src/Tools/quickcheck.ML
author bulwahn
Fri Nov 05 08:16:35 2010 +0100 (2010-11-05)
changeset 40366 a2866dbfbe6b
parent 40253 f99ec71de82d
child 40381 96c37a685a13
permissions -rw-r--r--
changing timeout to real value; handling Interrupt and Timeout more like nitpick does
wenzelm@30824
     1
(*  Title:      Tools/quickcheck.ML
bulwahn@40246
     2
    Author:     Stefan Berghofer, Florian Haftmann, Lukas Bulwahn, TU Muenchen
haftmann@28256
     3
haftmann@28256
     4
Generic counterexample search engine.
haftmann@28256
     5
*)
haftmann@28256
     6
haftmann@28256
     7
signature QUICKCHECK =
haftmann@28256
     8
sig
bulwahn@37910
     9
  val setup: theory -> theory
bulwahn@37910
    10
  (* configuration *)
wenzelm@32740
    11
  val auto: bool Unsynchronized.ref
bulwahn@34948
    12
  val timing : bool Unsynchronized.ref
bulwahn@35378
    13
  datatype report = Report of
bulwahn@35378
    14
    { iterations : int, raised_match_errors : int,
bulwahn@35378
    15
      satisfied_assms : int list, positive_concl_tests : int }
bulwahn@37929
    16
  datatype expectation = No_Expectation | No_Counterexample | Counterexample;
bulwahn@37910
    17
  datatype test_params = Test_Params of
bulwahn@37929
    18
  { size: int, iterations: int, default_type: typ list, no_assms: bool,
bulwahn@40366
    19
    expect : expectation, report: bool, quiet : bool, timeout : real};
bulwahn@40246
    20
  val test_params_of : Proof.context -> test_params
bulwahn@39253
    21
  val report : Proof.context -> bool
bulwahn@40246
    22
  val map_test_params :
bulwahn@40366
    23
    ((int * int) * ((typ list * bool) * ((expectation * bool) * (bool * real)))
bulwahn@40366
    24
      -> (int * int) * ((typ list * bool) * ((expectation * bool) * (bool * real))))
bulwahn@40246
    25
    -> Context.generic -> Context.generic
bulwahn@37910
    26
  val add_generator:
bulwahn@39253
    27
    string * (Proof.context -> term -> int -> term list option * (bool list * bool))
bulwahn@39252
    28
      -> Context.generic -> Context.generic
bulwahn@37910
    29
  (* testing terms and proof states *)
bulwahn@40246
    30
  val test_term: Proof.context -> string option -> term ->
bulwahn@35380
    31
    (string * term) list option * ((string * int) list * ((int * report list) list) option)
bulwahn@37909
    32
  val quickcheck: (string * string list) list -> int -> Proof.state -> (string * term) list option
haftmann@28256
    33
end;
haftmann@28256
    34
haftmann@28256
    35
structure Quickcheck : QUICKCHECK =
haftmann@28256
    36
struct
haftmann@28256
    37
wenzelm@30980
    38
(* preferences *)
wenzelm@30980
    39
wenzelm@32740
    40
val auto = Unsynchronized.ref false;
wenzelm@30980
    41
bulwahn@34948
    42
val timing = Unsynchronized.ref false;
bulwahn@34948
    43
wenzelm@30980
    44
val _ =
wenzelm@30980
    45
  ProofGeneralPgip.add_preference Preferences.category_tracing
wenzelm@39616
    46
  (Unsynchronized.setmp auto true (fn () =>
wenzelm@30980
    47
    Preferences.bool_pref auto
wenzelm@30980
    48
      "auto-quickcheck"
blanchet@39329
    49
      "Run Quickcheck automatically.") ());
wenzelm@30980
    50
bulwahn@35378
    51
(* quickcheck report *)
bulwahn@35378
    52
bulwahn@35378
    53
datatype single_report = Run of bool list * bool | MatchExc
bulwahn@35378
    54
bulwahn@35378
    55
datatype report = Report of
bulwahn@35378
    56
  { iterations : int, raised_match_errors : int,
bulwahn@35378
    57
    satisfied_assms : int list, positive_concl_tests : int }
bulwahn@35378
    58
bulwahn@35378
    59
fun collect_single_report single_report
bulwahn@35378
    60
    (Report {iterations = iterations, raised_match_errors = raised_match_errors,
bulwahn@35378
    61
    satisfied_assms = satisfied_assms, positive_concl_tests = positive_concl_tests}) =
bulwahn@35378
    62
  case single_report
bulwahn@35378
    63
  of MatchExc =>
bulwahn@35378
    64
