src/HOL/Tools/Sledgehammer/sledgehammer_mepo.ML
author blanchet
Thu Jan 31 17:54:05 2013 +0100 (2013-01-31 ago)
changeset 51004 5f2788c38127
parent 50985 23bb011a5832
child 51026 48e82e199df1
permissions -rw-r--r--
distinguish raw and non-raw facts, using raw for 10 000s of facts and non-raw after selection of some hundreds
blanchet@48380
     1
(*  Title:      HOL/Tools/Sledgehammer/sledgehammer_mepo.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@48380
     5
Sledgehammer's iterative relevance filter (MePo = Meng-Paulson).
wenzelm@33309
     6
*)
paulson@15452
     7
blanchet@48381
     8
signature SLEDGEHAMMER_MEPO =
wenzelm@16802
     9
sig
blanchet@46340
    10
  type stature = ATP_Problem_Generate.stature
blanchet@51004
    11
  type raw_fact = Sledgehammer_Fact.raw_fact
blanchet@48296
    12
  type fact = Sledgehammer_Fact.fact
blanchet@48288
    13
  type params = Sledgehammer_Provers.params
blanchet@48288
    14
  type relevance_fudge = Sledgehammer_Provers.relevance_fudge
blanchet@40070
    15
blanchet@42646
    16
  val trace : bool Config.T
blanchet@48219
    17
  val pseudo_abs_name : string
blanchet@48219
    18
  val pseudo_skolem_prefix : string
blanchet@43351
    19
  val const_names_in_fact :
blanchet@43351
    20
    theory -> (string * typ -> term list -> bool * term list) -> term
blanchet@43351
    21
    -> string list
blanchet@48406
    22
  val mepo_suggested_facts :
blanchet@48288
    23
    Proof.context -> params -> string -> int -> relevance_fudge option
blanchet@51004
    24
    -> term list -> term -> raw_fact list -> fact list
paulson@15347
    25
end;
paulson@15347
    26
blanchet@48381
    27
structure Sledgehammer_MePo : SLEDGEHAMMER_MEPO =
paulson@15347
    28
struct
paulson@15347
    29
blanchet@46320
    30
open ATP_Problem_Generate
blanchet@50608
    31
open Sledgehammer_Util
blanchet@48296
    32
open Sledgehammer_Fact
blanchet@48288
    33
open Sledgehammer_Provers
blanchet@38652
    34
blanchet@42646
    35
val trace =
blanchet@48308
    36
  Attrib.setup_config_bool @{binding sledgehammer_filter_iter_trace} (K false)
blanchet@42646
    37
fun trace_msg ctxt msg = if Config.get ctxt trace then tracing (msg ()) else ()
blanchet@35826
    38
blanchet@37616
    39
val sledgehammer_prefix = "Sledgehammer" ^ Long_Name.separator
blanchet@48219
    40
val pseudo_abs_name = sledgehammer_prefix ^ "abs"
blanchet@48219
    41
val pseudo_skolem_prefix = sledgehammer_prefix ^ "sko"
blanchet@38992
    42
val theory_const_suffix = Long_Name.separator ^ " 1"
blanchet@37616
    43
blanchet@46073
    44
fun order_of_type (Type (@{type_name fun}, [T1, T2])) =
blanchet@38939
    45
    Int.max (order_of_type T1 + 1, order_of_type T2)
blanchet@38939
    46
  | order_of_type (Type (_, Ts)) = fold (Integer.max o order_of_type) Ts 0
blanchet@38939
    47
  | order_of_type _ = 0
blanchet@38939
    48
blanchet@38823
    49
(* An abstraction of Isabelle types and first-order terms *)
blanchet@38823
    50
datatype pattern = PVar | PApp of string * pattern list
blanchet@38939
    51
datatype ptype = PType of int * pattern list
blanchet@38744
    52
blanchet@38823
    53
fun string_for_pattern PVar = "_"
blanchet@38823
    54
  | string_for_pattern (PApp (s, ps)) =
blanchet@38823
    55
    if null ps then s else s ^ string_for_patterns ps
blanchet@38823
    56
and string_for_patterns ps = "(" ^ commas (map string_for_pattern ps) ^ ")"
blanchet@38939
    57
fun string_for_ptype (PType (_, ps)) = string_for_patterns ps
paulson@24287
    58
paulson@24287
    59
(*Is the second type an instance of the first one?*)
blanchet@38824
    60
fun match_pattern (PVar, _) = true
blanchet@38824
    61
  | match_pattern (PApp _, PVar) = false
blanchet@38824
    62
  | match_pattern (PApp (s, ps), PApp (t, qs)) =
blanchet@38824
    63
    s = t andalso match_patterns (ps, qs)
blanchet@38824
    64
and match_patterns (_, []) = true
blanchet@38824
    65
  | match_patterns ([], _) = false
blanchet@38824
    66
