) => (x,y)::acc) xys ys;
paulson@24287
   218
blanchet@37505
   219
fun consts_typs_of_term thy t =
blanchet@37505
   220
  Symtab.fold add_expand_pairs (get_goal_consts_typs thy [t]) []
paulson@24287
   221
blanchet@37505
   222
fun pair_consts_typs_axiom theory_relevant thy axiom =
blanchet@37505
   223
  (axiom, axiom |> appropriate_prop_of theory_relevant
blanchet@37505
   224
                |> consts_typs_of_term thy)
paulson@24287
   225
blanchet@37505
   226
exception CONST_OR_FREE of unit
blanchet@37505
   227
blanchet@37505
   228
fun dest_Const_or_Free (Const x) = x
blanchet@37505
   229
  | dest_Const_or_Free (Free x) = x
blanchet@37505
   230
  | dest_Const_or_Free _ = raise CONST_OR_FREE ()
paulson@24287
   231
paulson@24287
   232
(*Look for definitions of the form f ?x1 ... ?xn = t, but not reversed.*)
wenzelm@32994
   233
fun defines thy thm gctypes =
paulson@24287
   234
    let val tm = prop_of thm
wenzelm@32960
   235
        fun defs lhs rhs =
paulson@24287
   236
            let val (rator,args) = strip_comb lhs
blanchet@37505
   237
                val ct = const_with_typ thy (dest_Const_or_Free rator)
haftmann@33037
   238
            in
haftmann@33037
   239
              forall is_Var args andalso uni_mem gctypes ct andalso
haftmann@33038
   240
                subset (op =) (Term.add_vars rhs [], Term.add_vars lhs [])
paulson@24287
   241
            end
blanchet@37505
   242
            handle CONST_OR_FREE () => false
blanchet@37505
   243
    in
blanchet@35963
   244
        case tm of
blanchet@37505
   245
          @{const Trueprop} $ (Const (@{const_name "op ="}, _) $ lhs $ rhs) =>
blanchet@37505
   246
            defs lhs rhs
blanchet@35963
   247
        | _ => false
paulson@24287
   248
    end;
paulson@24287
   249
blanchet@37500
   250
type annotated_clause = cnf_thm * ((string * const_typ list) list)
blanchet@37505
   251
paulson@24287
   252
(*For a reverse sort, putting the largest values first.*)
blanchet@37500
   253
fun compare_pairs ((_, w1), (_, w2)) = Real.compare (w2, w1)
paulson@24287
   254
paulson@24287
   255
(*Limit the number of new clauses, to prevent runaway acceptance.*)
blanchet@37500
   256
fun take_best max_new (newpairs : (annotated_clause * real) list) =
paulson@24287
   257
  let val nnew = length newpairs
paulson@24287
   258
  in
wenzelm@28477
   259
    if nnew <= max_new then (map #1 newpairs, [])
blanchet@37505
   260
    else
paulson@24287
   261
      let val cls = sort compare_pairs newpairs
wenzelm@28477
   262
          val accepted = List.take (cls, max_new)
paulson@24287
   263
      in
blanchet@37505
   264
        trace_msg (fn () => ("Number of candidates, " ^ Int.toString nnew ^
wenzelm@32960
   265
                       ", exceeds the limit of " ^ Int.toString (max_new)));
blanchet@35865
   266
        trace_msg (fn () => ("Effective pass mark: " ^ Real.toString (#2 (List.last accepted))));
blanchet@35865
   267
        trace_msg (fn () => "Actually passed: " ^
blanchet@37500
   268
          space_implode ", " (map (fn (((_,((name,_), _)),_),_) => name) accepted));
paulson@24287
   269
wenzelm@32960
   270
        (map #1 accepted, map #1 (List.drop (cls, max_new)))
paulson@24287
   271
      end
paulson@24287
   272
  end;
paulson@24287
   273
blanchet@36922
   274
fun relevant_clauses ctxt relevance_convergence defs_relevant max_new
blanchet@37505
   275
                     ({add, del, ...} : relevance_override) const_tab =
blanchet@36182
   276
  let
blanchet@37344
   277
    val thy = ProofContext.theory_of ctxt
blanchet@37501
   278
    val add_thms = maps (ProofContext.get_fact ctxt) add
blanchet@37501
   279
    val del_thms = maps (ProofContext.get_fact ctxt) del
blanchet@37505
   280
    fun iter threshold rel_const_tab =
blanchet@36182
   281
      let
blanchet@36182
   282
        fun relevant ([], _) [] = []  (* Nothing added this iteration *)
blanchet@37344
   283
          | relevant (newpairs, rejects) [] =
blanchet@36182
   284
            let
