src/HOL/Tools/Sledgehammer/sledgehammer_filter.ML
author blanchet
Thu Dec 16 15:12:17 2010 +0100 (2010-12-16 ago)
changeset 41204 bd57cf5944cb
parent 41202 bf00e0a578e8
child 41205 209546e0af2c
permissions -rw-r--r--
get rid of experimental feature of term patterns in relevance filter -- doesn't work well unless we take into consideration the equality theory entailed by the relevant facts
blanchet@38988
     1
(*  Title:      HOL/Tools/Sledgehammer/sledgehammer_filter.ML
blanchet@38027
     2
    Author:     Jia Meng, Cambridge University Computer Laboratory and NICTA
blanchet@36393
     3
    Author:     Jasmin Blanchette, TU Muenchen
blanchet@39958
     4
blanchet@39958
     5
Sledgehammer's relevance filter.
wenzelm@33309
     6
*)
paulson@15452
     7
blanchet@38988
     8
signature SLEDGEHAMMER_FILTER =
wenzelm@16802
     9
sig
blanchet@38993
    10
  datatype locality = General | Intro | Elim | Simp | Local | Assum | Chained
blanchet@38752
    11
blanchet@40070
    12
  type relevance_fudge =
blanchet@41159
    13
    {allow_ext : bool,
blanchet@41159
    14
     worse_irrel_freq : real,
blanchet@40070
    15
     higher_order_irrel_weight : real,
blanchet@40070
    16
     abs_rel_weight : real,
blanchet@40070
    17
     abs_irrel_weight : real,
blanchet@40070
    18
     skolem_irrel_weight : real,
blanchet@40070
    19
     theory_const_rel_weight : real,
blanchet@40070
    20
     theory_const_irrel_weight : real,
blanchet@40070
    21
     intro_bonus : real,
blanchet@40070
    22
     elim_bonus : real,
blanchet@40070
    23
     simp_bonus : real,
blanchet@40070
    24
     local_bonus : real,
blanchet@40070
    25
     assum_bonus : real,
blanchet@40070
    26
     chained_bonus : real,
blanchet@40070
    27
     max_imperfect : real,
blanchet@40070
    28
     max_imperfect_exp : real,
blanchet@40070
    29
     threshold_divisor : real,
blanchet@40070
    30
     ridiculous_threshold : real}
blanchet@40070
    31
blanchet@35966
    32
  type relevance_override =
blanchet@40070
    33
    {add : (Facts.ref * Attrib.src list) list,
blanchet@40070
    34
     del : (Facts.ref * Attrib.src list) list,
blanchet@40070
    35
     only : bool}
blanchet@35966
    36
blanchet@37616
    37
  val trace : bool Unsynchronized.ref
blanchet@40205
    38
  val fact_from_ref :
blanchet@38996
    39
    Proof.context -> unit Symtab.table -> thm list
blanchet@38996
    40
    -> Facts.ref * Attrib.src list -> ((string * locality) * thm) list
blanchet@37347
    41
  val relevant_facts :
blanchet@41066
    42
    Proof.context -> bool -> real * real -> int
blanchet@41066
    43
    -> (string * typ -> term list -> bool) -> relevance_fudge
blanchet@41066
    44
    -> relevance_override -> thm list -> term list -> term
blanchet@40070
    45
    -> ((string * locality) * thm) list
paulson@15347
    46
end;
paulson@15347
    47
blanchet@38988
    48
structure Sledgehammer_Filter : SLEDGEHAMMER_FILTER =
paulson@15347
    49
struct
paulson@15347
    50
blanchet@38652
    51
open Sledgehammer_Util
blanchet@38652
    52
blanchet@37616
    53
val trace = Unsynchronized.ref false
blanchet@37616
    54
fun trace_msg msg = if !trace then tracing (msg ()) else ()
blanchet@35826
    55
blanchet@41204
    56
(* experimental feature *)
blanchet@38997
    57
val respect_no_atp = true
blanchet@38827
    58
blanchet@38993
    59
datatype locality = General | Intro | Elim | Simp | Local | Assum | Chained
blanchet@38752
    60
blanchet@40070
    61
type relevance_fudge =
blanchet@41159
    62
  {allow_ext : bool,
blanchet@41159
    63
   worse_irrel_freq : real,
blanchet@40070
    64
   higher_order_irrel_weight : real,
blanchet@40070
    65
   abs_rel_weight : real,
blanchet@40070
    66
   abs_irrel_weight : real,
blanchet@40070
    67
   skolem_irrel_weight : real,
blanchet@40070
    68
   theory_const_rel_weight : real,
blanchet@40070
    69
   theory_const_irrel_weight : real,
blanchet@40070
    70
   intro_bonus : real,
blanchet@40070
    71
   elim_bonus : real,
blanchet@40070
    72
   simp_bonus : real,
blanchet@40070
    73
   local_bonus : real,
blanchet@40070
    74
   assum_bonus : real,
blanchet@40070
    75
   chained_bonus : real,
blanchet@40070
    76
   max_imperfect : real,
blanchet@40070
    77
   max_imperfect_exp : real,
blanchet@40070
    78
   threshold_divisor : real,
blanchet@40070
    79
   ridiculous_threshold : real}
blanchet@40070
    80
blanchet@35966
    81
type relevance_override =
blanchet@40070
    82
  {add : (Facts.ref * Attrib.src list) list,
blanchet@40070
    83
   del : (Facts.ref * Attrib.src list) list,
blanchet@40070
    84
   only : bool}
paulson@21070
    85
blanchet@37616
    86
val sledgehammer_prefix = "Sledgehammer" ^ Long_Name.separator
blanchet@39896
    87
val abs_name = sledgehammer_prefix ^ "abs"
blanchet@39896
    88
val skolem_prefix = sledgehammer_prefix ^ "sko"
blanchet@38992
    89
val theory_const_suffix = Long_Name.separator ^ " 1"
blanchet@37616
    90
blanchet@40227
    91
fun needs_quoting reserved s =
blanchet@40375
    92
  Symtab.defined reserved s orelse
blanchet@40375
    93
  exists (not o Syntax.is_identifier) (Long_Name.explode s)
blanchet@40227
    94
blanchet@40375
    95
fun make_name reserved multi j name =
blanchet@40227
    96
  (name |> needs_quoting reserved name ? quote) ^
blanchet@38744
    97
  (if multi then "(" ^ Int.toString j ^ ")" else "")
blanchet@38744
    98
blanchet@40375
    99
fun explode_interval _ (Facts.FromTo (i, j)) = i upto j
blanchet@40375
   100
  | explode_interval max (Facts.From i) = i upto i + max - 1
blanchet@40375
   101
  | explode_interval _ (Facts.Single i) = [i]
blanchet@40375
   102
blanchet@40205
   103
fun fact_from_ref ctxt reserved chained_ths (xthm as (xref, args)) =
blanchet@38744
   104
  let
blanchet@38996
   105
    val ths = Attrib.eval_thms ctxt [xthm]
blanchet@38996
   106
    val bracket =
blanchet@38996
   107
      implode (map (fn arg => "[" ^ Pretty.str_of (Args.pretty_src ctxt arg)
blanchet@38996
   108
                               ^ "]") args)
blanchet@40375
   109
    fun nth_name j =
blanchet@38996
   110
      case xref of
blanchet@39264
   111
        Facts.Fact s => "`" ^ s ^ "`" ^ bracket
blanchet@38996
   112
      | Facts.Named (("", _), _) => "[" ^ bracket ^ "]"
blanchet@40375
   113
      | Facts.Named ((name, _), NONE) =>
blanchet@40375
   114
        make_name reserved (length ths > 1) (j + 1) name ^ bracket
blanchet@40375
   115
      | Facts.Named ((name, _), SOME intervals) =>
blanchet@40375
   116
        make_name reserved true
blanchet@40375
   117
                 (nth (maps (explode_interval (length ths)) intervals) j) name ^
blanchet@40375
   118
        bracket
blanchet@38744
   119
  in
blanchet@40375
   120
    (ths, (0, []))
blanchet@38752
   121
    |-> fold (fn th => fn (j, rest) =>
blanchet@40375
   122
                 (j + 1, ((nth_name j,
blanchet@38752
   123
                          if member Thm.eq_thm chained_ths th then Chained
blanchet@38752
   124
                          else General), th) :: rest))
blanchet@38744
   125
    |> snd
blanchet@38699
   126
  end
blanchet@37616
   127
blanchet@41199
   128
(* This is a terrible hack. Free variables are sometimes code as "M__" when they
