wenzelm@30824: (*  Title:      Tools/quickcheck.ML
bulwahn@40246:     Author:     Stefan Berghofer, Florian Haftmann, Lukas Bulwahn, TU Muenchen
haftmann@28256: 
haftmann@28256: Generic counterexample search engine.
haftmann@28256: *)
haftmann@28256: 
haftmann@28256: signature QUICKCHECK =
haftmann@28256: sig
blanchet@43020:   val quickcheckN: string
blanchet@43020:   val genuineN: string
blanchet@43020:   val noneN: string
blanchet@43020:   val unknownN: string
wenzelm@51302:   (*configuration*)
bulwahn@43882:   val batch_tester : string Config.T
bulwahn@40644:   val size : int Config.T
bulwahn@40644:   val iterations : int Config.T
bulwahn@45213:   val depth : int Config.T
bulwahn@40644:   val no_assms : bool Config.T
bulwahn@40644:   val report : bool Config.T
bulwahn@46565:   val timeout : real Config.T
bulwahn@42088:   val timing : bool Config.T
bulwahn@45757:   val genuine_only : bool Config.T
wenzelm@51302:   val abort_potential : bool Config.T
bulwahn@40644:   val quiet : bool Config.T
bulwahn@45764:   val verbose : bool Config.T
bulwahn@46565:   val use_subtype : bool Config.T
bulwahn@46565:   val allow_function_inversion : bool Config.T
bulwahn@40648:   val finite_types : bool Config.T
bulwahn@40648:   val finite_type_size : int Config.T
bulwahn@46863:   val tag : string Config.T
bulwahn@47348:   val locale : string Config.T
bulwahn@43912:   val set_active_testers: string list -> Context.generic -> Context.generic
wenzelm@41517:   datatype expectation = No_Expectation | No_Counterexample | Counterexample;
bulwahn@40644:   datatype test_params = Test_Params of {default_type: typ list, expect : expectation};
bulwahn@40246:   val test_params_of : Proof.context -> test_params
bulwahn@40644:   val map_test_params : (typ list * expectation -> typ list * expectation)
bulwahn@40246:     -> Context.generic -> Context.generic
bulwahn@45159:   val default_type : Proof.context -> typ list
bulwahn@42089:   datatype report = Report of
bulwahn@42089:     { iterations : int, raised_match_errors : int,
bulwahn@42089:       satisfied_assms : int list, positive_concl_tests : int }
wenzelm@51302:   (*quickcheck's result*)
bulwahn@42088:   datatype result =
bulwahn@42088:     Result of
bulwahn@45727:      {counterexample : (bool * (string * term) list) option,
bulwahn@42088:       evaluation_terms : (term * term) list option,
bulwahn@42088:       timings : (string * int) list,
bulwahn@42088:       reports : (int * report) list}
bulwahn@43314:   val empty_result : result
bulwahn@45159:   val found_counterexample : result -> bool
bulwahn@43585:   val add_timing : (string * int) -> result Unsynchronized.ref -> unit
wenzelm@51302:   val add_response : string list -> term list -> (bool * term list) option ->
wenzelm@51302:     result Unsynchronized.ref -> unit
bulwahn@45159:   val add_report : int -> report option -> result Unsynchronized.ref -> unit
bulwahn@45727:   val counterexample_of : result -> (bool * (string * term) list) option
bulwahn@42089:   val timings_of : result -> (string * int) list
wenzelm@51302:   (*registering testers & generators*)
bulwahn@43878:   type tester =
bulwahn@45419:     Proof.context -> bool -> (string * typ) list -> (term * term list) list -> result list
bulwahn@43878:   val add_tester : string * (bool Config.T * tester) -> Context.generic -> Context.generic
bulwahn@43878:   val add_batch_generator :
bulwahn@43112:     string * (Proof.context -> term list -> (int -> term list option) list)
bulwahn@43112:       -> Context.generic -> Context.generic
bulwahn@43878:   val add_batch_validator :
bulwahn@43112:     string * (Proof.context -> term list -> (int -> bool) list)
bulwahn@43112:       -> Context.generic -> Context.generic
wenzelm@51302:   (*basic operations*)
bulwahn@45159:   val message : Proof.context -> string -> unit
bulwahn@45765:   val verbose_message : Proof.context -> string -> unit
