blanchet@39692
|
1 |
(* Title: HOL/Tools/ATP/atp_proof.ML
|
blanchet@39692
|
2 |
Author: Lawrence C. Paulson, Cambridge University Computer Laboratory
|
blanchet@39692
|
3 |
Author: Claire Quigley, Cambridge University Computer Laboratory
|
blanchet@39692
|
4 |
Author: Jasmin Blanchette, TU Muenchen
|
blanchet@39692
|
5 |
|
blanchet@43745
|
6 |
Abstract representation of ATP proofs and TSTP/SPASS syntax.
|
blanchet@39692
|
7 |
*)
|
blanchet@39692
|
8 |
|
blanchet@39692
|
9 |
signature ATP_PROOF =
|
blanchet@39692
|
10 |
sig
|
nik@44537
|
11 |
type ('a, 'b) ho_term = ('a, 'b) ATP_Problem.ho_term
|
blanchet@43402
|
12 |
type ('a, 'b, 'c) formula = ('a, 'b, 'c) ATP_Problem.formula
|
blanchet@43784
|
13 |
type 'a problem = 'a ATP_Problem.problem
|
blanchet@39692
|
14 |
|
blanchet@43806
|
15 |
exception UNRECOGNIZED_ATP_PROOF of unit
|
blanchet@43806
|
16 |
|
blanchet@39731
|
17 |
datatype failure =
|
blanchet@43458
|
18 |
Unprovable |
|
blanchet@43891
|
19 |
GaveUp |
|
blanchet@43458
|
20 |
ProofMissing |
|
blanchet@43751
|
21 |
ProofIncomplete |
|
blanchet@45786
|
22 |
UnsoundProof of bool * string list |
|
blanchet@43458
|
23 |
CantConnect |
|
blanchet@43458
|
24 |
TimedOut |
|
blanchet@43794
|
25 |
Inappropriate |
|
blanchet@43458
|
26 |
OutOfResources |
|
blanchet@43458
|
27 |
SpassTooOld |
|
blanchet@43458
|
28 |
VampireTooOld |
|
blanchet@43458
|
29 |
NoPerl |
|
blanchet@43458
|
30 |
NoLibwwwPerl |
|
blanchet@43458
|
31 |
MalformedInput |
|
blanchet@43458
|
32 |
MalformedOutput |
|
blanchet@43458
|
33 |
Interrupted |
|
blanchet@43458
|
34 |
Crashed |
|
blanchet@43458
|
35 |
InternalError |
|
blanchet@43458
|
36 |
UnknownError of string
|
blanchet@39731
|
37 |
|
blanchet@44352
|
38 |
type step_name = string * string list option
|
blanchet@39692
|
39 |
|
blanchet@39693
|
40 |
datatype 'a step =
|
blanchet@39693
|
41 |
Definition of step_name * 'a * 'a |
|
blanchet@39693
|
42 |
Inference of step_name * 'a * step_name list
|
blanchet@39692
|
43 |
|
nik@44537
|
44 |
type 'a proof = ('a, 'a, ('a, 'a) ho_term) formula step list
|
blanchet@39692
|
45 |
|
blanchet@41505
|
46 |
val short_output : bool -> string -> string
|
blanchet@42615
|
47 |
val string_for_failure : failure -> string
|
blanchet@39731
|
48 |
val extract_important_message : string -> string
|
blanchet@39731
|
49 |
val extract_known_failure :
|
blanchet@39731
|
50 |
(failure * string) list -> string -> failure option
|
blanchet@39731
|
51 |
val extract_tstplike_proof_and_outcome :
|
blanchet@44344
|
52 |
bool -> bool -> (string * string) list -> (failure * string) list -> string
|
blanchet@44344
|
53 |
-> string * failure option
|
blanchet@43809
|
54 |
val is_same_atp_step : step_name -> step_name -> bool
|
blanchet@43802
|
55 |
val scan_general_id : string list -> string * string list
|
blanchet@43802
|
56 |
val parse_formula :
|
nik@44537
|
57 |
string list -> (string, 'a, (string, 'a) ho_term) formula * string list
|
blanchet@44352
|
58 |
val atp_proof_from_tstplike_proof :
|
blanchet@44352
|
59 |
string problem -> string -> string -> string proof
|
blanchet@43809
|
60 |
val clean_up_atp_proof_dependencies : string proof -> string proof
|
blanchet@39694
|
61 |
val map_term_names_in_atp_proof :
|
blanchet@39694
|
62 |
(string -> string) -> string proof -> string proof
|
blanchet@39694
|
63 |
val nasty_atp_proof : string Symtab.table -> string proof -> string proof
|
blanchet@39692
|
64 |
end;
|
blanchet@39692
|
65 |
|
blanchet@39692
|
66 |
structure ATP_Proof : ATP_PROOF =
|
blanchet@39692
|
67 |
struct
|
blanchet@39692
|
68 |
|
blanchet@43926
|
69 |
open ATP_Util
|
blanchet@39731
|
70 |
open ATP_Problem
|
blanchet@39731
|
71 |
|
blanchet@43806
|
72 |
exception UNRECOGNIZED_ATP_PROOF of unit
|
blanchet@43806
|
73 |
|
blanchet@39731
|
74 |
datatype failure =
|
blanchet@43458
|
75 |
Unprovable |
|
blanchet@43891
|
76 |
GaveUp |
|
blanchet@43458
|
77 |
ProofMissing |
|
blanchet@43751
|
78 |
ProofIncomplete |
|
blanchet@45786
|
79 |
UnsoundProof of bool * string list |
|
blanchet@43458
|
80 |
CantConnect |
|
blanchet@43458
|
81 |
TimedOut |
|
blanchet@43794
|
82 |
Inappropriate |
|
blanchet@43458
|
83 |
OutOfResources |
|
blanchet@43458
|
84 |
SpassTooOld |
|
blanchet@43458
|
85 |
VampireTooOld |
|
blanchet@43458
|
86 |
NoPerl |
|
blanchet@43458
|
87 |
NoLibwwwPerl |
|
blanchet@43458
|
88 |
MalformedInput |
|
blanchet@43458
|
89 |
MalformedOutput |
|
blanchet@43458
|
90 |
Interrupted |
|
blanchet@43458
|
91 |
Crashed |
|
blanchet@43458
|
92 |
InternalError |
|
blanchet@43458
|
93 |
UnknownError of string
|
blanchet@39731
|
94 |
|
blanchet@41505
|
95 |
fun elide_string threshold s =
|
blanchet@41505
|
96 |
if size s > threshold then
|
blanchet@41505
|
97 |
String.extract (s, 0, SOME (threshold div 2 - 5)) ^ " ...... " ^
|
blanchet@41505
|
98 |
String.extract (s, size s - (threshold + 1) div 2 + 6, NONE)
|
blanchet@41505
|
99 |
else
|
blanchet@41505
|
100 |
s
|
blanchet@41505
|
101 |
fun short_output verbose output =
|
blanchet@42915
|
102 |
if verbose then
|
blanchet@42915
|
103 |
if output = "" then "No details available" else elide_string 1000 output
|
blanchet@42915
|
104 |
else
|
blanchet@42915
|
105 |
""
|
blanchet@41505
|
106 |
|
blanchet@42615
|
107 |
val missing_message_tail =
|
blanchet@42615
|
108 |
" appears to be missing. You will need to install it if you want to invoke \
|
blanchet@42615
|
109 |
\remote provers."
