src/HOL/Tools/res_atp.ML
author wenzelm
Fri Oct 03 16:37:09 2008 +0200 (2008-10-03)
changeset 28477 9339d4dcec8b
parent 28065 3899dff63cd7
child 29267 8615b4f54047
permissions -rw-r--r--
version of sledgehammer using threads instead of processes, misc cleanup;
(by Fabian Immler);
mengj@19194
     1
(*  Author: Jia Meng, Cambridge University Computer Laboratory, NICTA
paulson@15608
     2
    ID: $Id$
paulson@15608
     3
    Copyright 2004 University of Cambridge
paulson@15347
     4
*)
paulson@15452
     5
wenzelm@16802
     6
signature RES_ATP =
wenzelm@16802
     7
sig
mengj@19194
     8
  datatype mode = Auto | Fol | Hol
paulson@22989
     9
  val tvar_classes_of_terms : term list -> string list
paulson@22989
    10
  val tfree_classes_of_terms : term list -> string list
paulson@22989
    11
  val type_consts_of_terms : theory -> term list -> string list
wenzelm@28477
    12
  val write_problem_files : (theory -> bool -> Thm.thm list -> string ->
wenzelm@28477
    13
  (thm * (ResHolClause.axiom_name * ResHolClause.clause_id)) list * ResClause.classrelClause list *
wenzelm@28477
    14
    ResClause.arityClause list -> string list -> ResHolClause.axiom_name list)
wenzelm@28477
    15
    -> int -> bool
wenzelm@28477
    16
    -> (int -> Path.T) -> Proof.context * thm list * thm
wenzelm@28477
    17
    -> string list * ResHolClause.axiom_name Vector.vector list
paulson@15347
    18
end;
paulson@15347
    19
wenzelm@22865
    20
structure ResAtp: RES_ATP =
paulson@15347
    21
struct
paulson@15347
    22
paulson@21070
    23
mengj@19194
    24
(********************************************************************)
mengj@19194
    25
(* some settings for both background automatic ATP calling procedure*)
mengj@19194
    26
(* and also explicit ATP invocation methods                         *)
mengj@19194
    27
(********************************************************************)
mengj@19194
    28
wenzelm@28477
    29
(*Translation mode can be auto-detected, or forced to be first-order or higher-order*)
wenzelm@28477
    30
datatype mode = Auto | Fol | Hol;
wenzelm@28477
    31
wenzelm@28477
    32
val linkup_logic_mode = Auto;
wenzelm@28477
    33
mengj@19194
    34
(*** background linkup ***)
wenzelm@28477
    35
val run_blacklist_filter = true;
paulson@21224
    36
paulson@24287
    37
(*** relevance filter parameters ***)
wenzelm@28477
    38
val run_relevance_filter = true;
wenzelm@28477
    39
val pass_mark = 0.5;
wenzelm@28477
    40
val convergence = 3.2;    (*Higher numbers allow longer inference chains*)
wenzelm@28477
    41
val follow_defs = false;  (*Follow definitions. Makes problems bigger.*)
wenzelm@28477
    42
val include_all = true;
wenzelm@28477
    43
val include_simpset = false;
wenzelm@28477
    44
val include_claset = false;
wenzelm@28477
    45
val include_atpset = true;
wenzelm@28477
    46
  
wenzelm@28477
    47
(***************************************************************)
wenzelm@28477
    48
(* Relevance Filtering                                         *)
wenzelm@28477
    49
(***************************************************************)
mengj@19194
    50
paulson@24958
    51
fun strip_Trueprop (Const("Trueprop",_) $ t) = t
paulson@24958
    52
  | strip_Trueprop t = t;
mengj@19194
    53
paulson@24287
    54
(*A surprising number of theorems contain only a few significant constants.