blanchet@41199
   129
   are displayed as "M" and we want to avoid clashes with these. But sometimes
blanchet@41199
   130
   it's even worse: "Ma__" encodes "M". So we simply reserve all prefixes of all
blanchet@41199
   131
   free variables. In the worse case scenario, where the fact won't be resolved
blanchet@41199
   132
   correctly, the user can fix it manually, e.g., by naming the fact in
blanchet@41199
   133
   question. Ideally we would need nothing of it, but backticks just don't work
blanchet@41199
   134
   with schematic variables. *)
blanchet@41199
   135
fun all_prefixes_of s =
blanchet@41199
   136
  map (fn i => String.extract (s, 0, SOME i)) (1 upto size s - 1)
blanchet@41199
   137
fun close_form t =
blanchet@41199
   138
  (t, [] |> Term.add_free_names t |> maps all_prefixes_of)
blanchet@41199
   139
  |> fold (fn ((s, i), T) => fn (t', taken) =>
blanchet@41199
   140
              let val s' = Name.variant taken s in
blanchet@41199
   141
                ((if fastype_of t' = HOLogic.boolT then HOLogic.all_const
blanchet@41199
   142
                  else Term.all) T
blanchet@41199
   143
                 $ Abs (s', T, abstract_over (Var ((s, i), T), t')),
blanchet@41199
   144
                 s' :: taken)
blanchet@41199
   145
              end)
blanchet@41199
   146
          (Term.add_vars t [] |> sort_wrt (fst o fst))
blanchet@41199
   147
  |> fst
blanchet@41199
   148
blanchet@41199
   149
fun string_for_term ctxt t =
blanchet@41199
   150
  Print_Mode.setmp (filter (curry (op =) Symbol.xsymbolsN)
blanchet@41199
   151
                   (print_mode_value ())) (Syntax.string_of_term ctxt) t
blanchet@41199
   152
  |> String.translate (fn c => if Char.isPrint c then str c else "")
blanchet@41199
   153
  |> simplify_spaces
blanchet@41199
   154
blanchet@41199
   155
(** Structural induction rules **)
blanchet@41199
   156
blanchet@41200
   157
fun struct_induct_rule_on th =
blanchet@41199
   158
  case Logic.strip_horn (prop_of th) of
blanchet@41199
   159
    (prems, @{const Trueprop}
blanchet@41199
   160
            $ ((p as Var ((p_name, 0), _)) $ (a as Var (_, ind_T)))) =>
blanchet@41199
   161
    if not (is_TVar ind_T) andalso length prems > 1 andalso
blanchet@41199
   162
       exists (exists_subterm (curry (op aconv) p)) prems andalso
blanchet@41199
   163
       not (exists (exists_subterm (curry (op aconv) a)) prems) then
blanchet@41199
   164
      SOME (p_name, ind_T)
blanchet@41199
   165
    else
blanchet@41199
   166
      NONE
blanchet@41199
   167
  | _ => NONE
blanchet@41199
   168
blanchet@41202
   169
fun instantiate_induct_rule ctxt concl_prop p_name ((name, _), (multi, th))
blanchet@41199
   170
                            ind_x =
blanchet@41199
   171
  let
blanchet@41199
   172
    fun varify_noninducts (t as Free (s, T)) =
blanchet@41199
   173
        if (s, T) = ind_x orelse can dest_funT T then t else Var ((s, 0), T)
blanchet@41199
   174
      | varify_noninducts t = t
blanchet@41199
   175
    val p_inst =
blanchet@41199
   176
      concl_prop |> map_aterms varify_noninducts |> close_form
blanchet@41199
   177
                 |> lambda (Free ind_x)
blanchet@41199
   178
                 |> string_for_term ctxt
blanchet@41199
   179
  in
blanchet@41202
   180
    ((fn () => name () ^ "[where " ^ p_name ^ " = " ^ quote p_inst ^ "]",
blanchet@41202
   181
      Local),
blanchet@41199
   182
     (multi, th |> read_instantiate ctxt [((p_name, 0), p_inst)]))
blanchet@41199
   183
  end
blanchet@41199
   184
blanchet@41199
   185
fun type_match thy (T1, T2) =
blanchet@41199
   186
  (Sign.typ_match thy (T2, T1) Vartab.empty; true)
blanchet@41199
   187
  handle Type.TYPE_MATCH => false
blanchet@41199
   188
blanchet@41199
   189
fun instantiate_if_induct_rule ctxt stmt stmt_xs (ax as (_, (_, th))) =
blanchet@41200
   190
  case struct_induct_rule_on th of
blanchet@41199
   191
    SOME (p_name, ind_T) =>
blanchet@41199
   192
    let val thy = ProofContext.theory_of ctxt in
blanchet@41199
   193
      stmt_xs |> filter (fn (_, T) => type_match thy (T, ind_T))
blanchet@41199
   194
              |> map_filter (try (instantiate_induct_rule ctxt stmt p_name ax))
blanchet@41199
   195
    end
blanchet@41199
   196
  | NONE => [ax]
blanchet@41199
   197
wenzelm@28477
   198
(***************************************************************)
wenzelm@28477
   199
(* Relevance Filtering                                         *)
wenzelm@28477
   200
(***************************************************************)
mengj@19194
   201
paulson@24287
   202
(*** constants with types ***)
paulson@24287
   203
blanchet@38939
   204
fun order_of_type (Type (@{type_name fun}, [T1, @{typ bool}])) =
blanchet@38939
   205
    order_of_type T1 (* cheat: pretend sets are first-order *)
blanchet@38939
   206
  | order_of_type (Type (@{type_name fun}, [T1, T2])) =
blanchet@38939
   207
    Int.max (order_of_type T1 + 1, order_of_type T2)
blanchet@38939
   208
  | order_of_type (Type (_, Ts)) = fold (Integer.max o order_of_type) Ts 0
blanchet@38939
   209
  | order_of_type _ = 0
blanchet@38939
   210
blanchet@38823
   211
(* An abstraction of Isabelle types and first-order terms *)
blanchet@38823
   212
datatype pattern = PVar | PApp of string * pattern list
blanchet@38939
   213
datatype ptype = PType of int * pattern list
blanchet@38744
   214
blanchet@38823
   215
fun string_for_pattern PVar = "_"
blanchet@38823
   216
  | string_for_pattern (PApp (s, ps)) =
blanchet@38823
   217
    if null ps then s else s ^ string_for_patterns ps
blanchet@38823
   218
and string_for_patterns ps = "(" ^ commas (map string_for_pattern ps) ^ ")"
blanchet@38939
   219
fun string_for_ptype (PType (_, ps)) = string_for_patterns ps
paulson@24287
   220
paulson@24287
   221
(*Is the second type an instance of the first one?*)
blanchet@38824
   222
fun match_pattern (PVar, _) = true
blanchet@38824
   223
  | match_pattern (PApp _, PVar) = false
blanchet@38824
   224
  | match_pattern (PApp (s, ps), PApp (t, qs)) =
blanchet@38824
   225
    s = t andalso match_patterns (ps, qs)
blanchet@38824
   226
and match_patterns (_, []) = true
blanchet@38824
   227
  | match_patterns ([], _) = false
blanchet@38824
   228
  | match_patterns (p :: ps, q :: qs) =
blanchet@38824
   229
    match_pattern (p, q) andalso match_patterns (ps, qs)
blanchet@38939
   230
fun match_ptype (PType (_, ps), PType (_, qs)) = match_patterns (ps, qs)
paulson@24287
   231
blanchet@38823
   232
(* Is there a unifiable constant? *)
blanchet@38827
   233
fun pconst_mem f consts (s, ps) =
blanchet@38939
   234
  exists (curry (match_ptype o f) ps)
blanchet@38827
   235
         (map snd (filter (curry (op =) s o fst) consts))
blanchet@38827
   236
fun pconst_hyper_mem f const_tab (s, ps) =
blanchet@38939
   237
  exists (curry (match_ptype o f) ps) (these (Symtab.lookup const_tab s))
blanchet@37505
   238
blanchet@38939
   239
fun pattern_for_type (Type (s, Ts)) = PApp (s, map pattern_for_type Ts)
blanchet@38939
   240
  | pattern_for_type (TFree (s, _)) = PApp (s, [])
blanchet@38939
   241
  | pattern_for_type (TVar _) = PVar
blanchet@38827
   242
blanchet@38744
   243
(* Pairs a constant with the list of its type instantiations. *)
blanchet@41204
