src/HOL/Tools/Sledgehammer/sledgehammer_fact_filter.ML
author blanchet
Fri Mar 19 13:02:18 2010 +0100 (2010-03-19)
changeset 35865 2f8fb5242799
parent 35828 46cfc4b8112e
child 35963 943e2582dc87
permissions -rw-r--r--
more Sledgehammer refactoring
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@35865
    15
  val get_relevant_facts : int -> bool -> Proof.context * (thm list * 'a) -> thm list ->
immler@31409
    16
    (thm * (string * int)) list
immler@31410
    17
  val prepare_clauses : bool -> thm list -> thm list ->
blanchet@35865
    18
    (thm * (axiom_name * hol_clause_id)) list ->
blanchet@35865
    19
    (thm * (axiom_name * hol_clause_id)) list -> theory ->
blanchet@35826
    20
    axiom_name vector *
blanchet@35865
    21
      (hol_clause list * hol_clause list * hol_clause list *
blanchet@35865
    22
      hol_clause list * classrel_clause list * arity_clause list)
paulson@15347
    23
end;
paulson@15347
    24
blanchet@35826
    25
structure Sledgehammer_Fact_Filter : SLEDGEHAMMER_FACT_FILTER =
paulson@15347
    26
struct
paulson@15347
    27
blanchet@35865
    28
open Sledgehammer_FOL_Clause
blanchet@35865
    29
open Sledgehammer_Fact_Preprocessor
blanchet@35865
    30
open Sledgehammer_HOL_Clause
blanchet@35826
    31
blanchet@35865
    32
type axiom_name = axiom_name
blanchet@35865
    33
type hol_clause = hol_clause
blanchet@35865
    34
type hol_clause_id = hol_clause_id
paulson@21070
    35
mengj@19194
    36
(********************************************************************)
mengj@19194
    37
(* some settings for both background automatic ATP calling procedure*)
mengj@19194
    38
(* and also explicit ATP invocation methods                         *)
mengj@19194
    39
(********************************************************************)
mengj@19194
    40
blanchet@35865
    41
(* Translation mode can be auto-detected, or forced to be first-order or
blanchet@35865
    42
   higher-order *)
wenzelm@28477
    43
datatype mode = Auto | Fol | Hol;
wenzelm@28477
    44
blanchet@35865
    45
val translation_mode = Auto;
wenzelm@28477
    46
mengj@19194
    47
(*** background linkup ***)
wenzelm@28477
    48
val run_blacklist_filter = true;
paulson@21224
    49
paulson@24287
    50
(*** relevance filter parameters ***)
wenzelm@28477
    51
val run_relevance_filter = true;
wenzelm@28477
    52
val pass_mark = 0.5;
wenzelm@28477
    53
val convergence = 3.2;    (*Higher numbers allow longer inference chains*)
wenzelm@28477
    54
val follow_defs = false;  (*Follow definitions. Makes problems bigger.*)
wenzelm@28477
    55
  
wenzelm@28477
    56
(***************************************************************)
wenzelm@28477
    57
(* Relevance Filtering                                         *)
wenzelm@28477
    58
(***************************************************************)
mengj@19194
    59
blanchet@35865
    60
fun strip_Trueprop (@{const Trueprop} $ t) = t
paulson@24958
    61
  | strip_Trueprop t = t;
mengj@19194
    62
paulson@24287
    63
(*A surprising number of theorems contain only a few significant constants.
paulson@24287
    64
  These include all induction rules, and other general theorems. Filtering
paulson@24287
    65
  theorems in clause form reveals these complexities in the form of Skolem 
paulson@24287
    66
  functions. If we were instead to filter theorems in their natural form,
paulson@24287
    67
  some other method of measuring theorem complexity would become necessary.*)
paulson@24287
    68
paulson@24287
    69
fun log_weight2 (x:real) = 1.0 + 2.0/Math.ln (x+1.0);
paulson@24287
    70
paulson@24287
    71
(*The default seems best in practice. A constant function of one ignores
paulson@24287
    72
  the constant frequencies.*)
wenzelm@28477
    73
val weight_fn = log_weight2;
paulson@24287
    74
paulson@24287
    75
paulson@24287
    76
(*Including equality in this list might be expected to stop rules like subset_antisym from
paulson@24287
    77
  being chosen, but for some reason filtering works better with them listed. The
paulson@24287
    78
  logical signs All, Ex, &, and --> are omitted because any remaining occurrrences
paulson@24287
    79
  must be within comprehensions.*)
blanchet@35865
    80
val standard_consts =
blanchet@35865
    81
  [@{const_name Trueprop}, @{const_name "==>"}, @{const_name all},
blanchet@35865
    82
   @{const_name "=="}, @{const_name "op |"}, @{const_name Not},
