src/HOL/Tools/Sledgehammer/sledgehammer_filter.ML
author blanchet
Sun May 01 18:37:24 2011 +0200 (2011-05-01)
changeset 42546 8591fcc56c34
parent 42449 494e4ac5b0f8
child 42579 2552c09b1a72
permissions -rw-r--r--
make sure typing fact names are unique (needed e.g. by SNARK)
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@41790
    14
     local_const_multiplier : real,
blanchet@41159
    15
     worse_irrel_freq : real,
blanchet@40070
    16
     higher_order_irrel_weight : real,
blanchet@40070
    17
     abs_rel_weight : real,
blanchet@40070
    18
     abs_irrel_weight : real,
blanchet@40070
    19
     skolem_irrel_weight : real,
blanchet@40070
    20
     theory_const_rel_weight : real,
blanchet@40070
    21
     theory_const_irrel_weight : real,
blanchet@40070
    22
     intro_bonus : real,
blanchet@40070
    23
     elim_bonus : real,
blanchet@40070
    24
     simp_bonus : real,
blanchet@40070
    25
     local_bonus : real,
blanchet@40070
    26
     assum_bonus : real,
blanchet@40070
    27
     chained_bonus : real,
blanchet@40070
    28
     max_imperfect : real,
blanchet@40070
    29
     max_imperfect_exp : real,
blanchet@40070
    30
     threshold_divisor : real,
blanchet@40070
    31
     ridiculous_threshold : real}
blanchet@40070
    32
blanchet@35966
    33
  type relevance_override =
blanchet@40070
    34
    {add : (Facts.ref * Attrib.src list) list,
blanchet@40070
    35
     del : (Facts.ref * Attrib.src list) list,
blanchet@40070
    36
     only : bool}
blanchet@35966
    37
blanchet@37616
    38
  val trace : bool Unsynchronized.ref
blanchet@42449
    39
  val is_global_locality : locality -> bool
blanchet@40205
    40
  val fact_from_ref :
blanchet@38996
    41
    Proof.context -> unit Symtab.table -> thm list
blanchet@38996
    42
    -> Facts.ref * Attrib.src list -> ((string * locality) * thm) list
blanchet@41767
    43
  val all_facts :
blanchet@41989
    44
    Proof.context -> 'a Symtab.table -> bool -> bool -> relevance_fudge
blanchet@41989
    45
    -> thm list -> thm list -> (((unit -> string) * locality) * (bool * thm)) list
blanchet@41768
    46
  val const_names_in_fact :
blanchet@41768
    47
    theory -> (string * typ -> term list -> bool * term list) -> term
blanchet@41768
    48
    -> string list
blanchet@37347
    49
  val relevant_facts :
blanchet@41066
    50
    Proof.context -> bool -> real * real -> int
blanchet@41336
    51
    -> (string * typ -> term list -> bool * term list) -> relevance_fudge
blanchet@41066
    52
    -> relevance_override -> thm list -> term list -> term
blanchet@40070
    53
    -> ((string * locality) * thm) list
paulson@15347
    54
end;
paulson@15347
    55
blanchet@38988
    56
structure Sledgehammer_Filter : SLEDGEHAMMER_FILTER =
paulson@15347
    57
struct
paulson@15347
    58
blanchet@38652
    59
open Sledgehammer_Util
blanchet@38652
    60
blanchet@37616
    61
val trace = Unsynchronized.ref false
blanchet@37616
    62
fun trace_msg msg = if !trace then tracing (msg ()) else ()
blanchet@35826
    63
blanchet@41273
    64
(* experimental features *)
blanchet@38997
    65
val respect_no_atp = true
blanchet@41273
    66
val instantiate_inducts = false
blanchet@38827
    67
blanchet@38993
    68
datatype locality = General | Intro | Elim | Simp | Local | Assum | Chained
blanchet@38752
    69
blanchet@42449
    70
(* (quasi-)underapproximation of the truth *)
blanchet@42449
    71
fun is_global_locality Local = false
blanchet@42449
    72
  | is_global_locality Assum = false
blanchet@42449
    73
  | is_global_locality Chained = false
blanchet@42449
    74
  | is_global_locality _ = true
blanchet@42449
    75
blanchet@40070
    76
type relevance_fudge =
blanchet@41159
    77
  {allow_ext : bool,
blanchet@41790
    78
   local_const_multiplier : real,
blanchet@41159
    79
   worse_irrel_freq : real,
blanchet@40070
    80
   higher_order_irrel_weight : real,
blanchet@40070
    81
   abs_rel_weight : real,
blanchet@40070
    82
   abs_irrel_weight : real,
blanchet@40070
    83
   skolem_irrel_weight : real,
blanchet@40070
    84
   theory_const_rel_weight : real,
blanchet@40070
    85
   theory_const_irrel_weight : real,
blanchet@40070
    86
   intro_bonus : real,
blanchet@40070
    87
   elim_bonus : real,
blanchet@40070
    88
   simp_bonus : real,
blanchet@40070
    89
   local_bonus : real,
blanchet@40070
    90
   assum_bonus : real,
blanchet@40070
    91
   chained_bonus : real,
blanchet@40070
    92
   max_imperfect : real,
blanchet@40070
    93
   max_imperfect_exp : real,
blanchet@40070
    94
   threshold_divisor : real,
blanchet@40070
    95
   ridiculous_threshold : real}
blanchet@40070
    96
blanchet@35966
    97
type relevance_override =
blanchet@40070
    98
  {add : (Facts.ref * Attrib.src list) list,
blanchet@40070
    99
   del : (Facts.ref * Attrib.src list) list,
blanchet@40070
   100
   only : bool}
paulson@21070
   101
blanchet@37616
   102
val sledgehammer_prefix = "Sledgehammer" ^ Long_Name.separator
blanchet@39896
   103
val abs_name = sledgehammer_prefix ^ "abs"
blanchet@39896
   104
val skolem_prefix = sledgehammer_prefix ^ "sko"
blanchet@38992
   105
val theory_const_suffix = Long_Name.separator ^ " 1"
blanchet@37616
   106
blanchet@40227
   107
fun needs_quoting reserved s =
blanchet@40375
   108
  Symtab.defined reserved s orelse
wenzelm@42290
   109
  exists (not o Lexicon.is_identifier) (Long_Name.explode s)
blanchet@40227
   110
blanchet@40375
   111
fun make_name reserved multi j name =
blanchet@40227
   112
  (name |> needs_quoting reserved name ? quote) ^
wenzelm@41491
   113
  (if multi then "(" ^ string_of_int j ^ ")" else "")
blanchet@38744
   114
blanchet@40375
   115
fun explode_interval _ (Facts.FromTo (i, j)) = i upto j
blanchet@40375
   116
  | explode_interval max (Facts.From i) = i upto i + max - 1
blanchet@40375
   117
  | explode_interval _ (Facts.Single i) = [i]
blanchet@40375
   118
blanchet@41279
   119
val backquote =
blanchet@41279
   120
  raw_explode #> map (fn "`" => "\\`" | s => s) #> implode #> enclose "`" "`"
blanchet@40205
   121
fun fact_from_ref ctxt reserved chained_ths (xthm as (xref, args)) =
blanchet@38744
   122
  let
blanchet@38996
   123
    val ths = Attrib.eval_thms ctxt [xthm]
blanchet@38996
   124
    val bracket =
blanchet@41999
   125
      map (enclose "[" "]" o Pretty.str_of o Args.pretty_src ctxt) args
blanchet@41999
   126
      |> implode
blanchet@40375
   127
    fun nth_name j =
blanchet@38996
   128
      case xref of
blanchet@41279
   129
        Facts.Fact s => backquote s ^ bracket
blanchet@38996
   130
      | Facts.Named (("", _), _) => "[" ^ bracket ^ "]"
blanchet@40375
   131
      | Facts.Named ((name, _), NONE) =>
blanchet@40375
   132
        make_name reserved (length ths > 1) (j + 1) name ^ bracket
blanchet@40375
   133
      | Facts.Named ((name, _), SOME intervals) =>
blanchet@40375
   134
        make_name reserved true
blanchet@40375
   135
