src/HOL/Tools/Sledgehammer/sledgehammer_fact_filter.ML
author blanchet
Tue Mar 23 11:39:21 2010 +0100 (2010-03-23)
changeset 35963 943e2582dc87
parent 35865 2f8fb5242799
child 35966 f8c738abaed8
permissions -rw-r--r--
added options to Sledgehammer;
syntax: sledgehammer [options] goal_no, where "[options]" and "goal_no" are optional
blanchet@35826
     1
(*  Title:      HOL/Tools/Sledgehammer/sledgehammer_fact_filter.ML
wenzelm@33309
     2
    Author:     Jia Meng, Cambridge University Computer Laboratory, NICTA
wenzelm@33309
     3
*)
paulson@15452
     4
blanchet@35826
     5
signature SLEDGEHAMMER_FACT_FILTER =
wenzelm@16802
     6
sig
blanchet@35865
     7
  type classrel_clause = Sledgehammer_FOL_Clause.classrel_clause
blanchet@35865
     8
  type arity_clause = Sledgehammer_FOL_Clause.arity_clause
blanchet@35826
     9
  type axiom_name = Sledgehammer_HOL_Clause.axiom_name
blanchet@35865
    10
  type hol_clause = Sledgehammer_HOL_Clause.hol_clause
blanchet@35865
    11
  type hol_clause_id = Sledgehammer_HOL_Clause.hol_clause_id
paulson@22989
    12
  val tvar_classes_of_terms : term list -> string list
paulson@22989
    13
  val tfree_classes_of_terms : term list -> string list
paulson@22989
    14
  val type_consts_of_terms : theory -> term list -> string list
blanchet@35963
    15
  val get_relevant_facts :
blanchet@35963
    16
    real -> bool option -> bool -> int -> bool
blanchet@35963
    17
    -> Proof.context * (thm list * 'a) -> thm list
blanchet@35963
    18
    -> (thm * (string * int)) list
blanchet@35963
    19
  val prepare_clauses : bool option -> bool -> thm list -> thm list ->
blanchet@35865
    20
    (thm * (axiom_name * hol_clause_id)) list ->
blanchet@35865
    21
    (thm * (axiom_name * hol_clause_id)) list -> theory ->
blanchet@35826
    22
    axiom_name vector *
blanchet@35865
    23
      (hol_clause list * hol_clause list * hol_clause list *
blanchet@35865
    24
      hol_clause list * classrel_clause list * arity_clause list)
paulson@15347
    25
end;
paulson@15347
    26
blanchet@35826
    27
structure Sledgehammer_Fact_Filter : SLEDGEHAMMER_FACT_FILTER =
paulson@15347
    28
struct
paulson@15347
    29
blanchet@35865
    30
open Sledgehammer_FOL_Clause
blanchet@35865
    31
open Sledgehammer_Fact_Preprocessor
blanchet@35865
    32
open Sledgehammer_HOL_Clause
blanchet@35826
    33
blanchet@35865
    34
type axiom_name = axiom_name
blanchet@35865
    35
type hol_clause = hol_clause
blanchet@35865
    36
type hol_clause_id = hol_clause_id
paulson@21070
    37
mengj@19194
    38
(********************************************************************)
mengj@19194
    39
(* some settings for both background automatic ATP calling procedure*)
mengj@19194
    40
(* and also explicit ATP invocation methods                         *)
mengj@19194
    41
(********************************************************************)
mengj@19194
    42
mengj@19194
    43
(*** background linkup ***)
wenzelm@28477
    44
val run_blacklist_filter = true;
paulson@21224
    45
paulson@24287
    46
(*** relevance filter parameters ***)
wenzelm@28477
    47
val convergence = 3.2;    (*Higher numbers allow longer inference chains*)
wenzelm@28477
    48
  
wenzelm@28477
    49
(***************************************************************)
wenzelm@28477
    50
(* Relevance Filtering                                         *)
wenzelm@28477
    51
(***************************************************************)
mengj@19194
    52
blanchet@35865
    53
fun strip_Trueprop (@{const Trueprop} $ t) = t
paulson@24958
    54
  | strip_Trueprop t = t;
mengj@19194
    55
paulson@24287
    56
(*A surprising number of theorems contain only a few significant constants.