    Report {iterations = iterations + 1, raised_match_errors = raised_match_errors + 1,
bulwahn@35378
    65
      satisfied_assms = satisfied_assms, positive_concl_tests = positive_concl_tests}
bulwahn@35378
    66
   | Run (assms, concl) =>
bulwahn@35378
    67
    Report {iterations = iterations + 1, raised_match_errors = raised_match_errors,
bulwahn@35378
    68
      satisfied_assms =
bulwahn@35378
    69
        map2 (fn b => fn s => if b then s + 1 else s) assms
bulwahn@35378
    70
         (if null satisfied_assms then replicate (length assms) 0 else satisfied_assms),
bulwahn@35378
    71
      positive_concl_tests = if concl then positive_concl_tests + 1 else positive_concl_tests}
haftmann@30973
    72
bulwahn@37929
    73
(* expectation *)
bulwahn@37929
    74
bulwahn@37929
    75
datatype expectation = No_Expectation | No_Counterexample | Counterexample; 
bulwahn@37929
    76
bulwahn@37929
    77
fun merge_expectation (expect1, expect2) =
bulwahn@37929
    78
  if expect1 = expect2 then expect1 else No_Expectation
bulwahn@37929
    79
haftmann@28315
    80
(* quickcheck configuration -- default parameters, test generators *)
haftmann@28315
    81
haftmann@28309
    82
datatype test_params = Test_Params of
bulwahn@37929
    83
  { size: int, iterations: int, default_type: typ list, no_assms: bool,
bulwahn@40366
    84
    expect : expectation, report: bool, quiet : bool, timeout : real};
haftmann@28309
    85
bulwahn@40246
    86
fun dest_test_params
bulwahn@40246
    87
    (Test_Params { size, iterations, default_type, no_assms, expect, report, quiet, timeout }) =
bulwahn@40246
    88
  ((size, iterations), ((default_type, no_assms), ((expect, report), (quiet, timeout))));
wenzelm@38759
    89
bulwahn@40246
    90
fun make_test_params
bulwahn@40246
    91
    ((size, iterations), ((default_type, no_assms), ((expect, report), (quiet, timeout)))) =
blanchet@34128
    92
  Test_Params { size = size, iterations = iterations, default_type = default_type,
bulwahn@40246
    93
    no_assms = no_assms, expect = expect, report = report, quiet = quiet, timeout = timeout };
wenzelm@38759
    94
bulwahn@40246
    95
fun map_test_params' f
bulwahn@40246
    96
    (Test_Params { size, iterations, default_type, no_assms, expect, report, quiet, timeout }) =
bulwahn@40246
    97
  make_test_params
bulwahn@40246
    98
    (f ((size, iterations), ((default_type, no_assms), ((expect, report), (quiet, timeout)))));
wenzelm@38759
    99
wenzelm@38759
   100
fun merge_test_params
wenzelm@38759
   101
 (Test_Params { size = size1, iterations = iterations1, default_type = default_type1,
bulwahn@40246
   102
    no_assms = no_assms1, expect = expect1, report = report1, quiet = quiet1, timeout = timeout1 },
blanchet@34128
   103
  Test_Params { size = size2, iterations = iterations2, default_type = default_type2,
bulwahn@40246
   104
    no_assms = no_assms2, expect = expect2, report = report2, quiet = quiet2, timeout = timeout2}) =
haftmann@31599
   105
  make_test_params ((Int.max (size1, size2), Int.max (iterations1, iterations2)),
bulwahn@37911
   106
    ((merge (op =) (default_type1, default_type2), no_assms1 orelse no_assms2),
bulwahn@40246
   107
    ((merge_expectation (expect1, expect2), report1 orelse report2),
bulwahn@40366
   108
    (quiet1 orelse quiet2, Real.min (timeout1, timeout2)))));
haftmann@28309
   109
bulwahn@39252
   110
structure Data = Generic_Data
wenzelm@33522
   111
(
wenzelm@38759
   112
  type T =
bulwahn@39253
   113
    (string * (Proof.context -> term -> int -> term list option * (bool list * bool))) list
wenzelm@38759
   114
      * test_params;
bulwahn@35378
   115
  val empty = ([], Test_Params
bulwahn@37929
   116
    { size = 10, iterations = 100, default_type = [], no_assms = false,
bulwahn@40366
   117
      expect = No_Expectation, report = false, quiet = false, timeout = 30.0});
haftmann@28256
   118
  val extend = I;
wenzelm@33522
   119
  fun merge ((generators1, params1), (generators2, params2)) : T =