blanchet@35865
    83
   @{const_name "op ="}];
paulson@24287
    84
paulson@24287
    85
paulson@24287
    86
(*** constants with types ***)
paulson@24287
    87
paulson@24287
    88
(*An abstraction of Isabelle types*)
paulson@24287
    89
datatype const_typ =  CTVar | CType of string * const_typ list
paulson@24287
    90
paulson@24287
    91
(*Is the second type an instance of the first one?*)
paulson@24287
    92
fun match_type (CType(con1,args1)) (CType(con2,args2)) = 
paulson@24287
    93
      con1=con2 andalso match_types args1 args2
paulson@24287
    94
  | match_type CTVar _ = true
paulson@24287
    95
  | match_type _ CTVar = false
paulson@24287
    96
and match_types [] [] = true
paulson@24287
    97
  | match_types (a1::as1) (a2::as2) = match_type a1 a2 andalso match_types as1 as2;
paulson@24287
    98
paulson@24287
    99
(*Is there a unifiable constant?*)
paulson@24287
   100
fun uni_mem gctab (c,c_typ) =
paulson@24287
   101
  case Symtab.lookup gctab c of
paulson@24287
   102
      NONE => false
paulson@24287
   103
    | SOME ctyps_list => exists (match_types c_typ) ctyps_list;
paulson@24287
   104
  
paulson@24287
   105
(*Maps a "real" type to a const_typ*)
paulson@24287
   106
fun const_typ_of (Type (c,typs)) = CType (c, map const_typ_of typs) 
paulson@24287
   107
  | const_typ_of (TFree _) = CTVar
paulson@24287
   108
  | const_typ_of (TVar _) = CTVar
paulson@24287
   109
paulson@24287
   110
(*Pairs a constant with the list of its type instantiations (using const_typ)*)
paulson@24287
   111
fun const_with_typ thy (c,typ) = 
paulson@24287
   112
    let val tvars = Sign.const_typargs thy (c,typ)
paulson@24287
   113
    in (c, map const_typ_of tvars) end
paulson@24287
   114
    handle TYPE _ => (c,[]);   (*Variable (locale constant): monomorphic*)   
paulson@24287
   115
paulson@24287
   116
(*Add a const/type pair to the table, but a [] entry means a standard connective,
paulson@24287
   117
  which we ignore.*)
paulson@24287
   118
fun add_const_typ_table ((c,ctyps), tab) =
paulson@24287
   119
  Symtab.map_default (c, [ctyps]) (fn [] => [] | ctyps_list => insert (op =) ctyps ctyps_list) 
paulson@24287
   120
    tab;
paulson@24287
   121
paulson@24287
   122
(*Free variables are included, as well as constants, to handle locales*)
paulson@24287
   123
fun add_term_consts_typs_rm thy (Const(c, typ), tab) =
paulson@24287
   124
      add_const_typ_table (const_with_typ thy (c,typ), tab) 
paulson@24287
   125
  | add_term_consts_typs_rm thy (Free(c, typ), tab) =
paulson@24287
   126
      add_const_typ_table (const_with_typ thy (c,typ), tab) 
paulson@24287
   127
  | add_term_consts_typs_rm thy (t $ u, tab) =
paulson@24287
   128
      add_term_consts_typs_rm thy (t, add_term_consts_typs_rm thy (u, tab))
paulson@24287
   129
  | add_term_consts_typs_rm thy (Abs(_,_,t), tab) = add_term_consts_typs_rm thy (t, tab)
wenzelm@32994
   130
  | add_term_consts_typs_rm _ (_, tab) = tab;
paulson@24287
   131
paulson@24287
   132
(*The empty list here indicates that the constant is being ignored*)
paulson@24287
   133
fun add_standard_const (s,tab) = Symtab.update (s,[]) tab;
paulson@24287
   134
paulson@24287
   135
val null_const_tab : const_typ list list Symtab.table = 
wenzelm@30190
   136
    List.foldl add_standard_const Symtab.empty standard_consts;
paulson@24287
   137
wenzelm@30190
   138
fun get_goal_consts_typs thy = List.foldl (add_term_consts_typs_rm thy) null_const_tab;
paulson@24287
   139
paulson@24287
   140
(*Inserts a dummy "constant" referring to the theory name, so that relevance
paulson@24287
   141
  takes the given theory into account.*)
wenzelm@28477
   142
fun const_prop_of theory_const th =
wenzelm@28477
   143
 if theory_const then
paulson@24287
   144
  let val name = Context.theory_name (theory_of_thm th)
paulson@24287
   145
      val t = Const (name ^ ". 1", HOLogic.boolT)
paulson@24287
   146
  in  t $ prop_of th  end
paulson@24287
   147
 else prop_of th;
paulson@24287
   148
paulson@24287
   149
(**** Constant / Type Frequencies ****)
paulson@24287
   150
paulson@24287
   151
(*A two-dimensional symbol table counts frequencies of constants. It's keyed first by
paulson@24287
   152
  constant name and second by its list of type instantiations. For the latter, we need
paulson@24287
   153
  a linear ordering on type const_typ list.*)
paulson@24287
   154
  
paulson@24287