paulson@24287
    57
  These include all induction rules, and other general theorems. Filtering
paulson@24287
    58
  theorems in clause form reveals these complexities in the form of Skolem 
paulson@24287
    59
  functions. If we were instead to filter theorems in their natural form,
paulson@24287
    60
  some other method of measuring theorem complexity would become necessary.*)
paulson@24287
    61
paulson@24287
    62
fun log_weight2 (x:real) = 1.0 + 2.0/Math.ln (x+1.0);
paulson@24287
    63
paulson@24287
    64
(*The default seems best in practice. A constant function of one ignores
paulson@24287
    65
  the constant frequencies.*)
wenzelm@28477
    66
val weight_fn = log_weight2;
paulson@24287
    67
paulson@24287
    68
paulson@24287
    69
(*Including equality in this list might be expected to stop rules like subset_antisym from
paulson@24287
    70
  being chosen, but for some reason filtering works better with them listed. The
paulson@24287
    71
  logical signs All, Ex, &, and --> are omitted because any remaining occurrrences
paulson@24287
    72
  must be within comprehensions.*)
blanchet@35865
    73
val standard_consts =
blanchet@35865
    74
  [@{const_name Trueprop}, @{const_name "==>"}, @{const_name all},
blanchet@35865
    75
   @{const_name "=="}, @{const_name "op |"}, @{const_name Not},
blanchet@35865
    76
   @{const_name "op ="}];
paulson@24287
    77
paulson@24287
    78
paulson@24287
    79
(*** constants with types ***)
paulson@24287
    80
paulson@24287
    81
(*An abstraction of Isabelle types*)
paulson@24287
    82
datatype const_typ =  CTVar | CType of string * const_typ list
paulson@24287
    83
paulson@24287
    84
(*Is the second type an instance of the first one?*)
paulson@24287
    85
fun match_type (CType(con1,args1)) (CType(con2,args2)) = 
paulson@24287
    86
      con1=con2 andalso match_types args1 args2
paulson@24287
    87
  | match_type CTVar _ = true
paulson@24287
    88
  | match_type _ CTVar = false
paulson@24287
    89
and match_types [] [] = true
paulson@24287
    90
  | match_types (a1::as1) (a2::as2) = match_type a1 a2 andalso match_types as1 as2;
paulson@24287
    91
paulson@24287
    92
(*Is there a unifiable constant?*)
paulson@24287
    93
fun uni_mem gctab (c,c_typ) =
paulson@24287
    94
  case Symtab.lookup gctab c of
paulson@24287
    95
      NONE => false
paulson@24287
    96
    | SOME ctyps_list => exists (match_types c_typ) ctyps_list;
paulson@24287
    97
  
paulson@24287
    98
(*Maps a "real" type to a const_typ*)
paulson@24287
    99
fun const_typ_of (Type (c,typs)) = CType (c, map const_typ_of typs) 
paulson@24287
   100
  | const_typ_of (TFree _) = CTVar
paulson@24287
   101
  | const_typ_of (TVar _) = CTVar
paulson@24287
   102
paulson@24287
   103
(*Pairs a constant with the list of its type instantiations (using const_typ)*)
paulson@24287
   104
fun const_with_typ thy (c,typ) = 
paulson@24287
   105
    let val tvars = Sign.const_typargs thy (c,typ)
paulson@24287
   106
    in (c, map const_typ_of tvars) end
paulson@24287
   107
    handle TYPE _ => (c,[]);   (*Variable (locale constant): monomorphic*)   
paulson@24287
   108
paulson@24287
   109
(*Add a const/type pair to the table, but a [] entry means a standard connective,
paulson@24287
   110
  which we ignore.*)
paulson@24287
   111
fun add_const_typ_table ((c,ctyps), tab) =
paulson@24287
   112
  Symtab.map_default (c, [ctyps]) (fn [] => [] | ctyps_list => insert (op =) ctyps ctyps_list) 
paulson@24287
   113
    tab;
paulson@24287
   114
paulson@24287
   115
(*Free variables are included, as well as constants, to handle locales*)
paulson@24287
   116
fun add_term_consts_typs_rm thy (Const(c, typ), tab) =
paulson@24287
   117
      add_const_typ_table (const_with_typ thy (c,typ), tab) 
paulson@24287
   118
  | add_term_consts_typs_rm thy (Free(c, typ), tab) =
paulson@24287
   119
      add_const_typ_table (const_with_typ thy (c,typ), tab) 
paulson@24287
   120
  | add_term_consts_typs_rm thy (t $ u, tab) =
paulson@24287
   121
      add_term_consts_typs_rm thy (t, add_term_consts_typs_rm thy (u, tab))
paulson@24287
   122
  | add_term_consts_typs_rm thy (Abs(_,_,t), tab) = add_term_consts_typs_rm thy (t, tab)
wenzelm@32994
   123
  | add_term_consts_typs_rm _ (_, tab) = tab;
paulson@24287
   124
paulson@24287
   125
(*The empty list here indicates that the constant is being ignored*)
paulson@24287
   126
fun add_standard_const (s,tab) = Symtab.update (s,[]) tab;
paulson@24287
   127
paulson@24287
   128
val null_const_tab : const_typ list list Symtab.table = 
wenzelm@30190
   129
    List.foldl add_standard_const Symtab.empty standard_consts;
paulson@24287
   130
wenzelm@30190