wenzelm@33522
   120
    (AList.merge (op =) (K true) (generators1, generators2),
haftmann@28309
   121
      merge_test_params (params1, params2));
wenzelm@33522
   122
);
haftmann@28256
   123
bulwahn@39252
   124
val test_params_of = snd o Data.get o Context.Proof;
bulwahn@37910
   125
bulwahn@40246
   126
val size = fst o fst o dest_test_params o test_params_of
bulwahn@40246
   127
bulwahn@40246
   128
val iterations = snd o fst o dest_test_params o test_params_of
bulwahn@40246
   129
bulwahn@40246
   130
val default_type = fst o fst o snd o dest_test_params o test_params_of
bulwahn@40246
   131
bulwahn@40246
   132
val no_assms = snd o fst o snd o dest_test_params o test_params_of
bulwahn@40246
   133
bulwahn@40246
   134
val expect = fst o fst o snd o snd o dest_test_params o test_params_of
bulwahn@40246
   135
bulwahn@39253
   136
val report = snd o fst o snd o snd o dest_test_params o test_params_of
bulwahn@39253
   137
bulwahn@40246
   138
val quiet = fst o snd o snd o snd o dest_test_params o test_params_of
bulwahn@40246
   139
bulwahn@40246
   140
val timeout = snd o snd o snd o snd o dest_test_params o test_params_of
bulwahn@40246
   141
bulwahn@40246
   142
fun map_report f
bulwahn@40246
   143
  (Test_Params { size, iterations, default_type, no_assms, expect, report, quiet, timeout }) =
bulwahn@40246
   144
    make_test_params
bulwahn@40246
   145
      ((size, iterations), ((default_type, no_assms), ((expect, f report), (quiet, timeout))));
bulwahn@39253
   146
bulwahn@40246
   147
fun map_quiet f
bulwahn@40246
   148
  (Test_Params { size, iterations, default_type, no_assms, expect, report, quiet, timeout }) =
bulwahn@40246
   149
    make_test_params
bulwahn@40246
   150
      ((size, iterations), ((default_type, no_assms), ((expect, report), (f quiet, timeout))));
bulwahn@40246
   151
    
bulwahn@40246
   152
fun set_report report = Data.map (apsnd (map_report (K report)))
bulwahn@40246
   153
bulwahn@40246
   154
fun set_quiet quiet = Data.map (apsnd (map_quiet (K quiet)))
bulwahn@40246
   155
bulwahn@40246
   156
val map_test_params = Data.map o apsnd o map_test_params'
bulwahn@39253
   157
haftmann@28309
   158
val add_generator = Data.map o apfst o AList.update (op =);
haftmann@28309
   159
haftmann@28315
   160
(* generating tests *)
haftmann@28315
   161
haftmann@28309
   162
fun mk_tester_select name ctxt =
bulwahn@39252
   163
  case AList.lookup (op =) ((fst o Data.get o Context.Proof) ctxt) name
haftmann@28309
   164
   of NONE => error ("No such quickcheck generator: " ^ name)
haftmann@28309
   165
    | SOME generator => generator ctxt;
haftmann@28309
   166
bulwahn@39253
   167
fun mk_testers ctxt t =
bulwahn@39252
   168
  (map snd o fst o Data.get o Context.Proof) ctxt
bulwahn@39253
   169
  |> map_filter (fn generator => try (generator ctxt) t);
haftmann@28309
   170
bulwahn@39253
   171
fun mk_testers_strict ctxt t =
haftmann@28309
   172
  let
wenzelm@40235
   173
    val generators = ((map snd o fst o Data.get o Context.Proof) ctxt)
wenzelm@40235
   174
    val testers = map (fn generator => Exn.interruptible_capture (generator ctxt) t) generators;
wenzelm@40235
   175
  in
wenzelm@40235
   176
    if forall (is_none o Exn.get_result) testers
haftmann@28309
   177
    then [(Exn.release o snd o split_last) testers]
haftmann@28309
   178
    else map_filter Exn.get_result testers
haftmann@28309
   179
  end;
haftmann@28309
   180
haftmann@28315
   181
haftmann@28315
   182
(* testing propositions *)
haftmann@28315
   183
haftmann@28309
   184
fun prep_test_term t =
haftmann@28309
   185
  let
wenzelm@29266
   186
    val _ = (null (Term.add_tvars t []) andalso null (Term.add_tfrees t [])) orelse
haftmann@28309
   187
      error "Term to be tested contains type variables";
wenzelm@29266
   188
    val _ = null (Term.add_vars t []) orelse
haftmann@28309
   189
      error "Term to be tested contains schematic variables";
haftmann@31138