paulson@24287
    55
  These include all induction rules, and other general theorems. Filtering
paulson@24287
    56
  theorems in clause form reveals these complexities in the form of Skolem 
paulson@24287
    57
  functions. If we were instead to filter theorems in their natural form,
paulson@24287
    58
  some other method of measuring theorem complexity would become necessary.*)
paulson@24287
    59
paulson@24287
    60
fun log_weight2 (x:real) = 1.0 + 2.0/Math.ln (x+1.0);
paulson@24287
    61
paulson@24287
    62
(*The default seems best in practice. A constant function of one ignores
paulson@24287
    63
  the constant frequencies.*)
wenzelm@28477
    64
val weight_fn = log_weight2;
paulson@24287
    65
paulson@24287
    66
paulson@24287
    67
(*Including equality in this list might be expected to stop rules like subset_antisym from
paulson@24287
    68
  being chosen, but for some reason filtering works better with them listed. The
paulson@24287
    69
  logical signs All, Ex, &, and --> are omitted because any remaining occurrrences
paulson@24287
    70
  must be within comprehensions.*)
paulson@24287
    71
val standard_consts = ["Trueprop","==>","all","==","op |","Not","op ="];
paulson@24287
    72
paulson@24287
    73
paulson@24287
    74
(*** constants with types ***)
paulson@24287
    75
paulson@24287
    76
(*An abstraction of Isabelle types*)
paulson@24287
    77
datatype const_typ =  CTVar | CType of string * const_typ list
paulson@24287
    78
paulson@24287
    79
(*Is the second type an instance of the first one?*)
paulson@24287
    80
fun match_type (CType(con1,args1)) (CType(con2,args2)) = 
paulson@24287
    81
      con1=con2 andalso match_types args1 args2
paulson@24287
    82
  | match_type CTVar _ = true
paulson@24287
    83
  | match_type _ CTVar = false
paulson@24287
    84
and match_types [] [] = true
paulson@24287
    85
  | match_types (a1::as1) (a2::as2) = match_type a1 a2 andalso match_types as1 as2;
paulson@24287
    86
paulson@24287
    87
(*Is there a unifiable constant?*)
paulson@24287
    88
fun uni_mem gctab (c,c_typ) =
paulson@24287
    89
  case Symtab.lookup gctab c of
paulson@24287
    90
      NONE => false
paulson@24287
    91
    | SOME ctyps_list => exists (match_types c_typ) ctyps_list;
paulson@24287
    92
  
paulson@24287
    93
(*Maps a "real" type to a const_typ*)
paulson@24287
    94
fun const_typ_of (Type (c,typs)) = CType (c, map const_typ_of typs) 
paulson@24287
    95
  | const_typ_of (TFree _) = CTVar
paulson@24287
    96
  | const_typ_of (TVar _) = CTVar
paulson@24287
    97
paulson@24287
    98
(*Pairs a constant with the list of its type instantiations (using const_typ)*)
paulson@24287
    99
fun const_with_typ thy (c,typ) = 
paulson@24287
   100
    let val tvars = Sign.const_typargs thy (c,typ)
paulson@24287
   101
    in (c, map const_typ_of tvars) end
paulson@24287
   102
    handle TYPE _ => (c,[]);   (*Variable (locale constant): monomorphic*)   
paulson@24287
   103
paulson@24287
   104
(*Add a const/type pair to the table, but a [] entry means a standard connective,
paulson@24287
   105
  which we ignore.*)
paulson@24287
   106
fun add_const_typ_table ((c,ctyps), tab) =
paulson@24287
   107
  Symtab.map_default (c, [ctyps]) (fn [] => [] | ctyps_list => insert (op =) ctyps ctyps_list) 
paulson@24287
   108
    tab;
paulson@24287
   109
paulson@24287
   110
(*Free variables are included, as well as constants, to handle locales*)
paulson@24287
   111
fun add_term_consts_typs_rm thy (Const(c, typ), tab) =
paulson@24287
   112
      add_const_typ_table (const_with_typ thy (c,typ), tab) 
paulson@24287
   113
  | add_term_consts_typs_rm thy (Free(c, typ), tab) =
paulson@24287
   114
      add_const_typ_table (const_with_typ thy (c,typ), tab) 
paulson@24287
   115
  | add_term_consts_typs_rm thy (t $ u, tab) =
paulson@24287
   116
      add_term_consts_typs_rm thy (t, add_term_consts_typs_rm thy (u, tab))
paulson@24287
   117
  | add_term_consts_typs_rm thy (Abs(_,_,t), tab) = add_term_consts_typs_rm thy (t, tab)
paulson@24287
   118
  | add_term_consts_typs_rm thy (_, tab) = tab;
paulson@24287
   119
paulson@24287
   120
(*The empty list here indicates that the constant is being ignored*)
paulson@24287
   121
fun add_standard_const (s,tab) = Symtab.update (s,[]) tab;
paulson@24287
   122
paulson@24287
   123
val null_const_tab : const_typ list list Symtab.table = 
paulson@24287
   124
    foldl add_standard_const Symtab.empty standard_consts;
paulson@24287
   125
paulson@24287
   126
fun get_goal_consts_typs thy = foldl (add_term_consts_typs_rm thy) null_const_tab;
paulson@24287
   127
paulson@24287
   128
(*Inserts a dummy "constant" referring to the theory name, so that relevance
paulson@24287
   129