  | match_patterns (p :: ps, q :: qs) =
blanchet@38824
    67
    match_pattern (p, q) andalso match_patterns (ps, qs)
blanchet@38939
    68
fun match_ptype (PType (_, ps), PType (_, qs)) = match_patterns (ps, qs)
paulson@24287
    69
blanchet@38823
    70
(* Is there a unifiable constant? *)
blanchet@38827
    71
fun pconst_mem f consts (s, ps) =
blanchet@38939
    72
  exists (curry (match_ptype o f) ps)
blanchet@38827
    73
         (map snd (filter (curry (op =) s o fst) consts))
blanchet@38827
    74
fun pconst_hyper_mem f const_tab (s, ps) =
blanchet@38939
    75
  exists (curry (match_ptype o f) ps) (these (Symtab.lookup const_tab s))
blanchet@37505
    76
blanchet@38939
    77
fun pattern_for_type (Type (s, Ts)) = PApp (s, map pattern_for_type Ts)
blanchet@38939
    78
  | pattern_for_type (TFree (s, _)) = PApp (s, [])
blanchet@38939
    79
  | pattern_for_type (TVar _) = PVar
blanchet@38827
    80
blanchet@38744
    81
(* Pairs a constant with the list of its type instantiations. *)
blanchet@41204
    82
fun ptype thy const x =
blanchet@38939
    83
  (if const then map pattern_for_type (these (try (Sign.const_typargs thy) x))
blanchet@41204
    84
   else [])
blanchet@41204
    85
fun rich_ptype thy const (s, T) =
blanchet@41204
    86
  PType (order_of_type T, ptype thy const (s, T))
blanchet@41204
    87
fun rich_pconst thy const (s, T) = (s, rich_ptype thy const (s, T))
paulson@24287
    88
blanchet@38939
    89
fun string_for_hyper_pconst (s, ps) =
blanchet@38939
    90
  s ^ "{" ^ commas (map string_for_ptype ps) ^ "}"
paulson@24287
    91
blanchet@38823
    92
(* Add a pconstant to the table, but a [] entry means a standard
blanchet@38819
    93
   connective, which we ignore.*)
blanchet@41066
    94
fun add_pconst_to_table also_skolem (s, p) =
blanchet@48219
    95
  if (not also_skolem andalso String.isPrefix pseudo_skolem_prefix s) then I
blanchet@41066
    96
  else Symtab.map_default (s, [p]) (insert (op =) p)
blanchet@38819
    97
blanchet@47933
    98
(* Set constants tend to pull in too many irrelevant facts. We limit the damage
blanchet@47933
    99
   by treating them more or less as if they were built-in but add their
blanchet@47933
   100
   axiomatization at the end. *)
blanchet@47933
   101
val set_consts = [@{const_name Collect}, @{const_name Set.member}]
blanchet@47939
   102
val set_thms = @{thms Collect_mem_eq mem_Collect_eq Collect_cong}
blanchet@47933
   103
blanchet@42732
   104
fun add_pconsts_in_term thy is_built_in_const also_skolems pos =
blanchet@37505
   105
  let
blanchet@38819
   106
    val flip = Option.map not
blanchet@38587
   107
    (* We include free variables, as well as constants, to handle locales. For
blanchet@41205
   108
       each quantifiers that must necessarily be skolemized by the automatic
blanchet@41205
   109
       prover, we introduce a fresh constant to simulate the effect of
blanchet@41205
   110
       Skolemization. *)
blanchet@47933
   111
    fun do_const const ext_arg (x as (s, _)) ts =
blanchet@41336
   112
      let val (built_in, ts) = is_built_in_const x ts in
blanchet@47933
   113
        if member (op =) set_consts s then
blanchet@47933
   114
          fold (do_term ext_arg) ts
blanchet@47933
   115
        else
blanchet@47933
   116
          (not built_in
blanchet@47933
   117
           ? add_pconst_to_table also_skolems (rich_pconst thy const x))
blanchet@47933
   118
          #> fold (do_term false) ts
blanchet@41336
   119
      end
blanchet@42741
   120
    and do_term ext_arg t =
blanchet@38827
   121
      case strip_comb t of
blanchet@47933
   122
        (Const x, ts) => do_const true ext_arg x ts
blanchet@47933
   123
      | (Free x, ts) => do_const false ext_arg x ts
blanchet@38939
   124
      | (Abs (_, T, t'), ts) =>
blanchet@42741
   125
        ((null ts andalso not ext_arg)
blanchet@42741
   126
         (* Since lambdas on the right-hand side of equalities are usually
blanchet@47953
   127