|
blanchet@39731
|
110 |
|
blanchet@43745
|
111 |
fun involving [] = ""
|
blanchet@43846
|
112 |
| involving ss =
|
blanchet@43870
|
113 |
"involving " ^ space_implode " " (Try.serial_commas "and" (map quote ss)) ^
|
blanchet@43870
|
114 |
" "
|
blanchet@43745
|
115 |
|
blanchet@43809
|
116 |
fun string_for_failure Unprovable = "The problem is unprovable."
|
blanchet@43891
|
117 |
| string_for_failure GaveUp = "The prover gave up."
|
blanchet@42615
|
118 |
| string_for_failure ProofMissing =
|
blanchet@42615
|
119 |
"The prover claims the conjecture is a theorem but did not provide a proof."
|
blanchet@43751
|
120 |
| string_for_failure ProofIncomplete =
|
blanchet@43751
|
121 |
"The prover claims the conjecture is a theorem but provided an incomplete \
|
blanchet@43751
|
122 |
\proof."
|
blanchet@45786
|
123 |
| string_for_failure (UnsoundProof (false, ss)) =
|
blanchet@44687
|
124 |
"The prover found a type-unsound proof " ^ involving ss ^
|
blanchet@44687
|
125 |
"(or, less likely, your axioms are inconsistent). Specify a sound type \
|
blanchet@44687
|
126 |
\encoding or omit the \"type_enc\" option."
|
blanchet@45786
|
127 |
| string_for_failure (UnsoundProof (true, ss)) =
|
blanchet@43745
|
128 |
"The prover found a type-unsound proof " ^ involving ss ^
|
blanchet@43745
|
129 |
"even though a supposedly type-sound encoding was used (or, less likely, \
|
blanchet@44326
|
130 |
\your axioms are inconsistent). Please report this to the Isabelle \
|
blanchet@44326
|
131 |
\developers."
|
blanchet@42615
|
132 |
| string_for_failure CantConnect = "Cannot connect to remote server."
|
blanchet@42615
|
133 |
| string_for_failure TimedOut = "Timed out."
|
blanchet@43794
|
134 |
| string_for_failure Inappropriate =
|
blanchet@43794
|
135 |
"The problem lies outside the prover's scope."
|
blanchet@42615
|
136 |
| string_for_failure OutOfResources = "The prover ran out of resources."
|
blanchet@42615
|
137 |
| string_for_failure SpassTooOld =
|
blanchet@39731
|
138 |
"Isabelle requires a more recent version of SPASS with support for the \
|
blanchet@39731
|
139 |
\TPTP syntax. To install it, download and extract the package \
|
blanchet@39731
|
140 |
\\"http://isabelle.in.tum.de/dist/contrib/spass-3.7.tar.gz\" and add the \
|
blanchet@39731
|
141 |
\\"spass-3.7\" directory's absolute path to " ^
|
wenzelm@44475
|
142 |
Path.print (Path.expand (Path.explode "$ISABELLE_HOME_USER/etc/components")) ^
|
blanchet@39731
|
143 |
" on a line of its own."
|
blanchet@42615
|
144 |
| string_for_failure VampireTooOld =
|
blanchet@39731
|
145 |
"Isabelle requires a more recent version of Vampire. To install it, follow \
|
blanchet@39731
|
146 |
\the instructions from the Sledgehammer manual (\"isabelle doc\
|
blanchet@39731
|
147 |
\ sledgehammer\")."
|
blanchet@42615
|
148 |
| string_for_failure NoPerl = "Perl" ^ missing_message_tail
|
blanchet@42615
|
149 |
| string_for_failure NoLibwwwPerl =
|
blanchet@42615
|
150 |
"The Perl module \"libwww-perl\"" ^ missing_message_tail
|
blanchet@42615
|
151 |
| string_for_failure MalformedInput =
|
blanchet@42615
|
152 |
"The generated problem is malformed. Please report this to the Isabelle \
|
blanchet@42615
|
153 |
\developers."
|
blanchet@42615
|
154 |
| string_for_failure MalformedOutput = "The prover output is malformed."
|
blanchet@43926
|
155 |
| string_for_failure Interrupted = "The prover was interrupted."
|
blanchet@42615
|
156 |
| string_for_failure Crashed = "The prover crashed."