bulwahn@45159:   val limit : Time.time -> (bool * bool) -> (unit -> 'a) -> (unit -> 'a) -> unit -> 'a
bulwahn@45755:   val pretty_counterex : Proof.context -> bool ->
bulwahn@45755:     ((bool * (string * term) list) * (term * term) list) option -> Pretty.T
wenzelm@51302:   (*testing terms and proof states*)
bulwahn@45159:   val mk_batch_validator : Proof.context -> term list -> (int -> bool) list option
bulwahn@45159:   val mk_batch_tester : Proof.context -> term list -> (int -> term list option) list option
bulwahn@45159:   val active_testers : Proof.context -> tester list
wenzelm@51302:   val test_terms : Proof.context -> bool * bool -> (string * typ) list ->
wenzelm@51302:     (term * term list) list -> result list option
wenzelm@51302:   val quickcheck: (string * string list) list -> int -> Proof.state ->
wenzelm@51302:     (bool * (string * term) list) option
haftmann@28256: end;
haftmann@28256: 
haftmann@28256: structure Quickcheck : QUICKCHECK =
haftmann@28256: struct
haftmann@28256: 
wenzelm@55627: val quickcheckN = "quickcheck";
blanchet@43020: 
wenzelm@55627: val genuineN = "genuine";
wenzelm@55627: val noneN = "none";
wenzelm@55627: val unknownN = "unknown";
blanchet@43020: 
wenzelm@51302: 
bulwahn@35378: (* quickcheck report *)
bulwahn@35378: 
bulwahn@35378: datatype report = Report of
wenzelm@55627:  {iterations : int,
wenzelm@55627:   raised_match_errors : int,
wenzelm@55627:   satisfied_assms : int list,
wenzelm@55627:   positive_concl_tests : int};
bulwahn@35378: 
wenzelm@51302: 
bulwahn@42088: (* Quickcheck Result *)
bulwahn@42088: 
bulwahn@42088: datatype result = Result of
wenzelm@55627:  {counterexample : (bool * (string * term) list) option,
wenzelm@55627:   evaluation_terms : (term * term) list option,
wenzelm@55627:   timings : (string * int) list,
wenzelm@55627:   reports : (int * report) list};
bulwahn@42088: 
bulwahn@42088: val empty_result =
wenzelm@55627:   Result {counterexample = NONE, evaluation_terms = NONE, timings = [], reports = []};
bulwahn@42088: 
wenzelm@55627: fun counterexample_of (Result r) = #counterexample r;
bulwahn@42088: 
wenzelm@55627: fun found_counterexample (Result r) = is_some (#counterexample r);
bulwahn@42088: 
wenzelm@55627: fun response_of (Result r) =
wenzelm@55627:   (case (#counterexample r, #evaluation_terms r) of
bulwahn@45730:     (SOME ts, SOME evals) => SOME (ts, evals)
wenzelm@55627:   | (NONE, NONE) => NONE);
bulwahn@42088: 
wenzelm@55627: fun timings_of (Result r) = #timings r;
bulwahn@42088: 
bulwahn@45727: fun set_response names eval_terms (SOME (genuine, ts)) (Result r) =
wenzelm@55627:       let
wenzelm@55627:         val (ts1, ts2) = chop (length names) ts
wenzelm@55627:         val (eval_terms', _) = chop (length ts2) eval_terms
wenzelm@55627:       in
wenzelm@55627:         Result {counterexample = SOME (genuine, (names ~~ ts1)),
wenzelm@55627:           evaluation_terms = SOME (eval_terms' ~~ ts2),
wenzelm@55627:           timings = #timings r, reports = #reports r}
wenzelm@55627:       end
wenzelm@55627:   | set_response _ _ NONE result = result;
bulwahn@42088: 
bulwahn@43585: 
bulwahn@42088: fun cons_timing timing (Result r) =
bulwahn@42088:   Result {counterexample = #counterexample r, evaluation_terms = #evaluation_terms r,
wenzelm@55627:     timings = cons timing (#timings r), reports = #reports r};
bulwahn@42088: 
bulwahn@42088: fun cons_report size (SOME report) (Result r) =
wenzelm@55627:       Result {counterexample = #counterexample r, evaluation_terms = #evaluation_terms r,
wenzelm@55627:         timings = #timings r, reports = cons (size, report) (#reports r)}
wenzelm@55627:   | cons_report _ NONE result = result;
bulwahn@42088: 
wenzelm@42198: fun add_timing timing result_ref =
wenzelm@55627:   Unsynchronized.change result_ref (cons_timing timing);
bulwahn@42088: 
wenzelm@42198: fun add_report size report result_ref =
wenzelm@55627:   Unsynchronized.change result_ref (cons_report size report);