   155
local
paulson@24287
   156
paulson@24287
   157
fun cons_nr CTVar = 0
paulson@24287
   158
  | cons_nr (CType _) = 1;
paulson@24287
   159
paulson@24287
   160
in
paulson@24287
   161
paulson@24287
   162
fun const_typ_ord TU =
paulson@24287
   163
  case TU of
paulson@24287
   164
    (CType (a, Ts), CType (b, Us)) =>
paulson@24287
   165
      (case fast_string_ord(a,b) of EQUAL => dict_ord const_typ_ord (Ts,Us) | ord => ord)
paulson@24287
   166
  | (T, U) => int_ord (cons_nr T, cons_nr U);
paulson@24287
   167
paulson@24287
   168
end;
paulson@24287
   169
wenzelm@31971
   170
structure CTtab = Table(type key = const_typ list val ord = dict_ord const_typ_ord);
paulson@24287
   171
wenzelm@28477
   172
fun count_axiom_consts theory_const thy ((thm,_), tab) = 
paulson@24287
   173
  let fun count_const (a, T, tab) =
wenzelm@32960
   174
        let val (c, cts) = const_with_typ thy (a,T)
wenzelm@32960
   175
        in  (*Two-dimensional table update. Constant maps to types maps to count.*)
wenzelm@32960
   176
            Symtab.map_default (c, CTtab.empty) 
wenzelm@32960
   177
                               (CTtab.map_default (cts,0) (fn n => n+1)) tab
wenzelm@32960
   178
        end
paulson@24287
   179
      fun count_term_consts (Const(a,T), tab) = count_const(a,T,tab)
wenzelm@32960
   180
        | count_term_consts (Free(a,T), tab) = count_const(a,T,tab)
wenzelm@32960
   181
        | count_term_consts (t $ u, tab) =
wenzelm@32960
   182
            count_term_consts (t, count_term_consts (u, tab))
wenzelm@32960
   183
        | count_term_consts (Abs(_,_,t), tab) = count_term_consts (t, tab)
wenzelm@32960
   184
        | count_term_consts (_, tab) = tab
wenzelm@28477
   185
  in  count_term_consts (const_prop_of theory_const thm, tab)  end;
paulson@24287
   186
paulson@24287
   187
paulson@24287
   188
(**** Actual Filtering Code ****)
paulson@24287
   189
paulson@24287
   190
(*The frequency of a constant is the sum of those of all instances of its type.*)
paulson@24287
   191
fun const_frequency ctab (c, cts) =
paulson@24287
   192
  let val pairs = CTtab.dest (the (Symtab.lookup ctab c))
paulson@24287
   193
      fun add ((cts',m), n) = if match_types cts cts' then m+n else n
paulson@24287
   194
  in  List.foldl add 0 pairs  end;
paulson@24287
   195
paulson@24287
   196
(*Add in a constant's weight, as determined by its frequency.*)
paulson@24287
   197
fun add_ct_weight ctab ((c,T), w) =
wenzelm@28477
   198
  w + weight_fn (real (const_frequency ctab (c,T)));
paulson@24287
   199
paulson@24287
   200
(*Relevant constants are weighted according to frequency, 
paulson@24287
   201
  but irrelevant constants are simply counted. Otherwise, Skolem functions,
paulson@24287
   202
  which are rare, would harm a clause's chances of being picked.*)
paulson@24287
   203
fun clause_weight ctab gctyps consts_typs =
paulson@24287
   204
    let val rel = filter (uni_mem gctyps) consts_typs
paulson@24287
   205
        val rel_weight = List.foldl (add_ct_weight ctab) 0.0 rel
paulson@24287
   206
    in
wenzelm@32960
   207
        rel_weight / (rel_weight + real (length consts_typs - length rel))
paulson@24287
   208
    end;
paulson@24287
   209
    
paulson@24287
   210
(*Multiplies out to a list of pairs: 'a * 'b list -> ('a * 'b) list -> ('a * 'b) list*)
wenzelm@30190
   211
fun add_expand_pairs (x,ys) xys = List.foldl (fn (y,acc) => (x,y)::acc) xys ys;
paulson@24287
   212
paulson@24287
   213
fun consts_typs_of_term thy t = 
paulson@24287
   214
  let val tab = add_term_consts_typs_rm thy (t, null_const_tab)
paulson@24287
   215
  in  Symtab.fold add_expand_pairs tab []  end;
paulson@24287
   216
wenzelm@28477
   217
fun pair_consts_typs_axiom theory_const thy (thm,name) =
wenzelm@28477
   218
    ((thm,name), (consts_typs_of_term thy (const_prop_of theory_const thm)));
paulson@24287
   219
paulson@24287
   220
exception ConstFree;
paulson@24287
   221
fun dest_ConstFree (Const aT) = aT
paulson@24287
   222
  | dest_ConstFree (Free aT) = aT
paulson@24287
   223
  | dest_ConstFree _ = raise ConstFree;
paulson@24287
   224
paulson@24287
   225
(*Look for definitions of the form f ?x1 ... ?xn = t, but not reversed.*)
wenzelm@32994
   226
fun defines thy thm gctypes =
paulson@24287
   227
    let val tm = prop_of thm
wenzelm@32960
   228
        fun defs lhs rhs =
paulson@24287
   229
            let val (rator,args) = strip_comb lhs