   131
fun get_goal_consts_typs thy = List.foldl (add_term_consts_typs_rm thy) null_const_tab;
paulson@24287
   132
paulson@24287
   133
(*Inserts a dummy "constant" referring to the theory name, so that relevance
paulson@24287
   134
  takes the given theory into account.*)
blanchet@35963
   135
fun const_prop_of add_theory_const th =
blanchet@35963
   136
 if add_theory_const then
paulson@24287
   137
  let val name = Context.theory_name (theory_of_thm th)
paulson@24287
   138
      val t = Const (name ^ ". 1", HOLogic.boolT)
paulson@24287
   139
  in  t $ prop_of th  end
paulson@24287
   140
 else prop_of th;
paulson@24287
   141
paulson@24287
   142
(**** Constant / Type Frequencies ****)
paulson@24287
   143
paulson@24287
   144
(*A two-dimensional symbol table counts frequencies of constants. It's keyed first by
paulson@24287
   145
  constant name and second by its list of type instantiations. For the latter, we need
paulson@24287
   146
  a linear ordering on type const_typ list.*)
paulson@24287
   147
  
paulson@24287
   148
local
paulson@24287
   149
paulson@24287
   150
fun cons_nr CTVar = 0
paulson@24287
   151
  | cons_nr (CType _) = 1;
paulson@24287
   152
paulson@24287
   153
in
paulson@24287
   154
paulson@24287
   155
fun const_typ_ord TU =
paulson@24287
   156
  case TU of
paulson@24287
   157
    (CType (a, Ts), CType (b, Us)) =>
paulson@24287
   158
      (case fast_string_ord(a,b) of EQUAL => dict_ord const_typ_ord (Ts,Us) | ord => ord)
paulson@24287
   159
  | (T, U) => int_ord (cons_nr T, cons_nr U);
paulson@24287
   160
paulson@24287
   161
end;
paulson@24287
   162
wenzelm@31971
   163
structure CTtab = Table(type key = const_typ list val ord = dict_ord const_typ_ord);
paulson@24287
   164
blanchet@35963
   165
fun count_axiom_consts add_theory_const thy ((thm,_), tab) = 
paulson@24287
   166
  let fun count_const (a, T, tab) =
wenzelm@32960
   167
        let val (c, cts) = const_with_typ thy (a,T)
wenzelm@32960
   168
        in  (*Two-dimensional table update. Constant maps to types maps to count.*)
wenzelm@32960
   169
            Symtab.map_default (c, CTtab.empty) 
wenzelm@32960
   170
                               (CTtab.map_default (cts,0) (fn n => n+1)) tab
wenzelm@32960
   171
        end
paulson@24287
   172
      fun count_term_consts (Const(a,T), tab) = count_const(a,T,tab)
wenzelm@32960
   173
        | count_term_consts (Free(a,T), tab) = count_const(a,T,tab)
wenzelm@32960
   174
        | count_term_consts (t $ u, tab) =
wenzelm@32960
   175
            count_term_consts (t, count_term_consts (u, tab))
wenzelm@32960
   176
        | count_term_consts (Abs(_,_,t), tab) = count_term_consts (t, tab)
wenzelm@32960
   177
        | count_term_consts (_, tab) = tab
blanchet@35963
   178
  in  count_term_consts (const_prop_of add_theory_const thm, tab)  end;
paulson@24287
   179
paulson@24287
   180
paulson@24287
   181
(**** Actual Filtering Code ****)
paulson@24287
   182
paulson@24287
   183
(*The frequency of a constant is the sum of those of all instances of its type.*)
paulson@24287
   184
fun const_frequency ctab (c, cts) =
paulson@24287
   185
  let val pairs = CTtab.dest (the (Symtab.lookup ctab c))
paulson@24287
   186
      fun add ((cts',m), n) = if match_types cts cts' then m+n else n
paulson@24287
   187
  in  List.foldl add 0 pairs  end;
paulson@24287
   188
paulson@24287
   189
(*Add in a constant's weight, as determined by its frequency.*)
paulson@24287
   190
fun add_ct_weight ctab ((c,T), w) =
wenzelm@28477
   191
  w + weight_fn (real (const_frequency ctab (c,T)));
paulson@24287
   192
paulson@24287
   193
(*Relevant constants are weighted according to frequency, 
paulson@24287
   194
  but irrelevant constants are simply counted. Otherwise, Skolem functions,
paulson@24287
   195
  which are rare, would harm a clause's chances of being picked.*)
paulson@24287
   196
fun clause_weight ctab gctyps consts_typs =
paulson@24287
   197
    let val rel = filter (uni_mem gctyps) consts_typs
paulson@24287
   198
        val rel_weight = List.foldl (add_ct_weight ctab) 0.0 rel
paulson@24287
   199
    in
wenzelm@32960
   200
        rel_weight / (rel_weight + real (length consts_typs - length rel))
paulson@24287
   201
    end;
paulson@24287
   202
    
paulson@24287
   203
(*Multiplies out to a list of pairs: 'a * 'b list -> ('a * 'b) list -> ('a * 'b) list*)
wenzelm@30190
   204
fun add_expand_pairs (x,ys) xys = List.foldl (fn (y,acc) => (x,y)::acc) xys ys;
paulson@24287
   205
paulson@24287
   206
fun consts_typs_of_term thy t = 
paulson@24287
   207
  let val tab = add_term_consts_typs_rm thy (t, null_const_tab)
paulson@24287
   208
  in  Symtab.fold add_expand_pairs tab []  end;