                 (nth (maps (explode_interval (length ths)) intervals) j) name ^
blanchet@40375
   136
        bracket
blanchet@38744
   137
  in
blanchet@40375
   138
    (ths, (0, []))
blanchet@38752
   139
    |-> fold (fn th => fn (j, rest) =>
blanchet@40375
   140
                 (j + 1, ((nth_name j,
blanchet@38752
   141
                          if member Thm.eq_thm chained_ths th then Chained
blanchet@38752
   142
                          else General), th) :: rest))
blanchet@38744
   143
    |> snd
blanchet@38699
   144
  end
blanchet@37616
   145
blanchet@41199
   146
(* This is a terrible hack. Free variables are sometimes code as "M__" when they
blanchet@41199
   147
   are displayed as "M" and we want to avoid clashes with these. But sometimes
blanchet@41199
   148
   it's even worse: "Ma__" encodes "M". So we simply reserve all prefixes of all
blanchet@41199
   149
   free variables. In the worse case scenario, where the fact won't be resolved
blanchet@41199
   150
   correctly, the user can fix it manually, e.g., by naming the fact in
blanchet@41199
   151
   question. Ideally we would need nothing of it, but backticks just don't work
blanchet@41199
   152
   with schematic variables. *)
blanchet@41199
   153
fun all_prefixes_of s =
blanchet@41199
   154
  map (fn i => String.extract (s, 0, SOME i)) (1 upto size s - 1)
blanchet@41199
   155
fun close_form t =
blanchet@41199
   156
  (t, [] |> Term.add_free_names t |> maps all_prefixes_of)
blanchet@41199
   157
  |> fold (fn ((s, i), T) => fn (t', taken) =>
blanchet@41199
   158
              let val s' = Name.variant taken s in
blanchet@41199
   159
                ((if fastype_of t' = HOLogic.boolT then HOLogic.all_const
blanchet@41199
   160
                  else Term.all) T
blanchet@41199
   161
                 $ Abs (s', T, abstract_over (Var ((s, i), T), t')),
blanchet@41199
   162
                 s' :: taken)
blanchet@41199
   163
              end)
blanchet@41199
   164
          (Term.add_vars t [] |> sort_wrt (fst o fst))
blanchet@41199
   165
  |> fst
blanchet@41199
   166
blanchet@41199
   167
fun string_for_term ctxt t =
blanchet@41199
   168
  Print_Mode.setmp (filter (curry (op =) Symbol.xsymbolsN)
blanchet@41199
   169
                   (print_mode_value ())) (Syntax.string_of_term ctxt) t
blanchet@41199
   170
  |> String.translate (fn c => if Char.isPrint c then str c else "")
blanchet@41199
   171
  |> simplify_spaces
blanchet@41199
   172
blanchet@41199
   173
(** Structural induction rules **)
blanchet@41199
   174
blanchet@41200
   175
fun struct_induct_rule_on th =
blanchet@41199
   176
  case Logic.strip_horn (prop_of th) of
blanchet@41199
   177
    (prems, @{const Trueprop}
blanchet@41199
   178
            $ ((p as Var ((p_name, 0), _)) $ (a as Var (_, ind_T)))) =>
blanchet@41199
   179
    if not (is_TVar ind_T) andalso length prems > 1 andalso
blanchet@41199
   180
       exists (exists_subterm (curry (op aconv) p)) prems andalso
blanchet@41199
   181
       not (exists (exists_subterm (curry (op aconv) a)) prems) then
blanchet@41199
   182
      SOME (p_name, ind_T)
blanchet@41199
   183
    else
blanchet@41199
   184
      NONE
blanchet@41199
   185
  | _ => NONE
blanchet@41199
   186
blanchet@41207
   187
fun instantiate_induct_rule ctxt concl_prop p_name ((name, loc), (multi, th))
blanchet@41199
   188
                            ind_x =
blanchet@41199
   189
  let
blanchet@41199
   190
    fun varify_noninducts (t as Free (s, T)) =
blanchet@41199
   191
        if (s, T) = ind_x orelse can dest_funT T then t else Var ((s, 0), T)
blanchet@41199
   192
      | varify_noninducts t = t
blanchet@41199
   193
    val p_inst =
blanchet@41199
   194
      concl_prop |> map_aterms varify_noninducts |> close_form
blanchet@41199
   195
                 |> lambda (Free ind_x)
blanchet@41199
   196
                 |> string_for_term ctxt
blanchet@41199
   197
  in
blanchet@41207
   198
    ((fn () => name () ^ "[where " ^ p_name ^ " = " ^ quote p_inst ^ "]", loc),
blanchet@41199
   199
     (multi, th |> read_instantiate ctxt [((p_name, 0), p_inst)]))
blanchet@41199
   200
  end
blanchet@41199
   201
blanchet@41199
   202
fun type_match thy (T1, T2) =
blanchet@41199
   203
  (Sign.typ_match thy (T2, T1) Vartab.empty; true)
blanchet@41199
   204
  handle Type.TYPE_MATCH => false
blanchet@41199
   205
blanchet@41199
   206
fun instantiate_if_induct_rule ctxt stmt stmt_xs (ax as (_, (_, th))) =
blanchet@41200
   207
  case struct_induct_rule_on th of
blanchet@41199
   208
    SOME (p_name, ind_T) =>
wenzelm@42361
   209
    let val thy = Proof_Context.theory_of ctxt in
blanchet@41199
   210
      stmt_xs |> filter (fn (_, T) => type_match thy (T, ind_T))
blanchet@41199
   211
              |> map_filter (try (instantiate_induct_rule ctxt stmt p_name ax))
blanchet@41199
   212
    end
blanchet@41199
   213
  | NONE => [ax]
blanchet@41199
   214
wenzelm@28477
   215
(***************************************************************)
wenzelm@28477
   216
(* Relevance Filtering                                         *)
wenzelm@28477
   217
(***************************************************************)
mengj@19194
   218
paulson@24287
   219
(*** constants with types ***)
paulson@24287
   220
blanchet@38939
   221
fun order_of_type (Type (@{type_name fun}, [T1, @{typ bool}])) =
blanchet@38939
   222
    order_of_type T1 (* cheat: pretend sets are first-order *)
blanchet@38939
   223
  | order_of_type (Type (@{type_name fun}, [T1, T2])) =
blanchet@38939
   224
    Int.max (order_of_type T1 + 1, order_of_type T2)
blanchet@38939
   225
  | order_of_type (Type (_, Ts)) = fold (Integer.max o order_of_type) Ts 0
blanchet@38939
   226
  | order_of_type _ = 0
blanchet@38939
   227
blanchet@38823
   228
(* An abstraction of Isabelle types and first-order terms *)
blanchet@38823
   229
datatype pattern = PVar | PApp of string * pattern list
blanchet@38939
   230
datatype ptype = PType of int * pattern list
blanchet@38744
   231
blanchet@38823
   232
fun string_for_pattern PVar = "_"
blanchet@38823
   233
  | string_for_pattern (PApp (s, ps)) =
blanchet@38823
   234
    if null ps then s else s ^ string_for_patterns ps
blanchet@38823
   235
and string_for_patterns ps = "(" ^ commas (map string_for_pattern ps) ^ ")"
blanchet@38939
   236
fun string_for_ptype (PType (_, ps)) = string_for_patterns ps
paulson@24287
   237
paulson@24287
   238
(*Is the second type an instance of the first one?*)
blanchet@38824
   239
fun match_pattern (PVar, _) = true
blanchet@38824
   240
  | match_pattern (PApp _, PVar) = false
blanchet@38824
   241
  | match_pattern (PApp (s, ps), PApp (t, qs)) =
blanchet@38824
   242
    s = t andalso match_patterns (ps, qs)
blanchet@38824
   243
and match_patterns (_, []) = true
blanchet@38824
   244
  | match_patterns ([], _) = false
blanchet@38824
   245
  | match_patterns (p :: ps, q :: qs) =
blanchet@38824
   246