            extensionalized later by "abs_extensionalize_term", we don't
blanchet@47953
   128
            penalize them here. *)
blanchet@48219
   129
         ? add_pconst_to_table true (pseudo_abs_name,
blanchet@48219
   130
                                     PType (order_of_type T + 1, [])))
blanchet@42735
   131
        #> fold (do_term false) (t' :: ts)
blanchet@42735
   132
      | (_, ts) => fold (do_term false) ts
blanchet@38939
   133
    fun do_quantifier will_surely_be_skolemized abs_T body_t =
blanchet@37537
   134
      do_formula pos body_t
blanchet@38747
   135
      #> (if also_skolems andalso will_surely_be_skolemized then
blanchet@48219
   136
            add_pconst_to_table true (pseudo_skolem_prefix ^ serial_string (),
blanchet@48219
   137
                                      PType (order_of_type abs_T, []))
blanchet@38587
   138
          else
blanchet@38587
   139
            I)
blanchet@42741
   140
    and do_term_or_formula ext_arg T =
blanchet@42741
   141
      if T = HOLogic.boolT then do_formula NONE else do_term ext_arg
blanchet@37537
   142
    and do_formula pos t =
blanchet@37537
   143
      case t of
blanchet@38939
   144
        Const (@{const_name all}, _) $ Abs (_, T, t') =>
blanchet@38939
   145
        do_quantifier (pos = SOME false) T t'
blanchet@37537
   146
      | @{const "==>"} $ t1 $ t2 =>
blanchet@37537
   147
        do_formula (flip pos) t1 #> do_formula pos t2
blanchet@37537
   148
      | Const (@{const_name "=="}, Type (_, [T, _])) $ t1 $ t2 =>
blanchet@42741
   149
        do_term_or_formula false T t1 #> do_term_or_formula true T t2
blanchet@37537
   150
      | @{const Trueprop} $ t1 => do_formula pos t1
blanchet@41140
   151
      | @{const False} => I
blanchet@41140
   152
      | @{const True} => I
blanchet@37537
   153
      | @{const Not} $ t1 => do_formula (flip pos) t1
blanchet@38939
   154
      | Const (@{const_name All}, _) $ Abs (_, T, t') =>
blanchet@38939
   155
        do_quantifier (pos = SOME false) T t'
blanchet@38939
   156
      | Const (@{const_name Ex}, _) $ Abs (_, T, t') =>
blanchet@38939
   157
        do_quantifier (pos = SOME true) T t'
haftmann@38795
   158
      | @{const HOL.conj} $ t1 $ t2 => fold (do_formula pos) [t1, t2]
haftmann@38795
   159
      | @{const HOL.disj} $ t1 $ t2 => fold (do_formula pos) [t1, t2]
haftmann@38786
   160
      | @{const HOL.implies} $ t1 $ t2 =>
blanchet@37537
   161
        do_formula (flip pos) t1 #> do_formula pos t2
haftmann@38864
   162
      | Const (@{const_name HOL.eq}, Type (_, [T, _])) $ t1 $ t2 =>
blanchet@42741
   163
        do_term_or_formula false T t1 #> do_term_or_formula true T t2
blanchet@38587
   164
      | Const (@{const_name If}, Type (_, [_, Type (_, [T, _])]))
blanchet@38587
   165
        $ t1 $ t2 $ t3 =>
blanchet@42735
   166
        do_formula NONE t1 #> fold (do_term_or_formula false T) [t2, t3]
blanchet@38939
   167
      | Const (@{const_name Ex1}, _) $ Abs (_, T, t') =>
blanchet@38939
   168
        do_quantifier (is_some pos) T t'
blanchet@38939
   169
      | Const (@{const_name Ball}, _) $ t1 $ Abs (_, T, t') =>
blanchet@38939
   170
        do_quantifier (pos = SOME false) T
blanchet@38939
   171
                      (HOLogic.mk_imp (incr_boundvars 1 t1 $ Bound 0, t'))
blanchet@38939
   172
      | Const (@{const_name Bex}, _) $ t1 $ Abs (_, T, t') =>
blanchet@38939
   173
        do_quantifier (pos = SOME true) T
blanchet@38939
   174
                      (HOLogic.mk_conj (incr_boundvars 1 t1 $ Bound 0, t'))
blanchet@37537
   175
      | (t0 as Const (_, @{typ bool})) $ t1 =>
blanchet@42735
   176
        do_term false t0 #> do_formula pos t1  (* theory constant *)
blanchet@42735
   177
      | _ => do_term false t
blanchet@42732
   178
  in do_formula pos end
paulson@24287
   179
blanchet@48227
   180
fun pconsts_in_fact thy is_built_in_const t =
blanchet@48227
   181
  Symtab.fold (fn (s, pss) => fold (cons o pair s) pss)
blanchet@48227
   182