paulson@24287
   209
blanchet@35963
   210
fun pair_consts_typs_axiom add_theory_const thy (p as (thm, _)) =
blanchet@35963
   211
  (p, (consts_typs_of_term thy (const_prop_of add_theory_const thm)));
paulson@24287
   212
paulson@24287
   213
exception ConstFree;
paulson@24287
   214
fun dest_ConstFree (Const aT) = aT
paulson@24287
   215
  | dest_ConstFree (Free aT) = aT
paulson@24287
   216
  | dest_ConstFree _ = raise ConstFree;
paulson@24287
   217
paulson@24287
   218
(*Look for definitions of the form f ?x1 ... ?xn = t, but not reversed.*)
wenzelm@32994
   219
fun defines thy thm gctypes =
paulson@24287
   220
    let val tm = prop_of thm
wenzelm@32960
   221
        fun defs lhs rhs =
paulson@24287
   222
            let val (rator,args) = strip_comb lhs
wenzelm@32960
   223
                val ct = const_with_typ thy (dest_ConstFree rator)
haftmann@33037
   224
            in
haftmann@33037
   225
              forall is_Var args andalso uni_mem gctypes ct andalso
haftmann@33038
   226
                subset (op =) (Term.add_vars rhs [], Term.add_vars lhs [])
paulson@24287
   227
            end
wenzelm@32960
   228
            handle ConstFree => false
paulson@24287
   229
    in    
blanchet@35963
   230
        case tm of
blanchet@35963
   231
          @{const Trueprop} $ (Const (@{const_name "op ="}, _) $ lhs $ rhs) => 
blanchet@35963
   232
            defs lhs rhs 
blanchet@35963
   233
        | _ => false
paulson@24287
   234
    end;
paulson@24287
   235
paulson@24287
   236
type annotd_cls = (thm * (string * int)) * ((string * const_typ list) list);
paulson@24287
   237
       
paulson@24287
   238
(*For a reverse sort, putting the largest values first.*)
paulson@24287
   239
fun compare_pairs ((_,w1),(_,w2)) = Real.compare (w2,w1);
paulson@24287
   240
paulson@24287
   241
(*Limit the number of new clauses, to prevent runaway acceptance.*)
wenzelm@28477
   242
fun take_best max_new (newpairs : (annotd_cls*real) list) =
paulson@24287
   243
  let val nnew = length newpairs
paulson@24287
   244
  in
wenzelm@28477
   245
    if nnew <= max_new then (map #1 newpairs, [])
paulson@24287
   246
    else 
paulson@24287
   247
      let val cls = sort compare_pairs newpairs
wenzelm@28477
   248
          val accepted = List.take (cls, max_new)
paulson@24287
   249
      in
blanchet@35865
   250
        trace_msg (fn () => ("Number of candidates, " ^ Int.toString nnew ^ 
wenzelm@32960
   251
                       ", exceeds the limit of " ^ Int.toString (max_new)));
blanchet@35865
   252
        trace_msg (fn () => ("Effective pass mark: " ^ Real.toString (#2 (List.last accepted))));
blanchet@35865
   253
        trace_msg (fn () => "Actually passed: " ^
paulson@24287
   254
          space_implode ", " (map (fn (((_,(name,_)),_),_) => name) accepted));
paulson@24287
   255
wenzelm@32960
   256
        (map #1 accepted, map #1 (List.drop (cls, max_new)))
paulson@24287
   257
      end
paulson@24287
   258
  end;
paulson@24287
   259
blanchet@35963
   260
fun relevant_clauses follow_defs max_new thy ctab p rel_consts =
paulson@24287
   261
  let fun relevant ([],_) [] = [] : (thm * (string * int)) list  (*Nothing added this iteration*)
wenzelm@32960
   262
        | relevant (newpairs,rejects) [] =
wenzelm@32960
   263
            let val (newrels,more_rejects) = take_best max_new newpairs
wenzelm@32960
   264
                val new_consts = maps #2 newrels
wenzelm@32960
   265
                val rel_consts' = List.foldl add_const_typ_table rel_consts new_consts
wenzelm@32960
   266
                val newp = p + (1.0-p) / convergence
wenzelm@32960
   267
            in
blanchet@35865
   268
              trace_msg (fn () => "relevant this iteration: " ^ Int.toString (length newrels));
wenzelm@32960
   269
               (map #1 newrels) @ 
blanchet@35963
   270
               (relevant_clauses follow_defs max_new thy ctab newp rel_consts'
blanchet@35963
   271
                                 (more_rejects @ rejects))
wenzelm@32960
   272
            end
wenzelm@32960
   273
        | relevant (newrels,rejects) ((ax as (clsthm as (_,(name,n)),consts_typs)) :: axs) =
wenzelm@32960
   274
            let val weight = clause_weight ctab rel_consts consts_typs
wenzelm@32960
   275
            in
wenzelm@32994
   276
              if p <= weight orelse (follow_defs andalso defines thy (#1 clsthm) rel_consts)
blanchet@35865
   277
              then (trace_msg (fn () => (name ^ " clause " ^ Int.toString n ^ 
wenzelm@32960
   278
                                            " passes: " ^ Real.toString weight));
wenzelm@32960
   279
                    relevant ((ax,weight)::newrels, rejects) axs)