wenzelm@32960
   230
                val ct = const_with_typ thy (dest_ConstFree rator)
haftmann@33037
   231
            in
haftmann@33037
   232
              forall is_Var args andalso uni_mem gctypes ct andalso
haftmann@33038
   233
                subset (op =) (Term.add_vars rhs [], Term.add_vars lhs [])
paulson@24287
   234
            end
wenzelm@32960
   235
            handle ConstFree => false
paulson@24287
   236
    in    
blanchet@35865
   237
        case tm of @{const Trueprop} $ (Const(@{const_name "op ="}, _) $ lhs $ rhs) => 
wenzelm@32960
   238
                   defs lhs rhs 
wenzelm@32960
   239
                 | _ => false
paulson@24287
   240
    end;
paulson@24287
   241
paulson@24287
   242
type annotd_cls = (thm * (string * int)) * ((string * const_typ list) list);
paulson@24287
   243
       
paulson@24287
   244
(*For a reverse sort, putting the largest values first.*)
paulson@24287
   245
fun compare_pairs ((_,w1),(_,w2)) = Real.compare (w2,w1);
paulson@24287
   246
paulson@24287
   247
(*Limit the number of new clauses, to prevent runaway acceptance.*)
wenzelm@28477
   248
fun take_best max_new (newpairs : (annotd_cls*real) list) =
paulson@24287
   249
  let val nnew = length newpairs
paulson@24287
   250
  in
wenzelm@28477
   251
    if nnew <= max_new then (map #1 newpairs, [])
paulson@24287
   252
    else 
paulson@24287
   253
      let val cls = sort compare_pairs newpairs
wenzelm@28477
   254
          val accepted = List.take (cls, max_new)
paulson@24287
   255
      in
blanchet@35865
   256
        trace_msg (fn () => ("Number of candidates, " ^ Int.toString nnew ^ 
wenzelm@32960
   257
                       ", exceeds the limit of " ^ Int.toString (max_new)));
blanchet@35865
   258
        trace_msg (fn () => ("Effective pass mark: " ^ Real.toString (#2 (List.last accepted))));
blanchet@35865
   259
        trace_msg (fn () => "Actually passed: " ^
paulson@24287
   260
          space_implode ", " (map (fn (((_,(name,_)),_),_) => name) accepted));
paulson@24287
   261
wenzelm@32960
   262
        (map #1 accepted, map #1 (List.drop (cls, max_new)))
paulson@24287
   263
      end
paulson@24287
   264
  end;
paulson@24287
   265
wenzelm@28477
   266
fun relevant_clauses max_new thy ctab p rel_consts =
paulson@24287
   267
  let fun relevant ([],_) [] = [] : (thm * (string * int)) list  (*Nothing added this iteration*)
wenzelm@32960
   268
        | relevant (newpairs,rejects) [] =
wenzelm@32960
   269
            let val (newrels,more_rejects) = take_best max_new newpairs
wenzelm@32960
   270
                val new_consts = maps #2 newrels
wenzelm@32960
   271
                val rel_consts' = List.foldl add_const_typ_table rel_consts new_consts
wenzelm@32960
   272
                val newp = p + (1.0-p) / convergence
wenzelm@32960
   273
            in
blanchet@35865
   274
              trace_msg (fn () => "relevant this iteration: " ^ Int.toString (length newrels));
wenzelm@32960
   275
               (map #1 newrels) @ 
wenzelm@32960
   276
               (relevant_clauses max_new thy ctab newp rel_consts' (more_rejects@rejects))
wenzelm@32960
   277
            end
wenzelm@32960
   278
        | relevant (newrels,rejects) ((ax as (clsthm as (_,(name,n)),consts_typs)) :: axs) =
wenzelm@32960
   279
            let val weight = clause_weight ctab rel_consts consts_typs
wenzelm@32960
   280
            in
wenzelm@32994
   281
              if p <= weight orelse (follow_defs andalso defines thy (#1 clsthm) rel_consts)
blanchet@35865
   282
              then (trace_msg (fn () => (name ^ " clause " ^ Int.toString n ^ 
wenzelm@32960
   283
                                            " passes: " ^ Real.toString weight));
wenzelm@32960
   284
                    relevant ((ax,weight)::newrels, rejects) axs)
wenzelm@32960
   285
              else relevant (newrels, ax::rejects) axs
wenzelm@32960
   286
            end
blanchet@35865
   287
    in  trace_msg (fn () => ("relevant_clauses, current pass mark = " ^ Real.toString p));
paulson@24287
   288
        relevant ([],[]) 
paulson@24287
   289
    end;
wenzelm@32960
   290
        
wenzelm@28477
   291
fun relevance_filter max_new theory_const thy axioms goals = 
wenzelm@28477
   292
 if run_relevance_filter andalso pass_mark >= 0.1
paulson@24287
   293
 then
paulson@31910
   294
  let val const_tab = List.foldl (count_axiom_consts theory_const thy) Symtab.empty axioms