|
blanchet@42615
|
157 |
| string_for_failure InternalError = "An internal prover error occurred."
|
blanchet@42615
|
158 |
| string_for_failure (UnknownError string) =
|
blanchet@42615
|
159 |
"A prover error occurred" ^
|
blanchet@41582
|
160 |
(if string = "" then ". (Pass the \"verbose\" option for details.)"
|
blanchet@41582
|
161 |
else ":\n" ^ string)
|
blanchet@39731
|
162 |
|
blanchet@39731
|
163 |
fun extract_delimited (begin_delim, end_delim) output =
|
blanchet@39731
|
164 |
output |> first_field begin_delim |> the |> snd
|
blanchet@39731
|
165 |
|> first_field end_delim |> the |> fst
|
blanchet@39731
|
166 |
|> first_field "\n" |> the |> snd
|
blanchet@39731
|
167 |
handle Option.Option => ""
|
blanchet@39731
|
168 |
|
blanchet@39731
|
169 |
val tstp_important_message_delims =
|
blanchet@39731
|
170 |
("% SZS start RequiredInformation", "% SZS end RequiredInformation")
|
blanchet@39731
|
171 |
|
blanchet@39731
|
172 |
fun extract_important_message output =
|
blanchet@39731
|
173 |
case extract_delimited tstp_important_message_delims output of
|
blanchet@39731
|
174 |
"" => ""
|
blanchet@39731
|
175 |
| s => s |> space_explode "\n" |> filter_out (curry (op =) "")
|
blanchet@39731
|
176 |
|> map (perhaps (try (unprefix "%")))
|
blanchet@39731
|
177 |
|> map (perhaps (try (unprefix " ")))
|
blanchet@39731
|
178 |
|> space_implode "\n " |> quote
|
blanchet@39731
|
179 |
|
blanchet@39731
|
180 |
(* Splits by the first possible of a list of delimiters. *)
|
blanchet@39731
|
181 |
fun extract_tstplike_proof delims output =
|
blanchet@39731
|
182 |
case pairself (find_first (fn s => String.isSubstring s output))
|
blanchet@39731
|
183 |
(ListPair.unzip delims) of
|
blanchet@39731
|
184 |
(SOME begin_delim, SOME end_delim) =>
|
blanchet@39731
|
185 |
extract_delimited (begin_delim, end_delim) output
|
blanchet@39731
|
186 |
| _ => ""
|
blanchet@39731
|
187 |
|
blanchet@39731
|
188 |
fun extract_known_failure known_failures output =
|
blanchet@39731
|
189 |
known_failures
|
blanchet@39731
|
190 |
|> find_first (fn (_, pattern) => String.isSubstring pattern output)
|
blanchet@39731
|
191 |
|> Option.map fst
|
blanchet@39731
|
192 |
|
blanchet@44344
|
193 |
fun extract_tstplike_proof_and_outcome verbose complete proof_delims
|
blanchet@43717
|
194 |
known_failures output =
|
blanchet@43751
|
195 |
case (extract_tstplike_proof proof_delims output,
|
blanchet@43751
|
196 |
extract_known_failure known_failures output) of
|
blanchet@43751
|
197 |
(_, SOME ProofIncomplete) => ("", SOME ProofIncomplete)
|
blanchet@44087
|
198 |
| ("", SOME ProofMissing) => ("", NONE)
|
blanchet@43751
|
199 |
| ("", SOME failure) =>
|
blanchet@43891
|
200 |
("", SOME (if failure = GaveUp andalso complete then Unprovable
|
blanchet@43751
|
201 |
else failure))
|
blanchet@44087
|
202 |
| ("", NONE) => ("", SOME (UnknownError (short_output verbose output)))
|
blanchet@43751
|
203 |
| (tstplike_proof, _) => (tstplike_proof, NONE)
|
blanchet@39692
|
204 |
|
blanchet@44352
|
205 |
type step_name = string * string list option
|
blanchet@39692
|
206 |
|
blanchet@43809
|
207 |
fun is_same_atp_step (s1, _) (s2, _) = s1 = s2
|
blanchet@43809
|
208 |
|
blanchet@43809
|
209 |
val vampire_fact_prefix = "f"
|
blanchet@39692
|
210 |
|
blanchet@39692
|
211 |
fun step_name_ord p =
|
blanchet@39695
|
212 |
let val q = pairself fst p in
|
blanchet@39692
|
213 |
(* The "unprefix" part is to cope with remote Vampire's output. The proper
|
blanchet@39692
|
214 |
solution would be to perform a topological sort, e.g. using the nice
|
blanchet@39692
|
215 |
"Graph" functor. *)
|
blanchet@43809
|
216 |
case pairself (Int.fromString
|
blanchet@43809
|
217 |
o perhaps (try (unprefix vampire_fact_prefix))) q of
|
blanchet@39692
|
218 |
(NONE, NONE) => string_ord q
|
blanchet@39692
|
219 |
| (NONE, SOME _) => LESS
|
blanchet@39692
|
220 |
| (SOME _, NONE) => GREATER
|
blanchet@39692
|
221 |
| (SOME i, SOME j) => int_ord (i, j)
|
blanchet@39692
|
222 |
end
|
blanchet@39692
|
223 |
|
blanchet@39693
|
224 |
datatype 'a step =
|
blanchet@39693
|
225 |
Definition of step_name * 'a * 'a |
|
blanchet@39693
|
226 |
Inference of step_name * 'a * step_name list