              (Symtab.empty |> add_pconsts_in_term thy is_built_in_const true
blanchet@48227
   183
                                                   (SOME true) t) []
blanchet@48227
   184
blanchet@48227
   185
val const_names_in_fact = map fst ooo pconsts_in_fact
blanchet@48227
   186
blanchet@48227
   187
(* Inserts a dummy "constant" referring to the theory name, so that relevance
blanchet@48227
   188
   takes the given theory into account. *)
blanchet@41200
   189
fun theory_constify ({theory_const_rel_weight, theory_const_irrel_weight, ...}
blanchet@41200
   190
                     : relevance_fudge) thy_name t =
blanchet@40070
   191
  if exists (curry (op <) 0.0) [theory_const_rel_weight,
blanchet@40070
   192
                                theory_const_irrel_weight] then
blanchet@41200
   193
    Const (thy_name ^ theory_const_suffix, @{typ bool}) $ t
blanchet@37505
   194
  else
blanchet@41200
   195
    t
blanchet@41200
   196
blanchet@41200
   197
fun theory_const_prop_of fudge th =
blanchet@41200
   198
  theory_constify fudge (Context.theory_name (theory_of_thm th)) (prop_of th)
blanchet@37505
   199
blanchet@48227
   200
fun pair_consts_fact thy is_built_in_const fudge fact =
blanchet@48227
   201
  case fact |> snd |> theory_const_prop_of fudge
blanchet@48227
   202
            |> pconsts_in_fact thy is_built_in_const of
blanchet@48227
   203
    [] => NONE
blanchet@48227
   204
  | consts => SOME ((fact, consts), NONE)
blanchet@48227
   205
blanchet@38743
   206
(* A two-dimensional symbol table counts frequencies of constants. It's keyed
blanchet@38743
   207
   first by constant name and second by its list of type instantiations. For the
blanchet@38823
   208
   latter, we need a linear ordering on "pattern list". *)
paulson@24287
   209
blanchet@38823
   210
fun pattern_ord p =
blanchet@38743
   211
  case p of
blanchet@38744
   212
    (PVar, PVar) => EQUAL
blanchet@38823
   213
  | (PVar, PApp _) => LESS
blanchet@38823
   214
  | (PApp _, PVar) => GREATER
blanchet@38823
   215
  | (PApp q1, PApp q2) =>
blanchet@38823
   216
    prod_ord fast_string_ord (dict_ord pattern_ord) (q1, q2)
blanchet@38939
   217
fun ptype_ord (PType p, PType q) =
blanchet@38939
   218
  prod_ord (dict_ord pattern_ord) int_ord (swap p, swap q)
paulson@24287
   219
blanchet@38939
   220
structure PType_Tab = Table(type key = ptype val ord = ptype_ord)
paulson@24287
   221
blanchet@40204
   222
fun count_fact_consts thy fudge =
blanchet@37503
   223
  let
blanchet@38827
   224
    fun do_const const (s, T) ts =
blanchet@38827
   225
      (* Two-dimensional table update. Constant maps to types maps to count. *)
blanchet@41204
   226
      PType_Tab.map_default (rich_ptype thy const (s, T), 0) (Integer.add 1)
blanchet@38939
   227
      |> Symtab.map_default (s, PType_Tab.empty)
blanchet@38827
   228
      #> fold do_term ts
blanchet@38827
   229
    and do_term t =
blanchet@38827
   230
      case strip_comb t of
blanchet@38827
   231
        (Const x, ts) => do_const true x ts
blanchet@38827
   232
      | (Free x, ts) => do_const false x ts
blanchet@38827
   233
      | (Abs (_, _, t'), ts) => fold do_term (t' :: ts)
blanchet@38827
   234
      | (_, ts) => fold do_term ts
blanchet@40070
   235
  in do_term o theory_const_prop_of fudge o snd end
paulson@24287
   236
blanchet@39367
   237
fun pow_int _ 0 = 1.0
blanchet@38939
   238
  | pow_int x 1 = x
blanchet@38939
   239
  | pow_int x n = if n > 0 then x * pow_int x (n - 1) else pow_int x (n + 1) / x
blanchet@38939
   240
paulson@24287
   241
(*The frequency of a constant is the sum of those of all instances of its type.*)
blanchet@38824
   242
fun pconst_freq match const_tab (c, ps) =
blanchet@38939
   243
  PType_Tab.fold (fn (qs, m) => match (ps, qs) ? Integer.add m)
blanchet@38939
   244
                 (the (Symtab.lookup const_tab c)) 0
blanchet@38686
   245
paulson@24287
   246
blanchet@38085
   247
(* A surprising number of theorems contain only a few significant constants.