blanchet@36182
   285
              val (newrels, more_rejects) = take_best max_new newpairs
blanchet@36182
   286
              val new_consts = maps #2 newrels
blanchet@37505
   287
              val rel_const_tab =
blanchet@37505
   288
                rel_const_tab |> fold add_const_type_to_table new_consts
blanchet@37344
   289
              val threshold =
blanchet@37344
   290
                threshold + (1.0 - threshold) / relevance_convergence
wenzelm@32960
   291
            in
blanchet@36182
   292
              trace_msg (fn () => "relevant this iteration: " ^
blanchet@36182
   293
                                  Int.toString (length newrels));
blanchet@37505
   294
              map #1 newrels @ iter threshold rel_const_tab
blanchet@37344
   295
                  (more_rejects @ rejects)
wenzelm@32960
   296
            end
blanchet@36182
   297
          | relevant (newrels, rejects)
blanchet@37503
   298
                     ((ax as (clsthm as (_, ((name, n), orig_th)),
blanchet@37501
   299
                              consts_typs)) :: axs) =
blanchet@35966
   300
            let
blanchet@37505
   301
              val weight =
blanchet@37505
   302
                if member Thm.eq_thm add_thms orig_th then 1.0
blanchet@37505
   303
                else if member Thm.eq_thm del_thms orig_th then 0.0
blanchet@37505
   304
                else clause_weight const_tab rel_const_tab consts_typs
wenzelm@32960
   305
            in
blanchet@37344
   306
              if weight >= threshold orelse
blanchet@37505
   307
                 (defs_relevant andalso
blanchet@37505
   308
                  defines thy (#1 clsthm) rel_const_tab) then
blanchet@37505
   309
                (trace_msg (fn () => name ^ " clause " ^ Int.toString n ^
blanchet@36182
   310
                                     " passes: " ^ Real.toString weight);
blanchet@36182
   311
                relevant ((ax, weight) :: newrels, rejects) axs)
blanchet@36182
   312
              else
blanchet@36182
   313
                relevant (newrels, ax :: rejects) axs
wenzelm@32960
   314
            end
blanchet@36182
   315
        in
blanchet@37344
   316
          trace_msg (fn () => "relevant_clauses, current threshold: " ^
blanchet@37344
   317
                              Real.toString threshold);
blanchet@36182
   318
          relevant ([], [])
blanchet@36182
   319
        end
blanchet@36182
   320
  in iter end
blanchet@37505
   321
blanchet@36922
   322
fun relevance_filter ctxt relevance_threshold relevance_convergence
blanchet@36922
   323
                     defs_relevant max_new theory_relevant relevance_override
blanchet@37505
   324
                     thy (axioms : cnf_thm list) goals =
blanchet@35966
   325
  if relevance_threshold > 0.0 then
blanchet@35963
   326
    let
blanchet@37503
   327
      val const_tab = fold (count_axiom_consts theory_relevant thy) axioms
blanchet@37503
   328
                           Symtab.empty
paulson@24287
   329
      val goal_const_tab = get_goal_consts_typs thy goals
blanchet@35963
   330
      val _ =
blanchet@35963
   331
        trace_msg (fn () => "Initial constants: " ^
blanchet@35963
   332
                            commas (Symtab.keys goal_const_tab))
blanchet@35963
   333
      val relevant =
blanchet@36922
   334
        relevant_clauses ctxt relevance_convergence defs_relevant max_new
blanchet@37344
   335
                         relevance_override const_tab relevance_threshold
blanchet@36922
   336
                         goal_const_tab
blanchet@36220
   337
                         (map (pair_consts_typs_axiom theory_relevant thy)
blanchet@36220
   338
                              axioms)
blanchet@35963
   339
    in
blanchet@35963
   340
      trace_msg (fn () => "Total relevant: " ^ Int.toString (length relevant));
blanchet@35963
   341
      relevant
blanchet@35963
   342
    end
blanchet@35963
   343
  else
blanchet@35963
   344
    axioms;
paulson@24287
   345
paulson@24287