|
blanchet@39692
|
227 |
|
nik@44537
|
228 |
type 'a proof = ('a, 'a, ('a, 'a) ho_term) formula step list
|
blanchet@39692
|
229 |
|
blanchet@39692
|
230 |
fun step_name (Definition (name, _, _)) = name
|
blanchet@39692
|
231 |
| step_name (Inference (name, _, _)) = name
|
blanchet@39692
|
232 |
|
blanchet@39692
|
233 |
(**** PARSING OF TSTP FORMAT ****)
|
blanchet@39692
|
234 |
|
blanchet@43814
|
235 |
(* FIXME: temporary hack *)
|
blanchet@43814
|
236 |
fun repair_waldmeister_step_name s =
|
blanchet@43814
|
237 |
case space_explode "." s of
|
blanchet@43814
|
238 |
[a, b, c, d] =>
|
blanchet@43814
|
239 |
(case a of "0" => "X" | "1" => "Y" | _ => "Z" ^ a) ^
|
blanchet@43814
|
240 |
(if size b = 1 then "0" else "") ^ b ^ c ^ d
|
blanchet@43814
|
241 |
| _ => s
|
blanchet@43814
|
242 |
|
blanchet@43407
|
243 |
(* Strings enclosed in single quotes (e.g., file names) *)
|
blanchet@39692
|
244 |
val scan_general_id =
|
blanchet@43814
|
245 |
$$ "'" |-- Scan.repeat (~$$ "'") --| $$ "'"
|
blanchet@43814
|
246 |
>> implode >> repair_waldmeister_step_name
|
blanchet@39692
|
247 |
|| Scan.repeat ($$ "$") -- Scan.many1 Symbol.is_letdig
|
blanchet@39692
|
248 |
>> (fn (ss1, ss2) => implode ss1 ^ implode ss2)
|
blanchet@39692
|
249 |
|
blanchet@46079
|
250 |
val dummy_phi = AAtom (ATerm ("", []))
|
blanchet@46079
|
251 |
|
blanchet@46079
|
252 |
fun skip_formula ss =
|
blanchet@46079
|
253 |
let
|
blanchet@46079
|
254 |
fun skip _ [] = []
|
blanchet@46079
|
255 |
| skip 0 (ss as "," :: _) = ss
|
blanchet@46079
|
256 |
| skip 0 (ss as ")" :: _) = ss
|
blanchet@46079
|
257 |
| skip 0 (ss as "]" :: _) = ss
|
blanchet@46079
|
258 |
| skip n ("(" :: ss) = skip (n + 1) ss
|
blanchet@46079
|
259 |
| skip n ("[" :: ss) = skip (n + 1) ss
|
blanchet@46079
|
260 |
| skip n ("]" :: ss) = skip (n - 1) ss
|
blanchet@46079
|
261 |
| skip n (")" :: ss) = skip (n - 1) ss
|
blanchet@46079
|
262 |
| skip n (_ :: ss) = skip n ss
|
blanchet@46079
|
263 |
in (dummy_phi, skip 0 ss) end
|
blanchet@46079
|
264 |
|
blanchet@46079
|
265 |
datatype source =
|
blanchet@46079
|
266 |
File_Source of string * string option |
|
blanchet@46079
|
267 |
Inference_Source of string list
|
blanchet@46079
|
268 |
|
blanchet@46079
|
269 |
fun parse_dependencies x =
|
blanchet@46079
|
270 |
(scan_general_id ::: Scan.repeat ($$ "," |-- scan_general_id)) x
|
blanchet@46079
|
271 |
|
blanchet@46079
|
272 |
fun parse_source x =
|
blanchet@46079
|
273 |
(Scan.this_string "file" |-- $$ "(" |-- scan_general_id --
|
blanchet@46079
|
274 |
Scan.option ($$ "," |-- scan_general_id) --| $$ ")"
|
blanchet@46079
|
275 |
>> File_Source
|
blanchet@46079
|
276 |
|| (Scan.this_string "inference" |-- $$ "(" |-- scan_general_id
|
blanchet@46079
|
277 |
--| skip_formula --| $$ ",")
|
blanchet@46079
|
278 |
::: (skip_formula |-- $$ "," |-- $$ "[" |-- parse_dependencies --| $$ "]"
|
blanchet@46079
|
279 |
--| $$ ")")
|
blanchet@46079
|
280 |
>> Inference_Source) x
|
blanchet@39692
|
281 |
|
blanchet@43807
|
282 |
fun list_app (f, args) =
|
blanchet@43809
|
283 |
fold (fn arg => fn f => ATerm (tptp_app, [f, arg])) args f
|
blanchet@43807
|
284 |
|
blanchet@43809
|
285 |
(* We ignore TFF and THF types for now. *)
|
blanchet@43809
|
286 |
fun parse_type_stuff x =
|
blanchet@43809
|
287 |
Scan.repeat (($$ tptp_has_type || $$ tptp_fun_type) |-- parse_arg) x
|
blanchet@43809
|
288 |
and parse_arg x =
|
blanchet@43809
|
289 |
($$ "(" |-- parse_term --| $$ ")" --| parse_type_stuff
|
blanchet@43809
|
290 |
|| scan_general_id --| parse_type_stuff
|
blanchet@43809
|
291 |
-- Scan.optional ($$ "(" |-- parse_terms --| $$ ")") []
|
blanchet@43809
|
292 |
>> ATerm) x
|
blanchet@43807
|
293 |
and parse_app x =
|
blanchet@43809
|
294 |
(parse_arg -- Scan.repeat ($$ tptp_app |-- parse_arg) >> list_app) x
|
blanchet@43807
|
295 |
and parse_term x =
|
blanchet@43809
|
296 |
(parse_app -- Scan.option (Scan.option ($$ tptp_not_infix) --| $$ tptp_equal
|
blanchet@43809
|
297 |
-- parse_app)
|
blanchet@43809
|
298 |
>> (fn (u1, NONE) => u1
|
blanchet@43809
|
299 |
| (u1, SOME (NONE, u2)) => ATerm ("equal", [u1, u2])
|
blanchet@43809
|
300 |
| (u1, SOME (SOME _, u2)) =>
|
blanchet@43809