bulwahn@42088: 
bulwahn@42088: fun add_response names eval_terms response result_ref =
wenzelm@55627:   Unsynchronized.change result_ref (set_response names eval_terms response);
bulwahn@42088: 
wenzelm@51302: 
bulwahn@37929: (* expectation *)
bulwahn@37929: 
wenzelm@41517: datatype expectation = No_Expectation | No_Counterexample | Counterexample;
bulwahn@37929: 
bulwahn@37929: fun merge_expectation (expect1, expect2) =
wenzelm@55627:   if expect1 = expect2 then expect1 else No_Expectation;
bulwahn@37929: 
wenzelm@51302: (*quickcheck configuration -- default parameters, test generators*)
wenzelm@67149: val batch_tester = Attrib.setup_config_string \<^binding>\<open>quickcheck_batch_tester\<close> (K "");
wenzelm@67149: val size = Attrib.setup_config_int \<^binding>\<open>quickcheck_size\<close> (K 10);
wenzelm@67149: val iterations = Attrib.setup_config_int \<^binding>\<open>quickcheck_iterations\<close> (K 100);
wenzelm@67149: val depth = Attrib.setup_config_int \<^binding>\<open>quickcheck_depth\<close> (K 10);
bulwahn@45213: 
wenzelm@67149: val no_assms = Attrib.setup_config_bool \<^binding>\<open>quickcheck_no_assms\<close> (K false);
wenzelm@67149: val locale = Attrib.setup_config_string \<^binding>\<open>quickcheck_locale\<close> (K "interpret expand");
wenzelm@67149: val report = Attrib.setup_config_bool \<^binding>\<open>quickcheck_report\<close> (K true);
wenzelm@67149: val timing = Attrib.setup_config_bool \<^binding>\<open>quickcheck_timing\<close> (K false);
wenzelm@67149: val timeout = Attrib.setup_config_real \<^binding>\<open>quickcheck_timeout\<close> (K 30.0);
bulwahn@46565: 
wenzelm@67149: val genuine_only = Attrib.setup_config_bool \<^binding>\<open>quickcheck_genuine_only\<close> (K false);
wenzelm@67149: val abort_potential = Attrib.setup_config_bool \<^binding>\<open>quickcheck_abort_potential\<close> (K false);
bulwahn@46565: 
wenzelm@67149: val quiet = Attrib.setup_config_bool \<^binding>\<open>quickcheck_quiet\<close> (K false);
wenzelm@67149: val verbose = Attrib.setup_config_bool \<^binding>\<open>quickcheck_verbose\<close> (K false);
wenzelm@67149: val tag = Attrib.setup_config_string \<^binding>\<open>quickcheck_tag\<close> (K "");
bulwahn@46565: 
wenzelm@67149: val use_subtype = Attrib.setup_config_bool \<^binding>\<open>quickcheck_use_subtype\<close> (K false);
bulwahn@46565: 
bulwahn@45449: val allow_function_inversion =
wenzelm@67149:   Attrib.setup_config_bool \<^binding>\<open>quickcheck_allow_function_inversion\<close> (K false);
wenzelm@67149: val finite_types = Attrib.setup_config_bool \<^binding>\<open>quickcheck_finite_types\<close> (K true);
wenzelm@67149: val finite_type_size = Attrib.setup_config_int \<^binding>\<open>quickcheck_finite_type_size\<close> (K 3);
bulwahn@40646: 
bulwahn@40644: datatype test_params = Test_Params of
bulwahn@40644:   {default_type: typ list, expect : expectation};
wenzelm@38759: 
bulwahn@40644: fun dest_test_params (Test_Params {default_type, expect}) = (default_type, expect);
bulwahn@40644: 
wenzelm@41517: fun make_test_params (default_type, expect) =
wenzelm@41517:   Test_Params {default_type = default_type, expect = expect};
bulwahn@40644: 
wenzelm@41517: fun map_test_params' f (Test_Params {default_type, expect}) =
wenzelm@41517:   make_test_params (f (default_type, expect));
wenzelm@38759: 
wenzelm@38759: fun merge_test_params
wenzelm@41472:   (Test_Params {default_type = default_type1, expect = expect1},
wenzelm@41472:     Test_Params {default_type = default_type2, expect = expect2}) =
wenzelm@41472:   make_test_params
wenzelm@41472:     (merge (op =) (default_type1, default_type2), merge_expectation (expect1, expect2));
haftmann@28309: 
bulwahn@43878: type tester =
wenzelm@51302:   Proof.context -> bool -> (string * typ) list -> (term * term list) list -> result list;
bulwahn@43878: 
bulwahn@39252: structure Data = Generic_Data
wenzelm@33522: (