  takes the given theory into account.*)
wenzelm@28477
   130
fun const_prop_of theory_const th =
wenzelm@28477
   131
 if theory_const then
paulson@24287
   132
  let val name = Context.theory_name (theory_of_thm th)
paulson@24287
   133
      val t = Const (name ^ ". 1", HOLogic.boolT)
paulson@24287
   134
  in  t $ prop_of th  end
paulson@24287
   135
 else prop_of th;
paulson@24287
   136
paulson@24287
   137
(**** Constant / Type Frequencies ****)
paulson@24287
   138
paulson@24287
   139
(*A two-dimensional symbol table counts frequencies of constants. It's keyed first by
paulson@24287
   140
  constant name and second by its list of type instantiations. For the latter, we need
paulson@24287
   141
  a linear ordering on type const_typ list.*)
paulson@24287
   142
  
paulson@24287
   143
local
paulson@24287
   144
paulson@24287
   145
fun cons_nr CTVar = 0
paulson@24287
   146
  | cons_nr (CType _) = 1;
paulson@24287
   147
paulson@24287
   148
in
paulson@24287
   149
paulson@24287
   150
fun const_typ_ord TU =
paulson@24287
   151
  case TU of
paulson@24287
   152
    (CType (a, Ts), CType (b, Us)) =>
paulson@24287
   153
      (case fast_string_ord(a,b) of EQUAL => dict_ord const_typ_ord (Ts,Us) | ord => ord)
paulson@24287
   154
  | (T, U) => int_ord (cons_nr T, cons_nr U);
paulson@24287
   155
paulson@24287
   156
end;
paulson@24287
   157
paulson@24287
   158
structure CTtab = TableFun(type key = const_typ list val ord = dict_ord const_typ_ord);
paulson@24287
   159
wenzelm@28477
   160
fun count_axiom_consts theory_const thy ((thm,_), tab) = 
paulson@24287
   161
  let fun count_const (a, T, tab) =
paulson@24287
   162
	let val (c, cts) = const_with_typ thy (a,T)
paulson@24287
   163
	in  (*Two-dimensional table update. Constant maps to types maps to count.*)
paulson@24287
   164
	    Symtab.map_default (c, CTtab.empty) 
paulson@24287
   165
	                       (CTtab.map_default (cts,0) (fn n => n+1)) tab
paulson@24287
   166
	end
paulson@24287
   167
      fun count_term_consts (Const(a,T), tab) = count_const(a,T,tab)
paulson@24287
   168
	| count_term_consts (Free(a,T), tab) = count_const(a,T,tab)
paulson@24287
   169
	| count_term_consts (t $ u, tab) =
paulson@24287
   170
	    count_term_consts (t, count_term_consts (u, tab))
paulson@24287
   171
	| count_term_consts (Abs(_,_,t), tab) = count_term_consts (t, tab)
paulson@24287
   172
	| count_term_consts (_, tab) = tab
wenzelm@28477
   173
  in  count_term_consts (const_prop_of theory_const thm, tab)  end;
paulson@24287
   174
paulson@24287
   175
paulson@24287
   176
(**** Actual Filtering Code ****)
paulson@24287
   177
paulson@24287
   178
(*The frequency of a constant is the sum of those of all instances of its type.*)
paulson@24287
   179
fun const_frequency ctab (c, cts) =
paulson@24287
   180
  let val pairs = CTtab.dest (the (Symtab.lookup ctab c))
paulson@24287
   181
      fun add ((cts',m), n) = if match_types cts cts' then m+n else n
paulson@24287
   182
  in  List.foldl add 0 pairs  end;
paulson@24287
   183
paulson@24287
   184
(*Add in a constant's weight, as determined by its frequency.*)
paulson@24287
   185
fun add_ct_weight ctab ((c,T), w) =
wenzelm@28477
   186
  w + weight_fn (real (const_frequency ctab (c,T)));
paulson@24287
   187
paulson@24287
   188
(*Relevant constants are weighted according to frequency, 
paulson@24287
   189
  but irrelevant constants are simply counted. Otherwise, Skolem functions,
paulson@24287
   190
  which are rare, would harm a clause's chances of being picked.*)
paulson@24287
   191
fun clause_weight ctab gctyps consts_typs =
paulson@24287
   192
    let val rel = filter (uni_mem gctyps) consts_typs
paulson@24287
   193
        val rel_weight = List.foldl (add_ct_weight ctab) 0.0 rel
paulson@24287
   194
    in
paulson@24287
   195
	rel_weight / (rel_weight + real (length consts_typs - length rel))
paulson@24287
   196
    end;
paulson@24287
   197
    
paulson@24287
   198
(*Multiplies out to a list of pairs: 'a * 'b list -> ('a * 'b) list -> ('a * 'b) list*)
paulson@24287
   199
fun add_expand_pairs (x,ys) xys = foldl (fn (y,acc) => (x,y)::acc) xys ys;
paulson@24287
   200
paulson@24287
   201
fun consts_typs_of_term thy t = 
paulson@24287
   202
  let val tab = add_term_consts_typs_rm thy (t, null_const_tab)
paulson@24287
   203
  in  Symtab.fold add_expand_pairs tab []  end;
paulson@24287
   204
wenzelm@28477
   205
fun pair_consts_typs_axiom theory_const thy (thm,name) =
wenzelm@28477
   206
    ((thm,name), (consts_typs_of_term thy (const_prop_of theory_const thm)));
paulson@24287
   207
paulson@24287
   208
exception ConstFree;
paulson@24287
   209
fun dest_ConstFree (Const aT) = aT
paulson@24287
   210
  | dest_ConstFree (Free aT) = aT
paulson@24287
   211