blanchet@38085
   248
   These include all induction rules, and other general theorems. *)
blanchet@37503
   249
blanchet@37503
   250
(* "log" seems best in practice. A constant function of one ignores the constant
blanchet@38938
   251
   frequencies. Rare constants give more points if they are relevant than less
blanchet@38938
   252
   rare ones. *)
blanchet@39367
   253
fun rel_weight_for _ freq = 1.0 + 2.0 / Math.ln (Real.fromInt freq + 1.0)
blanchet@38938
   254
blanchet@38938
   255
(* Irrelevant constants are treated differently. We associate lower penalties to
blanchet@38938
   256
   very rare constants and very common ones -- the former because they can't
blanchet@38938
   257
   lead to the inclusion of too many new facts, and the latter because they are
blanchet@38938
   258
   so common as to be of little interest. *)
blanchet@40070
   259
fun irrel_weight_for ({worse_irrel_freq, higher_order_irrel_weight, ...}
blanchet@40070
   260
                      : relevance_fudge) order freq =
blanchet@40070
   261
  let val (k, x) = worse_irrel_freq |> `Real.ceil in
blanchet@38939
   262
    (if freq < k then Math.ln (Real.fromInt (freq + 1)) / Math.ln x
blanchet@38939
   263
     else rel_weight_for order freq / rel_weight_for order k)
blanchet@40070
   264
    * pow_int higher_order_irrel_weight (order - 1)
blanchet@38938
   265
  end
blanchet@37503
   266
blanchet@41790
   267
fun multiplier_for_const_name local_const_multiplier s =
blanchet@41790
   268
  if String.isSubstring "." s then 1.0 else local_const_multiplier
blanchet@38821
   269
blanchet@41790
   270
(* Computes a constant's weight, as determined by its frequency. *)
blanchet@41790
   271
fun generic_pconst_weight local_const_multiplier abs_weight skolem_weight
blanchet@42735
   272
                          theory_const_weight chained_const_weight weight_for f
blanchet@42735
   273
                          const_tab chained_const_tab (c as (s, PType (m, _))) =
blanchet@48219
   274
  if s = pseudo_abs_name then
blanchet@41790
   275
    abs_weight
blanchet@48219
   276
  else if String.isPrefix pseudo_skolem_prefix s then
blanchet@41790
   277
    skolem_weight
blanchet@41790
   278
  else if String.isSuffix theory_const_suffix s then
blanchet@41790
   279
    theory_const_weight
blanchet@41790
   280
  else
blanchet@41790
   281
    multiplier_for_const_name local_const_multiplier s
blanchet@41790
   282
    * weight_for m (pconst_freq (match_ptype o f) const_tab c)
blanchet@42735
   283
    |> (if chained_const_weight < 1.0 andalso
blanchet@42735
   284
           pconst_hyper_mem I chained_const_tab c then
blanchet@42735
   285
          curry (op *) chained_const_weight
blanchet@42735
   286
        else
blanchet@42735
   287
          I)
blanchet@41790
   288
blanchet@41790
   289
fun rel_pconst_weight ({local_const_multiplier, abs_rel_weight,
blanchet@41790
   290
                        theory_const_rel_weight, ...} : relevance_fudge)
blanchet@41790
   291
                      const_tab =
blanchet@41790
   292
  generic_pconst_weight local_const_multiplier abs_rel_weight 0.0
blanchet@42735
   293
                        theory_const_rel_weight 0.0 rel_weight_for I const_tab
blanchet@42735
   294
                        Symtab.empty
blanchet@42735
   295
blanchet@41790
   296
fun irrel_pconst_weight (fudge as {local_const_multiplier, abs_irrel_weight,
blanchet@41790
   297
                                   skolem_irrel_weight,
blanchet@42735
   298
                                   theory_const_irrel_weight,
blanchet@42735
   299
                                   chained_const_irrel_weight, ...})
blanchet@42735
   300
                        const_tab chained_const_tab =
blanchet@41790
   301
  generic_pconst_weight local_const_multiplier abs_irrel_weight
blanchet@41790
   302
                        skolem_irrel_weight theory_const_irrel_weight
blanchet@42735
   303
                        chained_const_irrel_weight (irrel_weight_for fudge) swap
blanchet@42735
   304
                        const_tab chained_const_tab
paulson@24287
   305
blanchet@46340
   306
fun stature_bonus ({intro_bonus, ...} : relevance_fudge) (_, Intro) =
blanchet@46340
   307
    intro_bonus
blanchet@46340
   308
  | stature_bonus {elim_bonus, ...} (_, Elim) = elim_bonus
blanchet@46340
   309
  | stature_bonus {simp_bonus, ...} (_, Simp) = simp_bonus
blanchet@46340
   310
  | stature_bonus {local_bonus, ...} (Local, _) = local_bonus
blanchet@46340
   311
  | stature_bonus {assum_bonus, ...} (Assum, _) = assum_bonus
blanchet@46340
   312
  | stature_bonus {chained_bonus, ...} (Chained, _) = chained_bonus
blanchet@46340
   313
  | stature_bonus _ _ = 0.0
blanchet@38751
   314
blanchet@40418
   315
fun is_odd_const_name s =
blanchet@48219
   316
  s = pseudo_abs_name orelse String.isPrefix pseudo_skolem_prefix s orelse
blanchet@40418