wenzelm@38759:   type T =
wenzelm@59436:     (string * (bool Config.T * tester)) list *
wenzelm@59436:     (string * (Proof.context -> term list -> (int -> term list option) list)) list *
wenzelm@59436:     (string * (Proof.context -> term list -> (int -> bool) list)) list *
wenzelm@59436:     test_params;
wenzelm@59436:   val empty = ([], [], [], Test_Params {default_type = [], expect = No_Expectation});
haftmann@28256:   val extend = I;
wenzelm@55627:   fun merge
wenzelm@59436:    ((testers1, batch_generators1, batch_validators1, params1),
wenzelm@59436:     (testers2, batch_generators2, batch_validators2, params2)) : T =
wenzelm@59436:     (AList.merge (op =) (K true) (testers1, testers2),
wenzelm@59436:      AList.merge (op =) (K true) (batch_generators1, batch_generators2),
wenzelm@59436:      AList.merge (op =) (K true) (batch_validators1, batch_validators2),
wenzelm@59436:      merge_test_params (params1, params2));
wenzelm@33522: );
haftmann@28256: 
wenzelm@59436: val test_params_of = #4 o Data.get o Context.Proof;
wenzelm@55627: val default_type = fst o dest_test_params o test_params_of;
wenzelm@55627: val expect = snd o dest_test_params o test_params_of;
wenzelm@59436: val map_test_params = Data.map o @{apply 4(4)} o map_test_params';
bulwahn@40246: 
wenzelm@59436: val add_tester = Data.map o @{apply 4(1)} o AList.update (op =);
wenzelm@59436: val add_batch_generator = Data.map o @{apply 4(2)} o AList.update (op =);
wenzelm@59436: val add_batch_validator = Data.map o @{apply 4(3)} o AList.update (op =);
haftmann@28309: 
bulwahn@43881: fun active_testers ctxt =
bulwahn@43881:   let
wenzelm@59436:     val testers = map snd (#1 (Data.get (Context.Proof ctxt)));
bulwahn@43881:   in
bulwahn@43881:     map snd (filter (fn (active, _) => Config.get ctxt active) testers)
wenzelm@55627:   end;
wenzelm@51302: 
wenzelm@55629: fun set_active_testers [] context = context
wenzelm@55629:   | set_active_testers testers context =
wenzelm@51302:       let
wenzelm@59436:         val registered_testers = #1 (Data.get context);
wenzelm@51302:       in
wenzelm@51302:         fold (fn (name, (config, _)) => Config.put_generic config (member (op =) testers name))
wenzelm@55629:           registered_testers context
wenzelm@51302:       end;
wenzelm@51302: 
wenzelm@51302: 
haftmann@28315: (* generating tests *)
haftmann@28315: 
bulwahn@41862: fun gen_mk_tester lookup ctxt v =
haftmann@28309:   let
bulwahn@43882:     val name = Config.get ctxt batch_tester
wenzelm@51302:     val tester =
wenzelm@51302:       (case lookup ctxt name of
wenzelm@51302:         NONE => error ("No such quickcheck batch-tester: " ^ name)
wenzelm@51302:       | SOME tester => tester ctxt);
wenzelm@40235:   in
bulwahn@40909:     if Config.get ctxt quiet then
bulwahn@41862:       try tester v
bulwahn@40909:     else
wenzelm@43761:       let (* FIXME !?!? *)
bulwahn@41862:         val tester = Exn.interruptible_capture tester v
wenzelm@51302:       in
wenzelm@51302:         (case Exn.get_res tester of
bulwahn@40909:           NONE => SOME (Exn.release tester)
wenzelm@51302:         | SOME tester => SOME tester)
bulwahn@40909:       end
wenzelm@51302:   end;
bulwahn@43882: 
wenzelm@59436: val mk_batch_tester = gen_mk_tester (AList.lookup (op =) o #2 o Data.get o Context.Proof);
wenzelm@59436: val mk_batch_validator = gen_mk_tester (AList.lookup (op =) o #3 o Data.get o Context.Proof);
wenzelm@51302: 
wenzelm@51302: 
haftmann@28315: (* testing propositions *)
haftmann@28315: 
bulwahn@43876: type compile_generator =
wenzelm@51302:   Proof.context -> (term * term list) list -> int list -> term list option * report option;
bulwahn@43876: 
bulwahn@43585: fun limit timeout (limit_time, is_interactive) f exc () =
bulwahn@41754:   if limit_time then
wenzelm@62519:     Timeout.apply timeout f ()
wenzelm@62519:       handle timeout_exn as Timeout.TIMEOUT _ =>