   190
    val frees = Term.add_frees t [];
haftmann@28309
   191
  in (map fst frees, list_abs_free (frees, t)) end
haftmann@28256
   192
bulwahn@35324
   193
fun cpu_time description f =
bulwahn@35324
   194
  let
bulwahn@35324
   195
    val start = start_timing ()
bulwahn@35324
   196
    val result = Exn.capture f ()
bulwahn@35324
   197
    val time = Time.toMilliseconds (#cpu (end_timing start))
bulwahn@35324
   198
  in (Exn.release result, (description, time)) end
bulwahn@35324
   199
bulwahn@40246
   200
fun test_term ctxt generator_name t =
haftmann@28309
   201
  let
haftmann@28309
   202
    val (names, t') = prep_test_term t;
bulwahn@40366
   203
    val current_size = Unsynchronized.ref 0
bulwahn@40366
   204
    fun excipit s =
bulwahn@40366
   205
      "Quickcheck " ^ s ^ " while testing at size " ^ string_of_int (!current_size)
bulwahn@35324
   206
    val (testers, comp_time) = cpu_time "quickcheck compilation"
bulwahn@35324
   207
      (fn () => (case generator_name
bulwahn@40246
   208
       of NONE => if quiet ctxt then mk_testers ctxt t' else mk_testers_strict ctxt t'
bulwahn@39253
   209
        | SOME name => [mk_tester_select name ctxt t']));
bulwahn@35378
   210
    fun iterate f 0 report = (NONE, report)
bulwahn@35378
   211
      | iterate f j report =
bulwahn@35378
   212
        let
bulwahn@40246
   213
          val (test_result, single_report) = apsnd Run (f ()) handle Match => (if quiet ctxt then ()
bulwahn@35378
   214
             else warning "Exception Match raised during quickcheck"; (NONE, MatchExc))
bulwahn@35378
   215
          val report = collect_single_report single_report report
bulwahn@35378
   216
        in
bulwahn@35378
   217
          case test_result of NONE => iterate f (j - 1) report | SOME q => (SOME q, report)
bulwahn@35378
   218
        end
bulwahn@35378
   219
    val empty_report = Report { iterations = 0, raised_match_errors = 0,
bulwahn@35378
   220
      satisfied_assms = [], positive_concl_tests = 0 }
bulwahn@35378
   221
    fun with_testers k [] = (NONE, [])
haftmann@28309
   222
      | with_testers k (tester :: testers) =
bulwahn@40246
   223
          case iterate (fn () => tester (k - 1)) (iterations ctxt) empty_report
bulwahn@35378
   224
           of (NONE, report) => apsnd (cons report) (with_testers k testers)
bulwahn@35378
   225
            | (SOME q, report) => (SOME q, [report]);
wenzelm@40132
   226
    fun with_size k reports =
bulwahn@40246
   227
      if k > size ctxt then (NONE, reports)
wenzelm@40132
   228
      else
bulwahn@40246
   229
       (if quiet ctxt then () else Output.urgent_message ("Test data size: " ^ string_of_int k);
bulwahn@35378
   230
        let
bulwahn@40366
   231
          val _ = current_size := k  
bulwahn@35378
   232
          val (result, new_report) = with_testers k testers
bulwahn@35378
   233
          val reports = ((k, new_report) :: reports)
bulwahn@35378
   234
        in case result of NONE => with_size (k + 1) reports | SOME q => (SOME q, reports) end);
bulwahn@40136
   235
    val ((result, reports), exec_time) =
bulwahn@40366
   236
      TimeLimit.timeLimit (seconds (timeout ctxt)) (fn () => cpu_time "quickcheck execution"
bulwahn@35378
   237
      (fn () => apfst
bulwahn@35378
   238
         (fn result => case result of NONE => NONE
bulwahn@40136
   239
        | SOME ts => SOME (names ~~ ts)) (with_size 1 []))) ()
bulwahn@40366
   240
      handle TimeLimit.TimeOut =>
bulwahn@40366
   241
        error (excipit "ran out of time")
bulwahn@40366
   242
     | Exn.Interrupt => error (excipit "was interrupted")
bulwahn@34948
   243
  in
bulwahn@39253
   244
    (result, ([exec_time, comp_time], if report ctxt then SOME reports else NONE))
haftmann@28309
   245
  end;
haftmann@28309
   246
bulwahn@37913
   247
exception WELLSORTED of string
bulwahn@37913
   248
haftmann@28309
   249
fun monomorphic_term thy insts default_T = 
haftmann@28309
   250
  let