  | dest_ConstFree _ = raise ConstFree;
paulson@24287
   212
paulson@24287
   213
(*Look for definitions of the form f ?x1 ... ?xn = t, but not reversed.*)
paulson@24287
   214
fun defines thy (thm,(name,n)) gctypes =
paulson@24287
   215
    let val tm = prop_of thm
paulson@24287
   216
	fun defs lhs rhs =
paulson@24287
   217
            let val (rator,args) = strip_comb lhs
paulson@24287
   218
		val ct = const_with_typ thy (dest_ConstFree rator)
paulson@24287
   219
            in  forall is_Var args andalso uni_mem gctypes ct andalso
paulson@24287
   220
                Term.add_vars rhs [] subset Term.add_vars lhs []
paulson@24287
   221
            end
paulson@24287
   222
	    handle ConstFree => false
paulson@24287
   223
    in    
paulson@24287
   224
	case tm of Const ("Trueprop",_) $ (Const("op =",_) $ lhs $ rhs) => 
paulson@24287
   225
		   defs lhs rhs andalso
paulson@24287
   226
		   (Output.debug (fn () => "Definition found: " ^ name ^ "_" ^ Int.toString n); true)
paulson@24287
   227
		 | _ => false
paulson@24287
   228
    end;
paulson@24287
   229
paulson@24287
   230
type annotd_cls = (thm * (string * int)) * ((string * const_typ list) list);
paulson@24287
   231
       
paulson@24287
   232
(*For a reverse sort, putting the largest values first.*)
paulson@24287
   233
fun compare_pairs ((_,w1),(_,w2)) = Real.compare (w2,w1);
paulson@24287
   234
paulson@24287
   235
(*Limit the number of new clauses, to prevent runaway acceptance.*)
wenzelm@28477
   236
fun take_best max_new (newpairs : (annotd_cls*real) list) =
paulson@24287
   237
  let val nnew = length newpairs
paulson@24287
   238
  in
wenzelm@28477
   239
    if nnew <= max_new then (map #1 newpairs, [])
paulson@24287
   240
    else 
paulson@24287
   241
      let val cls = sort compare_pairs newpairs
wenzelm@28477
   242
          val accepted = List.take (cls, max_new)
paulson@24287
   243
      in
paulson@24287
   244
        Output.debug (fn () => ("Number of candidates, " ^ Int.toString nnew ^ 
wenzelm@28477
   245
		       ", exceeds the limit of " ^ Int.toString (max_new)));
paulson@24287
   246
        Output.debug (fn () => ("Effective pass mark: " ^ Real.toString (#2 (List.last accepted))));
paulson@24287
   247
        Output.debug (fn () => "Actually passed: " ^
paulson@24287
   248
          space_implode ", " (map (fn (((_,(name,_)),_),_) => name) accepted));
paulson@24287
   249
wenzelm@28477
   250
	(map #1 accepted, map #1 (List.drop (cls, max_new)))
paulson@24287
   251
      end
paulson@24287
   252
  end;
paulson@24287
   253
wenzelm@28477
   254
fun relevant_clauses max_new thy ctab p rel_consts =
paulson@24287
   255
  let fun relevant ([],_) [] = [] : (thm * (string * int)) list  (*Nothing added this iteration*)
paulson@24287
   256
	| relevant (newpairs,rejects) [] =
wenzelm@28477
   257
	    let val (newrels,more_rejects) = take_best max_new newpairs
paulson@24287
   258
		val new_consts = List.concat (map #2 newrels)
paulson@24287
   259
		val rel_consts' = foldl add_const_typ_table rel_consts new_consts
wenzelm@28477
   260
		val newp = p + (1.0-p) / convergence
paulson@24287
   261
	    in
paulson@24287
   262
              Output.debug (fn () => ("relevant this iteration: " ^ Int.toString (length newrels)));
paulson@24287
   263
	       (map #1 newrels) @ 
wenzelm@28477
   264
	       (relevant_clauses max_new thy ctab newp rel_consts' (more_rejects@rejects))
paulson@24287
   265
	    end
paulson@24287
   266
	| relevant (newrels,rejects) ((ax as (clsthm as (_,(name,n)),consts_typs)) :: axs) =
paulson@24287
   267
	    let val weight = clause_weight ctab rel_consts consts_typs
paulson@24287
   268
	    in
wenzelm@28477
   269
	      if p <= weight orelse (follow_defs andalso defines thy clsthm rel_consts)
paulson@24287
   270
	      then (Output.debug (fn () => (name ^ " clause " ^ Int.toString n ^ 
paulson@24287
   271
	                                    " passes: " ^ Real.toString weight));
paulson@24287
   272
	            relevant ((ax,weight)::newrels, rejects) axs)
paulson@24287
   273
	      else relevant (newrels, ax::rejects) axs
paulson@24287
   274
	    end
paulson@24287
   275
    in  Output.debug (fn () => ("relevant_clauses, current pass mark = " ^ Real.toString p));
paulson@24287
   276
        relevant ([],[]) 
paulson@24287
   277
    end;
paulson@24287
   278
	
wenzelm@28477
   279
fun relevance_filter max_new theory_const thy axioms goals = 
wenzelm@28477
   280
 if run_relevance_filter andalso pass_mark >= 0.1
paulson@24287
   281
 then
paulson@24287
   282
  let val _ = Output.debug (fn () => "Start of relevance filtering");
wenzelm@28477
   283
      val const_tab = List.foldl (count_axiom_consts theory_const thy) Symtab.empty axioms
paulson@24287