wenzelm@62519:         if is_interactive then exc () else Exn.reraise timeout_exn
wenzelm@55627:   else f ();
wenzelm@51302: 
wenzelm@58843: fun message ctxt s = if Config.get ctxt quiet then () else writeln s;
bulwahn@41753: 
wenzelm@51302: fun verbose_message ctxt s =
wenzelm@51302:   if not (Config.get ctxt quiet) andalso Config.get ctxt verbose
wenzelm@58843:   then writeln s else ();
bulwahn@45765: 
wenzelm@62983: fun test_terms ctxt0 (limit_time, is_interactive) insts goals =
wenzelm@62983:   let val ctxt = Simplifier_Trace.disable ctxt0 in
wenzelm@62983:     (case active_testers ctxt of
wenzelm@62983:       [] => error "No active testers for quickcheck"
wenzelm@62983:     | testers =>
wenzelm@62983:         limit (seconds (Config.get ctxt timeout)) (limit_time, is_interactive)
wenzelm@62983:           (fn () =>
wenzelm@62983:             Par_List.get_some (fn tester =>
wenzelm@62983:               tester ctxt (length testers > 1) insts goals |>
wenzelm@62983:               (fn result => if exists found_counterexample result then SOME result else NONE))
wenzelm@62983:             testers)
wenzelm@62983:           (fn () => (message ctxt "Quickcheck ran out of time"; NONE)) ())
wenzelm@62983:   end
bulwahn@46759: 
bulwahn@46759: fun all_axioms_of ctxt t =
bulwahn@46759:   let
wenzelm@51302:     val intros = Locale.get_intros ctxt;
wenzelm@51302:     val unfolds = Locale.get_unfolds ctxt;
wenzelm@51302:     fun retrieve_prems thms t =
wenzelm@51302:        (case filter (fn th => Term.could_unify (Thm.concl_of th, t)) thms of
wenzelm@51302:          [] => NONE
bulwahn@46759:        | [th] =>
wenzelm@51302:            let
wenzelm@51302:              val (tyenv, tenv) =
wenzelm@51302:                Pattern.match (Proof_Context.theory_of ctxt)
wenzelm@51302:                 (Thm.concl_of th, t) (Vartab.empty, Vartab.empty)
wenzelm@51302:            in SOME (map (Envir.subst_term (tyenv, tenv)) (Thm.prems_of th)) end);
bulwahn@46759:     fun all t =
wenzelm@51302:       (case retrieve_prems intros t of
bulwahn@46759:         NONE => retrieve_prems unfolds t
wenzelm@51302:       | SOME ts => SOME (maps (fn t => the_default [t] (all t)) ts));
bulwahn@46759:   in
bulwahn@46759:     all t
wenzelm@51302:   end;
bulwahn@46759: 
bulwahn@47348: fun locale_config_of s =
bulwahn@47348:   let
wenzelm@51302:     val cs = space_explode " " s;
bulwahn@47348:   in
bulwahn@47348:     if forall (fn c => c = "expand" orelse c = "interpret") cs then cs
wenzelm@55627:     else
wenzelm@55627:      (warning ("Invalid quickcheck_locale setting: falling back to the default setting.");
bulwahn@47348:       ["interpret", "expand"])
wenzelm@51302:   end;
bulwahn@47348: 
bulwahn@42026: fun test_goal (time_limit, is_interactive) (insts, eval_terms) i state =
bulwahn@40648:   let
wenzelm@62982:     val ctxt = Proof.context_of state;
bulwahn@40648:     val thy = Proof.theory_of state;
wenzelm@62982: 
wenzelm@67149:     fun strip (Const (\<^const_name>\<open>Pure.all\<close>, _) $ Abs (_, _, t)) = strip t
bulwahn@40648:       | strip t = t;
bulwahn@40648:     val {goal = st, ...} = Proof.raw_goal state;
wenzelm@59582:     val (gi, frees) = Logic.goal_params (Thm.prop_of st) i;
wenzelm@62982:     val opt_locale = Named_Target.bottom_locale_of ctxt;
wenzelm@51302:     val assms =
wenzelm@62982:       if Config.get ctxt no_assms then []
wenzelm@51302:       else
wenzelm@62982:         (case opt_locale of
wenzelm@62982:           NONE => Assumption.all_assms_of ctxt
wenzelm@62982:         | SOME locale => Assumption.local_assms_of ctxt (Locale.init locale thy));
bulwahn@40648:     val proto_goal = Logic.list_implies (map Thm.term_of assms, subst_bounds (frees, strip gi));
wenzelm@51302:     fun axioms_of locale =
wenzelm@51302:       (case fst (Locale.specification_of thy locale) of
bulwahn@46759:         NONE => []
wenzelm@62982:       | SOME t => the_default [] (all_axioms_of ctxt t));