wenzelm@32960
   280
              else relevant (newrels, ax::rejects) axs
wenzelm@32960
   281
            end
blanchet@35865
   282
    in  trace_msg (fn () => ("relevant_clauses, current pass mark = " ^ Real.toString p));
paulson@24287
   283
        relevant ([],[]) 
paulson@24287
   284
    end;
wenzelm@32960
   285
        
blanchet@35963
   286
fun relevance_filter relevance_threshold follow_defs max_new add_theory_const
blanchet@35963
   287
                     thy axioms goals = 
blanchet@35963
   288
  if relevance_threshold >= 0.1 then
blanchet@35963
   289
    let
blanchet@35963
   290
      val const_tab = List.foldl (count_axiom_consts add_theory_const thy)
blanchet@35963
   291
                                 Symtab.empty axioms
paulson@24287
   292
      val goal_const_tab = get_goal_consts_typs thy goals
blanchet@35963
   293
      val _ =
blanchet@35963
   294
        trace_msg (fn () => "Initial constants: " ^
blanchet@35963
   295
                            commas (Symtab.keys goal_const_tab))
blanchet@35963
   296
      val relevant =
blanchet@35963
   297
        relevant_clauses follow_defs max_new thy const_tab
blanchet@35963
   298
                         relevance_threshold goal_const_tab
blanchet@35963
   299
                         (map (pair_consts_typs_axiom add_theory_const thy)
blanchet@35963
   300
                              axioms)
blanchet@35963
   301
    in
blanchet@35963
   302
      trace_msg (fn () => "Total relevant: " ^ Int.toString (length relevant));
blanchet@35963
   303
      relevant
blanchet@35963
   304
    end
blanchet@35963
   305
  else
blanchet@35963
   306
    axioms;
paulson@24287
   307
paulson@24287
   308
(***************************************************************)
mengj@19768
   309
(* Retrieving and filtering lemmas                             *)
mengj@19768
   310
(***************************************************************)
mengj@19768
   311
paulson@33022
   312
(*** retrieve lemmas and filter them ***)
mengj@19768
   313
mengj@19768
   314
(*Hashing to detect duplicate and variant clauses, e.g. from the [iff] attribute*)
mengj@19768
   315
paulson@22382
   316
fun setinsert (x,s) = Symtab.update (x,()) s;
mengj@19768
   317
paulson@20757
   318
(*Reject theorems with names like "List.filter.filter_list_def" or
paulson@21690
   319
  "Accessible_Part.acc.defs", as these are definitions arising from packages.*)
paulson@20757
   320
fun is_package_def a =
wenzelm@30364
   321
  let val names = Long_Name.explode a
paulson@21690
   322
  in
paulson@21690
   323
     length names > 2 andalso
paulson@21690
   324
     not (hd names = "local") andalso
paulson@21690
   325
     String.isSuffix "_def" a  orelse  String.isSuffix "_defs" a
paulson@21690
   326
  end;
paulson@20757
   327
mengj@19768
   328
(** a hash function from Term.term to int, and also a hash table **)
mengj@19768
   329
val xor_words = List.foldl Word.xorb 0w0;
mengj@19768
   330
mengj@19768
   331
fun hashw_term ((Const(c,_)), w) = Polyhash.hashw_string (c,w)
paulson@20661
   332
  | hashw_term ((Free(a,_)), w) = Polyhash.hashw_string (a,w)
mengj@19768
   333
  | hashw_term ((Var(_,_)), w) = w
paulson@20661
   334
  | hashw_term ((Bound i), w) = Polyhash.hashw_int(i,w)
mengj@19768
   335
  | hashw_term ((Abs(_,_,t)), w) = hashw_term (t, w)
mengj@19768
   336
  | hashw_term ((P$Q), w) = hashw_term (Q, (hashw_term (P, w)));
mengj@19768
   337
blanchet@35865
   338
fun hash_literal (@{const Not} $ P) = Word.notb(hashw_term(P,0w0))
paulson@21070
   339
  | hash_literal P = hashw_term(P,0w0);
mengj@19768
   340
paulson@24958
   341
fun hash_term t = Word.toIntX (xor_words (map hash_literal (HOLogic.disjuncts (strip_Trueprop t))));
mengj@19768
   342
mengj@19768
   343
fun equal_thm (thm1,thm2) = Term.aconv(prop_of thm1, prop_of thm2);
paulson@20457
   344
paulson@22382
   345
exception HASH_CLAUSE;
paulson@22382
   346
mengj@19768
   347
(*Create a hash table for clauses, of the given size*)
mengj@19768
   348
fun mk_clause_table n =
paulson@20457
   349
      Polyhash.mkTable (hash_term o prop_of, equal_thm)
mengj@19768
   350
                       (n, HASH_CLAUSE);
mengj@19768
   351
paulson@20457
   352
(*Use a hash table to eliminate duplicates from xs. Argument is a list of
paulson@20868
   353
  (thm * (string * int)) tuples. The theorems are hashed into the table. *)
wenzelm@21588
   354
fun make_unique xs =
paulson@20868
   355
  let val ht = mk_clause_table 7000
paulson@20457
   356
  in
blanchet@35865
   357
      trace_msg (fn () => ("make_unique gets " ^ Int.toString (length xs) ^ " clauses"));
wenzelm@21588
   358
      app (ignore o Polyhash.peekInsert ht) xs;