   346
(***************************************************************)
mengj@19768
   347
(* Retrieving and filtering lemmas                             *)
mengj@19768
   348
(***************************************************************)
mengj@19768
   349
paulson@33022
   350
(*** retrieve lemmas and filter them ***)
mengj@19768
   351
mengj@19768
   352
(*Hashing to detect duplicate and variant clauses, e.g. from the [iff] attribute*)
mengj@19768
   353
paulson@22382
   354
fun setinsert (x,s) = Symtab.update (x,()) s;
mengj@19768
   355
paulson@20757
   356
(*Reject theorems with names like "List.filter.filter_list_def" or
paulson@21690
   357
  "Accessible_Part.acc.defs", as these are definitions arising from packages.*)
paulson@20757
   358
fun is_package_def a =
wenzelm@30364
   359
  let val names = Long_Name.explode a
paulson@21690
   360
  in
paulson@21690
   361
     length names > 2 andalso
paulson@21690
   362
     not (hd names = "local") andalso
paulson@21690
   363
     String.isSuffix "_def" a  orelse  String.isSuffix "_defs" a
paulson@21690
   364
  end;
paulson@20757
   365
blanchet@36061
   366
fun mk_clause_table xs =
blanchet@36061
   367
  fold (Termtab.update o `(prop_of o fst)) xs Termtab.empty
paulson@22382
   368
blanchet@36061
   369
fun make_unique xs =
blanchet@36061
   370
  Termtab.fold (cons o snd) (mk_clause_table xs) []
mengj@19768
   371
blanchet@36061
   372
(* Remove existing axiom clauses from the conjecture clauses, as this can
blanchet@36061
   373
   dramatically boost an ATP's performance (for some reason). *)
blanchet@36061
   374
fun subtract_cls ax_clauses =
blanchet@36061
   375
  filter_out (Termtab.defined (mk_clause_table ax_clauses) o prop_of)
mengj@19768
   376
blanchet@37345
   377
fun all_name_thms_pairs respect_no_atp ctxt chained_ths =
paulson@22382
   378
  let
wenzelm@26675
   379
    val global_facts = PureThy.facts_of (ProofContext.theory_of ctxt);
wenzelm@26278
   380
    val local_facts = ProofContext.facts_of ctxt;
wenzelm@33641
   381
    val full_space =
wenzelm@33641
   382
      Name_Space.merge (Facts.space_of global_facts, Facts.space_of local_facts);
wenzelm@33641
   383
wenzelm@33641
   384
    fun valid_facts facts =
wenzelm@33641
   385
      (facts, []) |-> Facts.fold_static (fn (name, ths0) =>
blanchet@37399
   386
        if Facts.is_concealed facts name orelse
blanchet@37399
   387
           (respect_no_atp andalso is_package_def name) orelse
blanchet@37399
   388
           member (op =) multi_base_blacklist (Long_Name.base_name name) then
blanchet@37399
   389
          I
blanchet@37399
   390
        else
blanchet@37399
   391
          let
blanchet@37399
   392
            fun check_thms a =
blanchet@37399
   393
              (case try (ProofContext.get_thms ctxt) a of
blanchet@37399
   394
                NONE => false
blanchet@37399
   395
              | SOME ths1 => Thm.eq_thms (ths0, ths1));
wenzelm@33641
   396
blanchet@37399
   397
            val name1 = Facts.extern facts name;
blanchet@37399
   398
            val name2 = Name_Space.extern full_space name;
blanchet@37399
   399
            val ths = filter_out is_theorem_bad_for_atps ths0;
blanchet@37399
   400
          in
blanchet@37399
   401
            case find_first check_thms [name1, name2, name] of
blanchet@37399
   402
              NONE => I
blanchet@37399
   403
            | SOME name' =>
blanchet@37399
   404
              cons (name' |> forall (member Thm.eq_thm chained_ths) ths
blanchet@37399
   405
                             ? prefix chained_prefix, ths)
blanchet@37399
   406
          end)
wenzelm@26675
   407
  in valid_facts global_facts @ valid_facts local_facts end;
paulson@21224
   408
wenzelm@33309
   409
fun multi_name a th (n, pairs) =
wenzelm@33309
   410
  (n + 1, (a ^ "(" ^ Int.toString n ^ ")", th) :: pairs);
paulson@21224
   411
blanchet@37498
   412
fun add_names (_, []) pairs = pairs
blanchet@37399
   413
  | add_names (a, [th]) pairs = (a, th) :: pairs