    match_pattern (p, q) andalso match_patterns (ps, qs)
blanchet@38939
   247
fun match_ptype (PType (_, ps), PType (_, qs)) = match_patterns (ps, qs)
paulson@24287
   248
blanchet@38823
   249
(* Is there a unifiable constant? *)
blanchet@38827
   250
fun pconst_mem f consts (s, ps) =
blanchet@38939
   251
  exists (curry (match_ptype o f) ps)
blanchet@38827
   252
         (map snd (filter (curry (op =) s o fst) consts))
blanchet@38827
   253
fun pconst_hyper_mem f const_tab (s, ps) =
blanchet@38939
   254
  exists (curry (match_ptype o f) ps) (these (Symtab.lookup const_tab s))
blanchet@37505
   255
blanchet@38939
   256
fun pattern_for_type (Type (s, Ts)) = PApp (s, map pattern_for_type Ts)
blanchet@38939
   257
  | pattern_for_type (TFree (s, _)) = PApp (s, [])
blanchet@38939
   258
  | pattern_for_type (TVar _) = PVar
blanchet@38827
   259
blanchet@38744
   260
(* Pairs a constant with the list of its type instantiations. *)
blanchet@41204
   261
fun ptype thy const x =
blanchet@38939
   262
  (if const then map pattern_for_type (these (try (Sign.const_typargs thy) x))
blanchet@41204
   263
   else [])
blanchet@41204
   264
fun rich_ptype thy const (s, T) =
blanchet@41204
   265
  PType (order_of_type T, ptype thy const (s, T))
blanchet@41204
   266
fun rich_pconst thy const (s, T) = (s, rich_ptype thy const (s, T))
paulson@24287
   267
blanchet@38939
   268
fun string_for_hyper_pconst (s, ps) =
blanchet@38939
   269
  s ^ "{" ^ commas (map string_for_ptype ps) ^ "}"
paulson@24287
   270
blanchet@38823
   271
(* Add a pconstant to the table, but a [] entry means a standard
blanchet@38819
   272
   connective, which we ignore.*)
blanchet@41066
   273
fun add_pconst_to_table also_skolem (s, p) =
blanchet@41066
   274
  if (not also_skolem andalso String.isPrefix skolem_prefix s) then I
blanchet@41066
   275
  else Symtab.map_default (s, [p]) (insert (op =) p)
blanchet@38819
   276
blanchet@40369
   277
fun pconsts_in_terms thy is_built_in_const also_skolems pos ts =
blanchet@37505
   278
  let
blanchet@38819
   279
    val flip = Option.map not
blanchet@38587
   280
    (* We include free variables, as well as constants, to handle locales. For
blanchet@41205
   281
       each quantifiers that must necessarily be skolemized by the automatic
blanchet@41205
   282
       prover, we introduce a fresh constant to simulate the effect of
blanchet@41205
   283
       Skolemization. *)
blanchet@40373
   284
    fun do_const const x ts =
blanchet@41336
   285
      let val (built_in, ts) = is_built_in_const x ts in
blanchet@41336
   286
        (not built_in
blanchet@41336
   287
         ? add_pconst_to_table also_skolems (rich_pconst thy const x))
blanchet@41336
   288
        #> fold do_term ts
blanchet@41336
   289
      end
blanchet@38827
   290
    and do_term t =
blanchet@38827
   291
      case strip_comb t of
blanchet@38827
   292
        (Const x, ts) => do_const true x ts
blanchet@38827
   293
      | (Free x, ts) => do_const false x ts
blanchet@38939
   294
      | (Abs (_, T, t'), ts) =>
blanchet@38939
   295
        (null ts
blanchet@41066
   296
         ? add_pconst_to_table true (abs_name, PType (order_of_type T + 1, [])))
blanchet@38827
   297
        #> fold do_term (t' :: ts)
blanchet@38827
   298
      | (_, ts) => fold do_term ts
blanchet@38939
   299
    fun do_quantifier will_surely_be_skolemized abs_T body_t =
blanchet@37537
   300
      do_formula pos body_t
blanchet@38747
   301
      #> (if also_skolems andalso will_surely_be_skolemized then
blanchet@41066
   302
            add_pconst_to_table true
blanchet@41207
   303
                (gensym skolem_prefix, PType (order_of_type abs_T, []))
blanchet@38587
   304
          else
blanchet@38587
   305
            I)
blanchet@38587
   306
    and do_term_or_formula T =
blanchet@41273
   307
      if T = HOLogic.boolT then do_formula NONE else do_term
blanchet@37537
   308
    and do_formula pos t =
blanchet@37537
   309
      case t of
blanchet@38939
   310
        Const (@{const_name all}, _) $ Abs (_, T, t') =>
blanchet@38939
   311
        do_quantifier (pos = SOME false) T t'
blanchet@37537
   312
      | @{const "==>"} $ t1 $ t2 =>
blanchet@37537
   313
        do_formula (flip pos) t1 #> do_formula pos t2
blanchet@37537
   314
      | Const (@{const_name "=="}, Type (_, [T, _])) $ t1 $ t2 =>
blanchet@38587
   315
        fold (do_term_or_formula T) [t1, t2]
blanchet@37537
   316
      | @{const Trueprop} $ t1 => do_formula pos t1
blanchet@41140
   317
      | @{const False} => I
blanchet@41140
   318
      | @{const True} => I
blanchet@37537
   319
      | @{const Not} $ t1 => do_formula (flip pos) t1
blanchet@38939
   320
      | Const (@{const_name All}, _) $ Abs (_, T, t') =>
blanchet@38939
   321
        do_quantifier (pos = SOME false) T t'
blanchet@38939
   322
      | Const (@{const_name Ex}, _) $ Abs (_, T, t') =>
blanchet@38939
   323
        do_quantifier (pos = SOME true) T t'
haftmann@38795
   324
      | @{const HOL.conj} $ t1 $ t2 => fold (do_formula pos) [t1, t2]
haftmann@38795
   325
      | @{const HOL.disj} $ t1 $ t2 => fold (do_formula pos) [t1, t2]
haftmann@38786
   326
      | @{const HOL.implies} $ t1 $ t2 =>
blanchet@37537
   327
        do_formula (flip pos) t1 #> do_formula pos t2
haftmann@38864
   328
      | Const (@{const_name HOL.eq}, Type (_, [T, _])) $ t1 $ t2 =>
blanchet@38587
   329
        fold (do_term_or_formula T) [t1, t2]
blanchet@38587
   330
      | Const (@{const_name If}, Type (_, [_, Type (_, [T, _])]))
blanchet@38587
   331
        $ t1 $ t2 $ t3 =>
blanchet@38587
   332
        do_formula NONE t1 #> fold (do_term_or_formula T) [t2, t3]
blanchet@38939
   333
      | Const (@{const_name Ex1}, _) $ Abs (_, T, t') =>
blanchet@38939
   334
        do_quantifier (is_some pos) T t'
blanchet@38939
   335
      | Const (@{const_name Ball}, _) $ t1 $ Abs (_, T, t') =>
blanchet@38939
   336
        do_quantifier (pos = SOME false) T
blanchet@38939
   337
                      (HOLogic.mk_imp (incr_boundvars 1 t1 $ Bound 0, t'))
blanchet@38939
   338
      | Const (@{const_name Bex}, _) $ t1 $ Abs (_, T, t') =>
blanchet@38939
   339
        do_quantifier (pos = SOME true) T
blanchet@38939
   340
                      (HOLogic.mk_conj (incr_boundvars 1 t1 $ Bound 0, t'))
blanchet@37537
   341
      | (t0 as Const (_, @{typ bool})) $ t1 =>
blanchet@37537
   342
        do_term t0 #> do_formula pos t1  (* theory constant *)
blanchet@37537
   343
      | _ => do_term t
blanchet@38819
   344
  in Symtab.empty |> fold (do_formula pos) ts end
paulson@24287
   345
paulson@24287
   346
(*Inserts a dummy "constant" referring to the theory name, so that relevance
paulson@24287
   347
  takes the given theory into account.*)
blanchet@41200
   348
fun theory_constify ({theory_const_rel_weight, theory_const_irrel_weight, ...}
blanchet@41200
   349
                     : relevance_fudge) thy_name t =
blanchet@40070
   350