   244
fun ptype thy const x =
blanchet@38939
   245
  (if const then map pattern_for_type (these (try (Sign.const_typargs thy) x))
blanchet@41204
   246
   else [])
blanchet@41204
   247
fun rich_ptype thy const (s, T) =
blanchet@41204
   248
  PType (order_of_type T, ptype thy const (s, T))
blanchet@41204
   249
fun rich_pconst thy const (s, T) = (s, rich_ptype thy const (s, T))
paulson@24287
   250
blanchet@38939
   251
fun string_for_hyper_pconst (s, ps) =
blanchet@38939
   252
  s ^ "{" ^ commas (map string_for_ptype ps) ^ "}"
paulson@24287
   253
blanchet@38823
   254
(* Add a pconstant to the table, but a [] entry means a standard
blanchet@38819
   255
   connective, which we ignore.*)
blanchet@41066
   256
fun add_pconst_to_table also_skolem (s, p) =
blanchet@41066
   257
  if (not also_skolem andalso String.isPrefix skolem_prefix s) then I
blanchet@41066
   258
  else Symtab.map_default (s, [p]) (insert (op =) p)
blanchet@38819
   259
blanchet@38819
   260
fun is_formula_type T = (T = HOLogic.boolT orelse T = propT)
blanchet@38819
   261
blanchet@40369
   262
fun pconsts_in_terms thy is_built_in_const also_skolems pos ts =
blanchet@37505
   263
  let
blanchet@41202
   264
    (* Giving a penalty to Skolems is a good idea, but that penalty shouldn't
blanchet@41202
   265
       increase linearly with the number of Skolems in the fact. Hence, we
blanchet@41202
   266
       create equivalence classes of Skolems by reusing fresh Skolem names. *)
blanchet@41202
   267
    val skolem_cache = Unsynchronized.ref ""
blanchet@41202
   268
    fun get_skolem_name () =
blanchet@41202
   269
      (if !skolem_cache = "" then skolem_cache := gensym skolem_prefix else ();
blanchet@41202
   270
       !skolem_cache)
blanchet@38819
   271
    val flip = Option.map not
blanchet@38587
   272
    (* We include free variables, as well as constants, to handle locales. For
blanchet@38587
   273
       each quantifiers that must necessarily be skolemized by the ATP, we
blanchet@38587
   274
       introduce a fresh constant to simulate the effect of Skolemization. *)
blanchet@40373
   275
    fun do_const const x ts =
blanchet@41066
   276
      (not (is_built_in_const x ts)
blanchet@41204
   277
       ? add_pconst_to_table also_skolems (rich_pconst thy const x))
blanchet@38827
   278
      #> fold do_term ts
blanchet@38827
   279
    and do_term t =
blanchet@38827
   280
      case strip_comb t of
blanchet@38827
   281
        (Const x, ts) => do_const true x ts
blanchet@38827
   282
      | (Free x, ts) => do_const false x ts
blanchet@38939
   283
      | (Abs (_, T, t'), ts) =>
blanchet@38939
   284
        (null ts
blanchet@41066
   285
         ? add_pconst_to_table true (abs_name, PType (order_of_type T + 1, [])))
blanchet@38827
   286
        #> fold do_term (t' :: ts)
blanchet@38827
   287
      | (_, ts) => fold do_term ts
blanchet@38939
   288
    fun do_quantifier will_surely_be_skolemized abs_T body_t =
blanchet@37537
   289
      do_formula pos body_t
blanchet@38747
   290
      #> (if also_skolems andalso will_surely_be_skolemized then
blanchet@41066
   291
            add_pconst_to_table true
blanchet@41202
   292
                (get_skolem_name (), PType (order_of_type abs_T, []))
blanchet@38587
   293
          else
blanchet@38587
   294
            I)
blanchet@38587
   295
    and do_term_or_formula T =
blanchet@38692
   296
      if is_formula_type T then do_formula NONE else do_term
blanchet@37537
   297
    and do_formula pos t =
blanchet@37537
   298
      case t of
blanchet@38939
   299
        Const (@{const_name all}, _) $ Abs (_, T, t') =>
blanchet@38939
   300
        do_quantifier (pos = SOME false) T t'
blanchet@37537
   301
      | @{const "==>"} $ t1 $ t2 =>
blanchet@37537
   302
        do_formula (flip pos) t1 #> do_formula pos t2
blanchet@37537
   303
      | Const (@{const_name "=="}, Type (_, [T, _])) $ t1 $ t2 =>
blanchet@38587
   304
        fold (do_term_or_formula T) [t1, t2]
blanchet@37537
   305
      | @{const Trueprop} $ t1 => do_formula pos t1
blanchet@41140
   306
      | @{const False} => I
blanchet@41140
   307
      | @{const True} => I
blanchet@37537
   308
      | @{const Not} $ t1 => do_formula (flip pos) t1
blanchet@38939
   309
      | Const (@{const_name All}, _) $ Abs (_, T, t') =>
blanchet@38939
   310
        do_quantifier (pos = SOME false) T t'
blanchet@38939
   311
      | Const (@{const_name Ex}, _) $ Abs (_, T, t') =>
blanchet@38939
   312
        do_quantifier (pos = SOME true) T t'
haftmann@38795
   313
      | @{const HOL.conj} $ t1 $ t2 => fold (do_formula pos) [t1, t2]
haftmann@38795
   314
      | @{const HOL.disj} $ t1 $ t2 => fold (do_formula pos) [t1, t2]
haftmann@38786
   315
      | @{const HOL.implies} $ t1 $ t2 =>
blanchet@37537
   316
        do_formula (flip pos) t1 #> do_formula pos t2
haftmann@38864
   317
      | Const (@{const_name HOL.eq}, Type (_, [T, _])) $ t1 $ t2 =>
blanchet@38587
   318
        fold (do_term_or_formula T) [t1, t2]
blanchet@38587
   319
      | Const (@{const_name If}, Type (_, [_, Type (_, [T, _])]))
blanchet@38587
   320
        $ t1 $ t2 $ t3 =>
blanchet@38587
   321
        do_formula NONE t1 #> fold (do_term_or_formula T) [t2, t3]
blanchet@38939
   322
      | Const (@{const_name Ex1}, _) $ Abs (_, T, t') =>
blanchet@38939
   323
        do_quantifier (is_some pos) T t'
blanchet@38939
   324
      | Const (@{const_name Ball}, _) $ t1 $ Abs (_, T, t') =>
blanchet@38939
   325
        do_quantifier (pos = SOME false) T
blanchet@38939
   326
                      (HOLogic.mk_imp (incr_boundvars 1 t1 $ Bound 0, t'))
blanchet@38939
   327
      | Const (@{const_name Bex}, _) $ t1 $ Abs (_, T, t') =>
blanchet@38939
   328
        do_quantifier (pos = SOME true) T
blanchet@38939
   329
                      (HOLogic.mk_conj (incr_boundvars 1 t1 $ Bound 0, t'))
blanchet@37537
   330
      | (t0 as Const (_, @{typ bool})) $ t1 =>
blanchet@37537
   331
        do_term t0 #> do_formula pos t1  (* theory constant *)
blanchet@37537
   332
      | _ => do_term t
blanchet@38819
   333
  in Symtab.empty |> fold (do_formula pos) ts end
paulson@24287
   334
paulson@24287
   335
(*Inserts a dummy "constant" referring to the theory name, so that relevance
paulson@24287
   336
  takes the given theory into account.*)
blanchet@41200
   337
fun theory_constify ({theory_const_rel_weight, theory_const_irrel_weight, ...}
blanchet@41200
   338
                     : relevance_fudge) thy_name t =
blanchet@40070
   339
  if exists (curry (op <) 0.0) [theory_const_rel_weight,
blanchet@40070
   340
                                theory_const_irrel_weight] then
blanchet@41200
   341
    Const (thy_name ^ theory_const_suffix, @{typ bool}) $ t
blanchet@37505
   342
  else
blanchet@41200
   343
    t
blanchet@41200
   344
blanchet@41200
   345
fun theory_const_prop_of fudge th =
blanchet@41200
   346
  theory_constify fudge (Context.theory_name (theory_of_thm th)) (prop_of th)
blanchet@37505
   347
paulson@24287
   348
(**** Constant / Type Frequencies ****)
paulson@24287
   349
blanchet@38743
   350
(* A two-dimensional symbol table counts frequencies of constants. It's keyed
blanchet@38743
   351
   first by constant name and second by its list of type instantiations. For the
