wenzelm@30826
|
1 |
(* Title: Tools/quickcheck.ML
|
haftmann@28256
|
2 |
Author: Stefan Berghofer, Florian Haftmann, 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@38150
|
9 |
val setup: theory -> theory
|
bulwahn@38150
|
10 |
(* configuration *)
|
wenzelm@32740
|
11 |
val auto: bool Unsynchronized.ref
|
bulwahn@34935
|
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@38169
|
16 |
datatype expectation = No_Expectation | No_Counterexample | Counterexample;
|
bulwahn@38150
|
17 |
datatype test_params = Test_Params of
|
bulwahn@38169
|
18 |
{ size: int, iterations: int, default_type: typ list, no_assms: bool,
|
bulwahn@38169
|
19 |
expect : expectation, report: bool, quiet : bool};
|
bulwahn@39479
|
20 |
val test_params_of: Proof.context -> test_params
|
bulwahn@39480
|
21 |
val report : Proof.context -> bool
|
bulwahn@39480
|
22 |
val set_reporting : bool -> Proof.context -> Proof.context
|
bulwahn@38150
|
23 |
val add_generator:
|
bulwahn@39480
|
24 |
string * (Proof.context -> term -> int -> term list option * (bool list * bool))
|
bulwahn@39479
|
25 |
-> Context.generic -> Context.generic
|
bulwahn@38150
|
26 |
(* testing terms and proof states *)
|
bulwahn@39480
|
27 |
val gen_test_term: Proof.context -> bool -> string option -> int -> int -> term ->
|
bulwahn@35380
|
28 |
(string * term) list option * ((string * int) list * ((int * report list) list) option)
|
wenzelm@30981
|
29 |
val test_term: Proof.context -> bool -> string option -> int -> int -> term ->
|
wenzelm@30981
|
30 |
(string * term) list option
|
bulwahn@38149
|
31 |
val quickcheck: (string * string list) list -> int -> Proof.state -> (string * term) list option
|
haftmann@28256
|
32 |
end;
|
haftmann@28256
|
33 |
|
haftmann@28256
|
34 |
structure Quickcheck : QUICKCHECK =
|
haftmann@28256
|
35 |
struct
|
haftmann@28256
|
36 |
|
wenzelm@30981
|
37 |
(* preferences *)
|
wenzelm@30981
|
38 |
|
wenzelm@32740
|
39 |
val auto = Unsynchronized.ref false;
|
wenzelm@30981
|
40 |
|
bulwahn@34935
|
41 |
val timing = Unsynchronized.ref false;
|
bulwahn@34935
|
42 |
|
wenzelm@30981
|
43 |
val _ =
|
wenzelm@30981
|
44 |
ProofGeneralPgip.add_preference Preferences.category_tracing
|
wenzelm@39397
|
45 |
(setmp_noncritical auto true (fn () =>
|
wenzelm@30981
|
46 |
Preferences.bool_pref auto
|
wenzelm@30981
|
47 |
"auto-quickcheck"
|
blanchet@33552
|
48 |
"Whether to run Quickcheck automatically.") ());
|
wenzelm@30981
|
49 |
|
bulwahn@35378
|
50 |
(* quickcheck report *)
|
bulwahn@35378
|
51 |
|
bulwahn@35378
|
52 |
datatype single_report = Run of bool list * bool | MatchExc
|
bulwahn@35378
|
53 |
|
bulwahn@35378
|
54 |
datatype report = Report of
|
bulwahn@35378
|
55 |
{ iterations : int, raised_match_errors : int,
|
bulwahn@35378
|
56 |
satisfied_assms : int list, positive_concl_tests : int }
|
bulwahn@35378
|
57 |
|
bulwahn@35378
|
58 |
fun collect_single_report single_report
|
bulwahn@35378
|
59 |
(Report {iterations = iterations, raised_match_errors = raised_match_errors,
|
bulwahn@35378
|
60 |
satisfied_assms = satisfied_assms, positive_concl_tests = positive_concl_tests}) =
|
bulwahn@35378
|
61 |
case single_report
|
bulwahn@35378
|
62 |
of MatchExc =>
|
bulwahn@35378
|
63 |
Report {iterations = iterations + 1, raised_match_errors = raised_match_errors + 1,
|
bulwahn@35378
|
64 |