haftmann@28309
   251
    fun subst (T as TFree (v, S)) =
haftmann@28309
   252
          let
haftmann@28309
   253
            val T' = AList.lookup (op =) insts v
bulwahn@37912
   254
              |> the_default default_T
bulwahn@37913
   255
          in if Sign.of_sort thy (T', S) then T'
bulwahn@37913
   256
            else raise (WELLSORTED ("For instantiation with default_type " ^ Syntax.string_of_typ_global thy default_T ^
bulwahn@37913
   257
              ":\n" ^ Syntax.string_of_typ_global thy T' ^
haftmann@28309
   258
              " to be substituted for variable " ^
bulwahn@37913
   259
              Syntax.string_of_typ_global thy T ^ " does not have sort " ^
bulwahn@37913
   260
              Syntax.string_of_sort_global thy S))
haftmann@28309
   261
          end
haftmann@28309
   262
      | subst T = T;
haftmann@28309
   263
  in (map_types o map_atyps) subst end;
haftmann@28309
   264
bulwahn@37913
   265
datatype wellsorted_error = Wellsorted_Error of string | Term of term
bulwahn@37913
   266
bulwahn@40246
   267
fun test_goal generator_name insts i state =
haftmann@28309
   268
  let
haftmann@37974
   269
    val lthy = Proof.context_of state;
haftmann@28309
   270
    val thy = Proof.theory_of state;
haftmann@28309
   271
    fun strip (Const ("all", _) $ Abs (_, _, t)) = strip t
haftmann@28309
   272
      | strip t = t;
wenzelm@33291
   273
    val {goal = st, ...} = Proof.raw_goal state;
haftmann@28309
   274
    val (gi, frees) = Logic.goal_params (prop_of st) i;
haftmann@38390
   275
    val some_locale = case (Option.map #target o Named_Target.peek) lthy
haftmann@38390
   276
     of NONE => NONE
haftmann@38390
   277
      | SOME "" => NONE
haftmann@38390
   278
      | SOME locale => SOME locale;
bulwahn@40246
   279
    val assms = if no_assms lthy then [] else case some_locale
haftmann@37974
   280
     of NONE => Assumption.all_assms_of lthy
haftmann@37974
   281
      | SOME locale => Assumption.local_assms_of lthy (Locale.init locale thy);
haftmann@37974
   282
    val proto_goal = Logic.list_implies (map Thm.term_of assms, subst_bounds (frees, strip gi));
haftmann@37974
   283
    val check_goals = case some_locale
haftmann@37974
   284
     of NONE => [proto_goal]
bulwahn@40246
   285
      | SOME locale => map (fn (_, phi) => Morphism.term phi proto_goal)
bulwahn@40246
   286
        (Locale.registrations_of (Context.Theory thy) (*FIXME*) locale);
haftmann@37974
   287
    val inst_goals = maps (fn check_goal => map (fn T =>
haftmann@37974
   288
      Term ((Object_Logic.atomize_term thy o monomorphic_term thy insts T) check_goal)
bulwahn@40246
   289
        handle WELLSORTED s => Wellsorted_Error s) (default_type lthy)) check_goals
bulwahn@37913
   290
    val error_msg = cat_lines (map_filter (fn Term t => NONE | Wellsorted_Error s => SOME s) inst_goals)
bulwahn@37913
   291
    val correct_inst_goals =
bulwahn@37913
   292
      case map_filter (fn Term t => SOME t | Wellsorted_Error s => NONE) inst_goals of
bulwahn@37913
   293
        [] => error error_msg
bulwahn@37913
   294
      | xs => xs
bulwahn@40246
   295
    val _ = if quiet lthy then () else warning error_msg
bulwahn@37912
   296
    fun collect_results f reports [] = (NONE, rev reports)
bulwahn@37912
   297
      | collect_results f reports (t :: ts) =
bulwahn@37912
   298
        case f t of
bulwahn@37912
   299
          (SOME res, report) => (SOME res, rev (report :: reports))
bulwahn@37912
   300
        | (NONE, report) => collect_results f (report :: reports) ts
bulwahn@40246
   301
  in collect_results (test_term lthy generator_name) [] correct_inst_goals end;
bulwahn@37912
   302
bulwahn@37912
   303
(* pretty printing *)
haftmann@28315
   304
blanchet@40225
   305
fun tool_name auto = (if auto then "Auto " else "") ^ "Quickcheck"
blanchet@40225
   306
blanchet@40225
   307
fun pretty_counterex ctxt auto NONE = Pretty.str (tool_name auto ^ " found no counterexample.")