blanchet@38823
   352
   latter, we need a linear ordering on "pattern list". *)
paulson@24287
   353
blanchet@38823
   354
fun pattern_ord p =
blanchet@38743
   355
  case p of
blanchet@38744
   356
    (PVar, PVar) => EQUAL
blanchet@38823
   357
  | (PVar, PApp _) => LESS
blanchet@38823
   358
  | (PApp _, PVar) => GREATER
blanchet@38823
   359
  | (PApp q1, PApp q2) =>
blanchet@38823
   360
    prod_ord fast_string_ord (dict_ord pattern_ord) (q1, q2)
blanchet@38939
   361
fun ptype_ord (PType p, PType q) =
blanchet@38939
   362
  prod_ord (dict_ord pattern_ord) int_ord (swap p, swap q)
paulson@24287
   363
blanchet@38939
   364
structure PType_Tab = Table(type key = ptype val ord = ptype_ord)
paulson@24287
   365
blanchet@40204
   366
fun count_fact_consts thy fudge =
blanchet@37503
   367
  let
blanchet@38827
   368
    fun do_const const (s, T) ts =
blanchet@38827
   369
      (* Two-dimensional table update. Constant maps to types maps to count. *)
blanchet@41204
   370
      PType_Tab.map_default (rich_ptype thy const (s, T), 0) (Integer.add 1)
blanchet@38939
   371
      |> Symtab.map_default (s, PType_Tab.empty)
blanchet@38827
   372
      #> fold do_term ts
blanchet@38827
   373
    and do_term t =
blanchet@38827
   374
      case strip_comb t of
blanchet@38827
   375
        (Const x, ts) => do_const true x ts
blanchet@38827
   376
      | (Free x, ts) => do_const false x ts
blanchet@38827
   377
      | (Abs (_, _, t'), ts) => fold do_term (t' :: ts)
blanchet@38827
   378
      | (_, ts) => fold do_term ts
blanchet@40070
   379
  in do_term o theory_const_prop_of fudge o snd end
paulson@24287
   380
paulson@24287
   381
paulson@24287
   382
(**** Actual Filtering Code ****)
paulson@24287
   383
blanchet@39367
   384
fun pow_int _ 0 = 1.0
blanchet@38939
   385
  | pow_int x 1 = x
blanchet@38939
   386
  | pow_int x n = if n > 0 then x * pow_int x (n - 1) else pow_int x (n + 1) / x
blanchet@38939
   387
paulson@24287
   388
(*The frequency of a constant is the sum of those of all instances of its type.*)
blanchet@38824
   389
fun pconst_freq match const_tab (c, ps) =
blanchet@38939
   390
  PType_Tab.fold (fn (qs, m) => match (ps, qs) ? Integer.add m)
blanchet@38939
   391
                 (the (Symtab.lookup const_tab c)) 0
blanchet@38686
   392
paulson@24287
   393
blanchet@38085
   394
(* A surprising number of theorems contain only a few significant constants.
blanchet@38085
   395
   These include all induction rules, and other general theorems. *)
blanchet@37503
   396
blanchet@37503
   397
(* "log" seems best in practice. A constant function of one ignores the constant
blanchet@38938
   398
   frequencies. Rare constants give more points if they are relevant than less
blanchet@38938
   399
   rare ones. *)
blanchet@39367
   400
fun rel_weight_for _ freq = 1.0 + 2.0 / Math.ln (Real.fromInt freq + 1.0)
blanchet@38938
   401
blanchet@38938
   402
(* Irrelevant constants are treated differently. We associate lower penalties to
blanchet@38938
   403
   very rare constants and very common ones -- the former because they can't
blanchet@38938
   404
   lead to the inclusion of too many new facts, and the latter because they are
blanchet@38938
   405
   so common as to be of little interest. *)
blanchet@40070
   406
fun irrel_weight_for ({worse_irrel_freq, higher_order_irrel_weight, ...}
blanchet@40070
   407
                      : relevance_fudge) order freq =
blanchet@40070
   408
  let val (k, x) = worse_irrel_freq |> `Real.ceil in
blanchet@38939
   409
    (if freq < k then Math.ln (Real.fromInt (freq + 1)) / Math.ln x
blanchet@38939
   410
     else rel_weight_for order freq / rel_weight_for order k)
blanchet@40070
   411
    * pow_int higher_order_irrel_weight (order - 1)
blanchet@38938
   412
  end
blanchet@37503
   413
blanchet@37503
   414
(* Computes a constant's weight, as determined by its frequency. *)
blanchet@38992
   415
fun generic_pconst_weight abs_weight skolem_weight theory_const_weight
blanchet@38992
   416
                          weight_for f const_tab (c as (s, PType (m, _))) =
blanchet@38816
   417
  if s = abs_name then abs_weight
blanchet@38816
   418
  else if String.isPrefix skolem_prefix s then skolem_weight
blanchet@38992
   419
  else if String.isSuffix theory_const_suffix s then theory_const_weight
blanchet@38939
   420
  else weight_for m (pconst_freq (match_ptype o f) const_tab c)
blanchet@38821
   421
blanchet@40070
   422
fun rel_pconst_weight ({abs_rel_weight, theory_const_rel_weight, ...}
blanchet@40070
   423
                       : relevance_fudge) const_tab =
blanchet@40070
   424
  generic_pconst_weight abs_rel_weight 0.0 theory_const_rel_weight
blanchet@38992
   425
                        rel_weight_for I const_tab
blanchet@40070
   426
fun irrel_pconst_weight (fudge as {abs_irrel_weight, skolem_irrel_weight,
blanchet@40070
   427
                                   theory_const_irrel_weight, ...}) const_tab =
blanchet@40070
   428
  generic_pconst_weight abs_irrel_weight skolem_irrel_weight
blanchet@40070
   429
                        theory_const_irrel_weight (irrel_weight_for fudge) swap
blanchet@40070
   430
                        const_tab
paulson@24287
   431
blanchet@40070
   432
fun locality_bonus (_ : relevance_fudge) General = 0.0
blanchet@40070
   433
  | locality_bonus {intro_bonus, ...} Intro = intro_bonus
blanchet@40070
   434
  | locality_bonus {elim_bonus, ...} Elim = elim_bonus
blanchet@40070
   435
  | locality_bonus {simp_bonus, ...} Simp = simp_bonus
blanchet@40070
   436
  | locality_bonus {local_bonus, ...} Local = local_bonus
blanchet@40070
   437
  | locality_bonus {assum_bonus, ...} Assum = assum_bonus
blanchet@40070
   438
  | locality_bonus {chained_bonus, ...} Chained = chained_bonus
blanchet@38751
   439
blanchet@40418
   440
fun is_odd_const_name s =
blanchet@40418
   441
  s = abs_name orelse String.isPrefix skolem_prefix s orelse
blanchet@40418
   442
  String.isSuffix theory_const_suffix s
blanchet@40418
   443
blanchet@40204
   444
fun fact_weight fudge loc const_tab relevant_consts fact_consts =
blanchet@40204
   445
  case fact_consts |> List.partition (pconst_hyper_mem I relevant_consts)
blanchet@40204
   446
                   ||> filter_out (pconst_hyper_mem swap relevant_consts) of
blanchet@38827
   447
    ([], _) => 0.0
blanchet@38744
   448
  | (rel, irrel) =>
blanchet@40418
   449
    if forall (forall (is_odd_const_name o fst)) [rel, irrel] then
blanchet@40371
   450
      0.0
blanchet@40371
   451
    else
blanchet@40371
   452
      let
blanchet@40371
   453
        val irrel = irrel |> filter_out (pconst_mem swap rel)
blanchet@40371
   454
        val rel_weight =
blanchet@40371
   455
          0.0 |> fold (curry (op +) o rel_pconst_weight fudge const_tab) rel
blanchet@40371
   456
        val irrel_weight =
blanchet@40371
   457
          ~ (locality_bonus fudge loc)
blanchet@40371
   458
          |> fold (curry (op +) o irrel_pconst_weight fudge const_tab) irrel
blanchet@40371
   459
        val res = rel_weight / (rel_weight + irrel_weight)
blanchet@40371
   460
      in if Real.isFinite res then res else 0.0 end
blanchet@38747
   461
blanchet@40369
   462
fun pconsts_in_fact thy is_built_in_const t =