paulson@20868
   359
      Polyhash.listItems ht
paulson@20457
   360
  end;
mengj@19768
   361
paulson@20868
   362
(*Remove existing axiom clauses from the conjecture clauses, as this can dramatically
paulson@20868
   363
  boost an ATP's performance (for some reason)*)
wenzelm@21588
   364
fun subtract_cls c_clauses ax_clauses =
paulson@20868
   365
  let val ht = mk_clause_table 2200
wenzelm@33035
   366
      fun known x = is_some (Polyhash.peek ht x)
paulson@20868
   367
  in
wenzelm@21588
   368
      app (ignore o Polyhash.peekInsert ht) ax_clauses;
wenzelm@21588
   369
      filter (not o known) c_clauses
paulson@20868
   370
  end;
mengj@19768
   371
paulson@21224
   372
fun all_valid_thms ctxt =
paulson@22382
   373
  let
wenzelm@26675
   374
    val global_facts = PureThy.facts_of (ProofContext.theory_of ctxt);
wenzelm@26278
   375
    val local_facts = ProofContext.facts_of ctxt;
wenzelm@33641
   376
    val full_space =
wenzelm@33641
   377
      Name_Space.merge (Facts.space_of global_facts, Facts.space_of local_facts);
wenzelm@33641
   378
wenzelm@33641
   379
    fun valid_facts facts =
wenzelm@33641
   380
      (facts, []) |-> Facts.fold_static (fn (name, ths0) =>
wenzelm@33641
   381
        let
wenzelm@33641
   382
          fun check_thms a =
wenzelm@33641
   383
            (case try (ProofContext.get_thms ctxt) a of
wenzelm@33641
   384
              NONE => false
wenzelm@33641
   385
            | SOME ths1 => Thm.eq_thms (ths0, ths1));
wenzelm@33641
   386
wenzelm@33641
   387
          val name1 = Facts.extern facts name;
wenzelm@33641
   388
          val name2 = Name_Space.extern full_space name;
blanchet@35865
   389
          val ths = filter_out bad_for_atp ths0;
wenzelm@33641
   390
        in
wenzelm@33641
   391
          if Facts.is_concealed facts name orelse null ths orelse
wenzelm@33641
   392
            run_blacklist_filter andalso is_package_def name then I
wenzelm@33641
   393
          else
wenzelm@33641
   394
            (case find_first check_thms [name1, name2, name] of
wenzelm@33641
   395
              NONE => I
wenzelm@33641
   396
            | SOME a => cons (a, ths))
wenzelm@33641
   397
        end);
wenzelm@26675
   398
  in valid_facts global_facts @ valid_facts local_facts end;
paulson@21224
   399
wenzelm@33309
   400
fun multi_name a th (n, pairs) =
wenzelm@33309
   401
  (n + 1, (a ^ "(" ^ Int.toString n ^ ")", th) :: pairs);
paulson@21224
   402
wenzelm@33309
   403
fun add_single_names (a, []) pairs = pairs
wenzelm@33309
   404
  | add_single_names (a, [th]) pairs = (a, th) :: pairs
wenzelm@33309
   405
  | add_single_names (a, ths) pairs = #2 (fold (multi_name a) ths (1, pairs));
paulson@21431
   406
paulson@22382
   407
(*Ignore blacklisted basenames*)
wenzelm@33309
   408
fun add_multi_names (a, ths) pairs =
blanchet@35865
   409
  if (Long_Name.base_name a) mem_string multi_base_blacklist then pairs
wenzelm@33309
   410
  else add_single_names (a, ths) pairs;
paulson@21224
   411
paulson@21290
   412
fun is_multi (a, ths) = length ths > 1 orelse String.isSuffix ".axioms" a;
paulson@21290
   413
paulson@24286
   414
(*The single theorems go BEFORE the multiple ones. Blacklist is applied to all.*)
wenzelm@21588
   415
fun name_thm_pairs ctxt =
wenzelm@33309
   416
  let
wenzelm@33309
   417
    val (mults, singles) = List.partition is_multi (all_valid_thms ctxt)
blanchet@35865
   418
    val ps = [] |> fold add_multi_names mults
blanchet@35865
   419
                |> fold add_single_names singles
paulson@24286
   420
  in
blanchet@35865
   421
    if run_blacklist_filter then
blanchet@35865
   422
      let
blanchet@35865
   423
        val blacklist = No_ATPs.get ctxt
blanchet@35865
   424
                        |> map (`Thm.full_prop_of) |> Termtab.make
blanchet@35865
   425
        val is_blacklisted = Termtab.defined blacklist o Thm.full_prop_of o snd
blanchet@35865
   426
      in ps |> filter_out is_blacklisted end
blanchet@35865
   427
    else
blanchet@35865
   428
      ps
paulson@24286
   429
  end;
paulson@21224
   430
wenzelm@32091
   431
fun check_named ("", th) =
wenzelm@32091
   432
      (warning ("No name for theorem " ^ Display.string_of_thm_without_context th); false)
wenzelm@32994
   433
  | check_named _ = true;
paulson@19894
   434
paulson@33022
   435
fun get_all_lemmas ctxt =
paulson@19894
   436
  let val included_thms =
blanchet@35865
   437
        tap (fn ths => trace_msg
paulson@33022
   438
                     (fn () => ("Including all " ^ Int.toString (length ths) ^ " theorems")))
paulson@33022
   439
            (name_thm_pairs ctxt)
paulson@19894
   440
  in
immler@31410
   441
    filter check_named included_thms