   284
      val goal_const_tab = get_goal_consts_typs thy goals
paulson@24287
   285
      val _ = Output.debug (fn () => ("Initial constants: " ^
paulson@24287
   286
                                 space_implode ", " (Symtab.keys goal_const_tab)));
wenzelm@28477
   287
      val rels = relevant_clauses max_new thy const_tab (pass_mark) 
wenzelm@28477
   288
                   goal_const_tab  (map (pair_consts_typs_axiom theory_const thy) axioms)
paulson@24287
   289
  in
paulson@24287
   290
      Output.debug (fn () => ("Total relevant: " ^ Int.toString (length rels)));
paulson@24287
   291
      rels
paulson@24287
   292
  end
paulson@24287
   293
 else axioms;
paulson@24287
   294
paulson@24287
   295
(***************************************************************)
mengj@19768
   296
(* Retrieving and filtering lemmas                             *)
mengj@19768
   297
(***************************************************************)
mengj@19768
   298
mengj@19768
   299
(*** white list and black list of lemmas ***)
mengj@19768
   300
paulson@24215
   301
(*The rule subsetI is frequently omitted by the relevance filter. This could be theory data
paulson@24215
   302
  or identified with ATPset (which however is too big currently)*)
paulson@24215
   303
val whitelist = [subsetI];
wenzelm@21588
   304
mengj@19768
   305
(*** retrieve lemmas from clasimpset and atpset, may filter them ***)
mengj@19768
   306
mengj@19768
   307
(*Hashing to detect duplicate and variant clauses, e.g. from the [iff] attribute*)
mengj@19768
   308
paulson@22382
   309
fun setinsert (x,s) = Symtab.update (x,()) s;
mengj@19768
   310
paulson@20757
   311
(*Reject theorems with names like "List.filter.filter_list_def" or
paulson@21690
   312
  "Accessible_Part.acc.defs", as these are definitions arising from packages.*)
paulson@20757
   313
fun is_package_def a =
wenzelm@21858
   314
  let val names = NameSpace.explode a
paulson@21690
   315
  in
paulson@21690
   316
     length names > 2 andalso
paulson@21690
   317
     not (hd names = "local") andalso
paulson@21690
   318
     String.isSuffix "_def" a  orelse  String.isSuffix "_defs" a
paulson@21690
   319
  end;
paulson@20757
   320
mengj@19768
   321
(** a hash function from Term.term to int, and also a hash table **)
mengj@19768
   322
val xor_words = List.foldl Word.xorb 0w0;
mengj@19768
   323
mengj@19768
   324
fun hashw_term ((Const(c,_)), w) = Polyhash.hashw_string (c,w)
paulson@20661
   325
  | hashw_term ((Free(a,_)), w) = Polyhash.hashw_string (a,w)
mengj@19768
   326
  | hashw_term ((Var(_,_)), w) = w
paulson@20661
   327
  | hashw_term ((Bound i), w) = Polyhash.hashw_int(i,w)
mengj@19768
   328
  | hashw_term ((Abs(_,_,t)), w) = hashw_term (t, w)
mengj@19768
   329
  | hashw_term ((P$Q), w) = hashw_term (Q, (hashw_term (P, w)));
mengj@19768
   330
paulson@21070
   331
fun hash_literal (Const("Not",_)$P) = Word.notb(hashw_term(P,0w0))
paulson@21070
   332
  | hash_literal P = hashw_term(P,0w0);
mengj@19768
   333
paulson@24958
   334
fun hash_term t = Word.toIntX (xor_words (map hash_literal (HOLogic.disjuncts (strip_Trueprop t))));
mengj@19768
   335
mengj@19768
   336
fun equal_thm (thm1,thm2) = Term.aconv(prop_of thm1, prop_of thm2);
paulson@20457
   337
paulson@22382
   338
exception HASH_CLAUSE;
paulson@22382
   339
mengj@19768
   340
(*Create a hash table for clauses, of the given size*)
mengj@19768
   341
fun mk_clause_table n =
paulson@20457
   342
      Polyhash.mkTable (hash_term o prop_of, equal_thm)
mengj@19768
   343
                       (n, HASH_CLAUSE);
mengj@19768
   344
paulson@20457
   345
(*Use a hash table to eliminate duplicates from xs. Argument is a list of
paulson@20868
   346
  (thm * (string * int)) tuples. The theorems are hashed into the table. *)
wenzelm@21588
   347
fun make_unique xs =
paulson@20868
   348
  let val ht = mk_clause_table 7000
paulson@20457
   349
  in
wenzelm@22130
   350
      Output.debug (fn () => ("make_unique gets " ^ Int.toString (length xs) ^ " clauses"));
wenzelm@21588
   351
      app (ignore o Polyhash.peekInsert ht) xs;
paulson@20868
   352
      Polyhash.listItems ht
paulson@20457
   353
  end;
mengj@19768
   354
paulson@20868
   355
(*Remove existing axiom clauses from the conjecture clauses, as this can dramatically
paulson@20868
   356
  boost an ATP's performance (for some reason)*)
wenzelm@21588
   357
fun subtract_cls c_clauses ax_clauses =
paulson@20868
   358
  let val ht = mk_clause_table 2200
paulson@20868
   359
      fun known x = isSome (Polyhash.peek ht x)
paulson@20868
   360
  in
wenzelm@21588
   361
      app (ignore o Polyhash.peekInsert ht) ax_clauses;
wenzelm@21588
   362
      filter (not o known) c_clauses
paulson@20868
   363
  end;
mengj@19768
   364