blanchet@38825
   463
  Symtab.fold (fn (s, pss) => fold (cons o pair s) pss)
blanchet@40369
   464
              (pconsts_in_terms thy is_built_in_const true (SOME true) [t]) []
blanchet@40369
   465
fun pair_consts_fact thy is_built_in_const fudge fact =
blanchet@40204
   466
  case fact |> snd |> theory_const_prop_of fudge
blanchet@40369
   467
            |> pconsts_in_fact thy is_built_in_const of
blanchet@38827
   468
    [] => NONE
blanchet@40204
   469
  | consts => SOME ((fact, consts), NONE)
paulson@24287
   470
blanchet@38699
   471
type annotated_thm =
blanchet@38939
   472
  (((unit -> string) * locality) * thm) * (string * ptype) list
blanchet@37505
   473
blanchet@38904
   474
fun take_most_relevant max_relevant remaining_max
blanchet@40070
   475
        ({max_imperfect, max_imperfect_exp, ...} : relevance_fudge) 
blanchet@40070
   476
        (candidates : (annotated_thm * real) list) =
blanchet@38744
   477
  let
blanchet@38747
   478
    val max_imperfect =
blanchet@40070
   479
      Real.ceil (Math.pow (max_imperfect,
blanchet@38904
   480
                    Math.pow (Real.fromInt remaining_max
blanchet@40070
   481
                              / Real.fromInt max_relevant, max_imperfect_exp)))
blanchet@38747
   482
    val (perfect, imperfect) =
blanchet@38889
   483
      candidates |> sort (Real.compare o swap o pairself snd)
blanchet@38889
   484
                 |> take_prefix (fn (_, w) => w > 0.99999)
blanchet@38747
   485
    val ((accepts, more_rejects), rejects) =
blanchet@38747
   486
      chop max_imperfect imperfect |>> append perfect |>> chop remaining_max
blanchet@38744
   487
  in
blanchet@38889
   488
    trace_msg (fn () =>
blanchet@38889
   489
        "Actually passed (" ^ Int.toString (length accepts) ^ " of " ^
blanchet@38889
   490
        Int.toString (length candidates) ^ "): " ^
blanchet@38889
   491
        (accepts |> map (fn ((((name, _), _), _), weight) =>
blanchet@38752
   492
                            name () ^ " [" ^ Real.toString weight ^ "]")
blanchet@38745
   493
                 |> commas));
blanchet@38747
   494
    (accepts, more_rejects @ rejects)
blanchet@38744
   495
  end
paulson@24287
   496
blanchet@40369
   497
fun if_empty_replace_with_locality thy is_built_in_const facts loc tab =
blanchet@38819
   498
  if Symtab.is_empty tab then
blanchet@40369
   499
    pconsts_in_terms thy is_built_in_const false (SOME false)
blanchet@38819
   500
        (map_filter (fn ((_, loc'), th) =>
blanchet@40204
   501
                        if loc' = loc then SOME (prop_of th) else NONE) facts)
blanchet@38819
   502
  else
blanchet@38819
   503
    tab
blanchet@38819
   504
blanchet@41158
   505
fun add_arities is_built_in_const th =
blanchet@41158
   506
  let
blanchet@41158
   507
    fun aux _ _ NONE = NONE
blanchet@41158
   508
      | aux t args (SOME tab) =
blanchet@41158
   509
        case t of
blanchet@41158
   510
          t1 $ t2 => SOME tab |> aux t1 (t2 :: args) |> aux t2 []
blanchet@41158
   511
        | Const (x as (s, _)) =>
blanchet@41158
   512
          (if is_built_in_const x args then
blanchet@41158
   513
             SOME tab
blanchet@41158
   514
           else case Symtab.lookup tab s of
blanchet@41158
   515
             NONE => SOME (Symtab.update (s, length args) tab)
blanchet@41158
   516
           | SOME n => if n = length args then SOME tab else NONE)
blanchet@41158
   517
        | _ => SOME tab
blanchet@41158
   518
  in aux (prop_of th) [] end
blanchet@41158
   519
blanchet@41158
   520
fun needs_ext is_built_in_const facts =
blanchet@41158
   521
  fold (add_arities is_built_in_const o snd) facts (SOME Symtab.empty)
blanchet@41158
   522
  |> is_none
blanchet@41158
   523
blanchet@40369
   524
fun relevance_filter ctxt threshold0 decay max_relevant is_built_in_const
blanchet@40070
   525
        (fudge as {threshold_divisor, ridiculous_threshold, ...})
blanchet@40204
   526
        ({add, del, ...} : relevance_override) facts goal_ts =
blanchet@38739
   527
  let
blanchet@38739
   528
    val thy = ProofContext.theory_of ctxt
blanchet@40204
   529
    val const_tab = fold (count_fact_consts thy fudge) facts Symtab.empty
blanchet@38819
   530
    val goal_const_tab =
blanchet@40369
   531
      pconsts_in_terms thy is_built_in_const false (SOME false) goal_ts
blanchet@40369
   532
      |> fold (if_empty_replace_with_locality thy is_built_in_const facts)
blanchet@38993
   533
              [Chained, Assum, Local]
blanchet@39012
   534
    val add_ths = Attrib.eval_thms ctxt add
blanchet@39012
   535
    val del_ths = Attrib.eval_thms ctxt del
blanchet@40204
   536
    val facts = facts |> filter_out (member Thm.eq_thm del_ths o snd)
blanchet@38747
   537
    fun iter j remaining_max threshold rel_const_tab hopeless hopeful =
blanchet@38739
   538
      let
blanchet@40191
   539
        fun relevant [] _ [] =
blanchet@38747
   540
            (* Nothing has been added this iteration. *)
blanchet@40070
   541
            if j = 0 andalso threshold >= ridiculous_threshold then
blanchet@38747
   542
              (* First iteration? Try again. *)
blanchet@40070
   543
              iter 0 max_relevant (threshold / threshold_divisor) rel_const_tab
blanchet@38747
   544
                   hopeless hopeful
blanchet@38744
   545
            else
blanchet@40191
   546
              []
blanchet@38889
   547
          | relevant candidates rejects [] =
blanchet@38739
   548
            let
blanchet@38747
   549
              val (accepts, more_rejects) =
blanchet@40070
   550
                take_most_relevant max_relevant remaining_max fudge candidates
blanchet@38739
   551
              val rel_const_tab' =
blanchet@38745
   552
                rel_const_tab
blanchet@41066
   553
                |> fold (add_pconst_to_table false) (maps (snd o fst) accepts)
blanchet@38744
   554
              fun is_dirty (c, _) =
blanchet@38744
   555
                Symtab.lookup rel_const_tab' c <> Symtab.lookup rel_const_tab c
blanchet@38745
   556
              val (hopeful_rejects, hopeless_rejects) =
blanchet@38745
   557
                 (rejects @ hopeless, ([], []))
blanchet@38745
   558
                 |-> fold (fn (ax as (_, consts), old_weight) =>
blanchet@38745
   559
                              if exists is_dirty consts then
blanchet@38745
   560
                                apfst (cons (ax, NONE))
blanchet@38745
   561
                              else
blanchet@38745
   562
                                apsnd (cons (ax, old_weight)))
blanchet@38745
   563
                 |>> append (more_rejects
blanchet@38745
   564
                             |> map (fn (ax as (_, consts), old_weight) =>
blanchet@38745
   565
                                        (ax, if exists is_dirty consts then NONE
blanchet@38745
   566
                                             else SOME old_weight)))
blanchet@38747
   567
              val threshold =
blanchet@38822
   568
                1.0 - (1.0 - threshold)
blanchet@38822
   569
                      * Math.pow (decay, Real.fromInt (length accepts))
blanchet@38747
   570
              val remaining_max = remaining_max - length accepts
blanchet@38739
   571
            in
blanchet@38744
   572
