src/HOL/Tools/Sledgehammer/sledgehammer_mash.ML
author blanchet
Fri, 20 Jul 2012 22:19:45 +0200
changeset 49397 641af72b0059
parent 49396 1b7d798460bb
child 49398 df75b2d7e26a
permissions -rw-r--r--
added possibility of running external MaSh commands asynchronously
     1 (*  Title:      HOL/Tools/Sledgehammer/sledgehammer_mash.ML
     2     Author:     Jasmin Blanchette, TU Muenchen
     3 
     4 Sledgehammer's machine-learning-based relevance filter (MaSh).
     5 *)
     6 
     7 signature SLEDGEHAMMER_MASH =
     8 sig
     9   type status = ATP_Problem_Generate.status
    10   type stature = ATP_Problem_Generate.stature
    11   type fact = Sledgehammer_Fact.fact
    12   type fact_override = Sledgehammer_Fact.fact_override
    13   type params = Sledgehammer_Provers.params
    14   type relevance_fudge = Sledgehammer_Provers.relevance_fudge
    15   type prover_result = Sledgehammer_Provers.prover_result
    16 
    17   val trace : bool Config.T
    18   val MaShN : string
    19   val mepoN : string
    20   val mashN : string
    21   val meshN : string
    22   val fact_filters : string list
    23   val escape_meta : string -> string
    24   val escape_metas : string list -> string
    25   val unescape_meta : string -> string
    26   val unescape_metas : string -> string list
    27   val extract_query : string -> string * string list
    28   val nickname_of : thm -> string
    29   val suggested_facts : string list -> ('a * thm) list -> ('a * thm) list
    30   val mesh_facts :
    31     int -> (('a * thm) list * ('a * thm) list) list -> ('a * thm) list
    32   val is_likely_tautology_or_too_meta : thm -> bool
    33   val theory_ord : theory * theory -> order
    34   val thm_ord : thm * thm -> order
    35   val features_of :
    36     Proof.context -> string -> theory -> status -> term list -> string list
    37   val isabelle_dependencies_of : unit Symtab.table -> thm -> string list
    38   val goal_of_thm : theory -> thm -> thm
    39   val run_prover_for_mash :
    40     Proof.context -> params -> string -> fact list -> thm -> prover_result
    41   val mash_CLEAR : Proof.context -> unit
    42   val mash_INIT :
    43     Proof.context -> bool
    44     -> (string * string list * string list * string list) list -> unit
    45   val mash_ADD :
    46     Proof.context -> bool
    47     -> (string * string list * string list * string list) list -> unit
    48   val mash_QUERY :
    49     Proof.context -> bool -> int -> string list * string list -> string list
    50   val mash_unlearn : Proof.context -> unit
    51   val mash_could_suggest_facts : unit -> bool
    52   val mash_can_suggest_facts : Proof.context -> bool
    53   val mash_suggest_facts :
    54     Proof.context -> params -> string -> int -> term list -> term
    55     -> ('a * thm) list -> ('a * thm) list * ('a * thm) list
    56   val mash_learn_thy :
    57     Proof.context -> params -> theory -> Time.time -> fact list -> string
    58   val mash_learn_proof :
    59     Proof.context -> params -> term -> ('a * thm) list -> thm list -> unit
    60   val relevant_facts :
    61     Proof.context -> params -> string -> int -> fact_override -> term list
    62     -> term -> fact list -> fact list
    63   val kill_learners : unit -> unit
    64   val running_learners : unit -> unit
    65 end;
    66 
    67 structure Sledgehammer_MaSh : SLEDGEHAMMER_MASH =
    68 struct
    69 
    70 open ATP_Util
    71 open ATP_Problem_Generate
    72 open Sledgehammer_Util
    73 open Sledgehammer_Fact
    74 open Sledgehammer_Provers
    75 open Sledgehammer_Minimize
    76 open Sledgehammer_MePo
    77 
    78 val trace =
    79   Attrib.setup_config_bool @{binding sledgehammer_mash_trace} (K false)
    80 fun trace_msg ctxt msg = if Config.get ctxt trace then tracing (msg ()) else ()
    81 
    82 val MaShN = "MaSh"
    83 