satisfied_assms = satisfied_assms, positive_concl_tests = positive_concl_tests}
|
bulwahn@35378
|
65 |
| Run (assms, concl) =>
|
bulwahn@35378
|
66 |
Report {iterations = iterations + 1, raised_match_errors = raised_match_errors,
|
bulwahn@35378
|
67 |
satisfied_assms =
|
bulwahn@35378
|
68 |
map2 (fn b => fn s => if b then s + 1 else s) assms
|
bulwahn@35378
|
69 |
(if null satisfied_assms then replicate (length assms) 0 else satisfied_assms),
|
bulwahn@35378
|
70 |
positive_concl_tests = if concl then positive_concl_tests + 1 else positive_concl_tests}
|
haftmann@30973
|
71 |
|
bulwahn@38169
|
72 |
(* expectation *)
|
bulwahn@38169
|
73 |
|
bulwahn@38169
|
74 |
datatype expectation = No_Expectation | No_Counterexample | Counterexample;
|
bulwahn@38169
|
75 |
|
bulwahn@38169
|
76 |
fun merge_expectation (expect1, expect2) =
|
bulwahn@38169
|
77 |
if expect1 = expect2 then expect1 else No_Expectation
|
bulwahn@38169
|
78 |
|
haftmann@28315
|
79 |
(* quickcheck configuration -- default parameters, test generators *)
|
haftmann@28315
|
80 |
|
haftmann@28309
|
81 |
datatype test_params = Test_Params of
|
bulwahn@38169
|
82 |
{ size: int, iterations: int, default_type: typ list, no_assms: bool,
|
wenzelm@39034
|
83 |
expect : expectation, report: bool, quiet : bool};
|
haftmann@28309
|
84 |
|
bulwahn@38169
|
85 |
fun dest_test_params (Test_Params { size, iterations, default_type, no_assms, expect, report, quiet }) =
|
bulwahn@38169
|
86 |
((size, iterations), ((default_type, no_assms), ((expect, report), quiet)));
|
wenzelm@39034
|
87 |
|
bulwahn@38169
|
88 |
fun make_test_params ((size, iterations), ((default_type, no_assms), ((expect, report), quiet))) =
|
blanchet@34125
|
89 |
Test_Params { size = size, iterations = iterations, default_type = default_type,
|
wenzelm@39034
|
90 |
no_assms = no_assms, expect = expect, report = report, quiet = quiet };
|
wenzelm@39034
|
91 |
|
bulwahn@38169
|
92 |
fun map_test_params f (Test_Params { size, iterations, default_type, no_assms, expect, report, quiet }) =
|
bulwahn@38169
|
93 |
make_test_params (f ((size, iterations), ((default_type, no_assms), ((expect, report), quiet))));
|
wenzelm@39034
|
94 |
|
wenzelm@39034
|
95 |
fun merge_test_params
|
wenzelm@39034
|
96 |
(Test_Params { size = size1, iterations = iterations1, default_type = default_type1,
|
wenzelm@39034
|
97 |
no_assms = no_assms1, expect = expect1, report = report1, quiet = quiet1 },
|
blanchet@34125
|
98 |
Test_Params { size = size2, iterations = iterations2, default_type = default_type2,
|
wenzelm@39034
|
99 |
no_assms = no_assms2, expect = expect2, report = report2, quiet = quiet2 }) =
|
haftmann@31599
|
100 |
make_test_params ((Int.max (size1, size2), Int.max (iterations1, iterations2)),
|
bulwahn@38151
|
101 |
((merge (op =) (default_type1, default_type2), no_assms1 orelse no_assms2),
|
bulwahn@38169
|
102 |
((merge_expectation (expect1, expect2), report1 orelse report2), quiet1 orelse quiet2)));
|
haftmann@28309
|
103 |
|
bulwahn@39479
|
104 |
structure Data = Generic_Data
|
wenzelm@33522
|
105 |
(
|
wenzelm@39034
|
106 |
type T =
|
bulwahn@39480
|
107 |
(string * (Proof.context -> term -> int -> term list option * (bool list * bool))) list
|
wenzelm@39034
|
108 |
* test_params;
|
bulwahn@35378
|
109 |
val empty = ([], Test_Params
|
bulwahn@38169
|
110 |
{ size = 10, iterations = 100, default_type = [], no_assms = false,