  if exists (curry (op <) 0.0) [theory_const_rel_weight,
blanchet@40070
   351
                                theory_const_irrel_weight] then
blanchet@41200
   352
    Const (thy_name ^ theory_const_suffix, @{typ bool}) $ t
blanchet@37505
   353
  else
blanchet@41200
   354
    t
blanchet@41200
   355
blanchet@41200
   356
fun theory_const_prop_of fudge th =
blanchet@41200
   357
  theory_constify fudge (Context.theory_name (theory_of_thm th)) (prop_of th)
blanchet@37505
   358
paulson@24287
   359
(**** Constant / Type Frequencies ****)
paulson@24287
   360
blanchet@38743
   361
(* A two-dimensional symbol table counts frequencies of constants. It's keyed
blanchet@38743
   362
   first by constant name and second by its list of type instantiations. For the
blanchet@38823
   363
   latter, we need a linear ordering on "pattern list". *)
paulson@24287
   364
blanchet@38823
   365
fun pattern_ord p =
blanchet@38743
   366
  case p of
blanchet@38744
   367
    (PVar, PVar) => EQUAL
blanchet@38823
   368
  | (PVar, PApp _) => LESS
blanchet@38823
   369
  | (PApp _, PVar) => GREATER
blanchet@38823
   370
  | (PApp q1, PApp q2) =>
blanchet@38823
   371
    prod_ord fast_string_ord (dict_ord pattern_ord) (q1, q2)
blanchet@38939
   372
fun ptype_ord (PType p, PType q) =
blanchet@38939
   373
  prod_ord (dict_ord pattern_ord) int_ord (swap p, swap q)
paulson@24287
   374
blanchet@38939
   375
structure PType_Tab = Table(type key = ptype val ord = ptype_ord)
paulson@24287
   376
blanchet@40204
   377
fun count_fact_consts thy fudge =
blanchet@37503
   378
  let
blanchet@38827
   379
    fun do_const const (s, T) ts =
blanchet@38827
   380
      (* Two-dimensional table update. Constant maps to types maps to count. *)
blanchet@41204
   381
      PType_Tab.map_default (rich_ptype thy const (s, T), 0) (Integer.add 1)
blanchet@38939
   382
      |> Symtab.map_default (s, PType_Tab.empty)
blanchet@38827
   383
      #> fold do_term ts
blanchet@38827
   384
    and do_term t =
blanchet@38827
   385
      case strip_comb t of
blanchet@38827
   386
        (Const x, ts) => do_const true x ts
blanchet@38827
   387
      | (Free x, ts) => do_const false x ts
blanchet@38827
   388
      | (Abs (_, _, t'), ts) => fold do_term (t' :: ts)
blanchet@38827
   389
      | (_, ts) => fold do_term ts
blanchet@40070
   390
  in do_term o theory_const_prop_of fudge o snd end
paulson@24287
   391
paulson@24287
   392
paulson@24287
   393
(**** Actual Filtering Code ****)
paulson@24287
   394
blanchet@39367
   395
fun pow_int _ 0 = 1.0
blanchet@38939
   396
  | pow_int x 1 = x
blanchet@38939
   397
  | pow_int x n = if n > 0 then x * pow_int x (n - 1) else pow_int x (n + 1) / x
blanchet@38939
   398
paulson@24287
   399
(*The frequency of a constant is the sum of those of all instances of its type.*)
blanchet@38824
   400
fun pconst_freq match const_tab (c, ps) =
blanchet@38939
   401
  PType_Tab.fold (fn (qs, m) => match (ps, qs) ? Integer.add m)
blanchet@38939
   402
                 (the (Symtab.lookup const_tab c)) 0
blanchet@38686
   403
paulson@24287
   404
blanchet@38085
   405
(* A surprising number of theorems contain only a few significant constants.
blanchet@38085
   406
   These include all induction rules, and other general theorems. *)
blanchet@37503
   407
blanchet@37503
   408
(* "log" seems best in practice. A constant function of one ignores the constant
blanchet@38938
   409
   frequencies. Rare constants give more points if they are relevant than less
blanchet@38938
   410
   rare ones. *)
blanchet@39367
   411
fun rel_weight_for _ freq = 1.0 + 2.0 / Math.ln (Real.fromInt freq + 1.0)
blanchet@38938
   412
blanchet@38938
   413
(* Irrelevant constants are treated differently. We associate lower penalties to
blanchet@38938
   414
   very rare constants and very common ones -- the former because they can't
blanchet@38938
   415
   lead to the inclusion of too many new facts, and the latter because they are
blanchet@38938
   416
   so common as to be of little interest. *)
blanchet@40070
   417
fun irrel_weight_for ({worse_irrel_freq, higher_order_irrel_weight, ...}
blanchet@40070
   418
                      : relevance_fudge) order freq =
blanchet@40070
   419
  let val (k, x) = worse_irrel_freq |> `Real.ceil in
blanchet@38939
   420
    (if freq < k then Math.ln (Real.fromInt (freq + 1)) / Math.ln x
blanchet@38939
   421
     else rel_weight_for order freq / rel_weight_for order k)
blanchet@40070
   422
    * pow_int higher_order_irrel_weight (order - 1)
blanchet@38938
   423
  end
blanchet@37503
   424
blanchet@41790
   425
fun multiplier_for_const_name local_const_multiplier s =
blanchet@41790
   426
  if String.isSubstring "." s then 1.0 else local_const_multiplier
blanchet@38821
   427
blanchet@41790
   428
(* Computes a constant's weight, as determined by its frequency. *)
blanchet@41790
   429
fun generic_pconst_weight local_const_multiplier abs_weight skolem_weight
blanchet@41790
   430
                          theory_const_weight weight_for f const_tab
blanchet@41790
   431
                          (c as (s, PType (m, _))) =
blanchet@41790
   432
  if s = abs_name then
blanchet@41790
   433
    abs_weight
blanchet@41790
   434
  else if String.isPrefix skolem_prefix s then
blanchet@41790
   435
    skolem_weight
blanchet@41790
   436
  else if String.isSuffix theory_const_suffix s then
blanchet@41790
   437
    theory_const_weight
blanchet@41790
   438
  else
blanchet@41790
   439
    multiplier_for_const_name local_const_multiplier s
blanchet@41790
   440
    * weight_for m (pconst_freq (match_ptype o f) const_tab c)
blanchet@41790
   441
blanchet@41790
   442
fun rel_pconst_weight ({local_const_multiplier, abs_rel_weight,
blanchet@41790
   443
                        theory_const_rel_weight, ...} : relevance_fudge)
blanchet@41790
   444
                      const_tab =
blanchet@41790
   445
  generic_pconst_weight local_const_multiplier abs_rel_weight 0.0
blanchet@41790
   446
                        theory_const_rel_weight rel_weight_for I const_tab
blanchet@41790
   447
fun irrel_pconst_weight (fudge as {local_const_multiplier, abs_irrel_weight,
blanchet@41790
   448
                                   skolem_irrel_weight,
blanchet@40070
   449
                                   theory_const_irrel_weight, ...}) const_tab =
blanchet@41790
   450
  generic_pconst_weight local_const_multiplier abs_irrel_weight
blanchet@41790
   451
                        skolem_irrel_weight theory_const_irrel_weight
blanchet@41790
   452
                        (irrel_weight_for fudge) swap const_tab
paulson@24287
   453
blanchet@40070
   454
fun locality_bonus (_ : relevance_fudge) General = 0.0
blanchet@40070
   455
  | locality_bonus {intro_bonus, ...} Intro = intro_bonus
blanchet@40070
   456
  | locality_bonus {elim_bonus, ...} Elim = elim_bonus
blanchet@40070
   457
  | locality_bonus {simp_bonus, ...} Simp = simp_bonus
blanchet@40070
   458