    84 val mepoN = "mepo"
    85 val mashN = "mash"
    86 val meshN = "mesh"
    87 
    88 val fact_filters = [meshN, mepoN, mashN]
    89 
    90 fun mash_home () = getenv "MASH_HOME"
    91 fun mash_state_dir () =
    92   getenv "ISABELLE_HOME_USER" ^ "/mash"
    93   |> tap (Isabelle_System.mkdir o Path.explode)
    94 fun mash_state_path () = mash_state_dir () ^ "/state" |> Path.explode
    95 
    96 
    97 (*** Isabelle helpers ***)
    98 
    99 fun meta_char c =
   100   if Char.isAlphaNum c orelse c = #"_" orelse c = #"." orelse c = #"(" orelse
   101      c = #")" orelse c = #"," then
   102     String.str c
   103   else
   104     (* fixed width, in case more digits follow *)
   105     "\\" ^ stringN_of_int 3 (Char.ord c)
   106 
   107 fun unmeta_chars accum [] = String.implode (rev accum)
   108   | unmeta_chars accum (#"\\" :: d1 :: d2 :: d3 :: cs) =
   109     (case Int.fromString (String.implode [d1, d2, d3]) of
   110        SOME n => unmeta_chars (Char.chr n :: accum) cs
   111      | NONE => "" (* error *))
   112   | unmeta_chars _ (#"\\" :: _) = "" (* error *)
   113   | unmeta_chars accum (c :: cs) = unmeta_chars (c :: accum) cs
   114 
   115 val escape_meta = String.translate meta_char
   116 val escape_metas = map escape_meta #> space_implode " "
   117 val unescape_meta = String.explode #> unmeta_chars []
   118 val unescape_metas =
   119   space_explode " " #> filter_out (curry (op =) "") #> map unescape_meta
   120 
   121 fun extract_query line =
   122   case space_explode ":" line of
   123     [goal_name, suggs] => (unescape_meta goal_name, unescape_metas suggs)
   124   | _ => ("", [])
   125 
   126 fun parent_of_local_thm th =
   127   let
   128     val thy = th |> Thm.theory_of_thm
   129     val facts = thy |> Global_Theory.facts_of
   130     val space = facts |> Facts.space_of
   131     fun id_of s = #id (Name_Space.the_entry space s)
   132     fun max_id (s', _) (s, id) =
   133       let val id' = id_of s' in if id > id' then (s, id) else (s', id') end
   134   in ("", ~1) |> Facts.fold_static max_id facts |> fst end
   135 
   136 val local_prefix = "local" ^ Long_Name.separator
   137 
   138 fun nickname_of th =
   139   let val hint = Thm.get_name_hint th in
   140     (* FIXME: There must be a better way to detect local facts. *)
   141     case try (unprefix local_prefix) hint of
   142       SOME suff =>
   143       parent_of_local_thm th ^ Long_Name.separator ^ Long_Name.separator ^ suff
   144     | NONE => hint
   145   end
   146 
   147 fun suggested_facts suggs facts =
   148   let
   149     fun add_fact (fact as (_, th)) = Symtab.default (nickname_of th, fact)
   150     val tab = Symtab.empty |> fold add_fact facts
   151   in map_filter (Symtab.lookup tab) suggs end
   152 
   153 (* Ad hoc score function roughly based on Blanchette's Ringberg 2011 data. *)
   154 fun score x = Math.pow (1.5, 15.5 - 0.05 * Real.fromInt x) + 15.0
   155 
   156 fun sum_sq_avg [] = 0
   157   | sum_sq_avg xs =
   158     Real.ceil (100000.0 * fold (curry (op +) o score) xs 0.0) div length xs
   159 
   160 fun mesh_facts max_facts [(selected, unknown)] =
   161     take max_facts selected @ take (max_facts - length selected) unknown
   162   | mesh_facts max_facts mess =
   163     let
   164       val mess = mess |> map (apfst (`length))
   165       val fact_eq = Thm.eq_thm o pairself snd
   166       fun score_in fact ((sel_len, sels), unks) =
   167         case find_index (curry fact_eq fact) sels of
   168           ~1 => (case find_index (curry fact_eq fact) unks of
   169                    ~1 => SOME sel_len
   170                  | _ => NONE)
   171         | j => SOME j
   172       fun score_of fact = mess |> map_filter (score_in fact) |> sum_sq_avg
   173       val facts = fold (union fact_eq o take max_facts o snd o fst) mess []
   174     in