|
bulwahn@38169
|
111 |
expect = No_Expectation, report = false, quiet = false});
|
haftmann@28256
|
112 |
val extend = I;
|
wenzelm@33522
|
113 |
fun merge ((generators1, params1), (generators2, params2)) : T =
|
wenzelm@33522
|
114 |
(AList.merge (op =) (K true) (generators1, generators2),
|
haftmann@28309
|
115 |
merge_test_params (params1, params2));
|
wenzelm@33522
|
116 |
);
|
haftmann@28256
|
117 |
|
bulwahn@39479
|
118 |
val test_params_of = snd o Data.get o Context.Proof;
|
bulwahn@38150
|
119 |
|
bulwahn@39480
|
120 |
val report = snd o fst o snd o snd o dest_test_params o test_params_of
|
bulwahn@39480
|
121 |
|
bulwahn@39480
|
122 |
fun map_report f (Test_Params { size, iterations, default_type, no_assms, expect, report, quiet }) =
|
bulwahn@39480
|
123 |
make_test_params ((size, iterations), ((default_type, no_assms), ((expect, f report), quiet)));
|
bulwahn@39480
|
124 |
|
bulwahn@39480
|
125 |
fun set_reporting report = Context.proof_map (Data.map (apsnd (map_report (K report))))
|
bulwahn@39480
|
126 |
|
haftmann@28309
|
127 |
val add_generator = Data.map o apfst o AList.update (op =);
|
haftmann@28256
|
128 |
|
haftmann@28315
|
129 |
(* generating tests *)
|
haftmann@28315
|
130 |
|
haftmann@28309
|
131 |
fun mk_tester_select name ctxt =
|
bulwahn@39479
|
132 |
case AList.lookup (op =) ((fst o Data.get o Context.Proof) ctxt) name
|
haftmann@28309
|
133 |
of NONE => error ("No such quickcheck generator: " ^ name)
|
haftmann@28309
|
134 |
| SOME generator => generator ctxt;
|
haftmann@28256
|
135 |
|
bulwahn@39480
|
136 |
fun mk_testers ctxt t =
|
bulwahn@39479
|
137 |
(map snd o fst o Data.get o Context.Proof) ctxt
|
bulwahn@39480
|
138 |
|> map_filter (fn generator => try (generator ctxt) t);
|
haftmann@28256
|
139 |
|
bulwahn@39480
|
140 |
fun mk_testers_strict ctxt t =
|
haftmann@28309
|
141 |
let
|
bulwahn@39479
|
142 |
val generators = ((map snd o fst o Data.get o Context.Proof) ctxt)
|
bulwahn@39480
|
143 |
val testers = map (fn generator => Exn.capture (generator ctxt) t) generators;
|
haftmann@28309
|
144 |
in if forall (is_none o Exn.get_result) testers
|
haftmann@28309
|
145 |
then [(Exn.release o snd o split_last) testers]
|
haftmann@28309
|
146 |
else map_filter Exn.get_result testers
|
haftmann@28309
|
147 |
end;
|
haftmann@28309
|
148 |
|
haftmann@28315
|
149 |
|
haftmann@28315
|
150 |
(* testing propositions *)
|
haftmann@28315
|
151 |
|
haftmann@28309
|
152 |
fun prep_test_term t =
|
haftmann@28309
|
153 |
let
|
wenzelm@29266
|
154 |
val _ = (null (Term.add_tvars t []) andalso null (Term.add_tfrees t [])) orelse
|
haftmann@28309
|
155 |
error "Term to be tested contains type variables";
|
wenzelm@29266
|
156 |
val _ = null (Term.add_vars t []) orelse
|
haftmann@28309
|
157 |
error "Term to be tested contains schematic variables";
|
haftmann@31138
|
158 |
val frees = Term.add_frees t [];
|
haftmann@28309
|
159 |
in (map fst frees, list_abs_free (frees, t)) end
|
haftmann@28309
|
160 |
|
bulwahn@35324
|
161 |
fun cpu_time description f =
|
bulwahn@35324
|
162 |
let
|
bulwahn@35324
|
163 |
val start = start_timing ()
|
bulwahn@35324
|
164 |
val result = Exn.capture f ()
|
bulwahn@35324
|
165 |
val time = Time.toMilliseconds (#cpu (end_timing start))
|
bulwahn@35324
|
166 |
in (Exn.release result, (description, time)) end
|
bulwahn@35324
|