  | locality_bonus {local_bonus, ...} Local = local_bonus
blanchet@40070
   459
  | locality_bonus {assum_bonus, ...} Assum = assum_bonus
blanchet@40070
   460
  | locality_bonus {chained_bonus, ...} Chained = chained_bonus
blanchet@38751
   461
blanchet@40418
   462
fun is_odd_const_name s =
blanchet@40418
   463
  s = abs_name orelse String.isPrefix skolem_prefix s orelse
blanchet@40418
   464
  String.isSuffix theory_const_suffix s
blanchet@40418
   465
blanchet@40204
   466
fun fact_weight fudge loc const_tab relevant_consts fact_consts =
blanchet@40204
   467
  case fact_consts |> List.partition (pconst_hyper_mem I relevant_consts)
blanchet@40204
   468
                   ||> filter_out (pconst_hyper_mem swap relevant_consts) of
blanchet@38827
   469
    ([], _) => 0.0
blanchet@38744
   470
  | (rel, irrel) =>
blanchet@40418
   471
    if forall (forall (is_odd_const_name o fst)) [rel, irrel] then
blanchet@40371
   472
      0.0
blanchet@40371
   473
    else
blanchet@40371
   474
      let
blanchet@40371
   475
        val irrel = irrel |> filter_out (pconst_mem swap rel)
blanchet@40371
   476
        val rel_weight =
blanchet@40371
   477
          0.0 |> fold (curry (op +) o rel_pconst_weight fudge const_tab) rel
blanchet@40371
   478
        val irrel_weight =
blanchet@40371
   479
          ~ (locality_bonus fudge loc)
blanchet@40371
   480
          |> fold (curry (op +) o irrel_pconst_weight fudge const_tab) irrel
blanchet@40371
   481
        val res = rel_weight / (rel_weight + irrel_weight)
blanchet@40371
   482
      in if Real.isFinite res then res else 0.0 end
blanchet@38747
   483
blanchet@40369
   484
fun pconsts_in_fact thy is_built_in_const t =
blanchet@38825
   485
  Symtab.fold (fn (s, pss) => fold (cons o pair s) pss)
blanchet@40369
   486
              (pconsts_in_terms thy is_built_in_const true (SOME true) [t]) []
blanchet@40369
   487
fun pair_consts_fact thy is_built_in_const fudge fact =
blanchet@40204
   488
  case fact |> snd |> theory_const_prop_of fudge
blanchet@40369
   489
            |> pconsts_in_fact thy is_built_in_const of
blanchet@38827
   490
    [] => NONE
blanchet@40204
   491
  | consts => SOME ((fact, consts), NONE)
paulson@24287
   492
blanchet@41768
   493
val const_names_in_fact = map fst ooo pconsts_in_fact
blanchet@41768
   494
blanchet@38699
   495
type annotated_thm =
blanchet@38939
   496
  (((unit -> string) * locality) * thm) * (string * ptype) list
blanchet@37505
   497
blanchet@38904
   498
fun take_most_relevant max_relevant remaining_max
blanchet@40070
   499
        ({max_imperfect, max_imperfect_exp, ...} : relevance_fudge) 
blanchet@40070
   500
        (candidates : (annotated_thm * real) list) =
blanchet@38744
   501
  let
blanchet@38747
   502
    val max_imperfect =
blanchet@40070
   503
      Real.ceil (Math.pow (max_imperfect,
blanchet@38904
   504
                    Math.pow (Real.fromInt remaining_max
blanchet@40070
   505
                              / Real.fromInt max_relevant, max_imperfect_exp)))
blanchet@38747
   506
    val (perfect, imperfect) =
blanchet@38889
   507
      candidates |> sort (Real.compare o swap o pairself snd)
blanchet@38889
   508
                 |> take_prefix (fn (_, w) => w > 0.99999)
blanchet@38747
   509
    val ((accepts, more_rejects), rejects) =
blanchet@38747
   510
      chop max_imperfect imperfect |>> append perfect |>> chop remaining_max
blanchet@38744
   511
  in
blanchet@38889
   512
    trace_msg (fn () =>
wenzelm@41491
   513
        "Actually passed (" ^ string_of_int (length accepts) ^ " of " ^
wenzelm@41491
   514
        string_of_int (length candidates) ^ "): " ^
blanchet@38889
   515
        (accepts |> map (fn ((((name, _), _), _), weight) =>
blanchet@38752
   516
                            name () ^ " [" ^ Real.toString weight ^ "]")
blanchet@38745
   517
                 |> commas));
blanchet@38747
   518
    (accepts, more_rejects @ rejects)
blanchet@38744
   519
  end
paulson@24287
   520
blanchet@40369
   521
fun if_empty_replace_with_locality thy is_built_in_const facts loc tab =
blanchet@38819
   522
  if Symtab.is_empty tab then
blanchet@40369
   523
    pconsts_in_terms thy is_built_in_const false (SOME false)
blanchet@38819
   524
        (map_filter (fn ((_, loc'), th) =>
blanchet@40204
   525
                        if loc' = loc then SOME (prop_of th) else NONE) facts)
blanchet@38819
   526
  else
blanchet@38819
   527
    tab
blanchet@38819
   528
blanchet@41158
   529
fun add_arities is_built_in_const th =
blanchet@41158
   530
  let
blanchet@41158
   531
    fun aux _ _ NONE = NONE
blanchet@41158
   532
      | aux t args (SOME tab) =
blanchet@41158
   533
        case t of
blanchet@41158
   534
          t1 $ t2 => SOME tab |> aux t1 (t2 :: args) |> aux t2 []
blanchet@41158
   535
        | Const (x as (s, _)) =>
blanchet@41336
   536
          (if is_built_in_const x args |> fst then
blanchet@41158
   537
             SOME tab
blanchet@41158
   538
           else case Symtab.lookup tab s of
blanchet@41158
   539
             NONE => SOME (Symtab.update (s, length args) tab)
blanchet@41158
   540
           | SOME n => if n = length args then SOME tab else NONE)
blanchet@41158
   541
        | _ => SOME tab
blanchet@41158
   542
  in aux (prop_of th) [] end
blanchet@41158
   543
blanchet@41158
   544
fun needs_ext is_built_in_const facts =
blanchet@41158
   545
  fold (add_arities is_built_in_const o snd) facts (SOME Symtab.empty)
blanchet@41158
   546
  |> is_none
blanchet@41158
   547
blanchet@40369
   548
fun relevance_filter ctxt threshold0 decay max_relevant is_built_in_const
blanchet@40070
   549
        (fudge as {threshold_divisor, ridiculous_threshold, ...})
blanchet@40204
   550
        ({add, del, ...} : relevance_override) facts goal_ts =
blanchet@38739
   551
  let
wenzelm@42361
   552
    val thy = Proof_Context.theory_of ctxt
blanchet@40204
   553
    val const_tab = fold (count_fact_consts thy fudge) facts Symtab.empty
blanchet@38819
   554
    val goal_const_tab =
blanchet@40369
   555
      pconsts_in_terms thy is_built_in_const false (SOME false) goal_ts
blanchet@40369
   556
      |> fold (if_empty_replace_with_locality thy is_built_in_const facts)
blanchet@38993
   557
              [Chained, Assum, Local]
blanchet@39012
   558
    val add_ths = Attrib.eval_thms ctxt add
blanchet@39012
   559
    val del_ths = Attrib.eval_thms ctxt del
blanchet@40204
   560
    val facts = facts |> filter_out (member Thm.eq_thm del_ths o snd)
blanchet@38747
   561
    fun iter j remaining_max threshold rel_const_tab hopeless hopeful =
blanchet@38739
   562
      let
blanchet@40191
   563
        fun relevant [] _ [] =
blanchet@38747
   564
            (* Nothing has been added this iteration. *)
blanchet@40070
   565
            if j = 0 andalso threshold >= ridiculous_threshold then
blanchet@38747
   566
              (* First iteration? Try again. *)
blanchet@40070
   567