   175       facts |> map (`score_of) |> sort (int_ord o swap o pairself fst)
   176             |> map snd |> take max_facts
   177     end
   178 
   179 val thy_feature_prefix = "y_"
   180 
   181 val thy_feature_name_of = prefix thy_feature_prefix
   182 val const_name_of = prefix const_prefix
   183 val type_name_of = prefix type_const_prefix
   184 val class_name_of = prefix class_prefix
   185 
   186 fun is_likely_tautology_or_too_meta th =
   187   let
   188     val is_boring_const = member (op =) atp_widely_irrelevant_consts
   189     fun is_boring_bool t =
   190       not (exists_Const (not o is_boring_const o fst) t) orelse
   191       exists_type (exists_subtype (curry (op =) @{typ prop})) t
   192     fun is_boring_prop (@{const Trueprop} $ t) = is_boring_bool t
   193       | is_boring_prop (@{const "==>"} $ t $ u) =
   194         is_boring_prop t andalso is_boring_prop u
   195       | is_boring_prop (Const (@{const_name all}, _) $ (Abs (_, _, t)) $ u) =
   196         is_boring_prop t andalso is_boring_prop u
   197       | is_boring_prop (Const (@{const_name "=="}, _) $ t $ u) =
   198         is_boring_bool t andalso is_boring_bool u
   199       | is_boring_prop _ = true
   200   in
   201     is_boring_prop (prop_of th) andalso not (Thm.eq_thm_prop (@{thm ext}, th))
   202   end
   203 
   204 fun theory_ord p =
   205   if Theory.eq_thy p then
   206     EQUAL
   207   else if Theory.subthy p then
   208     LESS
   209   else if Theory.subthy (swap p) then
   210     GREATER
   211   else case int_ord (pairself (length o Theory.ancestors_of) p) of
   212     EQUAL => string_ord (pairself Context.theory_name p)
   213   | order => order
   214 
   215 val thm_ord = theory_ord o pairself theory_of_thm
   216 
   217 val bad_types = [@{type_name prop}, @{type_name bool}, @{type_name fun}]
   218 
   219 fun interesting_terms_types_and_classes ctxt prover term_max_depth
   220                                         type_max_depth ts =
   221   let
   222     fun is_bad_const (x as (s, _)) args =
   223       member (op =) atp_logical_consts s orelse
   224       fst (is_built_in_const_for_prover ctxt prover x args)
   225     fun add_classes @{sort type} = I
   226       | add_classes S = union (op =) (map class_name_of S)
   227     fun do_add_type (Type (s, Ts)) =
   228         (not (member (op =) bad_types s) ? insert (op =) (type_name_of s))
   229         #> fold do_add_type Ts
   230       | do_add_type (TFree (_, S)) = add_classes S
   231       | do_add_type (TVar (_, S)) = add_classes S
   232     fun add_type T = type_max_depth >= 0 ? do_add_type T
   233     fun mk_app s args =
   234       if member (op <>) args "" then s ^ "(" ^ space_implode "," args ^ ")"
   235       else s
   236     fun patternify ~1 _ = ""
   237       | patternify depth t =
   238         case strip_comb t of
   239           (Const (s, _), args) =>
   240           mk_app (const_name_of s) (map (patternify (depth - 1)) args)
   241         | _ => ""
   242     fun add_term_patterns ~1 _ = I
   243       | add_term_patterns depth t =
   244         insert (op =) (patternify depth t)
   245         #> add_term_patterns (depth - 1) t
   246     val add_term = add_term_patterns term_max_depth
   247     fun add_patterns t =
   248       let val (head, args) = strip_comb t in
   249         (case head of
   250            Const (x as (_, T)) =>
   251            not (is_bad_const x args) ? (add_term t #> add_type T)
   252          | Free (_, T) => add_type T
   253          | Var (_, T) => add_type T
   254          | Abs (_, T, body) => add_type T #> add_patterns body
   255          | _ => I)
   256         #> fold add_patterns args
   257       end
   258   in [] |> fold add_patterns ts end
   259 
   260 fun is_exists (s, _) = (s = @{const_name Ex} orelse s = @{const_name Ex1})
   261 
   262 val term_max_depth = 1