blanchet@40225
   308
  | pretty_counterex ctxt auto (SOME cex) =
blanchet@40225
   309
      Pretty.chunks (Pretty.str (tool_name auto ^ " found a counterexample:\n") ::
haftmann@28315
   310
        map (fn (s, t) =>
haftmann@28315
   311
          Pretty.block [Pretty.str (s ^ " ="), Pretty.brk 1, Syntax.pretty_term ctxt t]) cex);
haftmann@28315
   312
bulwahn@35378
   313
fun pretty_report (Report {iterations = iterations, raised_match_errors = raised_match_errors,
bulwahn@35378
   314
    satisfied_assms = satisfied_assms, positive_concl_tests = positive_concl_tests}) =
bulwahn@35378
   315
  let
bulwahn@35378
   316
    fun pretty_stat s i = Pretty.block ([Pretty.str (s ^ ": " ^ string_of_int i)])
bulwahn@35378
   317
  in
bulwahn@35378
   318
     ([pretty_stat "iterations" iterations,
bulwahn@35378
   319
     pretty_stat "match exceptions" raised_match_errors]
bulwahn@35378
   320
     @ map_index (fn (i, n) => pretty_stat ("satisfied " ^ string_of_int (i + 1) ^ ". assumption") n)
bulwahn@35378
   321
       satisfied_assms
bulwahn@35378
   322
     @ [pretty_stat "positive conclusion tests" positive_concl_tests])
bulwahn@35378
   323
  end
bulwahn@35378
   324
bulwahn@35378
   325
fun pretty_reports' [report] = [Pretty.chunks (pretty_report report)]
bulwahn@35378
   326
  | pretty_reports' reports =
bulwahn@35378
   327
  map_index (fn (i, report) =>
bulwahn@35378
   328
    Pretty.chunks (Pretty.str (string_of_int (i + 1) ^ ". generator:\n") :: pretty_report report))
bulwahn@35378
   329
    reports
bulwahn@35378
   330
bulwahn@35380
   331
fun pretty_reports ctxt (SOME reports) =
bulwahn@35378
   332
  Pretty.chunks (Pretty.str "Quickcheck report:" ::
bulwahn@35378
   333
    maps (fn (size, reports) =>
bulwahn@35378
   334
      Pretty.str ("size " ^ string_of_int size ^ ":") :: pretty_reports' reports @ [Pretty.brk 1])
bulwahn@35378
   335
      (rev reports))
bulwahn@35380
   336
  | pretty_reports ctxt NONE = Pretty.str ""
bulwahn@35378
   337
blanchet@40225
   338
fun pretty_counterex_and_reports ctxt auto (cex, timing_and_reports) =
blanchet@40225
   339
  Pretty.chunks (pretty_counterex ctxt auto cex ::
blanchet@40225
   340
    map (pretty_reports ctxt) (map snd timing_and_reports))
haftmann@28315
   341
haftmann@28315
   342
(* automatic testing *)
haftmann@28309
   343
blanchet@33561
   344
fun auto_quickcheck state =
blanchet@33561
   345
  if not (!auto) then
blanchet@33561
   346
    (false, state)
blanchet@33561
   347
  else
blanchet@33561
   348
    let
blanchet@33561
   349
      val ctxt = Proof.context_of state;
blanchet@33561
   350
      val res =
bulwahn@39253
   351
        state
bulwahn@40246
   352
        |> Proof.map_context (Context.proof_map (set_report false #> set_quiet true))
bulwahn@40246
   353
        |> try (test_goal NONE [] 1);
blanchet@33561
   354
    in
blanchet@33561
   355
      case res of
blanchet@33561
   356
        NONE => (false, state)
bulwahn@35378
   357
      | SOME (NONE, report) => (false, state)
bulwahn@35378
   358
      | SOME (cex, report) => (true, Proof.goal_message (K (Pretty.chunks [Pretty.str "",
blanchet@40225
   359