wenzelm@62982:     val config = locale_config_of (Config.get ctxt locale);
wenzelm@55627:     val goals =
wenzelm@62982:       (case opt_locale of
wenzelm@55627:         NONE => [(proto_goal, eval_terms)]
wenzelm@55627:       | SOME locale =>
wenzelm@55627:           fold (fn c =>
wenzelm@55627:             if c = "expand" then
wenzelm@55627:               cons (Logic.list_implies (axioms_of locale, proto_goal), eval_terms)
wenzelm@55627:             else if c = "interpret" then
wenzelm@55627:               append (map (fn (_, phi) =>
wenzelm@55627:                   (Morphism.term phi proto_goal, map (Morphism.term phi) eval_terms))
wenzelm@55627:                 (Locale.registrations_of (Context.Theory thy) (* FIXME !? *) locale))
wenzelm@55627:             else I) config []);
wenzelm@51302:     val _ =
wenzelm@62982:       verbose_message ctxt
wenzelm@51302:         (Pretty.string_of
wenzelm@62982:           (Pretty.big_list ("Checking goals: ") (map (Syntax.pretty_term ctxt o fst) goals)));
bulwahn@40648:   in
wenzelm@62982:     test_terms ctxt (time_limit, is_interactive) insts goals
wenzelm@51302:   end;
wenzelm@51302: 
bulwahn@40648: 
bulwahn@37912: (* pretty printing *)
haftmann@28315: 
wenzelm@59435: fun tool_name auto = if auto then "Auto Quickcheck" else "Quickcheck";
blanchet@40225: 
bulwahn@46863: fun pretty_counterex ctxt auto NONE =
wenzelm@59435:       Pretty.para (tool_name auto ^ " found no counterexample." ^ Config.get ctxt tag)
bulwahn@45730:   | pretty_counterex ctxt auto (SOME ((genuine, cex), eval_terms)) =
wenzelm@59438:       let
wenzelm@59438:         val header =
wenzelm@59438:           Pretty.para
wenzelm@59438:             (tool_name auto ^ " found a " ^
wenzelm@60666:               (if genuine then "counterexample"
wenzelm@60666:                else "potentially spurious counterexample due to underspecified functions") ^
wenzelm@60666:               (if null cex then "." else ":") ^
wenzelm@59438:               Config.get ctxt tag);
wenzelm@59439:         fun pretty_cex (x, t) =
wenzelm@59439:           Pretty.block [Pretty.str (x ^ " ="), Pretty.brk 1, Syntax.pretty_term ctxt t];
wenzelm@59438:       in
wenzelm@59439:         Pretty.chunks (Pretty.block (Pretty.fbreaks (header :: map pretty_cex (rev cex))) ::
wenzelm@59438:           (if null eval_terms then []
wenzelm@59438:            else
wenzelm@59438:             [Pretty.big_list "Evaluated terms:"
wenzelm@59438:               (map (fn (t, u) =>
wenzelm@59438:                 Pretty.block [Syntax.pretty_term ctxt t, Pretty.str " =", Pretty.brk 1,
wenzelm@59438:                   Syntax.pretty_term ctxt u]) (rev eval_terms))]))
wenzelm@59438:       end;
haftmann@28315: 
wenzelm@51302: 
wenzelm@30980: (* Isar commands *)
haftmann@28315: 
wenzelm@51302: fun read_nat s =
wenzelm@55627:   (case Library.read_int (Symbol.explode s) of
wenzelm@51302:     (k, []) =>
wenzelm@51302:       if k >= 0 then k
haftmann@28336:       else error ("Not a natural number: " ^ s)
wenzelm@55627:   | _ => error ("Not a natural number: " ^ s));
bulwahn@37909: 
blanchet@34128: fun read_bool "false" = false
blanchet@34128:   | read_bool "true" = true
wenzelm@51302:   | read_bool s = error ("Not a Boolean value: " ^ s);
haftmann@28315: 
bulwahn@40366: fun read_real s =
wenzelm@55627:   (case Real.fromString s of
bulwahn@40366:     SOME s => s
wenzelm@51302:   | NONE => error ("Not a real number: " ^ s));
bulwahn@40366: 
bulwahn@37929: fun read_expectation "no_expectation" = No_Expectation
wenzelm@41517:   | read_expectation "no_counterexample" = No_Counterexample
bulwahn@37929:   | read_expectation "counterexample" = Counterexample
wenzelm@51302:   | read_expectation s = error ("Not an expectation value: " ^ s);
bulwahn@37929: 
wenzelm@59437: fun valid_tester_name context name =
wenzelm@59437:   AList.defined (op =) (#1 (Data.get context)) name;
wenzelm@51302: 
wenzelm@59437: fun parse_tester name (testers, context) =