|
301 |
ATerm (tptp_not, [ATerm ("equal", [u1, u2])]))) x
|
blanchet@43807
|
302 |
and parse_terms x =
|
blanchet@43807
|
303 |
(parse_term ::: Scan.repeat ($$ "," |-- parse_term)) x
|
blanchet@39692
|
304 |
|
blanchet@43809
|
305 |
(* TODO: Avoid duplication with "parse_term" above. *)
|
blanchet@39822
|
306 |
fun parse_atom x =
|
blanchet@43809
|
307 |
(parse_term -- Scan.option (Scan.option ($$ tptp_not_infix) --| $$ tptp_equal
|
blanchet@43809
|
308 |
-- parse_term)
|
blanchet@39822
|
309 |
>> (fn (u1, NONE) => AAtom u1
|
blanchet@43784
|
310 |
| (u1, SOME (NONE, u2)) => AAtom (ATerm ("equal", [u1, u2]))
|
blanchet@39822
|
311 |
| (u1, SOME (SOME _, u2)) =>
|
blanchet@43784
|
312 |
mk_anot (AAtom (ATerm ("equal", [u1, u2]))))) x
|
blanchet@39692
|
313 |
|
nik@44537
|
314 |
fun ho_term_head (ATerm (s, _)) = s
|
blanchet@39692
|
315 |
|
blanchet@39692
|
316 |
(* TPTP formulas are fully parenthesized, so we don't need to worry about
|
blanchet@39692
|
317 |
operator precedence. *)
|
blanchet@43476
|
318 |
fun parse_literal x =
|
blanchet@43809
|
319 |
((Scan.repeat ($$ tptp_not) >> length)
|
blanchet@43476
|
320 |
-- ($$ "(" |-- parse_formula --| $$ ")"
|
blanchet@43476
|
321 |
|| parse_quantified_formula
|
blanchet@43476
|
322 |
|| parse_atom)
|
blanchet@43476
|
323 |
>> (fn (n, phi) => phi |> n mod 2 = 1 ? mk_anot)) x
|
blanchet@43476
|
324 |
and parse_formula x =
|
blanchet@43476
|
325 |
(parse_literal
|
blanchet@44004
|
326 |
-- Scan.option ((Scan.this_string tptp_implies
|
blanchet@44004
|
327 |
|| Scan.this_string tptp_iff
|
blanchet@44004
|
328 |
|| Scan.this_string tptp_not_iff
|
blanchet@44004
|
329 |
|| Scan.this_string tptp_if
|
blanchet@44004
|
330 |
|| $$ tptp_or
|
blanchet@44004
|
331 |
|| $$ tptp_and) -- parse_formula)
|
blanchet@39692
|
332 |
>> (fn (phi1, NONE) => phi1
|
blanchet@44004
|
333 |
| (phi1, SOME (c, phi2)) =>
|
blanchet@44004
|
334 |
if c = tptp_implies then mk_aconn AImplies phi1 phi2
|
blanchet@44004
|
335 |
else if c = tptp_iff then mk_aconn AIff phi1 phi2
|
blanchet@44004
|
336 |
else if c = tptp_not_iff then mk_anot (mk_aconn AIff phi1 phi2)
|
blanchet@44004
|
337 |
else if c = tptp_if then mk_aconn AImplies phi2 phi1
|
blanchet@44004
|
338 |
else if c = tptp_or then mk_aconn AOr phi1 phi2
|
blanchet@44004
|
339 |
else if c = tptp_and then mk_aconn AAnd phi1 phi2
|
blanchet@44004
|
340 |
else raise Fail ("impossible connective " ^ quote c))) x
|
blanchet@43476
|
341 |
and parse_quantified_formula x =
|
blanchet@43809
|
342 |
(($$ tptp_forall >> K AForall || $$ tptp_exists >> K AExists)
|
blanchet@43476
|
343 |
--| $$ "[" -- parse_terms --| $$ "]" --| $$ ":" -- parse_literal
|
blanchet@43476
|
344 |
>> (fn ((q, ts), phi) =>
|
blanchet@43807
|
345 |
(* We ignore TFF and THF types for now. *)
|
nik@44537
|
346 |
AQuant (q, map (rpair NONE o ho_term_head) ts, phi))) x
|
blanchet@39692
|
347 |
|
blanchet@39692
|
348 |
val parse_tstp_extra_arguments =
|
blanchet@46079
|
349 |
Scan.optional ($$ "," |-- parse_source
|
blanchet@46079
|
350 |
--| Scan.option ($$ "," |-- skip_formula))
|
blanchet@46079
|
351 |
(Inference_Source [])
|
blanchet@39692
|
352 |
|
blanchet@43784
|
353 |
val waldmeister_conjecture = "conjecture_1"
|
blanchet@43784
|
354 |
|
blanchet@43407
|
355 |
val tofof_fact_prefix = "fof_"
|
blanchet@41451
|
356 |
|
blanchet@43784
|
357 |
fun is_same_term subst tm1 tm2 =
|
blanchet@43784
|
358 |
let
|
blanchet@43784
|
359 |
fun do_term_pair _ NONE = NONE
|
blanchet@43784
|
360 |
| do_term_pair (ATerm (s1, tm1), ATerm (s2, tm2)) (SOME subst) =
|
blanchet@43839
|
361 |
case pairself is_tptp_variable (s1, s2) of
|
blanchet@43784
|
362 |
(true, true) =>
|
blanchet@43784
|
363 |
(case AList.lookup (op =) subst s1 of
|
blanchet@43784
|
364 |
SOME s2' => if s2' = s2 then SOME subst else NONE
|
blanchet@43784
|
365 |
| NONE =>
|
blanchet@43784
|
366 |
if null (AList.find (op =) subst s2) then SOME ((s1, s2) :: subst)
|
blanchet@43784
|
367 |
else NONE)
|
blanchet@43784
|
368 |
| (false, false) =>
|
blanchet@43784
|
369 |
if s1 = s2 andalso length tm1 = length tm2 then
|
blanchet@43784
|
370 |
SOME subst |> fold do_term_pair (tm1 ~~ tm2)
|
blanchet@43784
|
371 |