167 |
|
bulwahn@39480
|
168 |
fun gen_test_term ctxt quiet generator_name size i t =
|
haftmann@28309
|
169 |
let
|
haftmann@28309
|
170 |
val (names, t') = prep_test_term t;
|
bulwahn@35324
|
171 |
val (testers, comp_time) = cpu_time "quickcheck compilation"
|
bulwahn@35324
|
172 |
(fn () => (case generator_name
|
bulwahn@39480
|
173 |
of NONE => if quiet then mk_testers ctxt t' else mk_testers_strict ctxt t'
|
bulwahn@39480
|
174 |
| SOME name => [mk_tester_select name ctxt t']));
|
bulwahn@35378
|
175 |
fun iterate f 0 report = (NONE, report)
|
bulwahn@35378
|
176 |
| iterate f j report =
|
bulwahn@35378
|
177 |
let
|
bulwahn@35378
|
178 |
val (test_result, single_report) = apsnd Run (f ()) handle Match => (if quiet then ()
|
bulwahn@35378
|
179 |
else warning "Exception Match raised during quickcheck"; (NONE, MatchExc))
|
bulwahn@35378
|
180 |
val report = collect_single_report single_report report
|
bulwahn@35378
|
181 |
in
|
bulwahn@35378
|
182 |
case test_result of NONE => iterate f (j - 1) report | SOME q => (SOME q, report)
|
bulwahn@35378
|
183 |
end
|
bulwahn@35378
|
184 |
val empty_report = Report { iterations = 0, raised_match_errors = 0,
|
bulwahn@35378
|
185 |
satisfied_assms = [], positive_concl_tests = 0 }
|
bulwahn@35378
|
186 |
fun with_testers k [] = (NONE, [])
|
haftmann@28309
|
187 |
| with_testers k (tester :: testers) =
|
bulwahn@35378
|
188 |
case iterate (fn () => tester (k - 1)) i empty_report
|
bulwahn@35378
|
189 |
of (NONE, report) => apsnd (cons report) (with_testers k testers)
|
bulwahn@35378
|
190 |
| (SOME q, report) => (SOME q, [report]);
|
bulwahn@35378
|
191 |
fun with_size k reports = if k > size then (NONE, reports)
|
haftmann@28309
|
192 |
else (if quiet then () else priority ("Test data size: " ^ string_of_int k);
|
bulwahn@35378
|
193 |
let
|
bulwahn@35378
|
194 |
val (result, new_report) = with_testers k testers
|
bulwahn@35378
|
195 |
val reports = ((k, new_report) :: reports)
|
bulwahn@35378
|
196 |
in case result of NONE => with_size (k + 1) reports | SOME q => (SOME q, reports) end);
|
bulwahn@35378
|
197 |
val ((result, reports), exec_time) = cpu_time "quickcheck execution"
|
bulwahn@35378
|
198 |
(fn () => apfst
|
bulwahn@35378
|
199 |
(fn result => case result of NONE => NONE
|
bulwahn@35378
|
200 |
| SOME ts => SOME (names ~~ ts)) (with_size 1 []))
|
bulwahn@34935
|
201 |
in
|
bulwahn@39480
|
202 |
(result, ([exec_time, comp_time], if report ctxt then SOME reports else NONE))
|
haftmann@28309
|
203 |
end;
|
haftmann@28309
|
204 |
|
bulwahn@35324
|
205 |
fun test_term ctxt quiet generator_name size i t =
|
bulwahn@39480
|
206 |
ctxt
|
bulwahn@39480
|
207 |
|> set_reporting false
|
bulwahn@39480
|
208 |
|> (fn ctxt' => fst (gen_test_term ctxt' quiet generator_name size i t))
|
bulwahn@35324
|
209 |
|
bulwahn@38153
|
210 |
exception WELLSORTED of string
|
bulwahn@38153
|
211 |
|
haftmann@28309
|
212 |
fun monomorphic_term thy insts default_T =
|
haftmann@28309
|
213 |
let
|
haftmann@28309
|
214 |
fun subst (T as TFree (v, S)) =
|
haftmann@28309
|
215 |
let
|
haftmann@28309
|
216 |
val T' = AList.lookup (op =) insts v
|
bulwahn@38152
|
217 |
|> the_default default_T
|
bulwahn@38153
|
218 |
in if Sign.of_sort thy (T', S) then T'
|
bulwahn@38153
|
219 |
else raise (WELLSORTED ("For instantiation with default_type " ^ Syntax.string_of_typ_global thy default_T ^