wenzelm@26675
   365
fun valid_facts facts =
wenzelm@26691
   366
  Facts.fold_static (fn (name, ths) =>
wenzelm@28477
   367
    if run_blacklist_filter andalso is_package_def name then I
wenzelm@26691
   368
    else
wenzelm@26691
   369
      let val xname = Facts.extern facts name in
wenzelm@26691
   370
        if NameSpace.is_hidden xname then I
wenzelm@26691
   371
        else cons (xname, filter_out ResAxioms.bad_for_atp ths)
wenzelm@26691
   372
      end) facts [];
wenzelm@26675
   373
paulson@21224
   374
fun all_valid_thms ctxt =
paulson@22382
   375
  let
wenzelm@26675
   376
    val global_facts = PureThy.facts_of (ProofContext.theory_of ctxt);
wenzelm@26278
   377
    val local_facts = ProofContext.facts_of ctxt;
wenzelm@26675
   378
  in valid_facts global_facts @ valid_facts local_facts end;
paulson@21224
   379
wenzelm@21588
   380
fun multi_name a (th, (n,pairs)) =
paulson@21224
   381
  (n+1, (a ^ "(" ^ Int.toString n ^ ")", th) :: pairs)
paulson@21224
   382
paulson@22382
   383
fun add_single_names ((a, []), pairs) = pairs
paulson@22382
   384
  | add_single_names ((a, [th]), pairs) = (a,th)::pairs
paulson@22382
   385
  | add_single_names ((a, ths), pairs) = #2 (foldl (multi_name a) (1,pairs) ths);
paulson@21431
   386
paulson@22382
   387
(*Ignore blacklisted basenames*)
wenzelm@21588
   388
fun add_multi_names ((a, ths), pairs) =
paulson@24854
   389
  if (Sign.base_name a) mem_string ResAxioms.multi_base_blacklist  then pairs
paulson@22382
   390
  else add_single_names ((a, ths), pairs);
paulson@21224
   391
paulson@21290
   392
fun is_multi (a, ths) = length ths > 1 orelse String.isSuffix ".axioms" a;
paulson@21290
   393
paulson@24286
   394
(*The single theorems go BEFORE the multiple ones. Blacklist is applied to all.*)
wenzelm@21588
   395
fun name_thm_pairs ctxt =
paulson@21290
   396
  let val (mults,singles) = List.partition is_multi (all_valid_thms ctxt)
paulson@24286
   397
      val ht = mk_clause_table 900   (*ht for blacklisted theorems*)
wenzelm@28477
   398
      fun blacklisted x = run_blacklist_filter andalso isSome (Polyhash.peek ht x)
paulson@24286
   399
  in
paulson@24286
   400
      app (fn th => ignore (Polyhash.peekInsert ht (th,()))) (ResBlacklist.get ctxt);
paulson@24286
   401
      filter (not o blacklisted o #2)
paulson@24286
   402
        (foldl add_single_names (foldl add_multi_names [] mults) singles)
paulson@24286
   403
  end;
paulson@21224
   404
wenzelm@26928
   405
fun check_named ("",th) = (warning ("No name for theorem " ^ Display.string_of_thm th); false)
paulson@21224
   406
  | check_named (_,th) = true;
paulson@19894
   407
wenzelm@26928
   408
fun display_thm (name,th) = Output.debug (fn () => name ^ ": " ^ Display.string_of_thm th);
paulson@22193
   409
mengj@19768
   410
(* get lemmas from claset, simpset, atpset and extra supplied rules *)
wenzelm@21588
   411
fun get_clasimp_atp_lemmas ctxt user_thms =
paulson@19894
   412
  let val included_thms =
wenzelm@28477
   413
        if include_all
wenzelm@21588
   414
        then (tap (fn ths => Output.debug
wenzelm@22130
   415
                     (fn () => ("Including all " ^ Int.toString (length ths) ^ " theorems")))
wenzelm@21588
   416
                  (name_thm_pairs ctxt))
wenzelm@21588
   417
        else
wenzelm@21588
   418
        let val claset_thms =
wenzelm@28477
   419
                if include_claset then ResAxioms.claset_rules_of ctxt
wenzelm@21588
   420
                else []
wenzelm@21588
   421
            val simpset_thms =
wenzelm@28477
   422
                if include_simpset then ResAxioms.simpset_rules_of ctxt
wenzelm@21588
   423
                else []
wenzelm@21588
   424
            val atpset_thms =
wenzelm@28477
   425
                if include_atpset then ResAxioms.atpset_rules_of ctxt
wenzelm@21588
   426
                else []
paulson@22193
   427
            val _ = (Output.debug (fn () => "ATP theorems: ");  app display_thm atpset_thms)
wenzelm@21588
   428
        in  claset_thms @ simpset_thms @ atpset_thms  end
wenzelm@21588
   429
      val user_rules = filter check_named
paulson@22382
   430
                         (map ResAxioms.pairname
paulson@24215
   431
                           (if null user_thms then whitelist else user_thms))
paulson@19894
   432
  in
paulson@21224
   433
      (filter check_named included_thms, user_rules)
paulson@19894
   434
  end;
mengj@19768
   435
paulson@21290
   436
(***************************************************************)
paulson@21290
   437
(* Type Classes Present in the Axiom or Conjecture Clauses     *)
paulson@21290
   438
(***************************************************************)
paulson@21290
   439
paulson@21290
   440