else
|
blanchet@43784
|
372 |
NONE
|
blanchet@43784
|
373 |
| _ => NONE
|
blanchet@43784
|
374 |
in SOME subst |> do_term_pair (tm1, tm2) |> is_some end
|
blanchet@43784
|
375 |
|
blanchet@43784
|
376 |
fun is_same_formula subst (AQuant (q1, xs1, phi1)) (AQuant (q2, xs2, phi2)) =
|
blanchet@43784
|
377 |
q1 = q2 andalso length xs1 = length xs2 andalso
|
blanchet@43784
|
378 |
is_same_formula ((map fst xs1 ~~ map fst xs2) @ subst) phi1 phi2
|
blanchet@43784
|
379 |
| is_same_formula subst (AConn (c1, phis1)) (AConn (c2, phis2)) =
|
blanchet@43784
|
380 |
c1 = c2 andalso length phis1 = length phis2 andalso
|
blanchet@43784
|
381 |
forall (uncurry (is_same_formula subst)) (phis1 ~~ phis2)
|
blanchet@43784
|
382 |
| is_same_formula subst (AAtom (ATerm ("equal", [tm11, tm12]))) (AAtom tm2) =
|
blanchet@43784
|
383 |
is_same_term subst (ATerm ("equal", [tm11, tm12])) tm2 orelse
|
blanchet@43784
|
384 |
is_same_term subst (ATerm ("equal", [tm12, tm11])) tm2
|
blanchet@43784
|
385 |
| is_same_formula subst (AAtom tm1) (AAtom tm2) = is_same_term subst tm1 tm2
|
blanchet@43784
|
386 |
| is_same_formula _ _ _ = false
|
blanchet@43784
|
387 |
|
blanchet@43784
|
388 |
fun matching_formula_line_identifier phi (Formula (ident, _, phi', _, _)) =
|
blanchet@43784
|
389 |
if is_same_formula [] phi phi' then SOME ident else NONE
|
blanchet@43784
|
390 |
| matching_formula_line_identifier _ _ = NONE
|
blanchet@43784
|
391 |
|
blanchet@43784
|
392 |
fun find_formula_in_problem problem phi =
|
blanchet@43784
|
393 |
problem |> maps snd |> map_filter (matching_formula_line_identifier phi)
|
blanchet@44352
|
394 |
|> try (single o hd)
|
blanchet@43784
|
395 |
|
blanchet@43803
|
396 |
(* Syntax: (cnf|fof|tff|thf)\(<num>, <formula_role>,
|
blanchet@43803
|
397 |
<formula> <extra_arguments>\).
|
blanchet@39692
|
398 |
The <num> could be an identifier, but we assume integers. *)
|
blanchet@43784
|
399 |
fun parse_tstp_line problem =
|
blanchet@43809
|
400 |
((Scan.this_string tptp_cnf || Scan.this_string tptp_fof
|
blanchet@43809
|
401 |
|| Scan.this_string tptp_tff || Scan.this_string tptp_thf) -- $$ "(")
|
blanchet@43784
|
402 |
|-- scan_general_id --| $$ "," -- Symbol.scan_id --| $$ ","
|
blanchet@43809
|
403 |
-- (parse_formula || skip_formula) -- parse_tstp_extra_arguments --| $$ ")"
|
blanchet@43809
|
404 |
--| $$ "."
|
blanchet@43784
|
405 |
>> (fn (((num, role), phi), deps) =>
|
blanchet@43784
|
406 |
let
|
blanchet@43784
|
407 |
val (name, deps) =
|
blanchet@43784
|
408 |
(* Waldmeister isn't exactly helping. *)
|
blanchet@43784
|
409 |
case deps of
|
blanchet@46079
|
410 |
File_Source (_, SOME s) =>
|
blanchet@43784
|
411 |
((num,
|
blanchet@45276
|
412 |
if s = waldmeister_conjecture then
|
blanchet@43784
|
413 |
find_formula_in_problem problem (mk_anot phi)
|
blanchet@43784
|
414 |
else
|
blanchet@44352
|
415 |
SOME [s |> perhaps (try (unprefix tofof_fact_prefix))]),
|
blanchet@43784
|
416 |
[])
|
blanchet@46079
|
417 |
| File_Source _ =>
|
blanchet@46079
|
418 |
((num, find_formula_in_problem problem phi), [])
|
blanchet@46079
|
419 |
| Inference_Source deps => ((num, NONE), deps)
|
blanchet@43784
|
420 |
in
|
blanchet@43784
|
421 |
case role of
|
blanchet@43784
|
422 |
"definition" =>
|
blanchet@43784
|
423 |
(case phi of
|
blanchet@43784
|
424 |
AConn (AIff, [phi1 as AAtom _, phi2]) =>
|
blanchet@43784
|
425 |
Definition (name, phi1, phi2)
|
blanchet@43784
|
426 |
| AAtom (ATerm ("equal", _)) =>
|
blanchet@43784
|
427 |
(* Vampire's equality proxy axiom *)
|
blanchet@43784
|
428 |
Inference (name, phi, map (rpair NONE) deps)
|
blanchet@43809
|
429 |
| _ => raise UNRECOGNIZED_ATP_PROOF ())
|
blanchet@43784
|
430 |
| _ => Inference (name, phi, map (rpair NONE) deps)
|
blanchet@43784
|
431 |
end)
|
blanchet@39692
|
432 |
|
blanchet@39692
|
433 |
(**** PARSING OF SPASS OUTPUT ****)
|
blanchet@39692
|
434 |
|
blanchet@39692
|
435 |
(* SPASS returns clause references of the form "x.y". We ignore "y", whose role
|
blanchet@39692
|
436 |
is not clear anyway. *)
|
blanchet@39692
|
437 |
val parse_dot_name = scan_general_id --| $$ "." --| scan_general_id
|
blanchet@39692
|
438 |
|
blanchet@39692
|
439 |