          Pretty.mark Markup.hilite (pretty_counterex ctxt true cex)])) state)
blanchet@33561
   360
    end
blanchet@33561
   361
blanchet@39324
   362
val setup = Auto_Tools.register_tool ("quickcheck", auto_quickcheck)
haftmann@28315
   363
haftmann@28315
   364
wenzelm@30980
   365
(* Isar commands *)
haftmann@28315
   366
haftmann@28336
   367
fun read_nat s = case (Library.read_int o Symbol.explode) s
haftmann@28336
   368
 of (k, []) => if k >= 0 then k
haftmann@28336
   369
      else error ("Not a natural number: " ^ s)
haftmann@28336
   370
  | (_, _ :: _) => error ("Not a natural number: " ^ s);
bulwahn@37909
   371
blanchet@34128
   372
fun read_bool "false" = false
blanchet@34128
   373
  | read_bool "true" = true
blanchet@34128
   374
  | read_bool s = error ("Not a Boolean value: " ^ s)
haftmann@28315
   375
bulwahn@40366
   376
fun read_real s =
bulwahn@40366
   377
  case (Real.fromString s) of
bulwahn@40366
   378
    SOME s => s
bulwahn@40366
   379
  | NONE => error ("Not a real number: " ^ s)
bulwahn@40366
   380
bulwahn@37929
   381
fun read_expectation "no_expectation" = No_Expectation
bulwahn@37929
   382
  | read_expectation "no_counterexample" = No_Counterexample 
bulwahn@37929
   383
  | read_expectation "counterexample" = Counterexample
bulwahn@37929
   384
  | read_expectation s = error ("Not an expectation value: " ^ s)  
bulwahn@37929
   385
bulwahn@37909
   386
fun parse_test_param ctxt ("size", [arg]) =
haftmann@28336
   387
      (apfst o apfst o K) (read_nat arg)
bulwahn@37909
   388
  | parse_test_param ctxt ("iterations", [arg]) =
haftmann@28336
   389
      (apfst o apsnd o K) (read_nat arg)
haftmann@28336
   390
  | parse_test_param ctxt ("default_type", arg) =
bulwahn@37909
   391
      (apsnd o apfst o apfst o K) (map (ProofContext.read_typ ctxt) arg)
bulwahn@37909
   392
  | parse_test_param ctxt ("no_assms", [arg]) =
bulwahn@35378
   393
      (apsnd o apfst o apsnd o K) (read_bool arg)
bulwahn@37929
   394
  | parse_test_param ctxt ("expect", [arg]) =
bulwahn@37929
   395
      (apsnd o apsnd o apfst o apfst o K) (read_expectation arg)
bulwahn@37909
   396
  | parse_test_param ctxt ("report", [arg]) =
bulwahn@37929
   397
      (apsnd o apsnd o apfst o apsnd o K) (read_bool arg)
bulwahn@37909
   398
  | parse_test_param ctxt ("quiet", [arg]) =
bulwahn@40246
   399
      (apsnd o apsnd o apsnd o apfst o K) (read_bool arg)
bulwahn@40246
   400
  | parse_test_param ctxt ("timeout", [arg]) =
bulwahn@40366
   401
      (apsnd o apsnd o apsnd o apsnd o K) (read_real arg)
haftmann@28336
   402
  | parse_test_param ctxt (name, _) =
blanchet@34128
   403
      error ("Unknown test parameter: " ^ name);
haftmann@28315
   404
bulwahn@37909
   405
fun parse_test_param_inst ctxt ("generator", [arg]) =
haftmann@28336
   406
      (apsnd o apfst o K o SOME) arg
haftmann@28336
   407
  | parse_test_param_inst ctxt (name, arg) =
haftmann@28336
   408
      case try (ProofContext.read_typ ctxt) name
haftmann@28336
   409
       of SOME (TFree (v, _)) => (apsnd o apsnd o AList.update (op =))
bulwahn@37909
   410
              (v, ProofContext.read_typ ctxt (the_single arg))
haftmann@28336
   411