              iter 0 max_relevant (threshold / threshold_divisor) rel_const_tab
blanchet@38747
   568
                   hopeless hopeful
blanchet@38744
   569
            else
blanchet@40191
   570
              []
blanchet@38889
   571
          | relevant candidates rejects [] =
blanchet@38739
   572
            let
blanchet@38747
   573
              val (accepts, more_rejects) =
blanchet@40070
   574
                take_most_relevant max_relevant remaining_max fudge candidates
blanchet@38739
   575
              val rel_const_tab' =
blanchet@38745
   576
                rel_const_tab
blanchet@41066
   577
                |> fold (add_pconst_to_table false) (maps (snd o fst) accepts)
blanchet@38744
   578
              fun is_dirty (c, _) =
blanchet@38744
   579
                Symtab.lookup rel_const_tab' c <> Symtab.lookup rel_const_tab c
blanchet@38745
   580
              val (hopeful_rejects, hopeless_rejects) =
blanchet@38745
   581
                 (rejects @ hopeless, ([], []))
blanchet@38745
   582
                 |-> fold (fn (ax as (_, consts), old_weight) =>
blanchet@38745
   583
                              if exists is_dirty consts then
blanchet@38745
   584
                                apfst (cons (ax, NONE))
blanchet@38745
   585
                              else
blanchet@38745
   586
                                apsnd (cons (ax, old_weight)))
blanchet@38745
   587
                 |>> append (more_rejects
blanchet@38745
   588
                             |> map (fn (ax as (_, consts), old_weight) =>
blanchet@38745
   589
                                        (ax, if exists is_dirty consts then NONE
blanchet@38745
   590
                                             else SOME old_weight)))
blanchet@38747
   591
              val threshold =
blanchet@38822
   592
                1.0 - (1.0 - threshold)
blanchet@38822
   593
                      * Math.pow (decay, Real.fromInt (length accepts))
blanchet@38747
   594
              val remaining_max = remaining_max - length accepts
blanchet@38739
   595
            in
blanchet@38744
   596
              trace_msg (fn () => "New or updated constants: " ^
blanchet@38744
   597
                  commas (rel_const_tab' |> Symtab.dest
blanchet@38822
   598
                          |> subtract (op =) (rel_const_tab |> Symtab.dest)
blanchet@38827
   599
                          |> map string_for_hyper_pconst));
blanchet@38745
   600
              map (fst o fst) accepts @
blanchet@38747
   601
              (if remaining_max = 0 then
blanchet@40191
   602
                 []
blanchet@38745
   603
               else
blanchet@38747
   604
                 iter (j + 1) remaining_max threshold rel_const_tab'
blanchet@38747
   605
                      hopeless_rejects hopeful_rejects)
blanchet@38739
   606
            end
blanchet@38889
   607
          | relevant candidates rejects
blanchet@40204
   608
                     (((ax as (((_, loc), _), fact_consts)), cached_weight)
blanchet@38747
   609
                      :: hopeful) =
blanchet@38739
   610
            let
blanchet@38739
   611
              val weight =
blanchet@38739
   612
                case cached_weight of
blanchet@38739
   613
                  SOME w => w
blanchet@40204
   614
                | NONE => fact_weight fudge loc const_tab rel_const_tab
blanchet@40204
   615
                                      fact_consts
blanchet@38739
   616
            in
blanchet@38741
   617
              if weight >= threshold then
blanchet@38889
   618
                relevant ((ax, weight) :: candidates) rejects hopeful
blanchet@38739
   619
              else
blanchet@38889
   620
                relevant candidates ((ax, weight) :: rejects) hopeful
blanchet@38739
   621
            end
blanchet@38739
   622
        in
blanchet@38744
   623
          trace_msg (fn () =>
blanchet@38744
   624
              "ITERATION " ^ string_of_int j ^ ": current threshold: " ^
blanchet@38744
   625
              Real.toString threshold ^ ", constants: " ^
blanchet@38744
   626
              commas (rel_const_tab |> Symtab.dest
blanchet@38744
   627
                      |> filter (curry (op <>) [] o snd)
blanchet@38827
   628
                      |> map string_for_hyper_pconst));
blanchet@38889
   629
          relevant [] [] hopeful
blanchet@38739
   630
        end
blanchet@41158
   631
    fun add_facts ths accepts =
blanchet@41167
   632
      (facts |> filter (member Thm.eq_thm ths o snd)) @
blanchet@41167
   633
      (accepts |> filter_out (member Thm.eq_thm ths o snd))
blanchet@40408
   634
      |> take max_relevant
blanchet@38739
   635
  in
blanchet@40369
   636
    facts |> map_filter (pair_consts_fact thy is_built_in_const fudge)
blanchet@40204
   637
          |> iter 0 max_relevant threshold0 goal_const_tab []
blanchet@41158
   638
          |> not (null add_ths) ? add_facts add_ths
blanchet@41158
   639
          |> (fn accepts =>
blanchet@41158
   640
                 accepts |> needs_ext is_built_in_const accepts
blanchet@41158
   641
                            ? add_facts @{thms ext})
blanchet@41158
   642
          |> tap (fn accepts => trace_msg (fn () =>
wenzelm@41491
   643
                      "Total relevant: " ^ string_of_int (length accepts)))
blanchet@38739
   644
  end
paulson@24287
   645
blanchet@38744
   646
paulson@24287
   647
(***************************************************************)
mengj@19768
   648
(* Retrieving and filtering lemmas                             *)
mengj@19768
   649
(***************************************************************)
mengj@19768
   650
paulson@33022
   651
(*** retrieve lemmas and filter them ***)
mengj@19768
   652
paulson@20757
   653
(*Reject theorems with names like "List.filter.filter_list_def" or
paulson@21690
   654
  "Accessible_Part.acc.defs", as these are definitions arising from packages.*)
paulson@20757
   655
fun is_package_def a =
blanchet@40205
   656
  let val names = Long_Name.explode a in
blanchet@40205
   657
    (length names > 2 andalso not (hd names = "local") andalso
blanchet@40205
   658
     String.isSuffix "_def" a) orelse String.isSuffix "_defs" a
blanchet@40205
   659
  end
paulson@20757
   660
blanchet@38937
   661
fun mk_fact_table f xs =
blanchet@38937
   662
  fold (Termtab.update o `(prop_of o f)) xs Termtab.empty
blanchet@38937
   663
fun uniquify xs = Termtab.fold (cons o snd) (mk_fact_table snd xs) []
mengj@19768
   664
blanchet@37626
   665
(* FIXME: put other record thms here, or declare as "no_atp" *)
blanchet@37626
   666
val multi_base_blacklist =
blanchet@41199
   667
  ["defs", "select_defs", "update_defs", "split", "splits", "split_asm",
blanchet@41199
   668
   "cases", "ext_cases", "eq.simps", "eq.refl", "nchotomy", "case_cong",
blanchet@41199
   669
   "weak_case_cong"]
blanchet@41273
   670
  |> not instantiate_inducts ? append ["induct", "inducts"]
blanchet@38682
   671
  |> map (prefix ".")