|
bulwahn@38153
|
220 |
":\n" ^ Syntax.string_of_typ_global thy T' ^
|
haftmann@28309
|
221 |
" to be substituted for variable " ^
|
bulwahn@38153
|
222 |
Syntax.string_of_typ_global thy T ^ " does not have sort " ^
|
bulwahn@38153
|
223 |
Syntax.string_of_sort_global thy S))
|
haftmann@28309
|
224 |
end
|
haftmann@28309
|
225 |
| subst T = T;
|
haftmann@28309
|
226 |
in (map_types o map_atyps) subst end;
|
haftmann@28309
|
227 |
|
bulwahn@38153
|
228 |
datatype wellsorted_error = Wellsorted_Error of string | Term of term
|
bulwahn@38153
|
229 |
|
bulwahn@39480
|
230 |
fun test_goal quiet generator_name size iterations default_Ts no_assms insts i state =
|
haftmann@28309
|
231 |
let
|
haftmann@38211
|
232 |
val lthy = Proof.context_of state;
|
haftmann@28309
|
233 |
val thy = Proof.theory_of state;
|
haftmann@28309
|
234 |
fun strip (Const ("all", _) $ Abs (_, _, t)) = strip t
|
haftmann@28309
|
235 |
| strip t = t;
|
wenzelm@33291
|
236 |
val {goal = st, ...} = Proof.raw_goal state;
|
haftmann@28309
|
237 |
val (gi, frees) = Logic.goal_params (prop_of st) i;
|
haftmann@38616
|
238 |
val some_locale = case (Option.map #target o Named_Target.peek) lthy
|
haftmann@38616
|
239 |
of NONE => NONE
|
haftmann@38616
|
240 |
| SOME "" => NONE
|
haftmann@38616
|
241 |
| SOME locale => SOME locale;
|
haftmann@38211
|
242 |
val assms = if no_assms then [] else case some_locale
|
haftmann@38211
|
243 |
of NONE => Assumption.all_assms_of lthy
|
haftmann@38211
|
244 |
| SOME locale => Assumption.local_assms_of lthy (Locale.init locale thy);
|
haftmann@38211
|
245 |
val proto_goal = Logic.list_implies (map Thm.term_of assms, subst_bounds (frees, strip gi));
|
haftmann@38211
|
246 |
val check_goals = case some_locale
|
haftmann@38211
|
247 |
of NONE => [proto_goal]
|
ballarin@38357
|
248 |
| SOME locale => map (fn (_, phi) => Morphism.term phi proto_goal) (Locale.registrations_of (Context.Theory thy) (*FIXME*) locale);
|
haftmann@38211
|
249 |
val inst_goals = maps (fn check_goal => map (fn T =>
|
haftmann@38211
|
250 |
Term ((Object_Logic.atomize_term thy o monomorphic_term thy insts T) check_goal)
|
haftmann@38211
|
251 |
handle WELLSORTED s => Wellsorted_Error s) default_Ts) check_goals
|
bulwahn@38153
|
252 |
val error_msg = cat_lines (map_filter (fn Term t => NONE | Wellsorted_Error s => SOME s) inst_goals)
|
bulwahn@38153
|
253 |
val correct_inst_goals =
|
bulwahn@38153
|
254 |
case map_filter (fn Term t => SOME t | Wellsorted_Error s => NONE) inst_goals of
|
bulwahn@38153
|
255 |
[] => error error_msg
|
bulwahn@38153
|
256 |
| xs => xs
|
bulwahn@38153
|
257 |
val _ = if quiet then () else warning error_msg
|
bulwahn@38152
|
258 |
fun collect_results f reports [] = (NONE, rev reports)
|
bulwahn@38152
|
259 |
| collect_results f reports (t :: ts) =
|
bulwahn@38152
|
260 |
case f t of
|
bulwahn@38152
|
261 |
(SOME res, report) => (SOME res, rev (report :: reports))
|
bulwahn@38152
|
262 |
| (NONE, report) => collect_results f (report :: reports) ts
|
bulwahn@39480
|
263 |
in collect_results (gen_test_term lthy quiet generator_name size iterations) [] correct_inst_goals end;
|
bulwahn@38152
|
264 |
|
bulwahn@38152
|
265 |
(* pretty printing *)
|
haftmann@28315
|
266 |
|
blanchet@35074
|
267 |
fun pretty_counterex ctxt NONE = Pretty.str "Quickcheck found no counterexample."