fun add_classes (sorts, cset) = foldl setinsert cset (List.concat sorts);
paulson@21290
   441
paulson@21290
   442
(*Remove this trivial type class*)
paulson@21290
   443
fun delete_type cset = Symtab.delete_safe "HOL.type" cset;
paulson@21290
   444
paulson@21290
   445
fun tvar_classes_of_terms ts =
paulson@21290
   446
  let val sorts_list = map (map #2 o term_tvars) ts
paulson@21290
   447
  in  Symtab.keys (delete_type (foldl add_classes Symtab.empty sorts_list))  end;
paulson@21290
   448
paulson@21290
   449
fun tfree_classes_of_terms ts =
paulson@21290
   450
  let val sorts_list = map (map #2 o term_tfrees) ts
paulson@21290
   451
  in  Symtab.keys (delete_type (foldl add_classes Symtab.empty sorts_list))  end;
paulson@20526
   452
paulson@21373
   453
(*fold type constructors*)
paulson@21373
   454
fun fold_type_consts f (Type (a, Ts)) x = fold (fold_type_consts f) Ts (f (a,x))
paulson@21373
   455
  | fold_type_consts f T x = x;
paulson@21373
   456
paulson@21373
   457
val add_type_consts_in_type = fold_type_consts setinsert;
paulson@21373
   458
paulson@21397
   459
(*Type constructors used to instantiate overloaded constants are the only ones needed.*)
paulson@21397
   460
fun add_type_consts_in_term thy =
paulson@21397
   461
  let val const_typargs = Sign.const_typargs thy
paulson@21397
   462
      fun add_tcs (Const cT) x = fold add_type_consts_in_type (const_typargs cT) x
paulson@21397
   463
        | add_tcs (Abs (_, T, u)) x = add_tcs u x
paulson@21397
   464
        | add_tcs (t $ u) x = add_tcs t (add_tcs u x)
paulson@21397
   465
        | add_tcs _ x = x
paulson@21397
   466
  in  add_tcs  end
paulson@21373
   467
paulson@21397
   468
fun type_consts_of_terms thy ts =
paulson@21397
   469
  Symtab.keys (fold (add_type_consts_in_term thy) ts Symtab.empty);
paulson@21373
   470
paulson@21373
   471
mengj@19194
   472
(***************************************************************)
mengj@19194
   473
(* ATP invocation methods setup                                *)
mengj@19194
   474
(***************************************************************)
mengj@19194
   475
paulson@20526
   476
(*Ensures that no higher-order theorems "leak out"*)
paulson@24958
   477
fun restrict_to_logic thy true cls = filter (Meson.is_fol_term thy o prop_of o fst) cls
paulson@24958
   478
  | restrict_to_logic thy false cls = cls;
paulson@20526
   479
paulson@21470
   480
(**** Predicates to detect unwanted clauses (prolific or likely to cause unsoundness) ****)
paulson@21470
   481
paulson@21470
   482
(** Too general means, positive equality literal with a variable X as one operand,
paulson@21470
   483
  when X does not occur properly in the other operand. This rules out clearly
paulson@21470
   484
  inconsistent clauses such as V=a|V=b, though it by no means guarantees soundness. **)
wenzelm@21588
   485
paulson@21470
   486
fun occurs ix =
paulson@21470
   487
    let fun occ(Var (jx,_)) = (ix=jx)
paulson@21470
   488
          | occ(t1$t2)      = occ t1 orelse occ t2
paulson@21470
   489
          | occ(Abs(_,_,t)) = occ t
paulson@21470
   490
          | occ _           = false
paulson@21470
   491
    in occ end;
paulson@21470
   492
paulson@21470
   493
fun is_recordtype T = not (null (RecordPackage.dest_recTs T));
paulson@21470
   494
paulson@21470
   495
(*Unwanted equalities include
paulson@21470
   496
  (1) those between a variable that does not properly occur in the second operand,
paulson@21470
   497
  (2) those between a variable and a record, since these seem to be prolific "cases" thms
wenzelm@21588
   498
*)
paulson@21470
   499
fun too_general_eqterms (Var (ix,T), t) = not (occurs ix t) orelse is_recordtype T
paulson@21470
   500
  | too_general_eqterms _ = false;
paulson@21470
   501
paulson@21470
   502
fun too_general_equality (Const ("op =", _) $ x $ y) =
paulson@21470
   503
      too_general_eqterms (x,y) orelse too_general_eqterms(y,x)
paulson@21470
   504
  | too_general_equality _ = false;
paulson@21470
   505
paulson@21470
   506
(* tautologous? *)
paulson@21470
   507
fun is_taut (Const ("Trueprop", _) $ Const ("True", _)) = true
paulson@21470
   508
  | is_taut _ = false;
paulson@21470
   509
paulson@21431
   510
(*True if the term contains a variable whose (atomic) type is in the given list.*)
wenzelm@21588
   511
fun has_typed_var tycons =
paulson@21431
   512
  let fun var_tycon (Var (_, Type(a,_))) = a mem_string tycons
paulson@21431
   513
        | var_tycon _ = false
paulson@21431
   514
  in  exists var_tycon o term_vars  end;
paulson@21431
   515
paulson@22217
   516
(*Clauses are forbidden to contain variables of these types. The typical reason is that
paulson@22217
   517
  they lead to unsoundness. Note that "unit" satisfies numerous equations like ?X=().