wenzelm@59437:   if valid_tester_name context name then
wenzelm@59437:     (insert (op =) name testers, context)
wenzelm@55627:   else error ("Unknown tester: " ^ name);
bulwahn@40912: 
bulwahn@43881: fun parse_test_param ("tester", args) = fold parse_tester args
bulwahn@43881:   | parse_test_param ("size", [arg]) = apsnd (Config.put_generic size (read_nat arg))
bulwahn@43881:   | parse_test_param ("iterations", [arg]) = apsnd (Config.put_generic iterations (read_nat arg))
wenzelm@51302:   | parse_test_param ("depth", [arg]) = apsnd (Config.put_generic depth (read_nat arg))
wenzelm@55627:   | parse_test_param ("default_type", arg) =
wenzelm@55629:       (fn (testers, context) =>
wenzelm@55627:         (testers, map_test_params
wenzelm@55629:           (apfst (K (map (Proof_Context.read_typ (Context.proof_of context)) arg))) context))
bulwahn@43881:   | parse_test_param ("no_assms", [arg]) = apsnd (Config.put_generic no_assms (read_bool arg))
wenzelm@55627:   | parse_test_param ("expect", [arg]) = apsnd (map_test_params (apsnd (K (read_expectation arg))))
bulwahn@43881:   | parse_test_param ("report", [arg]) = apsnd (Config.put_generic report (read_bool arg))
wenzelm@55627:   | parse_test_param ("genuine_only", [arg]) =
wenzelm@55627:       apsnd (Config.put_generic genuine_only (read_bool arg))
wenzelm@55627:   | parse_test_param ("abort_potential", [arg]) =
wenzelm@55627:       apsnd (Config.put_generic abort_potential (read_bool arg))
bulwahn@43881:   | parse_test_param ("quiet", [arg]) = apsnd (Config.put_generic quiet (read_bool arg))
bulwahn@45764:   | parse_test_param ("verbose", [arg]) = apsnd (Config.put_generic verbose (read_bool arg))
bulwahn@46863:   | parse_test_param ("tag", [arg]) = apsnd (Config.put_generic tag arg)
wenzelm@55627:   | parse_test_param ("use_subtype", [arg]) =
wenzelm@55627:       apsnd (Config.put_generic use_subtype (read_bool arg))
wenzelm@55627:   | parse_test_param ("timeout", [arg]) =
wenzelm@55627:       apsnd (Config.put_generic timeout (read_real arg))
wenzelm@55627:   | parse_test_param ("finite_types", [arg]) =
wenzelm@55627:       apsnd (Config.put_generic finite_types (read_bool arg))
bulwahn@45449:   | parse_test_param ("allow_function_inversion", [arg]) =
bulwahn@45449:       apsnd (Config.put_generic allow_function_inversion (read_bool arg))
wenzelm@41517:   | parse_test_param ("finite_type_size", [arg]) =
wenzelm@55627:       apsnd (Config.put_generic finite_type_size (read_nat arg))
wenzelm@51302:   | parse_test_param (name, _) =
wenzelm@59437:       (fn (testers, context) =>
wenzelm@59437:         if valid_tester_name context name then
wenzelm@59437:           (insert (op =) name testers, context)
wenzelm@51302:         else error ("Unknown tester or test parameter: " ^ name));
haftmann@28315: 
bulwahn@43881: fun parse_test_param_inst (name, arg) ((insts, eval_terms), (testers, ctxt)) =
wenzelm@51302:   (case try (Proof_Context.read_typ ctxt) name of
wenzelm@51302:     SOME (TFree (v, _)) =>
wenzelm@51302:       ((AList.update (op =) (v, Proof_Context.read_typ ctxt (the_single arg)) insts, eval_terms),
wenzelm@51302:         (testers, ctxt))
wenzelm@51302:   | NONE =>
wenzelm@51302:       (case name of
wenzelm@51302:         "eval" => ((insts, eval_terms @ map (Syntax.read_term ctxt) arg), (testers, ctxt))
wenzelm@51302:       | _ =>
wenzelm@51302:         ((insts, eval_terms),
wenzelm@55630:           let
wenzelm@55630:             val (testers', Context.Proof ctxt') =
wenzelm@55630:               parse_test_param (name, arg) (testers, Context.Proof ctxt);
wenzelm@55630:           in (testers', ctxt') end)));
haftmann@28309: 
wenzelm@55627: fun quickcheck_params_cmd args =
wenzelm@55627:   Context.theory_map
wenzelm@55629:     (fn context => uncurry set_active_testers (fold parse_test_param args ([], context)));
wenzelm@41517: 
bulwahn@42088: fun check_expectation state results =