|
haftmann@28315
|
268 |
| pretty_counterex ctxt (SOME cex) =
|
blanchet@35074
|
269 |
Pretty.chunks (Pretty.str "Quickcheck found a counterexample:\n" ::
|
haftmann@28315
|
270 |
map (fn (s, t) =>
|
haftmann@28315
|
271 |
Pretty.block [Pretty.str (s ^ " ="), Pretty.brk 1, Syntax.pretty_term ctxt t]) cex);
|
haftmann@28315
|
272 |
|
bulwahn@35378
|
273 |
fun pretty_report (Report {iterations = iterations, raised_match_errors = raised_match_errors,
|
bulwahn@35378
|
274 |
satisfied_assms = satisfied_assms, positive_concl_tests = positive_concl_tests}) =
|
bulwahn@35378
|
275 |
let
|
bulwahn@35378
|
276 |
fun pretty_stat s i = Pretty.block ([Pretty.str (s ^ ": " ^ string_of_int i)])
|
bulwahn@35378
|
277 |
in
|
bulwahn@35378
|
278 |
([pretty_stat "iterations" iterations,
|
bulwahn@35378
|
279 |
pretty_stat "match exceptions" raised_match_errors]
|
bulwahn@35378
|
280 |
@ map_index (fn (i, n) => pretty_stat ("satisfied " ^ string_of_int (i + 1) ^ ". assumption") n)
|
bulwahn@35378
|
281 |
satisfied_assms
|
bulwahn@35378
|
282 |
@ [pretty_stat "positive conclusion tests" positive_concl_tests])
|
bulwahn@35378
|
283 |
end
|
bulwahn@35378
|
284 |
|
bulwahn@35378
|
285 |
fun pretty_reports' [report] = [Pretty.chunks (pretty_report report)]
|
bulwahn@35378
|
286 |
| pretty_reports' reports =
|
bulwahn@35378
|
287 |
map_index (fn (i, report) =>
|
bulwahn@35378
|
288 |
Pretty.chunks (Pretty.str (string_of_int (i + 1) ^ ". generator:\n") :: pretty_report report))
|
bulwahn@35378
|
289 |
reports
|
bulwahn@35378
|
290 |
|
bulwahn@35380
|
291 |
fun pretty_reports ctxt (SOME reports) =
|
bulwahn@35378
|
292 |
Pretty.chunks (Pretty.str "Quickcheck report:" ::
|
bulwahn@35378
|
293 |
maps (fn (size, reports) =>
|
bulwahn@35378
|
294 |
Pretty.str ("size " ^ string_of_int size ^ ":") :: pretty_reports' reports @ [Pretty.brk 1])
|
bulwahn@35378
|
295 |
(rev reports))
|
bulwahn@35380
|
296 |
| pretty_reports ctxt NONE = Pretty.str ""
|
bulwahn@35378
|
297 |
|
bulwahn@38152
|
298 |
fun pretty_counterex_and_reports ctxt (cex, timing_and_reports) =
|
bulwahn@38152
|
299 |
Pretty.chunks (pretty_counterex ctxt cex :: map (pretty_reports ctxt) (map snd timing_and_reports))
|
haftmann@28315
|
300 |
|
haftmann@28315
|
301 |
(* automatic testing *)
|
haftmann@28309
|
302 |
|
blanchet@33552
|
303 |
fun auto_quickcheck state =
|
blanchet@33552
|
304 |
if not (!auto) then
|
blanchet@33552
|
305 |
(false, state)
|
blanchet@33552
|
306 |
else
|
blanchet@33552
|
307 |
let
|
blanchet@33552
|
308 |
val ctxt = Proof.context_of state;
|
bulwahn@38169
|
309 |
val Test_Params {size, iterations, default_type, no_assms, ...} =
|
bulwahn@39479
|
310 |
(snd o Data.get o Context.Proof) ctxt;
|
blanchet@33552
|
311 |
val res =
|
bulwahn@39480
|
312 |
state
|
bulwahn@39480
|
313 |
|> Proof.map_context (set_reporting false)
|
bulwahn@39480
|
314 |
|> try (test_goal true NONE size iterations default_type no_assms [] 1);
|
blanchet@33552
|
315 |
in
|
blanchet@33552
|
316 |
case res of
|
blanchet@33552
|
317 |
NONE => (false, state)
|
bulwahn@35378
|
318 |
| SOME (NONE, report) => (false, state)