              trace_msg (fn () => "New or updated constants: " ^
blanchet@38744
   573
                  commas (rel_const_tab' |> Symtab.dest
blanchet@38822
   574
                          |> subtract (op =) (rel_const_tab |> Symtab.dest)
blanchet@38827
   575
                          |> map string_for_hyper_pconst));
blanchet@38745
   576
              map (fst o fst) accepts @
blanchet@38747
   577
              (if remaining_max = 0 then
blanchet@40191
   578
                 []
blanchet@38745
   579
               else
blanchet@38747
   580
                 iter (j + 1) remaining_max threshold rel_const_tab'
blanchet@38747
   581
                      hopeless_rejects hopeful_rejects)
blanchet@38739
   582
            end
blanchet@38889
   583
          | relevant candidates rejects
blanchet@40204
   584
                     (((ax as (((_, loc), _), fact_consts)), cached_weight)
blanchet@38747
   585
                      :: hopeful) =
blanchet@38739
   586
            let
blanchet@38739
   587
              val weight =
blanchet@38739
   588
                case cached_weight of
blanchet@38739
   589
                  SOME w => w
blanchet@40204
   590
                | NONE => fact_weight fudge loc const_tab rel_const_tab
blanchet@40204
   591
                                      fact_consts
blanchet@38739
   592
            in
blanchet@38741
   593
              if weight >= threshold then
blanchet@38889
   594
                relevant ((ax, weight) :: candidates) rejects hopeful
blanchet@38739
   595
              else
blanchet@38889
   596
                relevant candidates ((ax, weight) :: rejects) hopeful
blanchet@38739
   597
            end
blanchet@38739
   598
        in
blanchet@38744
   599
          trace_msg (fn () =>
blanchet@38744
   600
              "ITERATION " ^ string_of_int j ^ ": current threshold: " ^
blanchet@38744
   601
              Real.toString threshold ^ ", constants: " ^
blanchet@38744
   602
              commas (rel_const_tab |> Symtab.dest
blanchet@38744
   603
                      |> filter (curry (op <>) [] o snd)
blanchet@38827
   604
                      |> map string_for_hyper_pconst));
blanchet@38889
   605
          relevant [] [] hopeful
blanchet@38739
   606
        end
blanchet@41158
   607
    fun add_facts ths accepts =
blanchet@41167
   608
      (facts |> filter (member Thm.eq_thm ths o snd)) @
blanchet@41167
   609
      (accepts |> filter_out (member Thm.eq_thm ths o snd))
blanchet@40408
   610
      |> take max_relevant
blanchet@38739
   611
  in
blanchet@40369
   612
    facts |> map_filter (pair_consts_fact thy is_built_in_const fudge)
blanchet@40204
   613
          |> iter 0 max_relevant threshold0 goal_const_tab []
blanchet@41158
   614
          |> not (null add_ths) ? add_facts add_ths
blanchet@41158
   615
          |> (fn accepts =>
blanchet@41158
   616
                 accepts |> needs_ext is_built_in_const accepts
blanchet@41158
   617
                            ? add_facts @{thms ext})
blanchet@41158
   618
          |> tap (fn accepts => trace_msg (fn () =>
blanchet@41158
   619
                      "Total relevant: " ^ Int.toString (length accepts)))
blanchet@38739
   620
  end
paulson@24287
   621
blanchet@38744
   622
paulson@24287
   623
(***************************************************************)
mengj@19768
   624
(* Retrieving and filtering lemmas                             *)
mengj@19768
   625
(***************************************************************)
mengj@19768
   626
paulson@33022
   627
(*** retrieve lemmas and filter them ***)
mengj@19768
   628
paulson@20757
   629
(*Reject theorems with names like "List.filter.filter_list_def" or
paulson@21690
   630
  "Accessible_Part.acc.defs", as these are definitions arising from packages.*)
paulson@20757
   631
fun is_package_def a =
blanchet@40205
   632
  let val names = Long_Name.explode a in
blanchet@40205
   633
    (length names > 2 andalso not (hd names = "local") andalso
blanchet@40205
   634
     String.isSuffix "_def" a) orelse String.isSuffix "_defs" a
blanchet@40205
   635
  end
paulson@20757
   636
blanchet@38937
   637
fun mk_fact_table f xs =
blanchet@38937
   638
  fold (Termtab.update o `(prop_of o f)) xs Termtab.empty
blanchet@38937
   639
fun uniquify xs = Termtab.fold (cons o snd) (mk_fact_table snd xs) []
mengj@19768
   640
blanchet@37626
   641
(* FIXME: put other record thms here, or declare as "no_atp" *)
blanchet@37626
   642
val multi_base_blacklist =
blanchet@41199
   643
  ["defs", "select_defs", "update_defs", "split", "splits", "split_asm",
blanchet@41199
   644
   "cases", "ext_cases", "eq.simps", "eq.refl", "nchotomy", "case_cong",
blanchet@41199
   645
   "weak_case_cong"]
blanchet@38682
   646
  |> map (prefix ".")
blanchet@37626
   647
blanchet@37626
   648
val max_lambda_nesting = 3
blanchet@37626
   649
blanchet@37626
   650
fun term_has_too_many_lambdas max (t1 $ t2) =
blanchet@37626
   651
    exists (term_has_too_many_lambdas max) [t1, t2]
blanchet@37626
   652
  | term_has_too_many_lambdas max (Abs (_, _, t)) =
blanchet@37626
   653
    max = 0 orelse term_has_too_many_lambdas (max - 1) t
blanchet@37626
   654
  | term_has_too_many_lambdas _ _ = false
blanchet@37626
   655
blanchet@37626
   656
(* Don't count nested lambdas at the level of formulas, since they are
blanchet@37626
   657
   quantifiers. *)
blanchet@37626
   658
fun formula_has_too_many_lambdas Ts (Abs (_, T, t)) =
blanchet@37626
   659
    formula_has_too_many_lambdas (T :: Ts) t
blanchet@37626
   660
  | formula_has_too_many_lambdas Ts t =
blanchet@37626
   661
    if is_formula_type (fastype_of1 (Ts, t)) then
blanchet@37626
   662
      exists (formula_has_too_many_lambdas Ts) (#2 (strip_comb t))
blanchet@37626
   663
    else
blanchet@37626
   664
      term_has_too_many_lambdas max_lambda_nesting t
blanchet@37626
   665
blanchet@38692
   666
(* The max apply depth of any "metis" call in "Metis_Examples" (on 2007-10-31)
blanchet@37626
   667
   was 11. *)
blanchet@37626
   668
val max_apply_depth = 15
blanchet@37626
   669
blanchet@37626
   670
fun apply_depth (f $ t) = Int.max (apply_depth f, apply_depth t + 1)
blanchet@37626
   671
  | apply_depth (Abs (_, _, t)) = apply_depth t
blanchet@37626
   672
  | apply_depth _ = 0
blanchet@37626
   673
blanchet@37626
   674
fun is_formula_too_complex t =
blanchet@38085
   675
  apply_depth t > max_apply_depth orelse formula_has_too_many_lambdas [] t
blanchet@37626
   676
blanchet@39946
   677
(* FIXME: Extend to "Meson" and "Metis" *)
blanchet@37543
   678
val exists_sledgehammer_const =
blanchet@37626
   679
  exists_Const (fn (s, _) => String.isPrefix sledgehammer_prefix s)
blanchet@37626
   680
blanchet@38904
   681
(* FIXME: make more reliable *)
blanchet@38904
   682
val exists_low_level_class_const =
blanchet@38904
   683
  exists_Const (fn (s, _) =>
blanchet@38904
   684
     String.isSubstring (Long_Name.separator ^ "class" ^ Long_Name.separator) s)
blanchet@38904
   685
blanchet@38821
   686
fun is_metastrange_theorem th =
blanchet@37626
   687
  case head_of (concl_of th) of
blanchet@37626
   688
      Const (a, _) => (a <> @{const_name Trueprop} andalso
blanchet@37626
   689
                       a <> @{const_name "=="})