paulson@19894
   442
  end;
mengj@19768
   443
paulson@21290
   444
(***************************************************************)
paulson@21290
   445
(* Type Classes Present in the Axiom or Conjecture Clauses     *)
paulson@21290
   446
(***************************************************************)
paulson@21290
   447
wenzelm@32952
   448
fun add_classes (sorts, cset) = List.foldl setinsert cset (flat sorts);
paulson@21290
   449
paulson@21290
   450
(*Remove this trivial type class*)
blanchet@35865
   451
fun delete_type cset = Symtab.delete_safe (the_single @{sort HOL.type}) cset;
paulson@21290
   452
paulson@21290
   453
fun tvar_classes_of_terms ts =
wenzelm@29270
   454
  let val sorts_list = map (map #2 o OldTerm.term_tvars) ts
wenzelm@30190
   455
  in  Symtab.keys (delete_type (List.foldl add_classes Symtab.empty sorts_list))  end;
paulson@21290
   456
paulson@21290
   457
fun tfree_classes_of_terms ts =
wenzelm@29270
   458
  let val sorts_list = map (map #2 o OldTerm.term_tfrees) ts
wenzelm@30190
   459
  in  Symtab.keys (delete_type (List.foldl add_classes Symtab.empty sorts_list))  end;
paulson@20526
   460
paulson@21373
   461
(*fold type constructors*)
paulson@21373
   462
fun fold_type_consts f (Type (a, Ts)) x = fold (fold_type_consts f) Ts (f (a,x))
wenzelm@32994
   463
  | fold_type_consts _ _ x = x;
paulson@21373
   464
paulson@21373
   465
val add_type_consts_in_type = fold_type_consts setinsert;
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 =
paulson@21397
   469
  let val const_typargs = Sign.const_typargs thy
paulson@21397
   470
      fun add_tcs (Const cT) x = fold add_type_consts_in_type (const_typargs cT) x
wenzelm@32994
   471
        | add_tcs (Abs (_, _, u)) x = add_tcs u x
paulson@21397
   472
        | add_tcs (t $ u) x = add_tcs t (add_tcs u x)
paulson@21397
   473
        | add_tcs _ x = x
paulson@21397
   474
  in  add_tcs  end
paulson@21373
   475
paulson@21397
   476
fun type_consts_of_terms thy ts =
paulson@21397
   477
  Symtab.keys (fold (add_type_consts_in_term thy) ts Symtab.empty);
paulson@21373
   478
paulson@21373
   479
mengj@19194
   480
(***************************************************************)
mengj@19194
   481
(* ATP invocation methods setup                                *)
mengj@19194
   482
(***************************************************************)
mengj@19194
   483
paulson@20526
   484
(*Ensures that no higher-order theorems "leak out"*)
paulson@24958
   485
fun restrict_to_logic thy true cls = filter (Meson.is_fol_term thy o prop_of o fst) cls
paulson@24958
   486
  | restrict_to_logic thy false cls = cls;
paulson@20526
   487
paulson@21470
   488
(**** Predicates to detect unwanted clauses (prolific or likely to cause unsoundness) ****)
paulson@21470
   489
paulson@21470
   490
(** Too general means, positive equality literal with a variable X as one operand,
paulson@21470
   491
  when X does not occur properly in the other operand. This rules out clearly
paulson@21470
   492
  inconsistent clauses such as V=a|V=b, though it by no means guarantees soundness. **)
wenzelm@21588
   493
paulson@21470
   494
fun occurs ix =
paulson@21470
   495
    let fun occ(Var (jx,_)) = (ix=jx)
paulson@21470
   496
          | occ(t1$t2)      = occ t1 orelse occ t2
paulson@21470
   497
          | occ(Abs(_,_,t)) = occ t
paulson@21470
   498
          | occ _           = false
paulson@21470
   499
    in occ end;
paulson@21470
   500
haftmann@31723
   501
fun is_recordtype T = not (null (Record.dest_recTs T));
paulson@21470
   502
paulson@21470
   503
(*Unwanted equalities include
paulson@21470
   504
  (1) those between a variable that does not properly occur in the second operand,
paulson@21470
   505
  (2) those between a variable and a record, since these seem to be prolific "cases" thms
wenzelm@21588
   506
*)
paulson@21470
   507
fun too_general_eqterms (Var (ix,T), t) = not (occurs ix t) orelse is_recordtype T
paulson@21470
   508
  | too_general_eqterms _ = false;
paulson@21470
   509
blanchet@35865
   510
fun too_general_equality (Const (@{const_name "op ="}, _) $ x $ y) =
paulson@21470
   511
      too_general_eqterms (x,y) orelse too_general_eqterms(y,x)
paulson@21470
   512
  | too_general_equality _ = false;
paulson@21470
   513
wenzelm@29267
   514
fun has_typed_var tycons = exists_subterm
wenzelm@29267
   515
  (fn Var (_, Type (a, _)) => member (op =) tycons a | _ => false);
paulson@21431
   516
paulson@22217
   517
(*Clauses are forbidden to contain variables of these types. The typical reason is that
paulson@22217
   518
  they lead to unsoundness. Note that "unit" satisfies numerous equations like ?X=().