   317
  String.isSuffix theory_const_suffix s
blanchet@40418
   318
blanchet@46340
   319
fun fact_weight fudge stature const_tab relevant_consts chained_consts
blanchet@46340
   320
                fact_consts =
blanchet@40204
   321
  case fact_consts |> List.partition (pconst_hyper_mem I relevant_consts)
blanchet@40204
   322
                   ||> filter_out (pconst_hyper_mem swap relevant_consts) of
blanchet@38827
   323
    ([], _) => 0.0
blanchet@38744
   324
  | (rel, irrel) =>
blanchet@40418
   325
    if forall (forall (is_odd_const_name o fst)) [rel, irrel] then
blanchet@40371
   326
      0.0
blanchet@40371
   327
    else
blanchet@40371
   328
      let
blanchet@42735
   329
        val irrel = irrel |> filter_out (pconst_mem swap rel)
blanchet@40371
   330
        val rel_weight =
blanchet@40371
   331
          0.0 |> fold (curry (op +) o rel_pconst_weight fudge const_tab) rel
blanchet@40371
   332
        val irrel_weight =
blanchet@46340
   333
          ~ (stature_bonus fudge stature)
blanchet@42735
   334
          |> fold (curry (op +)
blanchet@42735
   335
                   o irrel_pconst_weight fudge const_tab chained_consts) irrel
blanchet@40371
   336
        val res = rel_weight / (rel_weight + irrel_weight)
blanchet@40371
   337
      in if Real.isFinite res then res else 0.0 end
blanchet@38747
   338
blanchet@48293
   339
fun take_most_relevant ctxt max_facts remaining_max
blanchet@42728
   340
        ({max_imperfect, max_imperfect_exp, ...} : relevance_fudge)
blanchet@51004
   341
        (candidates : ((raw_fact * (string * ptype) list) * real) list) =
blanchet@38744
   342
  let
blanchet@38747
   343
    val max_imperfect =
blanchet@40070
   344
      Real.ceil (Math.pow (max_imperfect,
blanchet@38904
   345
                    Math.pow (Real.fromInt remaining_max
blanchet@48293
   346
                              / Real.fromInt max_facts, max_imperfect_exp)))
blanchet@38747
   347
    val (perfect, imperfect) =
blanchet@38889
   348
      candidates |> sort (Real.compare o swap o pairself snd)
blanchet@38889
   349
                 |> take_prefix (fn (_, w) => w > 0.99999)
blanchet@38747
   350
    val ((accepts, more_rejects), rejects) =
blanchet@38747
   351
      chop max_imperfect imperfect |>> append perfect |>> chop remaining_max
blanchet@38744
   352
  in
blanchet@42646
   353
    trace_msg ctxt (fn () =>
wenzelm@41491
   354
        "Actually passed (" ^ string_of_int (length accepts) ^ " of " ^
wenzelm@41491
   355
        string_of_int (length candidates) ^ "): " ^
blanchet@38889
   356
        (accepts |> map (fn ((((name, _), _), _), weight) =>
blanchet@38752
   357
                            name () ^ " [" ^ Real.toString weight ^ "]")
blanchet@38745
   358
                 |> commas));
blanchet@38747
   359
    (accepts, more_rejects @ rejects)
blanchet@38744
   360
  end
paulson@24287
   361
blanchet@46340
   362
fun if_empty_replace_with_scope thy is_built_in_const facts sc tab =
blanchet@38819
   363
  if Symtab.is_empty tab then
blanchet@42732
   364
    Symtab.empty
blanchet@42732
   365
    |> fold (add_pconsts_in_term thy is_built_in_const false (SOME false))
blanchet@46340
   366
            (map_filter (fn ((_, (sc', _)), th) =>
blanchet@46340
   367
                            if sc' = sc then SOME (prop_of th) else NONE) facts)
blanchet@38819
   368
  else
blanchet@38819
   369
    tab
blanchet@38819
   370
blanchet@42702
   371
fun consider_arities is_built_in_const th =
blanchet@41158
   372
  let
blanchet@41158
   373
    fun aux _ _ NONE = NONE
blanchet@41158
   374
      | aux t args (SOME tab) =
blanchet@41158
   375
        case t of
blanchet@41158
   376
          t1 $ t2 => SOME tab |> aux t1 (t2 :: args) |> aux t2 []
blanchet@41158
   377
        | Const (x as (s, _)) =>
blanchet@41336
   378
          (if is_built_in_const x args |> fst then
blanchet@41158
   379
             SOME tab
blanchet@41158
   380
           else case Symtab.lookup tab s of
blanchet@41158
   381
             NONE => SOME (Symtab.update (s, length args) tab)
blanchet@41158
   382
           | SOME n => if n = length args then SOME tab else NONE)
blanchet@41158
   383
        | _ => SOME tab
blanchet@41158
   384
  in aux (prop_of th) [] end
blanchet@41158
   385
blanchet@44785
   386
(* FIXME: This is currently only useful for polymorphic type encodings. *)
blanchet@42702
   387
fun could_benefit_from_ext is_built_in_const facts =
blanchet@42702
   388
  fold (consider_arities is_built_in_const o snd) facts (SOME Symtab.empty)
blanchet@41158
   389
  |> is_none
blanchet@41158
   390
blanchet@43492
   391
(* High enough so that it isn't wrongly considered as very relevant (e.g., for E
blanchet@43492
   392
   weights), but low enough so that it is unlikely to be truncated away if few
blanchet@43492
   393
   facts are included. *)
blanchet@43492