|
bulwahn@35378
|
319 |
| SOME (cex, report) => (true, Proof.goal_message (K (Pretty.chunks [Pretty.str "",
|
blanchet@33552
|
320 |
Pretty.mark Markup.hilite (pretty_counterex ctxt cex)])) state)
|
blanchet@33552
|
321 |
end
|
blanchet@33552
|
322 |
|
blanchet@39570
|
323 |
val setup = Auto_Tools.register_tool ("quickcheck", auto_quickcheck)
|
haftmann@28309
|
324 |
|
haftmann@28309
|
325 |
|
wenzelm@30981
|
326 |
(* Isar commands *)
|
haftmann@28315
|
327 |
|
haftmann@28336
|
328 |
fun read_nat s = case (Library.read_int o Symbol.explode) s
|
haftmann@28336
|
329 |
of (k, []) => if k >= 0 then k
|
haftmann@28336
|
330 |
else error ("Not a natural number: " ^ s)
|
haftmann@28336
|
331 |
| (_, _ :: _) => error ("Not a natural number: " ^ s);
|
bulwahn@38149
|
332 |
|
blanchet@34125
|
333 |
fun read_bool "false" = false
|
blanchet@34125
|
334 |
| read_bool "true" = true
|
blanchet@34125
|
335 |
| read_bool s = error ("Not a Boolean value: " ^ s)
|
haftmann@28315
|
336 |
|
bulwahn@38169
|
337 |
fun read_expectation "no_expectation" = No_Expectation
|
bulwahn@38169
|
338 |
| read_expectation "no_counterexample" = No_Counterexample
|
bulwahn@38169
|
339 |
| read_expectation "counterexample" = Counterexample
|
bulwahn@38169
|
340 |
| read_expectation s = error ("Not an expectation value: " ^ s)
|
bulwahn@38169
|
341 |
|
bulwahn@38149
|
342 |
fun parse_test_param ctxt ("size", [arg]) =
|
haftmann@28336
|
343 |
(apfst o apfst o K) (read_nat arg)
|
bulwahn@38149
|
344 |
| parse_test_param ctxt ("iterations", [arg]) =
|
haftmann@28336
|
345 |
(apfst o apsnd o K) (read_nat arg)
|
haftmann@28336
|
346 |
| parse_test_param ctxt ("default_type", arg) =
|
bulwahn@38149
|
347 |
(apsnd o apfst o apfst o K) (map (ProofContext.read_typ ctxt) arg)
|
bulwahn@38149
|
348 |
| parse_test_param ctxt ("no_assms", [arg]) =
|
bulwahn@35378
|
349 |
(apsnd o apfst o apsnd o K) (read_bool arg)
|
bulwahn@38169
|
350 |
| parse_test_param ctxt ("expect", [arg]) =
|
bulwahn@38169
|
351 |
(apsnd o apsnd o apfst o apfst o K) (read_expectation arg)
|
bulwahn@38149
|
352 |
| parse_test_param ctxt ("report", [arg]) =
|
bulwahn@38169
|
353 |
(apsnd o apsnd o apfst o apsnd o K) (read_bool arg)
|
bulwahn@38149
|
354 |
| parse_test_param ctxt ("quiet", [arg]) =
|
bulwahn@38169
|
355 |
(apsnd o apsnd o apsnd o K) (read_bool arg)
|
haftmann@28336
|
356 |
| parse_test_param ctxt (name, _) =
|
blanchet@34125
|
357 |
error ("Unknown test parameter: " ^ name);
|
haftmann@28315
|
358 |
|
bulwahn@38149
|
359 |
fun parse_test_param_inst ctxt ("generator", [arg]) =
|
haftmann@28336
|
360 |
(apsnd o apfst o K o SOME) arg
|
haftmann@28336
|
361 |
| parse_test_param_inst ctxt (name, arg) =
|
haftmann@28336
|
362 |
case try (ProofContext.read_typ ctxt) name
|
haftmann@28336
|
363 |
of SOME (TFree (v, _)) => (apsnd o apsnd o AList.update (op =))
|
bulwahn@38149
|
364 |
(v, ProofContext.read_typ ctxt (the_single arg))
|
haftmann@28336
|
365 |
| _ => (apfst o parse_test_param ctxt) (name, arg);
|
haftmann@28315
|
366 |
|
haftmann@28336
|
367 |
fun quickcheck_params_cmd args thy =