        | _ => (apfst o parse_test_param ctxt) (name, arg);
haftmann@28309
   412
haftmann@28336
   413
fun quickcheck_params_cmd args thy =
haftmann@28315
   414
  let
bulwahn@39252
   415
    val ctxt = ProofContext.init_global thy
haftmann@28336
   416
    val f = fold (parse_test_param ctxt) args;
haftmann@28315
   417
  in
haftmann@28315
   418
    thy
bulwahn@40246
   419
    |> (Context.theory_map o map_test_params) f
haftmann@28315
   420
  end;
haftmann@28315
   421
bulwahn@35378
   422
fun gen_quickcheck args i state =
haftmann@28315
   423
  let
boehmes@32297
   424
    val ctxt = Proof.context_of state;
bulwahn@39252
   425
    val default_params = (dest_test_params o snd o Data.get o Context.Proof) ctxt;
haftmann@28336
   426
    val f = fold (parse_test_param_inst ctxt) args;
bulwahn@40246
   427
    val (test_params, (generator_name, insts)) = f (default_params, (NONE, []));
boehmes@32297
   428
  in
bulwahn@39253
   429
    state
bulwahn@40246
   430
    |> Proof.map_context (Context.proof_map (map_test_params (K test_params)))
bulwahn@40246
   431
    |> (fn state' => test_goal generator_name insts i state'
bulwahn@40246
   432
      |> tap (fn (SOME x, _) => if expect (Proof.context_of state') = No_Counterexample then
haftmann@37974
   433
                 error ("quickcheck expected to find no counterexample but found one") else ()
bulwahn@40246
   434
             | (NONE, _) => if expect (Proof.context_of state') = Counterexample then
bulwahn@40246
   435
                 error ("quickcheck expected to find a counterexample but did not find one") else ()))
boehmes@32297
   436
  end;
boehmes@32297
   437
wenzelm@36960
   438
fun quickcheck args i state = fst (gen_quickcheck args i state);
bulwahn@35378
   439
boehmes@32297
   440
fun quickcheck_cmd args i state =
bulwahn@35378
   441
  gen_quickcheck args i (Toplevel.proof_of state)
blanchet@40225
   442
  |> Pretty.writeln o pretty_counterex_and_reports (Toplevel.context_of state) false;
haftmann@28309
   443
bulwahn@37909
   444
val parse_arg = Parse.name -- (Scan.optional (Parse.$$$ "=" |-- 
bulwahn@40366
   445
  (((Parse.name || Parse.float_number) >> single) || (Parse.$$$ "[" |-- Parse.list1 Parse.name --| Parse.$$$ "]"))) ["true"]);
haftmann@28309
   446
wenzelm@36960
   447
val parse_args = Parse.$$$ "[" |-- Parse.list1 parse_arg --| Parse.$$$ "]"
haftmann@28336
   448
  || Scan.succeed [];
haftmann@28336
   449
wenzelm@36960
   450
val _ =
wenzelm@36960
   451
  Outer_Syntax.command "quickcheck_params" "set parameters for random testing" Keyword.thy_decl
wenzelm@36960
   452
    (parse_args >> (fn args => Toplevel.theory (quickcheck_params_cmd args)));
haftmann@28309
   453
wenzelm@36960
   454
val _ =
wenzelm@36960
   455
  Outer_Syntax.improper_command "quickcheck" "try to find counterexample for subgoal" Keyword.diag
wenzelm@36960
   456
    (parse_args -- Scan.optional Parse.nat 1
wenzelm@36960
   457
      >> (fn (args, i) => Toplevel.no_timing o Toplevel.keep (quickcheck_cmd args i)));
haftmann@28309
   458
haftmann@28315
   459
end;
haftmann@28309
   460
haftmann@28309
   461
haftmann@28315
   462
val auto_quickcheck = Quickcheck.auto;