   394
val special_fact_index = 75
blanchet@43492
   395
blanchet@48293
   396
fun relevance_filter ctxt thres0 decay max_facts is_built_in_const
blanchet@48292
   397
        (fudge as {threshold_divisor, ridiculous_threshold, ...}) facts hyp_ts
blanchet@48292
   398
        concl_t =
blanchet@38739
   399
  let
wenzelm@42361
   400
    val thy = Proof_Context.theory_of ctxt
blanchet@40204
   401
    val const_tab = fold (count_fact_consts thy fudge) facts Symtab.empty
blanchet@42732
   402
    val add_pconsts = add_pconsts_in_term thy is_built_in_const false o SOME
blanchet@48292
   403
    val chained_ts =
blanchet@48292
   404
      facts |> map_filter (fn ((_, (Chained, _)), th) => SOME (prop_of th)
blanchet@48292
   405
                            | _ => NONE)
blanchet@42732
   406
    val chained_const_tab = Symtab.empty |> fold (add_pconsts true) chained_ts
blanchet@38819
   407
    val goal_const_tab =
blanchet@42732
   408
      Symtab.empty |> fold (add_pconsts true) hyp_ts
blanchet@42732
   409
                   |> add_pconsts false concl_t
blanchet@42732
   410
      |> (fn tab => if Symtab.is_empty tab then chained_const_tab else tab)
blanchet@46340
   411
      |> fold (if_empty_replace_with_scope thy is_built_in_const facts)
blanchet@38993
   412
              [Chained, Assum, Local]
blanchet@48288
   413
    fun iter j remaining_max thres rel_const_tab hopeless hopeful =
blanchet@38739
   414
      let
blanchet@40191
   415
        fun relevant [] _ [] =
blanchet@38747
   416
            (* Nothing has been added this iteration. *)
blanchet@48288
   417
            if j = 0 andalso thres >= ridiculous_threshold then
blanchet@38747
   418
              (* First iteration? Try again. *)
blanchet@48293
   419
              iter 0 max_facts (thres / threshold_divisor) rel_const_tab
blanchet@38747
   420
                   hopeless hopeful
blanchet@38744
   421
            else
blanchet@40191
   422
              []
blanchet@38889
   423
          | relevant candidates rejects [] =
blanchet@38739
   424
            let
blanchet@38747
   425
              val (accepts, more_rejects) =
blanchet@48293
   426
                take_most_relevant ctxt max_facts remaining_max fudge candidates
blanchet@38739
   427
              val rel_const_tab' =
blanchet@38745
   428
                rel_const_tab
blanchet@41066
   429
                |> fold (add_pconst_to_table false) (maps (snd o fst) accepts)
blanchet@38744
   430
              fun is_dirty (c, _) =
blanchet@38744
   431
                Symtab.lookup rel_const_tab' c <> Symtab.lookup rel_const_tab c
blanchet@38745
   432
              val (hopeful_rejects, hopeless_rejects) =
blanchet@38745
   433
                 (rejects @ hopeless, ([], []))
blanchet@38745
   434
                 |-> fold (fn (ax as (_, consts), old_weight) =>
blanchet@38745
   435
                              if exists is_dirty consts then
blanchet@38745
   436
                                apfst (cons (ax, NONE))
blanchet@38745
   437
                              else
blanchet@38745
   438
                                apsnd (cons (ax, old_weight)))
blanchet@38745
   439
                 |>> append (more_rejects
blanchet@38745
   440
                             |> map (fn (ax as (_, consts), old_weight) =>
blanchet@38745
   441
                                        (ax, if exists is_dirty consts then NONE
blanchet@38745
   442
                                             else SOME old_weight)))
blanchet@48288
   443
              val thres =
blanchet@48288
   444
                1.0 - (1.0 - thres)
blanchet@38822
   445
                      * Math.pow (decay, Real.fromInt (length accepts))
blanchet@38747
   446
              val remaining_max = remaining_max - length accepts
blanchet@38739
   447
            in
blanchet@42646
   448
              trace_msg ctxt (fn () => "New or updated constants: " ^
blanchet@38744
   449
                  commas (rel_const_tab' |> Symtab.dest
blanchet@38822
   450
                          |> subtract (op =) (rel_const_tab |> Symtab.dest)
blanchet@38827
   451
                          |> map string_for_hyper_pconst));
blanchet@38745
   452
              map (fst o fst) accepts @
blanchet@38747
   453
              (if remaining_max = 0 then
blanchet@40191
   454
                 []
blanchet@38745
   455
               else
blanchet@48288
   456
                 iter (j + 1) remaining_max thres rel_const_tab'
blanchet@38747
   457
                      hopeless_rejects hopeful_rejects)
blanchet@38739
   458
            end
blanchet@38889
   459
          | relevant candidates rejects
blanchet@46340
   460
                     (((ax as (((_, stature), _), fact_consts)), cached_weight)
blanchet@38747
   461
                      :: hopeful) =
blanchet@38739
   462
            let
blanchet@38739
   463