blanchet@37626
   690
    | _ => false
blanchet@37626
   691
blanchet@38821
   692
fun is_that_fact th =
blanchet@38821
   693
  String.isSuffix (Long_Name.separator ^ Obtain.thatN) (Thm.get_name_hint th)
blanchet@38821
   694
  andalso exists_subterm (fn Free (s, _) => s = Name.skolem Auto_Bind.thesisN
blanchet@38821
   695
                           | _ => false) (prop_of th)
blanchet@38821
   696
blanchet@37626
   697
val type_has_top_sort =
blanchet@37626
   698
  exists_subtype (fn TFree (_, []) => true | TVar (_, []) => true | _ => false)
blanchet@37626
   699
blanchet@38085
   700
(**** Predicates to detect unwanted facts (prolific or likely to cause
blanchet@37347
   701
      unsoundness) ****)
paulson@21470
   702
blanchet@38289
   703
(* Too general means, positive equality literal with a variable X as one
blanchet@38289
   704
   operand, when X does not occur properly in the other operand. This rules out
blanchet@38289
   705
   clearly inconsistent facts such as X = a | X = b, though it by no means
blanchet@38289
   706
   guarantees soundness. *)
paulson@21470
   707
blanchet@38289
   708
(* Unwanted equalities are those between a (bound or schematic) variable that
blanchet@38289
   709
   does not properly occur in the second operand. *)
blanchet@38607
   710
val is_exhaustive_finite =
blanchet@38607
   711
  let
blanchet@38629
   712
    fun is_bad_equal (Var z) t =
blanchet@38629
   713
        not (exists_subterm (fn Var z' => z = z' | _ => false) t)
blanchet@38629
   714
      | is_bad_equal (Bound j) t = not (loose_bvar1 (t, j))
blanchet@38629
   715
      | is_bad_equal _ _ = false
blanchet@38629
   716
    fun do_equals t1 t2 = is_bad_equal t1 t2 orelse is_bad_equal t2 t1
blanchet@38607
   717
    fun do_formula pos t =
blanchet@38607
   718
      case (pos, t) of
blanchet@38615
   719
        (_, @{const Trueprop} $ t1) => do_formula pos t1
blanchet@38607
   720
      | (true, Const (@{const_name all}, _) $ Abs (_, _, t')) =>
blanchet@38607
   721
        do_formula pos t'
blanchet@38607
   722
      | (true, Const (@{const_name All}, _) $ Abs (_, _, t')) =>
blanchet@38607
   723
        do_formula pos t'
blanchet@38607
   724
      | (false, Const (@{const_name Ex}, _) $ Abs (_, _, t')) =>
blanchet@38607
   725
        do_formula pos t'
blanchet@38607
   726
      | (_, @{const "==>"} $ t1 $ t2) =>
blanchet@38629
   727
        do_formula (not pos) t1 andalso
blanchet@38629
   728
        (t2 = @{prop False} orelse do_formula pos t2)
haftmann@38786
   729
      | (_, @{const HOL.implies} $ t1 $ t2) =>
blanchet@38629
   730
        do_formula (not pos) t1 andalso
blanchet@38629
   731
        (t2 = @{const False} orelse do_formula pos t2)
blanchet@38607
   732
      | (_, @{const Not} $ t1) => do_formula (not pos) t1
haftmann@38795
   733
      | (true, @{const HOL.disj} $ t1 $ t2) => forall (do_formula pos) [t1, t2]
haftmann@38795
   734
      | (false, @{const HOL.conj} $ t1 $ t2) => forall (do_formula pos) [t1, t2]
haftmann@38864
   735
      | (true, Const (@{const_name HOL.eq}, _) $ t1 $ t2) => do_equals t1 t2
blanchet@38607
   736
      | (true, Const (@{const_name "=="}, _) $ t1 $ t2) => do_equals t1 t2
blanchet@38607
   737
      | _ => false
blanchet@38607
   738
  in do_formula true end
blanchet@38607
   739
blanchet@38592
   740
fun has_bound_or_var_of_type tycons =
blanchet@38592
   741
  exists_subterm (fn Var (_, Type (s, _)) => member (op =) tycons s
blanchet@38592
   742
                   | Abs (_, Type (s, _), _) => member (op =) tycons s
blanchet@38592
   743
                   | _ => false)
paulson@21431
   744
blanchet@38085
   745
(* Facts are forbidden to contain variables of these types. The typical reason
blanchet@37347
   746
   is that they lead to unsoundness. Note that "unit" satisfies numerous
blanchet@38085
   747
   equations like "?x = ()". The resulting clauses will have no type constraint,
blanchet@37347
   748
   yielding false proofs. Even "bool" leads to many unsound proofs, though only
blanchet@37347
   749
   for higher-order problems. *)
blanchet@38592
   750
val dangerous_types = [@{type_name unit}, @{type_name bool}, @{type_name prop}];
paulson@22217
   751
blanchet@38085
   752
(* Facts containing variables of type "unit" or "bool" or of the form
blanchet@38290
   753
   "ALL x. x = A | x = B | x = C" are likely to lead to unsound proofs if types
blanchet@38290
   754
   are omitted. *)
blanchet@41138
   755
fun is_dangerous_term no_dangerous_types t =
blanchet@41138
   756
  not no_dangerous_types andalso
blanchet@38679
   757
  let val t = transform_elim_term t in
blanchet@38679
   758
    has_bound_or_var_of_type dangerous_types t orelse
blanchet@38679
   759
    is_exhaustive_finite t
blanchet@38679
   760
  end
paulson@21470
   761
blanchet@41138
   762
fun is_theorem_bad_for_atps no_dangerous_types thm =
blanchet@38627
   763
  let val t = prop_of thm in
blanchet@38627
   764
    is_formula_too_complex t orelse exists_type type_has_top_sort t orelse
blanchet@41138
   765
    is_dangerous_term no_dangerous_types t orelse
blanchet@41138
   766
    exists_sledgehammer_const t orelse exists_low_level_class_const t orelse
blanchet@41138
   767
    is_metastrange_theorem thm orelse is_that_fact thm
blanchet@38627
   768
  end
blanchet@38627
   769
blanchet@38937
   770
fun clasimpset_rules_of ctxt =
blanchet@38937
   771
  let
blanchet@38937
   772
    val {safeIs, safeEs, hazIs, hazEs, ...} = ctxt |> claset_of |> rep_cs
blanchet@38937
   773
    val intros = safeIs @ hazIs
blanchet@38937
   774
    val elims = map Classical.classical_rule (safeEs @ hazEs)
blanchet@38937
   775
    val simps = ctxt |> simpset_of |> dest_ss |> #simps |> map snd
blanchet@38937
   776
  in (mk_fact_table I intros, mk_fact_table I elims, mk_fact_table I simps) end
blanchet@38937
   777
blanchet@41138
   778
fun all_facts ctxt reserved no_dangerous_types
blanchet@40205
   779
              ({intro_bonus, elim_bonus, simp_bonus, ...} : relevance_fudge)
blanchet@40205
   780
              add_ths chained_ths =
blanchet@38627
   781
  let
blanchet@38752
   782
    val thy = ProofContext.theory_of ctxt
wenzelm@39557
   783
    val global_facts = Global_Theory.facts_of thy
blanchet@38644
   784
    val local_facts = ProofContext.facts_of ctxt
blanchet@38644
   785
    val named_locals = local_facts |> Facts.dest_static []
blanchet@38993
   786
    val assms = Assumption.all_assms_of ctxt
blanchet@38993
   787
    fun is_assum th = exists (fn ct => prop_of th aconv term_of ct) assms
blanchet@38752
   788
    val is_chained = member Thm.eq_thm chained_ths
blanchet@38937
   789
    val (intros, elims, simps) =
blanchet@40070
   790
      if exists (curry (op <) 0.0) [intro_bonus, elim_bonus, simp_bonus] then
blanchet@38937
   791
        clasimpset_rules_of ctxt
blanchet@38937
   792
      else
blanchet@38937
   793
        (Termtab.empty, Termtab.empty, Termtab.empty)
blanchet@38738
   794
    fun is_good_unnamed_local th =
blanchet@38820
   795
      not (Thm.has_name_hint th) andalso
blanchet@38738
   796
      forall (fn (_, ths) => not (member Thm.eq_thm ths th)) named_locals
blanchet@38644
   797
    val unnamed_locals =
blanchet@38820
   798
      union Thm.eq_thm (Facts.props local_facts) chained_ths