paulson@24287
   295
      val goal_const_tab = get_goal_consts_typs thy goals
blanchet@35865
   296
      val _ = trace_msg (fn () => ("Initial constants: " ^
paulson@24287
   297
                                 space_implode ", " (Symtab.keys goal_const_tab)));
wenzelm@28477
   298
      val rels = relevant_clauses max_new thy const_tab (pass_mark) 
wenzelm@28477
   299
                   goal_const_tab  (map (pair_consts_typs_axiom theory_const thy) axioms)
paulson@24287
   300
  in
blanchet@35865
   301
      trace_msg (fn () => ("Total relevant: " ^ Int.toString (length rels)));
paulson@24287
   302
      rels
paulson@24287
   303
  end
paulson@24287
   304
 else axioms;
paulson@24287
   305
paulson@24287
   306
(***************************************************************)
mengj@19768
   307
(* Retrieving and filtering lemmas                             *)
mengj@19768
   308
(***************************************************************)
mengj@19768
   309
paulson@33022
   310
(*** retrieve lemmas and filter them ***)
mengj@19768
   311
mengj@19768
   312
(*Hashing to detect duplicate and variant clauses, e.g. from the [iff] attribute*)
mengj@19768
   313
paulson@22382
   314
fun setinsert (x,s) = Symtab.update (x,()) s;
mengj@19768
   315
paulson@20757
   316
(*Reject theorems with names like "List.filter.filter_list_def" or
paulson@21690
   317
  "Accessible_Part.acc.defs", as these are definitions arising from packages.*)
paulson@20757
   318
fun is_package_def a =
wenzelm@30364
   319
  let val names = Long_Name.explode a
paulson@21690
   320
  in
paulson@21690
   321
     length names > 2 andalso
paulson@21690
   322
     not (hd names = "local") andalso
paulson@21690
   323
     String.isSuffix "_def" a  orelse  String.isSuffix "_defs" a
paulson@21690
   324
  end;
paulson@20757
   325
mengj@19768
   326
(** a hash function from Term.term to int, and also a hash table **)
mengj@19768
   327
val xor_words = List.foldl Word.xorb 0w0;
mengj@19768
   328
mengj@19768
   329
fun hashw_term ((Const(c,_)), w) = Polyhash.hashw_string (c,w)
paulson@20661
   330
  | hashw_term ((Free(a,_)), w) = Polyhash.hashw_string (a,w)
mengj@19768
   331
  | hashw_term ((Var(_,_)), w) = w
paulson@20661
   332
  | hashw_term ((Bound i), w) = Polyhash.hashw_int(i,w)
mengj@19768
   333
  | hashw_term ((Abs(_,_,t)), w) = hashw_term (t, w)
mengj@19768
   334
  | hashw_term ((P$Q), w) = hashw_term (Q, (hashw_term (P, w)));
mengj@19768
   335
blanchet@35865
   336
fun hash_literal (@{const Not} $ P) = Word.notb(hashw_term(P,0w0))
paulson@21070
   337
  | hash_literal P = hashw_term(P,0w0);
mengj@19768
   338
paulson@24958
   339
fun hash_term t = Word.toIntX (xor_words (map hash_literal (HOLogic.disjuncts (strip_Trueprop t))));
mengj@19768
   340
mengj@19768
   341
fun equal_thm (thm1,thm2) = Term.aconv(prop_of thm1, prop_of thm2);
paulson@20457
   342
paulson@22382
   343
exception HASH_CLAUSE;
paulson@22382
   344
mengj@19768
   345
(*Create a hash table for clauses, of the given size*)
mengj@19768
   346
fun mk_clause_table n =
paulson@20457
   347
      Polyhash.mkTable (hash_term o prop_of, equal_thm)
mengj@19768
   348
                       (n, HASH_CLAUSE);
mengj@19768
   349
paulson@20457
   350
(*Use a hash table to eliminate duplicates from xs. Argument is a list of
paulson@20868
   351
  (thm * (string * int)) tuples. The theorems are hashed into the table. *)
wenzelm@21588
   352
fun make_unique xs =
paulson@20868
   353
  let val ht = mk_clause_table 7000
paulson@20457
   354
  in
blanchet@35865
   355
      trace_msg (fn () => ("make_unique gets " ^ Int.toString (length xs) ^ " clauses"));
wenzelm@21588
   356
      app (ignore o Polyhash.peekInsert ht) xs;
paulson@20868
   357
      Polyhash.listItems ht
paulson@20457
   358
  end;
mengj@19768
   359
paulson@20868
   360
(*Remove existing axiom clauses from the conjecture clauses, as this can dramatically
paulson@20868
   361
  boost an ATP's performance (for some reason)*)
wenzelm@21588
   362
fun subtract_cls c_clauses ax_clauses =
paulson@20868
   363
  let val ht = mk_clause_table 2200
wenzelm@33035
   364
      fun known x = is_some (Polyhash.peek ht x)
paulson@20868
   365
  in
wenzelm@21588
   366
      app (ignore o Polyhash.peekInsert ht) ax_clauses;
wenzelm@21588
   367
      filter (not o known) c_clauses
paulson@20868
   368
  end;
mengj@19768
   369