              val weight =
blanchet@38739
   464
                case cached_weight of
blanchet@38739
   465
                  SOME w => w
blanchet@46340
   466
                | NONE => fact_weight fudge stature const_tab rel_const_tab
blanchet@42732
   467
                                      chained_const_tab fact_consts
blanchet@38739
   468
            in
blanchet@48288
   469
              if weight >= thres then
blanchet@38889
   470
                relevant ((ax, weight) :: candidates) rejects hopeful
blanchet@38739
   471
              else
blanchet@38889
   472
                relevant candidates ((ax, weight) :: rejects) hopeful
blanchet@38739
   473
            end
blanchet@38739
   474
        in
blanchet@42646
   475
          trace_msg ctxt (fn () =>
blanchet@38744
   476
              "ITERATION " ^ string_of_int j ^ ": current threshold: " ^
blanchet@48288
   477
              Real.toString thres ^ ", constants: " ^
blanchet@38744
   478
              commas (rel_const_tab |> Symtab.dest
blanchet@38744
   479
                      |> filter (curry (op <>) [] o snd)
blanchet@38827
   480
                      |> map string_for_hyper_pconst));
blanchet@38889
   481
          relevant [] [] hopeful
blanchet@38739
   482
        end
blanchet@47933
   483
    fun uses_const s t =
blanchet@47933
   484
      fold_aterms (curry (fn (Const (s', _), false) => s' = s | (_, b) => b)) t
blanchet@47933
   485
                  false
blanchet@47933
   486
    fun uses_const_anywhere accepts s =
blanchet@47933
   487
      exists (uses_const s o prop_of o snd) accepts orelse
blanchet@47933
   488
      exists (uses_const s) (concl_t :: hyp_ts)
blanchet@47933
   489
    fun add_set_const_thms accepts =
blanchet@47933
   490
      exists (uses_const_anywhere accepts) set_consts ? append set_thms
blanchet@43492
   491
    fun insert_into_facts accepts [] = accepts
blanchet@43492
   492
      | insert_into_facts accepts ths =
blanchet@43492
   493
        let
blanchet@43492
   494
          val add = facts |> filter (member Thm.eq_thm_prop ths o snd)
blanchet@43492
   495
          val (bef, after) =
blanchet@43492
   496
            accepts |> filter_out (member Thm.eq_thm_prop ths o snd)
blanchet@48293
   497
                    |> take (max_facts - length add)
blanchet@43492
   498
                    |> chop special_fact_index
blanchet@43492
   499
        in bef @ add @ after end
blanchet@43492
   500
    fun insert_special_facts accepts =
blanchet@47939
   501
       (* FIXME: get rid of "ext" here once it is treated as a helper *)
blanchet@43066
   502
       [] |> could_benefit_from_ext is_built_in_const accepts ? cons @{thm ext}
blanchet@47933
   503
          |> add_set_const_thms accepts
blanchet@43492
   504
          |> insert_into_facts accepts
blanchet@38739
   505
  in
blanchet@40369
   506
    facts |> map_filter (pair_consts_fact thy is_built_in_const fudge)
blanchet@48293
   507
          |> iter 0 max_facts thres0 goal_const_tab []
blanchet@43492
   508
          |> insert_special_facts
blanchet@42646
   509
          |> tap (fn accepts => trace_msg ctxt (fn () =>
wenzelm@41491
   510
                      "Total relevant: " ^ string_of_int (length accepts)))
blanchet@38739
   511
  end
paulson@24287
   512
blanchet@48406
   513
fun mepo_suggested_facts ctxt
blanchet@48293
   514
        ({fact_thresholds = (thres0, thres1), ...} : params) prover
blanchet@48293
   515
        max_facts fudge hyp_ts concl_t facts =
blanchet@37538
   516
  let
wenzelm@42361
   517
    val thy = Proof_Context.theory_of ctxt
blanchet@48288
   518
    val is_built_in_const =
blanchet@48288
   519
      Sledgehammer_Provers.is_built_in_const_for_prover ctxt prover
blanchet@48288
   520
    val fudge =
blanchet@48288
   521
      case fudge of
blanchet@48288
   522
        SOME fudge => fudge
blanchet@48288
   523
      | NONE => Sledgehammer_Provers.relevance_fudge_for_prover ctxt prover
blanchet@48288
   524
    val decay = Math.pow ((1.0 - thres1) / (1.0 - thres0),
blanchet@48293
   525
                          1.0 / Real.fromInt (max_facts + 1))
blanchet@37538
   526
  in
blanchet@42646
   527
    trace_msg ctxt (fn () => "Considering " ^ string_of_int (length facts) ^
blanchet@42646
   528
                             " facts");
blanchet@48288
   529
    (if thres1 < 0.0 then
blanchet@40204
   530
       facts
blanchet@48288
   531
     else if thres0 > 1.0 orelse thres0 > thres1 then
blanchet@38739
   532
       []
blanchet@38739
   533
     else
blanchet@48293
   534
       relevance_filter ctxt thres0 decay max_facts is_built_in_const fudge
blanchet@48293
   535
           facts hyp_ts
blanchet@44625
   536
           (concl_t |> theory_constify fudge (Context.theory_name thy)))
blanchet@51004
   537
    |> map fact_of_raw_fact
blanchet@37538
   538
  end
immler@30536
   539
paulson@15347
   540
end;