val parse_spass_annotations =
|
blanchet@39692
|
440 |
Scan.optional ($$ ":" |-- Scan.repeat (parse_dot_name
|
blanchet@39692
|
441 |
--| Scan.option ($$ ","))) []
|
blanchet@39692
|
442 |
|
blanchet@39692
|
443 |
(* It is not clear why some literals are followed by sequences of stars and/or
|
blanchet@39692
|
444 |
pluses. We ignore them. *)
|
blanchet@39826
|
445 |
fun parse_decorated_atom x =
|
blanchet@39826
|
446 |
(parse_atom --| Scan.repeat ($$ "*" || $$ "+" || $$ " ")) x
|
blanchet@39692
|
447 |
|
blanchet@39692
|
448 |
fun mk_horn ([], []) = AAtom (ATerm ("c_False", []))
|
blanchet@43784
|
449 |
| mk_horn ([], pos_lits) = foldr1 (uncurry (mk_aconn AOr)) pos_lits
|
blanchet@43784
|
450 |
| mk_horn (neg_lits, []) = mk_anot (foldr1 (uncurry (mk_aconn AAnd)) neg_lits)
|
blanchet@39692
|
451 |
| mk_horn (neg_lits, pos_lits) =
|
blanchet@43784
|
452 |
mk_aconn AImplies (foldr1 (uncurry (mk_aconn AAnd)) neg_lits)
|
blanchet@43784
|
453 |
(foldr1 (uncurry (mk_aconn AOr)) pos_lits)
|
blanchet@39692
|
454 |
|
blanchet@39869
|
455 |
fun parse_horn_clause x =
|
blanchet@39869
|
456 |
(Scan.repeat parse_decorated_atom --| $$ "|" --| $$ "|"
|
blanchet@39869
|
457 |
-- Scan.repeat parse_decorated_atom --| $$ "-" --| $$ ">"
|
blanchet@39869
|
458 |
-- Scan.repeat parse_decorated_atom
|
blanchet@39869
|
459 |
>> (mk_horn o apfst (op @))) x
|
blanchet@39692
|
460 |
|
blanchet@44352
|
461 |
fun resolve_spass_num spass_names num =
|
blanchet@44352
|
462 |
case Int.fromString num of
|
blanchet@44352
|
463 |
SOME j => if j > 0 andalso j <= Vector.length spass_names then
|
blanchet@44352
|
464 |
SOME (Vector.sub (spass_names, j - 1))
|
blanchet@44352
|
465 |
else
|
blanchet@44352
|
466 |
NONE
|
blanchet@44352
|
467 |
| NONE => NONE
|
blanchet@44352
|
468 |
|
blanchet@46033
|
469 |
(* Syntax: <num>[0:<inference><annotations>] <atoms> || <atoms> -> <atoms>. *)
|
blanchet@46033
|
470 |
fun parse_spass_line spass_names =
|
blanchet@46033
|
471 |
scan_general_id --| $$ "[" --| $$ "0" --| $$ ":" --| Symbol.scan_id
|
blanchet@46033
|
472 |
-- parse_spass_annotations --| $$ "]" -- parse_horn_clause --| $$ "."
|
blanchet@46033
|
473 |
>> (fn ((num, deps), u) =>
|
blanchet@46033
|
474 |
Inference ((num, resolve_spass_num spass_names num), u,
|
blanchet@46033
|
475 |
map (swap o `(resolve_spass_num spass_names)) deps))
|
blanchet@46033
|
476 |
|
blanchet@46033
|
477 |
(* Syntax: <name> *)
|
blanchet@46071
|
478 |
fun parse_satallax_line x =
|
blanchet@46071
|
479 |
(scan_general_id --| Scan.option ($$ " ")
|
blanchet@46071
|
480 |
>> (fn s => Inference ((s, SOME [s]), dummy_phi, []))) x
|
blanchet@39692
|
481 |
|
blanchet@44352
|
482 |
fun parse_line problem spass_names =
|
blanchet@46071
|
483 |
parse_tstp_line problem || parse_spass_line spass_names || parse_satallax_line
|
blanchet@44352
|
484 |
fun parse_proof problem spass_names tstp =
|
blanchet@45652
|
485 |
tstp |> strip_spaces_except_between_idents
|
blanchet@44352
|
486 |
|> raw_explode
|
blanchet@44352
|
487 |
|> Scan.finite Symbol.stopper
|
blanchet@44352
|
488 |
(Scan.error (!! (fn _ => raise UNRECOGNIZED_ATP_PROOF ())
|
blanchet@44352
|
489 |
(Scan.repeat1 (parse_line problem spass_names))))
|
blanchet@44352
|
490 |
|> fst
|
blanchet@39692
|
491 |
|
blanchet@44352
|
492 |
(** SPASS's FLOTTER hack **)
|
blanchet@44352
|
493 |
|
blanchet@44352
|
494 |
(* This is a hack required for keeping track of facts after they have been
|
blanchet@44352
|
495 |
clausified by SPASS's FLOTTER preprocessor. The "ATP/scripts/spass" script is
|
blanchet@44352
|
496 |
also part of this hack. *)
|
blanchet@44352
|
497 |
|
blanchet@44352
|
498 |
val set_ClauseFormulaRelationN = "set_ClauseFormulaRelation"
|
blanchet@44352
|
499 |
|
blanchet@44352
|
500 |
fun extract_clause_sequence output =
|
blanchet@44352
|
501 |
let
|
blanchet@44352
|
502 |
val tokens_of = String.tokens (not o Char.isAlphaNum)
|
blanchet@44352
|
503 |
fun extract_num ("clause" :: (ss as _ :: _)) = Int.fromString (List.last ss)
|
blanchet@44352
|
504 |
| extract_num _ = NONE
|
blanchet@44352
|
505 |
in output |> split_lines |> map_filter (extract_num o tokens_of) end
|
blanchet@44352
|
506 |
|
blanchet@44352
|
507 |