paulson@21224
   370
fun all_valid_thms ctxt =
paulson@22382
   371
  let
wenzelm@26675
   372
    val global_facts = PureThy.facts_of (ProofContext.theory_of ctxt);
wenzelm@26278
   373
    val local_facts = ProofContext.facts_of ctxt;
wenzelm@33641
   374
    val full_space =
wenzelm@33641
   375
      Name_Space.merge (Facts.space_of global_facts, Facts.space_of local_facts);
wenzelm@33641
   376
wenzelm@33641
   377
    fun valid_facts facts =
wenzelm@33641
   378
      (facts, []) |-> Facts.fold_static (fn (name, ths0) =>
wenzelm@33641
   379
        let
wenzelm@33641
   380
          fun check_thms a =
wenzelm@33641
   381
            (case try (ProofContext.get_thms ctxt) a of
wenzelm@33641
   382
              NONE => false
wenzelm@33641
   383
            | SOME ths1 => Thm.eq_thms (ths0, ths1));
wenzelm@33641
   384
wenzelm@33641
   385
          val name1 = Facts.extern facts name;
wenzelm@33641
   386
          val name2 = Name_Space.extern full_space name;
blanchet@35865
   387
          val ths = filter_out bad_for_atp ths0;
wenzelm@33641
   388
        in
wenzelm@33641
   389
          if Facts.is_concealed facts name orelse null ths orelse
wenzelm@33641
   390
            run_blacklist_filter andalso is_package_def name then I
wenzelm@33641
   391
          else
wenzelm@33641
   392
            (case find_first check_thms [name1, name2, name] of
wenzelm@33641
   393
              NONE => I
wenzelm@33641
   394
            | SOME a => cons (a, ths))
wenzelm@33641
   395
        end);
wenzelm@26675
   396
  in valid_facts global_facts @ valid_facts local_facts end;
paulson@21224
   397
wenzelm@33309
   398
fun multi_name a th (n, pairs) =
wenzelm@33309
   399
  (n + 1, (a ^ "(" ^ Int.toString n ^ ")", th) :: pairs);
paulson@21224
   400
wenzelm@33309
   401
fun add_single_names (a, []) pairs = pairs
wenzelm@33309
   402
  | add_single_names (a, [th]) pairs = (a, th) :: pairs
wenzelm@33309
   403
  | add_single_names (a, ths) pairs = #2 (fold (multi_name a) ths (1, pairs));
paulson@21431
   404
paulson@22382
   405
(*Ignore blacklisted basenames*)
wenzelm@33309
   406
fun add_multi_names (a, ths) pairs =
blanchet@35865
   407
  if (Long_Name.base_name a) mem_string multi_base_blacklist then pairs
wenzelm@33309
   408
  else add_single_names (a, ths) pairs;
paulson@21224
   409
paulson@21290
   410
fun is_multi (a, ths) = length ths > 1 orelse String.isSuffix ".axioms" a;
paulson@21290
   411
paulson@24286
   412
(*The single theorems go BEFORE the multiple ones. Blacklist is applied to all.*)
wenzelm@21588
   413
fun name_thm_pairs ctxt =
wenzelm@33309
   414
  let
wenzelm@33309
   415
    val (mults, singles) = List.partition is_multi (all_valid_thms ctxt)
blanchet@35865
   416
    val ps = [] |> fold add_multi_names mults
blanchet@35865
   417
                |> fold add_single_names singles
paulson@24286
   418
  in
blanchet@35865
   419
    if run_blacklist_filter then
blanchet@35865
   420
      let
blanchet@35865
   421
        val blacklist = No_ATPs.get ctxt
blanchet@35865
   422
                        |> map (`Thm.full_prop_of) |> Termtab.make
blanchet@35865
   423
        val is_blacklisted = Termtab.defined blacklist o Thm.full_prop_of o snd
blanchet@35865
   424
      in ps |> filter_out is_blacklisted end
blanchet@35865
   425
    else
blanchet@35865
   426
      ps
paulson@24286
   427
  end;
paulson@21224
   428
wenzelm@32091
   429
fun check_named ("", th) =
wenzelm@32091
   430
      (warning ("No name for theorem " ^ Display.string_of_thm_without_context th); false)
wenzelm@32994
   431
  | check_named _ = true;
paulson@19894
   432
paulson@33022
   433
fun get_all_lemmas ctxt =
paulson@19894
   434
  let val included_thms =
blanchet@35865
   435
        tap (fn ths => trace_msg
paulson@33022
   436
                     (fn () => ("Including all " ^ Int.toString (length ths) ^ " theorems")))
paulson@33022
   437
            (name_thm_pairs ctxt)
paulson@19894
   438
  in
immler@31410
   439
    filter check_named included_thms
paulson@19894
   440
  end;
mengj@19768
   441
paulson@21290
   442
(***************************************************************)
paulson@21290
   443
(* Type Classes Present in the Axiom or Conjecture Clauses     *)
paulson@21290
   444
(***************************************************************)
paulson@21290
   445
wenzelm@32952
   446
fun add_classes (sorts, cset) = List.foldl setinsert cset (flat sorts);