blanchet@37399
   414
  | add_names (a, ths) pairs = #2 (fold (multi_name a) ths (1, pairs))
paulson@21224
   415
paulson@21290
   416
fun is_multi (a, ths) = length ths > 1 orelse String.isSuffix ".axioms" a;
paulson@21290
   417
blanchet@36550
   418
(* The single-name theorems go after the multiple-name ones, so that single
blanchet@36550
   419
   names are preferred when both are available. *)
blanchet@37344
   420
fun name_thm_pairs respect_no_atp ctxt name_thms_pairs =
wenzelm@33309
   421
  let
blanchet@37344
   422
    val (mults, singles) = List.partition is_multi name_thms_pairs
blanchet@37399
   423
    val ps = [] |> fold add_names singles |> fold add_names mults
blanchet@36060
   424
  in ps |> respect_no_atp ? filter_out (No_ATPs.member ctxt o snd) end;
paulson@21224
   425
blanchet@37344
   426
fun is_named ("", th) =
blanchet@37344
   427
    (warning ("No name for theorem " ^
blanchet@37344
   428
              Display.string_of_thm_without_context th); false)
blanchet@37344
   429
  | is_named _ = true
blanchet@37344
   430
fun checked_name_thm_pairs respect_no_atp ctxt =
blanchet@37344
   431
  name_thm_pairs respect_no_atp ctxt
blanchet@37344
   432
  #> tap (fn ps => trace_msg
blanchet@37344
   433
                        (fn () => ("Considering " ^ Int.toString (length ps) ^
blanchet@37344
   434
                                   " theorems")))
blanchet@37344
   435
  #> filter is_named
paulson@19894
   436
blanchet@37344
   437
fun name_thms_pair_from_ref ctxt chained_ths xref =
blanchet@37344
   438
  let
blanchet@37344
   439
    val ths = ProofContext.get_fact ctxt xref
blanchet@37344
   440
    val name = Facts.string_of_ref xref
blanchet@37344
   441
               |> forall (member Thm.eq_thm chained_ths) ths
blanchet@37344
   442
                  ? prefix chained_prefix
blanchet@37344
   443
  in (name, ths) end
blanchet@37344
   444
mengj@19768
   445
paulson@21290
   446
(***************************************************************)
paulson@21290
   447
(* Type Classes Present in the Axiom or Conjecture Clauses     *)
paulson@21290
   448
(***************************************************************)
paulson@21290
   449
wenzelm@32952
   450
fun add_classes (sorts, cset) = List.foldl setinsert cset (flat sorts);
paulson@21290
   451
paulson@21290
   452
(*Remove this trivial type class*)
blanchet@35865
   453
fun delete_type cset = Symtab.delete_safe (the_single @{sort HOL.type}) cset;
paulson@21290
   454
paulson@21290
   455
fun tvar_classes_of_terms ts =
wenzelm@29270
   456
  let val sorts_list = map (map #2 o OldTerm.term_tvars) ts
wenzelm@30190
   457
  in  Symtab.keys (delete_type (List.foldl add_classes Symtab.empty sorts_list))  end;
paulson@21290
   458
paulson@21290
   459
fun tfree_classes_of_terms ts =
wenzelm@29270
   460
  let val sorts_list = map (map #2 o OldTerm.term_tfrees) ts
wenzelm@30190
   461
  in  Symtab.keys (delete_type (List.foldl add_classes Symtab.empty sorts_list))  end;
paulson@20526
   462
paulson@21373
   463
(*fold type constructors*)
paulson@21373
   464
fun fold_type_consts f (Type (a, Ts)) x = fold (fold_type_consts f) Ts (f (a,x))
wenzelm@32994
   465
  | fold_type_consts _ _ x = x;
paulson@21373
   466
paulson@21397
   467
(*Type constructors used to instantiate overloaded constants are the only ones needed.*)
paulson@21397
   468
fun add_type_consts_in_term thy =
blanchet@37504
   469
  let
blanchet@37504
   470
    val const_typargs = Sign.const_typargs thy
blanchet@37504
   471
    fun aux (Const cT) = fold (fold_type_consts setinsert) (const_typargs cT)
blanchet@37504
   472
      | aux (Abs (_, _, u)) = aux u
blanchet@37504
   473
      | aux (Const (@{const_name skolem_id}, _) $ _) = I
blanchet@37504
   474
      | aux (t $ u) = aux t #> aux u
blanchet@37504
   475
      | aux _ = I
blanchet@37504
   476
  in aux end
paulson@21373
   477
paulson@21397
   478
fun type_consts_of_terms thy ts =
paulson@21397
   479
  Symtab.keys (fold (add_type_consts_in_term thy) ts Symtab.empty);