   263 val type_max_depth = 1
   264 
   265 (* TODO: Generate type classes for types? *)
   266 fun features_of ctxt prover thy status ts =
   267   thy_feature_name_of (Context.theory_name thy) ::
   268   interesting_terms_types_and_classes ctxt prover term_max_depth type_max_depth
   269                                       ts
   270   |> forall is_lambda_free ts ? cons "no_lams"
   271   |> forall (not o exists_Const is_exists) ts ? cons "no_skos"
   272   |> (case status of
   273         General => I
   274       | Induction => cons "induction"
   275       | Intro => cons "intro"
   276       | Inductive => cons "inductive"
   277       | Elim => cons "elim"
   278       | Simp => cons "simp"
   279       | Def => cons "def")
   280 
   281 fun isabelle_dependencies_of all_facts = thms_in_proof (SOME all_facts)
   282 
   283 val freezeT = Type.legacy_freeze_type
   284 
   285 fun freeze (t $ u) = freeze t $ freeze u
   286   | freeze (Abs (s, T, t)) = Abs (s, freezeT T, freeze t)
   287   | freeze (Var ((s, _), T)) = Free (s, freezeT T)
   288   | freeze (Const (s, T)) = Const (s, freezeT T)
   289   | freeze (Free (s, T)) = Free (s, freezeT T)
   290   | freeze t = t
   291 
   292 fun goal_of_thm thy = prop_of #> freeze #> cterm_of thy #> Goal.init
   293 
   294 fun run_prover_for_mash ctxt params prover facts goal =
   295   let
   296     val problem =
   297       {state = Proof.init ctxt, goal = goal, subgoal = 1, subgoal_count = 1,
   298        facts = facts |> map (apfst (apfst (fn name => name ())))
   299                      |> map Untranslated_Fact}
   300     val prover = get_minimizing_prover ctxt Normal (K ()) prover
   301   in prover params (K (K (K ""))) problem end
   302 
   303 
   304 (*** Low-level communication with MaSh ***)
   305 
   306 fun write_file (xs, f) file =
   307   let val path = Path.explode file in
   308     File.write path "";
   309     xs |> chunk_list 500
   310        |> List.app (File.append path o space_implode "" o map f)
   311   end
   312 
   313 fun mash_info overlord =
   314   if overlord then (getenv "ISABELLE_HOME_USER", "")
   315   else (getenv "ISABELLE_TMP", serial_string ())
   316 
   317 fun and_rm_files overlord flags files =
   318   if overlord then
   319     ""
   320   else
   321     " && rm -f" ^ flags ^ " -- " ^
   322     space_implode " " (map File.shell_quote files)
   323 
   324 fun run_mash ctxt overlord (temp_dir, serial) async core =
   325   let
   326     val log_file = temp_dir ^ "/mash_log" ^ serial
   327     val err_file = temp_dir ^ "/mash_err" ^ serial
   328     val command =
   329       "(" ^ mash_home () ^ "/mash --quiet --outputDir " ^ mash_state_dir () ^
   330       " --log " ^ log_file ^ " " ^ core ^ ") 2>&1 > " ^ err_file ^
   331       and_rm_files overlord "" [log_file, err_file] ^
   332       (if async then " &" else "")
   333   in
   334     trace_msg ctxt (fn () =>
   335         (if async then "Launching " else "Running ") ^ command);
   336     write_file ([], K "") log_file;
   337     write_file ([], K "") err_file;
   338     Isabelle_System.bash command;
   339     if not async then trace_msg ctxt (K "Done") else ()
   340   end
   341 
   342 (* TODO: Eliminate code once "mash.py" handles sequences of ADD commands as fast
   343    as a single INIT. *)
   344 fun run_mash_init ctxt overlord write_access write_feats write_deps =
   345   let
   346     val info as (temp_dir, serial) = mash_info overlord
   347     val in_dir = temp_dir ^ "/mash_init" ^ serial
   348                  |> tap (Path.explode #> Isabelle_System.mkdir)
   349   in
   350     write_file write_access (in_dir ^ "/mash_accessibility");
   351     write_file write_feats (in_dir ^ "/mash_features");
   352     write_file write_deps (in_dir ^ "/mash_dependencies");
   353     run_mash ctxt overlord info true