paulson@22217
   519
  The resulting clause will have no type constraint, yielding false proofs. Even "bool"
paulson@22217
   520
  leads to many unsound proofs, though (obviously) only for higher-order problems.*)
blanchet@35865
   521
val unwanted_types = [@{type_name unit}, @{type_name bool}];
paulson@22217
   522
paulson@21470
   523
fun unwanted t =
blanchet@35865
   524
  t = @{prop True} orelse has_typed_var unwanted_types t orelse
paulson@24958
   525
  forall too_general_equality (HOLogic.disjuncts (strip_Trueprop t));
paulson@21470
   526
paulson@21431
   527
(*Clauses containing variables of type "unit" or "bool" are unlikely to be useful and
paulson@21431
   528
  likely to lead to unsound proofs.*)
paulson@22217
   529
fun remove_unwanted_clauses cls = filter (not o unwanted o prop_of o fst) cls;
paulson@21431
   530
blanchet@35963
   531
fun is_first_order thy higher_order goal_cls =
blanchet@35963
   532
  case higher_order of
blanchet@35963
   533
    NONE => forall (Meson.is_fol_term thy) (map prop_of goal_cls)
blanchet@35963
   534
  | SOME b => not b
immler@30536
   535
blanchet@35963
   536
fun get_relevant_facts relevance_threshold higher_order follow_defs max_new
blanchet@35963
   537
                       add_theory_const (ctxt, (chain_ths, th)) goal_cls =
immler@30536
   538
  let
immler@30536
   539
    val thy = ProofContext.theory_of ctxt
blanchet@35963
   540
    val is_FO = is_first_order thy higher_order goal_cls
wenzelm@33306
   541
    val included_cls = get_all_lemmas ctxt
blanchet@35865
   542
      |> cnf_rules_pairs thy |> make_unique
blanchet@35865
   543
      |> restrict_to_logic thy is_FO
wenzelm@33306
   544
      |> remove_unwanted_clauses
mengj@19194
   545
  in
blanchet@35963
   546
    relevance_filter relevance_threshold follow_defs max_new add_theory_const
blanchet@35963
   547
                     thy included_cls (map prop_of goal_cls)
immler@30536
   548
  end;
immler@30536
   549
immler@31752
   550
(* prepare for passing to writer,
immler@31752
   551
   create additional clauses based on the information from extra_cls *)
blanchet@35963
   552
fun prepare_clauses higher_order dfg goal_cls chain_ths axcls extra_cls thy =
immler@31409
   553
  let
Philipp@32866
   554
    (* add chain thms *)
wenzelm@33306
   555
    val chain_cls =
blanchet@35865
   556
      cnf_rules_pairs thy (filter check_named (map pairname chain_ths))
Philipp@32866
   557
    val axcls = chain_cls @ axcls
Philipp@32866
   558
    val extra_cls = chain_cls @ extra_cls
blanchet@35963
   559
    val is_FO = is_first_order thy higher_order goal_cls
immler@31752
   560
    val ccls = subtract_cls goal_cls extra_cls
blanchet@35865
   561
    val _ = app (fn th => trace_msg (fn _ => Display.string_of_thm_global thy th)) ccls
immler@30536
   562
    val ccltms = map prop_of ccls
immler@31752
   563
    and axtms = map (prop_of o #1) extra_cls
immler@30536
   564
    val subs = tfree_classes_of_terms ccltms
immler@30536
   565
    and supers = tvar_classes_of_terms axtms
blanchet@35865
   566
    and tycons = type_consts_of_terms thy (ccltms @ axtms)
immler@30536
   567
    (*TFrees in conjecture clauses; TVars in axiom clauses*)
blanchet@35865
   568
    val conjectures = make_conjecture_clauses dfg thy ccls
blanchet@35865
   569
    val (_, extra_clauses) = ListPair.unzip (make_axiom_clauses dfg thy extra_cls)
blanchet@35865
   570
    val (clnames, axiom_clauses) = ListPair.unzip (make_axiom_clauses dfg thy axcls)
blanchet@35865
   571
    val helper_clauses = get_helper_clauses dfg thy is_FO (conjectures, extra_cls, [])
blanchet@35865
   572
    val (supers', arity_clauses) = make_arity_clauses_dfg dfg thy tycons supers
blanchet@35865
   573
    val classrel_clauses = make_classrel_clauses thy subs supers'
immler@30536
   574
  in
immler@31752
   575
    (Vector.fromList clnames,
immler@31865
   576
      (conjectures, axiom_clauses, extra_clauses, helper_clauses, classrel_clauses, arity_clauses))
immler@31409
   577
  end
quigley@15644
   578
paulson@15347
   579
end;