|
haftmann@28315
|
368 |
let
|
bulwahn@39479
|
369 |
val ctxt = ProofContext.init_global thy
|
haftmann@28336
|
370 |
val f = fold (parse_test_param ctxt) args;
|
haftmann@28315
|
371 |
in
|
haftmann@28315
|
372 |
thy
|
bulwahn@39479
|
373 |
|> (Context.theory_map o Data.map o apsnd o map_test_params) f
|
haftmann@28315
|
374 |
end;
|
haftmann@28315
|
375 |
|
bulwahn@35378
|
376 |
fun gen_quickcheck args i state =
|
haftmann@28315
|
377 |
let
|
boehmes@32295
|
378 |
val ctxt = Proof.context_of state;
|
bulwahn@39479
|
379 |
val default_params = (dest_test_params o snd o Data.get o Context.Proof) ctxt;
|
haftmann@28336
|
380 |
val f = fold (parse_test_param_inst ctxt) args;
|
bulwahn@38169
|
381 |
val (((size, iterations), ((default_type, no_assms), ((expect, report), quiet))), (generator_name, insts)) =
|
haftmann@28336
|
382 |
f (default_params, (NONE, []));
|
boehmes@32295
|
383 |
in
|
bulwahn@39480
|
384 |
state
|
bulwahn@39480
|
385 |
|> Proof.map_context (set_reporting report)
|
bulwahn@39480
|
386 |
|> test_goal quiet generator_name size iterations default_type no_assms insts i
|
bulwahn@38169
|
387 |
|> tap (fn (SOME x, _) => if expect = No_Counterexample then
|
haftmann@38211
|
388 |
error ("quickcheck expected to find no counterexample but found one") else ()
|
bulwahn@38169
|
389 |
| (NONE, _) => if expect = Counterexample then
|
bulwahn@38169
|
390 |
error ("quickcheck expected to find a counterexample but did not find one") else ())
|
boehmes@32295
|
391 |
end;
|
boehmes@32295
|
392 |
|
wenzelm@36970
|
393 |
fun quickcheck args i state = fst (gen_quickcheck args i state);
|
bulwahn@35378
|
394 |
|
boehmes@32295
|
395 |
fun quickcheck_cmd args i state =
|
bulwahn@35378
|
396 |
gen_quickcheck args i (Toplevel.proof_of state)
|
bulwahn@35378
|
397 |
|> Pretty.writeln o pretty_counterex_and_reports (Toplevel.context_of state);
|
haftmann@28309
|
398 |
|
bulwahn@38149
|
399 |
val parse_arg = Parse.name -- (Scan.optional (Parse.$$$ "=" |--
|
bulwahn@38149
|
400 |
((Parse.name >> single) || (Parse.$$$ "[" |-- Parse.list1 Parse.name --| Parse.$$$ "]"))) ["true"]);
|
haftmann@28309
|
401 |
|
wenzelm@36970
|
402 |
val parse_args = Parse.$$$ "[" |-- Parse.list1 parse_arg --| Parse.$$$ "]"
|
haftmann@28336
|
403 |
|| Scan.succeed [];
|
haftmann@28336
|
404 |
|
wenzelm@36970
|
405 |
val _ =
|
wenzelm@36970
|
406 |
Outer_Syntax.command "quickcheck_params" "set parameters for random testing" Keyword.thy_decl
|
wenzelm@36970
|
407 |
(parse_args >> (fn args => Toplevel.theory (quickcheck_params_cmd args)));
|
haftmann@28309
|
408 |
|
wenzelm@36970
|
409 |
val _ =
|
wenzelm@36970
|
410 |
Outer_Syntax.improper_command "quickcheck" "try to find counterexample for subgoal" Keyword.diag
|
wenzelm@36970
|
411 |
(parse_args -- Scan.optional Parse.nat 1
|
wenzelm@36970
|
412 |
>> (fn (args, i) => Toplevel.no_timing o Toplevel.keep (quickcheck_cmd args i)));
|
haftmann@28309
|
413 |
|
haftmann@28309
|
414 |
end;
|
haftmann@28256
|
415 |
|
haftmann@28315
|
416 |
|
haftmann@28315
|
417 |
val auto_quickcheck = Quickcheck.auto;
|