   354              ("--init --inputDir " ^ in_dir ^
   355               and_rm_files overlord " -r" [in_dir])
   356   end
   357 
   358 fun run_mash_commands ctxt overlord async save max_suggs write_cmds read_suggs =
   359   let
   360     val info as (temp_dir, serial) = mash_info overlord
   361     val sugg_file = temp_dir ^ "/mash_suggs" ^ serial
   362     val cmd_file = temp_dir ^ "/mash_commands" ^ serial
   363   in
   364     write_file ([], K "") sugg_file;
   365     write_file write_cmds cmd_file;
   366     run_mash ctxt overlord info async
   367              ("--inputFile " ^ cmd_file ^ " --predictions " ^ sugg_file ^
   368               " --numberOfPredictions " ^ string_of_int max_suggs ^
   369               (if save then " --saveModel" else "") ^
   370               (and_rm_files overlord "" [sugg_file, cmd_file]));
   371     read_suggs (fn () => File.read_lines (Path.explode sugg_file))
   372   end
   373 
   374 fun str_of_update (name, parents, feats, deps) =
   375   "! " ^ escape_meta name ^ ": " ^ escape_metas parents ^ "; " ^
   376   escape_metas feats ^ "; " ^ escape_metas deps ^ "\n"
   377 
   378 fun str_of_query (parents, feats) =
   379   "? " ^ escape_metas parents ^ "; " ^ escape_metas feats
   380 
   381 fun init_str_of_update get (upd as (name, _, _, _)) =
   382   escape_meta name ^ ": " ^ escape_metas (get upd) ^ "\n"
   383 
   384 fun mash_CLEAR ctxt =
   385   let val path = mash_state_dir () |> Path.explode in
   386     trace_msg ctxt (K "MaSh CLEAR");
   387     File.fold_dir (fn file => fn () =>
   388                       File.rm (Path.append path (Path.basic file)))
   389                   path ()
   390   end
   391 
   392 fun mash_INIT ctxt _ [] = mash_CLEAR ctxt
   393   | mash_INIT ctxt overlord upds =
   394     (trace_msg ctxt (fn () => "MaSh INIT " ^
   395          elide_string 1000 (space_implode " " (map #1 upds)));
   396      run_mash_init ctxt overlord (upds, init_str_of_update #2)
   397                    (upds, init_str_of_update #3) (upds, init_str_of_update #4))
   398 
   399 fun mash_ADD _ _ [] = ()
   400   | mash_ADD ctxt overlord upds =
   401     (trace_msg ctxt (fn () => "MaSh ADD " ^
   402          elide_string 1000 (space_implode " " (map #1 upds)));
   403      run_mash_commands ctxt overlord true true 0 (upds, str_of_update) (K ()))
   404 
   405 fun mash_QUERY ctxt overlord max_suggs (query as (_, feats)) =
   406   (trace_msg ctxt (fn () => "MaSh QUERY " ^ space_implode " " feats);
   407    run_mash_commands ctxt overlord false false max_suggs
   408        ([query], str_of_query)
   409        (fn suggs => snd (extract_query (List.last (suggs ()))))
   410    handle List.Empty => [])
   411 
   412 
   413 (*** High-level communication with MaSh ***)
   414 
   415 fun try_graph ctxt when def f =
   416   f ()
   417   handle Graph.CYCLES (cycle :: _) =>
   418          (trace_msg ctxt (fn () =>
   419               "Cycle involving " ^ commas cycle ^ " when " ^ when); def)
   420        | Graph.UNDEF name =>
   421          (trace_msg ctxt (fn () =>
   422               "Unknown fact " ^ quote name ^ " when " ^ when); def)
   423 
   424 type mash_state =
   425   {thys : bool Symtab.table,
   426    fact_graph : unit Graph.T}
   427 
   428 val empty_state = {thys = Symtab.empty, fact_graph = Graph.empty}
   429 
   430 local
   431 
   432 fun mash_load _ (state as (true, _)) = state
   433   | mash_load ctxt _ =
   434     let val path = mash_state_path () in
   435       (true,
   436        case try File.read_lines path of
   437          SOME (comp_thys :: incomp_thys :: fact_lines) =>
   438          let
   439            fun add_thy comp thy = Symtab.update (thy, comp)
   440            fun add_edge_to name parent =
   441              Graph.default_node (parent, ())
   442              #> Graph.add_edge (parent, name)