paulson@21290
   447
paulson@21290
   448
(*Remove this trivial type class*)
blanchet@35865
   449
fun delete_type cset = Symtab.delete_safe (the_single @{sort HOL.type}) cset;
paulson@21290
   450
paulson@21290
   451
fun tvar_classes_of_terms ts =
wenzelm@29270
   452
  let val sorts_list = map (map #2 o OldTerm.term_tvars) ts
wenzelm@30190
   453
  in  Symtab.keys (delete_type (List.foldl add_classes Symtab.empty sorts_list))  end;
paulson@21290
   454
paulson@21290
   455
fun tfree_classes_of_terms ts =
wenzelm@29270
   456
  let val sorts_list = map (map #2 o OldTerm.term_tfrees) ts
wenzelm@30190
   457
  in  Symtab.keys (delete_type (List.foldl add_classes Symtab.empty sorts_list))  end;
paulson@20526
   458
paulson@21373
   459
(*fold type constructors*)
paulson@21373
   460
fun fold_type_consts f (Type (a, Ts)) x = fold (fold_type_consts f) Ts (f (a,x))
wenzelm@32994
   461
  | fold_type_consts _ _ x = x;
paulson@21373
   462
paulson@21373
   463
val add_type_consts_in_type = fold_type_consts setinsert;
paulson@21373
   464
paulson@21397
   465
(*Type constructors used to instantiate overloaded constants are the only ones needed.*)
paulson@21397
   466
fun add_type_consts_in_term thy =
paulson@21397
   467
  let val const_typargs = Sign.const_typargs thy
paulson@21397
   468
      fun add_tcs (Const cT) x = fold add_type_consts_in_type (const_typargs cT) x
wenzelm@32994
   469
        | add_tcs (Abs (_, _, u)) x = add_tcs u x
paulson@21397
   470
        | add_tcs (t $ u) x = add_tcs t (add_tcs u x)
paulson@21397
   471
        | add_tcs _ x = x
paulson@21397
   472
  in  add_tcs  end
paulson@21373
   473
paulson@21397
   474
fun type_consts_of_terms thy ts =
paulson@21397
   475
  Symtab.keys (fold (add_type_consts_in_term thy) ts Symtab.empty);
paulson@21373
   476
paulson@21373
   477
mengj@19194
   478
(***************************************************************)
mengj@19194
   479
(* ATP invocation methods setup                                *)
mengj@19194
   480
(***************************************************************)
mengj@19194
   481
paulson@20526
   482
(*Ensures that no higher-order theorems "leak out"*)
paulson@24958
   483
fun restrict_to_logic thy true cls = filter (Meson.is_fol_term thy o prop_of o fst) cls
paulson@24958
   484
  | restrict_to_logic thy false cls = cls;
paulson@20526
   485
paulson@21470
   486
(**** Predicates to detect unwanted clauses (prolific or likely to cause unsoundness) ****)
paulson@21470
   487
paulson@21470
   488
(** Too general means, positive equality literal with a variable X as one operand,
paulson@21470
   489
  when X does not occur properly in the other operand. This rules out clearly
paulson@21470
   490
  inconsistent clauses such as V=a|V=b, though it by no means guarantees soundness. **)
wenzelm@21588
   491
paulson@21470
   492
fun occurs ix =
paulson@21470
   493
    let fun occ(Var (jx,_)) = (ix=jx)
paulson@21470
   494
          | occ(t1$t2)      = occ t1 orelse occ t2
paulson@21470
   495
          | occ(Abs(_,_,t)) = occ t
paulson@21470
   496
          | occ _           = false
paulson@21470
   497
    in occ end;
paulson@21470
   498
haftmann@31723
   499
fun is_recordtype T = not (null (Record.dest_recTs T));
paulson@21470
   500
paulson@21470
   501
(*Unwanted equalities include
paulson@21470
   502
  (1) those between a variable that does not properly occur in the second operand,
paulson@21470
   503
  (2) those between a variable and a record, since these seem to be prolific "cases" thms
wenzelm@21588
   504
*)
paulson@21470
   505
fun too_general_eqterms (Var (ix,T), t) = not (occurs ix t) orelse is_recordtype T
paulson@21470
   506
  | too_general_eqterms _ = false;
paulson@21470
   507
blanchet@35865
   508
fun too_general_equality (Const (@{const_name "op ="}, _) $ x $ y) =
paulson@21470
   509
      too_general_eqterms (x,y) orelse too_general_eqterms(y,x)
paulson@21470
   510
  | too_general_equality _ = false;
paulson@21470
   511
wenzelm@29267
   512
fun has_typed_var tycons = exists_subterm
wenzelm@29267
   513
  (fn Var (_, Type (a, _)) => member (op =) tycons a | _ => false);
paulson@21431
   514
paulson@22217
   515
(*Clauses are forbidden to contain variables of these types. The typical reason is that
paulson@22217
   516
  they lead to unsoundness. Note that "unit" satisfies numerous equations like ?X=().