paulson@21373
   480
paulson@21373
   481
mengj@19194
   482
(***************************************************************)
mengj@19194
   483
(* ATP invocation methods setup                                *)
mengj@19194
   484
(***************************************************************)
mengj@19194
   485
blanchet@37410
   486
fun is_quasi_fol_term thy =
blanchet@37410
   487
  Meson.is_fol_term thy o snd o conceal_skolem_somes ~1 []
blanchet@37399
   488
paulson@20526
   489
(*Ensures that no higher-order theorems "leak out"*)
blanchet@37399
   490
fun restrict_to_logic thy true cls =
blanchet@37399
   491
    filter (is_quasi_fol_term thy o prop_of o fst) cls
blanchet@37498
   492
  | restrict_to_logic _ false cls = cls
paulson@20526
   493
blanchet@37347
   494
(**** Predicates to detect unwanted clauses (prolific or likely to cause
blanchet@37347
   495
      unsoundness) ****)
paulson@21470
   496
paulson@21470
   497
(** Too general means, positive equality literal with a variable X as one operand,
paulson@21470
   498
  when X does not occur properly in the other operand. This rules out clearly
paulson@21470
   499
  inconsistent clauses such as V=a|V=b, though it by no means guarantees soundness. **)
wenzelm@21588
   500
blanchet@37348
   501
fun var_occurs_in_term ix =
blanchet@37348
   502
  let
blanchet@37348
   503
    fun aux (Var (jx, _)) = (ix = jx)
blanchet@37348
   504
      | aux (t1 $ t2) = aux t1 orelse aux t2
blanchet@37348
   505
      | aux (Abs (_, _, t)) = aux t
blanchet@37348
   506
      | aux _ = false
blanchet@37348
   507
  in aux end
paulson@21470
   508
blanchet@37348
   509
fun is_record_type T = not (null (Record.dest_recTs T))
paulson@21470
   510
paulson@21470
   511
(*Unwanted equalities include
paulson@21470
   512
  (1) those between a variable that does not properly occur in the second operand,
paulson@21470
   513
  (2) those between a variable and a record, since these seem to be prolific "cases" thms
wenzelm@21588
   514
*)
blanchet@37348
   515
fun too_general_eqterms (Var (ix,T), t) =
blanchet@37348
   516
    not (var_occurs_in_term ix t) orelse is_record_type T
paulson@21470
   517
  | too_general_eqterms _ = false;
paulson@21470
   518
blanchet@35865
   519
fun too_general_equality (Const (@{const_name "op ="}, _) $ x $ y) =
paulson@21470
   520
      too_general_eqterms (x,y) orelse too_general_eqterms(y,x)
paulson@21470
   521
  | too_general_equality _ = false;
paulson@21470
   522
wenzelm@29267
   523
fun has_typed_var tycons = exists_subterm
wenzelm@29267
   524
  (fn Var (_, Type (a, _)) => member (op =) tycons a | _ => false);
paulson@21431
   525
blanchet@37347
   526
(* Clauses are forbidden to contain variables of these types. The typical reason
blanchet@37347
   527
   is that they lead to unsoundness. Note that "unit" satisfies numerous
blanchet@37347
   528
   equations like "?x = ()". The resulting clause will have no type constraint,
blanchet@37347
   529
   yielding false proofs. Even "bool" leads to many unsound proofs, though only
blanchet@37347
   530
   for higher-order problems. *)
blanchet@37347
   531
val dangerous_types = [@{type_name unit}, @{type_name bool}];
paulson@22217
   532
blanchet@37347
   533
(* Clauses containing variables of type "unit" or "bool" or of the form
blanchet@37347
   534
   "?x = A | ?x = B | ?x = C" are likely to lead to unsound proofs if types are
blanchet@37347
   535
   omitted. *)
blanchet@37347
   536
fun is_dangerous_term _ @{prop True} = true
blanchet@37347
   537
  | is_dangerous_term full_types t =
blanchet@37505
   538
    not full_types andalso
blanchet@37347
   539
    (has_typed_var dangerous_types t orelse
blanchet@37347
   540
     forall too_general_equality (HOLogic.disjuncts (strip_Trueprop t)))
paulson@21470
   541
blanchet@37347
   542
fun remove_dangerous_clauses full_types add_thms =
blanchet@37501
   543
  filter_out (fn (cnf_th, (_, orig_th)) =>
blanchet@37501
   544