   443            fun add_fact_line line =
   444              case extract_query line of
   445                ("", _) => I (* shouldn't happen *)
   446              | (name, parents) =>
   447                Graph.default_node (name, ())
   448                #> fold (add_edge_to name) parents
   449            val thys =
   450              Symtab.empty |> fold (add_thy true) (unescape_metas comp_thys)
   451                           |> fold (add_thy false) (unescape_metas incomp_thys)
   452            val fact_graph =
   453              try_graph ctxt "loading state" Graph.empty (fn () =>
   454                  Graph.empty |> fold add_fact_line fact_lines)
   455          in {thys = thys, fact_graph = fact_graph} end
   456        | _ => empty_state)
   457     end
   458 
   459 fun mash_save ({thys, fact_graph, ...} : mash_state) =
   460   let
   461     val path = mash_state_path ()
   462     val thys = Symtab.dest thys
   463     val line_for_thys = escape_metas o AList.find (op =) thys
   464     fun fact_line_for name parents =
   465       escape_meta name ^ ": " ^ escape_metas parents
   466     val append_fact = File.append path o suffix "\n" oo fact_line_for
   467   in
   468     File.write path (line_for_thys true ^ "\n" ^ line_for_thys false ^ "\n");
   469     Graph.fold (fn (name, ((), (parents, _))) => fn () =>
   470                    append_fact name (Graph.Keys.dest parents))
   471         fact_graph ()
   472   end
   473 
   474 val global_state =
   475   Synchronized.var "Sledgehammer_MaSh.global_state" (false, empty_state)
   476 
   477 in
   478 
   479 fun mash_map ctxt f =
   480   Synchronized.change global_state (mash_load ctxt ##> (f #> tap mash_save))
   481 
   482 fun mash_get ctxt =
   483   Synchronized.change_result global_state (mash_load ctxt #> `snd)
   484 
   485 fun mash_unlearn ctxt =
   486   Synchronized.change global_state (fn _ =>
   487       (mash_CLEAR ctxt; File.write (mash_state_path ()) "";
   488        (true, empty_state)))
   489 
   490 end
   491 
   492 fun mash_could_suggest_facts () = mash_home () <> ""
   493 fun mash_can_suggest_facts ctxt =
   494   not (Graph.is_empty (#fact_graph (mash_get ctxt)))
   495 
   496 fun parents_wrt_facts facts fact_graph =
   497   let
   498     val facts = [] |> fold (cons o nickname_of o snd) facts
   499     val tab = Symtab.empty |> fold (fn name => Symtab.update (name, ())) facts
   500     fun insert_not_seen seen name =
   501       not (member (op =) seen name) ? insert (op =) name
   502     fun parents_of _ parents [] = parents
   503       | parents_of seen parents (name :: names) =
   504         if Symtab.defined tab name then
   505           parents_of (name :: seen) (name :: parents) names
   506         else
   507           parents_of (name :: seen) parents
   508                      (Graph.Keys.fold (insert_not_seen seen)
   509                                       (Graph.imm_preds fact_graph name) names)
   510   in parents_of [] [] (Graph.maximals fact_graph) end
   511 
   512 (* Generate more suggestions than requested, because some might be thrown out
   513    later for various reasons and "meshing" gives better results with some
   514    slack. *)
   515 fun max_suggs_of max_facts = max_facts + Int.min (200, max_facts)
   516 
   517 fun is_fact_in_graph fact_graph (_, th) =
   518   can (Graph.get_node fact_graph) (nickname_of th)
   519 
   520 fun mash_suggest_facts ctxt ({overlord, ...} : params) prover max_facts hyp_ts
   521                        concl_t facts =
   522   let
   523     val thy = Proof_Context.theory_of ctxt
   524     val fact_graph = #fact_graph (mash_get ctxt)
   525     val parents = parents_wrt_facts facts fact_graph
   526     val feats = features_of ctxt prover thy General (concl_t :: hyp_ts)
   527     val suggs =
   528       if Graph.is_empty fact_graph then []