paulson@22217
   517
  The resulting clause will have no type constraint, yielding false proofs. Even "bool"
paulson@22217
   518
  leads to many unsound proofs, though (obviously) only for higher-order problems.*)
blanchet@35865
   519
val unwanted_types = [@{type_name unit}, @{type_name bool}];
paulson@22217
   520
paulson@21470
   521
fun unwanted t =
blanchet@35865
   522
  t = @{prop True} orelse has_typed_var unwanted_types t orelse
paulson@24958
   523
  forall too_general_equality (HOLogic.disjuncts (strip_Trueprop t));
paulson@21470
   524
paulson@21431
   525
(*Clauses containing variables of type "unit" or "bool" are unlikely to be useful and
paulson@21431
   526
  likely to lead to unsound proofs.*)
paulson@22217
   527
fun remove_unwanted_clauses cls = filter (not o unwanted o prop_of o fst) cls;
paulson@21431
   528
blanchet@35865
   529
fun is_first_order thy goal_cls =
blanchet@35865
   530
  case translation_mode of
blanchet@35865
   531
    Auto => forall (Meson.is_fol_term thy) (map prop_of goal_cls)
blanchet@35865
   532
  | Fol => true
blanchet@35865
   533
  | Hol => false
immler@30536
   534
blanchet@35865
   535
fun get_relevant_facts max_new theory_const (ctxt, (chain_ths, th)) goal_cls =
immler@30536
   536
  let
immler@30536
   537
    val thy = ProofContext.theory_of ctxt
blanchet@35865
   538
    val is_FO = is_first_order thy goal_cls
wenzelm@33306
   539
    val included_cls = get_all_lemmas ctxt
blanchet@35865
   540
      |> cnf_rules_pairs thy |> make_unique
blanchet@35865
   541
      |> restrict_to_logic thy is_FO
wenzelm@33306
   542
      |> remove_unwanted_clauses
mengj@19194
   543
  in
paulson@33046
   544
    relevance_filter max_new theory_const thy included_cls (map prop_of goal_cls) 
immler@30536
   545
  end;
immler@30536
   546
immler@31752
   547
(* prepare for passing to writer,
immler@31752
   548
   create additional clauses based on the information from extra_cls *)
immler@31752
   549
fun prepare_clauses dfg goal_cls chain_ths axcls extra_cls thy =
immler@31409
   550
  let
Philipp@32866
   551
    (* add chain thms *)
wenzelm@33306
   552
    val chain_cls =
blanchet@35865
   553
      cnf_rules_pairs thy (filter check_named (map pairname chain_ths))
Philipp@32866
   554
    val axcls = chain_cls @ axcls
Philipp@32866
   555
    val extra_cls = chain_cls @ extra_cls
blanchet@35865
   556
    val is_FO = is_first_order thy goal_cls
immler@31752
   557
    val ccls = subtract_cls goal_cls extra_cls
blanchet@35865
   558
    val _ = app (fn th => trace_msg (fn _ => Display.string_of_thm_global thy th)) ccls
immler@30536
   559
    val ccltms = map prop_of ccls
immler@31752
   560
    and axtms = map (prop_of o #1) extra_cls
immler@30536
   561
    val subs = tfree_classes_of_terms ccltms
immler@30536
   562
    and supers = tvar_classes_of_terms axtms
blanchet@35865
   563
    and tycons = type_consts_of_terms thy (ccltms @ axtms)
immler@30536
   564
    (*TFrees in conjecture clauses; TVars in axiom clauses*)
blanchet@35865
   565
    val conjectures = make_conjecture_clauses dfg thy ccls
blanchet@35865
   566
    val (_, extra_clauses) = ListPair.unzip (make_axiom_clauses dfg thy extra_cls)
blanchet@35865
   567
    val (clnames, axiom_clauses) = ListPair.unzip (make_axiom_clauses dfg thy axcls)
blanchet@35865
   568
    val helper_clauses = get_helper_clauses dfg thy is_FO (conjectures, extra_cls, [])
blanchet@35865
   569
    val (supers', arity_clauses) = make_arity_clauses_dfg dfg thy tycons supers
blanchet@35865
   570
    val classrel_clauses = make_classrel_clauses thy subs supers'
immler@30536
   571
  in
immler@31752
   572
    (Vector.fromList clnames,
immler@31865
   573
      (conjectures, axiom_clauses, extra_clauses, helper_clauses, classrel_clauses, arity_clauses))
immler@31409
   574
  end
quigley@15644
   575
paulson@15347
   576
end;