                 not (member Thm.eq_thm add_thms orig_th) andalso
blanchet@37501
   545
                 is_dangerous_term full_types (prop_of cnf_th))
paulson@21431
   546
blanchet@37399
   547
fun is_fol_goal thy = forall (Meson.is_fol_term thy) o map prop_of
immler@30536
   548
blanchet@37347
   549
fun relevant_facts full_types respect_no_atp relevance_threshold
blanchet@37347
   550
                   relevance_convergence defs_relevant max_new theory_relevant
blanchet@37347
   551
                   (relevance_override as {add, del, only})
blanchet@37347
   552
                   (ctxt, (chained_ths, _)) goal_cls =
blanchet@36185
   553
  if (only andalso null add) orelse relevance_threshold > 1.0 then
blanchet@36185
   554
    []
blanchet@36185
   555
  else
blanchet@36185
   556
    let
blanchet@36185
   557
      val thy = ProofContext.theory_of ctxt
blanchet@37345
   558
      val has_override = not (null add) orelse not (null del)
blanchet@37399
   559
      val is_FO = is_fol_goal thy goal_cls
blanchet@37345
   560
      val axioms =
blanchet@37399
   561
        checked_name_thm_pairs (respect_no_atp andalso not only) ctxt
blanchet@37344
   562
            (if only then map (name_thms_pair_from_ref ctxt chained_ths) add
blanchet@37345
   563
             else all_name_thms_pairs respect_no_atp ctxt chained_ths)
blanchet@37345
   564
        |> cnf_rules_pairs thy
blanchet@37345
   565
        |> not has_override ? make_unique
blanchet@37399
   566
        |> not only ? restrict_to_logic thy is_FO
blanchet@37501
   567
        |> (if only then
blanchet@37501
   568
              I
blanchet@37501
   569
            else
blanchet@37501
   570
              remove_dangerous_clauses full_types
blanchet@37501
   571
                                       (maps (ProofContext.get_fact ctxt) add))
blanchet@36185
   572
    in
blanchet@36922
   573
      relevance_filter ctxt relevance_threshold relevance_convergence
blanchet@36922
   574
                       defs_relevant max_new theory_relevant relevance_override
blanchet@37345
   575
                       thy axioms (map prop_of goal_cls)
blanchet@37345
   576
      |> has_override ? make_unique
blanchet@36185
   577
    end
immler@30536
   578
immler@31752
   579
(* prepare for passing to writer,
immler@31752
   580
   create additional clauses based on the information from extra_cls *)
blanchet@37498
   581
fun prepare_clauses full_types goal_cls axcls extra_cls thy =
immler@31409
   582
  let
blanchet@37399
   583
    val is_FO = is_fol_goal thy goal_cls
blanchet@36061
   584
    val ccls = subtract_cls extra_cls goal_cls
blanchet@35865
   585
    val _ = app (fn th => trace_msg (fn _ => Display.string_of_thm_global thy th)) ccls
immler@30536
   586
    val ccltms = map prop_of ccls
immler@31752
   587
    and axtms = map (prop_of o #1) extra_cls
immler@30536
   588
    val subs = tfree_classes_of_terms ccltms
immler@30536
   589
    and supers = tvar_classes_of_terms axtms
blanchet@35865
   590
    and tycons = type_consts_of_terms thy (ccltms @ axtms)
immler@30536
   591
    (*TFrees in conjecture clauses; TVars in axiom clauses*)
blanchet@37498
   592
    val conjectures = make_conjecture_clauses thy ccls
blanchet@37498
   593
    val (_, extra_clauses) = ListPair.unzip (make_axiom_clauses thy extra_cls)
blanchet@37498
   594
    val (clnames, axiom_clauses) = ListPair.unzip (make_axiom_clauses thy axcls)
blanchet@37479
   595
    val helper_clauses =
blanchet@37498
   596
      get_helper_clauses thy is_FO full_types conjectures extra_cls
blanchet@37498
   597
    val (supers', arity_clauses) = make_arity_clauses thy tycons supers
blanchet@35865
   598
    val classrel_clauses = make_classrel_clauses thy subs supers'
immler@30536
   599
  in
immler@31752
   600
    (Vector.fromList clnames,
immler@31865
   601
      (conjectures, axiom_clauses, extra_clauses, helper_clauses, classrel_clauses, arity_clauses))
immler@31409
   602
  end
quigley@15644
   603
paulson@15347
   604
end;