blanchet@37626
   672
blanchet@37626
   673
val max_lambda_nesting = 3
blanchet@37626
   674
blanchet@37626
   675
fun term_has_too_many_lambdas max (t1 $ t2) =
blanchet@37626
   676
    exists (term_has_too_many_lambdas max) [t1, t2]
blanchet@37626
   677
  | term_has_too_many_lambdas max (Abs (_, _, t)) =
blanchet@37626
   678
    max = 0 orelse term_has_too_many_lambdas (max - 1) t
blanchet@37626
   679
  | term_has_too_many_lambdas _ _ = false
blanchet@37626
   680
blanchet@37626
   681
(* Don't count nested lambdas at the level of formulas, since they are
blanchet@37626
   682
   quantifiers. *)
blanchet@37626
   683
fun formula_has_too_many_lambdas Ts (Abs (_, T, t)) =
blanchet@37626
   684
    formula_has_too_many_lambdas (T :: Ts) t
blanchet@37626
   685
  | formula_has_too_many_lambdas Ts t =
blanchet@41273
   686
    if member (op =) [HOLogic.boolT, propT] (fastype_of1 (Ts, t)) then
blanchet@37626
   687
      exists (formula_has_too_many_lambdas Ts) (#2 (strip_comb t))
blanchet@37626
   688
    else
blanchet@37626
   689
      term_has_too_many_lambdas max_lambda_nesting t
blanchet@37626
   690
blanchet@38692
   691
(* The max apply depth of any "metis" call in "Metis_Examples" (on 2007-10-31)
blanchet@37626
   692
   was 11. *)
blanchet@37626
   693
val max_apply_depth = 15
blanchet@37626
   694
blanchet@37626
   695
fun apply_depth (f $ t) = Int.max (apply_depth f, apply_depth t + 1)
blanchet@37626
   696
  | apply_depth (Abs (_, _, t)) = apply_depth t
blanchet@37626
   697
  | apply_depth _ = 0
blanchet@37626
   698
blanchet@37626
   699
fun is_formula_too_complex t =
blanchet@38085
   700
  apply_depth t > max_apply_depth orelse formula_has_too_many_lambdas [] t
blanchet@37626
   701
blanchet@39946
   702
(* FIXME: Extend to "Meson" and "Metis" *)
blanchet@37543
   703
val exists_sledgehammer_const =
blanchet@37626
   704
  exists_Const (fn (s, _) => String.isPrefix sledgehammer_prefix s)
blanchet@37626
   705
blanchet@38904
   706
(* FIXME: make more reliable *)
blanchet@38904
   707
val exists_low_level_class_const =
blanchet@38904
   708
  exists_Const (fn (s, _) =>
blanchet@38904
   709
     String.isSubstring (Long_Name.separator ^ "class" ^ Long_Name.separator) s)
blanchet@38904
   710
blanchet@38821
   711
fun is_metastrange_theorem th =
blanchet@37626
   712
  case head_of (concl_of th) of
blanchet@37626
   713
      Const (a, _) => (a <> @{const_name Trueprop} andalso
blanchet@37626
   714
                       a <> @{const_name "=="})
blanchet@37626
   715
    | _ => false
blanchet@37626
   716
blanchet@38821
   717
fun is_that_fact th =
blanchet@38821
   718
  String.isSuffix (Long_Name.separator ^ Obtain.thatN) (Thm.get_name_hint th)
blanchet@38821
   719
  andalso exists_subterm (fn Free (s, _) => s = Name.skolem Auto_Bind.thesisN
blanchet@38821
   720
                           | _ => false) (prop_of th)
blanchet@38821
   721
blanchet@37626
   722
val type_has_top_sort =
blanchet@37626
   723
  exists_subtype (fn TFree (_, []) => true | TVar (_, []) => true | _ => false)
blanchet@37626
   724
blanchet@38085
   725
(**** Predicates to detect unwanted facts (prolific or likely to cause
blanchet@37347
   726
      unsoundness) ****)
paulson@21470
   727
blanchet@38289
   728
(* Too general means, positive equality literal with a variable X as one
blanchet@38289
   729
   operand, when X does not occur properly in the other operand. This rules out
blanchet@38289
   730
   clearly inconsistent facts such as X = a | X = b, though it by no means
blanchet@38289
   731
   guarantees soundness. *)
paulson@21470
   732
blanchet@38289
   733
(* Unwanted equalities are those between a (bound or schematic) variable that
blanchet@38289
   734
   does not properly occur in the second operand. *)
blanchet@38607
   735
val is_exhaustive_finite =
blanchet@38607
   736
  let
blanchet@38629
   737
    fun is_bad_equal (Var z) t =
blanchet@38629
   738
        not (exists_subterm (fn Var z' => z = z' | _ => false) t)
blanchet@38629
   739
      | is_bad_equal (Bound j) t = not (loose_bvar1 (t, j))
blanchet@38629
   740
      | is_bad_equal _ _ = false
blanchet@38629
   741
    fun do_equals t1 t2 = is_bad_equal t1 t2 orelse is_bad_equal t2 t1
blanchet@38607
   742
    fun do_formula pos t =
blanchet@38607
   743
      case (pos, t) of
blanchet@38615
   744
        (_, @{const Trueprop} $ t1) => do_formula pos t1
blanchet@38607
   745
      | (true, Const (@{const_name all}, _) $ Abs (_, _, t')) =>
blanchet@38607
   746
        do_formula pos t'
blanchet@38607
   747
      | (true, Const (@{const_name All}, _) $ Abs (_, _, t')) =>
blanchet@38607
   748
        do_formula pos t'
blanchet@38607
   749
      | (false, Const (@{const_name Ex}, _) $ Abs (_, _, t')) =>
blanchet@38607
   750
        do_formula pos t'
blanchet@38607
   751
      | (_, @{const "==>"} $ t1 $ t2) =>
blanchet@38629
   752
        do_formula (not pos) t1 andalso
blanchet@38629
   753
        (t2 = @{prop False} orelse do_formula pos t2)
haftmann@38786
   754
      | (_, @{const HOL.implies} $ t1 $ t2) =>
blanchet@38629
   755
        do_formula (not pos) t1 andalso
blanchet@38629
   756
        (t2 = @{const False} orelse do_formula pos t2)
blanchet@38607
   757
      | (_, @{const Not} $ t1) => do_formula (not pos) t1
haftmann@38795
   758
      | (true, @{const HOL.disj} $ t1 $ t2) => forall (do_formula pos) [t1, t2]
haftmann@38795
   759
      | (false, @{const HOL.conj} $ t1 $ t2) => forall (do_formula pos) [t1, t2]
haftmann@38864
   760
      | (true, Const (@{const_name HOL.eq}, _) $ t1 $ t2) => do_equals t1 t2
blanchet@38607
   761
      | (true, Const (@{const_name "=="}, _) $ t1 $ t2) => do_equals t1 t2
blanchet@38607
   762
      | _ => false
blanchet@38607
   763
  in do_formula true end
blanchet@38607
   764
blanchet@38592
   765
fun has_bound_or_var_of_type tycons =
blanchet@38592
   766
  exists_subterm (fn Var (_, Type (s, _)) => member (op =) tycons s
blanchet@38592
   767
                   | Abs (_, Type (s, _), _) => member (op =) tycons s
blanchet@38592
   768
                   | _ => false)
paulson@21431
   769
blanchet@38085
   770
(* Facts are forbidden to contain variables of these types. The typical reason
blanchet@37347
   771
   is that they lead to unsoundness. Note that "unit" satisfies numerous
blanchet@38085
   772
   equations like "?x = ()". The resulting clauses will have no type constraint,
blanchet@37347
   773
   yielding false proofs. Even "bool" leads to many unsound proofs, though only
blanchet@37347
   774
   for higher-order problems. *)
blanchet@38592
   775
val dangerous_types = [@{type_name unit}, @{type_name bool}, @{type_name prop}];
paulson@22217
   776
blanchet@38085
   777
(* Facts containing variables of type "unit" or "bool" or of the form
blanchet@38290
   778
   "ALL x. x = A | x = B | x = C" are likely to lead to unsound proofs if types
blanchet@38290
   779
   are omitted. *)
blanchet@41138
   780
fun is_dangerous_term no_dangerous_types t =
blanchet@41138
   781
  not no_dangerous_types andalso
blanchet@38679
   782
  let val t = transform_elim_term t in
blanchet@38679
   783
    has_bound_or_var_of_type dangerous_types t orelse
blanchet@38679
   784
    is_exhaustive_finite t
blanchet@38679
   785
  end
paulson@21470
   786
blanchet@41138
   787
fun is_theorem_bad_for_atps no_dangerous_types thm =
blanchet@38627
   788
  let val t = prop_of thm in
blanchet@38627
   789
    is_formula_too_complex t orelse exists_type type_has_top_sort t orelse
blanchet@41138
   790
    is_dangerous_term no_dangerous_types t orelse
blanchet@41138
   791
    exists_sledgehammer_const t orelse exists_low_level_class_const t orelse
blanchet@41138
   792
    is_metastrange_theorem thm orelse is_that_fact thm
blanchet@38627
   793
  end
blanchet@38627
   794
blanchet@38937
   795
fun clasimpset_rules_of ctxt =
blanchet@38937
   796
  let
blanchet@38937
   797
    val {safeIs, safeEs, hazIs, hazEs, ...} = ctxt |> claset_of |> rep_cs
blanchet@38937
   798
    val intros = safeIs @ hazIs
blanchet@38937
   799
    val elims = map Classical.classical_rule (safeEs @ hazEs)
blanchet@38937
   800
    val simps = ctxt |> simpset_of |> dest_ss |> #simps |> map snd
blanchet@38937
   801
  in (mk_fact_table I intros, mk_fact_table I elims, mk_fact_table I simps) end
blanchet@38937
   802
blanchet@41989
   803
fun all_facts ctxt reserved really_all no_dangerous_types
blanchet@40205
   804