paulson@22217
   518
  The resulting clause will have no type constraint, yielding false proofs. Even "bool"
paulson@22217
   519
  leads to many unsound proofs, though (obviously) only for higher-order problems.*)
paulson@24215
   520
val unwanted_types = ["Product_Type.unit","bool"];
paulson@22217
   521
paulson@21470
   522
fun unwanted t =
paulson@24958
   523
  is_taut t orelse has_typed_var unwanted_types t orelse
paulson@24958
   524
  forall too_general_equality (HOLogic.disjuncts (strip_Trueprop t));
paulson@21470
   525
paulson@21431
   526
(*Clauses containing variables of type "unit" or "bool" are unlikely to be useful and
paulson@21431
   527
  likely to lead to unsound proofs.*)
paulson@22217
   528
fun remove_unwanted_clauses cls = filter (not o unwanted o prop_of o fst) cls;
paulson@21431
   529
wenzelm@28477
   530
(* TODO: problem file for *one* subgoal would be sufficient *)
wenzelm@28477
   531
(*Write out problem files for each subgoal.
wenzelm@28477
   532
  Argument probfile generates filenames from subgoal-number
wenzelm@28477
   533
  Argument writer is either a tptp_write_file or dfg_write_file from ResHolClause
wenzelm@28477
   534
  Arguments max_new and theory_const are booleans controlling relevance_filter
wenzelm@28477
   535
    (necessary for different provers)
wenzelm@28477
   536
  *)
wenzelm@28477
   537
fun write_problem_files writer max_new theory_const probfile (ctxt, chain_ths, th) =
wenzelm@28477
   538
  let val goals = Thm.prems_of th
paulson@17717
   539
      val thy = ProofContext.theory_of ctxt
paulson@20457
   540
      fun get_neg_subgoals [] _ = []
wenzelm@22865
   541
        | get_neg_subgoals (gl::gls) n = #1 (ResAxioms.neg_conjecture_clauses th n) ::
paulson@21999
   542
                                         get_neg_subgoals gls (n+1)
paulson@20526
   543
      val goal_cls = get_neg_subgoals goals 1
paulson@25243
   544
                     handle THM ("assume: variables", _, _) => 
paulson@25243
   545
                       error "Sledgehammer: Goal contains type variables (TVars)"                       
wenzelm@28477
   546
      val isFO = case linkup_logic_mode of
paulson@24958
   547
			  Auto => forall (Meson.is_fol_term thy) (map prop_of (List.concat goal_cls))
paulson@24958
   548
			| Fol => true
paulson@24958
   549
			| Hol => false
paulson@24546
   550
      val (included_thms,white_thms) = get_clasimp_atp_lemmas ctxt chain_ths
wenzelm@27178
   551
      val included_cls = included_thms |> ResAxioms.cnf_rules_pairs thy |> make_unique
paulson@24958
   552
                                       |> restrict_to_logic thy isFO
paulson@21470
   553
                                       |> remove_unwanted_clauses
wenzelm@27178
   554
      val white_cls = ResAxioms.cnf_rules_pairs thy white_thms
paulson@20526
   555
      (*clauses relevant to goal gl*)
wenzelm@28477
   556
      val axcls_list = map (fn ngcls => white_cls @ relevance_filter max_new theory_const thy included_cls (map prop_of ngcls)) goal_cls
paulson@20526
   557
      fun write_all [] [] _ = []
wenzelm@21588
   558
        | write_all (ccls::ccls_list) (axcls::axcls_list) k =
wenzelm@28477
   559
            let val fname = File.platform_path (probfile k)
paulson@20868
   560
                val axcls = make_unique axcls
wenzelm@26928
   561
                val _ = app (fn th => Output.debug (fn _ => Display.string_of_thm th)) ccls
paulson@20868
   562
                val ccls = subtract_cls ccls axcls
wenzelm@26928
   563
                val _ = app (fn th => Output.debug (fn _ => Display.string_of_thm th)) ccls
paulson@21431
   564
                val ccltms = map prop_of ccls
paulson@21373
   565
                and axtms = map (prop_of o #1) axcls
paulson@21431
   566
                val subs = tfree_classes_of_terms ccltms
paulson@21431
   567
                and supers = tvar_classes_of_terms axtms
paulson@21431
   568
                and tycons = type_consts_of_terms thy (ccltms@axtms)
paulson@21290
   569
                (*TFrees in conjecture clauses; TVars in axiom clauses*)
paulson@22645
   570
                val (supers',arity_clauses) = ResClause.make_arity_clauses thy tycons supers
paulson@22645
   571
                val classrel_clauses = ResClause.make_classrel_clauses thy subs supers'
paulson@24958
   572
                val clnames = writer thy isFO ccls fname (axcls,classrel_clauses,arity_clauses) []
paulson@21888
   573
                val thm_names = Vector.fromList clnames
paulson@20870
   574
            in  (thm_names,fname) :: write_all ccls_list axcls_list (k+1)  end
paulson@20870
   575
      val (thm_names_list, filenames) = ListPair.unzip (write_all goal_cls axcls_list 1)
mengj@19194
   576
  in
paulson@20870
   577
      (filenames, thm_names_list)
mengj@19194
   578
  end;
quigley@15644
   579
paulson@15347
   580
end;