blanchet@38820
   799
      |> filter is_good_unnamed_local |> map (pair "" o single)
blanchet@38627
   800
    val full_space =
blanchet@38738
   801
      Name_Space.merge (Facts.space_of global_facts, Facts.space_of local_facts)
blanchet@38752
   802
    fun add_facts global foldx facts =
blanchet@38699
   803
      foldx (fn (name0, ths) =>
blanchet@38699
   804
        if name0 <> "" andalso
blanchet@39012
   805
           forall (not o member Thm.eq_thm add_ths) ths andalso
blanchet@38699
   806
           (Facts.is_concealed facts name0 orelse
blanchet@38699
   807
            (respect_no_atp andalso is_package_def name0) orelse
blanchet@38699
   808
            exists (fn s => String.isSuffix s name0) multi_base_blacklist orelse
blanchet@38699
   809
            String.isSuffix "_def_raw" (* FIXME: crude hack *) name0) then
blanchet@38627
   810
          I
blanchet@38627
   811
        else
blanchet@38627
   812
          let
blanchet@38699
   813
            val multi = length ths > 1
blanchet@41199
   814
            val backquotify =
blanchet@41199
   815
              enclose "`" "`" o string_for_term ctxt o close_form o prop_of
blanchet@38699
   816
            fun check_thms a =
blanchet@38699
   817
              case try (ProofContext.get_thms ctxt) a of
blanchet@38699
   818
                NONE => false
blanchet@38699
   819
              | SOME ths' => Thm.eq_thms (ths, ths')
blanchet@38627
   820
          in
blanchet@38699
   821
            pair 1
blanchet@38699
   822
            #> fold (fn th => fn (j, rest) =>
blanchet@38699
   823
                 (j + 1,
blanchet@41138
   824
                  if is_theorem_bad_for_atps no_dangerous_types th andalso
blanchet@39012
   825
                     not (member Thm.eq_thm add_ths th) then
blanchet@38699
   826
                    rest
blanchet@38699
   827
                  else
blanchet@38752
   828
                    (((fn () =>
blanchet@38752
   829
                          if name0 = "" then
blanchet@38752
   830
                            th |> backquotify
blanchet@38752
   831
                          else
blanchet@38752
   832
                            let
blanchet@38752
   833
                              val name1 = Facts.extern facts name0
blanchet@38752
   834
                              val name2 = Name_Space.extern full_space name0
blanchet@38752
   835
                            in
blanchet@38752
   836
                              case find_first check_thms [name1, name2, name0] of
blanchet@40375
   837
                                SOME name => make_name reserved multi j name
blanchet@38752
   838
                              | NONE => ""
blanchet@38937
   839
                            end),
blanchet@38937
   840
                      let val t = prop_of th in
blanchet@40071
   841
                        if is_chained th then
blanchet@40071
   842
                          Chained
blanchet@38993
   843
                        else if global then
blanchet@38993
   844
                          if Termtab.defined intros t then Intro
blanchet@38993
   845
                          else if Termtab.defined elims t then Elim
blanchet@38993
   846
                          else if Termtab.defined simps t then Simp
blanchet@38993
   847
                          else General
blanchet@38993
   848
                        else
blanchet@38993
   849
                          if is_assum th then Assum else Local
blanchet@38937
   850
                      end),
blanchet@38752
   851
                      (multi, th)) :: rest)) ths
blanchet@38699
   852
            #> snd
blanchet@38627
   853
          end)
blanchet@38644
   854
  in
blanchet@38752
   855
    [] |> add_facts false fold local_facts (unnamed_locals @ named_locals)
blanchet@38752
   856
       |> add_facts true Facts.fold_static global_facts global_facts
blanchet@38644
   857
  end
blanchet@38627
   858
blanchet@38627
   859
(* The single-name theorems go after the multiple-name ones, so that single
blanchet@38627
   860
   names are preferred when both are available. *)
blanchet@40205
   861
fun rearrange_facts ctxt respect_no_atp =
blanchet@38744
   862
  List.partition (fst o snd) #> op @ #> map (apsnd snd)
blanchet@38699
   863
  #> respect_no_atp ? filter_out (No_ATPs.member ctxt o snd)
blanchet@38627
   864
blanchet@38627
   865
(***************************************************************)
blanchet@38627
   866
(* ATP invocation methods setup                                *)
blanchet@38627
   867
(***************************************************************)
blanchet@38627
   868
blanchet@41199
   869
fun external_frees t =
blanchet@41199
   870
  [] |> Term.add_frees t |> filter_out (can Name.dest_internal o fst)
blanchet@41199
   871
blanchet@41138
   872
fun relevant_facts ctxt no_dangerous_types (threshold0, threshold1)
blanchet@41138
   873
                   max_relevant is_built_in_const fudge
blanchet@41138
   874
                   (override as {add, only, ...}) chained_ths hyp_ts concl_t =
blanchet@37538
   875
  let
blanchet@41200
   876
    val thy = ProofContext.theory_of ctxt
blanchet@38822
   877
    val decay = Math.pow ((1.0 - threshold1) / (1.0 - threshold0),
blanchet@38822
   878
                          1.0 / Real.fromInt (max_relevant + 1))
blanchet@39012
   879
    val add_ths = Attrib.eval_thms ctxt add
blanchet@38696
   880
    val reserved = reserved_isar_keyword_table ()
blanchet@41199
   881
    val ind_stmt =
blanchet@41199
   882
      Logic.list_implies (hyp_ts |> filter_out (null o external_frees), concl_t)
blanchet@41199
   883
      |> atomize_term
blanchet@41199
   884
    val ind_stmt_xs = external_frees ind_stmt
blanchet@40204
   885
    val facts =
blanchet@38699
   886
      (if only then
blanchet@38752
   887
         maps (map (fn ((name, loc), th) => ((K name, loc), (true, th)))
blanchet@40205
   888
               o fact_from_ref ctxt reserved chained_ths) add
blanchet@38699
   889
       else
blanchet@41138
   890
         all_facts ctxt reserved no_dangerous_types fudge add_ths chained_ths)
blanchet@41199
   891
      |> maps (instantiate_if_induct_rule ctxt ind_stmt ind_stmt_xs)
blanchet@40205
   892
      |> rearrange_facts ctxt (respect_no_atp andalso not only)
blanchet@38937
   893
      |> uniquify
blanchet@37538
   894
  in
blanchet@40204
   895
    trace_msg (fn () => "Considering " ^ Int.toString (length facts) ^
blanchet@40204
   896
                        " facts");
blanchet@39366
   897
    (if only orelse threshold1 < 0.0 then
blanchet@40204
   898
       facts
blanchet@39366
   899
     else if threshold0 > 1.0 orelse threshold0 > threshold1 orelse
blanchet@39366
   900
             max_relevant = 0 then
blanchet@38739
   901
       []
blanchet@38739
   902
     else
blanchet@41200
   903
       ((concl_t |> theory_constify fudge (Context.theory_name thy)) :: hyp_ts)
blanchet@41200
   904
       |> relevance_filter ctxt threshold0 decay max_relevant is_built_in_const
blanchet@41200
   905
                           fudge override facts)
blanchet@38822
   906
    |> map (apfst (apfst (fn f => f ())))
blanchet@37538
   907
  end
immler@30536
   908
paulson@15347
   909
end;