              ({intro_bonus, elim_bonus, simp_bonus, ...} : relevance_fudge)
blanchet@40205
   805
              add_ths chained_ths =
blanchet@38627
   806
  let
wenzelm@42361
   807
    val thy = Proof_Context.theory_of ctxt
wenzelm@39557
   808
    val global_facts = Global_Theory.facts_of thy
wenzelm@42361
   809
    val local_facts = Proof_Context.facts_of ctxt
blanchet@38644
   810
    val named_locals = local_facts |> Facts.dest_static []
blanchet@38993
   811
    val assms = Assumption.all_assms_of ctxt
blanchet@38993
   812
    fun is_assum th = exists (fn ct => prop_of th aconv term_of ct) assms
blanchet@38752
   813
    val is_chained = member Thm.eq_thm chained_ths
blanchet@38937
   814
    val (intros, elims, simps) =
blanchet@40070
   815
      if exists (curry (op <) 0.0) [intro_bonus, elim_bonus, simp_bonus] then
blanchet@38937
   816
        clasimpset_rules_of ctxt
blanchet@38937
   817
      else
blanchet@38937
   818
        (Termtab.empty, Termtab.empty, Termtab.empty)
blanchet@38738
   819
    fun is_good_unnamed_local th =
blanchet@38820
   820
      not (Thm.has_name_hint th) andalso
blanchet@38738
   821
      forall (fn (_, ths) => not (member Thm.eq_thm ths th)) named_locals
blanchet@38644
   822
    val unnamed_locals =
blanchet@38820
   823
      union Thm.eq_thm (Facts.props local_facts) chained_ths
blanchet@38820
   824
      |> filter is_good_unnamed_local |> map (pair "" o single)
blanchet@38627
   825
    val full_space =
blanchet@38738
   826
      Name_Space.merge (Facts.space_of global_facts, Facts.space_of local_facts)
blanchet@38752
   827
    fun add_facts global foldx facts =
blanchet@38699
   828
      foldx (fn (name0, ths) =>
blanchet@41989
   829
        if not really_all andalso name0 <> "" andalso
blanchet@39012
   830
           forall (not o member Thm.eq_thm add_ths) ths andalso
blanchet@38699
   831
           (Facts.is_concealed facts name0 orelse
blanchet@38699
   832
            (respect_no_atp andalso is_package_def name0) orelse
blanchet@38699
   833
            exists (fn s => String.isSuffix s name0) multi_base_blacklist orelse
blanchet@38699
   834
            String.isSuffix "_def_raw" (* FIXME: crude hack *) name0) then
blanchet@38627
   835
          I
blanchet@38627
   836
        else
blanchet@38627
   837
          let
blanchet@38699
   838
            val multi = length ths > 1
blanchet@41279
   839
            val backquote_thm =
blanchet@41279
   840
              backquote o string_for_term ctxt o close_form o prop_of
blanchet@38699
   841
            fun check_thms a =
wenzelm@42361
   842
              case try (Proof_Context.get_thms ctxt) a of
blanchet@38699
   843
                NONE => false
blanchet@38699
   844
              | SOME ths' => Thm.eq_thms (ths, ths')
blanchet@38627
   845
          in
blanchet@38699
   846
            pair 1
blanchet@38699
   847
            #> fold (fn th => fn (j, rest) =>
blanchet@38699
   848
                 (j + 1,
blanchet@41138
   849
                  if is_theorem_bad_for_atps no_dangerous_types th andalso
blanchet@39012
   850
                     not (member Thm.eq_thm add_ths th) then
blanchet@38699
   851
                    rest
blanchet@38699
   852
                  else
blanchet@38752
   853
                    (((fn () =>
blanchet@38752
   854
                          if name0 = "" then
blanchet@41279
   855
                            th |> backquote_thm
blanchet@38752
   856
                          else
blanchet@38752
   857
                            let
wenzelm@42358
   858
                              val name1 = Facts.extern ctxt facts name0
wenzelm@42358
   859
                              val name2 = Name_Space.extern ctxt full_space name0
blanchet@38752
   860
                            in
blanchet@38752
   861
                              case find_first check_thms [name1, name2, name0] of
blanchet@40375
   862
                                SOME name => make_name reserved multi j name
blanchet@38752
   863
                              | NONE => ""
blanchet@38937
   864
                            end),
blanchet@38937
   865
                      let val t = prop_of th in
blanchet@40071
   866
                        if is_chained th then
blanchet@40071
   867
                          Chained
blanchet@38993
   868
                        else if global then
blanchet@38993
   869
                          if Termtab.defined intros t then Intro
blanchet@38993
   870
                          else if Termtab.defined elims t then Elim
blanchet@38993
   871
                          else if Termtab.defined simps t then Simp
blanchet@38993
   872
                          else General
blanchet@38993
   873
                        else
blanchet@38993
   874
                          if is_assum th then Assum else Local
blanchet@38937
   875
                      end),
blanchet@38752
   876
                      (multi, th)) :: rest)) ths
blanchet@38699
   877
            #> snd
blanchet@38627
   878
          end)
blanchet@38644
   879
  in
blanchet@38752
   880
    [] |> add_facts false fold local_facts (unnamed_locals @ named_locals)
blanchet@38752
   881
       |> add_facts true Facts.fold_static global_facts global_facts
blanchet@38644
   882
  end
blanchet@38627
   883
blanchet@38627
   884
(* The single-name theorems go after the multiple-name ones, so that single
blanchet@38627
   885
   names are preferred when both are available. *)
blanchet@40205
   886
fun rearrange_facts ctxt respect_no_atp =
blanchet@38744
   887
  List.partition (fst o snd) #> op @ #> map (apsnd snd)
blanchet@38699
   888
  #> respect_no_atp ? filter_out (No_ATPs.member ctxt o snd)
blanchet@38627
   889
blanchet@41199
   890
fun external_frees t =
blanchet@41199
   891
  [] |> Term.add_frees t |> filter_out (can Name.dest_internal o fst)
blanchet@41199
   892
blanchet@41138
   893
fun relevant_facts ctxt no_dangerous_types (threshold0, threshold1)
blanchet@41138
   894
                   max_relevant is_built_in_const fudge
blanchet@41138
   895
                   (override as {add, only, ...}) chained_ths hyp_ts concl_t =
blanchet@37538
   896
  let
wenzelm@42361
   897
    val thy = Proof_Context.theory_of ctxt
blanchet@38822
   898
    val decay = Math.pow ((1.0 - threshold1) / (1.0 - threshold0),
blanchet@38822
   899
                          1.0 / Real.fromInt (max_relevant + 1))
blanchet@39012
   900
    val add_ths = Attrib.eval_thms ctxt add
blanchet@38696
   901
    val reserved = reserved_isar_keyword_table ()
blanchet@41199
   902
    val ind_stmt =
blanchet@41199
   903
      Logic.list_implies (hyp_ts |> filter_out (null o external_frees), concl_t)
blanchet@41211
   904
      |> Object_Logic.atomize_term thy
blanchet@41199
   905
    val ind_stmt_xs = external_frees ind_stmt
blanchet@40204
   906
    val facts =
blanchet@38699
   907
      (if only then
blanchet@38752
   908
         maps (map (fn ((name, loc), th) => ((K name, loc), (true, th)))
blanchet@40205
   909
               o fact_from_ref ctxt reserved chained_ths) add
blanchet@38699
   910
       else
blanchet@41989
   911
         all_facts ctxt reserved false no_dangerous_types fudge add_ths
blanchet@41989
   912
                   chained_ths)
blanchet@41273
   913
      |> instantiate_inducts
blanchet@41273
   914
         ? maps (instantiate_if_induct_rule ctxt ind_stmt ind_stmt_xs)
blanchet@40205
   915
      |> rearrange_facts ctxt (respect_no_atp andalso not only)
blanchet@38937
   916
      |> uniquify
blanchet@37538
   917
  in
wenzelm@41491
   918
    trace_msg (fn () => "Considering " ^ string_of_int (length facts) ^
blanchet@40204
   919
                        " facts");
blanchet@39366
   920
    (if only orelse threshold1 < 0.0 then
blanchet@40204
   921
       facts
blanchet@39366
   922
     else if threshold0 > 1.0 orelse threshold0 > threshold1 orelse
blanchet@39366
   923
             max_relevant = 0 then
blanchet@38739
   924
       []
blanchet@38739
   925
     else
blanchet@41200
   926
       ((concl_t |> theory_constify fudge (Context.theory_name thy)) :: hyp_ts)
blanchet@41200
   927
       |> relevance_filter ctxt threshold0 decay max_relevant is_built_in_const
blanchet@41200
   928
                           fudge override facts)
blanchet@38822
   929
    |> map (apfst (apfst (fn f => f ())))
blanchet@37538
   930
  end
immler@30536
   931
paulson@15347
   932
end;