fun is_head_digit s = Char.isDigit (String.sub (s, 0))
|
blanchet@44352
|
508 |
val scan_integer = Scan.many1 is_head_digit >> (the o Int.fromString o implode)
|
blanchet@44352
|
509 |
|
blanchet@44352
|
510 |
val parse_clause_formula_pair =
|
blanchet@44352
|
511 |
$$ "(" |-- scan_integer --| $$ ","
|
blanchet@44352
|
512 |
-- (Symbol.scan_id ::: Scan.repeat ($$ "," |-- Symbol.scan_id)) --| $$ ")"
|
blanchet@44352
|
513 |
--| Scan.option ($$ ",")
|
blanchet@44352
|
514 |
val parse_clause_formula_relation =
|
blanchet@44352
|
515 |
Scan.this_string set_ClauseFormulaRelationN |-- $$ "("
|
blanchet@44352
|
516 |
|-- Scan.repeat parse_clause_formula_pair
|
blanchet@44352
|
517 |
val extract_clause_formula_relation =
|
blanchet@44352
|
518 |
Substring.full #> Substring.position set_ClauseFormulaRelationN
|
blanchet@44352
|
519 |
#> snd #> Substring.position "." #> fst #> Substring.string
|
blanchet@44352
|
520 |
#> raw_explode #> filter_out Symbol.is_blank #> parse_clause_formula_relation
|
blanchet@44352
|
521 |
#> fst
|
blanchet@44352
|
522 |
|
blanchet@44352
|
523 |
fun extract_spass_name_vector output =
|
blanchet@44352
|
524 |
(if String.isSubstring set_ClauseFormulaRelationN output then
|
blanchet@44352
|
525 |
let
|
blanchet@44352
|
526 |
val num_seq = extract_clause_sequence output
|
blanchet@44352
|
527 |
val name_map = extract_clause_formula_relation output
|
blanchet@44352
|
528 |
val name_seq = num_seq |> map (these o AList.lookup (op =) name_map)
|
blanchet@44352
|
529 |
in name_seq end
|
blanchet@44352
|
530 |
else
|
blanchet@44352
|
531 |
[])
|
blanchet@44352
|
532 |
|> Vector.fromList
|
blanchet@44352
|
533 |
|
blanchet@44352
|
534 |
fun atp_proof_from_tstplike_proof _ _ "" = []
|
blanchet@44352
|
535 |
| atp_proof_from_tstplike_proof problem output tstp =
|
blanchet@44352
|
536 |
tstp ^ "$" (* the $ sign acts as a sentinel (FIXME: needed?) *)
|
blanchet@44352
|
537 |
|> parse_proof problem (extract_spass_name_vector output)
|
blanchet@43320
|
538 |
|> sort (step_name_ord o pairself step_name)
|
blanchet@43809
|
539 |
|
blanchet@43809
|
540 |
fun clean_up_dependencies _ [] = []
|
blanchet@43809
|
541 |
| clean_up_dependencies seen ((step as Definition (name, _, _)) :: steps) =
|
blanchet@43809
|
542 |
step :: clean_up_dependencies (name :: seen) steps
|
blanchet@43809
|
543 |
| clean_up_dependencies seen (Inference (name, u, deps) :: steps) =
|
blanchet@43809
|
544 |
Inference (name, u,
|
blanchet@43809
|
545 |
map_filter (fn dep => find_first (is_same_atp_step dep) seen)
|
blanchet@43809
|
546 |
deps) ::
|
blanchet@43809
|
547 |
clean_up_dependencies (name :: seen) steps
|
blanchet@43809
|
548 |
|
blanchet@43816
|
549 |
fun clean_up_atp_proof_dependencies proof = clean_up_dependencies [] proof
|
blanchet@39692
|
550 |
|
blanchet@39694
|
551 |
fun map_term_names_in_term f (ATerm (s, ts)) =
|
blanchet@39694
|
552 |
ATerm (f s, map (map_term_names_in_term f) ts)
|
blanchet@39694
|
553 |
fun map_term_names_in_formula f (AQuant (q, xs, phi)) =
|
blanchet@39694
|
554 |
AQuant (q, xs, map_term_names_in_formula f phi)
|
blanchet@39694
|
555 |
| map_term_names_in_formula f (AConn (c, phis)) =
|
blanchet@39694
|
556 |
AConn (c, map (map_term_names_in_formula f) phis)
|
blanchet@39694
|
557 |
| map_term_names_in_formula f (AAtom t) = AAtom (map_term_names_in_term f t)
|
blanchet@39694
|
558 |
fun map_term_names_in_step f (Definition (name, phi1, phi2)) =
|
blanchet@39694
|
559 |
Definition (name, map_term_names_in_formula f phi1,
|
blanchet@39694
|
560 |
map_term_names_in_formula f phi2)
|
blanchet@39694
|
561 |
| map_term_names_in_step f (Inference (name, phi, deps)) =
|
blanchet@39694
|
562 |
Inference (name, map_term_names_in_formula f phi, deps)
|
blanchet@39694
|
563 |
fun map_term_names_in_atp_proof f = map (map_term_names_in_step f)
|
blanchet@39694
|
564 |
|
blanchet@39694
|
565 |
fun nasty_name pool s = s |> Symtab.lookup pool |> the_default s
|
blanchet@39694
|
566 |
fun nasty_atp_proof pool =
|
blanchet@39694
|
567 |
if Symtab.is_empty pool then I
|
blanchet@39694
|
568 |
else map_term_names_in_atp_proof (nasty_name pool)
|
blanchet@39694
|
569 |
|
blanchet@39692
|
570 |
end;
|