   529       else mash_QUERY ctxt overlord (max_suggs_of max_facts) (parents, feats)
   530     val selected = facts |> suggested_facts suggs
   531     val unknown = facts |> filter_out (is_fact_in_graph fact_graph)
   532   in (selected, unknown) end
   533 
   534 fun add_thys_for thy =
   535   let fun add comp thy = Symtab.update (Context.theory_name thy, comp) in
   536     add false thy #> fold (add true) (Theory.ancestors_of thy)
   537   end
   538 
   539 fun update_fact_graph ctxt (name, parents, feats, deps) (upds, graph) =
   540   let
   541     fun maybe_add_from from (accum as (parents, graph)) =
   542       try_graph ctxt "updating graph" accum (fn () =>
   543           (from :: parents, Graph.add_edge_acyclic (from, name) graph))
   544     val graph = graph |> Graph.default_node (name, ())
   545     val (parents, graph) = ([], graph) |> fold maybe_add_from parents
   546     val (deps, graph) = ([], graph) |> fold maybe_add_from deps
   547   in ((name, parents, feats, deps) :: upds, graph) end
   548 
   549 val pass1_learn_timeout_factor = 0.5
   550 
   551 (* Too many dependencies is a sign that a decision procedure is at work. There
   552    isn't much too learn from such proofs. *)
   553 val max_dependencies = 10
   554 
   555 (* The timeout is understood in a very slack fashion. *)
   556 fun mash_learn_thy ctxt ({provers, verbose, overlord, ...} : params) thy timeout
   557                    facts =
   558   let
   559     val timer = Timer.startRealTimer ()
   560     val prover = hd provers
   561     fun timed_out frac =
   562       Time.> (Timer.checkRealTimer timer, time_mult frac timeout)
   563     val {fact_graph, ...} = mash_get ctxt
   564     val new_facts =
   565       facts |> filter_out (is_fact_in_graph fact_graph)
   566             |> sort (thm_ord o pairself snd)
   567   in
   568     if null new_facts then
   569       ""
   570     else
   571       let
   572         val ths = facts |> map snd
   573         val all_names =
   574           ths |> filter_out is_likely_tautology_or_too_meta
   575               |> map (rpair () o nickname_of)
   576               |> Symtab.make
   577         fun trim_deps deps = if length deps > max_dependencies then [] else deps
   578         fun do_fact _ (accum as (_, true)) = accum
   579           | do_fact ((_, (_, status)), th) ((parents, upds), false) =
   580             let
   581               val name = nickname_of th
   582               val feats =
   583                 features_of ctxt prover (theory_of_thm th) status [prop_of th]
   584               val deps = isabelle_dependencies_of all_names th |> trim_deps
   585               val upd = (name, parents, feats, deps)
   586             in (([name], upd :: upds), timed_out pass1_learn_timeout_factor) end
   587         val parents = parents_wrt_facts facts fact_graph
   588         val ((_, upds), _) =
   589           ((parents, []), false) |> fold do_fact new_facts |>> apsnd rev
   590         val n = length upds
   591         fun trans {thys, fact_graph} =
   592           let
   593             val mash_INIT_or_ADD =
   594               if Graph.is_empty fact_graph then mash_INIT else mash_ADD
   595             val (upds, fact_graph) =
   596               ([], fact_graph) |> fold (update_fact_graph ctxt) upds
   597           in
   598             (mash_INIT_or_ADD ctxt overlord (rev upds);
   599              {thys = thys |> add_thys_for thy,
   600               fact_graph = fact_graph})
   601           end
   602       in
   603         mash_map ctxt trans;
   604         if verbose then
   605           "Processed " ^ string_of_int n ^ " proof" ^ plural_s n ^
   606           (if verbose then
   607              " in " ^ string_from_time (Timer.checkRealTimer timer)
   608            else
   609              "") ^ "."