wenzelm@51302:   if is_some results andalso expect (Proof.context_of state) = No_Counterexample then
wenzelm@51302:     error "quickcheck expected to find no counterexample but found one"
wenzelm@51302:   else if is_none results andalso expect (Proof.context_of state) = Counterexample then
wenzelm@51302:     error "quickcheck expected to find a counterexample but did not find one"
wenzelm@51302:   else ();
bulwahn@42088: 
bulwahn@35378: fun gen_quickcheck args i state =
bulwahn@40644:   state
bulwahn@43881:   |> Proof.map_context_result (fn ctxt =>
bulwahn@43881:     apsnd (fn (testers, ctxt) => Context.proof_map (set_active_testers testers) ctxt)
bulwahn@43881:       (fold parse_test_param_inst args (([], []), ([], ctxt))))
wenzelm@55627:   |> (fn ((insts, eval_terms), state') =>
wenzelm@55627:       test_goal (true, true) (insts, eval_terms) i state'
wenzelm@55627:       |> tap (check_expectation state')
wenzelm@55627:       |> rpair state');
boehmes@32297: 
bulwahn@43879: fun quickcheck args i state =
wenzelm@51302:   Option.map (the o get_first counterexample_of) (fst (gen_quickcheck args i state));
bulwahn@35378: 
wenzelm@60190: fun quickcheck_cmd args i st =
wenzelm@60190:   gen_quickcheck args i (Toplevel.proof_of st)
bulwahn@46863:   |> apfst (Option.map (the o get_first response_of))
wenzelm@51302:   |> (fn (r, state) =>
wenzelm@58843:       writeln (Pretty.string_of
wenzelm@51302:         (pretty_counterex (Proof.context_of state) false r)));
haftmann@28309: 
wenzelm@41517: val parse_arg =
wenzelm@41517:   Parse.name --
wenzelm@67149:     (Scan.optional (\<^keyword>\<open>=\<close> |--
wenzelm@41517:       (((Parse.name || Parse.float_number) >> single) ||
wenzelm@67149:         (\<^keyword>\<open>[\<close> |-- Parse.list1 Parse.name --| \<^keyword>\<open>]\<close>))) ["true"]);
haftmann@28309: 
wenzelm@41517: val parse_args =
wenzelm@67149:   \<^keyword>\<open>[\<close> |-- Parse.list1 parse_arg --| \<^keyword>\<open>]\<close> || Scan.succeed [];
haftmann@28336: 
wenzelm@36960: val _ =
wenzelm@67149:   Outer_Syntax.command \<^command_keyword>\<open>quickcheck_params\<close> "set parameters for random testing"
wenzelm@36960:     (parse_args >> (fn args => Toplevel.theory (quickcheck_params_cmd args)));
haftmann@28309: 
wenzelm@36960: val _ =
wenzelm@67149:   Outer_Syntax.command \<^command_keyword>\<open>quickcheck\<close>
wenzelm@46961:     "try to find counterexample for subgoal"
wenzelm@60094:     (parse_args -- Scan.optional Parse.nat 1 >>
wenzelm@60190:       (fn (args, i) => Toplevel.keep_proof (quickcheck_cmd args i)));
haftmann@28309: 
wenzelm@51302: 
blanchet@43020: (* automatic testing *)
blanchet@43020: 
blanchet@43020: fun try_quickcheck auto state =
blanchet@43020:   let
blanchet@43020:     val ctxt = Proof.context_of state;
blanchet@43020:     val i = 1;
blanchet@43020:     val res =
blanchet@43020:       state
blanchet@43020:       |> Proof.map_context (Config.put report false #> Config.put quiet true)
blanchet@43020:       |> try (test_goal (false, false) ([], []) i);
blanchet@43020:   in
wenzelm@51302:     (case res of
wenzelm@58892:       NONE => (unknownN, [])
blanchet@43020:     | SOME results =>
blanchet@43020:         let
wenzelm@59184:           val msg =
wenzelm@59184:             Pretty.string_of
wenzelm@59184:               (pretty_counterex ctxt auto (Option.map (the o get_first response_of) results))
blanchet@43020:         in
wenzelm@59184:           if is_some results then (genuineN, if auto then [msg] else (writeln msg; []))
wenzelm@58892:           else (noneN, [])
wenzelm@51302:         end)
blanchet@43020:   end
wenzelm@51302:   |> `(fn (outcome_code, _) => outcome_code = genuineN);
blanchet@43020: 
wenzelm@67149: val _ = Try.tool_setup (quickcheckN, (20, \<^system_option>\<open>auto_quickcheck\<close>, try_quickcheck));
blanchet@43020: 
haftmann@28315: end;
haftmann@28309: