src/HOL/Tools/ATP/atp_systems.ML
author blanchet
Wed Aug 18 17:09:05 2010 +0200 (2010-08-18)
changeset 38589 b03f8fe043ec
parent 38588 6a5b104f92cb
child 38596 f881b865dcf4
permissions -rw-r--r--
added "max_relevant_per_iter" option to Sledgehammer
blanchet@38047
     1
(*  Title:      HOL/Tools/ATP/atp_systems.ML
wenzelm@28592
     2
    Author:     Fabian Immler, TU Muenchen
blanchet@36371
     3
    Author:     Jasmin Blanchette, TU Muenchen
wenzelm@28592
     4
blanchet@36376
     5
Setup for supported ATPs.
wenzelm@28592
     6
*)
wenzelm@28592
     7
blanchet@36376
     8
signature ATP_SYSTEMS =
wenzelm@28592
     9
sig
blanchet@38023
    10
  datatype failure =
blanchet@38519
    11
    Unprovable | IncompleteUnprovable | CantConnect | TimedOut |
blanchet@38519
    12
    OutOfResources | SpassTooOld | VampireTooOld | NoPerl | NoLibwwwPerl |
blanchet@38519
    13
    MalformedInput | MalformedOutput | UnknownError
blanchet@38023
    14
blanchet@38023
    15
  type prover_config =
blanchet@38092
    16
    {exec: string * string,
blanchet@38092
    17
     required_execs: (string * string) list,
blanchet@38023
    18
     arguments: bool -> Time.time -> string,
blanchet@38023
    19
     proof_delims: (string * string) list,
blanchet@38023
    20
     known_failures: (failure * string) list,
blanchet@38589
    21
     default_max_relevant_per_iter: int,
blanchet@38589
    22
     default_theory_relevant: bool,
blanchet@38023
    23
     explicit_forall: bool}
blanchet@38023
    24
blanchet@38061
    25
  val string_for_failure : failure -> string
blanchet@38061
    26
  val known_failure_in_output :
blanchet@38061
    27
    string -> (failure * string) list -> failure option
blanchet@38023
    28
  val add_prover: string * prover_config -> theory -> theory
blanchet@38023
    29
  val get_prover: theory -> string -> prover_config
blanchet@38023
    30
  val available_atps: theory -> unit
blanchet@35867
    31
  val refresh_systems_on_tptp : unit -> unit
blanchet@36371
    32
  val default_atps_param_value : unit -> string
blanchet@35867
    33
  val setup : theory -> theory
wenzelm@28592
    34
end;
wenzelm@28592
    35
blanchet@36376
    36
structure ATP_Systems : ATP_SYSTEMS =
wenzelm@28592
    37
struct
wenzelm@28596
    38
blanchet@38023
    39
(* prover configuration *)
boehmes@32864
    40
blanchet@38023
    41
datatype failure =
blanchet@38454
    42
  Unprovable | IncompleteUnprovable | CantConnect | TimedOut | OutOfResources |
blanchet@38519
    43
  SpassTooOld | VampireTooOld | NoPerl | NoLibwwwPerl | MalformedInput |
blanchet@38454
    44
  MalformedOutput | UnknownError
boehmes@32864
    45
wenzelm@32941
    46
type prover_config =
blanchet@38092
    47
  {exec: string * string,
blanchet@38092
    48
   required_execs: (string * string) list,
blanchet@37514
    49
   arguments: bool -> Time.time -> string,
blanchet@36370
    50
   proof_delims: (string * string) list,
blanchet@36370
    51
   known_failures: (failure * string) list,
blanchet@38589
    52
   default_max_relevant_per_iter: int,
blanchet@38589
    53
   default_theory_relevant: bool,
blanchet@37994
    54
   explicit_forall: bool}
wenzelm@28596
    55
blanchet@38061
    56
val missing_message_tail =
blanchet@38061
    57
  " appears to be missing. You will need to install it if you want to run \
blanchet@38061
    58
  \ATPs remotely."
blanchet@38061
    59
blanchet@38061
    60
fun string_for_failure Unprovable = "The ATP problem is unprovable."
blanchet@38061
    61
  | string_for_failure IncompleteUnprovable =
blanchet@38061
    62
    "The ATP cannot prove the problem."
blanchet@38094
    63
  | string_for_failure CantConnect = "Can't connect to remote server."
blanchet@38061
    64
  | string_for_failure TimedOut = "Timed out."
blanchet@38061
    65
  | string_for_failure OutOfResources = "The ATP ran out of resources."
blanchet@38519
    66
  | string_for_failure SpassTooOld =
blanchet@38096
    67
    "Isabelle requires a more recent version of SPASS with support for the \
blanchet@38096
    68
    \TPTP syntax. To install it, download and extract the package \
blanchet@38096
    69
    \\"http://isabelle.in.tum.de/dist/contrib/spass-3.7.tar.gz\" and add the \
blanchet@38096
    70
    \\"spass-3.7\" directory's absolute path to " ^
blanchet@38061
    71
    quote (Path.implode (Path.expand (Path.appends
blanchet@38061
    72
               (Path.variable "ISABELLE_HOME_USER" ::
blanchet@38061
    73
                map Path.basic ["etc", "components"])))) ^
blanchet@38061
    74
    " on a line of its own."
blanchet@38519
    75
  | string_for_failure VampireTooOld =
blanchet@38454
    76
    "Isabelle requires a more recent version of Vampire. To install it, follow \
blanchet@38454
    77
    \the instructions from the Sledgehammer manual (\"isabelle doc\
blanchet@38454
    78
    \ sledgehammer\")."
blanchet@38061
    79
  | string_for_failure NoPerl = "Perl" ^ missing_message_tail
blanchet@38061
    80
  | string_for_failure NoLibwwwPerl =
blanchet@38061
    81
    "The Perl module \"libwww-perl\"" ^ missing_message_tail
blanchet@38061
    82
  | string_for_failure MalformedInput =
blanchet@38096
    83
    "The ATP problem is malformed. Please report this to the Isabelle \
blanchet@38096
    84
    \developers."
blanchet@38096
    85
  | string_for_failure MalformedOutput = "The ATP output is malformed."
blanchet@38096
    86
  | string_for_failure UnknownError = "An unknown ATP error occurred."
blanchet@38061
    87
blanchet@38061
    88
fun known_failure_in_output output =
blanchet@38061
    89
  find_first (fn (_, pattern) => String.isSubstring pattern output)
blanchet@38061
    90
  #> Option.map fst
blanchet@38061
    91
blanchet@38061
    92
val known_perl_failures =
blanchet@38094
    93
  [(CantConnect, "HTTP error"),
blanchet@38094
    94
   (NoPerl, "env: perl"),
blanchet@38065
    95
   (NoLibwwwPerl, "Can't locate HTTP")]
wenzelm@28596
    96
blanchet@38023
    97
(* named provers *)
blanchet@38001
    98
blanchet@38023
    99
structure Data = Theory_Data
blanchet@38023
   100
(
blanchet@38023
   101
  type T = (prover_config * stamp) Symtab.table
blanchet@38023
   102
  val empty = Symtab.empty
blanchet@38023
   103
  val extend = I
blanchet@38023
   104
  fun merge data : T = Symtab.merge (eq_snd op =) data
blanchet@38023
   105
    handle Symtab.DUP name => error ("Duplicate ATP: " ^ quote name ^ ".")
blanchet@38023
   106
)
blanchet@38017
   107
blanchet@38023
   108
fun add_prover (name, config) thy =
blanchet@38023
   109
  Data.map (Symtab.update_new (name, (config, stamp ()))) thy
blanchet@38023
   110
  handle Symtab.DUP name => error ("Duplicate ATP: " ^ quote name ^ ".")
blanchet@38017
   111
blanchet@38023
   112
fun get_prover thy name =
blanchet@38023
   113
  the (Symtab.lookup (Data.get thy) name) |> fst
blanchet@38023
   114
  handle Option.Option => error ("Unknown ATP: " ^ name ^ ".")
blanchet@37962
   115
blanchet@38023
   116
fun available_atps thy =
blanchet@38023
   117
  priority ("Available ATPs: " ^
blanchet@38023
   118
            commas (sort_strings (Symtab.keys (Data.get thy))) ^ ".")
boehmes@32864
   119
blanchet@38023
   120
fun available_atps thy =
blanchet@38023
   121
  priority ("Available ATPs: " ^
blanchet@38023
   122
            commas (sort_strings (Symtab.keys (Data.get thy))) ^ ".")
wenzelm@28596
   123
blanchet@36382
   124
fun to_generous_secs time = (Time.toMilliseconds time + 999) div 1000
blanchet@36142
   125
wenzelm@28596
   126
(* E prover *)
wenzelm@28596
   127
blanchet@36369
   128
val tstp_proof_delims =
blanchet@36369
   129
  ("# SZS output start CNFRefutation.", "# SZS output end CNFRefutation")
blanchet@36369
   130
blanchet@35969
   131
val e_config : prover_config =
blanchet@38092
   132
  {exec = ("E_HOME", "eproof"),
blanchet@38092
   133
   required_execs = [],
blanchet@37514
   134
   arguments = fn _ => fn timeout =>
blanchet@36382
   135
     "--tstp-in --tstp-out -l5 -xAutoDev -tAutoDev --silent --cpu-limit=" ^
blanchet@36382
   136
     string_of_int (to_generous_secs timeout),
blanchet@36369
   137
   proof_delims = [tstp_proof_delims],
blanchet@36265
   138
   known_failures =
blanchet@37995
   139
     [(Unprovable, "SZS status: CounterSatisfiable"),
blanchet@37995
   140
      (Unprovable, "SZS status CounterSatisfiable"),
blanchet@36370
   141
      (TimedOut, "Failure: Resource limit exceeded (time)"),
blanchet@36370
   142
      (TimedOut, "time limit exceeded"),
blanchet@36370
   143
      (OutOfResources,
blanchet@36370
   144
       "# Cannot determine problem status within resource limit"),
blanchet@36370
   145
      (OutOfResources, "SZS status: ResourceOut"),
blanchet@36370
   146
      (OutOfResources, "SZS status ResourceOut")],
blanchet@38589
   147
   default_max_relevant_per_iter = 50 (* FIXME *),
blanchet@38589
   148
   default_theory_relevant = false,
blanchet@37994
   149
   explicit_forall = false}
blanchet@38454
   150
blanchet@38023
   151
val e = ("e", e_config)
wenzelm@28596
   152
wenzelm@28596
   153
blanchet@36219
   154
(* The "-VarWeight=3" option helps the higher-order problems, probably by
blanchet@36219
   155
   counteracting the presence of "hAPP". *)
blanchet@37498
   156
val spass_config : prover_config =
blanchet@38092
   157
  {exec = ("ISABELLE_ATP", "scripts/spass"),
blanchet@38092
   158
   required_execs = [("SPASS_HOME", "SPASS")],
blanchet@37550
   159
   (* "div 2" accounts for the fact that SPASS is often run twice. *)
blanchet@37514
   160
   arguments = fn complete => fn timeout =>
blanchet@37962
   161
     ("-Auto -PGiven=0 -PProblem=0 -Splits=0 -FullRed=0 -DocProof \
blanchet@37550
   162
      \-VarWeight=3 -TimeLimit=" ^
blanchet@38094
   163
      string_of_int ((to_generous_secs timeout + 1) div 2))
blanchet@37514
   164
     |> not complete ? prefix "-SOS=1 ",
blanchet@36369
   165
   proof_delims = [("Here is a proof", "Formulae used in the proof")],
blanchet@36289
   166
   known_failures =
blanchet@38061
   167
     known_perl_failures @
blanchet@37413
   168
     [(IncompleteUnprovable, "SPASS beiseite: Completion found"),
blanchet@36370
   169
      (TimedOut, "SPASS beiseite: Ran out of time"),
blanchet@36965
   170
      (OutOfResources, "SPASS beiseite: Maximal number of loops exceeded"),
blanchet@37413
   171
      (MalformedInput, "Undefined symbol"),
blanchet@37414
   172
      (MalformedInput, "Free Variable"),
blanchet@38519
   173
      (SpassTooOld, "tptp2dfg")],
blanchet@38589
   174
   default_max_relevant_per_iter = 35 (* FIXME *),
blanchet@38589
   175
   default_theory_relevant = true,
blanchet@37994
   176
   explicit_forall = true}
blanchet@38454
   177
blanchet@38023
   178
val spass = ("spass", spass_config)
wenzelm@28596
   179
blanchet@38454
   180
blanchet@37509
   181
(* Vampire *)
blanchet@37509
   182
blanchet@37509
   183
val vampire_config : prover_config =
blanchet@38092
   184
  {exec = ("VAMPIRE_HOME", "vampire"),
blanchet@38092
   185
   required_execs = [],
blanchet@37514
   186
   arguments = fn _ => fn timeout =>
blanchet@38033
   187
     "--mode casc -t " ^ string_of_int (to_generous_secs timeout) ^
blanchet@38588
   188
     " --thanks Andrei --input_file",
blanchet@37509
   189
   proof_delims =
blanchet@37509
   190
     [("=========== Refutation ==========",
blanchet@37509
   191
       "======= End of refutation ======="),
blanchet@38033
   192
      ("% SZS output start Refutation", "% SZS output end Refutation"),
blanchet@38033
   193
      ("% SZS output start Proof", "% SZS output end Proof")],
blanchet@37509
   194
   known_failures =
blanchet@37509
   195
     [(Unprovable, "UNPROVABLE"),
blanchet@37509
   196
      (IncompleteUnprovable, "CANNOT PROVE"),
blanchet@38092
   197
      (TimedOut, "SZS status Timeout"),
blanchet@37509
   198
      (Unprovable, "Satisfiability detected"),
blanchet@38454
   199
      (OutOfResources, "Refutation not found"),
blanchet@38519
   200
      (VampireTooOld, "not a valid option")],
blanchet@38589
   201
   default_max_relevant_per_iter = 45 (* FIXME *),
blanchet@38589
   202
   default_theory_relevant = false,
blanchet@37994
   203
   explicit_forall = false}
blanchet@38454
   204
blanchet@38023
   205
val vampire = ("vampire", vampire_config)
blanchet@37509
   206
blanchet@38454
   207
blanchet@37509
   208
(* Remote prover invocation via SystemOnTPTP *)
wenzelm@28596
   209
blanchet@38061
   210
val systems = Synchronized.var "atp_systems" ([] : string list)
immler@31835
   211
immler@31835
   212
fun get_systems () =
blanchet@38061
   213
  case bash_output "\"$ISABELLE_ATP/scripts/remote_atp\" -w 2>&1" of
blanchet@36370
   214
    (answer, 0) => split_lines answer
blanchet@36370
   215
  | (answer, _) =>
blanchet@38065
   216
    error (case known_failure_in_output answer known_perl_failures of
blanchet@38065
   217
             SOME failure => string_for_failure failure
blanchet@38065
   218
           | NONE => perhaps (try (unsuffix "\n")) answer ^ ".")
immler@31835
   219
blanchet@35867
   220
fun refresh_systems_on_tptp () =
blanchet@37509
   221
  Synchronized.change systems (fn _ => get_systems ())
immler@31835
   222
blanchet@38589
   223
fun get_system prefix =
blanchet@38589
   224
  Synchronized.change_result systems
blanchet@38589
   225
      (fn systems => (if null systems then get_systems () else systems)
blanchet@38589
   226
                     |> `(find_first (String.isPrefix prefix)))
boehmes@32864
   227
wenzelm@32948
   228
fun the_system prefix =
boehmes@32864
   229
  (case get_system prefix of
blanchet@37509
   230
    NONE => error ("System " ^ quote prefix ^ " not available at SystemOnTPTP.")
wenzelm@32942
   231
  | SOME sys => sys);
immler@31835
   232
blanchet@38041
   233
fun remote_config atp_prefix
blanchet@38589
   234
        ({proof_delims, known_failures, default_max_relevant_per_iter,
blanchet@38589
   235
          default_theory_relevant, ...} : prover_config) : prover_config =
blanchet@38092
   236
  {exec = ("ISABELLE_ATP", "scripts/remote_atp"),
blanchet@38092
   237
   required_execs = [],
blanchet@37514
   238
   arguments = fn _ => fn timeout =>
blanchet@38041
   239
     " -t " ^ string_of_int (to_generous_secs timeout) ^ " -s " ^
blanchet@36382
   240
     the_system atp_prefix,
blanchet@36369
   241
   proof_delims = insert (op =) tstp_proof_delims proof_delims,
blanchet@38061
   242
   known_failures =
blanchet@38061
   243
     known_failures @ known_perl_failures @
blanchet@38094
   244
     [(TimedOut, "says Timeout")],
blanchet@38589
   245
   default_max_relevant_per_iter = default_max_relevant_per_iter,
blanchet@38589
   246
   default_theory_relevant = default_theory_relevant,
blanchet@38433
   247
   explicit_forall = true}
wenzelm@28596
   248
blanchet@38023
   249
val remote_name = prefix "remote_"
blanchet@38041
   250
fun remote_prover (name, config) atp_prefix =
blanchet@38041
   251
  (remote_name name, remote_config atp_prefix config)
blanchet@38023
   252
blanchet@38041
   253
val remote_e = remote_prover e "EP---"
blanchet@38041
   254
val remote_vampire = remote_prover vampire "Vampire---9"
wenzelm@28592
   255
blanchet@38454
   256
blanchet@38454
   257
(* Setup *)
blanchet@38454
   258
blanchet@38092
   259
fun is_installed ({exec, required_execs, ...} : prover_config) =
blanchet@38092
   260
  forall (curry (op <>) "" o getenv o fst) (exec :: required_execs)
blanchet@38041
   261
fun maybe_remote (name, config) =
blanchet@38041
   262
  name |> not (is_installed config) ? remote_name
blanchet@36371
   263
blanchet@36371
   264
fun default_atps_param_value () =
blanchet@38041
   265
  space_implode " " ([maybe_remote e] @
blanchet@38041
   266
                     (if is_installed (snd spass) then [fst spass] else []) @
blanchet@38041
   267
                     [remote_name (fst vampire)])
blanchet@36371
   268
blanchet@38041
   269
val provers = [e, spass, vampire, remote_e, remote_vampire]
blanchet@38023
   270
val setup = fold add_prover provers
blanchet@35867
   271
wenzelm@28592
   272
end;