   610         else
   611           ""
   612       end
   613   end
   614 
   615 fun mash_learn_proof ctxt ({provers, overlord, ...} : params) t facts used_ths =
   616   let
   617     val thy = Proof_Context.theory_of ctxt
   618     val prover = hd provers
   619     val name = ATP_Util.timestamp () ^ " " ^ serial_string () (* fresh enough *)
   620     val feats = features_of ctxt prover thy General [t]
   621     val deps = used_ths |> map nickname_of
   622   in
   623     mash_map ctxt (fn {thys, fact_graph} =>
   624         let
   625           val parents = parents_wrt_facts facts fact_graph
   626           val upds = [(name, parents, feats, deps)]
   627           val (upds, fact_graph) =
   628             ([], fact_graph) |> fold (update_fact_graph ctxt) upds
   629         in
   630           mash_ADD ctxt overlord upds;
   631           {thys = thys, fact_graph = fact_graph}
   632         end)
   633   end
   634 
   635 (* The threshold should be large enough so that MaSh doesn't kick in for Auto
   636    Sledgehammer and Try. *)
   637 val min_secs_for_learning = 15
   638 val learn_timeout_factor = 2.0
   639 
   640 fun relevant_facts ctxt (params as {learn, fact_filter, timeout, ...}) prover
   641         max_facts ({add, only, ...} : fact_override) hyp_ts concl_t facts =
   642   if not (subset (op =) (the_list fact_filter, fact_filters)) then
   643     error ("Unknown fact filter: " ^ quote (the fact_filter) ^ ".")
   644   else if only then
   645     facts
   646   else if max_facts <= 0 orelse null facts then
   647     []
   648   else
   649     let
   650       val thy = Proof_Context.theory_of ctxt
   651       fun maybe_learn () =
   652         if not learn orelse Async_Manager.has_running_threads MaShN then
   653           ()
   654         else if Time.toSeconds timeout >= min_secs_for_learning then
   655           let
   656             val soft_timeout = time_mult learn_timeout_factor timeout
   657             val hard_timeout = time_mult 4.0 soft_timeout
   658             val birth_time = Time.now ()
   659             val death_time = Time.+ (birth_time, hard_timeout)
   660             val desc = ("machine learner for Sledgehammer", "")
   661           in
   662             Async_Manager.launch MaShN birth_time death_time desc
   663                 (fn () =>
   664                     (true, mash_learn_thy ctxt params thy soft_timeout facts))
   665           end
   666         else
   667           ()
   668       val fact_filter =
   669         case fact_filter of
   670           SOME ff => (() |> ff <> mepoN ? maybe_learn; ff)
   671         | NONE =>
   672           if mash_can_suggest_facts ctxt then (maybe_learn (); meshN)
   673           else if mash_could_suggest_facts () then (maybe_learn (); mepoN)
   674           else mepoN
   675       val add_ths = Attrib.eval_thms ctxt add
   676       fun prepend_facts ths accepts =
   677         ((facts |> filter (member Thm.eq_thm_prop ths o snd)) @
   678          (accepts |> filter_out (member Thm.eq_thm_prop ths o snd)))
   679         |> take max_facts
   680       fun iter () =
   681         iterative_relevant_facts ctxt params prover max_facts NONE hyp_ts
   682                                  concl_t facts
   683       fun mash () =
   684         mash_suggest_facts ctxt params prover max_facts hyp_ts concl_t facts
   685       val mess =
   686         [] |> (if fact_filter <> mashN then cons (iter (), []) else I)
   687            |> (if fact_filter <> mepoN then cons (mash ()) else I)
   688     in
   689       mesh_facts max_facts mess
   690       |> not (null add_ths) ? prepend_facts add_ths
   691     end
   692 
   693 fun kill_learners () = Async_Manager.kill_threads MaShN "learner"
   694 fun running_learners () = Async_Manager.running_threads MaShN "learner"
   695 
   696 end;